external/vulkancts/modules/vulkan/spirv_assembly/vktSpvAsmInstructionTests.cpp

   1 /*-------------------------------------------------------------------------
   2  * Vulkan Conformance Tests
   3  * ------------------------
   4  *
   5  * Copyright (c) 2015 Google Inc.
   6  * Copyright (c) 2016 The Khronos Group Inc.
   7  *
   8  * Licensed under the Apache License, Version 2.0 (the "License");
   9  * you may not use this file except in compliance with the License.
  10  * You may obtain a copy of the License at
  11  *
  12  *      http://www.apache.org/licenses/LICENSE-2.0
  13  *
  14  * Unless required by applicable law or agreed to in writing, software
  15  * distributed under the License is distributed on an "AS IS" BASIS,
  16  * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
  17  * See the License for the specific language governing permissions and
  18  * limitations under the License.
  19  *
  20  *//*!
  21  * \file
  22  * \brief SPIR-V Assembly Tests for Instructions (special opcode/operand)
  23  *//*--------------------------------------------------------------------*/
  24
  25 #include "vktSpvAsmInstructionTests.hpp"
  26
  27 #include "tcuCommandLine.hpp"
  28 #include "tcuFormatUtil.hpp"
  29 #include "tcuFloat.hpp"
  30 #include "tcuFloatFormat.hpp"
  31 #include "tcuRGBA.hpp"
  32 #include "tcuStringTemplate.hpp"
  33 #include "tcuTestLog.hpp"
  34 #include "tcuVectorUtil.hpp"
  35 #include "tcuInterval.hpp"
  36
  37 #include "vkDefs.hpp"
  38 #include "vkDeviceUtil.hpp"
  39 #include "vkMemUtil.hpp"
  40 #include "vkPlatform.hpp"
  41 #include "vkPrograms.hpp"
  42 #include "vkQueryUtil.hpp"
  43 #include "vkRef.hpp"
  44 #include "vkRefUtil.hpp"
  45 #include "vkStrUtil.hpp"
  46 #include "vkTypeUtil.hpp"
  47
  48 #include "deStringUtil.hpp"
  49 #include "deUniquePtr.hpp"
  50 #include "deMath.h"
  51 #include "tcuStringTemplate.hpp"
  52
  53 #include "vktSpvAsmCrossStageInterfaceTests.hpp"
  54 #include "vktSpvAsm8bitStorageTests.hpp"
  55 #include "vktSpvAsm16bitStorageTests.hpp"
  56 #include "vktSpvAsmUboMatrixPaddingTests.hpp"
  57 #include "vktSpvAsmConditionalBranchTests.hpp"
  58 #include "vktSpvAsmIndexingTests.hpp"
  59 #include "vktSpvAsmImageSamplerTests.hpp"
  60 #include "vktSpvAsmComputeShaderCase.hpp"
  61 #include "vktSpvAsmComputeShaderTestUtil.hpp"
  62 #include "vktSpvAsmFloatControlsTests.hpp"
  63 #include "vktSpvAsmGraphicsShaderTestUtil.hpp"
  64 #include "vktSpvAsmVariablePointersTests.hpp"
  65 #include "vktSpvAsmVariableInitTests.hpp"
  66 #include "vktSpvAsmPointerParameterTests.hpp"
  67 #include "vktSpvAsmSpirvVersionTests.hpp"
  68 #include "vktTestCaseUtil.hpp"
  69 #include "vktSpvAsmLoopDepLenTests.hpp"
  70 #include "vktSpvAsmLoopDepInfTests.hpp"
  71 #include "vktSpvAsmCompositeInsertTests.hpp"
  72 #include "vktSpvAsmVaryingNameTests.hpp"
  73 #include "vktSpvAsmWorkgroupMemoryTests.hpp"
  74 #include "vktSpvAsmSignedIntCompareTests.hpp"
  75
  76 #include <cmath>
  77 #include <limits>
  78 #include <map>
  79 #include <string>
  80 #include <sstream>
  81 #include <utility>
  82 #include <stack>
  83
  84 namespace vkt
  85 {
  86 namespace SpirVAssembly
  87 {
  88
  89 namespace
  90 {
  91
  92 using namespace vk;
  93 using std::map;
  94 using std::string;
  95 using std::vector;
  96 using tcu::IVec3;
  97 using tcu::IVec4;
  98 using tcu::RGBA;
  99 using tcu::TestLog;
 100 using tcu::TestStatus;
 101 using tcu::Vec4;
 102 using de::UniquePtr;
 103 using tcu::StringTemplate;
 104 using tcu::Vec4;
 105
 106 const bool TEST_WITH_NAN        = true;
 107 const bool TEST_WITHOUT_NAN     = false;
 108
 109 template<typename T>
 110 static void fillRandomScalars (de::Random& rnd, T minValue, T maxValue, void* dst, int numValues, int offset = 0)
 111 {
 112         T* const typedPtr = (T*)dst;
 113         for (int ndx = 0; ndx < numValues; ndx++)
 114                 typedPtr[offset + ndx] = randomScalar<T>(rnd, minValue, maxValue);
 115 }
 116
 117 // Filter is a function that returns true if a value should pass, false otherwise.
 118 template<typename T, typename FilterT>
 119 static void fillRandomScalars (de::Random& rnd, T minValue, T maxValue, void* dst, int numValues, FilterT filter, int offset = 0)
 120 {
 121         T* const typedPtr = (T*)dst;
 122         T value;
 123         for (int ndx = 0; ndx < numValues; ndx++)
 124         {
 125                 do
 126                         value = randomScalar<T>(rnd, minValue, maxValue);
 127                 while (!filter(value));
 128
 129                 typedPtr[offset + ndx] = value;
 130         }
 131 }
 132
 133 // Gets a 64-bit integer with a more logarithmic distribution
 134 deInt64 randomInt64LogDistributed (de::Random& rnd)
 135 {
 136         deInt64 val = rnd.getUint64();
 137         val &= (1ull << rnd.getInt(1, 63)) - 1;
 138         if (rnd.getBool())
 139                 val = -val;
 140         return val;
 141 }
 142
 143 static void fillRandomInt64sLogDistributed (de::Random& rnd, vector<deInt64>& dst, int numValues)
 144 {
 145         for (int ndx = 0; ndx < numValues; ndx++)
 146                 dst[ndx] = randomInt64LogDistributed(rnd);
 147 }
 148
 149 template<typename FilterT>
 150 static void fillRandomInt64sLogDistributed (de::Random& rnd, vector<deInt64>& dst, int numValues, FilterT filter)
 151 {
 152         for (int ndx = 0; ndx < numValues; ndx++)
 153         {
 154                 deInt64 value;
 155                 do {
 156                         value = randomInt64LogDistributed(rnd);
 157                 } while (!filter(value));
 158                 dst[ndx] = value;
 159         }
 160 }
 161
 162 inline bool filterNonNegative (const deInt64 value)
 163 {
 164         return value >= 0;
 165 }
 166
 167 inline bool filterPositive (const deInt64 value)
 168 {
 169         return value > 0;
 170 }
 171
 172 inline bool filterNotZero (const deInt64 value)
 173 {
 174         return value != 0;
 175 }
 176
 177 static void floorAll (vector<float>& values)
 178 {
 179         for (size_t i = 0; i < values.size(); i++)
 180                 values[i] = deFloatFloor(values[i]);
 181 }
 182
 183 static void floorAll (vector<Vec4>& values)
 184 {
 185         for (size_t i = 0; i < values.size(); i++)
 186                 values[i] = floor(values[i]);
 187 }
 188
 189 struct CaseParameter
 190 {
 191         const char*             name;
 192         string                  param;
 193
 194         CaseParameter   (const char* case_, const string& param_) : name(case_), param(param_) {}
 195 };
 196
 197 // Assembly code used for testing LocalSize, OpNop, OpConstant{Null|Composite}, Op[No]Line, OpSource[Continued], OpSourceExtension, OpUndef is based on GLSL source code:
 198 //
 199 // #version 430
 200 //
 201 // layout(std140, set = 0, binding = 0) readonly buffer Input {
 202 //   float elements[];
 203 // } input_data;
 204 // layout(std140, set = 0, binding = 1) writeonly buffer Output {
 205 //   float elements[];
 206 // } output_data;
 207 //
 208 // layout (local_size_x = 1, local_size_y = 1, local_size_z = 1) in;
 209 //
 210 // void main() {
 211 //   uint x = gl_GlobalInvocationID.x;
 212 //   output_data.elements[x] = -input_data.elements[x];
 213 // }
 214
 215 static string getAsmForLocalSizeTest(bool useLiteralLocalSize, bool useSpecConstantWorkgroupSize, IVec3 workGroupSize, deUint32 ndx)
 216 {
 217         std::ostringstream out;
 218         out << getComputeAsmShaderPreambleWithoutLocalSize();
 219
 220         if (useLiteralLocalSize)
 221         {
 222                 out << "OpExecutionMode %main LocalSize "
 223                         << workGroupSize.x() << " " << workGroupSize.y() << " " << workGroupSize.z() << "\n";
 224         }
 225
 226         out << "OpSource GLSL 430\n"
 227                 "OpName %main           \"main\"\n"
 228                 "OpName %id             \"gl_GlobalInvocationID\"\n"
 229                 "OpDecorate %id BuiltIn GlobalInvocationId\n";
 230
 231         if (useSpecConstantWorkgroupSize)
 232         {
 233                 out << "OpDecorate %spec_0 SpecId 100\n"
 234                         << "OpDecorate %spec_1 SpecId 101\n"
 235                         << "OpDecorate %spec_2 SpecId 102\n"
 236                         << "OpDecorate %gl_WorkGroupSize BuiltIn WorkgroupSize\n";
 237         }
 238
 239         out << getComputeAsmInputOutputBufferTraits()
 240                 << getComputeAsmCommonTypes()
 241                 << getComputeAsmInputOutputBuffer()
 242                 << "%id        = OpVariable %uvec3ptr Input\n"
 243                 << "%zero      = OpConstant %i32 0 \n";
 244
 245         if (useSpecConstantWorkgroupSize)
 246         {
 247                 out     << "%spec_0   = OpSpecConstant %u32 "<< workGroupSize.x() << "\n"
 248                         << "%spec_1   = OpSpecConstant %u32 "<< workGroupSize.y() << "\n"
 249                         << "%spec_2   = OpSpecConstant %u32 "<< workGroupSize.z() << "\n"
 250                         << "%gl_WorkGroupSize = OpSpecConstantComposite %uvec3 %spec_0 %spec_1 %spec_2\n";
 251         }
 252
 253         out << "%main      = OpFunction %void None %voidf\n"
 254                 << "%label     = OpLabel\n"
 255                 << "%idval     = OpLoad %uvec3 %id\n"
 256                 << "%ndx         = OpCompositeExtract %u32 %idval " << ndx << "\n"
 257
 258                         "%inloc     = OpAccessChain %f32ptr %indata %zero %ndx\n"
 259                         "%inval     = OpLoad %f32 %inloc\n"
 260                         "%neg       = OpFNegate %f32 %inval\n"
 261                         "%outloc    = OpAccessChain %f32ptr %outdata %zero %ndx\n"
 262                         "             OpStore %outloc %neg\n"
 263                         "             OpReturn\n"
 264                         "             OpFunctionEnd\n";
 265         return out.str();
 266 }
 267
 268 tcu::TestCaseGroup* createLocalSizeGroup (tcu::TestContext& testCtx)
 269 {
 270         de::MovePtr<tcu::TestCaseGroup> group                   (new tcu::TestCaseGroup(testCtx, "localsize", ""));
 271         ComputeShaderSpec                               spec;
 272         de::Random                                              rnd                             (deStringHash(group->getName()));
 273         const deUint32                                  numElements             = 64u;
 274         vector<float>                                   positiveFloats  (numElements, 0);
 275         vector<float>                                   negativeFloats  (numElements, 0);
 276
 277         fillRandomScalars(rnd, 1.f, 100.f, &positiveFloats[0], numElements);
 278
 279         for (size_t ndx = 0; ndx < numElements; ++ndx)
 280                 negativeFloats[ndx] = -positiveFloats[ndx];
 281
 282         spec.inputs.push_back(BufferSp(new Float32Buffer(positiveFloats)));
 283         spec.outputs.push_back(BufferSp(new Float32Buffer(negativeFloats)));
 284
 285         spec.numWorkGroups = IVec3(numElements, 1, 1);
 286
 287         spec.assembly = getAsmForLocalSizeTest(true, false, IVec3(1, 1, 1), 0u);
 288         group->addChild(new SpvAsmComputeShaderCase(testCtx, "literal_localsize", "", spec));
 289
 290         spec.assembly = getAsmForLocalSizeTest(true, true, IVec3(1, 1, 1), 0u);
 291         group->addChild(new SpvAsmComputeShaderCase(testCtx, "literal_and_specid_localsize", "", spec));
 292
 293         spec.assembly = getAsmForLocalSizeTest(false, true, IVec3(1, 1, 1), 0u);
 294         group->addChild(new SpvAsmComputeShaderCase(testCtx, "specid_localsize", "", spec));
 295
 296         spec.numWorkGroups = IVec3(1, 1, 1);
 297
 298         spec.assembly = getAsmForLocalSizeTest(true, false, IVec3(numElements, 1, 1), 0u);
 299         group->addChild(new SpvAsmComputeShaderCase(testCtx, "literal_localsize_x", "", spec));
 300
 301         spec.assembly = getAsmForLocalSizeTest(true, true, IVec3(numElements, 1, 1), 0u);
 302         group->addChild(new SpvAsmComputeShaderCase(testCtx, "literal_and_specid_localsize_x", "", spec));
 303
 304         spec.assembly = getAsmForLocalSizeTest(false, true, IVec3(numElements, 1, 1), 0u);
 305         group->addChild(new SpvAsmComputeShaderCase(testCtx, "specid_localsize_x", "", spec));
 306
 307         spec.assembly = getAsmForLocalSizeTest(true, false, IVec3(1, numElements, 1), 1u);
 308         group->addChild(new SpvAsmComputeShaderCase(testCtx, "literal_localsize_y", "", spec));
 309
 310         spec.assembly = getAsmForLocalSizeTest(true, true, IVec3(1, numElements, 1), 1u);
 311         group->addChild(new SpvAsmComputeShaderCase(testCtx, "literal_and_specid_localsize_y", "", spec));
 312
 313         spec.assembly = getAsmForLocalSizeTest(false, true, IVec3(1, numElements, 1), 1u);
 314         group->addChild(new SpvAsmComputeShaderCase(testCtx, "specid_localsize_y", "", spec));
 315
 316         spec.assembly = getAsmForLocalSizeTest(true, false, IVec3(1, 1, numElements), 2u);
 317         group->addChild(new SpvAsmComputeShaderCase(testCtx, "literal_localsize_z", "", spec));
 318
 319         spec.assembly = getAsmForLocalSizeTest(true, true, IVec3(1, 1, numElements), 2u);
 320         group->addChild(new SpvAsmComputeShaderCase(testCtx, "literal_and_specid_localsize_z", "", spec));
 321
 322         spec.assembly = getAsmForLocalSizeTest(false, true, IVec3(1, 1, numElements), 2u);
 323         group->addChild(new SpvAsmComputeShaderCase(testCtx, "specid_localsize_z", "", spec));
 324
 325         return group.release();
 326 }
 327
 328 tcu::TestCaseGroup* createOpNopGroup (tcu::TestContext& testCtx)
 329 {
 330         de::MovePtr<tcu::TestCaseGroup> group                   (new tcu::TestCaseGroup(testCtx, "opnop", "Test the OpNop instruction"));
 331         ComputeShaderSpec                               spec;
 332         de::Random                                              rnd                             (deStringHash(group->getName()));
 333         const int                                               numElements             = 100;
 334         vector<float>                                   positiveFloats  (numElements, 0);
 335         vector<float>                                   negativeFloats  (numElements, 0);
 336
 337         fillRandomScalars(rnd, 1.f, 100.f, &positiveFloats[0], numElements);
 338
 339         for (size_t ndx = 0; ndx < numElements; ++ndx)
 340                 negativeFloats[ndx] = -positiveFloats[ndx];
 341
 342         spec.assembly =
 343                 string(getComputeAsmShaderPreamble()) +
 344
 345                 "OpSource GLSL 430\n"
 346                 "OpName %main           \"main\"\n"
 347                 "OpName %id             \"gl_GlobalInvocationID\"\n"
 348
 349                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
 350
 351                 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes())
 352
 353                 + string(getComputeAsmInputOutputBuffer()) +
 354
 355                 "%id        = OpVariable %uvec3ptr Input\n"
 356                 "%zero      = OpConstant %i32 0\n"
 357
 358                 "%main      = OpFunction %void None %voidf\n"
 359                 "%label     = OpLabel\n"
 360                 "%idval     = OpLoad %uvec3 %id\n"
 361                 "%x         = OpCompositeExtract %u32 %idval 0\n"
 362
 363                 "             OpNop\n" // Inside a function body
 364
 365                 "%inloc     = OpAccessChain %f32ptr %indata %zero %x\n"
 366                 "%inval     = OpLoad %f32 %inloc\n"
 367                 "%neg       = OpFNegate %f32 %inval\n"
 368                 "%outloc    = OpAccessChain %f32ptr %outdata %zero %x\n"
 369                 "             OpStore %outloc %neg\n"
 370                 "             OpReturn\n"
 371                 "             OpFunctionEnd\n";
 372         spec.inputs.push_back(BufferSp(new Float32Buffer(positiveFloats)));
 373         spec.outputs.push_back(BufferSp(new Float32Buffer(negativeFloats)));
 374         spec.numWorkGroups = IVec3(numElements, 1, 1);
 375
 376         group->addChild(new SpvAsmComputeShaderCase(testCtx, "all", "OpNop appearing at different places", spec));
 377
 378         return group.release();
 379 }
 380
 381 template<bool nanSupported>
 382 bool compareFUnord (const std::vector<Resource>& inputs, const vector<AllocationSp>& outputAllocs, const std::vector<Resource>& expectedOutputs, TestLog& log)
 383 {
 384         if (outputAllocs.size() != 1)
 385                 return false;
 386
 387         vector<deUint8> input1Bytes;
 388         vector<deUint8> input2Bytes;
 389         vector<deUint8> expectedBytes;
 390
 391         inputs[0].getBytes(input1Bytes);
 392         inputs[1].getBytes(input2Bytes);
 393         expectedOutputs[0].getBytes(expectedBytes);
 394
 395         const deInt32* const    expectedOutputAsInt             = reinterpret_cast<const deInt32*>(&expectedBytes.front());
 396         const deInt32* const    outputAsInt                             = static_cast<const deInt32*>(outputAllocs[0]->getHostPtr());
 397         const float* const              input1AsFloat                   = reinterpret_cast<const float*>(&input1Bytes.front());
 398         const float* const              input2AsFloat                   = reinterpret_cast<const float*>(&input2Bytes.front());
 399         bool returnValue                                                                = true;
 400
 401         for (size_t idx = 0; idx < expectedBytes.size() / sizeof(deInt32); ++idx)
 402         {
 403                 if (!nanSupported && (tcu::Float32(input1AsFloat[idx]).isNaN() || tcu::Float32(input2AsFloat[idx]).isNaN()))
 404                         continue;
 405
 406                 if (outputAsInt[idx] != expectedOutputAsInt[idx])
 407                 {
 408                         log << TestLog::Message << "ERROR: Sub-case failed. inputs: " << input1AsFloat[idx] << "," << input2AsFloat[idx] << " output: " << outputAsInt[idx]<< " expected output: " << expectedOutputAsInt[idx] << TestLog::EndMessage;
 409                         returnValue = false;
 410                 }
 411         }
 412         return returnValue;
 413 }
 414
 415 typedef VkBool32 (*compareFuncType) (float, float);
 416
 417 struct OpFUnordCase
 418 {
 419         const char*             name;
 420         const char*             opCode;
 421         compareFuncType compareFunc;
 422
 423                                         OpFUnordCase                    (const char* _name, const char* _opCode, compareFuncType _compareFunc)
 424                                                 : name                          (_name)
 425                                                 , opCode                        (_opCode)
 426                                                 , compareFunc           (_compareFunc) {}
 427 };
 428
 429 #define ADD_OPFUNORD_CASE(NAME, OPCODE, OPERATOR) \
 430 do { \
 431         struct compare_##NAME { static VkBool32 compare(float x, float y) { return (x OPERATOR y) ? VK_TRUE : VK_FALSE; } }; \
 432         cases.push_back(OpFUnordCase(#NAME, OPCODE, compare_##NAME::compare)); \
 433 } while (deGetFalse())
 434
 435 tcu::TestCaseGroup* createOpFUnordGroup (tcu::TestContext& testCtx, const bool testWithNan)
 436 {
 437         const string                                    nan                             = testWithNan ? "_nan" : "";
 438         const string                                    groupName               = "opfunord" + nan;
 439         de::MovePtr<tcu::TestCaseGroup> group                   (new tcu::TestCaseGroup(testCtx, groupName.c_str(), "Test the OpFUnord* opcodes"));
 440         de::Random                                              rnd                             (deStringHash(group->getName()));
 441         const int                                               numElements             = 100;
 442         vector<OpFUnordCase>                    cases;
 443         string                                                  extensions              = testWithNan ? "OpExtension \"SPV_KHR_float_controls\"\n" : "";
 444         string                                                  capabilities    = testWithNan ? "OpCapability SignedZeroInfNanPreserve\n" : "";
 445         string                                                  exeModes                = testWithNan ? "OpExecutionMode %main SignedZeroInfNanPreserve 32\n" : "";
 446         const StringTemplate                    shaderTemplate  (
 447                 string(getComputeAsmShaderPreamble(capabilities, extensions, exeModes)) +
 448                 "OpSource GLSL 430\n"
 449                 "OpName %main           \"main\"\n"
 450                 "OpName %id             \"gl_GlobalInvocationID\"\n"
 451
 452                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
 453
 454                 "OpDecorate %buf BufferBlock\n"
 455                 "OpDecorate %buf2 BufferBlock\n"
 456                 "OpDecorate %indata1 DescriptorSet 0\n"
 457                 "OpDecorate %indata1 Binding 0\n"
 458                 "OpDecorate %indata2 DescriptorSet 0\n"
 459                 "OpDecorate %indata2 Binding 1\n"
 460                 "OpDecorate %outdata DescriptorSet 0\n"
 461                 "OpDecorate %outdata Binding 2\n"
 462                 "OpDecorate %f32arr ArrayStride 4\n"
 463                 "OpDecorate %i32arr ArrayStride 4\n"
 464                 "OpMemberDecorate %buf 0 Offset 0\n"
 465                 "OpMemberDecorate %buf2 0 Offset 0\n"
 466
 467                 + string(getComputeAsmCommonTypes()) +
 468
 469                 "%buf        = OpTypeStruct %f32arr\n"
 470                 "%bufptr     = OpTypePointer Uniform %buf\n"
 471                 "%indata1    = OpVariable %bufptr Uniform\n"
 472                 "%indata2    = OpVariable %bufptr Uniform\n"
 473
 474                 "%buf2       = OpTypeStruct %i32arr\n"
 475                 "%buf2ptr    = OpTypePointer Uniform %buf2\n"
 476                 "%outdata    = OpVariable %buf2ptr Uniform\n"
 477
 478                 "%id        = OpVariable %uvec3ptr Input\n"
 479                 "%zero      = OpConstant %i32 0\n"
 480                 "%consti1   = OpConstant %i32 1\n"
 481                 "%constf1   = OpConstant %f32 1.0\n"
 482
 483                 "%main      = OpFunction %void None %voidf\n"
 484                 "%label     = OpLabel\n"
 485                 "%idval     = OpLoad %uvec3 %id\n"
 486                 "%x         = OpCompositeExtract %u32 %idval 0\n"
 487
 488                 "%inloc1    = OpAccessChain %f32ptr %indata1 %zero %x\n"
 489                 "%inval1    = OpLoad %f32 %inloc1\n"
 490                 "%inloc2    = OpAccessChain %f32ptr %indata2 %zero %x\n"
 491                 "%inval2    = OpLoad %f32 %inloc2\n"
 492                 "%outloc    = OpAccessChain %i32ptr %outdata %zero %x\n"
 493
 494                 "%result    = ${OPCODE} %bool %inval1 %inval2\n"
 495                 "%int_res   = OpSelect %i32 %result %consti1 %zero\n"
 496                 "             OpStore %outloc %int_res\n"
 497
 498                 "             OpReturn\n"
 499                 "             OpFunctionEnd\n");
 500
 501         ADD_OPFUNORD_CASE(equal, "OpFUnordEqual", ==);
 502         ADD_OPFUNORD_CASE(less, "OpFUnordLessThan", <);
 503         ADD_OPFUNORD_CASE(lessequal, "OpFUnordLessThanEqual", <=);
 504         ADD_OPFUNORD_CASE(greater, "OpFUnordGreaterThan", >);
 505         ADD_OPFUNORD_CASE(greaterequal, "OpFUnordGreaterThanEqual", >=);
 506         ADD_OPFUNORD_CASE(notequal, "OpFUnordNotEqual", !=);
 507
 508         for (size_t caseNdx = 0; caseNdx < cases.size(); ++caseNdx)
 509         {
 510                 map<string, string>                     specializations;
 511                 ComputeShaderSpec                       spec;
 512                 const float                                     NaN                             = std::numeric_limits<float>::quiet_NaN();
 513                 vector<float>                           inputFloats1    (numElements, 0);
 514                 vector<float>                           inputFloats2    (numElements, 0);
 515                 vector<deInt32>                         expectedInts    (numElements, 0);
 516
 517                 specializations["OPCODE"]       = cases[caseNdx].opCode;
 518                 spec.assembly                           = shaderTemplate.specialize(specializations);
 519
 520                 fillRandomScalars(rnd, 1.f, 100.f, &inputFloats1[0], numElements);
 521                 for (size_t ndx = 0; ndx < numElements; ++ndx)
 522                 {
 523                         switch (ndx % 6)
 524                         {
 525                                 case 0:         inputFloats2[ndx] = inputFloats1[ndx] + 1.0f; break;
 526                                 case 1:         inputFloats2[ndx] = inputFloats1[ndx] - 1.0f; break;
 527                                 case 2:         inputFloats2[ndx] = inputFloats1[ndx]; break;
 528                                 case 3:         inputFloats2[ndx] = NaN; break;
 529                                 case 4:         inputFloats2[ndx] = inputFloats1[ndx];  inputFloats1[ndx] = NaN; break;
 530                                 case 5:         inputFloats2[ndx] = NaN;                                inputFloats1[ndx] = NaN; break;
 531                         }
 532                         expectedInts[ndx] = tcu::Float32(inputFloats1[ndx]).isNaN() || tcu::Float32(inputFloats2[ndx]).isNaN() || cases[caseNdx].compareFunc(inputFloats1[ndx], inputFloats2[ndx]);
 533                 }
 534
 535                 spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats1)));
 536                 spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats2)));
 537                 spec.outputs.push_back(BufferSp(new Int32Buffer(expectedInts)));
 538                 spec.numWorkGroups      = IVec3(numElements, 1, 1);
 539                 spec.verifyIO           = testWithNan ? &compareFUnord<true> : &compareFUnord<false>;
 540
 541                 if (testWithNan)
 542                 {
 543                         spec.extensions.push_back("VK_KHR_shader_float_controls");
 544                         spec.requestedVulkanFeatures.floatControlsProperties.shaderSignedZeroInfNanPreserveFloat32 = DE_TRUE;
 545                 }
 546
 547                 group->addChild(new SpvAsmComputeShaderCase(testCtx, cases[caseNdx].name, cases[caseNdx].name, spec));
 548         }
 549
 550         return group.release();
 551 }
 552
 553 struct OpAtomicCase
 554 {
 555         const char*             name;
 556         const char*             assembly;
 557         const char*             retValAssembly;
 558         OpAtomicType    opAtomic;
 559         deInt32                 numOutputElements;
 560
 561                                         OpAtomicCase(const char* _name, const char* _assembly, const char* _retValAssembly, OpAtomicType _opAtomic, deInt32 _numOutputElements)
 562                                                 : name                          (_name)
 563                                                 , assembly                      (_assembly)
 564                                                 , retValAssembly        (_retValAssembly)
 565                                                 , opAtomic                      (_opAtomic)
 566                                                 , numOutputElements     (_numOutputElements) {}
 567 };
 568
 569 tcu::TestCaseGroup* createOpAtomicGroup (tcu::TestContext& testCtx, bool useStorageBuffer, int numElements = 65535, bool verifyReturnValues = false)
 570 {
 571         std::string                                             groupName                       ("opatomic");
 572         if (useStorageBuffer)
 573                 groupName += "_storage_buffer";
 574         if (verifyReturnValues)
 575                 groupName += "_return_values";
 576         de::MovePtr<tcu::TestCaseGroup> group                           (new tcu::TestCaseGroup(testCtx, groupName.c_str(), "Test the OpAtomic* opcodes"));
 577         vector<OpAtomicCase>                    cases;
 578
 579         const StringTemplate                    shaderTemplate  (
 580
 581                 string("OpCapability Shader\n") +
 582                 (useStorageBuffer ? "OpExtension \"SPV_KHR_storage_buffer_storage_class\"\n" : "") +
 583                 "OpMemoryModel Logical GLSL450\n"
 584                 "OpEntryPoint GLCompute %main \"main\" %id\n"
 585                 "OpExecutionMode %main LocalSize 1 1 1\n" +
 586
 587                 "OpSource GLSL 430\n"
 588                 "OpName %main           \"main\"\n"
 589                 "OpName %id             \"gl_GlobalInvocationID\"\n"
 590
 591                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
 592
 593                 "OpDecorate %buf ${BLOCK_DECORATION}\n"
 594                 "OpDecorate %indata DescriptorSet 0\n"
 595                 "OpDecorate %indata Binding 0\n"
 596                 "OpDecorate %i32arr ArrayStride 4\n"
 597                 "OpMemberDecorate %buf 0 Offset 0\n"
 598
 599                 "OpDecorate %sumbuf ${BLOCK_DECORATION}\n"
 600                 "OpDecorate %sum DescriptorSet 0\n"
 601                 "OpDecorate %sum Binding 1\n"
 602                 "OpMemberDecorate %sumbuf 0 Coherent\n"
 603                 "OpMemberDecorate %sumbuf 0 Offset 0\n"
 604
 605                 "${RETVAL_BUF_DECORATE}"
 606
 607                 + getComputeAsmCommonTypes("${BLOCK_POINTER_TYPE}") +
 608
 609                 "%buf       = OpTypeStruct %i32arr\n"
 610                 "%bufptr    = OpTypePointer ${BLOCK_POINTER_TYPE} %buf\n"
 611                 "%indata    = OpVariable %bufptr ${BLOCK_POINTER_TYPE}\n"
 612
 613                 "%sumbuf    = OpTypeStruct %i32arr\n"
 614                 "%sumbufptr = OpTypePointer ${BLOCK_POINTER_TYPE} %sumbuf\n"
 615                 "%sum       = OpVariable %sumbufptr ${BLOCK_POINTER_TYPE}\n"
 616
 617                 "${RETVAL_BUF_DECL}"
 618
 619                 "%id        = OpVariable %uvec3ptr Input\n"
 620                 "%minusone  = OpConstant %i32 -1\n"
 621                 "%zero      = OpConstant %i32 0\n"
 622                 "%one       = OpConstant %u32 1\n"
 623                 "%two       = OpConstant %i32 2\n"
 624
 625                 "%main      = OpFunction %void None %voidf\n"
 626                 "%label     = OpLabel\n"
 627                 "%idval     = OpLoad %uvec3 %id\n"
 628                 "%x         = OpCompositeExtract %u32 %idval 0\n"
 629
 630                 "%inloc     = OpAccessChain %i32ptr %indata %zero %x\n"
 631                 "%inval     = OpLoad %i32 %inloc\n"
 632
 633                 "%outloc    = OpAccessChain %i32ptr %sum %zero ${INDEX}\n"
 634                 "${INSTRUCTION}"
 635                 "${RETVAL_ASSEMBLY}"
 636
 637                 "             OpReturn\n"
 638                 "             OpFunctionEnd\n");
 639
 640         #define ADD_OPATOMIC_CASE(NAME, ASSEMBLY, RETVAL_ASSEMBLY, OPATOMIC, NUM_OUTPUT_ELEMENTS) \
 641         do { \
 642                 DE_ASSERT((NUM_OUTPUT_ELEMENTS) == 1 || (NUM_OUTPUT_ELEMENTS) == numElements); \
 643                 cases.push_back(OpAtomicCase(#NAME, ASSEMBLY, RETVAL_ASSEMBLY, OPATOMIC, NUM_OUTPUT_ELEMENTS)); \
 644         } while (deGetFalse())
 645         #define ADD_OPATOMIC_CASE_1(NAME, ASSEMBLY, RETVAL_ASSEMBLY, OPATOMIC) ADD_OPATOMIC_CASE(NAME, ASSEMBLY, RETVAL_ASSEMBLY, OPATOMIC, 1)
 646         #define ADD_OPATOMIC_CASE_N(NAME, ASSEMBLY, RETVAL_ASSEMBLY, OPATOMIC) ADD_OPATOMIC_CASE(NAME, ASSEMBLY, RETVAL_ASSEMBLY, OPATOMIC, numElements)
 647
 648         ADD_OPATOMIC_CASE_1(iadd,       "%retv      = OpAtomicIAdd %i32 %outloc %one %zero %inval\n",
 649                                                                 "             OpStore %retloc %retv\n", OPATOMIC_IADD );
 650         ADD_OPATOMIC_CASE_1(isub,       "%retv      = OpAtomicISub %i32 %outloc %one %zero %inval\n",
 651                                                                 "             OpStore %retloc %retv\n", OPATOMIC_ISUB );
 652         ADD_OPATOMIC_CASE_1(iinc,       "%retv      = OpAtomicIIncrement %i32 %outloc %one %zero\n",
 653                                                                 "             OpStore %retloc %retv\n", OPATOMIC_IINC );
 654         ADD_OPATOMIC_CASE_1(idec,       "%retv      = OpAtomicIDecrement %i32 %outloc %one %zero\n",
 655                                                                 "             OpStore %retloc %retv\n", OPATOMIC_IDEC );
 656         if (!verifyReturnValues)
 657         {
 658                 ADD_OPATOMIC_CASE_N(load,       "%inval2    = OpAtomicLoad %i32 %inloc %one %zero\n"
 659                                                                         "             OpStore %outloc %inval2\n", "", OPATOMIC_LOAD );
 660                 ADD_OPATOMIC_CASE_N(store,      "             OpAtomicStore %outloc %one %zero %inval\n", "", OPATOMIC_STORE );
 661         }
 662
 663         ADD_OPATOMIC_CASE_N(compex, "%even      = OpSMod %i32 %inval %two\n"
 664                                                                 "             OpStore %outloc %even\n"
 665                                                                 "%retv      = OpAtomicCompareExchange %i32 %outloc %one %zero %zero %minusone %zero\n",
 666                                                                 "                         OpStore %retloc %retv\n", OPATOMIC_COMPEX );
 667
 668
 669         #undef ADD_OPATOMIC_CASE
 670         #undef ADD_OPATOMIC_CASE_1
 671         #undef ADD_OPATOMIC_CASE_N
 672
 673         for (size_t caseNdx = 0; caseNdx < cases.size(); ++caseNdx)
 674         {
 675                 map<string, string>                     specializations;
 676                 ComputeShaderSpec                       spec;
 677                 vector<deInt32>                         inputInts               (numElements, 0);
 678                 vector<deInt32>                         expected                (cases[caseNdx].numOutputElements, -1);
 679
 680                 specializations["INDEX"]                                = (cases[caseNdx].numOutputElements == 1) ? "%zero" : "%x";
 681                 specializations["INSTRUCTION"]                  = cases[caseNdx].assembly;
 682                 specializations["BLOCK_DECORATION"]             = useStorageBuffer ? "Block" : "BufferBlock";
 683                 specializations["BLOCK_POINTER_TYPE"]   = useStorageBuffer ? "StorageBuffer" : "Uniform";
 684
 685                 if (verifyReturnValues)
 686                 {
 687                         const StringTemplate blockDecoration    (
 688                                 "\n"
 689                                 "OpDecorate %retbuf ${BLOCK_DECORATION}\n"
 690                                 "OpDecorate %ret DescriptorSet 0\n"
 691                                 "OpDecorate %ret Binding 2\n"
 692                                 "OpMemberDecorate %retbuf 0 Offset 0\n\n");
 693
 694                         const StringTemplate blockDeclaration   (
 695                                 "\n"
 696                                 "%retbuf    = OpTypeStruct %i32arr\n"
 697                                 "%retbufptr = OpTypePointer ${BLOCK_POINTER_TYPE} %retbuf\n"
 698                                 "%ret       = OpVariable %retbufptr ${BLOCK_POINTER_TYPE}\n\n");
 699
 700                         specializations["RETVAL_ASSEMBLY"] =
 701                                 "%retloc    = OpAccessChain %i32ptr %ret %zero %x\n"
 702                                 + std::string(cases[caseNdx].retValAssembly);
 703
 704                         specializations["RETVAL_BUF_DECORATE"]  = blockDecoration.specialize(specializations);
 705                         specializations["RETVAL_BUF_DECL"]              = blockDeclaration.specialize(specializations);
 706                 }
 707                 else
 708                 {
 709                         specializations["RETVAL_ASSEMBLY"]              = "";
 710                         specializations["RETVAL_BUF_DECORATE"]  = "";
 711                         specializations["RETVAL_BUF_DECL"]              = "";
 712                 }
 713
 714                 spec.assembly                                                   = shaderTemplate.specialize(specializations);
 715
 716                 if (useStorageBuffer)
 717                         spec.extensions.push_back("VK_KHR_storage_buffer_storage_class");
 718
 719                 spec.inputs.push_back(BufferSp(new OpAtomicBuffer(numElements, cases[caseNdx].numOutputElements, cases[caseNdx].opAtomic, BUFFERTYPE_INPUT)));
 720                 spec.outputs.push_back(BufferSp(new OpAtomicBuffer(numElements, cases[caseNdx].numOutputElements, cases[caseNdx].opAtomic, BUFFERTYPE_EXPECTED)));
 721                 if (verifyReturnValues)
 722                         spec.outputs.push_back(BufferSp(new OpAtomicBuffer(numElements, cases[caseNdx].numOutputElements, cases[caseNdx].opAtomic, BUFFERTYPE_ATOMIC_RET)));
 723                 spec.numWorkGroups = IVec3(numElements, 1, 1);
 724
 725                 if (verifyReturnValues)
 726                 {
 727                         switch (cases[caseNdx].opAtomic)
 728                         {
 729                                 case OPATOMIC_IADD:
 730                                         spec.verifyIO = OpAtomicBuffer::compareWithRetvals<OPATOMIC_IADD>;
 731                                         break;
 732                                 case OPATOMIC_ISUB:
 733                                         spec.verifyIO = OpAtomicBuffer::compareWithRetvals<OPATOMIC_ISUB>;
 734                                         break;
 735                                 case OPATOMIC_IINC:
 736                                         spec.verifyIO = OpAtomicBuffer::compareWithRetvals<OPATOMIC_IINC>;
 737                                         break;
 738                                 case OPATOMIC_IDEC:
 739                                         spec.verifyIO = OpAtomicBuffer::compareWithRetvals<OPATOMIC_IDEC>;
 740                                         break;
 741                                 case OPATOMIC_COMPEX:
 742                                         spec.verifyIO = OpAtomicBuffer::compareWithRetvals<OPATOMIC_COMPEX>;
 743                                         break;
 744                                 default:
 745                                         DE_FATAL("Unsupported OpAtomic type for return value verification");
 746                         }
 747                 }
 748                 group->addChild(new SpvAsmComputeShaderCase(testCtx, cases[caseNdx].name, cases[caseNdx].name, spec));
 749         }
 750
 751         return group.release();
 752 }
 753
 754 tcu::TestCaseGroup* createOpLineGroup (tcu::TestContext& testCtx)
 755 {
 756         de::MovePtr<tcu::TestCaseGroup> group                   (new tcu::TestCaseGroup(testCtx, "opline", "Test the OpLine instruction"));
 757         ComputeShaderSpec                               spec;
 758         de::Random                                              rnd                             (deStringHash(group->getName()));
 759         const int                                               numElements             = 100;
 760         vector<float>                                   positiveFloats  (numElements, 0);
 761         vector<float>                                   negativeFloats  (numElements, 0);
 762
 763         fillRandomScalars(rnd, 1.f, 100.f, &positiveFloats[0], numElements);
 764
 765         for (size_t ndx = 0; ndx < numElements; ++ndx)
 766                 negativeFloats[ndx] = -positiveFloats[ndx];
 767
 768         spec.assembly =
 769                 string(getComputeAsmShaderPreamble()) +
 770
 771                 "%fname1 = OpString \"negateInputs.comp\"\n"
 772                 "%fname2 = OpString \"negateInputs\"\n"
 773
 774                 "OpSource GLSL 430\n"
 775                 "OpName %main           \"main\"\n"
 776                 "OpName %id             \"gl_GlobalInvocationID\"\n"
 777
 778                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
 779
 780                 + string(getComputeAsmInputOutputBufferTraits()) +
 781
 782                 "OpLine %fname1 0 0\n" // At the earliest possible position
 783
 784                 + string(getComputeAsmCommonTypes()) + string(getComputeAsmInputOutputBuffer()) +
 785
 786                 "OpLine %fname1 0 1\n" // Multiple OpLines in sequence
 787                 "OpLine %fname2 1 0\n" // Different filenames
 788                 "OpLine %fname1 1000 100000\n"
 789
 790                 "%id        = OpVariable %uvec3ptr Input\n"
 791                 "%zero      = OpConstant %i32 0\n"
 792
 793                 "OpLine %fname1 1 1\n" // Before a function
 794
 795                 "%main      = OpFunction %void None %voidf\n"
 796                 "%label     = OpLabel\n"
 797
 798                 "OpLine %fname1 1 1\n" // In a function
 799
 800                 "%idval     = OpLoad %uvec3 %id\n"
 801                 "%x         = OpCompositeExtract %u32 %idval 0\n"
 802                 "%inloc     = OpAccessChain %f32ptr %indata %zero %x\n"
 803                 "%inval     = OpLoad %f32 %inloc\n"
 804                 "%neg       = OpFNegate %f32 %inval\n"
 805                 "%outloc    = OpAccessChain %f32ptr %outdata %zero %x\n"
 806                 "             OpStore %outloc %neg\n"
 807                 "             OpReturn\n"
 808                 "             OpFunctionEnd\n";
 809         spec.inputs.push_back(BufferSp(new Float32Buffer(positiveFloats)));
 810         spec.outputs.push_back(BufferSp(new Float32Buffer(negativeFloats)));
 811         spec.numWorkGroups = IVec3(numElements, 1, 1);
 812
 813         group->addChild(new SpvAsmComputeShaderCase(testCtx, "all", "OpLine appearing at different places", spec));
 814
 815         return group.release();
 816 }
 817
 818 bool veryfiBinaryShader (const ProgramBinary& binary)
 819 {
 820         const size_t    paternCount                     = 3u;
 821         bool paternsCheck[paternCount]          =
 822         {
 823                 false, false, false
 824         };
 825         const string patersns[paternCount]      =
 826         {
 827                 "VULKAN CTS",
 828                 "Negative values",
 829                 "Date: 2017/09/21"
 830         };
 831         size_t                  paternNdx               = 0u;
 832
 833         for (size_t ndx = 0u; ndx < binary.getSize(); ++ndx)
 834         {
 835                 if (false == paternsCheck[paternNdx] &&
 836                         patersns[paternNdx][0] == static_cast<char>(binary.getBinary()[ndx]) &&
 837                         deMemoryEqual((const char*)&binary.getBinary()[ndx], &patersns[paternNdx][0], patersns[paternNdx].length()))
 838                 {
 839                         paternsCheck[paternNdx]= true;
 840                         paternNdx++;
 841                         if (paternNdx == paternCount)
 842                                 break;
 843                 }
 844         }
 845
 846         for (size_t ndx = 0u; ndx < paternCount; ++ndx)
 847         {
 848                 if (!paternsCheck[ndx])
 849                         return false;
 850         }
 851
 852         return true;
 853 }
 854
 855 tcu::TestCaseGroup* createOpModuleProcessedGroup (tcu::TestContext& testCtx)
 856 {
 857         de::MovePtr<tcu::TestCaseGroup> group                   (new tcu::TestCaseGroup(testCtx, "opmoduleprocessed", "Test the OpModuleProcessed instruction"));
 858         ComputeShaderSpec                               spec;
 859         de::Random                                              rnd                             (deStringHash(group->getName()));
 860         const int                                               numElements             = 10;
 861         vector<float>                                   positiveFloats  (numElements, 0);
 862         vector<float>                                   negativeFloats  (numElements, 0);
 863
 864         fillRandomScalars(rnd, 1.f, 100.f, &positiveFloats[0], numElements);
 865
 866         for (size_t ndx = 0; ndx < numElements; ++ndx)
 867                 negativeFloats[ndx] = -positiveFloats[ndx];
 868
 869         spec.assembly =
 870                 string(getComputeAsmShaderPreamble()) +
 871                 "%fname = OpString \"negateInputs.comp\"\n"
 872
 873                 "OpSource GLSL 430\n"
 874                 "OpName %main           \"main\"\n"
 875                 "OpName %id             \"gl_GlobalInvocationID\"\n"
 876                 "OpModuleProcessed \"VULKAN CTS\"\n"                                    //OpModuleProcessed;
 877                 "OpModuleProcessed \"Negative values\"\n"
 878                 "OpModuleProcessed \"Date: 2017/09/21\"\n"
 879                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
 880
 881                 + string(getComputeAsmInputOutputBufferTraits())
 882
 883                 + string(getComputeAsmCommonTypes()) + string(getComputeAsmInputOutputBuffer()) +
 884
 885                 "OpLine %fname 0 1\n"
 886
 887                 "OpLine %fname 1000 1\n"
 888
 889                 "%id        = OpVariable %uvec3ptr Input\n"
 890                 "%zero      = OpConstant %i32 0\n"
 891                 "%main      = OpFunction %void None %voidf\n"
 892
 893                 "%label     = OpLabel\n"
 894                 "%idval     = OpLoad %uvec3 %id\n"
 895                 "%x         = OpCompositeExtract %u32 %idval 0\n"
 896
 897                 "%inloc     = OpAccessChain %f32ptr %indata %zero %x\n"
 898                 "%inval     = OpLoad %f32 %inloc\n"
 899                 "%neg       = OpFNegate %f32 %inval\n"
 900                 "%outloc    = OpAccessChain %f32ptr %outdata %zero %x\n"
 901                 "             OpStore %outloc %neg\n"
 902                 "             OpReturn\n"
 903                 "             OpFunctionEnd\n";
 904         spec.inputs.push_back(BufferSp(new Float32Buffer(positiveFloats)));
 905         spec.outputs.push_back(BufferSp(new Float32Buffer(negativeFloats)));
 906         spec.numWorkGroups = IVec3(numElements, 1, 1);
 907         spec.verifyBinary = veryfiBinaryShader;
 908         spec.spirvVersion = SPIRV_VERSION_1_3;
 909
 910         group->addChild(new SpvAsmComputeShaderCase(testCtx, "all", "OpModuleProcessed Tests", spec));
 911
 912         return group.release();
 913 }
 914
 915 tcu::TestCaseGroup* createOpNoLineGroup (tcu::TestContext& testCtx)
 916 {
 917         de::MovePtr<tcu::TestCaseGroup> group                   (new tcu::TestCaseGroup(testCtx, "opnoline", "Test the OpNoLine instruction"));
 918         ComputeShaderSpec                               spec;
 919         de::Random                                              rnd                             (deStringHash(group->getName()));
 920         const int                                               numElements             = 100;
 921         vector<float>                                   positiveFloats  (numElements, 0);
 922         vector<float>                                   negativeFloats  (numElements, 0);
 923
 924         fillRandomScalars(rnd, 1.f, 100.f, &positiveFloats[0], numElements);
 925
 926         for (size_t ndx = 0; ndx < numElements; ++ndx)
 927                 negativeFloats[ndx] = -positiveFloats[ndx];
 928
 929         spec.assembly =
 930                 string(getComputeAsmShaderPreamble()) +
 931
 932                 "%fname = OpString \"negateInputs.comp\"\n"
 933
 934                 "OpSource GLSL 430\n"
 935                 "OpName %main           \"main\"\n"
 936                 "OpName %id             \"gl_GlobalInvocationID\"\n"
 937
 938                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
 939
 940                 + string(getComputeAsmInputOutputBufferTraits()) +
 941
 942                 "OpNoLine\n" // At the earliest possible position, without preceding OpLine
 943
 944                 + string(getComputeAsmCommonTypes()) + string(getComputeAsmInputOutputBuffer()) +
 945
 946                 "OpLine %fname 0 1\n"
 947                 "OpNoLine\n" // Immediately following a preceding OpLine
 948
 949                 "OpLine %fname 1000 1\n"
 950
 951                 "%id        = OpVariable %uvec3ptr Input\n"
 952                 "%zero      = OpConstant %i32 0\n"
 953
 954                 "OpNoLine\n" // Contents after the previous OpLine
 955
 956                 "%main      = OpFunction %void None %voidf\n"
 957                 "%label     = OpLabel\n"
 958                 "%idval     = OpLoad %uvec3 %id\n"
 959                 "%x         = OpCompositeExtract %u32 %idval 0\n"
 960
 961                 "OpNoLine\n" // Multiple OpNoLine
 962                 "OpNoLine\n"
 963                 "OpNoLine\n"
 964
 965                 "%inloc     = OpAccessChain %f32ptr %indata %zero %x\n"
 966                 "%inval     = OpLoad %f32 %inloc\n"
 967                 "%neg       = OpFNegate %f32 %inval\n"
 968                 "%outloc    = OpAccessChain %f32ptr %outdata %zero %x\n"
 969                 "             OpStore %outloc %neg\n"
 970                 "             OpReturn\n"
 971                 "             OpFunctionEnd\n";
 972         spec.inputs.push_back(BufferSp(new Float32Buffer(positiveFloats)));
 973         spec.outputs.push_back(BufferSp(new Float32Buffer(negativeFloats)));
 974         spec.numWorkGroups = IVec3(numElements, 1, 1);
 975
 976         group->addChild(new SpvAsmComputeShaderCase(testCtx, "all", "OpNoLine appearing at different places", spec));
 977
 978         return group.release();
 979 }
 980
 981 // Compare instruction for the contraction compute case.
 982 // Returns true if the output is what is expected from the test case.
 983 bool compareNoContractCase(const std::vector<Resource>&, const vector<AllocationSp>& outputAllocs, const std::vector<Resource>& expectedOutputs, TestLog&)
 984 {
 985         if (outputAllocs.size() != 1)
 986                 return false;
 987
 988         // Only size is needed because we are not comparing the exact values.
 989         size_t byteSize = expectedOutputs[0].getByteSize();
 990
 991         const float*    outputAsFloat   = static_cast<const float*>(outputAllocs[0]->getHostPtr());
 992
 993         for(size_t i = 0; i < byteSize / sizeof(float); ++i) {
 994                 if (outputAsFloat[i] != 0.f &&
 995                         outputAsFloat[i] != -ldexp(1, -24)) {
 996                         return false;
 997                 }
 998         }
 999
1000         return true;
1001 }
1002
1003 tcu::TestCaseGroup* createNoContractionGroup (tcu::TestContext& testCtx)
1004 {
1005         de::MovePtr<tcu::TestCaseGroup> group                   (new tcu::TestCaseGroup(testCtx, "nocontraction", "Test the NoContraction decoration"));
1006         vector<CaseParameter>                   cases;
1007         const int                                               numElements             = 100;
1008         vector<float>                                   inputFloats1    (numElements, 0);
1009         vector<float>                                   inputFloats2    (numElements, 0);
1010         vector<float>                                   outputFloats    (numElements, 0);
1011         const StringTemplate                    shaderTemplate  (
1012                 string(getComputeAsmShaderPreamble()) +
1013
1014                 "OpName %main           \"main\"\n"
1015                 "OpName %id             \"gl_GlobalInvocationID\"\n"
1016
1017                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
1018
1019                 "${DECORATION}\n"
1020
1021                 "OpDecorate %buf BufferBlock\n"
1022                 "OpDecorate %indata1 DescriptorSet 0\n"
1023                 "OpDecorate %indata1 Binding 0\n"
1024                 "OpDecorate %indata2 DescriptorSet 0\n"
1025                 "OpDecorate %indata2 Binding 1\n"
1026                 "OpDecorate %outdata DescriptorSet 0\n"
1027                 "OpDecorate %outdata Binding 2\n"
1028                 "OpDecorate %f32arr ArrayStride 4\n"
1029                 "OpMemberDecorate %buf 0 Offset 0\n"
1030
1031                 + string(getComputeAsmCommonTypes()) +
1032
1033                 "%buf        = OpTypeStruct %f32arr\n"
1034                 "%bufptr     = OpTypePointer Uniform %buf\n"
1035                 "%indata1    = OpVariable %bufptr Uniform\n"
1036                 "%indata2    = OpVariable %bufptr Uniform\n"
1037                 "%outdata    = OpVariable %bufptr Uniform\n"
1038
1039                 "%id         = OpVariable %uvec3ptr Input\n"
1040                 "%zero       = OpConstant %i32 0\n"
1041                 "%c_f_m1     = OpConstant %f32 -1.\n"
1042
1043                 "%main       = OpFunction %void None %voidf\n"
1044                 "%label      = OpLabel\n"
1045                 "%idval      = OpLoad %uvec3 %id\n"
1046                 "%x          = OpCompositeExtract %u32 %idval 0\n"
1047                 "%inloc1     = OpAccessChain %f32ptr %indata1 %zero %x\n"
1048                 "%inval1     = OpLoad %f32 %inloc1\n"
1049                 "%inloc2     = OpAccessChain %f32ptr %indata2 %zero %x\n"
1050                 "%inval2     = OpLoad %f32 %inloc2\n"
1051                 "%mul        = OpFMul %f32 %inval1 %inval2\n"
1052                 "%add        = OpFAdd %f32 %mul %c_f_m1\n"
1053                 "%outloc     = OpAccessChain %f32ptr %outdata %zero %x\n"
1054                 "              OpStore %outloc %add\n"
1055                 "              OpReturn\n"
1056                 "              OpFunctionEnd\n");
1057
1058         cases.push_back(CaseParameter("multiplication", "OpDecorate %mul NoContraction"));
1059         cases.push_back(CaseParameter("addition",               "OpDecorate %add NoContraction"));
1060         cases.push_back(CaseParameter("both",                   "OpDecorate %mul NoContraction\nOpDecorate %add NoContraction"));
1061
1062         for (size_t ndx = 0; ndx < numElements; ++ndx)
1063         {
1064                 inputFloats1[ndx]       = 1.f + std::ldexp(1.f, -23); // 1 + 2^-23.
1065                 inputFloats2[ndx]       = 1.f - std::ldexp(1.f, -23); // 1 - 2^-23.
1066                 // Result for (1 + 2^-23) * (1 - 2^-23) - 1. With NoContraction, the multiplication will be
1067                 // conducted separately and the result is rounded to 1, or 0x1.fffffcp-1
1068                 // So the final result will be 0.f or 0x1p-24.
1069                 // If the operation is combined into a precise fused multiply-add, then the result would be
1070                 // 2^-46 (0xa8800000).
1071                 outputFloats[ndx]       = 0.f;
1072         }
1073
1074         for (size_t caseNdx = 0; caseNdx < cases.size(); ++caseNdx)
1075         {
1076                 map<string, string>             specializations;
1077                 ComputeShaderSpec               spec;
1078
1079                 specializations["DECORATION"] = cases[caseNdx].param;
1080                 spec.assembly = shaderTemplate.specialize(specializations);
1081                 spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats1)));
1082                 spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats2)));
1083                 spec.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
1084                 spec.numWorkGroups = IVec3(numElements, 1, 1);
1085                 // Check against the two possible answers based on rounding mode.
1086                 spec.verifyIO = &compareNoContractCase;
1087
1088                 group->addChild(new SpvAsmComputeShaderCase(testCtx, cases[caseNdx].name, cases[caseNdx].name, spec));
1089         }
1090         return group.release();
1091 }
1092
1093 bool compareFRem(const std::vector<Resource>&, const vector<AllocationSp>& outputAllocs, const std::vector<Resource>& expectedOutputs, TestLog&)
1094 {
1095         if (outputAllocs.size() != 1)
1096                 return false;
1097
1098         vector<deUint8> expectedBytes;
1099         expectedOutputs[0].getBytes(expectedBytes);
1100
1101         const float*    expectedOutputAsFloat   = reinterpret_cast<const float*>(&expectedBytes.front());
1102         const float*    outputAsFloat                   = static_cast<const float*>(outputAllocs[0]->getHostPtr());
1103
1104         for (size_t idx = 0; idx < expectedBytes.size() / sizeof(float); ++idx)
1105         {
1106                 const float f0 = expectedOutputAsFloat[idx];
1107                 const float f1 = outputAsFloat[idx];
1108                 // \todo relative error needs to be fairly high because FRem may be implemented as
1109                 // (roughly) frac(a/b)*b, so LSB errors can be magnified. But this should be fine for now.
1110                 if (deFloatAbs((f1 - f0) / f0) > 0.02)
1111                         return false;
1112         }
1113
1114         return true;
1115 }
1116
1117 tcu::TestCaseGroup* createOpFRemGroup (tcu::TestContext& testCtx)
1118 {
1119         de::MovePtr<tcu::TestCaseGroup> group                   (new tcu::TestCaseGroup(testCtx, "opfrem", "Test the OpFRem instruction"));
1120         ComputeShaderSpec                               spec;
1121         de::Random                                              rnd                             (deStringHash(group->getName()));
1122         const int                                               numElements             = 200;
1123         vector<float>                                   inputFloats1    (numElements, 0);
1124         vector<float>                                   inputFloats2    (numElements, 0);
1125         vector<float>                                   outputFloats    (numElements, 0);
1126
1127         fillRandomScalars(rnd, -10000.f, 10000.f, &inputFloats1[0], numElements);
1128         fillRandomScalars(rnd, -100.f, 100.f, &inputFloats2[0], numElements);
1129
1130         for (size_t ndx = 0; ndx < numElements; ++ndx)
1131         {
1132                 // Guard against divisors near zero.
1133                 if (std::fabs(inputFloats2[ndx]) < 1e-3)
1134                         inputFloats2[ndx] = 8.f;
1135
1136                 // The return value of std::fmod() has the same sign as its first operand, which is how OpFRem spec'd.
1137                 outputFloats[ndx] = std::fmod(inputFloats1[ndx], inputFloats2[ndx]);
1138         }
1139
1140         spec.assembly =
1141                 string(getComputeAsmShaderPreamble()) +
1142
1143                 "OpName %main           \"main\"\n"
1144                 "OpName %id             \"gl_GlobalInvocationID\"\n"
1145
1146                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
1147
1148                 "OpDecorate %buf BufferBlock\n"
1149                 "OpDecorate %indata1 DescriptorSet 0\n"
1150                 "OpDecorate %indata1 Binding 0\n"
1151                 "OpDecorate %indata2 DescriptorSet 0\n"
1152                 "OpDecorate %indata2 Binding 1\n"
1153                 "OpDecorate %outdata DescriptorSet 0\n"
1154                 "OpDecorate %outdata Binding 2\n"
1155                 "OpDecorate %f32arr ArrayStride 4\n"
1156                 "OpMemberDecorate %buf 0 Offset 0\n"
1157
1158                 + string(getComputeAsmCommonTypes()) +
1159
1160                 "%buf        = OpTypeStruct %f32arr\n"
1161                 "%bufptr     = OpTypePointer Uniform %buf\n"
1162                 "%indata1    = OpVariable %bufptr Uniform\n"
1163                 "%indata2    = OpVariable %bufptr Uniform\n"
1164                 "%outdata    = OpVariable %bufptr Uniform\n"
1165
1166                 "%id        = OpVariable %uvec3ptr Input\n"
1167                 "%zero      = OpConstant %i32 0\n"
1168
1169                 "%main      = OpFunction %void None %voidf\n"
1170                 "%label     = OpLabel\n"
1171                 "%idval     = OpLoad %uvec3 %id\n"
1172                 "%x         = OpCompositeExtract %u32 %idval 0\n"
1173                 "%inloc1    = OpAccessChain %f32ptr %indata1 %zero %x\n"
1174                 "%inval1    = OpLoad %f32 %inloc1\n"
1175                 "%inloc2    = OpAccessChain %f32ptr %indata2 %zero %x\n"
1176                 "%inval2    = OpLoad %f32 %inloc2\n"
1177                 "%rem       = OpFRem %f32 %inval1 %inval2\n"
1178                 "%outloc    = OpAccessChain %f32ptr %outdata %zero %x\n"
1179                 "             OpStore %outloc %rem\n"
1180                 "             OpReturn\n"
1181                 "             OpFunctionEnd\n";
1182
1183         spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats1)));
1184         spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats2)));
1185         spec.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
1186         spec.numWorkGroups = IVec3(numElements, 1, 1);
1187         spec.verifyIO = &compareFRem;
1188
1189         group->addChild(new SpvAsmComputeShaderCase(testCtx, "all", "", spec));
1190
1191         return group.release();
1192 }
1193
1194 bool compareNMin (const std::vector<Resource>&, const vector<AllocationSp>& outputAllocs, const std::vector<Resource>& expectedOutputs, TestLog&)
1195 {
1196         if (outputAllocs.size() != 1)
1197                 return false;
1198
1199         const BufferSp&                 expectedOutput                  (expectedOutputs[0].getBuffer());
1200         std::vector<deUint8>    data;
1201         expectedOutput->getBytes(data);
1202
1203         const float* const              expectedOutputAsFloat   = reinterpret_cast<const float*>(&data.front());
1204         const float* const              outputAsFloat                   = static_cast<const float*>(outputAllocs[0]->getHostPtr());
1205
1206         for (size_t idx = 0; idx < expectedOutput->getByteSize() / sizeof(float); ++idx)
1207         {
1208                 const float f0 = expectedOutputAsFloat[idx];
1209                 const float f1 = outputAsFloat[idx];
1210
1211                 // For NMin, we accept NaN as output if both inputs were NaN.
1212                 // Otherwise the NaN is the wrong choise, as on architectures that
1213                 // do not handle NaN, those are huge values.
1214                 if (!(tcu::Float32(f1).isNaN() && tcu::Float32(f0).isNaN()) && deFloatAbs(f1 - f0) > 0.00001f)
1215                         return false;
1216         }
1217
1218         return true;
1219 }
1220
1221 tcu::TestCaseGroup* createOpNMinGroup (tcu::TestContext& testCtx)
1222 {
1223         de::MovePtr<tcu::TestCaseGroup> group                   (new tcu::TestCaseGroup(testCtx, "opnmin", "Test the OpNMin instruction"));
1224         ComputeShaderSpec                               spec;
1225         de::Random                                              rnd                             (deStringHash(group->getName()));
1226         const int                                               numElements             = 200;
1227         vector<float>                                   inputFloats1    (numElements, 0);
1228         vector<float>                                   inputFloats2    (numElements, 0);
1229         vector<float>                                   outputFloats    (numElements, 0);
1230
1231         fillRandomScalars(rnd, -10000.f, 10000.f, &inputFloats1[0], numElements);
1232         fillRandomScalars(rnd, -10000.f, 10000.f, &inputFloats2[0], numElements);
1233
1234         // Make the first case a full-NAN case.
1235         inputFloats1[0] = TCU_NAN;
1236         inputFloats2[0] = TCU_NAN;
1237
1238         for (size_t ndx = 0; ndx < numElements; ++ndx)
1239         {
1240                 // By default, pick the smallest
1241                 outputFloats[ndx] = std::min(inputFloats1[ndx], inputFloats2[ndx]);
1242
1243                 // Make half of the cases NaN cases
1244                 if ((ndx & 1) == 0)
1245                 {
1246                         // Alternate between the NaN operand
1247                         if ((ndx & 2) == 0)
1248                         {
1249                                 outputFloats[ndx] = inputFloats2[ndx];
1250                                 inputFloats1[ndx] = TCU_NAN;
1251                         }
1252                         else
1253                         {
1254                                 outputFloats[ndx] = inputFloats1[ndx];
1255                                 inputFloats2[ndx] = TCU_NAN;
1256                         }
1257                 }
1258         }
1259
1260         spec.assembly =
1261                 "OpCapability Shader\n"
1262                 "%std450        = OpExtInstImport \"GLSL.std.450\"\n"
1263                 "OpMemoryModel Logical GLSL450\n"
1264                 "OpEntryPoint GLCompute %main \"main\" %id\n"
1265                 "OpExecutionMode %main LocalSize 1 1 1\n"
1266
1267                 "OpName %main           \"main\"\n"
1268                 "OpName %id             \"gl_GlobalInvocationID\"\n"
1269
1270                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
1271
1272                 "OpDecorate %buf BufferBlock\n"
1273                 "OpDecorate %indata1 DescriptorSet 0\n"
1274                 "OpDecorate %indata1 Binding 0\n"
1275                 "OpDecorate %indata2 DescriptorSet 0\n"
1276                 "OpDecorate %indata2 Binding 1\n"
1277                 "OpDecorate %outdata DescriptorSet 0\n"
1278                 "OpDecorate %outdata Binding 2\n"
1279                 "OpDecorate %f32arr ArrayStride 4\n"
1280                 "OpMemberDecorate %buf 0 Offset 0\n"
1281
1282                 + string(getComputeAsmCommonTypes()) +
1283
1284                 "%buf        = OpTypeStruct %f32arr\n"
1285                 "%bufptr     = OpTypePointer Uniform %buf\n"
1286                 "%indata1    = OpVariable %bufptr Uniform\n"
1287                 "%indata2    = OpVariable %bufptr Uniform\n"
1288                 "%outdata    = OpVariable %bufptr Uniform\n"
1289
1290                 "%id        = OpVariable %uvec3ptr Input\n"
1291                 "%zero      = OpConstant %i32 0\n"
1292
1293                 "%main      = OpFunction %void None %voidf\n"
1294                 "%label     = OpLabel\n"
1295                 "%idval     = OpLoad %uvec3 %id\n"
1296                 "%x         = OpCompositeExtract %u32 %idval 0\n"
1297                 "%inloc1    = OpAccessChain %f32ptr %indata1 %zero %x\n"
1298                 "%inval1    = OpLoad %f32 %inloc1\n"
1299                 "%inloc2    = OpAccessChain %f32ptr %indata2 %zero %x\n"
1300                 "%inval2    = OpLoad %f32 %inloc2\n"
1301                 "%rem       = OpExtInst %f32 %std450 NMin %inval1 %inval2\n"
1302                 "%outloc    = OpAccessChain %f32ptr %outdata %zero %x\n"
1303                 "             OpStore %outloc %rem\n"
1304                 "             OpReturn\n"
1305                 "             OpFunctionEnd\n";
1306
1307         spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats1)));
1308         spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats2)));
1309         spec.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
1310         spec.numWorkGroups = IVec3(numElements, 1, 1);
1311         spec.verifyIO = &compareNMin;
1312
1313         group->addChild(new SpvAsmComputeShaderCase(testCtx, "all", "", spec));
1314
1315         return group.release();
1316 }
1317
1318 bool compareNMax (const std::vector<Resource>&, const vector<AllocationSp>& outputAllocs, const std::vector<Resource>& expectedOutputs, TestLog&)
1319 {
1320         if (outputAllocs.size() != 1)
1321                 return false;
1322
1323         const BufferSp&                 expectedOutput                  = expectedOutputs[0].getBuffer();
1324         std::vector<deUint8>    data;
1325         expectedOutput->getBytes(data);
1326
1327         const float* const              expectedOutputAsFloat   = reinterpret_cast<const float*>(&data.front());
1328         const float* const              outputAsFloat                   = static_cast<const float*>(outputAllocs[0]->getHostPtr());
1329
1330         for (size_t idx = 0; idx < expectedOutput->getByteSize() / sizeof(float); ++idx)
1331         {
1332                 const float f0 = expectedOutputAsFloat[idx];
1333                 const float f1 = outputAsFloat[idx];
1334
1335                 // For NMax, NaN is considered acceptable result, since in
1336                 // architectures that do not handle NaNs, those are huge values.
1337                 if (!tcu::Float32(f1).isNaN() && deFloatAbs(f1 - f0) > 0.00001f)
1338                         return false;
1339         }
1340
1341         return true;
1342 }
1343
1344 tcu::TestCaseGroup* createOpNMaxGroup (tcu::TestContext& testCtx)
1345 {
1346         de::MovePtr<tcu::TestCaseGroup> group(new tcu::TestCaseGroup(testCtx, "opnmax", "Test the OpNMax instruction"));
1347         ComputeShaderSpec                               spec;
1348         de::Random                                              rnd                             (deStringHash(group->getName()));
1349         const int                                               numElements             = 200;
1350         vector<float>                                   inputFloats1    (numElements, 0);
1351         vector<float>                                   inputFloats2    (numElements, 0);
1352         vector<float>                                   outputFloats    (numElements, 0);
1353
1354         fillRandomScalars(rnd, -10000.f, 10000.f, &inputFloats1[0], numElements);
1355         fillRandomScalars(rnd, -10000.f, 10000.f, &inputFloats2[0], numElements);
1356
1357         // Make the first case a full-NAN case.
1358         inputFloats1[0] = TCU_NAN;
1359         inputFloats2[0] = TCU_NAN;
1360
1361         for (size_t ndx = 0; ndx < numElements; ++ndx)
1362         {
1363                 // By default, pick the biggest
1364                 outputFloats[ndx] = std::max(inputFloats1[ndx], inputFloats2[ndx]);
1365
1366                 // Make half of the cases NaN cases
1367                 if ((ndx & 1) == 0)
1368                 {
1369                         // Alternate between the NaN operand
1370                         if ((ndx & 2) == 0)
1371                         {
1372                                 outputFloats[ndx] = inputFloats2[ndx];
1373                                 inputFloats1[ndx] = TCU_NAN;
1374                         }
1375                         else
1376                         {
1377                                 outputFloats[ndx] = inputFloats1[ndx];
1378                                 inputFloats2[ndx] = TCU_NAN;
1379                         }
1380                 }
1381         }
1382
1383         spec.assembly =
1384                 "OpCapability Shader\n"
1385                 "%std450        = OpExtInstImport \"GLSL.std.450\"\n"
1386                 "OpMemoryModel Logical GLSL450\n"
1387                 "OpEntryPoint GLCompute %main \"main\" %id\n"
1388                 "OpExecutionMode %main LocalSize 1 1 1\n"
1389
1390                 "OpName %main           \"main\"\n"
1391                 "OpName %id             \"gl_GlobalInvocationID\"\n"
1392
1393                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
1394
1395                 "OpDecorate %buf BufferBlock\n"
1396                 "OpDecorate %indata1 DescriptorSet 0\n"
1397                 "OpDecorate %indata1 Binding 0\n"
1398                 "OpDecorate %indata2 DescriptorSet 0\n"
1399                 "OpDecorate %indata2 Binding 1\n"
1400                 "OpDecorate %outdata DescriptorSet 0\n"
1401                 "OpDecorate %outdata Binding 2\n"
1402                 "OpDecorate %f32arr ArrayStride 4\n"
1403                 "OpMemberDecorate %buf 0 Offset 0\n"
1404
1405                 + string(getComputeAsmCommonTypes()) +
1406
1407                 "%buf        = OpTypeStruct %f32arr\n"
1408                 "%bufptr     = OpTypePointer Uniform %buf\n"
1409                 "%indata1    = OpVariable %bufptr Uniform\n"
1410                 "%indata2    = OpVariable %bufptr Uniform\n"
1411                 "%outdata    = OpVariable %bufptr Uniform\n"
1412
1413                 "%id        = OpVariable %uvec3ptr Input\n"
1414                 "%zero      = OpConstant %i32 0\n"
1415
1416                 "%main      = OpFunction %void None %voidf\n"
1417                 "%label     = OpLabel\n"
1418                 "%idval     = OpLoad %uvec3 %id\n"
1419                 "%x         = OpCompositeExtract %u32 %idval 0\n"
1420                 "%inloc1    = OpAccessChain %f32ptr %indata1 %zero %x\n"
1421                 "%inval1    = OpLoad %f32 %inloc1\n"
1422                 "%inloc2    = OpAccessChain %f32ptr %indata2 %zero %x\n"
1423                 "%inval2    = OpLoad %f32 %inloc2\n"
1424                 "%rem       = OpExtInst %f32 %std450 NMax %inval1 %inval2\n"
1425                 "%outloc    = OpAccessChain %f32ptr %outdata %zero %x\n"
1426                 "             OpStore %outloc %rem\n"
1427                 "             OpReturn\n"
1428                 "             OpFunctionEnd\n";
1429
1430         spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats1)));
1431         spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats2)));
1432         spec.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
1433         spec.numWorkGroups = IVec3(numElements, 1, 1);
1434         spec.verifyIO = &compareNMax;
1435
1436         group->addChild(new SpvAsmComputeShaderCase(testCtx, "all", "", spec));
1437
1438         return group.release();
1439 }
1440
1441 bool compareNClamp (const std::vector<Resource>&, const vector<AllocationSp>& outputAllocs, const std::vector<Resource>& expectedOutputs, TestLog&)
1442 {
1443         if (outputAllocs.size() != 1)
1444                 return false;
1445
1446         const BufferSp&                 expectedOutput                  = expectedOutputs[0].getBuffer();
1447         std::vector<deUint8>    data;
1448         expectedOutput->getBytes(data);
1449
1450         const float* const              expectedOutputAsFloat   = reinterpret_cast<const float*>(&data.front());
1451         const float* const              outputAsFloat                   = static_cast<const float*>(outputAllocs[0]->getHostPtr());
1452
1453         for (size_t idx = 0; idx < expectedOutput->getByteSize() / sizeof(float) / 2; ++idx)
1454         {
1455                 const float e0 = expectedOutputAsFloat[idx * 2];
1456                 const float e1 = expectedOutputAsFloat[idx * 2 + 1];
1457                 const float res = outputAsFloat[idx];
1458
1459                 // For NClamp, we have two possible outcomes based on
1460                 // whether NaNs are handled or not.
1461                 // If either min or max value is NaN, the result is undefined,
1462                 // so this test doesn't stress those. If the clamped value is
1463                 // NaN, and NaNs are handled, the result is min; if NaNs are not
1464                 // handled, they are big values that result in max.
1465                 // If all three parameters are NaN, the result should be NaN.
1466                 if (!((tcu::Float32(e0).isNaN() && tcu::Float32(res).isNaN()) ||
1467                          (deFloatAbs(e0 - res) < 0.00001f) ||
1468                          (deFloatAbs(e1 - res) < 0.00001f)))
1469                         return false;
1470         }
1471
1472         return true;
1473 }
1474
1475 tcu::TestCaseGroup* createOpNClampGroup (tcu::TestContext& testCtx)
1476 {
1477         de::MovePtr<tcu::TestCaseGroup> group                   (new tcu::TestCaseGroup(testCtx, "opnclamp", "Test the OpNClamp instruction"));
1478         ComputeShaderSpec                               spec;
1479         de::Random                                              rnd                             (deStringHash(group->getName()));
1480         const int                                               numElements             = 200;
1481         vector<float>                                   inputFloats1    (numElements, 0);
1482         vector<float>                                   inputFloats2    (numElements, 0);
1483         vector<float>                                   inputFloats3    (numElements, 0);
1484         vector<float>                                   outputFloats    (numElements * 2, 0);
1485
1486         fillRandomScalars(rnd, -10000.f, 10000.f, &inputFloats1[0], numElements);
1487         fillRandomScalars(rnd, -10000.f, 10000.f, &inputFloats2[0], numElements);
1488         fillRandomScalars(rnd, -10000.f, 10000.f, &inputFloats3[0], numElements);
1489
1490         for (size_t ndx = 0; ndx < numElements; ++ndx)
1491         {
1492                 // Results are only defined if max value is bigger than min value.
1493                 if (inputFloats2[ndx] > inputFloats3[ndx])
1494                 {
1495                         float t = inputFloats2[ndx];
1496                         inputFloats2[ndx] = inputFloats3[ndx];
1497                         inputFloats3[ndx] = t;
1498                 }
1499
1500                 // By default, do the clamp, setting both possible answers
1501                 float defaultRes = std::min(std::max(inputFloats1[ndx], inputFloats2[ndx]), inputFloats3[ndx]);
1502
1503                 float maxResA = std::max(inputFloats1[ndx], inputFloats2[ndx]);
1504                 float maxResB = maxResA;
1505
1506                 // Alternate between the NaN cases
1507                 if (ndx & 1)
1508                 {
1509                         inputFloats1[ndx] = TCU_NAN;
1510                         // If NaN is handled, the result should be same as the clamp minimum.
1511                         // If NaN is not handled, the result should clamp to the clamp maximum.
1512                         maxResA = inputFloats2[ndx];
1513                         maxResB = inputFloats3[ndx];
1514                 }
1515                 else
1516                 {
1517                         // Not a NaN case - only one legal result.
1518                         maxResA = defaultRes;
1519                         maxResB = defaultRes;
1520                 }
1521
1522                 outputFloats[ndx * 2] = maxResA;
1523                 outputFloats[ndx * 2 + 1] = maxResB;
1524         }
1525
1526         // Make the first case a full-NAN case.
1527         inputFloats1[0] = TCU_NAN;
1528         inputFloats2[0] = TCU_NAN;
1529         inputFloats3[0] = TCU_NAN;
1530         outputFloats[0] = TCU_NAN;
1531         outputFloats[1] = TCU_NAN;
1532
1533         spec.assembly =
1534                 "OpCapability Shader\n"
1535                 "%std450        = OpExtInstImport \"GLSL.std.450\"\n"
1536                 "OpMemoryModel Logical GLSL450\n"
1537                 "OpEntryPoint GLCompute %main \"main\" %id\n"
1538                 "OpExecutionMode %main LocalSize 1 1 1\n"
1539
1540                 "OpName %main           \"main\"\n"
1541                 "OpName %id             \"gl_GlobalInvocationID\"\n"
1542
1543                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
1544
1545                 "OpDecorate %buf BufferBlock\n"
1546                 "OpDecorate %indata1 DescriptorSet 0\n"
1547                 "OpDecorate %indata1 Binding 0\n"
1548                 "OpDecorate %indata2 DescriptorSet 0\n"
1549                 "OpDecorate %indata2 Binding 1\n"
1550                 "OpDecorate %indata3 DescriptorSet 0\n"
1551                 "OpDecorate %indata3 Binding 2\n"
1552                 "OpDecorate %outdata DescriptorSet 0\n"
1553                 "OpDecorate %outdata Binding 3\n"
1554                 "OpDecorate %f32arr ArrayStride 4\n"
1555                 "OpMemberDecorate %buf 0 Offset 0\n"
1556
1557                 + string(getComputeAsmCommonTypes()) +
1558
1559                 "%buf        = OpTypeStruct %f32arr\n"
1560                 "%bufptr     = OpTypePointer Uniform %buf\n"
1561                 "%indata1    = OpVariable %bufptr Uniform\n"
1562                 "%indata2    = OpVariable %bufptr Uniform\n"
1563                 "%indata3    = OpVariable %bufptr Uniform\n"
1564                 "%outdata    = OpVariable %bufptr Uniform\n"
1565
1566                 "%id        = OpVariable %uvec3ptr Input\n"
1567                 "%zero      = OpConstant %i32 0\n"
1568
1569                 "%main      = OpFunction %void None %voidf\n"
1570                 "%label     = OpLabel\n"
1571                 "%idval     = OpLoad %uvec3 %id\n"
1572                 "%x         = OpCompositeExtract %u32 %idval 0\n"
1573                 "%inloc1    = OpAccessChain %f32ptr %indata1 %zero %x\n"
1574                 "%inval1    = OpLoad %f32 %inloc1\n"
1575                 "%inloc2    = OpAccessChain %f32ptr %indata2 %zero %x\n"
1576                 "%inval2    = OpLoad %f32 %inloc2\n"
1577                 "%inloc3    = OpAccessChain %f32ptr %indata3 %zero %x\n"
1578                 "%inval3    = OpLoad %f32 %inloc3\n"
1579                 "%rem       = OpExtInst %f32 %std450 NClamp %inval1 %inval2 %inval3\n"
1580                 "%outloc    = OpAccessChain %f32ptr %outdata %zero %x\n"
1581                 "             OpStore %outloc %rem\n"
1582                 "             OpReturn\n"
1583                 "             OpFunctionEnd\n";
1584
1585         spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats1)));
1586         spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats2)));
1587         spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats3)));
1588         spec.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
1589         spec.numWorkGroups = IVec3(numElements, 1, 1);
1590         spec.verifyIO = &compareNClamp;
1591
1592         group->addChild(new SpvAsmComputeShaderCase(testCtx, "all", "", spec));
1593
1594         return group.release();
1595 }
1596
1597 tcu::TestCaseGroup* createOpSRemComputeGroup (tcu::TestContext& testCtx, qpTestResult negFailResult)
1598 {
1599         de::MovePtr<tcu::TestCaseGroup> group                   (new tcu::TestCaseGroup(testCtx, "opsrem", "Test the OpSRem instruction"));
1600         de::Random                                              rnd                             (deStringHash(group->getName()));
1601         const int                                               numElements             = 200;
1602
1603         const struct CaseParams
1604         {
1605                 const char*             name;
1606                 const char*             failMessage;            // customized status message
1607                 qpTestResult    failResult;                     // override status on failure
1608                 int                             op1Min, op1Max;         // operand ranges
1609                 int                             op2Min, op2Max;
1610         } cases[] =
1611         {
1612                 { "positive",   "Output doesn't match with expected",                           QP_TEST_RESULT_FAIL,    0,              65536,  0,              100 },
1613                 { "all",                "Inconsistent results, but within specification",       negFailResult,                  -65536, 65536,  -100,   100 },  // see below
1614         };
1615         // If either operand is negative the result is undefined. Some implementations may still return correct values.
1616
1617         for (int caseNdx = 0; caseNdx < DE_LENGTH_OF_ARRAY(cases); ++caseNdx)
1618         {
1619                 const CaseParams&       params          = cases[caseNdx];
1620                 ComputeShaderSpec       spec;
1621                 vector<deInt32>         inputInts1      (numElements, 0);
1622                 vector<deInt32>         inputInts2      (numElements, 0);
1623                 vector<deInt32>         outputInts      (numElements, 0);
1624
1625                 fillRandomScalars(rnd, params.op1Min, params.op1Max, &inputInts1[0], numElements);
1626                 fillRandomScalars(rnd, params.op2Min, params.op2Max, &inputInts2[0], numElements, filterNotZero);
1627
1628                 for (int ndx = 0; ndx < numElements; ++ndx)
1629                 {
1630                         // The return value of std::fmod() has the same sign as its first operand, which is how OpFRem spec'd.
1631                         outputInts[ndx] = inputInts1[ndx] % inputInts2[ndx];
1632                 }
1633
1634                 spec.assembly =
1635                         string(getComputeAsmShaderPreamble()) +
1636
1637                         "OpName %main           \"main\"\n"
1638                         "OpName %id             \"gl_GlobalInvocationID\"\n"
1639
1640                         "OpDecorate %id BuiltIn GlobalInvocationId\n"
1641
1642                         "OpDecorate %buf BufferBlock\n"
1643                         "OpDecorate %indata1 DescriptorSet 0\n"
1644                         "OpDecorate %indata1 Binding 0\n"
1645                         "OpDecorate %indata2 DescriptorSet 0\n"
1646                         "OpDecorate %indata2 Binding 1\n"
1647                         "OpDecorate %outdata DescriptorSet 0\n"
1648                         "OpDecorate %outdata Binding 2\n"
1649                         "OpDecorate %i32arr ArrayStride 4\n"
1650                         "OpMemberDecorate %buf 0 Offset 0\n"
1651
1652                         + string(getComputeAsmCommonTypes()) +
1653
1654                         "%buf        = OpTypeStruct %i32arr\n"
1655                         "%bufptr     = OpTypePointer Uniform %buf\n"
1656                         "%indata1    = OpVariable %bufptr Uniform\n"
1657                         "%indata2    = OpVariable %bufptr Uniform\n"
1658                         "%outdata    = OpVariable %bufptr Uniform\n"
1659
1660                         "%id        = OpVariable %uvec3ptr Input\n"
1661                         "%zero      = OpConstant %i32 0\n"
1662
1663                         "%main      = OpFunction %void None %voidf\n"
1664                         "%label     = OpLabel\n"
1665                         "%idval     = OpLoad %uvec3 %id\n"
1666                         "%x         = OpCompositeExtract %u32 %idval 0\n"
1667                         "%inloc1    = OpAccessChain %i32ptr %indata1 %zero %x\n"
1668                         "%inval1    = OpLoad %i32 %inloc1\n"
1669                         "%inloc2    = OpAccessChain %i32ptr %indata2 %zero %x\n"
1670                         "%inval2    = OpLoad %i32 %inloc2\n"
1671                         "%rem       = OpSRem %i32 %inval1 %inval2\n"
1672                         "%outloc    = OpAccessChain %i32ptr %outdata %zero %x\n"
1673                         "             OpStore %outloc %rem\n"
1674                         "             OpReturn\n"
1675                         "             OpFunctionEnd\n";
1676
1677                 spec.inputs.push_back   (BufferSp(new Int32Buffer(inputInts1)));
1678                 spec.inputs.push_back   (BufferSp(new Int32Buffer(inputInts2)));
1679                 spec.outputs.push_back  (BufferSp(new Int32Buffer(outputInts)));
1680                 spec.numWorkGroups              = IVec3(numElements, 1, 1);
1681                 spec.failResult                 = params.failResult;
1682                 spec.failMessage                = params.failMessage;
1683
1684                 group->addChild(new SpvAsmComputeShaderCase(testCtx, params.name, "", spec));
1685         }
1686
1687         return group.release();
1688 }
1689
1690 tcu::TestCaseGroup* createOpSRemComputeGroup64 (tcu::TestContext& testCtx, qpTestResult negFailResult)
1691 {
1692         de::MovePtr<tcu::TestCaseGroup> group                   (new tcu::TestCaseGroup(testCtx, "opsrem64", "Test the 64-bit OpSRem instruction"));
1693         de::Random                                              rnd                             (deStringHash(group->getName()));
1694         const int                                               numElements             = 200;
1695
1696         const struct CaseParams
1697         {
1698                 const char*             name;
1699                 const char*             failMessage;            // customized status message
1700                 qpTestResult    failResult;                     // override status on failure
1701                 bool                    positive;
1702         } cases[] =
1703         {
1704                 { "positive",   "Output doesn't match with expected",                           QP_TEST_RESULT_FAIL,    true },
1705                 { "all",                "Inconsistent results, but within specification",       negFailResult,                  false },        // see below
1706         };
1707         // If either operand is negative the result is undefined. Some implementations may still return correct values.
1708
1709         for (int caseNdx = 0; caseNdx < DE_LENGTH_OF_ARRAY(cases); ++caseNdx)
1710         {
1711                 const CaseParams&       params          = cases[caseNdx];
1712                 ComputeShaderSpec       spec;
1713                 vector<deInt64>         inputInts1      (numElements, 0);
1714                 vector<deInt64>         inputInts2      (numElements, 0);
1715                 vector<deInt64>         outputInts      (numElements, 0);
1716
1717                 if (params.positive)
1718                 {
1719                         fillRandomInt64sLogDistributed(rnd, inputInts1, numElements, filterNonNegative);
1720                         fillRandomInt64sLogDistributed(rnd, inputInts2, numElements, filterPositive);
1721                 }
1722                 else
1723                 {
1724                         fillRandomInt64sLogDistributed(rnd, inputInts1, numElements);
1725                         fillRandomInt64sLogDistributed(rnd, inputInts2, numElements, filterNotZero);
1726                 }
1727
1728                 for (int ndx = 0; ndx < numElements; ++ndx)
1729                 {
1730                         // The return value of std::fmod() has the same sign as its first operand, which is how OpFRem spec'd.
1731                         outputInts[ndx] = inputInts1[ndx] % inputInts2[ndx];
1732                 }
1733
1734                 spec.assembly =
1735                         "OpCapability Int64\n"
1736
1737                         + string(getComputeAsmShaderPreamble()) +
1738
1739                         "OpName %main           \"main\"\n"
1740                         "OpName %id             \"gl_GlobalInvocationID\"\n"
1741
1742                         "OpDecorate %id BuiltIn GlobalInvocationId\n"
1743
1744                         "OpDecorate %buf BufferBlock\n"
1745                         "OpDecorate %indata1 DescriptorSet 0\n"
1746                         "OpDecorate %indata1 Binding 0\n"
1747                         "OpDecorate %indata2 DescriptorSet 0\n"
1748                         "OpDecorate %indata2 Binding 1\n"
1749                         "OpDecorate %outdata DescriptorSet 0\n"
1750                         "OpDecorate %outdata Binding 2\n"
1751                         "OpDecorate %i64arr ArrayStride 8\n"
1752                         "OpMemberDecorate %buf 0 Offset 0\n"
1753
1754                         + string(getComputeAsmCommonTypes())
1755                         + string(getComputeAsmCommonInt64Types()) +
1756
1757                         "%buf        = OpTypeStruct %i64arr\n"
1758                         "%bufptr     = OpTypePointer Uniform %buf\n"
1759                         "%indata1    = OpVariable %bufptr Uniform\n"
1760                         "%indata2    = OpVariable %bufptr Uniform\n"
1761                         "%outdata    = OpVariable %bufptr Uniform\n"
1762
1763                         "%id        = OpVariable %uvec3ptr Input\n"
1764                         "%zero      = OpConstant %i64 0\n"
1765
1766                         "%main      = OpFunction %void None %voidf\n"
1767                         "%label     = OpLabel\n"
1768                         "%idval     = OpLoad %uvec3 %id\n"
1769                         "%x         = OpCompositeExtract %u32 %idval 0\n"
1770                         "%inloc1    = OpAccessChain %i64ptr %indata1 %zero %x\n"
1771                         "%inval1    = OpLoad %i64 %inloc1\n"
1772                         "%inloc2    = OpAccessChain %i64ptr %indata2 %zero %x\n"
1773                         "%inval2    = OpLoad %i64 %inloc2\n"
1774                         "%rem       = OpSRem %i64 %inval1 %inval2\n"
1775                         "%outloc    = OpAccessChain %i64ptr %outdata %zero %x\n"
1776                         "             OpStore %outloc %rem\n"
1777                         "             OpReturn\n"
1778                         "             OpFunctionEnd\n";
1779
1780                 spec.inputs.push_back   (BufferSp(new Int64Buffer(inputInts1)));
1781                 spec.inputs.push_back   (BufferSp(new Int64Buffer(inputInts2)));
1782                 spec.outputs.push_back  (BufferSp(new Int64Buffer(outputInts)));
1783                 spec.numWorkGroups              = IVec3(numElements, 1, 1);
1784                 spec.failResult                 = params.failResult;
1785                 spec.failMessage                = params.failMessage;
1786
1787                 spec.requestedVulkanFeatures.coreFeatures.shaderInt64 = VK_TRUE;
1788
1789                 group->addChild(new SpvAsmComputeShaderCase(testCtx, params.name, "", spec));
1790         }
1791
1792         return group.release();
1793 }
1794
1795 tcu::TestCaseGroup* createOpSModComputeGroup (tcu::TestContext& testCtx, qpTestResult negFailResult)
1796 {
1797         de::MovePtr<tcu::TestCaseGroup> group                   (new tcu::TestCaseGroup(testCtx, "opsmod", "Test the OpSMod instruction"));
1798         de::Random                                              rnd                             (deStringHash(group->getName()));
1799         const int                                               numElements             = 200;
1800
1801         const struct CaseParams
1802         {
1803                 const char*             name;
1804                 const char*             failMessage;            // customized status message
1805                 qpTestResult    failResult;                     // override status on failure
1806                 int                             op1Min, op1Max;         // operand ranges
1807                 int                             op2Min, op2Max;
1808         } cases[] =
1809         {
1810                 { "positive",   "Output doesn't match with expected",                           QP_TEST_RESULT_FAIL,    0,              65536,  0,              100 },
1811                 { "all",                "Inconsistent results, but within specification",       negFailResult,                  -65536, 65536,  -100,   100 },  // see below
1812         };
1813         // If either operand is negative the result is undefined. Some implementations may still return correct values.
1814
1815         for (int caseNdx = 0; caseNdx < DE_LENGTH_OF_ARRAY(cases); ++caseNdx)
1816         {
1817                 const CaseParams&       params          = cases[caseNdx];
1818
1819                 ComputeShaderSpec       spec;
1820                 vector<deInt32>         inputInts1      (numElements, 0);
1821                 vector<deInt32>         inputInts2      (numElements, 0);
1822                 vector<deInt32>         outputInts      (numElements, 0);
1823
1824                 fillRandomScalars(rnd, params.op1Min, params.op1Max, &inputInts1[0], numElements);
1825                 fillRandomScalars(rnd, params.op2Min, params.op2Max, &inputInts2[0], numElements, filterNotZero);
1826
1827                 for (int ndx = 0; ndx < numElements; ++ndx)
1828                 {
1829                         deInt32 rem = inputInts1[ndx] % inputInts2[ndx];
1830                         if (rem == 0)
1831                         {
1832                                 outputInts[ndx] = 0;
1833                         }
1834                         else if ((inputInts1[ndx] >= 0) == (inputInts2[ndx] >= 0))
1835                         {
1836                                 // They have the same sign
1837                                 outputInts[ndx] = rem;
1838                         }
1839                         else
1840                         {
1841                                 // They have opposite sign.  The remainder operation takes the
1842                                 // sign inputInts1[ndx] but OpSMod is supposed to take ths sign
1843                                 // of inputInts2[ndx].  Adding inputInts2[ndx] will ensure that
1844                                 // the result has the correct sign and that it is still
1845                                 // congruent to inputInts1[ndx] modulo inputInts2[ndx]
1846                                 //
1847                                 // See also http://mathforum.org/library/drmath/view/52343.html
1848                                 outputInts[ndx] = rem + inputInts2[ndx];
1849                         }
1850                 }
1851
1852                 spec.assembly =
1853                         string(getComputeAsmShaderPreamble()) +
1854
1855                         "OpName %main           \"main\"\n"
1856                         "OpName %id             \"gl_GlobalInvocationID\"\n"
1857
1858                         "OpDecorate %id BuiltIn GlobalInvocationId\n"
1859
1860                         "OpDecorate %buf BufferBlock\n"
1861                         "OpDecorate %indata1 DescriptorSet 0\n"
1862                         "OpDecorate %indata1 Binding 0\n"
1863                         "OpDecorate %indata2 DescriptorSet 0\n"
1864                         "OpDecorate %indata2 Binding 1\n"
1865                         "OpDecorate %outdata DescriptorSet 0\n"
1866                         "OpDecorate %outdata Binding 2\n"
1867                         "OpDecorate %i32arr ArrayStride 4\n"
1868                         "OpMemberDecorate %buf 0 Offset 0\n"
1869
1870                         + string(getComputeAsmCommonTypes()) +
1871
1872                         "%buf        = OpTypeStruct %i32arr\n"
1873                         "%bufptr     = OpTypePointer Uniform %buf\n"
1874                         "%indata1    = OpVariable %bufptr Uniform\n"
1875                         "%indata2    = OpVariable %bufptr Uniform\n"
1876                         "%outdata    = OpVariable %bufptr Uniform\n"
1877
1878                         "%id        = OpVariable %uvec3ptr Input\n"
1879                         "%zero      = OpConstant %i32 0\n"
1880
1881                         "%main      = OpFunction %void None %voidf\n"
1882                         "%label     = OpLabel\n"
1883                         "%idval     = OpLoad %uvec3 %id\n"
1884                         "%x         = OpCompositeExtract %u32 %idval 0\n"
1885                         "%inloc1    = OpAccessChain %i32ptr %indata1 %zero %x\n"
1886                         "%inval1    = OpLoad %i32 %inloc1\n"
1887                         "%inloc2    = OpAccessChain %i32ptr %indata2 %zero %x\n"
1888                         "%inval2    = OpLoad %i32 %inloc2\n"
1889                         "%rem       = OpSMod %i32 %inval1 %inval2\n"
1890                         "%outloc    = OpAccessChain %i32ptr %outdata %zero %x\n"
1891                         "             OpStore %outloc %rem\n"
1892                         "             OpReturn\n"
1893                         "             OpFunctionEnd\n";
1894
1895                 spec.inputs.push_back   (BufferSp(new Int32Buffer(inputInts1)));
1896                 spec.inputs.push_back   (BufferSp(new Int32Buffer(inputInts2)));
1897                 spec.outputs.push_back  (BufferSp(new Int32Buffer(outputInts)));
1898                 spec.numWorkGroups              = IVec3(numElements, 1, 1);
1899                 spec.failResult                 = params.failResult;
1900                 spec.failMessage                = params.failMessage;
1901
1902                 group->addChild(new SpvAsmComputeShaderCase(testCtx, params.name, "", spec));
1903         }
1904
1905         return group.release();
1906 }
1907
1908 tcu::TestCaseGroup* createOpSModComputeGroup64 (tcu::TestContext& testCtx, qpTestResult negFailResult)
1909 {
1910         de::MovePtr<tcu::TestCaseGroup> group                   (new tcu::TestCaseGroup(testCtx, "opsmod64", "Test the OpSMod instruction"));
1911         de::Random                                              rnd                             (deStringHash(group->getName()));
1912         const int                                               numElements             = 200;
1913
1914         const struct CaseParams
1915         {
1916                 const char*             name;
1917                 const char*             failMessage;            // customized status message
1918                 qpTestResult    failResult;                     // override status on failure
1919                 bool                    positive;
1920         } cases[] =
1921         {
1922                 { "positive",   "Output doesn't match with expected",                           QP_TEST_RESULT_FAIL,    true },
1923                 { "all",                "Inconsistent results, but within specification",       negFailResult,                  false },        // see below
1924         };
1925         // If either operand is negative the result is undefined. Some implementations may still return correct values.
1926
1927         for (int caseNdx = 0; caseNdx < DE_LENGTH_OF_ARRAY(cases); ++caseNdx)
1928         {
1929                 const CaseParams&       params          = cases[caseNdx];
1930
1931                 ComputeShaderSpec       spec;
1932                 vector<deInt64>         inputInts1      (numElements, 0);
1933                 vector<deInt64>         inputInts2      (numElements, 0);
1934                 vector<deInt64>         outputInts      (numElements, 0);
1935
1936
1937                 if (params.positive)
1938                 {
1939                         fillRandomInt64sLogDistributed(rnd, inputInts1, numElements, filterNonNegative);
1940                         fillRandomInt64sLogDistributed(rnd, inputInts2, numElements, filterPositive);
1941                 }
1942                 else
1943                 {
1944                         fillRandomInt64sLogDistributed(rnd, inputInts1, numElements);
1945                         fillRandomInt64sLogDistributed(rnd, inputInts2, numElements, filterNotZero);
1946                 }
1947
1948                 for (int ndx = 0; ndx < numElements; ++ndx)
1949                 {
1950                         deInt64 rem = inputInts1[ndx] % inputInts2[ndx];
1951                         if (rem == 0)
1952                         {
1953                                 outputInts[ndx] = 0;
1954                         }
1955                         else if ((inputInts1[ndx] >= 0) == (inputInts2[ndx] >= 0))
1956                         {
1957                                 // They have the same sign
1958                                 outputInts[ndx] = rem;
1959                         }
1960                         else
1961                         {
1962                                 // They have opposite sign.  The remainder operation takes the
1963                                 // sign inputInts1[ndx] but OpSMod is supposed to take ths sign
1964                                 // of inputInts2[ndx].  Adding inputInts2[ndx] will ensure that
1965                                 // the result has the correct sign and that it is still
1966                                 // congruent to inputInts1[ndx] modulo inputInts2[ndx]
1967                                 //
1968                                 // See also http://mathforum.org/library/drmath/view/52343.html
1969                                 outputInts[ndx] = rem + inputInts2[ndx];
1970                         }
1971                 }
1972
1973                 spec.assembly =
1974                         "OpCapability Int64\n"
1975
1976                         + string(getComputeAsmShaderPreamble()) +
1977
1978                         "OpName %main           \"main\"\n"
1979                         "OpName %id             \"gl_GlobalInvocationID\"\n"
1980
1981                         "OpDecorate %id BuiltIn GlobalInvocationId\n"
1982
1983                         "OpDecorate %buf BufferBlock\n"
1984                         "OpDecorate %indata1 DescriptorSet 0\n"
1985                         "OpDecorate %indata1 Binding 0\n"
1986                         "OpDecorate %indata2 DescriptorSet 0\n"
1987                         "OpDecorate %indata2 Binding 1\n"
1988                         "OpDecorate %outdata DescriptorSet 0\n"
1989                         "OpDecorate %outdata Binding 2\n"
1990                         "OpDecorate %i64arr ArrayStride 8\n"
1991                         "OpMemberDecorate %buf 0 Offset 0\n"
1992
1993                         + string(getComputeAsmCommonTypes())
1994                         + string(getComputeAsmCommonInt64Types()) +
1995
1996                         "%buf        = OpTypeStruct %i64arr\n"
1997                         "%bufptr     = OpTypePointer Uniform %buf\n"
1998                         "%indata1    = OpVariable %bufptr Uniform\n"
1999                         "%indata2    = OpVariable %bufptr Uniform\n"
2000                         "%outdata    = OpVariable %bufptr Uniform\n"
2001
2002                         "%id        = OpVariable %uvec3ptr Input\n"
2003                         "%zero      = OpConstant %i64 0\n"
2004
2005                         "%main      = OpFunction %void None %voidf\n"
2006                         "%label     = OpLabel\n"
2007                         "%idval     = OpLoad %uvec3 %id\n"
2008                         "%x         = OpCompositeExtract %u32 %idval 0\n"
2009                         "%inloc1    = OpAccessChain %i64ptr %indata1 %zero %x\n"
2010                         "%inval1    = OpLoad %i64 %inloc1\n"
2011                         "%inloc2    = OpAccessChain %i64ptr %indata2 %zero %x\n"
2012                         "%inval2    = OpLoad %i64 %inloc2\n"
2013                         "%rem       = OpSMod %i64 %inval1 %inval2\n"
2014                         "%outloc    = OpAccessChain %i64ptr %outdata %zero %x\n"
2015                         "             OpStore %outloc %rem\n"
2016                         "             OpReturn\n"
2017                         "             OpFunctionEnd\n";
2018
2019                 spec.inputs.push_back   (BufferSp(new Int64Buffer(inputInts1)));
2020                 spec.inputs.push_back   (BufferSp(new Int64Buffer(inputInts2)));
2021                 spec.outputs.push_back  (BufferSp(new Int64Buffer(outputInts)));
2022                 spec.numWorkGroups              = IVec3(numElements, 1, 1);
2023                 spec.failResult                 = params.failResult;
2024                 spec.failMessage                = params.failMessage;
2025
2026                 spec.requestedVulkanFeatures.coreFeatures.shaderInt64 = VK_TRUE;
2027
2028                 group->addChild(new SpvAsmComputeShaderCase(testCtx, params.name, "", spec));
2029         }
2030
2031         return group.release();
2032 }
2033
2034 // Copy contents in the input buffer to the output buffer.
2035 tcu::TestCaseGroup* createOpCopyMemoryGroup (tcu::TestContext& testCtx)
2036 {
2037         de::MovePtr<tcu::TestCaseGroup> group                   (new tcu::TestCaseGroup(testCtx, "opcopymemory", "Test the OpCopyMemory instruction"));
2038         de::Random                                              rnd                             (deStringHash(group->getName()));
2039         const int                                               numElements             = 100;
2040
2041         // The following case adds vec4(0., 0.5, 1.5, 2.5) to each of the elements in the input buffer and writes output to the output buffer.
2042         ComputeShaderSpec                               spec1;
2043         vector<Vec4>                                    inputFloats1    (numElements);
2044         vector<Vec4>                                    outputFloats1   (numElements);
2045
2046         fillRandomScalars(rnd, -200.f, 200.f, &inputFloats1[0], numElements * 4);
2047
2048         // CPU might not use the same rounding mode as the GPU. Use whole numbers to avoid rounding differences.
2049         floorAll(inputFloats1);
2050
2051         for (size_t ndx = 0; ndx < numElements; ++ndx)
2052                 outputFloats1[ndx] = inputFloats1[ndx] + Vec4(0.f, 0.5f, 1.5f, 2.5f);
2053
2054         spec1.assembly =
2055                 string(getComputeAsmShaderPreamble()) +
2056
2057                 "OpName %main           \"main\"\n"
2058                 "OpName %id             \"gl_GlobalInvocationID\"\n"
2059
2060                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
2061                 "OpDecorate %vec4arr ArrayStride 16\n"
2062
2063                 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) +
2064
2065                 "%vec4       = OpTypeVector %f32 4\n"
2066                 "%vec4ptr_u  = OpTypePointer Uniform %vec4\n"
2067                 "%vec4ptr_f  = OpTypePointer Function %vec4\n"
2068                 "%vec4arr    = OpTypeRuntimeArray %vec4\n"
2069                 "%buf        = OpTypeStruct %vec4arr\n"
2070                 "%bufptr     = OpTypePointer Uniform %buf\n"
2071                 "%indata     = OpVariable %bufptr Uniform\n"
2072                 "%outdata    = OpVariable %bufptr Uniform\n"
2073
2074                 "%id         = OpVariable %uvec3ptr Input\n"
2075                 "%zero       = OpConstant %i32 0\n"
2076                 "%c_f_0      = OpConstant %f32 0.\n"
2077                 "%c_f_0_5    = OpConstant %f32 0.5\n"
2078                 "%c_f_1_5    = OpConstant %f32 1.5\n"
2079                 "%c_f_2_5    = OpConstant %f32 2.5\n"
2080                 "%c_vec4     = OpConstantComposite %vec4 %c_f_0 %c_f_0_5 %c_f_1_5 %c_f_2_5\n"
2081
2082                 "%main       = OpFunction %void None %voidf\n"
2083                 "%label      = OpLabel\n"
2084                 "%v_vec4     = OpVariable %vec4ptr_f Function\n"
2085                 "%idval      = OpLoad %uvec3 %id\n"
2086                 "%x          = OpCompositeExtract %u32 %idval 0\n"
2087                 "%inloc      = OpAccessChain %vec4ptr_u %indata %zero %x\n"
2088                 "%outloc     = OpAccessChain %vec4ptr_u %outdata %zero %x\n"
2089                 "              OpCopyMemory %v_vec4 %inloc\n"
2090                 "%v_vec4_val = OpLoad %vec4 %v_vec4\n"
2091                 "%add        = OpFAdd %vec4 %v_vec4_val %c_vec4\n"
2092                 "              OpStore %outloc %add\n"
2093                 "              OpReturn\n"
2094                 "              OpFunctionEnd\n";
2095
2096         spec1.inputs.push_back(BufferSp(new Vec4Buffer(inputFloats1)));
2097         spec1.outputs.push_back(BufferSp(new Vec4Buffer(outputFloats1)));
2098         spec1.numWorkGroups = IVec3(numElements, 1, 1);
2099
2100         group->addChild(new SpvAsmComputeShaderCase(testCtx, "vector", "OpCopyMemory elements of vector type", spec1));
2101
2102         // The following case copies a float[100] variable from the input buffer to the output buffer.
2103         ComputeShaderSpec                               spec2;
2104         vector<float>                                   inputFloats2    (numElements);
2105         vector<float>                                   outputFloats2   (numElements);
2106
2107         fillRandomScalars(rnd, -200.f, 200.f, &inputFloats2[0], numElements);
2108
2109         for (size_t ndx = 0; ndx < numElements; ++ndx)
2110                 outputFloats2[ndx] = inputFloats2[ndx];
2111
2112         spec2.assembly =
2113                 string(getComputeAsmShaderPreamble()) +
2114
2115                 "OpName %main           \"main\"\n"
2116                 "OpName %id             \"gl_GlobalInvocationID\"\n"
2117
2118                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
2119                 "OpDecorate %f32arr100 ArrayStride 4\n"
2120
2121                 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) +
2122
2123                 "%hundred        = OpConstant %u32 100\n"
2124                 "%f32arr100      = OpTypeArray %f32 %hundred\n"
2125                 "%f32arr100ptr_f = OpTypePointer Function %f32arr100\n"
2126                 "%f32arr100ptr_u = OpTypePointer Uniform %f32arr100\n"
2127                 "%buf            = OpTypeStruct %f32arr100\n"
2128                 "%bufptr         = OpTypePointer Uniform %buf\n"
2129                 "%indata         = OpVariable %bufptr Uniform\n"
2130                 "%outdata        = OpVariable %bufptr Uniform\n"
2131
2132                 "%id             = OpVariable %uvec3ptr Input\n"
2133                 "%zero           = OpConstant %i32 0\n"
2134
2135                 "%main           = OpFunction %void None %voidf\n"
2136                 "%label          = OpLabel\n"
2137                 "%var            = OpVariable %f32arr100ptr_f Function\n"
2138                 "%inarr          = OpAccessChain %f32arr100ptr_u %indata %zero\n"
2139                 "%outarr         = OpAccessChain %f32arr100ptr_u %outdata %zero\n"
2140                 "                  OpCopyMemory %var %inarr\n"
2141                 "                  OpCopyMemory %outarr %var\n"
2142                 "                  OpReturn\n"
2143                 "                  OpFunctionEnd\n";
2144
2145         spec2.inputs.push_back(BufferSp(new Float32Buffer(inputFloats2)));
2146         spec2.outputs.push_back(BufferSp(new Float32Buffer(outputFloats2)));
2147         spec2.numWorkGroups = IVec3(1, 1, 1);
2148
2149         group->addChild(new SpvAsmComputeShaderCase(testCtx, "array", "OpCopyMemory elements of array type", spec2));
2150
2151         // The following case copies a struct{vec4, vec4, vec4, vec4} variable from the input buffer to the output buffer.
2152         ComputeShaderSpec                               spec3;
2153         vector<float>                                   inputFloats3    (16);
2154         vector<float>                                   outputFloats3   (16);
2155
2156         fillRandomScalars(rnd, -200.f, 200.f, &inputFloats3[0], 16);
2157
2158         for (size_t ndx = 0; ndx < 16; ++ndx)
2159                 outputFloats3[ndx] = inputFloats3[ndx];
2160
2161         spec3.assembly =
2162                 string(getComputeAsmShaderPreamble()) +
2163
2164                 "OpName %main           \"main\"\n"
2165                 "OpName %id             \"gl_GlobalInvocationID\"\n"
2166
2167                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
2168                 //"OpMemberDecorate %buf 0 Offset 0\n"  - exists in getComputeAsmInputOutputBufferTraits
2169                 "OpMemberDecorate %buf 1 Offset 16\n"
2170                 "OpMemberDecorate %buf 2 Offset 32\n"
2171                 "OpMemberDecorate %buf 3 Offset 48\n"
2172
2173                 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) +
2174
2175                 "%vec4      = OpTypeVector %f32 4\n"
2176                 "%buf       = OpTypeStruct %vec4 %vec4 %vec4 %vec4\n"
2177                 "%bufptr    = OpTypePointer Uniform %buf\n"
2178                 "%indata    = OpVariable %bufptr Uniform\n"
2179                 "%outdata   = OpVariable %bufptr Uniform\n"
2180                 "%vec4stptr = OpTypePointer Function %buf\n"
2181
2182                 "%id        = OpVariable %uvec3ptr Input\n"
2183                 "%zero      = OpConstant %i32 0\n"
2184
2185                 "%main      = OpFunction %void None %voidf\n"
2186                 "%label     = OpLabel\n"
2187                 "%var       = OpVariable %vec4stptr Function\n"
2188                 "             OpCopyMemory %var %indata\n"
2189                 "             OpCopyMemory %outdata %var\n"
2190                 "             OpReturn\n"
2191                 "             OpFunctionEnd\n";
2192
2193         spec3.inputs.push_back(BufferSp(new Float32Buffer(inputFloats3)));
2194         spec3.outputs.push_back(BufferSp(new Float32Buffer(outputFloats3)));
2195         spec3.numWorkGroups = IVec3(1, 1, 1);
2196
2197         group->addChild(new SpvAsmComputeShaderCase(testCtx, "struct", "OpCopyMemory elements of struct type", spec3));
2198
2199         // The following case negates multiple float variables from the input buffer and stores the results to the output buffer.
2200         ComputeShaderSpec                               spec4;
2201         vector<float>                                   inputFloats4    (numElements);
2202         vector<float>                                   outputFloats4   (numElements);
2203
2204         fillRandomScalars(rnd, -200.f, 200.f, &inputFloats4[0], numElements);
2205
2206         for (size_t ndx = 0; ndx < numElements; ++ndx)
2207                 outputFloats4[ndx] = -inputFloats4[ndx];
2208
2209         spec4.assembly =
2210                 string(getComputeAsmShaderPreamble()) +
2211
2212                 "OpName %main           \"main\"\n"
2213                 "OpName %id             \"gl_GlobalInvocationID\"\n"
2214
2215                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
2216
2217                 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) + string(getComputeAsmInputOutputBuffer()) +
2218
2219                 "%f32ptr_f  = OpTypePointer Function %f32\n"
2220                 "%id        = OpVariable %uvec3ptr Input\n"
2221                 "%zero      = OpConstant %i32 0\n"
2222
2223                 "%main      = OpFunction %void None %voidf\n"
2224                 "%label     = OpLabel\n"
2225                 "%var       = OpVariable %f32ptr_f Function\n"
2226                 "%idval     = OpLoad %uvec3 %id\n"
2227                 "%x         = OpCompositeExtract %u32 %idval 0\n"
2228                 "%inloc     = OpAccessChain %f32ptr %indata %zero %x\n"
2229                 "%outloc    = OpAccessChain %f32ptr %outdata %zero %x\n"
2230                 "             OpCopyMemory %var %inloc\n"
2231                 "%val       = OpLoad %f32 %var\n"
2232                 "%neg       = OpFNegate %f32 %val\n"
2233                 "             OpStore %outloc %neg\n"
2234                 "             OpReturn\n"
2235                 "             OpFunctionEnd\n";
2236
2237         spec4.inputs.push_back(BufferSp(new Float32Buffer(inputFloats4)));
2238         spec4.outputs.push_back(BufferSp(new Float32Buffer(outputFloats4)));
2239         spec4.numWorkGroups = IVec3(numElements, 1, 1);
2240
2241         group->addChild(new SpvAsmComputeShaderCase(testCtx, "float", "OpCopyMemory elements of float type", spec4));
2242
2243         return group.release();
2244 }
2245
2246 tcu::TestCaseGroup* createOpCopyObjectGroup (tcu::TestContext& testCtx)
2247 {
2248         de::MovePtr<tcu::TestCaseGroup> group                   (new tcu::TestCaseGroup(testCtx, "opcopyobject", "Test the OpCopyObject instruction"));
2249         ComputeShaderSpec                               spec;
2250         de::Random                                              rnd                             (deStringHash(group->getName()));
2251         const int                                               numElements             = 100;
2252         vector<float>                                   inputFloats             (numElements, 0);
2253         vector<float>                                   outputFloats    (numElements, 0);
2254
2255         fillRandomScalars(rnd, -200.f, 200.f, &inputFloats[0], numElements);
2256
2257         // CPU might not use the same rounding mode as the GPU. Use whole numbers to avoid rounding differences.
2258         floorAll(inputFloats);
2259
2260         for (size_t ndx = 0; ndx < numElements; ++ndx)
2261                 outputFloats[ndx] = inputFloats[ndx] + 7.5f;
2262
2263         spec.assembly =
2264                 string(getComputeAsmShaderPreamble()) +
2265
2266                 "OpName %main           \"main\"\n"
2267                 "OpName %id             \"gl_GlobalInvocationID\"\n"
2268
2269                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
2270
2271                 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) +
2272
2273                 "%fmat     = OpTypeMatrix %fvec3 3\n"
2274                 "%three    = OpConstant %u32 3\n"
2275                 "%farr     = OpTypeArray %f32 %three\n"
2276                 "%fst      = OpTypeStruct %f32 %f32\n"
2277
2278                 + string(getComputeAsmInputOutputBuffer()) +
2279
2280                 "%id            = OpVariable %uvec3ptr Input\n"
2281                 "%zero          = OpConstant %i32 0\n"
2282                 "%c_f           = OpConstant %f32 1.5\n"
2283                 "%c_fvec3       = OpConstantComposite %fvec3 %c_f %c_f %c_f\n"
2284                 "%c_fmat        = OpConstantComposite %fmat %c_fvec3 %c_fvec3 %c_fvec3\n"
2285                 "%c_farr        = OpConstantComposite %farr %c_f %c_f %c_f\n"
2286                 "%c_fst         = OpConstantComposite %fst %c_f %c_f\n"
2287
2288                 "%main          = OpFunction %void None %voidf\n"
2289                 "%label         = OpLabel\n"
2290                 "%c_f_copy      = OpCopyObject %f32   %c_f\n"
2291                 "%c_fvec3_copy  = OpCopyObject %fvec3 %c_fvec3\n"
2292                 "%c_fmat_copy   = OpCopyObject %fmat  %c_fmat\n"
2293                 "%c_farr_copy   = OpCopyObject %farr  %c_farr\n"
2294                 "%c_fst_copy    = OpCopyObject %fst   %c_fst\n"
2295                 "%fvec3_elem    = OpCompositeExtract %f32 %c_fvec3_copy 0\n"
2296                 "%fmat_elem     = OpCompositeExtract %f32 %c_fmat_copy 1 2\n"
2297                 "%farr_elem     = OpCompositeExtract %f32 %c_farr_copy 2\n"
2298                 "%fst_elem      = OpCompositeExtract %f32 %c_fst_copy 1\n"
2299                 // Add up. 1.5 * 5 = 7.5.
