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
  75 #include <cmath>
  76 #include <limits>
  77 #include <map>
  78 #include <string>
  79 #include <sstream>
  80 #include <utility>
  81 #include <stack>
  82
  83 namespace vkt
  84 {
  85 namespace SpirVAssembly
  86 {
  87
  88 namespace
  89 {
  90
  91 using namespace vk;
  92 using std::map;
  93 using std::string;
  94 using std::vector;
  95 using tcu::IVec3;
  96 using tcu::IVec4;
  97 using tcu::RGBA;
  98 using tcu::TestLog;
  99 using tcu::TestStatus;
 100 using tcu::Vec4;
 101 using de::UniquePtr;
 102 using tcu::StringTemplate;
 103 using tcu::Vec4;
 104
 105 const bool TEST_WITH_NAN        = true;
 106 const bool TEST_WITHOUT_NAN     = false;
 107
 108 template<typename T>
 109 static void fillRandomScalars (de::Random& rnd, T minValue, T maxValue, void* dst, int numValues, int offset = 0)
 110 {
 111         T* const typedPtr = (T*)dst;
 112         for (int ndx = 0; ndx < numValues; ndx++)
 113                 typedPtr[offset + ndx] = randomScalar<T>(rnd, minValue, maxValue);
 114 }
 115
 116 // Filter is a function that returns true if a value should pass, false otherwise.
 117 template<typename T, typename FilterT>
 118 static void fillRandomScalars (de::Random& rnd, T minValue, T maxValue, void* dst, int numValues, FilterT filter, int offset = 0)
 119 {
 120         T* const typedPtr = (T*)dst;
 121         T value;
 122         for (int ndx = 0; ndx < numValues; ndx++)
 123         {
 124                 do
 125                         value = randomScalar<T>(rnd, minValue, maxValue);
 126                 while (!filter(value));
 127
 128                 typedPtr[offset + ndx] = value;
 129         }
 130 }
 131
 132 // Gets a 64-bit integer with a more logarithmic distribution
 133 deInt64 randomInt64LogDistributed (de::Random& rnd)
 134 {
 135         deInt64 val = rnd.getUint64();
 136         val &= (1ull << rnd.getInt(1, 63)) - 1;
 137         if (rnd.getBool())
 138                 val = -val;
 139         return val;
 140 }
 141
 142 static void fillRandomInt64sLogDistributed (de::Random& rnd, vector<deInt64>& dst, int numValues)
 143 {
 144         for (int ndx = 0; ndx < numValues; ndx++)
 145                 dst[ndx] = randomInt64LogDistributed(rnd);
 146 }
 147
 148 template<typename FilterT>
 149 static void fillRandomInt64sLogDistributed (de::Random& rnd, vector<deInt64>& dst, int numValues, FilterT filter)
 150 {
 151         for (int ndx = 0; ndx < numValues; ndx++)
 152         {
 153                 deInt64 value;
 154                 do {
 155                         value = randomInt64LogDistributed(rnd);
 156                 } while (!filter(value));
 157                 dst[ndx] = value;
 158         }
 159 }
 160
 161 inline bool filterNonNegative (const deInt64 value)
 162 {
 163         return value >= 0;
 164 }
 165
 166 inline bool filterPositive (const deInt64 value)
 167 {
 168         return value > 0;
 169 }
 170
 171 inline bool filterNotZero (const deInt64 value)
 172 {
 173         return value != 0;
 174 }
 175
 176 static void floorAll (vector<float>& values)
 177 {
 178         for (size_t i = 0; i < values.size(); i++)
 179                 values[i] = deFloatFloor(values[i]);
 180 }
 181
 182 static void floorAll (vector<Vec4>& values)
 183 {
 184         for (size_t i = 0; i < values.size(); i++)
 185                 values[i] = floor(values[i]);
 186 }
 187
 188 struct CaseParameter
 189 {
 190         const char*             name;
 191         string                  param;
 192
 193         CaseParameter   (const char* case_, const string& param_) : name(case_), param(param_) {}
 194 };
 195
 196 // Assembly code used for testing LocalSize, OpNop, OpConstant{Null|Composite}, Op[No]Line, OpSource[Continued], OpSourceExtension, OpUndef is based on GLSL source code:
 197 //
 198 // #version 430
 199 //
 200 // layout(std140, set = 0, binding = 0) readonly buffer Input {
 201 //   float elements[];
 202 // } input_data;
 203 // layout(std140, set = 0, binding = 1) writeonly buffer Output {
 204 //   float elements[];
 205 // } output_data;
 206 //
 207 // layout (local_size_x = 1, local_size_y = 1, local_size_z = 1) in;
 208 //
 209 // void main() {
 210 //   uint x = gl_GlobalInvocationID.x;
 211 //   output_data.elements[x] = -input_data.elements[x];
 212 // }
 213
 214 static string getAsmForLocalSizeTest(bool useLiteralLocalSize, bool useSpecConstantWorkgroupSize, IVec3 workGroupSize, deUint32 ndx)
 215 {
 216         std::ostringstream out;
 217         out << getComputeAsmShaderPreambleWithoutLocalSize();
 218
 219         if (useLiteralLocalSize)
 220         {
 221                 out << "OpExecutionMode %main LocalSize "
 222                         << workGroupSize.x() << " " << workGroupSize.y() << " " << workGroupSize.z() << "\n";
 223         }
 224
 225         out << "OpSource GLSL 430\n"
 226                 "OpName %main           \"main\"\n"
 227                 "OpName %id             \"gl_GlobalInvocationID\"\n"
 228                 "OpDecorate %id BuiltIn GlobalInvocationId\n";
 229
 230         if (useSpecConstantWorkgroupSize)
 231         {
 232                 out << "OpDecorate %spec_0 SpecId 100\n"
 233                         << "OpDecorate %spec_1 SpecId 101\n"
 234                         << "OpDecorate %spec_2 SpecId 102\n"
 235                         << "OpDecorate %gl_WorkGroupSize BuiltIn WorkgroupSize\n";
 236         }
 237
 238         out << getComputeAsmInputOutputBufferTraits()
 239                 << getComputeAsmCommonTypes()
 240                 << getComputeAsmInputOutputBuffer()
 241                 << "%id        = OpVariable %uvec3ptr Input\n"
 242                 << "%zero      = OpConstant %i32 0 \n";
 243
 244         if (useSpecConstantWorkgroupSize)
 245         {
 246                 out     << "%spec_0   = OpSpecConstant %u32 "<< workGroupSize.x() << "\n"
 247                         << "%spec_1   = OpSpecConstant %u32 "<< workGroupSize.y() << "\n"
 248                         << "%spec_2   = OpSpecConstant %u32 "<< workGroupSize.z() << "\n"
 249                         << "%gl_WorkGroupSize = OpSpecConstantComposite %uvec3 %spec_0 %spec_1 %spec_2\n";
 250         }
 251
 252         out << "%main      = OpFunction %void None %voidf\n"
 253                 << "%label     = OpLabel\n"
 254                 << "%idval     = OpLoad %uvec3 %id\n"
 255                 << "%ndx         = OpCompositeExtract %u32 %idval " << ndx << "\n"
 256
 257                         "%inloc     = OpAccessChain %f32ptr %indata %zero %ndx\n"
 258                         "%inval     = OpLoad %f32 %inloc\n"
 259                         "%neg       = OpFNegate %f32 %inval\n"
 260                         "%outloc    = OpAccessChain %f32ptr %outdata %zero %ndx\n"
 261                         "             OpStore %outloc %neg\n"
 262                         "             OpReturn\n"
 263                         "             OpFunctionEnd\n";
 264         return out.str();
 265 }
 266
 267 tcu::TestCaseGroup* createLocalSizeGroup (tcu::TestContext& testCtx)
 268 {
 269         de::MovePtr<tcu::TestCaseGroup> group                   (new tcu::TestCaseGroup(testCtx, "localsize", ""));
 270         ComputeShaderSpec                               spec;
 271         de::Random                                              rnd                             (deStringHash(group->getName()));
 272         const deUint32                                  numElements             = 64u;
 273         vector<float>                                   positiveFloats  (numElements, 0);
 274         vector<float>                                   negativeFloats  (numElements, 0);
 275
 276         fillRandomScalars(rnd, 1.f, 100.f, &positiveFloats[0], numElements);
 277
 278         for (size_t ndx = 0; ndx < numElements; ++ndx)
 279                 negativeFloats[ndx] = -positiveFloats[ndx];
 280
 281         spec.inputs.push_back(BufferSp(new Float32Buffer(positiveFloats)));
 282         spec.outputs.push_back(BufferSp(new Float32Buffer(negativeFloats)));
 283
 284         spec.numWorkGroups = IVec3(numElements, 1, 1);
 285
 286         spec.assembly = getAsmForLocalSizeTest(true, false, IVec3(1, 1, 1), 0u);
 287         group->addChild(new SpvAsmComputeShaderCase(testCtx, "literal_localsize", "", spec));
 288
 289         spec.assembly = getAsmForLocalSizeTest(true, true, IVec3(1, 1, 1), 0u);
 290         group->addChild(new SpvAsmComputeShaderCase(testCtx, "literal_and_specid_localsize", "", spec));
 291
 292         spec.assembly = getAsmForLocalSizeTest(false, true, IVec3(1, 1, 1), 0u);
 293         group->addChild(new SpvAsmComputeShaderCase(testCtx, "specid_localsize", "", spec));
 294
 295         spec.numWorkGroups = IVec3(1, 1, 1);
 296
 297         spec.assembly = getAsmForLocalSizeTest(true, false, IVec3(numElements, 1, 1), 0u);
 298         group->addChild(new SpvAsmComputeShaderCase(testCtx, "literal_localsize_x", "", spec));
 299
 300         spec.assembly = getAsmForLocalSizeTest(true, true, IVec3(numElements, 1, 1), 0u);
 301         group->addChild(new SpvAsmComputeShaderCase(testCtx, "literal_and_specid_localsize_x", "", spec));
 302
 303         spec.assembly = getAsmForLocalSizeTest(false, true, IVec3(numElements, 1, 1), 0u);
 304         group->addChild(new SpvAsmComputeShaderCase(testCtx, "specid_localsize_x", "", spec));
 305
 306         spec.assembly = getAsmForLocalSizeTest(true, false, IVec3(1, numElements, 1), 1u);
 307         group->addChild(new SpvAsmComputeShaderCase(testCtx, "literal_localsize_y", "", spec));
 308
 309         spec.assembly = getAsmForLocalSizeTest(true, true, IVec3(1, numElements, 1), 1u);
 310         group->addChild(new SpvAsmComputeShaderCase(testCtx, "literal_and_specid_localsize_y", "", spec));
 311
 312         spec.assembly = getAsmForLocalSizeTest(false, true, IVec3(1, numElements, 1), 1u);
 313         group->addChild(new SpvAsmComputeShaderCase(testCtx, "specid_localsize_y", "", spec));
 314
 315         spec.assembly = getAsmForLocalSizeTest(true, false, IVec3(1, 1, numElements), 2u);
 316         group->addChild(new SpvAsmComputeShaderCase(testCtx, "literal_localsize_z", "", spec));
 317
 318         spec.assembly = getAsmForLocalSizeTest(true, true, IVec3(1, 1, numElements), 2u);
 319         group->addChild(new SpvAsmComputeShaderCase(testCtx, "literal_and_specid_localsize_z", "", spec));
 320
 321         spec.assembly = getAsmForLocalSizeTest(false, true, IVec3(1, 1, numElements), 2u);
 322         group->addChild(new SpvAsmComputeShaderCase(testCtx, "specid_localsize_z", "", spec));
 323
 324         return group.release();
 325 }
 326
 327 tcu::TestCaseGroup* createOpNopGroup (tcu::TestContext& testCtx)
 328 {
 329         de::MovePtr<tcu::TestCaseGroup> group                   (new tcu::TestCaseGroup(testCtx, "opnop", "Test the OpNop instruction"));
 330         ComputeShaderSpec                               spec;
 331         de::Random                                              rnd                             (deStringHash(group->getName()));
 332         const int                                               numElements             = 100;
 333         vector<float>                                   positiveFloats  (numElements, 0);
 334         vector<float>                                   negativeFloats  (numElements, 0);
 335
 336         fillRandomScalars(rnd, 1.f, 100.f, &positiveFloats[0], numElements);
 337
 338         for (size_t ndx = 0; ndx < numElements; ++ndx)
 339                 negativeFloats[ndx] = -positiveFloats[ndx];
 340
 341         spec.assembly =
 342                 string(getComputeAsmShaderPreamble()) +
 343
 344                 "OpSource GLSL 430\n"
 345                 "OpName %main           \"main\"\n"
 346                 "OpName %id             \"gl_GlobalInvocationID\"\n"
 347
 348                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
 349
 350                 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes())
 351
 352                 + string(getComputeAsmInputOutputBuffer()) +
 353
 354                 "%id        = OpVariable %uvec3ptr Input\n"
 355                 "%zero      = OpConstant %i32 0\n"
 356
 357                 "%main      = OpFunction %void None %voidf\n"
 358                 "%label     = OpLabel\n"
 359                 "%idval     = OpLoad %uvec3 %id\n"
 360                 "%x         = OpCompositeExtract %u32 %idval 0\n"
 361
 362                 "             OpNop\n" // Inside a function body
 363
 364                 "%inloc     = OpAccessChain %f32ptr %indata %zero %x\n"
 365                 "%inval     = OpLoad %f32 %inloc\n"
 366                 "%neg       = OpFNegate %f32 %inval\n"
 367                 "%outloc    = OpAccessChain %f32ptr %outdata %zero %x\n"
 368                 "             OpStore %outloc %neg\n"
 369                 "             OpReturn\n"
 370                 "             OpFunctionEnd\n";
 371         spec.inputs.push_back(BufferSp(new Float32Buffer(positiveFloats)));
 372         spec.outputs.push_back(BufferSp(new Float32Buffer(negativeFloats)));
 373         spec.numWorkGroups = IVec3(numElements, 1, 1);
 374
 375         group->addChild(new SpvAsmComputeShaderCase(testCtx, "all", "OpNop appearing at different places", spec));
 376
 377         return group.release();
 378 }
 379
 380 template<bool nanSupported>
 381 bool compareFUnord (const std::vector<Resource>& inputs, const vector<AllocationSp>& outputAllocs, const std::vector<Resource>& expectedOutputs, TestLog& log)
 382 {
 383         if (outputAllocs.size() != 1)
 384                 return false;
 385
 386         vector<deUint8> input1Bytes;
 387         vector<deUint8> input2Bytes;
 388         vector<deUint8> expectedBytes;
 389
 390         inputs[0].getBytes(input1Bytes);
 391         inputs[1].getBytes(input2Bytes);
 392         expectedOutputs[0].getBytes(expectedBytes);
 393
 394         const deInt32* const    expectedOutputAsInt             = reinterpret_cast<const deInt32*>(&expectedBytes.front());
 395         const deInt32* const    outputAsInt                             = static_cast<const deInt32*>(outputAllocs[0]->getHostPtr());
 396         const float* const              input1AsFloat                   = reinterpret_cast<const float*>(&input1Bytes.front());
 397         const float* const              input2AsFloat                   = reinterpret_cast<const float*>(&input2Bytes.front());
 398         bool returnValue                                                                = true;
 399
 400         for (size_t idx = 0; idx < expectedBytes.size() / sizeof(deInt32); ++idx)
 401         {
 402                 if (!nanSupported && (tcu::Float32(input1AsFloat[idx]).isNaN() || tcu::Float32(input2AsFloat[idx]).isNaN()))
 403                         continue;
 404
 405                 if (outputAsInt[idx] != expectedOutputAsInt[idx])
 406                 {
 407                         log << TestLog::Message << "ERROR: Sub-case failed. inputs: " << input1AsFloat[idx] << "," << input2AsFloat[idx] << " output: " << outputAsInt[idx]<< " expected output: " << expectedOutputAsInt[idx] << TestLog::EndMessage;
 408                         returnValue = false;
 409                 }
 410         }
 411         return returnValue;
 412 }
 413
 414 typedef VkBool32 (*compareFuncType) (float, float);
 415
 416 struct OpFUnordCase
 417 {
 418         const char*             name;
 419         const char*             opCode;
 420         compareFuncType compareFunc;
 421
 422                                         OpFUnordCase                    (const char* _name, const char* _opCode, compareFuncType _compareFunc)
 423                                                 : name                          (_name)
 424                                                 , opCode                        (_opCode)
 425                                                 , compareFunc           (_compareFunc) {}
 426 };
 427
 428 #define ADD_OPFUNORD_CASE(NAME, OPCODE, OPERATOR) \
 429 do { \
 430         struct compare_##NAME { static VkBool32 compare(float x, float y) { return (x OPERATOR y) ? VK_TRUE : VK_FALSE; } }; \
 431         cases.push_back(OpFUnordCase(#NAME, OPCODE, compare_##NAME::compare)); \
 432 } while (deGetFalse())
 433
 434 tcu::TestCaseGroup* createOpFUnordGroup (tcu::TestContext& testCtx, const bool nanSupported)
 435 {
 436         const string                                    nan                             = nanSupported ? "_nan" : "";
 437         const string                                    groupName               = "opfunord" + nan;
 438         de::MovePtr<tcu::TestCaseGroup> group                   (new tcu::TestCaseGroup(testCtx, groupName.c_str(), "Test the OpFUnord* opcodes"));
 439         de::Random                                              rnd                             (deStringHash(group->getName()));
 440         const int                                               numElements             = 100;
 441         vector<OpFUnordCase>                    cases;
 442         string                                                  extensions              = nanSupported ? "OpExtension \"SPV_KHR_float_controls\"\n" : "";
 443         string                                                  capabilities    = nanSupported ? "OpCapability SignedZeroInfNanPreserve\n" : "";
 444         string                          exeModes        = nanSupported ? "OpExecutionMode %main SignedZeroInfNanPreserve 32\n" : "";
 445         const StringTemplate                    shaderTemplate  (
 446                 string(getComputeAsmShaderPreamble(capabilities, extensions, exeModes)) +
 447                 "OpSource GLSL 430\n"
 448                 "OpName %main           \"main\"\n"
 449                 "OpName %id             \"gl_GlobalInvocationID\"\n"
 450
 451                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
 452
 453                 "OpDecorate %buf BufferBlock\n"
 454                 "OpDecorate %buf2 BufferBlock\n"
 455                 "OpDecorate %indata1 DescriptorSet 0\n"
 456                 "OpDecorate %indata1 Binding 0\n"
 457                 "OpDecorate %indata2 DescriptorSet 0\n"
 458                 "OpDecorate %indata2 Binding 1\n"
 459                 "OpDecorate %outdata DescriptorSet 0\n"
 460                 "OpDecorate %outdata Binding 2\n"
 461                 "OpDecorate %f32arr ArrayStride 4\n"
 462                 "OpDecorate %i32arr ArrayStride 4\n"
 463                 "OpMemberDecorate %buf 0 Offset 0\n"
 464                 "OpMemberDecorate %buf2 0 Offset 0\n"
 465
 466                 + string(getComputeAsmCommonTypes()) +
 467
 468                 "%buf        = OpTypeStruct %f32arr\n"
 469                 "%bufptr     = OpTypePointer Uniform %buf\n"
 470                 "%indata1    = OpVariable %bufptr Uniform\n"
 471                 "%indata2    = OpVariable %bufptr Uniform\n"
 472
 473                 "%buf2       = OpTypeStruct %i32arr\n"
 474                 "%buf2ptr    = OpTypePointer Uniform %buf2\n"
 475                 "%outdata    = OpVariable %buf2ptr Uniform\n"
 476
 477                 "%id        = OpVariable %uvec3ptr Input\n"
 478                 "%zero      = OpConstant %i32 0\n"
 479                 "%consti1   = OpConstant %i32 1\n"
 480                 "%constf1   = OpConstant %f32 1.0\n"
 481
 482                 "%main      = OpFunction %void None %voidf\n"
 483                 "%label     = OpLabel\n"
 484                 "%idval     = OpLoad %uvec3 %id\n"
 485                 "%x         = OpCompositeExtract %u32 %idval 0\n"
 486
 487                 "%inloc1    = OpAccessChain %f32ptr %indata1 %zero %x\n"
 488                 "%inval1    = OpLoad %f32 %inloc1\n"
 489                 "%inloc2    = OpAccessChain %f32ptr %indata2 %zero %x\n"
 490                 "%inval2    = OpLoad %f32 %inloc2\n"
 491                 "%outloc    = OpAccessChain %i32ptr %outdata %zero %x\n"
 492
 493                 "%result    = ${OPCODE} %bool %inval1 %inval2\n"
 494                 "%int_res   = OpSelect %i32 %result %consti1 %zero\n"
 495                 "             OpStore %outloc %int_res\n"
 496
 497                 "             OpReturn\n"
 498                 "             OpFunctionEnd\n");
 499
 500         ADD_OPFUNORD_CASE(equal, "OpFUnordEqual", ==);
 501         ADD_OPFUNORD_CASE(less, "OpFUnordLessThan", <);
 502         ADD_OPFUNORD_CASE(lessequal, "OpFUnordLessThanEqual", <=);
 503         ADD_OPFUNORD_CASE(greater, "OpFUnordGreaterThan", >);
 504         ADD_OPFUNORD_CASE(greaterequal, "OpFUnordGreaterThanEqual", >=);
 505         ADD_OPFUNORD_CASE(notequal, "OpFUnordNotEqual", !=);
 506
 507         for (size_t caseNdx = 0; caseNdx < cases.size(); ++caseNdx)
 508         {
 509                 map<string, string>                     specializations;
 510                 ComputeShaderSpec                       spec;
 511                 const float                                     NaN                             = std::numeric_limits<float>::quiet_NaN();
 512                 vector<float>                           inputFloats1    (numElements, 0);
 513                 vector<float>                           inputFloats2    (numElements, 0);
 514                 vector<deInt32>                         expectedInts    (numElements, 0);
 515
 516                 specializations["OPCODE"]       = cases[caseNdx].opCode;
 517                 spec.assembly                           = shaderTemplate.specialize(specializations);
 518
 519                 fillRandomScalars(rnd, 1.f, 100.f, &inputFloats1[0], numElements);
 520                 for (size_t ndx = 0; ndx < numElements; ++ndx)
 521                 {
 522                         switch (ndx % 6)
 523                         {
 524                                 case 0:         inputFloats2[ndx] = inputFloats1[ndx] + 1.0f; break;
 525                                 case 1:         inputFloats2[ndx] = inputFloats1[ndx] - 1.0f; break;
 526                                 case 2:         inputFloats2[ndx] = inputFloats1[ndx]; break;
 527                                 case 3:         inputFloats2[ndx] = NaN; break;
 528                                 case 4:         inputFloats2[ndx] = inputFloats1[ndx];  inputFloats1[ndx] = NaN; break;
 529                                 case 5:         inputFloats2[ndx] = NaN;                                inputFloats1[ndx] = NaN; break;
 530                         }
 531                         expectedInts[ndx] = tcu::Float32(inputFloats1[ndx]).isNaN() || tcu::Float32(inputFloats2[ndx]).isNaN() || cases[caseNdx].compareFunc(inputFloats1[ndx], inputFloats2[ndx]);
 532                 }
 533
 534                 spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats1)));
 535                 spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats2)));
 536                 spec.outputs.push_back(BufferSp(new Int32Buffer(expectedInts)));
 537                 spec.numWorkGroups = IVec3(numElements, 1, 1);
 538                 spec.verifyIO = nanSupported ? &compareFUnord<true> : &compareFUnord<false>;
 539                 if (nanSupported)
 540                 {
 541                         spec.extensions.push_back("VK_KHR_shader_float_controls");
 542                         spec.requestedVulkanFeatures.floatControlsProperties.shaderSignedZeroInfNanPreserveFloat32 = DE_TRUE;
 543                 }
 544                 group->addChild(new SpvAsmComputeShaderCase(testCtx, cases[caseNdx].name, cases[caseNdx].name, spec));
 545         }
 546
 547         return group.release();
 548 }
 549
 550 struct OpAtomicCase
 551 {
 552         const char*             name;
 553         const char*             assembly;
 554         const char*             retValAssembly;
 555         OpAtomicType    opAtomic;
 556         deInt32                 numOutputElements;
 557
 558                                         OpAtomicCase(const char* _name, const char* _assembly, const char* _retValAssembly, OpAtomicType _opAtomic, deInt32 _numOutputElements)
 559                                                 : name                          (_name)
 560                                                 , assembly                      (_assembly)
 561                                                 , retValAssembly        (_retValAssembly)
 562                                                 , opAtomic                      (_opAtomic)
 563                                                 , numOutputElements     (_numOutputElements) {}
 564 };
 565
 566 tcu::TestCaseGroup* createOpAtomicGroup (tcu::TestContext& testCtx, bool useStorageBuffer, int numElements = 65535, bool verifyReturnValues = false)
 567 {
 568         std::string                                             groupName                       ("opatomic");
 569         if (useStorageBuffer)
 570                 groupName += "_storage_buffer";
 571         if (verifyReturnValues)
 572                 groupName += "_return_values";
 573         de::MovePtr<tcu::TestCaseGroup> group                           (new tcu::TestCaseGroup(testCtx, groupName.c_str(), "Test the OpAtomic* opcodes"));
 574         vector<OpAtomicCase>                    cases;
 575
 576         const StringTemplate                    shaderTemplate  (
 577
 578                 string("OpCapability Shader\n") +
 579                 (useStorageBuffer ? "OpExtension \"SPV_KHR_storage_buffer_storage_class\"\n" : "") +
 580                 "OpMemoryModel Logical GLSL450\n"
 581                 "OpEntryPoint GLCompute %main \"main\" %id\n"
 582                 "OpExecutionMode %main LocalSize 1 1 1\n" +
 583
 584                 "OpSource GLSL 430\n"
 585                 "OpName %main           \"main\"\n"
 586                 "OpName %id             \"gl_GlobalInvocationID\"\n"
 587
 588                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
 589
 590                 "OpDecorate %buf ${BLOCK_DECORATION}\n"
 591                 "OpDecorate %indata DescriptorSet 0\n"
 592                 "OpDecorate %indata Binding 0\n"
 593                 "OpDecorate %i32arr ArrayStride 4\n"
 594                 "OpMemberDecorate %buf 0 Offset 0\n"
 595
 596                 "OpDecorate %sumbuf ${BLOCK_DECORATION}\n"
 597                 "OpDecorate %sum DescriptorSet 0\n"
 598                 "OpDecorate %sum Binding 1\n"
 599                 "OpMemberDecorate %sumbuf 0 Coherent\n"
 600                 "OpMemberDecorate %sumbuf 0 Offset 0\n"
 601
 602                 "${RETVAL_BUF_DECORATE}"
 603
 604                 + getComputeAsmCommonTypes("${BLOCK_POINTER_TYPE}") +
 605
 606                 "%buf       = OpTypeStruct %i32arr\n"
 607                 "%bufptr    = OpTypePointer ${BLOCK_POINTER_TYPE} %buf\n"
 608                 "%indata    = OpVariable %bufptr ${BLOCK_POINTER_TYPE}\n"
 609
 610                 "%sumbuf    = OpTypeStruct %i32arr\n"
 611                 "%sumbufptr = OpTypePointer ${BLOCK_POINTER_TYPE} %sumbuf\n"
 612                 "%sum       = OpVariable %sumbufptr ${BLOCK_POINTER_TYPE}\n"
 613
 614                 "${RETVAL_BUF_DECL}"
 615
 616                 "%id        = OpVariable %uvec3ptr Input\n"
 617                 "%minusone  = OpConstant %i32 -1\n"
 618                 "%zero      = OpConstant %i32 0\n"
 619                 "%one       = OpConstant %u32 1\n"
 620                 "%two       = OpConstant %i32 2\n"
 621
 622                 "%main      = OpFunction %void None %voidf\n"
 623                 "%label     = OpLabel\n"
 624                 "%idval     = OpLoad %uvec3 %id\n"
 625                 "%x         = OpCompositeExtract %u32 %idval 0\n"
 626
 627                 "%inloc     = OpAccessChain %i32ptr %indata %zero %x\n"
 628                 "%inval     = OpLoad %i32 %inloc\n"