2300                 "%add1          = OpFAdd %f32 %c_f_copy %fvec3_elem\n"
2301                 "%add2          = OpFAdd %f32 %add1     %fmat_elem\n"
2302                 "%add3          = OpFAdd %f32 %add2     %farr_elem\n"
2303                 "%add4          = OpFAdd %f32 %add3     %fst_elem\n"
2304
2305                 "%idval         = OpLoad %uvec3 %id\n"
2306                 "%x             = OpCompositeExtract %u32 %idval 0\n"
2307                 "%inloc         = OpAccessChain %f32ptr %indata %zero %x\n"
2308                 "%outloc        = OpAccessChain %f32ptr %outdata %zero %x\n"
2309                 "%inval         = OpLoad %f32 %inloc\n"
2310                 "%add           = OpFAdd %f32 %add4 %inval\n"
2311                 "                 OpStore %outloc %add\n"
2312                 "                 OpReturn\n"
2313                 "                 OpFunctionEnd\n";
2314         spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
2315         spec.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
2316         spec.numWorkGroups = IVec3(numElements, 1, 1);
2317
2318         group->addChild(new SpvAsmComputeShaderCase(testCtx, "spotcheck", "OpCopyObject on different types", spec));
2319
2320         return group.release();
2321 }
2322 // Assembly code used for testing OpUnreachable is based on GLSL source code:
2323 //
2324 // #version 430
2325 //
2326 // layout(std140, set = 0, binding = 0) readonly buffer Input {
2327 //   float elements[];
2328 // } input_data;
2329 // layout(std140, set = 0, binding = 1) writeonly buffer Output {
2330 //   float elements[];
2331 // } output_data;
2332 //
2333 // void not_called_func() {
2334 //   // place OpUnreachable here
2335 // }
2336 //
2337 // uint modulo4(uint val) {
2338 //   switch (val % uint(4)) {
2339 //     case 0:  return 3;
2340 //     case 1:  return 2;
2341 //     case 2:  return 1;
2342 //     case 3:  return 0;
2343 //     default: return 100; // place OpUnreachable here
2344 //   }
2345 // }
2346 //
2347 // uint const5() {
2348 //   return 5;
2349 //   // place OpUnreachable here
2350 // }
2351 //
2352 // void main() {
2353 //   uint x = gl_GlobalInvocationID.x;
2354 //   if (const5() > modulo4(1000)) {
2355 //     output_data.elements[x] = -input_data.elements[x];
2356 //   } else {
2357 //     // place OpUnreachable here
2358 //     output_data.elements[x] = input_data.elements[x];
2359 //   }
2360 // }
2361
2362 tcu::TestCaseGroup* createOpUnreachableGroup (tcu::TestContext& testCtx)
2363 {
2364         de::MovePtr<tcu::TestCaseGroup> group                   (new tcu::TestCaseGroup(testCtx, "opunreachable", "Test the OpUnreachable instruction"));
2365         ComputeShaderSpec                               spec;
2366         de::Random                                              rnd                             (deStringHash(group->getName()));
2367         const int                                               numElements             = 100;
2368         vector<float>                                   positiveFloats  (numElements, 0);
2369         vector<float>                                   negativeFloats  (numElements, 0);
2370
2371         fillRandomScalars(rnd, 1.f, 100.f, &positiveFloats[0], numElements);
2372
2373         for (size_t ndx = 0; ndx < numElements; ++ndx)
2374                 negativeFloats[ndx] = -positiveFloats[ndx];
2375
2376         spec.assembly =
2377                 string(getComputeAsmShaderPreamble()) +
2378
2379                 "OpSource GLSL 430\n"
2380                 "OpName %main            \"main\"\n"
2381                 "OpName %func_not_called_func \"not_called_func(\"\n"
2382                 "OpName %func_modulo4         \"modulo4(u1;\"\n"
2383                 "OpName %func_const5          \"const5(\"\n"
2384                 "OpName %id                   \"gl_GlobalInvocationID\"\n"
2385
2386                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
2387
2388                 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) +
2389
2390                 "%u32ptr    = OpTypePointer Function %u32\n"
2391                 "%uintfuint = OpTypeFunction %u32 %u32ptr\n"
2392                 "%unitf     = OpTypeFunction %u32\n"
2393
2394                 "%id        = OpVariable %uvec3ptr Input\n"
2395                 "%zero      = OpConstant %u32 0\n"
2396                 "%one       = OpConstant %u32 1\n"
2397                 "%two       = OpConstant %u32 2\n"
2398                 "%three     = OpConstant %u32 3\n"
2399                 "%four      = OpConstant %u32 4\n"
2400                 "%five      = OpConstant %u32 5\n"
2401                 "%hundred   = OpConstant %u32 100\n"
2402                 "%thousand  = OpConstant %u32 1000\n"
2403
2404                 + string(getComputeAsmInputOutputBuffer()) +
2405
2406                 // Main()
2407                 "%main   = OpFunction %void None %voidf\n"
2408                 "%main_entry  = OpLabel\n"
2409                 "%v_thousand  = OpVariable %u32ptr Function %thousand\n"
2410                 "%idval       = OpLoad %uvec3 %id\n"
2411                 "%x           = OpCompositeExtract %u32 %idval 0\n"
2412                 "%inloc       = OpAccessChain %f32ptr %indata %zero %x\n"
2413                 "%inval       = OpLoad %f32 %inloc\n"
2414                 "%outloc      = OpAccessChain %f32ptr %outdata %zero %x\n"
2415                 "%ret_const5  = OpFunctionCall %u32 %func_const5\n"
2416                 "%ret_modulo4 = OpFunctionCall %u32 %func_modulo4 %v_thousand\n"
2417                 "%cmp_gt      = OpUGreaterThan %bool %ret_const5 %ret_modulo4\n"
2418                 "               OpSelectionMerge %if_end None\n"
2419                 "               OpBranchConditional %cmp_gt %if_true %if_false\n"
2420                 "%if_true     = OpLabel\n"
2421                 "%negate      = OpFNegate %f32 %inval\n"
2422                 "               OpStore %outloc %negate\n"
2423                 "               OpBranch %if_end\n"
2424                 "%if_false    = OpLabel\n"
2425                 "               OpUnreachable\n" // Unreachable else branch for if statement
2426                 "%if_end      = OpLabel\n"
2427                 "               OpReturn\n"
2428                 "               OpFunctionEnd\n"
2429
2430                 // not_called_function()
2431                 "%func_not_called_func  = OpFunction %void None %voidf\n"
2432                 "%not_called_func_entry = OpLabel\n"
2433                 "                         OpUnreachable\n" // Unreachable entry block in not called static function
2434                 "                         OpFunctionEnd\n"
2435
2436                 // modulo4()
2437                 "%func_modulo4  = OpFunction %u32 None %uintfuint\n"
2438                 "%valptr        = OpFunctionParameter %u32ptr\n"
2439                 "%modulo4_entry = OpLabel\n"
2440                 "%val           = OpLoad %u32 %valptr\n"
2441                 "%modulo        = OpUMod %u32 %val %four\n"
2442                 "                 OpSelectionMerge %switch_merge None\n"
2443                 "                 OpSwitch %modulo %default 0 %case0 1 %case1 2 %case2 3 %case3\n"
2444                 "%case0         = OpLabel\n"
2445                 "                 OpReturnValue %three\n"
2446                 "%case1         = OpLabel\n"
2447                 "                 OpReturnValue %two\n"
2448                 "%case2         = OpLabel\n"
2449                 "                 OpReturnValue %one\n"
2450                 "%case3         = OpLabel\n"
2451                 "                 OpReturnValue %zero\n"
2452                 "%default       = OpLabel\n"
2453                 "                 OpUnreachable\n" // Unreachable default case for switch statement
2454                 "%switch_merge  = OpLabel\n"
2455                 "                 OpUnreachable\n" // Unreachable merge block for switch statement
2456                 "                 OpFunctionEnd\n"
2457
2458                 // const5()
2459                 "%func_const5  = OpFunction %u32 None %unitf\n"
2460                 "%const5_entry = OpLabel\n"
2461                 "                OpReturnValue %five\n"
2462                 "%unreachable  = OpLabel\n"
2463                 "                OpUnreachable\n" // Unreachable block in function
2464                 "                OpFunctionEnd\n";
2465         spec.inputs.push_back(BufferSp(new Float32Buffer(positiveFloats)));
2466         spec.outputs.push_back(BufferSp(new Float32Buffer(negativeFloats)));
2467         spec.numWorkGroups = IVec3(numElements, 1, 1);
2468
2469         group->addChild(new SpvAsmComputeShaderCase(testCtx, "all", "OpUnreachable appearing at different places", spec));
2470
2471         return group.release();
2472 }
2473
2474 // Assembly code used for testing decoration group is based on GLSL source code:
2475 //
2476 // #version 430
2477 //
2478 // layout(std140, set = 0, binding = 0) readonly buffer Input0 {
2479 //   float elements[];
2480 // } input_data0;
2481 // layout(std140, set = 0, binding = 1) readonly buffer Input1 {
2482 //   float elements[];
2483 // } input_data1;
2484 // layout(std140, set = 0, binding = 2) readonly buffer Input2 {
2485 //   float elements[];
2486 // } input_data2;
2487 // layout(std140, set = 0, binding = 3) readonly buffer Input3 {
2488 //   float elements[];
2489 // } input_data3;
2490 // layout(std140, set = 0, binding = 4) readonly buffer Input4 {
2491 //   float elements[];
2492 // } input_data4;
2493 // layout(std140, set = 0, binding = 5) writeonly buffer Output {
2494 //   float elements[];
2495 // } output_data;
2496 //
2497 // void main() {
2498 //   uint x = gl_GlobalInvocationID.x;
2499 //   output_data.elements[x] = input_data0.elements[x] + input_data1.elements[x] + input_data2.elements[x] + input_data3.elements[x] + input_data4.elements[x];
2500 // }
2501 tcu::TestCaseGroup* createDecorationGroupGroup (tcu::TestContext& testCtx)
2502 {
2503         de::MovePtr<tcu::TestCaseGroup> group                   (new tcu::TestCaseGroup(testCtx, "decoration_group", "Test the OpDecorationGroup & OpGroupDecorate instruction"));
2504         ComputeShaderSpec                               spec;
2505         de::Random                                              rnd                             (deStringHash(group->getName()));
2506         const int                                               numElements             = 100;
2507         vector<float>                                   inputFloats0    (numElements, 0);
2508         vector<float>                                   inputFloats1    (numElements, 0);
2509         vector<float>                                   inputFloats2    (numElements, 0);
2510         vector<float>                                   inputFloats3    (numElements, 0);
2511         vector<float>                                   inputFloats4    (numElements, 0);
2512         vector<float>                                   outputFloats    (numElements, 0);
2513
2514         fillRandomScalars(rnd, -300.f, 300.f, &inputFloats0[0], numElements);
2515         fillRandomScalars(rnd, -300.f, 300.f, &inputFloats1[0], numElements);
2516         fillRandomScalars(rnd, -300.f, 300.f, &inputFloats2[0], numElements);
2517         fillRandomScalars(rnd, -300.f, 300.f, &inputFloats3[0], numElements);
2518         fillRandomScalars(rnd, -300.f, 300.f, &inputFloats4[0], numElements);
2519
2520         // CPU might not use the same rounding mode as the GPU. Use whole numbers to avoid rounding differences.
2521         floorAll(inputFloats0);
2522         floorAll(inputFloats1);
2523         floorAll(inputFloats2);
2524         floorAll(inputFloats3);
2525         floorAll(inputFloats4);
2526
2527         for (size_t ndx = 0; ndx < numElements; ++ndx)
2528                 outputFloats[ndx] = inputFloats0[ndx] + inputFloats1[ndx] + inputFloats2[ndx] + inputFloats3[ndx] + inputFloats4[ndx];
2529
2530         spec.assembly =
2531                 string(getComputeAsmShaderPreamble()) +
2532
2533                 "OpSource GLSL 430\n"
2534                 "OpName %main \"main\"\n"
2535                 "OpName %id \"gl_GlobalInvocationID\"\n"
2536
2537                 // Not using group decoration on variable.
2538                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
2539                 // Not using group decoration on type.
2540                 "OpDecorate %f32arr ArrayStride 4\n"
2541
2542                 "OpDecorate %groups BufferBlock\n"
2543                 "OpDecorate %groupm Offset 0\n"
2544                 "%groups = OpDecorationGroup\n"
2545                 "%groupm = OpDecorationGroup\n"
2546
2547                 // Group decoration on multiple structs.
2548                 "OpGroupDecorate %groups %outbuf %inbuf0 %inbuf1 %inbuf2 %inbuf3 %inbuf4\n"
2549                 // Group decoration on multiple struct members.
2550                 "OpGroupMemberDecorate %groupm %outbuf 0 %inbuf0 0 %inbuf1 0 %inbuf2 0 %inbuf3 0 %inbuf4 0\n"
2551
2552                 "OpDecorate %group1 DescriptorSet 0\n"
2553                 "OpDecorate %group3 DescriptorSet 0\n"
2554                 "OpDecorate %group3 NonWritable\n"
2555                 "OpDecorate %group3 Restrict\n"
2556                 "%group0 = OpDecorationGroup\n"
2557                 "%group1 = OpDecorationGroup\n"
2558                 "%group3 = OpDecorationGroup\n"
2559
2560                 // Applying the same decoration group multiple times.
2561                 "OpGroupDecorate %group1 %outdata\n"
2562                 "OpGroupDecorate %group1 %outdata\n"
2563                 "OpGroupDecorate %group1 %outdata\n"
2564                 "OpDecorate %outdata DescriptorSet 0\n"
2565                 "OpDecorate %outdata Binding 5\n"
2566                 // Applying decoration group containing nothing.
2567                 "OpGroupDecorate %group0 %indata0\n"
2568                 "OpDecorate %indata0 DescriptorSet 0\n"
2569                 "OpDecorate %indata0 Binding 0\n"
2570                 // Applying decoration group containing one decoration.
2571                 "OpGroupDecorate %group1 %indata1\n"
2572                 "OpDecorate %indata1 Binding 1\n"
2573                 // Applying decoration group containing multiple decorations.
2574                 "OpGroupDecorate %group3 %indata2 %indata3\n"
2575                 "OpDecorate %indata2 Binding 2\n"
2576                 "OpDecorate %indata3 Binding 3\n"
2577                 // Applying multiple decoration groups (with overlapping).
2578                 "OpGroupDecorate %group0 %indata4\n"
2579                 "OpGroupDecorate %group1 %indata4\n"
2580                 "OpGroupDecorate %group3 %indata4\n"
2581                 "OpDecorate %indata4 Binding 4\n"
2582
2583                 + string(getComputeAsmCommonTypes()) +
2584
2585                 "%id   = OpVariable %uvec3ptr Input\n"
2586                 "%zero = OpConstant %i32 0\n"
2587
2588                 "%outbuf    = OpTypeStruct %f32arr\n"
2589                 "%outbufptr = OpTypePointer Uniform %outbuf\n"
2590                 "%outdata   = OpVariable %outbufptr Uniform\n"
2591                 "%inbuf0    = OpTypeStruct %f32arr\n"
2592                 "%inbuf0ptr = OpTypePointer Uniform %inbuf0\n"
2593                 "%indata0   = OpVariable %inbuf0ptr Uniform\n"
2594                 "%inbuf1    = OpTypeStruct %f32arr\n"
2595                 "%inbuf1ptr = OpTypePointer Uniform %inbuf1\n"
2596                 "%indata1   = OpVariable %inbuf1ptr Uniform\n"
2597                 "%inbuf2    = OpTypeStruct %f32arr\n"
2598                 "%inbuf2ptr = OpTypePointer Uniform %inbuf2\n"
2599                 "%indata2   = OpVariable %inbuf2ptr Uniform\n"
2600                 "%inbuf3    = OpTypeStruct %f32arr\n"
2601                 "%inbuf3ptr = OpTypePointer Uniform %inbuf3\n"
2602                 "%indata3   = OpVariable %inbuf3ptr Uniform\n"
2603                 "%inbuf4    = OpTypeStruct %f32arr\n"
2604                 "%inbufptr  = OpTypePointer Uniform %inbuf4\n"
2605                 "%indata4   = OpVariable %inbufptr Uniform\n"
2606
2607                 "%main   = OpFunction %void None %voidf\n"
2608                 "%label  = OpLabel\n"
2609                 "%idval  = OpLoad %uvec3 %id\n"
2610                 "%x      = OpCompositeExtract %u32 %idval 0\n"
2611                 "%inloc0 = OpAccessChain %f32ptr %indata0 %zero %x\n"
2612                 "%inloc1 = OpAccessChain %f32ptr %indata1 %zero %x\n"
2613                 "%inloc2 = OpAccessChain %f32ptr %indata2 %zero %x\n"
2614                 "%inloc3 = OpAccessChain %f32ptr %indata3 %zero %x\n"
2615                 "%inloc4 = OpAccessChain %f32ptr %indata4 %zero %x\n"
2616                 "%outloc = OpAccessChain %f32ptr %outdata %zero %x\n"
2617                 "%inval0 = OpLoad %f32 %inloc0\n"
2618                 "%inval1 = OpLoad %f32 %inloc1\n"
2619                 "%inval2 = OpLoad %f32 %inloc2\n"
2620                 "%inval3 = OpLoad %f32 %inloc3\n"
2621                 "%inval4 = OpLoad %f32 %inloc4\n"
2622                 "%add0   = OpFAdd %f32 %inval0 %inval1\n"
2623                 "%add1   = OpFAdd %f32 %add0 %inval2\n"
2624                 "%add2   = OpFAdd %f32 %add1 %inval3\n"
2625                 "%add    = OpFAdd %f32 %add2 %inval4\n"
2626                 "          OpStore %outloc %add\n"
2627                 "          OpReturn\n"
2628                 "          OpFunctionEnd\n";
2629         spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats0)));
2630         spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats1)));
2631         spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats2)));
2632         spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats3)));
2633         spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats4)));
2634         spec.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
2635         spec.numWorkGroups = IVec3(numElements, 1, 1);
2636
2637         group->addChild(new SpvAsmComputeShaderCase(testCtx, "all", "decoration group cases", spec));
2638
2639         return group.release();
2640 }
2641
2642 struct SpecConstantTwoIntCase
2643 {
2644         const char*             caseName;
2645         const char*             scDefinition0;
2646         const char*             scDefinition1;
2647         const char*             scResultType;
2648         const char*             scOperation;
2649         deInt32                 scActualValue0;
2650         deInt32                 scActualValue1;
2651         const char*             resultOperation;
2652         vector<deInt32> expectedOutput;
2653         deInt32                 scActualValueLength;
2654
2655                                         SpecConstantTwoIntCase (const char* name,
2656                                                                                         const char* definition0,
2657                                                                                         const char* definition1,
2658                                                                                         const char* resultType,
2659                                                                                         const char* operation,
2660                                                                                         deInt32 value0,
2661                                                                                         deInt32 value1,
2662                                                                                         const char* resultOp,
2663                                                                                         const vector<deInt32>& output,
2664                                                                                         const deInt32   valueLength = sizeof(deInt32))
2665                                                 : caseName                              (name)
2666                                                 , scDefinition0                 (definition0)
2667                                                 , scDefinition1                 (definition1)
2668                                                 , scResultType                  (resultType)
2669                                                 , scOperation                   (operation)
2670                                                 , scActualValue0                (value0)
2671                                                 , scActualValue1                (value1)
2672                                                 , resultOperation               (resultOp)
2673                                                 , expectedOutput                (output)
2674                                                 , scActualValueLength   (valueLength)
2675                                                 {}
2676 };
2677
2678 tcu::TestCaseGroup* createSpecConstantGroup (tcu::TestContext& testCtx)
2679 {
2680         de::MovePtr<tcu::TestCaseGroup> group                   (new tcu::TestCaseGroup(testCtx, "opspecconstantop", "Test the OpSpecConstantOp instruction"));
2681         vector<SpecConstantTwoIntCase>  cases;
2682         de::Random                                              rnd                             (deStringHash(group->getName()));
2683         const int                                               numElements             = 100;
2684         const deInt32                                   p1AsFloat16             = 0x3c00; // +1(fp16) == 0 01111 0000000000 == 0011 1100 0000 0000
2685         vector<deInt32>                                 inputInts               (numElements, 0);
2686         vector<deInt32>                                 outputInts1             (numElements, 0);
2687         vector<deInt32>                                 outputInts2             (numElements, 0);
2688         vector<deInt32>                                 outputInts3             (numElements, 0);
2689         vector<deInt32>                                 outputInts4             (numElements, 0);
2690         const StringTemplate                    shaderTemplate  (
2691                 "${CAPABILITIES:opt}"
2692                 + string(getComputeAsmShaderPreamble()) +
2693
2694                 "OpName %main           \"main\"\n"
2695                 "OpName %id             \"gl_GlobalInvocationID\"\n"
2696
2697                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
2698                 "OpDecorate %sc_0  SpecId 0\n"
2699                 "OpDecorate %sc_1  SpecId 1\n"
2700                 "OpDecorate %i32arr ArrayStride 4\n"
2701
2702                 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) +
2703
2704                 "${OPTYPE_DEFINITIONS:opt}"
2705                 "%buf     = OpTypeStruct %i32arr\n"
2706                 "%bufptr  = OpTypePointer Uniform %buf\n"
2707                 "%indata    = OpVariable %bufptr Uniform\n"
2708                 "%outdata   = OpVariable %bufptr Uniform\n"
2709
2710                 "%id        = OpVariable %uvec3ptr Input\n"
2711                 "%zero      = OpConstant %i32 0\n"
2712
2713                 "%sc_0      = OpSpecConstant${SC_DEF0}\n"
2714                 "%sc_1      = OpSpecConstant${SC_DEF1}\n"
2715                 "%sc_final  = OpSpecConstantOp ${SC_RESULT_TYPE} ${SC_OP}\n"
2716
2717                 "%main      = OpFunction %void None %voidf\n"
2718                 "%label     = OpLabel\n"
2719                 "${TYPE_CONVERT:opt}"
2720                 "%idval     = OpLoad %uvec3 %id\n"
2721                 "%x         = OpCompositeExtract %u32 %idval 0\n"
2722                 "%inloc     = OpAccessChain %i32ptr %indata %zero %x\n"
2723                 "%inval     = OpLoad %i32 %inloc\n"
2724                 "%final     = ${GEN_RESULT}\n"
2725                 "%outloc    = OpAccessChain %i32ptr %outdata %zero %x\n"
2726                 "             OpStore %outloc %final\n"
2727                 "             OpReturn\n"
2728                 "             OpFunctionEnd\n");
2729
2730         fillRandomScalars(rnd, -65536, 65536, &inputInts[0], numElements);
2731
2732         for (size_t ndx = 0; ndx < numElements; ++ndx)
2733         {
2734                 outputInts1[ndx] = inputInts[ndx] + 42;
2735                 outputInts2[ndx] = inputInts[ndx];
2736                 outputInts3[ndx] = inputInts[ndx] - 11200;
2737                 outputInts4[ndx] = inputInts[ndx] + 1;
2738         }
2739
2740         const char addScToInput[]               = "OpIAdd %i32 %inval %sc_final";
2741         const char addSc32ToInput[]             = "OpIAdd %i32 %inval %sc_final32";
2742         const char selectTrueUsingSc[]  = "OpSelect %i32 %sc_final %inval %zero";
2743         const char selectFalseUsingSc[] = "OpSelect %i32 %sc_final %zero %inval";
2744
2745         cases.push_back(SpecConstantTwoIntCase("iadd",                                  " %i32 0",              " %i32 0",              "%i32",         "IAdd                 %sc_0 %sc_1",                     62,             -20,    addScToInput,           outputInts1));
2746         cases.push_back(SpecConstantTwoIntCase("isub",                                  " %i32 0",              " %i32 0",              "%i32",         "ISub                 %sc_0 %sc_1",                     100,    58,             addScToInput,           outputInts1));
2747         cases.push_back(SpecConstantTwoIntCase("imul",                                  " %i32 0",              " %i32 0",              "%i32",         "IMul                 %sc_0 %sc_1",                     -2,             -21,    addScToInput,           outputInts1));
2748         cases.push_back(SpecConstantTwoIntCase("sdiv",                                  " %i32 0",              " %i32 0",              "%i32",         "SDiv                 %sc_0 %sc_1",                     -126,   -3,             addScToInput,           outputInts1));
2749         cases.push_back(SpecConstantTwoIntCase("udiv",                                  " %i32 0",              " %i32 0",              "%i32",         "UDiv                 %sc_0 %sc_1",                     126,    3,              addScToInput,           outputInts1));
2750         cases.push_back(SpecConstantTwoIntCase("srem",                                  " %i32 0",              " %i32 0",              "%i32",         "SRem                 %sc_0 %sc_1",                     7,              3,              addScToInput,           outputInts4));
2751         cases.push_back(SpecConstantTwoIntCase("smod",                                  " %i32 0",              " %i32 0",              "%i32",         "SMod                 %sc_0 %sc_1",                     7,              3,              addScToInput,           outputInts4));
2752         cases.push_back(SpecConstantTwoIntCase("umod",                                  " %i32 0",              " %i32 0",              "%i32",         "UMod                 %sc_0 %sc_1",                     342,    50,             addScToInput,           outputInts1));
2753         cases.push_back(SpecConstantTwoIntCase("bitwiseand",                    " %i32 0",              " %i32 0",              "%i32",         "BitwiseAnd           %sc_0 %sc_1",                     42,             63,             addScToInput,           outputInts1));
2754         cases.push_back(SpecConstantTwoIntCase("bitwiseor",                             " %i32 0",              " %i32 0",              "%i32",         "BitwiseOr            %sc_0 %sc_1",                     34,             8,              addScToInput,           outputInts1));
2755         cases.push_back(SpecConstantTwoIntCase("bitwisexor",                    " %i32 0",              " %i32 0",              "%i32",         "BitwiseXor           %sc_0 %sc_1",                     18,             56,             addScToInput,           outputInts1));
2756         cases.push_back(SpecConstantTwoIntCase("shiftrightlogical",             " %i32 0",              " %i32 0",              "%i32",         "ShiftRightLogical    %sc_0 %sc_1",                     168,    2,              addScToInput,           outputInts1));
2757         cases.push_back(SpecConstantTwoIntCase("shiftrightarithmetic",  " %i32 0",              " %i32 0",              "%i32",         "ShiftRightArithmetic %sc_0 %sc_1",                     168,    2,              addScToInput,           outputInts1));
2758         cases.push_back(SpecConstantTwoIntCase("shiftleftlogical",              " %i32 0",              " %i32 0",              "%i32",         "ShiftLeftLogical     %sc_0 %sc_1",                     21,             1,              addScToInput,           outputInts1));
2759         cases.push_back(SpecConstantTwoIntCase("slessthan",                             " %i32 0",              " %i32 0",              "%bool",        "SLessThan            %sc_0 %sc_1",                     -20,    -10,    selectTrueUsingSc,      outputInts2));
2760         cases.push_back(SpecConstantTwoIntCase("ulessthan",                             " %i32 0",              " %i32 0",              "%bool",        "ULessThan            %sc_0 %sc_1",                     10,             20,             selectTrueUsingSc,      outputInts2));
2761         cases.push_back(SpecConstantTwoIntCase("sgreaterthan",                  " %i32 0",              " %i32 0",              "%bool",        "SGreaterThan         %sc_0 %sc_1",                     -1000,  50,             selectFalseUsingSc,     outputInts2));
2762         cases.push_back(SpecConstantTwoIntCase("ugreaterthan",                  " %i32 0",              " %i32 0",              "%bool",        "UGreaterThan         %sc_0 %sc_1",                     10,             5,              selectTrueUsingSc,      outputInts2));
2763         cases.push_back(SpecConstantTwoIntCase("slessthanequal",                " %i32 0",              " %i32 0",              "%bool",        "SLessThanEqual       %sc_0 %sc_1",                     -10,    -10,    selectTrueUsingSc,      outputInts2));
2764         cases.push_back(SpecConstantTwoIntCase("ulessthanequal",                " %i32 0",              " %i32 0",              "%bool",        "ULessThanEqual       %sc_0 %sc_1",                     50,             100,    selectTrueUsingSc,      outputInts2));
2765         cases.push_back(SpecConstantTwoIntCase("sgreaterthanequal",             " %i32 0",              " %i32 0",              "%bool",        "SGreaterThanEqual    %sc_0 %sc_1",                     -1000,  50,             selectFalseUsingSc,     outputInts2));
2766         cases.push_back(SpecConstantTwoIntCase("ugreaterthanequal",             " %i32 0",              " %i32 0",              "%bool",        "UGreaterThanEqual    %sc_0 %sc_1",                     10,             10,             selectTrueUsingSc,      outputInts2));
2767         cases.push_back(SpecConstantTwoIntCase("iequal",                                " %i32 0",              " %i32 0",              "%bool",        "IEqual               %sc_0 %sc_1",                     42,             24,             selectFalseUsingSc,     outputInts2));
2768         cases.push_back(SpecConstantTwoIntCase("inotequal",                             " %i32 0",              " %i32 0",              "%bool",        "INotEqual            %sc_0 %sc_1",                     42,             24,             selectTrueUsingSc,      outputInts2));
2769         cases.push_back(SpecConstantTwoIntCase("logicaland",                    "True %bool",   "True %bool",   "%bool",        "LogicalAnd           %sc_0 %sc_1",                     0,              1,              selectFalseUsingSc,     outputInts2));
2770         cases.push_back(SpecConstantTwoIntCase("logicalor",                             "False %bool",  "False %bool",  "%bool",        "LogicalOr            %sc_0 %sc_1",                     1,              0,              selectTrueUsingSc,      outputInts2));
2771         cases.push_back(SpecConstantTwoIntCase("logicalequal",                  "True %bool",   "True %bool",   "%bool",        "LogicalEqual         %sc_0 %sc_1",                     0,              1,              selectFalseUsingSc,     outputInts2));
2772         cases.push_back(SpecConstantTwoIntCase("logicalnotequal",               "False %bool",  "False %bool",  "%bool",        "LogicalNotEqual      %sc_0 %sc_1",                     1,              0,              selectTrueUsingSc,      outputInts2));
2773         cases.push_back(SpecConstantTwoIntCase("snegate",                               " %i32 0",              " %i32 0",              "%i32",         "SNegate              %sc_0",                           -42,    0,              addScToInput,           outputInts1));
2774         cases.push_back(SpecConstantTwoIntCase("not",                                   " %i32 0",              " %i32 0",              "%i32",         "Not                  %sc_0",                           -43,    0,              addScToInput,           outputInts1));
2775         cases.push_back(SpecConstantTwoIntCase("logicalnot",                    "False %bool",  "False %bool",  "%bool",        "LogicalNot           %sc_0",                           1,              0,              selectFalseUsingSc,     outputInts2));
2776         cases.push_back(SpecConstantTwoIntCase("select",                                "False %bool",  " %i32 0",              "%i32",         "Select               %sc_0 %sc_1 %zero",       1,              42,             addScToInput,           outputInts1));
2777         cases.push_back(SpecConstantTwoIntCase("sconvert",                              " %i32 0",              " %i32 0",              "%i16",         "SConvert             %sc_0",                           -11200, 0,              addSc32ToInput,         outputInts3));
2778         // -969998336 stored as 32-bit two's complement is the binary representation of -11200 as IEEE-754 Float
2779         cases.push_back(SpecConstantTwoIntCase("fconvert",                              " %f32 0",              " %f32 0",              "%f64",         "FConvert             %sc_0",                           -969998336, 0,  addSc32ToInput,         outputInts3));
2780         cases.push_back(SpecConstantTwoIntCase("fconvert16",                    " %f16 0",              " %f16 0",              "%f32",         "FConvert             %sc_0",                           p1AsFloat16, 0, addSc32ToInput,         outputInts4, sizeof(deFloat16)));
2781
2782         for (size_t caseNdx = 0; caseNdx < cases.size(); ++caseNdx)
2783         {
2784                 map<string, string>             specializations;
2785                 ComputeShaderSpec               spec;
2786
2787                 specializations["SC_DEF0"]                      = cases[caseNdx].scDefinition0;
2788                 specializations["SC_DEF1"]                      = cases[caseNdx].scDefinition1;
2789                 specializations["SC_RESULT_TYPE"]       = cases[caseNdx].scResultType;
2790                 specializations["SC_OP"]                        = cases[caseNdx].scOperation;
2791                 specializations["GEN_RESULT"]           = cases[caseNdx].resultOperation;
2792
2793                 // Special SPIR-V code for SConvert-case
2794                 if (strcmp(cases[caseNdx].caseName, "sconvert") == 0)
2795                 {
2796                         spec.requestedVulkanFeatures.coreFeatures.shaderInt16   = VK_TRUE;
2797                         specializations["CAPABILITIES"]                                                 = "OpCapability Int16\n";                                                       // Adds 16-bit integer capability
2798                         specializations["OPTYPE_DEFINITIONS"]                                   = "%i16 = OpTypeInt 16 1\n";                                            // Adds 16-bit integer type
2799                         specializations["TYPE_CONVERT"]                                                 = "%sc_final32 = OpSConvert %i32 %sc_final\n";          // Converts 16-bit integer to 32-bit integer
2800                 }
2801
2802                 // Special SPIR-V code for FConvert-case
2803                 if (strcmp(cases[caseNdx].caseName, "fconvert") == 0)
2804                 {
2805                         spec.requestedVulkanFeatures.coreFeatures.shaderFloat64 = VK_TRUE;
2806                         specializations["CAPABILITIES"]                                                 = "OpCapability Float64\n";                                                     // Adds 64-bit float capability
2807                         specializations["OPTYPE_DEFINITIONS"]                                   = "%f64 = OpTypeFloat 64\n";                                            // Adds 64-bit float type
2808                         specializations["TYPE_CONVERT"]                                                 = "%sc_final32 = OpConvertFToS %i32 %sc_final\n";       // Converts 64-bit float to 32-bit integer
2809                 }
2810
2811                 // Special SPIR-V code for FConvert-case for 16-bit floats
2812                 if (strcmp(cases[caseNdx].caseName, "fconvert16") == 0)
2813                 {
2814                         spec.extensions.push_back("VK_KHR_shader_float16_int8");
2815                         spec.requestedVulkanFeatures.extFloat16Int8 = EXTFLOAT16INT8FEATURES_FLOAT16;
2816                         specializations["CAPABILITIES"]                 = "OpCapability Float16\n";                                                     // Adds 16-bit float capability
2817                         specializations["OPTYPE_DEFINITIONS"]   = "%f16 = OpTypeFloat 16\n";                                            // Adds 16-bit float type
2818                         specializations["TYPE_CONVERT"]                 = "%sc_final32 = OpConvertFToS %i32 %sc_final\n";       // Converts 16-bit float to 32-bit integer
2819                 }
2820
2821                 spec.assembly = shaderTemplate.specialize(specializations);
2822                 spec.inputs.push_back(BufferSp(new Int32Buffer(inputInts)));
2823                 spec.outputs.push_back(BufferSp(new Int32Buffer(cases[caseNdx].expectedOutput)));
2824                 spec.numWorkGroups = IVec3(numElements, 1, 1);
2825                 spec.specConstants.append(&cases[caseNdx].scActualValue0, cases[caseNdx].scActualValueLength);
2826                 spec.specConstants.append(&cases[caseNdx].scActualValue1, cases[caseNdx].scActualValueLength);
2827
2828                 group->addChild(new SpvAsmComputeShaderCase(testCtx, cases[caseNdx].caseName, cases[caseNdx].caseName, spec));
2829         }
2830
2831         ComputeShaderSpec                               spec;
2832
2833         spec.assembly =
2834                 string(getComputeAsmShaderPreamble()) +
2835
2836                 "OpName %main           \"main\"\n"
2837                 "OpName %id             \"gl_GlobalInvocationID\"\n"
2838
2839                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
2840                 "OpDecorate %sc_0  SpecId 0\n"
2841                 "OpDecorate %sc_1  SpecId 1\n"
2842                 "OpDecorate %sc_2  SpecId 2\n"
2843                 "OpDecorate %i32arr ArrayStride 4\n"
2844
2845                 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) +
2846
2847                 "%ivec3       = OpTypeVector %i32 3\n"
2848                 "%buf         = OpTypeStruct %i32arr\n"
2849                 "%bufptr      = OpTypePointer Uniform %buf\n"
2850                 "%indata      = OpVariable %bufptr Uniform\n"
2851                 "%outdata     = OpVariable %bufptr Uniform\n"
2852
2853                 "%id          = OpVariable %uvec3ptr Input\n"
2854                 "%zero        = OpConstant %i32 0\n"
2855                 "%ivec3_0     = OpConstantComposite %ivec3 %zero %zero %zero\n"
2856                 "%vec3_undef  = OpUndef %ivec3\n"
2857
2858                 "%sc_0        = OpSpecConstant %i32 0\n"
2859                 "%sc_1        = OpSpecConstant %i32 0\n"
2860                 "%sc_2        = OpSpecConstant %i32 0\n"
2861                 "%sc_vec3_0   = OpSpecConstantOp %ivec3 CompositeInsert  %sc_0        %ivec3_0     0\n"                                                 // (sc_0, 0, 0)
2862                 "%sc_vec3_1   = OpSpecConstantOp %ivec3 CompositeInsert  %sc_1        %ivec3_0     1\n"                                                 // (0, sc_1, 0)
2863                 "%sc_vec3_2   = OpSpecConstantOp %ivec3 CompositeInsert  %sc_2        %ivec3_0     2\n"                                                 // (0, 0, sc_2)
2864                 "%sc_vec3_0_s = OpSpecConstantOp %ivec3 VectorShuffle    %sc_vec3_0   %vec3_undef  0          0xFFFFFFFF 2\n"   // (sc_0, ???,  0)
2865                 "%sc_vec3_1_s = OpSpecConstantOp %ivec3 VectorShuffle    %sc_vec3_1   %vec3_undef  0xFFFFFFFF 1          0\n"   // (???,  sc_1, 0)
2866                 "%sc_vec3_2_s = OpSpecConstantOp %ivec3 VectorShuffle    %vec3_undef  %sc_vec3_2   5          0xFFFFFFFF 5\n"   // (sc_2, ???,  sc_2)
2867                 "%sc_vec3_01  = OpSpecConstantOp %ivec3 VectorShuffle    %sc_vec3_0_s %sc_vec3_1_s 1 0 4\n"                                             // (0,    sc_0, sc_1)
2868                 "%sc_vec3_012 = OpSpecConstantOp %ivec3 VectorShuffle    %sc_vec3_01  %sc_vec3_2_s 5 1 2\n"                                             // (sc_2, sc_0, sc_1)
2869                 "%sc_ext_0    = OpSpecConstantOp %i32   CompositeExtract %sc_vec3_012              0\n"                                                 // sc_2
2870                 "%sc_ext_1    = OpSpecConstantOp %i32   CompositeExtract %sc_vec3_012              1\n"                                                 // sc_0
2871                 "%sc_ext_2    = OpSpecConstantOp %i32   CompositeExtract %sc_vec3_012              2\n"                                                 // sc_1
2872                 "%sc_sub      = OpSpecConstantOp %i32   ISub             %sc_ext_0    %sc_ext_1\n"                                                              // (sc_2 - sc_0)
2873                 "%sc_final    = OpSpecConstantOp %i32   IMul             %sc_sub      %sc_ext_2\n"                                                              // (sc_2 - sc_0) * sc_1
2874
2875                 "%main      = OpFunction %void None %voidf\n"
2876                 "%label     = OpLabel\n"
2877                 "%idval     = OpLoad %uvec3 %id\n"
2878                 "%x         = OpCompositeExtract %u32 %idval 0\n"
2879                 "%inloc     = OpAccessChain %i32ptr %indata %zero %x\n"
2880                 "%inval     = OpLoad %i32 %inloc\n"
2881                 "%final     = OpIAdd %i32 %inval %sc_final\n"
2882                 "%outloc    = OpAccessChain %i32ptr %outdata %zero %x\n"
2883                 "             OpStore %outloc %final\n"
2884                 "             OpReturn\n"
2885                 "             OpFunctionEnd\n";
2886         spec.inputs.push_back(BufferSp(new Int32Buffer(inputInts)));
2887         spec.outputs.push_back(BufferSp(new Int32Buffer(outputInts3)));
2888         spec.numWorkGroups = IVec3(numElements, 1, 1);
2889         spec.specConstants.append<deInt32>(123);
2890         spec.specConstants.append<deInt32>(56);
2891         spec.specConstants.append<deInt32>(-77);
2892
2893         group->addChild(new SpvAsmComputeShaderCase(testCtx, "vector_related", "VectorShuffle, CompositeExtract, & CompositeInsert", spec));
2894
2895         return group.release();
2896 }
2897
2898 void createOpPhiVartypeTests (de::MovePtr<tcu::TestCaseGroup>& group, tcu::TestContext& testCtx)
2899 {
2900         ComputeShaderSpec       specInt;
2901         ComputeShaderSpec       specFloat;
2902         ComputeShaderSpec       specFloat16;
2903         ComputeShaderSpec       specVec3;
2904         ComputeShaderSpec       specMat4;
2905         ComputeShaderSpec       specArray;
2906         ComputeShaderSpec       specStruct;
2907         de::Random                      rnd                             (deStringHash(group->getName()));
2908         const int                       numElements             = 100;
2909         vector<float>           inputFloats             (numElements, 0);
2910         vector<float>           outputFloats    (numElements, 0);
2911         vector<deFloat16>       inputFloats16   (numElements, 0);
2912         vector<deFloat16>       outputFloats16  (numElements, 0);
2913
2914         fillRandomScalars(rnd, -300.f, 300.f, &inputFloats[0], numElements);
2915
2916         // CPU might not use the same rounding mode as the GPU. Use whole numbers to avoid rounding differences.
2917         floorAll(inputFloats);
2918
2919         for (size_t ndx = 0; ndx < numElements; ++ndx)
2920         {
2921                 // Just check if the value is positive or not
2922                 outputFloats[ndx] = (inputFloats[ndx] > 0) ? 1.0f : -1.0f;
2923         }
2924
2925         for (size_t ndx = 0; ndx < numElements; ++ndx)
2926         {
2927                 inputFloats16[ndx] = tcu::Float16(inputFloats[ndx]).bits();
2928                 outputFloats16[ndx] = tcu::Float16(outputFloats[ndx]).bits();
2929         }
2930
2931         // All of the tests are of the form:
2932         //
2933         // testtype r
2934         //
2935         // if (inputdata > 0)
2936         //   r = 1
2937         // else
2938         //   r = -1
2939         //
2940         // return (float)r
2941
2942         specFloat.assembly =
2943                 string(getComputeAsmShaderPreamble()) +
2944
2945                 "OpSource GLSL 430\n"
2946                 "OpName %main \"main\"\n"
2947                 "OpName %id \"gl_GlobalInvocationID\"\n"
2948
2949                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
2950
2951                 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) + string(getComputeAsmInputOutputBuffer()) +
2952
2953                 "%id = OpVariable %uvec3ptr Input\n"
2954                 "%zero       = OpConstant %i32 0\n"
2955                 "%float_0    = OpConstant %f32 0.0\n"
2956                 "%float_1    = OpConstant %f32 1.0\n"
2957                 "%float_n1   = OpConstant %f32 -1.0\n"
2958
2959                 "%main     = OpFunction %void None %voidf\n"
2960                 "%entry    = OpLabel\n"
2961                 "%idval    = OpLoad %uvec3 %id\n"
2962                 "%x        = OpCompositeExtract %u32 %idval 0\n"
2963                 "%inloc    = OpAccessChain %f32ptr %indata %zero %x\n"
2964                 "%inval    = OpLoad %f32 %inloc\n"
2965
2966                 "%comp     = OpFOrdGreaterThan %bool %inval %float_0\n"
2967                 "            OpSelectionMerge %cm None\n"
2968                 "            OpBranchConditional %comp %tb %fb\n"
2969                 "%tb       = OpLabel\n"
2970                 "            OpBranch %cm\n"
2971                 "%fb       = OpLabel\n"
2972                 "            OpBranch %cm\n"
2973                 "%cm       = OpLabel\n"
2974                 "%res      = OpPhi %f32 %float_1 %tb %float_n1 %fb\n"
2975
2976                 "%outloc   = OpAccessChain %f32ptr %outdata %zero %x\n"
2977                 "            OpStore %outloc %res\n"
2978                 "            OpReturn\n"
2979
2980                 "            OpFunctionEnd\n";
2981         specFloat.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
2982         specFloat.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
2983         specFloat.numWorkGroups = IVec3(numElements, 1, 1);
2984
2985         specFloat16.assembly =
2986                 "OpCapability Shader\n"
2987                 "OpCapability StorageUniformBufferBlock16\n"
2988                 "OpExtension \"SPV_KHR_16bit_storage\"\n"
2989                 "OpMemoryModel Logical GLSL450\n"
2990                 "OpEntryPoint GLCompute %main \"main\" %id\n"
2991                 "OpExecutionMode %main LocalSize 1 1 1\n"
2992
2993                 "OpSource GLSL 430\n"
2994                 "OpName %main \"main\"\n"
2995                 "OpName %id \"gl_GlobalInvocationID\"\n"
2996
2997                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
2998
2999                 "OpDecorate %buf BufferBlock\n"
3000                 "OpDecorate %indata DescriptorSet 0\n"
3001                 "OpDecorate %indata Binding 0\n"
3002                 "OpDecorate %outdata DescriptorSet 0\n"
3003                 "OpDecorate %outdata Binding 1\n"
3004                 "OpDecorate %f16arr ArrayStride 2\n"
3005                 "OpMemberDecorate %buf 0 Offset 0\n"
3006
3007                 "%f16      = OpTypeFloat 16\n"
3008                 "%f16ptr   = OpTypePointer Uniform %f16\n"
3009                 "%f16arr   = OpTypeRuntimeArray %f16\n"
3010
3011                 + string(getComputeAsmCommonTypes()) +
3012
3013                 "%buf      = OpTypeStruct %f16arr\n"
3014                 "%bufptr   = OpTypePointer Uniform %buf\n"
3015                 "%indata   = OpVariable %bufptr Uniform\n"
3016                 "%outdata  = OpVariable %bufptr Uniform\n"
3017
3018                 "%id       = OpVariable %uvec3ptr Input\n"
3019                 "%zero     = OpConstant %i32 0\n"
3020                 "%float_0  = OpConstant %f16 0.0\n"
3021                 "%float_1  = OpConstant %f16 1.0\n"
3022                 "%float_n1 = OpConstant %f16 -1.0\n"
3023
3024                 "%main     = OpFunction %void None %voidf\n"
3025                 "%entry    = OpLabel\n"
3026                 "%idval    = OpLoad %uvec3 %id\n"
3027                 "%x        = OpCompositeExtract %u32 %idval 0\n"
3028                 "%inloc    = OpAccessChain %f16ptr %indata %zero %x\n"
3029                 "%inval    = OpLoad %f16 %inloc\n"
3030
3031                 "%comp     = OpFOrdGreaterThan %bool %inval %float_0\n"
3032                 "            OpSelectionMerge %cm None\n"
3033                 "            OpBranchConditional %comp %tb %fb\n"
3034                 "%tb       = OpLabel\n"
3035                 "            OpBranch %cm\n"
3036                 "%fb       = OpLabel\n"
3037                 "            OpBranch %cm\n"
3038                 "%cm       = OpLabel\n"
3039                 "%res      = OpPhi %f16 %float_1 %tb %float_n1 %fb\n"
3040
3041                 "%outloc   = OpAccessChain %f16ptr %outdata %zero %x\n"
3042                 "            OpStore %outloc %res\n"
3043                 "            OpReturn\n"
3044
3045                 "            OpFunctionEnd\n";
3046         specFloat16.inputs.push_back(BufferSp(new Float16Buffer(inputFloats16)));
3047         specFloat16.outputs.push_back(BufferSp(new Float16Buffer(outputFloats16)));
3048         specFloat16.numWorkGroups = IVec3(numElements, 1, 1);
3049         specFloat16.requestedVulkanFeatures.ext16BitStorage = EXT16BITSTORAGEFEATURES_UNIFORM_BUFFER_BLOCK;
3050         specFloat16.requestedVulkanFeatures.extFloat16Int8 = EXTFLOAT16INT8FEATURES_FLOAT16;
3051
3052         specMat4.assembly =
3053                 string(getComputeAsmShaderPreamble()) +
3054
3055                 "OpSource GLSL 430\n"
3056                 "OpName %main \"main\"\n"
3057                 "OpName %id \"gl_GlobalInvocationID\"\n"
3058
3059                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
3060
3061                 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) + string(getComputeAsmInputOutputBuffer()) +
3062
3063                 "%id = OpVariable %uvec3ptr Input\n"
3064                 "%v4f32      = OpTypeVector %f32 4\n"
3065                 "%mat4v4f32  = OpTypeMatrix %v4f32 4\n"
3066                 "%zero       = OpConstant %i32 0\n"
3067                 "%float_0    = OpConstant %f32 0.0\n"
3068                 "%float_1    = OpConstant %f32 1.0\n"
3069                 "%float_n1   = OpConstant %f32 -1.0\n"
3070                 "%m11        = OpConstantComposite %v4f32 %float_1 %float_0 %float_0 %float_0\n"
3071                 "%m12        = OpConstantComposite %v4f32 %float_0 %float_1 %float_0 %float_0\n"
3072                 "%m13        = OpConstantComposite %v4f32 %float_0 %float_0 %float_1 %float_0\n"
3073                 "%m14        = OpConstantComposite %v4f32 %float_0 %float_0 %float_0 %float_1\n"
3074                 "%m1         = OpConstantComposite %mat4v4f32 %m11 %m12 %m13 %m14\n"
3075                 "%m21        = OpConstantComposite %v4f32 %float_n1 %float_0 %float_0 %float_0\n"
3076                 "%m22        = OpConstantComposite %v4f32 %float_0 %float_n1 %float_0 %float_0\n"
3077                 "%m23        = OpConstantComposite %v4f32 %float_0 %float_0 %float_n1 %float_0\n"
3078                 "%m24        = OpConstantComposite %v4f32 %float_0 %float_0 %float_0 %float_n1\n"
3079                 "%m2         = OpConstantComposite %mat4v4f32 %m21 %m22 %m23 %m24\n"
3080
3081                 "%main     = OpFunction %void None %voidf\n"
3082                 "%entry    = OpLabel\n"
3083                 "%idval    = OpLoad %uvec3 %id\n"
3084                 "%x        = OpCompositeExtract %u32 %idval 0\n"
3085                 "%inloc    = OpAccessChain %f32ptr %indata %zero %x\n"
3086                 "%inval    = OpLoad %f32 %inloc\n"
3087
3088                 "%comp     = OpFOrdGreaterThan %bool %inval %float_0\n"
3089                 "            OpSelectionMerge %cm None\n"
3090                 "            OpBranchConditional %comp %tb %fb\n"
3091                 "%tb       = OpLabel\n"
3092                 "            OpBranch %cm\n"
3093                 "%fb       = OpLabel\n"
3094                 "            OpBranch %cm\n"
3095                 "%cm       = OpLabel\n"
3096                 "%mres     = OpPhi %mat4v4f32 %m1 %tb %m2 %fb\n"
3097                 "%res      = OpCompositeExtract %f32 %mres 2 2\n"
3098
3099                 "%outloc   = OpAccessChain %f32ptr %outdata %zero %x\n"
3100                 "            OpStore %outloc %res\n"
3101                 "            OpReturn\n"
3102
3103                 "            OpFunctionEnd\n";
3104         specMat4.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
3105         specMat4.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
3106         specMat4.numWorkGroups = IVec3(numElements, 1, 1);
3107
3108         specVec3.assembly =
3109                 string(getComputeAsmShaderPreamble()) +
3110
3111                 "OpSource GLSL 430\n"
3112                 "OpName %main \"main\"\n"
3113                 "OpName %id \"gl_GlobalInvocationID\"\n"
3114
3115                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
3116
3117                 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) + string(getComputeAsmInputOutputBuffer()) +
3118
3119                 "%id = OpVariable %uvec3ptr Input\n"
3120                 "%zero       = OpConstant %i32 0\n"
3121                 "%float_0    = OpConstant %f32 0.0\n"
3122                 "%float_1    = OpConstant %f32 1.0\n"
3123                 "%float_n1   = OpConstant %f32 -1.0\n"
3124                 "%v1         = OpConstantComposite %fvec3 %float_1 %float_1 %float_1\n"
3125                 "%v2         = OpConstantComposite %fvec3 %float_n1 %float_n1 %float_n1\n"
3126
3127                 "%main     = OpFunction %void None %voidf\n"
3128                 "%entry    = OpLabel\n"
3129                 "%idval    = OpLoad %uvec3 %id\n"
3130                 "%x        = OpCompositeExtract %u32 %idval 0\n"
3131                 "%inloc    = OpAccessChain %f32ptr %indata %zero %x\n"
3132                 "%inval    = OpLoad %f32 %inloc\n"
3133
3134                 "%comp     = OpFOrdGreaterThan %bool %inval %float_0\n"
3135                 "            OpSelectionMerge %cm None\n"
3136                 "            OpBranchConditional %comp %tb %fb\n"
3137                 "%tb       = OpLabel\n"
3138                 "            OpBranch %cm\n"
3139                 "%fb       = OpLabel\n"
3140                 "            OpBranch %cm\n"
3141                 "%cm       = OpLabel\n"
3142                 "%vres     = OpPhi %fvec3 %v1 %tb %v2 %fb\n"
3143                 "%res      = OpCompositeExtract %f32 %vres 2\n"
3144
3145                 "%outloc   = OpAccessChain %f32ptr %outdata %zero %x\n"
3146                 "            OpStore %outloc %res\n"
3147                 "            OpReturn\n"
3148
3149                 "            OpFunctionEnd\n";
3150         specVec3.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
3151         specVec3.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
3152         specVec3.numWorkGroups = IVec3(numElements, 1, 1);
3153
3154         specInt.assembly =
3155                 string(getComputeAsmShaderPreamble()) +
3156
3157                 "OpSource GLSL 430\n"
3158                 "OpName %main \"main\"\n"
3159                 "OpName %id \"gl_GlobalInvocationID\"\n"
3160
3161                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
3162
3163                 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) + string(getComputeAsmInputOutputBuffer()) +
3164
3165                 "%id = OpVariable %uvec3ptr Input\n"
3166                 "%zero       = OpConstant %i32 0\n"
3167                 "%float_0    = OpConstant %f32 0.0\n"
3168                 "%i1         = OpConstant %i32 1\n"
3169                 "%i2         = OpConstant %i32 -1\n"
3170
3171                 "%main     = OpFunction %void None %voidf\n"
3172                 "%entry    = OpLabel\n"
3173                 "%idval    = OpLoad %uvec3 %id\n"
3174                 "%x        = OpCompositeExtract %u32 %idval 0\n"
3175                 "%inloc    = OpAccessChain %f32ptr %indata %zero %x\n"
3176                 "%inval    = OpLoad %f32 %inloc\n"
3177
3178                 "%comp     = OpFOrdGreaterThan %bool %inval %float_0\n"
3179                 "            OpSelectionMerge %cm None\n"
3180                 "            OpBranchConditional %comp %tb %fb\n"
3181                 "%tb       = OpLabel\n"
3182                 "            OpBranch %cm\n"
3183                 "%fb       = OpLabel\n"
3184                 "            OpBranch %cm\n"
3185                 "%cm       = OpLabel\n"
3186                 "%ires     = OpPhi %i32 %i1 %tb %i2 %fb\n"
3187                 "%res      = OpConvertSToF %f32 %ires\n"
3188
3189                 "%outloc   = OpAccessChain %f32ptr %outdata %zero %x\n"
3190                 "            OpStore %outloc %res\n"
3191                 "            OpReturn\n"
3192
3193                 "            OpFunctionEnd\n";
3194         specInt.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
3195         specInt.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
3196         specInt.numWorkGroups = IVec3(numElements, 1, 1);
3197
3198         specArray.assembly =
3199                 string(getComputeAsmShaderPreamble()) +
3200
3201                 "OpSource GLSL 430\n"
3202                 "OpName %main \"main\"\n"
3203                 "OpName %id \"gl_GlobalInvocationID\"\n"
3204
3205                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
3206
3207                 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) + string(getComputeAsmInputOutputBuffer()) +
3208
3209                 "%id = OpVariable %uvec3ptr Input\n"
3210                 "%zero       = OpConstant %i32 0\n"
3211                 "%u7         = OpConstant %u32 7\n"
3212                 "%float_0    = OpConstant %f32 0.0\n"
3213                 "%float_1    = OpConstant %f32 1.0\n"
3214                 "%float_n1   = OpConstant %f32 -1.0\n"
3215                 "%f32a7      = OpTypeArray %f32 %u7\n"
3216                 "%a1         = OpConstantComposite %f32a7 %float_1 %float_1 %float_1 %float_1 %float_1 %float_1 %float_1\n"
3217                 "%a2         = OpConstantComposite %f32a7 %float_n1 %float_n1 %float_n1 %float_n1 %float_n1 %float_n1 %float_n1\n"
3218                 "%main     = OpFunction %void None %voidf\n"
3219                 "%entry    = OpLabel\n"
3220                 "%idval    = OpLoad %uvec3 %id\n"
3221                 "%x        = OpCompositeExtract %u32 %idval 0\n"
3222                 "%inloc    = OpAccessChain %f32ptr %indata %zero %x\n"
3223                 "%inval    = OpLoad %f32 %inloc\n"
3224
3225                 "%comp     = OpFOrdGreaterThan %bool %inval %float_0\n"
3226                 "            OpSelectionMerge %cm None\n"
3227                 "            OpBranchConditional %comp %tb %fb\n"
3228                 "%tb       = OpLabel\n"
3229                 "            OpBranch %cm\n"
3230                 "%fb       = OpLabel\n"
3231                 "            OpBranch %cm\n"
3232                 "%cm       = OpLabel\n"
3233                 "%ares     = OpPhi %f32a7 %a1 %tb %a2 %fb\n"
3234                 "%res      = OpCompositeExtract %f32 %ares 5\n"
3235
3236                 "%outloc   = OpAccessChain %f32ptr %outdata %zero %x\n"
3237                 "            OpStore %outloc %res\n"
3238                 "            OpReturn\n"
3239
3240                 "            OpFunctionEnd\n";
3241         specArray.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
3242         specArray.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
3243         specArray.numWorkGroups = IVec3(numElements, 1, 1);
3244
3245         specStruct.assembly =
3246                 string(getComputeAsmShaderPreamble()) +
3247
3248                 "OpSource GLSL 430\n"
3249                 "OpName %main \"main\"\n"
3250                 "OpName %id \"gl_GlobalInvocationID\"\n"
3251
3252                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
3253
3254                 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) + string(getComputeAsmInputOutputBuffer()) +
3255
3256                 "%id = OpVariable %uvec3ptr Input\n"
3257                 "%zero       = OpConstant %i32 0\n"
3258                 "%float_0    = OpConstant %f32 0.0\n"
3259                 "%float_1    = OpConstant %f32 1.0\n"
3260                 "%float_n1   = OpConstant %f32 -1.0\n"
3261
3262                 "%v2f32      = OpTypeVector %f32 2\n"
3263                 "%Data2      = OpTypeStruct %f32 %v2f32\n"
3264                 "%Data       = OpTypeStruct %Data2 %f32\n"
3265
3266                 "%in1a       = OpConstantComposite %v2f32 %float_1 %float_1\n"
3267                 "%in1b       = OpConstantComposite %Data2 %float_1 %in1a\n"
3268                 "%s1         = OpConstantComposite %Data %in1b %float_1\n"
3269                 "%in2a       = OpConstantComposite %v2f32 %float_n1 %float_n1\n"
3270                 "%in2b       = OpConstantComposite %Data2 %float_n1 %in2a\n"
3271                 "%s2         = OpConstantComposite %Data %in2b %float_n1\n"
3272
3273                 "%main     = OpFunction %void None %voidf\n"
3274                 "%entry    = OpLabel\n"
3275                 "%idval    = OpLoad %uvec3 %id\n"
3276                 "%x        = OpCompositeExtract %u32 %idval 0\n"
3277                 "%inloc    = OpAccessChain %f32ptr %indata %zero %x\n"
3278                 "%inval    = OpLoad %f32 %inloc\n"
3279
3280                 "%comp     = OpFOrdGreaterThan %bool %inval %float_0\n"
3281                 "            OpSelectionMerge %cm None\n"
3282                 "            OpBranchConditional %comp %tb %fb\n"
3283                 "%tb       = OpLabel\n"
3284                 "            OpBranch %cm\n"
3285                 "%fb       = OpLabel\n"
3286                 "            OpBranch %cm\n"
3287                 "%cm       = OpLabel\n"
3288                 "%sres     = OpPhi %Data %s1 %tb %s2 %fb\n"
3289                 "%res      = OpCompositeExtract %f32 %sres 0 0\n"
3290
3291                 "%outloc   = OpAccessChain %f32ptr %outdata %zero %x\n"
3292                 "            OpStore %outloc %res\n"
3293                 "            OpReturn\n"
3294
3295                 "            OpFunctionEnd\n";
3296         specStruct.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
3297         specStruct.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
3298         specStruct.numWorkGroups = IVec3(numElements, 1, 1);
3299
3300         group->addChild(new SpvAsmComputeShaderCase(testCtx, "vartype_int", "OpPhi with int variables", specInt));
3301         group->addChild(new SpvAsmComputeShaderCase(testCtx, "vartype_float", "OpPhi with float variables", specFloat));
3302         group->addChild(new SpvAsmComputeShaderCase(testCtx, "vartype_float16", "OpPhi with 16bit float variables", specFloat16));
3303         group->addChild(new SpvAsmComputeShaderCase(testCtx, "vartype_vec3", "OpPhi with vec3 variables", specVec3));
3304         group->addChild(new SpvAsmComputeShaderCase(testCtx, "vartype_mat4", "OpPhi with mat4 variables", specMat4));
3305         group->addChild(new SpvAsmComputeShaderCase(testCtx, "vartype_array", "OpPhi with array variables", specArray));
3306         group->addChild(new SpvAsmComputeShaderCase(testCtx, "vartype_struct", "OpPhi with struct variables", specStruct));
3307 }
3308
3309 string generateConstantDefinitions (int count)
3310 {
3311         std::ostringstream      r;
3312         for (int i = 0; i < count; i++)
3313                 r << "%cf" << (i * 10 + 5) << " = OpConstant %f32 " <<(i * 10 + 5) << ".0\n";
3314         r << "\n";
3315         return r.str();
3316 }
3317
3318 string generateSwitchCases (int count)
3319 {
3320         std::ostringstream      r;
3321         for (int i = 0; i < count; i++)
3322                 r << " " << i << " %case" << i;
3323         r << "\n";
3324         return r.str();
3325 }
3326
3327 string generateSwitchTargets (int count)
3328 {
3329         std::ostringstream      r;
3330         for (int i = 0; i < count; i++)
3331                 r << "%case" << i << " = OpLabel\n            OpBranch %phi\n";
3332         r << "\n";
3333         return r.str();
3334 }
3335
3336 string generateOpPhiParams (int count)
3337 {
3338         std::ostringstream      r;
3339         for (int i = 0; i < count; i++)
3340                 r << " %cf" << (i * 10 + 5) << " %case" << i;
3341         r << "\n";
3342         return r.str();
3343 }
3344
3345 string generateIntWidth (int value)
3346 {
3347         std::ostringstream      r;
3348         r << value;
3349         return r.str();
3350 }
3351
3352 // Expand input string by injecting "ABC" between the input
3353 // string characters. The acc/add/treshold parameters are used
3354 // to skip some of the injections to make the result less
3355 // uniform (and a lot shorter).
3356 string expandOpPhiCase5 (const string& s, int &acc, int add, int treshold)
3357 {
3358         std::ostringstream      res;
3359         const char*                     p = s.c_str();
3360
3361         while (*p)
3362         {
3363                 res << *p;
3364                 acc += add;
3365                 if (acc > treshold)
3366                 {
3367                         acc -= treshold;
3368                         res << "ABC";
3369                 }
3370                 p++;
3371         }
3372         return res.str();
3373 }
3374
3375 // Calculate expected result based on the code string
3376 float calcOpPhiCase5 (float val, const string& s)
3377 {
3378         const char*             p               = s.c_str();
3379         float                   x[8];
3380         bool                    b[8];
3381         const float             tv[8]   = { 0.5f, 1.5f, 3.5f, 7.5f, 15.5f, 31.5f, 63.5f, 127.5f };
3382         const float             v               = deFloatAbs(val);
3383         float                   res             = 0;
3384         int                             depth   = -1;
3385         int                             skip    = 0;
3386
3387         for (int i = 7; i >= 0; --i)
3388                 x[i] = std::fmod((float)v, (float)(2 << i));
3389         for (int i = 7; i >= 0; --i)
3390                 b[i] = x[i] > tv[i];
3391
3392         while (*p)
3393         {
3394                 if (*p == 'A')
3395                 {
3396                         depth++;
3397                         if (skip == 0 && b[depth])
3398                         {
3399                                 res++;
3400                         }
3401                         else
3402                                 skip++;
3403                 }
3404                 if (*p == 'B')
3405                 {
3406                         if (skip)
3407                                 skip--;
3408                         if (b[depth] || skip)
3409                                 skip++;
3410                 }
3411                 if (*p == 'C')
3412                 {
3413                         depth--;
3414                         if (skip)
3415                                 skip--;
3416                 }
3417                 p++;
3418         }
3419         return res;
3420 }
3421
3422 // In the code string, the letters represent the following:
3423 //
3424 // A:
3425 //     if (certain bit is set)
3426 //     {
3427 //       result++;
3428 //
3429 // B:
3430 //     } else {
3431 //
3432 // C:
3433 //     }
3434 //
3435 // examples:
3436 // AABCBC leads to if(){r++;if(){r++;}else{}}else{}
3437 // ABABCC leads to if(){r++;}else{if(){r++;}else{}}
3438 // ABCABC leads to if(){r++;}else{}if(){r++;}else{}
3439 //
3440 // Code generation gets a bit complicated due to the else-branches,
3441 // which do not generate new values. Thus, the generator needs to
3442 // keep track of the previous variable change seen by the else
3443 // branch.
3444 string generateOpPhiCase5 (const string& s)
3445 {
3446         std::stack<int>                         idStack;
3447         std::stack<std::string>         value;
3448         std::stack<std::string>         valueLabel;
3449         std::stack<std::string>         mergeLeft;
3450         std::stack<std::string>         mergeRight;
3451         std::ostringstream                      res;
3452         const char*                                     p                       = s.c_str();
3453         int                                                     depth           = -1;
3454         int                                                     currId          = 0;
3455         int                                                     iter            = 0;
3456
3457         idStack.push(-1);
3458         value.push("%f32_0");
3459         valueLabel.push("%f32_0 %entry");
3460
3461         while (*p)
3462         {
3463                 if (*p == 'A')
3464                 {
3465                         depth++;
3466                         currId = iter;
3467                         idStack.push(currId);
3468                         res << "\tOpSelectionMerge %m" << currId << " None\n";
3469                         res << "\tOpBranchConditional %b" << depth << " %t" << currId << " %f" << currId << "\n";
3470                         res << "%t" << currId << " = OpLabel\n";
3471                         res << "%rt" << currId << " = OpFAdd %f32 " << value.top() << " %f32_1\n";
3472                         std::ostringstream tag;
3473                         tag << "%rt" << currId;
3474                         value.push(tag.str());
3475                         tag << " %t" << currId;
3476                         valueLabel.push(tag.str());
3477                 }
3478
3479                 if (*p == 'B')
3480                 {
3481                         mergeLeft.push(valueLabel.top());
3482                         value.pop();
3483                         valueLabel.pop();
3484                         res << "\tOpBranch %m" << currId << "\n";
3485                         res << "%f" << currId << " = OpLabel\n";
3486                         std::ostringstream tag;
3487                         tag << value.top() << " %f" << currId;
3488                         valueLabel.pop();
3489                         valueLabel.push(tag.str());
3490                 }
3491
3492                 if (*p == 'C')
3493                 {
3494                         mergeRight.push(valueLabel.top());
3495                         res << "\tOpBranch %m" << currId << "\n";
3496                         res << "%m" << currId << " = OpLabel\n";
3497                         if (*(p + 1) == 0)
3498                                 res << "%res"; // last result goes to %res
3499                         else
3500                                 res << "%rm" << currId;
3501                         res << " = OpPhi %f32  " << mergeLeft.top() << "  " << mergeRight.top() << "\n";
3502                         std::ostringstream tag;
3503                         tag << "%rm" << currId;
3504                         value.pop();
3505                         value.push(tag.str());
3506                         tag << " %m" << currId;
3507                         valueLabel.pop();
3508                         valueLabel.push(tag.str());
3509                         mergeLeft.pop();
3510                         mergeRight.pop();
3511                         depth--;
3512                         idStack.pop();
3513                         currId = idStack.top();
3514                 }
3515                 p++;
3516                 iter++;
3517         }
3518         return res.str();
3519 }
3520
3521 tcu::TestCaseGroup* createOpPhiGroup (tcu::TestContext& testCtx)
3522 {
3523         de::MovePtr<tcu::TestCaseGroup> group                   (new tcu::TestCaseGroup(testCtx, "opphi", "Test the OpPhi instruction"));
3524         ComputeShaderSpec                               spec1;
3525         ComputeShaderSpec                               spec2;
3526         ComputeShaderSpec                               spec3;
3527         ComputeShaderSpec                               spec4;
3528         ComputeShaderSpec                               spec5;
3529         de::Random                                              rnd                             (deStringHash(group->getName()));
3530         const int                                               numElements             = 100;
3531         vector<float>                                   inputFloats             (numElements, 0);
3532         vector<float>                                   outputFloats1   (numElements, 0);
3533         vector<float>                                   outputFloats2   (numElements, 0);
3534         vector<float>                                   outputFloats3   (numElements, 0);
3535         vector<float>                                   outputFloats4   (numElements, 0);
3536         vector<float>                                   outputFloats5   (numElements, 0);
3537         std::string                                             codestring              = "ABC";
3538         const int                                               test4Width              = 1024;
3539
3540         // Build case 5 code string. Each iteration makes the hierarchy more complicated.
3541         // 9 iterations with (7, 24) parameters makes the hierarchy 8 deep with about 1500 lines of
3542         // shader code.
3543         for (int i = 0, acc = 0; i < 9; i++)
3544                 codestring = expandOpPhiCase5(codestring, acc, 7, 24);
3545
3546         fillRandomScalars(rnd, -300.f, 300.f, &inputFloats[0], numElements);
3547
3548         // CPU might not use the same rounding mode as the GPU. Use whole numbers to avoid rounding differences.
3549         floorAll(inputFloats);
3550
3551         for (size_t ndx = 0; ndx < numElements; ++ndx)
3552         {
3553                 switch (ndx % 3)
3554                 {
3555                         case 0:         outputFloats1[ndx] = inputFloats[ndx] + 5.5f;   break;
3556                         case 1:         outputFloats1[ndx] = inputFloats[ndx] + 20.5f;  break;
3557                         case 2:         outputFloats1[ndx] = inputFloats[ndx] + 1.75f;  break;
3558                         default:        break;
3559                 }
3560                 outputFloats2[ndx] = inputFloats[ndx] + 6.5f * 3;
3561                 outputFloats3[ndx] = 8.5f - inputFloats[ndx];
3562
3563                 int index4 = (int)deFloor(deAbs((float)ndx * inputFloats[ndx]));
3564                 outputFloats4[ndx] = (float)(index4 % test4Width) * 10.0f + 5.0f;
3565
3566                 outputFloats5[ndx] = calcOpPhiCase5(inputFloats[ndx], codestring);
3567         }
3568
3569         spec1.assembly =
3570                 string(getComputeAsmShaderPreamble()) +
3571
3572                 "OpSource GLSL 430\n"
3573                 "OpName %main \"main\"\n"
3574                 "OpName %id \"gl_GlobalInvocationID\"\n"
3575
3576                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
3577
3578                 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) + string(getComputeAsmInputOutputBuffer()) +
3579
3580                 "%id = OpVariable %uvec3ptr Input\n"
3581                 "%zero       = OpConstant %i32 0\n"
3582                 "%three      = OpConstant %u32 3\n"
3583                 "%constf5p5  = OpConstant %f32 5.5\n"
3584                 "%constf20p5 = OpConstant %f32 20.5\n"
3585                 "%constf1p75 = OpConstant %f32 1.75\n"
3586                 "%constf8p5  = OpConstant %f32 8.5\n"
3587                 "%constf6p5  = OpConstant %f32 6.5\n"
3588
3589                 "%main     = OpFunction %void None %voidf\n"
3590                 "%entry    = OpLabel\n"
3591                 "%idval    = OpLoad %uvec3 %id\n"
3592                 "%x        = OpCompositeExtract %u32 %idval 0\n"
3593                 "%selector = OpUMod %u32 %x %three\n"
3594                 "            OpSelectionMerge %phi None\n"
3595                 "            OpSwitch %selector %default 0 %case0 1 %case1 2 %case2\n"
3596
3597                 // Case 1 before OpPhi.
3598                 "%case1    = OpLabel\n"
3599                 "            OpBranch %phi\n"
3600
3601                 "%default  = OpLabel\n"
3602                 "            OpUnreachable\n"
3603
3604                 "%phi      = OpLabel\n"
3605                 "%operand  = OpPhi %f32   %constf1p75 %case2   %constf20p5 %case1   %constf5p5 %case0\n" // not in the order of blocks
3606                 "%inloc    = OpAccessChain %f32ptr %indata %zero %x\n"
3607                 "%inval    = OpLoad %f32 %inloc\n"
3608                 "%add      = OpFAdd %f32 %inval %operand\n"
3609                 "%outloc   = OpAccessChain %f32ptr %outdata %zero %x\n"
3610                 "            OpStore %outloc %add\n"
3611                 "            OpReturn\n"
3612
3613                 // Case 0 after OpPhi.
3614                 "%case0    = OpLabel\n"
3615                 "            OpBranch %phi\n"
3616
3617
3618                 // Case 2 after OpPhi.
3619                 "%case2    = OpLabel\n"
3620                 "            OpBranch %phi\n"
3621
3622                 "            OpFunctionEnd\n";
3623         spec1.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
3624         spec1.outputs.push_back(BufferSp(new Float32Buffer(outputFloats1)));
3625         spec1.numWorkGroups = IVec3(numElements, 1, 1);
3626
3627         group->addChild(new SpvAsmComputeShaderCase(testCtx, "block", "out-of-order and unreachable blocks for OpPhi", spec1));
3628
3629         spec2.assembly =
3630                 string(getComputeAsmShaderPreamble()) +
3631
3632                 "OpName %main \"main\"\n"
3633                 "OpName %id \"gl_GlobalInvocationID\"\n"
3634
3635                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
3636
3637                 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) + string(getComputeAsmInputOutputBuffer()) +
3638
3639                 "%id         = OpVariable %uvec3ptr Input\n"
3640                 "%zero       = OpConstant %i32 0\n"
3641                 "%one        = OpConstant %i32 1\n"
3642                 "%three      = OpConstant %i32 3\n"
3643                 "%constf6p5  = OpConstant %f32 6.5\n"
3644
3645                 "%main       = OpFunction %void None %voidf\n"
3646                 "%entry      = OpLabel\n"
3647                 "%idval      = OpLoad %uvec3 %id\n"
3648                 "%x          = OpCompositeExtract %u32 %idval 0\n"
3649                 "%inloc      = OpAccessChain %f32ptr %indata %zero %x\n"
3650                 "%outloc     = OpAccessChain %f32ptr %outdata %zero %x\n"
3651                 "%inval      = OpLoad %f32 %inloc\n"
3652                 "              OpBranch %phi\n"
3653
3654                 "%phi        = OpLabel\n"
3655                 "%step       = OpPhi %i32 %zero  %entry %step_next  %phi\n"
3656                 "%accum      = OpPhi %f32 %inval %entry %accum_next %phi\n"
3657                 "%step_next  = OpIAdd %i32 %step %one\n"
3658                 "%accum_next = OpFAdd %f32 %accum %constf6p5\n"
3659                 "%still_loop = OpSLessThan %bool %step %three\n"
3660                 "              OpLoopMerge %exit %phi None\n"
3661                 "              OpBranchConditional %still_loop %phi %exit\n"
3662
3663                 "%exit       = OpLabel\n"
3664                 "              OpStore %outloc %accum\n"
3665                 "              OpReturn\n"
3666                 "              OpFunctionEnd\n";
3667         spec2.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
3668         spec2.outputs.push_back(BufferSp(new Float32Buffer(outputFloats2)));
3669         spec2.numWorkGroups = IVec3(numElements, 1, 1);
3670
3671         group->addChild(new SpvAsmComputeShaderCase(testCtx, "induction", "The usual way induction variables are handled in LLVM IR", spec2));
3672
3673         spec3.assembly =
3674                 string(getComputeAsmShaderPreamble()) +
3675
3676                 "OpName %main \"main\"\n"
3677                 "OpName %id \"gl_GlobalInvocationID\"\n"
3678
3679                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
3680
3681                 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) + string(getComputeAsmInputOutputBuffer()) +
3682
3683                 "%f32ptr_f   = OpTypePointer Function %f32\n"
3684                 "%id         = OpVariable %uvec3ptr Input\n"
3685                 "%true       = OpConstantTrue %bool\n"
3686                 "%false      = OpConstantFalse %bool\n"
3687                 "%zero       = OpConstant %i32 0\n"
3688                 "%constf8p5  = OpConstant %f32 8.5\n"
3689
3690                 "%main       = OpFunction %void None %voidf\n"
3691                 "%entry      = OpLabel\n"
3692                 "%b          = OpVariable %f32ptr_f Function %constf8p5\n"
3693                 "%idval      = OpLoad %uvec3 %id\n"
3694                 "%x          = OpCompositeExtract %u32 %idval 0\n"
3695                 "%inloc      = OpAccessChain %f32ptr %indata %zero %x\n"
3696                 "%outloc     = OpAccessChain %f32ptr %outdata %zero %x\n"
3697                 "%a_init     = OpLoad %f32 %inloc\n"
3698                 "%b_init     = OpLoad %f32 %b\n"
3699                 "              OpBranch %phi\n"
3700
3701                 "%phi        = OpLabel\n"
3702                 "%still_loop = OpPhi %bool %true   %entry %false  %phi\n"
3703                 "%a_next     = OpPhi %f32  %a_init %entry %b_next %phi\n"
3704                 "%b_next     = OpPhi %f32  %b_init %entry %a_next %phi\n"
3705                 "              OpLoopMerge %exit %phi None\n"
3706                 "              OpBranchConditional %still_loop %phi %exit\n"
3707
3708                 "%exit       = OpLabel\n"
3709                 "%sub        = OpFSub %f32 %a_next %b_next\n"
3710                 "              OpStore %outloc %sub\n"
3711                 "              OpReturn\n"
3712                 "              OpFunctionEnd\n";
3713         spec3.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
3714         spec3.outputs.push_back(BufferSp(new Float32Buffer(outputFloats3)));
3715         spec3.numWorkGroups = IVec3(numElements, 1, 1);
3716
3717         group->addChild(new SpvAsmComputeShaderCase(testCtx, "swap", "Swap the values of two variables using OpPhi", spec3));
3718
3719         spec4.assembly =
3720                 "OpCapability Shader\n"
3721                 "%ext = OpExtInstImport \"GLSL.std.450\"\n"
3722                 "OpMemoryModel Logical GLSL450\n"
3723                 "OpEntryPoint GLCompute %main \"main\" %id\n"
3724                 "OpExecutionMode %main LocalSize 1 1 1\n"
3725
3726                 "OpSource GLSL 430\n"
3727                 "OpName %main \"main\"\n"
3728                 "OpName %id \"gl_GlobalInvocationID\"\n"
3729
3730                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
3731
3732                 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) + string(getComputeAsmInputOutputBuffer()) +
3733
3734                 "%id       = OpVariable %uvec3ptr Input\n"
3735                 "%zero     = OpConstant %i32 0\n"
3736                 "%cimod    = OpConstant %u32 " + generateIntWidth(test4Width) + "\n"
3737
3738                 + generateConstantDefinitions(test4Width) +
3739
3740                 "%main     = OpFunction %void None %voidf\n"
3741                 "%entry    = OpLabel\n"
3742                 "%idval    = OpLoad %uvec3 %id\n"
3743                 "%x        = OpCompositeExtract %u32 %idval 0\n"
3744                 "%inloc    = OpAccessChain %f32ptr %indata %zero %x\n"
3745                 "%inval    = OpLoad %f32 %inloc\n"
3746                 "%xf       = OpConvertUToF %f32 %x\n"
3747                 "%xm       = OpFMul %f32 %xf %inval\n"
3748                 "%xa       = OpExtInst %f32 %ext FAbs %xm\n"
3749                 "%xi       = OpConvertFToU %u32 %xa\n"
3750                 "%selector = OpUMod %u32 %xi %cimod\n"
3751                 "            OpSelectionMerge %phi None\n"
3752                 "            OpSwitch %selector %default "
3753
3754                 + generateSwitchCases(test4Width) +
3755
3756                 "%default  = OpLabel\n"
3757                 "            OpUnreachable\n"
3758
3759                 + generateSwitchTargets(test4Width) +
3760
3761                 "%phi      = OpLabel\n"
3762                 "%result   = OpPhi %f32"
3763
3764                 + generateOpPhiParams(test4Width) +
3765
3766                 "%outloc   = OpAccessChain %f32ptr %outdata %zero %x\n"
3767                 "            OpStore %outloc %result\n"
3768                 "            OpReturn\n"
3769
3770                 "            OpFunctionEnd\n";
3771         spec4.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
3772         spec4.outputs.push_back(BufferSp(new Float32Buffer(outputFloats4)));
3773         spec4.numWorkGroups = IVec3(numElements, 1, 1);
3774
3775         group->addChild(new SpvAsmComputeShaderCase(testCtx, "wide", "OpPhi with a lot of parameters", spec4));
3776
3777         spec5.assembly =
3778                 "OpCapability Shader\n"
3779                 "%ext      = OpExtInstImport \"GLSL.std.450\"\n"
3780                 "OpMemoryModel Logical GLSL450\n"
3781                 "OpEntryPoint GLCompute %main \"main\" %id\n"
3782                 "OpExecutionMode %main LocalSize 1 1 1\n"
3783                 "%code     = OpString \"" + codestring + "\"\n"
3784
3785                 "OpSource GLSL 430\n"
3786                 "OpName %main \"main\"\n"
3787                 "OpName %id \"gl_GlobalInvocationID\"\n"
3788
3789                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
3790
3791                 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) + string(getComputeAsmInputOutputBuffer()) +
3792
3793                 "%id       = OpVariable %uvec3ptr Input\n"
3794                 "%zero     = OpConstant %i32 0\n"
3795                 "%f32_0    = OpConstant %f32 0.0\n"
3796                 "%f32_0_5  = OpConstant %f32 0.5\n"
3797                 "%f32_1    = OpConstant %f32 1.0\n"
3798                 "%f32_1_5  = OpConstant %f32 1.5\n"
3799                 "%f32_2    = OpConstant %f32 2.0\n"
3800                 "%f32_3_5  = OpConstant %f32 3.5\n"
3801                 "%f32_4    = OpConstant %f32 4.0\n"
3802                 "%f32_7_5  = OpConstant %f32 7.5\n"
3803                 "%f32_8    = OpConstant %f32 8.0\n"
3804                 "%f32_15_5 = OpConstant %f32 15.5\n"
3805                 "%f32_16   = OpConstant %f32 16.0\n"
3806                 "%f32_31_5 = OpConstant %f32 31.5\n"
3807                 "%f32_32   = OpConstant %f32 32.0\n"
3808                 "%f32_63_5 = OpConstant %f32 63.5\n"
3809                 "%f32_64   = OpConstant %f32 64.0\n"
3810                 "%f32_127_5 = OpConstant %f32 127.5\n"
3811                 "%f32_128  = OpConstant %f32 128.0\n"
3812                 "%f32_256  = OpConstant %f32 256.0\n"
3813
3814                 "%main     = OpFunction %void None %voidf\n"
3815                 "%entry    = OpLabel\n"
3816                 "%idval    = OpLoad %uvec3 %id\n"
3817                 "%x        = OpCompositeExtract %u32 %idval 0\n"
3818                 "%inloc    = OpAccessChain %f32ptr %indata %zero %x\n"
3819                 "%inval    = OpLoad %f32 %inloc\n"
3820
3821                 "%xabs     = OpExtInst %f32 %ext FAbs %inval\n"
3822                 "%x8       = OpFMod %f32 %xabs %f32_256\n"
3823                 "%x7       = OpFMod %f32 %xabs %f32_128\n"
3824                 "%x6       = OpFMod %f32 %xabs %f32_64\n"
3825                 "%x5       = OpFMod %f32 %xabs %f32_32\n"
3826                 "%x4       = OpFMod %f32 %xabs %f32_16\n"
3827                 "%x3       = OpFMod %f32 %xabs %f32_8\n"
3828                 "%x2       = OpFMod %f32 %xabs %f32_4\n"
3829                 "%x1       = OpFMod %f32 %xabs %f32_2\n"
3830
3831                 "%b7       = OpFOrdGreaterThanEqual %bool %x8 %f32_127_5\n"
3832                 "%b6       = OpFOrdGreaterThanEqual %bool %x7 %f32_63_5\n"
3833                 "%b5       = OpFOrdGreaterThanEqual %bool %x6 %f32_31_5\n"
3834                 "%b4       = OpFOrdGreaterThanEqual %bool %x5 %f32_15_5\n"
3835                 "%b3       = OpFOrdGreaterThanEqual %bool %x4 %f32_7_5\n"
3836                 "%b2       = OpFOrdGreaterThanEqual %bool %x3 %f32_3_5\n"
3837                 "%b1       = OpFOrdGreaterThanEqual %bool %x2 %f32_1_5\n"
3838                 "%b0       = OpFOrdGreaterThanEqual %bool %x1 %f32_0_5\n"
3839
3840                 + generateOpPhiCase5(codestring) +
3841
3842                 "%outloc   = OpAccessChain %f32ptr %outdata %zero %x\n"
3843                 "            OpStore %outloc %res\n"
3844                 "            OpReturn\n"
3845
3846                 "            OpFunctionEnd\n";
3847         spec5.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
3848         spec5.outputs.push_back(BufferSp(new Float32Buffer(outputFloats5)));
3849         spec5.numWorkGroups = IVec3(numElements, 1, 1);
3850
3851         group->addChild(new SpvAsmComputeShaderCase(testCtx, "nested", "Stress OpPhi with a lot of nesting", spec5));
3852
3853         createOpPhiVartypeTests(group, testCtx);
3854
3855         return group.release();
3856 }
3857
3858 // Assembly code used for testing block order is based on GLSL source code:
3859 //
3860 // #version 430
3861 //
3862 // layout(std140, set = 0, binding = 0) readonly buffer Input {
3863 //   float elements[];
3864 // } input_data;
3865 // layout(std140, set = 0, binding = 1) writeonly buffer Output {
3866 //   float elements[];
3867 // } output_data;
3868 //
3869 // void main() {
3870 //   uint x = gl_GlobalInvocationID.x;
3871 //   output_data.elements[x] = input_data.elements[x];
3872 //   if (x > uint(50)) {
3873 //     switch (x % uint(3)) {
3874 //       case 0: output_data.elements[x] += 1.5f; break;
3875 //       case 1: output_data.elements[x] += 42.f; break;
3876 //       case 2: output_data.elements[x] -= 27.f; break;
3877 //       default: break;
3878 //     }
3879 //   } else {
3880 //     output_data.elements[x] = -input_data.elements[x];
3881 //   }
3882 // }
3883 tcu::TestCaseGroup* createBlockOrderGroup (tcu::TestContext& testCtx)
3884 {
3885         de::MovePtr<tcu::TestCaseGroup> group                   (new tcu::TestCaseGroup(testCtx, "block_order", "Test block orders"));
3886         ComputeShaderSpec                               spec;
3887         de::Random                                              rnd                             (deStringHash(group->getName()));
3888         const int                                               numElements             = 100;
3889         vector<float>                                   inputFloats             (numElements, 0);
3890         vector<float>                                   outputFloats    (numElements, 0);
3891
3892         fillRandomScalars(rnd, -100.f, 100.f, &inputFloats[0], numElements);
3893
3894         // CPU might not use the same rounding mode as the GPU. Use whole numbers to avoid rounding differences.
3895         floorAll(inputFloats);
3896
3897         for (size_t ndx = 0; ndx <= 50; ++ndx)
3898                 outputFloats[ndx] = -inputFloats[ndx];
3899
3900         for (size_t ndx = 51; ndx < numElements; ++ndx)
3901         {
3902                 switch (ndx % 3)
3903                 {
3904                         case 0:         outputFloats[ndx] = inputFloats[ndx] + 1.5f; break;
3905                         case 1:         outputFloats[ndx] = inputFloats[ndx] + 42.f; break;
3906                         case 2:         outputFloats[ndx] = inputFloats[ndx] - 27.f; break;
3907                         default:        break;
3908                 }
3909         }
3910
3911         spec.assembly =
3912                 string(getComputeAsmShaderPreamble()) +
3913
3914                 "OpSource GLSL 430\n"
3915                 "OpName %main \"main\"\n"
3916                 "OpName %id \"gl_GlobalInvocationID\"\n"
3917
3918                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
3919
3920                 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) +
3921
3922                 "%u32ptr       = OpTypePointer Function %u32\n"
3923                 "%u32ptr_input = OpTypePointer Input %u32\n"
3924
3925                 + string(getComputeAsmInputOutputBuffer()) +
3926
3927                 "%id        = OpVariable %uvec3ptr Input\n"
3928                 "%zero      = OpConstant %i32 0\n"
3929                 "%const3    = OpConstant %u32 3\n"
3930                 "%const50   = OpConstant %u32 50\n"
3931                 "%constf1p5 = OpConstant %f32 1.5\n"
3932                 "%constf27  = OpConstant %f32 27.0\n"
3933                 "%constf42  = OpConstant %f32 42.0\n"
3934
3935                 "%main = OpFunction %void None %voidf\n"
3936
3937                 // entry block.
3938                 "%entry    = OpLabel\n"
3939
3940                 // Create a temporary variable to hold the value of gl_GlobalInvocationID.x.
3941                 "%xvar     = OpVariable %u32ptr Function\n"
3942                 "%xptr     = OpAccessChain %u32ptr_input %id %zero\n"
3943                 "%x        = OpLoad %u32 %xptr\n"
3944                 "            OpStore %xvar %x\n"
3945
3946                 "%cmp      = OpUGreaterThan %bool %x %const50\n"
3947                 "            OpSelectionMerge %if_merge None\n"
3948                 "            OpBranchConditional %cmp %if_true %if_false\n"
3949
3950                 // False branch for if-statement: placed in the middle of switch cases and before true branch.
3951                 "%if_false = OpLabel\n"
3952                 "%x_f      = OpLoad %u32 %xvar\n"
3953                 "%inloc_f  = OpAccessChain %f32ptr %indata %zero %x_f\n"
3954                 "%inval_f  = OpLoad %f32 %inloc_f\n"
3955                 "%negate   = OpFNegate %f32 %inval_f\n"
3956                 "%outloc_f = OpAccessChain %f32ptr %outdata %zero %x_f\n"
3957                 "            OpStore %outloc_f %negate\n"
3958                 "            OpBranch %if_merge\n"
3959
3960                 // Merge block for if-statement: placed in the middle of true and false branch.
3961                 "%if_merge = OpLabel\n"
3962                 "            OpReturn\n"
3963
3964                 // True branch for if-statement: placed in the middle of swtich cases and after the false branch.
3965                 "%if_true  = OpLabel\n"
3966                 "%xval_t   = OpLoad %u32 %xvar\n"
3967                 "%mod      = OpUMod %u32 %xval_t %const3\n"
3968                 "            OpSelectionMerge %switch_merge None\n"
3969                 "            OpSwitch %mod %default 0 %case0 1 %case1 2 %case2\n"
3970
3971                 // Merge block for switch-statement: placed before the case
3972                 // bodies.  But it must follow OpSwitch which dominates it.
3973                 "%switch_merge = OpLabel\n"
3974                 "                OpBranch %if_merge\n"
3975
3976                 // Case 1 for switch-statement: placed before case 0.
3977                 // It must follow the OpSwitch that dominates it.
3978                 "%case1    = OpLabel\n"
3979                 "%x_1      = OpLoad %u32 %xvar\n"
3980                 "%inloc_1  = OpAccessChain %f32ptr %indata %zero %x_1\n"
3981                 "%inval_1  = OpLoad %f32 %inloc_1\n"
3982                 "%addf42   = OpFAdd %f32 %inval_1 %constf42\n"
3983                 "%outloc_1 = OpAccessChain %f32ptr %outdata %zero %x_1\n"
3984                 "            OpStore %outloc_1 %addf42\n"
3985                 "            OpBranch %switch_merge\n"
3986
3987                 // Case 2 for switch-statement.
3988                 "%case2    = OpLabel\n"
3989                 "%x_2      = OpLoad %u32 %xvar\n"
3990                 "%inloc_2  = OpAccessChain %f32ptr %indata %zero %x_2\n"
3991                 "%inval_2  = OpLoad %f32 %inloc_2\n"
3992                 "%subf27   = OpFSub %f32 %inval_2 %constf27\n"
3993                 "%outloc_2 = OpAccessChain %f32ptr %outdata %zero %x_2\n"
3994                 "            OpStore %outloc_2 %subf27\n"
3995                 "            OpBranch %switch_merge\n"
3996
3997                 // Default case for switch-statement: placed in the middle of normal cases.
3998                 "%default = OpLabel\n"
3999                 "           OpBranch %switch_merge\n"
4000
4001                 // Case 0 for switch-statement: out of order.
4002                 "%case0    = OpLabel\n"
4003                 "%x_0      = OpLoad %u32 %xvar\n"
4004                 "%inloc_0  = OpAccessChain %f32ptr %indata %zero %x_0\n"
4005                 "%inval_0  = OpLoad %f32 %inloc_0\n"
4006                 "%addf1p5  = OpFAdd %f32 %inval_0 %constf1p5\n"
4007                 "%outloc_0 = OpAccessChain %f32ptr %outdata %zero %x_0\n"
4008                 "            OpStore %outloc_0 %addf1p5\n"
4009                 "            OpBranch %switch_merge\n"
4010
4011                 "            OpFunctionEnd\n";
4012         spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
4013         spec.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
4014         spec.numWorkGroups = IVec3(numElements, 1, 1);
4015
4016         group->addChild(new SpvAsmComputeShaderCase(testCtx, "all", "various out-of-order blocks", spec));
4017
4018         return group.release();
4019 }
4020
4021 tcu::TestCaseGroup* createMultipleShaderGroup (tcu::TestContext& testCtx)
4022 {
4023         de::MovePtr<tcu::TestCaseGroup> group                   (new tcu::TestCaseGroup(testCtx, "multiple_shaders", "Test multiple shaders in the same module"));
4024         ComputeShaderSpec                               spec1;
4025         ComputeShaderSpec                               spec2;
4026         de::Random                                              rnd                             (deStringHash(group->getName()));
4027         const int                                               numElements             = 100;
4028         vector<float>                                   inputFloats             (numElements, 0);
4029         vector<float>                                   outputFloats1   (numElements, 0);
4030         vector<float>                                   outputFloats2   (numElements, 0);
4031         fillRandomScalars(rnd, -500.f, 500.f, &inputFloats[0], numElements);
4032
4033         for (size_t ndx = 0; ndx < numElements; ++ndx)
4034         {
4035                 outputFloats1[ndx] = inputFloats[ndx] + inputFloats[ndx];
4036                 outputFloats2[ndx] = -inputFloats[ndx];
4037         }
4038
4039         const string assembly(
4040                 "OpCapability Shader\n"
4041                 "OpMemoryModel Logical GLSL450\n"
4042                 "OpEntryPoint GLCompute %comp_main1 \"entrypoint1\" %id\n"
4043                 "OpEntryPoint GLCompute %comp_main2 \"entrypoint2\" %id\n"
4044                 // A module cannot have two OpEntryPoint instructions with the same Execution Model and the same Name string.
4045                 "OpEntryPoint Vertex    %vert_main  \"entrypoint2\" %vert_builtins %vertexIndex %instanceIndex\n"
4046                 "OpExecutionMode %comp_main1 LocalSize 1 1 1\n"
4047                 "OpExecutionMode %comp_main2 LocalSize 1 1 1\n"
4048
4049                 "OpName %comp_main1              \"entrypoint1\"\n"
4050                 "OpName %comp_main2              \"entrypoint2\"\n"
4051                 "OpName %vert_main               \"entrypoint2\"\n"
4052                 "OpName %id                      \"gl_GlobalInvocationID\"\n"
4053                 "OpName %vert_builtin_st         \"gl_PerVertex\"\n"
4054                 "OpName %vertexIndex             \"gl_VertexIndex\"\n"
4055                 "OpName %instanceIndex           \"gl_InstanceIndex\"\n"
4056                 "OpMemberName %vert_builtin_st 0 \"gl_Position\"\n"
4057                 "OpMemberName %vert_builtin_st 1 \"gl_PointSize\"\n"
4058                 "OpMemberName %vert_builtin_st 2 \"gl_ClipDistance\"\n"
4059
4060                 "OpDecorate %id                      BuiltIn GlobalInvocationId\n"
4061                 "OpDecorate %vertexIndex             BuiltIn VertexIndex\n"
4062                 "OpDecorate %instanceIndex           BuiltIn InstanceIndex\n"
4063                 "OpDecorate %vert_builtin_st         Block\n"
4064                 "OpMemberDecorate %vert_builtin_st 0 BuiltIn Position\n"
4065                 "OpMemberDecorate %vert_builtin_st 1 BuiltIn PointSize\n"
4066                 "OpMemberDecorate %vert_builtin_st 2 BuiltIn ClipDistance\n"
4067
4068                 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) + string(getComputeAsmInputOutputBuffer()) +
4069
4070                 "%zero       = OpConstant %i32 0\n"
4071                 "%one        = OpConstant %u32 1\n"
4072                 "%c_f32_1    = OpConstant %f32 1\n"
4073
4074                 "%i32inputptr         = OpTypePointer Input %i32\n"
4075                 "%vec4                = OpTypeVector %f32 4\n"
4076                 "%vec4ptr             = OpTypePointer Output %vec4\n"
4077                 "%f32arr1             = OpTypeArray %f32 %one\n"
4078                 "%vert_builtin_st     = OpTypeStruct %vec4 %f32 %f32arr1\n"
4079                 "%vert_builtin_st_ptr = OpTypePointer Output %vert_builtin_st\n"
4080                 "%vert_builtins       = OpVariable %vert_builtin_st_ptr Output\n"
4081
4082                 "%id         = OpVariable %uvec3ptr Input\n"
4083                 "%vertexIndex = OpVariable %i32inputptr Input\n"
4084                 "%instanceIndex = OpVariable %i32inputptr Input\n"
4085                 "%c_vec4_1   = OpConstantComposite %vec4 %c_f32_1 %c_f32_1 %c_f32_1 %c_f32_1\n"
4086
4087                 // gl_Position = vec4(1.);
4088                 "%vert_main  = OpFunction %void None %voidf\n"
4089                 "%vert_entry = OpLabel\n"
4090                 "%position   = OpAccessChain %vec4ptr %vert_builtins %zero\n"
4091                 "              OpStore %position %c_vec4_1\n"
4092                 "              OpReturn\n"
4093                 "              OpFunctionEnd\n"
4094
4095                 // Double inputs.
4096                 "%comp_main1  = OpFunction %void None %voidf\n"
4097                 "%comp1_entry = OpLabel\n"
4098                 "%idval1      = OpLoad %uvec3 %id\n"
4099                 "%x1          = OpCompositeExtract %u32 %idval1 0\n"
4100                 "%inloc1      = OpAccessChain %f32ptr %indata %zero %x1\n"
4101                 "%inval1      = OpLoad %f32 %inloc1\n"
4102                 "%add         = OpFAdd %f32 %inval1 %inval1\n"
4103                 "%outloc1     = OpAccessChain %f32ptr %outdata %zero %x1\n"
4104                 "               OpStore %outloc1 %add\n"
4105                 "               OpReturn\n"
4106                 "               OpFunctionEnd\n"
4107
4108                 // Negate inputs.
4109                 "%comp_main2  = OpFunction %void None %voidf\n"
4110                 "%comp2_entry = OpLabel\n"
4111                 "%idval2      = OpLoad %uvec3 %id\n"
4112                 "%x2          = OpCompositeExtract %u32 %idval2 0\n"
4113                 "%inloc2      = OpAccessChain %f32ptr %indata %zero %x2\n"
4114                 "%inval2      = OpLoad %f32 %inloc2\n"
4115                 "%neg         = OpFNegate %f32 %inval2\n"
4116                 "%outloc2     = OpAccessChain %f32ptr %outdata %zero %x2\n"
4117                 "               OpStore %outloc2 %neg\n"
4118                 "               OpReturn\n"
4119                 "               OpFunctionEnd\n");
4120
4121         spec1.assembly = assembly;
4122         spec1.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
4123         spec1.outputs.push_back(BufferSp(new Float32Buffer(outputFloats1)));
4124         spec1.numWorkGroups = IVec3(numElements, 1, 1);
4125         spec1.entryPoint = "entrypoint1";
4126
4127         spec2.assembly = assembly;
4128         spec2.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
4129         spec2.outputs.push_back(BufferSp(new Float32Buffer(outputFloats2)));
4130         spec2.numWorkGroups = IVec3(numElements, 1, 1);
4131         spec2.entryPoint = "entrypoint2";
4132
4133         group->addChild(new SpvAsmComputeShaderCase(testCtx, "shader1", "multiple shaders in the same module", spec1));
4134         group->addChild(new SpvAsmComputeShaderCase(testCtx, "shader2", "multiple shaders in the same module", spec2));
4135
4136         return group.release();
4137 }
4138
4139 inline std::string makeLongUTF8String (size_t num4ByteChars)
4140 {
4141         // An example of a longest valid UTF-8 character.  Be explicit about the
4142         // character type because Microsoft compilers can otherwise interpret the
4143         // character string as being over wide (16-bit) characters. Ideally, we
4144         // would just use a C++11 UTF-8 string literal, but we want to support older
4145         // Microsoft compilers.
4146         const std::basic_string<char> earthAfrica("\xF0\x9F\x8C\x8D");
4147         std::string longString;
4148         longString.reserve(num4ByteChars * 4);
4149         for (size_t count = 0; count < num4ByteChars; count++)
4150         {
4151                 longString += earthAfrica;
4152         }
4153         return longString;
4154 }
4155
4156 tcu::TestCaseGroup* createOpSourceGroup (tcu::TestContext& testCtx)
4157 {
4158         de::MovePtr<tcu::TestCaseGroup> group                   (new tcu::TestCaseGroup(testCtx, "opsource", "Tests the OpSource & OpSourceContinued instruction"));
4159         vector<CaseParameter>                   cases;
4160         de::Random                                              rnd                             (deStringHash(group->getName()));
4161         const int                                               numElements             = 100;
4162         vector<float>                                   positiveFloats  (numElements, 0);
4163         vector<float>                                   negativeFloats  (numElements, 0);
4164         const StringTemplate                    shaderTemplate  (
4165                 "OpCapability Shader\n"
4166                 "OpMemoryModel Logical GLSL450\n"
4167
4168                 "OpEntryPoint GLCompute %main \"main\" %id\n"
4169                 "OpExecutionMode %main LocalSize 1 1 1\n"
4170
4171                 "${SOURCE}\n"
4172
4173                 "OpName %main           \"main\"\n"
4174                 "OpName %id             \"gl_GlobalInvocationID\"\n"
4175
4176                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
4177
4178                 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) + string(getComputeAsmInputOutputBuffer()) +
4179
4180                 "%id        = OpVariable %uvec3ptr Input\n"
4181                 "%zero      = OpConstant %i32 0\n"
4182
4183                 "%main      = OpFunction %void None %voidf\n"
4184                 "%label     = OpLabel\n"
4185                 "%idval     = OpLoad %uvec3 %id\n"
4186                 "%x         = OpCompositeExtract %u32 %idval 0\n"
4187                 "%inloc     = OpAccessChain %f32ptr %indata %zero %x\n"
4188                 "%inval     = OpLoad %f32 %inloc\n"
4189                 "%neg       = OpFNegate %f32 %inval\n"
4190                 "%outloc    = OpAccessChain %f32ptr %outdata %zero %x\n"
4191                 "             OpStore %outloc %neg\n"
4192                 "             OpReturn\n"
4193                 "             OpFunctionEnd\n");
4194
4195         cases.push_back(CaseParameter("unknown_source",                                                 "OpSource Unknown 0"));
4196         cases.push_back(CaseParameter("wrong_source",                                                   "OpSource OpenCL_C 210"));
4197         cases.push_back(CaseParameter("normal_filename",                                                "%fname = OpString \"filename\"\n"
4198                                                                                                                                                         "OpSource GLSL 430 %fname"));
4199         cases.push_back(CaseParameter("empty_filename",                                                 "%fname = OpString \"\"\n"
4200                                                                                                                                                         "OpSource GLSL 430 %fname"));
4201         cases.push_back(CaseParameter("normal_source_code",                                             "%fname = OpString \"filename\"\n"
4202                                                                                                                                                         "OpSource GLSL 430 %fname \"#version 430\nvoid main() {}\""));
4203         cases.push_back(CaseParameter("empty_source_code",                                              "%fname = OpString \"filename\"\n"
4204                                                                                                                                                         "OpSource GLSL 430 %fname \"\""));
4205         cases.push_back(CaseParameter("long_source_code",                                               "%fname = OpString \"filename\"\n"
4206                                                                                                                                                         "OpSource GLSL 430 %fname \"" + makeLongUTF8String(65530) + "ccc\"")); // word count: 65535
4207         cases.push_back(CaseParameter("utf8_source_code",                                               "%fname = OpString \"filename\"\n"
4208                                                                                                                                                         "OpSource GLSL 430 %fname \"\xE2\x98\x82\xE2\x98\x85\"")); // umbrella & black star symbol
4209         cases.push_back(CaseParameter("normal_sourcecontinued",                                 "%fname = OpString \"filename\"\n"
4210                                                                                                                                                         "OpSource GLSL 430 %fname \"#version 430\nvo\"\n"
4211                                                                                                                                                         "OpSourceContinued \"id main() {}\""));
4212         cases.push_back(CaseParameter("empty_sourcecontinued",                                  "%fname = OpString \"filename\"\n"
4213                                                                                                                                                         "OpSource GLSL 430 %fname \"#version 430\nvoid main() {}\"\n"
4214                                                                                                                                                         "OpSourceContinued \"\""));
4215         cases.push_back(CaseParameter("long_sourcecontinued",                                   "%fname = OpString \"filename\"\n"
4216                                                                                                                                                         "OpSource GLSL 430 %fname \"#version 430\nvoid main() {}\"\n"
4217                                                                                                                                                         "OpSourceContinued \"" + makeLongUTF8String(65533) + "ccc\"")); // word count: 65535
4218         cases.push_back(CaseParameter("utf8_sourcecontinued",                                   "%fname = OpString \"filename\"\n"
4219                                                                                                                                                         "OpSource GLSL 430 %fname \"#version 430\nvoid main() {}\"\n"
4220                                                                                                                                                         "OpSourceContinued \"\xE2\x98\x8E\xE2\x9A\x91\"")); // white telephone & black flag symbol
4221         cases.push_back(CaseParameter("multi_sourcecontinued",                                  "%fname = OpString \"filename\"\n"
4222                                                                                                                                                         "OpSource GLSL 430 %fname \"#version 430\n\"\n"
4223                                                                                                                                                         "OpSourceContinued \"void\"\n"
4224                                                                                                                                                         "OpSourceContinued \"main()\"\n"
4225                                                                                                                                                         "OpSourceContinued \"{}\""));
4226         cases.push_back(CaseParameter("empty_source_before_sourcecontinued",    "%fname = OpString \"filename\"\n"
4227                                                                                                                                                         "OpSource GLSL 430 %fname \"\"\n"
4228                                                                                                                                                         "OpSourceContinued \"#version 430\nvoid main() {}\""));
4229
4230         fillRandomScalars(rnd, 1.f, 100.f, &positiveFloats[0], numElements);
4231
4232         for (size_t ndx = 0; ndx < numElements; ++ndx)
4233                 negativeFloats[ndx] = -positiveFloats[ndx];
4234
4235         for (size_t caseNdx = 0; caseNdx < cases.size(); ++caseNdx)
4236         {
4237                 map<string, string>             specializations;
4238                 ComputeShaderSpec               spec;
4239
4240                 specializations["SOURCE"] = cases[caseNdx].param;
4241                 spec.assembly = shaderTemplate.specialize(specializations);
4242                 spec.inputs.push_back(BufferSp(new Float32Buffer(positiveFloats)));
4243                 spec.outputs.push_back(BufferSp(new Float32Buffer(negativeFloats)));
4244                 spec.numWorkGroups = IVec3(numElements, 1, 1);
4245
4246                 group->addChild(new SpvAsmComputeShaderCase(testCtx, cases[caseNdx].name, cases[caseNdx].name, spec));
4247         }
4248
4249         return group.release();
4250 }
4251
4252 tcu::TestCaseGroup* createOpSourceExtensionGroup (tcu::TestContext& testCtx)
4253 {
4254         de::MovePtr<tcu::TestCaseGroup> group                   (new tcu::TestCaseGroup(testCtx, "opsourceextension", "Tests the OpSource instruction"));
4255         vector<CaseParameter>                   cases;
4256         de::Random                                              rnd                             (deStringHash(group->getName()));
4257         const int                                               numElements             = 100;
4258         vector<float>                                   inputFloats             (numElements, 0);
4259         vector<float>                                   outputFloats    (numElements, 0);
4260         const StringTemplate                    shaderTemplate  (
4261                 string(getComputeAsmShaderPreamble()) +
4262
4263                 "OpSourceExtension \"${EXTENSION}\"\n"
4264
4265                 "OpName %main           \"main\"\n"
4266                 "OpName %id             \"gl_GlobalInvocationID\"\n"
4267
4268                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
4269
4270                 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) + string(getComputeAsmInputOutputBuffer()) +
4271
4272                 "%id        = OpVariable %uvec3ptr Input\n"
4273                 "%zero      = OpConstant %i32 0\n"
4274
4275                 "%main      = OpFunction %void None %voidf\n"
4276                 "%label     = OpLabel\n"
4277                 "%idval     = OpLoad %uvec3 %id\n"
4278                 "%x         = OpCompositeExtract %u32 %idval 0\n"
4279                 "%inloc     = OpAccessChain %f32ptr %indata %zero %x\n"
4280                 "%inval     = OpLoad %f32 %inloc\n"
4281                 "%neg       = OpFNegate %f32 %inval\n"
4282                 "%outloc    = OpAccessChain %f32ptr %outdata %zero %x\n"
4283                 "             OpStore %outloc %neg\n"
4284                 "             OpReturn\n"
4285                 "             OpFunctionEnd\n");
4286
4287         cases.push_back(CaseParameter("empty_extension",        ""));
4288         cases.push_back(CaseParameter("real_extension",         "GL_ARB_texture_rectangle"));
4289         cases.push_back(CaseParameter("fake_extension",         "GL_ARB_im_the_ultimate_extension"));
4290         cases.push_back(CaseParameter("utf8_extension",         "GL_ARB_\xE2\x98\x82\xE2\x98\x85"));
4291         cases.push_back(CaseParameter("long_extension",         makeLongUTF8String(65533) + "ccc")); // word count: 65535
4292
4293         fillRandomScalars(rnd, -200.f, 200.f, &inputFloats[0], numElements);
4294
4295         for (size_t ndx = 0; ndx < numElements; ++ndx)
4296                 outputFloats[ndx] = -inputFloats[ndx];
4297
4298         for (size_t caseNdx = 0; caseNdx < cases.size(); ++caseNdx)
4299         {
4300                 map<string, string>             specializations;
4301                 ComputeShaderSpec               spec;
4302
4303                 specializations["EXTENSION"] = cases[caseNdx].param;
4304                 spec.assembly = shaderTemplate.specialize(specializations);
4305                 spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
4306                 spec.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
4307                 spec.numWorkGroups = IVec3(numElements, 1, 1);
4308
4309                 group->addChild(new SpvAsmComputeShaderCase(testCtx, cases[caseNdx].name, cases[caseNdx].name, spec));
4310         }
4311
4312         return group.release();
4313 }
4314
4315 // Checks that a compute shader can generate a constant null value of various types, without exercising a computation on it.
4316 tcu::TestCaseGroup* createOpConstantNullGroup (tcu::TestContext& testCtx)
4317 {
4318         de::MovePtr<tcu::TestCaseGroup> group                   (new tcu::TestCaseGroup(testCtx, "opconstantnull", "Tests the OpConstantNull instruction"));
4319         vector<CaseParameter>                   cases;
4320         de::Random                                              rnd                             (deStringHash(group->getName()));
4321         const int                                               numElements             = 100;
4322         vector<float>                                   positiveFloats  (numElements, 0);
4323         vector<float>                                   negativeFloats  (numElements, 0);
4324         const StringTemplate                    shaderTemplate  (
4325                 string(getComputeAsmShaderPreamble()) +
4326
4327                 "OpSource GLSL 430\n"
4328                 "OpName %main           \"main\"\n"
4329                 "OpName %id             \"gl_GlobalInvocationID\"\n"
4330
4331                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
4332
4333                 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) +
4334                 "%uvec2     = OpTypeVector %u32 2\n"
4335                 "%bvec3     = OpTypeVector %bool 3\n"
4336                 "%fvec4     = OpTypeVector %f32 4\n"
4337                 "%fmat33    = OpTypeMatrix %fvec3 3\n"
4338                 "%const100  = OpConstant %u32 100\n"
4339                 "%uarr100   = OpTypeArray %i32 %const100\n"
4340                 "%struct    = OpTypeStruct %f32 %i32 %u32\n"
4341                 "%pointer   = OpTypePointer Function %i32\n"
4342                 + string(getComputeAsmInputOutputBuffer()) +
4343
4344                 "%null      = OpConstantNull ${TYPE}\n"
4345
4346                 "%id        = OpVariable %uvec3ptr Input\n"
4347                 "%zero      = OpConstant %i32 0\n"
4348
4349                 "%main      = OpFunction %void None %voidf\n"
4350                 "%label     = OpLabel\n"
4351                 "%idval     = OpLoad %uvec3 %id\n"
4352                 "%x         = OpCompositeExtract %u32 %idval 0\n"
4353                 "%inloc     = OpAccessChain %f32ptr %indata %zero %x\n"
4354                 "%inval     = OpLoad %f32 %inloc\n"
4355                 "%neg       = OpFNegate %f32 %inval\n"
4356                 "%outloc    = OpAccessChain %f32ptr %outdata %zero %x\n"
4357                 "             OpStore %outloc %neg\n"
4358                 "             OpReturn\n"
4359                 "             OpFunctionEnd\n");
4360
4361         cases.push_back(CaseParameter("bool",                   "%bool"));
4362         cases.push_back(CaseParameter("sint32",                 "%i32"));
4363         cases.push_back(CaseParameter("uint32",                 "%u32"));
4364         cases.push_back(CaseParameter("float32",                "%f32"));
4365         cases.push_back(CaseParameter("vec4float32",    "%fvec4"));
4366         cases.push_back(CaseParameter("vec3bool",               "%bvec3"));
4367         cases.push_back(CaseParameter("vec2uint32",             "%uvec2"));
4368         cases.push_back(CaseParameter("matrix",                 "%fmat33"));
4369         cases.push_back(CaseParameter("array",                  "%uarr100"));
4370         cases.push_back(CaseParameter("struct",                 "%struct"));
4371         cases.push_back(CaseParameter("pointer",                "%pointer"));
4372
4373         fillRandomScalars(rnd, 1.f, 100.f, &positiveFloats[0], numElements);
4374
4375         for (size_t ndx = 0; ndx < numElements; ++ndx)
4376                 negativeFloats[ndx] = -positiveFloats[ndx];
4377
4378         for (size_t caseNdx = 0; caseNdx < cases.size(); ++caseNdx)
4379         {
4380                 map<string, string>             specializations;
4381                 ComputeShaderSpec               spec;
4382
4383                 specializations["TYPE"] = cases[caseNdx].param;
4384                 spec.assembly = shaderTemplate.specialize(specializations);
4385                 spec.inputs.push_back(BufferSp(new Float32Buffer(positiveFloats)));
4386                 spec.outputs.push_back(BufferSp(new Float32Buffer(negativeFloats)));
4387                 spec.numWorkGroups = IVec3(numElements, 1, 1);
4388
4389                 group->addChild(new SpvAsmComputeShaderCase(testCtx, cases[caseNdx].name, cases[caseNdx].name, spec));
4390         }
4391
4392         return group.release();
4393 }
4394
4395 // Checks that a compute shader can generate a constant composite value of various types, without exercising a computation on it.
4396 tcu::TestCaseGroup* createOpConstantCompositeGroup (tcu::TestContext& testCtx)
4397 {
4398         de::MovePtr<tcu::TestCaseGroup> group                   (new tcu::TestCaseGroup(testCtx, "opconstantcomposite", "Tests the OpConstantComposite instruction"));
4399         vector<CaseParameter>                   cases;
4400         de::Random                                              rnd                             (deStringHash(group->getName()));
4401         const int                                               numElements             = 100;
4402         vector<float>                                   positiveFloats  (numElements, 0);
4403         vector<float>                                   negativeFloats  (numElements, 0);
4404         const StringTemplate                    shaderTemplate  (
4405                 string(getComputeAsmShaderPreamble()) +
4406
4407                 "OpSource GLSL 430\n"
4408                 "OpName %main           \"main\"\n"
4409                 "OpName %id             \"gl_GlobalInvocationID\"\n"
4410
4411                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
4412
4413                 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) + string(getComputeAsmInputOutputBuffer()) +
4414
4415                 "%id        = OpVariable %uvec3ptr Input\n"
4416                 "%zero      = OpConstant %i32 0\n"
4417
4418                 "${CONSTANT}\n"
4419
4420                 "%main      = OpFunction %void None %voidf\n"
4421                 "%label     = OpLabel\n"
4422                 "%idval     = OpLoad %uvec3 %id\n"
4423                 "%x         = OpCompositeExtract %u32 %idval 0\n"
4424                 "%inloc     = OpAccessChain %f32ptr %indata %zero %x\n"
4425                 "%inval     = OpLoad %f32 %inloc\n"
4426                 "%neg       = OpFNegate %f32 %inval\n"
4427                 "%outloc    = OpAccessChain %f32ptr %outdata %zero %x\n"
4428                 "             OpStore %outloc %neg\n"
4429                 "             OpReturn\n"
4430                 "             OpFunctionEnd\n");
4431
4432         cases.push_back(CaseParameter("vector",                 "%five = OpConstant %u32 5\n"
4433                                                                                                         "%const = OpConstantComposite %uvec3 %five %zero %five"));
4434         cases.push_back(CaseParameter("matrix",                 "%m3fvec3 = OpTypeMatrix %fvec3 3\n"
4435                                                                                                         "%ten = OpConstant %f32 10.\n"
4436                                                                                                         "%fzero = OpConstant %f32 0.\n"
4437                                                                                                         "%vec = OpConstantComposite %fvec3 %ten %fzero %ten\n"
4438                                                                                                         "%mat = OpConstantComposite %m3fvec3 %vec %vec %vec"));
4439         cases.push_back(CaseParameter("struct",                 "%m2vec3 = OpTypeMatrix %fvec3 2\n"
4440                                                                                                         "%struct = OpTypeStruct %i32 %f32 %fvec3 %m2vec3\n"
4441                                                                                                         "%fzero = OpConstant %f32 0.\n"
4442                                                                                                         "%one = OpConstant %f32 1.\n"
4443                                                                                                         "%point5 = OpConstant %f32 0.5\n"
4444                                                                                                         "%vec = OpConstantComposite %fvec3 %one %one %fzero\n"
4445                                                                                                         "%mat = OpConstantComposite %m2vec3 %vec %vec\n"
4446                                                                                                         "%const = OpConstantComposite %struct %zero %point5 %vec %mat"));
4447         cases.push_back(CaseParameter("nested_struct",  "%st1 = OpTypeStruct %u32 %f32\n"
4448                                                                                                         "%st2 = OpTypeStruct %i32 %i32\n"
4449                                                                                                         "%struct = OpTypeStruct %st1 %st2\n"
4450                                                                                                         "%point5 = OpConstant %f32 0.5\n"
4451                                                                                                         "%one = OpConstant %u32 1\n"
4452                                                                                                         "%ten = OpConstant %i32 10\n"
4453                                                                                                         "%st1val = OpConstantComposite %st1 %one %point5\n"
4454                                                                                                         "%st2val = OpConstantComposite %st2 %ten %ten\n"
4455                                                                                                         "%const = OpConstantComposite %struct %st1val %st2val"));
4456
4457         fillRandomScalars(rnd, 1.f, 100.f, &positiveFloats[0], numElements);
4458
4459         for (size_t ndx = 0; ndx < numElements; ++ndx)
4460                 negativeFloats[ndx] = -positiveFloats[ndx];
4461
4462         for (size_t caseNdx = 0; caseNdx < cases.size(); ++caseNdx)
4463         {
4464                 map<string, string>             specializations;
4465                 ComputeShaderSpec               spec;
4466
4467                 specializations["CONSTANT"] = cases[caseNdx].param;
4468                 spec.assembly = shaderTemplate.specialize(specializations);
4469                 spec.inputs.push_back(BufferSp(new Float32Buffer(positiveFloats)));
4470                 spec.outputs.push_back(BufferSp(new Float32Buffer(negativeFloats)));
4471                 spec.numWorkGroups = IVec3(numElements, 1, 1);
4472
4473                 group->addChild(new SpvAsmComputeShaderCase(testCtx, cases[caseNdx].name, cases[caseNdx].name, spec));
4474         }
4475
4476         return group.release();
4477 }
4478
4479 // Creates a floating point number with the given exponent, and significand
4480 // bits set. It can only create normalized numbers. Only the least significant
4481 // 24 bits of the significand will be examined. The final bit of the
4482 // significand will also be ignored. This allows alignment to be written
4483 // similarly to C99 hex-floats.
4484 // For example if you wanted to write 0x1.7f34p-12 you would call
4485 // constructNormalizedFloat(-12, 0x7f3400)
4486 float constructNormalizedFloat (deInt32 exponent, deUint32 significand)
4487 {
4488         float f = 1.0f;
4489
4490         for (deInt32 idx = 0; idx < 23; ++idx)
4491         {
4492                 f += ((significand & 0x800000) == 0) ? 0.f : std::ldexp(1.0f, -(idx + 1));
4493                 significand <<= 1;
4494         }
4495
4496         return std::ldexp(f, exponent);
4497 }
4498
4499 // Compare instruction for the OpQuantizeF16 compute exact case.
4500 // Returns true if the output is what is expected from the test case.
4501 bool compareOpQuantizeF16ComputeExactCase (const std::vector<Resource>&, const vector<AllocationSp>& outputAllocs, const std::vector<Resource>& expectedOutputs, TestLog&)
4502 {
4503         if (outputAllocs.size() != 1)
4504                 return false;
4505
4506         // Only size is needed because we cannot compare Nans.
4507         size_t byteSize = expectedOutputs[0].getByteSize();
4508
4509         const float*    outputAsFloat   = static_cast<const float*>(outputAllocs[0]->getHostPtr());
4510
4511         if (byteSize != 4*sizeof(float)) {
4512                 return false;
4513         }
4514
4515         if (*outputAsFloat != constructNormalizedFloat(8, 0x304000) &&
4516                 *outputAsFloat != constructNormalizedFloat(8, 0x300000)) {
4517                 return false;
4518         }
4519         outputAsFloat++;
4520
4521         if (*outputAsFloat != -constructNormalizedFloat(-7, 0x600000) &&
4522                 *outputAsFloat != -constructNormalizedFloat(-7, 0x604000)) {
4523                 return false;
4524         }
4525         outputAsFloat++;
4526
4527         if (*outputAsFloat != constructNormalizedFloat(2, 0x01C000) &&
4528                 *outputAsFloat != constructNormalizedFloat(2, 0x020000)) {
4529                 return false;
4530         }
4531         outputAsFloat++;
4532
4533         if (*outputAsFloat != constructNormalizedFloat(1, 0xFFC000) &&
4534                 *outputAsFloat != constructNormalizedFloat(2, 0x000000)) {
4535                 return false;
4536         }
4537
4538         return true;
4539 }
4540
4541 // Checks that every output from a test-case is a float NaN.
4542 bool compareNan (const std::vector<Resource>&, const vector<AllocationSp>& outputAllocs, const std::vector<Resource>& expectedOutputs, TestLog&)
4543 {
4544         if (outputAllocs.size() != 1)
4545                 return false;
4546
4547         // Only size is needed because we cannot compare Nans.
4548         size_t byteSize = expectedOutputs[0].getByteSize();
4549
4550         const float* const      output_as_float = static_cast<const float*>(outputAllocs[0]->getHostPtr());
4551
4552         for (size_t idx = 0; idx < byteSize / sizeof(float); ++idx)
4553         {
4554                 if (!deFloatIsNaN(output_as_float[idx]))
4555                 {
4556                         return false;
4557                 }
4558         }
4559
4560         return true;
4561 }
4562
4563 // Checks that a compute shader can generate a constant composite value of various types, without exercising a computation on it.
4564 tcu::TestCaseGroup* createOpQuantizeToF16Group (tcu::TestContext& testCtx)
4565 {
4566         de::MovePtr<tcu::TestCaseGroup> group                   (new tcu::TestCaseGroup(testCtx, "opquantize", "Tests the OpQuantizeToF16 instruction"));
4567
4568         const std::string shader (
4569                 string(getComputeAsmShaderPreamble()) +
4570
4571                 "OpSource GLSL 430\n"
4572                 "OpName %main           \"main\"\n"
4573                 "OpName %id             \"gl_GlobalInvocationID\"\n"
4574
4575                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
4576
4577                 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) + string(getComputeAsmInputOutputBuffer()) +
4578
4579                 "%id        = OpVariable %uvec3ptr Input\n"
4580                 "%zero      = OpConstant %i32 0\n"
4581
4582                 "%main      = OpFunction %void None %voidf\n"
4583                 "%label     = OpLabel\n"
4584                 "%idval     = OpLoad %uvec3 %id\n"
4585                 "%x         = OpCompositeExtract %u32 %idval 0\n"
4586                 "%inloc     = OpAccessChain %f32ptr %indata %zero %x\n"
4587                 "%inval     = OpLoad %f32 %inloc\n"
4588                 "%quant     = OpQuantizeToF16 %f32 %inval\n"
4589                 "%outloc    = OpAccessChain %f32ptr %outdata %zero %x\n"
4590                 "             OpStore %outloc %quant\n"
4591                 "             OpReturn\n"
4592                 "             OpFunctionEnd\n");
4593
4594         {
4595                 ComputeShaderSpec       spec;
4596                 const deUint32          numElements             = 100;
4597                 vector<float>           infinities;
4598                 vector<float>           results;
4599
4600                 infinities.reserve(numElements);
4601                 results.reserve(numElements);
4602
4603                 for (size_t idx = 0; idx < numElements; ++idx)
4604                 {
4605                         switch(idx % 4)
4606                         {
4607                                 case 0:
4608                                         infinities.push_back(std::numeric_limits<float>::infinity());
4609                                         results.push_back(std::numeric_limits<float>::infinity());
4610                                         break;
4611                                 case 1:
4612                                         infinities.push_back(-std::numeric_limits<float>::infinity());
4613                                         results.push_back(-std::numeric_limits<float>::infinity());
4614                                         break;
4615                                 case 2:
4616                                         infinities.push_back(std::ldexp(1.0f, 16));
4617                                         results.push_back(std::numeric_limits<float>::infinity());
4618                                         break;
4619                                 case 3:
4620                                         infinities.push_back(std::ldexp(-1.0f, 32));
4621                                         results.push_back(-std::numeric_limits<float>::infinity());
4622                                         break;
4623                         }
4624                 }
4625
4626                 spec.assembly = shader;
4627                 spec.inputs.push_back(BufferSp(new Float32Buffer(infinities)));
4628                 spec.outputs.push_back(BufferSp(new Float32Buffer(results)));
4629                 spec.numWorkGroups = IVec3(numElements, 1, 1);
4630
4631                 group->addChild(new SpvAsmComputeShaderCase(
4632                         testCtx, "infinities", "Check that infinities propagated and created", spec));
4633         }
4634
4635         {
4636                 ComputeShaderSpec       spec;
4637                 vector<float>           nans;
4638                 const deUint32          numElements             = 100;
4639
4640                 nans.reserve(numElements);
4641
4642                 for (size_t idx = 0; idx < numElements; ++idx)
4643                 {
4644                         if (idx % 2 == 0)
4645                         {
4646                                 nans.push_back(std::numeric_limits<float>::quiet_NaN());
4647                         }
4648                         else
4649                         {
4650                                 nans.push_back(-std::numeric_limits<float>::quiet_NaN());
4651                         }
4652                 }
4653
4654                 spec.assembly = shader;
4655                 spec.inputs.push_back(BufferSp(new Float32Buffer(nans)));
4656                 spec.outputs.push_back(BufferSp(new Float32Buffer(nans)));
4657                 spec.numWorkGroups = IVec3(numElements, 1, 1);
4658                 spec.verifyIO = &compareNan;
4659
4660                 group->addChild(new SpvAsmComputeShaderCase(
4661                         testCtx, "propagated_nans", "Check that nans are propagated", spec));
4662         }
4663
4664         {
4665                 ComputeShaderSpec       spec;
4666                 vector<float>           small;
4667                 vector<float>           zeros;
4668                 const deUint32          numElements             = 100;
4669
4670                 small.reserve(numElements);
4671                 zeros.reserve(numElements);
4672
4673                 for (size_t idx = 0; idx < numElements; ++idx)
4674                 {
4675                         switch(idx % 6)
4676                         {
4677                                 case 0:
4678                                         small.push_back(0.f);
4679                                         zeros.push_back(0.f);
4680                                         break;
4681                                 case 1:
4682                                         small.push_back(-0.f);
4683                                         zeros.push_back(-0.f);
4684                                         break;
4685                                 case 2:
4686                                         small.push_back(std::ldexp(1.0f, -16));
4687                                         zeros.push_back(0.f);
4688                                         break;
4689                                 case 3:
4690                                         small.push_back(std::ldexp(-1.0f, -32));
4691                                         zeros.push_back(-0.f);
4692                                         break;
4693                                 case 4:
4694                                         small.push_back(std::ldexp(1.0f, -127));
4695                                         zeros.push_back(0.f);
4696                                         break;
4697                                 case 5:
4698                                         small.push_back(-std::ldexp(1.0f, -128));
4699                                         zeros.push_back(-0.f);
4700                                         break;
4701                         }
4702                 }
4703
4704                 spec.assembly = shader;
4705                 spec.inputs.push_back(BufferSp(new Float32Buffer(small)));
4706                 spec.outputs.push_back(BufferSp(new Float32Buffer(zeros)));
4707                 spec.numWorkGroups = IVec3(numElements, 1, 1);
4708
4709                 group->addChild(new SpvAsmComputeShaderCase(
4710                         testCtx, "flush_to_zero", "Check that values are zeroed correctly", spec));
4711         }
4712
4713         {
4714                 ComputeShaderSpec       spec;
4715                 vector<float>           exact;
4716                 const deUint32          numElements             = 200;
4717
4718                 exact.reserve(numElements);
4719
4720                 for (size_t idx = 0; idx < numElements; ++idx)
4721                         exact.push_back(static_cast<float>(static_cast<int>(idx) - 100));
4722
4723                 spec.assembly = shader;
4724                 spec.inputs.push_back(BufferSp(new Float32Buffer(exact)));
4725                 spec.outputs.push_back(BufferSp(new Float32Buffer(exact)));
4726                 spec.numWorkGroups = IVec3(numElements, 1, 1);
4727
4728                 group->addChild(new SpvAsmComputeShaderCase(
4729                         testCtx, "exact", "Check that values exactly preserved where appropriate", spec));
4730         }
4731
4732         {
4733                 ComputeShaderSpec       spec;
4734                 vector<float>           inputs;
4735                 const deUint32          numElements             = 4;
4736
4737                 inputs.push_back(constructNormalizedFloat(8,    0x300300));
4738                 inputs.push_back(-constructNormalizedFloat(-7,  0x600800));
4739                 inputs.push_back(constructNormalizedFloat(2,    0x01E000));
4740                 inputs.push_back(constructNormalizedFloat(1,    0xFFE000));
4741
4742                 spec.assembly = shader;
4743                 spec.verifyIO = &compareOpQuantizeF16ComputeExactCase;
4744                 spec.inputs.push_back(BufferSp(new Float32Buffer(inputs)));
4745                 spec.outputs.push_back(BufferSp(new Float32Buffer(inputs)));
4746                 spec.numWorkGroups = IVec3(numElements, 1, 1);
4747
4748                 group->addChild(new SpvAsmComputeShaderCase(
4749                         testCtx, "rounded", "Check that are rounded when needed", spec));
4750         }
4751
4752         return group.release();
4753 }
4754
4755 tcu::TestCaseGroup* createSpecConstantOpQuantizeToF16Group (tcu::TestContext& testCtx)
4756 {
4757         de::MovePtr<tcu::TestCaseGroup> group                   (new tcu::TestCaseGroup(testCtx, "opspecconstantop_opquantize", "Tests the OpQuantizeToF16 opcode for the OpSpecConstantOp instruction"));
4758
4759         const std::string shader (
4760                 string(getComputeAsmShaderPreamble()) +
4761
4762                 "OpName %main           \"main\"\n"
4763                 "OpName %id             \"gl_GlobalInvocationID\"\n"
4764
4765                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
4766
4767                 "OpDecorate %sc_0  SpecId 0\n"
4768                 "OpDecorate %sc_1  SpecId 1\n"
4769                 "OpDecorate %sc_2  SpecId 2\n"
4770                 "OpDecorate %sc_3  SpecId 3\n"
4771                 "OpDecorate %sc_4  SpecId 4\n"
4772                 "OpDecorate %sc_5  SpecId 5\n"
4773
4774                 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) + string(getComputeAsmInputOutputBuffer()) +
4775
4776                 "%id        = OpVariable %uvec3ptr Input\n"
4777                 "%zero      = OpConstant %i32 0\n"
4778                 "%c_u32_6   = OpConstant %u32 6\n"
4779
4780                 "%sc_0      = OpSpecConstant %f32 0.\n"
4781                 "%sc_1      = OpSpecConstant %f32 0.\n"
4782                 "%sc_2      = OpSpecConstant %f32 0.\n"
4783                 "%sc_3      = OpSpecConstant %f32 0.\n"
4784                 "%sc_4      = OpSpecConstant %f32 0.\n"
4785                 "%sc_5      = OpSpecConstant %f32 0.\n"
4786
4787                 "%sc_0_quant = OpSpecConstantOp %f32 QuantizeToF16 %sc_0\n"
4788                 "%sc_1_quant = OpSpecConstantOp %f32 QuantizeToF16 %sc_1\n"
4789                 "%sc_2_quant = OpSpecConstantOp %f32 QuantizeToF16 %sc_2\n"
4790                 "%sc_3_quant = OpSpecConstantOp %f32 QuantizeToF16 %sc_3\n"
4791                 "%sc_4_quant = OpSpecConstantOp %f32 QuantizeToF16 %sc_4\n"
4792                 "%sc_5_quant = OpSpecConstantOp %f32 QuantizeToF16 %sc_5\n"
4793
4794                 "%main      = OpFunction %void None %voidf\n"
4795                 "%label     = OpLabel\n"
4796                 "%idval     = OpLoad %uvec3 %id\n"
4797                 "%x         = OpCompositeExtract %u32 %idval 0\n"
4798                 "%outloc    = OpAccessChain %f32ptr %outdata %zero %x\n"
4799                 "%selector  = OpUMod %u32 %x %c_u32_6\n"
4800                 "            OpSelectionMerge %exit None\n"
4801                 "            OpSwitch %selector %exit 0 %case0 1 %case1 2 %case2 3 %case3 4 %case4 5 %case5\n"
4802
4803                 "%case0     = OpLabel\n"
4804                 "             OpStore %outloc %sc_0_quant\n"
4805                 "             OpBranch %exit\n"
4806
4807                 "%case1     = OpLabel\n"
4808                 "             OpStore %outloc %sc_1_quant\n"
4809                 "             OpBranch %exit\n"
4810
4811                 "%case2     = OpLabel\n"
4812                 "             OpStore %outloc %sc_2_quant\n"
4813                 "             OpBranch %exit\n"
4814
4815                 "%case3     = OpLabel\n"
4816                 "             OpStore %outloc %sc_3_quant\n"
4817                 "             OpBranch %exit\n"
4818
4819                 "%case4     = OpLabel\n"
4820                 "             OpStore %outloc %sc_4_quant\n"
4821                 "             OpBranch %exit\n"
4822
4823                 "%case5     = OpLabel\n"
4824                 "             OpStore %outloc %sc_5_quant\n"
4825                 "             OpBranch %exit\n"
4826
4827                 "%exit      = OpLabel\n"
4828                 "             OpReturn\n"
4829
4830                 "             OpFunctionEnd\n");
4831
4832         {
4833                 ComputeShaderSpec       spec;
4834                 const deUint8           numCases        = 4;
4835                 vector<float>           inputs          (numCases, 0.f);
4836                 vector<float>           outputs;
4837
4838                 spec.assembly           = shader;
4839                 spec.numWorkGroups      = IVec3(numCases, 1, 1);
4840
4841                 spec.specConstants.append<deInt32>(bitwiseCast<deUint32>(std::numeric_limits<float>::infinity()));
4842                 spec.specConstants.append<deInt32>(bitwiseCast<deUint32>(-std::numeric_limits<float>::infinity()));
4843                 spec.specConstants.append<deInt32>(bitwiseCast<deUint32>(std::ldexp(1.0f, 16)));
4844                 spec.specConstants.append<deInt32>(bitwiseCast<deUint32>(std::ldexp(-1.0f, 32)));
4845
4846                 outputs.push_back(std::numeric_limits<float>::infinity());
4847                 outputs.push_back(-std::numeric_limits<float>::infinity());
4848                 outputs.push_back(std::numeric_limits<float>::infinity());
4849                 outputs.push_back(-std::numeric_limits<float>::infinity());
4850
4851                 spec.inputs.push_back(BufferSp(new Float32Buffer(inputs)));
4852                 spec.outputs.push_back(BufferSp(new Float32Buffer(outputs)));
4853
4854                 group->addChild(new SpvAsmComputeShaderCase(
4855                         testCtx, "infinities", "Check that infinities propagated and created", spec));
4856         }
4857
4858         {
4859                 ComputeShaderSpec       spec;
4860                 const deUint8           numCases        = 2;
4861                 vector<float>           inputs          (numCases, 0.f);
4862                 vector<float>           outputs;
4863
4864                 spec.assembly           = shader;
4865                 spec.numWorkGroups      = IVec3(numCases, 1, 1);
4866                 spec.verifyIO           = &compareNan;
4867
4868                 outputs.push_back(std::numeric_limits<float>::quiet_NaN());
4869                 outputs.push_back(-std::numeric_limits<float>::quiet_NaN());
4870
4871                 for (deUint8 idx = 0; idx < numCases; ++idx)
4872                         spec.specConstants.append<deInt32>(bitwiseCast<deUint32>(outputs[idx]));
4873
4874                 spec.inputs.push_back(BufferSp(new Float32Buffer(inputs)));
4875                 spec.outputs.push_back(BufferSp(new Float32Buffer(outputs)));
4876
4877                 group->addChild(new SpvAsmComputeShaderCase(
4878                         testCtx, "propagated_nans", "Check that nans are propagated", spec));
4879         }
4880
4881         {
4882                 ComputeShaderSpec       spec;
4883                 const deUint8           numCases        = 6;
4884                 vector<float>           inputs          (numCases, 0.f);
4885                 vector<float>           outputs;
4886
4887                 spec.assembly           = shader;
4888                 spec.numWorkGroups      = IVec3(numCases, 1, 1);
4889
4890                 spec.specConstants.append<deInt32>(bitwiseCast<deUint32>(0.f));
4891                 spec.specConstants.append<deInt32>(bitwiseCast<deUint32>(-0.f));
4892                 spec.specConstants.append<deInt32>(bitwiseCast<deUint32>(std::ldexp(1.0f, -16)));
4893                 spec.specConstants.append<deInt32>(bitwiseCast<deUint32>(std::ldexp(-1.0f, -32)));
4894                 spec.specConstants.append<deInt32>(bitwiseCast<deUint32>(std::ldexp(1.0f, -127)));
4895                 spec.specConstants.append<deInt32>(bitwiseCast<deUint32>(-std::ldexp(1.0f, -128)));
4896
4897                 outputs.push_back(0.f);
4898                 outputs.push_back(-0.f);
4899                 outputs.push_back(0.f);
4900                 outputs.push_back(-0.f);
4901                 outputs.push_back(0.f);
4902                 outputs.push_back(-0.f);
4903
4904                 spec.inputs.push_back(BufferSp(new Float32Buffer(inputs)));
4905                 spec.outputs.push_back(BufferSp(new Float32Buffer(outputs)));
4906
4907                 group->addChild(new SpvAsmComputeShaderCase(
4908                         testCtx, "flush_to_zero", "Check that values are zeroed correctly", spec));
4909         }
4910
4911         {
4912                 ComputeShaderSpec       spec;
4913                 const deUint8           numCases        = 6;
4914                 vector<float>           inputs          (numCases, 0.f);
4915                 vector<float>           outputs;
4916
4917                 spec.assembly           = shader;
4918                 spec.numWorkGroups      = IVec3(numCases, 1, 1);
4919
4920                 for (deUint8 idx = 0; idx < 6; ++idx)
4921                 {
4922                         const float f = static_cast<float>(idx * 10 - 30) / 4.f;
4923                         spec.specConstants.append<deInt32>(bitwiseCast<deUint32>(f));
4924                         outputs.push_back(f);
4925                 }
4926
4927                 spec.inputs.push_back(BufferSp(new Float32Buffer(inputs)));
4928                 spec.outputs.push_back(BufferSp(new Float32Buffer(outputs)));
4929
4930                 group->addChild(new SpvAsmComputeShaderCase(
4931                         testCtx, "exact", "Check that values exactly preserved where appropriate", spec));
4932         }
4933
4934         {
4935                 ComputeShaderSpec       spec;
4936                 const deUint8           numCases        = 4;
4937                 vector<float>           inputs          (numCases, 0.f);
4938                 vector<float>           outputs;
4939
4940                 spec.assembly           = shader;
4941                 spec.numWorkGroups      = IVec3(numCases, 1, 1);
4942                 spec.verifyIO           = &compareOpQuantizeF16ComputeExactCase;
4943
4944                 outputs.push_back(constructNormalizedFloat(8, 0x300300));
4945                 outputs.push_back(-constructNormalizedFloat(-7, 0x600800));
4946                 outputs.push_back(constructNormalizedFloat(2, 0x01E000));
4947                 outputs.push_back(constructNormalizedFloat(1, 0xFFE000));
4948
4949                 for (deUint8 idx = 0; idx < numCases; ++idx)
4950                         spec.specConstants.append<deInt32>(bitwiseCast<deUint32>(outputs[idx]));
4951
4952                 spec.inputs.push_back(BufferSp(new Float32Buffer(inputs)));
4953                 spec.outputs.push_back(BufferSp(new Float32Buffer(outputs)));
4954
4955                 group->addChild(new SpvAsmComputeShaderCase(
4956                         testCtx, "rounded", "Check that are rounded when needed", spec));
4957         }
4958
4959         return group.release();
4960 }
4961
4962 // Checks that constant null/composite values can be used in computation.
4963 tcu::TestCaseGroup* createOpConstantUsageGroup (tcu::TestContext& testCtx)
4964 {
4965         de::MovePtr<tcu::TestCaseGroup> group                   (new tcu::TestCaseGroup(testCtx, "opconstantnullcomposite", "Spotcheck the OpConstantNull & OpConstantComposite instruction"));
4966         ComputeShaderSpec                               spec;
4967         de::Random                                              rnd                             (deStringHash(group->getName()));
4968         const int                                               numElements             = 100;
4969         vector<float>                                   positiveFloats  (numElements, 0);
4970         vector<float>                                   negativeFloats  (numElements, 0);
4971
4972         fillRandomScalars(rnd, 1.f, 100.f, &positiveFloats[0], numElements);
4973
4974         for (size_t ndx = 0; ndx < numElements; ++ndx)
4975                 negativeFloats[ndx] = -positiveFloats[ndx];
4976
4977         spec.assembly =
4978                 "OpCapability Shader\n"
4979                 "%std450 = OpExtInstImport \"GLSL.std.450\"\n"
4980                 "OpMemoryModel Logical GLSL450\n"
4981                 "OpEntryPoint GLCompute %main \"main\" %id\n"
4982                 "OpExecutionMode %main LocalSize 1 1 1\n"
4983
4984                 "OpSource GLSL 430\n"
4985                 "OpName %main           \"main\"\n"
4986                 "OpName %id             \"gl_GlobalInvocationID\"\n"
4987
4988                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
4989
4990                 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) +
4991
4992                 "%fmat      = OpTypeMatrix %fvec3 3\n"
4993                 "%ten       = OpConstant %u32 10\n"
4994                 "%f32arr10  = OpTypeArray %f32 %ten\n"
4995                 "%fst       = OpTypeStruct %f32 %f32\n"
4996
4997                 + string(getComputeAsmInputOutputBuffer()) +
4998
4999                 "%id        = OpVariable %uvec3ptr Input\n"
5000                 "%zero      = OpConstant %i32 0\n"
5001
5002                 // Create a bunch of null values
5003                 "%unull     = OpConstantNull %u32\n"
5004                 "%fnull     = OpConstantNull %f32\n"
5005                 "%vnull     = OpConstantNull %fvec3\n"
5006                 "%mnull     = OpConstantNull %fmat\n"
5007                 "%anull     = OpConstantNull %f32arr10\n"
5008                 "%snull     = OpConstantComposite %fst %fnull %fnull\n"
5009
5010                 "%main      = OpFunction %void None %voidf\n"
5011                 "%label     = OpLabel\n"
5012                 "%idval     = OpLoad %uvec3 %id\n"
5013                 "%x         = OpCompositeExtract %u32 %idval 0\n"
5014                 "%inloc     = OpAccessChain %f32ptr %indata %zero %x\n"
5015                 "%inval     = OpLoad %f32 %inloc\n"
5016                 "%neg       = OpFNegate %f32 %inval\n"
5017
5018                 // Get the abs() of (a certain element of) those null values
5019                 "%unull_cov = OpConvertUToF %f32 %unull\n"
5020                 "%unull_abs = OpExtInst %f32 %std450 FAbs %unull_cov\n"
5021                 "%fnull_abs = OpExtInst %f32 %std450 FAbs %fnull\n"
5022                 "%vnull_0   = OpCompositeExtract %f32 %vnull 0\n"
5023                 "%vnull_abs = OpExtInst %f32 %std450 FAbs %vnull_0\n"
5024                 "%mnull_12  = OpCompositeExtract %f32 %mnull 1 2\n"
5025                 "%mnull_abs = OpExtInst %f32 %std450 FAbs %mnull_12\n"
5026                 "%anull_3   = OpCompositeExtract %f32 %anull 3\n"
5027                 "%anull_abs = OpExtInst %f32 %std450 FAbs %anull_3\n"
5028                 "%snull_1   = OpCompositeExtract %f32 %snull 1\n"
5029                 "%snull_abs = OpExtInst %f32 %std450 FAbs %snull_1\n"
5030
5031                 // Add them all
5032                 "%add1      = OpFAdd %f32 %neg  %unull_abs\n"
5033                 "%add2      = OpFAdd %f32 %add1 %fnull_abs\n"
5034                 "%add3      = OpFAdd %f32 %add2 %vnull_abs\n"
5035                 "%add4      = OpFAdd %f32 %add3 %mnull_abs\n"
5036                 "%add5      = OpFAdd %f32 %add4 %anull_abs\n"
5037                 "%final     = OpFAdd %f32 %add5 %snull_abs\n"
5038
5039                 "%outloc    = OpAccessChain %f32ptr %outdata %zero %x\n"
5040                 "             OpStore %outloc %final\n" // write to output
5041                 "             OpReturn\n"
5042                 "             OpFunctionEnd\n";
5043         spec.inputs.push_back(BufferSp(new Float32Buffer(positiveFloats)));
5044         spec.outputs.push_back(BufferSp(new Float32Buffer(negativeFloats)));
5045         spec.numWorkGroups = IVec3(numElements, 1, 1);
5046
5047         group->addChild(new SpvAsmComputeShaderCase(testCtx, "spotcheck", "Check that values constructed via OpConstantNull & OpConstantComposite can be used", spec));
5048
5049         return group.release();
5050 }
5051
5052 // Assembly code used for testing loop control is based on GLSL source code:
5053 // #version 430
5054 //
5055 // layout(std140, set = 0, binding = 0) readonly buffer Input {
5056 //   float elements[];
5057 // } input_data;
5058 // layout(std140, set = 0, binding = 1) writeonly buffer Output {
5059 //   float elements[];
5060 // } output_data;
5061 //
5062 // void main() {
5063 //   uint x = gl_GlobalInvocationID.x;
5064 //   output_data.elements[x] = input_data.elements[x];
5065 //   for (uint i = 0; i < 4; ++i)
5066 //     output_data.elements[x] += 1.f;
5067 // }
5068 tcu::TestCaseGroup* createLoopControlGroup (tcu::TestContext& testCtx)
5069 {
5070         de::MovePtr<tcu::TestCaseGroup> group                   (new tcu::TestCaseGroup(testCtx, "loop_control", "Tests loop control cases"));
5071         vector<CaseParameter>                   cases;
5072         de::Random                                              rnd                             (deStringHash(group->getName()));
5073         const int                                               numElements             = 100;
5074         vector<float>                                   inputFloats             (numElements, 0);
5075         vector<float>                                   outputFloats    (numElements, 0);
5076         const StringTemplate                    shaderTemplate  (
5077                 string(getComputeAsmShaderPreamble()) +
5078
5079                 "OpSource GLSL 430\n"
5080                 "OpName %main \"main\"\n"
5081                 "OpName %id \"gl_GlobalInvocationID\"\n"
5082
5083                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
5084
5085                 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) + string(getComputeAsmInputOutputBuffer()) +
5086
5087                 "%u32ptr      = OpTypePointer Function %u32\n"
5088
5089                 "%id          = OpVariable %uvec3ptr Input\n"
5090                 "%zero        = OpConstant %i32 0\n"
5091                 "%uzero       = OpConstant %u32 0\n"
5092                 "%one         = OpConstant %i32 1\n"
5093                 "%constf1     = OpConstant %f32 1.0\n"
5094                 "%four        = OpConstant %u32 4\n"
5095
5096                 "%main        = OpFunction %void None %voidf\n"
5097                 "%entry       = OpLabel\n"
5098                 "%i           = OpVariable %u32ptr Function\n"
5099                 "               OpStore %i %uzero\n"
5100
5101                 "%idval       = OpLoad %uvec3 %id\n"
5102                 "%x           = OpCompositeExtract %u32 %idval 0\n"
5103                 "%inloc       = OpAccessChain %f32ptr %indata %zero %x\n"
5104                 "%inval       = OpLoad %f32 %inloc\n"
5105                 "%outloc      = OpAccessChain %f32ptr %outdata %zero %x\n"
5106                 "               OpStore %outloc %inval\n"
5107                 "               OpBranch %loop_entry\n"
5108
5109                 "%loop_entry  = OpLabel\n"
5110                 "%i_val       = OpLoad %u32 %i\n"
5111                 "%cmp_lt      = OpULessThan %bool %i_val %four\n"
5112                 "               OpLoopMerge %loop_merge %loop_body ${CONTROL}\n"
5113                 "               OpBranchConditional %cmp_lt %loop_body %loop_merge\n"
5114                 "%loop_body   = OpLabel\n"
5115                 "%outval      = OpLoad %f32 %outloc\n"
5116                 "%addf1       = OpFAdd %f32 %outval %constf1\n"
5117                 "               OpStore %outloc %addf1\n"
5118                 "%new_i       = OpIAdd %u32 %i_val %one\n"
5119                 "               OpStore %i %new_i\n"
5120                 "               OpBranch %loop_entry\n"
5121                 "%loop_merge  = OpLabel\n"
5122                 "               OpReturn\n"
5123                 "               OpFunctionEnd\n");
5124
5125         cases.push_back(CaseParameter("none",                           "None"));
5126         cases.push_back(CaseParameter("unroll",                         "Unroll"));
5127         cases.push_back(CaseParameter("dont_unroll",            "DontUnroll"));
5128
5129         fillRandomScalars(rnd, -100.f, 100.f, &inputFloats[0], numElements);
5130
5131         for (size_t ndx = 0; ndx < numElements; ++ndx)
5132                 outputFloats[ndx] = inputFloats[ndx] + 4.f;
5133
5134         for (size_t caseNdx = 0; caseNdx < cases.size(); ++caseNdx)
5135         {
5136                 map<string, string>             specializations;
5137                 ComputeShaderSpec               spec;
5138
5139                 specializations["CONTROL"] = cases[caseNdx].param;
5140                 spec.assembly = shaderTemplate.specialize(specializations);
5141                 spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
5142                 spec.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
5143                 spec.numWorkGroups = IVec3(numElements, 1, 1);
5144
5145                 group->addChild(new SpvAsmComputeShaderCase(testCtx, cases[caseNdx].name, cases[caseNdx].name, spec));
5146         }
5147
5148         group->addChild(new SpvAsmLoopControlDependencyLengthCase(testCtx, "dependency_length", "dependency_length"));
5149         group->addChild(new SpvAsmLoopControlDependencyInfiniteCase(testCtx, "dependency_infinite", "dependency_infinite"));
5150
5151         return group.release();
5152 }
5153
5154 // Assembly code used for testing selection control is based on GLSL source code:
5155 // #version 430
5156 //
5157 // layout(std140, set = 0, binding = 0) readonly buffer Input {
5158 //   float elements[];
5159 // } input_data;
5160 // layout(std140, set = 0, binding = 1) writeonly buffer Output {
5161 //   float elements[];
5162 // } output_data;
5163 //
5164 // void main() {
5165 //   uint x = gl_GlobalInvocationID.x;
5166 //   float val = input_data.elements[x];
5167 //   if (val > 10.f)
5168 //     output_data.elements[x] = val + 1.f;
5169 //   else
5170 //     output_data.elements[x] = val - 1.f;
5171 // }
5172 tcu::TestCaseGroup* createSelectionControlGroup (tcu::TestContext& testCtx)
5173 {
5174         de::MovePtr<tcu::TestCaseGroup> group                   (new tcu::TestCaseGroup(testCtx, "selection_control", "Tests selection control cases"));
5175         vector<CaseParameter>                   cases;
5176         de::Random                                              rnd                             (deStringHash(group->getName()));
5177         const int                                               numElements             = 100;
5178         vector<float>                                   inputFloats             (numElements, 0);
5179         vector<float>                                   outputFloats    (numElements, 0);
5180         const StringTemplate                    shaderTemplate  (
5181                 string(getComputeAsmShaderPreamble()) +
5182
5183                 "OpSource GLSL 430\n"
5184                 "OpName %main \"main\"\n"
5185                 "OpName %id \"gl_GlobalInvocationID\"\n"
5186
5187                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
5188
5189                 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) + string(getComputeAsmInputOutputBuffer()) +
5190
5191                 "%id       = OpVariable %uvec3ptr Input\n"
5192                 "%zero     = OpConstant %i32 0\n"
5193                 "%constf1  = OpConstant %f32 1.0\n"
5194                 "%constf10 = OpConstant %f32 10.0\n"
5195
5196                 "%main     = OpFunction %void None %voidf\n"
5197                 "%entry    = OpLabel\n"
5198                 "%idval    = OpLoad %uvec3 %id\n"
5199                 "%x        = OpCompositeExtract %u32 %idval 0\n"
5200                 "%inloc    = OpAccessChain %f32ptr %indata %zero %x\n"
5201                 "%inval    = OpLoad %f32 %inloc\n"
5202                 "%outloc   = OpAccessChain %f32ptr %outdata %zero %x\n"
5203                 "%cmp_gt   = OpFOrdGreaterThan %bool %inval %constf10\n"
5204
5205                 "            OpSelectionMerge %if_end ${CONTROL}\n"
5206                 "            OpBranchConditional %cmp_gt %if_true %if_false\n"
5207                 "%if_true  = OpLabel\n"
5208                 "%addf1    = OpFAdd %f32 %inval %constf1\n"
5209                 "            OpStore %outloc %addf1\n"
5210                 "            OpBranch %if_end\n"
5211                 "%if_false = OpLabel\n"
5212                 "%subf1    = OpFSub %f32 %inval %constf1\n"
5213                 "            OpStore %outloc %subf1\n"
5214                 "            OpBranch %if_end\n"
5215                 "%if_end   = OpLabel\n"
5216                 "            OpReturn\n"
5217                 "            OpFunctionEnd\n");
5218
5219         cases.push_back(CaseParameter("none",                                   "None"));
5220         cases.push_back(CaseParameter("flatten",                                "Flatten"));
5221         cases.push_back(CaseParameter("dont_flatten",                   "DontFlatten"));
5222         cases.push_back(CaseParameter("flatten_dont_flatten",   "DontFlatten|Flatten"));
5223
5224         fillRandomScalars(rnd, -100.f, 100.f, &inputFloats[0], numElements);
5225
5226         // CPU might not use the same rounding mode as the GPU. Use whole numbers to avoid rounding differences.
5227         floorAll(inputFloats);
5228
5229         for (size_t ndx = 0; ndx < numElements; ++ndx)
5230                 outputFloats[ndx] = inputFloats[ndx] + (inputFloats[ndx] > 10.f ? 1.f : -1.f);
5231
5232         for (size_t caseNdx = 0; caseNdx < cases.size(); ++caseNdx)
5233         {
5234                 map<string, string>             specializations;
5235                 ComputeShaderSpec               spec;
5236
5237                 specializations["CONTROL"] = cases[caseNdx].param;
5238                 spec.assembly = shaderTemplate.specialize(specializations);
5239                 spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
5240                 spec.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
5241                 spec.numWorkGroups = IVec3(numElements, 1, 1);
5242
5243                 group->addChild(new SpvAsmComputeShaderCase(testCtx, cases[caseNdx].name, cases[caseNdx].name, spec));
5244         }
5245
5246         return group.release();
5247 }
5248
5249 void getOpNameAbuseCases (vector<CaseParameter> &abuseCases)
5250 {
5251         // Generate a long name.
5252         std::string longname;
5253         longname.resize(65535, 'k'); // max string literal, spir-v 2.17
5254
5255         // Some bad names, abusing utf-8 encoding. This may also cause problems
5256         // with the logs.
5257         // 1. Various illegal code points in utf-8
5258         std::string utf8illegal =
5259                 "Illegal bytes in UTF-8: "
5260                 "\xc0 \xc1 \xf5 \xf6 \xf7 \xf8 \xf9 \xfa \xfb \xfc \xfd \xfe \xff"
5261                 "illegal surrogates: \xed\xad\xbf \xed\xbe\x80";
5262
5263         // 2. Zero encoded as overlong, not exactly legal but often supported to differentiate from terminating zero
5264         std::string utf8nul = "UTF-8 encoded nul \xC0\x80 (should not end name)";
5265
5266         // 3. Some overlong encodings
5267         std::string utf8overlong =
5268                 "UTF-8 overlong \xF0\x82\x82\xAC \xfc\x83\xbf\xbf\xbf\xbf \xf8\x87\xbf\xbf\xbf "
5269                 "\xf0\x8f\xbf\xbf";
5270
5271         // 4. Internet "zalgo" meme "bleeding text"
5272         std::string utf8zalgo =
5273                 "\x56\xcc\xb5\xcc\x85\xcc\x94\xcc\x88\xcd\x8a\xcc\x91\xcc\x88\xcd\x91\xcc\x83\xcd\x82"
5274                 "\xcc\x83\xcd\x90\xcc\x8a\xcc\x92\xcc\x92\xcd\x8b\xcc\x94\xcd\x9d\xcc\x98\xcc\xab\xcc"
5275                 "\xae\xcc\xa9\xcc\xad\xcc\x97\xcc\xb0\x75\xcc\xb6\xcc\xbe\xcc\x80\xcc\x82\xcc\x84\xcd"
5276                 "\x84\xcc\x90\xcd\x86\xcc\x9a\xcd\x84\xcc\x9b\xcd\x86\xcd\x92\xcc\x9a\xcd\x99\xcd\x99"
5277                 "\xcc\xbb\xcc\x98\xcd\x8e\xcd\x88\xcd\x9a\xcc\xa6\xcc\x9c\xcc\xab\xcc\x99\xcd\x94\xcd"
5278                 "\x99\xcd\x95\xcc\xa5\xcc\xab\xcd\x89\x6c\xcc\xb8\xcc\x8e\xcc\x8b\xcc\x8b\xcc\x9a\xcc"
5279                 "\x8e\xcd\x9d\xcc\x80\xcc\xa1\xcc\xad\xcd\x9c\xcc\xba\xcc\x96\xcc\xb3\xcc\xa2\xcd\x8e"
5280                 "\xcc\xa2\xcd\x96\x6b\xcc\xb8\xcc\x84\xcd\x81\xcc\xbf\xcc\x8d\xcc\x89\xcc\x85\xcc\x92"
5281                 "\xcc\x84\xcc\x90\xcd\x81\xcc\x93\xcd\x90\xcd\x92\xcd\x9d\xcc\x84\xcd\x98\xcd\x9d\xcd"
5282                 "\xa0\xcd\x91\xcc\x94\xcc\xb9\xcd\x93\xcc\xa5\xcd\x87\xcc\xad\xcc\xa7\xcd\x96\xcd\x99"
5283                 "\xcc\x9d\xcc\xbc\xcd\x96\xcd\x93\xcc\x9d\xcc\x99\xcc\xa8\xcc\xb1\xcd\x85\xcc\xba\xcc"
5284                 "\xa7\x61\xcc\xb8\xcc\x8e\xcc\x81\xcd\x90\xcd\x84\xcd\x8c\xcc\x8c\xcc\x85\xcd\x86\xcc"
5285                 "\x84\xcd\x84\xcc\x90\xcc\x84\xcc\x8d\xcd\x99\xcd\x8d\xcc\xb0\xcc\xa3\xcc\xa6\xcd\x89"
5286                 "\xcd\x8d\xcd\x87\xcc\x98\xcd\x8d\xcc\xa4\xcd\x9a\xcd\x8e\xcc\xab\xcc\xb9\xcc\xac\xcc"
5287                 "\xa2\xcd\x87\xcc\xa0\xcc\xb3\xcd\x89\xcc\xb9\xcc\xa7\xcc\xa6\xcd\x89\xcd\x95\x6e\xcc"
5288                 "\xb8\xcd\x8a\xcc\x8a\xcd\x82\xcc\x9b\xcd\x81\xcd\x90\xcc\x85\xcc\x9b\xcd\x80\xcd\x91"
5289                 "\xcd\x9b\xcc\x81\xcd\x81\xcc\x9a\xcc\xb3\xcd\x9c\xcc\x9e\xcc\x9d\xcd\x99\xcc\xa2\xcd"
5290                 "\x93\xcd\x96\xcc\x97\xff";
5291
5292         // General name abuses
5293         abuseCases.push_back(CaseParameter("_has_very_long_name", longname));
5294         abuseCases.push_back(CaseParameter("_utf8_illegal", utf8illegal));
5295         abuseCases.push_back(CaseParameter("_utf8_nul", utf8nul));
5296         abuseCases.push_back(CaseParameter("_utf8_overlong", utf8overlong));
5297         abuseCases.push_back(CaseParameter("_utf8_zalgo", utf8zalgo));
5298
5299         // GL keywords
5300         abuseCases.push_back(CaseParameter("_is_gl_Position", "gl_Position"));
5301         abuseCases.push_back(CaseParameter("_is_gl_InstanceID", "gl_InstanceID"));
5302         abuseCases.push_back(CaseParameter("_is_gl_PrimitiveID", "gl_PrimitiveID"));
5303         abuseCases.push_back(CaseParameter("_is_gl_TessCoord", "gl_TessCoord"));
5304         abuseCases.push_back(CaseParameter("_is_gl_PerVertex", "gl_PerVertex"));
5305         abuseCases.push_back(CaseParameter("_is_gl_InvocationID", "gl_InvocationID"));
5306         abuseCases.push_back(CaseParameter("_is_gl_PointSize", "gl_PointSize"));
5307         abuseCases.push_back(CaseParameter("_is_gl_PointCoord", "gl_PointCoord"));
5308         abuseCases.push_back(CaseParameter("_is_gl_Layer", "gl_Layer"));
5309         abuseCases.push_back(CaseParameter("_is_gl_FragDepth", "gl_FragDepth"));
5310         abuseCases.push_back(CaseParameter("_is_gl_NumWorkGroups", "gl_NumWorkGroups"));
5311         abuseCases.push_back(CaseParameter("_is_gl_WorkGroupID", "gl_WorkGroupID"));
5312         abuseCases.push_back(CaseParameter("_is_gl_LocalInvocationID", "gl_LocalInvocationID"));
5313         abuseCases.push_back(CaseParameter("_is_gl_GlobalInvocationID", "gl_GlobalInvocationID"));
5314         abuseCases.push_back(CaseParameter("_is_gl_MaxVertexAttribs", "gl_MaxVertexAttribs"));
5315         abuseCases.push_back(CaseParameter("_is_gl_MaxViewports", "gl_MaxViewports"));
5316         abuseCases.push_back(CaseParameter("_is_gl_MaxComputeWorkGroupCount", "gl_MaxComputeWorkGroupCount"));
5317         abuseCases.push_back(CaseParameter("_is_mat3", "mat3"));
5318         abuseCases.push_back(CaseParameter("_is_volatile", "volatile"));
5319         abuseCases.push_back(CaseParameter("_is_inout", "inout"));
5320         abuseCases.push_back(CaseParameter("_is_isampler3d", "isampler3d"));
5321 }
5322
5323 tcu::TestCaseGroup* createOpNameGroup (tcu::TestContext& testCtx)
5324 {
5325         de::MovePtr<tcu::TestCaseGroup> group                   (new tcu::TestCaseGroup(testCtx, "opname", "Tests OpName cases"));
5326         de::MovePtr<tcu::TestCaseGroup> entryMainGroup  (new tcu::TestCaseGroup(testCtx, "entry_main", "OpName tests with entry main"));
5327         de::MovePtr<tcu::TestCaseGroup> entryNotGroup   (new tcu::TestCaseGroup(testCtx, "entry_rdc", "OpName tests with entry rdc"));
5328         de::MovePtr<tcu::TestCaseGroup> abuseGroup              (new tcu::TestCaseGroup(testCtx, "abuse", "OpName abuse tests"));
5329         vector<CaseParameter>                   cases;
5330         vector<CaseParameter>                   abuseCases;
5331         vector<string>                                  testFunc;
5332         de::Random                                              rnd                             (deStringHash(group->getName()));
5333         const int                                               numElements             = 128;
5334         vector<float>                                   inputFloats             (numElements, 0);
5335         vector<float>                                   outputFloats    (numElements, 0);
5336
5337         getOpNameAbuseCases(abuseCases);
5338
5339         fillRandomScalars(rnd, -100.0f, 100.0f, &inputFloats[0], numElements);
5340
5341         for(size_t ndx = 0; ndx < numElements; ++ndx)
5342                 outputFloats[ndx] = -inputFloats[ndx];
5343
5344         const string commonShaderHeader =
5345                 "OpCapability Shader\n"
5346                 "OpMemoryModel Logical GLSL450\n"
5347                 "OpEntryPoint GLCompute %main \"main\" %id\n"
5348                 "OpExecutionMode %main LocalSize 1 1 1\n";
5349
5350         const string commonShaderFooter =
5351                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
5352
5353                 + string(getComputeAsmInputOutputBufferTraits())
5354                 + string(getComputeAsmCommonTypes())
5355                 + string(getComputeAsmInputOutputBuffer()) +
5356
5357                 "%id        = OpVariable %uvec3ptr Input\n"
5358                 "%zero      = OpConstant %i32 0\n"
5359
5360                 "%func      = OpFunction %void None %voidf\n"
5361                 "%5         = OpLabel\n"
5362                 "             OpReturn\n"
5363                 "             OpFunctionEnd\n"
5364
5365                 "%main      = OpFunction %void None %voidf\n"
5366                 "%entry     = OpLabel\n"
5367                 "%7         = OpFunctionCall %void %func\n"
5368
5369                 "%idval     = OpLoad %uvec3 %id\n"
5370                 "%x         = OpCompositeExtract %u32 %idval 0\n"
5371
5372                 "%inloc     = OpAccessChain %f32ptr %indata %zero %x\n"
5373                 "%inval     = OpLoad %f32 %inloc\n"
5374                 "%neg       = OpFNegate %f32 %inval\n"
5375                 "%outloc    = OpAccessChain %f32ptr %outdata %zero %x\n"
5376                 "             OpStore %outloc %neg\n"
5377
5378                 "             OpReturn\n"
5379                 "             OpFunctionEnd\n";
5380
5381         const StringTemplate shaderTemplate (
5382                 "OpCapability Shader\n"
5383                 "OpMemoryModel Logical GLSL450\n"
5384                 "OpEntryPoint GLCompute %main \"${ENTRY}\" %id\n"
5385                 "OpExecutionMode %main LocalSize 1 1 1\n"
5386                 "OpName %${ID} \"${NAME}\"\n" +
5387                 commonShaderFooter);
5388
5389         const std::string multipleNames =
5390                 commonShaderHeader +
5391                 "OpName %main \"to_be\"\n"
5392                 "OpName %id   \"or_not\"\n"
5393                 "OpName %main \"to_be\"\n"
5394                 "OpName %main \"makes_no\"\n"
5395                 "OpName %func \"difference\"\n"
5396                 "OpName %5    \"to_me\"\n" +
5397                 commonShaderFooter;
5398
5399         {
5400                 ComputeShaderSpec       spec;
5401
5402                 spec.assembly           = multipleNames;
5403                 spec.numWorkGroups      = IVec3(numElements, 1, 1);
5404                 spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
5405                 spec.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
5406
5407                 abuseGroup->addChild(new SpvAsmComputeShaderCase(testCtx, "main_has_multiple_names", "multiple_names", spec));
5408         }
5409
5410         const std::string everythingNamed =
5411                 commonShaderHeader +
5412                 "OpName %main   \"name1\"\n"
5413                 "OpName %id     \"name2\"\n"
5414                 "OpName %zero   \"name3\"\n"
5415                 "OpName %entry  \"name4\"\n"
5416                 "OpName %func   \"name5\"\n"
5417                 "OpName %5      \"name6\"\n"
5418                 "OpName %7      \"name7\"\n"
5419                 "OpName %idval  \"name8\"\n"
5420                 "OpName %inloc  \"name9\"\n"
5421                 "OpName %inval  \"name10\"\n"
5422                 "OpName %neg    \"name11\"\n"
5423                 "OpName %outloc \"name12\"\n"+
5424                 commonShaderFooter;
5425         {
5426                 ComputeShaderSpec       spec;
5427
5428                 spec.assembly           = everythingNamed;
5429                 spec.numWorkGroups      = IVec3(numElements, 1, 1);
5430                 spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
5431                 spec.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
5432
5433                 abuseGroup->addChild(new SpvAsmComputeShaderCase(testCtx, "everything_named", "everything_named", spec));
5434         }
5435
5436         const std::string everythingNamedTheSame =
5437                 commonShaderHeader +
5438                 "OpName %main   \"the_same\"\n"
5439                 "OpName %id     \"the_same\"\n"
5440                 "OpName %zero   \"the_same\"\n"
5441                 "OpName %entry  \"the_same\"\n"
5442                 "OpName %func   \"the_same\"\n"
5443                 "OpName %5      \"the_same\"\n"
5444                 "OpName %7      \"the_same\"\n"
5445                 "OpName %idval  \"the_same\"\n"
5446                 "OpName %inloc  \"the_same\"\n"
5447                 "OpName %inval  \"the_same\"\n"
5448                 "OpName %neg    \"the_same\"\n"
5449                 "OpName %outloc \"the_same\"\n"+
5450                 commonShaderFooter;
5451         {
5452                 ComputeShaderSpec       spec;
5453
5454                 spec.assembly           = everythingNamedTheSame;
5455                 spec.numWorkGroups      = IVec3(numElements, 1, 1);
5456                 spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
5457                 spec.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
5458
5459                 abuseGroup->addChild(new SpvAsmComputeShaderCase(testCtx, "everything_named_the_same", "everything_named_the_same", spec));
5460         }
5461
5462         // main_is_...
5463         for (size_t ndx = 0; ndx < abuseCases.size(); ++ndx)
5464         {
5465                 map<string, string>     specializations;
5466                 ComputeShaderSpec       spec;
5467
5468                 specializations["ENTRY"]        = "main";
5469                 specializations["ID"]           = "main";
5470                 specializations["NAME"]         = abuseCases[ndx].param;
5471                 spec.assembly                           = shaderTemplate.specialize(specializations);
5472                 spec.numWorkGroups                      = IVec3(numElements, 1, 1);
5473                 spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
5474                 spec.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
5475
5476                 abuseGroup->addChild(new SpvAsmComputeShaderCase(testCtx, (std::string("main") + abuseCases[ndx].name).c_str(), abuseCases[ndx].name, spec));
5477         }
5478
5479         // x_is_....
5480         for (size_t ndx = 0; ndx < abuseCases.size(); ++ndx)
5481         {
5482                 map<string, string>     specializations;
5483                 ComputeShaderSpec       spec;
5484
5485                 specializations["ENTRY"]        = "main";
5486                 specializations["ID"]           = "x";
5487                 specializations["NAME"]         = abuseCases[ndx].param;
5488                 spec.assembly                           = shaderTemplate.specialize(specializations);
5489                 spec.numWorkGroups                      = IVec3(numElements, 1, 1);
5490                 spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
5491                 spec.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
5492
5493                 abuseGroup->addChild(new SpvAsmComputeShaderCase(testCtx, (std::string("x") + abuseCases[ndx].name).c_str(), abuseCases[ndx].name, spec));
5494         }
5495
5496         cases.push_back(CaseParameter("_is_main", "main"));
5497         cases.push_back(CaseParameter("_is_not_main", "not_main"));
5498         testFunc.push_back("main");
5499         testFunc.push_back("func");
5500
5501         for(size_t fNdx = 0; fNdx < testFunc.size(); ++fNdx)
5502         {
5503                 for(size_t ndx = 0; ndx < cases.size(); ++ndx)
5504                 {
5505                         map<string, string>     specializations;
5506                         ComputeShaderSpec       spec;
5507
5508                         specializations["ENTRY"]        = "main";
5509                         specializations["ID"]           = testFunc[fNdx];
5510                         specializations["NAME"]         = cases[ndx].param;
5511                         spec.assembly                           = shaderTemplate.specialize(specializations);
5512                         spec.numWorkGroups                      = IVec3(numElements, 1, 1);
5513                         spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
5514                         spec.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
5515
5516                         entryMainGroup->addChild(new SpvAsmComputeShaderCase(testCtx, (testFunc[fNdx] + cases[ndx].name).c_str(), cases[ndx].name, spec));
5517                 }
5518         }
5519
5520         cases.push_back(CaseParameter("_is_entry", "rdc"));
5521
5522         for(size_t fNdx = 0; fNdx < testFunc.size(); ++fNdx)
5523         {
5524                 for(size_t ndx = 0; ndx < cases.size(); ++ndx)
5525                 {
5526                         map<string, string>     specializations;
5527                         ComputeShaderSpec       spec;
5528
5529                         specializations["ENTRY"]        = "rdc";
5530                         specializations["ID"]           = testFunc[fNdx];
5531                         specializations["NAME"]         = cases[ndx].param;
5532                         spec.assembly                           = shaderTemplate.specialize(specializations);
5533                         spec.numWorkGroups                      = IVec3(numElements, 1, 1);
5534                         spec.entryPoint                         = "rdc";
5535                         spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
5536                         spec.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
5537
5538                         entryNotGroup->addChild(new SpvAsmComputeShaderCase(testCtx, (testFunc[fNdx] + cases[ndx].name).c_str(), cases[ndx].name, spec));
5539                 }
5540         }
5541
5542         group->addChild(entryMainGroup.release());
5543         group->addChild(entryNotGroup.release());
5544         group->addChild(abuseGroup.release());
5545
5546         return group.release();
5547 }
5548
5549 tcu::TestCaseGroup* createOpMemberNameGroup (tcu::TestContext& testCtx)
5550 {
5551         de::MovePtr<tcu::TestCaseGroup> group(new tcu::TestCaseGroup(testCtx, "opmembername", "Tests OpMemberName cases"));
5552         de::MovePtr<tcu::TestCaseGroup> abuseGroup(new tcu::TestCaseGroup(testCtx, "abuse", "OpMemberName abuse tests"));
5553         vector<CaseParameter>                   abuseCases;
5554         vector<string>                                  testFunc;
5555         de::Random                                              rnd(deStringHash(group->getName()));
5556         const int                                               numElements = 128;
5557         vector<float>                                   inputFloats(numElements, 0);
5558         vector<float>                                   outputFloats(numElements, 0);
5559
5560         getOpNameAbuseCases(abuseCases);
5561
5562         fillRandomScalars(rnd, -100.0f, 100.0f, &inputFloats[0], numElements);
5563
5564         for (size_t ndx = 0; ndx < numElements; ++ndx)
5565                 outputFloats[ndx] = -inputFloats[ndx];
5566
5567         const string commonShaderHeader =
5568                 "OpCapability Shader\n"
5569                 "OpMemoryModel Logical GLSL450\n"
5570                 "OpEntryPoint GLCompute %main \"main\" %id\n"
5571                 "OpExecutionMode %main LocalSize 1 1 1\n";
5572
5573         const string commonShaderFooter =
5574                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
5575
5576                 + string(getComputeAsmInputOutputBufferTraits())
5577                 + string(getComputeAsmCommonTypes())
5578                 + string(getComputeAsmInputOutputBuffer()) +
5579
5580                 "%u3str     = OpTypeStruct %u32 %u32 %u32\n"
5581
5582                 "%id        = OpVariable %uvec3ptr Input\n"
5583                 "%zero      = OpConstant %i32 0\n"
5584
5585                 "%main      = OpFunction %void None %voidf\n"
5586                 "%entry     = OpLabel\n"
5587
5588                 "%idval     = OpLoad %uvec3 %id\n"
5589                 "%x0        = OpCompositeExtract %u32 %idval 0\n"
5590
5591                 "%idstr     = OpCompositeConstruct %u3str %x0 %x0 %x0\n"
5592                 "%x         = OpCompositeExtract %u32 %idstr 0\n"
5593
5594                 "%inloc     = OpAccessChain %f32ptr %indata %zero %x\n"
5595                 "%inval     = OpLoad %f32 %inloc\n"
5596                 "%neg       = OpFNegate %f32 %inval\n"
5597                 "%outloc    = OpAccessChain %f32ptr %outdata %zero %x\n"
5598                 "             OpStore %outloc %neg\n"
5599
5600                 "             OpReturn\n"
5601                 "             OpFunctionEnd\n";
5602
5603         const StringTemplate shaderTemplate(
5604                 commonShaderHeader +
5605                 "OpMemberName %u3str 0 \"${NAME}\"\n" +
5606                 commonShaderFooter);
5607
5608         const std::string multipleNames =
5609                 commonShaderHeader +
5610                 "OpMemberName %u3str 0 \"to_be\"\n"
5611                 "OpMemberName %u3str 1 \"or_not\"\n"
5612                 "OpMemberName %u3str 0 \"to_be\"\n"
5613                 "OpMemberName %u3str 2 \"makes_no\"\n"
5614                 "OpMemberName %u3str 0 \"difference\"\n"
5615                 "OpMemberName %u3str 0 \"to_me\"\n" +
5616                 commonShaderFooter;
5617         {
5618                 ComputeShaderSpec       spec;
5619
5620                 spec.assembly = multipleNames;
5621                 spec.numWorkGroups = IVec3(numElements, 1, 1);
5622                 spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
5623                 spec.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
5624
5625                 abuseGroup->addChild(new SpvAsmComputeShaderCase(testCtx, "u3str_x_has_multiple_names", "multiple_names", spec));
5626         }
5627
5628         const std::string everythingNamedTheSame =
5629                 commonShaderHeader +
5630                 "OpMemberName %u3str 0 \"the_same\"\n"
5631                 "OpMemberName %u3str 1 \"the_same\"\n"
5632                 "OpMemberName %u3str 2 \"the_same\"\n" +
5633                 commonShaderFooter;
5634
5635         {
5636                 ComputeShaderSpec       spec;
5637
5638                 spec.assembly = everythingNamedTheSame;
5639                 spec.numWorkGroups = IVec3(numElements, 1, 1);
5640                 spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
5641                 spec.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
5642
5643                 abuseGroup->addChild(new SpvAsmComputeShaderCase(testCtx, "everything_named_the_same", "everything_named_the_same", spec));
5644         }
5645
5646         // u3str_x_is_....
5647         for (size_t ndx = 0; ndx < abuseCases.size(); ++ndx)
5648         {
5649                 map<string, string>     specializations;
5650                 ComputeShaderSpec       spec;
5651
5652                 specializations["NAME"] = abuseCases[ndx].param;
5653                 spec.assembly = shaderTemplate.specialize(specializations);
5654                 spec.numWorkGroups = IVec3(numElements, 1, 1);
5655                 spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
5656                 spec.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
5657
5658                 abuseGroup->addChild(new SpvAsmComputeShaderCase(testCtx, (std::string("u3str_x") + abuseCases[ndx].name).c_str(), abuseCases[ndx].name, spec));
5659         }
5660
5661         group->addChild(abuseGroup.release());
5662
5663         return group.release();
5664 }
5665
5666 // Assembly code used for testing function control is based on GLSL source code:
5667 //
5668 // #version 430
5669 //
5670 // layout(std140, set = 0, binding = 0) readonly buffer Input {
5671 //   float elements[];
5672 // } input_data;
5673 // layout(std140, set = 0, binding = 1) writeonly buffer Output {
5674 //   float elements[];
5675 // } output_data;
5676 //
5677 // float const10() { return 10.f; }
5678 //
5679 // void main() {
5680 //   uint x = gl_GlobalInvocationID.x;
5681 //   output_data.elements[x] = input_data.elements[x] + const10();
5682 // }
5683 tcu::TestCaseGroup* createFunctionControlGroup (tcu::TestContext& testCtx)
5684 {
5685         de::MovePtr<tcu::TestCaseGroup> group                   (new tcu::TestCaseGroup(testCtx, "function_control", "Tests function control cases"));
5686         vector<CaseParameter>                   cases;
5687         de::Random                                              rnd                             (deStringHash(group->getName()));
5688         const int                                               numElements             = 100;
5689         vector<float>                                   inputFloats             (numElements, 0);
5690         vector<float>                                   outputFloats    (numElements, 0);
5691         const StringTemplate                    shaderTemplate  (
5692                 string(getComputeAsmShaderPreamble()) +
5693
5694                 "OpSource GLSL 430\n"
5695                 "OpName %main \"main\"\n"
5696                 "OpName %func_const10 \"const10(\"\n"
5697                 "OpName %id \"gl_GlobalInvocationID\"\n"
5698
5699                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
5700
5701                 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) + string(getComputeAsmInputOutputBuffer()) +
5702
5703                 "%f32f = OpTypeFunction %f32\n"
5704                 "%id = OpVariable %uvec3ptr Input\n"
5705                 "%zero = OpConstant %i32 0\n"
5706                 "%constf10 = OpConstant %f32 10.0\n"
5707
5708                 "%main         = OpFunction %void None %voidf\n"
5709                 "%entry        = OpLabel\n"
5710                 "%idval        = OpLoad %uvec3 %id\n"
5711                 "%x            = OpCompositeExtract %u32 %idval 0\n"
5712                 "%inloc        = OpAccessChain %f32ptr %indata %zero %x\n"
5713                 "%inval        = OpLoad %f32 %inloc\n"
5714                 "%ret_10       = OpFunctionCall %f32 %func_const10\n"
5715                 "%fadd         = OpFAdd %f32 %inval %ret_10\n"
5716                 "%outloc       = OpAccessChain %f32ptr %outdata %zero %x\n"
5717                 "                OpStore %outloc %fadd\n"
5718                 "                OpReturn\n"
5719                 "                OpFunctionEnd\n"
5720
5721                 "%func_const10 = OpFunction %f32 ${CONTROL} %f32f\n"
5722                 "%label        = OpLabel\n"
5723                 "                OpReturnValue %constf10\n"
5724                 "                OpFunctionEnd\n");
5725
5726         cases.push_back(CaseParameter("none",                                           "None"));
5727         cases.push_back(CaseParameter("inline",                                         "Inline"));
5728         cases.push_back(CaseParameter("dont_inline",                            "DontInline"));
5729         cases.push_back(CaseParameter("pure",                                           "Pure"));
5730         cases.push_back(CaseParameter("const",                                          "Const"));
5731         cases.push_back(CaseParameter("inline_pure",                            "Inline|Pure"));
5732         cases.push_back(CaseParameter("const_dont_inline",                      "Const|DontInline"));
5733         cases.push_back(CaseParameter("inline_dont_inline",                     "Inline|DontInline"));
5734         cases.push_back(CaseParameter("pure_inline_dont_inline",        "Pure|Inline|DontInline"));
5735
5736         fillRandomScalars(rnd, -100.f, 100.f, &inputFloats[0], numElements);
5737
5738         // CPU might not use the same rounding mode as the GPU. Use whole numbers to avoid rounding differences.
5739         floorAll(inputFloats);
5740
5741         for (size_t ndx = 0; ndx < numElements; ++ndx)
5742                 outputFloats[ndx] = inputFloats[ndx] + 10.f;
5743
5744         for (size_t caseNdx = 0; caseNdx < cases.size(); ++caseNdx)
5745         {
5746                 map<string, string>             specializations;
5747                 ComputeShaderSpec               spec;
5748
5749                 specializations["CONTROL"] = cases[caseNdx].param;
5750                 spec.assembly = shaderTemplate.specialize(specializations);
5751                 spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
5752                 spec.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
5753                 spec.numWorkGroups = IVec3(numElements, 1, 1);
5754
5755                 group->addChild(new SpvAsmComputeShaderCase(testCtx, cases[caseNdx].name, cases[caseNdx].name, spec));
5756         }
5757
5758         return group.release();
5759 }
5760
5761 tcu::TestCaseGroup* createMemoryAccessGroup (tcu::TestContext& testCtx)
5762 {
5763         de::MovePtr<tcu::TestCaseGroup> group                   (new tcu::TestCaseGroup(testCtx, "memory_access", "Tests memory access cases"));
5764         vector<CaseParameter>                   cases;
5765         de::Random                                              rnd                             (deStringHash(group->getName()));
5766         const int                                               numElements             = 100;
5767         vector<float>                                   inputFloats             (numElements, 0);
5768         vector<float>                                   outputFloats    (numElements, 0);
5769         const StringTemplate                    shaderTemplate  (
5770                 string(getComputeAsmShaderPreamble()) +
5771
5772                 "OpSource GLSL 430\n"
5773                 "OpName %main           \"main\"\n"
5774                 "OpName %id             \"gl_GlobalInvocationID\"\n"
5775
5776                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
5777
5778                 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) + string(getComputeAsmInputOutputBuffer()) +
5779
5780                 "%f32ptr_f  = OpTypePointer Function %f32\n"
5781
5782                 "%id        = OpVariable %uvec3ptr Input\n"
5783                 "%zero      = OpConstant %i32 0\n"
5784                 "%four      = OpConstant %i32 4\n"
5785
5786                 "%main      = OpFunction %void None %voidf\n"
5787                 "%label     = OpLabel\n"
5788                 "%copy      = OpVariable %f32ptr_f Function\n"
5789                 "%idval     = OpLoad %uvec3 %id ${ACCESS}\n"
5790                 "%x         = OpCompositeExtract %u32 %idval 0\n"
5791                 "%inloc     = OpAccessChain %f32ptr %indata  %zero %x\n"
5792                 "%outloc    = OpAccessChain %f32ptr %outdata %zero %x\n"
5793                 "             OpCopyMemory %copy %inloc ${ACCESS}\n"
5794                 "%val1      = OpLoad %f32 %copy\n"
5795                 "%val2      = OpLoad %f32 %inloc\n"
5796                 "%add       = OpFAdd %f32 %val1 %val2\n"
5797                 "             OpStore %outloc %add ${ACCESS}\n"
5798                 "             OpReturn\n"
5799                 "             OpFunctionEnd\n");
5800
5801         cases.push_back(CaseParameter("null",                                   ""));
5802         cases.push_back(CaseParameter("none",                                   "None"));
5803         cases.push_back(CaseParameter("volatile",                               "Volatile"));
5804         cases.push_back(CaseParameter("aligned",                                "Aligned 4"));
5805         cases.push_back(CaseParameter("nontemporal",                    "Nontemporal"));
5806         cases.push_back(CaseParameter("aligned_nontemporal",    "Aligned|Nontemporal 4"));
5807         cases.push_back(CaseParameter("aligned_volatile",               "Volatile|Aligned 4"));
5808
5809         fillRandomScalars(rnd, -100.f, 100.f, &inputFloats[0], numElements);
5810
5811         for (size_t ndx = 0; ndx < numElements; ++ndx)
5812                 outputFloats[ndx] = inputFloats[ndx] + inputFloats[ndx];
5813
5814         for (size_t caseNdx = 0; caseNdx < cases.size(); ++caseNdx)
5815         {
5816                 map<string, string>             specializations;
5817                 ComputeShaderSpec               spec;
5818
5819                 specializations["ACCESS"] = cases[caseNdx].param;
5820                 spec.assembly = shaderTemplate.specialize(specializations);
5821                 spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
5822                 spec.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
5823                 spec.numWorkGroups = IVec3(numElements, 1, 1);
5824
5825                 group->addChild(new SpvAsmComputeShaderCase(testCtx, cases[caseNdx].name, cases[caseNdx].name, spec));
5826         }
5827
5828         return group.release();
5829 }
5830
5831 // Checks that we can get undefined values for various types, without exercising a computation with it.
5832 tcu::TestCaseGroup* createOpUndefGroup (tcu::TestContext& testCtx)
5833 {
5834         de::MovePtr<tcu::TestCaseGroup> group                   (new tcu::TestCaseGroup(testCtx, "opundef", "Tests the OpUndef instruction"));
5835         vector<CaseParameter>                   cases;
5836         de::Random                                              rnd                             (deStringHash(group->getName()));
5837         const int                                               numElements             = 100;
5838         vector<float>                                   positiveFloats  (numElements, 0);
5839         vector<float>                                   negativeFloats  (numElements, 0);
5840         const StringTemplate                    shaderTemplate  (
5841                 string(getComputeAsmShaderPreamble()) +
5842
5843                 "OpSource GLSL 430\n"
5844                 "OpName %main           \"main\"\n"
5845                 "OpName %id             \"gl_GlobalInvocationID\"\n"
5846
5847                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
5848
5849                 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) +
5850                 "%uvec2     = OpTypeVector %u32 2\n"
5851                 "%fvec4     = OpTypeVector %f32 4\n"
5852                 "%fmat33    = OpTypeMatrix %fvec3 3\n"
5853                 "%image     = OpTypeImage %f32 2D 0 0 0 1 Unknown\n"
5854                 "%sampler   = OpTypeSampler\n"
5855                 "%simage    = OpTypeSampledImage %image\n"
5856                 "%const100  = OpConstant %u32 100\n"
5857                 "%uarr100   = OpTypeArray %i32 %const100\n"
5858                 "%struct    = OpTypeStruct %f32 %i32 %u32\n"
5859                 "%pointer   = OpTypePointer Function %i32\n"
5860                 + string(getComputeAsmInputOutputBuffer()) +
5861
5862                 "%id        = OpVariable %uvec3ptr Input\n"
5863                 "%zero      = OpConstant %i32 0\n"
5864
5865                 "%main      = OpFunction %void None %voidf\n"
5866                 "%label     = OpLabel\n"
5867
5868                 "%undef     = OpUndef ${TYPE}\n"
5869
5870                 "%idval     = OpLoad %uvec3 %id\n"
5871                 "%x         = OpCompositeExtract %u32 %idval 0\n"
5872
5873                 "%inloc     = OpAccessChain %f32ptr %indata %zero %x\n"
5874                 "%inval     = OpLoad %f32 %inloc\n"
5875                 "%neg       = OpFNegate %f32 %inval\n"
5876                 "%outloc    = OpAccessChain %f32ptr %outdata %zero %x\n"
5877                 "             OpStore %outloc %neg\n"
5878                 "             OpReturn\n"
5879                 "             OpFunctionEnd\n");
5880
5881         cases.push_back(CaseParameter("bool",                   "%bool"));
5882         cases.push_back(CaseParameter("sint32",                 "%i32"));
5883         cases.push_back(CaseParameter("uint32",                 "%u32"));
5884         cases.push_back(CaseParameter("float32",                "%f32"));
5885         cases.push_back(CaseParameter("vec4float32",    "%fvec4"));
5886         cases.push_back(CaseParameter("vec2uint32",             "%uvec2"));
5887         cases.push_back(CaseParameter("matrix",                 "%fmat33"));
5888         cases.push_back(CaseParameter("image",                  "%image"));
5889         cases.push_back(CaseParameter("sampler",                "%sampler"));
5890         cases.push_back(CaseParameter("sampledimage",   "%simage"));
5891         cases.push_back(CaseParameter("array",                  "%uarr100"));
5892         cases.push_back(CaseParameter("runtimearray",   "%f32arr"));
5893         cases.push_back(CaseParameter("struct",                 "%struct"));
5894         cases.push_back(CaseParameter("pointer",                "%pointer"));
5895
5896         fillRandomScalars(rnd, 1.f, 100.f, &positiveFloats[0], numElements);
5897
5898         for (size_t ndx = 0; ndx < numElements; ++ndx)
5899                 negativeFloats[ndx] = -positiveFloats[ndx];
5900
5901         for (size_t caseNdx = 0; caseNdx < cases.size(); ++caseNdx)
5902         {
5903                 map<string, string>             specializations;
5904                 ComputeShaderSpec               spec;
5905
5906                 specializations["TYPE"] = cases[caseNdx].param;
5907                 spec.assembly = shaderTemplate.specialize(specializations);
5908                 spec.inputs.push_back(BufferSp(new Float32Buffer(positiveFloats)));
5909                 spec.outputs.push_back(BufferSp(new Float32Buffer(negativeFloats)));
5910                 spec.numWorkGroups = IVec3(numElements, 1, 1);
5911
5912                 group->addChild(new SpvAsmComputeShaderCase(testCtx, cases[caseNdx].name, cases[caseNdx].name, spec));
5913         }
5914
5915                 return group.release();
5916 }
5917
5918 // Checks that a compute shader can generate a constant composite value of various types, without exercising a computation on it.
5919 tcu::TestCaseGroup* createFloat16OpConstantCompositeGroup (tcu::TestContext& testCtx)
5920 {
5921         de::MovePtr<tcu::TestCaseGroup> group                   (new tcu::TestCaseGroup(testCtx, "opconstantcomposite", "Tests the OpConstantComposite instruction"));
5922         vector<CaseParameter>                   cases;
5923         de::Random                                              rnd                             (deStringHash(group->getName()));
5924         const int                                               numElements             = 100;
5925         vector<float>                                   positiveFloats  (numElements, 0);
5926         vector<float>                                   negativeFloats  (numElements, 0);
5927         const StringTemplate                    shaderTemplate  (
5928                 "OpCapability Shader\n"
5929                 "OpCapability Float16\n"
5930                 "OpMemoryModel Logical GLSL450\n"
5931                 "OpEntryPoint GLCompute %main \"main\" %id\n"
5932                 "OpExecutionMode %main LocalSize 1 1 1\n"
5933                 "OpSource GLSL 430\n"
5934                 "OpName %main           \"main\"\n"
5935                 "OpName %id             \"gl_GlobalInvocationID\"\n"
5936
5937                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
5938
5939                 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) + string(getComputeAsmInputOutputBuffer()) +
5940
5941                 "%id        = OpVariable %uvec3ptr Input\n"
5942                 "%zero      = OpConstant %i32 0\n"
5943                 "%f16       = OpTypeFloat 16\n"
5944                 "%c_f16_0   = OpConstant %f16 0.0\n"
5945                 "%c_f16_0_5 = OpConstant %f16 0.5\n"
5946                 "%c_f16_1   = OpConstant %f16 1.0\n"
5947                 "%v2f16     = OpTypeVector %f16 2\n"
5948                 "%v3f16     = OpTypeVector %f16 3\n"
5949                 "%v4f16     = OpTypeVector %f16 4\n"
5950
5951                 "${CONSTANT}\n"
5952
5953                 "%main      = OpFunction %void None %voidf\n"
5954                 "%label     = OpLabel\n"
5955                 "%idval     = OpLoad %uvec3 %id\n"
5956                 "%x         = OpCompositeExtract %u32 %idval 0\n"
5957                 "%inloc     = OpAccessChain %f32ptr %indata %zero %x\n"
5958                 "%inval     = OpLoad %f32 %inloc\n"
5959                 "%neg       = OpFNegate %f32 %inval\n"
5960                 "%outloc    = OpAccessChain %f32ptr %outdata %zero %x\n"
5961                 "             OpStore %outloc %neg\n"
5962                 "             OpReturn\n"
5963                 "             OpFunctionEnd\n");
5964
5965
5966         cases.push_back(CaseParameter("vector",                 "%const = OpConstantComposite %v3f16 %c_f16_0 %c_f16_0_5 %c_f16_1\n"));
5967         cases.push_back(CaseParameter("matrix",                 "%m3v3f16 = OpTypeMatrix %v3f16 3\n"
5968                                                                                                         "%vec = OpConstantComposite %v3f16 %c_f16_0 %c_f16_0_5 %c_f16_1\n"
5969                                                                                                         "%mat = OpConstantComposite %m3v3f16 %vec %vec %vec"));
5970         cases.push_back(CaseParameter("struct",                 "%m2v3f16 = OpTypeMatrix %v3f16 2\n"
5971                                                                                                         "%struct = OpTypeStruct %i32 %f16 %v3f16 %m2v3f16\n"
5972                                                                                                         "%vec = OpConstantComposite %v3f16 %c_f16_0 %c_f16_0_5 %c_f16_1\n"
5973                                                                                                         "%mat = OpConstantComposite %m2v3f16 %vec %vec\n"
5974                                                                                                         "%const = OpConstantComposite %struct %zero %c_f16_0_5 %vec %mat\n"));
5975         cases.push_back(CaseParameter("nested_struct",  "%st1 = OpTypeStruct %i32 %f16\n"
5976                                                                                                         "%st2 = OpTypeStruct %i32 %i32\n"
5977                                                                                                         "%struct = OpTypeStruct %st1 %st2\n"
5978                                                                                                         "%st1val = OpConstantComposite %st1 %zero %c_f16_0_5\n"
5979                                                                                                         "%st2val = OpConstantComposite %st2 %zero %zero\n"
5980                                                                                                         "%const = OpConstantComposite %struct %st1val %st2val"));
5981
5982         fillRandomScalars(rnd, 1.f, 100.f, &positiveFloats[0], numElements);
5983
5984         for (size_t ndx = 0; ndx < numElements; ++ndx)
5985                 negativeFloats[ndx] = -positiveFloats[ndx];
5986
5987         for (size_t caseNdx = 0; caseNdx < cases.size(); ++caseNdx)
5988         {
5989                 map<string, string>             specializations;
5990                 ComputeShaderSpec               spec;
5991
5992                 specializations["CONSTANT"] = cases[caseNdx].param;
5993                 spec.assembly = shaderTemplate.specialize(specializations);
5994                 spec.inputs.push_back(BufferSp(new Float32Buffer(positiveFloats)));
5995                 spec.outputs.push_back(BufferSp(new Float32Buffer(negativeFloats)));
5996                 spec.numWorkGroups = IVec3(numElements, 1, 1);
5997
5998                 spec.extensions.push_back("VK_KHR_16bit_storage");
5999                 spec.extensions.push_back("VK_KHR_shader_float16_int8");
6000
6001                 spec.requestedVulkanFeatures.ext16BitStorage = EXT16BITSTORAGEFEATURES_UNIFORM_BUFFER_BLOCK;
6002                 spec.requestedVulkanFeatures.extFloat16Int8 = EXTFLOAT16INT8FEATURES_FLOAT16;
6003
6004                 group->addChild(new SpvAsmComputeShaderCase(testCtx, cases[caseNdx].name, cases[caseNdx].name, spec));
6005         }
6006
6007         return group.release();
6008 }
6009
6010 const vector<deFloat16> squarize(const vector<deFloat16>& inData, const deUint32 argNo)
6011 {
6012         const size_t            inDataLength    = inData.size();
6013         vector<deFloat16>       result;
6014
6015         result.reserve(inDataLength * inDataLength);
6016
6017         if (argNo == 0)
6018         {
6019                 for (size_t numIdx = 0; numIdx < inDataLength; ++numIdx)
6020                         result.insert(result.end(), inData.begin(), inData.end());
6021         }
6022
6023         if (argNo == 1)
6024         {
6025                 for (size_t numIdx = 0; numIdx < inDataLength; ++numIdx)
6026                 {
6027                         const vector<deFloat16> tmp(inDataLength, inData[numIdx]);
6028
6029                         result.insert(result.end(), tmp.begin(), tmp.end());
6030                 }
6031         }
6032
6033         return result;
6034 }
6035
6036 const vector<deFloat16> squarizeVector(const vector<deFloat16>& inData, const deUint32 argNo)
6037 {
6038         vector<deFloat16>       vec;
6039         vector<deFloat16>       result;
6040
6041         // Create vectors. vec will contain each possible pair from inData
6042         {
6043                 const size_t    inDataLength    = inData.size();
6044
6045                 DE_ASSERT(inDataLength <= 64);
6046
6047                 vec.reserve(2 * inDataLength * inDataLength);
6048
6049                 for (size_t numIdxX = 0; numIdxX < inDataLength; ++numIdxX)
6050                 for (size_t numIdxY = 0; numIdxY < inDataLength; ++numIdxY)
6051                 {
6052                         vec.push_back(inData[numIdxX]);
6053                         vec.push_back(inData[numIdxY]);
6054                 }
6055         }
6056
6057         // Create vector pairs. result will contain each possible pair from vec
6058         {
6059                 const size_t    coordsPerVector = 2;
6060                 const size_t    vectorsCount    = vec.size() / coordsPerVector;
6061
6062                 result.reserve(coordsPerVector * vectorsCount * vectorsCount);
6063
6064                 if (argNo == 0)
6065                 {
6066                         for (size_t numIdxX = 0; numIdxX < vectorsCount; ++numIdxX)
6067                         for (size_t numIdxY = 0; numIdxY < vectorsCount; ++numIdxY)
6068                         {
6069                                 for (size_t coordNdx = 0; coordNdx < coordsPerVector; ++coordNdx)
6070                                         result.push_back(vec[coordsPerVector * numIdxY + coordNdx]);
6071                         }
6072                 }
6073
6074                 if (argNo == 1)
6075                 {
6076                         for (size_t numIdxX = 0; numIdxX < vectorsCount; ++numIdxX)
6077                         for (size_t numIdxY = 0; numIdxY < vectorsCount; ++numIdxY)
6078                         {
6079                                 for (size_t coordNdx = 0; coordNdx < coordsPerVector; ++coordNdx)
6080                                         result.push_back(vec[coordsPerVector * numIdxX + coordNdx]);
6081                         }
6082                 }
6083         }
6084
6085         return result;
6086 }
6087
6088 struct fp16isNan                        { bool operator()(const tcu::Float16 in1, const tcu::Float16)           { return in1.isNaN(); } };
6089 struct fp16isInf                        { bool operator()(const tcu::Float16 in1, const tcu::Float16)           { return in1.isInf(); } };
6090 struct fp16isEqual                      { bool operator()(const tcu::Float16 in1, const tcu::Float16 in2)       { return in1.asFloat() == in2.asFloat(); } };
6091 struct fp16isUnequal            { bool operator()(const tcu::Float16 in1, const tcu::Float16 in2)       { return in1.asFloat() != in2.asFloat(); } };
6092 struct fp16isLess                       { bool operator()(const tcu::Float16 in1, const tcu::Float16 in2)       { return in1.asFloat() <  in2.asFloat(); } };
6093 struct fp16isGreater            { bool operator()(const tcu::Float16 in1, const tcu::Float16 in2)       { return in1.asFloat() >  in2.asFloat(); } };
6094 struct fp16isLessOrEqual        { bool operator()(const tcu::Float16 in1, const tcu::Float16 in2)       { return in1.asFloat() <= in2.asFloat(); } };
6095 struct fp16isGreaterOrEqual     { bool operator()(const tcu::Float16 in1, const tcu::Float16 in2)       { return in1.asFloat() >= in2.asFloat(); } };
6096
6097 template <class TestedLogicalFunction, bool onlyTestFunc, bool unationModeAnd, bool nanSupported>
6098 bool compareFP16Logical (const std::vector<Resource>& inputs, const vector<AllocationSp>& outputAllocs, const std::vector<Resource>&, TestLog& log)
6099 {
6100         if (inputs.size() != 2 || outputAllocs.size() != 1)
6101                 return false;
6102
6103         vector<deUint8> input1Bytes;
6104         vector<deUint8> input2Bytes;
6105
6106         inputs[0].getBytes(input1Bytes);
6107         inputs[1].getBytes(input2Bytes);
6108
6109         const deUint32                  denormModesCount                        = 2;
6110         const deFloat16                 float16one                                      = tcu::Float16(1.0f).bits();
6111         const deFloat16                 float16zero                                     = tcu::Float16(0.0f).bits();
6112         const tcu::Float16              zero                                            = tcu::Float16::zero(1);
6113         const deFloat16* const  outputAsFP16                            = static_cast<deFloat16*>(outputAllocs[0]->getHostPtr());
6114         const deFloat16* const  input1AsFP16                            = reinterpret_cast<deFloat16* const>(&input1Bytes.front());
6115         const deFloat16* const  input2AsFP16                            = reinterpret_cast<deFloat16* const>(&input2Bytes.front());
6116         deUint32                                successfulRuns                          = denormModesCount;
6117         std::string                             results[denormModesCount];
6118         TestedLogicalFunction   testedLogicalFunction;
6119
6120         for (deUint32 denormMode = 0; denormMode < denormModesCount; denormMode++)
6121         {
6122                 const bool flushToZero = (denormMode == 1);
6123
6124                 for (size_t idx = 0; idx < input1Bytes.size() / sizeof(deFloat16); ++idx)
6125                 {
6126                         const tcu::Float16      f1pre                   = tcu::Float16(input1AsFP16[idx]);
6127                         const tcu::Float16      f2pre                   = tcu::Float16(input2AsFP16[idx]);
6128                         const tcu::Float16      f1                              = (flushToZero && f1pre.isDenorm()) ? zero : f1pre;
6129                         const tcu::Float16      f2                              = (flushToZero && f2pre.isDenorm()) ? zero : f2pre;
6130                         deFloat16                       expectedOutput  = float16zero;
6131
6132                         if (onlyTestFunc)
6133                         {
6134                                 if (testedLogicalFunction(f1, f2))
6135                                         expectedOutput = float16one;
6136                         }
6137                         else
6138                         {
6139                                 const bool      f1nan   = f1.isNaN();
6140                                 const bool      f2nan   = f2.isNaN();
6141
6142                                 // Skip NaN floats if not supported by implementation
6143                                 if (!nanSupported && (f1nan || f2nan))
6144                                         continue;
6145
6146                                 if (unationModeAnd)
6147                                 {
6148                                         const bool      ordered         = !f1nan && !f2nan;
6149
6150                                         if (ordered && testedLogicalFunction(f1, f2))
6151                                                 expectedOutput = float16one;
6152                                 }
6153                                 else
6154                                 {
6155                                         const bool      unordered       = f1nan || f2nan;
6156
6157                                         if (unordered || testedLogicalFunction(f1, f2))
6158                                                 expectedOutput = float16one;
6159                                 }
6160                         }
6161
6162                         if (outputAsFP16[idx] != expectedOutput)
6163                         {
6164                                 std::ostringstream str;
6165
6166                                 str << "ERROR: Sub-case #" << idx
6167                                         << " flushToZero:" << flushToZero
6168                                         << std::hex
6169                                         << " failed, inputs: 0x" << f1.bits()
6170                                         << ";0x" << f2.bits()
6171                                         << " output: 0x" << outputAsFP16[idx]
6172                                         << " expected output: 0x" << expectedOutput;
6173
6174                                 results[denormMode] = str.str();
6175
6176                                 successfulRuns--;
6177
6178                                 break;
6179                         }
6180                 }
6181         }
6182
6183         if (successfulRuns == 0)
6184                 for (deUint32 denormMode = 0; denormMode < denormModesCount; denormMode++)
6185                         log << TestLog::Message << results[denormMode] << TestLog::EndMessage;
6186
6187         return successfulRuns > 0;
6188 }
6189
6190 } // anonymous
6191
6192 tcu::TestCaseGroup* createOpSourceTests (tcu::TestContext& testCtx)
6193 {
6194         struct NameCodePair { string name, code; };
6195         RGBA                                                    defaultColors[4];
6196         de::MovePtr<tcu::TestCaseGroup> opSourceTests                   (new tcu::TestCaseGroup(testCtx, "opsource", "OpSource instruction"));
6197         const std::string                               opsourceGLSLWithFile    = "%opsrcfile = OpString \"foo.vert\"\nOpSource GLSL 450 %opsrcfile ";
6198         map<string, string>                             fragments                               = passthruFragments();
6199         const NameCodePair                              tests[]                                 =
6200         {
6201                 {"unknown", "OpSource Unknown 321"},
6202                 {"essl", "OpSource ESSL 310"},
6203                 {"glsl", "OpSource GLSL 450"},
6204                 {"opencl_cpp", "OpSource OpenCL_CPP 120"},
6205                 {"opencl_c", "OpSource OpenCL_C 120"},
6206                 {"multiple", "OpSource GLSL 450\nOpSource GLSL 450"},
6207                 {"file", opsourceGLSLWithFile},
6208                 {"source", opsourceGLSLWithFile + "\"void main(){}\""},
6209                 // Longest possible source string: SPIR-V limits instructions to 65535
6210                 // words, of which the first 4 are opsourceGLSLWithFile; the rest will
6211                 // contain 65530 UTF8 characters (one word each) plus one last word
6212                 // containing 3 ASCII characters and \0.
6213                 {"longsource", opsourceGLSLWithFile + '"' + makeLongUTF8String(65530) + "ccc" + '"'}
6214         };
6215
6216         getDefaultColors(defaultColors);
6217         for (size_t testNdx = 0; testNdx < sizeof(tests) / sizeof(NameCodePair); ++testNdx)
6218         {
6219                 fragments["debug"] = tests[testNdx].code;
6220                 createTestsForAllStages(tests[testNdx].name, defaultColors, defaultColors, fragments, opSourceTests.get());
6221         }
6222
6223         return opSourceTests.release();
6224 }
6225
6226 tcu::TestCaseGroup* createOpSourceContinuedTests (tcu::TestContext& testCtx)
6227 {
6228         struct NameCodePair { string name, code; };
6229         RGBA                                                            defaultColors[4];
6230         de::MovePtr<tcu::TestCaseGroup>         opSourceTests           (new tcu::TestCaseGroup(testCtx, "opsourcecontinued", "OpSourceContinued instruction"));
6231         map<string, string>                                     fragments                       = passthruFragments();
6232         const std::string                                       opsource                        = "%opsrcfile = OpString \"foo.vert\"\nOpSource GLSL 450 %opsrcfile \"void main(){}\"\n";
6233         const NameCodePair                                      tests[]                         =
6234         {
6235                 {"empty", opsource + "OpSourceContinued \"\""},
6236                 {"short", opsource + "OpSourceContinued \"abcde\""},
6237                 {"multiple", opsource + "OpSourceContinued \"abcde\"\nOpSourceContinued \"fghij\""},
6238                 // Longest possible source string: SPIR-V limits instructions to 65535
6239                 // words, of which the first one is OpSourceContinued/length; the rest
6240                 // will contain 65533 UTF8 characters (one word each) plus one last word
6241                 // containing 3 ASCII characters and \0.
6242                 {"long", opsource + "OpSourceContinued \"" + makeLongUTF8String(65533) + "ccc\""}
6243         };
6244
6245         getDefaultColors(defaultColors);
6246         for (size_t testNdx = 0; testNdx < sizeof(tests) / sizeof(NameCodePair); ++testNdx)
6247         {
6248                 fragments["debug"] = tests[testNdx].code;
6249                 createTestsForAllStages(tests[testNdx].name, defaultColors, defaultColors, fragments, opSourceTests.get());
6250         }
6251
6252         return opSourceTests.release();
6253 }
6254 tcu::TestCaseGroup* createOpNoLineTests(tcu::TestContext& testCtx)
6255 {
6256         RGBA                                                             defaultColors[4];
6257         de::MovePtr<tcu::TestCaseGroup>          opLineTests             (new tcu::TestCaseGroup(testCtx, "opnoline", "OpNoLine instruction"));
6258         map<string, string>                                      fragments;
6259         getDefaultColors(defaultColors);
6260         fragments["debug"]                      =
6261                 "%name = OpString \"name\"\n";
6262
6263         fragments["pre_main"]   =
6264                 "OpNoLine\n"
6265                 "OpNoLine\n"
6266                 "OpLine %name 1 1\n"
6267                 "OpNoLine\n"
6268                 "OpLine %name 1 1\n"
6269                 "OpLine %name 1 1\n"
6270                 "%second_function = OpFunction %v4f32 None %v4f32_v4f32_function\n"
6271                 "OpNoLine\n"
6272                 "OpLine %name 1 1\n"
6273                 "OpNoLine\n"
6274                 "OpLine %name 1 1\n"
6275                 "OpLine %name 1 1\n"
6276                 "%second_param1 = OpFunctionParameter %v4f32\n"
6277                 "OpNoLine\n"
6278                 "OpNoLine\n"
6279                 "%label_secondfunction = OpLabel\n"
6280                 "OpNoLine\n"
6281                 "OpReturnValue %second_param1\n"
6282                 "OpFunctionEnd\n"
6283                 "OpNoLine\n"
6284                 "OpNoLine\n";
6285
6286         fragments["testfun"]            =
6287                 // A %test_code function that returns its argument unchanged.
6288                 "OpNoLine\n"
6289                 "OpNoLine\n"
6290                 "OpLine %name 1 1\n"
6291                 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
6292                 "OpNoLine\n"
6293                 "%param1 = OpFunctionParameter %v4f32\n"
6294                 "OpNoLine\n"
6295                 "OpNoLine\n"
6296                 "%label_testfun = OpLabel\n"
6297                 "OpNoLine\n"
6298                 "%val1 = OpFunctionCall %v4f32 %second_function %param1\n"
6299                 "OpReturnValue %val1\n"
6300                 "OpFunctionEnd\n"
6301                 "OpLine %name 1 1\n"
6302                 "OpNoLine\n";
6303
6304         createTestsForAllStages("opnoline", defaultColors, defaultColors, fragments, opLineTests.get());
6305
6306         return opLineTests.release();
6307 }
6308
6309 tcu::TestCaseGroup* createOpModuleProcessedTests(tcu::TestContext& testCtx)
6310 {
6311         RGBA                                                            defaultColors[4];
6312         de::MovePtr<tcu::TestCaseGroup>         opModuleProcessedTests                  (new tcu::TestCaseGroup(testCtx, "opmoduleprocessed", "OpModuleProcessed instruction"));
6313         map<string, string>                                     fragments;
6314         std::vector<std::string>                        noExtensions;
6315         GraphicsResources                                       resources;
6316
6317         getDefaultColors(defaultColors);
6318         resources.verifyBinary = veryfiBinaryShader;
6319         resources.spirvVersion = SPIRV_VERSION_1_3;
6320
6321         fragments["moduleprocessed"]                                                    =
6322                 "OpModuleProcessed \"VULKAN CTS\"\n"
6323                 "OpModuleProcessed \"Negative values\"\n"
6324                 "OpModuleProcessed \"Date: 2017/09/21\"\n";
6325
6326         fragments["pre_main"]   =
6327                 "%second_function = OpFunction %v4f32 None %v4f32_v4f32_function\n"
6328                 "%second_param1 = OpFunctionParameter %v4f32\n"
6329                 "%label_secondfunction = OpLabel\n"
6330                 "OpReturnValue %second_param1\n"
6331                 "OpFunctionEnd\n";
6332
6333         fragments["testfun"]            =
6334                 // A %test_code function that returns its argument unchanged.
6335                 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
6336                 "%param1 = OpFunctionParameter %v4f32\n"
6337                 "%label_testfun = OpLabel\n"
6338                 "%val1 = OpFunctionCall %v4f32 %second_function %param1\n"
6339                 "OpReturnValue %val1\n"
6340                 "OpFunctionEnd\n";
6341
6342         createTestsForAllStages ("opmoduleprocessed", defaultColors, defaultColors, fragments, resources, noExtensions, opModuleProcessedTests.get());
6343
6344         return opModuleProcessedTests.release();
6345 }
6346
6347
6348 tcu::TestCaseGroup* createOpLineTests(tcu::TestContext& testCtx)
6349 {
6350         RGBA                                                                                                    defaultColors[4];
6351         de::MovePtr<tcu::TestCaseGroup>                                                 opLineTests                     (new tcu::TestCaseGroup(testCtx, "opline", "OpLine instruction"));
6352         map<string, string>                                                                             fragments;
6353         std::vector<std::pair<std::string, std::string> >               problemStrings;
6354
6355         problemStrings.push_back(std::make_pair<std::string, std::string>("empty_name", ""));
6356         problemStrings.push_back(std::make_pair<std::string, std::string>("short_name", "short_name"));
6357         problemStrings.push_back(std::make_pair<std::string, std::string>("long_name", makeLongUTF8String(65530) + "ccc"));
6358         getDefaultColors(defaultColors);
6359
6360         fragments["debug"]                      =
6361                 "%other_name = OpString \"other_name\"\n";
6362
6363         fragments["pre_main"]   =
6364                 "OpLine %file_name 32 0\n"
6365                 "OpLine %file_name 32 32\n"
6366                 "OpLine %file_name 32 40\n"
6367                 "OpLine %other_name 32 40\n"
6368                 "OpLine %other_name 0 100\n"
6369                 "OpLine %other_name 0 4294967295\n"
6370                 "OpLine %other_name 4294967295 0\n"
6371                 "OpLine %other_name 32 40\n"
6372                 "OpLine %file_name 0 0\n"
6373                 "%second_function = OpFunction %v4f32 None %v4f32_v4f32_function\n"
6374                 "OpLine %file_name 1 0\n"
6375                 "%second_param1 = OpFunctionParameter %v4f32\n"
6376                 "OpLine %file_name 1 3\n"
6377                 "OpLine %file_name 1 2\n"
6378                 "%label_secondfunction = OpLabel\n"
6379                 "OpLine %file_name 0 2\n"
6380                 "OpReturnValue %second_param1\n"
6381                 "OpFunctionEnd\n"
6382                 "OpLine %file_name 0 2\n"
6383                 "OpLine %file_name 0 2\n";
6384
6385         fragments["testfun"]            =
6386                 // A %test_code function that returns its argument unchanged.
6387                 "OpLine %file_name 1 0\n"
6388                 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
6389                 "OpLine %file_name 16 330\n"
6390                 "%param1 = OpFunctionParameter %v4f32\n"
6391                 "OpLine %file_name 14 442\n"
6392                 "%label_testfun = OpLabel\n"
6393                 "OpLine %file_name 11 1024\n"
6394                 "%val1 = OpFunctionCall %v4f32 %second_function %param1\n"
6395                 "OpLine %file_name 2 97\n"
6396                 "OpReturnValue %val1\n"
6397                 "OpFunctionEnd\n"
6398                 "OpLine %file_name 5 32\n";
6399
6400         for (size_t i = 0; i < problemStrings.size(); ++i)
6401         {
6402                 map<string, string> testFragments = fragments;
6403                 testFragments["debug"] += "%file_name = OpString \"" + problemStrings[i].second + "\"\n";
6404                 createTestsForAllStages(string("opline") + "_" + problemStrings[i].first, defaultColors, defaultColors, testFragments, opLineTests.get());
6405         }
6406
6407         return opLineTests.release();
6408 }
6409
6410 tcu::TestCaseGroup* createOpConstantNullTests(tcu::TestContext& testCtx)
6411 {
6412         de::MovePtr<tcu::TestCaseGroup> opConstantNullTests             (new tcu::TestCaseGroup(testCtx, "opconstantnull", "OpConstantNull instruction"));
6413         RGBA                                                    colors[4];
6414
6415
6416         const char                                              functionStart[] =
6417                 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
6418                 "%param1 = OpFunctionParameter %v4f32\n"
6419                 "%lbl    = OpLabel\n";
6420
6421         const char                                              functionEnd[]   =
6422                 "OpReturnValue %transformed_param\n"
6423                 "OpFunctionEnd\n";
6424
6425         struct NameConstantsCode
6426         {
6427                 string name;
6428                 string constants;
6429                 string code;
6430         };
6431
6432         NameConstantsCode tests[] =
6433         {
6434                 {
6435                         "vec4",
6436                         "%cnull = OpConstantNull %v4f32\n",
6437                         "%transformed_param = OpFAdd %v4f32 %param1 %cnull\n"
6438                 },
6439                 {
6440                         "float",
6441                         "%cnull = OpConstantNull %f32\n",
6442                         "%vp = OpVariable %fp_v4f32 Function\n"
6443                         "%v  = OpLoad %v4f32 %vp\n"
6444                         "%v0 = OpVectorInsertDynamic %v4f32 %v %cnull %c_i32_0\n"
6445                         "%v1 = OpVectorInsertDynamic %v4f32 %v0 %cnull %c_i32_1\n"
6446                         "%v2 = OpVectorInsertDynamic %v4f32 %v1 %cnull %c_i32_2\n"
6447                         "%v3 = OpVectorInsertDynamic %v4f32 %v2 %cnull %c_i32_3\n"
6448                         "%transformed_param = OpFAdd %v4f32 %param1 %v3\n"
6449                 },
6450                 {
6451                         "bool",
6452                         "%cnull             = OpConstantNull %bool\n",
6453                         "%v                 = OpVariable %fp_v4f32 Function\n"
6454                         "                     OpStore %v %param1\n"
6455                         "                     OpSelectionMerge %false_label None\n"
6456                         "                     OpBranchConditional %cnull %true_label %false_label\n"
6457                         "%true_label        = OpLabel\n"
6458                         "                     OpStore %v %c_v4f32_0_5_0_5_0_5_0_5\n"
6459                         "                     OpBranch %false_label\n"
6460                         "%false_label       = OpLabel\n"
6461                         "%transformed_param = OpLoad %v4f32 %v\n"
6462                 },
6463                 {
6464                         "i32",
6465                         "%cnull             = OpConstantNull %i32\n",
6466                         "%v                 = OpVariable %fp_v4f32 Function %c_v4f32_0_5_0_5_0_5_0_5\n"
6467                         "%b                 = OpIEqual %bool %cnull %c_i32_0\n"
6468                         "                     OpSelectionMerge %false_label None\n"
6469                         "                     OpBranchConditional %b %true_label %false_label\n"
6470                         "%true_label        = OpLabel\n"
6471                         "                     OpStore %v %param1\n"
6472                         "                     OpBranch %false_label\n"
6473                         "%false_label       = OpLabel\n"
6474                         "%transformed_param = OpLoad %v4f32 %v\n"
6475                 },
6476                 {
6477                         "struct",
6478                         "%stype             = OpTypeStruct %f32 %v4f32\n"
6479                         "%fp_stype          = OpTypePointer Function %stype\n"
6480                         "%cnull             = OpConstantNull %stype\n",
6481                         "%v                 = OpVariable %fp_stype Function %cnull\n"
6482                         "%f                 = OpAccessChain %fp_v4f32 %v %c_i32_1\n"
6483                         "%f_val             = OpLoad %v4f32 %f\n"
6484                         "%transformed_param = OpFAdd %v4f32 %param1 %f_val\n"
6485                 },
6486                 {
6487                         "array",
6488                         "%a4_v4f32          = OpTypeArray %v4f32 %c_u32_4\n"
6489                         "%fp_a4_v4f32       = OpTypePointer Function %a4_v4f32\n"
6490                         "%cnull             = OpConstantNull %a4_v4f32\n",
6491                         "%v                 = OpVariable %fp_a4_v4f32 Function %cnull\n"
6492                         "%f                 = OpAccessChain %fp_v4f32 %v %c_u32_0\n"
6493                         "%f1                = OpAccessChain %fp_v4f32 %v %c_u32_1\n"
6494                         "%f2                = OpAccessChain %fp_v4f32 %v %c_u32_2\n"
6495                         "%f3                = OpAccessChain %fp_v4f32 %v %c_u32_3\n"
6496                         "%f_val             = OpLoad %v4f32 %f\n"
6497                         "%f1_val            = OpLoad %v4f32 %f1\n"
6498                         "%f2_val            = OpLoad %v4f32 %f2\n"
6499                         "%f3_val            = OpLoad %v4f32 %f3\n"
6500                         "%t0                = OpFAdd %v4f32 %param1 %f_val\n"
6501                         "%t1                = OpFAdd %v4f32 %t0 %f1_val\n"
6502                         "%t2                = OpFAdd %v4f32 %t1 %f2_val\n"
6503                         "%transformed_param = OpFAdd %v4f32 %t2 %f3_val\n"
6504                 },
6505                 {
6506                         "matrix",
6507                         "%mat4x4_f32        = OpTypeMatrix %v4f32 4\n"
6508                         "%cnull             = OpConstantNull %mat4x4_f32\n",
6509                         // Our null matrix * any vector should result in a zero vector.
6510                         "%v                 = OpVectorTimesMatrix %v4f32 %param1 %cnull\n"
6511                         "%transformed_param = OpFAdd %v4f32 %param1 %v\n"
6512                 }
6513         };
6514
6515         getHalfColorsFullAlpha(colors);
6516
6517         for (size_t testNdx = 0; testNdx < sizeof(tests) / sizeof(NameConstantsCode); ++testNdx)
6518         {
6519                 map<string, string> fragments;
6520                 fragments["pre_main"] = tests[testNdx].constants;
6521                 fragments["testfun"] = string(functionStart) + tests[testNdx].code + functionEnd;
6522                 createTestsForAllStages(tests[testNdx].name, colors, colors, fragments, opConstantNullTests.get());
6523         }
6524         return opConstantNullTests.release();
6525 }
6526 tcu::TestCaseGroup* createOpConstantCompositeTests(tcu::TestContext& testCtx)
6527 {
6528         de::MovePtr<tcu::TestCaseGroup> opConstantCompositeTests                (new tcu::TestCaseGroup(testCtx, "opconstantcomposite", "OpConstantComposite instruction"));
6529         RGBA                                                    inputColors[4];
6530         RGBA                                                    outputColors[4];
6531
6532
6533         const char                                              functionStart[]  =
6534                 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
6535                 "%param1 = OpFunctionParameter %v4f32\n"
6536                 "%lbl    = OpLabel\n";
6537
6538         const char                                              functionEnd[]           =
6539                 "OpReturnValue %transformed_param\n"
6540                 "OpFunctionEnd\n";
6541
6542         struct NameConstantsCode
6543         {
6544                 string name;
6545                 string constants;
6546                 string code;
6547         };
6548
6549         NameConstantsCode tests[] =
6550         {
6551                 {
6552                         "vec4",
6553
6554                         "%cval              = OpConstantComposite %v4f32 %c_f32_0_5 %c_f32_0_5 %c_f32_0_5 %c_f32_0\n",
6555                         "%transformed_param = OpFAdd %v4f32 %param1 %cval\n"
6556                 },
6557                 {
6558                         "struct",
6559
6560                         "%stype             = OpTypeStruct %v4f32 %f32\n"
6561                         "%fp_stype          = OpTypePointer Function %stype\n"
6562                         "%f32_n_1           = OpConstant %f32 -1.0\n"
6563                         "%f32_1_5           = OpConstant %f32 !0x3fc00000\n" // +1.5
6564                         "%cvec              = OpConstantComposite %v4f32 %f32_1_5 %f32_1_5 %f32_1_5 %c_f32_1\n"
6565                         "%cval              = OpConstantComposite %stype %cvec %f32_n_1\n",
6566
6567                         "%v                 = OpVariable %fp_stype Function %cval\n"
6568                         "%vec_ptr           = OpAccessChain %fp_v4f32 %v %c_u32_0\n"
6569                         "%f32_ptr           = OpAccessChain %fp_f32 %v %c_u32_1\n"
6570                         "%vec_val           = OpLoad %v4f32 %vec_ptr\n"
6571                         "%f32_val           = OpLoad %f32 %f32_ptr\n"
6572                         "%tmp1              = OpVectorTimesScalar %v4f32 %c_v4f32_1_1_1_1 %f32_val\n" // vec4(-1)
6573                         "%tmp2              = OpFAdd %v4f32 %tmp1 %param1\n" // param1 + vec4(-1)
6574                         "%transformed_param = OpFAdd %v4f32 %tmp2 %vec_val\n" // param1 + vec4(-1) + vec4(1.5, 1.5, 1.5, 1.0)
6575                 },
6576                 {
6577                         // [1|0|0|0.5] [x] = x + 0.5
6578                         // [0|1|0|0.5] [y] = y + 0.5
6579                         // [0|0|1|0.5] [z] = z + 0.5
6580                         // [0|0|0|1  ] [1] = 1
6581                         "matrix",
6582
6583                         "%mat4x4_f32          = OpTypeMatrix %v4f32 4\n"
6584                         "%v4f32_1_0_0_0       = OpConstantComposite %v4f32 %c_f32_1 %c_f32_0 %c_f32_0 %c_f32_0\n"
6585                         "%v4f32_0_1_0_0       = OpConstantComposite %v4f32 %c_f32_0 %c_f32_1 %c_f32_0 %c_f32_0\n"
6586                         "%v4f32_0_0_1_0       = OpConstantComposite %v4f32 %c_f32_0 %c_f32_0 %c_f32_1 %c_f32_0\n"
6587                         "%v4f32_0_5_0_5_0_5_1 = OpConstantComposite %v4f32 %c_f32_0_5 %c_f32_0_5 %c_f32_0_5 %c_f32_1\n"
6588                         "%cval                = OpConstantComposite %mat4x4_f32 %v4f32_1_0_0_0 %v4f32_0_1_0_0 %v4f32_0_0_1_0 %v4f32_0_5_0_5_0_5_1\n",
6589
6590                         "%transformed_param   = OpMatrixTimesVector %v4f32 %cval %param1\n"
6591                 },
6592                 {
6593                         "array",
6594
6595                         "%c_v4f32_1_1_1_0     = OpConstantComposite %v4f32 %c_f32_1 %c_f32_1 %c_f32_1 %c_f32_0\n"
6596                         "%fp_a4f32            = OpTypePointer Function %a4f32\n"
6597                         "%f32_n_1             = OpConstant %f32 -1.0\n"
6598                         "%f32_1_5             = OpConstant %f32 !0x3fc00000\n" // +1.5
6599                         "%carr                = OpConstantComposite %a4f32 %c_f32_0 %f32_n_1 %f32_1_5 %c_f32_0\n",
6600
6601                         "%v                   = OpVariable %fp_a4f32 Function %carr\n"
6602                         "%f                   = OpAccessChain %fp_f32 %v %c_u32_0\n"
6603                         "%f1                  = OpAccessChain %fp_f32 %v %c_u32_1\n"
6604                         "%f2                  = OpAccessChain %fp_f32 %v %c_u32_2\n"
6605                         "%f3                  = OpAccessChain %fp_f32 %v %c_u32_3\n"
6606                         "%f_val               = OpLoad %f32 %f\n"
6607                         "%f1_val              = OpLoad %f32 %f1\n"
6608                         "%f2_val              = OpLoad %f32 %f2\n"
6609                         "%f3_val              = OpLoad %f32 %f3\n"
6610                         "%ftot1               = OpFAdd %f32 %f_val %f1_val\n"
6611                         "%ftot2               = OpFAdd %f32 %ftot1 %f2_val\n"
6612                         "%ftot3               = OpFAdd %f32 %ftot2 %f3_val\n"  // 0 - 1 + 1.5 + 0
6613                         "%add_vec             = OpVectorTimesScalar %v4f32 %c_v4f32_1_1_1_0 %ftot3\n"
6614                         "%transformed_param   = OpFAdd %v4f32 %param1 %add_vec\n"
6615                 },
6616                 {
6617                         //
6618                         // [
6619                         //   {
6620                         //      0.0,
6621                         //      [ 1.0, 1.0, 1.0, 1.0]
6622                         //   },
6623                         //   {
6624                         //      1.0,
6625                         //      [ 0.0, 0.5, 0.0, 0.0]
6626                         //   }, //     ^^^
6627                         //   {
6628                         //      0.0,
6629                         //      [ 1.0, 1.0, 1.0, 1.0]
6630                         //   }
6631                         // ]
6632                         "array_of_struct_of_array",
6633
6634                         "%c_v4f32_1_1_1_0     = OpConstantComposite %v4f32 %c_f32_1 %c_f32_1 %c_f32_1 %c_f32_0\n"
6635                         "%fp_a4f32            = OpTypePointer Function %a4f32\n"
6636                         "%stype               = OpTypeStruct %f32 %a4f32\n"
6637                         "%a3stype             = OpTypeArray %stype %c_u32_3\n"
6638                         "%fp_a3stype          = OpTypePointer Function %a3stype\n"
6639                         "%ca4f32_0            = OpConstantComposite %a4f32 %c_f32_0 %c_f32_0_5 %c_f32_0 %c_f32_0\n"
6640                         "%ca4f32_1            = OpConstantComposite %a4f32 %c_f32_1 %c_f32_1 %c_f32_1 %c_f32_1\n"
6641                         "%cstype1             = OpConstantComposite %stype %c_f32_0 %ca4f32_1\n"
6642                         "%cstype2             = OpConstantComposite %stype %c_f32_1 %ca4f32_0\n"
6643                         "%carr                = OpConstantComposite %a3stype %cstype1 %cstype2 %cstype1",
6644
6645                         "%v                   = OpVariable %fp_a3stype Function %carr\n"
6646                         "%f                   = OpAccessChain %fp_f32 %v %c_u32_1 %c_u32_1 %c_u32_1\n"
6647                         "%f_l                 = OpLoad %f32 %f\n"
6648                         "%add_vec             = OpVectorTimesScalar %v4f32 %c_v4f32_1_1_1_0 %f_l\n"
6649                         "%transformed_param   = OpFAdd %v4f32 %param1 %add_vec\n"
6650                 }
6651         };
6652
6653         getHalfColorsFullAlpha(inputColors);
6654         outputColors[0] = RGBA(255, 255, 255, 255);
6655         outputColors[1] = RGBA(255, 127, 127, 255);
6656         outputColors[2] = RGBA(127, 255, 127, 255);
6657         outputColors[3] = RGBA(127, 127, 255, 255);
6658
6659         for (size_t testNdx = 0; testNdx < sizeof(tests) / sizeof(NameConstantsCode); ++testNdx)
6660         {
6661                 map<string, string> fragments;
6662                 fragments["pre_main"] = tests[testNdx].constants;
6663                 fragments["testfun"] = string(functionStart) + tests[testNdx].code + functionEnd;
6664                 createTestsForAllStages(tests[testNdx].name, inputColors, outputColors, fragments, opConstantCompositeTests.get());
6665         }
6666         return opConstantCompositeTests.release();
6667 }
6668
6669 tcu::TestCaseGroup* createSelectionBlockOrderTests(tcu::TestContext& testCtx)
6670 {
6671         de::MovePtr<tcu::TestCaseGroup> group                           (new tcu::TestCaseGroup(testCtx, "selection_block_order", "Out-of-order blocks for selection"));
6672         RGBA                                                    inputColors[4];
6673         RGBA                                                    outputColors[4];
6674         map<string, string>                             fragments;
6675
6676         // vec4 test_code(vec4 param) {
6677         //   vec4 result = param;
6678         //   for (int i = 0; i < 4; ++i) {
6679         //     if (i == 0) result[i] = 0.;
6680         //     else        result[i] = 1. - result[i];
6681         //   }
6682         //   return result;
6683         // }
6684         const char                                              function[]                      =
6685                 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
6686                 "%param1    = OpFunctionParameter %v4f32\n"
6687                 "%lbl       = OpLabel\n"
6688                 "%iptr      = OpVariable %fp_i32 Function\n"
6689                 "%result    = OpVariable %fp_v4f32 Function\n"
6690                 "             OpStore %iptr %c_i32_0\n"
6691                 "             OpStore %result %param1\n"
6692                 "             OpBranch %loop\n"
6693
6694                 // Loop entry block.
6695                 "%loop      = OpLabel\n"
6696                 "%ival      = OpLoad %i32 %iptr\n"
6697                 "%lt_4      = OpSLessThan %bool %ival %c_i32_4\n"
6698                 "             OpLoopMerge %exit %if_entry None\n"
6699                 "             OpBranchConditional %lt_4 %if_entry %exit\n"
6700
6701                 // Merge block for loop.
6702                 "%exit      = OpLabel\n"
6703                 "%ret       = OpLoad %v4f32 %result\n"
6704                 "             OpReturnValue %ret\n"
6705
6706                 // If-statement entry block.
6707                 "%if_entry  = OpLabel\n"
6708                 "%loc       = OpAccessChain %fp_f32 %result %ival\n"
6709                 "%eq_0      = OpIEqual %bool %ival %c_i32_0\n"
6710                 "             OpSelectionMerge %if_exit None\n"
6711                 "             OpBranchConditional %eq_0 %if_true %if_false\n"
6712
6713                 // False branch for if-statement.
6714                 "%if_false  = OpLabel\n"
6715                 "%val       = OpLoad %f32 %loc\n"
6716                 "%sub       = OpFSub %f32 %c_f32_1 %val\n"
6717                 "             OpStore %loc %sub\n"
6718                 "             OpBranch %if_exit\n"
6719
6720                 // Merge block for if-statement.
6721                 "%if_exit   = OpLabel\n"
6722                 "%ival_next = OpIAdd %i32 %ival %c_i32_1\n"
6723                 "             OpStore %iptr %ival_next\n"
6724                 "             OpBranch %loop\n"
6725
6726                 // True branch for if-statement.
6727                 "%if_true   = OpLabel\n"
6728                 "             OpStore %loc %c_f32_0\n"
6729                 "             OpBranch %if_exit\n"
6730
6731                 "             OpFunctionEnd\n";
6732
6733         fragments["testfun"]    = function;
6734
6735         inputColors[0]                  = RGBA(127, 127, 127, 0);
6736         inputColors[1]                  = RGBA(127, 0,   0,   0);
6737         inputColors[2]                  = RGBA(0,   127, 0,   0);
6738         inputColors[3]                  = RGBA(0,   0,   127, 0);
6739
6740         outputColors[0]                 = RGBA(0, 128, 128, 255);
6741         outputColors[1]                 = RGBA(0, 255, 255, 255);
6742         outputColors[2]                 = RGBA(0, 128, 255, 255);
6743         outputColors[3]                 = RGBA(0, 255, 128, 255);
6744
6745         createTestsForAllStages("out_of_order", inputColors, outputColors, fragments, group.get());
6746
6747         return group.release();
6748 }
6749
6750 tcu::TestCaseGroup* createSwitchBlockOrderTests(tcu::TestContext& testCtx)
6751 {
6752         de::MovePtr<tcu::TestCaseGroup> group                           (new tcu::TestCaseGroup(testCtx, "switch_block_order", "Out-of-order blocks for switch"));
6753         RGBA                                                    inputColors[4];
6754         RGBA                                                    outputColors[4];
6755         map<string, string>                             fragments;
6756
6757         const char                                              typesAndConstants[]     =
6758                 "%c_f32_p2  = OpConstant %f32 0.2\n"
6759                 "%c_f32_p4  = OpConstant %f32 0.4\n"
6760                 "%c_f32_p6  = OpConstant %f32 0.6\n"
6761                 "%c_f32_p8  = OpConstant %f32 0.8\n";
6762
6763         // vec4 test_code(vec4 param) {
6764         //   vec4 result = param;
6765         //   for (int i = 0; i < 4; ++i) {
6766         //     switch (i) {
6767         //       case 0: result[i] += .2; break;
6768         //       case 1: result[i] += .6; break;
6769         //       case 2: result[i] += .4; break;
6770         //       case 3: result[i] += .8; break;
6771         //       default: break; // unreachable
6772         //     }
6773         //   }
6774         //   return result;
6775         // }
6776         const char                                              function[]                      =
6777                 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
6778                 "%param1    = OpFunctionParameter %v4f32\n"
6779                 "%lbl       = OpLabel\n"
6780                 "%iptr      = OpVariable %fp_i32 Function\n"
6781                 "%result    = OpVariable %fp_v4f32 Function\n"
6782                 "             OpStore %iptr %c_i32_0\n"
6783                 "             OpStore %result %param1\n"
6784                 "             OpBranch %loop\n"
6785
6786                 // Loop entry block.
6787                 "%loop      = OpLabel\n"
6788                 "%ival      = OpLoad %i32 %iptr\n"
6789                 "%lt_4      = OpSLessThan %bool %ival %c_i32_4\n"
6790                 "             OpLoopMerge %exit %switch_exit None\n"
6791                 "             OpBranchConditional %lt_4 %switch_entry %exit\n"
6792
6793                 // Merge block for loop.
6794                 "%exit      = OpLabel\n"
6795                 "%ret       = OpLoad %v4f32 %result\n"
6796                 "             OpReturnValue %ret\n"
6797
6798                 // Switch-statement entry block.
6799                 "%switch_entry   = OpLabel\n"
6800                 "%loc            = OpAccessChain %fp_f32 %result %ival\n"
6801                 "%val            = OpLoad %f32 %loc\n"
6802                 "                  OpSelectionMerge %switch_exit None\n"
6803                 "                  OpSwitch %ival %switch_default 0 %case0 1 %case1 2 %case2 3 %case3\n"
6804
6805                 "%case2          = OpLabel\n"
6806                 "%addp4          = OpFAdd %f32 %val %c_f32_p4\n"
6807                 "                  OpStore %loc %addp4\n"
6808                 "                  OpBranch %switch_exit\n"
6809
6810                 "%switch_default = OpLabel\n"
6811                 "                  OpUnreachable\n"
6812
6813                 "%case3          = OpLabel\n"
6814                 "%addp8          = OpFAdd %f32 %val %c_f32_p8\n"
6815                 "                  OpStore %loc %addp8\n"
6816                 "                  OpBranch %switch_exit\n"
6817
6818                 "%case0          = OpLabel\n"
6819                 "%addp2          = OpFAdd %f32 %val %c_f32_p2\n"
6820                 "                  OpStore %loc %addp2\n"
6821                 "                  OpBranch %switch_exit\n"
6822
6823                 // Merge block for switch-statement.