 629
 630                 "%outloc    = OpAccessChain %i32ptr %sum %zero ${INDEX}\n"
 631                 "${INSTRUCTION}"
 632                 "${RETVAL_ASSEMBLY}"
 633
 634                 "             OpReturn\n"
 635                 "             OpFunctionEnd\n");
 636
 637         #define ADD_OPATOMIC_CASE(NAME, ASSEMBLY, RETVAL_ASSEMBLY, OPATOMIC, NUM_OUTPUT_ELEMENTS) \
 638         do { \
 639                 DE_ASSERT((NUM_OUTPUT_ELEMENTS) == 1 || (NUM_OUTPUT_ELEMENTS) == numElements); \
 640                 cases.push_back(OpAtomicCase(#NAME, ASSEMBLY, RETVAL_ASSEMBLY, OPATOMIC, NUM_OUTPUT_ELEMENTS)); \
 641         } while (deGetFalse())
 642         #define ADD_OPATOMIC_CASE_1(NAME, ASSEMBLY, RETVAL_ASSEMBLY, OPATOMIC) ADD_OPATOMIC_CASE(NAME, ASSEMBLY, RETVAL_ASSEMBLY, OPATOMIC, 1)
 643         #define ADD_OPATOMIC_CASE_N(NAME, ASSEMBLY, RETVAL_ASSEMBLY, OPATOMIC) ADD_OPATOMIC_CASE(NAME, ASSEMBLY, RETVAL_ASSEMBLY, OPATOMIC, numElements)
 644
 645         ADD_OPATOMIC_CASE_1(iadd,       "%retv      = OpAtomicIAdd %i32 %outloc %one %zero %inval\n",
 646                                                                 "             OpStore %retloc %retv\n", OPATOMIC_IADD );
 647         ADD_OPATOMIC_CASE_1(isub,       "%retv      = OpAtomicISub %i32 %outloc %one %zero %inval\n",
 648                                                                 "             OpStore %retloc %retv\n", OPATOMIC_ISUB );
 649         ADD_OPATOMIC_CASE_1(iinc,       "%retv      = OpAtomicIIncrement %i32 %outloc %one %zero\n",
 650                                                                 "             OpStore %retloc %retv\n", OPATOMIC_IINC );
 651         ADD_OPATOMIC_CASE_1(idec,       "%retv      = OpAtomicIDecrement %i32 %outloc %one %zero\n",
 652                                                                 "             OpStore %retloc %retv\n", OPATOMIC_IDEC );
 653         if (!verifyReturnValues)
 654         {
 655                 ADD_OPATOMIC_CASE_N(load,       "%inval2    = OpAtomicLoad %i32 %inloc %one %zero\n"
 656                                                                         "             OpStore %outloc %inval2\n", "", OPATOMIC_LOAD );
 657                 ADD_OPATOMIC_CASE_N(store,      "             OpAtomicStore %outloc %one %zero %inval\n", "", OPATOMIC_STORE );
 658         }
 659
 660         ADD_OPATOMIC_CASE_N(compex, "%even      = OpSMod %i32 %inval %two\n"
 661                                                                 "             OpStore %outloc %even\n"
 662                                                                 "%retv      = OpAtomicCompareExchange %i32 %outloc %one %zero %zero %minusone %zero\n",
 663                                                                 "                         OpStore %retloc %retv\n", OPATOMIC_COMPEX );
 664
 665
 666         #undef ADD_OPATOMIC_CASE
 667         #undef ADD_OPATOMIC_CASE_1
 668         #undef ADD_OPATOMIC_CASE_N
 669
 670         for (size_t caseNdx = 0; caseNdx < cases.size(); ++caseNdx)
 671         {
 672                 map<string, string>                     specializations;
 673                 ComputeShaderSpec                       spec;
 674                 vector<deInt32>                         inputInts               (numElements, 0);
 675                 vector<deInt32>                         expected                (cases[caseNdx].numOutputElements, -1);
 676
 677                 specializations["INDEX"]                                = (cases[caseNdx].numOutputElements == 1) ? "%zero" : "%x";
 678                 specializations["INSTRUCTION"]                  = cases[caseNdx].assembly;
 679                 specializations["BLOCK_DECORATION"]             = useStorageBuffer ? "Block" : "BufferBlock";
 680                 specializations["BLOCK_POINTER_TYPE"]   = useStorageBuffer ? "StorageBuffer" : "Uniform";
 681
 682                 if (verifyReturnValues)
 683                 {
 684                         const StringTemplate blockDecoration    (
 685                                 "\n"
 686                                 "OpDecorate %retbuf ${BLOCK_DECORATION}\n"
 687                                 "OpDecorate %ret DescriptorSet 0\n"
 688                                 "OpDecorate %ret Binding 2\n"
 689                                 "OpMemberDecorate %retbuf 0 Offset 0\n\n");
 690
 691                         const StringTemplate blockDeclaration   (
 692                                 "\n"
 693                                 "%retbuf    = OpTypeStruct %i32arr\n"
 694                                 "%retbufptr = OpTypePointer ${BLOCK_POINTER_TYPE} %retbuf\n"
 695                                 "%ret       = OpVariable %retbufptr ${BLOCK_POINTER_TYPE}\n\n");
 696
 697                         specializations["RETVAL_ASSEMBLY"] =
 698                                 "%retloc    = OpAccessChain %i32ptr %ret %zero %x\n"
 699                                 + std::string(cases[caseNdx].retValAssembly);
 700
 701                         specializations["RETVAL_BUF_DECORATE"]  = blockDecoration.specialize(specializations);
 702                         specializations["RETVAL_BUF_DECL"]              = blockDeclaration.specialize(specializations);
 703                 }
 704                 else
 705                 {
 706                         specializations["RETVAL_ASSEMBLY"]              = "";
 707                         specializations["RETVAL_BUF_DECORATE"]  = "";
 708                         specializations["RETVAL_BUF_DECL"]              = "";
 709                 }
 710
 711                 spec.assembly                                                   = shaderTemplate.specialize(specializations);
 712
 713                 if (useStorageBuffer)
 714                         spec.extensions.push_back("VK_KHR_storage_buffer_storage_class");
 715
 716                 spec.inputs.push_back(BufferSp(new OpAtomicBuffer(numElements, cases[caseNdx].numOutputElements, cases[caseNdx].opAtomic, BUFFERTYPE_INPUT)));
 717                 spec.outputs.push_back(BufferSp(new OpAtomicBuffer(numElements, cases[caseNdx].numOutputElements, cases[caseNdx].opAtomic, BUFFERTYPE_EXPECTED)));
 718                 if (verifyReturnValues)
 719                         spec.outputs.push_back(BufferSp(new OpAtomicBuffer(numElements, cases[caseNdx].numOutputElements, cases[caseNdx].opAtomic, BUFFERTYPE_ATOMIC_RET)));
 720                 spec.numWorkGroups = IVec3(numElements, 1, 1);
 721
 722                 if (verifyReturnValues)
 723                 {
 724                         switch (cases[caseNdx].opAtomic)
 725                         {
 726                                 case OPATOMIC_IADD:
 727                                         spec.verifyIO = OpAtomicBuffer::compareWithRetvals<OPATOMIC_IADD>;
 728                                         break;
 729                                 case OPATOMIC_ISUB:
 730                                         spec.verifyIO = OpAtomicBuffer::compareWithRetvals<OPATOMIC_ISUB>;
 731                                         break;
 732                                 case OPATOMIC_IINC:
 733                                         spec.verifyIO = OpAtomicBuffer::compareWithRetvals<OPATOMIC_IINC>;
 734                                         break;
 735                                 case OPATOMIC_IDEC:
 736                                         spec.verifyIO = OpAtomicBuffer::compareWithRetvals<OPATOMIC_IDEC>;
 737                                         break;
 738                                 case OPATOMIC_COMPEX:
 739                                         spec.verifyIO = OpAtomicBuffer::compareWithRetvals<OPATOMIC_COMPEX>;
 740                                         break;
 741                                 default:
 742                                         DE_FATAL("Unsupported OpAtomic type for return value verification");
 743                         }
 744                 }
 745                 group->addChild(new SpvAsmComputeShaderCase(testCtx, cases[caseNdx].name, cases[caseNdx].name, spec));
 746         }
 747
 748         return group.release();
 749 }
 750
 751 tcu::TestCaseGroup* createOpLineGroup (tcu::TestContext& testCtx)
 752 {
 753         de::MovePtr<tcu::TestCaseGroup> group                   (new tcu::TestCaseGroup(testCtx, "opline", "Test the OpLine instruction"));
 754         ComputeShaderSpec                               spec;
 755         de::Random                                              rnd                             (deStringHash(group->getName()));
 756         const int                                               numElements             = 100;
 757         vector<float>                                   positiveFloats  (numElements, 0);
 758         vector<float>                                   negativeFloats  (numElements, 0);
 759
 760         fillRandomScalars(rnd, 1.f, 100.f, &positiveFloats[0], numElements);
 761
 762         for (size_t ndx = 0; ndx < numElements; ++ndx)
 763                 negativeFloats[ndx] = -positiveFloats[ndx];
 764
 765         spec.assembly =
 766                 string(getComputeAsmShaderPreamble()) +
 767
 768                 "%fname1 = OpString \"negateInputs.comp\"\n"
 769                 "%fname2 = OpString \"negateInputs\"\n"
 770
 771                 "OpSource GLSL 430\n"
 772                 "OpName %main           \"main\"\n"
 773                 "OpName %id             \"gl_GlobalInvocationID\"\n"
 774
 775                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
 776
 777                 + string(getComputeAsmInputOutputBufferTraits()) +
 778
 779                 "OpLine %fname1 0 0\n" // At the earliest possible position
 780
 781                 + string(getComputeAsmCommonTypes()) + string(getComputeAsmInputOutputBuffer()) +
 782
 783                 "OpLine %fname1 0 1\n" // Multiple OpLines in sequence
 784                 "OpLine %fname2 1 0\n" // Different filenames
 785                 "OpLine %fname1 1000 100000\n"
 786
 787                 "%id        = OpVariable %uvec3ptr Input\n"
 788                 "%zero      = OpConstant %i32 0\n"
 789
 790                 "OpLine %fname1 1 1\n" // Before a function
 791
 792                 "%main      = OpFunction %void None %voidf\n"
 793                 "%label     = OpLabel\n"
 794
 795                 "OpLine %fname1 1 1\n" // In a function
 796
 797                 "%idval     = OpLoad %uvec3 %id\n"
 798                 "%x         = OpCompositeExtract %u32 %idval 0\n"
 799                 "%inloc     = OpAccessChain %f32ptr %indata %zero %x\n"
 800                 "%inval     = OpLoad %f32 %inloc\n"
 801                 "%neg       = OpFNegate %f32 %inval\n"
 802                 "%outloc    = OpAccessChain %f32ptr %outdata %zero %x\n"
 803                 "             OpStore %outloc %neg\n"
 804                 "             OpReturn\n"
 805                 "             OpFunctionEnd\n";
 806         spec.inputs.push_back(BufferSp(new Float32Buffer(positiveFloats)));
 807         spec.outputs.push_back(BufferSp(new Float32Buffer(negativeFloats)));
 808         spec.numWorkGroups = IVec3(numElements, 1, 1);
 809
 810         group->addChild(new SpvAsmComputeShaderCase(testCtx, "all", "OpLine appearing at different places", spec));
 811
 812         return group.release();
 813 }
 814
 815 bool veryfiBinaryShader (const ProgramBinary& binary)
 816 {
 817         const size_t    paternCount                     = 3u;
 818         bool paternsCheck[paternCount]          =
 819         {
 820                 false, false, false
 821         };
 822         const string patersns[paternCount]      =
 823         {
 824                 "VULKAN CTS",
 825                 "Negative values",
 826                 "Date: 2017/09/21"
 827         };
 828         size_t                  paternNdx               = 0u;
 829
 830         for (size_t ndx = 0u; ndx < binary.getSize(); ++ndx)
 831         {
 832                 if (false == paternsCheck[paternNdx] &&
 833                         patersns[paternNdx][0] == static_cast<char>(binary.getBinary()[ndx]) &&
 834                         deMemoryEqual((const char*)&binary.getBinary()[ndx], &patersns[paternNdx][0], patersns[paternNdx].length()))
 835                 {
 836                         paternsCheck[paternNdx]= true;
 837                         paternNdx++;
 838                         if (paternNdx == paternCount)
 839                                 break;
 840                 }
 841         }
 842
 843         for (size_t ndx = 0u; ndx < paternCount; ++ndx)
 844         {
 845                 if (!paternsCheck[ndx])
 846                         return false;
 847         }
 848
 849         return true;
 850 }
 851
 852 tcu::TestCaseGroup* createOpModuleProcessedGroup (tcu::TestContext& testCtx)
 853 {
 854         de::MovePtr<tcu::TestCaseGroup> group                   (new tcu::TestCaseGroup(testCtx, "opmoduleprocessed", "Test the OpModuleProcessed instruction"));
 855         ComputeShaderSpec                               spec;
 856         de::Random                                              rnd                             (deStringHash(group->getName()));
 857         const int                                               numElements             = 10;
 858         vector<float>                                   positiveFloats  (numElements, 0);
 859         vector<float>                                   negativeFloats  (numElements, 0);
 860
 861         fillRandomScalars(rnd, 1.f, 100.f, &positiveFloats[0], numElements);
 862
 863         for (size_t ndx = 0; ndx < numElements; ++ndx)
 864                 negativeFloats[ndx] = -positiveFloats[ndx];
 865
 866         spec.assembly =
 867                 string(getComputeAsmShaderPreamble()) +
 868                 "%fname = OpString \"negateInputs.comp\"\n"
 869
 870                 "OpSource GLSL 430\n"
 871                 "OpName %main           \"main\"\n"
 872                 "OpName %id             \"gl_GlobalInvocationID\"\n"
 873                 "OpModuleProcessed \"VULKAN CTS\"\n"                                    //OpModuleProcessed;
 874                 "OpModuleProcessed \"Negative values\"\n"
 875                 "OpModuleProcessed \"Date: 2017/09/21\"\n"
 876                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
 877
 878                 + string(getComputeAsmInputOutputBufferTraits())
 879
 880                 + string(getComputeAsmCommonTypes()) + string(getComputeAsmInputOutputBuffer()) +
 881
 882                 "OpLine %fname 0 1\n"
 883
 884                 "OpLine %fname 1000 1\n"
 885
 886                 "%id        = OpVariable %uvec3ptr Input\n"
 887                 "%zero      = OpConstant %i32 0\n"
 888                 "%main      = OpFunction %void None %voidf\n"
 889
 890                 "%label     = OpLabel\n"
 891                 "%idval     = OpLoad %uvec3 %id\n"
 892                 "%x         = OpCompositeExtract %u32 %idval 0\n"
 893
 894                 "%inloc     = OpAccessChain %f32ptr %indata %zero %x\n"
 895                 "%inval     = OpLoad %f32 %inloc\n"
 896                 "%neg       = OpFNegate %f32 %inval\n"
 897                 "%outloc    = OpAccessChain %f32ptr %outdata %zero %x\n"
 898                 "             OpStore %outloc %neg\n"
 899                 "             OpReturn\n"
 900                 "             OpFunctionEnd\n";
 901         spec.inputs.push_back(BufferSp(new Float32Buffer(positiveFloats)));
 902         spec.outputs.push_back(BufferSp(new Float32Buffer(negativeFloats)));
 903         spec.numWorkGroups = IVec3(numElements, 1, 1);
 904         spec.verifyBinary = veryfiBinaryShader;
 905         spec.spirvVersion = SPIRV_VERSION_1_3;
 906
 907         group->addChild(new SpvAsmComputeShaderCase(testCtx, "all", "OpModuleProcessed Tests", spec));
 908
 909         return group.release();
 910 }
 911
 912 tcu::TestCaseGroup* createOpNoLineGroup (tcu::TestContext& testCtx)
 913 {
 914         de::MovePtr<tcu::TestCaseGroup> group                   (new tcu::TestCaseGroup(testCtx, "opnoline", "Test the OpNoLine instruction"));
 915         ComputeShaderSpec                               spec;
 916         de::Random                                              rnd                             (deStringHash(group->getName()));
 917         const int                                               numElements             = 100;
 918         vector<float>                                   positiveFloats  (numElements, 0);
 919         vector<float>                                   negativeFloats  (numElements, 0);
 920
 921         fillRandomScalars(rnd, 1.f, 100.f, &positiveFloats[0], numElements);
 922
 923         for (size_t ndx = 0; ndx < numElements; ++ndx)
 924                 negativeFloats[ndx] = -positiveFloats[ndx];
 925
 926         spec.assembly =
 927                 string(getComputeAsmShaderPreamble()) +
 928
 929                 "%fname = OpString \"negateInputs.comp\"\n"
 930
 931                 "OpSource GLSL 430\n"
 932                 "OpName %main           \"main\"\n"
 933                 "OpName %id             \"gl_GlobalInvocationID\"\n"
 934
 935                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
 936
 937                 + string(getComputeAsmInputOutputBufferTraits()) +
 938
 939                 "OpNoLine\n" // At the earliest possible position, without preceding OpLine
 940
 941                 + string(getComputeAsmCommonTypes()) + string(getComputeAsmInputOutputBuffer()) +
 942
 943                 "OpLine %fname 0 1\n"
 944                 "OpNoLine\n" // Immediately following a preceding OpLine
 945
 946                 "OpLine %fname 1000 1\n"
 947
 948                 "%id        = OpVariable %uvec3ptr Input\n"
 949                 "%zero      = OpConstant %i32 0\n"
 950
 951                 "OpNoLine\n" // Contents after the previous OpLine
 952
 953                 "%main      = OpFunction %void None %voidf\n"
 954                 "%label     = OpLabel\n"
 955                 "%idval     = OpLoad %uvec3 %id\n"
 956                 "%x         = OpCompositeExtract %u32 %idval 0\n"
 957
 958                 "OpNoLine\n" // Multiple OpNoLine
 959                 "OpNoLine\n"
 960                 "OpNoLine\n"
 961
 962                 "%inloc     = OpAccessChain %f32ptr %indata %zero %x\n"
 963                 "%inval     = OpLoad %f32 %inloc\n"
 964                 "%neg       = OpFNegate %f32 %inval\n"
 965                 "%outloc    = OpAccessChain %f32ptr %outdata %zero %x\n"
 966                 "             OpStore %outloc %neg\n"
 967                 "             OpReturn\n"
 968                 "             OpFunctionEnd\n";
 969         spec.inputs.push_back(BufferSp(new Float32Buffer(positiveFloats)));
 970         spec.outputs.push_back(BufferSp(new Float32Buffer(negativeFloats)));
 971         spec.numWorkGroups = IVec3(numElements, 1, 1);
 972
 973         group->addChild(new SpvAsmComputeShaderCase(testCtx, "all", "OpNoLine appearing at different places", spec));
 974
 975         return group.release();
 976 }
 977
 978 // Compare instruction for the contraction compute case.
 979 // Returns true if the output is what is expected from the test case.
 980 bool compareNoContractCase(const std::vector<Resource>&, const vector<AllocationSp>& outputAllocs, const std::vector<Resource>& expectedOutputs, TestLog&)
 981 {
 982         if (outputAllocs.size() != 1)
 983                 return false;
 984
 985         // Only size is needed because we are not comparing the exact values.
 986         size_t byteSize = expectedOutputs[0].getByteSize();
 987
 988         const float*    outputAsFloat   = static_cast<const float*>(outputAllocs[0]->getHostPtr());
 989
 990         for(size_t i = 0; i < byteSize / sizeof(float); ++i) {
 991                 if (outputAsFloat[i] != 0.f &&
 992                         outputAsFloat[i] != -ldexp(1, -24)) {
 993                         return false;
 994                 }
 995         }
 996
 997         return true;
 998 }
 999
1000 tcu::TestCaseGroup* createNoContractionGroup (tcu::TestContext& testCtx)
1001 {
1002         de::MovePtr<tcu::TestCaseGroup> group                   (new tcu::TestCaseGroup(testCtx, "nocontraction", "Test the NoContraction decoration"));
1003         vector<CaseParameter>                   cases;
1004         const int                                               numElements             = 100;
1005         vector<float>                                   inputFloats1    (numElements, 0);
1006         vector<float>                                   inputFloats2    (numElements, 0);
1007         vector<float>                                   outputFloats    (numElements, 0);
1008         const StringTemplate                    shaderTemplate  (
1009                 string(getComputeAsmShaderPreamble()) +
1010
1011                 "OpName %main           \"main\"\n"
1012                 "OpName %id             \"gl_GlobalInvocationID\"\n"
1013
1014                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
1015
1016                 "${DECORATION}\n"
1017
1018                 "OpDecorate %buf BufferBlock\n"
1019                 "OpDecorate %indata1 DescriptorSet 0\n"
1020                 "OpDecorate %indata1 Binding 0\n"
1021                 "OpDecorate %indata2 DescriptorSet 0\n"
1022                 "OpDecorate %indata2 Binding 1\n"
1023                 "OpDecorate %outdata DescriptorSet 0\n"
1024                 "OpDecorate %outdata Binding 2\n"
1025                 "OpDecorate %f32arr ArrayStride 4\n"
1026                 "OpMemberDecorate %buf 0 Offset 0\n"
1027
1028                 + string(getComputeAsmCommonTypes()) +
1029
1030                 "%buf        = OpTypeStruct %f32arr\n"
1031                 "%bufptr     = OpTypePointer Uniform %buf\n"
1032                 "%indata1    = OpVariable %bufptr Uniform\n"
1033                 "%indata2    = OpVariable %bufptr Uniform\n"
1034                 "%outdata    = OpVariable %bufptr Uniform\n"
1035
1036                 "%id         = OpVariable %uvec3ptr Input\n"
1037                 "%zero       = OpConstant %i32 0\n"
1038                 "%c_f_m1     = OpConstant %f32 -1.\n"
1039
1040                 "%main       = OpFunction %void None %voidf\n"
1041                 "%label      = OpLabel\n"
1042                 "%idval      = OpLoad %uvec3 %id\n"
1043                 "%x          = OpCompositeExtract %u32 %idval 0\n"
1044                 "%inloc1     = OpAccessChain %f32ptr %indata1 %zero %x\n"
1045                 "%inval1     = OpLoad %f32 %inloc1\n"
1046                 "%inloc2     = OpAccessChain %f32ptr %indata2 %zero %x\n"
1047                 "%inval2     = OpLoad %f32 %inloc2\n"
1048                 "%mul        = OpFMul %f32 %inval1 %inval2\n"
1049                 "%add        = OpFAdd %f32 %mul %c_f_m1\n"
1050                 "%outloc     = OpAccessChain %f32ptr %outdata %zero %x\n"
1051                 "              OpStore %outloc %add\n"
1052                 "              OpReturn\n"
1053                 "              OpFunctionEnd\n");
1054
1055         cases.push_back(CaseParameter("multiplication", "OpDecorate %mul NoContraction"));
1056         cases.push_back(CaseParameter("addition",               "OpDecorate %add NoContraction"));
1057         cases.push_back(CaseParameter("both",                   "OpDecorate %mul NoContraction\nOpDecorate %add NoContraction"));
1058
1059         for (size_t ndx = 0; ndx < numElements; ++ndx)
1060         {
1061                 inputFloats1[ndx]       = 1.f + std::ldexp(1.f, -23); // 1 + 2^-23.
1062                 inputFloats2[ndx]       = 1.f - std::ldexp(1.f, -23); // 1 - 2^-23.
1063                 // Result for (1 + 2^-23) * (1 - 2^-23) - 1. With NoContraction, the multiplication will be
1064                 // conducted separately and the result is rounded to 1, or 0x1.fffffcp-1
1065                 // So the final result will be 0.f or 0x1p-24.
1066                 // If the operation is combined into a precise fused multiply-add, then the result would be
1067                 // 2^-46 (0xa8800000).
1068                 outputFloats[ndx]       = 0.f;
1069         }
1070
1071         for (size_t caseNdx = 0; caseNdx < cases.size(); ++caseNdx)
1072         {
1073                 map<string, string>             specializations;
1074                 ComputeShaderSpec               spec;
1075
1076                 specializations["DECORATION"] = cases[caseNdx].param;
1077                 spec.assembly = shaderTemplate.specialize(specializations);
1078                 spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats1)));
1079                 spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats2)));
1080                 spec.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
1081                 spec.numWorkGroups = IVec3(numElements, 1, 1);
1082                 // Check against the two possible answers based on rounding mode.
1083                 spec.verifyIO = &compareNoContractCase;
1084
1085                 group->addChild(new SpvAsmComputeShaderCase(testCtx, cases[caseNdx].name, cases[caseNdx].name, spec));
1086         }
1087         return group.release();
1088 }
1089
1090 bool compareFRem(const std::vector<Resource>&, const vector<AllocationSp>& outputAllocs, const std::vector<Resource>& expectedOutputs, TestLog&)
1091 {
1092         if (outputAllocs.size() != 1)
1093                 return false;
1094
1095         vector<deUint8> expectedBytes;
1096         expectedOutputs[0].getBytes(expectedBytes);
1097
1098         const float*    expectedOutputAsFloat   = reinterpret_cast<const float*>(&expectedBytes.front());
1099         const float*    outputAsFloat                   = static_cast<const float*>(outputAllocs[0]->getHostPtr());
1100
1101         for (size_t idx = 0; idx < expectedBytes.size() / sizeof(float); ++idx)
1102         {
1103                 const float f0 = expectedOutputAsFloat[idx];
1104                 const float f1 = outputAsFloat[idx];
1105                 // \todo relative error needs to be fairly high because FRem may be implemented as
1106                 // (roughly) frac(a/b)*b, so LSB errors can be magnified. But this should be fine for now.
1107                 if (deFloatAbs((f1 - f0) / f0) > 0.02)
1108                         return false;
1109         }
1110
1111         return true;
1112 }
1113
1114 tcu::TestCaseGroup* createOpFRemGroup (tcu::TestContext& testCtx)
1115 {
1116         de::MovePtr<tcu::TestCaseGroup> group                   (new tcu::TestCaseGroup(testCtx, "opfrem", "Test the OpFRem instruction"));
1117         ComputeShaderSpec                               spec;
1118         de::Random                                              rnd                             (deStringHash(group->getName()));
1119         const int                                               numElements             = 200;
1120         vector<float>                                   inputFloats1    (numElements, 0);
1121         vector<float>                                   inputFloats2    (numElements, 0);
1122         vector<float>                                   outputFloats    (numElements, 0);
1123
1124         fillRandomScalars(rnd, -10000.f, 10000.f, &inputFloats1[0], numElements);
1125         fillRandomScalars(rnd, -100.f, 100.f, &inputFloats2[0], numElements);
1126
1127         for (size_t ndx = 0; ndx < numElements; ++ndx)
1128         {
1129                 // Guard against divisors near zero.
1130                 if (std::fabs(inputFloats2[ndx]) < 1e-3)
1131                         inputFloats2[ndx] = 8.f;
1132
1133                 // The return value of std::fmod() has the same sign as its first operand, which is how OpFRem spec'd.
1134                 outputFloats[ndx] = std::fmod(inputFloats1[ndx], inputFloats2[ndx]);
1135         }
1136
1137         spec.assembly =
1138                 string(getComputeAsmShaderPreamble()) +
1139
1140                 "OpName %main           \"main\"\n"
1141                 "OpName %id             \"gl_GlobalInvocationID\"\n"
1142
1143                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
1144
1145                 "OpDecorate %buf BufferBlock\n"
1146                 "OpDecorate %indata1 DescriptorSet 0\n"
1147                 "OpDecorate %indata1 Binding 0\n"
1148                 "OpDecorate %indata2 DescriptorSet 0\n"
1149                 "OpDecorate %indata2 Binding 1\n"
1150                 "OpDecorate %outdata DescriptorSet 0\n"
1151                 "OpDecorate %outdata Binding 2\n"
1152                 "OpDecorate %f32arr ArrayStride 4\n"
1153                 "OpMemberDecorate %buf 0 Offset 0\n"
1154
1155                 + string(getComputeAsmCommonTypes()) +
1156
1157                 "%buf        = OpTypeStruct %f32arr\n"
1158                 "%bufptr     = OpTypePointer Uniform %buf\n"
1159                 "%indata1    = OpVariable %bufptr Uniform\n"
1160                 "%indata2    = OpVariable %bufptr Uniform\n"
1161                 "%outdata    = OpVariable %bufptr Uniform\n"
1162
1163                 "%id        = OpVariable %uvec3ptr Input\n"
1164                 "%zero      = OpConstant %i32 0\n"
1165
1166                 "%main      = OpFunction %void None %voidf\n"
1167                 "%label     = OpLabel\n"
1168                 "%idval     = OpLoad %uvec3 %id\n"
1169                 "%x         = OpCompositeExtract %u32 %idval 0\n"
1170                 "%inloc1    = OpAccessChain %f32ptr %indata1 %zero %x\n"
1171                 "%inval1    = OpLoad %f32 %inloc1\n"
1172                 "%inloc2    = OpAccessChain %f32ptr %indata2 %zero %x\n"
1173                 "%inval2    = OpLoad %f32 %inloc2\n"
1174                 "%rem       = OpFRem %f32 %inval1 %inval2\n"
1175                 "%outloc    = OpAccessChain %f32ptr %outdata %zero %x\n"
1176                 "             OpStore %outloc %rem\n"
1177                 "             OpReturn\n"
1178                 "             OpFunctionEnd\n";
1179
1180         spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats1)));
1181         spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats2)));
1182         spec.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
1183         spec.numWorkGroups = IVec3(numElements, 1, 1);
1184         spec.verifyIO = &compareFRem;
1185
1186         group->addChild(new SpvAsmComputeShaderCase(testCtx, "all", "", spec));
1187
1188         return group.release();
1189 }
1190
1191 bool compareNMin (const std::vector<Resource>&, const vector<AllocationSp>& outputAllocs, const std::vector<Resource>& expectedOutputs, TestLog&)
1192 {
1193         if (outputAllocs.size() != 1)
1194                 return false;
1195
1196         const BufferSp&                 expectedOutput                  (expectedOutputs[0].getBuffer());
1197         std::vector<deUint8>    data;
1198         expectedOutput->getBytes(data);
1199
1200         const float* const              expectedOutputAsFloat   = reinterpret_cast<const float*>(&data.front());
1201         const float* const              outputAsFloat                   = static_cast<const float*>(outputAllocs[0]->getHostPtr());
1202
1203         for (size_t idx = 0; idx < expectedOutput->getByteSize() / sizeof(float); ++idx)
1204         {
1205                 const float f0 = expectedOutputAsFloat[idx];
1206                 const float f1 = outputAsFloat[idx];
1207
1208                 // For NMin, we accept NaN as output if both inputs were NaN.
1209                 // Otherwise the NaN is the wrong choise, as on architectures that
1210                 // do not handle NaN, those are huge values.
1211                 if (!(tcu::Float32(f1).isNaN() && tcu::Float32(f0).isNaN()) && deFloatAbs(f1 - f0) > 0.00001f)
1212                         return false;
1213         }
1214
1215         return true;
1216 }
1217
1218 tcu::TestCaseGroup* createOpNMinGroup (tcu::TestContext& testCtx)
1219 {
1220         de::MovePtr<tcu::TestCaseGroup> group                   (new tcu::TestCaseGroup(testCtx, "opnmin", "Test the OpNMin instruction"));
1221         ComputeShaderSpec                               spec;
1222         de::Random                                              rnd                             (deStringHash(group->getName()));
1223         const int                                               numElements             = 200;
1224         vector<float>                                   inputFloats1    (numElements, 0);
1225         vector<float>                                   inputFloats2    (numElements, 0);
1226         vector<float>                                   outputFloats    (numElements, 0);
1227
1228         fillRandomScalars(rnd, -10000.f, 10000.f, &inputFloats1[0], numElements);
1229         fillRandomScalars(rnd, -10000.f, 10000.f, &inputFloats2[0], numElements);
1230
1231         // Make the first case a full-NAN case.
1232         inputFloats1[0] = TCU_NAN;
1233         inputFloats2[0] = TCU_NAN;
1234
1235         for (size_t ndx = 0; ndx < numElements; ++ndx)
1236         {
1237                 // By default, pick the smallest
1238                 outputFloats[ndx] = std::min(inputFloats1[ndx], inputFloats2[ndx]);
1239
1240                 // Make half of the cases NaN cases
1241                 if ((ndx & 1) == 0)
1242                 {
1243                         // Alternate between the NaN operand
1244                         if ((ndx & 2) == 0)
1245                         {
1246                                 outputFloats[ndx] = inputFloats2[ndx];
1247                                 inputFloats1[ndx] = TCU_NAN;
1248                         }
1249                         else
1250                         {
1251                                 outputFloats[ndx] = inputFloats1[ndx];
1252                                 inputFloats2[ndx] = TCU_NAN;
1253                         }
1254                 }
1255         }
1256
1257         spec.assembly =
1258                 "OpCapability Shader\n"
1259                 "%std450        = OpExtInstImport \"GLSL.std.450\"\n"
1260                 "OpMemoryModel Logical GLSL450\n"
1261                 "OpEntryPoint GLCompute %main \"main\" %id\n"
1262                 "OpExecutionMode %main LocalSize 1 1 1\n"
1263
1264                 "OpName %main           \"main\"\n"
1265                 "OpName %id             \"gl_GlobalInvocationID\"\n"
1266
1267                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
1268
1269                 "OpDecorate %buf BufferBlock\n"
1270                 "OpDecorate %indata1 DescriptorSet 0\n"
1271                 "OpDecorate %indata1 Binding 0\n"
1272                 "OpDecorate %indata2 DescriptorSet 0\n"
1273                 "OpDecorate %indata2 Binding 1\n"
1274                 "OpDecorate %outdata DescriptorSet 0\n"
1275                 "OpDecorate %outdata Binding 2\n"
1276                 "OpDecorate %f32arr ArrayStride 4\n"
1277                 "OpMemberDecorate %buf 0 Offset 0\n"
1278
1279                 + string(getComputeAsmCommonTypes()) +
1280
1281                 "%buf        = OpTypeStruct %f32arr\n"
1282                 "%bufptr     = OpTypePointer Uniform %buf\n"
1283                 "%indata1    = OpVariable %bufptr Uniform\n"
1284                 "%indata2    = OpVariable %bufptr Uniform\n"
1285                 "%outdata    = OpVariable %bufptr Uniform\n"
1286
1287                 "%id        = OpVariable %uvec3ptr Input\n"
1288                 "%zero      = OpConstant %i32 0\n"
1289
1290                 "%main      = OpFunction %void None %voidf\n"
1291                 "%label     = OpLabel\n"
1292                 "%idval     = OpLoad %uvec3 %id\n"
1293                 "%x         = OpCompositeExtract %u32 %idval 0\n"
1294                 "%inloc1    = OpAccessChain %f32ptr %indata1 %zero %x\n"
1295                 "%inval1    = OpLoad %f32 %inloc1\n"
1296                 "%inloc2    = OpAccessChain %f32ptr %indata2 %zero %x\n"
1297                 "%inval2    = OpLoad %f32 %inloc2\n"
1298                 "%rem       = OpExtInst %f32 %std450 NMin %inval1 %inval2\n"
1299                 "%outloc    = OpAccessChain %f32ptr %outdata %zero %x\n"
1300                 "             OpStore %outloc %rem\n"
1301                 "             OpReturn\n"
1302                 "             OpFunctionEnd\n";
1303
1304         spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats1)));
1305         spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats2)));
1306         spec.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
1307         spec.numWorkGroups = IVec3(numElements, 1, 1);
1308         spec.verifyIO = &compareNMin;
1309
1310         group->addChild(new SpvAsmComputeShaderCase(testCtx, "all", "", spec));
1311
1312         return group.release();
1313 }
1314
1315 bool compareNMax (const std::vector<Resource>&, const vector<AllocationSp>& outputAllocs, const std::vector<Resource>& expectedOutputs, TestLog&)
1316 {
1317         if (outputAllocs.size() != 1)
1318                 return false;
1319
1320         const BufferSp&                 expectedOutput                  = expectedOutputs[0].getBuffer();
1321         std::vector<deUint8>    data;
1322         expectedOutput->getBytes(data);
1323
1324         const float* const              expectedOutputAsFloat   = reinterpret_cast<const float*>(&data.front());
1325         const float* const              outputAsFloat                   = static_cast<const float*>(outputAllocs[0]->getHostPtr());
1326
1327         for (size_t idx = 0; idx < expectedOutput->getByteSize() / sizeof(float); ++idx)
1328         {
1329                 const float f0 = expectedOutputAsFloat[idx];
1330                 const float f1 = outputAsFloat[idx];
1331
1332                 // For NMax, NaN is considered acceptable result, since in
1333                 // architectures that do not handle NaNs, those are huge values.
1334                 if (!tcu::Float32(f1).isNaN() && deFloatAbs(f1 - f0) > 0.00001f)
1335                         return false;
1336         }
1337
1338         return true;
1339 }
1340
1341 tcu::TestCaseGroup* createOpNMaxGroup (tcu::TestContext& testCtx)
1342 {
1343         de::MovePtr<tcu::TestCaseGroup> group(new tcu::TestCaseGroup(testCtx, "opnmax", "Test the OpNMax instruction"));
1344         ComputeShaderSpec                               spec;
1345         de::Random                                              rnd                             (deStringHash(group->getName()));
1346         const int                                               numElements             = 200;
1347         vector<float>                                   inputFloats1    (numElements, 0);
1348         vector<float>                                   inputFloats2    (numElements, 0);
1349         vector<float>                                   outputFloats    (numElements, 0);
1350
1351         fillRandomScalars(rnd, -10000.f, 10000.f, &inputFloats1[0], numElements);
1352         fillRandomScalars(rnd, -10000.f, 10000.f, &inputFloats2[0], numElements);
1353
1354         // Make the first case a full-NAN case.
1355         inputFloats1[0] = TCU_NAN;
1356         inputFloats2[0] = TCU_NAN;
1357
1358         for (size_t ndx = 0; ndx < numElements; ++ndx)
1359         {
1360                 // By default, pick the biggest
1361                 outputFloats[ndx] = std::max(inputFloats1[ndx], inputFloats2[ndx]);
1362
1363                 // Make half of the cases NaN cases
1364                 if ((ndx & 1) == 0)
1365                 {
1366                         // Alternate between the NaN operand
1367                         if ((ndx & 2) == 0)
1368                         {
1369                                 outputFloats[ndx] = inputFloats2[ndx];
1370                                 inputFloats1[ndx] = TCU_NAN;
1371                         }
1372                         else
1373                         {
1374                                 outputFloats[ndx] = inputFloats1[ndx];
1375                                 inputFloats2[ndx] = TCU_NAN;
1376                         }
1377                 }
1378         }
1379
1380         spec.assembly =
1381                 "OpCapability Shader\n"
1382                 "%std450        = OpExtInstImport \"GLSL.std.450\"\n"
1383                 "OpMemoryModel Logical GLSL450\n"
1384                 "OpEntryPoint GLCompute %main \"main\" %id\n"
1385                 "OpExecutionMode %main LocalSize 1 1 1\n"
1386
1387                 "OpName %main           \"main\"\n"
1388                 "OpName %id             \"gl_GlobalInvocationID\"\n"
1389
1390                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
1391
1392                 "OpDecorate %buf BufferBlock\n"
1393                 "OpDecorate %indata1 DescriptorSet 0\n"
1394                 "OpDecorate %indata1 Binding 0\n"
1395                 "OpDecorate %indata2 DescriptorSet 0\n"
1396                 "OpDecorate %indata2 Binding 1\n"
1397                 "OpDecorate %outdata DescriptorSet 0\n"
1398                 "OpDecorate %outdata Binding 2\n"
1399                 "OpDecorate %f32arr ArrayStride 4\n"
1400                 "OpMemberDecorate %buf 0 Offset 0\n"
1401
1402                 + string(getComputeAsmCommonTypes()) +
1403
1404                 "%buf        = OpTypeStruct %f32arr\n"
1405                 "%bufptr     = OpTypePointer Uniform %buf\n"
1406                 "%indata1    = OpVariable %bufptr Uniform\n"
1407                 "%indata2    = OpVariable %bufptr Uniform\n"
1408                 "%outdata    = OpVariable %bufptr Uniform\n"
1409
1410                 "%id        = OpVariable %uvec3ptr Input\n"
1411                 "%zero      = OpConstant %i32 0\n"
1412
1413                 "%main      = OpFunction %void None %voidf\n"
1414                 "%label     = OpLabel\n"
1415                 "%idval     = OpLoad %uvec3 %id\n"
1416                 "%x         = OpCompositeExtract %u32 %idval 0\n"
1417                 "%inloc1    = OpAccessChain %f32ptr %indata1 %zero %x\n"
1418                 "%inval1    = OpLoad %f32 %inloc1\n"
1419                 "%inloc2    = OpAccessChain %f32ptr %indata2 %zero %x\n"
1420                 "%inval2    = OpLoad %f32 %inloc2\n"
1421                 "%rem       = OpExtInst %f32 %std450 NMax %inval1 %inval2\n"
1422                 "%outloc    = OpAccessChain %f32ptr %outdata %zero %x\n"
1423                 "             OpStore %outloc %rem\n"
1424                 "             OpReturn\n"
1425                 "             OpFunctionEnd\n";
1426
1427         spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats1)));
1428         spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats2)));
1429         spec.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
1430         spec.numWorkGroups = IVec3(numElements, 1, 1);
1431         spec.verifyIO = &compareNMax;
1432
1433         group->addChild(new SpvAsmComputeShaderCase(testCtx, "all", "", spec));
1434
1435         return group.release();
1436 }
1437
1438 bool compareNClamp (const std::vector<Resource>&, const vector<AllocationSp>& outputAllocs, const std::vector<Resource>& expectedOutputs, TestLog&)
1439 {
1440         if (outputAllocs.size() != 1)
1441                 return false;
1442
1443         const BufferSp&                 expectedOutput                  = expectedOutputs[0].getBuffer();
1444         std::vector<deUint8>    data;
1445         expectedOutput->getBytes(data);
1446
1447         const float* const              expectedOutputAsFloat   = reinterpret_cast<const float*>(&data.front());
1448         const float* const              outputAsFloat                   = static_cast<const float*>(outputAllocs[0]->getHostPtr());
1449
1450         for (size_t idx = 0; idx < expectedOutput->getByteSize() / sizeof(float) / 2; ++idx)
1451         {
1452                 const float e0 = expectedOutputAsFloat[idx * 2];
1453                 const float e1 = expectedOutputAsFloat[idx * 2 + 1];
1454                 const float res = outputAsFloat[idx];
1455
1456                 // For NClamp, we have two possible outcomes based on
1457                 // whether NaNs are handled or not.
1458                 // If either min or max value is NaN, the result is undefined,
1459                 // so this test doesn't stress those. If the clamped value is
1460                 // NaN, and NaNs are handled, the result is min; if NaNs are not
1461                 // handled, they are big values that result in max.
1462                 // If all three parameters are NaN, the result should be NaN.
1463                 if (!((tcu::Float32(e0).isNaN() && tcu::Float32(res).isNaN()) ||
1464                          (deFloatAbs(e0 - res) < 0.00001f) ||
1465                          (deFloatAbs(e1 - res) < 0.00001f)))
1466                         return false;
1467         }
1468
1469         return true;
1470 }
1471
1472 tcu::TestCaseGroup* createOpNClampGroup (tcu::TestContext& testCtx)
1473 {
1474         de::MovePtr<tcu::TestCaseGroup> group                   (new tcu::TestCaseGroup(testCtx, "opnclamp", "Test the OpNClamp instruction"));
1475         ComputeShaderSpec                               spec;
1476         de::Random                                              rnd                             (deStringHash(group->getName()));
1477         const int                                               numElements             = 200;
1478         vector<float>                                   inputFloats1    (numElements, 0);
1479         vector<float>                                   inputFloats2    (numElements, 0);
1480         vector<float>                                   inputFloats3    (numElements, 0);
1481         vector<float>                                   outputFloats    (numElements * 2, 0);
1482
1483         fillRandomScalars(rnd, -10000.f, 10000.f, &inputFloats1[0], numElements);
1484         fillRandomScalars(rnd, -10000.f, 10000.f, &inputFloats2[0], numElements);
1485         fillRandomScalars(rnd, -10000.f, 10000.f, &inputFloats3[0], numElements);
1486
1487         for (size_t ndx = 0; ndx < numElements; ++ndx)
1488         {
1489                 // Results are only defined if max value is bigger than min value.
1490                 if (inputFloats2[ndx] > inputFloats3[ndx])
1491                 {
1492                         float t = inputFloats2[ndx];
1493                         inputFloats2[ndx] = inputFloats3[ndx];
1494                         inputFloats3[ndx] = t;
1495                 }
1496
1497                 // By default, do the clamp, setting both possible answers
1498                 float defaultRes = std::min(std::max(inputFloats1[ndx], inputFloats2[ndx]), inputFloats3[ndx]);
1499
1500                 float maxResA = std::max(inputFloats1[ndx], inputFloats2[ndx]);
1501                 float maxResB = maxResA;
1502
1503                 // Alternate between the NaN cases
1504                 if (ndx & 1)
1505                 {
1506                         inputFloats1[ndx] = TCU_NAN;
1507                         // If NaN is handled, the result should be same as the clamp minimum.
1508                         // If NaN is not handled, the result should clamp to the clamp maximum.
1509                         maxResA = inputFloats2[ndx];
1510                         maxResB = inputFloats3[ndx];
1511                 }
1512                 else
1513                 {
1514                         // Not a NaN case - only one legal result.
1515                         maxResA = defaultRes;
1516                         maxResB = defaultRes;
1517                 }
1518
1519                 outputFloats[ndx * 2] = maxResA;
1520                 outputFloats[ndx * 2 + 1] = maxResB;
1521         }
1522
1523         // Make the first case a full-NAN case.
1524         inputFloats1[0] = TCU_NAN;
1525         inputFloats2[0] = TCU_NAN;
1526         inputFloats3[0] = TCU_NAN;
1527         outputFloats[0] = TCU_NAN;
1528         outputFloats[1] = TCU_NAN;
1529
1530         spec.assembly =
1531                 "OpCapability Shader\n"
1532                 "%std450        = OpExtInstImport \"GLSL.std.450\"\n"
1533                 "OpMemoryModel Logical GLSL450\n"
1534                 "OpEntryPoint GLCompute %main \"main\" %id\n"
1535                 "OpExecutionMode %main LocalSize 1 1 1\n"
1536
1537                 "OpName %main           \"main\"\n"
1538                 "OpName %id             \"gl_GlobalInvocationID\"\n"
1539
1540                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
1541
1542                 "OpDecorate %buf BufferBlock\n"
1543                 "OpDecorate %indata1 DescriptorSet 0\n"
1544                 "OpDecorate %indata1 Binding 0\n"
1545                 "OpDecorate %indata2 DescriptorSet 0\n"
1546                 "OpDecorate %indata2 Binding 1\n"
1547                 "OpDecorate %indata3 DescriptorSet 0\n"
1548                 "OpDecorate %indata3 Binding 2\n"
1549                 "OpDecorate %outdata DescriptorSet 0\n"
1550                 "OpDecorate %outdata Binding 3\n"
1551                 "OpDecorate %f32arr ArrayStride 4\n"
1552                 "OpMemberDecorate %buf 0 Offset 0\n"
1553
1554                 + string(getComputeAsmCommonTypes()) +
1555
1556                 "%buf        = OpTypeStruct %f32arr\n"
1557                 "%bufptr     = OpTypePointer Uniform %buf\n"
1558                 "%indata1    = OpVariable %bufptr Uniform\n"
1559                 "%indata2    = OpVariable %bufptr Uniform\n"
1560                 "%indata3    = OpVariable %bufptr Uniform\n"
1561                 "%outdata    = OpVariable %bufptr Uniform\n"
1562
1563                 "%id        = OpVariable %uvec3ptr Input\n"
1564                 "%zero      = OpConstant %i32 0\n"
1565
1566                 "%main      = OpFunction %void None %voidf\n"
1567                 "%label     = OpLabel\n"
1568                 "%idval     = OpLoad %uvec3 %id\n"
1569                 "%x         = OpCompositeExtract %u32 %idval 0\n"
1570                 "%inloc1    = OpAccessChain %f32ptr %indata1 %zero %x\n"
1571                 "%inval1    = OpLoad %f32 %inloc1\n"
1572                 "%inloc2    = OpAccessChain %f32ptr %indata2 %zero %x\n"
1573                 "%inval2    = OpLoad %f32 %inloc2\n"
1574                 "%inloc3    = OpAccessChain %f32ptr %indata3 %zero %x\n"
1575                 "%inval3    = OpLoad %f32 %inloc3\n"
1576                 "%rem       = OpExtInst %f32 %std450 NClamp %inval1 %inval2 %inval3\n"
1577                 "%outloc    = OpAccessChain %f32ptr %outdata %zero %x\n"
1578                 "             OpStore %outloc %rem\n"
1579                 "             OpReturn\n"
1580                 "             OpFunctionEnd\n";
1581
1582         spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats1)));
1583         spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats2)));
1584         spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats3)));
1585         spec.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
1586         spec.numWorkGroups = IVec3(numElements, 1, 1);
1587         spec.verifyIO = &compareNClamp;
1588
1589         group->addChild(new SpvAsmComputeShaderCase(testCtx, "all", "", spec));
1590
1591         return group.release();
1592 }
1593
1594 tcu::TestCaseGroup* createOpSRemComputeGroup (tcu::TestContext& testCtx, qpTestResult negFailResult)
1595 {
1596         de::MovePtr<tcu::TestCaseGroup> group                   (new tcu::TestCaseGroup(testCtx, "opsrem", "Test the OpSRem instruction"));
1597         de::Random                                              rnd                             (deStringHash(group->getName()));
1598         const int                                               numElements             = 200;
1599
1600         const struct CaseParams
1601         {
1602                 const char*             name;
1603                 const char*             failMessage;            // customized status message
1604                 qpTestResult    failResult;                     // override status on failure
1605                 int                             op1Min, op1Max;         // operand ranges
1606                 int                             op2Min, op2Max;
1607         } cases[] =
1608         {
1609                 { "positive",   "Output doesn't match with expected",                           QP_TEST_RESULT_FAIL,    0,              65536,  0,              100 },
1610                 { "all",                "Inconsistent results, but within specification",       negFailResult,                  -65536, 65536,  -100,   100 },  // see below
1611         };
1612         // If either operand is negative the result is undefined. Some implementations may still return correct values.
1613
1614         for (int caseNdx = 0; caseNdx < DE_LENGTH_OF_ARRAY(cases); ++caseNdx)
1615         {
1616                 const CaseParams&       params          = cases[caseNdx];
1617                 ComputeShaderSpec       spec;
1618                 vector<deInt32>         inputInts1      (numElements, 0);
1619                 vector<deInt32>         inputInts2      (numElements, 0);
1620                 vector<deInt32>         outputInts      (numElements, 0);
1621
1622                 fillRandomScalars(rnd, params.op1Min, params.op1Max, &inputInts1[0], numElements);
1623                 fillRandomScalars(rnd, params.op2Min, params.op2Max, &inputInts2[0], numElements, filterNotZero);
1624
1625                 for (int ndx = 0; ndx < numElements; ++ndx)
1626                 {
1627                         // The return value of std::fmod() has the same sign as its first operand, which is how OpFRem spec'd.
1628                         outputInts[ndx] = inputInts1[ndx] % inputInts2[ndx];
1629                 }
1630
1631                 spec.assembly =
1632                         string(getComputeAsmShaderPreamble()) +
1633
1634                         "OpName %main           \"main\"\n"
1635                         "OpName %id             \"gl_GlobalInvocationID\"\n"
1636
1637                         "OpDecorate %id BuiltIn GlobalInvocationId\n"
1638
1639                         "OpDecorate %buf BufferBlock\n"
1640                         "OpDecorate %indata1 DescriptorSet 0\n"
1641                         "OpDecorate %indata1 Binding 0\n"
1642                         "OpDecorate %indata2 DescriptorSet 0\n"
1643                         "OpDecorate %indata2 Binding 1\n"
1644                         "OpDecorate %outdata DescriptorSet 0\n"
1645                         "OpDecorate %outdata Binding 2\n"
1646                         "OpDecorate %i32arr ArrayStride 4\n"
1647                         "OpMemberDecorate %buf 0 Offset 0\n"
1648
1649                         + string(getComputeAsmCommonTypes()) +
1650
1651                         "%buf        = OpTypeStruct %i32arr\n"
1652                         "%bufptr     = OpTypePointer Uniform %buf\n"
1653                         "%indata1    = OpVariable %bufptr Uniform\n"
1654                         "%indata2    = OpVariable %bufptr Uniform\n"
1655                         "%outdata    = OpVariable %bufptr Uniform\n"
1656
1657                         "%id        = OpVariable %uvec3ptr Input\n"
1658                         "%zero      = OpConstant %i32 0\n"
1659
1660                         "%main      = OpFunction %void None %voidf\n"
1661                         "%label     = OpLabel\n"
1662                         "%idval     = OpLoad %uvec3 %id\n"
1663                         "%x         = OpCompositeExtract %u32 %idval 0\n"
1664                         "%inloc1    = OpAccessChain %i32ptr %indata1 %zero %x\n"
1665                         "%inval1    = OpLoad %i32 %inloc1\n"
1666                         "%inloc2    = OpAccessChain %i32ptr %indata2 %zero %x\n"
1667                         "%inval2    = OpLoad %i32 %inloc2\n"
1668                         "%rem       = OpSRem %i32 %inval1 %inval2\n"
1669                         "%outloc    = OpAccessChain %i32ptr %outdata %zero %x\n"
1670                         "             OpStore %outloc %rem\n"
1671                         "             OpReturn\n"
1672                         "             OpFunctionEnd\n";
1673
1674                 spec.inputs.push_back   (BufferSp(new Int32Buffer(inputInts1)));
1675                 spec.inputs.push_back   (BufferSp(new Int32Buffer(inputInts2)));
1676                 spec.outputs.push_back  (BufferSp(new Int32Buffer(outputInts)));
1677                 spec.numWorkGroups              = IVec3(numElements, 1, 1);
1678                 spec.failResult                 = params.failResult;
1679                 spec.failMessage                = params.failMessage;
1680
1681                 group->addChild(new SpvAsmComputeShaderCase(testCtx, params.name, "", spec));
1682         }
1683
1684         return group.release();
1685 }
1686
1687 tcu::TestCaseGroup* createOpSRemComputeGroup64 (tcu::TestContext& testCtx, qpTestResult negFailResult)
1688 {
1689         de::MovePtr<tcu::TestCaseGroup> group                   (new tcu::TestCaseGroup(testCtx, "opsrem64", "Test the 64-bit OpSRem instruction"));
1690         de::Random                                              rnd                             (deStringHash(group->getName()));
1691         const int                                               numElements             = 200;
1692
1693         const struct CaseParams
1694         {
1695                 const char*             name;
1696                 const char*             failMessage;            // customized status message
1697                 qpTestResult    failResult;                     // override status on failure
1698                 bool                    positive;
1699         } cases[] =
1700         {
1701                 { "positive",   "Output doesn't match with expected",                           QP_TEST_RESULT_FAIL,    true },
1702                 { "all",                "Inconsistent results, but within specification",       negFailResult,                  false },        // see below
1703         };
1704         // If either operand is negative the result is undefined. Some implementations may still return correct values.