6824                 "%switch_exit    = OpLabel\n"
6825                 "%ival_next      = OpIAdd %i32 %ival %c_i32_1\n"
6826                 "                  OpStore %iptr %ival_next\n"
6827                 "                  OpBranch %loop\n"
6828
6829                 "%case1          = OpLabel\n"
6830                 "%addp6          = OpFAdd %f32 %val %c_f32_p6\n"
6831                 "                  OpStore %loc %addp6\n"
6832                 "                  OpBranch %switch_exit\n"
6833
6834                 "                  OpFunctionEnd\n";
6835
6836         fragments["pre_main"]   = typesAndConstants;
6837         fragments["testfun"]    = function;
6838
6839         inputColors[0]                  = RGBA(127, 27,  127, 51);
6840         inputColors[1]                  = RGBA(127, 0,   0,   51);
6841         inputColors[2]                  = RGBA(0,   27,  0,   51);
6842         inputColors[3]                  = RGBA(0,   0,   127, 51);
6843
6844         outputColors[0]                 = RGBA(178, 180, 229, 255);
6845         outputColors[1]                 = RGBA(178, 153, 102, 255);
6846         outputColors[2]                 = RGBA(51,  180, 102, 255);
6847         outputColors[3]                 = RGBA(51,  153, 229, 255);
6848
6849         createTestsForAllStages("out_of_order", inputColors, outputColors, fragments, group.get());
6850
6851         return group.release();
6852 }
6853
6854 tcu::TestCaseGroup* createDecorationGroupTests(tcu::TestContext& testCtx)
6855 {
6856         de::MovePtr<tcu::TestCaseGroup> group                           (new tcu::TestCaseGroup(testCtx, "decoration_group", "Decoration group tests"));
6857         RGBA                                                    inputColors[4];
6858         RGBA                                                    outputColors[4];
6859         map<string, string>                             fragments;
6860
6861         const char                                              decorations[]           =
6862                 "OpDecorate %array_group         ArrayStride 4\n"
6863                 "OpDecorate %struct_member_group Offset 0\n"
6864                 "%array_group         = OpDecorationGroup\n"
6865                 "%struct_member_group = OpDecorationGroup\n"
6866
6867                 "OpDecorate %group1 RelaxedPrecision\n"
6868                 "OpDecorate %group3 RelaxedPrecision\n"
6869                 "OpDecorate %group3 Invariant\n"
6870                 "OpDecorate %group3 Restrict\n"
6871                 "%group0 = OpDecorationGroup\n"
6872                 "%group1 = OpDecorationGroup\n"
6873                 "%group3 = OpDecorationGroup\n";
6874
6875         const char                                              typesAndConstants[]     =
6876                 "%a3f32     = OpTypeArray %f32 %c_u32_3\n"
6877                 "%struct1   = OpTypeStruct %a3f32\n"
6878                 "%struct2   = OpTypeStruct %a3f32\n"
6879                 "%fp_struct1 = OpTypePointer Function %struct1\n"
6880                 "%fp_struct2 = OpTypePointer Function %struct2\n"
6881                 "%c_f32_2    = OpConstant %f32 2.\n"
6882                 "%c_f32_n2   = OpConstant %f32 -2.\n"
6883
6884                 "%c_a3f32_1 = OpConstantComposite %a3f32 %c_f32_1 %c_f32_2 %c_f32_1\n"
6885                 "%c_a3f32_2 = OpConstantComposite %a3f32 %c_f32_n1 %c_f32_n2 %c_f32_n1\n"
6886                 "%c_struct1 = OpConstantComposite %struct1 %c_a3f32_1\n"
6887                 "%c_struct2 = OpConstantComposite %struct2 %c_a3f32_2\n";
6888
6889         const char                                              function[]                      =
6890                 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
6891                 "%param     = OpFunctionParameter %v4f32\n"
6892                 "%entry     = OpLabel\n"
6893                 "%result    = OpVariable %fp_v4f32 Function\n"
6894                 "%v_struct1 = OpVariable %fp_struct1 Function\n"
6895                 "%v_struct2 = OpVariable %fp_struct2 Function\n"
6896                 "             OpStore %result %param\n"
6897                 "             OpStore %v_struct1 %c_struct1\n"
6898                 "             OpStore %v_struct2 %c_struct2\n"
6899                 "%ptr1      = OpAccessChain %fp_f32 %v_struct1 %c_i32_0 %c_i32_2\n"
6900                 "%val1      = OpLoad %f32 %ptr1\n"
6901                 "%ptr2      = OpAccessChain %fp_f32 %v_struct2 %c_i32_0 %c_i32_2\n"
6902                 "%val2      = OpLoad %f32 %ptr2\n"
6903                 "%addvalues = OpFAdd %f32 %val1 %val2\n"
6904                 "%ptr       = OpAccessChain %fp_f32 %result %c_i32_1\n"
6905                 "%val       = OpLoad %f32 %ptr\n"
6906                 "%addresult = OpFAdd %f32 %addvalues %val\n"
6907                 "             OpStore %ptr %addresult\n"
6908                 "%ret       = OpLoad %v4f32 %result\n"
6909                 "             OpReturnValue %ret\n"
6910                 "             OpFunctionEnd\n";
6911
6912         struct CaseNameDecoration
6913         {
6914                 string name;
6915                 string decoration;
6916         };
6917
6918         CaseNameDecoration tests[] =
6919         {
6920                 {
6921                         "same_decoration_group_on_multiple_types",
6922                         "OpGroupMemberDecorate %struct_member_group %struct1 0 %struct2 0\n"
6923                 },
6924                 {
6925                         "empty_decoration_group",
6926                         "OpGroupDecorate %group0      %a3f32\n"
6927                         "OpGroupDecorate %group0      %result\n"
6928                 },
6929                 {
6930                         "one_element_decoration_group",
6931                         "OpGroupDecorate %array_group %a3f32\n"
6932                 },
6933                 {
6934                         "multiple_elements_decoration_group",
6935                         "OpGroupDecorate %group3      %v_struct1\n"
6936                 },
6937                 {
6938                         "multiple_decoration_groups_on_same_variable",
6939                         "OpGroupDecorate %group0      %v_struct2\n"
6940                         "OpGroupDecorate %group1      %v_struct2\n"
6941                         "OpGroupDecorate %group3      %v_struct2\n"
6942                 },
6943                 {
6944                         "same_decoration_group_multiple_times",
6945                         "OpGroupDecorate %group1      %addvalues\n"
6946                         "OpGroupDecorate %group1      %addvalues\n"
6947                         "OpGroupDecorate %group1      %addvalues\n"
6948                 },
6949
6950         };
6951
6952         getHalfColorsFullAlpha(inputColors);
6953         getHalfColorsFullAlpha(outputColors);
6954
6955         for (size_t idx = 0; idx < (sizeof(tests) / sizeof(tests[0])); ++idx)
6956         {
6957                 fragments["decoration"] = decorations + tests[idx].decoration;
6958                 fragments["pre_main"]   = typesAndConstants;
6959                 fragments["testfun"]    = function;
6960
6961                 createTestsForAllStages(tests[idx].name, inputColors, outputColors, fragments, group.get());
6962         }
6963
6964         return group.release();
6965 }
6966
6967 struct SpecConstantTwoIntGraphicsCase
6968 {
6969         const char*             caseName;
6970         const char*             scDefinition0;
6971         const char*             scDefinition1;
6972         const char*             scResultType;
6973         const char*             scOperation;
6974         deInt32                 scActualValue0;
6975         deInt32                 scActualValue1;
6976         const char*             resultOperation;
6977         RGBA                    expectedColors[4];
6978         deInt32                 scActualValueLength;
6979
6980                                         SpecConstantTwoIntGraphicsCase (const char*             name,
6981                                                                                                         const char*             definition0,
6982                                                                                                         const char*             definition1,
6983                                                                                                         const char*             resultType,
6984                                                                                                         const char*             operation,
6985                                                                                                         const deInt32   value0,
6986                                                                                                         const deInt32   value1,
6987                                                                                                         const char*             resultOp,
6988                                                                                                         const RGBA              (&output)[4],
6989                                                                                                         const deInt32   valueLength = sizeof(deInt32))
6990                                                 : caseName                              (name)
6991                                                 , scDefinition0                 (definition0)
6992                                                 , scDefinition1                 (definition1)
6993                                                 , scResultType                  (resultType)
6994                                                 , scOperation                   (operation)
6995                                                 , scActualValue0                (value0)
6996                                                 , scActualValue1                (value1)
6997                                                 , resultOperation               (resultOp)
6998                                                 , scActualValueLength   (valueLength)
6999         {
7000                 expectedColors[0] = output[0];
7001                 expectedColors[1] = output[1];
7002                 expectedColors[2] = output[2];
7003                 expectedColors[3] = output[3];
7004         }
7005 };
7006
7007 tcu::TestCaseGroup* createSpecConstantTests (tcu::TestContext& testCtx)
7008 {
7009         de::MovePtr<tcu::TestCaseGroup> group                           (new tcu::TestCaseGroup(testCtx, "opspecconstantop", "Test the OpSpecConstantOp instruction"));
7010         vector<SpecConstantTwoIntGraphicsCase>  cases;
7011         RGBA                                                    inputColors[4];
7012         RGBA                                                    outputColors0[4];
7013         RGBA                                                    outputColors1[4];
7014         RGBA                                                    outputColors2[4];
7015
7016         const deInt32                                   m1AsFloat16                     = 0xbc00; // -1(fp16) == 1 01111 0000000000 == 1011 1100 0000 0000
7017
7018         const char      decorations1[]                  =
7019                 "OpDecorate %sc_0  SpecId 0\n"
7020                 "OpDecorate %sc_1  SpecId 1\n";
7021
7022         const char      typesAndConstants1[]    =
7023                 "${OPTYPE_DEFINITIONS:opt}"
7024                 "%sc_0      = OpSpecConstant${SC_DEF0}\n"
7025                 "%sc_1      = OpSpecConstant${SC_DEF1}\n"
7026                 "%sc_op     = OpSpecConstantOp ${SC_RESULT_TYPE} ${SC_OP}\n";
7027
7028         const char      function1[]                             =
7029                 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
7030                 "%param     = OpFunctionParameter %v4f32\n"
7031                 "%label     = OpLabel\n"
7032                 "%result    = OpVariable %fp_v4f32 Function\n"
7033                 "${TYPE_CONVERT:opt}"
7034                 "             OpStore %result %param\n"
7035                 "%gen       = ${GEN_RESULT}\n"
7036                 "%index     = OpIAdd %i32 %gen %c_i32_1\n"
7037                 "%loc       = OpAccessChain %fp_f32 %result %index\n"
7038                 "%val       = OpLoad %f32 %loc\n"
7039                 "%add       = OpFAdd %f32 %val %c_f32_0_5\n"
7040                 "             OpStore %loc %add\n"
7041                 "%ret       = OpLoad %v4f32 %result\n"
7042                 "             OpReturnValue %ret\n"
7043                 "             OpFunctionEnd\n";
7044
7045         inputColors[0] = RGBA(127, 127, 127, 255);
7046         inputColors[1] = RGBA(127, 0,   0,   255);
7047         inputColors[2] = RGBA(0,   127, 0,   255);
7048         inputColors[3] = RGBA(0,   0,   127, 255);
7049
7050         // Derived from inputColors[x] by adding 128 to inputColors[x][0].
7051         outputColors0[0] = RGBA(255, 127, 127, 255);
7052         outputColors0[1] = RGBA(255, 0,   0,   255);
7053         outputColors0[2] = RGBA(128, 127, 0,   255);
7054         outputColors0[3] = RGBA(128, 0,   127, 255);
7055
7056         // Derived from inputColors[x] by adding 128 to inputColors[x][1].
7057         outputColors1[0] = RGBA(127, 255, 127, 255);
7058         outputColors1[1] = RGBA(127, 128, 0,   255);
7059         outputColors1[2] = RGBA(0,   255, 0,   255);
7060         outputColors1[3] = RGBA(0,   128, 127, 255);
7061
7062         // Derived from inputColors[x] by adding 128 to inputColors[x][2].
7063         outputColors2[0] = RGBA(127, 127, 255, 255);
7064         outputColors2[1] = RGBA(127, 0,   128, 255);
7065         outputColors2[2] = RGBA(0,   127, 128, 255);
7066         outputColors2[3] = RGBA(0,   0,   255, 255);
7067
7068         const char addZeroToSc[]                = "OpIAdd %i32 %c_i32_0 %sc_op";
7069         const char addZeroToSc32[]              = "OpIAdd %i32 %c_i32_0 %sc_op32";
7070         const char selectTrueUsingSc[]  = "OpSelect %i32 %sc_op %c_i32_1 %c_i32_0";
7071         const char selectFalseUsingSc[] = "OpSelect %i32 %sc_op %c_i32_0 %c_i32_1";
7072
7073         cases.push_back(SpecConstantTwoIntGraphicsCase("iadd",                                  " %i32 0",              " %i32 0",              "%i32",         "IAdd                 %sc_0 %sc_1",                             19,             -20,    addZeroToSc,            outputColors0));
7074         cases.push_back(SpecConstantTwoIntGraphicsCase("isub",                                  " %i32 0",              " %i32 0",              "%i32",         "ISub                 %sc_0 %sc_1",                             19,             20,             addZeroToSc,            outputColors0));
7075         cases.push_back(SpecConstantTwoIntGraphicsCase("imul",                                  " %i32 0",              " %i32 0",              "%i32",         "IMul                 %sc_0 %sc_1",                             -1,             -1,             addZeroToSc,            outputColors2));
7076         cases.push_back(SpecConstantTwoIntGraphicsCase("sdiv",                                  " %i32 0",              " %i32 0",              "%i32",         "SDiv                 %sc_0 %sc_1",                             -126,   126,    addZeroToSc,            outputColors0));
7077         cases.push_back(SpecConstantTwoIntGraphicsCase("udiv",                                  " %i32 0",              " %i32 0",              "%i32",         "UDiv                 %sc_0 %sc_1",                             126,    126,    addZeroToSc,            outputColors2));
7078         cases.push_back(SpecConstantTwoIntGraphicsCase("srem",                                  " %i32 0",              " %i32 0",              "%i32",         "SRem                 %sc_0 %sc_1",                             3,              2,              addZeroToSc,            outputColors2));
7079         cases.push_back(SpecConstantTwoIntGraphicsCase("smod",                                  " %i32 0",              " %i32 0",              "%i32",         "SMod                 %sc_0 %sc_1",                             3,              2,              addZeroToSc,            outputColors2));
7080         cases.push_back(SpecConstantTwoIntGraphicsCase("umod",                                  " %i32 0",              " %i32 0",              "%i32",         "UMod                 %sc_0 %sc_1",                             1001,   500,    addZeroToSc,            outputColors2));
7081         cases.push_back(SpecConstantTwoIntGraphicsCase("bitwiseand",                    " %i32 0",              " %i32 0",              "%i32",         "BitwiseAnd           %sc_0 %sc_1",                             0x33,   0x0d,   addZeroToSc,            outputColors2));
7082         cases.push_back(SpecConstantTwoIntGraphicsCase("bitwiseor",                             " %i32 0",              " %i32 0",              "%i32",         "BitwiseOr            %sc_0 %sc_1",                             0,              1,              addZeroToSc,            outputColors2));
7083         cases.push_back(SpecConstantTwoIntGraphicsCase("bitwisexor",                    " %i32 0",              " %i32 0",              "%i32",         "BitwiseXor           %sc_0 %sc_1",                             0x2e,   0x2f,   addZeroToSc,            outputColors2));
7084         cases.push_back(SpecConstantTwoIntGraphicsCase("shiftrightlogical",             " %i32 0",              " %i32 0",              "%i32",         "ShiftRightLogical    %sc_0 %sc_1",                             2,              1,              addZeroToSc,            outputColors2));
7085         cases.push_back(SpecConstantTwoIntGraphicsCase("shiftrightarithmetic",  " %i32 0",              " %i32 0",              "%i32",         "ShiftRightArithmetic %sc_0 %sc_1",                             -4,             2,              addZeroToSc,            outputColors0));
7086         cases.push_back(SpecConstantTwoIntGraphicsCase("shiftleftlogical",              " %i32 0",              " %i32 0",              "%i32",         "ShiftLeftLogical     %sc_0 %sc_1",                             1,              0,              addZeroToSc,            outputColors2));
7087         cases.push_back(SpecConstantTwoIntGraphicsCase("slessthan",                             " %i32 0",              " %i32 0",              "%bool",        "SLessThan            %sc_0 %sc_1",                             -20,    -10,    selectTrueUsingSc,      outputColors2));
7088         cases.push_back(SpecConstantTwoIntGraphicsCase("ulessthan",                             " %i32 0",              " %i32 0",              "%bool",        "ULessThan            %sc_0 %sc_1",                             10,             20,             selectTrueUsingSc,      outputColors2));
7089         cases.push_back(SpecConstantTwoIntGraphicsCase("sgreaterthan",                  " %i32 0",              " %i32 0",              "%bool",        "SGreaterThan         %sc_0 %sc_1",                             -1000,  50,             selectFalseUsingSc,     outputColors2));
7090         cases.push_back(SpecConstantTwoIntGraphicsCase("ugreaterthan",                  " %i32 0",              " %i32 0",              "%bool",        "UGreaterThan         %sc_0 %sc_1",                             10,             5,              selectTrueUsingSc,      outputColors2));
7091         cases.push_back(SpecConstantTwoIntGraphicsCase("slessthanequal",                " %i32 0",              " %i32 0",              "%bool",        "SLessThanEqual       %sc_0 %sc_1",                             -10,    -10,    selectTrueUsingSc,      outputColors2));
7092         cases.push_back(SpecConstantTwoIntGraphicsCase("ulessthanequal",                " %i32 0",              " %i32 0",              "%bool",        "ULessThanEqual       %sc_0 %sc_1",                             50,             100,    selectTrueUsingSc,      outputColors2));
7093         cases.push_back(SpecConstantTwoIntGraphicsCase("sgreaterthanequal",             " %i32 0",              " %i32 0",              "%bool",        "SGreaterThanEqual    %sc_0 %sc_1",                             -1000,  50,             selectFalseUsingSc,     outputColors2));
7094         cases.push_back(SpecConstantTwoIntGraphicsCase("ugreaterthanequal",             " %i32 0",              " %i32 0",              "%bool",        "UGreaterThanEqual    %sc_0 %sc_1",                             10,             10,             selectTrueUsingSc,      outputColors2));
7095         cases.push_back(SpecConstantTwoIntGraphicsCase("iequal",                                " %i32 0",              " %i32 0",              "%bool",        "IEqual               %sc_0 %sc_1",                             42,             24,             selectFalseUsingSc,     outputColors2));
7096         cases.push_back(SpecConstantTwoIntGraphicsCase("inotequal",                             " %i32 0",              " %i32 0",              "%bool",        "INotEqual            %sc_0 %sc_1",                             42,             24,             selectTrueUsingSc,      outputColors2));
7097         cases.push_back(SpecConstantTwoIntGraphicsCase("logicaland",                    "True %bool",   "True %bool",   "%bool",        "LogicalAnd           %sc_0 %sc_1",                             0,              1,              selectFalseUsingSc,     outputColors2));
7098         cases.push_back(SpecConstantTwoIntGraphicsCase("logicalor",                             "False %bool",  "False %bool",  "%bool",        "LogicalOr            %sc_0 %sc_1",                             1,              0,              selectTrueUsingSc,      outputColors2));
7099         cases.push_back(SpecConstantTwoIntGraphicsCase("logicalequal",                  "True %bool",   "True %bool",   "%bool",        "LogicalEqual         %sc_0 %sc_1",                             0,              1,              selectFalseUsingSc,     outputColors2));
7100         cases.push_back(SpecConstantTwoIntGraphicsCase("logicalnotequal",               "False %bool",  "False %bool",  "%bool",        "LogicalNotEqual      %sc_0 %sc_1",                             1,              0,              selectTrueUsingSc,      outputColors2));
7101         cases.push_back(SpecConstantTwoIntGraphicsCase("snegate",                               " %i32 0",              " %i32 0",              "%i32",         "SNegate              %sc_0",                                   -1,             0,              addZeroToSc,            outputColors2));
7102         cases.push_back(SpecConstantTwoIntGraphicsCase("not",                                   " %i32 0",              " %i32 0",              "%i32",         "Not                  %sc_0",                                   -2,             0,              addZeroToSc,            outputColors2));
7103         cases.push_back(SpecConstantTwoIntGraphicsCase("logicalnot",                    "False %bool",  "False %bool",  "%bool",        "LogicalNot           %sc_0",                                   1,              0,              selectFalseUsingSc,     outputColors2));
7104         cases.push_back(SpecConstantTwoIntGraphicsCase("select",                                "False %bool",  " %i32 0",              "%i32",         "Select               %sc_0 %sc_1 %c_i32_0",    1,              1,              addZeroToSc,            outputColors2));
7105         cases.push_back(SpecConstantTwoIntGraphicsCase("sconvert",                              " %i32 0",              " %i32 0",              "%i16",         "SConvert             %sc_0",                                   -1,             0,              addZeroToSc32,          outputColors0));
7106         // -1082130432 stored as 32-bit two's complement is the binary representation of -1 as IEEE-754 Float
7107         cases.push_back(SpecConstantTwoIntGraphicsCase("fconvert",                              " %f32 0",              " %f32 0",              "%f64",         "FConvert             %sc_0",                                   -1082130432, 0, addZeroToSc32,          outputColors0));
7108         cases.push_back(SpecConstantTwoIntGraphicsCase("fconvert16",                    " %f16 0",              " %f16 0",              "%f32",         "FConvert             %sc_0",                                   m1AsFloat16, 0, addZeroToSc32,          outputColors0, sizeof(deFloat16)));
7109         // \todo[2015-12-1 antiagainst] OpQuantizeToF16
7110
7111         for (size_t caseNdx = 0; caseNdx < cases.size(); ++caseNdx)
7112         {
7113                 map<string, string>                     specializations;
7114                 map<string, string>                     fragments;
7115                 SpecConstants                           specConstants;
7116                 PushConstants                           noPushConstants;
7117                 GraphicsResources                       noResources;
7118                 GraphicsInterfaces                      noInterfaces;
7119                 vector<string>                          extensions;
7120                 VulkanFeatures                          requiredFeatures;
7121
7122                 // Special SPIR-V code for SConvert-case
7123                 if (strcmp(cases[caseNdx].caseName, "sconvert") == 0)
7124                 {
7125                         requiredFeatures.coreFeatures.shaderInt16 = VK_TRUE;
7126                         fragments["capability"]                                 = "OpCapability Int16\n";                                       // Adds 16-bit integer capability
7127                         specializations["OPTYPE_DEFINITIONS"]   = "%i16 = OpTypeInt 16 1\n";                            // Adds 16-bit integer type
7128                         specializations["TYPE_CONVERT"]                 = "%sc_op32 = OpSConvert %i32 %sc_op\n";        // Converts 16-bit integer to 32-bit integer
7129                 }
7130
7131                 // Special SPIR-V code for FConvert-case
7132                 if (strcmp(cases[caseNdx].caseName, "fconvert") == 0)
7133                 {
7134                         requiredFeatures.coreFeatures.shaderFloat64 = VK_TRUE;
7135                         fragments["capability"]                                 = "OpCapability Float64\n";                                     // Adds 64-bit float capability
7136                         specializations["OPTYPE_DEFINITIONS"]   = "%f64 = OpTypeFloat 64\n";                            // Adds 64-bit float type
7137                         specializations["TYPE_CONVERT"]                 = "%sc_op32 = OpConvertFToS %i32 %sc_op\n";     // Converts 64-bit float to 32-bit integer
7138                 }
7139
7140                 // Special SPIR-V code for FConvert-case for 16-bit floats
7141                 if (strcmp(cases[caseNdx].caseName, "fconvert16") == 0)
7142                 {
7143                         extensions.push_back("VK_KHR_shader_float16_int8");
7144                         requiredFeatures.extFloat16Int8 = EXTFLOAT16INT8FEATURES_FLOAT16;
7145                         fragments["capability"]                                 = "OpCapability Float16\n";                                     // Adds 16-bit float capability
7146                         specializations["OPTYPE_DEFINITIONS"]   = "%f16 = OpTypeFloat 16\n";                            // Adds 16-bit float type
7147                         specializations["TYPE_CONVERT"]                 = "%sc_op32 = OpConvertFToS %i32 %sc_op\n";     // Converts 16-bit float to 32-bit integer
7148                 }
7149
7150                 specializations["SC_DEF0"]                      = cases[caseNdx].scDefinition0;
7151                 specializations["SC_DEF1"]                      = cases[caseNdx].scDefinition1;
7152                 specializations["SC_RESULT_TYPE"]       = cases[caseNdx].scResultType;
7153                 specializations["SC_OP"]                        = cases[caseNdx].scOperation;
7154                 specializations["GEN_RESULT"]           = cases[caseNdx].resultOperation;
7155
7156                 fragments["decoration"]                         = tcu::StringTemplate(decorations1).specialize(specializations);
7157                 fragments["pre_main"]                           = tcu::StringTemplate(typesAndConstants1).specialize(specializations);
7158                 fragments["testfun"]                            = tcu::StringTemplate(function1).specialize(specializations);
7159
7160                 specConstants.append(&cases[caseNdx].scActualValue0, cases[caseNdx].scActualValueLength);
7161                 specConstants.append(&cases[caseNdx].scActualValue1, cases[caseNdx].scActualValueLength);
7162
7163                 createTestsForAllStages(
7164                         cases[caseNdx].caseName, inputColors, cases[caseNdx].expectedColors, fragments, specConstants,
7165                         noPushConstants, noResources, noInterfaces, extensions, requiredFeatures, group.get());
7166         }
7167
7168         const char      decorations2[]                  =
7169                 "OpDecorate %sc_0  SpecId 0\n"
7170                 "OpDecorate %sc_1  SpecId 1\n"
7171                 "OpDecorate %sc_2  SpecId 2\n";
7172
7173         const char      typesAndConstants2[]    =
7174                 "%vec3_0      = OpConstantComposite %v3i32 %c_i32_0 %c_i32_0 %c_i32_0\n"
7175                 "%vec3_undef  = OpUndef %v3i32\n"
7176
7177                 "%sc_0        = OpSpecConstant %i32 0\n"
7178                 "%sc_1        = OpSpecConstant %i32 0\n"
7179                 "%sc_2        = OpSpecConstant %i32 0\n"
7180                 "%sc_vec3_0   = OpSpecConstantOp %v3i32 CompositeInsert  %sc_0        %vec3_0      0\n"                                                 // (sc_0, 0,    0)
7181                 "%sc_vec3_1   = OpSpecConstantOp %v3i32 CompositeInsert  %sc_1        %vec3_0      1\n"                                                 // (0,    sc_1, 0)
7182                 "%sc_vec3_2   = OpSpecConstantOp %v3i32 CompositeInsert  %sc_2        %vec3_0      2\n"                                                 // (0,    0,    sc_2)
7183                 "%sc_vec3_0_s = OpSpecConstantOp %v3i32 VectorShuffle    %sc_vec3_0   %vec3_undef  0          0xFFFFFFFF 2\n"   // (sc_0, ???,  0)
7184                 "%sc_vec3_1_s = OpSpecConstantOp %v3i32 VectorShuffle    %sc_vec3_1   %vec3_undef  0xFFFFFFFF 1          0\n"   // (???,  sc_1, 0)
7185                 "%sc_vec3_2_s = OpSpecConstantOp %v3i32 VectorShuffle    %vec3_undef  %sc_vec3_2   5          0xFFFFFFFF 5\n"   // (sc_2, ???,  sc_2)
7186                 "%sc_vec3_01  = OpSpecConstantOp %v3i32 VectorShuffle    %sc_vec3_0_s %sc_vec3_1_s 1 0 4\n"                                             // (0,    sc_0, sc_1)
7187                 "%sc_vec3_012 = OpSpecConstantOp %v3i32 VectorShuffle    %sc_vec3_01  %sc_vec3_2_s 5 1 2\n"                                             // (sc_2, sc_0, sc_1)
7188                 "%sc_ext_0    = OpSpecConstantOp %i32   CompositeExtract %sc_vec3_012              0\n"                                                 // sc_2
7189                 "%sc_ext_1    = OpSpecConstantOp %i32   CompositeExtract %sc_vec3_012              1\n"                                                 // sc_0
7190                 "%sc_ext_2    = OpSpecConstantOp %i32   CompositeExtract %sc_vec3_012              2\n"                                                 // sc_1
7191                 "%sc_sub      = OpSpecConstantOp %i32   ISub             %sc_ext_0    %sc_ext_1\n"                                                              // (sc_2 - sc_0)
7192                 "%sc_final    = OpSpecConstantOp %i32   IMul             %sc_sub      %sc_ext_2\n";                                                             // (sc_2 - sc_0) * sc_1
7193
7194         const char      function2[]                             =
7195                 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
7196                 "%param     = OpFunctionParameter %v4f32\n"
7197                 "%label     = OpLabel\n"
7198                 "%result    = OpVariable %fp_v4f32 Function\n"
7199                 "             OpStore %result %param\n"
7200                 "%loc       = OpAccessChain %fp_f32 %result %sc_final\n"
7201                 "%val       = OpLoad %f32 %loc\n"
7202                 "%add       = OpFAdd %f32 %val %c_f32_0_5\n"
7203                 "             OpStore %loc %add\n"
7204                 "%ret       = OpLoad %v4f32 %result\n"
7205                 "             OpReturnValue %ret\n"
7206                 "             OpFunctionEnd\n";
7207
7208         map<string, string>     fragments;
7209         SpecConstants           specConstants;
7210
7211         fragments["decoration"] = decorations2;
7212         fragments["pre_main"]   = typesAndConstants2;
7213         fragments["testfun"]    = function2;
7214
7215         specConstants.append<deInt32>(56789);
7216         specConstants.append<deInt32>(-2);
7217         specConstants.append<deInt32>(56788);
7218
7219         createTestsForAllStages("vector_related", inputColors, outputColors2, fragments, specConstants, group.get());
7220
7221         return group.release();
7222 }
7223
7224 tcu::TestCaseGroup* createOpPhiTests(tcu::TestContext& testCtx)
7225 {
7226         de::MovePtr<tcu::TestCaseGroup> group                           (new tcu::TestCaseGroup(testCtx, "opphi", "Test the OpPhi instruction"));
7227         RGBA                                                    inputColors[4];
7228         RGBA                                                    outputColors1[4];
7229         RGBA                                                    outputColors2[4];
7230         RGBA                                                    outputColors3[4];
7231         RGBA                                                    outputColors4[4];
7232         map<string, string>                             fragments1;
7233         map<string, string>                             fragments2;
7234         map<string, string>                             fragments3;
7235         map<string, string>                             fragments4;
7236         std::vector<std::string>                extensions4;
7237         GraphicsResources                               resources4;
7238         VulkanFeatures                                  vulkanFeatures4;
7239
7240         const char      typesAndConstants1[]    =
7241                 "%c_f32_p2  = OpConstant %f32 0.2\n"
7242                 "%c_f32_p4  = OpConstant %f32 0.4\n"
7243                 "%c_f32_p5  = OpConstant %f32 0.5\n"
7244                 "%c_f32_p8  = OpConstant %f32 0.8\n";
7245
7246         // vec4 test_code(vec4 param) {
7247         //   vec4 result = param;
7248         //   for (int i = 0; i < 4; ++i) {
7249         //     float operand;
7250         //     switch (i) {
7251         //       case 0: operand = .2; break;
7252         //       case 1: operand = .5; break;
7253         //       case 2: operand = .4; break;
7254         //       case 3: operand = .0; break;
7255         //       default: break; // unreachable
7256         //     }
7257         //     result[i] += operand;
7258         //   }
7259         //   return result;
7260         // }
7261         const char      function1[]                             =
7262                 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
7263                 "%param1    = OpFunctionParameter %v4f32\n"
7264                 "%lbl       = OpLabel\n"
7265                 "%iptr      = OpVariable %fp_i32 Function\n"
7266                 "%result    = OpVariable %fp_v4f32 Function\n"
7267                 "             OpStore %iptr %c_i32_0\n"
7268                 "             OpStore %result %param1\n"
7269                 "             OpBranch %loop\n"
7270
7271                 "%loop      = OpLabel\n"
7272                 "%ival      = OpLoad %i32 %iptr\n"
7273                 "%lt_4      = OpSLessThan %bool %ival %c_i32_4\n"
7274                 "             OpLoopMerge %exit %phi None\n"
7275                 "             OpBranchConditional %lt_4 %entry %exit\n"
7276
7277                 "%entry     = OpLabel\n"
7278                 "%loc       = OpAccessChain %fp_f32 %result %ival\n"
7279                 "%val       = OpLoad %f32 %loc\n"
7280                 "             OpSelectionMerge %phi None\n"
7281                 "             OpSwitch %ival %default 0 %case0 1 %case1 2 %case2 3 %case3\n"
7282
7283                 "%case0     = OpLabel\n"
7284                 "             OpBranch %phi\n"
7285                 "%case1     = OpLabel\n"
7286                 "             OpBranch %phi\n"
7287                 "%case2     = OpLabel\n"
7288                 "             OpBranch %phi\n"
7289                 "%case3     = OpLabel\n"
7290                 "             OpBranch %phi\n"
7291
7292                 "%default   = OpLabel\n"
7293                 "             OpUnreachable\n"
7294
7295                 "%phi       = OpLabel\n"
7296                 "%operand   = OpPhi %f32 %c_f32_p4 %case2 %c_f32_p5 %case1 %c_f32_p2 %case0 %c_f32_0 %case3\n" // not in the order of blocks
7297                 "%add       = OpFAdd %f32 %val %operand\n"
7298                 "             OpStore %loc %add\n"
7299                 "%ival_next = OpIAdd %i32 %ival %c_i32_1\n"
7300                 "             OpStore %iptr %ival_next\n"
7301                 "             OpBranch %loop\n"
7302
7303                 "%exit      = OpLabel\n"
7304                 "%ret       = OpLoad %v4f32 %result\n"
7305                 "             OpReturnValue %ret\n"
7306
7307                 "             OpFunctionEnd\n";
7308
7309         fragments1["pre_main"]  = typesAndConstants1;
7310         fragments1["testfun"]   = function1;
7311
7312         getHalfColorsFullAlpha(inputColors);
7313
7314         outputColors1[0]                = RGBA(178, 255, 229, 255);
7315         outputColors1[1]                = RGBA(178, 127, 102, 255);
7316         outputColors1[2]                = RGBA(51,  255, 102, 255);
7317         outputColors1[3]                = RGBA(51,  127, 229, 255);
7318
7319         createTestsForAllStages("out_of_order", inputColors, outputColors1, fragments1, group.get());
7320
7321         const char      typesAndConstants2[]    =
7322                 "%c_f32_p2  = OpConstant %f32 0.2\n";
7323
7324         // Add .4 to the second element of the given parameter.
7325         const char      function2[]                             =
7326                 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
7327                 "%param     = OpFunctionParameter %v4f32\n"
7328                 "%entry     = OpLabel\n"
7329                 "%result    = OpVariable %fp_v4f32 Function\n"
7330                 "             OpStore %result %param\n"
7331                 "%loc       = OpAccessChain %fp_f32 %result %c_i32_1\n"
7332                 "%val       = OpLoad %f32 %loc\n"
7333                 "             OpBranch %phi\n"
7334
7335                 "%phi        = OpLabel\n"
7336                 "%step       = OpPhi %i32 %c_i32_0  %entry %step_next  %phi\n"
7337                 "%accum      = OpPhi %f32 %val      %entry %accum_next %phi\n"
7338                 "%step_next  = OpIAdd %i32 %step  %c_i32_1\n"
7339                 "%accum_next = OpFAdd %f32 %accum %c_f32_p2\n"
7340                 "%still_loop = OpSLessThan %bool %step %c_i32_2\n"
7341                 "              OpLoopMerge %exit %phi None\n"
7342                 "              OpBranchConditional %still_loop %phi %exit\n"
7343
7344                 "%exit       = OpLabel\n"
7345                 "              OpStore %loc %accum\n"
7346                 "%ret        = OpLoad %v4f32 %result\n"
7347                 "              OpReturnValue %ret\n"
7348
7349                 "              OpFunctionEnd\n";
7350
7351         fragments2["pre_main"]  = typesAndConstants2;
7352         fragments2["testfun"]   = function2;
7353
7354         outputColors2[0]                        = RGBA(127, 229, 127, 255);
7355         outputColors2[1]                        = RGBA(127, 102, 0,   255);
7356         outputColors2[2]                        = RGBA(0,   229, 0,   255);
7357         outputColors2[3]                        = RGBA(0,   102, 127, 255);
7358
7359         createTestsForAllStages("induction", inputColors, outputColors2, fragments2, group.get());
7360
7361         const char      typesAndConstants3[]    =
7362                 "%true      = OpConstantTrue %bool\n"
7363                 "%false     = OpConstantFalse %bool\n"
7364                 "%c_f32_p2  = OpConstant %f32 0.2\n";
7365
7366         // Swap the second and the third element of the given parameter.
7367         const char      function3[]                             =
7368                 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
7369                 "%param     = OpFunctionParameter %v4f32\n"
7370                 "%entry     = OpLabel\n"
7371                 "%result    = OpVariable %fp_v4f32 Function\n"
7372                 "             OpStore %result %param\n"
7373                 "%a_loc     = OpAccessChain %fp_f32 %result %c_i32_1\n"
7374                 "%a_init    = OpLoad %f32 %a_loc\n"
7375                 "%b_loc     = OpAccessChain %fp_f32 %result %c_i32_2\n"
7376                 "%b_init    = OpLoad %f32 %b_loc\n"
7377                 "             OpBranch %phi\n"
7378
7379                 "%phi        = OpLabel\n"
7380                 "%still_loop = OpPhi %bool %true   %entry %false  %phi\n"
7381                 "%a_next     = OpPhi %f32  %a_init %entry %b_next %phi\n"
7382                 "%b_next     = OpPhi %f32  %b_init %entry %a_next %phi\n"
7383                 "              OpLoopMerge %exit %phi None\n"
7384                 "              OpBranchConditional %still_loop %phi %exit\n"
7385
7386                 "%exit       = OpLabel\n"
7387                 "              OpStore %a_loc %a_next\n"
7388                 "              OpStore %b_loc %b_next\n"
7389                 "%ret        = OpLoad %v4f32 %result\n"
7390                 "              OpReturnValue %ret\n"
7391
7392                 "              OpFunctionEnd\n";
7393
7394         fragments3["pre_main"]  = typesAndConstants3;
7395         fragments3["testfun"]   = function3;
7396
7397         outputColors3[0]                        = RGBA(127, 127, 127, 255);
7398         outputColors3[1]                        = RGBA(127, 0,   0,   255);
7399         outputColors3[2]                        = RGBA(0,   0,   127, 255);
7400         outputColors3[3]                        = RGBA(0,   127, 0,   255);
7401
7402         createTestsForAllStages("swap", inputColors, outputColors3, fragments3, group.get());
7403
7404         const char      typesAndConstants4[]    =
7405                 "%f16        = OpTypeFloat 16\n"
7406                 "%v4f16      = OpTypeVector %f16 4\n"
7407                 "%fp_f16     = OpTypePointer Function %f16\n"
7408                 "%fp_v4f16   = OpTypePointer Function %v4f16\n"
7409                 "%true       = OpConstantTrue %bool\n"
7410                 "%false      = OpConstantFalse %bool\n"
7411                 "%c_f32_p2   = OpConstant %f32 0.2\n";
7412
7413         // Swap the second and the third element of the given parameter.
7414         const char      function4[]                             =
7415                 "%test_code  = OpFunction %v4f32 None %v4f32_v4f32_function\n"
7416                 "%param      = OpFunctionParameter %v4f32\n"
7417                 "%entry      = OpLabel\n"
7418                 "%result     = OpVariable %fp_v4f16 Function\n"
7419                 "%param16    = OpFConvert %v4f16 %param\n"
7420                 "              OpStore %result %param16\n"
7421                 "%a_loc      = OpAccessChain %fp_f16 %result %c_i32_1\n"
7422                 "%a_init     = OpLoad %f16 %a_loc\n"
7423                 "%b_loc      = OpAccessChain %fp_f16 %result %c_i32_2\n"
7424                 "%b_init     = OpLoad %f16 %b_loc\n"
7425                 "              OpBranch %phi\n"
7426
7427                 "%phi        = OpLabel\n"
7428                 "%still_loop = OpPhi %bool %true   %entry %false  %phi\n"
7429                 "%a_next     = OpPhi %f16  %a_init %entry %b_next %phi\n"
7430                 "%b_next     = OpPhi %f16  %b_init %entry %a_next %phi\n"
7431                 "              OpLoopMerge %exit %phi None\n"
7432                 "              OpBranchConditional %still_loop %phi %exit\n"
7433
7434                 "%exit       = OpLabel\n"
7435                 "              OpStore %a_loc %a_next\n"
7436                 "              OpStore %b_loc %b_next\n"
7437                 "%ret16      = OpLoad %v4f16 %result\n"
7438                 "%ret        = OpFConvert %v4f32 %ret16\n"
7439                 "              OpReturnValue %ret\n"
7440
7441                 "              OpFunctionEnd\n";
7442
7443         fragments4["pre_main"]          = typesAndConstants4;
7444         fragments4["testfun"]           = function4;
7445         fragments4["capability"]        = "OpCapability StorageUniformBufferBlock16\n";
7446         fragments4["extension"]         = "OpExtension \"SPV_KHR_16bit_storage\"";
7447
7448         extensions4.push_back("VK_KHR_16bit_storage");
7449         extensions4.push_back("VK_KHR_shader_float16_int8");
7450
7451         vulkanFeatures4.ext16BitStorage = EXT16BITSTORAGEFEATURES_UNIFORM_BUFFER_BLOCK;
7452         vulkanFeatures4.extFloat16Int8  = EXTFLOAT16INT8FEATURES_FLOAT16;
7453
7454         outputColors4[0]                        = RGBA(127, 127, 127, 255);
7455         outputColors4[1]                        = RGBA(127, 0,   0,   255);
7456         outputColors4[2]                        = RGBA(0,   0,   127, 255);
7457         outputColors4[3]                        = RGBA(0,   127, 0,   255);
7458
7459         createTestsForAllStages("swap16", inputColors, outputColors4, fragments4, resources4, extensions4, group.get(), vulkanFeatures4);
7460
7461         return group.release();
7462 }
7463
7464 tcu::TestCaseGroup* createNoContractionTests(tcu::TestContext& testCtx)
7465 {
7466         de::MovePtr<tcu::TestCaseGroup> group                   (new tcu::TestCaseGroup(testCtx, "nocontraction", "Test the NoContraction decoration"));
7467         RGBA                                                    inputColors[4];
7468         RGBA                                                    outputColors[4];
7469
7470         // With NoContraction, (1 + 2^-23) * (1 - 2^-23) - 1 should be conducted as a multiplication and an addition separately.
7471         // For the multiplication, the result is 1 - 2^-46, which is out of the precision range for 32-bit float. (32-bit float
7472         // only have 23-bit fraction.) So it will be rounded to 1. Or 0x1.fffffc. Then the final result is 0 or -0x1p-24.
7473         // On the contrary, the result will be 2^-46, which is a normalized number perfectly representable as 32-bit float.
7474         const char                                              constantsAndTypes[]      =
7475                 "%c_vec4_0       = OpConstantComposite %v4f32 %c_f32_0 %c_f32_0 %c_f32_0 %c_f32_1\n"
7476                 "%c_vec4_1       = OpConstantComposite %v4f32 %c_f32_1 %c_f32_1 %c_f32_1 %c_f32_1\n"
7477                 "%c_f32_1pl2_23  = OpConstant %f32 0x1.000002p+0\n" // 1 + 2^-23
7478                 "%c_f32_1mi2_23  = OpConstant %f32 0x1.fffffcp-1\n" // 1 - 2^-23
7479                 "%c_f32_n1pn24   = OpConstant %f32 -0x1p-24\n";
7480
7481         const char                                              function[]       =
7482                 "%test_code      = OpFunction %v4f32 None %v4f32_v4f32_function\n"
7483                 "%param          = OpFunctionParameter %v4f32\n"
7484                 "%label          = OpLabel\n"
7485                 "%var1           = OpVariable %fp_f32 Function %c_f32_1pl2_23\n"
7486                 "%var2           = OpVariable %fp_f32 Function\n"
7487                 "%red            = OpCompositeExtract %f32 %param 0\n"
7488                 "%plus_red       = OpFAdd %f32 %c_f32_1mi2_23 %red\n"
7489                 "                  OpStore %var2 %plus_red\n"
7490                 "%val1           = OpLoad %f32 %var1\n"
7491                 "%val2           = OpLoad %f32 %var2\n"
7492                 "%mul            = OpFMul %f32 %val1 %val2\n"
7493                 "%add            = OpFAdd %f32 %mul %c_f32_n1\n"
7494                 "%is0            = OpFOrdEqual %bool %add %c_f32_0\n"
7495                 "%isn1n24         = OpFOrdEqual %bool %add %c_f32_n1pn24\n"
7496                 "%success        = OpLogicalOr %bool %is0 %isn1n24\n"
7497                 "%v4success      = OpCompositeConstruct %v4bool %success %success %success %success\n"
7498                 "%ret            = OpSelect %v4f32 %v4success %c_vec4_0 %c_vec4_1\n"
7499                 "                  OpReturnValue %ret\n"
7500                 "                  OpFunctionEnd\n";
7501
7502         struct CaseNameDecoration
7503         {
7504                 string name;
7505                 string decoration;
7506         };
7507
7508
7509         CaseNameDecoration tests[] = {
7510                 {"multiplication",      "OpDecorate %mul NoContraction"},
7511                 {"addition",            "OpDecorate %add NoContraction"},
7512                 {"both",                        "OpDecorate %mul NoContraction\nOpDecorate %add NoContraction"},
7513         };
7514
7515         getHalfColorsFullAlpha(inputColors);
7516
7517         for (deUint8 idx = 0; idx < 4; ++idx)
7518         {
7519                 inputColors[idx].setRed(0);
7520                 outputColors[idx] = RGBA(0, 0, 0, 255);
7521         }
7522
7523         for (size_t testNdx = 0; testNdx < sizeof(tests) / sizeof(CaseNameDecoration); ++testNdx)
7524         {
7525                 map<string, string> fragments;
7526
7527                 fragments["decoration"] = tests[testNdx].decoration;
7528                 fragments["pre_main"] = constantsAndTypes;
7529                 fragments["testfun"] = function;
7530
7531                 createTestsForAllStages(tests[testNdx].name, inputColors, outputColors, fragments, group.get());
7532         }
7533
7534         return group.release();
7535 }
7536
7537 tcu::TestCaseGroup* createMemoryAccessTests(tcu::TestContext& testCtx)
7538 {
7539         de::MovePtr<tcu::TestCaseGroup> memoryAccessTests (new tcu::TestCaseGroup(testCtx, "opmemoryaccess", "Memory Semantics"));
7540         RGBA                                                    colors[4];
7541
7542         const char                                              constantsAndTypes[]      =
7543                 "%c_a2f32_1         = OpConstantComposite %a2f32 %c_f32_1 %c_f32_1\n"
7544                 "%fp_a2f32          = OpTypePointer Function %a2f32\n"
7545                 "%stype             = OpTypeStruct  %v4f32 %a2f32 %f32\n"
7546                 "%fp_stype          = OpTypePointer Function %stype\n";
7547
7548         const char                                              function[]       =
7549                 "%test_code         = OpFunction %v4f32 None %v4f32_v4f32_function\n"
7550                 "%param1            = OpFunctionParameter %v4f32\n"
7551                 "%lbl               = OpLabel\n"
7552                 "%v1                = OpVariable %fp_v4f32 Function\n"
7553                 "%v2                = OpVariable %fp_a2f32 Function\n"
7554                 "%v3                = OpVariable %fp_f32 Function\n"
7555                 "%v                 = OpVariable %fp_stype Function\n"
7556                 "%vv                = OpVariable %fp_stype Function\n"
7557                 "%vvv               = OpVariable %fp_f32 Function\n"
7558
7559                 "                     OpStore %v1 %c_v4f32_1_1_1_1\n"
7560                 "                     OpStore %v2 %c_a2f32_1\n"
7561                 "                     OpStore %v3 %c_f32_1\n"
7562
7563                 "%p_v4f32          = OpAccessChain %fp_v4f32 %v %c_u32_0\n"
7564                 "%p_a2f32          = OpAccessChain %fp_a2f32 %v %c_u32_1\n"
7565                 "%p_f32            = OpAccessChain %fp_f32 %v %c_u32_2\n"
7566                 "%v1_v             = OpLoad %v4f32 %v1 ${access_type}\n"
7567                 "%v2_v             = OpLoad %a2f32 %v2 ${access_type}\n"
7568                 "%v3_v             = OpLoad %f32 %v3 ${access_type}\n"
7569
7570                 "                    OpStore %p_v4f32 %v1_v ${access_type}\n"
7571                 "                    OpStore %p_a2f32 %v2_v ${access_type}\n"
7572                 "                    OpStore %p_f32 %v3_v ${access_type}\n"
7573
7574                 "                    OpCopyMemory %vv %v ${access_type}\n"
7575                 "                    OpCopyMemory %vvv %p_f32 ${access_type}\n"
7576
7577                 "%p_f32_2          = OpAccessChain %fp_f32 %vv %c_u32_2\n"
7578                 "%v_f32_2          = OpLoad %f32 %p_f32_2\n"
7579                 "%v_f32_3          = OpLoad %f32 %vvv\n"
7580
7581                 "%ret1             = OpVectorTimesScalar %v4f32 %param1 %v_f32_2\n"
7582                 "%ret2             = OpVectorTimesScalar %v4f32 %ret1 %v_f32_3\n"
7583                 "                    OpReturnValue %ret2\n"
7584                 "                    OpFunctionEnd\n";
7585
7586         struct NameMemoryAccess
7587         {
7588                 string name;
7589                 string accessType;
7590         };
7591
7592
7593         NameMemoryAccess tests[] =
7594         {
7595                 { "none", "" },
7596                 { "volatile", "Volatile" },
7597                 { "aligned",  "Aligned 1" },
7598                 { "volatile_aligned",  "Volatile|Aligned 1" },
7599                 { "nontemporal_aligned",  "Nontemporal|Aligned 1" },
7600                 { "volatile_nontemporal",  "Volatile|Nontemporal" },
7601                 { "volatile_nontermporal_aligned",  "Volatile|Nontemporal|Aligned 1" },
7602         };
7603
7604         getHalfColorsFullAlpha(colors);
7605
7606         for (size_t testNdx = 0; testNdx < sizeof(tests) / sizeof(NameMemoryAccess); ++testNdx)
7607         {
7608                 map<string, string> fragments;
7609                 map<string, string> memoryAccess;
7610                 memoryAccess["access_type"] = tests[testNdx].accessType;
7611
7612                 fragments["pre_main"] = constantsAndTypes;
7613                 fragments["testfun"] = tcu::StringTemplate(function).specialize(memoryAccess);
7614                 createTestsForAllStages(tests[testNdx].name, colors, colors, fragments, memoryAccessTests.get());
7615         }
7616         return memoryAccessTests.release();
7617 }
7618 tcu::TestCaseGroup* createOpUndefTests(tcu::TestContext& testCtx)
7619 {
7620         de::MovePtr<tcu::TestCaseGroup>         opUndefTests             (new tcu::TestCaseGroup(testCtx, "opundef", "Test OpUndef"));
7621         RGBA                                                            defaultColors[4];
7622         map<string, string>                                     fragments;
7623         getDefaultColors(defaultColors);
7624
7625         // First, simple cases that don't do anything with the OpUndef result.
7626         struct NameCodePair { string name, decl, type; };
7627         const NameCodePair tests[] =
7628         {
7629                 {"bool", "", "%bool"},
7630                 {"vec2uint32", "", "%v2u32"},
7631                 {"image", "%type = OpTypeImage %f32 2D 0 0 0 1 Unknown", "%type"},
7632                 {"sampler", "%type = OpTypeSampler", "%type"},
7633                 {"sampledimage", "%img = OpTypeImage %f32 2D 0 0 0 1 Unknown\n" "%type = OpTypeSampledImage %img", "%type"},
7634                 {"pointer", "", "%fp_i32"},
7635                 {"runtimearray", "%type = OpTypeRuntimeArray %f32", "%type"},
7636                 {"array", "%c_u32_100 = OpConstant %u32 100\n" "%type = OpTypeArray %i32 %c_u32_100", "%type"},
7637                 {"struct", "%type = OpTypeStruct %f32 %i32 %u32", "%type"}};
7638         for (size_t testNdx = 0; testNdx < sizeof(tests) / sizeof(NameCodePair); ++testNdx)
7639         {
7640                 fragments["undef_type"] = tests[testNdx].type;
7641                 fragments["testfun"] = StringTemplate(
7642                         "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
7643                         "%param1 = OpFunctionParameter %v4f32\n"
7644                         "%label_testfun = OpLabel\n"
7645                         "%undef = OpUndef ${undef_type}\n"
7646                         "OpReturnValue %param1\n"
7647                         "OpFunctionEnd\n").specialize(fragments);
7648                 fragments["pre_main"] = tests[testNdx].decl;
7649                 createTestsForAllStages(tests[testNdx].name, defaultColors, defaultColors, fragments, opUndefTests.get());
7650         }
7651         fragments.clear();
7652
7653         fragments["testfun"] =
7654                 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
7655                 "%param1 = OpFunctionParameter %v4f32\n"
7656                 "%label_testfun = OpLabel\n"
7657                 "%undef = OpUndef %f32\n"
7658                 "%zero = OpFMul %f32 %undef %c_f32_0\n"
7659                 "%is_nan = OpIsNan %bool %zero\n" //OpUndef may result in NaN which may turn %zero into Nan.
7660                 "%actually_zero = OpSelect %f32 %is_nan %c_f32_0 %zero\n"
7661                 "%a = OpVectorExtractDynamic %f32 %param1 %c_i32_0\n"
7662                 "%b = OpFAdd %f32 %a %actually_zero\n"
7663                 "%ret = OpVectorInsertDynamic %v4f32 %param1 %b %c_i32_0\n"
7664                 "OpReturnValue %ret\n"
7665                 "OpFunctionEnd\n";
7666
7667         createTestsForAllStages("float32", defaultColors, defaultColors, fragments, opUndefTests.get());
7668
7669         fragments["testfun"] =
7670                 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
7671                 "%param1 = OpFunctionParameter %v4f32\n"
7672                 "%label_testfun = OpLabel\n"
7673                 "%undef = OpUndef %i32\n"
7674                 "%zero = OpIMul %i32 %undef %c_i32_0\n"
7675                 "%a = OpVectorExtractDynamic %f32 %param1 %zero\n"
7676                 "%ret = OpVectorInsertDynamic %v4f32 %param1 %a %c_i32_0\n"
7677                 "OpReturnValue %ret\n"
7678                 "OpFunctionEnd\n";
7679
7680         createTestsForAllStages("sint32", defaultColors, defaultColors, fragments, opUndefTests.get());
7681
7682         fragments["testfun"] =
7683                 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
7684                 "%param1 = OpFunctionParameter %v4f32\n"
7685                 "%label_testfun = OpLabel\n"
7686                 "%undef = OpUndef %u32\n"
7687                 "%zero = OpIMul %u32 %undef %c_i32_0\n"
7688                 "%a = OpVectorExtractDynamic %f32 %param1 %zero\n"
7689                 "%ret = OpVectorInsertDynamic %v4f32 %param1 %a %c_i32_0\n"
7690                 "OpReturnValue %ret\n"
7691                 "OpFunctionEnd\n";
7692
7693         createTestsForAllStages("uint32", defaultColors, defaultColors, fragments, opUndefTests.get());
7694
7695         fragments["testfun"] =
7696                 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
7697                 "%param1 = OpFunctionParameter %v4f32\n"
7698                 "%label_testfun = OpLabel\n"
7699                 "%undef = OpUndef %v4f32\n"
7700                 "%vzero = OpVectorTimesScalar %v4f32 %undef %c_f32_0\n"
7701                 "%zero_0 = OpVectorExtractDynamic %f32 %vzero %c_i32_0\n"
7702                 "%zero_1 = OpVectorExtractDynamic %f32 %vzero %c_i32_1\n"
7703                 "%zero_2 = OpVectorExtractDynamic %f32 %vzero %c_i32_2\n"
7704                 "%zero_3 = OpVectorExtractDynamic %f32 %vzero %c_i32_3\n"
7705                 "%is_nan_0 = OpIsNan %bool %zero_0\n"
7706                 "%is_nan_1 = OpIsNan %bool %zero_1\n"
7707                 "%is_nan_2 = OpIsNan %bool %zero_2\n"
7708                 "%is_nan_3 = OpIsNan %bool %zero_3\n"
7709                 "%actually_zero_0 = OpSelect %f32 %is_nan_0 %c_f32_0 %zero_0\n"
7710                 "%actually_zero_1 = OpSelect %f32 %is_nan_1 %c_f32_0 %zero_1\n"
7711                 "%actually_zero_2 = OpSelect %f32 %is_nan_2 %c_f32_0 %zero_2\n"
7712                 "%actually_zero_3 = OpSelect %f32 %is_nan_3 %c_f32_0 %zero_3\n"
7713                 "%param1_0 = OpVectorExtractDynamic %f32 %param1 %c_i32_0\n"
7714                 "%param1_1 = OpVectorExtractDynamic %f32 %param1 %c_i32_1\n"
7715                 "%param1_2 = OpVectorExtractDynamic %f32 %param1 %c_i32_2\n"
7716                 "%param1_3 = OpVectorExtractDynamic %f32 %param1 %c_i32_3\n"
7717                 "%sum_0 = OpFAdd %f32 %param1_0 %actually_zero_0\n"
7718                 "%sum_1 = OpFAdd %f32 %param1_1 %actually_zero_1\n"
7719                 "%sum_2 = OpFAdd %f32 %param1_2 %actually_zero_2\n"
7720                 "%sum_3 = OpFAdd %f32 %param1_3 %actually_zero_3\n"
7721                 "%ret3 = OpVectorInsertDynamic %v4f32 %param1 %sum_3 %c_i32_3\n"
7722                 "%ret2 = OpVectorInsertDynamic %v4f32 %ret3 %sum_2 %c_i32_2\n"
7723                 "%ret1 = OpVectorInsertDynamic %v4f32 %ret2 %sum_1 %c_i32_1\n"
7724                 "%ret = OpVectorInsertDynamic %v4f32 %ret1 %sum_0 %c_i32_0\n"
7725                 "OpReturnValue %ret\n"
7726                 "OpFunctionEnd\n";
7727
7728         createTestsForAllStages("vec4float32", defaultColors, defaultColors, fragments, opUndefTests.get());
7729
7730         fragments["pre_main"] =
7731                 "%m2x2f32 = OpTypeMatrix %v2f32 2\n";
7732         fragments["testfun"] =
7733                 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
7734                 "%param1 = OpFunctionParameter %v4f32\n"
7735                 "%label_testfun = OpLabel\n"
7736                 "%undef = OpUndef %m2x2f32\n"
7737                 "%mzero = OpMatrixTimesScalar %m2x2f32 %undef %c_f32_0\n"
7738                 "%zero_0 = OpCompositeExtract %f32 %mzero 0 0\n"
7739                 "%zero_1 = OpCompositeExtract %f32 %mzero 0 1\n"
7740                 "%zero_2 = OpCompositeExtract %f32 %mzero 1 0\n"
7741                 "%zero_3 = OpCompositeExtract %f32 %mzero 1 1\n"
7742                 "%is_nan_0 = OpIsNan %bool %zero_0\n"
7743                 "%is_nan_1 = OpIsNan %bool %zero_1\n"
7744                 "%is_nan_2 = OpIsNan %bool %zero_2\n"
7745                 "%is_nan_3 = OpIsNan %bool %zero_3\n"
7746                 "%actually_zero_0 = OpSelect %f32 %is_nan_0 %c_f32_0 %zero_0\n"
7747                 "%actually_zero_1 = OpSelect %f32 %is_nan_1 %c_f32_0 %zero_1\n"
7748                 "%actually_zero_2 = OpSelect %f32 %is_nan_2 %c_f32_0 %zero_2\n"
7749                 "%actually_zero_3 = OpSelect %f32 %is_nan_3 %c_f32_0 %zero_3\n"
7750                 "%param1_0 = OpVectorExtractDynamic %f32 %param1 %c_i32_0\n"
7751                 "%param1_1 = OpVectorExtractDynamic %f32 %param1 %c_i32_1\n"
7752                 "%param1_2 = OpVectorExtractDynamic %f32 %param1 %c_i32_2\n"
7753                 "%param1_3 = OpVectorExtractDynamic %f32 %param1 %c_i32_3\n"
7754                 "%sum_0 = OpFAdd %f32 %param1_0 %actually_zero_0\n"
7755                 "%sum_1 = OpFAdd %f32 %param1_1 %actually_zero_1\n"
7756                 "%sum_2 = OpFAdd %f32 %param1_2 %actually_zero_2\n"
7757                 "%sum_3 = OpFAdd %f32 %param1_3 %actually_zero_3\n"
7758                 "%ret3 = OpVectorInsertDynamic %v4f32 %param1 %sum_3 %c_i32_3\n"
7759                 "%ret2 = OpVectorInsertDynamic %v4f32 %ret3 %sum_2 %c_i32_2\n"
7760                 "%ret1 = OpVectorInsertDynamic %v4f32 %ret2 %sum_1 %c_i32_1\n"
7761                 "%ret = OpVectorInsertDynamic %v4f32 %ret1 %sum_0 %c_i32_0\n"
7762                 "OpReturnValue %ret\n"
7763                 "OpFunctionEnd\n";
7764
7765         createTestsForAllStages("matrix", defaultColors, defaultColors, fragments, opUndefTests.get());
7766
7767         return opUndefTests.release();
7768 }
7769
7770 void createOpQuantizeSingleOptionTests(tcu::TestCaseGroup* testCtx)
7771 {
7772         const RGBA              inputColors[4]          =
7773         {
7774                 RGBA(0,         0,              0,              255),
7775                 RGBA(0,         0,              255,    255),
7776                 RGBA(0,         255,    0,              255),
7777                 RGBA(0,         255,    255,    255)
7778         };
7779
7780         const RGBA              expectedColors[4]       =
7781         {
7782                 RGBA(255,        0,              0,              255),
7783                 RGBA(255,        0,              0,              255),
7784                 RGBA(255,        0,              0,              255),
7785                 RGBA(255,        0,              0,              255)
7786         };
7787
7788         const struct SingleFP16Possibility
7789         {
7790                 const char* name;
7791                 const char* constant;  // Value to assign to %test_constant.
7792                 float           valueAsFloat;
7793                 const char* condition; // Must assign to %cond an expression that evaluates to true after %c = OpQuantizeToF16(%test_constant + 0).
7794         }                               tests[]                         =
7795         {
7796                 {
7797                         "negative",
7798                         "-0x1.3p1\n",
7799                         -constructNormalizedFloat(1, 0x300000),
7800                         "%cond = OpFOrdEqual %bool %c %test_constant\n"
7801                 }, // -19
7802                 {
7803                         "positive",
7804                         "0x1.0p7\n",
7805                         constructNormalizedFloat(7, 0x000000),
7806                         "%cond = OpFOrdEqual %bool %c %test_constant\n"
7807                 },  // +128
7808                 // SPIR-V requires that OpQuantizeToF16 flushes
7809                 // any numbers that would end up denormalized in F16 to zero.
7810                 {
7811                         "denorm",
7812                         "0x0.0006p-126\n",
7813                         std::ldexp(1.5f, -140),
7814                         "%cond = OpFOrdEqual %bool %c %c_f32_0\n"
7815                 },  // denorm
7816                 {
7817                         "negative_denorm",
7818                         "-0x0.0006p-126\n",
7819                         -std::ldexp(1.5f, -140),
7820                         "%cond = OpFOrdEqual %bool %c %c_f32_0\n"
7821                 }, // -denorm
7822                 {
7823                         "too_small",
7824                         "0x1.0p-16\n",
7825                         std::ldexp(1.0f, -16),
7826                         "%cond = OpFOrdEqual %bool %c %c_f32_0\n"
7827                 },     // too small positive
7828                 {
7829                         "negative_too_small",
7830                         "-0x1.0p-32\n",
7831                         -std::ldexp(1.0f, -32),
7832                         "%cond = OpFOrdEqual %bool %c %c_f32_0\n"
7833                 },      // too small negative
7834                 {
7835                         "negative_inf",
7836                         "-0x1.0p128\n",
7837                         -std::ldexp(1.0f, 128),
7838
7839                         "%gz = OpFOrdLessThan %bool %c %c_f32_0\n"
7840                         "%inf = OpIsInf %bool %c\n"
7841                         "%cond = OpLogicalAnd %bool %gz %inf\n"
7842                 },     // -inf to -inf
7843                 {
7844                         "inf",
7845                         "0x1.0p128\n",
7846                         std::ldexp(1.0f, 128),
7847
7848                         "%gz = OpFOrdGreaterThan %bool %c %c_f32_0\n"
7849                         "%inf = OpIsInf %bool %c\n"
7850                         "%cond = OpLogicalAnd %bool %gz %inf\n"
7851                 },     // +inf to +inf
7852                 {
7853                         "round_to_negative_inf",
7854                         "-0x1.0p32\n",
7855                         -std::ldexp(1.0f, 32),
7856
7857                         "%gz = OpFOrdLessThan %bool %c %c_f32_0\n"
7858                         "%inf = OpIsInf %bool %c\n"
7859                         "%cond = OpLogicalAnd %bool %gz %inf\n"
7860                 },     // round to -inf
7861                 {
7862                         "round_to_inf",
7863                         "0x1.0p16\n",
7864                         std::ldexp(1.0f, 16),
7865
7866                         "%gz = OpFOrdGreaterThan %bool %c %c_f32_0\n"
7867                         "%inf = OpIsInf %bool %c\n"
7868                         "%cond = OpLogicalAnd %bool %gz %inf\n"
7869                 },     // round to +inf
7870                 {
7871                         "nan",
7872                         "0x1.1p128\n",
7873                         std::numeric_limits<float>::quiet_NaN(),
7874
7875                         // Test for any NaN value, as NaNs are not preserved
7876                         "%direct_quant = OpQuantizeToF16 %f32 %test_constant\n"
7877                         "%cond = OpIsNan %bool %direct_quant\n"
7878                 }, // nan
7879                 {
7880                         "negative_nan",
7881                         "-0x1.0001p128\n",
7882                         std::numeric_limits<float>::quiet_NaN(),
7883
7884                         // Test for any NaN value, as NaNs are not preserved
7885                         "%direct_quant = OpQuantizeToF16 %f32 %test_constant\n"
7886                         "%cond = OpIsNan %bool %direct_quant\n"
7887                 } // -nan
7888         };
7889         const char*             constants                       =
7890                 "%test_constant = OpConstant %f32 ";  // The value will be test.constant.
7891
7892         StringTemplate  function                        (
7893                 "%test_code     = OpFunction %v4f32 None %v4f32_v4f32_function\n"
7894                 "%param1        = OpFunctionParameter %v4f32\n"
7895                 "%label_testfun = OpLabel\n"
7896                 "%a             = OpVectorExtractDynamic %f32 %param1 %c_i32_0\n"
7897                 "%b             = OpFAdd %f32 %test_constant %a\n"
7898                 "%c             = OpQuantizeToF16 %f32 %b\n"
7899                 "${condition}\n"
7900                 "%v4cond        = OpCompositeConstruct %v4bool %cond %cond %cond %cond\n"
7901                 "%retval        = OpSelect %v4f32 %v4cond %c_v4f32_1_0_0_1 %param1\n"
7902                 "                 OpReturnValue %retval\n"
7903                 "OpFunctionEnd\n"
7904         );
7905
7906         const char*             specDecorations         = "OpDecorate %test_constant SpecId 0\n";
7907         const char*             specConstants           =
7908                         "%test_constant = OpSpecConstant %f32 0.\n"
7909                         "%c             = OpSpecConstantOp %f32 QuantizeToF16 %test_constant\n";
7910
7911         StringTemplate  specConstantFunction(
7912                 "%test_code     = OpFunction %v4f32 None %v4f32_v4f32_function\n"
7913                 "%param1        = OpFunctionParameter %v4f32\n"
7914                 "%label_testfun = OpLabel\n"
7915                 "${condition}\n"
7916                 "%v4cond        = OpCompositeConstruct %v4bool %cond %cond %cond %cond\n"
7917                 "%retval        = OpSelect %v4f32 %v4cond %c_v4f32_1_0_0_1 %param1\n"
7918                 "                 OpReturnValue %retval\n"
7919                 "OpFunctionEnd\n"
7920         );
7921
7922         for (size_t idx = 0; idx < (sizeof(tests)/sizeof(tests[0])); ++idx)
7923         {
7924                 map<string, string>                                                             codeSpecialization;
7925                 map<string, string>                                                             fragments;
7926                 codeSpecialization["condition"]                                 = tests[idx].condition;
7927                 fragments["testfun"]                                                    = function.specialize(codeSpecialization);
7928                 fragments["pre_main"]                                                   = string(constants) + tests[idx].constant + "\n";
7929                 createTestsForAllStages(tests[idx].name, inputColors, expectedColors, fragments, testCtx);
7930         }
7931
7932         for (size_t idx = 0; idx < (sizeof(tests)/sizeof(tests[0])); ++idx)
7933         {
7934                 map<string, string>                                                             codeSpecialization;
7935                 map<string, string>                                                             fragments;
7936                 SpecConstants                                                                   passConstants;
7937
7938                 codeSpecialization["condition"]                                 = tests[idx].condition;
7939                 fragments["testfun"]                                                    = specConstantFunction.specialize(codeSpecialization);
7940                 fragments["decoration"]                                                 = specDecorations;
7941                 fragments["pre_main"]                                                   = specConstants;
7942
7943                 passConstants.append<float>(tests[idx].valueAsFloat);
7944
7945                 createTestsForAllStages(string("spec_const_") + tests[idx].name, inputColors, expectedColors, fragments, passConstants, testCtx);
7946         }
7947 }
7948
7949 void createOpQuantizeTwoPossibilityTests(tcu::TestCaseGroup* testCtx)
7950 {
7951         RGBA inputColors[4] =  {
7952                 RGBA(0,         0,              0,              255),
7953                 RGBA(0,         0,              255,    255),
7954                 RGBA(0,         255,    0,              255),
7955                 RGBA(0,         255,    255,    255)
7956         };
7957
7958         RGBA expectedColors[4] =
7959         {
7960                 RGBA(255,        0,              0,              255),
7961                 RGBA(255,        0,              0,              255),
7962                 RGBA(255,        0,              0,              255),
7963                 RGBA(255,        0,              0,              255)
7964         };
7965
7966         struct DualFP16Possibility
7967         {
7968                 const char* name;
7969                 const char* input;
7970                 float           inputAsFloat;
7971                 const char* possibleOutput1;
7972                 const char* possibleOutput2;
7973         } tests[] = {
7974                 {
7975                         "positive_round_up_or_round_down",
7976                         "0x1.3003p8",
7977                         constructNormalizedFloat(8, 0x300300),
7978                         "0x1.304p8",
7979                         "0x1.3p8"
7980                 },
7981                 {
7982                         "negative_round_up_or_round_down",
7983                         "-0x1.6008p-7",
7984                         -constructNormalizedFloat(-7, 0x600800),
7985                         "-0x1.6p-7",
7986                         "-0x1.604p-7"
7987                 },
7988                 {
7989                         "carry_bit",
7990                         "0x1.01ep2",
7991                         constructNormalizedFloat(2, 0x01e000),
7992                         "0x1.01cp2",
7993                         "0x1.02p2"
7994                 },
7995                 {
7996                         "carry_to_exponent",
7997                         "0x1.ffep1",
7998                         constructNormalizedFloat(1, 0xffe000),
7999                         "0x1.ffcp1",
8000                         "0x1.0p2"
8001                 },
8002         };
8003         StringTemplate constants (
8004                 "%input_const = OpConstant %f32 ${input}\n"
8005                 "%possible_solution1 = OpConstant %f32 ${output1}\n"
8006                 "%possible_solution2 = OpConstant %f32 ${output2}\n"
8007                 );
8008
8009         StringTemplate specConstants (
8010                 "%input_const = OpSpecConstant %f32 0.\n"
8011                 "%possible_solution1 = OpConstant %f32 ${output1}\n"
8012                 "%possible_solution2 = OpConstant %f32 ${output2}\n"
8013         );
8014
8015         const char* specDecorations = "OpDecorate %input_const  SpecId 0\n";
8016
8017         const char* function  =
8018                 "%test_code     = OpFunction %v4f32 None %v4f32_v4f32_function\n"
8019                 "%param1        = OpFunctionParameter %v4f32\n"
8020                 "%label_testfun = OpLabel\n"
8021                 "%a             = OpVectorExtractDynamic %f32 %param1 %c_i32_0\n"
8022                 // For the purposes of this test we assume that 0.f will always get
8023                 // faithfully passed through the pipeline stages.
8024                 "%b             = OpFAdd %f32 %input_const %a\n"
8025                 "%c             = OpQuantizeToF16 %f32 %b\n"
8026                 "%eq_1          = OpFOrdEqual %bool %c %possible_solution1\n"
8027                 "%eq_2          = OpFOrdEqual %bool %c %possible_solution2\n"
8028                 "%cond          = OpLogicalOr %bool %eq_1 %eq_2\n"
8029                 "%v4cond        = OpCompositeConstruct %v4bool %cond %cond %cond %cond\n"
8030                 "%retval        = OpSelect %v4f32 %v4cond %c_v4f32_1_0_0_1 %param1"
8031                 "                 OpReturnValue %retval\n"
8032                 "OpFunctionEnd\n";
8033
8034         for(size_t idx = 0; idx < (sizeof(tests)/sizeof(tests[0])); ++idx) {
8035                 map<string, string>                                                                     fragments;
8036                 map<string, string>                                                                     constantSpecialization;
8037
8038                 constantSpecialization["input"]                                         = tests[idx].input;
8039                 constantSpecialization["output1"]                                       = tests[idx].possibleOutput1;
8040                 constantSpecialization["output2"]                                       = tests[idx].possibleOutput2;
8041                 fragments["testfun"]                                                            = function;
8042                 fragments["pre_main"]                                                           = constants.specialize(constantSpecialization);
8043                 createTestsForAllStages(tests[idx].name, inputColors, expectedColors, fragments, testCtx);
8044         }
8045
8046         for(size_t idx = 0; idx < (sizeof(tests)/sizeof(tests[0])); ++idx) {
8047                 map<string, string>                                                                     fragments;
8048                 map<string, string>                                                                     constantSpecialization;
8049                 SpecConstants                                                                           passConstants;
8050
8051                 constantSpecialization["output1"]                                       = tests[idx].possibleOutput1;
8052                 constantSpecialization["output2"]                                       = tests[idx].possibleOutput2;
8053                 fragments["testfun"]                                                            = function;
8054                 fragments["decoration"]                                                         = specDecorations;
8055                 fragments["pre_main"]                                                           = specConstants.specialize(constantSpecialization);
8056
8057                 passConstants.append<float>(tests[idx].inputAsFloat);
8058
8059                 createTestsForAllStages(string("spec_const_") + tests[idx].name, inputColors, expectedColors, fragments, passConstants, testCtx);
8060         }
8061 }
8062
8063 tcu::TestCaseGroup* createOpQuantizeTests(tcu::TestContext& testCtx)
8064 {
8065         de::MovePtr<tcu::TestCaseGroup> opQuantizeTests (new tcu::TestCaseGroup(testCtx, "opquantize", "Test OpQuantizeToF16"));
8066         createOpQuantizeSingleOptionTests(opQuantizeTests.get());
8067         createOpQuantizeTwoPossibilityTests(opQuantizeTests.get());
8068         return opQuantizeTests.release();
8069 }
8070
8071 struct ShaderPermutation
8072 {
8073         deUint8 vertexPermutation;
8074         deUint8 geometryPermutation;
8075         deUint8 tesscPermutation;
8076         deUint8 tessePermutation;
8077         deUint8 fragmentPermutation;
8078 };
8079
8080 ShaderPermutation getShaderPermutation(deUint8 inputValue)
8081 {
8082         ShaderPermutation       permutation =
8083         {
8084                 static_cast<deUint8>(inputValue & 0x10? 1u: 0u),
8085                 static_cast<deUint8>(inputValue & 0x08? 1u: 0u),
8086                 static_cast<deUint8>(inputValue & 0x04? 1u: 0u),
8087                 static_cast<deUint8>(inputValue & 0x02? 1u: 0u),
8088                 static_cast<deUint8>(inputValue & 0x01? 1u: 0u)
8089         };
8090         return permutation;
8091 }
8092
8093 tcu::TestCaseGroup* createModuleTests(tcu::TestContext& testCtx)
8094 {
8095         RGBA                                                            defaultColors[4];
8096         RGBA                                                            invertedColors[4];
8097         de::MovePtr<tcu::TestCaseGroup>         moduleTests                     (new tcu::TestCaseGroup(testCtx, "module", "Multiple entry points into shaders"));
8098
8099         getDefaultColors(defaultColors);
8100         getInvertedDefaultColors(invertedColors);
8101
8102         // Combined module tests
8103         {
8104                 // Shader stages: vertex and fragment
8105                 {
8106                         const ShaderElement combinedPipeline[]  =
8107                         {
8108                                 ShaderElement("module", "main", VK_SHADER_STAGE_VERTEX_BIT),
8109                                 ShaderElement("module", "main", VK_SHADER_STAGE_FRAGMENT_BIT)
8110                         };
8111
8112                         addFunctionCaseWithPrograms<InstanceContext>(
8113                                 moduleTests.get(), "same_module", "", createCombinedModule, runAndVerifyDefaultPipeline,
8114                                 createInstanceContext(combinedPipeline, map<string, string>()));
8115                 }
8116
8117                 // Shader stages: vertex, geometry and fragment
8118                 {
8119                         const ShaderElement combinedPipeline[]  =
8120                         {
8121                                 ShaderElement("module", "main", VK_SHADER_STAGE_VERTEX_BIT),
8122                                 ShaderElement("module", "main", VK_SHADER_STAGE_GEOMETRY_BIT),
8123                                 ShaderElement("module", "main", VK_SHADER_STAGE_FRAGMENT_BIT)
8124                         };
8125
8126                         addFunctionCaseWithPrograms<InstanceContext>(
8127                                 moduleTests.get(), "same_module_geom", "", createCombinedModule, runAndVerifyDefaultPipeline,
8128                                 createInstanceContext(combinedPipeline, map<string, string>()));
8129                 }
8130
8131                 // Shader stages: vertex, tessellation control, tessellation evaluation and fragment
8132                 {
8133                         const ShaderElement combinedPipeline[]  =
8134                         {
8135                                 ShaderElement("module", "main", VK_SHADER_STAGE_VERTEX_BIT),
8136                                 ShaderElement("module", "main", VK_SHADER_STAGE_TESSELLATION_CONTROL_BIT),
8137                                 ShaderElement("module", "main", VK_SHADER_STAGE_TESSELLATION_EVALUATION_BIT),
8138                                 ShaderElement("module", "main", VK_SHADER_STAGE_FRAGMENT_BIT)
8139                         };
8140
8141                         addFunctionCaseWithPrograms<InstanceContext>(
8142                                 moduleTests.get(), "same_module_tessc_tesse", "", createCombinedModule, runAndVerifyDefaultPipeline,
8143                                 createInstanceContext(combinedPipeline, map<string, string>()));
8144                 }
8145
8146                 // Shader stages: vertex, tessellation control, tessellation evaluation, geometry and fragment
8147                 {
8148                         const ShaderElement combinedPipeline[]  =
8149                         {
8150                                 ShaderElement("module", "main", VK_SHADER_STAGE_VERTEX_BIT),
8151                                 ShaderElement("module", "main", VK_SHADER_STAGE_TESSELLATION_CONTROL_BIT),
8152                                 ShaderElement("module", "main", VK_SHADER_STAGE_TESSELLATION_EVALUATION_BIT),
8153                                 ShaderElement("module", "main", VK_SHADER_STAGE_GEOMETRY_BIT),
8154                                 ShaderElement("module", "main", VK_SHADER_STAGE_FRAGMENT_BIT)
8155                         };
8156
8157                         addFunctionCaseWithPrograms<InstanceContext>(
8158                                 moduleTests.get(), "same_module_tessc_tesse_geom", "", createCombinedModule, runAndVerifyDefaultPipeline,
8159                                 createInstanceContext(combinedPipeline, map<string, string>()));
8160                 }
8161         }
8162
8163         const char* numbers[] =
8164         {
8165                 "1", "2"
8166         };
8167
8168         for (deInt8 idx = 0; idx < 32; ++idx)
8169         {
8170                 ShaderPermutation                       permutation             = getShaderPermutation(idx);
8171                 string                                          name                    = string("vert") + numbers[permutation.vertexPermutation] + "_geom" + numbers[permutation.geometryPermutation] + "_tessc" + numbers[permutation.tesscPermutation] + "_tesse" + numbers[permutation.tessePermutation] + "_frag" + numbers[permutation.fragmentPermutation];
8172                 const ShaderElement                     pipeline[]              =
8173                 {
8174                         ShaderElement("vert",   string("vert") +        numbers[permutation.vertexPermutation],         VK_SHADER_STAGE_VERTEX_BIT),
8175                         ShaderElement("geom",   string("geom") +        numbers[permutation.geometryPermutation],       VK_SHADER_STAGE_GEOMETRY_BIT),
8176                         ShaderElement("tessc",  string("tessc") +       numbers[permutation.tesscPermutation],          VK_SHADER_STAGE_TESSELLATION_CONTROL_BIT),
8177                         ShaderElement("tesse",  string("tesse") +       numbers[permutation.tessePermutation],          VK_SHADER_STAGE_TESSELLATION_EVALUATION_BIT),
8178                         ShaderElement("frag",   string("frag") +        numbers[permutation.fragmentPermutation],       VK_SHADER_STAGE_FRAGMENT_BIT)
8179                 };
8180
8181                 // If there are an even number of swaps, then it should be no-op.
8182                 // If there are an odd number, the color should be flipped.
8183                 if ((permutation.vertexPermutation + permutation.geometryPermutation + permutation.tesscPermutation + permutation.tessePermutation + permutation.fragmentPermutation) % 2 == 0)
8184                 {
8185                         addFunctionCaseWithPrograms<InstanceContext>(
8186                                         moduleTests.get(), name, "", createMultipleEntries, runAndVerifyDefaultPipeline,
8187                                         createInstanceContext(pipeline, defaultColors, defaultColors, map<string, string>()));
8188                 }
8189                 else
8190                 {
8191                         addFunctionCaseWithPrograms<InstanceContext>(
8192                                         moduleTests.get(), name, "", createMultipleEntries, runAndVerifyDefaultPipeline,
8193                                         createInstanceContext(pipeline, defaultColors, invertedColors, map<string, string>()));
8194                 }
8195         }
8196         return moduleTests.release();
8197 }
8198
8199 tcu::TestCaseGroup* createLoopTests(tcu::TestContext& testCtx)
8200 {
8201         de::MovePtr<tcu::TestCaseGroup> testGroup(new tcu::TestCaseGroup(testCtx, "loop", "Looping control flow"));
8202         RGBA defaultColors[4];
8203         getDefaultColors(defaultColors);
8204         map<string, string> fragments;
8205         fragments["pre_main"] =
8206                 "%c_f32_5 = OpConstant %f32 5.\n";
8207
8208         // A loop with a single block. The Continue Target is the loop block
8209         // itself. In SPIR-V terms, the "loop construct" contains no blocks at all
8210         // -- the "continue construct" forms the entire loop.
8211         fragments["testfun"] =
8212                 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
8213                 "%param1 = OpFunctionParameter %v4f32\n"
8214
8215                 "%entry = OpLabel\n"
8216                 "%val0 = OpVectorExtractDynamic %f32 %param1 %c_i32_0\n"
8217                 "OpBranch %loop\n"
8218
8219                 ";adds and subtracts 1.0 to %val in alternate iterations\n"
8220                 "%loop = OpLabel\n"
8221                 "%count = OpPhi %i32 %c_i32_4 %entry %count__ %loop\n"
8222                 "%delta = OpPhi %f32 %c_f32_1 %entry %minus_delta %loop\n"
8223                 "%val1 = OpPhi %f32 %val0 %entry %val %loop\n"
8224                 "%val = OpFAdd %f32 %val1 %delta\n"
8225                 "%minus_delta = OpFSub %f32 %c_f32_0 %delta\n"
8226                 "%count__ = OpISub %i32 %count %c_i32_1\n"
8227                 "%again = OpSGreaterThan %bool %count__ %c_i32_0\n"
8228                 "OpLoopMerge %exit %loop None\n"
8229                 "OpBranchConditional %again %loop %exit\n"
8230
8231                 "%exit = OpLabel\n"
8232                 "%result = OpVectorInsertDynamic %v4f32 %param1 %val %c_i32_0\n"
8233                 "OpReturnValue %result\n"
8234
8235                 "OpFunctionEnd\n";
8236
8237         createTestsForAllStages("single_block", defaultColors, defaultColors, fragments, testGroup.get());
8238
8239         // Body comprised of multiple basic blocks.
8240         const StringTemplate multiBlock(
8241                 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
8242                 "%param1 = OpFunctionParameter %v4f32\n"
8243
8244                 "%entry = OpLabel\n"
8245                 "%val0 = OpVectorExtractDynamic %f32 %param1 %c_i32_0\n"
8246                 "OpBranch %loop\n"
8247
8248                 ";adds and subtracts 1.0 to %val in alternate iterations\n"
8249                 "%loop = OpLabel\n"
8250                 "%count = OpPhi %i32 %c_i32_4 %entry %count__ %gather\n"
8251                 "%delta = OpPhi %f32 %c_f32_1 %entry %delta_next %gather\n"
8252                 "%val1 = OpPhi %f32 %val0 %entry %val %gather\n"
8253                 // There are several possibilities for the Continue Target below.  Each
8254                 // will be specialized into a separate test case.
8255                 "OpLoopMerge %exit ${continue_target} None\n"
8256                 "OpBranch %if\n"
8257
8258                 "%if = OpLabel\n"
8259                 ";delta_next = (delta > 0) ? -1 : 1;\n"
8260                 "%gt0 = OpFOrdGreaterThan %bool %delta %c_f32_0\n"
8261                 "OpSelectionMerge %gather DontFlatten\n"
8262                 "OpBranchConditional %gt0 %even %odd ;tells us if %count is even or odd\n"
8263
8264                 "%odd = OpLabel\n"
8265                 "OpBranch %gather\n"
8266
8267                 "%even = OpLabel\n"
8268                 "OpBranch %gather\n"
8269
8270                 "%gather = OpLabel\n"
8271                 "%delta_next = OpPhi %f32 %c_f32_n1 %even %c_f32_1 %odd\n"
8272                 "%val = OpFAdd %f32 %val1 %delta\n"
8273                 "%count__ = OpISub %i32 %count %c_i32_1\n"
8274                 "%again = OpSGreaterThan %bool %count__ %c_i32_0\n"
8275                 "OpBranchConditional %again %loop %exit\n"
8276
8277                 "%exit = OpLabel\n"
8278                 "%result = OpVectorInsertDynamic %v4f32 %param1 %val %c_i32_0\n"
8279                 "OpReturnValue %result\n"
8280
8281                 "OpFunctionEnd\n");
8282
8283         map<string, string> continue_target;
8284
8285         // The Continue Target is the loop block itself.
8286         continue_target["continue_target"] = "%loop";
8287         fragments["testfun"] = multiBlock.specialize(continue_target);
8288         createTestsForAllStages("multi_block_continue_construct", defaultColors, defaultColors, fragments, testGroup.get());
8289
8290         // The Continue Target is at the end of the loop.
8291         continue_target["continue_target"] = "%gather";
8292         fragments["testfun"] = multiBlock.specialize(continue_target);
8293         createTestsForAllStages("multi_block_loop_construct", defaultColors, defaultColors, fragments, testGroup.get());
8294
8295         // A loop with continue statement.
8296         fragments["testfun"] =
8297                 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
8298                 "%param1 = OpFunctionParameter %v4f32\n"
8299
8300                 "%entry = OpLabel\n"
8301                 "%val0 = OpVectorExtractDynamic %f32 %param1 %c_i32_0\n"
8302                 "OpBranch %loop\n"
8303
8304                 ";adds 4, 3, and 1 to %val0 (skips 2)\n"
8305                 "%loop = OpLabel\n"
8306                 "%count = OpPhi %i32 %c_i32_4 %entry %count__ %continue\n"
8307                 "%val1 = OpPhi %f32 %val0 %entry %val %continue\n"
8308                 "OpLoopMerge %exit %continue None\n"
8309                 "OpBranch %if\n"
8310
8311                 "%if = OpLabel\n"
8312                 ";skip if %count==2\n"
8313                 "%eq2 = OpIEqual %bool %count %c_i32_2\n"
8314                 "OpSelectionMerge %continue DontFlatten\n"
8315                 "OpBranchConditional %eq2 %continue %body\n"
8316
8317                 "%body = OpLabel\n"
8318                 "%fcount = OpConvertSToF %f32 %count\n"
8319                 "%val2 = OpFAdd %f32 %val1 %fcount\n"
8320                 "OpBranch %continue\n"
8321
8322                 "%continue = OpLabel\n"
8323                 "%val = OpPhi %f32 %val2 %body %val1 %if\n"
8324                 "%count__ = OpISub %i32 %count %c_i32_1\n"
8325                 "%again = OpSGreaterThan %bool %count__ %c_i32_0\n"
8326                 "OpBranchConditional %again %loop %exit\n"
8327
8328                 "%exit = OpLabel\n"
8329                 "%same = OpFSub %f32 %val %c_f32_8\n"
8330                 "%result = OpVectorInsertDynamic %v4f32 %param1 %same %c_i32_0\n"
8331                 "OpReturnValue %result\n"
8332                 "OpFunctionEnd\n";
8333         createTestsForAllStages("continue", defaultColors, defaultColors, fragments, testGroup.get());
8334
8335         // A loop with break.
8336         fragments["testfun"] =
8337                 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
8338                 "%param1 = OpFunctionParameter %v4f32\n"
8339
8340                 "%entry = OpLabel\n"
8341                 ";param1 components are between 0 and 1, so dot product is 4 or less\n"
8342                 "%dot = OpDot %f32 %param1 %param1\n"
8343                 "%div = OpFDiv %f32 %dot %c_f32_5\n"
8344                 "%zero = OpConvertFToU %u32 %div\n"
8345                 "%two = OpIAdd %i32 %zero %c_i32_2\n"
8346                 "%val0 = OpVectorExtractDynamic %f32 %param1 %c_i32_0\n"
8347                 "OpBranch %loop\n"
8348
8349                 ";adds 4 and 3 to %val0 (exits early)\n"
8350                 "%loop = OpLabel\n"
8351                 "%count = OpPhi %i32 %c_i32_4 %entry %count__ %continue\n"
8352                 "%val1 = OpPhi %f32 %val0 %entry %val2 %continue\n"
8353                 "OpLoopMerge %exit %continue None\n"
8354                 "OpBranch %if\n"
8355
8356                 "%if = OpLabel\n"
8357                 ";end loop if %count==%two\n"
8358                 "%above2 = OpSGreaterThan %bool %count %two\n"
8359                 "OpSelectionMerge %continue DontFlatten\n"
8360                 "OpBranchConditional %above2 %body %exit\n"
8361
8362                 "%body = OpLabel\n"
8363                 "%fcount = OpConvertSToF %f32 %count\n"
8364                 "%val2 = OpFAdd %f32 %val1 %fcount\n"
8365                 "OpBranch %continue\n"
8366
8367                 "%continue = OpLabel\n"
8368                 "%count__ = OpISub %i32 %count %c_i32_1\n"
8369                 "%again = OpSGreaterThan %bool %count__ %c_i32_0\n"
8370                 "OpBranchConditional %again %loop %exit\n"
8371
8372                 "%exit = OpLabel\n"
8373                 "%val_post = OpPhi %f32 %val2 %continue %val1 %if\n"
8374                 "%same = OpFSub %f32 %val_post %c_f32_7\n"
8375                 "%result = OpVectorInsertDynamic %v4f32 %param1 %same %c_i32_0\n"
8376                 "OpReturnValue %result\n"
8377                 "OpFunctionEnd\n";
8378         createTestsForAllStages("break", defaultColors, defaultColors, fragments, testGroup.get());
8379
8380         // A loop with return.
8381         fragments["testfun"] =
8382                 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
8383                 "%param1 = OpFunctionParameter %v4f32\n"
8384
8385                 "%entry = OpLabel\n"
8386                 ";param1 components are between 0 and 1, so dot product is 4 or less\n"
8387                 "%dot = OpDot %f32 %param1 %param1\n"
8388                 "%div = OpFDiv %f32 %dot %c_f32_5\n"
8389                 "%zero = OpConvertFToU %u32 %div\n"
8390                 "%two = OpIAdd %i32 %zero %c_i32_2\n"
8391                 "%val0 = OpVectorExtractDynamic %f32 %param1 %c_i32_0\n"
8392                 "OpBranch %loop\n"
8393
8394                 ";returns early without modifying %param1\n"
8395                 "%loop = OpLabel\n"
8396                 "%count = OpPhi %i32 %c_i32_4 %entry %count__ %continue\n"
8397                 "%val1 = OpPhi %f32 %val0 %entry %val2 %continue\n"
8398                 "OpLoopMerge %exit %continue None\n"
8399                 "OpBranch %if\n"
8400
8401                 "%if = OpLabel\n"
8402                 ";return if %count==%two\n"
8403                 "%above2 = OpSGreaterThan %bool %count %two\n"
8404                 "OpSelectionMerge %continue DontFlatten\n"
8405                 "OpBranchConditional %above2 %body %early_exit\n"
8406
8407                 "%early_exit = OpLabel\n"
8408                 "OpReturnValue %param1\n"
8409
8410                 "%body = OpLabel\n"
8411                 "%fcount = OpConvertSToF %f32 %count\n"
8412                 "%val2 = OpFAdd %f32 %val1 %fcount\n"
8413                 "OpBranch %continue\n"
8414
8415                 "%continue = OpLabel\n"
8416                 "%count__ = OpISub %i32 %count %c_i32_1\n"
8417                 "%again = OpSGreaterThan %bool %count__ %c_i32_0\n"
8418                 "OpBranchConditional %again %loop %exit\n"
8419
8420                 "%exit = OpLabel\n"
8421                 ";should never get here, so return an incorrect result\n"
8422                 "%result = OpVectorInsertDynamic %v4f32 %param1 %val2 %c_i32_0\n"
8423                 "OpReturnValue %result\n"
8424                 "OpFunctionEnd\n";
8425         createTestsForAllStages("return", defaultColors, defaultColors, fragments, testGroup.get());
8426
8427         // Continue inside a switch block to break to enclosing loop's merge block.
8428         // Matches roughly the following GLSL code:
8429         // for (; keep_going; keep_going = false)
8430         // {
8431         //     switch (int(param1.x))
8432         //     {
8433         //         case 0: continue;
8434         //         case 1: continue;
8435         //         default: continue;
8436         //     }
8437         //     dead code: modify return value to invalid result.
8438         // }
8439         fragments["pre_main"] =
8440                 "%fp_bool = OpTypePointer Function %bool\n"
8441                 "%true = OpConstantTrue %bool\n"
8442                 "%false = OpConstantFalse %bool\n";
8443
8444         fragments["testfun"] =
8445                 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
8446                 "%param1 = OpFunctionParameter %v4f32\n"
8447
8448                 "%entry = OpLabel\n"
8449                 "%keep_going = OpVariable %fp_bool Function\n"
8450                 "%val_ptr = OpVariable %fp_f32 Function\n"
8451                 "%param1_x = OpCompositeExtract %f32 %param1 0\n"
8452                 "OpStore %keep_going %true\n"
8453                 "OpBranch %forloop_begin\n"
8454
8455                 "%forloop_begin = OpLabel\n"
8456                 "OpLoopMerge %forloop_merge %forloop_continue None\n"
8457                 "OpBranch %forloop\n"
8458
8459                 "%forloop = OpLabel\n"
8460                 "%for_condition = OpLoad %bool %keep_going\n"
8461                 "OpBranchConditional %for_condition %forloop_body %forloop_merge\n"
8462
8463                 "%forloop_body = OpLabel\n"
8464                 "OpStore %val_ptr %param1_x\n"
8465                 "%param1_x_int = OpConvertFToS %i32 %param1_x\n"
8466
8467                 "OpSelectionMerge %switch_merge None\n"
8468                 "OpSwitch %param1_x_int %default 0 %case_0 1 %case_1\n"
8469                 "%case_0 = OpLabel\n"
8470                 "OpBranch %forloop_continue\n"
8471                 "%case_1 = OpLabel\n"
8472                 "OpBranch %forloop_continue\n"
8473                 "%default = OpLabel\n"
8474                 "OpBranch %forloop_continue\n"
8475                 "%switch_merge = OpLabel\n"
8476                 ";should never get here, so change the return value to invalid result\n"
8477                 "OpStore %val_ptr %c_f32_1\n"
8478                 "OpBranch %forloop_continue\n"
8479
8480                 "%forloop_continue = OpLabel\n"
8481                 "OpStore %keep_going %false\n"
8482                 "OpBranch %forloop_begin\n"
8483                 "%forloop_merge = OpLabel\n"
8484
8485                 "%val = OpLoad %f32 %val_ptr\n"
8486                 "%result = OpVectorInsertDynamic %v4f32 %param1 %val %c_i32_0\n"
8487                 "OpReturnValue %result\n"
8488                 "OpFunctionEnd\n";
8489         createTestsForAllStages("switch_continue", defaultColors, defaultColors, fragments, testGroup.get());
8490
8491         return testGroup.release();
8492 }
8493
8494 // A collection of tests putting OpControlBarrier in places GLSL forbids but SPIR-V allows.
8495 tcu::TestCaseGroup* createBarrierTests(tcu::TestContext& testCtx)
8496 {
8497         de::MovePtr<tcu::TestCaseGroup> testGroup(new tcu::TestCaseGroup(testCtx, "barrier", "OpControlBarrier"));
8498         map<string, string> fragments;
8499
8500         // A barrier inside a function body.
8501         fragments["pre_main"] =
8502                 "%Workgroup = OpConstant %i32 2\n"
8503                 "%WorkgroupAcquireRelease = OpConstant %i32 0x108\n";
8504         fragments["testfun"] =
8505                 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
8506                 "%param1 = OpFunctionParameter %v4f32\n"
8507                 "%label_testfun = OpLabel\n"
8508                 "OpControlBarrier %Workgroup %Workgroup %WorkgroupAcquireRelease\n"
8509                 "OpReturnValue %param1\n"
8510                 "OpFunctionEnd\n";
8511         addTessCtrlTest(testGroup.get(), "in_function", fragments);
8512
8513         // Common setup code for the following tests.
8514         fragments["pre_main"] =
8515                 "%Workgroup = OpConstant %i32 2\n"
8516                 "%WorkgroupAcquireRelease = OpConstant %i32 0x108\n"
8517                 "%c_f32_5 = OpConstant %f32 5.\n";
8518         const string setupPercentZero =  // Begins %test_code function with code that sets %zero to 0u but cannot be optimized away.
8519                 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
8520                 "%param1 = OpFunctionParameter %v4f32\n"
8521                 "%entry = OpLabel\n"
8522                 ";param1 components are between 0 and 1, so dot product is 4 or less\n"
8523                 "%dot = OpDot %f32 %param1 %param1\n"
8524                 "%div = OpFDiv %f32 %dot %c_f32_5\n"
8525                 "%zero = OpConvertFToU %u32 %div\n";
8526
8527         // Barriers inside OpSwitch branches.
8528         fragments["testfun"] =
8529                 setupPercentZero +
8530                 "OpSelectionMerge %switch_exit None\n"
8531                 "OpSwitch %zero %switch_default 0 %case0 1 %case1 ;should always go to %case0\n"
8532
8533                 "%case1 = OpLabel\n"
8534                 ";This barrier should never be executed, but its presence makes test failure more likely when there's a bug.\n"
8535                 "OpControlBarrier %Workgroup %Workgroup %WorkgroupAcquireRelease\n"
8536                 "%wrong_branch_alert1 = OpVectorInsertDynamic %v4f32 %param1 %c_f32_0_5 %c_i32_0\n"
8537                 "OpBranch %switch_exit\n"
8538
8539                 "%switch_default = OpLabel\n"
8540                 "%wrong_branch_alert2 = OpVectorInsertDynamic %v4f32 %param1 %c_f32_0_5 %c_i32_0\n"
8541                 ";This barrier should never be executed, but its presence makes test failure more likely when there's a bug.\n"
8542                 "OpControlBarrier %Workgroup %Workgroup %WorkgroupAcquireRelease\n"
8543                 "OpBranch %switch_exit\n"
8544
8545                 "%case0 = OpLabel\n"
8546                 "OpControlBarrier %Workgroup %Workgroup %WorkgroupAcquireRelease\n"
8547                 "OpBranch %switch_exit\n"
8548
8549                 "%switch_exit = OpLabel\n"
8550                 "%ret = OpPhi %v4f32 %param1 %case0 %wrong_branch_alert1 %case1 %wrong_branch_alert2 %switch_default\n"
8551                 "OpReturnValue %ret\n"
8552                 "OpFunctionEnd\n";
8553         addTessCtrlTest(testGroup.get(), "in_switch", fragments);
8554
8555         // Barriers inside if-then-else.
8556         fragments["testfun"] =
8557                 setupPercentZero +
8558                 "%eq0 = OpIEqual %bool %zero %c_u32_0\n"
8559                 "OpSelectionMerge %exit DontFlatten\n"
8560                 "OpBranchConditional %eq0 %then %else\n"
8561
8562                 "%else = OpLabel\n"
8563                 ";This barrier should never be executed, but its presence makes test failure more likely when there's a bug.\n"
8564                 "OpControlBarrier %Workgroup %Workgroup %WorkgroupAcquireRelease\n"
8565                 "%wrong_branch_alert = OpVectorInsertDynamic %v4f32 %param1 %c_f32_0_5 %c_i32_0\n"
8566                 "OpBranch %exit\n"
8567
8568                 "%then = OpLabel\n"
8569                 "OpControlBarrier %Workgroup %Workgroup %WorkgroupAcquireRelease\n"
8570                 "OpBranch %exit\n"
8571                 "%exit = OpLabel\n"
8572                 "%ret = OpPhi %v4f32 %param1 %then %wrong_branch_alert %else\n"
8573                 "OpReturnValue %ret\n"
8574                 "OpFunctionEnd\n";
8575         addTessCtrlTest(testGroup.get(), "in_if", fragments);
8576
8577         // A barrier after control-flow reconvergence, tempting the compiler to attempt something like this:
8578         // http://lists.llvm.org/pipermail/llvm-dev/2009-October/026317.html.
8579         fragments["testfun"] =
8580                 setupPercentZero +
8581                 "%thread_id = OpLoad %i32 %BP_gl_InvocationID\n"
8582                 "%thread0 = OpIEqual %bool %thread_id %c_i32_0\n"
8583                 "OpSelectionMerge %exit DontFlatten\n"
8584                 "OpBranchConditional %thread0 %then %else\n"
8585
8586                 "%else = OpLabel\n"
8587                 "%val0 = OpVectorExtractDynamic %f32 %param1 %c_i32_0\n"
8588                 "OpBranch %exit\n"
8589
8590                 "%then = OpLabel\n"
8591                 "%val1 = OpVectorExtractDynamic %f32 %param1 %zero\n"
8592                 "OpBranch %exit\n"
8593
8594                 "%exit = OpLabel\n"
8595                 "%val = OpPhi %f32 %val0 %else %val1 %then\n"
8596                 "OpControlBarrier %Workgroup %Workgroup %WorkgroupAcquireRelease\n"
8597                 "%ret = OpVectorInsertDynamic %v4f32 %param1 %val %zero\n"
8598                 "OpReturnValue %ret\n"
8599                 "OpFunctionEnd\n";
8600         addTessCtrlTest(testGroup.get(), "after_divergent_if", fragments);
8601
8602         // A barrier inside a loop.
8603         fragments["pre_main"] =
8604                 "%Workgroup = OpConstant %i32 2\n"
8605                 "%WorkgroupAcquireRelease = OpConstant %i32 0x108\n"
8606                 "%c_f32_10 = OpConstant %f32 10.\n";
8607         fragments["testfun"] =
8608                 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
8609                 "%param1 = OpFunctionParameter %v4f32\n"
8610                 "%entry = OpLabel\n"
8611                 "%val0 = OpVectorExtractDynamic %f32 %param1 %c_i32_0\n"
8612                 "OpBranch %loop\n"
8613
8614                 ";adds 4, 3, 2, and 1 to %val0\n"
8615                 "%loop = OpLabel\n"
8616                 "%count = OpPhi %i32 %c_i32_4 %entry %count__ %loop\n"
8617                 "%val1 = OpPhi %f32 %val0 %entry %val %loop\n"
8618                 "OpControlBarrier %Workgroup %Workgroup %WorkgroupAcquireRelease\n"
8619                 "%fcount = OpConvertSToF %f32 %count\n"
8620                 "%val = OpFAdd %f32 %val1 %fcount\n"
8621                 "%count__ = OpISub %i32 %count %c_i32_1\n"
8622                 "%again = OpSGreaterThan %bool %count__ %c_i32_0\n"
8623                 "OpLoopMerge %exit %loop None\n"
8624                 "OpBranchConditional %again %loop %exit\n"
8625
8626                 "%exit = OpLabel\n"
8627                 "%same = OpFSub %f32 %val %c_f32_10\n"
8628                 "%ret = OpVectorInsertDynamic %v4f32 %param1 %same %c_i32_0\n"
8629                 "OpReturnValue %ret\n"
8630                 "OpFunctionEnd\n";
8631         addTessCtrlTest(testGroup.get(), "in_loop", fragments);
8632
8633         return testGroup.release();
8634 }
8635
8636 // Test for the OpFRem instruction.
8637 tcu::TestCaseGroup* createFRemTests(tcu::TestContext& testCtx)
8638 {
8639         de::MovePtr<tcu::TestCaseGroup>         testGroup(new tcu::TestCaseGroup(testCtx, "frem", "OpFRem"));
8640         map<string, string>                                     fragments;
8641         RGBA                                                            inputColors[4];
8642         RGBA                                                            outputColors[4];
8643
8644         fragments["pre_main"]                            =
8645                 "%c_f32_3 = OpConstant %f32 3.0\n"
8646                 "%c_f32_n3 = OpConstant %f32 -3.0\n"
8647                 "%c_f32_4 = OpConstant %f32 4.0\n"
8648                 "%c_f32_p75 = OpConstant %f32 0.75\n"
8649                 "%c_v4f32_p75_p75_p75_p75 = OpConstantComposite %v4f32 %c_f32_p75 %c_f32_p75 %c_f32_p75 %c_f32_p75 \n"
8650                 "%c_v4f32_4_4_4_4 = OpConstantComposite %v4f32 %c_f32_4 %c_f32_4 %c_f32_4 %c_f32_4\n"
8651                 "%c_v4f32_3_n3_3_n3 = OpConstantComposite %v4f32 %c_f32_3 %c_f32_n3 %c_f32_3 %c_f32_n3\n";
8652
8653         // The test does the following.
8654         // vec4 result = (param1 * 8.0) - 4.0;
8655         // return (frem(result.x,3) + 0.75, frem(result.y, -3) + 0.75, 0, 1)
8656         fragments["testfun"]                             =
8657                 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
8658                 "%param1 = OpFunctionParameter %v4f32\n"
8659                 "%label_testfun = OpLabel\n"
8660                 "%v_times_8 = OpVectorTimesScalar %v4f32 %param1 %c_f32_8\n"
8661                 "%minus_4 = OpFSub %v4f32 %v_times_8 %c_v4f32_4_4_4_4\n"
8662                 "%frem = OpFRem %v4f32 %minus_4 %c_v4f32_3_n3_3_n3\n"
8663                 "%added = OpFAdd %v4f32 %frem %c_v4f32_p75_p75_p75_p75\n"
8664                 "%xyz_1 = OpVectorInsertDynamic %v4f32 %added %c_f32_1 %c_i32_3\n"
8665                 "%xy_0_1 = OpVectorInsertDynamic %v4f32 %xyz_1 %c_f32_0 %c_i32_2\n"
8666                 "OpReturnValue %xy_0_1\n"
8667                 "OpFunctionEnd\n";
8668
8669
8670         inputColors[0]          = RGBA(16,      16,             0, 255);
8671         inputColors[1]          = RGBA(232, 232,        0, 255);
8672         inputColors[2]          = RGBA(232, 16,         0, 255);
8673         inputColors[3]          = RGBA(16,      232,    0, 255);
8674
8675         outputColors[0]         = RGBA(64,      64,             0, 255);
8676         outputColors[1]         = RGBA(255, 255,        0, 255);
8677         outputColors[2]         = RGBA(255, 64,         0, 255);
8678         outputColors[3]         = RGBA(64,      255,    0, 255);
8679
8680         createTestsForAllStages("frem", inputColors, outputColors, fragments, testGroup.get());
8681         return testGroup.release();
8682 }
8683
8684 // Test for the OpSRem instruction.
8685 tcu::TestCaseGroup* createOpSRemGraphicsTests(tcu::TestContext& testCtx, qpTestResult negFailResult)
8686 {
8687         de::MovePtr<tcu::TestCaseGroup>         testGroup(new tcu::TestCaseGroup(testCtx, "srem", "OpSRem"));
8688         map<string, string>                                     fragments;
8689
8690         fragments["pre_main"]                            =
8691                 "%c_f32_255 = OpConstant %f32 255.0\n"
8692                 "%c_i32_128 = OpConstant %i32 128\n"
8693                 "%c_i32_255 = OpConstant %i32 255\n"
8694                 "%c_v4f32_255 = OpConstantComposite %v4f32 %c_f32_255 %c_f32_255 %c_f32_255 %c_f32_255 \n"
8695                 "%c_v4f32_0_5 = OpConstantComposite %v4f32 %c_f32_0_5 %c_f32_0_5 %c_f32_0_5 %c_f32_0_5 \n"
8696                 "%c_v4i32_128 = OpConstantComposite %v4i32 %c_i32_128 %c_i32_128 %c_i32_128 %c_i32_128 \n";
8697
8698         // The test does the following.
8699         // ivec4 ints = int(param1 * 255.0 + 0.5) - 128;
8700         // ivec4 result = ivec4(srem(ints.x, ints.y), srem(ints.y, ints.z), srem(ints.z, ints.x), 255);
8701         // return float(result + 128) / 255.0;
8702         fragments["testfun"]                             =
8703                 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
8704                 "%param1 = OpFunctionParameter %v4f32\n"
8705                 "%label_testfun = OpLabel\n"
8706                 "%div255 = OpFMul %v4f32 %param1 %c_v4f32_255\n"
8707                 "%add0_5 = OpFAdd %v4f32 %div255 %c_v4f32_0_5\n"
8708                 "%uints_in = OpConvertFToS %v4i32 %add0_5\n"
8709                 "%ints_in = OpISub %v4i32 %uints_in %c_v4i32_128\n"
8710                 "%x_in = OpCompositeExtract %i32 %ints_in 0\n"
8711                 "%y_in = OpCompositeExtract %i32 %ints_in 1\n"
8712                 "%z_in = OpCompositeExtract %i32 %ints_in 2\n"
8713                 "%x_out = OpSRem %i32 %x_in %y_in\n"
8714                 "%y_out = OpSRem %i32 %y_in %z_in\n"
8715                 "%z_out = OpSRem %i32 %z_in %x_in\n"
8716                 "%ints_out = OpCompositeConstruct %v4i32 %x_out %y_out %z_out %c_i32_255\n"
8717                 "%ints_offset = OpIAdd %v4i32 %ints_out %c_v4i32_128\n"
8718                 "%f_ints_offset = OpConvertSToF %v4f32 %ints_offset\n"
8719                 "%float_out = OpFDiv %v4f32 %f_ints_offset %c_v4f32_255\n"
8720                 "OpReturnValue %float_out\n"
8721                 "OpFunctionEnd\n";
8722
8723         const struct CaseParams
8724         {
8725                 const char*             name;
8726                 const char*             failMessageTemplate;    // customized status message
8727                 qpTestResult    failResult;                             // override status on failure
8728                 int                             operands[4][3];                 // four (x, y, z) vectors of operands
8729                 int                             results[4][3];                  // four (x, y, z) vectors of results
8730         } cases[] =
8731         {
8732                 {
8733                         "positive",
8734                         "${reason}",
8735                         QP_TEST_RESULT_FAIL,
8736                         { { 5, 12, 17 }, { 5, 5, 7 }, { 75, 8, 81 }, { 25, 60, 100 } },                 // operands
8737                         { { 5, 12,  2 }, { 0, 5, 2 }, {  3, 8,  6 }, { 25, 60,   0 } },                 // results
8738                 },
8739                 {
8740                         "all",
8741                         "Inconsistent results, but within specification: ${reason}",
8742                         negFailResult,                                                                                                                  // negative operands, not required by the spec
8743                         { { 5, 12, -17 }, { -5, -5, 7 }, { 75, 8, -81 }, { 25, -60, 100 } },    // operands
8744                         { { 5, 12,  -2 }, {  0, -5, 2 }, {  3, 8,  -6 }, { 25, -60,   0 } },    // results
8745                 },
8746         };
8747         // If either operand is negative the result is undefined. Some implementations may still return correct values.
8748
8749         for (int caseNdx = 0; caseNdx < DE_LENGTH_OF_ARRAY(cases); ++caseNdx)
8750         {
8751                 const CaseParams&       params                  = cases[caseNdx];
8752                 RGBA                            inputColors[4];
8753                 RGBA                            outputColors[4];
8754
8755                 for (int i = 0; i < 4; ++i)
8756                 {
8757                         inputColors [i] = RGBA(params.operands[i][0] + 128, params.operands[i][1] + 128, params.operands[i][2] + 128, 255);
8758                         outputColors[i] = RGBA(params.results [i][0] + 128, params.results [i][1] + 128, params.results [i][2] + 128, 255);
8759                 }
8760
8761                 createTestsForAllStages(params.name, inputColors, outputColors, fragments, testGroup.get(), params.failResult, params.failMessageTemplate);
8762         }
8763
8764         return testGroup.release();
8765 }
8766
8767 // Test for the OpSMod instruction.
8768 tcu::TestCaseGroup* createOpSModGraphicsTests(tcu::TestContext& testCtx, qpTestResult negFailResult)
8769 {
8770         de::MovePtr<tcu::TestCaseGroup>         testGroup(new tcu::TestCaseGroup(testCtx, "smod", "OpSMod"));
8771         map<string, string>                                     fragments;
8772
8773         fragments["pre_main"]                            =
8774                 "%c_f32_255 = OpConstant %f32 255.0\n"
8775                 "%c_i32_128 = OpConstant %i32 128\n"
8776                 "%c_i32_255 = OpConstant %i32 255\n"
8777                 "%c_v4f32_255 = OpConstantComposite %v4f32 %c_f32_255 %c_f32_255 %c_f32_255 %c_f32_255 \n"
8778                 "%c_v4f32_0_5 = OpConstantComposite %v4f32 %c_f32_0_5 %c_f32_0_5 %c_f32_0_5 %c_f32_0_5 \n"
8779                 "%c_v4i32_128 = OpConstantComposite %v4i32 %c_i32_128 %c_i32_128 %c_i32_128 %c_i32_128 \n";
8780
8781         // The test does the following.
8782         // ivec4 ints = int(param1 * 255.0 + 0.5) - 128;
8783         // ivec4 result = ivec4(smod(ints.x, ints.y), smod(ints.y, ints.z), smod(ints.z, ints.x), 255);
8784         // return float(result + 128) / 255.0;
8785         fragments["testfun"]                             =
8786                 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
8787                 "%param1 = OpFunctionParameter %v4f32\n"
8788                 "%label_testfun = OpLabel\n"
8789                 "%div255 = OpFMul %v4f32 %param1 %c_v4f32_255\n"
8790                 "%add0_5 = OpFAdd %v4f32 %div255 %c_v4f32_0_5\n"
8791                 "%uints_in = OpConvertFToS %v4i32 %add0_5\n"
8792                 "%ints_in = OpISub %v4i32 %uints_in %c_v4i32_128\n"
8793                 "%x_in = OpCompositeExtract %i32 %ints_in 0\n"
8794                 "%y_in = OpCompositeExtract %i32 %ints_in 1\n"
8795                 "%z_in = OpCompositeExtract %i32 %ints_in 2\n"
8796                 "%x_out = OpSMod %i32 %x_in %y_in\n"
8797                 "%y_out = OpSMod %i32 %y_in %z_in\n"
8798                 "%z_out = OpSMod %i32 %z_in %x_in\n"
8799                 "%ints_out = OpCompositeConstruct %v4i32 %x_out %y_out %z_out %c_i32_255\n"
8800                 "%ints_offset = OpIAdd %v4i32 %ints_out %c_v4i32_128\n"
8801                 "%f_ints_offset = OpConvertSToF %v4f32 %ints_offset\n"
8802                 "%float_out = OpFDiv %v4f32 %f_ints_offset %c_v4f32_255\n"
8803                 "OpReturnValue %float_out\n"
8804                 "OpFunctionEnd\n";
8805
8806         const struct CaseParams
8807         {
8808                 const char*             name;
8809                 const char*             failMessageTemplate;    // customized status message
8810                 qpTestResult    failResult;                             // override status on failure
8811                 int                             operands[4][3];                 // four (x, y, z) vectors of operands
8812                 int                             results[4][3];                  // four (x, y, z) vectors of results
8813         } cases[] =
8814         {
8815                 {
8816                         "positive",
8817                         "${reason}",
8818                         QP_TEST_RESULT_FAIL,
8819                         { { 5, 12, 17 }, { 5, 5, 7 }, { 75, 8, 81 }, { 25, 60, 100 } },                         // operands
8820                         { { 5, 12,  2 }, { 0, 5, 2 }, {  3, 8,  6 }, { 25, 60,   0 } },                         // results
8821                 },
8822                 {
8823                         "all",
8824                         "Inconsistent results, but within specification: ${reason}",
8825                         negFailResult,                                                                                                                          // negative operands, not required by the spec
8826                         { { 5, 12, -17 }, { -5, -5,  7 }, { 75,   8, -81 }, {  25, -60, 100 } },        // operands
8827                         { { 5, -5,   3 }, {  0,  2, -3 }, {  3, -73,  69 }, { -35,  40,   0 } },        // results
8828                 },
8829         };
8830         // If either operand is negative the result is undefined. Some implementations may still return correct values.