1705
1706         for (int caseNdx = 0; caseNdx < DE_LENGTH_OF_ARRAY(cases); ++caseNdx)
1707         {
1708                 const CaseParams&       params          = cases[caseNdx];
1709                 ComputeShaderSpec       spec;
1710                 vector<deInt64>         inputInts1      (numElements, 0);
1711                 vector<deInt64>         inputInts2      (numElements, 0);
1712                 vector<deInt64>         outputInts      (numElements, 0);
1713
1714                 if (params.positive)
1715                 {
1716                         fillRandomInt64sLogDistributed(rnd, inputInts1, numElements, filterNonNegative);
1717                         fillRandomInt64sLogDistributed(rnd, inputInts2, numElements, filterPositive);
1718                 }
1719                 else
1720                 {
1721                         fillRandomInt64sLogDistributed(rnd, inputInts1, numElements);
1722                         fillRandomInt64sLogDistributed(rnd, inputInts2, numElements, filterNotZero);
1723                 }
1724
1725                 for (int ndx = 0; ndx < numElements; ++ndx)
1726                 {
1727                         // The return value of std::fmod() has the same sign as its first operand, which is how OpFRem spec'd.
1728                         outputInts[ndx] = inputInts1[ndx] % inputInts2[ndx];
1729                 }
1730
1731                 spec.assembly =
1732                         "OpCapability Int64\n"
1733
1734                         + string(getComputeAsmShaderPreamble()) +
1735
1736                         "OpName %main           \"main\"\n"
1737                         "OpName %id             \"gl_GlobalInvocationID\"\n"
1738
1739                         "OpDecorate %id BuiltIn GlobalInvocationId\n"
1740
1741                         "OpDecorate %buf BufferBlock\n"
1742                         "OpDecorate %indata1 DescriptorSet 0\n"
1743                         "OpDecorate %indata1 Binding 0\n"
1744                         "OpDecorate %indata2 DescriptorSet 0\n"
1745                         "OpDecorate %indata2 Binding 1\n"
1746                         "OpDecorate %outdata DescriptorSet 0\n"
1747                         "OpDecorate %outdata Binding 2\n"
1748                         "OpDecorate %i64arr ArrayStride 8\n"
1749                         "OpMemberDecorate %buf 0 Offset 0\n"
1750
1751                         + string(getComputeAsmCommonTypes())
1752                         + string(getComputeAsmCommonInt64Types()) +
1753
1754                         "%buf        = OpTypeStruct %i64arr\n"
1755                         "%bufptr     = OpTypePointer Uniform %buf\n"
1756                         "%indata1    = OpVariable %bufptr Uniform\n"
1757                         "%indata2    = OpVariable %bufptr Uniform\n"
1758                         "%outdata    = OpVariable %bufptr Uniform\n"
1759
1760                         "%id        = OpVariable %uvec3ptr Input\n"
1761                         "%zero      = OpConstant %i64 0\n"
1762
1763                         "%main      = OpFunction %void None %voidf\n"
1764                         "%label     = OpLabel\n"
1765                         "%idval     = OpLoad %uvec3 %id\n"
1766                         "%x         = OpCompositeExtract %u32 %idval 0\n"
1767                         "%inloc1    = OpAccessChain %i64ptr %indata1 %zero %x\n"
1768                         "%inval1    = OpLoad %i64 %inloc1\n"
1769                         "%inloc2    = OpAccessChain %i64ptr %indata2 %zero %x\n"
1770                         "%inval2    = OpLoad %i64 %inloc2\n"
1771                         "%rem       = OpSRem %i64 %inval1 %inval2\n"
1772                         "%outloc    = OpAccessChain %i64ptr %outdata %zero %x\n"
1773                         "             OpStore %outloc %rem\n"
1774                         "             OpReturn\n"
1775                         "             OpFunctionEnd\n";
1776
1777                 spec.inputs.push_back   (BufferSp(new Int64Buffer(inputInts1)));
1778                 spec.inputs.push_back   (BufferSp(new Int64Buffer(inputInts2)));
1779                 spec.outputs.push_back  (BufferSp(new Int64Buffer(outputInts)));
1780                 spec.numWorkGroups              = IVec3(numElements, 1, 1);
1781                 spec.failResult                 = params.failResult;
1782                 spec.failMessage                = params.failMessage;
1783
1784                 spec.requestedVulkanFeatures.coreFeatures.shaderInt64 = VK_TRUE;
1785
1786                 group->addChild(new SpvAsmComputeShaderCase(testCtx, params.name, "", spec));
1787         }
1788
1789         return group.release();
1790 }
1791
1792 tcu::TestCaseGroup* createOpSModComputeGroup (tcu::TestContext& testCtx, qpTestResult negFailResult)
1793 {
1794         de::MovePtr<tcu::TestCaseGroup> group                   (new tcu::TestCaseGroup(testCtx, "opsmod", "Test the OpSMod instruction"));
1795         de::Random                                              rnd                             (deStringHash(group->getName()));
1796         const int                                               numElements             = 200;
1797
1798         const struct CaseParams
1799         {
1800                 const char*             name;
1801                 const char*             failMessage;            // customized status message
1802                 qpTestResult    failResult;                     // override status on failure
1803                 int                             op1Min, op1Max;         // operand ranges
1804                 int                             op2Min, op2Max;
1805         } cases[] =
1806         {
1807                 { "positive",   "Output doesn't match with expected",                           QP_TEST_RESULT_FAIL,    0,              65536,  0,              100 },
1808                 { "all",                "Inconsistent results, but within specification",       negFailResult,                  -65536, 65536,  -100,   100 },  // see below
1809         };
1810         // If either operand is negative the result is undefined. Some implementations may still return correct values.
1811
1812         for (int caseNdx = 0; caseNdx < DE_LENGTH_OF_ARRAY(cases); ++caseNdx)
1813         {
1814                 const CaseParams&       params          = cases[caseNdx];
1815
1816                 ComputeShaderSpec       spec;
1817                 vector<deInt32>         inputInts1      (numElements, 0);
1818                 vector<deInt32>         inputInts2      (numElements, 0);
1819                 vector<deInt32>         outputInts      (numElements, 0);
1820
1821                 fillRandomScalars(rnd, params.op1Min, params.op1Max, &inputInts1[0], numElements);
1822                 fillRandomScalars(rnd, params.op2Min, params.op2Max, &inputInts2[0], numElements, filterNotZero);
1823
1824                 for (int ndx = 0; ndx < numElements; ++ndx)
1825                 {
1826                         deInt32 rem = inputInts1[ndx] % inputInts2[ndx];
1827                         if (rem == 0)
1828                         {
1829                                 outputInts[ndx] = 0;
1830                         }
1831                         else if ((inputInts1[ndx] >= 0) == (inputInts2[ndx] >= 0))
1832                         {
1833                                 // They have the same sign
1834                                 outputInts[ndx] = rem;
1835                         }
1836                         else
1837                         {
1838                                 // They have opposite sign.  The remainder operation takes the
1839                                 // sign inputInts1[ndx] but OpSMod is supposed to take ths sign
1840                                 // of inputInts2[ndx].  Adding inputInts2[ndx] will ensure that
1841                                 // the result has the correct sign and that it is still
1842                                 // congruent to inputInts1[ndx] modulo inputInts2[ndx]
1843                                 //
1844                                 // See also http://mathforum.org/library/drmath/view/52343.html
1845                                 outputInts[ndx] = rem + inputInts2[ndx];
1846                         }
1847                 }
1848
1849                 spec.assembly =
1850                         string(getComputeAsmShaderPreamble()) +
1851
1852                         "OpName %main           \"main\"\n"
1853                         "OpName %id             \"gl_GlobalInvocationID\"\n"
1854
1855                         "OpDecorate %id BuiltIn GlobalInvocationId\n"
1856
1857                         "OpDecorate %buf BufferBlock\n"
1858                         "OpDecorate %indata1 DescriptorSet 0\n"
1859                         "OpDecorate %indata1 Binding 0\n"
1860                         "OpDecorate %indata2 DescriptorSet 0\n"
1861                         "OpDecorate %indata2 Binding 1\n"
1862                         "OpDecorate %outdata DescriptorSet 0\n"
1863                         "OpDecorate %outdata Binding 2\n"
1864                         "OpDecorate %i32arr ArrayStride 4\n"
1865                         "OpMemberDecorate %buf 0 Offset 0\n"
1866
1867                         + string(getComputeAsmCommonTypes()) +
1868
1869                         "%buf        = OpTypeStruct %i32arr\n"
1870                         "%bufptr     = OpTypePointer Uniform %buf\n"
1871                         "%indata1    = OpVariable %bufptr Uniform\n"
1872                         "%indata2    = OpVariable %bufptr Uniform\n"
1873                         "%outdata    = OpVariable %bufptr Uniform\n"
1874
1875                         "%id        = OpVariable %uvec3ptr Input\n"
1876                         "%zero      = OpConstant %i32 0\n"
1877
1878                         "%main      = OpFunction %void None %voidf\n"
1879                         "%label     = OpLabel\n"
1880                         "%idval     = OpLoad %uvec3 %id\n"
1881                         "%x         = OpCompositeExtract %u32 %idval 0\n"
1882                         "%inloc1    = OpAccessChain %i32ptr %indata1 %zero %x\n"
1883                         "%inval1    = OpLoad %i32 %inloc1\n"
1884                         "%inloc2    = OpAccessChain %i32ptr %indata2 %zero %x\n"
1885                         "%inval2    = OpLoad %i32 %inloc2\n"
1886                         "%rem       = OpSMod %i32 %inval1 %inval2\n"
1887                         "%outloc    = OpAccessChain %i32ptr %outdata %zero %x\n"
1888                         "             OpStore %outloc %rem\n"
1889                         "             OpReturn\n"
1890                         "             OpFunctionEnd\n";
1891
1892                 spec.inputs.push_back   (BufferSp(new Int32Buffer(inputInts1)));
1893                 spec.inputs.push_back   (BufferSp(new Int32Buffer(inputInts2)));
1894                 spec.outputs.push_back  (BufferSp(new Int32Buffer(outputInts)));
1895                 spec.numWorkGroups              = IVec3(numElements, 1, 1);
1896                 spec.failResult                 = params.failResult;
1897                 spec.failMessage                = params.failMessage;
1898
1899                 group->addChild(new SpvAsmComputeShaderCase(testCtx, params.name, "", spec));
1900         }
1901
1902         return group.release();
1903 }
1904
1905 tcu::TestCaseGroup* createOpSModComputeGroup64 (tcu::TestContext& testCtx, qpTestResult negFailResult)
1906 {
1907         de::MovePtr<tcu::TestCaseGroup> group                   (new tcu::TestCaseGroup(testCtx, "opsmod64", "Test the OpSMod instruction"));
1908         de::Random                                              rnd                             (deStringHash(group->getName()));
1909         const int                                               numElements             = 200;
1910
1911         const struct CaseParams
1912         {
1913                 const char*             name;
1914                 const char*             failMessage;            // customized status message
1915                 qpTestResult    failResult;                     // override status on failure
1916                 bool                    positive;
1917         } cases[] =
1918         {
1919                 { "positive",   "Output doesn't match with expected",                           QP_TEST_RESULT_FAIL,    true },
1920                 { "all",                "Inconsistent results, but within specification",       negFailResult,                  false },        // see below
1921         };
1922         // If either operand is negative the result is undefined. Some implementations may still return correct values.
1923
1924         for (int caseNdx = 0; caseNdx < DE_LENGTH_OF_ARRAY(cases); ++caseNdx)
1925         {
1926                 const CaseParams&       params          = cases[caseNdx];
1927
1928                 ComputeShaderSpec       spec;
1929                 vector<deInt64>         inputInts1      (numElements, 0);
1930                 vector<deInt64>         inputInts2      (numElements, 0);
1931                 vector<deInt64>         outputInts      (numElements, 0);
1932
1933
1934                 if (params.positive)
1935                 {
1936                         fillRandomInt64sLogDistributed(rnd, inputInts1, numElements, filterNonNegative);
1937                         fillRandomInt64sLogDistributed(rnd, inputInts2, numElements, filterPositive);
1938                 }
1939                 else
1940                 {
1941                         fillRandomInt64sLogDistributed(rnd, inputInts1, numElements);
1942                         fillRandomInt64sLogDistributed(rnd, inputInts2, numElements, filterNotZero);
1943                 }
1944
1945                 for (int ndx = 0; ndx < numElements; ++ndx)
1946                 {
1947                         deInt64 rem = inputInts1[ndx] % inputInts2[ndx];
1948                         if (rem == 0)
1949                         {
1950                                 outputInts[ndx] = 0;
1951                         }
1952                         else if ((inputInts1[ndx] >= 0) == (inputInts2[ndx] >= 0))
1953                         {
1954                                 // They have the same sign
1955                                 outputInts[ndx] = rem;
1956                         }
1957                         else
1958                         {
1959                                 // They have opposite sign.  The remainder operation takes the
1960                                 // sign inputInts1[ndx] but OpSMod is supposed to take ths sign
1961                                 // of inputInts2[ndx].  Adding inputInts2[ndx] will ensure that
1962                                 // the result has the correct sign and that it is still
1963                                 // congruent to inputInts1[ndx] modulo inputInts2[ndx]
1964                                 //
1965                                 // See also http://mathforum.org/library/drmath/view/52343.html
1966                                 outputInts[ndx] = rem + inputInts2[ndx];
1967                         }
1968                 }
1969
1970                 spec.assembly =
1971                         "OpCapability Int64\n"
1972
1973                         + string(getComputeAsmShaderPreamble()) +
1974
1975                         "OpName %main           \"main\"\n"
1976                         "OpName %id             \"gl_GlobalInvocationID\"\n"
1977
1978                         "OpDecorate %id BuiltIn GlobalInvocationId\n"
1979
1980                         "OpDecorate %buf BufferBlock\n"
1981                         "OpDecorate %indata1 DescriptorSet 0\n"
1982                         "OpDecorate %indata1 Binding 0\n"
1983                         "OpDecorate %indata2 DescriptorSet 0\n"
1984                         "OpDecorate %indata2 Binding 1\n"
1985                         "OpDecorate %outdata DescriptorSet 0\n"
1986                         "OpDecorate %outdata Binding 2\n"
1987                         "OpDecorate %i64arr ArrayStride 8\n"
1988                         "OpMemberDecorate %buf 0 Offset 0\n"
1989
1990                         + string(getComputeAsmCommonTypes())
1991                         + string(getComputeAsmCommonInt64Types()) +
1992
1993                         "%buf        = OpTypeStruct %i64arr\n"
1994                         "%bufptr     = OpTypePointer Uniform %buf\n"
1995                         "%indata1    = OpVariable %bufptr Uniform\n"
1996                         "%indata2    = OpVariable %bufptr Uniform\n"
1997                         "%outdata    = OpVariable %bufptr Uniform\n"
1998
1999                         "%id        = OpVariable %uvec3ptr Input\n"
2000                         "%zero      = OpConstant %i64 0\n"
2001
2002                         "%main      = OpFunction %void None %voidf\n"
2003                         "%label     = OpLabel\n"
2004                         "%idval     = OpLoad %uvec3 %id\n"
2005                         "%x         = OpCompositeExtract %u32 %idval 0\n"
2006                         "%inloc1    = OpAccessChain %i64ptr %indata1 %zero %x\n"
2007                         "%inval1    = OpLoad %i64 %inloc1\n"
2008                         "%inloc2    = OpAccessChain %i64ptr %indata2 %zero %x\n"
2009                         "%inval2    = OpLoad %i64 %inloc2\n"
2010                         "%rem       = OpSMod %i64 %inval1 %inval2\n"
2011                         "%outloc    = OpAccessChain %i64ptr %outdata %zero %x\n"
2012                         "             OpStore %outloc %rem\n"
2013                         "             OpReturn\n"
2014                         "             OpFunctionEnd\n";
2015
2016                 spec.inputs.push_back   (BufferSp(new Int64Buffer(inputInts1)));
2017                 spec.inputs.push_back   (BufferSp(new Int64Buffer(inputInts2)));
2018                 spec.outputs.push_back  (BufferSp(new Int64Buffer(outputInts)));
2019                 spec.numWorkGroups              = IVec3(numElements, 1, 1);
2020                 spec.failResult                 = params.failResult;
2021                 spec.failMessage                = params.failMessage;
2022
2023                 spec.requestedVulkanFeatures.coreFeatures.shaderInt64 = VK_TRUE;
2024
2025                 group->addChild(new SpvAsmComputeShaderCase(testCtx, params.name, "", spec));
2026         }
2027
2028         return group.release();
2029 }
2030
2031 // Copy contents in the input buffer to the output buffer.
2032 tcu::TestCaseGroup* createOpCopyMemoryGroup (tcu::TestContext& testCtx)
2033 {
2034         de::MovePtr<tcu::TestCaseGroup> group                   (new tcu::TestCaseGroup(testCtx, "opcopymemory", "Test the OpCopyMemory instruction"));
2035         de::Random                                              rnd                             (deStringHash(group->getName()));
2036         const int                                               numElements             = 100;
2037
2038         // 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.
2039         ComputeShaderSpec                               spec1;
2040         vector<Vec4>                                    inputFloats1    (numElements);
2041         vector<Vec4>                                    outputFloats1   (numElements);
2042
2043         fillRandomScalars(rnd, -200.f, 200.f, &inputFloats1[0], numElements * 4);
2044
2045         // CPU might not use the same rounding mode as the GPU. Use whole numbers to avoid rounding differences.
2046         floorAll(inputFloats1);
2047
2048         for (size_t ndx = 0; ndx < numElements; ++ndx)
2049                 outputFloats1[ndx] = inputFloats1[ndx] + Vec4(0.f, 0.5f, 1.5f, 2.5f);
2050
2051         spec1.assembly =
2052                 string(getComputeAsmShaderPreamble()) +
2053
2054                 "OpName %main           \"main\"\n"
2055                 "OpName %id             \"gl_GlobalInvocationID\"\n"
2056
2057                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
2058                 "OpDecorate %vec4arr ArrayStride 16\n"
2059
2060                 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) +
2061
2062                 "%vec4       = OpTypeVector %f32 4\n"
2063                 "%vec4ptr_u  = OpTypePointer Uniform %vec4\n"
2064                 "%vec4ptr_f  = OpTypePointer Function %vec4\n"
2065                 "%vec4arr    = OpTypeRuntimeArray %vec4\n"
2066                 "%buf        = OpTypeStruct %vec4arr\n"
2067                 "%bufptr     = OpTypePointer Uniform %buf\n"
2068                 "%indata     = OpVariable %bufptr Uniform\n"
2069                 "%outdata    = OpVariable %bufptr Uniform\n"
2070
2071                 "%id         = OpVariable %uvec3ptr Input\n"
2072                 "%zero       = OpConstant %i32 0\n"
2073                 "%c_f_0      = OpConstant %f32 0.\n"
2074                 "%c_f_0_5    = OpConstant %f32 0.5\n"
2075                 "%c_f_1_5    = OpConstant %f32 1.5\n"
2076                 "%c_f_2_5    = OpConstant %f32 2.5\n"
2077                 "%c_vec4     = OpConstantComposite %vec4 %c_f_0 %c_f_0_5 %c_f_1_5 %c_f_2_5\n"
2078
2079                 "%main       = OpFunction %void None %voidf\n"
2080                 "%label      = OpLabel\n"
2081                 "%v_vec4     = OpVariable %vec4ptr_f Function\n"
2082                 "%idval      = OpLoad %uvec3 %id\n"
2083                 "%x          = OpCompositeExtract %u32 %idval 0\n"
2084                 "%inloc      = OpAccessChain %vec4ptr_u %indata %zero %x\n"
2085                 "%outloc     = OpAccessChain %vec4ptr_u %outdata %zero %x\n"
2086                 "              OpCopyMemory %v_vec4 %inloc\n"
2087                 "%v_vec4_val = OpLoad %vec4 %v_vec4\n"
2088                 "%add        = OpFAdd %vec4 %v_vec4_val %c_vec4\n"
2089                 "              OpStore %outloc %add\n"
2090                 "              OpReturn\n"
2091                 "              OpFunctionEnd\n";
2092
2093         spec1.inputs.push_back(BufferSp(new Vec4Buffer(inputFloats1)));
2094         spec1.outputs.push_back(BufferSp(new Vec4Buffer(outputFloats1)));
2095         spec1.numWorkGroups = IVec3(numElements, 1, 1);
2096
2097         group->addChild(new SpvAsmComputeShaderCase(testCtx, "vector", "OpCopyMemory elements of vector type", spec1));
2098
2099         // The following case copies a float[100] variable from the input buffer to the output buffer.
2100         ComputeShaderSpec                               spec2;
2101         vector<float>                                   inputFloats2    (numElements);
2102         vector<float>                                   outputFloats2   (numElements);
2103
2104         fillRandomScalars(rnd, -200.f, 200.f, &inputFloats2[0], numElements);
2105
2106         for (size_t ndx = 0; ndx < numElements; ++ndx)
2107                 outputFloats2[ndx] = inputFloats2[ndx];
2108
2109         spec2.assembly =
2110                 string(getComputeAsmShaderPreamble()) +
2111
2112                 "OpName %main           \"main\"\n"
2113                 "OpName %id             \"gl_GlobalInvocationID\"\n"
2114
2115                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
2116                 "OpDecorate %f32arr100 ArrayStride 4\n"
2117
2118                 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) +
2119
2120                 "%hundred        = OpConstant %u32 100\n"
2121                 "%f32arr100      = OpTypeArray %f32 %hundred\n"
2122                 "%f32arr100ptr_f = OpTypePointer Function %f32arr100\n"
2123                 "%f32arr100ptr_u = OpTypePointer Uniform %f32arr100\n"
2124                 "%buf            = OpTypeStruct %f32arr100\n"
2125                 "%bufptr         = OpTypePointer Uniform %buf\n"
2126                 "%indata         = OpVariable %bufptr Uniform\n"
2127                 "%outdata        = OpVariable %bufptr Uniform\n"
2128
2129                 "%id             = OpVariable %uvec3ptr Input\n"
2130                 "%zero           = OpConstant %i32 0\n"
2131
2132                 "%main           = OpFunction %void None %voidf\n"
2133                 "%label          = OpLabel\n"
2134                 "%var            = OpVariable %f32arr100ptr_f Function\n"
2135                 "%inarr          = OpAccessChain %f32arr100ptr_u %indata %zero\n"
2136                 "%outarr         = OpAccessChain %f32arr100ptr_u %outdata %zero\n"
2137                 "                  OpCopyMemory %var %inarr\n"
2138                 "                  OpCopyMemory %outarr %var\n"
2139                 "                  OpReturn\n"
2140                 "                  OpFunctionEnd\n";
2141
2142         spec2.inputs.push_back(BufferSp(new Float32Buffer(inputFloats2)));
2143         spec2.outputs.push_back(BufferSp(new Float32Buffer(outputFloats2)));
2144         spec2.numWorkGroups = IVec3(1, 1, 1);
2145
2146         group->addChild(new SpvAsmComputeShaderCase(testCtx, "array", "OpCopyMemory elements of array type", spec2));
2147
2148         // The following case copies a struct{vec4, vec4, vec4, vec4} variable from the input buffer to the output buffer.
2149         ComputeShaderSpec                               spec3;
2150         vector<float>                                   inputFloats3    (16);
2151         vector<float>                                   outputFloats3   (16);
2152
2153         fillRandomScalars(rnd, -200.f, 200.f, &inputFloats3[0], 16);
2154
2155         for (size_t ndx = 0; ndx < 16; ++ndx)
2156                 outputFloats3[ndx] = inputFloats3[ndx];
2157
2158         spec3.assembly =
2159                 string(getComputeAsmShaderPreamble()) +
2160
2161                 "OpName %main           \"main\"\n"
2162                 "OpName %id             \"gl_GlobalInvocationID\"\n"
2163
2164                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
2165                 //"OpMemberDecorate %buf 0 Offset 0\n"  - exists in getComputeAsmInputOutputBufferTraits
2166                 "OpMemberDecorate %buf 1 Offset 16\n"
2167                 "OpMemberDecorate %buf 2 Offset 32\n"
2168                 "OpMemberDecorate %buf 3 Offset 48\n"
2169
2170                 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) +
2171
2172                 "%vec4      = OpTypeVector %f32 4\n"
2173                 "%buf       = OpTypeStruct %vec4 %vec4 %vec4 %vec4\n"
2174                 "%bufptr    = OpTypePointer Uniform %buf\n"
2175                 "%indata    = OpVariable %bufptr Uniform\n"
2176                 "%outdata   = OpVariable %bufptr Uniform\n"
2177                 "%vec4stptr = OpTypePointer Function %buf\n"
2178
2179                 "%id        = OpVariable %uvec3ptr Input\n"
2180                 "%zero      = OpConstant %i32 0\n"
2181
2182                 "%main      = OpFunction %void None %voidf\n"
2183                 "%label     = OpLabel\n"
2184                 "%var       = OpVariable %vec4stptr Function\n"
2185                 "             OpCopyMemory %var %indata\n"
2186                 "             OpCopyMemory %outdata %var\n"
2187                 "             OpReturn\n"
2188                 "             OpFunctionEnd\n";
2189
2190         spec3.inputs.push_back(BufferSp(new Float32Buffer(inputFloats3)));
2191         spec3.outputs.push_back(BufferSp(new Float32Buffer(outputFloats3)));
2192         spec3.numWorkGroups = IVec3(1, 1, 1);
2193
2194         group->addChild(new SpvAsmComputeShaderCase(testCtx, "struct", "OpCopyMemory elements of struct type", spec3));
2195
2196         // The following case negates multiple float variables from the input buffer and stores the results to the output buffer.
2197         ComputeShaderSpec                               spec4;
2198         vector<float>                                   inputFloats4    (numElements);
2199         vector<float>                                   outputFloats4   (numElements);
2200
2201         fillRandomScalars(rnd, -200.f, 200.f, &inputFloats4[0], numElements);
2202
2203         for (size_t ndx = 0; ndx < numElements; ++ndx)
2204                 outputFloats4[ndx] = -inputFloats4[ndx];
2205
2206         spec4.assembly =
2207                 string(getComputeAsmShaderPreamble()) +
2208
2209                 "OpName %main           \"main\"\n"
2210                 "OpName %id             \"gl_GlobalInvocationID\"\n"
2211
2212                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
2213
2214                 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) + string(getComputeAsmInputOutputBuffer()) +
2215
2216                 "%f32ptr_f  = OpTypePointer Function %f32\n"
2217                 "%id        = OpVariable %uvec3ptr Input\n"
2218                 "%zero      = OpConstant %i32 0\n"
2219
2220                 "%main      = OpFunction %void None %voidf\n"
2221                 "%label     = OpLabel\n"
2222                 "%var       = OpVariable %f32ptr_f Function\n"
2223                 "%idval     = OpLoad %uvec3 %id\n"
2224                 "%x         = OpCompositeExtract %u32 %idval 0\n"
2225                 "%inloc     = OpAccessChain %f32ptr %indata %zero %x\n"
2226                 "%outloc    = OpAccessChain %f32ptr %outdata %zero %x\n"
2227                 "             OpCopyMemory %var %inloc\n"
2228                 "%val       = OpLoad %f32 %var\n"
2229                 "%neg       = OpFNegate %f32 %val\n"
2230                 "             OpStore %outloc %neg\n"
2231                 "             OpReturn\n"
2232                 "             OpFunctionEnd\n";
2233
2234         spec4.inputs.push_back(BufferSp(new Float32Buffer(inputFloats4)));
2235         spec4.outputs.push_back(BufferSp(new Float32Buffer(outputFloats4)));
2236         spec4.numWorkGroups = IVec3(numElements, 1, 1);
2237
2238         group->addChild(new SpvAsmComputeShaderCase(testCtx, "float", "OpCopyMemory elements of float type", spec4));
2239
2240         return group.release();
2241 }
2242
2243 tcu::TestCaseGroup* createOpCopyObjectGroup (tcu::TestContext& testCtx)
2244 {
2245         de::MovePtr<tcu::TestCaseGroup> group                   (new tcu::TestCaseGroup(testCtx, "opcopyobject", "Test the OpCopyObject instruction"));
2246         ComputeShaderSpec                               spec;
2247         de::Random                                              rnd                             (deStringHash(group->getName()));
2248         const int                                               numElements             = 100;
2249         vector<float>                                   inputFloats             (numElements, 0);
2250         vector<float>                                   outputFloats    (numElements, 0);
2251
2252         fillRandomScalars(rnd, -200.f, 200.f, &inputFloats[0], numElements);
2253
2254         // CPU might not use the same rounding mode as the GPU. Use whole numbers to avoid rounding differences.
2255         floorAll(inputFloats);
2256
2257         for (size_t ndx = 0; ndx < numElements; ++ndx)
2258                 outputFloats[ndx] = inputFloats[ndx] + 7.5f;
2259
2260         spec.assembly =
2261                 string(getComputeAsmShaderPreamble()) +
2262
2263                 "OpName %main           \"main\"\n"
2264                 "OpName %id             \"gl_GlobalInvocationID\"\n"
2265
2266                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
2267
2268                 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) +
2269
2270                 "%fmat     = OpTypeMatrix %fvec3 3\n"
2271                 "%three    = OpConstant %u32 3\n"
2272                 "%farr     = OpTypeArray %f32 %three\n"
2273                 "%fst      = OpTypeStruct %f32 %f32\n"
2274
2275                 + string(getComputeAsmInputOutputBuffer()) +
2276
2277                 "%id            = OpVariable %uvec3ptr Input\n"
2278                 "%zero          = OpConstant %i32 0\n"
2279                 "%c_f           = OpConstant %f32 1.5\n"
2280                 "%c_fvec3       = OpConstantComposite %fvec3 %c_f %c_f %c_f\n"
2281                 "%c_fmat        = OpConstantComposite %fmat %c_fvec3 %c_fvec3 %c_fvec3\n"
2282                 "%c_farr        = OpConstantComposite %farr %c_f %c_f %c_f\n"
2283                 "%c_fst         = OpConstantComposite %fst %c_f %c_f\n"
2284
2285                 "%main          = OpFunction %void None %voidf\n"
2286                 "%label         = OpLabel\n"
2287                 "%c_f_copy      = OpCopyObject %f32   %c_f\n"
2288                 "%c_fvec3_copy  = OpCopyObject %fvec3 %c_fvec3\n"
2289                 "%c_fmat_copy   = OpCopyObject %fmat  %c_fmat\n"
2290                 "%c_farr_copy   = OpCopyObject %farr  %c_farr\n"
2291                 "%c_fst_copy    = OpCopyObject %fst   %c_fst\n"
2292                 "%fvec3_elem    = OpCompositeExtract %f32 %c_fvec3_copy 0\n"
2293                 "%fmat_elem     = OpCompositeExtract %f32 %c_fmat_copy 1 2\n"
2294                 "%farr_elem     = OpCompositeExtract %f32 %c_farr_copy 2\n"
2295                 "%fst_elem      = OpCompositeExtract %f32 %c_fst_copy 1\n"
2296                 // Add up. 1.5 * 5 = 7.5.
2297                 "%add1          = OpFAdd %f32 %c_f_copy %fvec3_elem\n"
2298                 "%add2          = OpFAdd %f32 %add1     %fmat_elem\n"
2299                 "%add3          = OpFAdd %f32 %add2     %farr_elem\n"
2300                 "%add4          = OpFAdd %f32 %add3     %fst_elem\n"
2301
2302                 "%idval         = OpLoad %uvec3 %id\n"
2303                 "%x             = OpCompositeExtract %u32 %idval 0\n"
2304                 "%inloc         = OpAccessChain %f32ptr %indata %zero %x\n"
2305                 "%outloc        = OpAccessChain %f32ptr %outdata %zero %x\n"
2306                 "%inval         = OpLoad %f32 %inloc\n"
2307                 "%add           = OpFAdd %f32 %add4 %inval\n"
2308                 "                 OpStore %outloc %add\n"
2309                 "                 OpReturn\n"
2310                 "                 OpFunctionEnd\n";
2311         spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
2312         spec.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
2313         spec.numWorkGroups = IVec3(numElements, 1, 1);
2314
2315         group->addChild(new SpvAsmComputeShaderCase(testCtx, "spotcheck", "OpCopyObject on different types", spec));
2316
2317         return group.release();
2318 }
2319 // Assembly code used for testing OpUnreachable is based on GLSL source code:
2320 //
2321 // #version 430
2322 //
2323 // layout(std140, set = 0, binding = 0) readonly buffer Input {
2324 //   float elements[];
2325 // } input_data;
2326 // layout(std140, set = 0, binding = 1) writeonly buffer Output {
2327 //   float elements[];
2328 // } output_data;
2329 //
2330 // void not_called_func() {
2331 //   // place OpUnreachable here
2332 // }
2333 //
2334 // uint modulo4(uint val) {
2335 //   switch (val % uint(4)) {
2336 //     case 0:  return 3;
2337 //     case 1:  return 2;
2338 //     case 2:  return 1;
2339 //     case 3:  return 0;
2340 //     default: return 100; // place OpUnreachable here
2341 //   }
2342 // }
2343 //
2344 // uint const5() {
2345 //   return 5;
2346 //   // place OpUnreachable here
2347 // }
2348 //
2349 // void main() {
2350 //   uint x = gl_GlobalInvocationID.x;
2351 //   if (const5() > modulo4(1000)) {
2352 //     output_data.elements[x] = -input_data.elements[x];
2353 //   } else {
2354 //     // place OpUnreachable here
2355 //     output_data.elements[x] = input_data.elements[x];
2356 //   }
2357 // }
2358
2359 tcu::TestCaseGroup* createOpUnreachableGroup (tcu::TestContext& testCtx)
2360 {
2361         de::MovePtr<tcu::TestCaseGroup> group                   (new tcu::TestCaseGroup(testCtx, "opunreachable", "Test the OpUnreachable instruction"));
2362         ComputeShaderSpec                               spec;
2363         de::Random                                              rnd                             (deStringHash(group->getName()));
2364         const int                                               numElements             = 100;
2365         vector<float>                                   positiveFloats  (numElements, 0);
2366         vector<float>                                   negativeFloats  (numElements, 0);
2367
2368         fillRandomScalars(rnd, 1.f, 100.f, &positiveFloats[0], numElements);
2369
2370         for (size_t ndx = 0; ndx < numElements; ++ndx)
2371                 negativeFloats[ndx] = -positiveFloats[ndx];
2372
2373         spec.assembly =
2374                 string(getComputeAsmShaderPreamble()) +
2375
2376                 "OpSource GLSL 430\n"
2377                 "OpName %main            \"main\"\n"
2378                 "OpName %func_not_called_func \"not_called_func(\"\n"
2379                 "OpName %func_modulo4         \"modulo4(u1;\"\n"
2380                 "OpName %func_const5          \"const5(\"\n"
2381                 "OpName %id                   \"gl_GlobalInvocationID\"\n"
2382
2383                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
2384
2385                 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) +
2386
2387                 "%u32ptr    = OpTypePointer Function %u32\n"
2388                 "%uintfuint = OpTypeFunction %u32 %u32ptr\n"
2389                 "%unitf     = OpTypeFunction %u32\n"
2390
2391                 "%id        = OpVariable %uvec3ptr Input\n"
2392                 "%zero      = OpConstant %u32 0\n"
2393                 "%one       = OpConstant %u32 1\n"
2394                 "%two       = OpConstant %u32 2\n"
2395                 "%three     = OpConstant %u32 3\n"
2396                 "%four      = OpConstant %u32 4\n"
2397                 "%five      = OpConstant %u32 5\n"
2398                 "%hundred   = OpConstant %u32 100\n"
2399                 "%thousand  = OpConstant %u32 1000\n"
2400
2401                 + string(getComputeAsmInputOutputBuffer()) +
2402
2403                 // Main()
2404                 "%main   = OpFunction %void None %voidf\n"
2405                 "%main_entry  = OpLabel\n"
2406                 "%v_thousand  = OpVariable %u32ptr Function %thousand\n"
2407                 "%idval       = OpLoad %uvec3 %id\n"
2408                 "%x           = OpCompositeExtract %u32 %idval 0\n"
2409                 "%inloc       = OpAccessChain %f32ptr %indata %zero %x\n"
2410                 "%inval       = OpLoad %f32 %inloc\n"
2411                 "%outloc      = OpAccessChain %f32ptr %outdata %zero %x\n"
2412                 "%ret_const5  = OpFunctionCall %u32 %func_const5\n"
2413                 "%ret_modulo4 = OpFunctionCall %u32 %func_modulo4 %v_thousand\n"
2414                 "%cmp_gt      = OpUGreaterThan %bool %ret_const5 %ret_modulo4\n"
2415                 "               OpSelectionMerge %if_end None\n"
2416                 "               OpBranchConditional %cmp_gt %if_true %if_false\n"
2417                 "%if_true     = OpLabel\n"
2418                 "%negate      = OpFNegate %f32 %inval\n"
2419                 "               OpStore %outloc %negate\n"
2420                 "               OpBranch %if_end\n"
2421                 "%if_false    = OpLabel\n"
2422                 "               OpUnreachable\n" // Unreachable else branch for if statement
2423                 "%if_end      = OpLabel\n"
2424                 "               OpReturn\n"
2425                 "               OpFunctionEnd\n"
2426
2427                 // not_called_function()
2428                 "%func_not_called_func  = OpFunction %void None %voidf\n"
2429                 "%not_called_func_entry = OpLabel\n"
2430                 "                         OpUnreachable\n" // Unreachable entry block in not called static function
2431                 "                         OpFunctionEnd\n"
2432
2433                 // modulo4()
2434                 "%func_modulo4  = OpFunction %u32 None %uintfuint\n"
2435                 "%valptr        = OpFunctionParameter %u32ptr\n"
2436                 "%modulo4_entry = OpLabel\n"
2437                 "%val           = OpLoad %u32 %valptr\n"
2438                 "%modulo        = OpUMod %u32 %val %four\n"
2439                 "                 OpSelectionMerge %switch_merge None\n"
2440                 "                 OpSwitch %modulo %default 0 %case0 1 %case1 2 %case2 3 %case3\n"
2441                 "%case0         = OpLabel\n"
2442                 "                 OpReturnValue %three\n"
2443                 "%case1         = OpLabel\n"
2444                 "                 OpReturnValue %two\n"
2445                 "%case2         = OpLabel\n"
2446                 "                 OpReturnValue %one\n"
2447                 "%case3         = OpLabel\n"
2448                 "                 OpReturnValue %zero\n"
2449                 "%default       = OpLabel\n"
2450                 "                 OpUnreachable\n" // Unreachable default case for switch statement
2451                 "%switch_merge  = OpLabel\n"
2452                 "                 OpUnreachable\n" // Unreachable merge block for switch statement
2453                 "                 OpFunctionEnd\n"
2454
2455                 // const5()
2456                 "%func_const5  = OpFunction %u32 None %unitf\n"
2457                 "%const5_entry = OpLabel\n"
2458                 "                OpReturnValue %five\n"
2459                 "%unreachable  = OpLabel\n"
2460                 "                OpUnreachable\n" // Unreachable block in function
2461                 "                OpFunctionEnd\n";
2462         spec.inputs.push_back(BufferSp(new Float32Buffer(positiveFloats)));
2463         spec.outputs.push_back(BufferSp(new Float32Buffer(negativeFloats)));
2464         spec.numWorkGroups = IVec3(numElements, 1, 1);
2465
2466         group->addChild(new SpvAsmComputeShaderCase(testCtx, "all", "OpUnreachable appearing at different places", spec));
2467
2468         return group.release();
2469 }
2470
2471 // Assembly code used for testing decoration group is based on GLSL source code:
2472 //
2473 // #version 430
2474 //
2475 // layout(std140, set = 0, binding = 0) readonly buffer Input0 {
2476 //   float elements[];
2477 // } input_data0;
2478 // layout(std140, set = 0, binding = 1) readonly buffer Input1 {
2479 //   float elements[];
2480 // } input_data1;
2481 // layout(std140, set = 0, binding = 2) readonly buffer Input2 {
2482 //   float elements[];
2483 // } input_data2;
2484 // layout(std140, set = 0, binding = 3) readonly buffer Input3 {
2485 //   float elements[];
2486 // } input_data3;
2487 // layout(std140, set = 0, binding = 4) readonly buffer Input4 {
2488 //   float elements[];
2489 // } input_data4;
2490 // layout(std140, set = 0, binding = 5) writeonly buffer Output {
2491 //   float elements[];
2492 // } output_data;
2493 //
2494 // void main() {
2495 //   uint x = gl_GlobalInvocationID.x;
2496 //   output_data.elements[x] = input_data0.elements[x] + input_data1.elements[x] + input_data2.elements[x] + input_data3.elements[x] + input_data4.elements[x];
2497 // }
2498 tcu::TestCaseGroup* createDecorationGroupGroup (tcu::TestContext& testCtx)
2499 {
2500         de::MovePtr<tcu::TestCaseGroup> group                   (new tcu::TestCaseGroup(testCtx, "decoration_group", "Test the OpDecorationGroup & OpGroupDecorate instruction"));
2501         ComputeShaderSpec                               spec;
2502         de::Random                                              rnd                             (deStringHash(group->getName()));
2503         const int                                               numElements             = 100;
2504         vector<float>                                   inputFloats0    (numElements, 0);
2505         vector<float>                                   inputFloats1    (numElements, 0);
2506         vector<float>                                   inputFloats2    (numElements, 0);
2507         vector<float>                                   inputFloats3    (numElements, 0);
2508         vector<float>                                   inputFloats4    (numElements, 0);
2509         vector<float>                                   outputFloats    (numElements, 0);
2510
2511         fillRandomScalars(rnd, -300.f, 300.f, &inputFloats0[0], numElements);
2512         fillRandomScalars(rnd, -300.f, 300.f, &inputFloats1[0], numElements);
2513         fillRandomScalars(rnd, -300.f, 300.f, &inputFloats2[0], numElements);
2514         fillRandomScalars(rnd, -300.f, 300.f, &inputFloats3[0], numElements);
2515         fillRandomScalars(rnd, -300.f, 300.f, &inputFloats4[0], numElements);
2516
2517         // CPU might not use the same rounding mode as the GPU. Use whole numbers to avoid rounding differences.
2518         floorAll(inputFloats0);
2519         floorAll(inputFloats1);
2520         floorAll(inputFloats2);
2521         floorAll(inputFloats3);
2522         floorAll(inputFloats4);
2523
2524         for (size_t ndx = 0; ndx < numElements; ++ndx)
2525                 outputFloats[ndx] = inputFloats0[ndx] + inputFloats1[ndx] + inputFloats2[ndx] + inputFloats3[ndx] + inputFloats4[ndx];
2526
2527         spec.assembly =
2528                 string(getComputeAsmShaderPreamble()) +
2529
2530                 "OpSource GLSL 430\n"
2531                 "OpName %main \"main\"\n"
2532                 "OpName %id \"gl_GlobalInvocationID\"\n"
2533
2534                 // Not using group decoration on variable.
2535                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
2536                 // Not using group decoration on type.
2537                 "OpDecorate %f32arr ArrayStride 4\n"
2538
2539                 "OpDecorate %groups BufferBlock\n"
2540                 "OpDecorate %groupm Offset 0\n"
2541                 "%groups = OpDecorationGroup\n"
2542                 "%groupm = OpDecorationGroup\n"
2543
2544                 // Group decoration on multiple structs.
2545                 "OpGroupDecorate %groups %outbuf %inbuf0 %inbuf1 %inbuf2 %inbuf3 %inbuf4\n"
2546                 // Group decoration on multiple struct members.
2547                 "OpGroupMemberDecorate %groupm %outbuf 0 %inbuf0 0 %inbuf1 0 %inbuf2 0 %inbuf3 0 %inbuf4 0\n"
2548
2549                 "OpDecorate %group1 DescriptorSet 0\n"
2550                 "OpDecorate %group3 DescriptorSet 0\n"
2551                 "OpDecorate %group3 NonWritable\n"
2552                 "OpDecorate %group3 Restrict\n"
2553                 "%group0 = OpDecorationGroup\n"
2554                 "%group1 = OpDecorationGroup\n"
2555                 "%group3 = OpDecorationGroup\n"
2556
2557                 // Applying the same decoration group multiple times.
2558                 "OpGroupDecorate %group1 %outdata\n"
2559                 "OpGroupDecorate %group1 %outdata\n"
2560                 "OpGroupDecorate %group1 %outdata\n"
2561                 "OpDecorate %outdata DescriptorSet 0\n"
2562                 "OpDecorate %outdata Binding 5\n"
2563                 // Applying decoration group containing nothing.
2564                 "OpGroupDecorate %group0 %indata0\n"
2565                 "OpDecorate %indata0 DescriptorSet 0\n"
2566                 "OpDecorate %indata0 Binding 0\n"
2567                 // Applying decoration group containing one decoration.
2568                 "OpGroupDecorate %group1 %indata1\n"
2569                 "OpDecorate %indata1 Binding 1\n"
2570                 // Applying decoration group containing multiple decorations.
2571                 "OpGroupDecorate %group3 %indata2 %indata3\n"
2572                 "OpDecorate %indata2 Binding 2\n"
2573                 "OpDecorate %indata3 Binding 3\n"
2574                 // Applying multiple decoration groups (with overlapping).
2575                 "OpGroupDecorate %group0 %indata4\n"
2576                 "OpGroupDecorate %group1 %indata4\n"
2577                 "OpGroupDecorate %group3 %indata4\n"
2578                 "OpDecorate %indata4 Binding 4\n"
2579
2580                 + string(getComputeAsmCommonTypes()) +
2581
2582                 "%id   = OpVariable %uvec3ptr Input\n"
2583                 "%zero = OpConstant %i32 0\n"
2584
2585                 "%outbuf    = OpTypeStruct %f32arr\n"
2586                 "%outbufptr = OpTypePointer Uniform %outbuf\n"
2587                 "%outdata   = OpVariable %outbufptr Uniform\n"
2588                 "%inbuf0    = OpTypeStruct %f32arr\n"
2589                 "%inbuf0ptr = OpTypePointer Uniform %inbuf0\n"
2590                 "%indata0   = OpVariable %inbuf0ptr Uniform\n"
2591                 "%inbuf1    = OpTypeStruct %f32arr\n"
2592                 "%inbuf1ptr = OpTypePointer Uniform %inbuf1\n"
2593                 "%indata1   = OpVariable %inbuf1ptr Uniform\n"
2594                 "%inbuf2    = OpTypeStruct %f32arr\n"
2595                 "%inbuf2ptr = OpTypePointer Uniform %inbuf2\n"
2596                 "%indata2   = OpVariable %inbuf2ptr Uniform\n"
2597                 "%inbuf3    = OpTypeStruct %f32arr\n"
2598                 "%inbuf3ptr = OpTypePointer Uniform %inbuf3\n"
2599                 "%indata3   = OpVariable %inbuf3ptr Uniform\n"
2600                 "%inbuf4    = OpTypeStruct %f32arr\n"
2601                 "%inbufptr  = OpTypePointer Uniform %inbuf4\n"
2602                 "%indata4   = OpVariable %inbufptr Uniform\n"
2603
2604                 "%main   = OpFunction %void None %voidf\n"
2605                 "%label  = OpLabel\n"
2606                 "%idval  = OpLoad %uvec3 %id\n"
2607                 "%x      = OpCompositeExtract %u32 %idval 0\n"
2608                 "%inloc0 = OpAccessChain %f32ptr %indata0 %zero %x\n"
2609                 "%inloc1 = OpAccessChain %f32ptr %indata1 %zero %x\n"
2610                 "%inloc2 = OpAccessChain %f32ptr %indata2 %zero %x\n"
2611                 "%inloc3 = OpAccessChain %f32ptr %indata3 %zero %x\n"
2612                 "%inloc4 = OpAccessChain %f32ptr %indata4 %zero %x\n"
2613                 "%outloc = OpAccessChain %f32ptr %outdata %zero %x\n"
2614                 "%inval0 = OpLoad %f32 %inloc0\n"
2615                 "%inval1 = OpLoad %f32 %inloc1\n"
2616                 "%inval2 = OpLoad %f32 %inloc2\n"
2617                 "%inval3 = OpLoad %f32 %inloc3\n"
2618                 "%inval4 = OpLoad %f32 %inloc4\n"
2619                 "%add0   = OpFAdd %f32 %inval0 %inval1\n"
2620                 "%add1   = OpFAdd %f32 %add0 %inval2\n"
2621                 "%add2   = OpFAdd %f32 %add1 %inval3\n"
2622                 "%add    = OpFAdd %f32 %add2 %inval4\n"
2623                 "          OpStore %outloc %add\n"
2624                 "          OpReturn\n"
2625                 "          OpFunctionEnd\n";
2626         spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats0)));
2627         spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats1)));
2628         spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats2)));
2629         spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats3)));
2630         spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats4)));
2631         spec.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
2632         spec.numWorkGroups = IVec3(numElements, 1, 1);
2633
2634         group->addChild(new SpvAsmComputeShaderCase(testCtx, "all", "decoration group cases", spec));
2635
2636         return group.release();
2637 }
2638
2639 struct SpecConstantTwoIntCase
2640 {
2641         const char*             caseName;
2642         const char*             scDefinition0;
2643         const char*             scDefinition1;
2644         const char*             scResultType;
2645         const char*             scOperation;
2646         deInt32                 scActualValue0;
2647         deInt32                 scActualValue1;
2648         const char*             resultOperation;
2649         vector<deInt32> expectedOutput;
2650         deInt32                 scActualValueLength;
2651
2652                                         SpecConstantTwoIntCase (const char* name,
2653                                                                                         const char* definition0,
2654                                                                                         const char* definition1,
2655                                                                                         const char* resultType,
2656                                                                                         const char* operation,
2657                                                                                         deInt32 value0,
2658                                                                                         deInt32 value1,
2659                                                                                         const char* resultOp,
2660                                                                                         const vector<deInt32>& output,
2661                                                                                         const deInt32   valueLength = sizeof(deInt32))
2662                                                 : caseName                              (name)
2663                                                 , scDefinition0                 (definition0)
2664                                                 , scDefinition1                 (definition1)
2665                                                 , scResultType                  (resultType)
2666                                                 , scOperation                   (operation)
2667                                                 , scActualValue0                (value0)
2668                                                 , scActualValue1                (value1)
2669                                                 , resultOperation               (resultOp)
2670                                                 , expectedOutput                (output)
2671                                                 , scActualValueLength   (valueLength)
2672                                                 {}
2673 };
2674
2675 tcu::TestCaseGroup* createSpecConstantGroup (tcu::TestContext& testCtx)
2676 {
2677         de::MovePtr<tcu::TestCaseGroup> group                   (new tcu::TestCaseGroup(testCtx, "opspecconstantop", "Test the OpSpecConstantOp instruction"));
2678         vector<SpecConstantTwoIntCase>  cases;
2679         de::Random                                              rnd                             (deStringHash(group->getName()));
2680         const int                                               numElements             = 100;
2681         const deInt32                                   p1AsFloat16             = 0x3c00; // +1(fp16) == 0 01111 0000000000 == 0011 1100 0000 0000
2682         vector<deInt32>                                 inputInts               (numElements, 0);
2683         vector<deInt32>                                 outputInts1             (numElements, 0);
2684         vector<deInt32>                                 outputInts2             (numElements, 0);
2685         vector<deInt32>                                 outputInts3             (numElements, 0);
2686         vector<deInt32>                                 outputInts4             (numElements, 0);
2687         const StringTemplate                    shaderTemplate  (
2688                 "${CAPABILITIES:opt}"
2689                 + string(getComputeAsmShaderPreamble()) +
2690
2691                 "OpName %main           \"main\"\n"
2692                 "OpName %id             \"gl_GlobalInvocationID\"\n"
2693
2694                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
2695                 "OpDecorate %sc_0  SpecId 0\n"
2696                 "OpDecorate %sc_1  SpecId 1\n"