8831
8832         for (int caseNdx = 0; caseNdx < DE_LENGTH_OF_ARRAY(cases); ++caseNdx)
8833         {
8834                 const CaseParams&       params                  = cases[caseNdx];
8835                 RGBA                            inputColors[4];
8836                 RGBA                            outputColors[4];
8837
8838                 for (int i = 0; i < 4; ++i)
8839                 {
8840                         inputColors [i] = RGBA(params.operands[i][0] + 128, params.operands[i][1] + 128, params.operands[i][2] + 128, 255);
8841                         outputColors[i] = RGBA(params.results [i][0] + 128, params.results [i][1] + 128, params.results [i][2] + 128, 255);
8842                 }
8843
8844                 createTestsForAllStages(params.name, inputColors, outputColors, fragments, testGroup.get(), params.failResult, params.failMessageTemplate);
8845         }
8846         return testGroup.release();
8847 }
8848
8849 enum ConversionDataType
8850 {
8851         DATA_TYPE_SIGNED_8,
8852         DATA_TYPE_SIGNED_16,
8853         DATA_TYPE_SIGNED_32,
8854         DATA_TYPE_SIGNED_64,
8855         DATA_TYPE_UNSIGNED_8,
8856         DATA_TYPE_UNSIGNED_16,
8857         DATA_TYPE_UNSIGNED_32,
8858         DATA_TYPE_UNSIGNED_64,
8859         DATA_TYPE_FLOAT_16,
8860         DATA_TYPE_FLOAT_32,
8861         DATA_TYPE_FLOAT_64,
8862         DATA_TYPE_VEC2_SIGNED_16,
8863         DATA_TYPE_VEC2_SIGNED_32
8864 };
8865
8866 const string getBitWidthStr (ConversionDataType type)
8867 {
8868         switch (type)
8869         {
8870                 case DATA_TYPE_SIGNED_8:
8871                 case DATA_TYPE_UNSIGNED_8:
8872                         return "8";
8873
8874                 case DATA_TYPE_SIGNED_16:
8875                 case DATA_TYPE_UNSIGNED_16:
8876                 case DATA_TYPE_FLOAT_16:
8877                         return "16";
8878
8879                 case DATA_TYPE_SIGNED_32:
8880                 case DATA_TYPE_UNSIGNED_32:
8881                 case DATA_TYPE_FLOAT_32:
8882                 case DATA_TYPE_VEC2_SIGNED_16:
8883                         return "32";
8884
8885                 case DATA_TYPE_SIGNED_64:
8886                 case DATA_TYPE_UNSIGNED_64:
8887                 case DATA_TYPE_FLOAT_64:
8888                 case DATA_TYPE_VEC2_SIGNED_32:
8889                         return "64";
8890
8891                 default:
8892                         DE_ASSERT(false);
8893         }
8894         return "";
8895 }
8896
8897 const string getByteWidthStr (ConversionDataType type)
8898 {
8899         switch (type)
8900         {
8901                 case DATA_TYPE_SIGNED_8:
8902                 case DATA_TYPE_UNSIGNED_8:
8903                         return "1";
8904
8905                 case DATA_TYPE_SIGNED_16:
8906                 case DATA_TYPE_UNSIGNED_16:
8907                 case DATA_TYPE_FLOAT_16:
8908                         return "2";
8909
8910                 case DATA_TYPE_SIGNED_32:
8911                 case DATA_TYPE_UNSIGNED_32:
8912                 case DATA_TYPE_FLOAT_32:
8913                 case DATA_TYPE_VEC2_SIGNED_16:
8914                         return "4";
8915
8916                 case DATA_TYPE_SIGNED_64:
8917                 case DATA_TYPE_UNSIGNED_64:
8918                 case DATA_TYPE_FLOAT_64:
8919                 case DATA_TYPE_VEC2_SIGNED_32:
8920                         return "8";
8921
8922                 default:
8923                         DE_ASSERT(false);
8924         }
8925         return "";
8926 }
8927
8928 bool isSigned (ConversionDataType type)
8929 {
8930         switch (type)
8931         {
8932                 case DATA_TYPE_SIGNED_8:
8933                 case DATA_TYPE_SIGNED_16:
8934                 case DATA_TYPE_SIGNED_32:
8935                 case DATA_TYPE_SIGNED_64:
8936                 case DATA_TYPE_FLOAT_16:
8937                 case DATA_TYPE_FLOAT_32:
8938                 case DATA_TYPE_FLOAT_64:
8939                 case DATA_TYPE_VEC2_SIGNED_16:
8940                 case DATA_TYPE_VEC2_SIGNED_32:
8941                         return true;
8942
8943                 case DATA_TYPE_UNSIGNED_8:
8944                 case DATA_TYPE_UNSIGNED_16:
8945                 case DATA_TYPE_UNSIGNED_32:
8946                 case DATA_TYPE_UNSIGNED_64:
8947                         return false;
8948
8949                 default:
8950                         DE_ASSERT(false);
8951         }
8952         return false;
8953 }
8954
8955 bool isInt (ConversionDataType type)
8956 {
8957         switch (type)
8958         {
8959                 case DATA_TYPE_SIGNED_8:
8960                 case DATA_TYPE_SIGNED_16:
8961                 case DATA_TYPE_SIGNED_32:
8962                 case DATA_TYPE_SIGNED_64:
8963                 case DATA_TYPE_UNSIGNED_8:
8964                 case DATA_TYPE_UNSIGNED_16:
8965                 case DATA_TYPE_UNSIGNED_32:
8966                 case DATA_TYPE_UNSIGNED_64:
8967                         return true;
8968
8969                 case DATA_TYPE_FLOAT_16:
8970                 case DATA_TYPE_FLOAT_32:
8971                 case DATA_TYPE_FLOAT_64:
8972                 case DATA_TYPE_VEC2_SIGNED_16:
8973                 case DATA_TYPE_VEC2_SIGNED_32:
8974                         return false;
8975
8976                 default:
8977                         DE_ASSERT(false);
8978         }
8979         return false;
8980 }
8981
8982 bool isFloat (ConversionDataType type)
8983 {
8984         switch (type)
8985         {
8986                 case DATA_TYPE_SIGNED_8:
8987                 case DATA_TYPE_SIGNED_16:
8988                 case DATA_TYPE_SIGNED_32:
8989                 case DATA_TYPE_SIGNED_64:
8990                 case DATA_TYPE_UNSIGNED_8:
8991                 case DATA_TYPE_UNSIGNED_16:
8992                 case DATA_TYPE_UNSIGNED_32:
8993                 case DATA_TYPE_UNSIGNED_64:
8994                 case DATA_TYPE_VEC2_SIGNED_16:
8995                 case DATA_TYPE_VEC2_SIGNED_32:
8996                         return false;
8997
8998                 case DATA_TYPE_FLOAT_16:
8999                 case DATA_TYPE_FLOAT_32:
9000                 case DATA_TYPE_FLOAT_64:
9001                         return true;
9002
9003                 default:
9004                         DE_ASSERT(false);
9005         }
9006         return false;
9007 }
9008
9009 const string getTypeName (ConversionDataType type)
9010 {
9011         string prefix = isSigned(type) ? "" : "u";
9012
9013         if              (isInt(type))                                           return prefix + "int"   + getBitWidthStr(type);
9014         else if (isFloat(type))                                         return prefix + "float" + getBitWidthStr(type);
9015         else if (type == DATA_TYPE_VEC2_SIGNED_16)      return "i16vec2";
9016         else if (type == DATA_TYPE_VEC2_SIGNED_32)      return "i32vec2";
9017         else                                                                            DE_ASSERT(false);
9018
9019         return "";
9020 }
9021
9022 const string getTestName (ConversionDataType from, ConversionDataType to, const char* suffix)
9023 {
9024         const string fullSuffix(suffix == DE_NULL ? "" : string("_") + string(suffix));
9025
9026         return getTypeName(from) + "_to_" + getTypeName(to) + fullSuffix;
9027 }
9028
9029 const string getAsmTypeName (ConversionDataType type)
9030 {
9031         string prefix;
9032
9033         if              (isInt(type))                                           prefix = isSigned(type) ? "i" : "u";
9034         else if (isFloat(type))                                         prefix = "f";
9035         else if (type == DATA_TYPE_VEC2_SIGNED_16)      return "i16vec2";
9036         else if (type == DATA_TYPE_VEC2_SIGNED_32)      return "v2i32";
9037         else                                                                            DE_ASSERT(false);
9038
9039         return prefix + getBitWidthStr(type);
9040 }
9041
9042 template<typename T>
9043 BufferSp getSpecializedBuffer (deInt64 number)
9044 {
9045         return BufferSp(new Buffer<T>(vector<T>(1, (T)number)));
9046 }
9047
9048 BufferSp getBuffer (ConversionDataType type, deInt64 number)
9049 {
9050         switch (type)
9051         {
9052                 case DATA_TYPE_SIGNED_8:                return getSpecializedBuffer<deInt8>(number);
9053                 case DATA_TYPE_SIGNED_16:               return getSpecializedBuffer<deInt16>(number);
9054                 case DATA_TYPE_SIGNED_32:               return getSpecializedBuffer<deInt32>(number);
9055                 case DATA_TYPE_SIGNED_64:               return getSpecializedBuffer<deInt64>(number);
9056                 case DATA_TYPE_UNSIGNED_8:              return getSpecializedBuffer<deUint8>(number);
9057                 case DATA_TYPE_UNSIGNED_16:             return getSpecializedBuffer<deUint16>(number);
9058                 case DATA_TYPE_UNSIGNED_32:             return getSpecializedBuffer<deUint32>(number);
9059                 case DATA_TYPE_UNSIGNED_64:             return getSpecializedBuffer<deUint64>(number);
9060                 case DATA_TYPE_FLOAT_16:                return getSpecializedBuffer<deUint16>(number);
9061                 case DATA_TYPE_FLOAT_32:                return getSpecializedBuffer<deUint32>(number);
9062                 case DATA_TYPE_FLOAT_64:                return getSpecializedBuffer<deUint64>(number);
9063                 case DATA_TYPE_VEC2_SIGNED_16:  return getSpecializedBuffer<deUint32>(number);
9064                 case DATA_TYPE_VEC2_SIGNED_32:  return getSpecializedBuffer<deUint64>(number);
9065
9066                 default:                                                TCU_THROW(InternalError, "Unimplemented type passed");
9067         }
9068 }
9069
9070 bool usesInt8 (ConversionDataType from, ConversionDataType to)
9071 {
9072         return (from == DATA_TYPE_SIGNED_8 || to == DATA_TYPE_SIGNED_8 ||
9073                         from == DATA_TYPE_UNSIGNED_8 || to == DATA_TYPE_UNSIGNED_8);
9074 }
9075
9076 bool usesInt16 (ConversionDataType from, ConversionDataType to)
9077 {
9078         return (from == DATA_TYPE_SIGNED_16 || to == DATA_TYPE_SIGNED_16 ||
9079                         from == DATA_TYPE_UNSIGNED_16 || to == DATA_TYPE_UNSIGNED_16 ||
9080                         from == DATA_TYPE_VEC2_SIGNED_16 || to == DATA_TYPE_VEC2_SIGNED_16);
9081 }
9082
9083 bool usesInt32 (ConversionDataType from, ConversionDataType to)
9084 {
9085         return (from == DATA_TYPE_SIGNED_32 || to == DATA_TYPE_SIGNED_32 ||
9086                         from == DATA_TYPE_UNSIGNED_32 || to == DATA_TYPE_UNSIGNED_32 ||
9087                         from == DATA_TYPE_VEC2_SIGNED_32|| to == DATA_TYPE_VEC2_SIGNED_32);
9088 }
9089
9090 bool usesInt64 (ConversionDataType from, ConversionDataType to)
9091 {
9092         return (from == DATA_TYPE_SIGNED_64 || to == DATA_TYPE_SIGNED_64 ||
9093                         from == DATA_TYPE_UNSIGNED_64 || to == DATA_TYPE_UNSIGNED_64);
9094 }
9095
9096 bool usesFloat16 (ConversionDataType from, ConversionDataType to)
9097 {
9098         return (from == DATA_TYPE_FLOAT_16 || to == DATA_TYPE_FLOAT_16);
9099 }
9100
9101 bool usesFloat32 (ConversionDataType from, ConversionDataType to)
9102 {
9103         return (from == DATA_TYPE_FLOAT_32 || to == DATA_TYPE_FLOAT_32);
9104 }
9105
9106 bool usesFloat64 (ConversionDataType from, ConversionDataType to)
9107 {
9108         return (from == DATA_TYPE_FLOAT_64 || to == DATA_TYPE_FLOAT_64);
9109 }
9110
9111 void getVulkanFeaturesAndExtensions (ConversionDataType from, ConversionDataType to, VulkanFeatures& vulkanFeatures, vector<string>& extensions)
9112 {
9113         if (usesInt16(from, to) && !usesInt32(from, to))
9114                 vulkanFeatures.coreFeatures.shaderInt16 = DE_TRUE;
9115
9116         if (usesInt64(from, to))
9117                 vulkanFeatures.coreFeatures.shaderInt64 = DE_TRUE;
9118
9119         if (usesFloat64(from, to))
9120                 vulkanFeatures.coreFeatures.shaderFloat64 = DE_TRUE;
9121
9122         if (usesInt16(from, to) || usesFloat16(from, to))
9123         {
9124                 extensions.push_back("VK_KHR_16bit_storage");
9125                 vulkanFeatures.ext16BitStorage |= EXT16BITSTORAGEFEATURES_UNIFORM_BUFFER_BLOCK;
9126         }
9127
9128         if (usesFloat16(from, to) || usesInt8(from, to))
9129         {
9130                 extensions.push_back("VK_KHR_shader_float16_int8");
9131
9132                 if (usesFloat16(from, to))
9133                 {
9134                         vulkanFeatures.extFloat16Int8 |= EXTFLOAT16INT8FEATURES_FLOAT16;
9135                 }
9136
9137                 if (usesInt8(from, to))
9138                 {
9139                         vulkanFeatures.extFloat16Int8 |= EXTFLOAT16INT8FEATURES_INT8;
9140
9141                         extensions.push_back("VK_KHR_8bit_storage");
9142                         vulkanFeatures.ext8BitStorage |= EXT8BITSTORAGEFEATURES_STORAGE_BUFFER;
9143                 }
9144         }
9145 }
9146
9147 struct ConvertCase
9148 {
9149         ConvertCase (const string& instruction, ConversionDataType from, ConversionDataType to, deInt64 number, bool separateOutput = false, deInt64 outputNumber = 0, const char* suffix = DE_NULL)
9150         : m_fromType            (from)
9151         , m_toType                      (to)
9152         , m_name                        (getTestName(from, to, suffix))
9153         , m_inputBuffer         (getBuffer(from, number))
9154         {
9155                 string caps;
9156                 string decl;
9157                 string exts;
9158
9159                 m_asmTypes["inputType"]         = getAsmTypeName(from);
9160                 m_asmTypes["outputType"]        = getAsmTypeName(to);
9161
9162                 if (separateOutput)
9163                         m_outputBuffer = getBuffer(to, outputNumber);
9164                 else
9165                         m_outputBuffer = getBuffer(to, number);
9166
9167                 if (usesInt8(from, to))
9168                 {
9169                         bool requiresInt8Capability = true;
9170                         if (instruction == "OpUConvert" || instruction == "OpSConvert")
9171                         {
9172                                 // Conversions between 8 and 32 bit are provided by SPV_KHR_8bit_storage. The rest requires explicit Int8
9173                                 if (usesInt32(from, to))
9174                                         requiresInt8Capability = false;
9175                         }
9176
9177                         caps += "OpCapability StorageBuffer8BitAccess\n";
9178                         if (requiresInt8Capability)
9179                                 caps += "OpCapability Int8\n";
9180
9181                         decl += "%i8         = OpTypeInt 8 1\n"
9182                                         "%u8         = OpTypeInt 8 0\n";
9183                         exts += "OpExtension \"SPV_KHR_8bit_storage\"\n";
9184                 }
9185
9186                 if (usesInt16(from, to))
9187                 {
9188                         bool requiresInt16Capability = true;
9189
9190                         if (instruction == "OpUConvert" || instruction == "OpSConvert" || instruction == "OpFConvert")
9191                         {
9192                                 // Conversions between 16 and 32 bit are provided by SPV_KHR_16bit_storage. The rest requires explicit Int16
9193                                 if (usesInt32(from, to) || usesFloat32(from, to))
9194                                         requiresInt16Capability = false;
9195                         }
9196
9197                         decl += "%i16        = OpTypeInt 16 1\n"
9198                                         "%u16        = OpTypeInt 16 0\n"
9199                                         "%i16vec2    = OpTypeVector %i16 2\n";
9200
9201                         // Conversions between 16 and 32 bit are provided by SPV_KHR_16bit_storage. The rest requires explicit Int16
9202                         if (requiresInt16Capability)
9203                                 caps += "OpCapability Int16\n";
9204                 }
9205
9206                 if (usesFloat16(from, to))
9207                 {
9208                         decl += "%f16        = OpTypeFloat 16\n";
9209
9210                         // Conversions between 16 and 32 bit are provided by SPV_KHR_16bit_storage. The rest requires explicit Float16
9211                         if (!(usesInt32(from, to) || usesFloat32(from, to)))
9212                                 caps += "OpCapability Float16\n";
9213                 }
9214
9215                 if (usesInt16(from, to) || usesFloat16(from, to))
9216                 {
9217                         caps += "OpCapability StorageUniformBufferBlock16\n";
9218                         exts += "OpExtension \"SPV_KHR_16bit_storage\"\n";
9219                 }
9220
9221                 if (usesInt64(from, to))
9222                 {
9223                         caps += "OpCapability Int64\n";
9224                         decl += "%i64        = OpTypeInt 64 1\n"
9225                                         "%u64        = OpTypeInt 64 0\n";
9226                 }
9227
9228                 if (usesFloat64(from, to))
9229                 {
9230                         caps += "OpCapability Float64\n";
9231                         decl += "%f64        = OpTypeFloat 64\n";
9232                 }
9233
9234                 m_asmTypes["datatype_capabilities"]             = caps;
9235                 m_asmTypes["datatype_additional_decl"]  = decl;
9236                 m_asmTypes["datatype_extensions"]               = exts;
9237         }
9238
9239         ConversionDataType              m_fromType;
9240         ConversionDataType              m_toType;
9241         string                                  m_name;
9242         map<string, string>             m_asmTypes;
9243         BufferSp                                m_inputBuffer;
9244         BufferSp                                m_outputBuffer;
9245 };
9246
9247 const string getConvertCaseShaderStr (const string& instruction, const ConvertCase& convertCase)
9248 {
9249         map<string, string> params = convertCase.m_asmTypes;
9250
9251         params["instruction"]   = instruction;
9252         params["inDecorator"]   = getByteWidthStr(convertCase.m_fromType);
9253         params["outDecorator"]  = getByteWidthStr(convertCase.m_toType);
9254
9255         const StringTemplate shader (
9256                 "OpCapability Shader\n"
9257                 "${datatype_capabilities}"
9258                 "${datatype_extensions:opt}"
9259                 "OpMemoryModel Logical GLSL450\n"
9260                 "OpEntryPoint GLCompute %main \"main\"\n"
9261                 "OpExecutionMode %main LocalSize 1 1 1\n"
9262                 "OpSource GLSL 430\n"
9263                 "OpName %main           \"main\"\n"
9264                 // Decorators
9265                 "OpDecorate %indata DescriptorSet 0\n"
9266                 "OpDecorate %indata Binding 0\n"
9267                 "OpDecorate %outdata DescriptorSet 0\n"
9268                 "OpDecorate %outdata Binding 1\n"
9269                 "OpDecorate %in_buf BufferBlock\n"
9270                 "OpDecorate %out_buf BufferBlock\n"
9271                 "OpMemberDecorate %in_buf 0 Offset 0\n"
9272                 "OpMemberDecorate %out_buf 0 Offset 0\n"
9273                 // Base types
9274                 "%void       = OpTypeVoid\n"
9275                 "%voidf      = OpTypeFunction %void\n"
9276                 "%u32        = OpTypeInt 32 0\n"
9277                 "%i32        = OpTypeInt 32 1\n"
9278                 "%f32        = OpTypeFloat 32\n"
9279                 "%v2i32      = OpTypeVector %i32 2\n"
9280                 "${datatype_additional_decl}"
9281                 "%uvec3      = OpTypeVector %u32 3\n"
9282                 // Derived types
9283                 "%in_ptr     = OpTypePointer Uniform %${inputType}\n"
9284                 "%out_ptr    = OpTypePointer Uniform %${outputType}\n"
9285                 "%in_buf     = OpTypeStruct %${inputType}\n"
9286                 "%out_buf    = OpTypeStruct %${outputType}\n"
9287                 "%in_bufptr  = OpTypePointer Uniform %in_buf\n"
9288                 "%out_bufptr = OpTypePointer Uniform %out_buf\n"
9289                 "%indata     = OpVariable %in_bufptr Uniform\n"
9290                 "%outdata    = OpVariable %out_bufptr Uniform\n"
9291                 // Constants
9292                 "%zero       = OpConstant %i32 0\n"
9293                 // Main function
9294                 "%main       = OpFunction %void None %voidf\n"
9295                 "%label      = OpLabel\n"
9296                 "%inloc      = OpAccessChain %in_ptr %indata %zero\n"
9297                 "%outloc     = OpAccessChain %out_ptr %outdata %zero\n"
9298                 "%inval      = OpLoad %${inputType} %inloc\n"
9299                 "%conv       = ${instruction} %${outputType} %inval\n"
9300                 "              OpStore %outloc %conv\n"
9301                 "              OpReturn\n"
9302                 "              OpFunctionEnd\n"
9303         );
9304
9305         return shader.specialize(params);
9306 }
9307
9308 void createConvertCases (vector<ConvertCase>& testCases, const string& instruction)
9309 {
9310         if (instruction == "OpUConvert")
9311         {
9312                 // Convert unsigned int to unsigned int
9313                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_8,           DATA_TYPE_UNSIGNED_16,          42));
9314                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_8,           DATA_TYPE_UNSIGNED_32,          73));
9315                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_8,           DATA_TYPE_UNSIGNED_64,          121));
9316
9317                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_16,          DATA_TYPE_UNSIGNED_8,           33));
9318                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_16,          DATA_TYPE_UNSIGNED_32,          60653));
9319                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_16,          DATA_TYPE_UNSIGNED_64,          17991));
9320
9321                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_32,          DATA_TYPE_UNSIGNED_64,          904256275));
9322                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_32,          DATA_TYPE_UNSIGNED_16,          6275));
9323                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_32,          DATA_TYPE_UNSIGNED_8,           17));
9324
9325                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_64,          DATA_TYPE_UNSIGNED_32,          701256243));
9326                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_64,          DATA_TYPE_UNSIGNED_16,          4741));
9327                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_64,          DATA_TYPE_UNSIGNED_8,           65));
9328         }
9329         else if (instruction == "OpSConvert")
9330         {
9331                 // Sign extension int->int
9332                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_8,                     DATA_TYPE_SIGNED_16,            -30));
9333                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_8,                     DATA_TYPE_SIGNED_32,            55));
9334                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_8,                     DATA_TYPE_SIGNED_64,            -3));
9335                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_16,            DATA_TYPE_SIGNED_32,            14669));
9336                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_16,            DATA_TYPE_SIGNED_64,            -3341));
9337                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_32,            DATA_TYPE_SIGNED_64,            973610259));
9338
9339                 // Truncate for int->int
9340                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_16,            DATA_TYPE_SIGNED_8,                     81));
9341                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_32,            DATA_TYPE_SIGNED_8,                     -93));
9342                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_64,            DATA_TYPE_SIGNED_8,                     3182748172687672ll,                                     true,   56));
9343                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_32,            DATA_TYPE_SIGNED_16,            12382));
9344                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_64,            DATA_TYPE_SIGNED_32,            -972812359));
9345                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_64,            DATA_TYPE_SIGNED_16,            -1067742499291926803ll,                         true,   -4371));
9346
9347                 // Sign extension for int->uint
9348                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_8,                     DATA_TYPE_UNSIGNED_16,          56));
9349                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_8,                     DATA_TYPE_UNSIGNED_32,          -47,                                                            true,   4294967249u));
9350                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_8,                     DATA_TYPE_UNSIGNED_64,          -5,                                                                     true,   18446744073709551611ull));
9351                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_16,            DATA_TYPE_UNSIGNED_32,          14669));
9352                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_16,            DATA_TYPE_UNSIGNED_64,          -3341,                                                          true,   18446744073709548275ull));
9353                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_32,            DATA_TYPE_UNSIGNED_64,          973610259));
9354
9355                 // Truncate for int->uint
9356                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_16,            DATA_TYPE_UNSIGNED_8,           -25711,                                                         true,   145));
9357                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_32,            DATA_TYPE_UNSIGNED_8,           103));
9358                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_64,            DATA_TYPE_UNSIGNED_8,           -1067742499291926803ll,                         true,   61165));
9359                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_32,            DATA_TYPE_UNSIGNED_16,          12382));
9360                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_64,            DATA_TYPE_UNSIGNED_32,          -972812359,                                                     true,   3322154937u));
9361                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_64,            DATA_TYPE_UNSIGNED_16,          -1067742499291926803ll,                         true,   61165));
9362
9363                 // Sign extension for uint->int
9364                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_8,           DATA_TYPE_SIGNED_16,            71));
9365                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_8,           DATA_TYPE_SIGNED_32,            201,                                                            true,   -55));
9366                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_8,           DATA_TYPE_SIGNED_64,            188,                                                            true,   -68));
9367                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_16,          DATA_TYPE_SIGNED_32,            14669));
9368                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_16,          DATA_TYPE_SIGNED_64,            62195,                                                          true,   -3341));
9369                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_32,          DATA_TYPE_SIGNED_64,            973610259));
9370
9371                 // Truncate for uint->int
9372                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_16,          DATA_TYPE_SIGNED_8,                     67));
9373                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_32,          DATA_TYPE_SIGNED_8,                     133,                                                            true,   -123));
9374                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_64,          DATA_TYPE_SIGNED_8,                     836927654193256494ull,                          true,   46));
9375                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_32,          DATA_TYPE_SIGNED_16,            12382));
9376                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_64,          DATA_TYPE_SIGNED_32,            18446744072736739257ull,                        true,   -972812359));
9377                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_64,          DATA_TYPE_SIGNED_16,            17379001574417624813ull,                        true,   -4371));
9378
9379                 // Convert i16vec2 to i32vec2 and vice versa
9380                 // Unsigned values are used here to represent negative signed values and to allow defined shifting behaviour.
9381                 // The actual signed value -32123 is used here as uint16 value 33413 and uint32 value 4294935173
9382                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_VEC2_SIGNED_16,       DATA_TYPE_VEC2_SIGNED_32,       (33413u << 16)                  | 27593,        true,   (4294935173ull << 32)   | 27593));
9383                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_VEC2_SIGNED_32,       DATA_TYPE_VEC2_SIGNED_16,       (4294935173ull << 32)   | 27593,        true,   (33413u << 16)                  | 27593));
9384         }
9385         else if (instruction == "OpFConvert")
9386         {
9387                 // All hexadecimal values below represent 1234.0 as 16/32/64-bit IEEE 754 float
9388                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_32,                     DATA_TYPE_FLOAT_64,                     0x449a4000,                                                     true,   0x4093480000000000));
9389                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_64,                     DATA_TYPE_FLOAT_32,                     0x4093480000000000,                                     true,   0x449a4000));
9390
9391                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_32,                     DATA_TYPE_FLOAT_16,                     0x449a4000,                                                     true,   0x64D2));
9392                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_16,                     DATA_TYPE_FLOAT_32,                     0x64D2,                                                         true,   0x449a4000));
9393
9394                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_16,                     DATA_TYPE_FLOAT_64,                     0x64D2,                                                         true,   0x4093480000000000));
9395                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_64,                     DATA_TYPE_FLOAT_16,                     0x4093480000000000,                                     true,   0x64D2));
9396         }
9397         else if (instruction == "OpConvertFToU")
9398         {
9399                 // Normal numbers from uint8 range
9400                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_16,                     DATA_TYPE_UNSIGNED_8,           0x5020,                                                         true,   33,                                                                     "33"));
9401                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_32,                     DATA_TYPE_UNSIGNED_8,           0x42280000,                                                     true,   42,                                                                     "42"));
9402                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_64,                     DATA_TYPE_UNSIGNED_8,           0x4067800000000000ull,                          true,   188,                                                            "188"));
9403
9404                 // Maximum uint8 value
9405                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_16,                     DATA_TYPE_UNSIGNED_8,           0x5BF8,                                                         true,   255,                                                            "max"));
9406                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_32,                     DATA_TYPE_UNSIGNED_8,           0x437F0000,                                                     true,   255,                                                            "max"));
9407                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_64,                     DATA_TYPE_UNSIGNED_8,           0x406FE00000000000ull,                          true,   255,                                                            "max"));
9408
9409                 // +0
9410                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_16,                     DATA_TYPE_UNSIGNED_8,           0x0000,                                                         true,   0,                                                                      "p0"));
9411                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_32,                     DATA_TYPE_UNSIGNED_8,           0x00000000,                                                     true,   0,                                                                      "p0"));
9412                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_64,                     DATA_TYPE_UNSIGNED_8,           0x0000000000000000ull,                          true,   0,                                                                      "p0"));
9413
9414                 // -0
9415                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_16,                     DATA_TYPE_UNSIGNED_8,           0x8000,                                                         true,   0,                                                                      "m0"));
9416                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_32,                     DATA_TYPE_UNSIGNED_8,           0x80000000,                                                     true,   0,                                                                      "m0"));
9417                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_64,                     DATA_TYPE_UNSIGNED_8,           0x8000000000000000ull,                          true,   0,                                                                      "m0"));
9418
9419                 // All hexadecimal values below represent 1234.0 as 16/32/64-bit IEEE 754 float
9420                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_16,                     DATA_TYPE_UNSIGNED_16,          0x64D2,                                                         true,   1234,                                                           "1234"));
9421                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_16,                     DATA_TYPE_UNSIGNED_32,          0x64D2,                                                         true,   1234,                                                           "1234"));
9422                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_16,                     DATA_TYPE_UNSIGNED_64,          0x64D2,                                                         true,   1234,                                                           "1234"));
9423
9424                 // 0x7BFF = 0111 1011 1111 1111 = 0 11110 1111111111 = 65504
9425                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_16,                     DATA_TYPE_UNSIGNED_16,          0x7BFF,                                                         true,   65504,                                                          "max"));
9426                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_16,                     DATA_TYPE_UNSIGNED_32,          0x7BFF,                                                         true,   65504,                                                          "max"));
9427                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_16,                     DATA_TYPE_UNSIGNED_64,          0x7BFF,                                                         true,   65504,                                                          "max"));
9428
9429                 // +0
9430                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_16,                     DATA_TYPE_UNSIGNED_32,          0x0000,                                                         true,   0,                                                                      "p0"));
9431                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_16,                     DATA_TYPE_UNSIGNED_16,          0x0000,                                                         true,   0,                                                                      "p0"));
9432                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_16,                     DATA_TYPE_UNSIGNED_64,          0x0000,                                                         true,   0,                                                                      "p0"));
9433
9434                 // -0
9435                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_16,                     DATA_TYPE_UNSIGNED_16,          0x8000,                                                         true,   0,                                                                      "m0"));
9436                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_16,                     DATA_TYPE_UNSIGNED_32,          0x8000,                                                         true,   0,                                                                      "m0"));
9437                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_16,                     DATA_TYPE_UNSIGNED_64,          0x8000,                                                         true,   0,                                                                      "m0"));
9438
9439                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_32,                     DATA_TYPE_UNSIGNED_16,          0x449a4000,                                                     true,   1234));
9440                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_32,                     DATA_TYPE_UNSIGNED_32,          0x449a4000,                                                     true,   1234));
9441                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_32,                     DATA_TYPE_UNSIGNED_64,          0x449a4000,                                                     true,   1234));
9442                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_64,                     DATA_TYPE_UNSIGNED_16,          0x4093480000000000,                                     true,   1234));
9443                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_64,                     DATA_TYPE_UNSIGNED_32,          0x4093480000000000,                                     true,   1234));
9444                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_64,                     DATA_TYPE_UNSIGNED_64,          0x4093480000000000,                                     true,   1234));
9445         }
9446         else if (instruction == "OpConvertUToF")
9447         {
9448                 // Normal numbers from uint8 range
9449                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_8,           DATA_TYPE_FLOAT_16,                     116,                                                            true,   0x5740,                                                         "116"));
9450                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_8,           DATA_TYPE_FLOAT_32,                     232,                                                            true,   0x43680000,                                                     "232"));
9451                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_8,           DATA_TYPE_FLOAT_64,                     164,                                                            true,   0x4064800000000000ull,                          "164"));
9452
9453                 // Maximum uint8 value
9454                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_8,           DATA_TYPE_FLOAT_16,                     255,                                                            true,   0x5BF8,                                                         "max"));
9455                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_8,           DATA_TYPE_FLOAT_32,                     255,                                                            true,   0x437F0000,                                                     "max"));
9456                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_8,           DATA_TYPE_FLOAT_64,                     255,                                                            true,   0x406FE00000000000ull,                          "max"));
9457
9458                 // All hexadecimal values below represent 1234.0 as 32/64-bit IEEE 754 float
9459                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_16,          DATA_TYPE_FLOAT_16,                     1234,                                                           true,   0x64D2,                                                         "1234"));
9460                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_32,          DATA_TYPE_FLOAT_16,                     1234,                                                           true,   0x64D2,                                                         "1234"));
9461                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_64,          DATA_TYPE_FLOAT_16,                     1234,                                                           true,   0x64D2,                                                         "1234"));
9462
9463                 // 0x7BFF = 0111 1011 1111 1111 = 0 11110 1111111111 = 65504
9464                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_16,          DATA_TYPE_FLOAT_16,                     65504,                                                          true,   0x7BFF,                                                         "max"));
9465                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_32,          DATA_TYPE_FLOAT_16,                     65504,                                                          true,   0x7BFF,                                                         "max"));
9466                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_64,          DATA_TYPE_FLOAT_16,                     65504,                                                          true,   0x7BFF,                                                         "max"));
9467
9468                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_16,          DATA_TYPE_FLOAT_32,                     1234,                                                           true,   0x449a4000));
9469                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_16,          DATA_TYPE_FLOAT_64,                     1234,                                                           true,   0x4093480000000000));
9470                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_32,          DATA_TYPE_FLOAT_32,                     1234,                                                           true,   0x449a4000));
9471                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_32,          DATA_TYPE_FLOAT_64,                     1234,                                                           true,   0x4093480000000000));
9472                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_64,          DATA_TYPE_FLOAT_32,                     1234,                                                           true,   0x449a4000));
9473                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_64,          DATA_TYPE_FLOAT_64,                     1234,                                                           true,   0x4093480000000000));
9474         }
9475         else if (instruction == "OpConvertFToS")
9476         {
9477                 // Normal numbers from int8 range
9478                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_16,                     DATA_TYPE_SIGNED_8,                     0xC980,                                                         true,   -11,                                                            "m11"));
9479                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_32,                     DATA_TYPE_SIGNED_8,                     0xC2140000,                                                     true,   -37,                                                            "m37"));
9480                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_64,                     DATA_TYPE_SIGNED_8,                     0xC050800000000000ull,                          true,   -66,                                                            "m66"));
9481
9482                 // Minimum int8 value
9483                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_16,                     DATA_TYPE_SIGNED_8,                     0xD800,                                                         true,   -128,                                                           "min"));
9484                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_32,                     DATA_TYPE_SIGNED_8,                     0xC3000000,                                                     true,   -128,                                                           "min"));
9485                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_64,                     DATA_TYPE_SIGNED_8,                     0xC060000000000000ull,                          true,   -128,                                                           "min"));
9486
9487                 // Maximum int8 value
9488                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_16,                     DATA_TYPE_SIGNED_8,                     0x57F0,                                                         true,   127,                                                            "max"));
9489                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_32,                     DATA_TYPE_SIGNED_8,                     0x42FE0000,                                                     true,   127,                                                            "max"));
9490                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_64,                     DATA_TYPE_SIGNED_8,                     0x405FC00000000000ull,                          true,   127,                                                            "max"));
9491
9492                 // +0
9493                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_16,                     DATA_TYPE_SIGNED_8,                     0x0000,                                                         true,   0,                                                                      "p0"));
9494                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_32,                     DATA_TYPE_SIGNED_8,                     0x00000000,                                                     true,   0,                                                                      "p0"));
9495                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_64,                     DATA_TYPE_SIGNED_8,                     0x0000000000000000ull,                          true,   0,                                                                      "p0"));
9496
9497                 // -0
9498                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_16,                     DATA_TYPE_SIGNED_8,                     0x8000,                                                         true,   0,                                                                      "m0"));
9499                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_32,                     DATA_TYPE_SIGNED_8,                     0x80000000,                                                     true,   0,                                                                      "m0"));
9500                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_64,                     DATA_TYPE_SIGNED_8,                     0x8000000000000000ull,                          true,   0,                                                                      "m0"));
9501
9502                 // All hexadecimal values below represent -1234.0 as 32/64-bit IEEE 754 float
9503                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_16,                     DATA_TYPE_SIGNED_16,            0xE4D2,                                                         true,   -1234,                                                          "m1234"));
9504                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_16,                     DATA_TYPE_SIGNED_32,            0xE4D2,                                                         true,   -1234,                                                          "m1234"));
9505                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_16,                     DATA_TYPE_SIGNED_64,            0xE4D2,                                                         true,   -1234,                                                          "m1234"));
9506
9507                 // 0xF800 = 1111 1000 0000 0000 = 1 11110 0000000000 = -32768
9508                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_16,                     DATA_TYPE_SIGNED_16,            0xF800,                                                         true,   -32768,                                                         "min"));
9509                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_16,                     DATA_TYPE_SIGNED_32,            0xF800,                                                         true,   -32768,                                                         "min"));
9510                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_16,                     DATA_TYPE_SIGNED_64,            0xF800,                                                         true,   -32768,                                                         "min"));
9511
9512                 // 0x77FF = 0111 0111 1111 1111 = 0 11101 1111111111 = 32752
9513                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_16,                     DATA_TYPE_SIGNED_16,            0x77FF,                                                         true,   32752,                                                          "max"));
9514                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_16,                     DATA_TYPE_SIGNED_32,            0x77FF,                                                         true,   32752,                                                          "max"));
9515                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_16,                     DATA_TYPE_SIGNED_64,            0x77FF,                                                         true,   32752,                                                          "max"));
9516
9517                 // +0
9518                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_16,                     DATA_TYPE_SIGNED_16,            0x0000,                                                         true,   0,                                                                      "p0"));
9519                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_16,                     DATA_TYPE_SIGNED_32,            0x0000,                                                         true,   0,                                                                      "p0"));
9520                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_16,                     DATA_TYPE_SIGNED_64,            0x0000,                                                         true,   0,                                                                      "p0"));
9521
9522                 // -0
9523                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_16,                     DATA_TYPE_SIGNED_16,            0x8000,                                                         true,   0,                                                                      "m0"));
9524                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_16,                     DATA_TYPE_SIGNED_32,            0x8000,                                                         true,   0,                                                                      "m0"));
9525                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_16,                     DATA_TYPE_SIGNED_64,            0x8000,                                                         true,   0,                                                                      "m0"));
9526
9527                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_32,                     DATA_TYPE_SIGNED_16,            0xc49a4000,                                                     true,   -1234));
9528                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_32,                     DATA_TYPE_SIGNED_32,            0xc49a4000,                                                     true,   -1234));
9529                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_32,                     DATA_TYPE_SIGNED_64,            0xc49a4000,                                                     true,   -1234));
9530                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_64,                     DATA_TYPE_SIGNED_16,            0xc093480000000000,                                     true,   -1234));
9531                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_64,                     DATA_TYPE_SIGNED_32,            0xc093480000000000,                                     true,   -1234));
9532                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_64,                     DATA_TYPE_SIGNED_64,            0xc093480000000000,                                     true,   -1234));
9533                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_32,                     DATA_TYPE_SIGNED_16,            0x453b9000,                                                     true,    3001,                                                          "p3001"));
9534                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_32,                     DATA_TYPE_SIGNED_16,            0xc53b9000,                                                     true,   -3001,                                                          "m3001"));
9535         }
9536         else if (instruction == "OpConvertSToF")
9537         {
9538                 // Normal numbers from int8 range
9539                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_8,                     DATA_TYPE_FLOAT_16,                     -12,                                                            true,   0xCA00,                                                         "m21"));
9540                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_8,                     DATA_TYPE_FLOAT_32,                     -21,                                                            true,   0xC1A80000,                                                     "m21"));
9541                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_8,                     DATA_TYPE_FLOAT_64,                     -99,                                                            true,   0xC058C00000000000ull,                          "m99"));
9542
9543                 // Minimum int8 value
9544                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_8,                     DATA_TYPE_FLOAT_16,                     -128,                                                           true,   0xD800,                                                         "min"));
9545                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_8,                     DATA_TYPE_FLOAT_32,                     -128,                                                           true,   0xC3000000,                                                     "min"));
9546                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_8,                     DATA_TYPE_FLOAT_64,                     -128,                                                           true,   0xC060000000000000ull,                          "min"));
9547
9548                 // Maximum int8 value
9549                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_8,                     DATA_TYPE_FLOAT_16,                     127,                                                            true,   0x57F0,                                                         "max"));
9550                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_8,                     DATA_TYPE_FLOAT_32,                     127,                                                            true,   0x42FE0000,                                                     "max"));
9551                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_8,                     DATA_TYPE_FLOAT_64,                     127,                                                            true,   0x405FC00000000000ull,                          "max"));
9552
9553                 // All hexadecimal values below represent 1234.0 as 32/64-bit IEEE 754 float
9554                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_16,            DATA_TYPE_FLOAT_16,                     -1234,                                                          true,   0xE4D2,                                                         "m1234"));
9555                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_32,            DATA_TYPE_FLOAT_16,                     -1234,                                                          true,   0xE4D2,                                                         "m1234"));
9556                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_64,            DATA_TYPE_FLOAT_16,                     -1234,                                                          true,   0xE4D2,                                                         "m1234"));
9557
9558                 // 0xF800 = 1111 1000 0000 0000 = 1 11110 0000000000 = -32768
9559                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_16,            DATA_TYPE_FLOAT_16,                     -32768,                                                         true,   0xF800,                                                         "min"));
9560                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_32,            DATA_TYPE_FLOAT_16,                     -32768,                                                         true,   0xF800,                                                         "min"));
9561                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_64,            DATA_TYPE_FLOAT_16,                     -32768,                                                         true,   0xF800,                                                         "min"));
9562
9563                 // 0x77FF = 0111 0111 1111 1111 = 0 11101 1111111111 = 32752
9564                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_16,            DATA_TYPE_FLOAT_16,                     32752,                                                          true,   0x77FF,                                                         "max"));
9565                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_32,            DATA_TYPE_FLOAT_16,                     32752,                                                          true,   0x77FF,                                                         "max"));
9566                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_64,            DATA_TYPE_FLOAT_16,                     32752,                                                          true,   0x77FF,                                                         "max"));
9567
9568                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_16,            DATA_TYPE_FLOAT_32,                     -1234,                                                          true,   0xc49a4000));
9569                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_16,            DATA_TYPE_FLOAT_64,                     -1234,                                                          true,   0xc093480000000000));
9570                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_32,            DATA_TYPE_FLOAT_32,                     -1234,                                                          true,   0xc49a4000));
9571                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_32,            DATA_TYPE_FLOAT_64,                     -1234,                                                          true,   0xc093480000000000));
9572                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_64,            DATA_TYPE_FLOAT_32,                     -1234,                                                          true,   0xc49a4000));
9573                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_64,            DATA_TYPE_FLOAT_64,                     -1234,                                                          true,   0xc093480000000000));
9574         }
9575         else
9576                 DE_FATAL("Unknown instruction");
9577 }
9578
9579 const map<string, string> getConvertCaseFragments (string instruction, const ConvertCase& convertCase)
9580 {
9581         map<string, string> params = convertCase.m_asmTypes;
9582         map<string, string> fragments;
9583
9584         params["instruction"] = instruction;
9585         params["inDecorator"] = getByteWidthStr(convertCase.m_fromType);
9586
9587         const StringTemplate decoration (
9588                 "      OpDecorate %SSBOi DescriptorSet 0\n"
9589                 "      OpDecorate %SSBOo DescriptorSet 0\n"
9590                 "      OpDecorate %SSBOi Binding 0\n"
9591                 "      OpDecorate %SSBOo Binding 1\n"
9592                 "      OpDecorate %s_SSBOi Block\n"
9593                 "      OpDecorate %s_SSBOo Block\n"
9594                 "OpMemberDecorate %s_SSBOi 0 Offset 0\n"
9595                 "OpMemberDecorate %s_SSBOo 0 Offset 0\n");
9596
9597         const StringTemplate pre_main (
9598                 "${datatype_additional_decl:opt}"
9599                 "    %ptr_in = OpTypePointer StorageBuffer %${inputType}\n"
9600                 "   %ptr_out = OpTypePointer StorageBuffer %${outputType}\n"
9601                 "   %s_SSBOi = OpTypeStruct %${inputType}\n"
9602                 "   %s_SSBOo = OpTypeStruct %${outputType}\n"
9603                 " %ptr_SSBOi = OpTypePointer StorageBuffer %s_SSBOi\n"
9604                 " %ptr_SSBOo = OpTypePointer StorageBuffer %s_SSBOo\n"
9605                 "     %SSBOi = OpVariable %ptr_SSBOi StorageBuffer\n"
9606                 "     %SSBOo = OpVariable %ptr_SSBOo StorageBuffer\n");
9607
9608         const StringTemplate testfun (
9609                 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
9610                 "%param     = OpFunctionParameter %v4f32\n"
9611                 "%label     = OpLabel\n"
9612                 "%iLoc      = OpAccessChain %ptr_in %SSBOi %c_u32_0\n"
9613                 "%oLoc      = OpAccessChain %ptr_out %SSBOo %c_u32_0\n"
9614                 "%valIn     = OpLoad %${inputType} %iLoc\n"
9615                 "%valOut    = ${instruction} %${outputType} %valIn\n"
9616                 "             OpStore %oLoc %valOut\n"
9617                 "             OpReturnValue %param\n"
9618                 "             OpFunctionEnd\n");
9619
9620         params["datatype_extensions"] =
9621                 params["datatype_extensions"] +
9622                 "OpExtension \"SPV_KHR_storage_buffer_storage_class\"\n";
9623
9624         fragments["capability"] = params["datatype_capabilities"];
9625         fragments["extension"]  = params["datatype_extensions"];
9626         fragments["decoration"] = decoration.specialize(params);
9627         fragments["pre_main"]   = pre_main.specialize(params);
9628         fragments["testfun"]    = testfun.specialize(params);
9629
9630         return fragments;
9631 }
9632
9633 // Test for OpSConvert, OpUConvert, OpFConvert and OpConvert* in compute shaders
9634 tcu::TestCaseGroup* createConvertComputeTests (tcu::TestContext& testCtx, const string& instruction, const string& name)
9635 {
9636         de::MovePtr<tcu::TestCaseGroup>         group(new tcu::TestCaseGroup(testCtx, name.c_str(), instruction.c_str()));
9637         vector<ConvertCase>                                     testCases;
9638         createConvertCases(testCases, instruction);
9639
9640         for (vector<ConvertCase>::const_iterator test = testCases.begin(); test != testCases.end(); ++test)
9641         {
9642                 ComputeShaderSpec spec;
9643                 spec.assembly                   = getConvertCaseShaderStr(instruction, *test);
9644                 spec.numWorkGroups              = IVec3(1, 1, 1);
9645                 spec.inputs.push_back   (test->m_inputBuffer);
9646                 spec.outputs.push_back  (test->m_outputBuffer);
9647
9648                 getVulkanFeaturesAndExtensions(test->m_fromType, test->m_toType, spec.requestedVulkanFeatures, spec.extensions);
9649
9650                 group->addChild(new SpvAsmComputeShaderCase(testCtx, test->m_name.c_str(), "", spec));
9651         }
9652         return group.release();
9653 }
9654
9655 // Test for OpSConvert, OpUConvert, OpFConvert and OpConvert* in graphics shaders
9656 tcu::TestCaseGroup* createConvertGraphicsTests (tcu::TestContext& testCtx, const string& instruction, const string& name)
9657 {
9658         de::MovePtr<tcu::TestCaseGroup>         group(new tcu::TestCaseGroup(testCtx, name.c_str(), instruction.c_str()));
9659         vector<ConvertCase>                                     testCases;
9660         createConvertCases(testCases, instruction);
9661
9662         for (vector<ConvertCase>::const_iterator test = testCases.begin(); test != testCases.end(); ++test)
9663         {
9664                 map<string, string>     fragments               = getConvertCaseFragments(instruction, *test);
9665                 VulkanFeatures          vulkanFeatures;
9666                 GraphicsResources       resources;
9667                 vector<string>          extensions;
9668                 SpecConstants           noSpecConstants;
9669                 PushConstants           noPushConstants;
9670                 GraphicsInterfaces      noInterfaces;
9671                 tcu::RGBA                       defaultColors[4];
9672
9673                 getDefaultColors                        (defaultColors);
9674                 resources.inputs.push_back      (Resource(test->m_inputBuffer, VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
9675                 resources.outputs.push_back     (Resource(test->m_outputBuffer, VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
9676                 extensions.push_back            ("VK_KHR_storage_buffer_storage_class");
9677
9678                 getVulkanFeaturesAndExtensions(test->m_fromType, test->m_toType, vulkanFeatures, extensions);
9679
9680                 vulkanFeatures.coreFeatures.vertexPipelineStoresAndAtomics      = true;
9681                 vulkanFeatures.coreFeatures.fragmentStoresAndAtomics            = true;
9682
9683                 createTestsForAllStages(
9684                         test->m_name, defaultColors, defaultColors, fragments, noSpecConstants,
9685                         noPushConstants, resources, noInterfaces, extensions, vulkanFeatures, group.get());
9686         }
9687         return group.release();
9688 }
9689
9690 // Constant-Creation Instructions: OpConstant, OpConstantComposite
9691 tcu::TestCaseGroup* createOpConstantFloat16Tests(tcu::TestContext& testCtx)
9692 {
9693         de::MovePtr<tcu::TestCaseGroup> opConstantCompositeTests                (new tcu::TestCaseGroup(testCtx, "opconstant", "OpConstant and OpConstantComposite instruction"));
9694         RGBA                                                    inputColors[4];
9695         RGBA                                                    outputColors[4];
9696         vector<string>                                  extensions;
9697         GraphicsResources                               resources;
9698         VulkanFeatures                                  features;
9699
9700         const char                                              functionStart[]  =
9701                 "%test_code             = OpFunction %v4f32 None %v4f32_v4f32_function\n"
9702                 "%param1                = OpFunctionParameter %v4f32\n"
9703                 "%lbl                   = OpLabel\n";
9704
9705         const char                                              functionEnd[]           =
9706                 "%transformed_param_32  = OpFConvert %v4f32 %transformed_param\n"
9707                 "                         OpReturnValue %transformed_param_32\n"
9708                 "                         OpFunctionEnd\n";
9709
9710         struct NameConstantsCode
9711         {
9712                 string name;
9713                 string constants;
9714                 string code;
9715         };
9716
9717 #define FLOAT_16_COMMON_TYPES_AND_CONSTS \
9718                         "%f16                  = OpTypeFloat 16\n"                                                 \
9719                         "%c_f16_0              = OpConstant %f16 0.0\n"                                            \
9720                         "%c_f16_0_5            = OpConstant %f16 0.5\n"                                            \
9721                         "%c_f16_1              = OpConstant %f16 1.0\n"                                            \
9722                         "%v4f16                = OpTypeVector %f16 4\n"                                            \
9723                         "%fp_f16               = OpTypePointer Function %f16\n"                                    \
9724                         "%fp_v4f16             = OpTypePointer Function %v4f16\n"                                  \
9725                         "%c_v4f16_1_1_1_1      = OpConstantComposite %v4f16 %c_f16_1 %c_f16_1 %c_f16_1 %c_f16_1\n" \
9726                         "%a4f16                = OpTypeArray %f16 %c_u32_4\n"                                      \
9727
9728         NameConstantsCode                               tests[] =
9729         {
9730                 {
9731                         "vec4",
9732
9733                         FLOAT_16_COMMON_TYPES_AND_CONSTS
9734                         "%cval                 = OpConstantComposite %v4f16 %c_f16_0_5 %c_f16_0_5 %c_f16_0_5 %c_f16_0\n",
9735                         "%param1_16            = OpFConvert %v4f16 %param1\n"
9736                         "%transformed_param    = OpFAdd %v4f16 %param1_16 %cval\n"
9737                 },
9738                 {
9739                         "struct",
9740
9741                         FLOAT_16_COMMON_TYPES_AND_CONSTS
9742                         "%stype                = OpTypeStruct %v4f16 %f16\n"
9743                         "%fp_stype             = OpTypePointer Function %stype\n"
9744                         "%f16_n_1              = OpConstant %f16 -1.0\n"
9745                         "%f16_1_5              = OpConstant %f16 !0x3e00\n" // +1.5
9746                         "%cvec                 = OpConstantComposite %v4f16 %f16_1_5 %f16_1_5 %f16_1_5 %c_f16_1\n"
9747                         "%cval                 = OpConstantComposite %stype %cvec %f16_n_1\n",
9748
9749                         "%v                    = OpVariable %fp_stype Function %cval\n"
9750                         "%vec_ptr              = OpAccessChain %fp_v4f16 %v %c_u32_0\n"
9751                         "%f16_ptr              = OpAccessChain %fp_f16 %v %c_u32_1\n"
9752                         "%vec_val              = OpLoad %v4f16 %vec_ptr\n"
9753                         "%f16_val              = OpLoad %f16 %f16_ptr\n"
9754                         "%tmp1                 = OpVectorTimesScalar %v4f16 %c_v4f16_1_1_1_1 %f16_val\n" // vec4(-1)
9755                         "%param1_16            = OpFConvert %v4f16 %param1\n"
9756                         "%tmp2                 = OpFAdd %v4f16 %tmp1 %param1_16\n" // param1 + vec4(-1)
9757                         "%transformed_param    = OpFAdd %v4f16 %tmp2 %vec_val\n" // param1 + vec4(-1) + vec4(1.5, 1.5, 1.5, 1.0)
9758                 },
9759                 {
9760                         // [1|0|0|0.5] [x] = x + 0.5
9761                         // [0|1|0|0.5] [y] = y + 0.5
9762                         // [0|0|1|0.5] [z] = z + 0.5
9763                         // [0|0|0|1  ] [1] = 1
9764                         "matrix",
9765
9766                         FLOAT_16_COMMON_TYPES_AND_CONSTS
9767                         "%mat4x4_f16           = OpTypeMatrix %v4f16 4\n"
9768                         "%v4f16_1_0_0_0        = OpConstantComposite %v4f16 %c_f16_1 %c_f16_0 %c_f16_0 %c_f16_0\n"
9769                         "%v4f16_0_1_0_0        = OpConstantComposite %v4f16 %c_f16_0 %c_f16_1 %c_f16_0 %c_f16_0\n"
9770                         "%v4f16_0_0_1_0        = OpConstantComposite %v4f16 %c_f16_0 %c_f16_0 %c_f16_1 %c_f16_0\n"
9771                         "%v4f16_0_5_0_5_0_5_1  = OpConstantComposite %v4f16 %c_f16_0_5 %c_f16_0_5 %c_f16_0_5 %c_f16_1\n"
9772                         "%cval                 = OpConstantComposite %mat4x4_f16 %v4f16_1_0_0_0 %v4f16_0_1_0_0 %v4f16_0_0_1_0 %v4f16_0_5_0_5_0_5_1\n",
9773
9774                         "%param1_16            = OpFConvert %v4f16 %param1\n"
9775                         "%transformed_param    = OpMatrixTimesVector %v4f16 %cval %param1_16\n"
9776                 },
9777                 {
9778                         "array",
9779
9780                         FLOAT_16_COMMON_TYPES_AND_CONSTS
9781                         "%c_v4f16_1_1_1_0      = OpConstantComposite %v4f16 %c_f16_1 %c_f16_1 %c_f16_1 %c_f16_0\n"
9782                         "%fp_a4f16             = OpTypePointer Function %a4f16\n"
9783                         "%f16_n_1              = OpConstant %f16 -1.0\n"
9784                         "%f16_1_5              = OpConstant %f16 !0x3e00\n" // +1.5
9785                         "%carr                 = OpConstantComposite %a4f16 %c_f16_0 %f16_n_1 %f16_1_5 %c_f16_0\n",
9786
9787                         "%v                    = OpVariable %fp_a4f16 Function %carr\n"
9788                         "%f                    = OpAccessChain %fp_f16 %v %c_u32_0\n"
9789                         "%f1                   = OpAccessChain %fp_f16 %v %c_u32_1\n"
9790                         "%f2                   = OpAccessChain %fp_f16 %v %c_u32_2\n"
9791                         "%f3                   = OpAccessChain %fp_f16 %v %c_u32_3\n"
9792                         "%f_val                = OpLoad %f16 %f\n"
9793                         "%f1_val               = OpLoad %f16 %f1\n"
9794                         "%f2_val               = OpLoad %f16 %f2\n"
9795                         "%f3_val               = OpLoad %f16 %f3\n"
9796                         "%ftot1                = OpFAdd %f16 %f_val %f1_val\n"
9797                         "%ftot2                = OpFAdd %f16 %ftot1 %f2_val\n"
9798                         "%ftot3                = OpFAdd %f16 %ftot2 %f3_val\n"  // 0 - 1 + 1.5 + 0
9799                         "%add_vec              = OpVectorTimesScalar %v4f16 %c_v4f16_1_1_1_0 %ftot3\n"
9800                         "%param1_16            = OpFConvert %v4f16 %param1\n"
9801                         "%transformed_param    = OpFAdd %v4f16 %param1_16 %add_vec\n"
9802                 },
9803                 {
9804                         //
9805                         // [
9806                         //   {
9807                         //      0.0,
9808                         //      [ 1.0, 1.0, 1.0, 1.0]
9809                         //   },
9810                         //   {
9811                         //      1.0,
9812                         //      [ 0.0, 0.5, 0.0, 0.0]
9813                         //   }, //     ^^^
9814                         //   {
9815                         //      0.0,
9816                         //      [ 1.0, 1.0, 1.0, 1.0]
9817                         //   }
9818                         // ]
9819                         "array_of_struct_of_array",
9820
9821                         FLOAT_16_COMMON_TYPES_AND_CONSTS
9822                         "%c_v4f16_1_1_1_0      = OpConstantComposite %v4f16 %c_f16_1 %c_f16_1 %c_f16_1 %c_f16_0\n"
9823                         "%fp_a4f16             = OpTypePointer Function %a4f16\n"
9824                         "%stype                = OpTypeStruct %f16 %a4f16\n"
9825                         "%a3stype              = OpTypeArray %stype %c_u32_3\n"
9826                         "%fp_a3stype           = OpTypePointer Function %a3stype\n"
9827                         "%ca4f16_0             = OpConstantComposite %a4f16 %c_f16_0 %c_f16_0_5 %c_f16_0 %c_f16_0\n"
9828                         "%ca4f16_1             = OpConstantComposite %a4f16 %c_f16_1 %c_f16_1 %c_f16_1 %c_f16_1\n"
9829                         "%cstype1              = OpConstantComposite %stype %c_f16_0 %ca4f16_1\n"
9830                         "%cstype2              = OpConstantComposite %stype %c_f16_1 %ca4f16_0\n"
9831                         "%carr                 = OpConstantComposite %a3stype %cstype1 %cstype2 %cstype1",
9832
9833                         "%v                    = OpVariable %fp_a3stype Function %carr\n"
9834                         "%f                    = OpAccessChain %fp_f16 %v %c_u32_1 %c_u32_1 %c_u32_1\n"
9835                         "%f_l                  = OpLoad %f16 %f\n"
9836                         "%add_vec              = OpVectorTimesScalar %v4f16 %c_v4f16_1_1_1_0 %f_l\n"
9837                         "%param1_16            = OpFConvert %v4f16 %param1\n"
9838                         "%transformed_param    = OpFAdd %v4f16 %param1_16 %add_vec\n"
9839                 }
9840         };
9841
9842         getHalfColorsFullAlpha(inputColors);
9843         outputColors[0] = RGBA(255, 255, 255, 255);
9844         outputColors[1] = RGBA(255, 127, 127, 255);
9845         outputColors[2] = RGBA(127, 255, 127, 255);
9846         outputColors[3] = RGBA(127, 127, 255, 255);
9847
9848         extensions.push_back("VK_KHR_16bit_storage");
9849         extensions.push_back("VK_KHR_shader_float16_int8");
9850         features.extFloat16Int8 = EXTFLOAT16INT8FEATURES_FLOAT16;
9851
9852         for (size_t testNdx = 0; testNdx < sizeof(tests) / sizeof(NameConstantsCode); ++testNdx)
9853         {
9854                 map<string, string> fragments;
9855
9856                 fragments["extension"]  = "OpExtension \"SPV_KHR_16bit_storage\"";
9857                 fragments["capability"] = "OpCapability Float16\n";
9858                 fragments["pre_main"]   = tests[testNdx].constants;
9859                 fragments["testfun"]    = string(functionStart) + tests[testNdx].code + functionEnd;
9860
9861                 createTestsForAllStages(tests[testNdx].name, inputColors, outputColors, fragments, resources, extensions, opConstantCompositeTests.get(), features);
9862         }
9863         return opConstantCompositeTests.release();
9864 }
9865
9866 template<typename T>
9867 void finalizeTestsCreation (T&                                                  specResource,
9868                                                         const map<string, string>&      fragments,
9869                                                         tcu::TestContext&                       testCtx,
9870                                                         tcu::TestCaseGroup&                     testGroup,
9871                                                         const std::string&                      testName,
9872                                                         const VulkanFeatures&           vulkanFeatures,
9873                                                         const vector<string>&           extensions,
9874                                                         const IVec3&                            numWorkGroups);
9875
9876 template<>
9877 void finalizeTestsCreation (GraphicsResources&                  specResource,
9878                                                         const map<string, string>&      fragments,
9879                                                         tcu::TestContext&                       ,
9880                                                         tcu::TestCaseGroup&                     testGroup,
9881                                                         const std::string&                      testName,
9882                                                         const VulkanFeatures&           vulkanFeatures,
9883                                                         const vector<string>&           extensions,
9884                                                         const IVec3&                            )
9885 {
9886         RGBA defaultColors[4];
9887         getDefaultColors(defaultColors);
9888
9889         createTestsForAllStages(testName, defaultColors, defaultColors, fragments, specResource, extensions, &testGroup, vulkanFeatures);
9890 }
9891
9892 template<>
9893 void finalizeTestsCreation (ComputeShaderSpec&                  specResource,
9894                                                         const map<string, string>&      fragments,
9895                                                         tcu::TestContext&                       testCtx,
9896                                                         tcu::TestCaseGroup&                     testGroup,
9897                                                         const std::string&                      testName,
9898                                                         const VulkanFeatures&           vulkanFeatures,
9899                                                         const vector<string>&           extensions,
9900                                                         const IVec3&                            numWorkGroups)
9901 {
9902         specResource.numWorkGroups = numWorkGroups;
9903         specResource.requestedVulkanFeatures = vulkanFeatures;
9904         specResource.extensions = extensions;
9905
9906         specResource.assembly = makeComputeShaderAssembly(fragments);
9907
9908         testGroup.addChild(new SpvAsmComputeShaderCase(testCtx, testName.c_str(), "", specResource));
9909 }
9910
9911 template<class SpecResource>
9912 tcu::TestCaseGroup* createFloat16LogicalSet (tcu::TestContext& testCtx, const bool nanSupported)
9913 {
9914         const string                                            nan                                     = nanSupported ? "_nan" : "";
9915         const string                                            groupName                       = "logical" + nan;
9916         de::MovePtr<tcu::TestCaseGroup>         testGroup                       (new tcu::TestCaseGroup(testCtx, groupName.c_str(), "Float 16 logical tests"));
9917
9918         de::Random                                                      rnd                                     (deStringHash(testGroup->getName()));
9919         const string                                            spvCapabilities         = string("OpCapability StorageUniformBufferBlock16\n") + (nanSupported ? "OpCapability SignedZeroInfNanPreserve\n" : "");
9920         const string                                            spvExtensions           = string("OpExtension \"SPV_KHR_16bit_storage\"\n") + (nanSupported ? "OpExtension \"SPV_KHR_float_controls\"\n" : "");
9921         const string                                            spvExecutionMode        = nanSupported ? "OpExecutionMode %BP_main SignedZeroInfNanPreserve 16\n" : "";
9922         const deUint32                                          numDataPoints           = 16;
9923         const vector<deFloat16>                         float16Data                     = getFloat16s(rnd, numDataPoints);
9924         const vector<deFloat16>                         float16Data1            = squarize(float16Data, 0);
9925         const vector<deFloat16>                         float16Data2            = squarize(float16Data, 1);
9926         const vector<deFloat16>                         float16DataVec1         = squarizeVector(float16Data, 0);
9927         const vector<deFloat16>                         float16DataVec2         = squarizeVector(float16Data, 1);
9928         const vector<deFloat16>                         float16OutDummy         (float16Data1.size(), 0);
9929         const vector<deFloat16>                         float16OutVecDummy      (float16DataVec1.size(), 0);
9930
9931         struct TestOp
9932         {
9933                 const char*             opCode;
9934                 VerifyIOFunc    verifyFuncNan;
9935                 VerifyIOFunc    verifyFuncNonNan;
9936                 const deUint32  argCount;
9937         };
9938
9939         const TestOp    testOps[]       =
9940         {
9941                 { "OpIsNan"                                             ,       compareFP16Logical<fp16isNan,                           true,  false, true>,    compareFP16Logical<fp16isNan,                           true,  false, false>,   1       },
9942                 { "OpIsInf"                                             ,       compareFP16Logical<fp16isInf,                           true,  false, true>,    compareFP16Logical<fp16isInf,                           true,  false, false>,   1       },
9943                 { "OpFOrdEqual"                                 ,       compareFP16Logical<fp16isEqual,                         false, true,  true>,    compareFP16Logical<fp16isEqual,                         false, true,  false>,   2       },
9944                 { "OpFUnordEqual"                               ,       compareFP16Logical<fp16isEqual,                         false, false, true>,    compareFP16Logical<fp16isEqual,                         false, false, false>,   2       },
9945                 { "OpFOrdNotEqual"                              ,       compareFP16Logical<fp16isUnequal,                       false, true,  true>,    compareFP16Logical<fp16isUnequal,                       false, true,  false>,   2       },
9946                 { "OpFUnordNotEqual"                    ,       compareFP16Logical<fp16isUnequal,                       false, false, true>,    compareFP16Logical<fp16isUnequal,                       false, false, false>,   2       },
9947                 { "OpFOrdLessThan"                              ,       compareFP16Logical<fp16isLess,                          false, true,  true>,    compareFP16Logical<fp16isLess,                          false, true,  false>,   2       },
9948                 { "OpFUnordLessThan"                    ,       compareFP16Logical<fp16isLess,                          false, false, true>,    compareFP16Logical<fp16isLess,                          false, false, false>,   2       },
9949                 { "OpFOrdGreaterThan"                   ,       compareFP16Logical<fp16isGreater,                       false, true,  true>,    compareFP16Logical<fp16isGreater,                       false, true,  false>,   2       },
9950                 { "OpFUnordGreaterThan"                 ,       compareFP16Logical<fp16isGreater,                       false, false, true>,    compareFP16Logical<fp16isGreater,                       false, false, false>,   2       },
9951                 { "OpFOrdLessThanEqual"                 ,       compareFP16Logical<fp16isLessOrEqual,           false, true,  true>,    compareFP16Logical<fp16isLessOrEqual,           false, true,  false>,   2       },
9952                 { "OpFUnordLessThanEqual"               ,       compareFP16Logical<fp16isLessOrEqual,           false, false, true>,    compareFP16Logical<fp16isLessOrEqual,           false, false, false>,   2       },
9953                 { "OpFOrdGreaterThanEqual"              ,       compareFP16Logical<fp16isGreaterOrEqual,        false, true,  true>,    compareFP16Logical<fp16isGreaterOrEqual,        false, true,  false>,   2       },
9954                 { "OpFUnordGreaterThanEqual"    ,       compareFP16Logical<fp16isGreaterOrEqual,        false, false, true>,    compareFP16Logical<fp16isGreaterOrEqual,        false, false, false>,   2       },
9955         };
9956
9957         { // scalar cases
9958                 const StringTemplate preMain
9959                 (
9960                         "%c_i32_ndp = OpConstant %i32 ${num_data_points}\n"
9961                         "      %f16 = OpTypeFloat 16\n"
9962                         "  %c_f16_0 = OpConstant %f16 0.0\n"
9963                         "  %c_f16_1 = OpConstant %f16 1.0\n"
9964                         "   %up_f16 = OpTypePointer Uniform %f16\n"
9965                         "   %ra_f16 = OpTypeArray %f16 %c_i32_ndp\n"
9966                         "   %SSBO16 = OpTypeStruct %ra_f16\n"
9967                         "%up_SSBO16 = OpTypePointer Uniform %SSBO16\n"
9968                         "%ssbo_src0 = OpVariable %up_SSBO16 Uniform\n"
9969                         "%ssbo_src1 = OpVariable %up_SSBO16 Uniform\n"
9970                         " %ssbo_dst = OpVariable %up_SSBO16 Uniform\n"
9971                 );
9972
9973                 const StringTemplate decoration
9974                 (
9975                         "OpDecorate %ra_f16 ArrayStride 2\n"
9976                         "OpMemberDecorate %SSBO16 0 Offset 0\n"
9977                         "OpDecorate %SSBO16 BufferBlock\n"
9978                         "OpDecorate %ssbo_src0 DescriptorSet 0\n"
9979                         "OpDecorate %ssbo_src0 Binding 0\n"
9980                         "OpDecorate %ssbo_src1 DescriptorSet 0\n"
9981                         "OpDecorate %ssbo_src1 Binding 1\n"
9982                         "OpDecorate %ssbo_dst DescriptorSet 0\n"
9983                         "OpDecorate %ssbo_dst Binding 2\n"
9984                 );
9985
9986                 const StringTemplate testFun
9987                 (
9988                         "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
9989                         "    %param = OpFunctionParameter %v4f32\n"
9990
9991                         "    %entry = OpLabel\n"
9992                         "        %i = OpVariable %fp_i32 Function\n"
9993                         "             OpStore %i %c_i32_0\n"
9994                         "             OpBranch %loop\n"
9995
9996                         "     %loop = OpLabel\n"
9997                         "    %i_cmp = OpLoad %i32 %i\n"
9998                         "       %lt = OpSLessThan %bool %i_cmp %c_i32_ndp\n"
9999                         "             OpLoopMerge %merge %next None\n"
10000                         "             OpBranchConditional %lt %write %merge\n"
10001
10002                         "    %write = OpLabel\n"
10003                         "      %ndx = OpLoad %i32 %i\n"
10004
10005                         "     %src0 = OpAccessChain %up_f16 %ssbo_src0 %c_i32_0 %ndx\n"
10006                         " %val_src0 = OpLoad %f16 %src0\n"
10007
10008                         "${op_arg1_calc}"
10009
10010                         " %val_bdst = ${op_code} %bool %val_src0 ${op_arg1}\n"
10011                         "  %val_dst = OpSelect %f16 %val_bdst %c_f16_1 %c_f16_0\n"
10012                         "      %dst = OpAccessChain %up_f16 %ssbo_dst %c_i32_0 %ndx\n"
10013                         "             OpStore %dst %val_dst\n"
10014                         "             OpBranch %next\n"
10015
10016                         "     %next = OpLabel\n"
10017                         "    %i_cur = OpLoad %i32 %i\n"
10018                         "    %i_new = OpIAdd %i32 %i_cur %c_i32_1\n"
10019                         "             OpStore %i %i_new\n"
10020                         "             OpBranch %loop\n"
10021
10022                         "    %merge = OpLabel\n"
10023                         "             OpReturnValue %param\n"
10024
10025                         "             OpFunctionEnd\n"
10026                 );
10027
10028                 const StringTemplate arg1Calc
10029                 (
10030                         "     %src1 = OpAccessChain %up_f16 %ssbo_src1 %c_i32_0 %ndx\n"
10031                         " %val_src1 = OpLoad %f16 %src1\n"
10032                 );
10033
10034                 for (deUint32 testOpsIdx = 0; testOpsIdx < DE_LENGTH_OF_ARRAY(testOps); ++testOpsIdx)
10035                 {
10036                         const size_t            iterations              = float16Data1.size();
10037                         const TestOp&           testOp                  = testOps[testOpsIdx];
10038                         const string            testName                = de::toLower(string(testOp.opCode)) + "_scalar";
10039                         SpecResource            specResource;
10040                         map<string, string>     specs;
10041                         VulkanFeatures          features;
10042                         map<string, string>     fragments;
10043                         vector<string>          extensions;
10044
10045                         specs["num_data_points"]        = de::toString(iterations);
10046                         specs["op_code"]                        = testOp.opCode;
10047                         specs["op_arg1"]                        = (testOp.argCount == 1) ? "" : "%val_src1";
10048                         specs["op_arg1_calc"]           = (testOp.argCount == 1) ? "" : arg1Calc.specialize(specs);
10049
10050                         fragments["extension"]          = spvExtensions;
10051                         fragments["capability"]         = spvCapabilities;
10052                         fragments["execution_mode"]     = spvExecutionMode;
10053                         fragments["decoration"]         = decoration.specialize(specs);
10054                         fragments["pre_main"]           = preMain.specialize(specs);
10055                         fragments["testfun"]            = testFun.specialize(specs);
10056
10057                         specResource.inputs.push_back(Resource(BufferSp(new Float16Buffer(float16Data1)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
10058                         specResource.inputs.push_back(Resource(BufferSp(new Float16Buffer(float16Data2)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
10059                         specResource.outputs.push_back(Resource(BufferSp(new Float16Buffer(float16OutDummy)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
10060                         specResource.verifyIO = nanSupported ? testOp.verifyFuncNan : testOp.verifyFuncNonNan;
10061
10062                         extensions.push_back("VK_KHR_16bit_storage");
10063                         extensions.push_back("VK_KHR_shader_float16_int8");
10064
10065                         if (nanSupported)
10066                         {
10067                                 extensions.push_back("VK_KHR_shader_float_controls");
10068
10069                                 features.floatControlsProperties.shaderSignedZeroInfNanPreserveFloat16 = DE_TRUE;
10070                         }
10071
10072                         features.ext16BitStorage = EXT16BITSTORAGEFEATURES_UNIFORM_BUFFER_BLOCK;
10073                         features.extFloat16Int8 = EXTFLOAT16INT8FEATURES_FLOAT16;
10074
10075                         finalizeTestsCreation(specResource, fragments, testCtx, *testGroup.get(), testName, features, extensions, IVec3(1, 1, 1));
10076                 }
10077         }
10078         { // vector cases
10079                 const StringTemplate preMain
10080                 (
10081                         "  %c_i32_ndp = OpConstant %i32 ${num_data_points}\n"
10082                         "     %v2bool = OpTypeVector %bool 2\n"
10083                         "        %f16 = OpTypeFloat 16\n"
10084                         "    %c_f16_0 = OpConstant %f16 0.0\n"
10085                         "    %c_f16_1 = OpConstant %f16 1.0\n"
10086                         "      %v2f16 = OpTypeVector %f16 2\n"
10087                         "%c_v2f16_0_0 = OpConstantComposite %v2f16 %c_f16_0 %c_f16_0\n"
10088                         "%c_v2f16_1_1 = OpConstantComposite %v2f16 %c_f16_1 %c_f16_1\n"
10089                         "   %up_v2f16 = OpTypePointer Uniform %v2f16\n"
10090                         "   %ra_v2f16 = OpTypeArray %v2f16 %c_i32_ndp\n"
10091                         "     %SSBO16 = OpTypeStruct %ra_v2f16\n"
10092                         "  %up_SSBO16 = OpTypePointer Uniform %SSBO16\n"
10093                         "  %ssbo_src0 = OpVariable %up_SSBO16 Uniform\n"
10094                         "  %ssbo_src1 = OpVariable %up_SSBO16 Uniform\n"
10095                         "   %ssbo_dst = OpVariable %up_SSBO16 Uniform\n"
10096                 );
10097
10098                 const StringTemplate decoration
10099                 (
10100                         "OpDecorate %ra_v2f16 ArrayStride 4\n"
10101                         "OpMemberDecorate %SSBO16 0 Offset 0\n"
10102                         "OpDecorate %SSBO16 BufferBlock\n"
10103                         "OpDecorate %ssbo_src0 DescriptorSet 0\n"
10104                         "OpDecorate %ssbo_src0 Binding 0\n"
10105                         "OpDecorate %ssbo_src1 DescriptorSet 0\n"
10106                         "OpDecorate %ssbo_src1 Binding 1\n"
10107                         "OpDecorate %ssbo_dst DescriptorSet 0\n"
10108                         "OpDecorate %ssbo_dst Binding 2\n"
10109                 );
10110
10111                 const StringTemplate testFun
10112                 (
10113                         "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
10114                         "    %param = OpFunctionParameter %v4f32\n"
10115
10116                         "    %entry = OpLabel\n"
10117                         "        %i = OpVariable %fp_i32 Function\n"
10118                         "             OpStore %i %c_i32_0\n"
10119                         "             OpBranch %loop\n"
10120
10121                         "     %loop = OpLabel\n"
10122                         "    %i_cmp = OpLoad %i32 %i\n"
10123                         "       %lt = OpSLessThan %bool %i_cmp %c_i32_ndp\n"
10124                         "             OpLoopMerge %merge %next None\n"
10125                         "             OpBranchConditional %lt %write %merge\n"
10126
10127                         "    %write = OpLabel\n"
10128                         "      %ndx = OpLoad %i32 %i\n"
10129
10130                         "     %src0 = OpAccessChain %up_v2f16 %ssbo_src0 %c_i32_0 %ndx\n"
10131                         " %val_src0 = OpLoad %v2f16 %src0\n"
10132
10133                         "${op_arg1_calc}"
10134
10135                         " %val_bdst = ${op_code} %v2bool %val_src0 ${op_arg1}\n"
10136                         "  %val_dst = OpSelect %v2f16 %val_bdst %c_v2f16_1_1 %c_v2f16_0_0\n"
10137                         "      %dst = OpAccessChain %up_v2f16 %ssbo_dst %c_i32_0 %ndx\n"
10138                         "             OpStore %dst %val_dst\n"
10139                         "             OpBranch %next\n"
10140
10141                         "     %next = OpLabel\n"
10142                         "    %i_cur = OpLoad %i32 %i\n"
10143                         "    %i_new = OpIAdd %i32 %i_cur %c_i32_1\n"
10144                         "             OpStore %i %i_new\n"
10145                         "             OpBranch %loop\n"
10146
10147                         "    %merge = OpLabel\n"
10148                         "             OpReturnValue %param\n"
10149
10150                         "             OpFunctionEnd\n"
10151                 );
10152
10153                 const StringTemplate arg1Calc
10154                 (
10155                         "     %src1 = OpAccessChain %up_v2f16 %ssbo_src1 %c_i32_0 %ndx\n"
10156                         " %val_src1 = OpLoad %v2f16 %src1\n"
10157                 );
10158
10159                 for (deUint32 testOpsIdx = 0; testOpsIdx < DE_LENGTH_OF_ARRAY(testOps); ++testOpsIdx)
10160                 {
10161                         const deUint32          itemsPerVec     = 2;
10162                         const size_t            iterations      = float16DataVec1.size() / itemsPerVec;
10163                         const TestOp&           testOp          = testOps[testOpsIdx];
10164                         const string            testName        = de::toLower(string(testOp.opCode)) + "_vector";
10165                         SpecResource            specResource;
10166                         map<string, string>     specs;
10167                         vector<string>          extensions;
10168                         VulkanFeatures          features;
10169                         map<string, string>     fragments;
10170
10171                         specs["num_data_points"]        = de::toString(iterations);
10172                         specs["op_code"]                        = testOp.opCode;
10173                         specs["op_arg1"]                        = (testOp.argCount == 1) ? "" : "%val_src1";
10174                         specs["op_arg1_calc"]           = (testOp.argCount == 1) ? "" : arg1Calc.specialize(specs);
10175
10176                         fragments["extension"]          = spvExtensions;
10177                         fragments["capability"]         = spvCapabilities;
10178                         fragments["execution_mode"]     = spvExecutionMode;
10179                         fragments["decoration"]         = decoration.specialize(specs);
10180                         fragments["pre_main"]           = preMain.specialize(specs);
10181                         fragments["testfun"]            = testFun.specialize(specs);
10182
10183                         specResource.inputs.push_back(Resource(BufferSp(new Float16Buffer(float16DataVec1)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
10184                         specResource.inputs.push_back(Resource(BufferSp(new Float16Buffer(float16DataVec2)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
10185                         specResource.outputs.push_back(Resource(BufferSp(new Float16Buffer(float16OutVecDummy)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
10186                         specResource.verifyIO = nanSupported ? testOp.verifyFuncNan : testOp.verifyFuncNonNan;
10187
10188                         extensions.push_back("VK_KHR_16bit_storage");
10189                         extensions.push_back("VK_KHR_shader_float16_int8");
10190
10191                         if (nanSupported)
10192                         {
10193                                 extensions.push_back("VK_KHR_shader_float_controls");
10194
10195                                 features.floatControlsProperties.shaderSignedZeroInfNanPreserveFloat16 = DE_TRUE;
10196                         }
10197
10198                         features.ext16BitStorage = EXT16BITSTORAGEFEATURES_UNIFORM_BUFFER_BLOCK;
10199                         features.extFloat16Int8 = EXTFLOAT16INT8FEATURES_FLOAT16;
10200
10201                         finalizeTestsCreation(specResource, fragments, testCtx, *testGroup.get(), testName, features, extensions, IVec3(1, 1, 1));
10202                 }
10203         }
10204
10205         return testGroup.release();
10206 }
10207
10208 bool compareFP16FunctionSetFunc (const std::vector<Resource>& inputs, const vector<AllocationSp>& outputAllocs, const std::vector<Resource>&, TestLog& log)
10209 {
10210         if (inputs.size() != 1 || outputAllocs.size() != 1)
10211                 return false;
10212
10213         vector<deUint8> input1Bytes;
10214
10215         inputs[0].getBytes(input1Bytes);
10216
10217         const deUint16* const   input1AsFP16    = (const deUint16*)&input1Bytes[0];
10218         const deUint16* const   outputAsFP16    = (const deUint16*)outputAllocs[0]->getHostPtr();
10219         std::string                             error;
10220
10221         for (size_t idx = 0; idx < input1Bytes.size() / sizeof(deUint16); ++idx)
10222         {
10223                 if (!compare16BitFloat(input1AsFP16[idx], outputAsFP16[idx], error))
10224                 {
10225                         log << TestLog::Message << error << TestLog::EndMessage;
10226
10227                         return false;
10228                 }
10229         }
10230
10231         return true;
10232 }
10233
10234 template<class SpecResource>
10235 tcu::TestCaseGroup* createFloat16FuncSet (tcu::TestContext& testCtx)
10236 {
10237         de::MovePtr<tcu::TestCaseGroup>         testGroup                       (new tcu::TestCaseGroup(testCtx, "function", "Float 16 function call related tests"));
10238
10239         de::Random                                                      rnd                                     (deStringHash(testGroup->getName()));
10240         const StringTemplate                            capabilities            ("OpCapability ${cap}\n");
10241         const deUint32                                          numDataPoints           = 256;
10242         const vector<deFloat16>                         float16InputData        = getFloat16s(rnd, numDataPoints);
10243         const vector<deFloat16>                         float16OutputDummy      (float16InputData.size(), 0);
10244         map<string, string>                                     fragments;
10245
10246         struct TestType
10247         {
10248                 const deUint32  typeComponents;
10249                 const char*             typeName;
10250                 const char*             typeDecls;
10251         };
10252
10253         const TestType  testTypes[]     =
10254         {
10255                 {
10256                         1,
10257                         "f16",
10258                         ""
10259                 },
10260                 {
10261                         2,
10262                         "v2f16",
10263                         "      %v2f16 = OpTypeVector %f16 2\n"
10264                         "  %c_v2f16_0 = OpConstantComposite %v2f16 %c_f16_0 %c_f16_0\n"
10265                 },
10266                 {
10267                         4,
10268                         "v4f16",
10269                         "      %v4f16 = OpTypeVector %f16 4\n"
10270                         "  %c_v4f16_0 = OpConstantComposite %v4f16 %c_f16_0 %c_f16_0 %c_f16_0 %c_f16_0\n"
10271                 },
10272         };
10273
10274         const StringTemplate preMain
10275         (
10276                 "  %c_i32_ndp = OpConstant %i32 ${num_data_points}\n"
10277                 "     %v2bool = OpTypeVector %bool 2\n"
10278                 "        %f16 = OpTypeFloat 16\n"
10279                 "    %c_f16_0 = OpConstant %f16 0.0\n"
10280
10281                 "${type_decls}"
10282
10283                 "  %${tt}_fun = OpTypeFunction %${tt} %${tt}\n"
10284                 "   %up_${tt} = OpTypePointer Uniform %${tt}\n"
10285                 "   %ra_${tt} = OpTypeArray %${tt} %c_i32_ndp\n"
10286                 "     %SSBO16 = OpTypeStruct %ra_${tt}\n"
10287                 "  %up_SSBO16 = OpTypePointer Uniform %SSBO16\n"
10288                 "   %ssbo_src = OpVariable %up_SSBO16 Uniform\n"
10289                 "   %ssbo_dst = OpVariable %up_SSBO16 Uniform\n"
10290         );
10291
10292         const StringTemplate decoration
10293         (
10294                 "OpDecorate %ra_${tt} ArrayStride ${tt_stride}\n"
10295                 "OpMemberDecorate %SSBO16 0 Offset 0\n"
10296                 "OpDecorate %SSBO16 BufferBlock\n"
10297                 "OpDecorate %ssbo_src DescriptorSet 0\n"
10298                 "OpDecorate %ssbo_src Binding 0\n"
10299                 "OpDecorate %ssbo_dst DescriptorSet 0\n"
10300                 "OpDecorate %ssbo_dst Binding 1\n"
10301         );
10302
10303         const StringTemplate testFun
10304         (
10305                 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
10306                 "    %param = OpFunctionParameter %v4f32\n"
10307                 "    %entry = OpLabel\n"
10308
10309                 "        %i = OpVariable %fp_i32 Function\n"
10310                 "             OpStore %i %c_i32_0\n"
10311                 "             OpBranch %loop\n"
10312
10313                 "     %loop = OpLabel\n"
10314                 "    %i_cmp = OpLoad %i32 %i\n"
10315                 "       %lt = OpSLessThan %bool %i_cmp %c_i32_ndp\n"
10316                 "             OpLoopMerge %merge %next None\n"
10317                 "             OpBranchConditional %lt %write %merge\n"
10318
10319                 "    %write = OpLabel\n"
10320                 "      %ndx = OpLoad %i32 %i\n"
10321
10322                 "      %src = OpAccessChain %up_${tt} %ssbo_src %c_i32_0 %ndx\n"
10323                 "  %val_src = OpLoad %${tt} %src\n"
10324
10325                 "  %val_dst = OpFunctionCall %${tt} %pass_fun %val_src\n"
10326                 "      %dst = OpAccessChain %up_${tt} %ssbo_dst %c_i32_0 %ndx\n"
10327                 "             OpStore %dst %val_dst\n"
10328                 "             OpBranch %next\n"
10329
10330                 "     %next = OpLabel\n"
10331                 "    %i_cur = OpLoad %i32 %i\n"
10332                 "    %i_new = OpIAdd %i32 %i_cur %c_i32_1\n"
10333                 "             OpStore %i %i_new\n"
10334                 "             OpBranch %loop\n"
10335
10336                 "    %merge = OpLabel\n"
10337                 "             OpReturnValue %param\n"
10338
10339                 "             OpFunctionEnd\n"
10340
10341                 " %pass_fun = OpFunction %${tt} None %${tt}_fun\n"
10342                 "   %param0 = OpFunctionParameter %${tt}\n"
10343                 " %entry_pf = OpLabel\n"
10344                 "     %res0 = OpFAdd %${tt} %param0 %c_${tt}_0\n"
10345                 "             OpReturnValue %res0\n"
10346                 "             OpFunctionEnd\n"
10347         );
10348
10349         for (deUint32 testTypeIdx = 0; testTypeIdx < DE_LENGTH_OF_ARRAY(testTypes); ++testTypeIdx)
10350         {
10351                 const TestType&         testType                = testTypes[testTypeIdx];
10352                 const string            testName                = testType.typeName;
10353                 const deUint32          itemsPerType    = testType.typeComponents;
10354                 const size_t            iterations              = float16InputData.size() / itemsPerType;
10355                 const size_t            typeStride              = itemsPerType * sizeof(deFloat16);
10356                 SpecResource            specResource;
10357                 map<string, string>     specs;
10358                 VulkanFeatures          features;
10359                 vector<string>          extensions;
10360
10361                 specs["cap"]                            = "StorageUniformBufferBlock16";
10362                 specs["num_data_points"]        = de::toString(iterations);
10363                 specs["tt"]                                     = testType.typeName;
10364                 specs["tt_stride"]                      = de::toString(typeStride);
10365                 specs["type_decls"]                     = testType.typeDecls;
10366
10367                 fragments["extension"]          = "OpExtension \"SPV_KHR_16bit_storage\"";
10368                 fragments["capability"]         = capabilities.specialize(specs);
10369                 fragments["decoration"]         = decoration.specialize(specs);
10370                 fragments["pre_main"]           = preMain.specialize(specs);
10371                 fragments["testfun"]            = testFun.specialize(specs);
10372
10373                 specResource.inputs.push_back(Resource(BufferSp(new Float16Buffer(float16InputData)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
10374                 specResource.outputs.push_back(Resource(BufferSp(new Float16Buffer(float16OutputDummy)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
10375                 specResource.verifyIO = compareFP16FunctionSetFunc;
10376
10377                 extensions.push_back("VK_KHR_16bit_storage");
10378                 extensions.push_back("VK_KHR_shader_float16_int8");
10379
10380                 features.ext16BitStorage = EXT16BITSTORAGEFEATURES_UNIFORM_BUFFER_BLOCK;
10381                 features.extFloat16Int8 = EXTFLOAT16INT8FEATURES_FLOAT16;
10382
10383                 finalizeTestsCreation(specResource, fragments, testCtx, *testGroup.get(), testName, features, extensions, IVec3(1, 1, 1));
10384         }
10385
10386         return testGroup.release();
10387 }
10388
10389 struct getV_    { deUint32 inline operator()(deUint32 v) const  { return v;        } getV_(){} };
10390 struct getV0    { deUint32 inline operator()(deUint32 v) const  { return v & (~1); } getV0(){} };
10391 struct getV1    { deUint32 inline operator()(deUint32 v) const  { return v | ( 1); } getV1(){} };
10392
10393 template<deUint32 R, deUint32 N>
10394 inline static deUint32 getOffset(deUint32 x, deUint32 y, deUint32 n)
10395 {
10396         return N * ((R * y) + x) + n;
10397 }
10398
10399 template<deUint32 R, deUint32 N, class X0, class X1, class Y0, class Y1>
10400 struct getFDelta
10401 {
10402         float operator() (const deFloat16* data, deUint32 x, deUint32 y, deUint32 n, deUint32 flavor) const
10403         {
10404                 DE_STATIC_ASSERT(R%2 == 0);
10405                 DE_ASSERT(flavor == 0);
10406                 DE_UNREF(flavor);
10407
10408                 const X0                        x0;
10409                 const X1                        x1;
10410                 const Y0                        y0;
10411                 const Y1                        y1;
10412                 const deFloat16         v0      = data[getOffset<R, N>(x0(x), y0(y), n)];
10413                 const deFloat16         v1      = data[getOffset<R, N>(x1(x), y1(y), n)];
10414                 const tcu::Float16      f0      = tcu::Float16(v0);
10415                 const tcu::Float16      f1      = tcu::Float16(v1);
10416                 const float                     d0      = f0.asFloat();
10417                 const float                     d1      = f1.asFloat();
10418                 const float                     d       = d1 - d0;
10419
10420                 return d;
10421         }
10422
10423         getFDelta(){}
10424 };
10425
10426 template<deUint32 F, class Class0, class Class1>
10427 struct getFOneOf
10428 {
10429         float operator() (const deFloat16* data, deUint32 x, deUint32 y, deUint32 n, deUint32 flavor) const
10430         {
10431                 DE_ASSERT(flavor < F);
10432
10433                 if (flavor == 0)
10434                 {
10435                         Class0 c;
10436
10437                         return c(data, x, y, n, flavor);
10438                 }
10439                 else
10440                 {
10441                         Class1 c;
10442
10443                         return c(data, x, y, n, flavor - 1);
10444                 }
10445         }
10446
10447         getFOneOf(){}
10448 };
10449
10450 template<class FineX0, class FineX1, class FineY0, class FineY1>
10451 struct calcWidthOf4
10452 {
10453         float operator() (const deFloat16* data, deUint32 x, deUint32 y, deUint32 n, deUint32 flavor) const
10454         {
10455                 DE_ASSERT(flavor < 4);
10456
10457                 const deUint32                                          flavorX = (flavor & 1) == 0 ? 0 : 1;
10458                 const deUint32                                          flavorY = (flavor & 2) == 0 ? 0 : 1;
10459                 const getFOneOf<2, FineX0, FineX1>      cx;
10460                 const getFOneOf<2, FineY0, FineY1>      cy;
10461                 float                                                           v               = 0;
10462
10463                 v += fabsf(cx(data, x, y, n, flavorX));
10464                 v += fabsf(cy(data, x, y, n, flavorY));
10465
10466                 return v;
10467         }
10468
10469         calcWidthOf4(){}
10470 };
10471
10472 template<deUint32 R, deUint32 N, class Derivative>
10473 bool compareDerivativeWithFlavor (const deFloat16* inputAsFP16, const deFloat16* outputAsFP16, deUint32 flavor, std::string& error)
10474 {
10475         const deUint32          numDataPointsByAxis     = R;
10476         const Derivative        derivativeFunc;
10477
10478         for (deUint32 y = 0; y < numDataPointsByAxis; ++y)
10479         for (deUint32 x = 0; x < numDataPointsByAxis; ++x)
10480         for (deUint32 n = 0; n < N; ++n)
10481         {
10482                 const float             expectedFloat   = derivativeFunc(inputAsFP16, x, y, n, flavor);
10483                 deFloat16               expected                = deFloat32To16Round(expectedFloat, DE_ROUNDINGMODE_TO_NEAREST_EVEN);
10484                 const deFloat16 output                  = outputAsFP16[getOffset<R, N>(x, y, n)];
10485
10486                 bool                    reportError             = !compare16BitFloat(expected, output, error);
10487
10488                 if (reportError)
10489                 {
10490                         expected        = deFloat32To16Round(expectedFloat, DE_ROUNDINGMODE_TO_ZERO);
10491                         reportError     = !compare16BitFloat(expected, output, error);
10492                 }
10493
10494                 if (reportError)
10495                 {
10496                         error = "subcase at " + de::toString(x) + "," + de::toString(y) + "," + de::toString(n) + ": " + error;
10497
10498                         return false;
10499                 }
10500         }
10501
10502         return true;
10503 }
10504
10505 template<deUint32 R, deUint32 N, deUint32 FLAVOUR_COUNT, class Derivative>
10506 bool compareDerivative (const std::vector<Resource>& inputs, const vector<AllocationSp>& outputAllocs, const std::vector<Resource>&, TestLog& log)
10507 {
10508         if (inputs.size() != 1 || outputAllocs.size() != 1)
10509                 return false;
10510
10511         deUint32                        successfulRuns                  = FLAVOUR_COUNT;
10512         std::string                     results[FLAVOUR_COUNT];
10513         vector<deUint8>         inputBytes;
10514
10515         inputs[0].getBytes(inputBytes);
10516
10517         const deFloat16*        inputAsFP16             = reinterpret_cast<deFloat16* const>(&inputBytes.front());
10518         const deFloat16*        outputAsFP16    = static_cast<deFloat16*>(outputAllocs[0]->getHostPtr());
10519
10520         DE_ASSERT(inputBytes.size() ==  R * R * N * sizeof(deFloat16));
10521
10522         for (deUint32 flavor = 0; flavor < FLAVOUR_COUNT; ++flavor)
10523                 if (compareDerivativeWithFlavor<R, N, Derivative> (inputAsFP16, outputAsFP16, flavor, results[flavor]))
10524                 {
10525                         break;
10526                 }
10527                 else
10528                 {
10529                         successfulRuns--;
10530                 }
10531
10532         if (successfulRuns == 0)
10533                 for (deUint32 flavor = 0; flavor < FLAVOUR_COUNT; flavor++)
10534                         log << TestLog::Message << "At flavor #" << flavor << " " << results[flavor] << TestLog::EndMessage;
10535
10536         return successfulRuns > 0;
10537 }
10538
10539 template<deUint32 R, deUint32 N>
10540 tcu::TestCaseGroup* createDerivativeTests (tcu::TestContext& testCtx)
10541 {
10542         typedef getFDelta<R, N, getV0, getV1, getV_, getV_> getFDxFine;
10543         typedef getFDelta<R, N, getV_, getV_, getV0, getV1> getFDyFine;
10544
10545         typedef getFDelta<R, N, getV0, getV1, getV0, getV0> getFdxCoarse0;
10546         typedef getFDelta<R, N, getV0, getV1, getV1, getV1> getFdxCoarse1;
10547         typedef getFDelta<R, N, getV0, getV0, getV0, getV1> getFdyCoarse0;
10548         typedef getFDelta<R, N, getV1, getV1, getV0, getV1> getFdyCoarse1;
10549         typedef getFOneOf<2, getFdxCoarse0, getFdxCoarse1> getFDxCoarse;
10550         typedef getFOneOf<2, getFdyCoarse0, getFdyCoarse1> getFDyCoarse;
10551
10552         typedef calcWidthOf4<getFDxFine, getFDxFine, getFDyFine, getFDyFine> getFWidthFine;
10553         typedef calcWidthOf4<getFdxCoarse0, getFdxCoarse1, getFdyCoarse0, getFdyCoarse1> getFWidthCoarse;
10554
10555         typedef getFOneOf<3, getFDxFine, getFDxCoarse> getFDx;
10556         typedef getFOneOf<3, getFDyFine, getFDyCoarse> getFDy;
10557         typedef getFOneOf<5, getFWidthFine, getFWidthCoarse> getFWidth;
10558
10559         const std::string                                       testGroupName           (std::string("derivative_") + de::toString(N));
10560         de::MovePtr<tcu::TestCaseGroup>         testGroup                       (new tcu::TestCaseGroup(testCtx, testGroupName.c_str(), "Derivative instruction tests"));
10561
10562         de::Random                                                      rnd                                     (deStringHash(testGroup->getName()));
10563         const deUint32                                          numDataPointsByAxis     = R;
10564         const deUint32                                          numDataPoints           = N * numDataPointsByAxis * numDataPointsByAxis;
10565         vector<deFloat16>                                       float16InputX;
10566         vector<deFloat16>                                       float16InputY;
10567         vector<deFloat16>                                       float16InputW;
10568         vector<deFloat16>                                       float16OutputDummy      (numDataPoints, 0);
10569         RGBA                                                            defaultColors[4];
10570
10571         getDefaultColors(defaultColors);
10572
10573         float16InputX.reserve(numDataPoints);
10574         for (deUint32 y = 0; y < numDataPointsByAxis; ++y)
10575         for (deUint32 x = 0; x < numDataPointsByAxis; ++x)
10576         for (deUint32 n = 0; n < N; ++n)
10577         {
10578                 const float arg = static_cast<float>(2 * DE_PI) * static_cast<float>(x * (n + 1)) / static_cast<float>(1 * numDataPointsByAxis);
10579
10580                 if (y%2 == 0)
10581                         float16InputX.push_back(tcu::Float16(sin(arg)).bits());
10582                 else
10583                         float16InputX.push_back(tcu::Float16(cos(arg)).bits());
10584         }
10585
10586         float16InputY.reserve(numDataPoints);
10587         for (deUint32 y = 0; y < numDataPointsByAxis; ++y)
10588         for (deUint32 x = 0; x < numDataPointsByAxis; ++x)
10589         for (deUint32 n = 0; n < N; ++n)
10590         {
10591                 const float arg = static_cast<float>(2 * DE_PI) * static_cast<float>(y * (n + 1)) / static_cast<float>(1 * numDataPointsByAxis);
10592
10593                 if (x%2 == 0)
10594                         float16InputY.push_back(tcu::Float16(sin(arg)).bits());
10595                 else
10596                         float16InputY.push_back(tcu::Float16(cos(arg)).bits());
10597         }
10598
10599         const deFloat16 testNumbers[]   =
10600         {
10601                 tcu::Float16( 2.0  ).bits(),
10602                 tcu::Float16( 4.0  ).bits(),
10603                 tcu::Float16( 8.0  ).bits(),
10604                 tcu::Float16( 16.0 ).bits(),
10605                 tcu::Float16( 32.0 ).bits(),
10606                 tcu::Float16( 64.0 ).bits(),
10607                 tcu::Float16( 128.0).bits(),
10608                 tcu::Float16( 256.0).bits(),
10609                 tcu::Float16( 512.0).bits(),
10610                 tcu::Float16(-2.0  ).bits(),
10611                 tcu::Float16(-4.0  ).bits(),
10612                 tcu::Float16(-8.0  ).bits(),
10613                 tcu::Float16(-16.0 ).bits(),
10614                 tcu::Float16(-32.0 ).bits(),
10615                 tcu::Float16(-64.0 ).bits(),
10616                 tcu::Float16(-128.0).bits(),
10617                 tcu::Float16(-256.0).bits(),
10618                 tcu::Float16(-512.0).bits(),
10619         };
10620
10621         float16InputW.reserve(numDataPoints);
10622         for (deUint32 y = 0; y < numDataPointsByAxis; ++y)
10623         for (deUint32 x = 0; x < numDataPointsByAxis; ++x)
10624         for (deUint32 n = 0; n < N; ++n)
10625                 float16InputW.push_back(testNumbers[rnd.getInt(0, DE_LENGTH_OF_ARRAY(testNumbers) - 1)]);
10626
10627         struct TestOp
10628         {
10629                 const char*                     opCode;
10630                 vector<deFloat16>&      inputData;
10631                 VerifyIOFunc            verifyFunc;
10632         };
10633
10634         const TestOp    testOps[]       =
10635         {
10636                 { "OpDPdxFine"          ,       float16InputX   ,       compareDerivative<R, N, 1, getFDxFine           >       },
10637                 { "OpDPdyFine"          ,       float16InputY   ,       compareDerivative<R, N, 1, getFDyFine           >       },
10638                 { "OpFwidthFine"        ,       float16InputW   ,       compareDerivative<R, N, 1, getFWidthFine        >       },
10639                 { "OpDPdxCoarse"        ,       float16InputX   ,       compareDerivative<R, N, 3, getFDx                       >       },
10640                 { "OpDPdyCoarse"        ,       float16InputY   ,       compareDerivative<R, N, 3, getFDy                       >       },
10641                 { "OpFwidthCoarse"      ,       float16InputW   ,       compareDerivative<R, N, 5, getFWidth            >       },
10642                 { "OpDPdx"                      ,       float16InputX   ,       compareDerivative<R, N, 3, getFDx                       >       },
10643                 { "OpDPdy"                      ,       float16InputY   ,       compareDerivative<R, N, 3, getFDy                       >       },
10644                 { "OpFwidth"            ,       float16InputW   ,       compareDerivative<R, N, 5, getFWidth            >       },
10645         };
10646
10647         struct TestType
10648         {
10649                 const deUint32  typeComponents;
10650                 const char*             typeName;
10651                 const char*             typeDecls;
10652         };
10653
10654         const TestType  testTypes[]     =
10655         {
10656                 {
10657                         1,
10658                         "f16",
10659                         ""
10660                 },
10661                 {
10662                         2,
10663                         "v2f16",
10664                         "      %v2f16 = OpTypeVector %f16 2\n"
10665                 },
10666                 {
10667                         4,
10668                         "v4f16",
10669                         "      %v4f16 = OpTypeVector %f16 4\n"
10670                 },
10671         };
10672
10673         const deUint32  testTypeNdx     = (N == 1) ? 0
10674                                                                 : (N == 2) ? 1
10675                                                                 : (N == 4) ? 2
10676                                                                 : DE_LENGTH_OF_ARRAY(testTypes);
10677         const TestType& testType        =       testTypes[testTypeNdx];
10678
10679         DE_ASSERT(testTypeNdx < DE_LENGTH_OF_ARRAY(testTypes));
10680         DE_ASSERT(testType.typeComponents == N);
10681
10682         const StringTemplate preMain
10683         (
10684                 "%c_i32_ndp = OpConstant %i32 ${num_data_points}\n"
10685                 " %c_u32_xw = OpConstant %u32 ${items_by_x}\n"
10686                 "      %f16 = OpTypeFloat 16\n"
10687                 "${type_decls}"
10688                 " %up_${tt} = OpTypePointer Uniform %${tt}\n"
10689                 " %ra_${tt} = OpTypeArray %${tt} %c_i32_ndp\n"
10690                 "   %SSBO16 = OpTypeStruct %ra_${tt}\n"
10691                 "%up_SSBO16 = OpTypePointer Uniform %SSBO16\n"
10692                 " %ssbo_src = OpVariable %up_SSBO16 Uniform\n"
10693                 " %ssbo_dst = OpVariable %up_SSBO16 Uniform\n"
10694         );
10695
10696         const StringTemplate decoration
10697         (
10698                 "OpDecorate %ra_${tt} ArrayStride ${tt_stride}\n"
10699                 "OpMemberDecorate %SSBO16 0 Offset 0\n"
10700                 "OpDecorate %SSBO16 BufferBlock\n"
10701                 "OpDecorate %ssbo_src DescriptorSet 0\n"
10702                 "OpDecorate %ssbo_src Binding 0\n"
10703                 "OpDecorate %ssbo_dst DescriptorSet 0\n"
10704                 "OpDecorate %ssbo_dst Binding 1\n"
10705         );
10706
10707         const StringTemplate testFun
10708         (
10709                 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
10710                 "    %param = OpFunctionParameter %v4f32\n"
10711                 "    %entry = OpLabel\n"
10712
10713                 "  %loc_x_c = OpAccessChain %ip_f32 %BP_gl_FragCoord %c_i32_0\n"
10714                 "  %loc_y_c = OpAccessChain %ip_f32 %BP_gl_FragCoord %c_i32_1\n"
10715                 "      %x_c = OpLoad %f32 %loc_x_c\n"
10716                 "      %y_c = OpLoad %f32 %loc_y_c\n"
10717                 "    %x_idx = OpConvertFToU %u32 %x_c\n"
10718                 "    %y_idx = OpConvertFToU %u32 %y_c\n"
10719                 "    %ndx_y = OpIMul %u32 %y_idx %c_u32_xw\n"
10720                 "      %ndx = OpIAdd %u32 %ndx_y %x_idx\n"
10721
10722                 "      %src = OpAccessChain %up_${tt} %ssbo_src %c_i32_0 %ndx\n"
10723                 "  %val_src = OpLoad %${tt} %src\n"
10724                 "  %val_dst = ${op_code} %${tt} %val_src\n"
10725                 "      %dst = OpAccessChain %up_${tt} %ssbo_dst %c_i32_0 %ndx\n"
10726                 "             OpStore %dst %val_dst\n"
10727                 "             OpBranch %merge\n"
10728
10729                 "    %merge = OpLabel\n"
10730                 "             OpReturnValue %param\n"
10731
10732                 "             OpFunctionEnd\n"
10733         );
10734
10735         for (deUint32 testOpsIdx = 0; testOpsIdx < DE_LENGTH_OF_ARRAY(testOps); ++testOpsIdx)
10736         {
10737                 const TestOp&           testOp                  = testOps[testOpsIdx];
10738                 const string            testName                = de::toLower(string(testOp.opCode));
10739                 const size_t            typeStride              = N * sizeof(deFloat16);
10740                 GraphicsResources       specResource;
10741                 map<string, string>     specs;
10742                 VulkanFeatures          features;
10743                 vector<string>          extensions;
10744                 map<string, string>     fragments;
10745                 SpecConstants           noSpecConstants;
10746                 PushConstants           noPushConstants;
10747                 GraphicsInterfaces      noInterfaces;
10748
10749                 specs["op_code"]                        = testOp.opCode;
10750                 specs["num_data_points"]        = de::toString(testOp.inputData.size() / N);
10751                 specs["items_by_x"]                     = de::toString(numDataPointsByAxis);
10752                 specs["tt"]                                     = testType.typeName;
10753                 specs["tt_stride"]                      = de::toString(typeStride);
10754                 specs["type_decls"]                     = testType.typeDecls;
10755
10756                 fragments["extension"]          = "OpExtension \"SPV_KHR_16bit_storage\"";
10757                 fragments["capability"]         = "OpCapability DerivativeControl\nOpCapability StorageUniformBufferBlock16\n";
10758                 fragments["decoration"]         = decoration.specialize(specs);
10759                 fragments["pre_main"]           = preMain.specialize(specs);
10760                 fragments["testfun"]            = testFun.specialize(specs);
10761
10762                 specResource.inputs.push_back(Resource(BufferSp(new Float16Buffer(testOp.inputData)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
10763                 specResource.outputs.push_back(Resource(BufferSp(new Float16Buffer(float16OutputDummy)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
10764                 specResource.verifyIO = testOp.verifyFunc;
10765
10766                 extensions.push_back("VK_KHR_16bit_storage");
10767                 extensions.push_back("VK_KHR_shader_float16_int8");
10768
10769                 features.extFloat16Int8         = EXTFLOAT16INT8FEATURES_FLOAT16;
10770                 features.ext16BitStorage        = EXT16BITSTORAGEFEATURES_UNIFORM_BUFFER_BLOCK;
10771
10772                 createTestForStage(VK_SHADER_STAGE_FRAGMENT_BIT, testName.c_str(), defaultColors, defaultColors, fragments, noSpecConstants,
10773                                                         noPushConstants, specResource, noInterfaces, extensions, features, testGroup.get(), QP_TEST_RESULT_FAIL, string(), true);
10774         }
10775
10776         return testGroup.release();
10777 }
10778
10779 bool compareFP16VectorExtractFunc (const std::vector<Resource>& inputs, const vector<AllocationSp>& outputAllocs, const std::vector<Resource>&, TestLog& log)
10780 {
10781         if (inputs.size() != 2 || outputAllocs.size() != 1)
10782                 return false;
10783
10784         vector<deUint8> input1Bytes;
10785         vector<deUint8> input2Bytes;
10786
10787         inputs[0].getBytes(input1Bytes);
10788         inputs[1].getBytes(input2Bytes);
10789
10790         DE_ASSERT(input1Bytes.size() > 0);
10791         DE_ASSERT(input2Bytes.size() > 0);
10792         DE_ASSERT(input2Bytes.size() % sizeof(deUint32) == 0);
10793
10794         const size_t                    iterations              = input2Bytes.size() / sizeof(deUint32);
10795         const size_t                    components              = input1Bytes.size() / (sizeof(deFloat16) * iterations);
10796         const deFloat16* const  input1AsFP16    = (const deFloat16*)&input1Bytes[0];
10797         const deUint32* const   inputIndices    = (const deUint32*)&input2Bytes[0];
10798         const deFloat16* const  outputAsFP16    = (const deFloat16*)outputAllocs[0]->getHostPtr();
10799         std::string                             error;
10800
10801         DE_ASSERT(components == 2 || components == 4);
10802         DE_ASSERT(input1Bytes.size() == iterations * components * sizeof(deFloat16));
10803
10804         for (size_t idx = 0; idx < iterations; ++idx)
10805         {
10806                 const deUint32  componentNdx    = inputIndices[idx];
10807
10808                 DE_ASSERT(componentNdx < components);
10809
10810                 const deFloat16 expected                = input1AsFP16[components * idx + componentNdx];
10811
10812                 if (!compare16BitFloat(expected, outputAsFP16[idx], error))
10813                 {
10814                         log << TestLog::Message << "At " << idx << error << TestLog::EndMessage;
10815
10816                         return false;
10817                 }
10818         }
10819
10820         return true;
10821 }
10822
10823 template<class SpecResource>
10824 tcu::TestCaseGroup* createFloat16VectorExtractSet (tcu::TestContext& testCtx)
10825 {
10826         de::MovePtr<tcu::TestCaseGroup>         testGroup                       (new tcu::TestCaseGroup(testCtx, "opvectorextractdynamic", "OpVectorExtractDynamic tests"));
10827
10828         de::Random                                                      rnd                                     (deStringHash(testGroup->getName()));
10829         const deUint32                                          numDataPoints           = 256;
10830         const vector<deFloat16>                         float16InputData        = getFloat16s(rnd, numDataPoints);
10831         const vector<deFloat16>                         float16OutputDummy      (float16InputData.size(), 0);
10832
10833         struct TestType
10834         {
10835                 const deUint32  typeComponents;
10836                 const size_t    typeStride;
10837                 const char*             typeName;
10838                 const char*             typeDecls;
10839         };
10840
10841         const TestType  testTypes[]     =
10842         {
10843                 {
10844                         2,
10845                         2 * sizeof(deFloat16),
10846                         "v2f16",
10847                         "      %v2f16 = OpTypeVector %f16 2\n"
10848                 },
10849                 {
10850                         3,
10851                         4 * sizeof(deFloat16),
10852                         "v3f16",
10853                         "      %v3f16 = OpTypeVector %f16 3\n"
10854                 },
10855                 {
10856                         4,
10857                         4 * sizeof(deFloat16),
10858                         "v4f16",
10859                         "      %v4f16 = OpTypeVector %f16 4\n"
10860                 },
10861         };
10862
10863         const StringTemplate preMain
10864         (
10865                 "  %c_i32_ndp = OpConstant %i32 ${num_data_points}\n"
10866                 "        %f16 = OpTypeFloat 16\n"
10867
10868                 "${type_decl}"
10869
10870                 "   %up_${tt} = OpTypePointer Uniform %${tt}\n"
10871                 "   %ra_${tt} = OpTypeArray %${tt} %c_i32_ndp\n"
10872                 "   %SSBO_SRC = OpTypeStruct %ra_${tt}\n"
10873                 "%up_SSBO_SRC = OpTypePointer Uniform %SSBO_SRC\n"
10874
10875                 "     %up_u32 = OpTypePointer Uniform %u32\n"
10876                 "     %ra_u32 = OpTypeArray %u32 %c_i32_ndp\n"
10877                 "   %SSBO_IDX = OpTypeStruct %ra_u32\n"
10878                 "%up_SSBO_IDX = OpTypePointer Uniform %SSBO_IDX\n"
10879
10880                 "     %up_f16 = OpTypePointer Uniform %f16\n"
10881                 "     %ra_f16 = OpTypeArray %f16 %c_i32_ndp\n"
10882                 "   %SSBO_DST = OpTypeStruct %ra_f16\n"
10883                 "%up_SSBO_DST = OpTypePointer Uniform %SSBO_DST\n"
10884
10885                 "   %ssbo_src = OpVariable %up_SSBO_SRC Uniform\n"
10886                 "   %ssbo_idx = OpVariable %up_SSBO_IDX Uniform\n"
10887                 "   %ssbo_dst = OpVariable %up_SSBO_DST Uniform\n"
10888         );
10889
10890         const StringTemplate decoration
10891         (
10892                 "OpDecorate %ra_${tt} ArrayStride ${tt_stride}\n"
10893                 "OpMemberDecorate %SSBO_SRC 0 Offset 0\n"
10894                 "OpDecorate %SSBO_SRC BufferBlock\n"
10895                 "OpDecorate %ssbo_src DescriptorSet 0\n"
10896                 "OpDecorate %ssbo_src Binding 0\n"
10897
10898                 "OpDecorate %ra_u32 ArrayStride 4\n"
10899                 "OpMemberDecorate %SSBO_IDX 0 Offset 0\n"
10900                 "OpDecorate %SSBO_IDX BufferBlock\n"
10901                 "OpDecorate %ssbo_idx DescriptorSet 0\n"
10902                 "OpDecorate %ssbo_idx Binding 1\n"
10903
10904                 "OpDecorate %ra_f16 ArrayStride 2\n"
10905                 "OpMemberDecorate %SSBO_DST 0 Offset 0\n"
10906                 "OpDecorate %SSBO_DST BufferBlock\n"
10907                 "OpDecorate %ssbo_dst DescriptorSet 0\n"
10908                 "OpDecorate %ssbo_dst Binding 2\n"
10909         );
10910
10911         const StringTemplate testFun
10912         (
10913                 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
10914                 "    %param = OpFunctionParameter %v4f32\n"
10915                 "    %entry = OpLabel\n"
10916
10917                 "        %i = OpVariable %fp_i32 Function\n"
10918                 "             OpStore %i %c_i32_0\n"
10919
10920                 " %will_run = OpFunctionCall %bool %isUniqueIdZero\n"
10921                 "             OpSelectionMerge %end_if None\n"
10922                 "             OpBranchConditional %will_run %run_test %end_if\n"
10923
10924                 " %run_test = OpLabel\n"
10925                 "             OpBranch %loop\n"
10926
10927                 "     %loop = OpLabel\n"
10928                 "    %i_cmp = OpLoad %i32 %i\n"
10929                 "       %lt = OpSLessThan %bool %i_cmp %c_i32_ndp\n"
10930                 "             OpLoopMerge %merge %next None\n"
10931                 "             OpBranchConditional %lt %write %merge\n"
10932
10933                 "    %write = OpLabel\n"
10934                 "      %ndx = OpLoad %i32 %i\n"
10935
10936                 "      %src = OpAccessChain %up_${tt} %ssbo_src %c_i32_0 %ndx\n"
10937                 "  %val_src = OpLoad %${tt} %src\n"
10938
10939                 "  %src_idx = OpAccessChain %up_u32 %ssbo_idx %c_i32_0 %ndx\n"
10940                 "  %val_idx = OpLoad %u32 %src_idx\n"
10941
10942                 "  %val_dst = OpVectorExtractDynamic %f16 %val_src %val_idx\n"
10943                 "      %dst = OpAccessChain %up_f16 %ssbo_dst %c_i32_0 %ndx\n"
10944
10945                 "             OpStore %dst %val_dst\n"
10946                 "             OpBranch %next\n"
10947
10948                 "     %next = OpLabel\n"
10949                 "    %i_cur = OpLoad %i32 %i\n"
10950                 "    %i_new = OpIAdd %i32 %i_cur %c_i32_1\n"
10951                 "             OpStore %i %i_new\n"
10952                 "             OpBranch %loop\n"
10953
10954                 "    %merge = OpLabel\n"
10955                 "             OpBranch %end_if\n"
10956                 "   %end_if = OpLabel\n"
10957                 "             OpReturnValue %param\n"
10958
10959                 "             OpFunctionEnd\n"
10960         );
10961
10962         for (deUint32 testTypeIdx = 0; testTypeIdx < DE_LENGTH_OF_ARRAY(testTypes); ++testTypeIdx)
10963         {
10964                 const TestType&         testType                = testTypes[testTypeIdx];
10965                 const string            testName                = testType.typeName;
10966                 const size_t            itemsPerType    = testType.typeStride / sizeof(deFloat16);
10967                 const size_t            iterations              = float16InputData.size() / itemsPerType;
10968                 SpecResource            specResource;
10969                 map<string, string>     specs;
10970                 VulkanFeatures          features;
10971                 vector<deUint32>        inputDataNdx;
10972                 map<string, string>     fragments;
10973                 vector<string>          extensions;
10974
10975                 for (deUint32 ndx = 0; ndx < iterations; ++ndx)
10976                         inputDataNdx.push_back(rnd.getUint32() % testType.typeComponents);
10977
10978                 specs["num_data_points"]        = de::toString(iterations);
10979                 specs["tt"]                                     = testType.typeName;
10980                 specs["tt_stride"]                      = de::toString(testType.typeStride);
10981                 specs["type_decl"]                      = testType.typeDecls;
10982
10983                 fragments["extension"]          = "OpExtension \"SPV_KHR_16bit_storage\"";
10984                 fragments["capability"]         = "OpCapability StorageUniformBufferBlock16\n";
10985                 fragments["decoration"]         = decoration.specialize(specs);
10986                 fragments["pre_main"]           = preMain.specialize(specs);
10987                 fragments["testfun"]            = testFun.specialize(specs);
10988
10989                 specResource.inputs.push_back(Resource(BufferSp(new Float16Buffer(float16InputData)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
10990                 specResource.inputs.push_back(Resource(BufferSp(new Uint32Buffer(inputDataNdx)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
10991                 specResource.outputs.push_back(Resource(BufferSp(new Float16Buffer(float16OutputDummy)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
10992                 specResource.verifyIO = compareFP16VectorExtractFunc;
10993
10994                 extensions.push_back("VK_KHR_16bit_storage");
10995                 extensions.push_back("VK_KHR_shader_float16_int8");
10996
10997                 features.extFloat16Int8         = EXTFLOAT16INT8FEATURES_FLOAT16;
10998                 features.ext16BitStorage        = EXT16BITSTORAGEFEATURES_UNIFORM_BUFFER_BLOCK;
10999
11000                 finalizeTestsCreation(specResource, fragments, testCtx, *testGroup.get(), testName, features, extensions, IVec3(1, 1, 1));
11001         }
11002
11003         return testGroup.release();
11004 }
11005
11006 template<deUint32 COMPONENTS_COUNT, deUint32 REPLACEMENT>
11007 bool compareFP16VectorInsertFunc (const std::vector<Resource>& inputs, const vector<AllocationSp>& outputAllocs, const std::vector<Resource>&, TestLog& log)