2697                 "OpDecorate %i32arr ArrayStride 4\n"
2698
2699                 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) +
2700
2701                 "${OPTYPE_DEFINITIONS:opt}"
2702                 "%buf     = OpTypeStruct %i32arr\n"
2703                 "%bufptr  = OpTypePointer Uniform %buf\n"
2704                 "%indata    = OpVariable %bufptr Uniform\n"
2705                 "%outdata   = OpVariable %bufptr Uniform\n"
2706
2707                 "%id        = OpVariable %uvec3ptr Input\n"
2708                 "%zero      = OpConstant %i32 0\n"
2709
2710                 "%sc_0      = OpSpecConstant${SC_DEF0}\n"
2711                 "%sc_1      = OpSpecConstant${SC_DEF1}\n"
2712                 "%sc_final  = OpSpecConstantOp ${SC_RESULT_TYPE} ${SC_OP}\n"
2713
2714                 "%main      = OpFunction %void None %voidf\n"
2715                 "%label     = OpLabel\n"
2716                 "${TYPE_CONVERT:opt}"
2717                 "%idval     = OpLoad %uvec3 %id\n"
2718                 "%x         = OpCompositeExtract %u32 %idval 0\n"
2719                 "%inloc     = OpAccessChain %i32ptr %indata %zero %x\n"
2720                 "%inval     = OpLoad %i32 %inloc\n"
2721                 "%final     = ${GEN_RESULT}\n"
2722                 "%outloc    = OpAccessChain %i32ptr %outdata %zero %x\n"
2723                 "             OpStore %outloc %final\n"
2724                 "             OpReturn\n"
2725                 "             OpFunctionEnd\n");
2726
2727         fillRandomScalars(rnd, -65536, 65536, &inputInts[0], numElements);
2728
2729         for (size_t ndx = 0; ndx < numElements; ++ndx)
2730         {
2731                 outputInts1[ndx] = inputInts[ndx] + 42;
2732                 outputInts2[ndx] = inputInts[ndx];
2733                 outputInts3[ndx] = inputInts[ndx] - 11200;
2734                 outputInts4[ndx] = inputInts[ndx] + 1;
2735         }
2736
2737         const char addScToInput[]               = "OpIAdd %i32 %inval %sc_final";
2738         const char addSc32ToInput[]             = "OpIAdd %i32 %inval %sc_final32";
2739         const char selectTrueUsingSc[]  = "OpSelect %i32 %sc_final %inval %zero";
2740         const char selectFalseUsingSc[] = "OpSelect %i32 %sc_final %zero %inval";
2741
2742         cases.push_back(SpecConstantTwoIntCase("iadd",                                  " %i32 0",              " %i32 0",              "%i32",         "IAdd                 %sc_0 %sc_1",                     62,             -20,    addScToInput,           outputInts1));
2743         cases.push_back(SpecConstantTwoIntCase("isub",                                  " %i32 0",              " %i32 0",              "%i32",         "ISub                 %sc_0 %sc_1",                     100,    58,             addScToInput,           outputInts1));
2744         cases.push_back(SpecConstantTwoIntCase("imul",                                  " %i32 0",              " %i32 0",              "%i32",         "IMul                 %sc_0 %sc_1",                     -2,             -21,    addScToInput,           outputInts1));
2745         cases.push_back(SpecConstantTwoIntCase("sdiv",                                  " %i32 0",              " %i32 0",              "%i32",         "SDiv                 %sc_0 %sc_1",                     -126,   -3,             addScToInput,           outputInts1));
2746         cases.push_back(SpecConstantTwoIntCase("udiv",                                  " %i32 0",              " %i32 0",              "%i32",         "UDiv                 %sc_0 %sc_1",                     126,    3,              addScToInput,           outputInts1));
2747         cases.push_back(SpecConstantTwoIntCase("srem",                                  " %i32 0",              " %i32 0",              "%i32",         "SRem                 %sc_0 %sc_1",                     7,              3,              addScToInput,           outputInts4));
2748         cases.push_back(SpecConstantTwoIntCase("smod",                                  " %i32 0",              " %i32 0",              "%i32",         "SMod                 %sc_0 %sc_1",                     7,              3,              addScToInput,           outputInts4));
2749         cases.push_back(SpecConstantTwoIntCase("umod",                                  " %i32 0",              " %i32 0",              "%i32",         "UMod                 %sc_0 %sc_1",                     342,    50,             addScToInput,           outputInts1));
2750         cases.push_back(SpecConstantTwoIntCase("bitwiseand",                    " %i32 0",              " %i32 0",              "%i32",         "BitwiseAnd           %sc_0 %sc_1",                     42,             63,             addScToInput,           outputInts1));
2751         cases.push_back(SpecConstantTwoIntCase("bitwiseor",                             " %i32 0",              " %i32 0",              "%i32",         "BitwiseOr            %sc_0 %sc_1",                     34,             8,              addScToInput,           outputInts1));
2752         cases.push_back(SpecConstantTwoIntCase("bitwisexor",                    " %i32 0",              " %i32 0",              "%i32",         "BitwiseXor           %sc_0 %sc_1",                     18,             56,             addScToInput,           outputInts1));
2753         cases.push_back(SpecConstantTwoIntCase("shiftrightlogical",             " %i32 0",              " %i32 0",              "%i32",         "ShiftRightLogical    %sc_0 %sc_1",                     168,    2,              addScToInput,           outputInts1));
2754         cases.push_back(SpecConstantTwoIntCase("shiftrightarithmetic",  " %i32 0",              " %i32 0",              "%i32",         "ShiftRightArithmetic %sc_0 %sc_1",                     168,    2,              addScToInput,           outputInts1));
2755         cases.push_back(SpecConstantTwoIntCase("shiftleftlogical",              " %i32 0",              " %i32 0",              "%i32",         "ShiftLeftLogical     %sc_0 %sc_1",                     21,             1,              addScToInput,           outputInts1));
2756         cases.push_back(SpecConstantTwoIntCase("slessthan",                             " %i32 0",              " %i32 0",              "%bool",        "SLessThan            %sc_0 %sc_1",                     -20,    -10,    selectTrueUsingSc,      outputInts2));
2757         cases.push_back(SpecConstantTwoIntCase("ulessthan",                             " %i32 0",              " %i32 0",              "%bool",        "ULessThan            %sc_0 %sc_1",                     10,             20,             selectTrueUsingSc,      outputInts2));
2758         cases.push_back(SpecConstantTwoIntCase("sgreaterthan",                  " %i32 0",              " %i32 0",              "%bool",        "SGreaterThan         %sc_0 %sc_1",                     -1000,  50,             selectFalseUsingSc,     outputInts2));
2759         cases.push_back(SpecConstantTwoIntCase("ugreaterthan",                  " %i32 0",              " %i32 0",              "%bool",        "UGreaterThan         %sc_0 %sc_1",                     10,             5,              selectTrueUsingSc,      outputInts2));
2760         cases.push_back(SpecConstantTwoIntCase("slessthanequal",                " %i32 0",              " %i32 0",              "%bool",        "SLessThanEqual       %sc_0 %sc_1",                     -10,    -10,    selectTrueUsingSc,      outputInts2));
2761         cases.push_back(SpecConstantTwoIntCase("ulessthanequal",                " %i32 0",              " %i32 0",              "%bool",        "ULessThanEqual       %sc_0 %sc_1",                     50,             100,    selectTrueUsingSc,      outputInts2));
2762         cases.push_back(SpecConstantTwoIntCase("sgreaterthanequal",             " %i32 0",              " %i32 0",              "%bool",        "SGreaterThanEqual    %sc_0 %sc_1",                     -1000,  50,             selectFalseUsingSc,     outputInts2));
2763         cases.push_back(SpecConstantTwoIntCase("ugreaterthanequal",             " %i32 0",              " %i32 0",              "%bool",        "UGreaterThanEqual    %sc_0 %sc_1",                     10,             10,             selectTrueUsingSc,      outputInts2));
2764         cases.push_back(SpecConstantTwoIntCase("iequal",                                " %i32 0",              " %i32 0",              "%bool",        "IEqual               %sc_0 %sc_1",                     42,             24,             selectFalseUsingSc,     outputInts2));
2765         cases.push_back(SpecConstantTwoIntCase("inotequal",                             " %i32 0",              " %i32 0",              "%bool",        "INotEqual            %sc_0 %sc_1",                     42,             24,             selectTrueUsingSc,      outputInts2));
2766         cases.push_back(SpecConstantTwoIntCase("logicaland",                    "True %bool",   "True %bool",   "%bool",        "LogicalAnd           %sc_0 %sc_1",                     0,              1,              selectFalseUsingSc,     outputInts2));
2767         cases.push_back(SpecConstantTwoIntCase("logicalor",                             "False %bool",  "False %bool",  "%bool",        "LogicalOr            %sc_0 %sc_1",                     1,              0,              selectTrueUsingSc,      outputInts2));
2768         cases.push_back(SpecConstantTwoIntCase("logicalequal",                  "True %bool",   "True %bool",   "%bool",        "LogicalEqual         %sc_0 %sc_1",                     0,              1,              selectFalseUsingSc,     outputInts2));
2769         cases.push_back(SpecConstantTwoIntCase("logicalnotequal",               "False %bool",  "False %bool",  "%bool",        "LogicalNotEqual      %sc_0 %sc_1",                     1,              0,              selectTrueUsingSc,      outputInts2));
2770         cases.push_back(SpecConstantTwoIntCase("snegate",                               " %i32 0",              " %i32 0",              "%i32",         "SNegate              %sc_0",                           -42,    0,              addScToInput,           outputInts1));
2771         cases.push_back(SpecConstantTwoIntCase("not",                                   " %i32 0",              " %i32 0",              "%i32",         "Not                  %sc_0",                           -43,    0,              addScToInput,           outputInts1));
2772         cases.push_back(SpecConstantTwoIntCase("logicalnot",                    "False %bool",  "False %bool",  "%bool",        "LogicalNot           %sc_0",                           1,              0,              selectFalseUsingSc,     outputInts2));
2773         cases.push_back(SpecConstantTwoIntCase("select",                                "False %bool",  " %i32 0",              "%i32",         "Select               %sc_0 %sc_1 %zero",       1,              42,             addScToInput,           outputInts1));
2774         cases.push_back(SpecConstantTwoIntCase("sconvert",                              " %i32 0",              " %i32 0",              "%i16",         "SConvert             %sc_0",                           -11200, 0,              addSc32ToInput,         outputInts3));
2775         // -969998336 stored as 32-bit two's complement is the binary representation of -11200 as IEEE-754 Float
2776         cases.push_back(SpecConstantTwoIntCase("fconvert",                              " %f32 0",              " %f32 0",              "%f64",         "FConvert             %sc_0",                           -969998336, 0,  addSc32ToInput,         outputInts3));
2777         cases.push_back(SpecConstantTwoIntCase("fconvert16",                    " %f16 0",              " %f16 0",              "%f32",         "FConvert             %sc_0",                           p1AsFloat16, 0, addSc32ToInput,         outputInts4, sizeof(deFloat16)));
2778
2779         for (size_t caseNdx = 0; caseNdx < cases.size(); ++caseNdx)
2780         {
2781                 map<string, string>             specializations;
2782                 ComputeShaderSpec               spec;
2783
2784                 specializations["SC_DEF0"]                      = cases[caseNdx].scDefinition0;
2785                 specializations["SC_DEF1"]                      = cases[caseNdx].scDefinition1;
2786                 specializations["SC_RESULT_TYPE"]       = cases[caseNdx].scResultType;
2787                 specializations["SC_OP"]                        = cases[caseNdx].scOperation;
2788                 specializations["GEN_RESULT"]           = cases[caseNdx].resultOperation;
2789
2790                 // Special SPIR-V code for SConvert-case
2791                 if (strcmp(cases[caseNdx].caseName, "sconvert") == 0)
2792                 {
2793                         spec.requestedVulkanFeatures.coreFeatures.shaderInt16   = VK_TRUE;
2794                         specializations["CAPABILITIES"]                                                 = "OpCapability Int16\n";                                                       // Adds 16-bit integer capability
2795                         specializations["OPTYPE_DEFINITIONS"]                                   = "%i16 = OpTypeInt 16 1\n";                                            // Adds 16-bit integer type
2796                         specializations["TYPE_CONVERT"]                                                 = "%sc_final32 = OpSConvert %i32 %sc_final\n";          // Converts 16-bit integer to 32-bit integer
2797                 }
2798
2799                 // Special SPIR-V code for FConvert-case
2800                 if (strcmp(cases[caseNdx].caseName, "fconvert") == 0)
2801                 {
2802                         spec.requestedVulkanFeatures.coreFeatures.shaderFloat64 = VK_TRUE;
2803                         specializations["CAPABILITIES"]                                                 = "OpCapability Float64\n";                                                     // Adds 64-bit float capability
2804                         specializations["OPTYPE_DEFINITIONS"]                                   = "%f64 = OpTypeFloat 64\n";                                            // Adds 64-bit float type
2805                         specializations["TYPE_CONVERT"]                                                 = "%sc_final32 = OpConvertFToS %i32 %sc_final\n";       // Converts 64-bit float to 32-bit integer
2806                 }
2807
2808                 // Special SPIR-V code for FConvert-case for 16-bit floats
2809                 if (strcmp(cases[caseNdx].caseName, "fconvert16") == 0)
2810                 {
2811                         spec.extensions.push_back("VK_KHR_shader_float16_int8");
2812                         spec.requestedVulkanFeatures.extFloat16Int8 = EXTFLOAT16INT8FEATURES_FLOAT16;
2813                         specializations["CAPABILITIES"]                 = "OpCapability Float16\n";                                                     // Adds 16-bit float capability
2814                         specializations["OPTYPE_DEFINITIONS"]   = "%f16 = OpTypeFloat 16\n";                                            // Adds 16-bit float type
2815                         specializations["TYPE_CONVERT"]                 = "%sc_final32 = OpConvertFToS %i32 %sc_final\n";       // Converts 16-bit float to 32-bit integer
2816                 }
2817
2818                 spec.assembly = shaderTemplate.specialize(specializations);
2819                 spec.inputs.push_back(BufferSp(new Int32Buffer(inputInts)));
2820                 spec.outputs.push_back(BufferSp(new Int32Buffer(cases[caseNdx].expectedOutput)));
2821                 spec.numWorkGroups = IVec3(numElements, 1, 1);
2822                 spec.specConstants.append(&cases[caseNdx].scActualValue0, cases[caseNdx].scActualValueLength);
2823                 spec.specConstants.append(&cases[caseNdx].scActualValue1, cases[caseNdx].scActualValueLength);
2824
2825                 group->addChild(new SpvAsmComputeShaderCase(testCtx, cases[caseNdx].caseName, cases[caseNdx].caseName, spec));
2826         }
2827
2828         ComputeShaderSpec                               spec;
2829
2830         spec.assembly =
2831                 string(getComputeAsmShaderPreamble()) +
2832
2833                 "OpName %main           \"main\"\n"
2834                 "OpName %id             \"gl_GlobalInvocationID\"\n"
2835
2836                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
2837                 "OpDecorate %sc_0  SpecId 0\n"
2838                 "OpDecorate %sc_1  SpecId 1\n"
2839                 "OpDecorate %sc_2  SpecId 2\n"
2840                 "OpDecorate %i32arr ArrayStride 4\n"
2841
2842                 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) +
2843
2844                 "%ivec3       = OpTypeVector %i32 3\n"
2845                 "%buf         = OpTypeStruct %i32arr\n"
2846                 "%bufptr      = OpTypePointer Uniform %buf\n"
2847                 "%indata      = OpVariable %bufptr Uniform\n"
2848                 "%outdata     = OpVariable %bufptr Uniform\n"
2849
2850                 "%id          = OpVariable %uvec3ptr Input\n"
2851                 "%zero        = OpConstant %i32 0\n"
2852                 "%ivec3_0     = OpConstantComposite %ivec3 %zero %zero %zero\n"
2853                 "%vec3_undef  = OpUndef %ivec3\n"
2854
2855                 "%sc_0        = OpSpecConstant %i32 0\n"
2856                 "%sc_1        = OpSpecConstant %i32 0\n"
2857                 "%sc_2        = OpSpecConstant %i32 0\n"
2858                 "%sc_vec3_0   = OpSpecConstantOp %ivec3 CompositeInsert  %sc_0        %ivec3_0     0\n"                                                 // (sc_0, 0, 0)
2859                 "%sc_vec3_1   = OpSpecConstantOp %ivec3 CompositeInsert  %sc_1        %ivec3_0     1\n"                                                 // (0, sc_1, 0)
2860                 "%sc_vec3_2   = OpSpecConstantOp %ivec3 CompositeInsert  %sc_2        %ivec3_0     2\n"                                                 // (0, 0, sc_2)
2861                 "%sc_vec3_0_s = OpSpecConstantOp %ivec3 VectorShuffle    %sc_vec3_0   %vec3_undef  0          0xFFFFFFFF 2\n"   // (sc_0, ???,  0)
2862                 "%sc_vec3_1_s = OpSpecConstantOp %ivec3 VectorShuffle    %sc_vec3_1   %vec3_undef  0xFFFFFFFF 1          0\n"   // (???,  sc_1, 0)
2863                 "%sc_vec3_2_s = OpSpecConstantOp %ivec3 VectorShuffle    %vec3_undef  %sc_vec3_2   5          0xFFFFFFFF 5\n"   // (sc_2, ???,  sc_2)
2864                 "%sc_vec3_01  = OpSpecConstantOp %ivec3 VectorShuffle    %sc_vec3_0_s %sc_vec3_1_s 1 0 4\n"                                             // (0,    sc_0, sc_1)
2865                 "%sc_vec3_012 = OpSpecConstantOp %ivec3 VectorShuffle    %sc_vec3_01  %sc_vec3_2_s 5 1 2\n"                                             // (sc_2, sc_0, sc_1)
2866                 "%sc_ext_0    = OpSpecConstantOp %i32   CompositeExtract %sc_vec3_012              0\n"                                                 // sc_2
2867                 "%sc_ext_1    = OpSpecConstantOp %i32   CompositeExtract %sc_vec3_012              1\n"                                                 // sc_0
2868                 "%sc_ext_2    = OpSpecConstantOp %i32   CompositeExtract %sc_vec3_012              2\n"                                                 // sc_1
2869                 "%sc_sub      = OpSpecConstantOp %i32   ISub             %sc_ext_0    %sc_ext_1\n"                                                              // (sc_2 - sc_0)
2870                 "%sc_final    = OpSpecConstantOp %i32   IMul             %sc_sub      %sc_ext_2\n"                                                              // (sc_2 - sc_0) * sc_1
2871
2872                 "%main      = OpFunction %void None %voidf\n"
2873                 "%label     = OpLabel\n"
2874                 "%idval     = OpLoad %uvec3 %id\n"
2875                 "%x         = OpCompositeExtract %u32 %idval 0\n"
2876                 "%inloc     = OpAccessChain %i32ptr %indata %zero %x\n"
2877                 "%inval     = OpLoad %i32 %inloc\n"
2878                 "%final     = OpIAdd %i32 %inval %sc_final\n"
2879                 "%outloc    = OpAccessChain %i32ptr %outdata %zero %x\n"
2880                 "             OpStore %outloc %final\n"
2881                 "             OpReturn\n"
2882                 "             OpFunctionEnd\n";
2883         spec.inputs.push_back(BufferSp(new Int32Buffer(inputInts)));
2884         spec.outputs.push_back(BufferSp(new Int32Buffer(outputInts3)));
2885         spec.numWorkGroups = IVec3(numElements, 1, 1);
2886         spec.specConstants.append<deInt32>(123);
2887         spec.specConstants.append<deInt32>(56);
2888         spec.specConstants.append<deInt32>(-77);
2889
2890         group->addChild(new SpvAsmComputeShaderCase(testCtx, "vector_related", "VectorShuffle, CompositeExtract, & CompositeInsert", spec));
2891
2892         return group.release();
2893 }
2894
2895 void createOpPhiVartypeTests (de::MovePtr<tcu::TestCaseGroup>& group, tcu::TestContext& testCtx)
2896 {
2897         ComputeShaderSpec       specInt;
2898         ComputeShaderSpec       specFloat;
2899         ComputeShaderSpec       specFloat16;
2900         ComputeShaderSpec       specVec3;
2901         ComputeShaderSpec       specMat4;
2902         ComputeShaderSpec       specArray;
2903         ComputeShaderSpec       specStruct;
2904         de::Random                      rnd                             (deStringHash(group->getName()));
2905         const int                       numElements             = 100;
2906         vector<float>           inputFloats             (numElements, 0);
2907         vector<float>           outputFloats    (numElements, 0);
2908         vector<deFloat16>       inputFloats16   (numElements, 0);
2909         vector<deFloat16>       outputFloats16  (numElements, 0);
2910
2911         fillRandomScalars(rnd, -300.f, 300.f, &inputFloats[0], numElements);
2912
2913         // CPU might not use the same rounding mode as the GPU. Use whole numbers to avoid rounding differences.
2914         floorAll(inputFloats);
2915
2916         for (size_t ndx = 0; ndx < numElements; ++ndx)
2917         {
2918                 // Just check if the value is positive or not
2919                 outputFloats[ndx] = (inputFloats[ndx] > 0) ? 1.0f : -1.0f;
2920         }
2921
2922         for (size_t ndx = 0; ndx < numElements; ++ndx)
2923         {
2924                 inputFloats16[ndx] = tcu::Float16(inputFloats[ndx]).bits();
2925                 outputFloats16[ndx] = tcu::Float16(outputFloats[ndx]).bits();
2926         }
2927
2928         // All of the tests are of the form:
2929         //
2930         // testtype r
2931         //
2932         // if (inputdata > 0)
2933         //   r = 1
2934         // else
2935         //   r = -1
2936         //
2937         // return (float)r
2938
2939         specFloat.assembly =
2940                 string(getComputeAsmShaderPreamble()) +
2941
2942                 "OpSource GLSL 430\n"
2943                 "OpName %main \"main\"\n"
2944                 "OpName %id \"gl_GlobalInvocationID\"\n"
2945
2946                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
2947
2948                 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) + string(getComputeAsmInputOutputBuffer()) +
2949
2950                 "%id = OpVariable %uvec3ptr Input\n"
2951                 "%zero       = OpConstant %i32 0\n"
2952                 "%float_0    = OpConstant %f32 0.0\n"
2953                 "%float_1    = OpConstant %f32 1.0\n"
2954                 "%float_n1   = OpConstant %f32 -1.0\n"
2955
2956                 "%main     = OpFunction %void None %voidf\n"
2957                 "%entry    = OpLabel\n"
2958                 "%idval    = OpLoad %uvec3 %id\n"
2959                 "%x        = OpCompositeExtract %u32 %idval 0\n"
2960                 "%inloc    = OpAccessChain %f32ptr %indata %zero %x\n"
2961                 "%inval    = OpLoad %f32 %inloc\n"
2962
2963                 "%comp     = OpFOrdGreaterThan %bool %inval %float_0\n"
2964                 "            OpSelectionMerge %cm None\n"
2965                 "            OpBranchConditional %comp %tb %fb\n"
2966                 "%tb       = OpLabel\n"
2967                 "            OpBranch %cm\n"
2968                 "%fb       = OpLabel\n"
2969                 "            OpBranch %cm\n"
2970                 "%cm       = OpLabel\n"
2971                 "%res      = OpPhi %f32 %float_1 %tb %float_n1 %fb\n"
2972
2973                 "%outloc   = OpAccessChain %f32ptr %outdata %zero %x\n"
2974                 "            OpStore %outloc %res\n"
2975                 "            OpReturn\n"
2976
2977                 "            OpFunctionEnd\n";
2978         specFloat.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
2979         specFloat.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
2980         specFloat.numWorkGroups = IVec3(numElements, 1, 1);
2981
2982         specFloat16.assembly =
2983                 "OpCapability Shader\n"
2984                 "OpCapability StorageUniformBufferBlock16\n"
2985                 "OpExtension \"SPV_KHR_16bit_storage\"\n"
2986                 "OpMemoryModel Logical GLSL450\n"
2987                 "OpEntryPoint GLCompute %main \"main\" %id\n"
2988                 "OpExecutionMode %main LocalSize 1 1 1\n"
2989
2990                 "OpSource GLSL 430\n"
2991                 "OpName %main \"main\"\n"
2992                 "OpName %id \"gl_GlobalInvocationID\"\n"
2993
2994                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
2995
2996                 "OpDecorate %buf BufferBlock\n"
2997                 "OpDecorate %indata DescriptorSet 0\n"
2998                 "OpDecorate %indata Binding 0\n"
2999                 "OpDecorate %outdata DescriptorSet 0\n"
3000                 "OpDecorate %outdata Binding 1\n"
3001                 "OpDecorate %f16arr ArrayStride 2\n"
3002                 "OpMemberDecorate %buf 0 Offset 0\n"
3003
3004                 "%f16      = OpTypeFloat 16\n"
3005                 "%f16ptr   = OpTypePointer Uniform %f16\n"
3006                 "%f16arr   = OpTypeRuntimeArray %f16\n"
3007
3008                 + string(getComputeAsmCommonTypes()) +
3009
3010                 "%buf      = OpTypeStruct %f16arr\n"
3011                 "%bufptr   = OpTypePointer Uniform %buf\n"
3012                 "%indata   = OpVariable %bufptr Uniform\n"
3013                 "%outdata  = OpVariable %bufptr Uniform\n"
3014
3015                 "%id       = OpVariable %uvec3ptr Input\n"
3016                 "%zero     = OpConstant %i32 0\n"
3017                 "%float_0  = OpConstant %f16 0.0\n"
3018                 "%float_1  = OpConstant %f16 1.0\n"
3019                 "%float_n1 = OpConstant %f16 -1.0\n"
3020
3021                 "%main     = OpFunction %void None %voidf\n"
3022                 "%entry    = OpLabel\n"
3023                 "%idval    = OpLoad %uvec3 %id\n"
3024                 "%x        = OpCompositeExtract %u32 %idval 0\n"
3025                 "%inloc    = OpAccessChain %f16ptr %indata %zero %x\n"
3026                 "%inval    = OpLoad %f16 %inloc\n"
3027
3028                 "%comp     = OpFOrdGreaterThan %bool %inval %float_0\n"
3029                 "            OpSelectionMerge %cm None\n"
3030                 "            OpBranchConditional %comp %tb %fb\n"
3031                 "%tb       = OpLabel\n"
3032                 "            OpBranch %cm\n"
3033                 "%fb       = OpLabel\n"
3034                 "            OpBranch %cm\n"
3035                 "%cm       = OpLabel\n"
3036                 "%res      = OpPhi %f16 %float_1 %tb %float_n1 %fb\n"
3037
3038                 "%outloc   = OpAccessChain %f16ptr %outdata %zero %x\n"
3039                 "            OpStore %outloc %res\n"
3040                 "            OpReturn\n"
3041
3042                 "            OpFunctionEnd\n";
3043         specFloat16.inputs.push_back(BufferSp(new Float16Buffer(inputFloats16)));