11008 {
11009         if (inputs.size() != 2 || outputAllocs.size() != 1)
11010                 return false;
11011
11012         vector<deUint8> input1Bytes;
11013         vector<deUint8> input2Bytes;
11014
11015         inputs[0].getBytes(input1Bytes);
11016         inputs[1].getBytes(input2Bytes);
11017
11018         DE_ASSERT(input1Bytes.size() > 0);
11019         DE_ASSERT(input2Bytes.size() > 0);
11020         DE_ASSERT(input2Bytes.size() % sizeof(deUint32) == 0);
11021
11022         const size_t                    iterations                      = input2Bytes.size() / sizeof(deUint32);
11023         const size_t                    componentsStride        = input1Bytes.size() / (sizeof(deFloat16) * iterations);
11024         const deFloat16* const  input1AsFP16            = (const deFloat16*)&input1Bytes[0];
11025         const deUint32* const   inputIndices            = (const deUint32*)&input2Bytes[0];
11026         const deFloat16* const  outputAsFP16            = (const deFloat16*)outputAllocs[0]->getHostPtr();
11027         const deFloat16                 magic                           = tcu::Float16(float(REPLACEMENT)).bits();
11028         std::string                             error;
11029
11030         DE_ASSERT(componentsStride == 2 || componentsStride == 4);
11031         DE_ASSERT(input1Bytes.size() == iterations * componentsStride * sizeof(deFloat16));
11032
11033         for (size_t idx = 0; idx < iterations; ++idx)
11034         {
11035                 const deFloat16*        inputVec                = &input1AsFP16[componentsStride * idx];
11036                 const deFloat16*        outputVec               = &outputAsFP16[componentsStride * idx];
11037                 const deUint32          replacedCompNdx = inputIndices[idx];
11038
11039                 DE_ASSERT(replacedCompNdx < COMPONENTS_COUNT);
11040
11041                 for (size_t compNdx = 0; compNdx < COMPONENTS_COUNT; ++compNdx)
11042                 {
11043                         const deFloat16 expected        = (compNdx == replacedCompNdx) ? magic : inputVec[compNdx];
11044
11045                         if (!compare16BitFloat(expected, outputVec[compNdx], error))
11046                         {
11047                                 log << TestLog::Message << "At " << idx << "[" << compNdx << "]: " << error << TestLog::EndMessage;
11048
11049                                 return false;
11050                         }
11051                 }
11052         }
11053
11054         return true;
11055 }
11056
11057 template<class SpecResource>
11058 tcu::TestCaseGroup* createFloat16VectorInsertSet (tcu::TestContext& testCtx)
11059 {
11060         de::MovePtr<tcu::TestCaseGroup>         testGroup                       (new tcu::TestCaseGroup(testCtx, "opvectorinsertdynamic", "OpVectorInsertDynamic tests"));
11061
11062         de::Random                                                      rnd                                     (deStringHash(testGroup->getName()));
11063         const deUint32                                          replacement                     = 42;
11064         const deUint32                                          numDataPoints           = 256;
11065         const vector<deFloat16>                         float16InputData        = getFloat16s(rnd, numDataPoints);
11066         const vector<deFloat16>                         float16OutputDummy      (float16InputData.size(), 0);
11067
11068         struct TestType
11069         {
11070                 const deUint32  typeComponents;
11071                 const size_t    typeStride;
11072                 const char*             typeName;
11073                 const char*             typeDecls;
11074                 VerifyIOFunc    verifyIOFunc;
11075         };
11076
11077         const TestType  testTypes[]     =
11078         {
11079                 {
11080                         2,
11081                         2 * sizeof(deFloat16),
11082                         "v2f16",
11083                         "      %v2f16 = OpTypeVector %f16 2\n",
11084                         compareFP16VectorInsertFunc<2, replacement>
11085                 },
11086                 {
11087                         3,
11088                         4 * sizeof(deFloat16),
11089                         "v3f16",
11090                         "      %v3f16 = OpTypeVector %f16 3\n",
11091                         compareFP16VectorInsertFunc<3, replacement>
11092                 },
11093                 {
11094                         4,
11095                         4 * sizeof(deFloat16),
11096                         "v4f16",
11097                         "      %v4f16 = OpTypeVector %f16 4\n",
11098                         compareFP16VectorInsertFunc<4, replacement>
11099                 },
11100         };
11101
11102         const StringTemplate preMain
11103         (
11104                 "  %c_i32_ndp = OpConstant %i32 ${num_data_points}\n"
11105                 "        %f16 = OpTypeFloat 16\n"
11106                 "  %c_f16_ins = OpConstant %f16 ${replacement}\n"
11107
11108                 "${type_decl}"
11109
11110                 "   %up_${tt} = OpTypePointer Uniform %${tt}\n"
11111                 "   %ra_${tt} = OpTypeArray %${tt} %c_i32_ndp\n"
11112                 "   %SSBO_SRC = OpTypeStruct %ra_${tt}\n"
11113                 "%up_SSBO_SRC = OpTypePointer Uniform %SSBO_SRC\n"
11114
11115                 "     %up_u32 = OpTypePointer Uniform %u32\n"
11116                 "     %ra_u32 = OpTypeArray %u32 %c_i32_ndp\n"
11117                 "   %SSBO_IDX = OpTypeStruct %ra_u32\n"
11118                 "%up_SSBO_IDX = OpTypePointer Uniform %SSBO_IDX\n"
11119
11120                 "   %SSBO_DST = OpTypeStruct %ra_${tt}\n"
11121                 "%up_SSBO_DST = OpTypePointer Uniform %SSBO_DST\n"
11122
11123                 "   %ssbo_src = OpVariable %up_SSBO_SRC Uniform\n"
11124                 "   %ssbo_idx = OpVariable %up_SSBO_IDX Uniform\n"
11125                 "   %ssbo_dst = OpVariable %up_SSBO_DST Uniform\n"
11126         );
11127
11128         const StringTemplate decoration
11129         (
11130                 "OpDecorate %ra_${tt} ArrayStride ${tt_stride}\n"
11131                 "OpMemberDecorate %SSBO_SRC 0 Offset 0\n"
11132                 "OpDecorate %SSBO_SRC BufferBlock\n"
11133                 "OpDecorate %ssbo_src DescriptorSet 0\n"
11134                 "OpDecorate %ssbo_src Binding 0\n"
11135
11136                 "OpDecorate %ra_u32 ArrayStride 4\n"
11137                 "OpMemberDecorate %SSBO_IDX 0 Offset 0\n"
11138                 "OpDecorate %SSBO_IDX BufferBlock\n"
11139                 "OpDecorate %ssbo_idx DescriptorSet 0\n"
11140                 "OpDecorate %ssbo_idx Binding 1\n"
11141
11142                 "OpMemberDecorate %SSBO_DST 0 Offset 0\n"
11143                 "OpDecorate %SSBO_DST BufferBlock\n"
11144                 "OpDecorate %ssbo_dst DescriptorSet 0\n"
11145                 "OpDecorate %ssbo_dst Binding 2\n"
11146         );
11147
11148         const StringTemplate testFun
11149         (
11150                 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
11151                 "    %param = OpFunctionParameter %v4f32\n"
11152                 "    %entry = OpLabel\n"
11153
11154                 "        %i = OpVariable %fp_i32 Function\n"
11155                 "             OpStore %i %c_i32_0\n"
11156
11157                 " %will_run = OpFunctionCall %bool %isUniqueIdZero\n"
11158                 "             OpSelectionMerge %end_if None\n"
11159                 "             OpBranchConditional %will_run %run_test %end_if\n"
11160
11161                 " %run_test = OpLabel\n"
11162                 "             OpBranch %loop\n"
11163
11164                 "     %loop = OpLabel\n"
11165                 "    %i_cmp = OpLoad %i32 %i\n"
11166                 "       %lt = OpSLessThan %bool %i_cmp %c_i32_ndp\n"
11167                 "             OpLoopMerge %merge %next None\n"
11168                 "             OpBranchConditional %lt %write %merge\n"
11169
11170                 "    %write = OpLabel\n"
11171                 "      %ndx = OpLoad %i32 %i\n"
11172
11173                 "      %src = OpAccessChain %up_${tt} %ssbo_src %c_i32_0 %ndx\n"
11174                 "  %val_src = OpLoad %${tt} %src\n"
11175
11176                 "  %src_idx = OpAccessChain %up_u32 %ssbo_idx %c_i32_0 %ndx\n"
11177                 "  %val_idx = OpLoad %u32 %src_idx\n"
11178
11179                 "  %val_dst = OpVectorInsertDynamic %${tt} %val_src %c_f16_ins %val_idx\n"
11180                 "      %dst = OpAccessChain %up_${tt} %ssbo_dst %c_i32_0 %ndx\n"
11181
11182                 "             OpStore %dst %val_dst\n"
11183                 "             OpBranch %next\n"
11184
11185                 "     %next = OpLabel\n"
11186                 "    %i_cur = OpLoad %i32 %i\n"
11187                 "    %i_new = OpIAdd %i32 %i_cur %c_i32_1\n"
11188                 "             OpStore %i %i_new\n"
11189                 "             OpBranch %loop\n"
11190
11191                 "    %merge = OpLabel\n"
11192                 "             OpBranch %end_if\n"
11193                 "   %end_if = OpLabel\n"
11194                 "             OpReturnValue %param\n"
11195
11196                 "             OpFunctionEnd\n"
11197         );
11198
11199         for (deUint32 testTypeIdx = 0; testTypeIdx < DE_LENGTH_OF_ARRAY(testTypes); ++testTypeIdx)
11200         {
11201                 const TestType&         testType                = testTypes[testTypeIdx];
11202                 const string            testName                = testType.typeName;
11203                 const size_t            itemsPerType    = testType.typeStride / sizeof(deFloat16);
11204                 const size_t            iterations              = float16InputData.size() / itemsPerType;
11205                 SpecResource            specResource;
11206                 map<string, string>     specs;
11207                 VulkanFeatures          features;
11208                 vector<deUint32>        inputDataNdx;
11209                 map<string, string>     fragments;
11210                 vector<string>          extensions;
11211
11212                 for (deUint32 ndx = 0; ndx < iterations; ++ndx)
11213                         inputDataNdx.push_back(rnd.getUint32() % testType.typeComponents);
11214
11215                 specs["num_data_points"]        = de::toString(iterations);
11216                 specs["tt"]                                     = testType.typeName;
11217                 specs["tt_stride"]                      = de::toString(testType.typeStride);
11218                 specs["type_decl"]                      = testType.typeDecls;
11219                 specs["replacement"]            = de::toString(replacement);
11220
11221                 fragments["extension"]          = "OpExtension \"SPV_KHR_16bit_storage\"";
11222                 fragments["capability"]         = "OpCapability StorageUniformBufferBlock16\n";
11223                 fragments["decoration"]         = decoration.specialize(specs);
11224                 fragments["pre_main"]           = preMain.specialize(specs);
11225                 fragments["testfun"]            = testFun.specialize(specs);
11226
11227                 specResource.inputs.push_back(Resource(BufferSp(new Float16Buffer(float16InputData)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
11228                 specResource.inputs.push_back(Resource(BufferSp(new Uint32Buffer(inputDataNdx)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
11229                 specResource.outputs.push_back(Resource(BufferSp(new Float16Buffer(float16OutputDummy)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
11230                 specResource.verifyIO = testType.verifyIOFunc;
11231
11232                 extensions.push_back("VK_KHR_16bit_storage");
11233                 extensions.push_back("VK_KHR_shader_float16_int8");
11234
11235                 features.extFloat16Int8         = EXTFLOAT16INT8FEATURES_FLOAT16;
11236                 features.ext16BitStorage        = EXT16BITSTORAGEFEATURES_UNIFORM_BUFFER_BLOCK;
11237
11238                 finalizeTestsCreation(specResource, fragments, testCtx, *testGroup.get(), testName, features, extensions, IVec3(1, 1, 1));
11239         }
11240
11241         return testGroup.release();
11242 }
11243
11244 inline deFloat16 getShuffledComponent (const size_t iteration, const size_t componentNdx, const deFloat16* input1Vec, const deFloat16* input2Vec, size_t vec1Len, size_t vec2Len, bool& validate)
11245 {
11246         const size_t    compNdxCount    = (vec1Len + vec2Len + 1);
11247         const size_t    compNdxLimited  = iteration % (compNdxCount * compNdxCount);
11248         size_t                  comp;
11249
11250         switch (componentNdx)
11251         {
11252                 case 0: comp = compNdxLimited / compNdxCount; break;
11253                 case 1: comp = compNdxLimited % compNdxCount; break;
11254                 case 2: comp = 0; break;
11255                 case 3: comp = 1; break;
11256                 default: TCU_THROW(InternalError, "Impossible");
11257         }
11258
11259         if (comp >= vec1Len + vec2Len)
11260         {
11261                 validate = false;
11262                 return 0;
11263         }
11264         else
11265         {
11266                 validate = true;
11267                 return (comp < vec1Len) ? input1Vec[comp] : input2Vec[comp - vec1Len];
11268         }
11269 }
11270
11271 template<deUint32 DST_COMPONENTS_COUNT, deUint32 SRC0_COMPONENTS_COUNT, deUint32 SRC1_COMPONENTS_COUNT>
11272 bool compareFP16VectorShuffleFunc (const std::vector<Resource>& inputs, const vector<AllocationSp>& outputAllocs, const std::vector<Resource>&, TestLog& log)
11273 {
11274         DE_STATIC_ASSERT(DST_COMPONENTS_COUNT == 2 || DST_COMPONENTS_COUNT == 3 || DST_COMPONENTS_COUNT == 4);
11275         DE_STATIC_ASSERT(SRC0_COMPONENTS_COUNT == 2 || SRC0_COMPONENTS_COUNT == 3 || SRC0_COMPONENTS_COUNT == 4);
11276         DE_STATIC_ASSERT(SRC1_COMPONENTS_COUNT == 2 || SRC1_COMPONENTS_COUNT == 3 || SRC1_COMPONENTS_COUNT == 4);
11277
11278         if (inputs.size() != 2 || outputAllocs.size() != 1)
11279                 return false;
11280
11281         vector<deUint8> input1Bytes;
11282         vector<deUint8> input2Bytes;
11283
11284         inputs[0].getBytes(input1Bytes);
11285         inputs[1].getBytes(input2Bytes);
11286
11287         DE_ASSERT(input1Bytes.size() > 0);
11288         DE_ASSERT(input2Bytes.size() > 0);
11289         DE_ASSERT(input2Bytes.size() % sizeof(deFloat16) == 0);
11290
11291         const size_t                    componentsStrideDst             = (DST_COMPONENTS_COUNT == 3) ? 4 : DST_COMPONENTS_COUNT;
11292         const size_t                    componentsStrideSrc0    = (SRC0_COMPONENTS_COUNT == 3) ? 4 : SRC0_COMPONENTS_COUNT;
11293         const size_t                    componentsStrideSrc1    = (SRC1_COMPONENTS_COUNT == 3) ? 4 : SRC1_COMPONENTS_COUNT;
11294         const size_t                    iterations                              = input1Bytes.size() / (componentsStrideSrc0 * sizeof(deFloat16));
11295         const deFloat16* const  input1AsFP16                    = (const deFloat16*)&input1Bytes[0];
11296         const deFloat16* const  input2AsFP16                    = (const deFloat16*)&input2Bytes[0];
11297         const deFloat16* const  outputAsFP16                    = (const deFloat16*)outputAllocs[0]->getHostPtr();
11298         std::string                             error;
11299
11300         DE_ASSERT(input1Bytes.size() == iterations * componentsStrideSrc0 * sizeof(deFloat16));
11301         DE_ASSERT(input2Bytes.size() == iterations * componentsStrideSrc1 * sizeof(deFloat16));
11302
11303         for (size_t idx = 0; idx < iterations; ++idx)
11304         {
11305                 const deFloat16*        input1Vec       = &input1AsFP16[componentsStrideSrc0 * idx];
11306                 const deFloat16*        input2Vec       = &input2AsFP16[componentsStrideSrc1 * idx];
11307                 const deFloat16*        outputVec       = &outputAsFP16[componentsStrideDst * idx];
11308
11309                 for (size_t compNdx = 0; compNdx < DST_COMPONENTS_COUNT; ++compNdx)
11310                 {
11311                         bool            validate        = true;
11312                         deFloat16       expected        = getShuffledComponent(idx, compNdx, input1Vec, input2Vec, SRC0_COMPONENTS_COUNT, SRC1_COMPONENTS_COUNT, validate);
11313
11314                         if (validate && !compare16BitFloat(expected, outputVec[compNdx], error))
11315                         {
11316                                 log << TestLog::Message << "At " << idx << "[" << compNdx << "]: " << error << TestLog::EndMessage;
11317
11318                                 return false;
11319                         }
11320                 }
11321         }
11322
11323         return true;
11324 }
11325
11326 VerifyIOFunc getFloat16VectorShuffleVerifyIOFunc (deUint32 dstComponentsCount, deUint32 src0ComponentsCount, deUint32 src1ComponentsCount)
11327 {
11328         DE_ASSERT(dstComponentsCount <= 4);
11329         DE_ASSERT(src0ComponentsCount <= 4);
11330         DE_ASSERT(src1ComponentsCount <= 4);
11331         deUint32 funcCode = 100 * dstComponentsCount + 10 * src0ComponentsCount + src1ComponentsCount;
11332
11333         switch (funcCode)
11334         {
11335                 case 222:return compareFP16VectorShuffleFunc<2, 2, 2>;
11336                 case 223:return compareFP16VectorShuffleFunc<2, 2, 3>;
11337                 case 224:return compareFP16VectorShuffleFunc<2, 2, 4>;
11338                 case 232:return compareFP16VectorShuffleFunc<2, 3, 2>;
11339                 case 233:return compareFP16VectorShuffleFunc<2, 3, 3>;
11340                 case 234:return compareFP16VectorShuffleFunc<2, 3, 4>;
11341                 case 242:return compareFP16VectorShuffleFunc<2, 4, 2>;
11342                 case 243:return compareFP16VectorShuffleFunc<2, 4, 3>;
11343                 case 244:return compareFP16VectorShuffleFunc<2, 4, 4>;
11344                 case 322:return compareFP16VectorShuffleFunc<3, 2, 2>;
11345                 case 323:return compareFP16VectorShuffleFunc<3, 2, 3>;
11346                 case 324:return compareFP16VectorShuffleFunc<3, 2, 4>;
11347                 case 332:return compareFP16VectorShuffleFunc<3, 3, 2>;
11348                 case 333:return compareFP16VectorShuffleFunc<3, 3, 3>;
11349                 case 334:return compareFP16VectorShuffleFunc<3, 3, 4>;
11350                 case 342:return compareFP16VectorShuffleFunc<3, 4, 2>;
11351                 case 343:return compareFP16VectorShuffleFunc<3, 4, 3>;
11352                 case 344:return compareFP16VectorShuffleFunc<3, 4, 4>;
11353                 case 422:return compareFP16VectorShuffleFunc<4, 2, 2>;
11354                 case 423:return compareFP16VectorShuffleFunc<4, 2, 3>;
11355                 case 424:return compareFP16VectorShuffleFunc<4, 2, 4>;
11356                 case 432:return compareFP16VectorShuffleFunc<4, 3, 2>;
11357                 case 433:return compareFP16VectorShuffleFunc<4, 3, 3>;
11358                 case 434:return compareFP16VectorShuffleFunc<4, 3, 4>;
11359                 case 442:return compareFP16VectorShuffleFunc<4, 4, 2>;
11360                 case 443:return compareFP16VectorShuffleFunc<4, 4, 3>;
11361                 case 444:return compareFP16VectorShuffleFunc<4, 4, 4>;
11362                 default: TCU_THROW(InternalError, "Invalid number of components specified.");
11363         }
11364 }
11365
11366 template<class SpecResource>
11367 tcu::TestCaseGroup* createFloat16VectorShuffleSet (tcu::TestContext& testCtx)
11368 {
11369         de::MovePtr<tcu::TestCaseGroup>         testGroup                       (new tcu::TestCaseGroup(testCtx, "opvectorshuffle", "OpVectorShuffle tests"));
11370         const int                                                       testSpecificSeed        = deStringHash(testGroup->getName());
11371         const int                                                       seed                            = testCtx.getCommandLine().getBaseSeed() ^ testSpecificSeed;
11372         de::Random                                                      rnd                                     (seed);
11373         const deUint32                                          numDataPoints           = 128;
11374         map<string, string>                                     fragments;
11375
11376         struct TestType
11377         {
11378                 const deUint32  typeComponents;
11379                 const char*             typeName;
11380         };
11381
11382         const TestType  testTypes[]     =
11383         {
11384                 {
11385                         2,
11386                         "v2f16",
11387                 },
11388                 {
11389                         3,
11390                         "v3f16",
11391                 },
11392                 {
11393                         4,
11394                         "v4f16",
11395                 },
11396         };
11397
11398         const StringTemplate preMain
11399         (
11400                 "    %c_i32_ndp = OpConstant %i32 ${num_data_points}\n"
11401                 "     %c_i32_cc = OpConstant %i32 ${case_count}\n"
11402                 "          %f16 = OpTypeFloat 16\n"
11403                 "        %v2f16 = OpTypeVector %f16 2\n"
11404                 "        %v3f16 = OpTypeVector %f16 3\n"
11405                 "        %v4f16 = OpTypeVector %f16 4\n"
11406
11407                 "     %up_v2f16 = OpTypePointer Uniform %v2f16\n"
11408                 "     %ra_v2f16 = OpTypeArray %v2f16 %c_i32_ndp\n"
11409                 "   %SSBO_v2f16 = OpTypeStruct %ra_v2f16\n"
11410                 "%up_SSBO_v2f16 = OpTypePointer Uniform %SSBO_v2f16\n"
11411
11412                 "     %up_v3f16 = OpTypePointer Uniform %v3f16\n"
11413                 "     %ra_v3f16 = OpTypeArray %v3f16 %c_i32_ndp\n"
11414                 "   %SSBO_v3f16 = OpTypeStruct %ra_v3f16\n"
11415                 "%up_SSBO_v3f16 = OpTypePointer Uniform %SSBO_v3f16\n"
11416
11417                 "     %up_v4f16 = OpTypePointer Uniform %v4f16\n"
11418                 "     %ra_v4f16 = OpTypeArray %v4f16 %c_i32_ndp\n"
11419                 "   %SSBO_v4f16 = OpTypeStruct %ra_v4f16\n"
11420                 "%up_SSBO_v4f16 = OpTypePointer Uniform %SSBO_v4f16\n"
11421
11422                 "        %fun_t = OpTypeFunction %${tt_dst} %${tt_src0} %${tt_src1} %i32\n"
11423
11424                 "    %ssbo_src0 = OpVariable %up_SSBO_${tt_src0} Uniform\n"
11425                 "    %ssbo_src1 = OpVariable %up_SSBO_${tt_src1} Uniform\n"
11426                 "     %ssbo_dst = OpVariable %up_SSBO_${tt_dst} Uniform\n"
11427         );
11428
11429         const StringTemplate decoration
11430         (
11431                 "OpDecorate %ra_v2f16 ArrayStride 4\n"
11432                 "OpDecorate %ra_v3f16 ArrayStride 8\n"
11433                 "OpDecorate %ra_v4f16 ArrayStride 8\n"
11434
11435                 "OpMemberDecorate %SSBO_v2f16 0 Offset 0\n"
11436                 "OpDecorate %SSBO_v2f16 BufferBlock\n"
11437
11438                 "OpMemberDecorate %SSBO_v3f16 0 Offset 0\n"
11439                 "OpDecorate %SSBO_v3f16 BufferBlock\n"
11440
11441                 "OpMemberDecorate %SSBO_v4f16 0 Offset 0\n"
11442                 "OpDecorate %SSBO_v4f16 BufferBlock\n"
11443
11444                 "OpDecorate %ssbo_src0 DescriptorSet 0\n"
11445                 "OpDecorate %ssbo_src0 Binding 0\n"
11446                 "OpDecorate %ssbo_src1 DescriptorSet 0\n"
11447                 "OpDecorate %ssbo_src1 Binding 1\n"
11448                 "OpDecorate %ssbo_dst DescriptorSet 0\n"
11449                 "OpDecorate %ssbo_dst Binding 2\n"
11450         );
11451
11452         const StringTemplate testFun
11453         (
11454                 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
11455                 "    %param = OpFunctionParameter %v4f32\n"
11456                 "    %entry = OpLabel\n"
11457
11458                 "        %i = OpVariable %fp_i32 Function\n"
11459                 "             OpStore %i %c_i32_0\n"
11460
11461                 " %will_run = OpFunctionCall %bool %isUniqueIdZero\n"
11462                 "             OpSelectionMerge %end_if None\n"
11463                 "             OpBranchConditional %will_run %run_test %end_if\n"
11464
11465                 " %run_test = OpLabel\n"
11466                 "             OpBranch %loop\n"
11467
11468                 "     %loop = OpLabel\n"
11469                 "    %i_cmp = OpLoad %i32 %i\n"
11470                 "       %lt = OpSLessThan %bool %i_cmp %c_i32_ndp\n"
11471                 "             OpLoopMerge %merge %next None\n"
11472                 "             OpBranchConditional %lt %write %merge\n"
11473
11474                 "    %write = OpLabel\n"
11475                 "      %ndx = OpLoad %i32 %i\n"
11476                 "     %src0 = OpAccessChain %up_${tt_src0} %ssbo_src0 %c_i32_0 %ndx\n"
11477                 " %val_src0 = OpLoad %${tt_src0} %src0\n"
11478                 "     %src1 = OpAccessChain %up_${tt_src1} %ssbo_src1 %c_i32_0 %ndx\n"
11479                 " %val_src1 = OpLoad %${tt_src1} %src1\n"
11480                 "  %val_dst = OpFunctionCall %${tt_dst} %sw_fun %val_src0 %val_src1 %ndx\n"
11481                 "      %dst = OpAccessChain %up_${tt_dst} %ssbo_dst %c_i32_0 %ndx\n"
11482                 "             OpStore %dst %val_dst\n"
11483                 "             OpBranch %next\n"
11484
11485                 "     %next = OpLabel\n"
11486                 "    %i_cur = OpLoad %i32 %i\n"
11487                 "    %i_new = OpIAdd %i32 %i_cur %c_i32_1\n"
11488                 "             OpStore %i %i_new\n"
11489                 "             OpBranch %loop\n"
11490
11491                 "    %merge = OpLabel\n"
11492                 "             OpBranch %end_if\n"
11493                 "   %end_if = OpLabel\n"
11494                 "             OpReturnValue %param\n"
11495                 "             OpFunctionEnd\n"
11496                 "\n"
11497
11498                 "   %sw_fun = OpFunction %${tt_dst} None %fun_t\n"
11499                 "%sw_param0 = OpFunctionParameter %${tt_src0}\n"
11500                 "%sw_param1 = OpFunctionParameter %${tt_src1}\n"
11501                 "%sw_paramn = OpFunctionParameter %i32\n"
11502                 " %sw_entry = OpLabel\n"
11503                 "   %modulo = OpSMod %i32 %sw_paramn %c_i32_cc\n"
11504                 "             OpSelectionMerge %switch_e None\n"
11505                 "             OpSwitch %modulo %default ${case_list}\n"
11506                 "${case_bodies}"
11507                 "%default   = OpLabel\n"
11508                 "             OpUnreachable\n" // Unreachable default case for switch statement
11509                 "%switch_e  = OpLabel\n"
11510                 "             OpUnreachable\n" // Unreachable merge block for switch statement
11511                 "             OpFunctionEnd\n"
11512         );
11513
11514         const StringTemplate testCaseBody
11515         (
11516                 "%case_${case_ndx}    = OpLabel\n"
11517                 "%val_dst_${case_ndx} = OpVectorShuffle %${tt_dst} %sw_param0 %sw_param1 ${shuffle}\n"
11518                 "             OpReturnValue %val_dst_${case_ndx}\n"
11519         );
11520
11521         for (deUint32 dstTypeIdx = 0; dstTypeIdx < DE_LENGTH_OF_ARRAY(testTypes); ++dstTypeIdx)
11522         {
11523                 const TestType& dstType                 = testTypes[dstTypeIdx];
11524
11525                 for (deUint32 comp0Idx = 0; comp0Idx < DE_LENGTH_OF_ARRAY(testTypes); ++comp0Idx)
11526                 {
11527                         const TestType& src0Type        = testTypes[comp0Idx];
11528
11529                         for (deUint32 comp1Idx = 0; comp1Idx < DE_LENGTH_OF_ARRAY(testTypes); ++comp1Idx)
11530                         {
11531                                 const TestType&                 src1Type                        = testTypes[comp1Idx];
11532                                 const deUint32                  input0Stride            = (src0Type.typeComponents == 3) ? 4 : src0Type.typeComponents;
11533                                 const deUint32                  input1Stride            = (src1Type.typeComponents == 3) ? 4 : src1Type.typeComponents;
11534                                 const deUint32                  outputStride            = (dstType.typeComponents == 3) ? 4 : dstType.typeComponents;
11535                                 const vector<deFloat16> float16Input0Data       = getFloat16s(rnd, input0Stride * numDataPoints);
11536                                 const vector<deFloat16> float16Input1Data       = getFloat16s(rnd, input1Stride * numDataPoints);
11537                                 const vector<deFloat16> float16OutputDummy      (outputStride * numDataPoints, 0);
11538                                 const string                    testName                        = de::toString(dstType.typeComponents) + de::toString(src0Type.typeComponents) + de::toString(src1Type.typeComponents);
11539                                 deUint32                                caseCount                       = 0;
11540                                 SpecResource                    specResource;
11541                                 map<string, string>             specs;
11542                                 vector<string>                  extensions;
11543                                 VulkanFeatures                  features;
11544                                 string                                  caseBodies;
11545                                 string                                  caseList;
11546
11547                                 // Generate case
11548                                 {
11549                                         vector<string>  componentList;
11550
11551                                         // Generate component possible indices for OpVectorShuffle for components 0 and 1 in output vector
11552                                         {
11553                                                 deUint32                caseNo          = 0;
11554
11555                                                 for (deUint32 comp0IdxLocal = 0; comp0IdxLocal < src0Type.typeComponents; ++comp0IdxLocal)
11556                                                         componentList.push_back(de::toString(caseNo++));
11557                                                 for (deUint32 comp1IdxLocal = 0; comp1IdxLocal < src1Type.typeComponents; ++comp1IdxLocal)
11558                                                         componentList.push_back(de::toString(caseNo++));
11559                                                 componentList.push_back("0xFFFFFFFF");
11560                                         }
11561
11562                                         for (deUint32 comp0IdxLocal = 0; comp0IdxLocal < componentList.size(); ++comp0IdxLocal)
11563                                         {
11564                                                 for (deUint32 comp1IdxLocal = 0; comp1IdxLocal < componentList.size(); ++comp1IdxLocal)
11565                                                 {
11566                                                         map<string, string>     specCase;
11567                                                         string                          shuffle         = componentList[comp0IdxLocal] + " " + componentList[comp1IdxLocal];
11568
11569                                                         for (deUint32 compIdx = 2; compIdx < dstType.typeComponents; ++compIdx)
11570                                                                 shuffle += " " + de::toString(compIdx - 2);
11571
11572                                                         specCase["case_ndx"]    = de::toString(caseCount);
11573                                                         specCase["shuffle"]             = shuffle;
11574                                                         specCase["tt_dst"]              = dstType.typeName;
11575
11576                                                         caseBodies      += testCaseBody.specialize(specCase);
11577                                                         caseList        += de::toString(caseCount) + " %case_" + de::toString(caseCount) + " ";
11578
11579                                                         caseCount++;
11580                                                 }
11581                                         }
11582                                 }
11583
11584                                 specs["num_data_points"]        = de::toString(numDataPoints);
11585                                 specs["tt_dst"]                         = dstType.typeName;
11586                                 specs["tt_src0"]                        = src0Type.typeName;
11587                                 specs["tt_src1"]                        = src1Type.typeName;
11588                                 specs["case_bodies"]            = caseBodies;
11589                                 specs["case_list"]                      = caseList;
11590                                 specs["case_count"]                     = de::toString(caseCount);
11591
11592                                 fragments["extension"]          = "OpExtension \"SPV_KHR_16bit_storage\"";
11593                                 fragments["capability"]         = "OpCapability StorageUniformBufferBlock16\n";
11594                                 fragments["decoration"]         = decoration.specialize(specs);
11595                                 fragments["pre_main"]           = preMain.specialize(specs);
11596                                 fragments["testfun"]            = testFun.specialize(specs);
11597
11598                                 specResource.inputs.push_back(Resource(BufferSp(new Float16Buffer(float16Input0Data)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
11599                                 specResource.inputs.push_back(Resource(BufferSp(new Float16Buffer(float16Input1Data)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
11600                                 specResource.outputs.push_back(Resource(BufferSp(new Float16Buffer(float16OutputDummy)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
11601                                 specResource.verifyIO = getFloat16VectorShuffleVerifyIOFunc(dstType.typeComponents, src0Type.typeComponents, src1Type.typeComponents);
11602
11603                                 extensions.push_back("VK_KHR_16bit_storage");
11604                                 extensions.push_back("VK_KHR_shader_float16_int8");
11605
11606                                 features.extFloat16Int8         = EXTFLOAT16INT8FEATURES_FLOAT16;
11607                                 features.ext16BitStorage        = EXT16BITSTORAGEFEATURES_UNIFORM_BUFFER_BLOCK;
11608
11609                                 finalizeTestsCreation(specResource, fragments, testCtx, *testGroup.get(), testName, features, extensions, IVec3(1, 1, 1));
11610                         }
11611                 }
11612         }
11613
11614         return testGroup.release();
11615 }
11616
11617 bool compareFP16CompositeFunc (const std::vector<Resource>& inputs, const vector<AllocationSp>& outputAllocs, const std::vector<Resource>&, TestLog& log)
11618 {
11619         if (inputs.size() != 1 || outputAllocs.size() != 1)
11620                 return false;
11621
11622         vector<deUint8> input1Bytes;
11623
11624         inputs[0].getBytes(input1Bytes);
11625
11626         DE_ASSERT(input1Bytes.size() > 0);
11627         DE_ASSERT(input1Bytes.size() % sizeof(deFloat16) == 0);
11628
11629         const size_t                    iterations              = input1Bytes.size() / sizeof(deFloat16);
11630         const deFloat16* const  input1AsFP16    = (const deFloat16*)&input1Bytes[0];
11631         const deFloat16* const  outputAsFP16    = (const deFloat16*)outputAllocs[0]->getHostPtr();
11632         const deFloat16                 exceptionValue  = tcu::Float16(-1.0).bits();
11633         std::string                             error;
11634
11635         for (size_t idx = 0; idx < iterations; ++idx)
11636         {
11637                 if (input1AsFP16[idx] == exceptionValue)
11638                         continue;
11639
11640                 if (!compare16BitFloat(input1AsFP16[idx], outputAsFP16[idx], error))
11641                 {
11642                         log << TestLog::Message << "At " << idx << ":" << error << TestLog::EndMessage;
11643
11644                         return false;
11645                 }
11646         }
11647
11648         return true;
11649 }
11650
11651 template<class SpecResource>
11652 tcu::TestCaseGroup* createFloat16CompositeConstructSet (tcu::TestContext& testCtx)
11653 {
11654         de::MovePtr<tcu::TestCaseGroup>         testGroup                               (new tcu::TestCaseGroup(testCtx, "opcompositeconstruct", "OpCompositeConstruct tests"));
11655         const deUint32                                          numElements                             = 8;
11656         const string                                            testName                                = "struct";
11657         const deUint32                                          structItemsCount                = 88;
11658         const deUint32                                          exceptionIndices[]              = { 1, 7, 15, 17, 25, 33, 51, 55, 59, 63, 67, 71, 84, 85, 86, 87 };
11659         const deFloat16                                         exceptionValue                  = tcu::Float16(-1.0).bits();
11660         const deUint32                                          fieldModifier                   = 2;
11661         const deUint32                                          fieldModifiedMulIndex   = 60;
11662         const deUint32                                          fieldModifiedAddIndex   = 66;
11663
11664         const StringTemplate preMain
11665         (
11666                 "    %c_i32_ndp = OpConstant %i32 ${num_elements}\n"
11667                 "          %f16 = OpTypeFloat 16\n"
11668                 "        %v2f16 = OpTypeVector %f16 2\n"
11669                 "        %v3f16 = OpTypeVector %f16 3\n"
11670                 "        %v4f16 = OpTypeVector %f16 4\n"
11671                 "    %c_f16_mod = OpConstant %f16 ${field_modifier}\n"
11672
11673                 "${consts}"
11674
11675                 "      %c_u32_5 = OpConstant %u32 5\n"
11676
11677                 " %f16arr3      = OpTypeArray %f16 %c_u32_3\n"
11678                 " %v2f16arr3    = OpTypeArray %v2f16 %c_u32_3\n"
11679                 " %v2f16arr5    = OpTypeArray %v2f16 %c_u32_5\n"
11680                 " %v3f16arr5    = OpTypeArray %v3f16 %c_u32_5\n"
11681                 " %v4f16arr3    = OpTypeArray %v4f16 %c_u32_3\n"
11682                 " %struct16     = OpTypeStruct %f16 %v2f16arr3\n"
11683                 " %struct16arr3 = OpTypeArray %struct16 %c_u32_3\n"
11684                 " %st_test      = OpTypeStruct %f16 %v2f16 %v3f16 %v4f16 %f16arr3 %struct16arr3 %v2f16arr5 %f16 %v3f16arr5 %v4f16arr3\n"
11685
11686                 "        %up_st = OpTypePointer Uniform %st_test\n"
11687                 "        %ra_st = OpTypeArray %st_test %c_i32_ndp\n"
11688                 "      %SSBO_st = OpTypeStruct %ra_st\n"
11689                 "   %up_SSBO_st = OpTypePointer Uniform %SSBO_st\n"
11690
11691                 "     %ssbo_dst = OpVariable %up_SSBO_st Uniform\n"
11692         );
11693
11694         const StringTemplate decoration
11695         (
11696                 "OpDecorate %SSBO_st BufferBlock\n"
11697                 "OpDecorate %ra_st ArrayStride ${struct_item_size}\n"
11698                 "OpDecorate %ssbo_dst DescriptorSet 0\n"
11699                 "OpDecorate %ssbo_dst Binding 1\n"
11700
11701                 "OpMemberDecorate %SSBO_st 0 Offset 0\n"
11702
11703                 "OpDecorate %v2f16arr3 ArrayStride 4\n"
11704                 "OpMemberDecorate %struct16 0 Offset 0\n"
11705                 "OpMemberDecorate %struct16 1 Offset 4\n"
11706                 "OpDecorate %struct16arr3 ArrayStride 16\n"
11707                 "OpDecorate %f16arr3 ArrayStride 2\n"
11708                 "OpDecorate %v2f16arr5 ArrayStride 4\n"
11709                 "OpDecorate %v3f16arr5 ArrayStride 8\n"
11710                 "OpDecorate %v4f16arr3 ArrayStride 8\n"
11711
11712                 "OpMemberDecorate %st_test 0 Offset 0\n"
11713                 "OpMemberDecorate %st_test 1 Offset 4\n"
11714                 "OpMemberDecorate %st_test 2 Offset 8\n"
11715                 "OpMemberDecorate %st_test 3 Offset 16\n"
11716                 "OpMemberDecorate %st_test 4 Offset 24\n"
11717                 "OpMemberDecorate %st_test 5 Offset 32\n"
11718                 "OpMemberDecorate %st_test 6 Offset 80\n"
11719                 "OpMemberDecorate %st_test 7 Offset 100\n"
11720                 "OpMemberDecorate %st_test 8 Offset 104\n"
11721                 "OpMemberDecorate %st_test 9 Offset 144\n"
11722         );
11723
11724         const StringTemplate testFun
11725         (
11726                 " %test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
11727                 "     %param = OpFunctionParameter %v4f32\n"
11728                 "     %entry = OpLabel\n"
11729
11730                 "         %i = OpVariable %fp_i32 Function\n"
11731                 "              OpStore %i %c_i32_0\n"
11732
11733                 "  %will_run = OpFunctionCall %bool %isUniqueIdZero\n"
11734                 "              OpSelectionMerge %end_if None\n"
11735                 "              OpBranchConditional %will_run %run_test %end_if\n"
11736
11737                 "  %run_test = OpLabel\n"
11738                 "              OpBranch %loop\n"
11739
11740                 "      %loop = OpLabel\n"
11741                 "     %i_cmp = OpLoad %i32 %i\n"
11742                 "        %lt = OpSLessThan %bool %i_cmp %c_i32_ndp\n"
11743                 "              OpLoopMerge %merge %next None\n"
11744                 "              OpBranchConditional %lt %write %merge\n"
11745
11746                 "     %write = OpLabel\n"
11747                 "       %ndx = OpLoad %i32 %i\n"
11748
11749                 "      %fld1 = OpCompositeConstruct %v2f16 %c_f16_2 %c_f16_3\n"
11750                 "      %fld2 = OpCompositeConstruct %v3f16 %c_f16_4 %c_f16_5 %c_f16_6\n"
11751                 "      %fld3 = OpCompositeConstruct %v4f16 %c_f16_8 %c_f16_9 %c_f16_10 %c_f16_11\n"
11752
11753                 "      %fld4 = OpCompositeConstruct %f16arr3 %c_f16_12 %c_f16_13 %c_f16_14\n"
11754
11755                 "%fld5_0_1_0 = OpCompositeConstruct %v2f16 %c_f16_18 %c_f16_19\n"
11756                 "%fld5_0_1_1 = OpCompositeConstruct %v2f16 %c_f16_20 %c_f16_21\n"
11757                 "%fld5_0_1_2 = OpCompositeConstruct %v2f16 %c_f16_22 %c_f16_23\n"
11758                 "  %fld5_0_1 = OpCompositeConstruct %v2f16arr3 %fld5_0_1_0 %fld5_0_1_1 %fld5_0_1_2\n"
11759                 "    %fld5_0 = OpCompositeConstruct %struct16 %c_f16_16 %fld5_0_1\n"
11760
11761                 "%fld5_1_1_0 = OpCompositeConstruct %v2f16 %c_f16_26 %c_f16_27\n"
11762                 "%fld5_1_1_1 = OpCompositeConstruct %v2f16 %c_f16_28 %c_f16_29\n"
11763                 "%fld5_1_1_2 = OpCompositeConstruct %v2f16 %c_f16_30 %c_f16_31\n"
11764                 "  %fld5_1_1 = OpCompositeConstruct %v2f16arr3 %fld5_1_1_0 %fld5_1_1_1 %fld5_1_1_2\n"
11765                 "    %fld5_1 = OpCompositeConstruct %struct16 %c_f16_24 %fld5_1_1\n"
11766
11767                 "%fld5_2_1_0 = OpCompositeConstruct %v2f16 %c_f16_34 %c_f16_35\n"
11768                 "%fld5_2_1_1 = OpCompositeConstruct %v2f16 %c_f16_36 %c_f16_37\n"
11769                 "%fld5_2_1_2 = OpCompositeConstruct %v2f16 %c_f16_38 %c_f16_39\n"
11770                 "  %fld5_2_1 = OpCompositeConstruct %v2f16arr3 %fld5_2_1_0 %fld5_2_1_1 %fld5_2_1_2\n"
11771                 "    %fld5_2 = OpCompositeConstruct %struct16 %c_f16_32 %fld5_2_1\n"
11772
11773                 "      %fld5 = OpCompositeConstruct %struct16arr3 %fld5_0 %fld5_1 %fld5_2\n"
11774
11775                 "    %fld6_0 = OpCompositeConstruct %v2f16 %c_f16_40 %c_f16_41\n"
11776                 "    %fld6_1 = OpCompositeConstruct %v2f16 %c_f16_42 %c_f16_43\n"
11777                 "    %fld6_2 = OpCompositeConstruct %v2f16 %c_f16_44 %c_f16_45\n"
11778                 "    %fld6_3 = OpCompositeConstruct %v2f16 %c_f16_46 %c_f16_47\n"
11779                 "    %fld6_4 = OpCompositeConstruct %v2f16 %c_f16_48 %c_f16_49\n"
11780                 "      %fld6 = OpCompositeConstruct %v2f16arr5 %fld6_0 %fld6_1 %fld6_2 %fld6_3 %fld6_4\n"
11781
11782                 "      %fndx = OpConvertSToF %f16 %ndx\n"
11783                 "  %fld8_2a0 = OpFMul %f16 %fndx %c_f16_mod\n"
11784                 "  %fld8_3b1 = OpFAdd %f16 %fndx %c_f16_mod\n"
11785
11786                 "   %fld8_2a = OpCompositeConstruct %v2f16 %fld8_2a0 %c_f16_61\n"
11787                 "   %fld8_3b = OpCompositeConstruct %v2f16 %c_f16_65 %fld8_3b1\n"
11788                 "    %fld8_0 = OpCompositeConstruct %v3f16 %c_f16_52 %c_f16_53 %c_f16_54\n"
11789                 "    %fld8_1 = OpCompositeConstruct %v3f16 %c_f16_56 %c_f16_57 %c_f16_58\n"
11790                 "    %fld8_2 = OpCompositeConstruct %v3f16 %fld8_2a %c_f16_62\n"
11791                 "    %fld8_3 = OpCompositeConstruct %v3f16 %c_f16_64 %fld8_3b\n"
11792                 "    %fld8_4 = OpCompositeConstruct %v3f16 %c_f16_68 %c_f16_69 %c_f16_70\n"
11793                 "      %fld8 = OpCompositeConstruct %v3f16arr5 %fld8_0 %fld8_1 %fld8_2 %fld8_3 %fld8_4\n"
11794
11795                 "    %fld9_0 = OpCompositeConstruct %v4f16 %c_f16_72 %c_f16_73 %c_f16_74 %c_f16_75\n"
11796                 "    %fld9_1 = OpCompositeConstruct %v4f16 %c_f16_76 %c_f16_77 %c_f16_78 %c_f16_79\n"
11797                 "    %fld9_2 = OpCompositeConstruct %v4f16 %c_f16_80 %c_f16_81 %c_f16_82 %c_f16_83\n"
11798                 "      %fld9 = OpCompositeConstruct %v4f16arr3 %fld9_0 %fld9_1 %fld9_2\n"
11799
11800                 "    %st_val = OpCompositeConstruct %st_test %c_f16_0 %fld1 %fld2 %fld3 %fld4 %fld5 %fld6 %c_f16_50 %fld8 %fld9\n"
11801                 "       %dst = OpAccessChain %up_st %ssbo_dst %c_i32_0 %ndx\n"
11802                 "              OpStore %dst %st_val\n"
11803
11804                 "              OpBranch %next\n"
11805
11806                 "      %next = OpLabel\n"
11807                 "     %i_cur = OpLoad %i32 %i\n"
11808                 "     %i_new = OpIAdd %i32 %i_cur %c_i32_1\n"
11809                 "              OpStore %i %i_new\n"
11810                 "              OpBranch %loop\n"
11811
11812                 "     %merge = OpLabel\n"
11813                 "              OpBranch %end_if\n"
11814                 "    %end_if = OpLabel\n"
11815                 "              OpReturnValue %param\n"
11816                 "              OpFunctionEnd\n"
11817         );
11818
11819         {
11820                 SpecResource            specResource;
11821                 map<string, string>     specs;
11822                 VulkanFeatures          features;
11823                 map<string, string>     fragments;
11824                 vector<string>          extensions;
11825                 vector<deFloat16>       expectedOutput;
11826                 string                          consts;
11827
11828                 for (deUint32 elementNdx = 0; elementNdx < numElements; ++elementNdx)
11829                 {
11830                         vector<deFloat16>       expectedIterationOutput;
11831
11832                         for (deUint32 structItemNdx = 0; structItemNdx < structItemsCount; ++structItemNdx)
11833                                 expectedIterationOutput.push_back(tcu::Float16(float(structItemNdx)).bits());
11834
11835                         for (deUint32 structItemNdx = 0; structItemNdx < DE_LENGTH_OF_ARRAY(exceptionIndices); ++structItemNdx)
11836                                 expectedIterationOutput[exceptionIndices[structItemNdx]] = exceptionValue;
11837
11838                         expectedIterationOutput[fieldModifiedMulIndex] = tcu::Float16(float(elementNdx * fieldModifier)).bits();
11839                         expectedIterationOutput[fieldModifiedAddIndex] = tcu::Float16(float(elementNdx + fieldModifier)).bits();
11840
11841                         expectedOutput.insert(expectedOutput.end(), expectedIterationOutput.begin(), expectedIterationOutput.end());
11842                 }
11843
11844                 for (deUint32 i = 0; i < structItemsCount; ++i)
11845                         consts += "     %c_f16_" + de::toString(i) + " = OpConstant %f16 "  + de::toString(i) + "\n";
11846
11847                 specs["num_elements"]           = de::toString(numElements);
11848                 specs["struct_item_size"]       = de::toString(structItemsCount * sizeof(deFloat16));
11849                 specs["field_modifier"]         = de::toString(fieldModifier);
11850                 specs["consts"]                         = consts;
11851
11852                 fragments["extension"]          = "OpExtension \"SPV_KHR_16bit_storage\"";
11853                 fragments["capability"]         = "OpCapability StorageUniformBufferBlock16\n";
11854                 fragments["decoration"]         = decoration.specialize(specs);
11855                 fragments["pre_main"]           = preMain.specialize(specs);
11856                 fragments["testfun"]            = testFun.specialize(specs);
11857
11858                 specResource.inputs.push_back(Resource(BufferSp(new Float16Buffer(expectedOutput)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
11859                 specResource.outputs.push_back(Resource(BufferSp(new Float16Buffer(expectedOutput)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
11860                 specResource.verifyIO = compareFP16CompositeFunc;
11861
11862                 extensions.push_back("VK_KHR_16bit_storage");
11863                 extensions.push_back("VK_KHR_shader_float16_int8");
11864
11865                 features.extFloat16Int8         = EXTFLOAT16INT8FEATURES_FLOAT16;
11866                 features.ext16BitStorage        = EXT16BITSTORAGEFEATURES_UNIFORM_BUFFER_BLOCK;
11867
11868                 finalizeTestsCreation(specResource, fragments, testCtx, *testGroup.get(), testName, features, extensions, IVec3(1, 1, 1));
11869         }
11870
11871         return testGroup.release();
11872 }
11873
11874 template<class SpecResource>
11875 tcu::TestCaseGroup* createFloat16CompositeInsertExtractSet (tcu::TestContext& testCtx, const char* op)
11876 {
11877         de::MovePtr<tcu::TestCaseGroup>         testGroup               (new tcu::TestCaseGroup(testCtx, de::toLower(op).c_str(), op));
11878         const deFloat16                                         exceptionValue  = tcu::Float16(-1.0).bits();
11879         const string                                            opName                  (op);
11880         const deUint32                                          opIndex                 = (opName == "OpCompositeInsert") ? 0
11881                                                                                                                 : (opName == "OpCompositeExtract") ? 1
11882                                                                                                                 : -1;
11883
11884         const StringTemplate preMain
11885         (
11886                 "   %c_i32_ndp = OpConstant %i32 ${num_elements}\n"
11887                 "         %f16 = OpTypeFloat 16\n"
11888                 "       %v2f16 = OpTypeVector %f16 2\n"
11889                 "       %v3f16 = OpTypeVector %f16 3\n"
11890                 "       %v4f16 = OpTypeVector %f16 4\n"
11891                 "    %c_f16_na = OpConstant %f16 -1.0\n"
11892                 "     %c_u32_5 = OpConstant %u32 5\n"
11893
11894                 "%f16arr3      = OpTypeArray %f16 %c_u32_3\n"
11895                 "%v2f16arr3    = OpTypeArray %v2f16 %c_u32_3\n"
11896                 "%v2f16arr5    = OpTypeArray %v2f16 %c_u32_5\n"
11897                 "%v3f16arr5    = OpTypeArray %v3f16 %c_u32_5\n"
11898                 "%v4f16arr3    = OpTypeArray %v4f16 %c_u32_3\n"
11899                 "%struct16     = OpTypeStruct %f16 %v2f16arr3\n"
11900                 "%struct16arr3 = OpTypeArray %struct16 %c_u32_3\n"
11901                 "%st_test      = OpTypeStruct %${field_type}\n"
11902
11903                 "      %up_f16 = OpTypePointer Uniform %f16\n"
11904                 "       %up_st = OpTypePointer Uniform %st_test\n"
11905                 "      %ra_f16 = OpTypeArray %f16 %c_i32_ndp\n"
11906                 "       %ra_st = OpTypeArray %st_test %c_i32_1\n"
11907
11908                 "${op_premain_decls}"
11909
11910                 " %up_SSBO_src = OpTypePointer Uniform %SSBO_src\n"
11911                 " %up_SSBO_dst = OpTypePointer Uniform %SSBO_dst\n"
11912
11913                 "    %ssbo_src = OpVariable %up_SSBO_src Uniform\n"
11914                 "    %ssbo_dst = OpVariable %up_SSBO_dst Uniform\n"
11915         );
11916
11917         const StringTemplate decoration
11918         (
11919                 "OpDecorate %SSBO_src BufferBlock\n"
11920                 "OpDecorate %SSBO_dst BufferBlock\n"
11921                 "OpDecorate %ra_f16 ArrayStride 2\n"
11922                 "OpDecorate %ra_st ArrayStride ${struct_item_size}\n"
11923                 "OpDecorate %ssbo_src DescriptorSet 0\n"
11924                 "OpDecorate %ssbo_src Binding 0\n"
11925                 "OpDecorate %ssbo_dst DescriptorSet 0\n"
11926                 "OpDecorate %ssbo_dst Binding 1\n"
11927
11928                 "OpMemberDecorate %SSBO_src 0 Offset 0\n"
11929                 "OpMemberDecorate %SSBO_dst 0 Offset 0\n"
11930
11931                 "OpDecorate %v2f16arr3 ArrayStride 4\n"
11932                 "OpMemberDecorate %struct16 0 Offset 0\n"
11933                 "OpMemberDecorate %struct16 1 Offset 4\n"
11934                 "OpDecorate %struct16arr3 ArrayStride 16\n"
11935                 "OpDecorate %f16arr3 ArrayStride 2\n"
11936                 "OpDecorate %v2f16arr5 ArrayStride 4\n"
11937                 "OpDecorate %v3f16arr5 ArrayStride 8\n"
11938                 "OpDecorate %v4f16arr3 ArrayStride 8\n"
11939
11940                 "OpMemberDecorate %st_test 0 Offset 0\n"
11941         );
11942
11943         const StringTemplate testFun
11944         (
11945                 " %test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
11946                 "     %param = OpFunctionParameter %v4f32\n"
11947                 "     %entry = OpLabel\n"
11948
11949                 "         %i = OpVariable %fp_i32 Function\n"
11950                 "              OpStore %i %c_i32_0\n"
11951
11952                 "  %will_run = OpFunctionCall %bool %isUniqueIdZero\n"
11953                 "              OpSelectionMerge %end_if None\n"
11954                 "              OpBranchConditional %will_run %run_test %end_if\n"
11955
11956                 "  %run_test = OpLabel\n"
11957                 "              OpBranch %loop\n"
11958
11959                 "      %loop = OpLabel\n"
11960                 "     %i_cmp = OpLoad %i32 %i\n"
11961                 "        %lt = OpSLessThan %bool %i_cmp %c_i32_ndp\n"
11962                 "              OpLoopMerge %merge %next None\n"
11963                 "              OpBranchConditional %lt %write %merge\n"
11964
11965                 "     %write = OpLabel\n"
11966                 "       %ndx = OpLoad %i32 %i\n"
11967
11968                 "${op_sw_fun_call}"
11969
11970                 "              OpStore %dst %val_dst\n"
11971                 "              OpBranch %next\n"
11972
11973                 "      %next = OpLabel\n"
11974                 "     %i_cur = OpLoad %i32 %i\n"
11975                 "     %i_new = OpIAdd %i32 %i_cur %c_i32_1\n"
11976                 "              OpStore %i %i_new\n"
11977                 "              OpBranch %loop\n"
11978
11979                 "     %merge = OpLabel\n"
11980                 "              OpBranch %end_if\n"
11981                 "    %end_if = OpLabel\n"
11982                 "              OpReturnValue %param\n"
11983                 "              OpFunctionEnd\n"
11984
11985                 "${op_sw_fun_header}"
11986                 " %sw_param = OpFunctionParameter %st_test\n"
11987                 "%sw_paramn = OpFunctionParameter %i32\n"
11988                 " %sw_entry = OpLabel\n"
11989                 "             OpSelectionMerge %switch_e None\n"
11990                 "             OpSwitch %sw_paramn %default ${case_list}\n"
11991
11992                 "${case_bodies}"
11993
11994                 "%default   = OpLabel\n"
11995                 "             OpReturnValue ${op_case_default_value}\n"
11996                 "%switch_e  = OpLabel\n"
11997                 "             OpUnreachable\n" // Unreachable merge block for switch statement
11998                 "             OpFunctionEnd\n"
11999         );
12000
12001         const StringTemplate testCaseBody
12002         (
12003                 "%case_${case_ndx}    = OpLabel\n"
12004                 "%val_ret_${case_ndx} = ${op_name} ${op_args_part} ${access_path}\n"
12005                 "             OpReturnValue %val_ret_${case_ndx}\n"
12006         );
12007
12008         struct OpParts
12009         {
12010                 const char*     premainDecls;
12011                 const char*     swFunCall;
12012                 const char*     swFunHeader;
12013                 const char*     caseDefaultValue;
12014                 const char*     argsPartial;
12015         };
12016
12017         OpParts                                                         opPartsArray[]                  =
12018         {
12019                 // OpCompositeInsert
12020                 {
12021                         "       %fun_t = OpTypeFunction %st_test %f16 %st_test %i32\n"
12022                         "    %SSBO_src = OpTypeStruct %ra_f16\n"
12023                         "    %SSBO_dst = OpTypeStruct %ra_st\n",
12024
12025                         "       %src = OpAccessChain %up_f16 %ssbo_src %c_i32_0 %ndx\n"
12026                         "       %dst = OpAccessChain %up_st %ssbo_dst %c_i32_0 %c_i32_0\n"
12027                         "   %val_new = OpLoad %f16 %src\n"
12028                         "   %val_old = OpLoad %st_test %dst\n"
12029                         "   %val_dst = OpFunctionCall %st_test %sw_fun %val_new %val_old %ndx\n",
12030
12031                         "   %sw_fun = OpFunction %st_test None %fun_t\n"
12032                         "%sw_paramv = OpFunctionParameter %f16\n",
12033
12034                         "%sw_param",
12035
12036                         "%st_test %sw_paramv %sw_param",
12037                 },
12038                 // OpCompositeExtract
12039                 {
12040                         "       %fun_t = OpTypeFunction %f16 %st_test %i32\n"
12041                         "    %SSBO_src = OpTypeStruct %ra_st\n"
12042                         "    %SSBO_dst = OpTypeStruct %ra_f16\n",
12043
12044                         "       %src = OpAccessChain %up_st %ssbo_src %c_i32_0 %c_i32_0\n"
12045                         "       %dst = OpAccessChain %up_f16 %ssbo_dst %c_i32_0 %ndx\n"
12046                         "   %val_src = OpLoad %st_test %src\n"
12047                         "   %val_dst = OpFunctionCall %f16 %sw_fun %val_src %ndx\n",
12048
12049                         "   %sw_fun = OpFunction %f16 None %fun_t\n",
12050
12051                         "%c_f16_na",
12052
12053                         "%f16 %sw_param",
12054                 },
12055         };
12056
12057         DE_ASSERT(opIndex >= 0 && opIndex < DE_LENGTH_OF_ARRAY(opPartsArray));
12058
12059         const char*     accessPathF16[] =
12060         {
12061                 "0",                    // %f16
12062                 DE_NULL,
12063         };
12064         const char*     accessPathV2F16[] =
12065         {
12066                 "0 0",                  // %v2f16
12067                 "0 1",
12068         };
12069         const char*     accessPathV3F16[] =
12070         {
12071                 "0 0",                  // %v3f16
12072                 "0 1",
12073                 "0 2",
12074                 DE_NULL,
12075         };
12076         const char*     accessPathV4F16[] =
12077         {
12078                 "0 0",                  // %v4f16"
12079                 "0 1",
12080                 "0 2",
12081                 "0 3",
12082         };
12083         const char*     accessPathF16Arr3[] =
12084         {
12085                 "0 0",                  // %f16arr3
12086                 "0 1",
12087                 "0 2",
12088                 DE_NULL,
12089         };
12090         const char*     accessPathStruct16Arr3[] =
12091         {
12092                 "0 0 0",                // %struct16arr3
12093                 DE_NULL,
12094                 "0 0 1 0 0",
12095                 "0 0 1 0 1",
12096                 "0 0 1 1 0",
12097                 "0 0 1 1 1",
12098                 "0 0 1 2 0",
12099                 "0 0 1 2 1",
12100                 "0 1 0",
12101                 DE_NULL,
12102                 "0 1 1 0 0",
12103                 "0 1 1 0 1",
12104                 "0 1 1 1 0",
12105                 "0 1 1 1 1",
12106                 "0 1 1 2 0",
12107                 "0 1 1 2 1",
12108                 "0 2 0",
12109                 DE_NULL,
12110                 "0 2 1 0 0",
12111                 "0 2 1 0 1",
12112                 "0 2 1 1 0",
12113                 "0 2 1 1 1",
12114                 "0 2 1 2 0",
12115                 "0 2 1 2 1",
12116         };
12117         const char*     accessPathV2F16Arr5[] =
12118         {
12119                 "0 0 0",                // %v2f16arr5
12120                 "0 0 1",
12121                 "0 1 0",
12122                 "0 1 1",
12123                 "0 2 0",
12124                 "0 2 1",
12125                 "0 3 0",
12126                 "0 3 1",
12127                 "0 4 0",
12128                 "0 4 1",
12129         };
12130         const char*     accessPathV3F16Arr5[] =
12131         {
12132                 "0 0 0",                // %v3f16arr5
12133                 "0 0 1",
12134                 "0 0 2",
12135                 DE_NULL,
12136                 "0 1 0",
12137                 "0 1 1",
12138                 "0 1 2",
12139                 DE_NULL,
12140                 "0 2 0",
12141                 "0 2 1",
12142                 "0 2 2",
12143                 DE_NULL,
12144                 "0 3 0",
12145                 "0 3 1",
12146                 "0 3 2",
12147                 DE_NULL,
12148                 "0 4 0",
12149                 "0 4 1",
12150                 "0 4 2",
12151                 DE_NULL,
12152         };
12153         const char*     accessPathV4F16Arr3[] =
12154         {
12155                 "0 0 0",                // %v4f16arr3
12156                 "0 0 1",
12157                 "0 0 2",
12158                 "0 0 3",
12159                 "0 1 0",
12160                 "0 1 1",
12161                 "0 1 2",
12162                 "0 1 3",
12163                 "0 2 0",
12164                 "0 2 1",
12165                 "0 2 2",
12166                 "0 2 3",
12167                 DE_NULL,
12168                 DE_NULL,
12169                 DE_NULL,
12170                 DE_NULL,
12171         };
12172
12173         struct TypeTestParameters
12174         {
12175                 const char*             name;
12176                 size_t                  accessPathLength;
12177                 const char**    accessPath;
12178         };
12179
12180         const TypeTestParameters typeTestParameters[] =
12181         {
12182                 {       "f16",                  DE_LENGTH_OF_ARRAY(accessPathF16),                      accessPathF16                   },
12183                 {       "v2f16",                DE_LENGTH_OF_ARRAY(accessPathV2F16),            accessPathV2F16                 },
12184                 {       "v3f16",                DE_LENGTH_OF_ARRAY(accessPathV3F16),            accessPathV3F16                 },
12185                 {       "v4f16",                DE_LENGTH_OF_ARRAY(accessPathV4F16),            accessPathV4F16                 },
12186                 {       "f16arr3",              DE_LENGTH_OF_ARRAY(accessPathF16Arr3),          accessPathF16Arr3               },
12187                 {       "v2f16arr5",    DE_LENGTH_OF_ARRAY(accessPathV2F16Arr5),        accessPathV2F16Arr5             },
12188                 {       "v3f16arr5",    DE_LENGTH_OF_ARRAY(accessPathV3F16Arr5),        accessPathV3F16Arr5             },
12189                 {       "v4f16arr3",    DE_LENGTH_OF_ARRAY(accessPathV4F16Arr3),        accessPathV4F16Arr3             },
12190                 {       "struct16arr3", DE_LENGTH_OF_ARRAY(accessPathStruct16Arr3),     accessPathStruct16Arr3  },
12191         };
12192
12193         for (size_t typeTestNdx = 0; typeTestNdx < DE_LENGTH_OF_ARRAY(typeTestParameters); ++typeTestNdx)
12194         {
12195                 const OpParts           opParts                         = opPartsArray[opIndex];
12196                 const string            testName                        = typeTestParameters[typeTestNdx].name;
12197                 const size_t            structItemsCount        = typeTestParameters[typeTestNdx].accessPathLength;
12198                 const char**            accessPath                      = typeTestParameters[typeTestNdx].accessPath;
12199                 SpecResource            specResource;
12200                 map<string, string>     specs;
12201                 VulkanFeatures          features;
12202                 map<string, string>     fragments;
12203                 vector<string>          extensions;
12204                 vector<deFloat16>       inputFP16;
12205                 vector<deFloat16>       dummyFP16Output;
12206
12207                 // Generate values for input
12208                 inputFP16.reserve(structItemsCount);
12209                 for (deUint32 structItemNdx = 0; structItemNdx < structItemsCount; ++structItemNdx)
12210                         inputFP16.push_back((accessPath[structItemNdx] == DE_NULL) ? exceptionValue : tcu::Float16(float(structItemNdx)).bits());
12211
12212                 dummyFP16Output.resize(structItemsCount);
12213
12214                 // Generate cases for OpSwitch
12215                 {
12216                         string  caseBodies;
12217                         string  caseList;
12218
12219                         for (deUint32 caseNdx = 0; caseNdx < structItemsCount; ++caseNdx)
12220                                 if (accessPath[caseNdx] != DE_NULL)
12221                                 {
12222                                         map<string, string>     specCase;
12223
12224                                         specCase["case_ndx"]            = de::toString(caseNdx);
12225                                         specCase["access_path"]         = accessPath[caseNdx];
12226                                         specCase["op_args_part"]        = opParts.argsPartial;
12227                                         specCase["op_name"]                     = opName;
12228
12229                                         caseBodies      += testCaseBody.specialize(specCase);
12230                                         caseList        += de::toString(caseNdx) + " %case_" + de::toString(caseNdx) + " ";
12231                                 }
12232
12233                         specs["case_bodies"]    = caseBodies;
12234                         specs["case_list"]              = caseList;
12235                 }
12236
12237                 specs["num_elements"]                   = de::toString(structItemsCount);
12238                 specs["field_type"]                             = typeTestParameters[typeTestNdx].name;
12239                 specs["struct_item_size"]               = de::toString(structItemsCount * sizeof(deFloat16));
12240                 specs["op_premain_decls"]               = opParts.premainDecls;
12241                 specs["op_sw_fun_call"]                 = opParts.swFunCall;
12242                 specs["op_sw_fun_header"]               = opParts.swFunHeader;
12243                 specs["op_case_default_value"]  = opParts.caseDefaultValue;
12244
12245                 fragments["extension"]          = "OpExtension \"SPV_KHR_16bit_storage\"";
12246                 fragments["capability"]         = "OpCapability StorageUniformBufferBlock16\n";
12247                 fragments["decoration"]         = decoration.specialize(specs);
12248                 fragments["pre_main"]           = preMain.specialize(specs);
12249                 fragments["testfun"]            = testFun.specialize(specs);
12250
12251                 specResource.inputs.push_back(Resource(BufferSp(new Float16Buffer(inputFP16)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
12252                 specResource.outputs.push_back(Resource(BufferSp(new Float16Buffer(dummyFP16Output)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
12253                 specResource.verifyIO = compareFP16CompositeFunc;
12254
12255                 extensions.push_back("VK_KHR_16bit_storage");
12256                 extensions.push_back("VK_KHR_shader_float16_int8");
12257
12258                 features.extFloat16Int8         = EXTFLOAT16INT8FEATURES_FLOAT16;
12259                 features.ext16BitStorage        = EXT16BITSTORAGEFEATURES_UNIFORM_BUFFER_BLOCK;
12260
12261                 finalizeTestsCreation(specResource, fragments, testCtx, *testGroup.get(), testName, features, extensions, IVec3(1, 1, 1));
12262         }
12263
12264         return testGroup.release();
12265 }
12266
12267 struct fp16PerComponent
12268 {
12269         fp16PerComponent()
12270                 : flavor(0)
12271                 , floatFormat16 (-14, 15, 10, true)
12272                 , outCompCount(0)
12273                 , argCompCount(3, 0)
12274         {
12275         }
12276
12277         bool                    callOncePerComponent    ()                                                                      { return true; }
12278         deUint32                getComponentValidity    ()                                                                      { return static_cast<deUint32>(-1); }
12279
12280         virtual double  getULPs                                 (vector<const deFloat16*>&)                     { return 1.0; }
12281         virtual double  getMin                                  (double value, double ulps)                     { return value - floatFormat16.ulp(deAbs(value), ulps); }
12282         virtual double  getMax                                  (double value, double ulps)                     { return value + floatFormat16.ulp(deAbs(value), ulps); }
12283
12284         virtual size_t  getFlavorCount                  ()                                                                      { return flavorNames.empty() ? 1 : flavorNames.size(); }
12285         virtual void    setFlavor                               (size_t flavorNo)                                       { DE_ASSERT(flavorNo < getFlavorCount()); flavor = flavorNo; }
12286         virtual size_t  getFlavor                               ()                                                                      { return flavor; }
12287         virtual string  getCurrentFlavorName    ()                                                                      { return flavorNames.empty() ? string("") : flavorNames[getFlavor()]; }
12288
12289         virtual void    setOutCompCount                 (size_t compCount)                                      { outCompCount = compCount; }
12290         virtual size_t  getOutCompCount                 ()                                                                      { return outCompCount; }
12291
12292         virtual void    setArgCompCount                 (size_t argNo, size_t compCount)        { argCompCount[argNo] = compCount; }
12293         virtual size_t  getArgCompCount                 (size_t argNo)                                          { return argCompCount[argNo]; }
12294
12295 protected:
12296         size_t                          flavor;
12297         tcu::FloatFormat        floatFormat16;
12298         size_t                          outCompCount;
12299         vector<size_t>          argCompCount;
12300         vector<string>          flavorNames;
12301 };
12302
12303 struct fp16OpFNegate : public fp16PerComponent
12304 {
12305         template <class fp16type>
12306         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
12307         {
12308                 const fp16type  x               (*in[0]);
12309                 const double    d               (x.asDouble());
12310                 const double    result  (0.0 - d);
12311
12312                 out[0] = fp16type(result).bits();
12313                 min[0] = getMin(result, getULPs(in));
12314                 max[0] = getMax(result, getULPs(in));
12315
12316                 return true;
12317         }
12318 };
12319
12320 struct fp16Round : public fp16PerComponent
12321 {
12322         fp16Round() : fp16PerComponent()
12323         {
12324                 flavorNames.push_back("Floor(x+0.5)");
12325                 flavorNames.push_back("Floor(x-0.5)");
12326                 flavorNames.push_back("RoundEven");
12327         }
12328
12329         template<class fp16type>
12330         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
12331         {
12332                 const fp16type  x               (*in[0]);
12333                 const double    d               (x.asDouble());
12334                 double                  result  (0.0);
12335
12336                 switch (flavor)
12337                 {
12338                         case 0:         result = deRound(d);            break;
12339                         case 1:         result = deFloor(d - 0.5);      break;
12340                         case 2:         result = deRoundEven(d);        break;
12341                         default:        TCU_THROW(InternalError, "Invalid flavor specified");
12342                 }
12343
12344                 out[0] = fp16type(result).bits();
12345                 min[0] = getMin(result, getULPs(in));
12346                 max[0] = getMax(result, getULPs(in));
12347
12348                 return true;
12349         }
12350 };
12351
12352 struct fp16RoundEven : public fp16PerComponent
12353 {
12354         template<class fp16type>
12355         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
12356         {
12357                 const fp16type  x               (*in[0]);
12358                 const double    d               (x.asDouble());
12359                 const double    result  (deRoundEven(d));
12360
12361                 out[0] = fp16type(result).bits();
12362                 min[0] = getMin(result, getULPs(in));
12363                 max[0] = getMax(result, getULPs(in));
12364
12365                 return true;
12366         }
12367 };
12368
12369 struct fp16Trunc : public fp16PerComponent
12370 {
12371         template<class fp16type>
12372         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
12373         {
12374                 const fp16type  x               (*in[0]);
12375                 const double    d               (x.asDouble());
12376                 const double    result  (deTrunc(d));
12377
12378                 out[0] = fp16type(result).bits();
12379                 min[0] = getMin(result, getULPs(in));
12380                 max[0] = getMax(result, getULPs(in));
12381
12382                 return true;
12383         }
12384 };
12385
12386 struct fp16FAbs : public fp16PerComponent
12387 {
12388         template<class fp16type>
12389         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
12390         {
12391                 const fp16type  x               (*in[0]);
12392                 const double    d               (x.asDouble());
12393                 const double    result  (deAbs(d));
12394
12395                 out[0] = fp16type(result).bits();
12396                 min[0] = getMin(result, getULPs(in));
12397                 max[0] = getMax(result, getULPs(in));
12398
12399                 return true;
12400         }
12401 };
12402
12403 struct fp16FSign : public fp16PerComponent
12404 {
12405         template<class fp16type>
12406         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
12407         {
12408                 const fp16type  x               (*in[0]);
12409                 const double    d               (x.asDouble());
12410                 const double    result  (deSign(d));
12411
12412                 if (x.isNaN())
12413                         return false;
12414
12415                 out[0] = fp16type(result).bits();
12416                 min[0] = getMin(result, getULPs(in));
12417                 max[0] = getMax(result, getULPs(in));
12418
12419                 return true;
12420         }
12421 };
12422
12423 struct fp16Floor : public fp16PerComponent
12424 {
12425         template<class fp16type>
12426         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
12427         {
12428                 const fp16type  x               (*in[0]);
12429                 const double    d               (x.asDouble());
12430                 const double    result  (deFloor(d));
12431
12432                 out[0] = fp16type(result).bits();
12433                 min[0] = getMin(result, getULPs(in));
12434                 max[0] = getMax(result, getULPs(in));
12435
12436                 return true;
12437         }
12438 };
12439
12440 struct fp16Ceil : public fp16PerComponent
12441 {
12442         template<class fp16type>
12443         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
12444         {
12445                 const fp16type  x               (*in[0]);
12446                 const double    d               (x.asDouble());
12447                 const double    result  (deCeil(d));
12448
12449                 out[0] = fp16type(result).bits();
12450                 min[0] = getMin(result, getULPs(in));
12451                 max[0] = getMax(result, getULPs(in));
12452
12453                 return true;
12454         }
12455 };
12456
12457 struct fp16Fract : public fp16PerComponent
12458 {
12459         template<class fp16type>
12460         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
12461         {
12462                 const fp16type  x               (*in[0]);
12463                 const double    d               (x.asDouble());
12464                 const double    result  (deFrac(d));
12465
12466                 out[0] = fp16type(result).bits();
12467                 min[0] = getMin(result, getULPs(in));
12468                 max[0] = getMax(result, getULPs(in));
12469
12470                 return true;
12471         }
12472 };
12473
12474 struct fp16Radians : public fp16PerComponent
12475 {
12476         virtual double getULPs (vector<const deFloat16*>& in)
12477         {
12478                 DE_UNREF(in);
12479
12480                 return 2.5;
12481         }
12482
12483         template<class fp16type>
12484         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
12485         {
12486                 const fp16type  x               (*in[0]);
12487                 const float             d               (x.asFloat());
12488                 const float             result  (deFloatRadians(d));
12489
12490                 out[0] = fp16type(result).bits();
12491                 min[0] = getMin(result, getULPs(in));
12492                 max[0] = getMax(result, getULPs(in));
12493
12494                 return true;
12495         }
12496 };
12497
12498 struct fp16Degrees : public fp16PerComponent
12499 {
12500         virtual double getULPs (vector<const deFloat16*>& in)
12501         {
12502                 DE_UNREF(in);
12503
12504                 return 2.5;
12505         }
12506
12507         template<class fp16type>
12508         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
12509         {
12510                 const fp16type  x               (*in[0]);
12511                 const float             d               (x.asFloat());
12512                 const float             result  (deFloatDegrees(d));
12513
12514                 out[0] = fp16type(result).bits();
12515                 min[0] = getMin(result, getULPs(in));
12516                 max[0] = getMax(result, getULPs(in));
12517
12518                 return true;
12519         }
12520 };
12521
12522 struct fp16Sin : public fp16PerComponent
12523 {
12524         template<class fp16type>
12525         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
12526         {
12527                 const fp16type  x                       (*in[0]);
12528                 const double    d                       (x.asDouble());
12529                 const double    result          (deSin(d));
12530                 const double    unspecUlp       (16.0);
12531                 const double    err                     (de::inRange(d, -DE_PI_DOUBLE, DE_PI_DOUBLE) ? deLdExp(1.0, -7) : floatFormat16.ulp(deAbs(result), unspecUlp));
12532
12533                 if (!de::inRange(d, -DE_PI_DOUBLE, DE_PI_DOUBLE))
12534                         return false;
12535
12536                 out[0] = fp16type(result).bits();
12537                 min[0] = result - err;
12538                 max[0] = result + err;
12539
12540                 return true;
12541         }
12542 };
12543
12544 struct fp16Cos : public fp16PerComponent
12545 {
12546         template<class fp16type>
12547         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
12548         {
12549                 const fp16type  x                       (*in[0]);
12550                 const double    d                       (x.asDouble());
12551                 const double    result          (deCos(d));
12552                 const double    unspecUlp       (16.0);
12553                 const double    err                     (de::inRange(d, -DE_PI_DOUBLE, DE_PI_DOUBLE) ? deLdExp(1.0, -7) : floatFormat16.ulp(deAbs(result), unspecUlp));
12554
12555                 if (!de::inRange(d, -DE_PI_DOUBLE, DE_PI_DOUBLE))
12556                         return false;
12557
12558                 out[0] = fp16type(result).bits();
12559                 min[0] = result - err;
12560                 max[0] = result + err;
12561
12562                 return true;
12563         }
12564 };
12565
12566 struct fp16Tan : public fp16PerComponent
12567 {
12568         template<class fp16type>
12569         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
12570         {
12571                 const fp16type  x               (*in[0]);
12572                 const double    d               (x.asDouble());
12573                 const double    result  (deTan(d));
12574
12575                 if (!de::inRange(d, -DE_PI_DOUBLE, DE_PI_DOUBLE))
12576                         return false;
12577
12578                 out[0] = fp16type(result).bits();
12579                 {
12580                         const double    err                     = deLdExp(1.0, -7);
12581                         const double    s1                      = deSin(d) + err;
12582                         const double    s2                      = deSin(d) - err;
12583                         const double    c1                      = deCos(d) + err;
12584                         const double    c2                      = deCos(d) - err;
12585                         const double    edgeVals[]      = {s1/c1, s1/c2, s2/c1, s2/c2};
12586                         double                  edgeLeft        = out[0];
12587                         double                  edgeRight       = out[0];
12588
12589                         if (deSign(c1 * c2) < 0.0)
12590                         {
12591                                 edgeLeft        = -std::numeric_limits<double>::infinity();
12592                                 edgeRight       = +std::numeric_limits<double>::infinity();
12593                         }
12594                         else
12595                         {
12596                                 edgeLeft        = *std::min_element(&edgeVals[0], &edgeVals[DE_LENGTH_OF_ARRAY(edgeVals)]);
12597                                 edgeRight       = *std::max_element(&edgeVals[0], &edgeVals[DE_LENGTH_OF_ARRAY(edgeVals)]);
12598                         }
12599
12600                         min[0] = edgeLeft;
12601                         max[0] = edgeRight;
12602                 }
12603
12604                 return true;
12605         }
12606 };
12607
12608 struct fp16Asin : public fp16PerComponent
12609 {
12610         template<class fp16type>
12611         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
12612         {
12613                 const fp16type  x               (*in[0]);
12614                 const double    d               (x.asDouble());
12615                 const double    result  (deAsin(d));
12616                 const double    error   (deAtan2(d, sqrt(1.0 - d * d)));
12617
12618                 if (!x.isNaN() && deAbs(d) > 1.0)
12619                         return false;
12620
12621                 out[0] = fp16type(result).bits();
12622                 min[0] = result - floatFormat16.ulp(deAbs(error), 2 * 5.0); // This is not a precision test. Value is not from spec
12623                 max[0] = result + floatFormat16.ulp(deAbs(error), 2 * 5.0); // This is not a precision test. Value is not from spec
12624
12625                 return true;
12626         }
12627 };
12628
12629 struct fp16Acos : public fp16PerComponent
12630 {
12631         template<class fp16type>
12632         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
12633         {
12634                 const fp16type  x               (*in[0]);
12635                 const double    d               (x.asDouble());
12636                 const double    result  (deAcos(d));
12637                 const double    error   (deAtan2(sqrt(1.0 - d * d), d));
12638
12639                 if (!x.isNaN() && deAbs(d) > 1.0)
12640                         return false;
12641
12642                 out[0] = fp16type(result).bits();
12643                 min[0] = result - floatFormat16.ulp(deAbs(error), 2 * 5.0); // This is not a precision test. Value is not from spec
12644                 max[0] = result + floatFormat16.ulp(deAbs(error), 2 * 5.0); // This is not a precision test. Value is not from spec
12645
12646                 return true;
12647         }
12648 };
12649
12650 struct fp16Atan : public fp16PerComponent
12651 {
12652         virtual double getULPs(vector<const deFloat16*>& in)
12653         {
12654                 DE_UNREF(in);
12655
12656                 return 2 * 5.0; // This is not a precision test. Value is not from spec
12657         }
12658
12659         template<class fp16type>
12660         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
12661         {
12662                 const fp16type  x               (*in[0]);
12663                 const double    d               (x.asDouble());
12664                 const double    result  (deAtanOver(d));
12665
12666                 out[0] = fp16type(result).bits();
12667                 min[0] = getMin(result, getULPs(in));
12668                 max[0] = getMax(result, getULPs(in));
12669
12670                 return true;
12671         }
12672 };
12673
12674 struct fp16Sinh : public fp16PerComponent
12675 {
12676         fp16Sinh() : fp16PerComponent()
12677         {
12678                 flavorNames.push_back("Double");
12679                 flavorNames.push_back("ExpFP16");
12680         }
12681
12682         template<class fp16type>
12683         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
12684         {
12685                 const fp16type  x               (*in[0]);
12686                 const double    d               (x.asDouble());
12687                 const double    ulps    (64 * (1.0 + 2 * deAbs(d))); // This is not a precision test. Value is not from spec
12688                 double                  result  (0.0);
12689                 double                  error   (0.0);
12690
12691                 if (getFlavor() == 0)
12692                 {
12693                         result  = deSinh(d);
12694                         error   = floatFormat16.ulp(deAbs(result), ulps);
12695                 }
12696                 else if (getFlavor() == 1)
12697                 {
12698                         const fp16type  epx     (deExp(d));
12699                         const fp16type  enx     (deExp(-d));
12700                         const fp16type  esx     (epx.asDouble() - enx.asDouble());
12701                         const fp16type  sx2     (esx.asDouble() / 2.0);
12702
12703                         result  = sx2.asDouble();
12704                         error   = deAbs(floatFormat16.ulp(epx.asDouble(), ulps)) + deAbs(floatFormat16.ulp(enx.asDouble(), ulps));
12705                 }
12706                 else
12707                 {
12708                         TCU_THROW(InternalError, "Unknown flavor");
12709                 }
12710
12711                 out[0] = fp16type(result).bits();
12712                 min[0] = result - error;
12713                 max[0] = result + error;
12714
12715                 return true;
12716         }
12717 };
12718
12719 struct fp16Cosh : public fp16PerComponent
12720 {
12721         fp16Cosh() : fp16PerComponent()
12722         {
12723                 flavorNames.push_back("Double");
12724                 flavorNames.push_back("ExpFP16");
12725         }
12726
12727         template<class fp16type>
12728         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
12729         {
12730                 const fp16type  x               (*in[0]);
12731                 const double    d               (x.asDouble());
12732                 const double    ulps    (64 * (1.0 + 2 * deAbs(d))); // This is not a precision test. Value is not from spec
12733                 double                  result  (0.0);
12734
12735                 if (getFlavor() == 0)
12736                 {
12737                         result = deCosh(d);
12738                 }
12739                 else if (getFlavor() == 1)
12740                 {
12741                         const fp16type  epx     (deExp(d));
12742                         const fp16type  enx     (deExp(-d));
12743                         const fp16type  esx     (epx.asDouble() + enx.asDouble());
12744                         const fp16type  sx2     (esx.asDouble() / 2.0);
12745
12746                         result = sx2.asDouble();
12747                 }
12748                 else
12749                 {
12750                         TCU_THROW(InternalError, "Unknown flavor");
12751                 }
12752
12753                 out[0] = fp16type(result).bits();
12754                 min[0] = result - floatFormat16.ulp(deAbs(result), ulps);
12755                 max[0] = result + floatFormat16.ulp(deAbs(result), ulps);
12756
12757                 return true;
12758         }
12759 };
12760
12761 struct fp16Tanh : public fp16PerComponent
12762 {
12763         fp16Tanh() : fp16PerComponent()
12764         {
12765                 flavorNames.push_back("Tanh");
12766                 flavorNames.push_back("SinhCosh");
12767                 flavorNames.push_back("SinhCoshFP16");
12768                 flavorNames.push_back("PolyFP16");
12769         }
12770
12771         virtual double getULPs (vector<const deFloat16*>& in)
12772         {
12773                 const tcu::Float16      x       (*in[0]);
12774                 const double            d       (x.asDouble());
12775
12776                 return 2 * (1.0 + 2 * deAbs(d)); // This is not a precision test. Value is not from spec
12777         }
12778
12779         template<class fp16type>
12780         inline double calcPoly (const fp16type& espx, const fp16type& esnx, const fp16type& ecpx, const fp16type& ecnx)
12781         {
12782                 const fp16type  esx     (espx.asDouble() - esnx.asDouble());
12783                 const fp16type  sx2     (esx.asDouble() / 2.0);
12784                 const fp16type  ecx     (ecpx.asDouble() + ecnx.asDouble());
12785                 const fp16type  cx2     (ecx.asDouble() / 2.0);
12786                 const fp16type  tg      (sx2.asDouble() / cx2.asDouble());
12787                 const double    rez     (tg.asDouble());
12788
12789                 return rez;
12790         }
12791
12792         template<class fp16type>
12793         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
12794         {
12795                 const fp16type  x               (*in[0]);
12796                 const double    d               (x.asDouble());
12797                 double                  result  (0.0);
12798
12799                 if (getFlavor() == 0)
12800                 {
12801                         result  = deTanh(d);
12802                         min[0]  = getMin(result, getULPs(in));
12803                         max[0]  = getMax(result, getULPs(in));
12804                 }
12805                 else if (getFlavor() == 1)
12806                 {
12807                         result  = deSinh(d) / deCosh(d);
12808                         min[0]  = getMin(result, getULPs(in));
12809                         max[0]  = getMax(result, getULPs(in));
12810                 }
12811                 else if (getFlavor() == 2)
12812                 {
12813                         const fp16type  s       (deSinh(d));
12814                         const fp16type  c       (deCosh(d));
12815
12816                         result  = s.asDouble() / c.asDouble();
12817                         min[0]  = getMin(result, getULPs(in));
12818                         max[0]  = getMax(result, getULPs(in));
12819                 }
12820                 else if (getFlavor() == 3)
12821                 {
12822                         const double    ulps    (getULPs(in));
12823                         const double    epxm    (deExp( d));
12824                         const double    enxm    (deExp(-d));
12825                         const double    epxmerr = floatFormat16.ulp(epxm, ulps);
12826                         const double    enxmerr = floatFormat16.ulp(enxm, ulps);
12827                         const fp16type  epx[]   = { fp16type(epxm - epxmerr), fp16type(epxm + epxmerr) };
12828                         const fp16type  enx[]   = { fp16type(enxm - enxmerr), fp16type(enxm + enxmerr) };
12829                         const fp16type  epxm16  (epxm);
12830                         const fp16type  enxm16  (enxm);
12831                         vector<double>  tgs;
12832
12833                         for (size_t spNdx = 0; spNdx < DE_LENGTH_OF_ARRAY(epx); ++spNdx)
12834                         for (size_t snNdx = 0; snNdx < DE_LENGTH_OF_ARRAY(enx); ++snNdx)
12835                         for (size_t cpNdx = 0; cpNdx < DE_LENGTH_OF_ARRAY(epx); ++cpNdx)
12836                         for (size_t cnNdx = 0; cnNdx < DE_LENGTH_OF_ARRAY(enx); ++cnNdx)
12837                         {
12838                                 const double tgh = calcPoly(epx[spNdx], enx[snNdx], epx[cpNdx], enx[cnNdx]);
12839
12840                                 tgs.push_back(tgh);
12841                         }
12842
12843                         result = calcPoly(epxm16, enxm16, epxm16, enxm16);
12844                         min[0] = *std::min_element(tgs.begin(), tgs.end());
12845                         max[0] = *std::max_element(tgs.begin(), tgs.end());
12846                 }
12847                 else
12848                 {
12849                         TCU_THROW(InternalError, "Unknown flavor");
12850                 }
12851
12852                 out[0] = fp16type(result).bits();
12853
12854                 return true;
12855         }
12856 };
12857
12858 struct fp16Asinh : public fp16PerComponent
12859 {
12860         fp16Asinh() : fp16PerComponent()
12861         {
12862                 flavorNames.push_back("Double");
12863                 flavorNames.push_back("PolyFP16Wiki");
12864                 flavorNames.push_back("PolyFP16Abs");
12865         }
12866
12867         virtual double getULPs (vector<const deFloat16*>& in)
12868         {
12869                 DE_UNREF(in);
12870
12871                 return 256.0; // This is not a precision test. Value is not from spec
12872         }
12873
12874         template<class fp16type>
12875         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
12876         {
12877                 const fp16type  x               (*in[0]);
12878                 const double    d               (x.asDouble());
12879                 double                  result  (0.0);
12880
12881                 if (getFlavor() == 0)
12882                 {
12883                         result = deAsinh(d);
12884                 }
12885                 else if (getFlavor() == 1)
12886                 {
12887                         const fp16type  x2              (d * d);
12888                         const fp16type  x2p1    (x2.asDouble() + 1.0);
12889                         const fp16type  sq              (deSqrt(x2p1.asDouble()));
12890                         const fp16type  sxsq    (d + sq.asDouble());
12891                         const fp16type  lsxsq   (deLog(sxsq.asDouble()));
12892
12893                         if (lsxsq.isInf())
12894                                 return false;
12895
12896                         result = lsxsq.asDouble();
12897                 }
12898                 else if (getFlavor() == 2)
12899                 {
12900                         const fp16type  x2              (d * d);
12901                         const fp16type  x2p1    (x2.asDouble() + 1.0);
12902                         const fp16type  sq              (deSqrt(x2p1.asDouble()));
12903                         const fp16type  sxsq    (deAbs(d) + sq.asDouble());
12904                         const fp16type  lsxsq   (deLog(sxsq.asDouble()));
12905
12906                         result = deSign(d) * lsxsq.asDouble();
12907                 }
12908                 else
12909                 {
12910                         TCU_THROW(InternalError, "Unknown flavor");
12911                 }
12912
12913                 out[0] = fp16type(result).bits();
12914                 min[0] = getMin(result, getULPs(in));
12915                 max[0] = getMax(result, getULPs(in));
12916
12917                 return true;
12918         }
12919 };
12920
12921 struct fp16Acosh : public fp16PerComponent
12922 {
12923         fp16Acosh() : fp16PerComponent()
12924         {
12925                 flavorNames.push_back("Double");
12926                 flavorNames.push_back("PolyFP16");
12927         }
12928
12929         virtual double getULPs (vector<const deFloat16*>& in)
12930         {
12931                 DE_UNREF(in);
12932
12933                 return 16.0; // This is not a precision test. Value is not from spec
12934         }
12935
12936         template<class fp16type>
12937         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
12938         {
12939                 const fp16type  x               (*in[0]);
12940                 const double    d               (x.asDouble());
12941                 double                  result  (0.0);
12942
12943                 if (!x.isNaN() && d < 1.0)
12944                         return false;
12945
12946                 if (getFlavor() == 0)
12947                 {
12948                         result = deAcosh(d);
12949                 }
12950                 else if (getFlavor() == 1)
12951                 {
12952                         const fp16type  x2              (d * d);
12953                         const fp16type  x2m1    (x2.asDouble() - 1.0);
12954                         const fp16type  sq              (deSqrt(x2m1.asDouble()));
12955                         const fp16type  sxsq    (d + sq.asDouble());
12956                         const fp16type  lsxsq   (deLog(sxsq.asDouble()));
12957
12958                         result = lsxsq.asDouble();
12959                 }
12960                 else
12961                 {
12962                         TCU_THROW(InternalError, "Unknown flavor");
12963                 }
12964
12965                 out[0] = fp16type(result).bits();
12966                 min[0] = getMin(result, getULPs(in));
12967                 max[0] = getMax(result, getULPs(in));
12968
12969                 return true;
12970         }
12971 };
12972
12973 struct fp16Atanh : public fp16PerComponent
12974 {
12975         fp16Atanh() : fp16PerComponent()
12976         {
12977                 flavorNames.push_back("Double");
12978                 flavorNames.push_back("PolyFP16");
12979         }
12980
12981         template<class fp16type>
12982         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
12983         {
12984                 const fp16type  x               (*in[0]);
12985                 const double    d               (x.asDouble());
12986                 double                  result  (0.0);
12987
12988                 if (deAbs(d) >= 1.0)
12989                         return false;
12990
12991                 if (getFlavor() == 0)
12992                 {
12993                         const double    ulps    (16.0); // This is not a precision test. Value is not from spec
12994
12995                         result = deAtanh(d);
12996                         min[0] = getMin(result, ulps);
12997                         max[0] = getMax(result, ulps);
12998                 }
12999                 else if (getFlavor() == 1)
13000                 {
13001                         const fp16type  x1a             (1.0 + d);
13002                         const fp16type  x1b             (1.0 - d);
13003                         const fp16type  x1d             (x1a.asDouble() / x1b.asDouble());
13004                         const fp16type  lx1d    (deLog(x1d.asDouble()));
13005                         const fp16type  lx1d2   (0.5 * lx1d.asDouble());
13006                         const double    error   (2 * (de::inRange(deAbs(x1d.asDouble()), 0.5, 2.0) ? deLdExp(2.0, -7) : floatFormat16.ulp(deAbs(x1d.asDouble()), 3.0)));
13007
13008                         result = lx1d2.asDouble();
13009                         min[0] = result - error;
13010                         max[0] = result + error;
13011                 }
13012                 else
13013                 {
13014                         TCU_THROW(InternalError, "Unknown flavor");
13015                 }
13016
13017                 out[0] = fp16type(result).bits();
13018
13019                 return true;
13020         }
13021 };
13022
13023 struct fp16Exp : public fp16PerComponent
13024 {
13025         template<class fp16type>
13026         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
13027         {
13028                 const fp16type  x               (*in[0]);
13029                 const double    d               (x.asDouble());
13030                 const double    ulps    (10.0 * (1.0 + 2.0 * deAbs(d)));
13031                 const double    result  (deExp(d));
13032
13033                 out[0] = fp16type(result).bits();
13034                 min[0] = getMin(result, ulps);
13035                 max[0] = getMax(result, ulps);
13036
13037                 return true;
13038         }
13039 };
13040
13041 struct fp16Log : public fp16PerComponent
13042 {
13043         template<class fp16type>
13044         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
13045         {
13046                 const fp16type  x               (*in[0]);
13047                 const double    d               (x.asDouble());
13048                 const double    result  (deLog(d));
13049                 const double    error   (de::inRange(deAbs(d), 0.5, 2.0) ? deLdExp(2.0, -7) : floatFormat16.ulp(deAbs(result), 3.0));
13050
13051                 if (d <= 0.0)
13052                         return false;
13053
13054                 out[0] = fp16type(result).bits();
13055                 min[0] = result - error;
13056                 max[0] = result + error;
13057
13058                 return true;
13059         }
13060 };
13061
13062 struct fp16Exp2 : public fp16PerComponent
13063 {
13064         template<class fp16type>
13065         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
13066         {
13067                 const fp16type  x               (*in[0]);
13068                 const double    d               (x.asDouble());
13069                 const double    result  (deExp2(d));
13070                 const double    ulps    (1.0 + 2.0 * deAbs(fp16type(in[0][0]).asDouble()));
13071
13072                 out[0] = fp16type(result).bits();
13073                 min[0] = getMin(result, ulps);
13074                 max[0] = getMax(result, ulps);
13075
13076                 return true;
13077         }
13078 };
13079
13080 struct fp16Log2 : public fp16PerComponent
13081 {
13082         template<class fp16type>
13083         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
13084         {
13085                 const fp16type  x               (*in[0]);
13086                 const double    d               (x.asDouble());
13087                 const double    result  (deLog2(d));
13088                 const double    error   (de::inRange(deAbs(d), 0.5, 2.0) ? deLdExp(2.0, -7) : floatFormat16.ulp(deAbs(result), 3.0));
13089
13090                 if (d <= 0.0)
13091                         return false;
13092
13093                 out[0] = fp16type(result).bits();
13094                 min[0] = result - error;
13095                 max[0] = result + error;
13096
13097                 return true;
13098         }
13099 };
13100
13101 struct fp16Sqrt : public fp16PerComponent
13102 {
13103         virtual double getULPs (vector<const deFloat16*>& in)
13104         {
13105                 DE_UNREF(in);
13106
13107                 return 6.0;
13108         }
13109
13110         template<class fp16type>
13111         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
13112         {
13113                 const fp16type  x               (*in[0]);
13114                 const double    d               (x.asDouble());
13115                 const double    result  (deSqrt(d));
13116
13117                 if (!x.isNaN() && d < 0.0)
13118                         return false;
13119
13120                 out[0] = fp16type(result).bits();
13121                 min[0] = getMin(result, getULPs(in));
13122                 max[0] = getMax(result, getULPs(in));
13123
13124                 return true;
13125         }
13126 };
13127
13128 struct fp16InverseSqrt : public fp16PerComponent
13129 {
13130         virtual double getULPs (vector<const deFloat16*>& in)
13131         {
13132                 DE_UNREF(in);
13133
13134                 return 2.0;
13135         }
13136
13137         template<class fp16type>
13138         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
13139         {
13140                 const fp16type  x               (*in[0]);
13141                 const double    d               (x.asDouble());
13142                 const double    result  (1.0/deSqrt(d));
13143
13144                 if (!x.isNaN() && d <= 0.0)
13145                         return false;
13146
13147                 out[0] = fp16type(result).bits();
13148                 min[0] = getMin(result, getULPs(in));
13149                 max[0] = getMax(result, getULPs(in));
13150
13151                 return true;
13152         }
13153 };
13154
13155 struct fp16ModfFrac : public fp16PerComponent
13156 {
13157         template<class fp16type>
13158         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
13159         {
13160                 const fp16type  x               (*in[0]);
13161                 const double    d               (x.asDouble());
13162                 double                  i               (0.0);
13163                 const double    result  (deModf(d, &i));
13164
13165                 if (x.isInf() || x.isNaN())
13166                         return false;
13167
13168                 out[0] = fp16type(result).bits();
13169                 min[0] = getMin(result, getULPs(in));
13170                 max[0] = getMax(result, getULPs(in));
13171
13172                 return true;
13173         }
13174 };
13175
13176 struct fp16ModfInt : public fp16PerComponent
13177 {
13178         template<class fp16type>
13179         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
13180         {
13181                 const fp16type  x               (*in[0]);
13182                 const double    d               (x.asDouble());
13183                 double                  i               (0.0);
13184                 const double    dummy   (deModf(d, &i));
13185                 const double    result  (i);
13186
13187                 DE_UNREF(dummy);
13188
13189                 if (x.isInf() || x.isNaN())
13190                         return false;
13191
13192                 out[0] = fp16type(result).bits();
13193                 min[0] = getMin(result, getULPs(in));
13194                 max[0] = getMax(result, getULPs(in));
13195
13196                 return true;
13197         }
13198 };
13199
13200 struct fp16FrexpS : public fp16PerComponent
13201 {
13202         template<class fp16type>
13203         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
13204         {
13205                 const fp16type  x               (*in[0]);
13206                 const double    d               (x.asDouble());
13207                 int                             e               (0);
13208                 const double    result  (deFrExp(d, &e));
13209
13210                 if (x.isNaN() || x.isInf())
13211                         return false;
13212
13213                 out[0] = fp16type(result).bits();
13214                 min[0] = getMin(result, getULPs(in));
13215                 max[0] = getMax(result, getULPs(in));
13216
13217                 return true;
13218         }
13219 };
13220
13221 struct fp16FrexpE : public fp16PerComponent
13222 {
13223         template<class fp16type>
13224         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
13225         {
13226                 const fp16type  x               (*in[0]);
13227                 const double    d               (x.asDouble());
13228                 int                             e               (0);
13229                 const double    dummy   (deFrExp(d, &e));
13230                 const double    result  (static_cast<double>(e));
13231
13232                 DE_UNREF(dummy);
13233
13234                 if (x.isNaN() || x.isInf())
13235                         return false;
13236
13237                 out[0] = fp16type(result).bits();
13238                 min[0] = getMin(result, getULPs(in));
13239                 max[0] = getMax(result, getULPs(in));
13240
13241                 return true;
13242         }
13243 };
13244
13245 struct fp16OpFAdd : public fp16PerComponent
13246 {
13247         template<class fp16type>
13248         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
13249         {
13250                 const fp16type  x               (*in[0]);
13251                 const fp16type  y               (*in[1]);
13252                 const double    xd              (x.asDouble());
13253                 const double    yd              (y.asDouble());
13254                 const double    result  (xd + yd);
13255
13256                 out[0] = fp16type(result).bits();
13257                 min[0] = getMin(result, getULPs(in));
13258                 max[0] = getMax(result, getULPs(in));
13259
13260                 return true;
13261         }
13262 };
13263
13264 struct fp16OpFSub : public fp16PerComponent
13265 {
13266         template<class fp16type>
13267         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
13268         {
13269                 const fp16type  x               (*in[0]);
13270                 const fp16type  y               (*in[1]);
13271                 const double    xd              (x.asDouble());
13272                 const double    yd              (y.asDouble());
13273                 const double    result  (xd - yd);
13274
13275                 out[0] = fp16type(result).bits();
13276                 min[0] = getMin(result, getULPs(in));
13277                 max[0] = getMax(result, getULPs(in));
13278
13279                 return true;
13280         }
13281 };
13282
13283 struct fp16OpFMul : public fp16PerComponent
13284 {
13285         template<class fp16type>
13286         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
13287         {
13288                 const fp16type  x               (*in[0]);
13289                 const fp16type  y               (*in[1]);
13290                 const double    xd              (x.asDouble());
13291                 const double    yd              (y.asDouble());
13292                 const double    result  (xd * yd);
13293
13294                 out[0] = fp16type(result).bits();
13295                 min[0] = getMin(result, getULPs(in));
13296                 max[0] = getMax(result, getULPs(in));
13297
13298                 return true;
13299         }
13300 };
13301
13302 struct fp16OpFDiv : public fp16PerComponent
13303 {
13304         fp16OpFDiv() : fp16PerComponent()
13305         {
13306                 flavorNames.push_back("DirectDiv");
13307                 flavorNames.push_back("InverseDiv");
13308         }
13309
13310         template<class fp16type>
13311         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
13312         {
13313                 const fp16type  x                       (*in[0]);
13314                 const fp16type  y                       (*in[1]);
13315                 const double    xd                      (x.asDouble());
13316                 const double    yd                      (y.asDouble());
13317                 const double    unspecUlp       (16.0);
13318                 const double    ulpCnt          (de::inRange(deAbs(yd), deLdExp(1, -14), deLdExp(1, 14)) ? 2.5 : unspecUlp);
13319                 double                  result          (0.0);
13320
13321                 if (y.isZero())
13322                         return false;
13323
13324                 if (getFlavor() == 0)
13325                 {
13326                         result = (xd / yd);
13327                 }
13328                 else if (getFlavor() == 1)
13329                 {
13330                         const double    invyd   (1.0 / yd);
13331                         const fp16type  invy    (invyd);
13332
13333                         result = (xd * invy.asDouble());
13334                 }
13335                 else
13336                 {
13337                         TCU_THROW(InternalError, "Unknown flavor");
13338                 }
13339
13340                 out[0] = fp16type(result).bits();
13341                 min[0] = getMin(result, ulpCnt);
13342                 max[0] = getMax(result, ulpCnt);
13343
13344                 return true;
13345         }
13346 };
13347
13348 struct fp16Atan2 : public fp16PerComponent
13349 {
13350         fp16Atan2() : fp16PerComponent()
13351         {
13352                 flavorNames.push_back("DoubleCalc");
13353                 flavorNames.push_back("DoubleCalc_PI");
13354         }
13355
13356         virtual double getULPs(vector<const deFloat16*>& in)
13357         {
13358                 DE_UNREF(in);
13359
13360                 return 2 * 5.0; // This is not a precision test. Value is not from spec
13361         }
13362
13363         template<class fp16type>
13364         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
13365         {
13366                 const fp16type  x               (*in[0]);
13367                 const fp16type  y               (*in[1]);
13368                 const double    xd              (x.asDouble());
13369                 const double    yd              (y.asDouble());
13370                 double                  result  (0.0);
13371
13372                 if (x.isZero() && y.isZero())
13373                         return false;
13374
13375                 if (getFlavor() == 0)
13376                 {
13377                         result  = deAtan2(xd, yd);
13378                 }
13379                 else if (getFlavor() == 1)
13380                 {
13381                         const double    ulps    (2.0 * 5.0); // This is not a precision test. Value is not from spec
13382                         const double    eps             (floatFormat16.ulp(DE_PI_DOUBLE, ulps));
13383
13384                         result  = deAtan2(xd, yd);
13385
13386                         if (de::inRange(deAbs(result), DE_PI_DOUBLE - eps, DE_PI_DOUBLE + eps))
13387                                 result  = -result;
13388                 }
13389                 else
13390                 {
13391                         TCU_THROW(InternalError, "Unknown flavor");
13392                 }
13393
13394                 out[0] = fp16type(result).bits();
13395                 min[0] = getMin(result, getULPs(in));
13396                 max[0] = getMax(result, getULPs(in));
13397
13398                 return true;
13399         }
13400 };
13401
13402 struct fp16Pow : public fp16PerComponent
13403 {
13404         fp16Pow() : fp16PerComponent()
13405         {
13406                 flavorNames.push_back("Pow");
13407                 flavorNames.push_back("PowLog2");
13408                 flavorNames.push_back("PowLog2FP16");
13409         }
13410
13411         template<class fp16type>
13412         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
13413         {
13414                 const fp16type  x               (*in[0]);
13415                 const fp16type  y               (*in[1]);
13416                 const double    xd              (x.asDouble());
13417                 const double    yd              (y.asDouble());
13418                 const double    logxeps (de::inRange(deAbs(xd), 0.5, 2.0) ? deLdExp(1.0, -7) : floatFormat16.ulp(deLog2(xd), 3.0));
13419                 const double    ulps1   (1.0 + 4.0 * deAbs(yd * (deLog2(xd) - logxeps)));
13420                 const double    ulps2   (1.0 + 4.0 * deAbs(yd * (deLog2(xd) + logxeps)));
13421                 const double    ulps    (deMax(deAbs(ulps1), deAbs(ulps2)));
13422                 double                  result  (0.0);
13423
13424                 if (xd < 0.0)
13425                         return false;
13426
13427                 if (x.isZero() && yd <= 0.0)
13428                         return false;
13429
13430                 if (getFlavor() == 0)
13431                 {
13432                         result = dePow(xd, yd);
13433                 }
13434                 else if (getFlavor() == 1)
13435                 {
13436                         const double    l2d     (deLog2(xd));
13437                         const double    e2d     (deExp2(yd * l2d));
13438
13439                         result = e2d;
13440                 }
13441                 else if (getFlavor() == 2)
13442                 {
13443                         const double    l2d     (deLog2(xd));
13444                         const fp16type  l2      (l2d);
13445                         const double    e2d     (deExp2(yd * l2.asDouble()));
13446                         const fp16type  e2      (e2d);
13447
13448                         result = e2.asDouble();
13449                 }
13450                 else
13451                 {
13452                         TCU_THROW(InternalError, "Unknown flavor");
13453                 }
13454
13455                 out[0] = fp16type(result).bits();
13456                 min[0] = getMin(result, ulps);
13457                 max[0] = getMax(result, ulps);
13458
13459                 return true;
13460         }
13461 };
13462
13463 struct fp16FMin : public fp16PerComponent
13464 {
13465         template<class fp16type>
13466         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
13467         {
13468                 const fp16type  x               (*in[0]);
13469                 const fp16type  y               (*in[1]);
13470                 const double    xd              (x.asDouble());
13471                 const double    yd              (y.asDouble());
13472                 const double    result  (deMin(xd, yd));
13473
13474                 if (x.isNaN() || y.isNaN())
13475                         return false;
13476
13477                 out[0] = fp16type(result).bits();
13478                 min[0] = getMin(result, getULPs(in));
13479                 max[0] = getMax(result, getULPs(in));
13480
13481                 return true;
13482         }
13483 };
13484
13485 struct fp16FMax : public fp16PerComponent
13486 {
13487         template<class fp16type>
13488         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
13489         {
13490                 const fp16type  x               (*in[0]);
13491                 const fp16type  y               (*in[1]);
13492                 const double    xd              (x.asDouble());
13493                 const double    yd              (y.asDouble());
13494                 const double    result  (deMax(xd, yd));
13495
13496                 if (x.isNaN() || y.isNaN())
13497                         return false;
13498
13499                 out[0] = fp16type(result).bits();
13500                 min[0] = getMin(result, getULPs(in));
13501                 max[0] = getMax(result, getULPs(in));
13502
13503                 return true;
13504         }
13505 };
13506
13507 struct fp16Step : public fp16PerComponent
13508 {
13509         template<class fp16type>
13510         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
13511         {
13512                 const fp16type  edge    (*in[0]);
13513                 const fp16type  x               (*in[1]);
13514                 const double    edged   (edge.asDouble());
13515                 const double    xd              (x.asDouble());
13516                 const double    result  (deStep(edged, xd));
13517
13518                 out[0] = fp16type(result).bits();
13519                 min[0] = getMin(result, getULPs(in));
13520                 max[0] = getMax(result, getULPs(in));
13521
13522                 return true;
13523         }
13524 };
13525
13526 struct fp16Ldexp : public fp16PerComponent
13527 {
13528         template<class fp16type>
13529         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
13530         {
13531                 const fp16type  x               (*in[0]);
13532                 const fp16type  y               (*in[1]);
13533                 const double    xd              (x.asDouble());
13534                 const int               yd              (static_cast<int>(deTrunc(y.asDouble())));
13535                 const double    result  (deLdExp(xd, yd));
13536
13537                 if (y.isNaN() || y.isInf() || y.isDenorm() || yd < -14 || yd > 15)
13538                         return false;
13539
13540                 // Spec: "If this product is too large to be represented in the floating-point type, the result is undefined."
13541                 if (fp16type(result).isInf())
13542                         return false;
13543
13544                 out[0] = fp16type(result).bits();
13545                 min[0] = getMin(result, getULPs(in));
13546                 max[0] = getMax(result, getULPs(in));
13547
13548                 return true;
13549         }
13550 };
13551
13552 struct fp16FClamp : public fp16PerComponent
13553 {
13554         template<class fp16type>
13555         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
13556         {
13557                 const fp16type  x               (*in[0]);
13558                 const fp16type  minVal  (*in[1]);
13559                 const fp16type  maxVal  (*in[2]);
13560                 const double    xd              (x.asDouble());
13561                 const double    minVald (minVal.asDouble());
13562                 const double    maxVald (maxVal.asDouble());
13563                 const double    result  (deClamp(xd, minVald, maxVald));
13564
13565                 if (minVal.isNaN() || maxVal.isNaN() || minVald > maxVald)
13566                         return false;
13567
13568                 out[0] = fp16type(result).bits();
13569                 min[0] = getMin(result, getULPs(in));
13570                 max[0] = getMax(result, getULPs(in));
13571
13572                 return true;
13573         }
13574 };
13575
13576 struct fp16FMix : public fp16PerComponent
13577 {
13578         fp16FMix() : fp16PerComponent()
13579         {
13580                 flavorNames.push_back("DoubleCalc");
13581                 flavorNames.push_back("EmulatingFP16");
13582                 flavorNames.push_back("EmulatingFP16YminusX");
13583         }
13584
13585         template<class fp16type>
13586         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
13587         {
13588                 const fp16type  x               (*in[0]);
13589                 const fp16type  y               (*in[1]);
13590                 const fp16type  a               (*in[2]);
13591                 const double    ulps    (8.0); // This is not a precision test. Value is not from spec
13592                 double                  result  (0.0);
13593
13594                 if (getFlavor() == 0)
13595                 {
13596                         const double    xd              (x.asDouble());
13597                         const double    yd              (y.asDouble());
13598                         const double    ad              (a.asDouble());
13599                         const double    xeps    (floatFormat16.ulp(deAbs(xd * (1.0 - ad)), ulps));
13600                         const double    yeps    (floatFormat16.ulp(deAbs(yd * ad), ulps));
13601                         const double    eps             (xeps + yeps);
13602
13603                         result = deMix(xd, yd, ad);
13604                         min[0] = result - eps;
13605                         max[0] = result + eps;
13606                 }
13607                 else if (getFlavor() == 1)
13608                 {
13609                         const double    xd              (x.asDouble());
13610                         const double    yd              (y.asDouble());
13611                         const double    ad              (a.asDouble());
13612                         const fp16type  am              (1.0 - ad);
13613                         const double    amd             (am.asDouble());
13614                         const fp16type  xam             (xd * amd);
13615                         const double    xamd    (xam.asDouble());
13616                         const fp16type  ya              (yd * ad);
13617                         const double    yad             (ya.asDouble());
13618                         const double    xeps    (floatFormat16.ulp(deAbs(xd * (1.0 - ad)), ulps));
13619                         const double    yeps    (floatFormat16.ulp(deAbs(yd * ad), ulps));
13620                         const double    eps             (xeps + yeps);
13621
13622                         result = xamd + yad;
13623                         min[0] = result - eps;
13624                         max[0] = result + eps;
13625                 }
13626                 else if (getFlavor() == 2)
13627                 {
13628                         const double    xd              (x.asDouble());
13629                         const double    yd              (y.asDouble());
13630                         const double    ad              (a.asDouble());
13631                         const fp16type  ymx             (yd - xd);
13632                         const double    ymxd    (ymx.asDouble());
13633                         const fp16type  ymxa    (ymxd * ad);
13634                         const double    ymxad   (ymxa.asDouble());
13635                         const double    xeps    (floatFormat16.ulp(deAbs(xd * (1.0 - ad)), ulps));
13636                         const double    yeps    (floatFormat16.ulp(deAbs(yd * ad), ulps));
13637                         const double    eps             (xeps + yeps);
13638
13639                         result = xd + ymxad;
13640                         min[0] = result - eps;
13641                         max[0] = result + eps;
13642                 }
13643                 else
13644                 {
13645                         TCU_THROW(InternalError, "Unknown flavor");
13646                 }
13647
13648                 out[0] = fp16type(result).bits();
13649
13650                 return true;
13651         }
13652 };
13653
13654 struct fp16SmoothStep : public fp16PerComponent
13655 {
13656         fp16SmoothStep() : fp16PerComponent()
13657         {
13658                 flavorNames.push_back("FloatCalc");
13659                 flavorNames.push_back("EmulatingFP16");
13660                 flavorNames.push_back("EmulatingFP16WClamp");
13661         }
13662
13663         virtual double getULPs(vector<const deFloat16*>& in)
13664         {
13665                 DE_UNREF(in);
13666
13667                 return 4.0; // This is not a precision test. Value is not from spec
13668         }
13669
13670         template<class fp16type>
13671         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
13672         {
13673                 const fp16type  edge0   (*in[0]);
13674                 const fp16type  edge1   (*in[1]);
13675                 const fp16type  x               (*in[2]);
13676                 double                  result  (0.0);
13677
13678                 if (edge0.isNaN() || edge1.isNaN() || x.isNaN() || edge0.asDouble() >= edge1.asDouble())
13679                         return false;
13680
13681                 if (edge0.isInf() || edge1.isInf() || x.isInf())
13682                         return false;
13683
13684                 if (getFlavor() == 0)
13685                 {
13686                         const float     edge0d  (edge0.asFloat());
13687                         const float     edge1d  (edge1.asFloat());
13688                         const float     xd              (x.asFloat());
13689                         const float     sstep   (deFloatSmoothStep(edge0d, edge1d, xd));
13690
13691                         result = sstep;
13692                 }
13693                 else if (getFlavor() == 1)
13694                 {
13695                         const double    edge0d  (edge0.asDouble());
13696                         const double    edge1d  (edge1.asDouble());
13697                         const double    xd              (x.asDouble());
13698
13699                         if (xd <= edge0d)
13700                                 result = 0.0;
13701                         else if (xd >= edge1d)
13702                                 result = 1.0;
13703                         else
13704                         {
13705                                 const fp16type  a       (xd - edge0d);
13706                                 const fp16type  b       (edge1d - edge0d);
13707                                 const fp16type  t       (a.asDouble() / b.asDouble());
13708                                 const fp16type  t2      (2.0 * t.asDouble());
13709                                 const fp16type  t3      (3.0 - t2.asDouble());
13710                                 const fp16type  t4      (t.asDouble() * t3.asDouble());
13711                                 const fp16type  t5      (t.asDouble() * t4.asDouble());
13712
13713                                 result = t5.asDouble();
13714                         }
13715                 }
13716                 else if (getFlavor() == 2)
13717                 {
13718                         const double    edge0d  (edge0.asDouble());
13719                         const double    edge1d  (edge1.asDouble());
13720                         const double    xd              (x.asDouble());
13721                         const fp16type  a       (xd - edge0d);
13722                         const fp16type  b       (edge1d - edge0d);
13723                         const fp16type  bi      (1.0 / b.asDouble());
13724                         const fp16type  t0      (a.asDouble() * bi.asDouble());
13725                         const double    tc      (deClamp(t0.asDouble(), 0.0, 1.0));
13726                         const fp16type  t       (tc);
13727                         const fp16type  t2      (2.0 * t.asDouble());
13728                         const fp16type  t3      (3.0 - t2.asDouble());
13729                         const fp16type  t4      (t.asDouble() * t3.asDouble());
13730                         const fp16type  t5      (t.asDouble() * t4.asDouble());
13731
13732                         result = t5.asDouble();
13733                 }
13734                 else
13735                 {
13736                         TCU_THROW(InternalError, "Unknown flavor");
13737                 }
13738
13739                 out[0] = fp16type(result).bits();
13740                 min[0] = getMin(result, getULPs(in));
13741                 max[0] = getMax(result, getULPs(in));
13742
13743                 return true;
13744         }
13745 };
13746
13747 struct fp16Fma : public fp16PerComponent
13748 {
13749         fp16Fma()
13750         {
13751                 flavorNames.push_back("DoubleCalc");
13752                 flavorNames.push_back("EmulatingFP16");
13753         }
13754
13755         virtual double getULPs(vector<const deFloat16*>& in)
13756         {
13757                 DE_UNREF(in);
13758
13759                 return 16.0;
13760         }
13761
13762         template<class fp16type>
13763         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
13764         {
13765                 DE_ASSERT(in.size() == 3);
13766                 DE_ASSERT(getArgCompCount(0) == getOutCompCount());
13767                 DE_ASSERT(getArgCompCount(1) == getOutCompCount());
13768                 DE_ASSERT(getArgCompCount(2) == getOutCompCount());
13769                 DE_ASSERT(getOutCompCount() > 0);
13770
13771                 const fp16type  a               (*in[0]);
13772                 const fp16type  b               (*in[1]);
13773                 const fp16type  c               (*in[2]);
13774                 double                  result  (0.0);
13775
13776                 if (getFlavor() == 0)
13777                 {
13778                         const double    ad      (a.asDouble());
13779                         const double    bd      (b.asDouble());
13780                         const double    cd      (c.asDouble());
13781
13782                         result  = deMadd(ad, bd, cd);
13783                 }
13784                 else if (getFlavor() == 1)
13785                 {
13786                         const double    ad      (a.asDouble());
13787                         const double    bd      (b.asDouble());
13788                         const double    cd      (c.asDouble());
13789                         const fp16type  ab      (ad * bd);
13790                         const fp16type  r       (ab.asDouble() + cd);
13791
13792                         result  = r.asDouble();
13793                 }
13794                 else
13795                 {
13796                         TCU_THROW(InternalError, "Unknown flavor");
13797                 }
13798
13799                 out[0] = fp16type(result).bits();
13800                 min[0] = getMin(result, getULPs(in));
13801                 max[0] = getMax(result, getULPs(in));
13802
13803                 return true;
13804         }
13805 };
13806
13807
13808 struct fp16AllComponents : public fp16PerComponent
13809 {
13810         bool            callOncePerComponent    ()      { return false; }
13811 };
13812
13813 struct fp16Length : public fp16AllComponents
13814 {
13815         fp16Length() : fp16AllComponents()
13816         {
13817                 flavorNames.push_back("EmulatingFP16");
13818                 flavorNames.push_back("DoubleCalc");
13819         }
13820
13821         virtual double getULPs(vector<const deFloat16*>& in)
13822         {
13823                 DE_UNREF(in);
13824
13825                 return 4.0;
13826         }
13827
13828         template<class fp16type>
13829         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
13830         {
13831                 DE_ASSERT(getOutCompCount() == 1);
13832                 DE_ASSERT(in.size() == 1);
13833
13834                 double  result  (0.0);
13835
13836                 if (getFlavor() == 0)
13837                 {
13838                         fp16type        r       (0.0);
13839
13840                         for (size_t componentNdx = 0; componentNdx < getArgCompCount(0); ++componentNdx)
13841                         {
13842                                 const fp16type  x       (in[0][componentNdx]);
13843                                 const fp16type  q       (x.asDouble() * x.asDouble());
13844
13845                                 r = fp16type(r.asDouble() + q.asDouble());
13846                         }
13847
13848                         result = deSqrt(r.asDouble());
13849
13850                         out[0] = fp16type(result).bits();
13851                 }
13852                 else if (getFlavor() == 1)
13853                 {
13854                         double  r       (0.0);
13855
13856                         for (size_t componentNdx = 0; componentNdx < getArgCompCount(0); ++componentNdx)
13857                         {
13858                                 const fp16type  x       (in[0][componentNdx]);
13859                                 const double    q       (x.asDouble() * x.asDouble());
13860
13861                                 r += q;
13862                         }
13863
13864                         result = deSqrt(r);
13865
13866                         out[0] = fp16type(result).bits();
13867                 }
13868                 else
13869                 {
13870                         TCU_THROW(InternalError, "Unknown flavor");
13871                 }
13872
13873                 min[0] = getMin(result, getULPs(in));
13874                 max[0] = getMax(result, getULPs(in));
13875
13876                 return true;
13877         }
13878 };
13879
13880 struct fp16Distance : public fp16AllComponents
13881 {
13882         fp16Distance() : fp16AllComponents()
13883         {
13884                 flavorNames.push_back("EmulatingFP16");
13885                 flavorNames.push_back("DoubleCalc");
13886         }
13887
13888         virtual double getULPs(vector<const deFloat16*>& in)
13889         {
13890                 DE_UNREF(in);
13891
13892                 return 4.0;
13893         }
13894
13895         template<class fp16type>
13896         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
13897         {
13898                 DE_ASSERT(getOutCompCount() == 1);
13899                 DE_ASSERT(in.size() == 2);
13900                 DE_ASSERT(getArgCompCount(0) == getArgCompCount(1));
13901
13902                 double  result  (0.0);
13903
13904                 if (getFlavor() == 0)
13905                 {
13906                         fp16type        r       (0.0);
13907
13908                         for (size_t componentNdx = 0; componentNdx < getArgCompCount(0); ++componentNdx)
13909                         {
13910                                 const fp16type  x       (in[0][componentNdx]);
13911                                 const fp16type  y       (in[1][componentNdx]);
13912                                 const fp16type  d       (x.asDouble() - y.asDouble());
13913                                 const fp16type  q       (d.asDouble() * d.asDouble());
13914
13915                                 r = fp16type(r.asDouble() + q.asDouble());
13916                         }
13917
13918                         result = deSqrt(r.asDouble());
13919                 }
13920                 else if (getFlavor() == 1)
13921                 {
13922                         double  r       (0.0);
13923
13924                         for (size_t componentNdx = 0; componentNdx < getArgCompCount(0); ++componentNdx)
13925                         {
13926                                 const fp16type  x       (in[0][componentNdx]);
13927                                 const fp16type  y       (in[1][componentNdx]);
13928                                 const double    d       (x.asDouble() - y.asDouble());
13929                                 const double    q       (d * d);
13930
13931                                 r += q;
13932                         }
13933
13934                         result = deSqrt(r);
13935                 }
13936                 else
13937                 {
13938                         TCU_THROW(InternalError, "Unknown flavor");
13939                 }
13940
13941                 out[0] = fp16type(result).bits();
13942                 min[0] = getMin(result, getULPs(in));
13943                 max[0] = getMax(result, getULPs(in));
13944
13945                 return true;
13946         }
13947 };
13948
13949 struct fp16Cross : public fp16AllComponents
13950 {
13951         fp16Cross() : fp16AllComponents()
13952         {
13953                 flavorNames.push_back("EmulatingFP16");
13954                 flavorNames.push_back("DoubleCalc");
13955         }
13956
13957         virtual double getULPs(vector<const deFloat16*>& in)
13958         {
13959                 DE_UNREF(in);
13960
13961                 return 4.0;
13962         }
13963
13964         template<class fp16type>
13965         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
13966         {
13967                 DE_ASSERT(getOutCompCount() == 3);
13968                 DE_ASSERT(in.size() == 2);
13969                 DE_ASSERT(getArgCompCount(0) == 3);
13970                 DE_ASSERT(getArgCompCount(1) == 3);
13971
13972                 if (getFlavor() == 0)
13973                 {
13974                         const fp16type  x0              (in[0][0]);
13975                         const fp16type  x1              (in[0][1]);
13976                         const fp16type  x2              (in[0][2]);
13977                         const fp16type  y0              (in[1][0]);
13978                         const fp16type  y1              (in[1][1]);
13979                         const fp16type  y2              (in[1][2]);
13980                         const fp16type  x1y2    (x1.asDouble() * y2.asDouble());
13981                         const fp16type  y1x2    (y1.asDouble() * x2.asDouble());
13982                         const fp16type  x2y0    (x2.asDouble() * y0.asDouble());
13983                         const fp16type  y2x0    (y2.asDouble() * x0.asDouble());
13984                         const fp16type  x0y1    (x0.asDouble() * y1.asDouble());
13985                         const fp16type  y0x1    (y0.asDouble() * x1.asDouble());
13986
13987                         out[0] = fp16type(x1y2.asDouble() - y1x2.asDouble()).bits();
13988                         out[1] = fp16type(x2y0.asDouble() - y2x0.asDouble()).bits();
13989                         out[2] = fp16type(x0y1.asDouble() - y0x1.asDouble()).bits();
13990                 }
13991                 else if (getFlavor() == 1)
13992                 {
13993                         const fp16type  x0              (in[0][0]);
13994                         const fp16type  x1              (in[0][1]);
13995                         const fp16type  x2              (in[0][2]);
13996                         const fp16type  y0              (in[1][0]);
13997                         const fp16type  y1              (in[1][1]);
13998                         const fp16type  y2              (in[1][2]);
13999                         const double    x1y2    (x1.asDouble() * y2.asDouble());
14000                         const double    y1x2    (y1.asDouble() * x2.asDouble());
14001                         const double    x2y0    (x2.asDouble() * y0.asDouble());
14002                         const double    y2x0    (y2.asDouble() * x0.asDouble());
14003                         const double    x0y1    (x0.asDouble() * y1.asDouble());
14004                         const double    y0x1    (y0.asDouble() * x1.asDouble());
14005
14006                         out[0] = fp16type(x1y2 - y1x2).bits();
14007                         out[1] = fp16type(x2y0 - y2x0).bits();
14008                         out[2] = fp16type(x0y1 - y0x1).bits();
14009                 }
14010                 else
14011                 {
14012                         TCU_THROW(InternalError, "Unknown flavor");
14013                 }
14014
14015                 for (size_t ndx = 0; ndx < getOutCompCount(); ++ndx)
14016                         min[ndx] = getMin(fp16type(out[ndx]).asDouble(), getULPs(in));
14017                 for (size_t ndx = 0; ndx < getOutCompCount(); ++ndx)
14018                         max[ndx] = getMax(fp16type(out[ndx]).asDouble(), getULPs(in));
14019
14020                 return true;
14021         }
14022 };
14023
14024 struct fp16Normalize : public fp16AllComponents
14025 {
14026         fp16Normalize() : fp16AllComponents()
14027         {
14028                 flavorNames.push_back("EmulatingFP16");
14029                 flavorNames.push_back("DoubleCalc");
14030
14031                 // flavorNames will be extended later
14032         }
14033
14034         virtual void    setArgCompCount                 (size_t argNo, size_t compCount)
14035         {
14036                 DE_ASSERT(argCompCount[argNo] == 0); // Once only
14037
14038                 if (argNo == 0 && argCompCount[argNo] == 0)
14039                 {
14040                         const size_t            maxPermutationsCount    = 24u; // Equal to 4!
14041                         std::vector<int>        indices;
14042
14043                         for (size_t componentNdx = 0; componentNdx < compCount; ++componentNdx)
14044                                 indices.push_back(static_cast<int>(componentNdx));
14045
14046                         m_permutations.reserve(maxPermutationsCount);
14047
14048                         permutationsFlavorStart = flavorNames.size();
14049
14050                         do
14051                         {
14052                                 tcu::UVec4      permutation;
14053                                 std::string     name            = "Permutted_";
14054
14055                                 for (size_t componentNdx = 0; componentNdx < compCount; ++componentNdx)
14056                                 {
14057                                         permutation[static_cast<int>(componentNdx)] = indices[componentNdx];
14058                                         name += de::toString(indices[componentNdx]);
14059                                 }
14060
14061                                 m_permutations.push_back(permutation);
14062                                 flavorNames.push_back(name);
14063
14064                         } while(std::next_permutation(indices.begin(), indices.end()));
14065
14066                         permutationsFlavorEnd = flavorNames.size();
14067                 }
14068
14069                 fp16AllComponents::setArgCompCount(argNo, compCount);
14070         }
14071         virtual double getULPs(vector<const deFloat16*>& in)
14072         {
14073                 DE_UNREF(in);
14074
14075                 return 8.0;
14076         }
14077
14078         template<class fp16type>
14079         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
14080         {
14081                 DE_ASSERT(in.size() == 1);
14082                 DE_ASSERT(getArgCompCount(0) == getOutCompCount());
14083
14084                 if (getFlavor() == 0)
14085                 {
14086                         fp16type        r(0.0);
14087
14088                         for (size_t componentNdx = 0; componentNdx < getArgCompCount(0); ++componentNdx)
14089                         {
14090                                 const fp16type  x       (in[0][componentNdx]);
14091                                 const fp16type  q       (x.asDouble() * x.asDouble());
14092
14093                                 r = fp16type(r.asDouble() + q.asDouble());
14094                         }
14095
14096                         r = fp16type(deSqrt(r.asDouble()));
14097
14098                         if (r.isZero())
14099                                 return false;
14100
14101                         for (size_t componentNdx = 0; componentNdx < getArgCompCount(0); ++componentNdx)
14102                         {
14103                                 const fp16type  x       (in[0][componentNdx]);
14104
14105                                 out[componentNdx] = fp16type(x.asDouble() / r.asDouble()).bits();
14106                         }
14107                 }
14108                 else if (getFlavor() == 1)
14109                 {
14110                         double  r(0.0);
14111
14112                         for (size_t componentNdx = 0; componentNdx < getArgCompCount(0); ++componentNdx)
14113                         {
14114                                 const fp16type  x       (in[0][componentNdx]);
14115                                 const double    q       (x.asDouble() * x.asDouble());
14116
14117                                 r += q;
14118                         }
14119
14120                         r = deSqrt(r);
14121
14122                         if (r == 0)
14123                                 return false;
14124
14125                         for (size_t componentNdx = 0; componentNdx < getArgCompCount(0); ++componentNdx)
14126                         {
14127                                 const fp16type  x       (in[0][componentNdx]);
14128
14129                                 out[componentNdx] = fp16type(x.asDouble() / r).bits();
14130                         }
14131                 }
14132                 else if (de::inBounds<size_t>(getFlavor(), permutationsFlavorStart, permutationsFlavorEnd))
14133                 {
14134                         const int                       compCount               (static_cast<int>(getArgCompCount(0)));
14135                         const size_t            permutationNdx  (getFlavor() - permutationsFlavorStart);
14136                         const tcu::UVec4&       permutation             (m_permutations[permutationNdx]);
14137                         fp16type                        r                               (0.0);
14138
14139                         for (int permComponentNdx = 0; permComponentNdx < compCount; ++permComponentNdx)
14140                         {
14141                                 const size_t    componentNdx    (permutation[permComponentNdx]);
14142                                 const fp16type  x                               (in[0][componentNdx]);
14143                                 const fp16type  q                               (x.asDouble() * x.asDouble());
14144
14145                                 r = fp16type(r.asDouble() + q.asDouble());
14146                         }
14147
14148                         r = fp16type(deSqrt(r.asDouble()));
14149
14150                         if (r.isZero())
14151                                 return false;
14152
14153                         for (int permComponentNdx = 0; permComponentNdx < compCount; ++permComponentNdx)
14154                         {
14155                                 const size_t    componentNdx    (permutation[permComponentNdx]);
14156                                 const fp16type  x                               (in[0][componentNdx]);
14157
14158                                 out[componentNdx] = fp16type(x.asDouble() / r.asDouble()).bits();
14159                         }
14160                 }
14161                 else
14162                 {
14163                         TCU_THROW(InternalError, "Unknown flavor");
14164                 }
14165
14166                 for (size_t ndx = 0; ndx < getOutCompCount(); ++ndx)
14167                         min[ndx] = getMin(fp16type(out[ndx]).asDouble(), getULPs(in));
14168                 for (size_t ndx = 0; ndx < getOutCompCount(); ++ndx)
14169                         max[ndx] = getMax(fp16type(out[ndx]).asDouble(), getULPs(in));
14170
14171                 return true;
14172         }
14173
14174 private:
14175         std::vector<tcu::UVec4> m_permutations;
14176         size_t                                  permutationsFlavorStart;
14177         size_t                                  permutationsFlavorEnd;
14178 };
14179
14180 struct fp16FaceForward : public fp16AllComponents
14181 {
14182         virtual double getULPs(vector<const deFloat16*>& in)
14183         {
14184                 DE_UNREF(in);
14185
14186                 return 4.0;
14187         }
14188
14189         template<class fp16type>
14190         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
14191         {
14192                 DE_ASSERT(in.size() == 3);
14193                 DE_ASSERT(getArgCompCount(0) == getOutCompCount());
14194                 DE_ASSERT(getArgCompCount(1) == getOutCompCount());
14195                 DE_ASSERT(getArgCompCount(2) == getOutCompCount());
14196
14197                 fp16type        dp(0.0);
14198
14199                 for (size_t componentNdx = 0; componentNdx < getArgCompCount(1); ++componentNdx)
14200                 {
14201                         const fp16type  x       (in[1][componentNdx]);
14202                         const fp16type  y       (in[2][componentNdx]);
14203                         const double    xd      (x.asDouble());
14204                         const double    yd      (y.asDouble());
14205                         const fp16type  q       (xd * yd);
14206
14207                         dp = fp16type(dp.asDouble() + q.asDouble());
14208                 }
14209
14210                 if (dp.isNaN() || dp.isZero())
14211                         return false;
14212
14213                 for (size_t componentNdx = 0; componentNdx < getOutCompCount(); ++componentNdx)
14214                 {
14215                         const fp16type  n       (in[0][componentNdx]);
14216
14217                         out[componentNdx] = (dp.signBit() == 1) ? n.bits() : fp16type(-n.asDouble()).bits();
14218                 }
14219
14220                 for (size_t ndx = 0; ndx < getOutCompCount(); ++ndx)
14221                         min[ndx] = getMin(fp16type(out[ndx]).asDouble(), getULPs(in));
14222                 for (size_t ndx = 0; ndx < getOutCompCount(); ++ndx)
14223                         max[ndx] = getMax(fp16type(out[ndx]).asDouble(), getULPs(in));
14224
14225                 return true;
14226         }
14227 };
14228
14229 struct fp16Reflect : public fp16AllComponents
14230 {
14231         fp16Reflect() : fp16AllComponents()
14232         {
14233                 flavorNames.push_back("EmulatingFP16");
14234                 flavorNames.push_back("EmulatingFP16+KeepZeroSign");
14235                 flavorNames.push_back("FloatCalc");
14236                 flavorNames.push_back("FloatCalc+KeepZeroSign");
14237                 flavorNames.push_back("EmulatingFP16+2Nfirst");
14238                 flavorNames.push_back("EmulatingFP16+2Ifirst");
14239         }
14240
14241         virtual double getULPs(vector<const deFloat16*>& in)
14242         {
14243                 DE_UNREF(in);
14244
14245                 return 256.0; // This is not a precision test. Value is not from spec
14246         }
14247
14248         template<class fp16type>
14249         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
14250         {
14251                 DE_ASSERT(in.size() == 2);
14252                 DE_ASSERT(getArgCompCount(0) == getOutCompCount());
14253                 DE_ASSERT(getArgCompCount(1) == getOutCompCount());
14254
14255                 if (getFlavor() < 4)
14256                 {
14257                         const bool      keepZeroSign    ((flavor & 1) != 0 ? true : false);
14258                         const bool      floatCalc               ((flavor & 2) != 0 ? true : false);
14259
14260                         if (floatCalc)
14261                         {
14262                                 float   dp(0.0f);
14263
14264                                 for (size_t componentNdx = 0; componentNdx < getArgCompCount(1); ++componentNdx)
14265                                 {
14266                                         const fp16type  i       (in[0][componentNdx]);
14267                                         const fp16type  n       (in[1][componentNdx]);
14268                                         const float             id      (i.asFloat());
14269                                         const float             nd      (n.asFloat());
14270                                         const float             qd      (id * nd);
14271
14272                                         if (keepZeroSign)
14273                                                 dp = (componentNdx == 0) ? qd : dp + qd;
14274                                         else
14275                                                 dp = dp + qd;
14276                                 }
14277
14278                                 for (size_t componentNdx = 0; componentNdx < getOutCompCount(); ++componentNdx)
14279                                 {
14280                                         const fp16type  i               (in[0][componentNdx]);
14281                                         const fp16type  n               (in[1][componentNdx]);
14282                                         const float             dpnd    (dp * n.asFloat());
14283                                         const float             dpn2d   (2.0f * dpnd);
14284                                         const float             idpn2d  (i.asFloat() - dpn2d);
14285                                         const fp16type  result  (idpn2d);
14286
14287                                         out[componentNdx] = result.bits();
14288                                 }
14289                         }
14290                         else
14291                         {
14292                                 fp16type        dp(0.0);
14293
14294                                 for (size_t componentNdx = 0; componentNdx < getArgCompCount(1); ++componentNdx)
14295                                 {
14296                                         const fp16type  i       (in[0][componentNdx]);
14297                                         const fp16type  n       (in[1][componentNdx]);
14298                                         const double    id      (i.asDouble());
14299                                         const double    nd      (n.asDouble());
14300                                         const fp16type  q       (id * nd);
14301
14302                                         if (keepZeroSign)
14303                                                 dp = (componentNdx == 0) ? q : fp16type(dp.asDouble() + q.asDouble());
14304                                         else
14305                                                 dp = fp16type(dp.asDouble() + q.asDouble());
14306                                 }
14307
14308                                 if (dp.isNaN())
14309                                         return false;
14310
14311                                 for (size_t componentNdx = 0; componentNdx < getOutCompCount(); ++componentNdx)
14312                                 {
14313                                         const fp16type  i               (in[0][componentNdx]);
14314                                         const fp16type  n               (in[1][componentNdx]);
14315                                         const fp16type  dpn             (dp.asDouble() * n.asDouble());
14316                                         const fp16type  dpn2    (2 * dpn.asDouble());
14317                                         const fp16type  idpn2   (i.asDouble() - dpn2.asDouble());
14318
14319                                         out[componentNdx] = idpn2.bits();
14320                                 }
14321                         }
14322                 }
14323                 else if (getFlavor() == 4)
14324                 {
14325                         fp16type        dp(0.0);
14326
14327                         for (size_t componentNdx = 0; componentNdx < getArgCompCount(1); ++componentNdx)
14328                         {
14329                                 const fp16type  i       (in[0][componentNdx]);
14330                                 const fp16type  n       (in[1][componentNdx]);
14331                                 const double    id      (i.asDouble());
14332                                 const double    nd      (n.asDouble());
14333                                 const fp16type  q       (id * nd);
14334
14335                                 dp = fp16type(dp.asDouble() + q.asDouble());
14336                         }
14337
14338                         if (dp.isNaN())
14339                                 return false;
14340
14341                         for (size_t componentNdx = 0; componentNdx < getOutCompCount(); ++componentNdx)
14342                         {
14343                                 const fp16type  i               (in[0][componentNdx]);
14344                                 const fp16type  n               (in[1][componentNdx]);
14345                                 const fp16type  n2              (2 * n.asDouble());
14346                                 const fp16type  dpn2    (dp.asDouble() * n2.asDouble());
14347                                 const fp16type  idpn2   (i.asDouble() - dpn2.asDouble());
14348
14349                                 out[componentNdx] = idpn2.bits();
14350                         }
14351                 }
14352                 else if (getFlavor() == 5)
14353                 {
14354                         fp16type        dp2(0.0);
14355
14356                         for (size_t componentNdx = 0; componentNdx < getArgCompCount(1); ++componentNdx)
14357                         {
14358                                 const fp16type  i       (in[0][componentNdx]);
14359                                 const fp16type  n       (in[1][componentNdx]);
14360                                 const fp16type  i2      (2.0 * i.asDouble());
14361                                 const double    i2d     (i2.asDouble());
14362                                 const double    nd      (n.asDouble());
14363                                 const fp16type  q       (i2d * nd);
14364
14365                                 dp2 = fp16type(dp2.asDouble() + q.asDouble());
14366                         }
14367
14368                         if (dp2.isNaN())
14369                                 return false;
14370
14371                         for (size_t componentNdx = 0; componentNdx < getOutCompCount(); ++componentNdx)
14372                         {
14373                                 const fp16type  i               (in[0][componentNdx]);
14374                                 const fp16type  n               (in[1][componentNdx]);
14375                                 const fp16type  dpn2    (dp2.asDouble() * n.asDouble());
14376                                 const fp16type  idpn2   (i.asDouble() - dpn2.asDouble());
14377
14378                                 out[componentNdx] = idpn2.bits();
14379                         }
14380                 }
14381                 else
14382                 {
14383                         TCU_THROW(InternalError, "Unknown flavor");
14384                 }
14385
14386                 for (size_t ndx = 0; ndx < getOutCompCount(); ++ndx)
14387                         min[ndx] = getMin(fp16type(out[ndx]).asDouble(), getULPs(in));
14388                 for (size_t ndx = 0; ndx < getOutCompCount(); ++ndx)
14389                         max[ndx] = getMax(fp16type(out[ndx]).asDouble(), getULPs(in));
14390
14391                 return true;
14392         }
14393 };
14394
14395 struct fp16Refract : public fp16AllComponents
14396 {
14397         fp16Refract() : fp16AllComponents()
14398         {
14399                 flavorNames.push_back("EmulatingFP16");
14400                 flavorNames.push_back("EmulatingFP16+KeepZeroSign");
14401                 flavorNames.push_back("FloatCalc");
14402                 flavorNames.push_back("FloatCalc+KeepZeroSign");
14403         }
14404
14405         virtual double getULPs(vector<const deFloat16*>& in)
14406         {
14407                 DE_UNREF(in);
14408
14409                 return 8192.0; // This is not a precision test. Value is not from spec
14410         }
14411
14412         template<class fp16type>
14413         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
14414         {
14415                 DE_ASSERT(in.size() == 3);
14416                 DE_ASSERT(getArgCompCount(0) == getOutCompCount());
14417                 DE_ASSERT(getArgCompCount(1) == getOutCompCount());
14418                 DE_ASSERT(getArgCompCount(2) == 1);
14419
14420                 const bool              keepZeroSign    ((flavor & 1) != 0 ? true : false);
14421                 const bool              doubleCalc              ((flavor & 2) != 0 ? true : false);
14422                 const fp16type  eta                             (*in[2]);
14423
14424                 if (doubleCalc)
14425                 {
14426                         double  dp      (0.0);
14427
14428                         for (size_t componentNdx = 0; componentNdx < getArgCompCount(1); ++componentNdx)
14429                         {
14430                                 const fp16type  i       (in[0][componentNdx]);
14431                                 const fp16type  n       (in[1][componentNdx]);
14432                                 const double    id      (i.asDouble());
14433                                 const double    nd      (n.asDouble());
14434                                 const double    qd      (id * nd);
14435
14436                                 if (keepZeroSign)
14437                                         dp = (componentNdx == 0) ? qd : dp + qd;
14438                                 else
14439                                         dp = dp + qd;
14440                         }
14441
14442                         const double    eta2    (eta.asDouble() * eta.asDouble());
14443                         const double    dp2             (dp * dp);
14444                         const double    dp1             (1.0 - dp2);
14445                         const double    dpe             (eta2 * dp1);
14446                         const double    k               (1.0 - dpe);
14447
14448                         if (k < 0.0)
14449                         {
14450                                 const fp16type  zero    (0.0);
14451
14452                                 for (size_t componentNdx = 0; componentNdx < getOutCompCount(); ++componentNdx)
14453                                         out[componentNdx] = zero.bits();
14454                         }
14455                         else
14456                         {
14457                                 const double    sk      (deSqrt(k));
14458
14459                                 for (size_t componentNdx = 0; componentNdx < getOutCompCount(); ++componentNdx)
14460                                 {
14461                                         const fp16type  i               (in[0][componentNdx]);
14462                                         const fp16type  n               (in[1][componentNdx]);
14463                                         const double    etai    (i.asDouble() * eta.asDouble());
14464                                         const double    etadp   (eta.asDouble() * dp);
14465                                         const double    etadpk  (etadp + sk);
14466                                         const double    etadpkn (etadpk * n.asDouble());
14467                                         const double    full    (etai - etadpkn);
14468                                         const fp16type  result  (full);
14469
14470                                         if (result.isInf())
14471                                                 return false;
14472
14473                                         out[componentNdx] = result.bits();
14474                                 }
14475                         }
14476                 }
14477                 else
14478                 {
14479                         fp16type        dp      (0.0);
14480
14481                         for (size_t componentNdx = 0; componentNdx < getArgCompCount(1); ++componentNdx)
14482                         {
14483                                 const fp16type  i       (in[0][componentNdx]);
14484                                 const fp16type  n       (in[1][componentNdx]);
14485                                 const double    id      (i.asDouble());
14486                                 const double    nd      (n.asDouble());
14487                                 const fp16type  q       (id * nd);
14488
14489                                 if (keepZeroSign)
14490                                         dp = (componentNdx == 0) ? q : fp16type(dp.asDouble() + q.asDouble());
14491                                 else
14492                                         dp = fp16type(dp.asDouble() + q.asDouble());
14493                         }
14494
14495                         if (dp.isNaN())
14496                                 return false;
14497
14498                         const fp16type  eta2(eta.asDouble() * eta.asDouble());
14499                         const fp16type  dp2     (dp.asDouble() * dp.asDouble());
14500                         const fp16type  dp1     (1.0 - dp2.asDouble());
14501                         const fp16type  dpe     (eta2.asDouble() * dp1.asDouble());
14502                         const fp16type  k       (1.0 - dpe.asDouble());
14503
14504                         if (k.asDouble() < 0.0)
14505                         {
14506                                 const fp16type  zero    (0.0);
14507
14508                                 for (size_t componentNdx = 0; componentNdx < getOutCompCount(); ++componentNdx)
14509                                         out[componentNdx] = zero.bits();
14510                         }
14511                         else
14512                         {
14513                                 const fp16type  sk      (deSqrt(k.asDouble()));
14514
14515                                 for (size_t componentNdx = 0; componentNdx < getOutCompCount(); ++componentNdx)
14516                                 {
14517                                         const fp16type  i               (in[0][componentNdx]);
14518                                         const fp16type  n               (in[1][componentNdx]);
14519                                         const fp16type  etai    (i.asDouble() * eta.asDouble());
14520                                         const fp16type  etadp   (eta.asDouble() * dp.asDouble());
14521                                         const fp16type  etadpk  (etadp.asDouble() + sk.asDouble());
14522                                         const fp16type  etadpkn (etadpk.asDouble() * n.asDouble());
14523                                         const fp16type  full    (etai.asDouble() - etadpkn.asDouble());
14524
14525                                         if (full.isNaN() || full.isInf())
14526                                                 return false;
14527
14528                                         out[componentNdx] = full.bits();
14529                                 }
14530                         }
14531                 }
14532
14533                 for (size_t ndx = 0; ndx < getOutCompCount(); ++ndx)
14534                         min[ndx] = getMin(fp16type(out[ndx]).asDouble(), getULPs(in));
14535                 for (size_t ndx = 0; ndx < getOutCompCount(); ++ndx)
14536                         max[ndx] = getMax(fp16type(out[ndx]).asDouble(), getULPs(in));
14537
14538                 return true;
14539         }
14540 };
14541
14542 struct fp16Dot : public fp16AllComponents
14543 {
14544         fp16Dot() : fp16AllComponents()
14545         {
14546                 flavorNames.push_back("EmulatingFP16");
14547                 flavorNames.push_back("FloatCalc");
14548                 flavorNames.push_back("DoubleCalc");
14549
14550                 // flavorNames will be extended later
14551         }
14552
14553         virtual void    setArgCompCount                 (size_t argNo, size_t compCount)
14554         {
14555                 DE_ASSERT(argCompCount[argNo] == 0); // Once only
14556
14557                 if (argNo == 0 && argCompCount[argNo] == 0)
14558                 {
14559                         const size_t            maxPermutationsCount    = 24u; // Equal to 4!
14560                         std::vector<int>        indices;
14561
14562                         for (size_t componentNdx = 0; componentNdx < compCount; ++componentNdx)
14563                                 indices.push_back(static_cast<int>(componentNdx));
14564
14565                         m_permutations.reserve(maxPermutationsCount);
14566
14567                         permutationsFlavorStart = flavorNames.size();
14568
14569                         do
14570                         {
14571                                 tcu::UVec4      permutation;
14572                                 std::string     name            = "Permutted_";
14573
14574                                 for (size_t componentNdx = 0; componentNdx < compCount; ++componentNdx)
14575                                 {
14576                                         permutation[static_cast<int>(componentNdx)] = indices[componentNdx];
14577                                         name += de::toString(indices[componentNdx]);
14578                                 }
14579
14580                                 m_permutations.push_back(permutation);
14581                                 flavorNames.push_back(name);
14582
14583                         } while(std::next_permutation(indices.begin(), indices.end()));
14584
14585                         permutationsFlavorEnd = flavorNames.size();
14586                 }
14587
14588                 fp16AllComponents::setArgCompCount(argNo, compCount);
14589         }
14590
14591         virtual double  getULPs(vector<const deFloat16*>& in)
14592         {
14593                 DE_UNREF(in);
14594
14595                 return 16.0; // This is not a precision test. Value is not from spec
14596         }
14597
14598         template<class fp16type>
14599         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
14600         {
14601                 DE_ASSERT(in.size() == 2);
14602                 DE_ASSERT(getArgCompCount(0) == getArgCompCount(1));
14603                 DE_ASSERT(getOutCompCount() == 1);
14604
14605                 double  result  (0.0);
14606                 double  eps             (0.0);
14607
14608                 if (getFlavor() == 0)
14609                 {
14610                         fp16type        dp      (0.0);
14611
14612                         for (size_t componentNdx = 0; componentNdx < getArgCompCount(1); ++componentNdx)
14613                         {
14614                                 const fp16type  x       (in[0][componentNdx]);
14615                                 const fp16type  y       (in[1][componentNdx]);
14616                                 const fp16type  q       (x.asDouble() * y.asDouble());
14617
14618                                 dp = fp16type(dp.asDouble() + q.asDouble());
14619                                 eps += floatFormat16.ulp(q.asDouble(), 2.0);
14620                         }
14621
14622                         result = dp.asDouble();
14623                 }
14624                 else if (getFlavor() == 1)
14625                 {
14626                         float   dp      (0.0);
14627
14628                         for (size_t componentNdx = 0; componentNdx < getArgCompCount(1); ++componentNdx)
14629                         {
14630                                 const fp16type  x       (in[0][componentNdx]);
14631                                 const fp16type  y       (in[1][componentNdx]);
14632                                 const float             q       (x.asFloat() * y.asFloat());
14633
14634                                 dp += q;
14635                                 eps += floatFormat16.ulp(static_cast<double>(q), 2.0);
14636                         }
14637
14638                         result = dp;
14639                 }
14640                 else if (getFlavor() == 2)
14641                 {
14642                         double  dp      (0.0);
14643
14644                         for (size_t componentNdx = 0; componentNdx < getArgCompCount(1); ++componentNdx)
14645                         {
14646                                 const fp16type  x       (in[0][componentNdx]);
14647                                 const fp16type  y       (in[1][componentNdx]);
14648                                 const double    q       (x.asDouble() * y.asDouble());
14649
14650                                 dp += q;
14651                                 eps += floatFormat16.ulp(q, 2.0);
14652                         }
14653
14654                         result = dp;
14655                 }
14656                 else if (de::inBounds<size_t>(getFlavor(), permutationsFlavorStart, permutationsFlavorEnd))
14657                 {
14658                         const int                       compCount               (static_cast<int>(getArgCompCount(1)));
14659                         const size_t            permutationNdx  (getFlavor() - permutationsFlavorStart);
14660                         const tcu::UVec4&       permutation             (m_permutations[permutationNdx]);
14661                         fp16type                        dp                              (0.0);
14662
14663                         for (int permComponentNdx = 0; permComponentNdx < compCount; ++permComponentNdx)
14664                         {
14665                                 const size_t            componentNdx    (permutation[permComponentNdx]);
14666                                 const fp16type          x                               (in[0][componentNdx]);
14667                                 const fp16type          y                               (in[1][componentNdx]);
14668                                 const fp16type          q                               (x.asDouble() * y.asDouble());
14669
14670                                 dp = fp16type(dp.asDouble() + q.asDouble());
14671                                 eps += floatFormat16.ulp(q.asDouble(), 2.0);
14672                         }
14673
14674                         result = dp.asDouble();
14675                 }
14676                 else
14677                 {
14678                         TCU_THROW(InternalError, "Unknown flavor");
14679                 }
14680
14681                 out[0] = fp16type(result).bits();
14682                 min[0] = result - eps;
14683                 max[0] = result + eps;
14684
14685                 return true;
14686         }
14687
14688 private:
14689         std::vector<tcu::UVec4> m_permutations;
14690         size_t                                  permutationsFlavorStart;
14691         size_t                                  permutationsFlavorEnd;
14692 };
14693
14694 struct fp16VectorTimesScalar : public fp16AllComponents
14695 {
14696         virtual double getULPs(vector<const deFloat16*>& in)
14697         {
14698                 DE_UNREF(in);
14699
14700                 return 2.0;
14701         }
14702
14703         template<class fp16type>
14704         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
14705         {
14706                 DE_ASSERT(in.size() == 2);
14707                 DE_ASSERT(getArgCompCount(0) == getOutCompCount());
14708                 DE_ASSERT(getArgCompCount(1) == 1);
14709
14710                 fp16type        s       (*in[1]);
14711
14712                 for (size_t componentNdx = 0; componentNdx < getArgCompCount(0); ++componentNdx)
14713                 {
14714                         const fp16type  x          (in[0][componentNdx]);
14715                         const double    result (s.asDouble() * x.asDouble());
14716                         const fp16type  m          (result);
14717
14718                         out[componentNdx] = m.bits();
14719                         min[componentNdx] = getMin(result, getULPs(in));
14720                         max[componentNdx] = getMax(result, getULPs(in));
14721                 }
14722
14723                 return true;
14724         }
14725 };
14726
14727 struct fp16MatrixBase : public fp16AllComponents
14728 {
14729         deUint32                getComponentValidity                    ()
14730         {
14731                 return static_cast<deUint32>(-1);
14732         }
14733
14734         inline size_t   getNdx                                                  (const size_t rowCount, const size_t col, const size_t row)
14735         {
14736                 const size_t minComponentCount  = 0;
14737                 const size_t maxComponentCount  = 3;
14738                 const size_t alignedRowsCount   = (rowCount == 3) ? 4 : rowCount;
14739
14740                 DE_ASSERT(de::inRange(rowCount, minComponentCount + 1, maxComponentCount + 1));
14741                 DE_ASSERT(de::inRange(col, minComponentCount, maxComponentCount));
14742                 DE_ASSERT(de::inBounds(row, minComponentCount, rowCount));
14743                 DE_UNREF(minComponentCount);
14744                 DE_UNREF(maxComponentCount);
14745
14746                 return col * alignedRowsCount + row;
14747         }
14748
14749         deUint32                getComponentMatrixValidityMask  (size_t cols, size_t rows)
14750         {
14751                 deUint32        result  = 0u;
14752
14753                 for (size_t rowNdx = 0; rowNdx < rows; ++rowNdx)
14754                         for (size_t colNdx = 0; colNdx < cols; ++colNdx)
14755                         {
14756                                 const size_t bitNdx = getNdx(rows, colNdx, rowNdx);
14757
14758                                 DE_ASSERT(bitNdx < sizeof(result) * 8);
14759
14760                                 result |= (1<<bitNdx);
14761                         }
14762
14763                 return result;
14764         }
14765 };
14766
14767 template<size_t cols, size_t rows>
14768 struct fp16Transpose : public fp16MatrixBase
14769 {
14770         virtual double getULPs(vector<const deFloat16*>& in)
14771         {
14772                 DE_UNREF(in);
14773
14774                 return 1.0;
14775         }
14776
14777         deUint32        getComponentValidity    ()
14778         {
14779                 return getComponentMatrixValidityMask(rows, cols);
14780         }
14781
14782         template<class fp16type>
14783         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
14784         {
14785                 DE_ASSERT(in.size() == 1);
14786
14787                 const size_t            alignedCols     = (cols == 3) ? 4 : cols;
14788                 const size_t            alignedRows     = (rows == 3) ? 4 : rows;
14789                 vector<deFloat16>       output          (alignedCols * alignedRows, 0);
14790
14791                 DE_ASSERT(output.size() == alignedCols * alignedRows);
14792
14793                 for (size_t rowNdx = 0; rowNdx < rows; ++rowNdx)
14794                         for (size_t colNdx = 0; colNdx < cols; ++colNdx)
14795                                 output[rowNdx * alignedCols + colNdx] = in[0][colNdx * alignedRows + rowNdx];
14796
14797                 deMemcpy(out, &output[0], sizeof(deFloat16) * output.size());
14798                 deMemcpy(min, &output[0], sizeof(deFloat16) * output.size());
14799                 deMemcpy(max, &output[0], sizeof(deFloat16) * output.size());
14800
14801                 return true;
14802         }
14803 };
14804
14805 template<size_t cols, size_t rows>
14806 struct fp16MatrixTimesScalar : public fp16MatrixBase
14807 {
14808         virtual double getULPs(vector<const deFloat16*>& in)
14809         {
14810                 DE_UNREF(in);
14811
14812                 return 4.0;
14813         }
14814
14815         deUint32        getComponentValidity    ()
14816         {
14817                 return getComponentMatrixValidityMask(cols, rows);
14818         }
14819
14820         template<class fp16type>
14821         bool calc(vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
14822         {
14823                 DE_ASSERT(in.size() == 2);
14824                 DE_ASSERT(getArgCompCount(1) == 1);
14825
14826                 const fp16type  y                       (in[1][0]);
14827                 const float             scalar          (y.asFloat());
14828                 const size_t    alignedCols     = (cols == 3) ? 4 : cols;
14829                 const size_t    alignedRows     = (rows == 3) ? 4 : rows;
14830
14831                 DE_ASSERT(getArgCompCount(0) == alignedCols * alignedRows);
14832                 DE_ASSERT(getOutCompCount() == alignedCols * alignedRows);
14833                 DE_UNREF(alignedCols);
14834
14835                 for (size_t rowNdx = 0; rowNdx < rows; ++rowNdx)
14836                         for (size_t colNdx = 0; colNdx < cols; ++colNdx)
14837                         {
14838                                 const size_t    ndx     (colNdx * alignedRows + rowNdx);
14839                                 const fp16type  x       (in[0][ndx]);
14840                                 const double    result  (scalar * x.asFloat());
14841
14842                                 out[ndx] = fp16type(result).bits();
14843                                 min[ndx] = getMin(result, getULPs(in));
14844                                 max[ndx] = getMax(result, getULPs(in));
14845                         }
14846
14847                 return true;
14848         }
14849 };
14850
14851 template<size_t cols, size_t rows>
14852 struct fp16VectorTimesMatrix : public fp16MatrixBase
14853 {
14854         fp16VectorTimesMatrix() : fp16MatrixBase()
14855         {
14856                 flavorNames.push_back("EmulatingFP16");
14857                 flavorNames.push_back("FloatCalc");
14858         }
14859
14860         virtual double getULPs (vector<const deFloat16*>& in)
14861         {
14862                 DE_UNREF(in);
14863
14864                 return (8.0 * cols);
14865         }
14866
14867         deUint32 getComponentValidity ()
14868         {
14869                 return getComponentMatrixValidityMask(cols, 1);
14870         }
14871
14872         template<class fp16type>
14873         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
14874         {
14875                 DE_ASSERT(in.size() == 2);
14876
14877                 const size_t    alignedCols     = (cols == 3) ? 4 : cols;
14878                 const size_t    alignedRows     = (rows == 3) ? 4 : rows;
14879
14880                 DE_ASSERT(getOutCompCount() == cols);
14881                 DE_ASSERT(getArgCompCount(0) == rows);
14882                 DE_ASSERT(getArgCompCount(1) == alignedCols * alignedRows);
14883                 DE_UNREF(alignedCols);
14884
14885                 if (getFlavor() == 0)
14886                 {
14887                         for (size_t colNdx = 0; colNdx < cols; ++colNdx)
14888                         {
14889                                 fp16type        s       (fp16type::zero(1));
14890
14891                                 for (size_t rowNdx = 0; rowNdx < rows; ++rowNdx)
14892                                 {
14893                                         const fp16type  v       (in[0][rowNdx]);
14894                                         const float             vf      (v.asFloat());
14895                                         const size_t    ndx     (colNdx * alignedRows + rowNdx);
14896                                         const fp16type  x       (in[1][ndx]);
14897                                         const float             xf      (x.asFloat());
14898                                         const fp16type  m       (vf * xf);
14899
14900                                         s = fp16type(s.asFloat() + m.asFloat());
14901                                 }
14902
14903                                 out[colNdx] = s.bits();
14904                                 min[colNdx] = getMin(s.asDouble(), getULPs(in));
14905                                 max[colNdx] = getMax(s.asDouble(), getULPs(in));
14906                         }
14907                 }
14908                 else if (getFlavor() == 1)
14909                 {
14910                         for (size_t colNdx = 0; colNdx < cols; ++colNdx)
14911                         {
14912                                 float   s       (0.0f);
14913
14914                                 for (size_t rowNdx = 0; rowNdx < rows; ++rowNdx)
14915                                 {
14916                                         const fp16type  v       (in[0][rowNdx]);
14917                                         const float             vf      (v.asFloat());
14918                                         const size_t    ndx     (colNdx * alignedRows + rowNdx);
14919                                         const fp16type  x       (in[1][ndx]);
14920                                         const float             xf      (x.asFloat());
14921                                         const float             m       (vf * xf);
14922
14923                                         s += m;
14924                                 }
14925
14926                                 out[colNdx] = fp16type(s).bits();
14927                                 min[colNdx] = getMin(static_cast<double>(s), getULPs(in));
14928                                 max[colNdx] = getMax(static_cast<double>(s), getULPs(in));
14929                         }
14930                 }
14931                 else
14932                 {
14933                         TCU_THROW(InternalError, "Unknown flavor");
14934                 }
14935
14936                 return true;
14937         }
14938 };
14939
14940 template<size_t cols, size_t rows>
14941 struct fp16MatrixTimesVector : public fp16MatrixBase
14942 {
14943         fp16MatrixTimesVector() : fp16MatrixBase()
14944         {
14945                 flavorNames.push_back("EmulatingFP16");
14946                 flavorNames.push_back("FloatCalc");
14947         }
14948
14949         virtual double getULPs (vector<const deFloat16*>& in)
14950         {
14951                 DE_UNREF(in);
14952
14953                 return (8.0 * rows);
14954         }
14955
14956         deUint32 getComponentValidity ()
14957         {
14958                 return getComponentMatrixValidityMask(rows, 1);
14959         }
14960
14961         template<class fp16type>
14962         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
14963         {
14964                 DE_ASSERT(in.size() == 2);
14965
14966                 const size_t    alignedCols     = (cols == 3) ? 4 : cols;
14967                 const size_t    alignedRows     = (rows == 3) ? 4 : rows;
14968
14969                 DE_ASSERT(getOutCompCount() == rows);
14970                 DE_ASSERT(getArgCompCount(0) == alignedCols * alignedRows);
14971                 DE_ASSERT(getArgCompCount(1) == cols);
14972                 DE_UNREF(alignedCols);
14973
14974                 if (getFlavor() == 0)
14975                 {
14976                         for (size_t rowNdx = 0; rowNdx < rows; ++rowNdx)
14977                         {
14978                                 fp16type        s       (fp16type::zero(1));
14979
14980                                 for (size_t colNdx = 0; colNdx < cols; ++colNdx)
14981                                 {
14982                                         const size_t    ndx     (colNdx * alignedRows + rowNdx);
14983                                         const fp16type  x       (in[0][ndx]);
14984                                         const float             xf      (x.asFloat());
14985                                         const fp16type  v       (in[1][colNdx]);
14986                                         const float             vf      (v.asFloat());
14987                                         const fp16type  m       (vf * xf);
14988
14989                                         s = fp16type(s.asFloat() + m.asFloat());
14990                                 }
14991
14992                                 out[rowNdx] = s.bits();
14993                                 min[rowNdx] = getMin(s.asDouble(), getULPs(in));
14994                                 max[rowNdx] = getMax(s.asDouble(), getULPs(in));
14995                         }
14996                 }
14997                 else if (getFlavor() == 1)
14998                 {
14999                         for (size_t rowNdx = 0; rowNdx < rows; ++rowNdx)
15000                         {
15001                                 float   s       (0.0f);
15002
15003                                 for (size_t colNdx = 0; colNdx < cols; ++colNdx)
15004                                 {
15005                                         const size_t    ndx     (colNdx * alignedRows + rowNdx);
15006                                         const fp16type  x       (in[0][ndx]);
15007                                         const float             xf      (x.asFloat());
15008                                         const fp16type  v       (in[1][colNdx]);
15009                                         const float             vf      (v.asFloat());
15010                                         const float             m       (vf * xf);
15011
15012                                         s += m;
15013                                 }
15014
15015                                 out[rowNdx] = fp16type(s).bits();
15016                                 min[rowNdx] = getMin(static_cast<double>(s), getULPs(in));
15017                                 max[rowNdx] = getMax(static_cast<double>(s), getULPs(in));
15018                         }
15019                 }
15020                 else
15021                 {
15022                         TCU_THROW(InternalError, "Unknown flavor");
15023                 }
15024
15025                 return true;
15026         }
15027 };
15028
15029 template<size_t colsL, size_t rowsL, size_t colsR, size_t rowsR>
15030 struct fp16MatrixTimesMatrix : public fp16MatrixBase
15031 {
15032         fp16MatrixTimesMatrix() : fp16MatrixBase()
15033         {
15034                 flavorNames.push_back("EmulatingFP16");
15035                 flavorNames.push_back("FloatCalc");
15036         }
15037
15038         virtual double getULPs (vector<const deFloat16*>& in)
15039         {
15040                 DE_UNREF(in);
15041
15042                 return 32.0;
15043         }
15044
15045         deUint32 getComponentValidity ()
15046         {
15047                 return getComponentMatrixValidityMask(colsR, rowsL);
15048         }
15049
15050         template<class fp16type>
15051         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
15052         {
15053                 DE_STATIC_ASSERT(colsL == rowsR);
15054
15055                 DE_ASSERT(in.size() == 2);
15056
15057                 const size_t    alignedColsL    = (colsL == 3) ? 4 : colsL;
15058                 const size_t    alignedRowsL    = (rowsL == 3) ? 4 : rowsL;
15059                 const size_t    alignedColsR    = (colsR == 3) ? 4 : colsR;
15060                 const size_t    alignedRowsR    = (rowsR == 3) ? 4 : rowsR;
15061
15062                 DE_ASSERT(getOutCompCount() == alignedColsR * alignedRowsL);
15063                 DE_ASSERT(getArgCompCount(0) == alignedColsL * alignedRowsL);
15064                 DE_ASSERT(getArgCompCount(1) == alignedColsR * alignedRowsR);
15065                 DE_UNREF(alignedColsL);
15066                 DE_UNREF(alignedColsR);
15067
15068                 if (getFlavor() == 0)
15069                 {
15070                         for (size_t rowNdx = 0; rowNdx < rowsL; ++rowNdx)
15071                         {
15072                                 for (size_t colNdx = 0; colNdx < colsR; ++colNdx)
15073                                 {
15074                                         const size_t    ndx     (colNdx * alignedRowsL + rowNdx);
15075                                         fp16type                s       (fp16type::zero(1));
15076
15077                                         for (size_t commonNdx = 0; commonNdx < colsL; ++commonNdx)
15078                                         {
15079                                                 const size_t    ndxl    (commonNdx * alignedRowsL + rowNdx);
15080                                                 const fp16type  l               (in[0][ndxl]);
15081                                                 const float             lf              (l.asFloat());
15082                                                 const size_t    ndxr    (colNdx * alignedRowsR + commonNdx);
15083                                                 const fp16type  r               (in[1][ndxr]);
15084                                                 const float             rf              (r.asFloat());
15085                                                 const fp16type  m               (lf * rf);
15086
15087                                                 s = fp16type(s.asFloat() + m.asFloat());
15088                                         }
15089
15090                                         out[ndx] = s.bits();
15091                                         min[ndx] = getMin(s.asDouble(), getULPs(in));
15092                                         max[ndx] = getMax(s.asDouble(), getULPs(in));
15093                                 }
15094                         }
15095                 }
15096                 else if (getFlavor() == 1)
15097                 {
15098                         for (size_t rowNdx = 0; rowNdx < rowsL; ++rowNdx)
15099                         {
15100                                 for (size_t colNdx = 0; colNdx < colsR; ++colNdx)
15101                                 {
15102                                         const size_t    ndx     (colNdx * alignedRowsL + rowNdx);
15103                                         float                   s       (0.0f);
15104
15105                                         for (size_t commonNdx = 0; commonNdx < colsL; ++commonNdx)
15106                                         {
15107                                                 const size_t    ndxl    (commonNdx * alignedRowsL + rowNdx);
15108                                                 const fp16type  l               (in[0][ndxl]);
15109                                                 const float             lf              (l.asFloat());
15110                                                 const size_t    ndxr    (colNdx * alignedRowsR + commonNdx);
15111                                                 const fp16type  r               (in[1][ndxr]);
15112                                                 const float             rf              (r.asFloat());
15113                                                 const float             m               (lf * rf);
15114
15115                                                 s += m;
15116                                         }
15117
15118                                         out[ndx] = fp16type(s).bits();
15119                                         min[ndx] = getMin(static_cast<double>(s), getULPs(in));
15120                                         max[ndx] = getMax(static_cast<double>(s), getULPs(in));
15121                                 }
15122                         }
15123                 }
15124                 else
15125                 {
15126                         TCU_THROW(InternalError, "Unknown flavor");
15127                 }
15128
15129                 return true;
15130         }
15131 };
15132
15133 template<size_t cols, size_t rows>
15134 struct fp16OuterProduct : public fp16MatrixBase
15135 {
15136         virtual double getULPs (vector<const deFloat16*>& in)
15137         {
15138                 DE_UNREF(in);
15139
15140                 return 2.0;
15141         }
15142
15143         deUint32 getComponentValidity ()
15144         {
15145                 return getComponentMatrixValidityMask(cols, rows);
15146         }
15147
15148         template<class fp16type>
15149         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
15150         {
15151                 DE_ASSERT(in.size() == 2);
15152
15153                 const size_t    alignedCols     = (cols == 3) ? 4 : cols;
15154                 const size_t    alignedRows     = (rows == 3) ? 4 : rows;
15155
15156                 DE_ASSERT(getArgCompCount(0) == rows);
15157                 DE_ASSERT(getArgCompCount(1) == cols);
15158                 DE_ASSERT(getOutCompCount() == alignedCols * alignedRows);
15159                 DE_UNREF(alignedCols);
15160
15161                 for (size_t rowNdx = 0; rowNdx < rows; ++rowNdx)
15162                 {
15163                         for (size_t colNdx = 0; colNdx < cols; ++colNdx)
15164                         {
15165                                 const size_t    ndx     (colNdx * alignedRows + rowNdx);
15166                                 const fp16type  x       (in[0][rowNdx]);
15167                                 const float             xf      (x.asFloat());
15168                                 const fp16type  y       (in[1][colNdx]);
15169                                 const float             yf      (y.asFloat());
15170                                 const fp16type  m       (xf * yf);
15171
15172                                 out[ndx] = m.bits();
15173                                 min[ndx] = getMin(m.asDouble(), getULPs(in));
15174                                 max[ndx] = getMax(m.asDouble(), getULPs(in));
15175                         }
15176                 }
15177
15178                 return true;
15179         }
15180 };
15181
15182 template<size_t size>
15183 struct fp16Determinant;
15184
15185 template<>
15186 struct fp16Determinant<2> : public fp16MatrixBase
15187 {
15188         virtual double getULPs (vector<const deFloat16*>& in)
15189         {
15190                 DE_UNREF(in);
15191
15192                 return 128.0; // This is not a precision test. Value is not from spec
15193         }
15194
15195         deUint32 getComponentValidity ()
15196         {
15197                 return 1;
15198         }
15199
15200         template<class fp16type>
15201         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
15202         {
15203                 const size_t    cols            = 2;
15204                 const size_t    rows            = 2;
15205                 const size_t    alignedCols     = (cols == 3) ? 4 : cols;
15206                 const size_t    alignedRows     = (rows == 3) ? 4 : rows;
15207
15208                 DE_ASSERT(in.size() == 1);
15209                 DE_ASSERT(getOutCompCount() == 1);
15210                 DE_ASSERT(getArgCompCount(0) == alignedRows * alignedCols);
15211                 DE_UNREF(alignedCols);
15212                 DE_UNREF(alignedRows);
15213
15214                 // [ a b ]
15215                 // [ c d ]
15216                 const float             a               (fp16type(in[0][getNdx(rows, 0, 0)]).asFloat());
15217                 const float             b               (fp16type(in[0][getNdx(rows, 1, 0)]).asFloat());
15218                 const float             c               (fp16type(in[0][getNdx(rows, 0, 1)]).asFloat());
15219                 const float             d               (fp16type(in[0][getNdx(rows, 1, 1)]).asFloat());
15220                 const float             ad              (a * d);
15221                 const fp16type  adf16   (ad);
15222                 const float             bc              (b * c);
15223                 const fp16type  bcf16   (bc);
15224                 const float             r               (adf16.asFloat() - bcf16.asFloat());
15225                 const fp16type  rf16    (r);
15226
15227                 out[0] = rf16.bits();
15228                 min[0] = getMin(r, getULPs(in));
15229                 max[0] = getMax(r, getULPs(in));
15230
15231                 return true;
15232         }
15233 };
15234
15235 template<>
15236 struct fp16Determinant<3> : public fp16MatrixBase
15237 {
15238         virtual double getULPs (vector<const deFloat16*>& in)
15239         {
15240                 DE_UNREF(in);
15241
15242                 return 128.0; // This is not a precision test. Value is not from spec
15243         }
15244
15245         deUint32 getComponentValidity ()
15246         {
15247                 return 1;
15248         }
15249
15250         template<class fp16type>
15251         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
15252         {
15253                 const size_t    cols            = 3;
15254                 const size_t    rows            = 3;
15255                 const size_t    alignedCols     = (cols == 3) ? 4 : cols;
15256                 const size_t    alignedRows     = (rows == 3) ? 4 : rows;
15257
15258                 DE_ASSERT(in.size() == 1);
15259                 DE_ASSERT(getOutCompCount() == 1);
15260                 DE_ASSERT(getArgCompCount(0) == alignedRows * alignedCols);
15261                 DE_UNREF(alignedCols);
15262                 DE_UNREF(alignedRows);
15263
15264                 // [ a b c ]
15265                 // [ d e f ]
15266                 // [ g h i ]
15267                 const float             a               (fp16type(in[0][getNdx(rows, 0, 0)]).asFloat());
15268                 const float             b               (fp16type(in[0][getNdx(rows, 1, 0)]).asFloat());
15269                 const float             c               (fp16type(in[0][getNdx(rows, 2, 0)]).asFloat());
15270                 const float             d               (fp16type(in[0][getNdx(rows, 0, 1)]).asFloat());
15271                 const float             e               (fp16type(in[0][getNdx(rows, 1, 1)]).asFloat());
15272                 const float             f               (fp16type(in[0][getNdx(rows, 2, 1)]).asFloat());
15273                 const float             g               (fp16type(in[0][getNdx(rows, 0, 2)]).asFloat());
15274                 const float             h               (fp16type(in[0][getNdx(rows, 1, 2)]).asFloat());
15275                 const float             i               (fp16type(in[0][getNdx(rows, 2, 2)]).asFloat());
15276                 const fp16type  aei             (a * e * i);
15277                 const fp16type  bfg             (b * f * g);
15278                 const fp16type  cdh             (c * d * h);
15279                 const fp16type  ceg             (c * e * g);
15280                 const fp16type  bdi             (b * d * i);
15281                 const fp16type  afh             (a * f * h);
15282                 const float             r               (aei.asFloat() + bfg.asFloat() + cdh.asFloat() - ceg.asFloat() - bdi.asFloat() - afh.asFloat());
15283                 const fp16type  rf16    (r);
15284
15285                 out[0] = rf16.bits();
15286                 min[0] = getMin(r, getULPs(in));
15287                 max[0] = getMax(r, getULPs(in));
15288
15289                 return true;
15290         }
15291 };
15292
15293 template<>
15294 struct fp16Determinant<4> : public fp16MatrixBase
15295 {
15296         virtual double getULPs (vector<const deFloat16*>& in)
15297         {
15298                 DE_UNREF(in);
15299
15300                 return 128.0; // This is not a precision test. Value is not from spec
15301         }
15302
15303         deUint32 getComponentValidity ()
15304         {
15305                 return 1;
15306         }
15307
15308         template<class fp16type>
15309         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
15310         {
15311                 const size_t    rows            = 4;
15312                 const size_t    cols            = 4;
15313                 const size_t    alignedCols     = (cols == 3) ? 4 : cols;
15314                 const size_t    alignedRows     = (rows == 3) ? 4 : rows;
15315
15316                 DE_ASSERT(in.size() == 1);
15317                 DE_ASSERT(getOutCompCount() == 1);
15318                 DE_ASSERT(getArgCompCount(0) == alignedRows * alignedCols);
15319                 DE_UNREF(alignedCols);
15320                 DE_UNREF(alignedRows);
15321
15322                 // [ a b c d ]
15323                 // [ e f g h ]
15324                 // [ i j k l ]
15325                 // [ m n o p ]
15326                 const float             a               (fp16type(in[0][getNdx(rows, 0, 0)]).asFloat());
15327                 const float             b               (fp16type(in[0][getNdx(rows, 1, 0)]).asFloat());
15328                 const float             c               (fp16type(in[0][getNdx(rows, 2, 0)]).asFloat());
15329                 const float             d               (fp16type(in[0][getNdx(rows, 3, 0)]).asFloat());
15330                 const float             e               (fp16type(in[0][getNdx(rows, 0, 1)]).asFloat());
15331                 const float             f               (fp16type(in[0][getNdx(rows, 1, 1)]).asFloat());
15332                 const float             g               (fp16type(in[0][getNdx(rows, 2, 1)]).asFloat());
15333                 const float             h               (fp16type(in[0][getNdx(rows, 3, 1)]).asFloat());
15334                 const float             i               (fp16type(in[0][getNdx(rows, 0, 2)]).asFloat());
15335                 const float             j               (fp16type(in[0][getNdx(rows, 1, 2)]).asFloat());
15336                 const float             k               (fp16type(in[0][getNdx(rows, 2, 2)]).asFloat());
15337                 const float             l               (fp16type(in[0][getNdx(rows, 3, 2)]).asFloat());
15338                 const float             m               (fp16type(in[0][getNdx(rows, 0, 3)]).asFloat());
15339                 const float             n               (fp16type(in[0][getNdx(rows, 1, 3)]).asFloat());
15340                 const float             o               (fp16type(in[0][getNdx(rows, 2, 3)]).asFloat());
15341                 const float             p               (fp16type(in[0][getNdx(rows, 3, 3)]).asFloat());
15342
15343                 // [ f g h ]
15344                 // [ j k l ]
15345                 // [ n o p ]
15346                 const fp16type  fkp             (f * k * p);
15347                 const fp16type  gln             (g * l * n);
15348                 const fp16type  hjo             (h * j * o);
15349                 const fp16type  hkn             (h * k * n);
15350                 const fp16type  gjp             (g * j * p);
15351                 const fp16type  flo             (f * l * o);
15352                 const fp16type  detA    (a * (fkp.asFloat() + gln.asFloat() + hjo.asFloat() - hkn.asFloat() - gjp.asFloat() - flo.asFloat()));
15353
15354                 // [ e g h ]
15355                 // [ i k l ]
15356                 // [ m o p ]
15357                 const fp16type  ekp             (e * k * p);
15358                 const fp16type  glm             (g * l * m);
15359                 const fp16type  hio             (h * i * o);
15360                 const fp16type  hkm             (h * k * m);
15361                 const fp16type  gip             (g * i * p);
15362                 const fp16type  elo             (e * l * o);
15363                 const fp16type  detB    (b * (ekp.asFloat() + glm.asFloat() + hio.asFloat() - hkm.asFloat() - gip.asFloat() - elo.asFloat()));
15364
15365                 // [ e f h ]
15366                 // [ i j l ]
15367                 // [ m n p ]
15368                 const fp16type  ejp             (e * j * p);
15369                 const fp16type  flm             (f * l * m);
15370                 const fp16type  hin             (h * i * n);
15371                 const fp16type  hjm             (h * j * m);
15372                 const fp16type  fip             (f * i * p);
15373                 const fp16type  eln             (e * l * n);
15374                 const fp16type  detC    (c * (ejp.asFloat() + flm.asFloat() + hin.asFloat() - hjm.asFloat() - fip.asFloat() - eln.asFloat()));
15375
15376                 // [ e f g ]
15377                 // [ i j k ]
15378                 // [ m n o ]
15379                 const fp16type  ejo             (e * j * o);
15380                 const fp16type  fkm             (f * k * m);
15381                 const fp16type  gin             (g * i * n);
15382                 const fp16type  gjm             (g * j * m);
15383                 const fp16type  fio             (f * i * o);
15384                 const fp16type  ekn             (e * k * n);
15385                 const fp16type  detD    (d * (ejo.asFloat() + fkm.asFloat() + gin.asFloat() - gjm.asFloat() - fio.asFloat() - ekn.asFloat()));
15386
15387                 const float             r               (detA.asFloat() - detB.asFloat() + detC.asFloat() - detD.asFloat());
15388                 const fp16type  rf16    (r);
15389
15390                 out[0] = rf16.bits();
15391                 min[0] = getMin(r, getULPs(in));
15392                 max[0] = getMax(r, getULPs(in));
15393
15394                 return true;
15395         }
15396 };
15397
15398 template<size_t size>
15399 struct fp16Inverse;
15400
15401 template<>
15402 struct fp16Inverse<2> : public fp16MatrixBase
15403 {
15404         virtual double getULPs (vector<const deFloat16*>& in)
15405         {
15406                 DE_UNREF(in);
15407
15408                 return 128.0; // This is not a precision test. Value is not from spec
15409         }
15410
15411         deUint32 getComponentValidity ()
15412         {
15413                 return getComponentMatrixValidityMask(2, 2);
15414         }
15415
15416         template<class fp16type>
15417         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
15418         {
15419                 const size_t    cols            = 2;
15420                 const size_t    rows            = 2;
15421                 const size_t    alignedCols     = (cols == 3) ? 4 : cols;
15422                 const size_t    alignedRows     = (rows == 3) ? 4 : rows;
15423
15424                 DE_ASSERT(in.size() == 1);
15425                 DE_ASSERT(getOutCompCount() == alignedRows * alignedCols);
15426                 DE_ASSERT(getArgCompCount(0) == alignedRows * alignedCols);
15427                 DE_UNREF(alignedCols);
15428
15429                 // [ a b ]
15430                 // [ c d ]
15431                 const float             a               (fp16type(in[0][getNdx(rows, 0, 0)]).asFloat());
15432                 const float             b               (fp16type(in[0][getNdx(rows, 1, 0)]).asFloat());
15433                 const float             c               (fp16type(in[0][getNdx(rows, 0, 1)]).asFloat());
15434                 const float             d               (fp16type(in[0][getNdx(rows, 1, 1)]).asFloat());
15435                 const float             ad              (a * d);
15436                 const fp16type  adf16   (ad);
15437                 const float             bc              (b * c);
15438                 const fp16type  bcf16   (bc);
15439                 const float             det             (adf16.asFloat() - bcf16.asFloat());
15440                 const fp16type  det16   (det);
15441
15442                 out[0] = fp16type( d / det16.asFloat()).bits();
15443                 out[1] = fp16type(-c / det16.asFloat()).bits();
15444                 out[2] = fp16type(-b / det16.asFloat()).bits();
15445                 out[3] = fp16type( a / det16.asFloat()).bits();
15446
15447                 for (size_t rowNdx = 0; rowNdx < rows; ++rowNdx)
15448                         for (size_t colNdx = 0; colNdx < cols; ++colNdx)
15449                         {
15450                                 const size_t    ndx     (colNdx * alignedRows + rowNdx);
15451                                 const fp16type  s       (out[ndx]);
15452
15453                                 min[ndx] = getMin(s.asDouble(), getULPs(in));
15454                                 max[ndx] = getMax(s.asDouble(), getULPs(in));
15455                         }
15456
15457                 return true;
15458         }
15459 };
15460
15461 inline std::string fp16ToString(deFloat16 val)
15462 {
15463         return tcu::toHex<4>(val).toString() + " (" + de::floatToString(tcu::Float16(val).asFloat(), 10) + ")";
15464 }
15465
15466 template <size_t RES_COMPONENTS, size_t ARG0_COMPONENTS, size_t ARG1_COMPONENTS, size_t ARG2_COMPONENTS, class TestedArithmeticFunction>
15467 bool compareFP16ArithmeticFunc (const std::vector<Resource>& inputs, const vector<AllocationSp>& outputAllocs, const std::vector<Resource>& expectedOutputs, TestLog& log)
15468 {
15469         if (inputs.size() < 1 || inputs.size() > 3 || outputAllocs.size() != 1 || expectedOutputs.size() != 1)
15470                 return false;
15471
15472         const size_t    resultStep                      = (RES_COMPONENTS == 3) ? 4 : RES_COMPONENTS;
15473         const size_t    iterationsCount         = expectedOutputs[0].getByteSize() / (sizeof(deFloat16) * resultStep);
15474         const size_t    inputsSteps[3]          =
15475         {
15476                 (ARG0_COMPONENTS == 3) ? 4 : ARG0_COMPONENTS,
15477                 (ARG1_COMPONENTS == 3) ? 4 : ARG1_COMPONENTS,
15478                 (ARG2_COMPONENTS == 3) ? 4 : ARG2_COMPONENTS,
15479         };
15480
15481         DE_ASSERT(expectedOutputs[0].getByteSize() > 0);
15482         DE_ASSERT(expectedOutputs[0].getByteSize() == sizeof(deFloat16) * iterationsCount * resultStep);
15483
15484         for (size_t inputNdx = 0; inputNdx < inputs.size(); ++inputNdx)
15485         {
15486                 DE_ASSERT(inputs[inputNdx].getByteSize() > 0);
15487                 DE_ASSERT(inputs[inputNdx].getByteSize() == sizeof(deFloat16) * iterationsCount * inputsSteps[inputNdx]);
15488         }
15489
15490         const deFloat16* const          outputAsFP16                                    = (const deFloat16*)outputAllocs[0]->getHostPtr();
15491         TestedArithmeticFunction        func;
15492
15493         func.setOutCompCount(RES_COMPONENTS);
15494         func.setArgCompCount(0, ARG0_COMPONENTS);
15495         func.setArgCompCount(1, ARG1_COMPONENTS);
15496         func.setArgCompCount(2, ARG2_COMPONENTS);
15497
15498         const bool                                      callOncePerComponent                    = func.callOncePerComponent();
15499         const deUint32                          componentValidityMask                   = func.getComponentValidity();
15500         const size_t                            denormModesCount                                = 2;
15501         const char*                                     denormModes[denormModesCount]   = { "keep denormal numbers", "flush to zero" };
15502         const size_t                            successfulRunsPerComponent              = denormModesCount * func.getFlavorCount();
15503         bool                                            success                                                 = true;
15504         size_t                                          validatedCount                                  = 0;
15505
15506         vector<deUint8> inputBytes[3];
15507
15508         for (size_t inputNdx = 0; inputNdx < inputs.size(); ++inputNdx)
15509                 inputs[inputNdx].getBytes(inputBytes[inputNdx]);
15510
15511         const deFloat16* const                  inputsAsFP16[3]                 =
15512         {
15513                 inputs.size() >= 1 ? (const deFloat16*)&inputBytes[0][0] : DE_NULL,
15514                 inputs.size() >= 2 ? (const deFloat16*)&inputBytes[1][0] : DE_NULL,
15515                 inputs.size() >= 3 ? (const deFloat16*)&inputBytes[2][0] : DE_NULL,
15516         };
15517
15518         for (size_t idx = 0; idx < iterationsCount; ++idx)
15519         {
15520                 std::vector<size_t>                     successfulRuns          (RES_COMPONENTS, successfulRunsPerComponent);
15521                 std::vector<std::string>        errors                          (RES_COMPONENTS);
15522                 bool                                            iterationValidated      (true);
15523
15524                 for (size_t denormNdx = 0; denormNdx < 2; ++denormNdx)
15525                 {
15526                         for (size_t flavorNdx = 0; flavorNdx < func.getFlavorCount(); ++flavorNdx)
15527                         {
15528                                 func.setFlavor(flavorNdx);
15529
15530                                 const deFloat16*                        iterationOutputFP16             = &outputAsFP16[idx * resultStep];
15531                                 vector<deFloat16>                       iterationCalculatedFP16 (resultStep, 0);
15532                                 vector<double>                          iterationEdgeMin                (resultStep, 0.0);
15533                                 vector<double>                          iterationEdgeMax                (resultStep, 0.0);
15534                                 vector<const deFloat16*>        arguments;
15535
15536                                 for (size_t componentNdx = 0; componentNdx < RES_COMPONENTS; ++componentNdx)
15537                                 {
15538                                         std::string     error;
15539                                         bool            reportError = false;
15540
15541                                         if (callOncePerComponent || componentNdx == 0)
15542                                         {
15543                                                 bool funcCallResult;
15544
15545                                                 arguments.clear();
15546
15547                                                 for (size_t inputNdx = 0; inputNdx < inputs.size(); ++inputNdx)
15548                                                         arguments.push_back(&inputsAsFP16[inputNdx][idx * inputsSteps[inputNdx] + componentNdx]);
15549
15550                                                 if (denormNdx == 0)
15551                                                         funcCallResult = func.template calc<tcu::Float16>(arguments, &iterationCalculatedFP16[componentNdx], &iterationEdgeMin[componentNdx], &iterationEdgeMax[componentNdx]);
15552                                                 else
15553                                                         funcCallResult = func.template calc<tcu::Float16Denormless>(arguments, &iterationCalculatedFP16[componentNdx], &iterationEdgeMin[componentNdx], &iterationEdgeMax[componentNdx]);
15554
15555                                                 if (!funcCallResult)
15556                                                 {
15557                                                         iterationValidated = false;
15558
15559                                                         if (callOncePerComponent)
15560                                                                 continue;
15561                                                         else
15562                                                                 break;
15563                                                 }
15564                                         }
15565
15566                                         if ((componentValidityMask != 0) && (componentValidityMask & (1<<componentNdx)) == 0)
15567                                                 continue;
15568
15569                                         reportError = !compare16BitFloat(iterationCalculatedFP16[componentNdx], iterationOutputFP16[componentNdx], error);
15570
15571                                         if (reportError)
15572                                         {
15573                                                 tcu::Float16 expected   (iterationCalculatedFP16[componentNdx]);
15574                                                 tcu::Float16 outputted  (iterationOutputFP16[componentNdx]);
15575
15576                                                 if (reportError && expected.isNaN())
15577                                                         reportError = false;
15578
15579                                                 if (reportError && !expected.isNaN() && !outputted.isNaN())
15580                                                 {
15581                                                         if (reportError && !expected.isInf() && !outputted.isInf())
15582                                                         {
15583                                                                 // Ignore rounding
15584                                                                 if (expected.bits() == outputted.bits() + 1 || expected.bits() + 1 == outputted.bits())
15585                                                                         reportError = false;
15586                                                         }
15587
15588                                                         if (reportError && expected.isInf())
15589                                                         {
15590                                                                 // RTZ rounding mode returns +/-65504 instead of Inf on overflow
15591                                                                 if (expected.sign() == 1 && outputted.bits() == 0x7bff && iterationEdgeMin[componentNdx] <= std::numeric_limits<double>::max())
15592                                                                         reportError = false;
15593                                                                 else if (expected.sign() == -1 && outputted.bits() == 0xfbff && iterationEdgeMax[componentNdx] >= -std::numeric_limits<double>::max())
15594                                                                         reportError = false;
15595                                                         }
15596
15597                                                         if (reportError)
15598                                                         {
15599                                                                 const double    outputtedDouble = outputted.asDouble();
15600
15601                                                                 DE_ASSERT(iterationEdgeMin[componentNdx] <= iterationEdgeMax[componentNdx]);
15602
15603                                                                 if (de::inRange(outputtedDouble, iterationEdgeMin[componentNdx], iterationEdgeMax[componentNdx]))
15604                                                                         reportError = false;
15605                                                         }
15606                                                 }
15607
15608                                                 if (reportError)
15609                                                 {
15610                                                         const size_t            inputsComps[3]  =
15611                                                         {
15612                                                                 ARG0_COMPONENTS,
15613                                                                 ARG1_COMPONENTS,
15614                                                                 ARG2_COMPONENTS,
15615                                                         };
15616                                                         string                          inputsValues    ("Inputs:");
15617                                                         string                          flavorName              (func.getFlavorCount() == 1 ? "" : string(" flavor ") + de::toString(flavorNdx) + " (" + func.getCurrentFlavorName() + ")");
15618                                                         std::stringstream       errStream;
15619
15620                                                         for (size_t inputNdx = 0; inputNdx < inputs.size(); ++inputNdx)
15621                                                         {
15622                                                                 const size_t    inputCompsCount = inputsComps[inputNdx];
15623
15624                                                                 inputsValues += " [" + de::toString(inputNdx) + "]=(";
15625
15626                                                                 for (size_t compNdx = 0; compNdx < inputCompsCount; ++compNdx)
15627                                                                 {
15628                                                                         const deFloat16 inputComponentValue = inputsAsFP16[inputNdx][idx * inputsSteps[inputNdx] + compNdx];
15629
15630                                                                         inputsValues += fp16ToString(inputComponentValue) + ((compNdx + 1 == inputCompsCount) ? ")": ", ");
15631                                                                 }
15632                                                         }
15633
15634                                                         errStream       << "At"
15635                                                                                 << " iteration " << de::toString(idx)
15636                                                                                 << " component " << de::toString(componentNdx)
15637                                                                                 << " denormMode " << de::toString(denormNdx)
15638                                                                                 << " (" << denormModes[denormNdx] << ")"
15639                                                                                 << " " << flavorName
15640                                                                                 << " " << inputsValues
15641                                                                                 << " outputted:" + fp16ToString(iterationOutputFP16[componentNdx])
15642                                                                                 << " expected:" + fp16ToString(iterationCalculatedFP16[componentNdx])
15643                                                                                 << " or in range: [" << iterationEdgeMin[componentNdx] << ", " << iterationEdgeMax[componentNdx] << "]."
15644                                                                                 << " " << error << "."
15645                                                                                 << std::endl;
15646
15647                                                         errors[componentNdx] += errStream.str();
15648
15649                                                         successfulRuns[componentNdx]--;
15650                                                 }
15651                                         }
15652                                 }
15653                         }
15654                 }
15655
15656                 for (size_t componentNdx = 0; componentNdx < RES_COMPONENTS; ++componentNdx)
15657                 {
15658                         // Check if any component has total failure
15659                         if (successfulRuns[componentNdx] == 0)
15660                         {
15661                                 // Test failed in all denorm modes and all flavors for certain component: dump errors
15662                                 log << TestLog::Message << errors[componentNdx] << TestLog::EndMessage;
15663
15664                                 success = false;
15665                         }
15666                 }
15667
15668                 if (iterationValidated)
15669                         validatedCount++;
15670         }
15671
15672         if (validatedCount < 16)
15673                 TCU_THROW(InternalError, "Too few samples has been validated.");
15674
15675         return success;
15676 }
15677
15678 // IEEE-754 floating point numbers:
15679 // +--------+------+----------+-------------+
15680 // | binary | sign | exponent | significand |
15681 // +--------+------+----------+-------------+
15682 // | 16-bit |  1   |    5     |     10      |
15683 // +--------+------+----------+-------------+
15684 // | 32-bit |  1   |    8     |     23      |
15685 // +--------+------+----------+-------------+
15686 //
15687 // 16-bit floats:
15688 //
15689 // 0   000 00   00 0000 0001 (0x0001: 2e-24:         minimum positive denormalized)
15690 // 0   000 00   11 1111 1111 (0x03ff: 2e-14 - 2e-24: maximum positive denormalized)
15691 // 0   000 01   00 0000 0000 (0x0400: 2e-14:         minimum positive normalized)
15692 // 0   111 10   11 1111 1111 (0x7bff: 65504:         maximum positive normalized)
15693 //
15694 // 0   000 00   00 0000 0000 (0x0000: +0)
15695 // 0   111 11   00 0000 0000 (0x7c00: +Inf)
15696 // 0   000 00   11 1111 0000 (0x03f0: +Denorm)
15697 // 0   000 01   00 0000 0001 (0x0401: +Norm)
15698 // 0   111 11   00 0000 1111 (0x7c0f: +SNaN)
15699 // 0   111 11   11 1111 0000 (0x7ff0: +QNaN)
15700 // Generate and return 16-bit floats and their corresponding 32-bit values.
15701 //
15702 // The first 14 number pairs are manually picked, while the rest are randomly generated.
15703 // Expected count to be at least 14 (numPicks).
15704 vector<deFloat16> getFloat16a (de::Random& rnd, deUint32 count)
15705 {
15706         vector<deFloat16>       float16;
15707
15708         float16.reserve(count);
15709
15710         // Zero
15711         float16.push_back(deUint16(0x0000));
15712         float16.push_back(deUint16(0x8000));
15713         // Infinity
15714         float16.push_back(deUint16(0x7c00));
15715         float16.push_back(deUint16(0xfc00));
15716         // Normalized
15717         float16.push_back(deUint16(0x0401));
15718         float16.push_back(deUint16(0x8401));
15719         // Some normal number
15720         float16.push_back(deUint16(0x14cb));
15721         float16.push_back(deUint16(0x94cb));
15722         // Min/max positive normal
15723         float16.push_back(deUint16(0x0400));
15724         float16.push_back(deUint16(0x7bff));
15725         // Min/max negative normal
15726         float16.push_back(deUint16(0x8400));
15727         float16.push_back(deUint16(0xfbff));
15728         // PI
15729         float16.push_back(deUint16(0x4248)); // 3.140625
15730         float16.push_back(deUint16(0xb248)); // -3.140625
15731         // PI/2
15732         float16.push_back(deUint16(0x3e48)); // 1.5703125
15733         float16.push_back(deUint16(0xbe48)); // -1.5703125
15734         float16.push_back(deUint16(0x3c00)); // 1.0
15735         float16.push_back(deUint16(0x3800)); // 0.5
15736         // Some useful constants
15737         float16.push_back(tcu::Float16(-2.5f).bits());
15738         float16.push_back(tcu::Float16(-1.0f).bits());
15739         float16.push_back(tcu::Float16( 0.4f).bits());
15740         float16.push_back(tcu::Float16( 2.5f).bits());
15741
15742         const deUint32          numPicks        = static_cast<deUint32>(float16.size());
15743
15744         DE_ASSERT(count >= numPicks);
15745         count -= numPicks;
15746
15747         for (deUint32 numIdx = 0; numIdx < count; ++numIdx)
15748         {
15749                 int                     sign            = (rnd.getUint16() % 2 == 0) ? +1 : -1;
15750                 int                     exponent        = (rnd.getUint16() % 29) - 14 + 1;
15751                 deUint16        mantissa        = static_cast<deUint16>(2 * (rnd.getUint16() % 512));
15752
15753                 // Exclude power of -14 to avoid denorms
15754                 DE_ASSERT(de::inRange(exponent, -13, 15));
15755
15756                 float16.push_back(tcu::Float16::constructBits(sign, exponent, mantissa).bits());
15757         }
15758
15759         return float16;
15760 }
15761
15762 static inline vector<deFloat16> getInputData1 (deUint32 seed, size_t count, size_t argNo)
15763 {
15764         DE_UNREF(argNo);
15765
15766         de::Random      rnd(seed);
15767
15768         return getFloat16a(rnd, static_cast<deUint32>(count));
15769 }
15770
15771 static inline vector<deFloat16> getInputData2 (deUint32 seed, size_t count, size_t argNo)
15772 {
15773         de::Random      rnd             (seed);
15774         size_t          newCount = static_cast<size_t>(deSqrt(double(count)));
15775
15776         DE_ASSERT(newCount * newCount == count);
15777
15778         vector<deFloat16>       float16 = getFloat16a(rnd, static_cast<deUint32>(newCount));
15779
15780         return squarize(float16, static_cast<deUint32>(argNo));
15781 }
15782
15783 static inline vector<deFloat16> getInputData3 (deUint32 seed, size_t count, size_t argNo)
15784 {
15785         if (argNo == 0 || argNo == 1)
15786                 return getInputData2(seed, count, argNo);
15787         else
15788                 return getInputData1(seed<<argNo, count, argNo);
15789 }
15790
15791 vector<deFloat16> getInputData (deUint32 seed, size_t count, size_t compCount, size_t stride, size_t argCount, size_t argNo)
15792 {
15793         DE_UNREF(stride);
15794
15795         vector<deFloat16>       result;
15796
15797         switch (argCount)
15798         {
15799                 case 1:result = getInputData1(seed, count, argNo); break;
15800                 case 2:result = getInputData2(seed, count, argNo); break;
15801                 case 3:result = getInputData3(seed, count, argNo); break;
15802                 default: TCU_THROW(InternalError, "Invalid argument count specified");
15803         }
15804
15805         if (compCount == 3)
15806         {
15807                 const size_t            newCount = (3 * count) / 4;
15808                 vector<deFloat16>       newResult;
15809
15810                 newResult.reserve(result.size());
15811
15812                 for (size_t ndx = 0; ndx < newCount; ++ndx)
15813                 {
15814                         newResult.push_back(result[ndx]);
15815
15816                         if (ndx % 3 == 2)
15817                                 newResult.push_back(0);
15818                 }
15819
15820                 result = newResult;
15821         }
15822
15823         DE_ASSERT(result.size() == count);
15824
15825         return result;
15826 }
15827
15828 // Generator for functions requiring data in range [1, inf]
15829 vector<deFloat16> getInputDataAC (deUint32 seed, size_t count, size_t compCount, size_t stride, size_t argCount, size_t argNo)
15830 {
15831         vector<deFloat16>       result;
15832
15833         result = getInputData(seed, count, compCount, stride, argCount, argNo);
15834
15835         // Filter out values below 1.0 from upper half of numbers
15836         for (size_t idx = result.size() / 2; idx < result.size(); ++idx)
15837         {
15838                 const float f = tcu::Float16(result[idx]).asFloat();
15839
15840                 if (f < 1.0f)
15841                         result[idx] = tcu::Float16(1.0f - f).bits();
15842         }
15843
15844         return result;
15845 }
15846
15847 // Generator for functions requiring data in range [-1, 1]
15848 vector<deFloat16> getInputDataA (deUint32 seed, size_t count, size_t compCount, size_t stride, size_t argCount, size_t argNo)
15849 {
15850         vector<deFloat16>       result;
15851
15852         result = getInputData(seed, count, compCount, stride, argCount, argNo);
15853
15854         for (size_t idx = result.size() / 2; idx < result.size(); ++idx)
15855         {
15856                 const float f = tcu::Float16(result[idx]).asFloat();
15857
15858                 if (!de::inRange(f, -1.0f, 1.0f))
15859                         result[idx] = tcu::Float16(deFloatFrac(f)).bits();
15860         }
15861
15862         return result;
15863 }
15864
15865 // Generator for functions requiring data in range [-pi, pi]
15866 vector<deFloat16> getInputDataPI (deUint32 seed, size_t count, size_t compCount, size_t stride, size_t argCount, size_t argNo)
15867 {
15868         vector<deFloat16>       result;
15869
15870         result = getInputData(seed, count, compCount, stride, argCount, argNo);
15871
15872         for (size_t idx = result.size() / 2; idx < result.size(); ++idx)
15873         {
15874                 const float f = tcu::Float16(result[idx]).asFloat();
15875
15876                 if (!de::inRange(f, -DE_PI, DE_PI))
15877                         result[idx] = tcu::Float16(fmodf(f, DE_PI)).bits();
15878         }
15879
15880         return result;
15881 }
15882
15883 // Generator for functions requiring data in range [0, inf]
15884 vector<deFloat16> getInputDataP (deUint32 seed, size_t count, size_t compCount, size_t stride, size_t argCount, size_t argNo)
15885 {
15886         vector<deFloat16>       result;
15887
15888         result = getInputData(seed, count, compCount, stride, argCount, argNo);
15889
15890         if (argNo == 0)
15891         {
15892                 for (size_t idx = result.size() / 2; idx < result.size(); ++idx)
15893                         result[idx] &= static_cast<deFloat16>(~0x8000);
15894         }
15895
15896         return result;
15897 }
15898
15899 vector<deFloat16> getInputDataV (deUint32 seed, size_t count, size_t compCount, size_t stride, size_t argCount, size_t argNo)
15900 {
15901         DE_UNREF(stride);
15902         DE_UNREF(argCount);
15903
15904         vector<deFloat16>       result;
15905
15906         if (argNo == 0)
15907                 result = getInputData2(seed, count, argNo);
15908         else
15909         {
15910                 const size_t            alignedCount    = (compCount == 3) ? 4 : compCount;
15911                 const size_t            newCountX               = static_cast<size_t>(deSqrt(double(count * alignedCount)));
15912                 const size_t            newCountY               = count / newCountX;
15913                 de::Random                      rnd                             (seed);
15914                 vector<deFloat16>       float16                 = getFloat16a(rnd, static_cast<deUint32>(newCountX));
15915
15916                 DE_ASSERT(newCountX * newCountX == alignedCount * count);
15917
15918                 for (size_t numIdx = 0; numIdx < newCountX; ++numIdx)
15919                 {
15920                         const vector<deFloat16> tmp(newCountY, float16[numIdx]);
15921
15922                         result.insert(result.end(), tmp.begin(), tmp.end());
15923                 }
15924         }
15925
15926         DE_ASSERT(result.size() == count);
15927
15928         return result;
15929 }
15930
15931 vector<deFloat16> getInputDataM (deUint32 seed, size_t count, size_t compCount, size_t stride, size_t argCount, size_t argNo)
15932 {
15933         DE_UNREF(compCount);
15934         DE_UNREF(stride);
15935         DE_UNREF(argCount);
15936
15937         de::Random                      rnd             (seed << argNo);
15938         vector<deFloat16>       result;
15939
15940         result = getFloat16a(rnd, static_cast<deUint32>(count));
15941
15942         DE_ASSERT(result.size() == count);
15943
15944         return result;
15945 }
15946
15947 vector<deFloat16> getInputDataD (deUint32 seed, size_t count, size_t compCount, size_t stride, size_t argCount, size_t argNo)
15948 {
15949         DE_UNREF(compCount);
15950         DE_UNREF(argCount);
15951
15952         de::Random                      rnd             (seed << argNo);
15953         vector<deFloat16>       result;
15954
15955         for (deUint32 numIdx = 0; numIdx < count; ++numIdx)
15956         {
15957                 int num = (rnd.getUint16() % 16) - 8;
15958
15959                 result.push_back(tcu::Float16(float(num)).bits());
15960         }
15961
15962         result[0 * stride] = deUint16(0x7c00); // +Inf
15963         result[1 * stride] = deUint16(0xfc00); // -Inf
15964
15965         DE_ASSERT(result.size() == count);
15966
15967         return result;
15968 }
15969
15970 // Generator for smoothstep function
15971 vector<deFloat16> getInputDataSS (deUint32 seed, size_t count, size_t compCount, size_t stride, size_t argCount, size_t argNo)
15972 {
15973         vector<deFloat16>       result;
15974
15975         result = getInputDataD(seed, count, compCount, stride, argCount, argNo);
15976
15977         if (argNo == 0)
15978         {
15979                 for (size_t idx = result.size() / 2; idx < result.size(); ++idx)
15980                 {
15981                         const float f = tcu::Float16(result[idx]).asFloat();
15982
15983                         if (f > 4.0f)
15984                                 result[idx] = tcu::Float16(-f).bits();
15985                 }
15986         }
15987
15988         if (argNo == 1)
15989         {
15990                 for (size_t idx = result.size() / 2; idx < result.size(); ++idx)
15991                 {
15992                         const float f = tcu::Float16(result[idx]).asFloat();
15993
15994                         if (f < 4.0f)
15995                                 result[idx] = tcu::Float16(-f).bits();
15996                 }
15997         }
15998
15999         return result;
16000 }
16001
16002 // Generates normalized vectors for arguments 0 and 1
16003 vector<deFloat16> getInputDataN (deUint32 seed, size_t count, size_t compCount, size_t stride, size_t argCount, size_t argNo)
16004 {
16005         DE_UNREF(compCount);
16006         DE_UNREF(argCount);
16007
16008         de::Random                      rnd             (seed << argNo);
16009         vector<deFloat16>       result;
16010
16011         if (argNo == 0 || argNo == 1)
16012         {
16013                 // The input parameters for the incident vector I and the surface normal N must already be normalized
16014                 for (size_t numIdx = 0; numIdx < count; numIdx += stride)
16015                 {
16016                         vector <float>  unnormolized;
16017                         float                   sum                             = 0;
16018
16019                         for (size_t compIdx = 0; compIdx < compCount; ++compIdx)
16020                                 unnormolized.push_back(float((rnd.getUint16() % 16) - 8));
16021
16022                         for (size_t compIdx = 0; compIdx < compCount; ++compIdx)
16023                                 sum += unnormolized[compIdx] * unnormolized[compIdx];
16024
16025                         sum = deFloatSqrt(sum);
16026                         if (sum == 0.0f)
16027                                 unnormolized[0] = sum = 1.0f;
16028
16029                         for (size_t compIdx = 0; compIdx < compCount; ++compIdx)
16030                                 result.push_back(tcu::Float16(unnormolized[compIdx] / sum).bits());
16031
16032                         for (size_t compIdx = compCount; compIdx < stride; ++compIdx)
16033                                 result.push_back(0);
16034                 }
16035         }
16036         else
16037         {
16038                 // Input parameter eta
16039                 for (deUint32 numIdx = 0; numIdx < count; ++numIdx)
16040                 {
16041                         int num = (rnd.getUint16() % 16) - 8;
16042
16043                         result.push_back(tcu::Float16(float(num)).bits());
16044                 }
16045         }
16046
16047         DE_ASSERT(result.size() == count);
16048
16049         return result;
16050 }
16051
16052 // Data generator for complex matrix functions like determinant and inverse
16053 vector<deFloat16> getInputDataC (deUint32 seed, size_t count, size_t compCount, size_t stride, size_t argCount, size_t argNo)
16054 {
16055         DE_UNREF(compCount);
16056         DE_UNREF(stride);
16057         DE_UNREF(argCount);
16058
16059         de::Random                      rnd             (seed << argNo);
16060         vector<deFloat16>       result;
16061
16062         for (deUint32 numIdx = 0; numIdx < count; ++numIdx)
16063         {
16064                 int num = (rnd.getUint16() % 16) - 8;
16065
16066                 result.push_back(tcu::Float16(float(num)).bits());
16067         }
16068
16069         DE_ASSERT(result.size() == count);
16070
16071         return result;
16072 }
16073
16074 struct Math16TestType
16075 {
16076         const char*             typePrefix;
16077         const size_t    typeComponents;
16078         const size_t    typeArrayStride;
16079         const size_t    typeStructStride;
16080 };
16081
16082 enum Math16DataTypes
16083 {
16084         NONE    = 0,
16085         SCALAR  = 1,
16086         VEC2    = 2,
16087         VEC3    = 3,
16088         VEC4    = 4,
16089         MAT2X2,
16090         MAT2X3,
16091         MAT2X4,
16092         MAT3X2,
16093         MAT3X3,
16094         MAT3X4,
16095         MAT4X2,
16096         MAT4X3,
16097         MAT4X4,
16098         MATH16_TYPE_LAST
16099 };
16100
16101 struct Math16ArgFragments
16102 {
16103         const char*     bodies;
16104         const char*     variables;
16105         const char*     decorations;
16106         const char*     funcVariables;
16107 };
16108
16109 typedef vector<deFloat16> Math16GetInputData (deUint32 seed, size_t count, size_t compCount, size_t stride, size_t argCount, size_t argNo);
16110
16111 struct Math16TestFunc
16112 {
16113         const char*                                     funcName;
16114         const char*                                     funcSuffix;
16115         size_t                                          funcArgsCount;
16116         size_t                                          typeResult;
16117         size_t                                          typeArg0;
16118         size_t                                          typeArg1;
16119         size_t                                          typeArg2;
16120         Math16GetInputData*                     getInputDataFunc;
16121         VerifyIOFunc                            verifyFunc;
16122 };
16123
16124 template<class SpecResource>
16125 void createFloat16ArithmeticFuncTest (tcu::TestContext& testCtx, tcu::TestCaseGroup& testGroup, const size_t testTypeIdx, const Math16TestFunc& testFunc)
16126 {
16127         const int                                       testSpecificSeed                        = deStringHash(testGroup.getName());
16128         const int                                       seed                                            = testCtx.getCommandLine().getBaseSeed() ^ testSpecificSeed;
16129         const size_t                            numDataPointsByAxis                     = 32;
16130         const size_t                            numDataPoints                           = numDataPointsByAxis * numDataPointsByAxis;
16131         const char*                                     componentType                           = "f16";
16132         const Math16TestType            testTypes[MATH16_TYPE_LAST]     =
16133         {
16134                 { "",           0,       0,                                              0,                                             },
16135                 { "",           1,       1 * sizeof(deFloat16),  2 * sizeof(deFloat16)  },
16136                 { "v2",         2,       2 * sizeof(deFloat16),  2 * sizeof(deFloat16)  },
16137                 { "v3",         3,       4 * sizeof(deFloat16),  4 * sizeof(deFloat16)  },
16138                 { "v4",         4,       4 * sizeof(deFloat16),  4 * sizeof(deFloat16)  },
16139                 { "m2x2",       0,       4 * sizeof(deFloat16),  4 * sizeof(deFloat16)  },
16140                 { "m2x3",       0,       8 * sizeof(deFloat16),  8 * sizeof(deFloat16)  },
16141                 { "m2x4",       0,       8 * sizeof(deFloat16),  8 * sizeof(deFloat16)  },
16142                 { "m3x2",       0,       8 * sizeof(deFloat16),  8 * sizeof(deFloat16)  },
16143                 { "m3x3",       0,      16 * sizeof(deFloat16), 16 * sizeof(deFloat16)  },
16144                 { "m3x4",       0,      16 * sizeof(deFloat16), 16 * sizeof(deFloat16)  },
16145                 { "m4x2",       0,       8 * sizeof(deFloat16),  8 * sizeof(deFloat16)  },
16146                 { "m4x3",       0,      16 * sizeof(deFloat16), 16 * sizeof(deFloat16)  },
16147                 { "m4x4",       0,      16 * sizeof(deFloat16), 16 * sizeof(deFloat16)  },
16148         };
16149
16150         DE_ASSERT(testTypeIdx == testTypes[testTypeIdx].typeComponents);
16151
16152
16153         const StringTemplate preMain
16154         (
16155                 "     %c_i32_ndp  = OpConstant %i32 ${num_data_points}\n"
16156
16157                 "        %f16     = OpTypeFloat 16\n"
16158                 "        %v2f16   = OpTypeVector %f16 2\n"
16159                 "        %v3f16   = OpTypeVector %f16 3\n"
16160                 "        %v4f16   = OpTypeVector %f16 4\n"
16161                 "        %m2x2f16 = OpTypeMatrix %v2f16 2\n"
16162                 "        %m2x3f16 = OpTypeMatrix %v3f16 2\n"
16163                 "        %m2x4f16 = OpTypeMatrix %v4f16 2\n"
16164                 "        %m3x2f16 = OpTypeMatrix %v2f16 3\n"
16165                 "        %m3x3f16 = OpTypeMatrix %v3f16 3\n"
16166                 "        %m3x4f16 = OpTypeMatrix %v4f16 3\n"
16167                 "        %m4x2f16 = OpTypeMatrix %v2f16 4\n"
16168                 "        %m4x3f16 = OpTypeMatrix %v3f16 4\n"
16169                 "        %m4x4f16 = OpTypeMatrix %v4f16 4\n"
16170
16171                 "     %up_f16     = OpTypePointer Uniform %f16    \n"
16172                 "     %up_v2f16   = OpTypePointer Uniform %v2f16  \n"
16173                 "     %up_v3f16   = OpTypePointer Uniform %v3f16  \n"
16174                 "     %up_v4f16   = OpTypePointer Uniform %v4f16  \n"
16175                 "     %up_m2x2f16 = OpTypePointer Uniform %m2x2f16\n"
16176                 "     %up_m2x3f16 = OpTypePointer Uniform %m2x3f16\n"
16177                 "     %up_m2x4f16 = OpTypePointer Uniform %m2x4f16\n"
16178                 "     %up_m3x2f16 = OpTypePointer Uniform %m3x2f16\n"
16179                 "     %up_m3x3f16 = OpTypePointer Uniform %m3x3f16\n"
16180                 "     %up_m3x4f16 = OpTypePointer Uniform %m3x4f16\n"
16181                 "     %up_m4x2f16 = OpTypePointer Uniform %m4x2f16\n"
16182                 "     %up_m4x3f16 = OpTypePointer Uniform %m4x3f16\n"
16183                 "     %up_m4x4f16 = OpTypePointer Uniform %m4x4f16\n"
16184
16185                 "     %ra_f16     = OpTypeArray %f16     %c_i32_ndp\n"
16186                 "     %ra_v2f16   = OpTypeArray %v2f16   %c_i32_ndp\n"
16187                 "     %ra_v3f16   = OpTypeArray %v3f16   %c_i32_ndp\n"
16188                 "     %ra_v4f16   = OpTypeArray %v4f16   %c_i32_ndp\n"
16189                 "     %ra_m2x2f16 = OpTypeArray %m2x2f16 %c_i32_ndp\n"
16190                 "     %ra_m2x3f16 = OpTypeArray %m2x3f16 %c_i32_ndp\n"
16191                 "     %ra_m2x4f16 = OpTypeArray %m2x4f16 %c_i32_ndp\n"
16192                 "     %ra_m3x2f16 = OpTypeArray %m3x2f16 %c_i32_ndp\n"
16193                 "     %ra_m3x3f16 = OpTypeArray %m3x3f16 %c_i32_ndp\n"
16194                 "     %ra_m3x4f16 = OpTypeArray %m3x4f16 %c_i32_ndp\n"
16195                 "     %ra_m4x2f16 = OpTypeArray %m4x2f16 %c_i32_ndp\n"
16196                 "     %ra_m4x3f16 = OpTypeArray %m4x3f16 %c_i32_ndp\n"
16197                 "     %ra_m4x4f16 = OpTypeArray %m4x4f16 %c_i32_ndp\n"
16198
16199                 "   %SSBO_f16     = OpTypeStruct %ra_f16    \n"
16200                 "   %SSBO_v2f16   = OpTypeStruct %ra_v2f16  \n"
16201                 "   %SSBO_v3f16   = OpTypeStruct %ra_v3f16  \n"
16202                 "   %SSBO_v4f16   = OpTypeStruct %ra_v4f16  \n"
16203                 "   %SSBO_m2x2f16 = OpTypeStruct %ra_m2x2f16\n"
16204                 "   %SSBO_m2x3f16 = OpTypeStruct %ra_m2x3f16\n"
16205                 "   %SSBO_m2x4f16 = OpTypeStruct %ra_m2x4f16\n"
16206                 "   %SSBO_m3x2f16 = OpTypeStruct %ra_m3x2f16\n"
16207                 "   %SSBO_m3x3f16 = OpTypeStruct %ra_m3x3f16\n"
16208                 "   %SSBO_m3x4f16 = OpTypeStruct %ra_m3x4f16\n"
16209                 "   %SSBO_m4x2f16 = OpTypeStruct %ra_m4x2f16\n"
16210                 "   %SSBO_m4x3f16 = OpTypeStruct %ra_m4x3f16\n"
16211                 "   %SSBO_m4x4f16 = OpTypeStruct %ra_m4x4f16\n"
16212
16213                 "%up_SSBO_f16     = OpTypePointer Uniform %SSBO_f16    \n"
16214                 "%up_SSBO_v2f16   = OpTypePointer Uniform %SSBO_v2f16  \n"
16215                 "%up_SSBO_v3f16   = OpTypePointer Uniform %SSBO_v3f16  \n"
16216                 "%up_SSBO_v4f16   = OpTypePointer Uniform %SSBO_v4f16  \n"
16217                 "%up_SSBO_m2x2f16 = OpTypePointer Uniform %SSBO_m2x2f16\n"
16218                 "%up_SSBO_m2x3f16 = OpTypePointer Uniform %SSBO_m2x3f16\n"
16219                 "%up_SSBO_m2x4f16 = OpTypePointer Uniform %SSBO_m2x4f16\n"
16220                 "%up_SSBO_m3x2f16 = OpTypePointer Uniform %SSBO_m3x2f16\n"
16221                 "%up_SSBO_m3x3f16 = OpTypePointer Uniform %SSBO_m3x3f16\n"
16222                 "%up_SSBO_m3x4f16 = OpTypePointer Uniform %SSBO_m3x4f16\n"
16223                 "%up_SSBO_m4x2f16 = OpTypePointer Uniform %SSBO_m4x2f16\n"
16224                 "%up_SSBO_m4x3f16 = OpTypePointer Uniform %SSBO_m4x3f16\n"
16225                 "%up_SSBO_m4x4f16 = OpTypePointer Uniform %SSBO_m4x4f16\n"
16226
16227                 "       %fp_v2i32 = OpTypePointer Function %v2i32\n"
16228                 "       %fp_v3i32 = OpTypePointer Function %v3i32\n"
16229                 "       %fp_v4i32 = OpTypePointer Function %v4i32\n"
16230                 "${arg_vars}"
16231         );
16232
16233         const StringTemplate decoration
16234         (
16235                 "OpDecorate %ra_f16     ArrayStride 2 \n"
16236                 "OpDecorate %ra_v2f16   ArrayStride 4 \n"
16237                 "OpDecorate %ra_v3f16   ArrayStride 8 \n"
16238                 "OpDecorate %ra_v4f16   ArrayStride 8 \n"
16239                 "OpDecorate %ra_m2x2f16 ArrayStride 8 \n"
16240                 "OpDecorate %ra_m2x3f16 ArrayStride 16\n"
16241                 "OpDecorate %ra_m2x4f16 ArrayStride 16\n"
16242                 "OpDecorate %ra_m3x2f16 ArrayStride 16\n"
16243                 "OpDecorate %ra_m3x3f16 ArrayStride 32\n"
16244                 "OpDecorate %ra_m3x4f16 ArrayStride 32\n"
16245                 "OpDecorate %ra_m4x2f16 ArrayStride 16\n"
16246                 "OpDecorate %ra_m4x3f16 ArrayStride 32\n"
16247                 "OpDecorate %ra_m4x4f16 ArrayStride 32\n"
16248
16249                 "OpMemberDecorate %SSBO_f16     0 Offset 0\n"
16250                 "OpMemberDecorate %SSBO_v2f16   0 Offset 0\n"
16251                 "OpMemberDecorate %SSBO_v3f16   0 Offset 0\n"
16252                 "OpMemberDecorate %SSBO_v4f16   0 Offset 0\n"
16253                 "OpMemberDecorate %SSBO_m2x2f16 0 Offset 0\n"
16254                 "OpMemberDecorate %SSBO_m2x3f16 0 Offset 0\n"
16255                 "OpMemberDecorate %SSBO_m2x4f16 0 Offset 0\n"
16256                 "OpMemberDecorate %SSBO_m3x2f16 0 Offset 0\n"
16257                 "OpMemberDecorate %SSBO_m3x3f16 0 Offset 0\n"
16258                 "OpMemberDecorate %SSBO_m3x4f16 0 Offset 0\n"
16259                 "OpMemberDecorate %SSBO_m4x2f16 0 Offset 0\n"
16260                 "OpMemberDecorate %SSBO_m4x3f16 0 Offset 0\n"
16261                 "OpMemberDecorate %SSBO_m4x4f16 0 Offset 0\n"
16262
16263                 "OpDecorate %SSBO_f16     BufferBlock\n"
16264                 "OpDecorate %SSBO_v2f16   BufferBlock\n"
16265                 "OpDecorate %SSBO_v3f16   BufferBlock\n"
16266                 "OpDecorate %SSBO_v4f16   BufferBlock\n"
16267                 "OpDecorate %SSBO_m2x2f16 BufferBlock\n"
16268                 "OpDecorate %SSBO_m2x3f16 BufferBlock\n"
16269                 "OpDecorate %SSBO_m2x4f16 BufferBlock\n"
16270                 "OpDecorate %SSBO_m3x2f16 BufferBlock\n"
16271                 "OpDecorate %SSBO_m3x3f16 BufferBlock\n"
16272                 "OpDecorate %SSBO_m3x4f16 BufferBlock\n"
16273                 "OpDecorate %SSBO_m4x2f16 BufferBlock\n"
16274                 "OpDecorate %SSBO_m4x3f16 BufferBlock\n"
16275                 "OpDecorate %SSBO_m4x4f16 BufferBlock\n"
16276
16277                 "OpMemberDecorate %SSBO_m2x2f16 0 ColMajor\n"
16278                 "OpMemberDecorate %SSBO_m2x3f16 0 ColMajor\n"
16279                 "OpMemberDecorate %SSBO_m2x4f16 0 ColMajor\n"
16280                 "OpMemberDecorate %SSBO_m3x2f16 0 ColMajor\n"
16281                 "OpMemberDecorate %SSBO_m3x3f16 0 ColMajor\n"
16282                 "OpMemberDecorate %SSBO_m3x4f16 0 ColMajor\n"
16283                 "OpMemberDecorate %SSBO_m4x2f16 0 ColMajor\n"
16284                 "OpMemberDecorate %SSBO_m4x3f16 0 ColMajor\n"
16285                 "OpMemberDecorate %SSBO_m4x4f16 0 ColMajor\n"
16286
16287                 "OpMemberDecorate %SSBO_m2x2f16 0 MatrixStride 4\n"
16288                 "OpMemberDecorate %SSBO_m2x3f16 0 MatrixStride 8\n"
16289                 "OpMemberDecorate %SSBO_m2x4f16 0 MatrixStride 8\n"
16290                 "OpMemberDecorate %SSBO_m3x2f16 0 MatrixStride 4\n"
16291                 "OpMemberDecorate %SSBO_m3x3f16 0 MatrixStride 8\n"
16292                 "OpMemberDecorate %SSBO_m3x4f16 0 MatrixStride 8\n"
16293                 "OpMemberDecorate %SSBO_m4x2f16 0 MatrixStride 4\n"
16294                 "OpMemberDecorate %SSBO_m4x3f16 0 MatrixStride 8\n"
16295                 "OpMemberDecorate %SSBO_m4x4f16 0 MatrixStride 8\n"
16296
16297                 "${arg_decorations}"
16298         );
16299
16300         const StringTemplate testFun
16301         (
16302                 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
16303                 "    %param = OpFunctionParameter %v4f32\n"
16304                 "    %entry = OpLabel\n"
16305
16306                 "        %i = OpVariable %fp_i32 Function\n"
16307                 "${arg_infunc_vars}"
16308                 "             OpStore %i %c_i32_0\n"
16309                 "             OpBranch %loop\n"
16310
16311                 "     %loop = OpLabel\n"
16312                 "    %i_cmp = OpLoad %i32 %i\n"
16313                 "       %lt = OpSLessThan %bool %i_cmp %c_i32_ndp\n"
16314                 "             OpLoopMerge %merge %next None\n"
16315                 "             OpBranchConditional %lt %write %merge\n"
16316
16317                 "    %write = OpLabel\n"
16318                 "      %ndx = OpLoad %i32 %i\n"
16319
16320                 "${arg_func_call}"
16321
16322                 "             OpBranch %next\n"
16323
16324                 "     %next = OpLabel\n"
16325                 "    %i_cur = OpLoad %i32 %i\n"
16326                 "    %i_new = OpIAdd %i32 %i_cur %c_i32_1\n"
16327                 "             OpStore %i %i_new\n"
16328                 "             OpBranch %loop\n"
16329
16330                 "    %merge = OpLabel\n"
16331                 "             OpReturnValue %param\n"
16332                 "             OpFunctionEnd\n"
16333         );
16334
16335         const Math16ArgFragments        argFragment1    =
16336         {
16337                 "     %src0 = OpAccessChain %up_${t0} %ssbo_src0 %c_i32_0 %ndx\n"
16338                 " %val_src0 = OpLoad %${t0} %src0\n"
16339                 "  %val_dst = ${op} %${tr} ${ext_inst} %val_src0\n"
16340                 "      %dst = OpAccessChain %up_${tr} %ssbo_dst %c_i32_0 %ndx\n"
16341                 "             OpStore %dst %val_dst\n",
16342                 "",
16343                 "",
16344                 "",
16345         };
16346
16347         const Math16ArgFragments        argFragment2    =
16348         {
16349                 "     %src0 = OpAccessChain %up_${t0} %ssbo_src0 %c_i32_0 %ndx\n"
16350                 " %val_src0 = OpLoad %${t0} %src0\n"
16351                 "     %src1 = OpAccessChain %up_${t1} %ssbo_src1 %c_i32_0 %ndx\n"
16352                 " %val_src1 = OpLoad %${t1} %src1\n"
16353                 "  %val_dst = ${op} %${tr} ${ext_inst} %val_src0 %val_src1\n"
16354                 "      %dst = OpAccessChain %up_${tr} %ssbo_dst %c_i32_0 %ndx\n"
16355                 "             OpStore %dst %val_dst\n",
16356                 "",
16357                 "",
16358                 "",
16359         };
16360
16361         const Math16ArgFragments        argFragment3    =
16362         {
16363                 "     %src0 = OpAccessChain %up_${t0} %ssbo_src0 %c_i32_0 %ndx\n"
16364                 " %val_src0 = OpLoad %${t0} %src0\n"
16365                 "     %src1 = OpAccessChain %up_${t1} %ssbo_src1 %c_i32_0 %ndx\n"
16366                 " %val_src1 = OpLoad %${t1} %src1\n"
16367                 "     %src2 = OpAccessChain %up_${t2} %ssbo_src2 %c_i32_0 %ndx\n"
16368                 " %val_src2 = OpLoad %${t2} %src2\n"
16369                 "  %val_dst = ${op} %${tr} ${ext_inst} %val_src0 %val_src1 %val_src2\n"
16370                 "      %dst = OpAccessChain %up_${tr} %ssbo_dst %c_i32_0 %ndx\n"
16371                 "             OpStore %dst %val_dst\n",
16372                 "",
16373                 "",
16374                 "",
16375         };
16376
16377         const Math16ArgFragments        argFragmentLdExp        =
16378         {
16379                 "     %src0 = OpAccessChain %up_${t0} %ssbo_src0 %c_i32_0 %ndx\n"
16380                 " %val_src0 = OpLoad %${t0} %src0\n"
16381                 "     %src1 = OpAccessChain %up_${t1} %ssbo_src1 %c_i32_0 %ndx\n"
16382                 " %val_src1 = OpLoad %${t1} %src1\n"
16383                 "%val_src1i = OpConvertFToS %${dr}i32 %val_src1\n"
16384                 "  %val_dst = ${op} %${tr} ${ext_inst} %val_src0 %val_src1i\n"
16385                 "      %dst = OpAccessChain %up_${tr} %ssbo_dst %c_i32_0 %ndx\n"
16386                 "             OpStore %dst %val_dst\n",
16387
16388                 "",
16389
16390                 "",
16391
16392                 "",
16393         };
16394
16395         const Math16ArgFragments        argFragmentModfFrac     =
16396         {
16397                 "     %src0 = OpAccessChain %up_${t0} %ssbo_src0 %c_i32_0 %ndx\n"
16398                 " %val_src0 = OpLoad %${t0} %src0\n"
16399                 "  %val_dst = ${op} %${tr} ${ext_inst} %val_src0 %tmp\n"
16400                 "      %dst = OpAccessChain %up_${tr} %ssbo_dst %c_i32_0 %ndx\n"
16401                 "             OpStore %dst %val_dst\n",
16402
16403                 "   %fp_tmp = OpTypePointer Function %${tr}\n",
16404
16405                 "",
16406
16407                 "      %tmp = OpVariable %fp_tmp Function\n",
16408         };
16409
16410         const Math16ArgFragments        argFragmentModfInt      =
16411         {
16412                 "     %src0 = OpAccessChain %up_${t0} %ssbo_src0 %c_i32_0 %ndx\n"
16413                 " %val_src0 = OpLoad %${t0} %src0\n"
16414                 "%val_dummy = ${op} %${tr} ${ext_inst} %val_src0 %tmp\n"
16415                 "     %tmp0 = OpAccessChain %fp_tmp %tmp\n"
16416                 "  %val_dst = OpLoad %${tr} %tmp0\n"
16417                 "      %dst = OpAccessChain %up_${tr} %ssbo_dst %c_i32_0 %ndx\n"
16418                 "             OpStore %dst %val_dst\n",
16419
16420                 "   %fp_tmp = OpTypePointer Function %${tr}\n",
16421
16422                 "",
16423
16424                 "      %tmp = OpVariable %fp_tmp Function\n",
16425         };
16426
16427         const Math16ArgFragments        argFragmentModfStruct   =
16428         {
16429                 "     %src0 = OpAccessChain %up_${t0} %ssbo_src0 %c_i32_0 %ndx\n"
16430                 " %val_src0 = OpLoad %${t0} %src0\n"
16431                 "  %val_tmp = ${op} %st_tmp ${ext_inst} %val_src0\n"
16432                 "%tmp_ptr_s = OpAccessChain %fp_tmp %tmp\n"
16433                 "             OpStore %tmp_ptr_s %val_tmp\n"
16434                 "%tmp_ptr_l = OpAccessChain %fp_${tr} %tmp %c_${struct_member}\n"
16435                 "  %val_dst = OpLoad %${tr} %tmp_ptr_l\n"
16436                 "      %dst = OpAccessChain %up_${tr} %ssbo_dst %c_i32_0 %ndx\n"
16437                 "             OpStore %dst %val_dst\n",
16438
16439                 "  %fp_${tr} = OpTypePointer Function %${tr}\n"
16440                 "   %st_tmp = OpTypeStruct %${tr} %${tr}\n"
16441                 "   %fp_tmp = OpTypePointer Function %st_tmp\n"
16442                 "   %c_frac = OpConstant %i32 0\n"
16443                 "    %c_int = OpConstant %i32 1\n",
16444
16445                 "OpMemberDecorate %st_tmp 0 Offset 0\n"
16446                 "OpMemberDecorate %st_tmp 1 Offset ${struct_stride}\n",
16447
16448                 "      %tmp = OpVariable %fp_tmp Function\n",
16449         };
16450
16451         const Math16ArgFragments        argFragmentFrexpStructS =
16452         {
16453                 "     %src0 = OpAccessChain %up_${t0} %ssbo_src0 %c_i32_0 %ndx\n"
16454                 " %val_src0 = OpLoad %${t0} %src0\n"
16455                 "  %val_tmp = ${op} %st_tmp ${ext_inst} %val_src0\n"
16456                 "%tmp_ptr_s = OpAccessChain %fp_tmp %tmp\n"
16457                 "             OpStore %tmp_ptr_s %val_tmp\n"
16458                 "%tmp_ptr_l = OpAccessChain %fp_${tr} %tmp %c_i32_0\n"
16459                 "  %val_dst = OpLoad %${tr} %tmp_ptr_l\n"
16460                 "      %dst = OpAccessChain %up_${tr} %ssbo_dst %c_i32_0 %ndx\n"
16461                 "             OpStore %dst %val_dst\n",
16462
16463                 "  %fp_${tr} = OpTypePointer Function %${tr}\n"
16464                 "   %st_tmp = OpTypeStruct %${tr} %${dr}i32\n"
16465                 "   %fp_tmp = OpTypePointer Function %st_tmp\n",
16466
16467                 "OpMemberDecorate %st_tmp 0 Offset 0\n"
16468                 "OpMemberDecorate %st_tmp 1 Offset ${struct_stride}\n",
16469
16470                 "      %tmp = OpVariable %fp_tmp Function\n",
16471         };
16472
16473         const Math16ArgFragments        argFragmentFrexpStructE =
16474         {
16475                 "     %src0 = OpAccessChain %up_${t0} %ssbo_src0 %c_i32_0 %ndx\n"
16476                 " %val_src0 = OpLoad %${t0} %src0\n"
16477                 "  %val_tmp = ${op} %st_tmp ${ext_inst} %val_src0\n"
16478                 "%tmp_ptr_s = OpAccessChain %fp_tmp %tmp\n"
16479                 "             OpStore %tmp_ptr_s %val_tmp\n"
16480                 "%tmp_ptr_l = OpAccessChain %fp_${dr}i32 %tmp %c_i32_1\n"
16481                 "%val_dst_i = OpLoad %${dr}i32 %tmp_ptr_l\n"
16482                 "  %val_dst = OpConvertSToF %${tr} %val_dst_i\n"
16483                 "      %dst = OpAccessChain %up_${tr} %ssbo_dst %c_i32_0 %ndx\n"
16484                 "             OpStore %dst %val_dst\n",
16485
16486                 "   %st_tmp = OpTypeStruct %${tr} %${dr}i32\n"
16487                 "   %fp_tmp = OpTypePointer Function %st_tmp\n",
16488
16489                 "OpMemberDecorate %st_tmp 0 Offset 0\n"
16490                 "OpMemberDecorate %st_tmp 1 Offset ${struct_stride}\n",
16491
16492                 "      %tmp = OpVariable %fp_tmp Function\n",
16493         };
16494
16495         const Math16ArgFragments        argFragmentFrexpS               =
16496         {
16497                 "     %src0 = OpAccessChain %up_${t0} %ssbo_src0 %c_i32_0 %ndx\n"
16498                 " %val_src0 = OpLoad %${t0} %src0\n"
16499                 "  %out_exp = OpAccessChain %fp_${dr}i32 %tmp\n"
16500                 "  %val_dst = ${op} %${tr} ${ext_inst} %val_src0 %out_exp\n"
16501                 "      %dst = OpAccessChain %up_${tr} %ssbo_dst %c_i32_0 %ndx\n"
16502                 "             OpStore %dst %val_dst\n",
16503
16504                 "",
16505
16506                 "",
16507
16508                 "      %tmp = OpVariable %fp_${dr}i32 Function\n",
16509         };
16510
16511         const Math16ArgFragments        argFragmentFrexpE               =
16512         {
16513                 "     %src0 = OpAccessChain %up_${t0} %ssbo_src0 %c_i32_0 %ndx\n"
16514                 " %val_src0 = OpLoad %${t0} %src0\n"
16515                 "  %out_exp = OpAccessChain %fp_${dr}i32 %tmp\n"
16516                 "%val_dummy = ${op} %${tr} ${ext_inst} %val_src0 %out_exp\n"
16517                 "%val_dst_i = OpLoad %${dr}i32 %out_exp\n"
16518                 "  %val_dst = OpConvertSToF %${tr} %val_dst_i\n"
16519                 "      %dst = OpAccessChain %up_${tr} %ssbo_dst %c_i32_0 %ndx\n"
16520                 "             OpStore %dst %val_dst\n",
16521
16522                 "",
16523
16524                 "",
16525
16526                 "      %tmp = OpVariable %fp_${dr}i32 Function\n",
16527         };
16528
16529         const Math16TestType&           testType                                = testTypes[testTypeIdx];
16530         const string                            funcNameString                  = string(testFunc.funcName) + string(testFunc.funcSuffix);
16531         const string                            testName                                = de::toLower(funcNameString);
16532         const Math16ArgFragments*       argFragments                    = DE_NULL;
16533         const size_t                            typeStructStride                = testType.typeStructStride;
16534         const bool                                      extInst                                 = !(testFunc.funcName[0] == 'O' && testFunc.funcName[1] == 'p');
16535         const size_t                            numFloatsPerArg0Type    = testTypes[testFunc.typeArg0].typeArrayStride / sizeof(deFloat16);
16536         const size_t                            iterations                              = numDataPoints / numFloatsPerArg0Type;
16537         const size_t                            numFloatsPerResultType  = testTypes[testFunc.typeResult].typeArrayStride / sizeof(deFloat16);
16538         const vector<deFloat16>         float16DummyOutput              (iterations * numFloatsPerResultType, 0);
16539         VulkanFeatures                          features;
16540         SpecResource                            specResource;
16541         map<string, string>                     specs;
16542         map<string, string>                     fragments;
16543         vector<string>                          extensions;
16544         string                                          funcCall;
16545         string                                          funcVariables;
16546         string                                          variables;
16547         string                                          declarations;
16548         string                                          decorations;
16549
16550         switch (testFunc.funcArgsCount)
16551         {
16552                 case 1:
16553                 {
16554                         argFragments = &argFragment1;
16555
16556                         if (funcNameString == "ModfFrac")               argFragments = &argFragmentModfFrac;
16557                         if (funcNameString == "ModfInt")                argFragments = &argFragmentModfInt;
16558                         if (funcNameString == "ModfStructFrac") argFragments = &argFragmentModfStruct;
16559                         if (funcNameString == "ModfStructInt")  argFragments = &argFragmentModfStruct;
16560                         if (funcNameString == "FrexpS")                 argFragments = &argFragmentFrexpS;
16561                         if (funcNameString == "FrexpE")                 argFragments = &argFragmentFrexpE;
16562                         if (funcNameString == "FrexpStructS")   argFragments = &argFragmentFrexpStructS;
16563                         if (funcNameString == "FrexpStructE")   argFragments = &argFragmentFrexpStructE;
16564
16565                         break;
16566                 }
16567                 case 2:
16568                 {
16569                         argFragments = &argFragment2;
16570
16571                         if (funcNameString == "Ldexp")                  argFragments = &argFragmentLdExp;
16572
16573                         break;
16574                 }
16575                 case 3:
16576                 {
16577                         argFragments = &argFragment3;
16578
16579                         break;
16580                 }
16581                 default:
16582                 {
16583                         TCU_THROW(InternalError, "Invalid number of arguments");
16584                 }
16585         }
16586
16587         if (testFunc.funcArgsCount == 1)
16588         {
16589                 variables +=
16590                         " %ssbo_src0 = OpVariable %up_SSBO_${t0} Uniform\n"
16591                         "  %ssbo_dst = OpVariable %up_SSBO_${tr} Uniform\n";
16592
16593                 decorations +=
16594                         "OpDecorate %ssbo_src0 DescriptorSet 0\n"
16595                         "OpDecorate %ssbo_src0 Binding 0\n"
16596                         "OpDecorate %ssbo_dst DescriptorSet 0\n"
16597                         "OpDecorate %ssbo_dst Binding 1\n";
16598         }
16599         else if (testFunc.funcArgsCount == 2)
16600         {
16601                 variables +=
16602                         " %ssbo_src0 = OpVariable %up_SSBO_${t0} Uniform\n"
16603                         " %ssbo_src1 = OpVariable %up_SSBO_${t1} Uniform\n"
16604                         "  %ssbo_dst = OpVariable %up_SSBO_${tr} Uniform\n";
16605
16606                 decorations +=
16607                         "OpDecorate %ssbo_src0 DescriptorSet 0\n"
16608                         "OpDecorate %ssbo_src0 Binding 0\n"
16609                         "OpDecorate %ssbo_src1 DescriptorSet 0\n"
16610                         "OpDecorate %ssbo_src1 Binding 1\n"
16611                         "OpDecorate %ssbo_dst DescriptorSet 0\n"
16612                         "OpDecorate %ssbo_dst Binding 2\n";
16613         }
16614         else if (testFunc.funcArgsCount == 3)
16615         {
16616                 variables +=
16617                         " %ssbo_src0 = OpVariable %up_SSBO_${t0} Uniform\n"