3044         specFloat16.outputs.push_back(BufferSp(new Float16Buffer(outputFloats16)));
3045         specFloat16.numWorkGroups = IVec3(numElements, 1, 1);
3046         specFloat16.requestedVulkanFeatures.ext16BitStorage = EXT16BITSTORAGEFEATURES_UNIFORM_BUFFER_BLOCK;
3047         specFloat16.requestedVulkanFeatures.extFloat16Int8 = EXTFLOAT16INT8FEATURES_FLOAT16;
3048
3049         specMat4.assembly =
3050                 string(getComputeAsmShaderPreamble()) +
3051
3052                 "OpSource GLSL 430\n"
3053                 "OpName %main \"main\"\n"
3054                 "OpName %id \"gl_GlobalInvocationID\"\n"
3055
3056                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
3057
3058                 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) + string(getComputeAsmInputOutputBuffer()) +
3059
3060                 "%id = OpVariable %uvec3ptr Input\n"
3061                 "%v4f32      = OpTypeVector %f32 4\n"
3062                 "%mat4v4f32  = OpTypeMatrix %v4f32 4\n"
3063                 "%zero       = OpConstant %i32 0\n"
3064                 "%float_0    = OpConstant %f32 0.0\n"
3065                 "%float_1    = OpConstant %f32 1.0\n"
3066                 "%float_n1   = OpConstant %f32 -1.0\n"
3067                 "%m11        = OpConstantComposite %v4f32 %float_1 %float_0 %float_0 %float_0\n"
3068                 "%m12        = OpConstantComposite %v4f32 %float_0 %float_1 %float_0 %float_0\n"
3069                 "%m13        = OpConstantComposite %v4f32 %float_0 %float_0 %float_1 %float_0\n"
3070                 "%m14        = OpConstantComposite %v4f32 %float_0 %float_0 %float_0 %float_1\n"
3071                 "%m1         = OpConstantComposite %mat4v4f32 %m11 %m12 %m13 %m14\n"
3072                 "%m21        = OpConstantComposite %v4f32 %float_n1 %float_0 %float_0 %float_0\n"
3073                 "%m22        = OpConstantComposite %v4f32 %float_0 %float_n1 %float_0 %float_0\n"
3074                 "%m23        = OpConstantComposite %v4f32 %float_0 %float_0 %float_n1 %float_0\n"
3075                 "%m24        = OpConstantComposite %v4f32 %float_0 %float_0 %float_0 %float_n1\n"
3076                 "%m2         = OpConstantComposite %mat4v4f32 %m21 %m22 %m23 %m24\n"
3077
3078                 "%main     = OpFunction %void None %voidf\n"
3079                 "%entry    = OpLabel\n"
3080                 "%idval    = OpLoad %uvec3 %id\n"
3081                 "%x        = OpCompositeExtract %u32 %idval 0\n"
3082                 "%inloc    = OpAccessChain %f32ptr %indata %zero %x\n"
3083                 "%inval    = OpLoad %f32 %inloc\n"
3084
3085                 "%comp     = OpFOrdGreaterThan %bool %inval %float_0\n"
3086                 "            OpSelectionMerge %cm None\n"
3087                 "            OpBranchConditional %comp %tb %fb\n"
3088                 "%tb       = OpLabel\n"
3089                 "            OpBranch %cm\n"
3090                 "%fb       = OpLabel\n"
3091                 "            OpBranch %cm\n"
3092                 "%cm       = OpLabel\n"
3093                 "%mres     = OpPhi %mat4v4f32 %m1 %tb %m2 %fb\n"
3094                 "%res      = OpCompositeExtract %f32 %mres 2 2\n"
3095
3096                 "%outloc   = OpAccessChain %f32ptr %outdata %zero %x\n"
3097                 "            OpStore %outloc %res\n"
3098                 "            OpReturn\n"
3099
3100                 "            OpFunctionEnd\n";
3101         specMat4.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
3102         specMat4.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
3103         specMat4.numWorkGroups = IVec3(numElements, 1, 1);
3104
3105         specVec3.assembly =
3106                 string(getComputeAsmShaderPreamble()) +
3107
3108                 "OpSource GLSL 430\n"
3109                 "OpName %main \"main\"\n"
3110                 "OpName %id \"gl_GlobalInvocationID\"\n"
3111
3112                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
3113
3114                 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) + string(getComputeAsmInputOutputBuffer()) +
3115
3116                 "%id = OpVariable %uvec3ptr Input\n"
3117                 "%zero       = OpConstant %i32 0\n"
3118                 "%float_0    = OpConstant %f32 0.0\n"
3119                 "%float_1    = OpConstant %f32 1.0\n"
3120                 "%float_n1   = OpConstant %f32 -1.0\n"
3121                 "%v1         = OpConstantComposite %fvec3 %float_1 %float_1 %float_1\n"
3122                 "%v2         = OpConstantComposite %fvec3 %float_n1 %float_n1 %float_n1\n"
3123
3124                 "%main     = OpFunction %void None %voidf\n"
3125                 "%entry    = OpLabel\n"
3126                 "%idval    = OpLoad %uvec3 %id\n"
3127                 "%x        = OpCompositeExtract %u32 %idval 0\n"
3128                 "%inloc    = OpAccessChain %f32ptr %indata %zero %x\n"
3129                 "%inval    = OpLoad %f32 %inloc\n"
3130
3131                 "%comp     = OpFOrdGreaterThan %bool %inval %float_0\n"
3132                 "            OpSelectionMerge %cm None\n"
3133                 "            OpBranchConditional %comp %tb %fb\n"
3134                 "%tb       = OpLabel\n"
3135                 "            OpBranch %cm\n"
3136                 "%fb       = OpLabel\n"
3137                 "            OpBranch %cm\n"
3138                 "%cm       = OpLabel\n"
3139                 "%vres     = OpPhi %fvec3 %v1 %tb %v2 %fb\n"
3140                 "%res      = OpCompositeExtract %f32 %vres 2\n"
3141
3142                 "%outloc   = OpAccessChain %f32ptr %outdata %zero %x\n"
3143                 "            OpStore %outloc %res\n"
3144                 "            OpReturn\n"
3145
3146                 "            OpFunctionEnd\n";
3147         specVec3.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
3148         specVec3.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
3149         specVec3.numWorkGroups = IVec3(numElements, 1, 1);
3150
3151         specInt.assembly =
3152                 string(getComputeAsmShaderPreamble()) +
3153
3154                 "OpSource GLSL 430\n"
3155                 "OpName %main \"main\"\n"
3156                 "OpName %id \"gl_GlobalInvocationID\"\n"
3157
3158                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
3159
3160                 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) + string(getComputeAsmInputOutputBuffer()) +
3161
3162                 "%id = OpVariable %uvec3ptr Input\n"
3163                 "%zero       = OpConstant %i32 0\n"
3164                 "%float_0    = OpConstant %f32 0.0\n"
3165                 "%i1         = OpConstant %i32 1\n"
3166                 "%i2         = OpConstant %i32 -1\n"
3167
3168                 "%main     = OpFunction %void None %voidf\n"
3169                 "%entry    = OpLabel\n"
3170                 "%idval    = OpLoad %uvec3 %id\n"
3171                 "%x        = OpCompositeExtract %u32 %idval 0\n"
3172                 "%inloc    = OpAccessChain %f32ptr %indata %zero %x\n"
3173                 "%inval    = OpLoad %f32 %inloc\n"
3174
3175                 "%comp     = OpFOrdGreaterThan %bool %inval %float_0\n"
3176                 "            OpSelectionMerge %cm None\n"
3177                 "            OpBranchConditional %comp %tb %fb\n"
3178                 "%tb       = OpLabel\n"
3179                 "            OpBranch %cm\n"
3180                 "%fb       = OpLabel\n"
3181                 "            OpBranch %cm\n"
3182                 "%cm       = OpLabel\n"
3183                 "%ires     = OpPhi %i32 %i1 %tb %i2 %fb\n"
3184                 "%res      = OpConvertSToF %f32 %ires\n"
3185
3186                 "%outloc   = OpAccessChain %f32ptr %outdata %zero %x\n"
3187                 "            OpStore %outloc %res\n"
3188                 "            OpReturn\n"
3189
3190                 "            OpFunctionEnd\n";
3191         specInt.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
3192         specInt.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
3193         specInt.numWorkGroups = IVec3(numElements, 1, 1);
3194
3195         specArray.assembly =
3196                 string(getComputeAsmShaderPreamble()) +
3197
3198                 "OpSource GLSL 430\n"
3199                 "OpName %main \"main\"\n"
3200                 "OpName %id \"gl_GlobalInvocationID\"\n"
3201
3202                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
3203
3204                 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) + string(getComputeAsmInputOutputBuffer()) +
3205
3206                 "%id = OpVariable %uvec3ptr Input\n"
3207                 "%zero       = OpConstant %i32 0\n"
3208                 "%u7         = OpConstant %u32 7\n"
3209                 "%float_0    = OpConstant %f32 0.0\n"
3210                 "%float_1    = OpConstant %f32 1.0\n"
3211                 "%float_n1   = OpConstant %f32 -1.0\n"
3212                 "%f32a7      = OpTypeArray %f32 %u7\n"
3213                 "%a1         = OpConstantComposite %f32a7 %float_1 %float_1 %float_1 %float_1 %float_1 %float_1 %float_1\n"
3214                 "%a2         = OpConstantComposite %f32a7 %float_n1 %float_n1 %float_n1 %float_n1 %float_n1 %float_n1 %float_n1\n"
3215                 "%main     = OpFunction %void None %voidf\n"
3216                 "%entry    = OpLabel\n"
3217                 "%idval    = OpLoad %uvec3 %id\n"
3218                 "%x        = OpCompositeExtract %u32 %idval 0\n"
3219                 "%inloc    = OpAccessChain %f32ptr %indata %zero %x\n"
3220                 "%inval    = OpLoad %f32 %inloc\n"
3221
3222                 "%comp     = OpFOrdGreaterThan %bool %inval %float_0\n"
3223                 "            OpSelectionMerge %cm None\n"
3224                 "            OpBranchConditional %comp %tb %fb\n"
3225                 "%tb       = OpLabel\n"
3226                 "            OpBranch %cm\n"
3227                 "%fb       = OpLabel\n"
3228                 "            OpBranch %cm\n"
3229                 "%cm       = OpLabel\n"
3230                 "%ares     = OpPhi %f32a7 %a1 %tb %a2 %fb\n"
3231                 "%res      = OpCompositeExtract %f32 %ares 5\n"
3232
3233                 "%outloc   = OpAccessChain %f32ptr %outdata %zero %x\n"
3234                 "            OpStore %outloc %res\n"
3235                 "            OpReturn\n"
3236
3237                 "            OpFunctionEnd\n";
3238         specArray.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
3239         specArray.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
3240         specArray.numWorkGroups = IVec3(numElements, 1, 1);
3241
3242         specStruct.assembly =
3243                 string(getComputeAsmShaderPreamble()) +
3244
3245                 "OpSource GLSL 430\n"
3246                 "OpName %main \"main\"\n"
3247                 "OpName %id \"gl_GlobalInvocationID\"\n"
3248
3249                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
3250
3251                 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) + string(getComputeAsmInputOutputBuffer()) +
3252
3253                 "%id = OpVariable %uvec3ptr Input\n"
3254                 "%zero       = OpConstant %i32 0\n"
3255                 "%float_0    = OpConstant %f32 0.0\n"
3256                 "%float_1    = OpConstant %f32 1.0\n"
3257                 "%float_n1   = OpConstant %f32 -1.0\n"
3258
3259                 "%v2f32      = OpTypeVector %f32 2\n"
3260                 "%Data2      = OpTypeStruct %f32 %v2f32\n"
3261                 "%Data       = OpTypeStruct %Data2 %f32\n"
3262
3263                 "%in1a       = OpConstantComposite %v2f32 %float_1 %float_1\n"
3264                 "%in1b       = OpConstantComposite %Data2 %float_1 %in1a\n"
3265                 "%s1         = OpConstantComposite %Data %in1b %float_1\n"
3266                 "%in2a       = OpConstantComposite %v2f32 %float_n1 %float_n1\n"
3267                 "%in2b       = OpConstantComposite %Data2 %float_n1 %in2a\n"
3268                 "%s2         = OpConstantComposite %Data %in2b %float_n1\n"
3269
3270                 "%main     = OpFunction %void None %voidf\n"
3271                 "%entry    = OpLabel\n"
3272                 "%idval    = OpLoad %uvec3 %id\n"
3273                 "%x        = OpCompositeExtract %u32 %idval 0\n"
3274                 "%inloc    = OpAccessChain %f32ptr %indata %zero %x\n"
3275                 "%inval    = OpLoad %f32 %inloc\n"
3276
3277                 "%comp     = OpFOrdGreaterThan %bool %inval %float_0\n"
3278                 "            OpSelectionMerge %cm None\n"
3279                 "            OpBranchConditional %comp %tb %fb\n"
3280                 "%tb       = OpLabel\n"
3281                 "            OpBranch %cm\n"
3282                 "%fb       = OpLabel\n"
3283                 "            OpBranch %cm\n"
3284                 "%cm       = OpLabel\n"
3285                 "%sres     = OpPhi %Data %s1 %tb %s2 %fb\n"
3286                 "%res      = OpCompositeExtract %f32 %sres 0 0\n"
3287
3288                 "%outloc   = OpAccessChain %f32ptr %outdata %zero %x\n"
3289                 "            OpStore %outloc %res\n"
3290                 "            OpReturn\n"
3291
3292                 "            OpFunctionEnd\n";
3293         specStruct.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
3294         specStruct.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
3295         specStruct.numWorkGroups = IVec3(numElements, 1, 1);
3296
3297         group->addChild(new SpvAsmComputeShaderCase(testCtx, "vartype_int", "OpPhi with int variables", specInt));
3298         group->addChild(new SpvAsmComputeShaderCase(testCtx, "vartype_float", "OpPhi with float variables", specFloat));
3299         group->addChild(new SpvAsmComputeShaderCase(testCtx, "vartype_float16", "OpPhi with 16bit float variables", specFloat16));
3300         group->addChild(new SpvAsmComputeShaderCase(testCtx, "vartype_vec3", "OpPhi with vec3 variables", specVec3));
3301         group->addChild(new SpvAsmComputeShaderCase(testCtx, "vartype_mat4", "OpPhi with mat4 variables", specMat4));
3302         group->addChild(new SpvAsmComputeShaderCase(testCtx, "vartype_array", "OpPhi with array variables", specArray));
3303         group->addChild(new SpvAsmComputeShaderCase(testCtx, "vartype_struct", "OpPhi with struct variables", specStruct));
3304 }
3305
3306 string generateConstantDefinitions (int count)
3307 {
3308         std::ostringstream      r;
3309         for (int i = 0; i < count; i++)
3310                 r << "%cf" << (i * 10 + 5) << " = OpConstant %f32 " <<(i * 10 + 5) << ".0\n";
3311         r << "\n";
3312         return r.str();
3313 }
3314
3315 string generateSwitchCases (int count)
3316 {
3317         std::ostringstream      r;
3318         for (int i = 0; i < count; i++)
3319                 r << " " << i << " %case" << i;
3320         r << "\n";
3321         return r.str();
3322 }
3323
3324 string generateSwitchTargets (int count)
3325 {
3326         std::ostringstream      r;
3327         for (int i = 0; i < count; i++)
3328                 r << "%case" << i << " = OpLabel\n            OpBranch %phi\n";
3329         r << "\n";
3330         return r.str();
3331 }
3332
3333 string generateOpPhiParams (int count)
3334 {
3335         std::ostringstream      r;
3336         for (int i = 0; i < count; i++)
3337                 r << " %cf" << (i * 10 + 5) << " %case" << i;
3338         r << "\n";
3339         return r.str();
3340 }
3341
3342 string generateIntWidth (int value)
3343 {
3344         std::ostringstream      r;
3345         r << value;
3346         return r.str();
3347 }
3348
3349 // Expand input string by injecting "ABC" between the input
3350 // string characters. The acc/add/treshold parameters are used
3351 // to skip some of the injections to make the result less
3352 // uniform (and a lot shorter).
3353 string expandOpPhiCase5 (const string& s, int &acc, int add, int treshold)
3354 {
3355         std::ostringstream      res;
3356         const char*                     p = s.c_str();
3357
3358         while (*p)
3359         {
3360                 res << *p;
3361                 acc += add;
3362                 if (acc > treshold)
3363                 {
3364                         acc -= treshold;
3365                         res << "ABC";
3366                 }
3367                 p++;
3368         }
3369         return res.str();
3370 }
3371
3372 // Calculate expected result based on the code string
3373 float calcOpPhiCase5 (float val, const string& s)
3374 {
3375         const char*             p               = s.c_str();
3376         float                   x[8];
3377         bool                    b[8];
3378         const float             tv[8]   = { 0.5f, 1.5f, 3.5f, 7.5f, 15.5f, 31.5f, 63.5f, 127.5f };
3379         const float             v               = deFloatAbs(val);
3380         float                   res             = 0;
3381         int                             depth   = -1;
3382         int                             skip    = 0;
3383
3384         for (int i = 7; i >= 0; --i)
3385                 x[i] = std::fmod((float)v, (float)(2 << i));
3386         for (int i = 7; i >= 0; --i)
3387                 b[i] = x[i] > tv[i];
3388
3389         while (*p)
3390         {
3391                 if (*p == 'A')
3392                 {
3393                         depth++;
3394                         if (skip == 0 && b[depth])
3395                         {
3396                                 res++;
3397                         }
3398                         else
3399                                 skip++;
3400                 }
3401                 if (*p == 'B')
3402                 {
3403                         if (skip)
3404                                 skip--;
3405                         if (b[depth] || skip)
3406                                 skip++;
3407                 }
3408                 if (*p == 'C')
3409                 {
3410                         depth--;
3411                         if (skip)
3412                                 skip--;
3413                 }
3414                 p++;
3415         }
3416         return res;
3417 }
3418
3419 // In the code string, the letters represent the following:
3420 //
3421 // A:
3422 //     if (certain bit is set)
3423 //     {
3424 //       result++;
3425 //
3426 // B:
3427 //     } else {
3428 //
3429 // C:
3430 //     }
3431 //
3432 // examples:
3433 // AABCBC leads to if(){r++;if(){r++;}else{}}else{}
3434 // ABABCC leads to if(){r++;}else{if(){r++;}else{}}
3435 // ABCABC leads to if(){r++;}else{}if(){r++;}else{}
3436 //
3437 // Code generation gets a bit complicated due to the else-branches,
3438 // which do not generate new values. Thus, the generator needs to
3439 // keep track of the previous variable change seen by the else
3440 // branch.
3441 string generateOpPhiCase5 (const string& s)
3442 {
3443         std::stack<int>                         idStack;
3444         std::stack<std::string>         value;
3445         std::stack<std::string>         valueLabel;
3446         std::stack<std::string>         mergeLeft;
3447         std::stack<std::string>         mergeRight;
3448         std::ostringstream                      res;
3449         const char*                                     p                       = s.c_str();
3450         int                                                     depth           = -1;
3451         int                                                     currId          = 0;
3452         int                                                     iter            = 0;
3453
3454         idStack.push(-1);
3455         value.push("%f32_0");
3456         valueLabel.push("%f32_0 %entry");
3457
3458         while (*p)
3459         {
3460                 if (*p == 'A')
3461                 {
3462                         depth++;
3463                         currId = iter;
3464                         idStack.push(currId);
3465                         res << "\tOpSelectionMerge %m" << currId << " None\n";
3466                         res << "\tOpBranchConditional %b" << depth << " %t" << currId << " %f" << currId << "\n";
3467                         res << "%t" << currId << " = OpLabel\n";
3468                         res << "%rt" << currId << " = OpFAdd %f32 " << value.top() << " %f32_1\n";
3469                         std::ostringstream tag;
3470                         tag << "%rt" << currId;
3471                         value.push(tag.str());
3472                         tag << " %t" << currId;
3473                         valueLabel.push(tag.str());
3474                 }
3475
3476                 if (*p == 'B')
3477                 {
3478                         mergeLeft.push(valueLabel.top());
3479                         value.pop();
3480                         valueLabel.pop();
3481                         res << "\tOpBranch %m" << currId << "\n";
3482                         res << "%f" << currId << " = OpLabel\n";
3483                         std::ostringstream tag;
3484                         tag << value.top() << " %f" << currId;
3485                         valueLabel.pop();
3486                         valueLabel.push(tag.str());
3487                 }
3488
3489                 if (*p == 'C')
3490                 {
3491                         mergeRight.push(valueLabel.top());
3492                         res << "\tOpBranch %m" << currId << "\n";
3493                         res << "%m" << currId << " = OpLabel\n";
3494                         if (*(p + 1) == 0)
3495                                 res << "%res"; // last result goes to %res
3496                         else
3497                                 res << "%rm" << currId;
3498                         res << " = OpPhi %f32  " << mergeLeft.top() << "  " << mergeRight.top() << "\n";
3499                         std::ostringstream tag;
3500                         tag << "%rm" << currId;
3501                         value.pop();
3502                         value.push(tag.str());
3503                         tag << " %m" << currId;
3504                         valueLabel.pop();
3505                         valueLabel.push(tag.str());
3506                         mergeLeft.pop();
3507                         mergeRight.pop();
3508                         depth--;
3509                         idStack.pop();
3510                         currId = idStack.top();
3511                 }
3512                 p++;
3513                 iter++;
3514         }
3515         return res.str();
3516 }
3517
3518 tcu::TestCaseGroup* createOpPhiGroup (tcu::TestContext& testCtx)
3519 {
3520         de::MovePtr<tcu::TestCaseGroup> group                   (new tcu::TestCaseGroup(testCtx, "opphi", "Test the OpPhi instruction"));
3521         ComputeShaderSpec                               spec1;
3522         ComputeShaderSpec                               spec2;
3523         ComputeShaderSpec                               spec3;
3524         ComputeShaderSpec                               spec4;
3525         ComputeShaderSpec                               spec5;
3526         de::Random                                              rnd                             (deStringHash(group->getName()));
3527         const int                                               numElements             = 100;
3528         vector<float>                                   inputFloats             (numElements, 0);
3529         vector<float>                                   outputFloats1   (numElements, 0);
3530         vector<float>                                   outputFloats2   (numElements, 0);
3531         vector<float>                                   outputFloats3   (numElements, 0);
3532         vector<float>                                   outputFloats4   (numElements, 0);
3533         vector<float>                                   outputFloats5   (numElements, 0);
3534         std::string                                             codestring              = "ABC";
3535         const int                                               test4Width              = 1024;
3536
3537         // Build case 5 code string. Each iteration makes the hierarchy more complicated.
3538         // 9 iterations with (7, 24) parameters makes the hierarchy 8 deep with about 1500 lines of
3539         // shader code.
3540         for (int i = 0, acc = 0; i < 9; i++)
3541                 codestring = expandOpPhiCase5(codestring, acc, 7, 24);
3542
3543         fillRandomScalars(rnd, -300.f, 300.f, &inputFloats[0], numElements);
3544
3545         // CPU might not use the same rounding mode as the GPU. Use whole numbers to avoid rounding differences.
3546         floorAll(inputFloats);
3547
3548         for (size_t ndx = 0; ndx < numElements; ++ndx)
3549         {
3550                 switch (ndx % 3)
3551                 {
3552                         case 0:         outputFloats1[ndx] = inputFloats[ndx] + 5.5f;   break;
3553                         case 1:         outputFloats1[ndx] = inputFloats[ndx] + 20.5f;  break;
3554                         case 2:         outputFloats1[ndx] = inputFloats[ndx] + 1.75f;  break;
3555                         default:        break;
3556                 }
3557                 outputFloats2[ndx] = inputFloats[ndx] + 6.5f * 3;
3558                 outputFloats3[ndx] = 8.5f - inputFloats[ndx];
3559
3560                 int index4 = (int)deFloor(deAbs((float)ndx * inputFloats[ndx]));
3561                 outputFloats4[ndx] = (float)(index4 % test4Width) * 10.0f + 5.0f;
3562
3563                 outputFloats5[ndx] = calcOpPhiCase5(inputFloats[ndx], codestring);
3564         }
3565
3566         spec1.assembly =
3567                 string(getComputeAsmShaderPreamble()) +
3568
3569                 "OpSource GLSL 430\n"
3570                 "OpName %main \"main\"\n"
3571                 "OpName %id \"gl_GlobalInvocationID\"\n"
3572
3573                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
3574
3575                 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) + string(getComputeAsmInputOutputBuffer()) +
3576
3577                 "%id = OpVariable %uvec3ptr Input\n"
3578                 "%zero       = OpConstant %i32 0\n"
3579                 "%three      = OpConstant %u32 3\n"
3580                 "%constf5p5  = OpConstant %f32 5.5\n"
3581                 "%constf20p5 = OpConstant %f32 20.5\n"
3582                 "%constf1p75 = OpConstant %f32 1.75\n"
3583                 "%constf8p5  = OpConstant %f32 8.5\n"
3584                 "%constf6p5  = OpConstant %f32 6.5\n"
3585
3586                 "%main     = OpFunction %void None %voidf\n"
3587                 "%entry    = OpLabel\n"
3588                 "%idval    = OpLoad %uvec3 %id\n"
3589                 "%x        = OpCompositeExtract %u32 %idval 0\n"
3590                 "%selector = OpUMod %u32 %x %three\n"
3591                 "            OpSelectionMerge %phi None\n"
3592                 "            OpSwitch %selector %default 0 %case0 1 %case1 2 %case2\n"
3593
3594                 // Case 1 before OpPhi.
3595                 "%case1    = OpLabel\n"
3596                 "            OpBranch %phi\n"
3597
3598                 "%default  = OpLabel\n"
3599                 "            OpUnreachable\n"
3600
3601                 "%phi      = OpLabel\n"
3602                 "%operand  = OpPhi %f32   %constf1p75 %case2   %constf20p5 %case1   %constf5p5 %case0\n" // not in the order of blocks
3603                 "%inloc    = OpAccessChain %f32ptr %indata %zero %x\n"
3604                 "%inval    = OpLoad %f32 %inloc\n"
3605                 "%add      = OpFAdd %f32 %inval %operand\n"
3606                 "%outloc   = OpAccessChain %f32ptr %outdata %zero %x\n"
3607                 "            OpStore %outloc %add\n"
3608                 "            OpReturn\n"
3609
3610                 // Case 0 after OpPhi.
3611                 "%case0    = OpLabel\n"
3612                 "            OpBranch %phi\n"
3613
3614
3615                 // Case 2 after OpPhi.