16618                         " %ssbo_src1 = OpVariable %up_SSBO_${t1} Uniform\n"
16619                         " %ssbo_src2 = OpVariable %up_SSBO_${t2} Uniform\n"
16620                         "  %ssbo_dst = OpVariable %up_SSBO_${tr} Uniform\n";
16621
16622                 decorations +=
16623                         "OpDecorate %ssbo_src0 DescriptorSet 0\n"
16624                         "OpDecorate %ssbo_src0 Binding 0\n"
16625                         "OpDecorate %ssbo_src1 DescriptorSet 0\n"
16626                         "OpDecorate %ssbo_src1 Binding 1\n"
16627                         "OpDecorate %ssbo_src2 DescriptorSet 0\n"
16628                         "OpDecorate %ssbo_src2 Binding 2\n"
16629                         "OpDecorate %ssbo_dst DescriptorSet 0\n"
16630                         "OpDecorate %ssbo_dst Binding 3\n";
16631         }
16632         else
16633         {
16634                 TCU_THROW(InternalError, "Invalid number of function arguments");
16635         }
16636
16637         variables       += argFragments->variables;
16638         decorations     += argFragments->decorations;
16639
16640         specs["dr"]                                     = testTypes[testFunc.typeResult].typePrefix;
16641         specs["d0"]                                     = testTypes[testFunc.typeArg0].typePrefix;
16642         specs["d1"]                                     = testTypes[testFunc.typeArg1].typePrefix;
16643         specs["d2"]                                     = testTypes[testFunc.typeArg2].typePrefix;
16644         specs["tr"]                                     = string(testTypes[testFunc.typeResult].typePrefix) + componentType;
16645         specs["t0"]                                     = string(testTypes[testFunc.typeArg0].typePrefix) + componentType;
16646         specs["t1"]                                     = string(testTypes[testFunc.typeArg1].typePrefix) + componentType;
16647         specs["t2"]                                     = string(testTypes[testFunc.typeArg2].typePrefix) + componentType;
16648         specs["struct_stride"]          = de::toString(typeStructStride);
16649         specs["op"]                                     = extInst ? "OpExtInst" : testFunc.funcName;
16650         specs["ext_inst"]                       = extInst ? string("%ext_import ") + testFunc.funcName : "";
16651         specs["struct_member"]          = de::toLower(testFunc.funcSuffix);
16652
16653         variables                                       = StringTemplate(variables).specialize(specs);
16654         decorations                                     = StringTemplate(decorations).specialize(specs);
16655         funcVariables                           = StringTemplate(argFragments->funcVariables).specialize(specs);
16656         funcCall                                        = StringTemplate(argFragments->bodies).specialize(specs);
16657
16658         specs["num_data_points"]        = de::toString(iterations);
16659         specs["arg_vars"]                       = variables;
16660         specs["arg_decorations"]        = decorations;
16661         specs["arg_infunc_vars"]        = funcVariables;
16662         specs["arg_func_call"]          = funcCall;
16663
16664         fragments["extension"]          = "OpExtension \"SPV_KHR_16bit_storage\"\n%ext_import = OpExtInstImport \"GLSL.std.450\"";
16665         fragments["capability"]         = "OpCapability Matrix\nOpCapability StorageUniformBufferBlock16";
16666         fragments["decoration"]         = decoration.specialize(specs);
16667         fragments["pre_main"]           = preMain.specialize(specs);
16668         fragments["testfun"]            = testFun.specialize(specs);
16669
16670         for (size_t inputArgNdx = 0; inputArgNdx < testFunc.funcArgsCount; ++inputArgNdx)
16671         {
16672                 const size_t                    numFloatsPerItem        = (inputArgNdx == 0) ? testTypes[testFunc.typeArg0].typeArrayStride / sizeof(deFloat16)
16673                                                                                                         : (inputArgNdx == 1) ? testTypes[testFunc.typeArg1].typeArrayStride / sizeof(deFloat16)
16674                                                                                                         : (inputArgNdx == 2) ? testTypes[testFunc.typeArg2].typeArrayStride / sizeof(deFloat16)
16675                                                                                                         : -1;
16676                 const vector<deFloat16> inputData                       = testFunc.getInputDataFunc(seed, numFloatsPerItem * iterations, testTypeIdx, numFloatsPerItem, testFunc.funcArgsCount, inputArgNdx);
16677
16678                 specResource.inputs.push_back(Resource(BufferSp(new Float16Buffer(inputData)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
16679         }
16680
16681         specResource.outputs.push_back(Resource(BufferSp(new Float16Buffer(float16DummyOutput)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
16682         specResource.verifyIO = testFunc.verifyFunc;
16683
16684         extensions.push_back("VK_KHR_16bit_storage");
16685         extensions.push_back("VK_KHR_shader_float16_int8");
16686
16687         features.ext16BitStorage        = EXT16BITSTORAGEFEATURES_UNIFORM_BUFFER_BLOCK;
16688         features.extFloat16Int8         = EXTFLOAT16INT8FEATURES_FLOAT16;
16689
16690         finalizeTestsCreation(specResource, fragments, testCtx, testGroup, testName, features, extensions, IVec3(1, 1, 1));
16691 }
16692
16693 template<size_t C, class SpecResource>
16694 tcu::TestCaseGroup* createFloat16ArithmeticSet (tcu::TestContext& testCtx)
16695 {
16696         DE_STATIC_ASSERT(C >= 1 && C <= 4);
16697
16698         const std::string                               testGroupName   (string("arithmetic_") + de::toString(C));
16699         de::MovePtr<tcu::TestCaseGroup> testGroup               (new tcu::TestCaseGroup(testCtx, testGroupName.c_str(), "Float 16 arithmetic and related tests"));
16700         const Math16TestFunc                    testFuncs[]             =
16701         {
16702                 {       "OpFNegate",                    "",                     1,      C,              C,              0,              0, &getInputData,       compareFP16ArithmeticFunc<  C,  C,  0,  0, fp16OpFNegate>                                       },
16703                 {       "Round",                                "",                     1,      C,              C,              0,              0, &getInputData,       compareFP16ArithmeticFunc<  C,  C,  0,  0, fp16Round>                                           },
16704                 {       "RoundEven",                    "",                     1,      C,              C,              0,              0, &getInputData,       compareFP16ArithmeticFunc<  C,  C,  0,  0, fp16RoundEven>                                       },
16705                 {       "Trunc",                                "",                     1,      C,              C,              0,              0, &getInputData,       compareFP16ArithmeticFunc<  C,  C,  0,  0, fp16Trunc>                                           },
16706                 {       "FAbs",                                 "",                     1,      C,              C,              0,              0, &getInputData,       compareFP16ArithmeticFunc<  C,  C,  0,  0, fp16FAbs>                                            },
16707                 {       "FSign",                                "",                     1,      C,              C,              0,              0, &getInputData,       compareFP16ArithmeticFunc<  C,  C,  0,  0, fp16FSign>                                           },
16708                 {       "Floor",                                "",                     1,      C,              C,              0,              0, &getInputData,       compareFP16ArithmeticFunc<  C,  C,  0,  0, fp16Floor>                                           },
16709                 {       "Ceil",                                 "",                     1,      C,              C,              0,              0, &getInputData,       compareFP16ArithmeticFunc<  C,  C,  0,  0, fp16Ceil>                                            },
16710                 {       "Fract",                                "",                     1,      C,              C,              0,              0, &getInputData,       compareFP16ArithmeticFunc<  C,  C,  0,  0, fp16Fract>                                           },
16711                 {       "Radians",                              "",                     1,      C,              C,              0,              0, &getInputData,       compareFP16ArithmeticFunc<  C,  C,  0,  0, fp16Radians>                                         },
16712                 {       "Degrees",                              "",                     1,      C,              C,              0,              0, &getInputData,       compareFP16ArithmeticFunc<  C,  C,  0,  0, fp16Degrees>                                         },
16713                 {       "Sin",                                  "",                     1,      C,              C,              0,              0, &getInputDataPI,     compareFP16ArithmeticFunc<  C,  C,  0,  0, fp16Sin>                                                     },
16714                 {       "Cos",                                  "",                     1,      C,              C,              0,              0, &getInputDataPI,     compareFP16ArithmeticFunc<  C,  C,  0,  0, fp16Cos>                                                     },
16715                 {       "Tan",                                  "",                     1,      C,              C,              0,              0, &getInputDataPI,     compareFP16ArithmeticFunc<  C,  C,  0,  0, fp16Tan>                                                     },
16716                 {       "Asin",                                 "",                     1,      C,              C,              0,              0, &getInputDataA,      compareFP16ArithmeticFunc<  C,  C,  0,  0, fp16Asin>                                            },
16717                 {       "Acos",                                 "",                     1,      C,              C,              0,              0, &getInputDataA,      compareFP16ArithmeticFunc<  C,  C,  0,  0, fp16Acos>                                            },
16718                 {       "Atan",                                 "",                     1,      C,              C,              0,              0, &getInputData,       compareFP16ArithmeticFunc<  C,  C,  0,  0, fp16Atan>                                            },
16719                 {       "Sinh",                                 "",                     1,      C,              C,              0,              0, &getInputData,       compareFP16ArithmeticFunc<  C,  C,  0,  0, fp16Sinh>                                            },
16720                 {       "Cosh",                                 "",                     1,      C,              C,              0,              0, &getInputData,       compareFP16ArithmeticFunc<  C,  C,  0,  0, fp16Cosh>                                            },
16721                 {       "Tanh",                                 "",                     1,      C,              C,              0,              0, &getInputData,       compareFP16ArithmeticFunc<  C,  C,  0,  0, fp16Tanh>                                            },
16722                 {       "Asinh",                                "",                     1,      C,              C,              0,              0, &getInputData,       compareFP16ArithmeticFunc<  C,  C,  0,  0, fp16Asinh>                                           },
16723                 {       "Acosh",                                "",                     1,      C,              C,              0,              0, &getInputDataAC,     compareFP16ArithmeticFunc<  C,  C,  0,  0, fp16Acosh>                                           },
16724                 {       "Atanh",                                "",                     1,      C,              C,              0,              0, &getInputDataA,      compareFP16ArithmeticFunc<  C,  C,  0,  0, fp16Atanh>                                           },
16725                 {       "Exp",                                  "",                     1,      C,              C,              0,              0, &getInputData,       compareFP16ArithmeticFunc<  C,  C,  0,  0, fp16Exp>                                                     },
16726                 {       "Log",                                  "",                     1,      C,              C,              0,              0, &getInputDataP,      compareFP16ArithmeticFunc<  C,  C,  0,  0, fp16Log>                                                     },
16727                 {       "Exp2",                                 "",                     1,      C,              C,              0,              0, &getInputData,       compareFP16ArithmeticFunc<  C,  C,  0,  0, fp16Exp2>                                            },
16728                 {       "Log2",                                 "",                     1,      C,              C,              0,              0, &getInputDataP,      compareFP16ArithmeticFunc<  C,  C,  0,  0, fp16Log2>                                            },
16729                 {       "Sqrt",                                 "",                     1,      C,              C,              0,              0, &getInputDataP,      compareFP16ArithmeticFunc<  C,  C,  0,  0, fp16Sqrt>                                            },
16730                 {       "InverseSqrt",                  "",                     1,      C,              C,              0,              0, &getInputDataP,      compareFP16ArithmeticFunc<  C,  C,  0,  0, fp16InverseSqrt>                                     },
16731                 {       "Modf",                                 "Frac",         1,      C,              C,              0,              0, &getInputData,       compareFP16ArithmeticFunc<  C,  C,  0,  0, fp16ModfFrac>                                        },
16732                 {       "Modf",                                 "Int",          1,      C,              C,              0,              0, &getInputData,       compareFP16ArithmeticFunc<  C,  C,  0,  0, fp16ModfInt>                                         },
16733                 {       "ModfStruct",                   "Frac",         1,      C,              C,              0,              0, &getInputData,       compareFP16ArithmeticFunc<  C,  C,  0,  0, fp16ModfFrac>                                        },
16734                 {       "ModfStruct",                   "Int",          1,      C,              C,              0,              0, &getInputData,       compareFP16ArithmeticFunc<  C,  C,  0,  0, fp16ModfInt>                                         },
16735                 {       "Frexp",                                "S",            1,      C,              C,              0,              0, &getInputData,       compareFP16ArithmeticFunc<  C,  C,  0,  0, fp16FrexpS>                                          },
16736                 {       "Frexp",                                "E",            1,      C,              C,              0,              0, &getInputData,       compareFP16ArithmeticFunc<  C,  C,  0,  0, fp16FrexpE>                                          },
16737                 {       "FrexpStruct",                  "S",            1,      C,              C,              0,              0, &getInputData,       compareFP16ArithmeticFunc<  C,  C,  0,  0, fp16FrexpS>                                          },
16738                 {       "FrexpStruct",                  "E",            1,      C,              C,              0,              0, &getInputData,       compareFP16ArithmeticFunc<  C,  C,  0,  0, fp16FrexpE>                                          },
16739                 {       "OpFAdd",                               "",                     2,      C,              C,              C,              0, &getInputData,       compareFP16ArithmeticFunc<  C,  C,  C,  0, fp16OpFAdd>                                          },
16740                 {       "OpFSub",                               "",                     2,      C,              C,              C,              0, &getInputData,       compareFP16ArithmeticFunc<  C,  C,  C,  0, fp16OpFSub>                                          },
16741                 {       "OpFMul",                               "",                     2,      C,              C,              C,              0, &getInputData,       compareFP16ArithmeticFunc<  C,  C,  C,  0, fp16OpFMul>                                          },
16742                 {       "OpFDiv",                               "",                     2,      C,              C,              C,              0, &getInputData,       compareFP16ArithmeticFunc<  C,  C,  C,  0, fp16OpFDiv>                                          },
16743                 {       "Atan2",                                "",                     2,      C,              C,              C,              0, &getInputData,       compareFP16ArithmeticFunc<  C,  C,  C,  0, fp16Atan2>                                           },
16744                 {       "Pow",                                  "",                     2,      C,              C,              C,              0, &getInputDataP,      compareFP16ArithmeticFunc<  C,  C,  C,  0, fp16Pow>                                                     },
16745                 {       "FMin",                                 "",                     2,      C,              C,              C,              0, &getInputData,       compareFP16ArithmeticFunc<  C,  C,  C,  0, fp16FMin>                                            },
16746                 {       "FMax",                                 "",                     2,      C,              C,              C,              0, &getInputData,       compareFP16ArithmeticFunc<  C,  C,  C,  0, fp16FMax>                                            },
16747                 {       "Step",                                 "",                     2,      C,              C,              C,              0, &getInputData,       compareFP16ArithmeticFunc<  C,  C,  C,  0, fp16Step>                                            },
16748                 {       "Ldexp",                                "",                     2,      C,              C,              C,              0, &getInputData,       compareFP16ArithmeticFunc<  C,  C,  C,  0, fp16Ldexp>                                           },
16749                 {       "FClamp",                               "",                     3,      C,              C,              C,              C, &getInputData,       compareFP16ArithmeticFunc<  C,  C,  C,  C, fp16FClamp>                                          },
16750                 {       "FMix",                                 "",                     3,      C,              C,              C,              C, &getInputDataD,      compareFP16ArithmeticFunc<  C,  C,  C,  C, fp16FMix>                                            },
16751                 {       "SmoothStep",                   "",                     3,      C,              C,              C,              C, &getInputDataSS,     compareFP16ArithmeticFunc<  C,  C,  C,  C, fp16SmoothStep>                                      },
16752                 {       "Fma",                                  "",                     3,      C,              C,              C,              C, &getInputData,       compareFP16ArithmeticFunc<  C,  C,  C,  C, fp16Fma>                                                     },
16753                 {       "Length",                               "",                     1,      1,              C,              0,              0, &getInputData,       compareFP16ArithmeticFunc<  1,  C,  0,  0, fp16Length>                                          },
16754                 {       "Distance",                             "",                     2,      1,              C,              C,              0, &getInputData,       compareFP16ArithmeticFunc<  1,  C,  C,  0, fp16Distance>                                        },
16755                 {       "Cross",                                "",                     2,      C,              C,              C,              0, &getInputDataD,      compareFP16ArithmeticFunc<  C,  C,  C,  0, fp16Cross>                                           },
16756                 {       "Normalize",                    "",                     1,      C,              C,              0,              0, &getInputData,       compareFP16ArithmeticFunc<  C,  C,  0,  0, fp16Normalize>                                       },
16757                 {       "FaceForward",                  "",                     3,      C,              C,              C,              C, &getInputDataD,      compareFP16ArithmeticFunc<  C,  C,  C,  C, fp16FaceForward>                                     },
16758                 {       "Reflect",                              "",                     2,      C,              C,              C,              0, &getInputDataD,      compareFP16ArithmeticFunc<  C,  C,  C,  0, fp16Reflect>                                         },
16759                 {       "Refract",                              "",                     3,      C,              C,              C,              1, &getInputDataN,      compareFP16ArithmeticFunc<  C,  C,  C,  1, fp16Refract>                                         },
16760                 {       "OpDot",                                "",                     2,      1,              C,              C,              0, &getInputDataD,      compareFP16ArithmeticFunc<  1,  C,  C,  0, fp16Dot>                                                     },
16761                 {       "OpVectorTimesScalar",  "",                     2,      C,              C,              1,              0, &getInputDataV,      compareFP16ArithmeticFunc<  C,  C,  1,  0, fp16VectorTimesScalar>                       },
16762         };
16763
16764         for (deUint32 testFuncIdx = 0; testFuncIdx < DE_LENGTH_OF_ARRAY(testFuncs); ++testFuncIdx)
16765         {
16766                 const Math16TestFunc&   testFunc                = testFuncs[testFuncIdx];
16767                 const string                    funcNameString  = testFunc.funcName;
16768
16769                 if ((C != 3) && funcNameString == "Cross")
16770                         continue;
16771
16772                 if ((C < 2) && funcNameString == "OpDot")
16773                         continue;
16774
16775                 if ((C < 2) && funcNameString == "OpVectorTimesScalar")
16776                         continue;
16777
16778                 createFloat16ArithmeticFuncTest<SpecResource>(testCtx, *testGroup.get(), C, testFunc);
16779         }
16780
16781         return testGroup.release();
16782 }
16783
16784 template<class SpecResource>
16785 tcu::TestCaseGroup* createFloat16ArithmeticSet (tcu::TestContext& testCtx)
16786 {
16787         const std::string                               testGroupName   ("arithmetic");
16788         de::MovePtr<tcu::TestCaseGroup> testGroup               (new tcu::TestCaseGroup(testCtx, testGroupName.c_str(), "Float 16 arithmetic and related tests"));
16789         const Math16TestFunc                    testFuncs[]             =
16790         {
16791                 {       "OpTranspose",                  "2x2",          1,      MAT2X2, MAT2X2, 0,              0, &getInputDataM,      compareFP16ArithmeticFunc<  4,  4,  0,  0, fp16Transpose<2,2> >                         },
16792                 {       "OpTranspose",                  "3x2",          1,      MAT2X3, MAT3X2, 0,              0, &getInputDataM,      compareFP16ArithmeticFunc<  8,  8,  0,  0, fp16Transpose<3,2> >                         },
16793                 {       "OpTranspose",                  "4x2",          1,      MAT2X4, MAT4X2, 0,              0, &getInputDataM,      compareFP16ArithmeticFunc<  8,  8,  0,  0, fp16Transpose<4,2> >                         },
16794                 {       "OpTranspose",                  "2x3",          1,      MAT3X2, MAT2X3, 0,              0, &getInputDataM,      compareFP16ArithmeticFunc<  8,  8,  0,  0, fp16Transpose<2,3> >                         },
16795                 {       "OpTranspose",                  "3x3",          1,      MAT3X3, MAT3X3, 0,              0, &getInputDataM,      compareFP16ArithmeticFunc< 16, 16,  0,  0, fp16Transpose<3,3> >                         },
16796                 {       "OpTranspose",                  "4x3",          1,      MAT3X4, MAT4X3, 0,              0, &getInputDataM,      compareFP16ArithmeticFunc< 16, 16,  0,  0, fp16Transpose<4,3> >                         },
16797                 {       "OpTranspose",                  "2x4",          1,      MAT4X2, MAT2X4, 0,              0, &getInputDataM,      compareFP16ArithmeticFunc<  8,  8,  0,  0, fp16Transpose<2,4> >                         },
16798                 {       "OpTranspose",                  "3x4",          1,      MAT4X3, MAT3X4, 0,              0, &getInputDataM,      compareFP16ArithmeticFunc< 16, 16,  0,  0, fp16Transpose<3,4> >                         },
16799                 {       "OpTranspose",                  "4x4",          1,      MAT4X4, MAT4X4, 0,              0, &getInputDataM,      compareFP16ArithmeticFunc< 16, 16,  0,  0, fp16Transpose<4,4> >                         },
16800                 {       "OpMatrixTimesScalar",  "2x2",          2,      MAT2X2, MAT2X2, 1,              0, &getInputDataD,      compareFP16ArithmeticFunc<  4,  4,  1,  0, fp16MatrixTimesScalar<2,2> >         },
16801                 {       "OpMatrixTimesScalar",  "2x3",          2,      MAT2X3, MAT2X3, 1,              0, &getInputDataD,      compareFP16ArithmeticFunc<  8,  8,  1,  0, fp16MatrixTimesScalar<2,3> >         },
16802                 {       "OpMatrixTimesScalar",  "2x4",          2,      MAT2X4, MAT2X4, 1,              0, &getInputDataD,      compareFP16ArithmeticFunc<  8,  8,  1,  0, fp16MatrixTimesScalar<2,4> >         },
16803                 {       "OpMatrixTimesScalar",  "3x2",          2,      MAT3X2, MAT3X2, 1,              0, &getInputDataD,      compareFP16ArithmeticFunc<  8,  8,  1,  0, fp16MatrixTimesScalar<3,2> >         },
16804                 {       "OpMatrixTimesScalar",  "3x3",          2,      MAT3X3, MAT3X3, 1,              0, &getInputDataD,      compareFP16ArithmeticFunc< 16, 16,  1,  0, fp16MatrixTimesScalar<3,3> >         },
16805                 {       "OpMatrixTimesScalar",  "3x4",          2,      MAT3X4, MAT3X4, 1,              0, &getInputDataD,      compareFP16ArithmeticFunc< 16, 16,  1,  0, fp16MatrixTimesScalar<3,4> >         },
16806                 {       "OpMatrixTimesScalar",  "4x2",          2,      MAT4X2, MAT4X2, 1,              0, &getInputDataD,      compareFP16ArithmeticFunc<  8,  8,  1,  0, fp16MatrixTimesScalar<4,2> >         },
16807                 {       "OpMatrixTimesScalar",  "4x3",          2,      MAT4X3, MAT4X3, 1,              0, &getInputDataD,      compareFP16ArithmeticFunc< 16, 16,  1,  0, fp16MatrixTimesScalar<4,3> >         },
16808                 {       "OpMatrixTimesScalar",  "4x4",          2,      MAT4X4, MAT4X4, 1,              0, &getInputDataD,      compareFP16ArithmeticFunc< 16, 16,  1,  0, fp16MatrixTimesScalar<4,4> >         },
16809                 {       "OpVectorTimesMatrix",  "2x2",          2,      VEC2,   VEC2,   MAT2X2, 0, &getInputDataD,      compareFP16ArithmeticFunc<  2,  2,  4,  0, fp16VectorTimesMatrix<2,2> >         },
16810                 {       "OpVectorTimesMatrix",  "2x3",          2,      VEC2,   VEC3,   MAT2X3, 0, &getInputDataD,      compareFP16ArithmeticFunc<  2,  3,  8,  0, fp16VectorTimesMatrix<2,3> >         },
16811                 {       "OpVectorTimesMatrix",  "2x4",          2,      VEC2,   VEC4,   MAT2X4, 0, &getInputDataD,      compareFP16ArithmeticFunc<  2,  4,  8,  0, fp16VectorTimesMatrix<2,4> >         },
16812                 {       "OpVectorTimesMatrix",  "3x2",          2,      VEC3,   VEC2,   MAT3X2, 0, &getInputDataD,      compareFP16ArithmeticFunc<  3,  2,  8,  0, fp16VectorTimesMatrix<3,2> >         },
16813                 {       "OpVectorTimesMatrix",  "3x3",          2,      VEC3,   VEC3,   MAT3X3, 0, &getInputDataD,      compareFP16ArithmeticFunc<  3,  3, 16,  0, fp16VectorTimesMatrix<3,3> >         },
16814                 {       "OpVectorTimesMatrix",  "3x4",          2,      VEC3,   VEC4,   MAT3X4, 0, &getInputDataD,      compareFP16ArithmeticFunc<  3,  4, 16,  0, fp16VectorTimesMatrix<3,4> >         },
16815                 {       "OpVectorTimesMatrix",  "4x2",          2,      VEC4,   VEC2,   MAT4X2, 0, &getInputDataD,      compareFP16ArithmeticFunc<  4,  2,  8,  0, fp16VectorTimesMatrix<4,2> >         },
16816                 {       "OpVectorTimesMatrix",  "4x3",          2,      VEC4,   VEC3,   MAT4X3, 0, &getInputDataD,      compareFP16ArithmeticFunc<  4,  3, 16,  0, fp16VectorTimesMatrix<4,3> >         },
16817                 {       "OpVectorTimesMatrix",  "4x4",          2,      VEC4,   VEC4,   MAT4X4, 0, &getInputDataD,      compareFP16ArithmeticFunc<  4,  4, 16,  0, fp16VectorTimesMatrix<4,4> >         },
16818                 {       "OpMatrixTimesVector",  "2x2",          2,      VEC2,   MAT2X2, VEC2,   0, &getInputDataD,      compareFP16ArithmeticFunc<  2,  4,  2,  0, fp16MatrixTimesVector<2,2> >         },
16819                 {       "OpMatrixTimesVector",  "2x3",          2,      VEC3,   MAT2X3, VEC2,   0, &getInputDataD,      compareFP16ArithmeticFunc<  3,  8,  2,  0, fp16MatrixTimesVector<2,3> >         },
16820                 {       "OpMatrixTimesVector",  "2x4",          2,      VEC4,   MAT2X4, VEC2,   0, &getInputDataD,      compareFP16ArithmeticFunc<  4,  8,  2,  0, fp16MatrixTimesVector<2,4> >         },
16821                 {       "OpMatrixTimesVector",  "3x2",          2,      VEC2,   MAT3X2, VEC3,   0, &getInputDataD,      compareFP16ArithmeticFunc<  2,  8,  3,  0, fp16MatrixTimesVector<3,2> >         },
16822                 {       "OpMatrixTimesVector",  "3x3",          2,      VEC3,   MAT3X3, VEC3,   0, &getInputDataD,      compareFP16ArithmeticFunc<  3, 16,  3,  0, fp16MatrixTimesVector<3,3> >         },
16823                 {       "OpMatrixTimesVector",  "3x4",          2,      VEC4,   MAT3X4, VEC3,   0, &getInputDataD,      compareFP16ArithmeticFunc<  4, 16,  3,  0, fp16MatrixTimesVector<3,4> >         },
16824                 {       "OpMatrixTimesVector",  "4x2",          2,      VEC2,   MAT4X2, VEC4,   0, &getInputDataD,      compareFP16ArithmeticFunc<  2,  8,  4,  0, fp16MatrixTimesVector<4,2> >         },
16825                 {       "OpMatrixTimesVector",  "4x3",          2,      VEC3,   MAT4X3, VEC4,   0, &getInputDataD,      compareFP16ArithmeticFunc<  3, 16,  4,  0, fp16MatrixTimesVector<4,3> >         },
16826                 {       "OpMatrixTimesVector",  "4x4",          2,      VEC4,   MAT4X4, VEC4,   0, &getInputDataD,      compareFP16ArithmeticFunc<  4, 16,  4,  0, fp16MatrixTimesVector<4,4> >         },
16827                 {       "OpMatrixTimesMatrix",  "2x2_2x2",      2,      MAT2X2, MAT2X2, MAT2X2, 0, &getInputDataD,      compareFP16ArithmeticFunc<  4,  4,  4,  0, fp16MatrixTimesMatrix<2,2,2,2> >     },
16828                 {       "OpMatrixTimesMatrix",  "2x2_3x2",      2,      MAT3X2, MAT2X2, MAT3X2, 0, &getInputDataD,      compareFP16ArithmeticFunc<  8,  4,  8,  0, fp16MatrixTimesMatrix<2,2,3,2> >     },
16829                 {       "OpMatrixTimesMatrix",  "2x2_4x2",      2,      MAT4X2, MAT2X2, MAT4X2, 0, &getInputDataD,      compareFP16ArithmeticFunc<  8,  4,  8,  0, fp16MatrixTimesMatrix<2,2,4,2> >     },
16830                 {       "OpMatrixTimesMatrix",  "2x3_2x2",      2,      MAT2X3, MAT2X3, MAT2X2, 0, &getInputDataD,      compareFP16ArithmeticFunc<  8,  8,  4,  0, fp16MatrixTimesMatrix<2,3,2,2> >     },
16831                 {       "OpMatrixTimesMatrix",  "2x3_3x2",      2,      MAT3X3, MAT2X3, MAT3X2, 0, &getInputDataD,      compareFP16ArithmeticFunc< 16,  8,  8,  0, fp16MatrixTimesMatrix<2,3,3,2> >     },
16832                 {       "OpMatrixTimesMatrix",  "2x3_4x2",      2,      MAT4X3, MAT2X3, MAT4X2, 0, &getInputDataD,      compareFP16ArithmeticFunc< 16,  8,  8,  0, fp16MatrixTimesMatrix<2,3,4,2> >     },
16833                 {       "OpMatrixTimesMatrix",  "2x4_2x2",      2,      MAT2X4, MAT2X4, MAT2X2, 0, &getInputDataD,      compareFP16ArithmeticFunc<  8,  8,  4,  0, fp16MatrixTimesMatrix<2,4,2,2> >     },
16834                 {       "OpMatrixTimesMatrix",  "2x4_3x2",      2,      MAT3X4, MAT2X4, MAT3X2, 0, &getInputDataD,      compareFP16ArithmeticFunc< 16,  8,  8,  0, fp16MatrixTimesMatrix<2,4,3,2> >     },
16835                 {       "OpMatrixTimesMatrix",  "2x4_4x2",      2,      MAT4X4, MAT2X4, MAT4X2, 0, &getInputDataD,      compareFP16ArithmeticFunc< 16,  8,  8,  0, fp16MatrixTimesMatrix<2,4,4,2> >     },
16836                 {       "OpMatrixTimesMatrix",  "3x2_2x3",      2,      MAT2X2, MAT3X2, MAT2X3, 0, &getInputDataD,      compareFP16ArithmeticFunc<  4,  8,  8,  0, fp16MatrixTimesMatrix<3,2,2,3> >     },
16837                 {       "OpMatrixTimesMatrix",  "3x2_3x3",      2,      MAT3X2, MAT3X2, MAT3X3, 0, &getInputDataD,      compareFP16ArithmeticFunc<  8,  8, 16,  0, fp16MatrixTimesMatrix<3,2,3,3> >     },
16838                 {       "OpMatrixTimesMatrix",  "3x2_4x3",      2,      MAT4X2, MAT3X2, MAT4X3, 0, &getInputDataD,      compareFP16ArithmeticFunc<  8,  8, 16,  0, fp16MatrixTimesMatrix<3,2,4,3> >     },
16839                 {       "OpMatrixTimesMatrix",  "3x3_2x3",      2,      MAT2X3, MAT3X3, MAT2X3, 0, &getInputDataD,      compareFP16ArithmeticFunc<  8, 16,  8,  0, fp16MatrixTimesMatrix<3,3,2,3> >     },
16840                 {       "OpMatrixTimesMatrix",  "3x3_3x3",      2,      MAT3X3, MAT3X3, MAT3X3, 0, &getInputDataD,      compareFP16ArithmeticFunc< 16, 16, 16,  0, fp16MatrixTimesMatrix<3,3,3,3> >     },
16841                 {       "OpMatrixTimesMatrix",  "3x3_4x3",      2,      MAT4X3, MAT3X3, MAT4X3, 0, &getInputDataD,      compareFP16ArithmeticFunc< 16, 16, 16,  0, fp16MatrixTimesMatrix<3,3,4,3> >     },
16842                 {       "OpMatrixTimesMatrix",  "3x4_2x3",      2,      MAT2X4, MAT3X4, MAT2X3, 0, &getInputDataD,      compareFP16ArithmeticFunc<  8, 16,  8,  0, fp16MatrixTimesMatrix<3,4,2,3> >     },
16843                 {       "OpMatrixTimesMatrix",  "3x4_3x3",      2,      MAT3X4, MAT3X4, MAT3X3, 0, &getInputDataD,      compareFP16ArithmeticFunc< 16, 16, 16,  0, fp16MatrixTimesMatrix<3,4,3,3> >     },
16844                 {       "OpMatrixTimesMatrix",  "3x4_4x3",      2,      MAT4X4, MAT3X4, MAT4X3, 0, &getInputDataD,      compareFP16ArithmeticFunc< 16, 16, 16,  0, fp16MatrixTimesMatrix<3,4,4,3> >     },
16845                 {       "OpMatrixTimesMatrix",  "4x2_2x4",      2,      MAT2X2, MAT4X2, MAT2X4, 0, &getInputDataD,      compareFP16ArithmeticFunc<  4,  8,  8,  0, fp16MatrixTimesMatrix<4,2,2,4> >     },
16846                 {       "OpMatrixTimesMatrix",  "4x2_3x4",      2,      MAT3X2, MAT4X2, MAT3X4, 0, &getInputDataD,      compareFP16ArithmeticFunc<  8,  8, 16,  0, fp16MatrixTimesMatrix<4,2,3,4> >     },
16847                 {       "OpMatrixTimesMatrix",  "4x2_4x4",      2,      MAT4X2, MAT4X2, MAT4X4, 0, &getInputDataD,      compareFP16ArithmeticFunc<  8,  8, 16,  0, fp16MatrixTimesMatrix<4,2,4,4> >     },
16848                 {       "OpMatrixTimesMatrix",  "4x3_2x4",      2,      MAT2X3, MAT4X3, MAT2X4, 0, &getInputDataD,      compareFP16ArithmeticFunc<  8, 16,  8,  0, fp16MatrixTimesMatrix<4,3,2,4> >     },
16849                 {       "OpMatrixTimesMatrix",  "4x3_3x4",      2,      MAT3X3, MAT4X3, MAT3X4, 0, &getInputDataD,      compareFP16ArithmeticFunc< 16, 16, 16,  0, fp16MatrixTimesMatrix<4,3,3,4> >     },
16850                 {       "OpMatrixTimesMatrix",  "4x3_4x4",      2,      MAT4X3, MAT4X3, MAT4X4, 0, &getInputDataD,      compareFP16ArithmeticFunc< 16, 16, 16,  0, fp16MatrixTimesMatrix<4,3,4,4> >     },
16851                 {       "OpMatrixTimesMatrix",  "4x4_2x4",      2,      MAT2X4, MAT4X4, MAT2X4, 0, &getInputDataD,      compareFP16ArithmeticFunc<  8, 16,  8,  0, fp16MatrixTimesMatrix<4,4,2,4> >     },
16852                 {       "OpMatrixTimesMatrix",  "4x4_3x4",      2,      MAT3X4, MAT4X4, MAT3X4, 0, &getInputDataD,      compareFP16ArithmeticFunc< 16, 16, 16,  0, fp16MatrixTimesMatrix<4,4,3,4> >     },
16853                 {       "OpMatrixTimesMatrix",  "4x4_4x4",      2,      MAT4X4, MAT4X4, MAT4X4, 0, &getInputDataD,      compareFP16ArithmeticFunc< 16, 16, 16,  0, fp16MatrixTimesMatrix<4,4,4,4> >     },
16854                 {       "OpOuterProduct",               "2x2",          2,      MAT2X2, VEC2,   VEC2,   0, &getInputDataD,      compareFP16ArithmeticFunc<  4,  2,  2,  0, fp16OuterProduct<2,2> >                      },
16855                 {       "OpOuterProduct",               "2x3",          2,      MAT2X3, VEC3,   VEC2,   0, &getInputDataD,      compareFP16ArithmeticFunc<  8,  3,  2,  0, fp16OuterProduct<2,3> >                      },
16856                 {       "OpOuterProduct",               "2x4",          2,      MAT2X4, VEC4,   VEC2,   0, &getInputDataD,      compareFP16ArithmeticFunc<  8,  4,  2,  0, fp16OuterProduct<2,4> >                      },
16857                 {       "OpOuterProduct",               "3x2",          2,      MAT3X2, VEC2,   VEC3,   0, &getInputDataD,      compareFP16ArithmeticFunc<  8,  2,  3,  0, fp16OuterProduct<3,2> >                      },
16858                 {       "OpOuterProduct",               "3x3",          2,      MAT3X3, VEC3,   VEC3,   0, &getInputDataD,      compareFP16ArithmeticFunc< 16,  3,  3,  0, fp16OuterProduct<3,3> >                      },
16859                 {       "OpOuterProduct",               "3x4",          2,      MAT3X4, VEC4,   VEC3,   0, &getInputDataD,      compareFP16ArithmeticFunc< 16,  4,  3,  0, fp16OuterProduct<3,4> >                      },
16860                 {       "OpOuterProduct",               "4x2",          2,      MAT4X2, VEC2,   VEC4,   0, &getInputDataD,      compareFP16ArithmeticFunc<  8,  2,  4,  0, fp16OuterProduct<4,2> >                      },
16861                 {       "OpOuterProduct",               "4x3",          2,      MAT4X3, VEC3,   VEC4,   0, &getInputDataD,      compareFP16ArithmeticFunc< 16,  3,  4,  0, fp16OuterProduct<4,3> >                      },
16862                 {       "OpOuterProduct",               "4x4",          2,      MAT4X4, VEC4,   VEC4,   0, &getInputDataD,      compareFP16ArithmeticFunc< 16,  4,  4,  0, fp16OuterProduct<4,4> >                      },
16863                 {       "Determinant",                  "2x2",          1,      SCALAR, MAT2X2, NONE,   0, &getInputDataC,      compareFP16ArithmeticFunc<  1,  4,  0,  0, fp16Determinant<2> >                         },
16864                 {       "Determinant",                  "3x3",          1,      SCALAR, MAT3X3, NONE,   0, &getInputDataC,      compareFP16ArithmeticFunc<  1, 16,  0,  0, fp16Determinant<3> >                         },
16865                 {       "Determinant",                  "4x4",          1,      SCALAR, MAT4X4, NONE,   0, &getInputDataC,      compareFP16ArithmeticFunc<  1, 16,  0,  0, fp16Determinant<4> >                         },
16866                 {       "MatrixInverse",                "2x2",          1,      MAT2X2, MAT2X2, NONE,   0, &getInputDataC,      compareFP16ArithmeticFunc<  4,  4,  0,  0, fp16Inverse<2> >                                     },
16867         };
16868
16869         for (deUint32 testFuncIdx = 0; testFuncIdx < DE_LENGTH_OF_ARRAY(testFuncs); ++testFuncIdx)
16870         {
16871                 const Math16TestFunc&   testFunc        = testFuncs[testFuncIdx];
16872
16873                 createFloat16ArithmeticFuncTest<SpecResource>(testCtx, *testGroup.get(), 0, testFunc);
16874         }
16875
16876         return testGroup.release();
16877 }
16878
16879 const string getNumberTypeName (const NumberType type)
16880 {
16881         if (type == NUMBERTYPE_INT32)
16882         {
16883                 return "int";
16884         }
16885         else if (type == NUMBERTYPE_UINT32)
16886         {
16887                 return "uint";
16888         }
16889         else if (type == NUMBERTYPE_FLOAT32)
16890         {
16891                 return "float";
16892         }
16893         else
16894         {
16895                 DE_ASSERT(false);
16896                 return "";
16897         }
16898 }
16899
16900 deInt32 getInt(de::Random& rnd)
16901 {
16902         return rnd.getInt(std::numeric_limits<int>::min(), std::numeric_limits<int>::max());
16903 }
16904
16905 const string repeatString (const string& str, int times)
16906 {
16907         string filler;
16908         for (int i = 0; i < times; ++i)
16909         {
16910                 filler += str;
16911         }
16912         return filler;
16913 }
16914
16915 const string getRandomConstantString (const NumberType type, de::Random& rnd)
16916 {
16917         if (type == NUMBERTYPE_INT32)
16918         {
16919                 return numberToString<deInt32>(getInt(rnd));
16920         }
16921         else if (type == NUMBERTYPE_UINT32)
16922         {
16923                 return numberToString<deUint32>(rnd.getUint32());
16924         }
16925         else if (type == NUMBERTYPE_FLOAT32)
16926         {
16927                 return numberToString<float>(rnd.getFloat());
16928         }
16929         else
16930         {
16931                 DE_ASSERT(false);
16932                 return "";
16933         }
16934 }
16935
16936 void createVectorCompositeCases (vector<map<string, string> >& testCases, de::Random& rnd, const NumberType type)
16937 {
16938         map<string, string> params;
16939
16940         // Vec2 to Vec4
16941         for (int width = 2; width <= 4; ++width)
16942         {
16943                 const string randomConst = numberToString(getInt(rnd));
16944                 const string widthStr = numberToString(width);
16945                 const string composite_type = "${customType}vec" + widthStr;
16946                 const int index = rnd.getInt(0, width-1);
16947
16948                 params["type"]                  = "vec";
16949                 params["name"]                  = params["type"] + "_" + widthStr;
16950                 params["compositeDecl"]         = composite_type + " = OpTypeVector ${customType} " + widthStr +"\n";
16951                 params["compositeType"]         = composite_type;
16952                 params["filler"]                = string("%filler    = OpConstant ${customType} ") + getRandomConstantString(type, rnd) + "\n";
16953                 params["compositeConstruct"]    = "%instance  = OpCompositeConstruct " + composite_type + repeatString(" %filler", width) + "\n";
16954                 params["indexes"]               = numberToString(index);
16955                 testCases.push_back(params);
16956         }
16957 }
16958
16959 void createArrayCompositeCases (vector<map<string, string> >& testCases, de::Random& rnd, const NumberType type)
16960 {
16961         const int limit = 10;
16962         map<string, string> params;
16963
16964         for (int width = 2; width <= limit; ++width)
16965         {
16966                 string randomConst = numberToString(getInt(rnd));
16967                 string widthStr = numberToString(width);
16968                 int index = rnd.getInt(0, width-1);
16969
16970                 params["type"]                  = "array";
16971                 params["name"]                  = params["type"] + "_" + widthStr;
16972                 params["compositeDecl"]         = string("%arraywidth = OpConstant %u32 " + widthStr + "\n")
16973                                                                                         +        "%composite = OpTypeArray ${customType} %arraywidth\n";
16974                 params["compositeType"]         = "%composite";
16975                 params["filler"]                = string("%filler    = OpConstant ${customType} ") + getRandomConstantString(type, rnd) + "\n";
16976                 params["compositeConstruct"]    = "%instance  = OpCompositeConstruct %composite" + repeatString(" %filler", width) + "\n";
16977                 params["indexes"]               = numberToString(index);
16978                 testCases.push_back(params);
16979         }
16980 }
16981
16982 void createStructCompositeCases (vector<map<string, string> >& testCases, de::Random& rnd, const NumberType type)
16983 {
16984         const int limit = 10;
16985         map<string, string> params;
16986
16987         for (int width = 2; width <= limit; ++width)
16988         {
16989                 string randomConst = numberToString(getInt(rnd));
16990                 int index = rnd.getInt(0, width-1);
16991
16992                 params["type"]                  = "struct";
16993                 params["name"]                  = params["type"] + "_" + numberToString(width);
16994                 params["compositeDecl"]         = "%composite = OpTypeStruct" + repeatString(" ${customType}", width) + "\n";
16995                 params["compositeType"]         = "%composite";
16996                 params["filler"]                = string("%filler    = OpConstant ${customType} ") + getRandomConstantString(type, rnd) + "\n";
16997                 params["compositeConstruct"]    = "%instance  = OpCompositeConstruct %composite" + repeatString(" %filler", width) + "\n";
16998                 params["indexes"]               = numberToString(index);
16999                 testCases.push_back(params);
17000         }
17001 }
17002
17003 void createMatrixCompositeCases (vector<map<string, string> >& testCases, de::Random& rnd, const NumberType type)
17004 {
17005         map<string, string> params;
17006
17007         // Vec2 to Vec4
17008         for (int width = 2; width <= 4; ++width)
17009         {
17010                 string widthStr = numberToString(width);
17011
17012                 for (int column = 2 ; column <= 4; ++column)
17013                 {
17014                         int index_0 = rnd.getInt(0, column-1);
17015                         int index_1 = rnd.getInt(0, width-1);
17016                         string columnStr = numberToString(column);
17017
17018                         params["type"]          = "matrix";
17019                         params["name"]          = params["type"] + "_" + widthStr + "x" + columnStr;
17020                         params["compositeDecl"] = string("%vectype   = OpTypeVector ${customType} " + widthStr + "\n")
17021                                                                                                 +        "%composite = OpTypeMatrix %vectype " + columnStr + "\n";
17022                         params["compositeType"] = "%composite";
17023
17024                         params["filler"]        = string("%filler    = OpConstant ${customType} ") + getRandomConstantString(type, rnd) + "\n"
17025                                                                                                 +        "%fillerVec = OpConstantComposite %vectype" + repeatString(" %filler", width) + "\n";
17026
17027                         params["compositeConstruct"]    = "%instance  = OpCompositeConstruct %composite" + repeatString(" %fillerVec", column) + "\n";
17028                         params["indexes"]       = numberToString(index_0) + " " + numberToString(index_1);
17029                         testCases.push_back(params);
17030                 }
17031         }
17032 }
17033
17034 void createCompositeCases (vector<map<string, string> >& testCases, de::Random& rnd, const NumberType type)
17035 {
17036         createVectorCompositeCases(testCases, rnd, type);
17037         createArrayCompositeCases(testCases, rnd, type);
17038         createStructCompositeCases(testCases, rnd, type);
17039         // Matrix only supports float types
17040         if (type == NUMBERTYPE_FLOAT32)
17041         {
17042                 createMatrixCompositeCases(testCases, rnd, type);
17043         }
17044 }
17045
17046 const string getAssemblyTypeDeclaration (const NumberType type)
17047 {
17048         switch (type)
17049         {
17050                 case NUMBERTYPE_INT32:          return "OpTypeInt 32 1";
17051                 case NUMBERTYPE_UINT32:         return "OpTypeInt 32 0";
17052                 case NUMBERTYPE_FLOAT32:        return "OpTypeFloat 32";
17053                 default:                        DE_ASSERT(false); return "";
17054         }
17055 }
17056
17057 const string getAssemblyTypeName (const NumberType type)
17058 {
17059         switch (type)
17060         {
17061                 case NUMBERTYPE_INT32:          return "%i32";
17062                 case NUMBERTYPE_UINT32:         return "%u32";
17063                 case NUMBERTYPE_FLOAT32:        return "%f32";
17064                 default:                        DE_ASSERT(false); return "";
17065         }
17066 }
17067
17068 const string specializeCompositeInsertShaderTemplate (const NumberType type, const map<string, string>& params)
17069 {
17070         map<string, string>     parameters(params);
17071
17072         const string customType = getAssemblyTypeName(type);
17073         map<string, string> substCustomType;
17074         substCustomType["customType"] = customType;
17075         parameters["compositeDecl"] = StringTemplate(parameters.at("compositeDecl")).specialize(substCustomType);
17076         parameters["compositeType"] = StringTemplate(parameters.at("compositeType")).specialize(substCustomType);
17077         parameters["compositeConstruct"] = StringTemplate(parameters.at("compositeConstruct")).specialize(substCustomType);
17078         parameters["filler"] = StringTemplate(parameters.at("filler")).specialize(substCustomType);
17079         parameters["customType"] = customType;
17080         parameters["compositeDecorator"] = (parameters["type"] == "array") ? "OpDecorate %composite ArrayStride 4\n" : "";
17081
17082         if (parameters.at("compositeType") != "%u32vec3")
17083         {
17084                 parameters["u32vec3Decl"] = "%u32vec3   = OpTypeVector %u32 3\n";
17085         }
17086
17087         return StringTemplate(
17088                 "OpCapability Shader\n"
17089                 "OpCapability Matrix\n"
17090                 "OpMemoryModel Logical GLSL450\n"
17091                 "OpEntryPoint GLCompute %main \"main\" %id\n"
17092                 "OpExecutionMode %main LocalSize 1 1 1\n"
17093
17094                 "OpSource GLSL 430\n"
17095                 "OpName %main           \"main\"\n"
17096                 "OpName %id             \"gl_GlobalInvocationID\"\n"
17097
17098                 // Decorators
17099                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
17100                 "OpDecorate %buf BufferBlock\n"
17101                 "OpDecorate %indata DescriptorSet 0\n"
17102                 "OpDecorate %indata Binding 0\n"
17103                 "OpDecorate %outdata DescriptorSet 0\n"
17104                 "OpDecorate %outdata Binding 1\n"
17105                 "OpDecorate %customarr ArrayStride 4\n"
17106                 "${compositeDecorator}"
17107                 "OpMemberDecorate %buf 0 Offset 0\n"
17108
17109                 // General types
17110                 "%void      = OpTypeVoid\n"
17111                 "%voidf     = OpTypeFunction %void\n"
17112                 "%u32       = OpTypeInt 32 0\n"
17113                 "%i32       = OpTypeInt 32 1\n"
17114                 "%f32       = OpTypeFloat 32\n"
17115
17116                 // Composite declaration
17117                 "${compositeDecl}"
17118
17119                 // Constants
17120                 "${filler}"
17121
17122                 "${u32vec3Decl:opt}"
17123                 "%uvec3ptr  = OpTypePointer Input %u32vec3\n"
17124
17125                 // Inherited from custom
17126                 "%customptr = OpTypePointer Uniform ${customType}\n"
17127                 "%customarr = OpTypeRuntimeArray ${customType}\n"
17128                 "%buf       = OpTypeStruct %customarr\n"
17129                 "%bufptr    = OpTypePointer Uniform %buf\n"
17130
17131                 "%indata    = OpVariable %bufptr Uniform\n"
17132                 "%outdata   = OpVariable %bufptr Uniform\n"
17133
17134                 "%id        = OpVariable %uvec3ptr Input\n"
17135                 "%zero      = OpConstant %i32 0\n"
17136
17137                 "%main      = OpFunction %void None %voidf\n"
17138                 "%label     = OpLabel\n"
17139                 "%idval     = OpLoad %u32vec3 %id\n"
17140                 "%x         = OpCompositeExtract %u32 %idval 0\n"
17141
17142                 "%inloc     = OpAccessChain %customptr %indata %zero %x\n"
17143                 "%outloc    = OpAccessChain %customptr %outdata %zero %x\n"
17144                 // Read the input value
17145                 "%inval     = OpLoad ${customType} %inloc\n"
17146                 // Create the composite and fill it
17147                 "${compositeConstruct}"
17148                 // Insert the input value to a place
17149                 "%instance2 = OpCompositeInsert ${compositeType} %inval %instance ${indexes}\n"
17150                 // Read back the value from the position
17151                 "%out_val   = OpCompositeExtract ${customType} %instance2 ${indexes}\n"
17152                 // Store it in the output position
17153                 "             OpStore %outloc %out_val\n"
17154                 "             OpReturn\n"
17155                 "             OpFunctionEnd\n"
17156         ).specialize(parameters);
17157 }
17158
17159 template<typename T>
17160 BufferSp createCompositeBuffer(T number)
17161 {
17162         return BufferSp(new Buffer<T>(vector<T>(1, number)));
17163 }
17164
17165 tcu::TestCaseGroup* createOpCompositeInsertGroup (tcu::TestContext& testCtx)
17166 {
17167         de::MovePtr<tcu::TestCaseGroup> group   (new tcu::TestCaseGroup(testCtx, "opcompositeinsert", "Test the OpCompositeInsert instruction"));
17168         de::Random                                              rnd             (deStringHash(group->getName()));
17169
17170         for (int type = NUMBERTYPE_INT32; type != NUMBERTYPE_END32; ++type)
17171         {
17172                 NumberType                                              numberType              = NumberType(type);
17173                 const string                                    typeName                = getNumberTypeName(numberType);
17174                 const string                                    description             = "Test the OpCompositeInsert instruction with " + typeName + "s";
17175                 de::MovePtr<tcu::TestCaseGroup> subGroup                (new tcu::TestCaseGroup(testCtx, typeName.c_str(), description.c_str()));
17176                 vector<map<string, string> >    testCases;
17177
17178                 createCompositeCases(testCases, rnd, numberType);
17179
17180                 for (vector<map<string, string> >::const_iterator test = testCases.begin(); test != testCases.end(); ++test)
17181                 {
17182                         ComputeShaderSpec       spec;
17183
17184                         spec.assembly = specializeCompositeInsertShaderTemplate(numberType, *test);
17185
17186                         switch (numberType)
17187                         {
17188                                 case NUMBERTYPE_INT32:
17189                                 {
17190                                         deInt32 number = getInt(rnd);
17191                                         spec.inputs.push_back(createCompositeBuffer<deInt32>(number));
17192                                         spec.outputs.push_back(createCompositeBuffer<deInt32>(number));
17193                                         break;
17194                                 }
17195                                 case NUMBERTYPE_UINT32:
17196                                 {
17197                                         deUint32 number = rnd.getUint32();
17198                                         spec.inputs.push_back(createCompositeBuffer<deUint32>(number));
17199                                         spec.outputs.push_back(createCompositeBuffer<deUint32>(number));
17200                                         break;
17201                                 }
17202                                 case NUMBERTYPE_FLOAT32:
17203                                 {
17204                                         float number = rnd.getFloat();
17205                                         spec.inputs.push_back(createCompositeBuffer<float>(number));
17206                                         spec.outputs.push_back(createCompositeBuffer<float>(number));
17207                                         break;
17208                                 }
17209                                 default:
17210                                         DE_ASSERT(false);
17211                         }
17212
17213                         spec.numWorkGroups = IVec3(1, 1, 1);
17214                         subGroup->addChild(new SpvAsmComputeShaderCase(testCtx, test->at("name").c_str(), "OpCompositeInsert test", spec));
17215                 }
17216                 group->addChild(subGroup.release());
17217         }
17218         return group.release();
17219 }
17220
17221 struct AssemblyStructInfo
17222 {
17223         AssemblyStructInfo (const deUint32 comp, const deUint32 idx)
17224         : components    (comp)
17225         , index                 (idx)
17226         {}
17227
17228         deUint32 components;
17229         deUint32 index;
17230 };
17231
17232 const string specializeInBoundsShaderTemplate (const NumberType type, const AssemblyStructInfo& structInfo, const map<string, string>& params)
17233 {
17234         // Create the full index string
17235         string                          fullIndex       = numberToString(structInfo.index) + " " + params.at("indexes");
17236         // Convert it to list of indexes
17237         vector<string>          indexes         = de::splitString(fullIndex, ' ');
17238
17239         map<string, string>     parameters      (params);
17240         parameters["structType"]        = repeatString(" ${compositeType}", structInfo.components);
17241         parameters["structConstruct"]   = repeatString(" %instance", structInfo.components);
17242         parameters["insertIndexes"]     = fullIndex;
17243
17244         // In matrix cases the last two index is the CompositeExtract indexes
17245         const deUint32 extractIndexes = (parameters["type"] == "matrix") ? 2 : 1;
17246
17247         // Construct the extractIndex
17248         for (vector<string>::const_iterator index = indexes.end() - extractIndexes; index != indexes.end(); ++index)
17249         {
17250                 parameters["extractIndexes"] += " " + *index;
17251         }
17252
17253         // Remove the last 1 or 2 element depends on matrix case or not
17254         indexes.erase(indexes.end() - extractIndexes, indexes.end());
17255
17256         deUint32 id = 0;
17257         // Generate AccessChain index expressions (except for the last one, because we use ptr to the composite)
17258         for (vector<string>::const_iterator index = indexes.begin(); index != indexes.end(); ++index)
17259         {
17260                 string indexId = "%index_" + numberToString(id++);
17261                 parameters["accessChainConstDeclaration"] += indexId + "   = OpConstant %u32 " + *index + "\n";
17262                 parameters["accessChainIndexes"] += " " + indexId;
17263         }
17264
17265         parameters["compositeDecorator"] = (parameters["type"] == "array") ? "OpDecorate %composite ArrayStride 4\n" : "";
17266
17267         const string customType = getAssemblyTypeName(type);
17268         map<string, string> substCustomType;
17269         substCustomType["customType"] = customType;
17270         parameters["compositeDecl"] = StringTemplate(parameters.at("compositeDecl")).specialize(substCustomType);
17271         parameters["compositeType"] = StringTemplate(parameters.at("compositeType")).specialize(substCustomType);
17272         parameters["compositeConstruct"] = StringTemplate(parameters.at("compositeConstruct")).specialize(substCustomType);
17273         parameters["filler"] = StringTemplate(parameters.at("filler")).specialize(substCustomType);
17274         parameters["customType"] = customType;
17275
17276         const string compositeType = parameters.at("compositeType");
17277         map<string, string> substCompositeType;
17278         substCompositeType["compositeType"] = compositeType;
17279         parameters["structType"] = StringTemplate(parameters.at("structType")).specialize(substCompositeType);
17280         if (compositeType != "%u32vec3")
17281         {
17282                 parameters["u32vec3Decl"] = "%u32vec3   = OpTypeVector %u32 3\n";
17283         }
17284
17285         return StringTemplate(
17286                 "OpCapability Shader\n"
17287                 "OpCapability Matrix\n"
17288                 "OpMemoryModel Logical GLSL450\n"
17289                 "OpEntryPoint GLCompute %main \"main\" %id\n"
17290                 "OpExecutionMode %main LocalSize 1 1 1\n"
17291
17292                 "OpSource GLSL 430\n"
17293                 "OpName %main           \"main\"\n"
17294                 "OpName %id             \"gl_GlobalInvocationID\"\n"
17295                 // Decorators
17296                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
17297                 "OpDecorate %buf BufferBlock\n"
17298                 "OpDecorate %indata DescriptorSet 0\n"
17299                 "OpDecorate %indata Binding 0\n"
17300                 "OpDecorate %outdata DescriptorSet 0\n"
17301                 "OpDecorate %outdata Binding 1\n"
17302                 "OpDecorate %customarr ArrayStride 4\n"
17303                 "${compositeDecorator}"
17304                 "OpMemberDecorate %buf 0 Offset 0\n"
17305                 // General types
17306                 "%void      = OpTypeVoid\n"
17307                 "%voidf     = OpTypeFunction %void\n"
17308                 "%i32       = OpTypeInt 32 1\n"
17309                 "%u32       = OpTypeInt 32 0\n"
17310                 "%f32       = OpTypeFloat 32\n"
17311                 // Custom types
17312                 "${compositeDecl}"
17313                 // %u32vec3 if not already declared in ${compositeDecl}
17314                 "${u32vec3Decl:opt}"
17315                 "%uvec3ptr  = OpTypePointer Input %u32vec3\n"
17316                 // Inherited from composite
17317                 "%composite_p = OpTypePointer Function ${compositeType}\n"
17318                 "%struct_t  = OpTypeStruct${structType}\n"
17319                 "%struct_p  = OpTypePointer Function %struct_t\n"
17320                 // Constants
17321                 "${filler}"
17322                 "${accessChainConstDeclaration}"
17323                 // Inherited from custom
17324                 "%customptr = OpTypePointer Uniform ${customType}\n"
17325                 "%customarr = OpTypeRuntimeArray ${customType}\n"
17326                 "%buf       = OpTypeStruct %customarr\n"
17327                 "%bufptr    = OpTypePointer Uniform %buf\n"
17328                 "%indata    = OpVariable %bufptr Uniform\n"
17329                 "%outdata   = OpVariable %bufptr Uniform\n"
17330
17331                 "%id        = OpVariable %uvec3ptr Input\n"
17332                 "%zero      = OpConstant %u32 0\n"
17333                 "%main      = OpFunction %void None %voidf\n"
17334                 "%label     = OpLabel\n"
17335                 "%struct_v  = OpVariable %struct_p Function\n"
17336                 "%idval     = OpLoad %u32vec3 %id\n"
17337                 "%x         = OpCompositeExtract %u32 %idval 0\n"
17338                 // Create the input/output type
17339                 "%inloc     = OpInBoundsAccessChain %customptr %indata %zero %x\n"
17340                 "%outloc    = OpInBoundsAccessChain %customptr %outdata %zero %x\n"
17341                 // Read the input value
17342                 "%inval     = OpLoad ${customType} %inloc\n"
17343                 // Create the composite and fill it
17344                 "${compositeConstruct}"
17345                 // Create the struct and fill it with the composite
17346                 "%struct    = OpCompositeConstruct %struct_t${structConstruct}\n"
17347                 // Insert the value
17348                 "%comp_obj  = OpCompositeInsert %struct_t %inval %struct ${insertIndexes}\n"
17349                 // Store the object
17350                 "             OpStore %struct_v %comp_obj\n"
17351                 // Get deepest possible composite pointer
17352                 "%inner_ptr = OpInBoundsAccessChain %composite_p %struct_v${accessChainIndexes}\n"
17353                 "%read_obj  = OpLoad ${compositeType} %inner_ptr\n"
17354                 // Read back the stored value
17355                 "%read_val  = OpCompositeExtract ${customType} %read_obj${extractIndexes}\n"
17356                 "             OpStore %outloc %read_val\n"
17357                 "             OpReturn\n"
17358                 "             OpFunctionEnd\n"
17359         ).specialize(parameters);
17360 }
17361
17362 tcu::TestCaseGroup* createOpInBoundsAccessChainGroup (tcu::TestContext& testCtx)
17363 {
17364         de::MovePtr<tcu::TestCaseGroup> group                   (new tcu::TestCaseGroup(testCtx, "opinboundsaccesschain", "Test the OpInBoundsAccessChain instruction"));
17365         de::Random                                              rnd                             (deStringHash(group->getName()));
17366
17367         for (int type = NUMBERTYPE_INT32; type != NUMBERTYPE_END32; ++type)
17368         {
17369                 NumberType                                              numberType      = NumberType(type);
17370                 const string                                    typeName        = getNumberTypeName(numberType);
17371                 const string                                    description     = "Test the OpInBoundsAccessChain instruction with " + typeName + "s";
17372                 de::MovePtr<tcu::TestCaseGroup> subGroup        (new tcu::TestCaseGroup(testCtx, typeName.c_str(), description.c_str()));
17373
17374                 vector<map<string, string> >    testCases;
17375                 createCompositeCases(testCases, rnd, numberType);
17376
17377                 for (vector<map<string, string> >::const_iterator test = testCases.begin(); test != testCases.end(); ++test)
17378                 {
17379                         ComputeShaderSpec       spec;
17380
17381                         // Number of components inside of a struct
17382                         deUint32 structComponents = rnd.getInt(2, 8);
17383                         // Component index value
17384                         deUint32 structIndex = rnd.getInt(0, structComponents - 1);
17385                         AssemblyStructInfo structInfo(structComponents, structIndex);
17386
17387                         spec.assembly = specializeInBoundsShaderTemplate(numberType, structInfo, *test);
17388
17389                         switch (numberType)
17390                         {
17391                                 case NUMBERTYPE_INT32:
17392                                 {
17393                                         deInt32 number = getInt(rnd);
17394                                         spec.inputs.push_back(createCompositeBuffer<deInt32>(number));
17395                                         spec.outputs.push_back(createCompositeBuffer<deInt32>(number));
17396                                         break;
17397                                 }
17398                                 case NUMBERTYPE_UINT32:
17399                                 {
17400                                         deUint32 number = rnd.getUint32();
17401                                         spec.inputs.push_back(createCompositeBuffer<deUint32>(number));
17402                                         spec.outputs.push_back(createCompositeBuffer<deUint32>(number));
17403                                         break;
17404                                 }
17405                                 case NUMBERTYPE_FLOAT32:
17406                                 {
17407                                         float number = rnd.getFloat();
17408                                         spec.inputs.push_back(createCompositeBuffer<float>(number));
17409                                         spec.outputs.push_back(createCompositeBuffer<float>(number));
17410                                         break;
17411                                 }
17412                                 default:
17413                                         DE_ASSERT(false);
17414                         }
17415                         spec.numWorkGroups = IVec3(1, 1, 1);
17416                         subGroup->addChild(new SpvAsmComputeShaderCase(testCtx, test->at("name").c_str(), "OpInBoundsAccessChain test", spec));
17417                 }
17418                 group->addChild(subGroup.release());
17419         }
17420         return group.release();
17421 }
17422
17423 // If the params missing, uninitialized case
17424 const string specializeDefaultOutputShaderTemplate (const NumberType type, const map<string, string>& params = map<string, string>())
17425 {
17426         map<string, string> parameters(params);
17427
17428         parameters["customType"]        = getAssemblyTypeName(type);
17429
17430         // Declare the const value, and use it in the initializer
17431         if (params.find("constValue") != params.end())
17432         {
17433                 parameters["variableInitializer"]       = " %const";
17434         }
17435         // Uninitialized case
17436         else
17437         {
17438                 parameters["commentDecl"]       = ";";
17439         }
17440
17441         return StringTemplate(
17442                 "OpCapability Shader\n"
17443                 "OpMemoryModel Logical GLSL450\n"
17444                 "OpEntryPoint GLCompute %main \"main\" %id\n"
17445                 "OpExecutionMode %main LocalSize 1 1 1\n"
17446                 "OpSource GLSL 430\n"
17447                 "OpName %main           \"main\"\n"
17448                 "OpName %id             \"gl_GlobalInvocationID\"\n"
17449                 // Decorators
17450                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
17451                 "OpDecorate %indata DescriptorSet 0\n"
17452                 "OpDecorate %indata Binding 0\n"
17453                 "OpDecorate %outdata DescriptorSet 0\n"
17454                 "OpDecorate %outdata Binding 1\n"
17455                 "OpDecorate %in_arr ArrayStride 4\n"
17456                 "OpDecorate %in_buf BufferBlock\n"
17457                 "OpMemberDecorate %in_buf 0 Offset 0\n"
17458                 // Base types
17459                 "%void       = OpTypeVoid\n"
17460                 "%voidf      = OpTypeFunction %void\n"
17461                 "%u32        = OpTypeInt 32 0\n"
17462                 "%i32        = OpTypeInt 32 1\n"
17463                 "%f32        = OpTypeFloat 32\n"
17464                 "%uvec3      = OpTypeVector %u32 3\n"
17465                 "%uvec3ptr   = OpTypePointer Input %uvec3\n"
17466                 "${commentDecl:opt}%const      = OpConstant ${customType} ${constValue:opt}\n"
17467                 // Derived types
17468                 "%in_ptr     = OpTypePointer Uniform ${customType}\n"
17469                 "%in_arr     = OpTypeRuntimeArray ${customType}\n"
17470                 "%in_buf     = OpTypeStruct %in_arr\n"
17471                 "%in_bufptr  = OpTypePointer Uniform %in_buf\n"
17472                 "%indata     = OpVariable %in_bufptr Uniform\n"
17473                 "%outdata    = OpVariable %in_bufptr Uniform\n"
17474                 "%id         = OpVariable %uvec3ptr Input\n"
17475                 "%var_ptr    = OpTypePointer Function ${customType}\n"
17476                 // Constants
17477                 "%zero       = OpConstant %i32 0\n"
17478                 // Main function
17479                 "%main       = OpFunction %void None %voidf\n"
17480                 "%label      = OpLabel\n"
17481                 "%out_var    = OpVariable %var_ptr Function${variableInitializer:opt}\n"
17482                 "%idval      = OpLoad %uvec3 %id\n"
17483                 "%x          = OpCompositeExtract %u32 %idval 0\n"
17484                 "%inloc      = OpAccessChain %in_ptr %indata %zero %x\n"
17485                 "%outloc     = OpAccessChain %in_ptr %outdata %zero %x\n"
17486
17487                 "%outval     = OpLoad ${customType} %out_var\n"
17488                 "              OpStore %outloc %outval\n"
17489                 "              OpReturn\n"
17490                 "              OpFunctionEnd\n"
17491         ).specialize(parameters);
17492 }
17493
17494 bool compareFloats (const std::vector<Resource>&, const vector<AllocationSp>& outputAllocs, const std::vector<Resource>& expectedOutputs, TestLog& log)
17495 {
17496         DE_ASSERT(outputAllocs.size() != 0);
17497         DE_ASSERT(outputAllocs.size() == expectedOutputs.size());
17498
17499         // Use custom epsilon because of the float->string conversion
17500         const float     epsilon = 0.00001f;
17501
17502         for (size_t outputNdx = 0; outputNdx < outputAllocs.size(); ++outputNdx)
17503         {
17504                 vector<deUint8> expectedBytes;
17505                 float                   expected;
17506                 float                   actual;
17507
17508                 expectedOutputs[outputNdx].getBytes(expectedBytes);
17509                 memcpy(&expected, &expectedBytes.front(), expectedBytes.size());
17510                 memcpy(&actual, outputAllocs[outputNdx]->getHostPtr(), expectedBytes.size());
17511
17512                 // Test with epsilon
17513                 if (fabs(expected - actual) > epsilon)
17514                 {
17515                         log << TestLog::Message << "Error: The actual and expected values not matching."
17516                                 << " Expected: " << expected << " Actual: " << actual << " Epsilon: " << epsilon << TestLog::EndMessage;
17517                         return false;
17518                 }
17519         }
17520         return true;
17521 }
17522
17523 // Checks if the driver crash with uninitialized cases
17524 bool passthruVerify (const std::vector<Resource>&, const vector<AllocationSp>& outputAllocs, const std::vector<Resource>& expectedOutputs, TestLog&)
17525 {
17526         DE_ASSERT(outputAllocs.size() != 0);
17527         DE_ASSERT(outputAllocs.size() == expectedOutputs.size());
17528
17529         // Copy and discard the result.
17530         for (size_t outputNdx = 0; outputNdx < outputAllocs.size(); ++outputNdx)
17531         {
17532                 vector<deUint8> expectedBytes;
17533                 expectedOutputs[outputNdx].getBytes(expectedBytes);
17534
17535                 const size_t    width                   = expectedBytes.size();
17536                 vector<char>    data                    (width);
17537
17538                 memcpy(&data[0], outputAllocs[outputNdx]->getHostPtr(), width);
17539         }
17540         return true;
17541 }
17542
17543 tcu::TestCaseGroup* createShaderDefaultOutputGroup (tcu::TestContext& testCtx)
17544 {
17545         de::MovePtr<tcu::TestCaseGroup> group   (new tcu::TestCaseGroup(testCtx, "shader_default_output", "Test shader default output."));
17546         de::Random                                              rnd             (deStringHash(group->getName()));
17547
17548         for (int type = NUMBERTYPE_INT32; type != NUMBERTYPE_END32; ++type)
17549         {
17550                 NumberType                                              numberType      = NumberType(type);
17551                 const string                                    typeName        = getNumberTypeName(numberType);
17552                 const string                                    description     = "Test the OpVariable initializer with " + typeName + ".";
17553                 de::MovePtr<tcu::TestCaseGroup> subGroup        (new tcu::TestCaseGroup(testCtx, typeName.c_str(), description.c_str()));
17554
17555                 // 2 similar subcases (initialized and uninitialized)
17556                 for (int subCase = 0; subCase < 2; ++subCase)
17557                 {
17558                         ComputeShaderSpec spec;
17559                         spec.numWorkGroups = IVec3(1, 1, 1);
17560
17561                         map<string, string>                             params;
17562
17563                         switch (numberType)
17564                         {
17565                                 case NUMBERTYPE_INT32:
17566                                 {
17567                                         deInt32 number = getInt(rnd);
17568                                         spec.inputs.push_back(createCompositeBuffer<deInt32>(number));
17569                                         spec.outputs.push_back(createCompositeBuffer<deInt32>(number));
17570                                         params["constValue"] = numberToString(number);
17571                                         break;
17572                                 }
17573                                 case NUMBERTYPE_UINT32:
17574                                 {
17575                                         deUint32 number = rnd.getUint32();
17576                                         spec.inputs.push_back(createCompositeBuffer<deUint32>(number));
17577                                         spec.outputs.push_back(createCompositeBuffer<deUint32>(number));
17578                                         params["constValue"] = numberToString(number);
17579                                         break;
17580                                 }
17581                                 case NUMBERTYPE_FLOAT32:
17582                                 {
17583                                         float number = rnd.getFloat();
17584                                         spec.inputs.push_back(createCompositeBuffer<float>(number));
17585                                         spec.outputs.push_back(createCompositeBuffer<float>(number));
17586                                         spec.verifyIO = &compareFloats;
17587                                         params["constValue"] = numberToString(number);
17588                                         break;
17589                                 }
17590                                 default:
17591                                         DE_ASSERT(false);
17592                         }
17593
17594                         // Initialized subcase
17595                         if (!subCase)
17596                         {
17597                                 spec.assembly = specializeDefaultOutputShaderTemplate(numberType, params);
17598                                 subGroup->addChild(new SpvAsmComputeShaderCase(testCtx, "initialized", "OpVariable initializer tests.", spec));
17599                         }
17600                         // Uninitialized subcase
17601                         else
17602                         {
17603                                 spec.assembly = specializeDefaultOutputShaderTemplate(numberType);
17604                                 spec.verifyIO = &passthruVerify;
17605                                 subGroup->addChild(new SpvAsmComputeShaderCase(testCtx, "uninitialized", "OpVariable initializer tests.", spec));
17606                         }
17607                 }
17608                 group->addChild(subGroup.release());
17609         }
17610         return group.release();
17611 }
17612
17613 tcu::TestCaseGroup* createOpNopTests (tcu::TestContext& testCtx)
17614 {
17615         de::MovePtr<tcu::TestCaseGroup> testGroup (new tcu::TestCaseGroup(testCtx, "opnop", "Test OpNop"));
17616         RGBA                                                    defaultColors[4];
17617         map<string, string>                             opNopFragments;
17618
17619         getDefaultColors(defaultColors);
17620
17621         opNopFragments["testfun"]               =
17622                 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
17623                 "%param1 = OpFunctionParameter %v4f32\n"
17624                 "%label_testfun = OpLabel\n"
17625                 "OpNop\n"
17626                 "OpNop\n"
17627                 "OpNop\n"
17628                 "OpNop\n"
17629                 "OpNop\n"
17630                 "OpNop\n"
17631                 "OpNop\n"
17632                 "OpNop\n"
17633                 "%a = OpVectorExtractDynamic %f32 %param1 %c_i32_0\n"
17634                 "%b = OpFAdd %f32 %a %a\n"
17635                 "OpNop\n"
17636                 "%c = OpFSub %f32 %b %a\n"
17637                 "%ret = OpVectorInsertDynamic %v4f32 %param1 %c %c_i32_0\n"
17638                 "OpNop\n"
17639                 "OpNop\n"
17640                 "OpReturnValue %ret\n"
17641                 "OpFunctionEnd\n";
17642
17643         createTestsForAllStages("opnop", defaultColors, defaultColors, opNopFragments, testGroup.get());
17644
17645         return testGroup.release();
17646 }
17647
17648 tcu::TestCaseGroup* createOpNameTests (tcu::TestContext& testCtx)
17649 {
17650         de::MovePtr<tcu::TestCaseGroup> testGroup       (new tcu::TestCaseGroup(testCtx, "opname","Test OpName"));
17651         RGBA                                                    defaultColors[4];
17652         map<string, string>                             opNameFragments;
17653
17654         getDefaultColors(defaultColors);
17655
17656         opNameFragments["testfun"] =
17657                 "%test_code  = OpFunction %v4f32 None %v4f32_v4f32_function\n"
17658                 "%param1     = OpFunctionParameter %v4f32\n"
17659                 "%label_func = OpLabel\n"
17660                 "%a          = OpVectorExtractDynamic %f32 %param1 %c_i32_0\n"
17661                 "%b          = OpFAdd %f32 %a %a\n"
17662                 "%c          = OpFSub %f32 %b %a\n"
17663                 "%ret        = OpVectorInsertDynamic %v4f32 %param1 %c %c_i32_0\n"
17664                 "OpReturnValue %ret\n"
17665                 "OpFunctionEnd\n";
17666
17667         opNameFragments["debug"] =
17668                 "OpName %BP_main \"not_main\"";
17669
17670         createTestsForAllStages("opname", defaultColors, defaultColors, opNameFragments, testGroup.get());
17671
17672         return testGroup.release();
17673 }
17674
17675 tcu::TestCaseGroup* createFloat16Tests (tcu::TestContext& testCtx)
17676 {
17677         de::MovePtr<tcu::TestCaseGroup>         testGroup                       (new tcu::TestCaseGroup(testCtx, "float16", "Float 16 tests"));
17678
17679         testGroup->addChild(createOpConstantFloat16Tests(testCtx));
17680         testGroup->addChild(createFloat16LogicalSet<GraphicsResources>(testCtx, TEST_WITH_NAN));
17681         testGroup->addChild(createFloat16LogicalSet<GraphicsResources>(testCtx, TEST_WITHOUT_NAN));
17682         testGroup->addChild(createFloat16FuncSet<GraphicsResources>(testCtx));
17683         testGroup->addChild(createDerivativeTests<256, 1>(testCtx));
17684         testGroup->addChild(createDerivativeTests<256, 2>(testCtx));
17685         testGroup->addChild(createDerivativeTests<256, 4>(testCtx));
17686         testGroup->addChild(createFloat16VectorExtractSet<GraphicsResources>(testCtx));
17687         testGroup->addChild(createFloat16VectorInsertSet<GraphicsResources>(testCtx));
17688         testGroup->addChild(createFloat16VectorShuffleSet<GraphicsResources>(testCtx));
17689         testGroup->addChild(createFloat16CompositeConstructSet<GraphicsResources>(testCtx));
17690         testGroup->addChild(createFloat16CompositeInsertExtractSet<GraphicsResources>(testCtx, "OpCompositeExtract"));
17691         testGroup->addChild(createFloat16CompositeInsertExtractSet<GraphicsResources>(testCtx, "OpCompositeInsert"));
17692         testGroup->addChild(createFloat16ArithmeticSet<GraphicsResources>(testCtx));
17693         testGroup->addChild(createFloat16ArithmeticSet<1, GraphicsResources>(testCtx));
17694         testGroup->addChild(createFloat16ArithmeticSet<2, GraphicsResources>(testCtx));
17695         testGroup->addChild(createFloat16ArithmeticSet<3, GraphicsResources>(testCtx));
17696         testGroup->addChild(createFloat16ArithmeticSet<4, GraphicsResources>(testCtx));
17697
17698         return testGroup.release();
17699 }
17700
17701 tcu::TestCaseGroup* createFloat16Group (tcu::TestContext& testCtx)
17702 {
17703         de::MovePtr<tcu::TestCaseGroup>         testGroup                       (new tcu::TestCaseGroup(testCtx, "float16", "Float 16 tests"));
17704
17705         testGroup->addChild(createFloat16OpConstantCompositeGroup(testCtx));
17706         testGroup->addChild(createFloat16LogicalSet<ComputeShaderSpec>(testCtx, TEST_WITH_NAN));
17707         testGroup->addChild(createFloat16LogicalSet<ComputeShaderSpec>(testCtx, TEST_WITHOUT_NAN));
17708         testGroup->addChild(createFloat16FuncSet<ComputeShaderSpec>(testCtx));
17709         testGroup->addChild(createFloat16VectorExtractSet<ComputeShaderSpec>(testCtx));
17710         testGroup->addChild(createFloat16VectorInsertSet<ComputeShaderSpec>(testCtx));
17711         testGroup->addChild(createFloat16VectorShuffleSet<ComputeShaderSpec>(testCtx));
17712         testGroup->addChild(createFloat16CompositeConstructSet<ComputeShaderSpec>(testCtx));
17713         testGroup->addChild(createFloat16CompositeInsertExtractSet<ComputeShaderSpec>(testCtx, "OpCompositeExtract"));
17714         testGroup->addChild(createFloat16CompositeInsertExtractSet<ComputeShaderSpec>(testCtx, "OpCompositeInsert"));
17715         testGroup->addChild(createFloat16ArithmeticSet<ComputeShaderSpec>(testCtx));
17716         testGroup->addChild(createFloat16ArithmeticSet<1, ComputeShaderSpec>(testCtx));
17717         testGroup->addChild(createFloat16ArithmeticSet<2, ComputeShaderSpec>(testCtx));
17718         testGroup->addChild(createFloat16ArithmeticSet<3, ComputeShaderSpec>(testCtx));
17719         testGroup->addChild(createFloat16ArithmeticSet<4, ComputeShaderSpec>(testCtx));
17720
17721         return testGroup.release();
17722 }
17723
17724 tcu::TestCaseGroup* createBoolMixedBitSizeGroup (tcu::TestContext& testCtx)
17725 {
17726         de::MovePtr<tcu::TestCaseGroup> group                   (new tcu::TestCaseGroup(testCtx, "mixed_bitsize", "Tests boolean operands produced from instructions of different bit-sizes"));
17727
17728         de::Random                                              rnd                             (deStringHash(group->getName()));
17729         const int               numElements             = 100;
17730         vector<float>   inputData               (numElements, 0);
17731         vector<float>   outputData              (numElements, 0);
17732         fillRandomScalars(rnd, 0.0f, 100.0f, &inputData[0], 100);
17733
17734         const StringTemplate                    shaderTemplate  (
17735                 "${CAPS}\n"
17736                 "OpMemoryModel Logical GLSL450\n"
17737                 "OpEntryPoint GLCompute %main \"main\" %id\n"
17738                 "OpExecutionMode %main LocalSize 1 1 1\n"
17739                 "OpSource GLSL 430\n"
17740                 "OpName %main           \"main\"\n"
17741                 "OpName %id             \"gl_GlobalInvocationID\"\n"
17742
17743                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
17744
17745                 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) + string(getComputeAsmInputOutputBuffer()) +
17746
17747                 "%id        = OpVariable %uvec3ptr Input\n"
17748                 "${CONST}\n"
17749                 "%main      = OpFunction %void None %voidf\n"
17750                 "%label     = OpLabel\n"
17751                 "%idval     = OpLoad %uvec3 %id\n"
17752                 "%x         = OpCompositeExtract %u32 %idval 0\n"
17753                 "%inloc     = OpAccessChain %f32ptr %indata %c0i32 %x\n"
17754
17755                 "${TEST}\n"
17756
17757                 "%outloc    = OpAccessChain %f32ptr %outdata %c0i32 %x\n"
17758                 "             OpStore %outloc %res\n"
17759                 "             OpReturn\n"
17760                 "             OpFunctionEnd\n"
17761         );
17762
17763         // Each test case produces 4 boolean values, and we want each of these values
17764         // to come froma different combination of the available bit-sizes, so compute
17765         // all possible combinations here.
17766         vector<deUint32>        widths;
17767         widths.push_back(32);
17768         widths.push_back(16);
17769         widths.push_back(8);
17770
17771         vector<IVec4>   cases;
17772         for (size_t width0 = 0; width0 < widths.size(); width0++)
17773         {
17774                 for (size_t width1 = 0; width1 < widths.size(); width1++)
17775                 {
17776                         for (size_t width2 = 0; width2 < widths.size(); width2++)
17777                         {
17778                                 for (size_t width3 = 0; width3 < widths.size(); width3++)
17779                                 {
17780                                         cases.push_back(IVec4(widths[width0], widths[width1], widths[width2], widths[width3]));
17781                                 }
17782                         }
17783                 }
17784         }
17785
17786         for (size_t caseNdx = 0; caseNdx < cases.size(); caseNdx++)
17787         {
17788                 /// Skip cases where all bitsizes are the same, we are only interested in testing booleans produced from instructions with different native bit-sizes
17789                 if (cases[caseNdx][0] == cases[caseNdx][1] && cases[caseNdx][0] == cases[caseNdx][2] && cases[caseNdx][0] == cases[caseNdx][3])
17790                         continue;
17791
17792                 map<string, string>     specializations;
17793                 ComputeShaderSpec       spec;
17794
17795                 // Inject appropriate capabilities and reference constants depending
17796                 // on the bit-sizes required by this test case
17797                 bool hasFloat32 = cases[caseNdx][0] == 32 || cases[caseNdx][1] == 32 || cases[caseNdx][2] == 32 || cases[caseNdx][3] == 32;
17798                 bool hasFloat16 = cases[caseNdx][0] == 16 || cases[caseNdx][1] == 16 || cases[caseNdx][2] == 16 || cases[caseNdx][3] == 16;
17799                 bool hasInt8    = cases[caseNdx][0] == 8 || cases[caseNdx][1] == 8 || cases[caseNdx][2] == 8 || cases[caseNdx][3] == 8;
17800
17801                 string capsStr  = "OpCapability Shader\n";
17802                 string constStr =
17803                         "%c0i32     = OpConstant %i32 0\n"
17804                         "%c1f32     = OpConstant %f32 1.0\n"
17805                         "%c0f32     = OpConstant %f32 0.0\n";
17806
17807                 if (hasFloat32)
17808                 {
17809                         constStr        +=
17810                                 "%c10f32    = OpConstant %f32 10.0\n"
17811                                 "%c25f32    = OpConstant %f32 25.0\n"
17812                                 "%c50f32    = OpConstant %f32 50.0\n"
17813                                 "%c90f32    = OpConstant %f32 90.0\n";
17814                 }
17815
17816                 if (hasFloat16)
17817                 {
17818                         capsStr         += "OpCapability Float16\n";
17819                         constStr        +=
17820                                 "%f16       = OpTypeFloat 16\n"
17821                                 "%c10f16    = OpConstant %f16 10.0\n"
17822                                 "%c25f16    = OpConstant %f16 25.0\n"
17823                                 "%c50f16    = OpConstant %f16 50.0\n"
17824                                 "%c90f16    = OpConstant %f16 90.0\n";
17825                 }
17826
17827                 if (hasInt8)
17828                 {
17829                         capsStr         += "OpCapability Int8\n";
17830                         constStr        +=
17831                                 "%i8        = OpTypeInt 8 1\n"
17832                                 "%c10i8     = OpConstant %i8 10\n"
17833                                 "%c25i8     = OpConstant %i8 25\n"
17834                                 "%c50i8     = OpConstant %i8 50\n"
17835                                 "%c90i8     = OpConstant %i8 90\n";
17836                 }
17837
17838                 // Each invocation reads a different float32 value as input. Depending on
17839                 // the bit-sizes required by the particular test case, we also produce
17840                 // float16 and/or and int8 values by converting from the 32-bit float.
17841                 string testStr  = "";
17842                 testStr                 += "%inval32   = OpLoad %f32 %inloc\n";
17843                 if (hasFloat16)
17844                         testStr         += "%inval16   = OpFConvert %f16 %inval32\n";
17845                 if (hasInt8)
17846                         testStr         += "%inval8    = OpConvertFToS %i8 %inval32\n";
17847
17848                 // Because conversions from Float to Int round towards 0 we want our "greater" comparisons to be >=,
17849                 // that way a float32/float16 comparison such as 50.6f >= 50.0f will preserve its result
17850                 // when converted to int8, since FtoS(50.6f) results in 50. For "less" comparisons, it is the
17851                 // other way around, so in this case we want < instead of <=.
17852                 if (cases[caseNdx][0] == 32)
17853                         testStr         += "%cmp1      = OpFOrdGreaterThanEqual %bool %inval32 %c25f32\n";
17854                 else if (cases[caseNdx][0] == 16)
17855                         testStr         += "%cmp1      = OpFOrdGreaterThanEqual %bool %inval16 %c25f16\n";
17856                 else
17857                         testStr         += "%cmp1      = OpSGreaterThanEqual %bool %inval8 %c25i8\n";
17858
17859                 if (cases[caseNdx][1] == 32)
17860                         testStr         += "%cmp2      = OpFOrdLessThan %bool %inval32 %c50f32\n";
17861                 else if (cases[caseNdx][1] == 16)
17862                         testStr         += "%cmp2      = OpFOrdLessThan %bool %inval16 %c50f16\n";
17863                 else
17864                         testStr         += "%cmp2      = OpSLessThan %bool %inval8 %c50i8\n";
17865
17866                 if (cases[caseNdx][2] == 32)
17867                         testStr         += "%cmp3      = OpFOrdLessThan %bool %inval32 %c10f32\n";
17868                 else if (cases[caseNdx][2] == 16)
17869                         testStr         += "%cmp3      = OpFOrdLessThan %bool %inval16 %c10f16\n";
17870                 else
17871                         testStr         += "%cmp3      = OpSLessThan %bool %inval8 %c10i8\n";
17872
17873                 if (cases[caseNdx][3] == 32)
17874                         testStr         += "%cmp4      = OpFOrdGreaterThanEqual %bool %inval32 %c90f32\n";
17875                 else if (cases[caseNdx][3] == 16)
17876                         testStr         += "%cmp4      = OpFOrdGreaterThanEqual %bool %inval16 %c90f16\n";
17877                 else
17878                         testStr         += "%cmp4      = OpSGreaterThanEqual %bool %inval8 %c90i8\n";
17879
17880                 testStr                 += "%and1      = OpLogicalAnd %bool %cmp1 %cmp2\n";
17881                 testStr                 += "%or1       = OpLogicalOr %bool %cmp3 %cmp4\n";
17882                 testStr                 += "%or2       = OpLogicalOr %bool %and1 %or1\n";
17883                 testStr                 += "%not1      = OpLogicalNot %bool %or2\n";
17884                 testStr                 += "%res       = OpSelect %f32 %not1 %c1f32 %c0f32\n";
17885
17886                 specializations["CAPS"]         = capsStr;
17887                 specializations["CONST"]        = constStr;
17888                 specializations["TEST"]         = testStr;
17889
17890                 // Compute expected result by evaluating the boolean expression computed in the shader for each input value
17891                 for (size_t ndx = 0; ndx < numElements; ++ndx)
17892                         outputData[ndx] = !((inputData[ndx] >= 25.0f && inputData[ndx] < 50.0f) || (inputData[ndx] < 10.0f || inputData[ndx] >= 90.0f));
17893
17894                 spec.assembly = shaderTemplate.specialize(specializations);
17895                 spec.inputs.push_back(BufferSp(new Float32Buffer(inputData)));
17896                 spec.outputs.push_back(BufferSp(new Float32Buffer(outputData)));
17897                 spec.numWorkGroups = IVec3(numElements, 1, 1);
17898                 if (hasFloat16)
17899                         spec.requestedVulkanFeatures.extFloat16Int8 |= EXTFLOAT16INT8FEATURES_FLOAT16;
17900                 if (hasInt8)
17901                         spec.requestedVulkanFeatures.extFloat16Int8 |= EXTFLOAT16INT8FEATURES_INT8;
17902                 spec.extensions.push_back("VK_KHR_shader_float16_int8");
17903
17904                 string testName = "b" + de::toString(cases[caseNdx][0]) + "b" + de::toString(cases[caseNdx][1]) + "b" + de::toString(cases[caseNdx][2]) + "b" + de::toString(cases[caseNdx][3]);
17905                 group->addChild(new SpvAsmComputeShaderCase(testCtx, testName.c_str(), "", spec));
17906         }
17907
17908         return group.release();
17909 }
17910
17911 tcu::TestCaseGroup* createBoolGroup (tcu::TestContext& testCtx)
17912 {
17913         de::MovePtr<tcu::TestCaseGroup>         testGroup                       (new tcu::TestCaseGroup(testCtx, "bool", "Boolean tests"));
17914
17915         testGroup->addChild(createBoolMixedBitSizeGroup(testCtx));
17916
17917         return testGroup.release();
17918 }
17919
17920 tcu::TestCaseGroup* createOpNameAbuseTests (tcu::TestContext& testCtx)
17921 {
17922         de::MovePtr<tcu::TestCaseGroup> abuseGroup(new tcu::TestCaseGroup(testCtx, "opname_abuse", "OpName abuse tests"));
17923         vector<CaseParameter>                   abuseCases;
17924         RGBA                                                    defaultColors[4];
17925         map<string, string>                             opNameFragments;
17926
17927         getOpNameAbuseCases(abuseCases);
17928         getDefaultColors(defaultColors);
17929
17930         opNameFragments["testfun"] =
17931                 "%test_code  = OpFunction %v4f32 None %v4f32_v4f32_function\n"
17932                 "%param1     = OpFunctionParameter %v4f32\n"
17933                 "%label_func = OpLabel\n"
17934                 "%a          = OpVectorExtractDynamic %f32 %param1 %c_i32_0\n"
17935                 "%b          = OpFAdd %f32 %a %a\n"
17936                 "%c          = OpFSub %f32 %b %a\n"
17937                 "%ret        = OpVectorInsertDynamic %v4f32 %param1 %c %c_i32_0\n"
17938                 "OpReturnValue %ret\n"
17939                 "OpFunctionEnd\n";
17940
17941         for (unsigned int i = 0; i < abuseCases.size(); i++)
17942         {
17943                 string casename;
17944                 casename = string("main") + abuseCases[i].name;
17945
17946                 opNameFragments["debug"] =
17947                         "OpName %BP_main \"" + abuseCases[i].param + "\"";
17948
17949                 createTestsForAllStages(casename, defaultColors, defaultColors, opNameFragments, abuseGroup.get());
17950         }
17951
17952         for (unsigned int i = 0; i < abuseCases.size(); i++)
17953         {
17954                 string casename;
17955                 casename = string("b") + abuseCases[i].name;
17956
17957                 opNameFragments["debug"] =
17958                         "OpName %b \"" + abuseCases[i].param + "\"";
17959
17960                 createTestsForAllStages(casename, defaultColors, defaultColors, opNameFragments, abuseGroup.get());
17961         }
17962
17963         {
17964                 opNameFragments["debug"] =
17965                         "OpName %test_code \"name1\"\n"
17966                         "OpName %param1    \"name2\"\n"
17967                         "OpName %a         \"name3\"\n"
17968                         "OpName %b         \"name4\"\n"
17969                         "OpName %c         \"name5\"\n"
17970                         "OpName %ret       \"name6\"\n";
17971
17972                 createTestsForAllStages("everything_named", defaultColors, defaultColors, opNameFragments, abuseGroup.get());
17973         }
17974
17975         {
17976                 opNameFragments["debug"] =
17977                         "OpName %test_code \"the_same\"\n"
17978                         "OpName %param1    \"the_same\"\n"
17979                         "OpName %a         \"the_same\"\n"
17980                         "OpName %b         \"the_same\"\n"
17981                         "OpName %c         \"the_same\"\n"
17982                         "OpName %ret       \"the_same\"\n";
17983
17984                 createTestsForAllStages("everything_named_the_same", defaultColors, defaultColors, opNameFragments, abuseGroup.get());
17985         }
17986
17987         {
17988                 opNameFragments["debug"] =
17989                         "OpName %BP_main \"to_be\"\n"
17990                         "OpName %BP_main \"or_not\"\n"
17991                         "OpName %BP_main \"to_be\"\n";
17992
17993                 createTestsForAllStages("main_has_multiple_names", defaultColors, defaultColors, opNameFragments, abuseGroup.get());
17994         }
17995
17996         {
17997                 opNameFragments["debug"] =
17998                         "OpName %b \"to_be\"\n"
17999                         "OpName %b \"or_not\"\n"
18000                         "OpName %b \"to_be\"\n";
18001
18002                 createTestsForAllStages("b_has_multiple_names", defaultColors, defaultColors, opNameFragments, abuseGroup.get());
18003         }
18004
18005         return abuseGroup.release();
18006 }
18007
18008
18009 tcu::TestCaseGroup* createOpMemberNameAbuseTests (tcu::TestContext& testCtx)
18010 {
18011         de::MovePtr<tcu::TestCaseGroup> abuseGroup(new tcu::TestCaseGroup(testCtx, "opmembername_abuse", "OpName abuse tests"));
18012         vector<CaseParameter>                   abuseCases;
18013         RGBA                                                    defaultColors[4];
18014         map<string, string>                             opMemberNameFragments;
18015
18016         getOpNameAbuseCases(abuseCases);
18017         getDefaultColors(defaultColors);
18018
18019         opMemberNameFragments["pre_main"] =
18020                 "%f3str = OpTypeStruct %f32 %f32 %f32\n";
18021
18022         opMemberNameFragments["testfun"] =
18023                 "%test_code  = OpFunction %v4f32 None %v4f32_v4f32_function\n"
18024                 "%param1     = OpFunctionParameter %v4f32\n"
18025                 "%label_func = OpLabel\n"
18026                 "%a          = OpVectorExtractDynamic %f32 %param1 %c_i32_0\n"
18027                 "%b          = OpFAdd %f32 %a %a\n"
18028                 "%c          = OpFSub %f32 %b %a\n"
18029                 "%cstr       = OpCompositeConstruct %f3str %c %c %c\n"
18030                 "%d          = OpCompositeExtract %f32 %cstr 0\n"
18031                 "%ret        = OpVectorInsertDynamic %v4f32 %param1 %d %c_i32_0\n"
18032                 "OpReturnValue %ret\n"
18033                 "OpFunctionEnd\n";
18034
18035         for (unsigned int i = 0; i < abuseCases.size(); i++)
18036         {
18037                 string casename;
18038                 casename = string("f3str_x") + abuseCases[i].name;
18039
18040                 opMemberNameFragments["debug"] =
18041                         "OpMemberName %f3str 0 \"" + abuseCases[i].param + "\"";
18042
18043                 createTestsForAllStages(casename, defaultColors, defaultColors, opMemberNameFragments, abuseGroup.get());
18044         }
18045
18046         {
18047                 opMemberNameFragments["debug"] =
18048                         "OpMemberName %f3str 0 \"name1\"\n"
18049                         "OpMemberName %f3str 1 \"name2\"\n"
18050                         "OpMemberName %f3str 2 \"name3\"\n";
18051
18052                 createTestsForAllStages("everything_named", defaultColors, defaultColors, opMemberNameFragments, abuseGroup.get());
18053         }
18054
18055         {
18056                 opMemberNameFragments["debug"] =
18057                         "OpMemberName %f3str 0 \"the_same\"\n"
18058                         "OpMemberName %f3str 1 \"the_same\"\n"
18059                         "OpMemberName %f3str 2 \"the_same\"\n";
18060
18061                 createTestsForAllStages("everything_named_the_same", defaultColors, defaultColors, opMemberNameFragments, abuseGroup.get());
18062         }
18063
18064         {
18065                 opMemberNameFragments["debug"] =
18066                         "OpMemberName %f3str 0 \"to_be\"\n"
18067                         "OpMemberName %f3str 1 \"or_not\"\n"
18068                         "OpMemberName %f3str 0 \"to_be\"\n"
18069                         "OpMemberName %f3str 2 \"makes_no\"\n"
18070                         "OpMemberName %f3str 0 \"difference\"\n"
18071                         "OpMemberName %f3str 0 \"to_me\"\n";
18072
18073
18074                 createTestsForAllStages("f3str_x_has_multiple_names", defaultColors, defaultColors, opMemberNameFragments, abuseGroup.get());
18075         }
18076
18077         return abuseGroup.release();
18078 }
18079
18080 vector<deUint32> getSparseIdsAbuseData (const deUint32 numDataPoints, const deUint32 seed)
18081 {
18082         vector<deUint32>        result;
18083         de::Random                      rnd             (seed);
18084
18085         result.reserve(numDataPoints);
18086
18087         for (deUint32 dataPointNdx = 0; dataPointNdx < numDataPoints; ++dataPointNdx)
18088                 result.push_back(rnd.getUint32());
18089
18090         return result;
18091 }
18092
18093 vector<deUint32> getSparseIdsAbuseResults (const vector<deUint32>& inData1, const vector<deUint32>& inData2)
18094 {
18095         vector<deUint32>        result;
18096
18097         result.reserve(inData1.size());
18098
18099         for (size_t dataPointNdx = 0; dataPointNdx < inData1.size(); ++dataPointNdx)
18100                 result.push_back(inData1[dataPointNdx] + inData2[dataPointNdx]);
18101
18102         return result;
18103 }
18104
18105 template<class SpecResource>
18106 void createSparseIdsAbuseTest (tcu::TestContext& testCtx, de::MovePtr<tcu::TestCaseGroup>& testGroup)
18107 {
18108         const deUint32                  numDataPoints   = 16;
18109         const std::string               testName                ("sparse_ids");
18110         const deUint32                  seed                    (deStringHash(testName.c_str()));
18111         const vector<deUint32>  inData1                 (getSparseIdsAbuseData(numDataPoints, seed + 1));
18112         const vector<deUint32>  inData2                 (getSparseIdsAbuseData(numDataPoints, seed + 2));
18113         const vector<deUint32>  outData                 (getSparseIdsAbuseResults(inData1, inData2));
18114         const StringTemplate    preMain
18115         (
18116                 "%c_i32_ndp = OpConstant %i32 ${num_data_points}\n"
18117                 "   %up_u32 = OpTypePointer Uniform %u32\n"
18118                 "   %ra_u32 = OpTypeArray %u32 %c_i32_ndp\n"
18119                 "   %SSBO32 = OpTypeStruct %ra_u32\n"
18120                 "%up_SSBO32 = OpTypePointer Uniform %SSBO32\n"
18121                 "%ssbo_src0 = OpVariable %up_SSBO32 Uniform\n"
18122                 "%ssbo_src1 = OpVariable %up_SSBO32 Uniform\n"
18123                 " %ssbo_dst = OpVariable %up_SSBO32 Uniform\n"
18124         );
18125         const StringTemplate    decoration
18126         (
18127                 "OpDecorate %ra_u32 ArrayStride 4\n"
18128                 "OpMemberDecorate %SSBO32 0 Offset 0\n"
18129                 "OpDecorate %SSBO32 BufferBlock\n"
18130                 "OpDecorate %ssbo_src0 DescriptorSet 0\n"
18131                 "OpDecorate %ssbo_src0 Binding 0\n"
18132                 "OpDecorate %ssbo_src1 DescriptorSet 0\n"
18133                 "OpDecorate %ssbo_src1 Binding 1\n"
18134                 "OpDecorate %ssbo_dst DescriptorSet 0\n"
18135                 "OpDecorate %ssbo_dst Binding 2\n"
18136         );
18137         const StringTemplate    testFun
18138         (
18139                 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
18140                 "    %param = OpFunctionParameter %v4f32\n"
18141
18142                 "    %entry = OpLabel\n"
18143                 "        %i = OpVariable %fp_i32 Function\n"
18144                 "             OpStore %i %c_i32_0\n"
18145                 "             OpBranch %loop\n"
18146
18147                 "     %loop = OpLabel\n"
18148                 "    %i_cmp = OpLoad %i32 %i\n"
18149                 "       %lt = OpSLessThan %bool %i_cmp %c_i32_ndp\n"
18150                 "             OpLoopMerge %merge %next None\n"
18151                 "             OpBranchConditional %lt %write %merge\n"
18152
18153                 "    %write = OpLabel\n"
18154                 "      %ndx = OpLoad %i32 %i\n"
18155
18156                 "      %127 = OpAccessChain %up_u32 %ssbo_src0 %c_i32_0 %ndx\n"
18157                 "      %128 = OpLoad %u32 %127\n"
18158
18159                 // The test relies on SPIR-V compiler option SPV_TEXT_TO_BINARY_OPTION_PRESERVE_NUMERIC_IDS set in assembleSpirV()
18160                 "  %4194000 = OpAccessChain %up_u32 %ssbo_src1 %c_i32_0 %ndx\n"
18161                 "  %4194001 = OpLoad %u32 %4194000\n"
18162
18163                 "  %2097151 = OpIAdd %u32 %128 %4194001\n"
18164                 "  %2097152 = OpAccessChain %up_u32 %ssbo_dst %c_i32_0 %ndx\n"
18165                 "             OpStore %2097152 %2097151\n"
18166                 "             OpBranch %next\n"
18167
18168                 "     %next = OpLabel\n"
18169                 "    %i_cur = OpLoad %i32 %i\n"
18170                 "    %i_new = OpIAdd %i32 %i_cur %c_i32_1\n"
18171                 "             OpStore %i %i_new\n"
18172                 "             OpBranch %loop\n"
18173
18174                 "    %merge = OpLabel\n"
18175                 "             OpReturnValue %param\n"
18176
18177                 "             OpFunctionEnd\n"
18178         );
18179         SpecResource                    specResource;
18180         map<string, string>             specs;
18181         VulkanFeatures                  features;
18182         map<string, string>             fragments;
18183         vector<string>                  extensions;
18184
18185         specs["num_data_points"]        = de::toString(numDataPoints);
18186
18187         fragments["decoration"]         = decoration.specialize(specs);
18188         fragments["pre_main"]           = preMain.specialize(specs);
18189         fragments["testfun"]            = testFun.specialize(specs);
18190
18191         specResource.inputs.push_back(Resource(BufferSp(new Uint32Buffer(inData1)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
18192         specResource.inputs.push_back(Resource(BufferSp(new Uint32Buffer(inData2)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
18193         specResource.outputs.push_back(Resource(BufferSp(new Uint32Buffer(outData)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
18194
18195         features.coreFeatures.vertexPipelineStoresAndAtomics    = true;
18196         features.coreFeatures.fragmentStoresAndAtomics                  = true;
18197
18198         finalizeTestsCreation(specResource, fragments, testCtx, *testGroup.get(), testName, features, extensions, IVec3(1, 1, 1));
18199 }
18200
18201 vector<deUint32> getLotsIdsAbuseData (const deUint32 numDataPoints, const deUint32 seed)
18202 {
18203         vector<deUint32>        result;
18204         de::Random                      rnd             (seed);
18205
18206         result.reserve(numDataPoints);
18207
18208         // Fixed value
18209         result.push_back(1u);
18210
18211         // Random values
18212         for (deUint32 dataPointNdx = 1; dataPointNdx < numDataPoints; ++dataPointNdx)
18213                 result.push_back(rnd.getUint8());
18214
18215         return result;
18216 }
18217
18218 vector<deUint32> getLotsIdsAbuseResults (const vector<deUint32>& inData1, const vector<deUint32>& inData2, const deUint32 count)
18219 {
18220         vector<deUint32>        result;
18221
18222         result.reserve(inData1.size());
18223
18224         for (size_t dataPointNdx = 0; dataPointNdx < inData1.size(); ++dataPointNdx)
18225                 result.push_back(inData1[dataPointNdx] + count * inData2[dataPointNdx]);
18226
18227         return result;
18228 }
18229
18230 template<class SpecResource>
18231 void createLotsIdsAbuseTest (tcu::TestContext& testCtx, de::MovePtr<tcu::TestCaseGroup>& testGroup)
18232 {
18233         const deUint32                  numDataPoints   = 16;
18234         const deUint32                  firstNdx                = 100u;
18235         const deUint32                  sequenceCount   = 10000u;
18236         const std::string               testName                ("lots_ids");
18237         const deUint32                  seed                    (deStringHash(testName.c_str()));
18238         const vector<deUint32>  inData1                 (getLotsIdsAbuseData(numDataPoints, seed + 1));
18239         const vector<deUint32>  inData2                 (getLotsIdsAbuseData(numDataPoints, seed + 2));
18240         const vector<deUint32>  outData                 (getLotsIdsAbuseResults(inData1, inData2, sequenceCount));
18241         const StringTemplate preMain
18242         (
18243                 "%c_i32_ndp = OpConstant %i32 ${num_data_points}\n"
18244                 "   %up_u32 = OpTypePointer Uniform %u32\n"
18245                 "   %ra_u32 = OpTypeArray %u32 %c_i32_ndp\n"
18246                 "   %SSBO32 = OpTypeStruct %ra_u32\n"
18247                 "%up_SSBO32 = OpTypePointer Uniform %SSBO32\n"
18248                 "%ssbo_src0 = OpVariable %up_SSBO32 Uniform\n"
18249                 "%ssbo_src1 = OpVariable %up_SSBO32 Uniform\n"
18250                 " %ssbo_dst = OpVariable %up_SSBO32 Uniform\n"
18251         );
18252         const StringTemplate decoration
18253         (
18254                 "OpDecorate %ra_u32 ArrayStride 4\n"
18255                 "OpMemberDecorate %SSBO32 0 Offset 0\n"
18256                 "OpDecorate %SSBO32 BufferBlock\n"
18257                 "OpDecorate %ssbo_src0 DescriptorSet 0\n"
18258                 "OpDecorate %ssbo_src0 Binding 0\n"
18259                 "OpDecorate %ssbo_src1 DescriptorSet 0\n"
18260                 "OpDecorate %ssbo_src1 Binding 1\n"
18261                 "OpDecorate %ssbo_dst DescriptorSet 0\n"
18262                 "OpDecorate %ssbo_dst Binding 2\n"
18263         );
18264         const StringTemplate testFun
18265         (
18266                 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
18267                 "    %param = OpFunctionParameter %v4f32\n"
18268
18269                 "    %entry = OpLabel\n"
18270                 "        %i = OpVariable %fp_i32 Function\n"
18271                 "             OpStore %i %c_i32_0\n"
18272                 "             OpBranch %loop\n"
18273
18274                 "     %loop = OpLabel\n"
18275                 "    %i_cmp = OpLoad %i32 %i\n"
18276                 "       %lt = OpSLessThan %bool %i_cmp %c_i32_ndp\n"
18277                 "             OpLoopMerge %merge %next None\n"
18278                 "             OpBranchConditional %lt %write %merge\n"
18279
18280                 "    %write = OpLabel\n"
18281                 "      %ndx = OpLoad %i32 %i\n"
18282
18283                 "       %90 = OpAccessChain %up_u32 %ssbo_src1 %c_i32_0 %ndx\n"
18284                 "       %91 = OpLoad %u32 %90\n"
18285
18286                 "       %98 = OpAccessChain %up_u32 %ssbo_src0 %c_i32_0 %ndx\n"
18287                 "       %${zeroth_id} = OpLoad %u32 %98\n"
18288
18289                 "${seq}\n"
18290
18291                 // The test relies on SPIR-V compiler option SPV_TEXT_TO_BINARY_OPTION_PRESERVE_NUMERIC_IDS set in assembleSpirV()
18292                 "      %dst = OpAccessChain %up_u32 %ssbo_dst %c_i32_0 %ndx\n"
18293                 "             OpStore %dst %${last_id}\n"
18294                 "             OpBranch %next\n"
18295
18296                 "     %next = OpLabel\n"
18297                 "    %i_cur = OpLoad %i32 %i\n"
18298                 "    %i_new = OpIAdd %i32 %i_cur %c_i32_1\n"
18299                 "             OpStore %i %i_new\n"
18300                 "             OpBranch %loop\n"
18301
18302                 "    %merge = OpLabel\n"
18303                 "             OpReturnValue %param\n"
18304
18305                 "             OpFunctionEnd\n"
18306         );
18307         deUint32                                lastId                  = firstNdx;
18308         SpecResource                    specResource;
18309         map<string, string>             specs;
18310         VulkanFeatures                  features;
18311         map<string, string>             fragments;
18312         vector<string>                  extensions;
18313         std::string                             sequence;
18314
18315         for (deUint32 sequenceNdx = 0; sequenceNdx < sequenceCount; ++sequenceNdx)
18316         {
18317                 const deUint32          sequenceId              = sequenceNdx + firstNdx;
18318                 const std::string       sequenceIdStr   = de::toString(sequenceId);
18319
18320                 sequence += "%" + sequenceIdStr + " = OpIAdd %u32 %91 %" + de::toString(sequenceId - 1) + "\n";
18321                 lastId = sequenceId;
18322
18323                 if (sequenceNdx == 0)
18324                         sequence.reserve((10 + sequence.length()) * sequenceCount);
18325         }
18326
18327         specs["num_data_points"]        = de::toString(numDataPoints);
18328         specs["zeroth_id"]                      = de::toString(firstNdx - 1);
18329         specs["last_id"]                        = de::toString(lastId);
18330         specs["seq"]                            = sequence;
18331
18332         fragments["decoration"]         = decoration.specialize(specs);
18333         fragments["pre_main"]           = preMain.specialize(specs);
18334         fragments["testfun"]            = testFun.specialize(specs);
18335
18336         specResource.inputs.push_back(Resource(BufferSp(new Uint32Buffer(inData1)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
18337         specResource.inputs.push_back(Resource(BufferSp(new Uint32Buffer(inData2)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
18338         specResource.outputs.push_back(Resource(BufferSp(new Uint32Buffer(outData)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
18339
18340         features.coreFeatures.vertexPipelineStoresAndAtomics    = true;
18341         features.coreFeatures.fragmentStoresAndAtomics                  = true;
18342
18343         finalizeTestsCreation(specResource, fragments, testCtx, *testGroup.get(), testName, features, extensions, IVec3(1, 1, 1));
18344 }
18345
18346 tcu::TestCaseGroup* createSpirvIdsAbuseTests (tcu::TestContext& testCtx)
18347 {
18348         de::MovePtr<tcu::TestCaseGroup> testGroup       (new tcu::TestCaseGroup(testCtx, "spirv_ids_abuse", "SPIR-V abuse tests"));
18349
18350         createSparseIdsAbuseTest<GraphicsResources>(testCtx, testGroup);
18351         createLotsIdsAbuseTest<GraphicsResources>(testCtx, testGroup);
18352
18353         return testGroup.release();
18354 }
18355
18356 tcu::TestCaseGroup* createSpirvIdsAbuseGroup (tcu::TestContext& testCtx)
18357 {
18358         de::MovePtr<tcu::TestCaseGroup> testGroup       (new tcu::TestCaseGroup(testCtx, "spirv_ids_abuse", "SPIR-V abuse tests"));
18359
18360         createSparseIdsAbuseTest<ComputeShaderSpec>(testCtx, testGroup);
18361         createLotsIdsAbuseTest<ComputeShaderSpec>(testCtx, testGroup);
18362
18363         return testGroup.release();
18364 }
18365
18366 tcu::TestCaseGroup* createInstructionTests (tcu::TestContext& testCtx)
18367 {
18368         const bool testComputePipeline = true;
18369
18370         de::MovePtr<tcu::TestCaseGroup> instructionTests        (new tcu::TestCaseGroup(testCtx, "instruction", "Instructions with special opcodes/operands"));
18371         de::MovePtr<tcu::TestCaseGroup> computeTests            (new tcu::TestCaseGroup(testCtx, "compute", "Compute Instructions with special opcodes/operands"));
18372         de::MovePtr<tcu::TestCaseGroup> graphicsTests           (new tcu::TestCaseGroup(testCtx, "graphics", "Graphics Instructions with special opcodes/operands"));
18373
18374         computeTests->addChild(createSpivVersionCheckTests(testCtx, testComputePipeline));
18375         computeTests->addChild(createLocalSizeGroup(testCtx));
18376         computeTests->addChild(createOpNopGroup(testCtx));
18377         computeTests->addChild(createOpFUnordGroup(testCtx, TEST_WITHOUT_NAN));
18378         computeTests->addChild(createOpFUnordGroup(testCtx, TEST_WITH_NAN));
18379         computeTests->addChild(createOpAtomicGroup(testCtx, false));
18380         computeTests->addChild(createOpAtomicGroup(testCtx, true));                                     // Using new StorageBuffer decoration
18381         computeTests->addChild(createOpAtomicGroup(testCtx, false, 1024, true));        // Return value validation
18382         computeTests->addChild(createOpLineGroup(testCtx));
18383         computeTests->addChild(createOpModuleProcessedGroup(testCtx));
18384         computeTests->addChild(createOpNoLineGroup(testCtx));
18385         computeTests->addChild(createOpConstantNullGroup(testCtx));
18386         computeTests->addChild(createOpConstantCompositeGroup(testCtx));
18387         computeTests->addChild(createOpConstantUsageGroup(testCtx));
18388         computeTests->addChild(createSpecConstantGroup(testCtx));
18389         computeTests->addChild(createOpSourceGroup(testCtx));
18390         computeTests->addChild(createOpSourceExtensionGroup(testCtx));
18391         computeTests->addChild(createDecorationGroupGroup(testCtx));
18392         computeTests->addChild(createOpPhiGroup(testCtx));
18393         computeTests->addChild(createLoopControlGroup(testCtx));
18394         computeTests->addChild(createFunctionControlGroup(testCtx));
18395         computeTests->addChild(createSelectionControlGroup(testCtx));
18396         computeTests->addChild(createBlockOrderGroup(testCtx));
18397         computeTests->addChild(createMultipleShaderGroup(testCtx));
18398         computeTests->addChild(createMemoryAccessGroup(testCtx));
18399         computeTests->addChild(createOpCopyMemoryGroup(testCtx));
18400         computeTests->addChild(createOpCopyObjectGroup(testCtx));
18401         computeTests->addChild(createNoContractionGroup(testCtx));
18402         computeTests->addChild(createOpUndefGroup(testCtx));
18403         computeTests->addChild(createOpUnreachableGroup(testCtx));
18404         computeTests->addChild(createOpQuantizeToF16Group(testCtx));
18405         computeTests->addChild(createOpFRemGroup(testCtx));
18406         computeTests->addChild(createOpSRemComputeGroup(testCtx, QP_TEST_RESULT_PASS));
18407         computeTests->addChild(createOpSRemComputeGroup64(testCtx, QP_TEST_RESULT_PASS));
18408         computeTests->addChild(createOpSModComputeGroup(testCtx, QP_TEST_RESULT_PASS));
18409         computeTests->addChild(createOpSModComputeGroup64(testCtx, QP_TEST_RESULT_PASS));
18410         computeTests->addChild(createConvertComputeTests(testCtx, "OpSConvert", "sconvert"));
18411         computeTests->addChild(createConvertComputeTests(testCtx, "OpUConvert", "uconvert"));
18412         computeTests->addChild(createConvertComputeTests(testCtx, "OpFConvert", "fconvert"));
18413         computeTests->addChild(createConvertComputeTests(testCtx, "OpConvertSToF", "convertstof"));
18414         computeTests->addChild(createConvertComputeTests(testCtx, "OpConvertFToS", "convertftos"));
18415         computeTests->addChild(createConvertComputeTests(testCtx, "OpConvertUToF", "convertutof"));
18416         computeTests->addChild(createConvertComputeTests(testCtx, "OpConvertFToU", "convertftou"));
18417         computeTests->addChild(createOpCompositeInsertGroup(testCtx));
18418         computeTests->addChild(createOpInBoundsAccessChainGroup(testCtx));
18419         computeTests->addChild(createShaderDefaultOutputGroup(testCtx));
18420         computeTests->addChild(createOpNMinGroup(testCtx));
18421         computeTests->addChild(createOpNMaxGroup(testCtx));
18422         computeTests->addChild(createOpNClampGroup(testCtx));
18423         {
18424                 de::MovePtr<tcu::TestCaseGroup> computeAndroidTests     (new tcu::TestCaseGroup(testCtx, "android", "Android CTS Tests"));
18425
18426                 computeAndroidTests->addChild(createOpSRemComputeGroup(testCtx, QP_TEST_RESULT_QUALITY_WARNING));
18427                 computeAndroidTests->addChild(createOpSModComputeGroup(testCtx, QP_TEST_RESULT_QUALITY_WARNING));
18428
18429                 computeTests->addChild(computeAndroidTests.release());
18430         }
18431
18432         computeTests->addChild(create8BitStorageComputeGroup(testCtx));
18433         computeTests->addChild(create16BitStorageComputeGroup(testCtx));
18434         computeTests->addChild(createFloatControlsComputeGroup(testCtx));
18435         computeTests->addChild(createUboMatrixPaddingComputeGroup(testCtx));
18436         computeTests->addChild(createCompositeInsertComputeGroup(testCtx));
18437         computeTests->addChild(createVariableInitComputeGroup(testCtx));
18438         computeTests->addChild(createConditionalBranchComputeGroup(testCtx));
18439         computeTests->addChild(createIndexingComputeGroup(testCtx));
18440         computeTests->addChild(createVariablePointersComputeGroup(testCtx));
18441         computeTests->addChild(createPhysicalPointersComputeGroup(testCtx));
18442         computeTests->addChild(createImageSamplerComputeGroup(testCtx));
18443         computeTests->addChild(createOpNameGroup(testCtx));
18444         computeTests->addChild(createOpMemberNameGroup(testCtx));
18445         computeTests->addChild(createPointerParameterComputeGroup(testCtx));
18446         computeTests->addChild(createFloat16Group(testCtx));
18447         computeTests->addChild(createBoolGroup(testCtx));
18448         computeTests->addChild(createWorkgroupMemoryComputeGroup(testCtx));
18449         computeTests->addChild(createSpirvIdsAbuseGroup(testCtx));
18450         computeTests->addChild(createSignedIntCompareGroup(testCtx));
18451
18452         graphicsTests->addChild(createCrossStageInterfaceTests(testCtx));
18453         graphicsTests->addChild(createSpivVersionCheckTests(testCtx, !testComputePipeline));
18454         graphicsTests->addChild(createOpNopTests(testCtx));
18455         graphicsTests->addChild(createOpSourceTests(testCtx));
18456         graphicsTests->addChild(createOpSourceContinuedTests(testCtx));
18457         graphicsTests->addChild(createOpModuleProcessedTests(testCtx));
18458         graphicsTests->addChild(createOpLineTests(testCtx));
18459         graphicsTests->addChild(createOpNoLineTests(testCtx));
18460         graphicsTests->addChild(createOpConstantNullTests(testCtx));
18461         graphicsTests->addChild(createOpConstantCompositeTests(testCtx));
18462         graphicsTests->addChild(createMemoryAccessTests(testCtx));
18463         graphicsTests->addChild(createOpUndefTests(testCtx));
18464         graphicsTests->addChild(createSelectionBlockOrderTests(testCtx));
18465         graphicsTests->addChild(createModuleTests(testCtx));
18466         graphicsTests->addChild(createSwitchBlockOrderTests(testCtx));
18467         graphicsTests->addChild(createOpPhiTests(testCtx));
18468         graphicsTests->addChild(createNoContractionTests(testCtx));
18469         graphicsTests->addChild(createOpQuantizeTests(testCtx));
18470         graphicsTests->addChild(createLoopTests(testCtx));
18471         graphicsTests->addChild(createSpecConstantTests(testCtx));
18472         graphicsTests->addChild(createSpecConstantOpQuantizeToF16Group(testCtx));
18473         graphicsTests->addChild(createBarrierTests(testCtx));
18474         graphicsTests->addChild(createDecorationGroupTests(testCtx));
18475         graphicsTests->addChild(createFRemTests(testCtx));
18476         graphicsTests->addChild(createOpSRemGraphicsTests(testCtx, QP_TEST_RESULT_PASS));
18477         graphicsTests->addChild(createOpSModGraphicsTests(testCtx, QP_TEST_RESULT_PASS));
18478
18479         {
18480                 de::MovePtr<tcu::TestCaseGroup> graphicsAndroidTests    (new tcu::TestCaseGroup(testCtx, "android", "Android CTS Tests"));
18481
18482                 graphicsAndroidTests->addChild(createOpSRemGraphicsTests(testCtx, QP_TEST_RESULT_QUALITY_WARNING));
18483                 graphicsAndroidTests->addChild(createOpSModGraphicsTests(testCtx, QP_TEST_RESULT_QUALITY_WARNING));
18484
18485                 graphicsTests->addChild(graphicsAndroidTests.release());
18486         }
18487         graphicsTests->addChild(createOpNameTests(testCtx));
18488         graphicsTests->addChild(createOpNameAbuseTests(testCtx));
18489         graphicsTests->addChild(createOpMemberNameAbuseTests(testCtx));
18490
18491         graphicsTests->addChild(create8BitStorageGraphicsGroup(testCtx));
18492         graphicsTests->addChild(create16BitStorageGraphicsGroup(testCtx));
18493         graphicsTests->addChild(createFloatControlsGraphicsGroup(testCtx));
18494         graphicsTests->addChild(createUboMatrixPaddingGraphicsGroup(testCtx));
18495         graphicsTests->addChild(createCompositeInsertGraphicsGroup(testCtx));
18496         graphicsTests->addChild(createVariableInitGraphicsGroup(testCtx));
18497         graphicsTests->addChild(createConditionalBranchGraphicsGroup(testCtx));
18498         graphicsTests->addChild(createIndexingGraphicsGroup(testCtx));
18499         graphicsTests->addChild(createVariablePointersGraphicsGroup(testCtx));
18500         graphicsTests->addChild(createImageSamplerGraphicsGroup(testCtx));
18501         graphicsTests->addChild(createConvertGraphicsTests(testCtx, "OpSConvert", "sconvert"));
18502         graphicsTests->addChild(createConvertGraphicsTests(testCtx, "OpUConvert", "uconvert"));
18503         graphicsTests->addChild(createConvertGraphicsTests(testCtx, "OpFConvert", "fconvert"));
18504         graphicsTests->addChild(createConvertGraphicsTests(testCtx, "OpConvertSToF", "convertstof"));
18505         graphicsTests->addChild(createConvertGraphicsTests(testCtx, "OpConvertFToS", "convertftos"));
18506         graphicsTests->addChild(createConvertGraphicsTests(testCtx, "OpConvertUToF", "convertutof"));
18507         graphicsTests->addChild(createConvertGraphicsTests(testCtx, "OpConvertFToU", "convertftou"));
18508         graphicsTests->addChild(createPointerParameterGraphicsGroup(testCtx));
18509         graphicsTests->addChild(createVaryingNameGraphicsGroup(testCtx));
18510         graphicsTests->addChild(createFloat16Tests(testCtx));
18511         graphicsTests->addChild(createSpirvIdsAbuseTests(testCtx));
18512
18513         instructionTests->addChild(computeTests.release());
18514         instructionTests->addChild(graphicsTests.release());
18515
18516         return instructionTests.release();
18517 }
18518
18519 } // SpirVAssembly
18520 } // vkt