3616                 "%case2    = OpLabel\n"
3617                 "            OpBranch %phi\n"
3618
3619                 "            OpFunctionEnd\n";
3620         spec1.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
3621         spec1.outputs.push_back(BufferSp(new Float32Buffer(outputFloats1)));
3622         spec1.numWorkGroups = IVec3(numElements, 1, 1);
3623
3624         group->addChild(new SpvAsmComputeShaderCase(testCtx, "block", "out-of-order and unreachable blocks for OpPhi", spec1));
3625
3626         spec2.assembly =
3627                 string(getComputeAsmShaderPreamble()) +
3628
3629                 "OpName %main \"main\"\n"
3630                 "OpName %id \"gl_GlobalInvocationID\"\n"
3631
3632                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
3633
3634                 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) + string(getComputeAsmInputOutputBuffer()) +
3635
3636                 "%id         = OpVariable %uvec3ptr Input\n"
3637                 "%zero       = OpConstant %i32 0\n"
3638                 "%one        = OpConstant %i32 1\n"
3639                 "%three      = OpConstant %i32 3\n"
3640                 "%constf6p5  = OpConstant %f32 6.5\n"
3641
3642                 "%main       = OpFunction %void None %voidf\n"
3643                 "%entry      = OpLabel\n"
3644                 "%idval      = OpLoad %uvec3 %id\n"
3645                 "%x          = OpCompositeExtract %u32 %idval 0\n"
3646                 "%inloc      = OpAccessChain %f32ptr %indata %zero %x\n"
3647                 "%outloc     = OpAccessChain %f32ptr %outdata %zero %x\n"
3648                 "%inval      = OpLoad %f32 %inloc\n"
3649                 "              OpBranch %phi\n"
3650
3651                 "%phi        = OpLabel\n"
3652                 "%step       = OpPhi %i32 %zero  %entry %step_next  %phi\n"
3653                 "%accum      = OpPhi %f32 %inval %entry %accum_next %phi\n"
3654                 "%step_next  = OpIAdd %i32 %step %one\n"
3655                 "%accum_next = OpFAdd %f32 %accum %constf6p5\n"
3656                 "%still_loop = OpSLessThan %bool %step %three\n"
3657                 "              OpLoopMerge %exit %phi None\n"
3658                 "              OpBranchConditional %still_loop %phi %exit\n"
3659
3660                 "%exit       = OpLabel\n"
3661                 "              OpStore %outloc %accum\n"
3662                 "              OpReturn\n"
3663                 "              OpFunctionEnd\n";
3664         spec2.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
3665         spec2.outputs.push_back(BufferSp(new Float32Buffer(outputFloats2)));
3666         spec2.numWorkGroups = IVec3(numElements, 1, 1);
3667
3668         group->addChild(new SpvAsmComputeShaderCase(testCtx, "induction", "The usual way induction variables are handled in LLVM IR", spec2));
3669
3670         spec3.assembly =
3671                 string(getComputeAsmShaderPreamble()) +
3672
3673                 "OpName %main \"main\"\n"
3674                 "OpName %id \"gl_GlobalInvocationID\"\n"
3675
3676                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
3677
3678                 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) + string(getComputeAsmInputOutputBuffer()) +
3679
3680                 "%f32ptr_f   = OpTypePointer Function %f32\n"
3681                 "%id         = OpVariable %uvec3ptr Input\n"
3682                 "%true       = OpConstantTrue %bool\n"
3683                 "%false      = OpConstantFalse %bool\n"
3684                 "%zero       = OpConstant %i32 0\n"
3685                 "%constf8p5  = OpConstant %f32 8.5\n"
3686
3687                 "%main       = OpFunction %void None %voidf\n"
3688                 "%entry      = OpLabel\n"
3689                 "%b          = OpVariable %f32ptr_f Function %constf8p5\n"
3690                 "%idval      = OpLoad %uvec3 %id\n"
3691                 "%x          = OpCompositeExtract %u32 %idval 0\n"
3692                 "%inloc      = OpAccessChain %f32ptr %indata %zero %x\n"
3693                 "%outloc     = OpAccessChain %f32ptr %outdata %zero %x\n"
3694                 "%a_init     = OpLoad %f32 %inloc\n"
3695                 "%b_init     = OpLoad %f32 %b\n"
3696                 "              OpBranch %phi\n"
3697
3698                 "%phi        = OpLabel\n"
3699                 "%still_loop = OpPhi %bool %true   %entry %false  %phi\n"
3700                 "%a_next     = OpPhi %f32  %a_init %entry %b_next %phi\n"
3701                 "%b_next     = OpPhi %f32  %b_init %entry %a_next %phi\n"
3702                 "              OpLoopMerge %exit %phi None\n"
3703                 "              OpBranchConditional %still_loop %phi %exit\n"
3704
3705                 "%exit       = OpLabel\n"
3706                 "%sub        = OpFSub %f32 %a_next %b_next\n"
3707                 "              OpStore %outloc %sub\n"
3708                 "              OpReturn\n"
3709                 "              OpFunctionEnd\n";
3710         spec3.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
3711         spec3.outputs.push_back(BufferSp(new Float32Buffer(outputFloats3)));
3712         spec3.numWorkGroups = IVec3(numElements, 1, 1);
3713
3714         group->addChild(new SpvAsmComputeShaderCase(testCtx, "swap", "Swap the values of two variables using OpPhi", spec3));
3715
3716         spec4.assembly =
3717                 "OpCapability Shader\n"
3718                 "%ext = OpExtInstImport \"GLSL.std.450\"\n"
3719                 "OpMemoryModel Logical GLSL450\n"
3720                 "OpEntryPoint GLCompute %main \"main\" %id\n"
3721                 "OpExecutionMode %main LocalSize 1 1 1\n"
3722
3723                 "OpSource GLSL 430\n"
3724                 "OpName %main \"main\"\n"
3725                 "OpName %id \"gl_GlobalInvocationID\"\n"
3726
3727                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
3728
3729                 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) + string(getComputeAsmInputOutputBuffer()) +
3730
3731                 "%id       = OpVariable %uvec3ptr Input\n"
3732                 "%zero     = OpConstant %i32 0\n"
3733                 "%cimod    = OpConstant %u32 " + generateIntWidth(test4Width) + "\n"
3734
3735                 + generateConstantDefinitions(test4Width) +
3736
3737                 "%main     = OpFunction %void None %voidf\n"
3738                 "%entry    = OpLabel\n"
3739                 "%idval    = OpLoad %uvec3 %id\n"
3740                 "%x        = OpCompositeExtract %u32 %idval 0\n"
3741                 "%inloc    = OpAccessChain %f32ptr %indata %zero %x\n"
3742                 "%inval    = OpLoad %f32 %inloc\n"
3743                 "%xf       = OpConvertUToF %f32 %x\n"
3744                 "%xm       = OpFMul %f32 %xf %inval\n"
3745                 "%xa       = OpExtInst %f32 %ext FAbs %xm\n"
3746                 "%xi       = OpConvertFToU %u32 %xa\n"
3747                 "%selector = OpUMod %u32 %xi %cimod\n"
3748                 "            OpSelectionMerge %phi None\n"
3749                 "            OpSwitch %selector %default "
3750
3751                 + generateSwitchCases(test4Width) +
3752
3753                 "%default  = OpLabel\n"
3754                 "            OpUnreachable\n"
3755
3756                 + generateSwitchTargets(test4Width) +
3757
3758                 "%phi      = OpLabel\n"
3759                 "%result   = OpPhi %f32"
3760
3761                 + generateOpPhiParams(test4Width) +
3762
3763                 "%outloc   = OpAccessChain %f32ptr %outdata %zero %x\n"
3764                 "            OpStore %outloc %result\n"
3765                 "            OpReturn\n"
3766
3767                 "            OpFunctionEnd\n";
3768         spec4.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
3769         spec4.outputs.push_back(BufferSp(new Float32Buffer(outputFloats4)));
3770         spec4.numWorkGroups = IVec3(numElements, 1, 1);
3771
3772         group->addChild(new SpvAsmComputeShaderCase(testCtx, "wide", "OpPhi with a lot of parameters", spec4));
3773
3774         spec5.assembly =
3775                 "OpCapability Shader\n"
3776                 "%ext      = OpExtInstImport \"GLSL.std.450\"\n"
3777                 "OpMemoryModel Logical GLSL450\n"
3778                 "OpEntryPoint GLCompute %main \"main\" %id\n"
3779                 "OpExecutionMode %main LocalSize 1 1 1\n"
3780                 "%code     = OpString \"" + codestring + "\"\n"
3781
3782                 "OpSource GLSL 430\n"
3783                 "OpName %main \"main\"\n"
3784                 "OpName %id \"gl_GlobalInvocationID\"\n"
3785
3786                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
3787
3788                 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) + string(getComputeAsmInputOutputBuffer()) +
3789
3790                 "%id       = OpVariable %uvec3ptr Input\n"
3791                 "%zero     = OpConstant %i32 0\n"
3792                 "%f32_0    = OpConstant %f32 0.0\n"
3793                 "%f32_0_5  = OpConstant %f32 0.5\n"
3794                 "%f32_1    = OpConstant %f32 1.0\n"
3795                 "%f32_1_5  = OpConstant %f32 1.5\n"
3796                 "%f32_2    = OpConstant %f32 2.0\n"
3797                 "%f32_3_5  = OpConstant %f32 3.5\n"
3798                 "%f32_4    = OpConstant %f32 4.0\n"
3799                 "%f32_7_5  = OpConstant %f32 7.5\n"
3800                 "%f32_8    = OpConstant %f32 8.0\n"
3801                 "%f32_15_5 = OpConstant %f32 15.5\n"
3802                 "%f32_16   = OpConstant %f32 16.0\n"
3803                 "%f32_31_5 = OpConstant %f32 31.5\n"
3804                 "%f32_32   = OpConstant %f32 32.0\n"
3805                 "%f32_63_5 = OpConstant %f32 63.5\n"
3806                 "%f32_64   = OpConstant %f32 64.0\n"
3807                 "%f32_127_5 = OpConstant %f32 127.5\n"
3808                 "%f32_128  = OpConstant %f32 128.0\n"
3809                 "%f32_256  = OpConstant %f32 256.0\n"
3810
3811                 "%main     = OpFunction %void None %voidf\n"
3812                 "%entry    = OpLabel\n"
3813                 "%idval    = OpLoad %uvec3 %id\n"
3814                 "%x        = OpCompositeExtract %u32 %idval 0\n"
3815                 "%inloc    = OpAccessChain %f32ptr %indata %zero %x\n"
3816                 "%inval    = OpLoad %f32 %inloc\n"
3817
3818                 "%xabs     = OpExtInst %f32 %ext FAbs %inval\n"
3819                 "%x8       = OpFMod %f32 %xabs %f32_256\n"
3820                 "%x7       = OpFMod %f32 %xabs %f32_128\n"
3821                 "%x6       = OpFMod %f32 %xabs %f32_64\n"
3822                 "%x5       = OpFMod %f32 %xabs %f32_32\n"
3823                 "%x4       = OpFMod %f32 %xabs %f32_16\n"
3824                 "%x3       = OpFMod %f32 %xabs %f32_8\n"
3825                 "%x2       = OpFMod %f32 %xabs %f32_4\n"
3826                 "%x1       = OpFMod %f32 %xabs %f32_2\n"
3827
3828                 "%b7       = OpFOrdGreaterThanEqual %bool %x8 %f32_127_5\n"
3829                 "%b6       = OpFOrdGreaterThanEqual %bool %x7 %f32_63_5\n"
3830                 "%b5       = OpFOrdGreaterThanEqual %bool %x6 %f32_31_5\n"
3831                 "%b4       = OpFOrdGreaterThanEqual %bool %x5 %f32_15_5\n"
3832                 "%b3       = OpFOrdGreaterThanEqual %bool %x4 %f32_7_5\n"
3833                 "%b2       = OpFOrdGreaterThanEqual %bool %x3 %f32_3_5\n"
3834                 "%b1       = OpFOrdGreaterThanEqual %bool %x2 %f32_1_5\n"
3835                 "%b0       = OpFOrdGreaterThanEqual %bool %x1 %f32_0_5\n"
3836
3837                 + generateOpPhiCase5(codestring) +
3838
3839                 "%outloc   = OpAccessChain %f32ptr %outdata %zero %x\n"
3840                 "            OpStore %outloc %res\n"
3841                 "            OpReturn\n"
3842
3843                 "            OpFunctionEnd\n";
3844         spec5.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
3845         spec5.outputs.push_back(BufferSp(new Float32Buffer(outputFloats5)));
3846         spec5.numWorkGroups = IVec3(numElements, 1, 1);
3847
3848         group->addChild(new SpvAsmComputeShaderCase(testCtx, "nested", "Stress OpPhi with a lot of nesting", spec5));
3849
3850         createOpPhiVartypeTests(group, testCtx);
3851
3852         return group.release();
3853 }
3854
3855 // Assembly code used for testing block order is based on GLSL source code:
3856 //
3857 // #version 430
3858 //
3859 // layout(std140, set = 0, binding = 0) readonly buffer Input {
3860 //   float elements[];
3861 // } input_data;
3862 // layout(std140, set = 0, binding = 1) writeonly buffer Output {
3863 //   float elements[];
3864 // } output_data;
3865 //
3866 // void main() {
3867 //   uint x = gl_GlobalInvocationID.x;
3868 //   output_data.elements[x] = input_data.elements[x];
3869 //   if (x > uint(50)) {
3870 //     switch (x % uint(3)) {
3871 //       case 0: output_data.elements[x] += 1.5f; break;
3872 //       case 1: output_data.elements[x] += 42.f; break;
3873 //       case 2: output_data.elements[x] -= 27.f; break;
3874 //       default: break;
3875 //     }
3876 //   } else {
3877 //     output_data.elements[x] = -input_data.elements[x];
3878 //   }
3879 // }
3880 tcu::TestCaseGroup* createBlockOrderGroup (tcu::TestContext& testCtx)
3881 {
3882         de::MovePtr<tcu::TestCaseGroup> group                   (new tcu::TestCaseGroup(testCtx, "block_order", "Test block orders"));
3883         ComputeShaderSpec                               spec;
3884         de::Random                                              rnd                             (deStringHash(group->getName()));
3885         const int                                               numElements             = 100;
3886         vector<float>                                   inputFloats             (numElements, 0);
3887         vector<float>                                   outputFloats    (numElements, 0);
3888
3889         fillRandomScalars(rnd, -100.f, 100.f, &inputFloats[0], numElements);
3890
3891         // CPU might not use the same rounding mode as the GPU. Use whole numbers to avoid rounding differences.
3892         floorAll(inputFloats);
3893
3894         for (size_t ndx = 0; ndx <= 50; ++ndx)
3895                 outputFloats[ndx] = -inputFloats[ndx];
3896
3897         for (size_t ndx = 51; ndx < numElements; ++ndx)
3898         {
3899                 switch (ndx % 3)
3900                 {
3901                         case 0:         outputFloats[ndx] = inputFloats[ndx] + 1.5f; break;
3902                         case 1:         outputFloats[ndx] = inputFloats[ndx] + 42.f; break;
3903                         case 2:         outputFloats[ndx] = inputFloats[ndx] - 27.f; break;
3904                         default:        break;
3905                 }
3906         }
3907
3908         spec.assembly =
3909                 string(getComputeAsmShaderPreamble()) +
3910
3911                 "OpSource GLSL 430\n"
3912                 "OpName %main \"main\"\n"
3913                 "OpName %id \"gl_GlobalInvocationID\"\n"
3914
3915                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
3916
3917                 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) +
3918
3919                 "%u32ptr       = OpTypePointer Function %u32\n"
3920                 "%u32ptr_input = OpTypePointer Input %u32\n"
3921
3922                 + string(getComputeAsmInputOutputBuffer()) +
3923
3924                 "%id        = OpVariable %uvec3ptr Input\n"
3925                 "%zero      = OpConstant %i32 0\n"
3926                 "%const3    = OpConstant %u32 3\n"
3927                 "%const50   = OpConstant %u32 50\n"
3928                 "%constf1p5 = OpConstant %f32 1.5\n"
3929                 "%constf27  = OpConstant %f32 27.0\n"
3930                 "%constf42  = OpConstant %f32 42.0\n"
3931
3932                 "%main = OpFunction %void None %voidf\n"
3933
3934                 // entry block.
3935                 "%entry    = OpLabel\n"
3936
3937                 // Create a temporary variable to hold the value of gl_GlobalInvocationID.x.
3938                 "%xvar     = OpVariable %u32ptr Function\n"
3939                 "%xptr     = OpAccessChain %u32ptr_input %id %zero\n"
3940                 "%x        = OpLoad %u32 %xptr\n"
3941                 "            OpStore %xvar %x\n"
3942
3943                 "%cmp      = OpUGreaterThan %bool %x %const50\n"
3944                 "            OpSelectionMerge %if_merge None\n"
3945                 "            OpBranchConditional %cmp %if_true %if_false\n"
3946
3947                 // False branch for if-statement: placed in the middle of switch cases and before true branch.
3948                 "%if_false = OpLabel\n"
3949                 "%x_f      = OpLoad %u32 %xvar\n"
3950                 "%inloc_f  = OpAccessChain %f32ptr %indata %zero %x_f\n"
3951                 "%inval_f  = OpLoad %f32 %inloc_f\n"
3952                 "%negate   = OpFNegate %f32 %inval_f\n"
3953                 "%outloc_f = OpAccessChain %f32ptr %outdata %zero %x_f\n"
3954                 "            OpStore %outloc_f %negate\n"
3955                 "            OpBranch %if_merge\n"
3956
3957                 // Merge block for if-statement: placed in the middle of true and false branch.
3958                 "%if_merge = OpLabel\n"
3959                 "            OpReturn\n"
3960
3961                 // True branch for if-statement: placed in the middle of swtich cases and after the false branch.
3962                 "%if_true  = OpLabel\n"
3963                 "%xval_t   = OpLoad %u32 %xvar\n"
3964                 "%mod      = OpUMod %u32 %xval_t %const3\n"
3965                 "            OpSelectionMerge %switch_merge None\n"
3966                 "            OpSwitch %mod %default 0 %case0 1 %case1 2 %case2\n"
3967
3968                 // Merge block for switch-statement: placed before the case
3969                 // bodies.  But it must follow OpSwitch which dominates it.
3970                 "%switch_merge = OpLabel\n"
3971                 "                OpBranch %if_merge\n"
3972
3973                 // Case 1 for switch-statement: placed before case 0.
3974                 // It must follow the OpSwitch that dominates it.
3975                 "%case1    = OpLabel\n"
3976                 "%x_1      = OpLoad %u32 %xvar\n"
3977                 "%inloc_1  = OpAccessChain %f32ptr %indata %zero %x_1\n"
3978                 "%inval_1  = OpLoad %f32 %inloc_1\n"
3979                 "%addf42   = OpFAdd %f32 %inval_1 %constf42\n"
3980                 "%outloc_1 = OpAccessChain %f32ptr %outdata %zero %x_1\n"
3981                 "            OpStore %outloc_1 %addf42\n"
3982                 "            OpBranch %switch_merge\n"
3983
3984                 // Case 2 for switch-statement.
3985                 "%case2    = OpLabel\n"
3986                 "%x_2      = OpLoad %u32 %xvar\n"
3987                 "%inloc_2  = OpAccessChain %f32ptr %indata %zero %x_2\n"
3988                 "%inval_2  = OpLoad %f32 %inloc_2\n"
3989                 "%subf27   = OpFSub %f32 %inval_2 %constf27\n"
3990                 "%outloc_2 = OpAccessChain %f32ptr %outdata %zero %x_2\n"
3991                 "            OpStore %outloc_2 %subf27\n"
3992                 "            OpBranch %switch_merge\n"
3993
3994                 // Default case for switch-statement: placed in the middle of normal cases.
3995                 "%default = OpLabel\n"
3996                 "           OpBranch %switch_merge\n"
3997
3998                 // Case 0 for switch-statement: out of order.
3999                 "%case0    = OpLabel\n"
4000                 "%x_0      = OpLoad %u32 %xvar\n"
4001                 "%inloc_0  = OpAccessChain %f32ptr %indata %zero %x_0\n"
4002                 "%inval_0  = OpLoad %f32 %inloc_0\n"
4003                 "%addf1p5  = OpFAdd %f32 %inval_0 %constf1p5\n"
4004                 "%outloc_0 = OpAccessChain %f32ptr %outdata %zero %x_0\n"
4005                 "            OpStore %outloc_0 %addf1p5\n"
4006                 "            OpBranch %switch_merge\n"
4007
4008                 "            OpFunctionEnd\n";
4009         spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
4010         spec.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
4011         spec.numWorkGroups = IVec3(numElements, 1, 1);
4012
4013         group->addChild(new SpvAsmComputeShaderCase(testCtx, "all", "various out-of-order blocks", spec));
4014
4015         return group.release();
4016 }
4017
4018 tcu::TestCaseGroup* createMultipleShaderGroup (tcu::TestContext& testCtx)
4019 {
4020         de::MovePtr<tcu::TestCaseGroup> group                   (new tcu::TestCaseGroup(testCtx, "multiple_shaders", "Test multiple shaders in the same module"));
4021         ComputeShaderSpec                               spec1;
4022         ComputeShaderSpec                               spec2;
4023         de::Random                                              rnd                             (deStringHash(group->getName()));
4024         const int                                               numElements             = 100;
4025         vector<float>                                   inputFloats             (numElements, 0);
4026         vector<float>                                   outputFloats1   (numElements, 0);
4027         vector<float>                                   outputFloats2   (numElements, 0);
4028         fillRandomScalars(rnd, -500.f, 500.f, &inputFloats[0], numElements);
4029
4030         for (size_t ndx = 0; ndx < numElements; ++ndx)
4031         {
4032                 outputFloats1[ndx] = inputFloats[ndx] + inputFloats[ndx];
4033                 outputFloats2[ndx] = -inputFloats[ndx];
4034         }
4035
4036         const string assembly(
4037                 "OpCapability Shader\n"
4038                 "OpMemoryModel Logical GLSL450\n"
4039                 "OpEntryPoint GLCompute %comp_main1 \"entrypoint1\" %id\n"
4040                 "OpEntryPoint GLCompute %comp_main2 \"entrypoint2\" %id\n"
4041                 // A module cannot have two OpEntryPoint instructions with the same Execution Model and the same Name string.
4042                 "OpEntryPoint Vertex    %vert_main  \"entrypoint2\" %vert_builtins %vertexIndex %instanceIndex\n"
4043                 "OpExecutionMode %comp_main1 LocalSize 1 1 1\n"
4044                 "OpExecutionMode %comp_main2 LocalSize 1 1 1\n"
4045
4046                 "OpName %comp_main1              \"entrypoint1\"\n"
4047                 "OpName %comp_main2              \"entrypoint2\"\n"
4048                 "OpName %vert_main               \"entrypoint2\"\n"
4049                 "OpName %id                      \"gl_GlobalInvocationID\"\n"
4050                 "OpName %vert_builtin_st         \"gl_PerVertex\"\n"
4051                 "OpName %vertexIndex             \"gl_VertexIndex\"\n"
4052                 "OpName %instanceIndex           \"gl_InstanceIndex\"\n"
4053                 "OpMemberName %vert_builtin_st 0 \"gl_Position\"\n"
4054                 "OpMemberName %vert_builtin_st 1 \"gl_PointSize\"\n"
4055                 "OpMemberName %vert_builtin_st 2 \"gl_ClipDistance\"\n"
4056
4057                 "OpDecorate %id                      BuiltIn GlobalInvocationId\n"
4058                 "OpDecorate %vertexIndex             BuiltIn VertexIndex\n"
4059                 "OpDecorate %instanceIndex           BuiltIn InstanceIndex\n"
4060                 "OpDecorate %vert_builtin_st         Block\n"
4061                 "OpMemberDecorate %vert_builtin_st 0 BuiltIn Position\n"
4062                 "OpMemberDecorate %vert_builtin_st 1 BuiltIn PointSize\n"
4063                 "OpMemberDecorate %vert_builtin_st 2 BuiltIn ClipDistance\n"
4064
4065                 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) + string(getComputeAsmInputOutputBuffer()) +
4066
4067                 "%zero       = OpConstant %i32 0\n"
4068                 "%one        = OpConstant %u32 1\n"
4069                 "%c_f32_1    = OpConstant %f32 1\n"
4070
4071                 "%i32inputptr         = OpTypePointer Input %i32\n"
4072                 "%vec4                = OpTypeVector %f32 4\n"
4073                 "%vec4ptr             = OpTypePointer Output %vec4\n"
4074                 "%f32arr1             = OpTypeArray %f32 %one\n"
4075                 "%vert_builtin_st     = OpTypeStruct %vec4 %f32 %f32arr1\n"
4076                 "%vert_builtin_st_ptr = OpTypePointer Output %vert_builtin_st\n"
4077                 "%vert_builtins       = OpVariable %vert_builtin_st_ptr Output\n"
4078
4079                 "%id         = OpVariable %uvec3ptr Input\n"
4080                 "%vertexIndex = OpVariable %i32inputptr Input\n"
4081                 "%instanceIndex = OpVariable %i32inputptr Input\n"
4082                 "%c_vec4_1   = OpConstantComposite %vec4 %c_f32_1 %c_f32_1 %c_f32_1 %c_f32_1\n"
4083
4084                 // gl_Position = vec4(1.);
4085                 "%vert_main  = OpFunction %void None %voidf\n"
4086                 "%vert_entry = OpLabel\n"
4087                 "%position   = OpAccessChain %vec4ptr %vert_builtins %zero\n"
4088                 "              OpStore %position %c_vec4_1\n"
4089                 "              OpReturn\n"
4090                 "              OpFunctionEnd\n"
4091
4092                 // Double inputs.
4093                 "%comp_main1  = OpFunction %void None %voidf\n"
4094                 "%comp1_entry = OpLabel\n"
4095                 "%idval1      = OpLoad %uvec3 %id\n"
4096                 "%x1          = OpCompositeExtract %u32 %idval1 0\n"
4097                 "%inloc1      = OpAccessChain %f32ptr %indata %zero %x1\n"
4098                 "%inval1      = OpLoad %f32 %inloc1\n"
4099                 "%add         = OpFAdd %f32 %inval1 %inval1\n"
4100                 "%outloc1     = OpAccessChain %f32ptr %outdata %zero %x1\n"
4101                 "               OpStore %outloc1 %add\n"
4102                 "               OpReturn\n"
4103                 "               OpFunctionEnd\n"
4104
4105                 // Negate inputs.
4106                 "%comp_main2  = OpFunction %void None %voidf\n"
4107                 "%comp2_entry = OpLabel\n"
4108                 "%idval2      = OpLoad %uvec3 %id\n"
4109                 "%x2          = OpCompositeExtract %u32 %idval2 0\n"
4110                 "%inloc2      = OpAccessChain %f32ptr %indata %zero %x2\n"
4111                 "%inval2      = OpLoad %f32 %inloc2\n"
4112                 "%neg         = OpFNegate %f32 %inval2\n"
4113                 "%outloc2     = OpAccessChain %f32ptr %outdata %zero %x2\n"
4114                 "               OpStore %outloc2 %neg\n"
4115                 "               OpReturn\n"
4116                 "               OpFunctionEnd\n");
4117
4118         spec1.assembly = assembly;
4119         spec1.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
4120         spec1.outputs.push_back(BufferSp(new Float32Buffer(outputFloats1)));
4121         spec1.numWorkGroups = IVec3(numElements, 1, 1);
4122         spec1.entryPoint = "entrypoint1";
4123
4124         spec2.assembly = assembly;
4125         spec2.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
4126         spec2.outputs.push_back(BufferSp(new Float32Buffer(outputFloats2)));
4127         spec2.numWorkGroups = IVec3(numElements, 1, 1);
4128         spec2.entryPoint = "entrypoint2";
4129
4130         group->addChild(new SpvAsmComputeShaderCase(testCtx, "shader1", "multiple shaders in the same module", spec1));
4131         group->addChild(new SpvAsmComputeShaderCase(testCtx, "shader2", "multiple shaders in the same module", spec2));
4132
4133         return group.release();
4134 }
4135
4136 inline std::string makeLongUTF8String (size_t num4ByteChars)
4137 {
4138         // An example of a longest valid UTF-8 character.  Be explicit about the
4139         // character type because Microsoft compilers can otherwise interpret the
4140         // character string as being over wide (16-bit) characters. Ideally, we
4141         // would just use a C++11 UTF-8 string literal, but we want to support older
4142         // Microsoft compilers.
4143         const std::basic_string<char> earthAfrica("\xF0\x9F\x8C\x8D");
4144         std::string longString;
4145         longString.reserve(num4ByteChars * 4);
4146         for (size_t count = 0; count < num4ByteChars; count++)
4147         {
4148                 longString += earthAfrica;
4149         }
4150         return longString;
4151 }
4152
4153 tcu::TestCaseGroup* createOpSourceGroup (tcu::TestContext& testCtx)
4154 {
4155         de::MovePtr<tcu::TestCaseGroup> group                   (new tcu::TestCaseGroup(testCtx, "opsource", "Tests the OpSource & OpSourceContinued instruction"));
4156         vector<CaseParameter>                   cases;
4157         de::Random                                              rnd                             (deStringHash(group->getName()));
4158         const int                                               numElements             = 100;
4159         vector<float>                                   positiveFloats  (numElements, 0);
4160         vector<float>                                   negativeFloats  (numElements, 0);
4161         const StringTemplate                    shaderTemplate  (
4162                 "OpCapability Shader\n"
4163                 "OpMemoryModel Logical GLSL450\n"
4164
4165                 "OpEntryPoint GLCompute %main \"main\" %id\n"
4166                 "OpExecutionMode %main LocalSize 1 1 1\n"
4167
4168                 "${SOURCE}\n"
4169
4170                 "OpName %main           \"main\"\n"
4171                 "OpName %id             \"gl_GlobalInvocationID\"\n"
4172
4173                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
4174
4175                 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) + string(getComputeAsmInputOutputBuffer()) +
4176
4177                 "%id        = OpVariable %uvec3ptr Input\n"
4178                 "%zero      = OpConstant %i32 0\n"
4179
4180                 "%main      = OpFunction %void None %voidf\n"
4181                 "%label     = OpLabel\n"
4182                 "%idval     = OpLoad %uvec3 %id\n"
4183                 "%x         = OpCompositeExtract %u32 %idval 0\n"
4184                 "%inloc     = OpAccessChain %f32ptr %indata %zero %x\n"
4185                 "%inval     = OpLoad %f32 %inloc\n"
4186                 "%neg       = OpFNegate %f32 %inval\n"
4187                 "%outloc    = OpAccessChain %f32ptr %outdata %zero %x\n"
4188                 "             OpStore %outloc %neg\n"
4189                 "             OpReturn\n"
4190                 "             OpFunctionEnd\n");
4191
4192         cases.push_back(CaseParameter("unknown_source",                                                 "OpSource Unknown 0"));
4193         cases.push_back(CaseParameter("wrong_source",                                                   "OpSource OpenCL_C 210"));
4194         cases.push_back(CaseParameter("normal_filename",                                                "%fname = OpString \"filename\"\n"
4195                                                                                                                                                         "OpSource GLSL 430 %fname"));
4196         cases.push_back(CaseParameter("empty_filename",                                                 "%fname = OpString \"\"\n"
4197                                                                                                                                                         "OpSource GLSL 430 %fname"));
4198         cases.push_back(CaseParameter("normal_source_code",                                             "%fname = OpString \"filename\"\n"
4199                                                                                                                                                         "OpSource GLSL 430 %fname \"#version 430\nvoid main() {}\""));
4200         cases.push_back(CaseParameter("empty_source_code",                                              "%fname = OpString \"filename\"\n"
4201                                                                                                                                                         "OpSource GLSL 430 %fname \"\""));
4202         cases.push_back(CaseParameter("long_source_code",                                               "%fname = OpString \"filename\"\n"
4203                                                                                                                                                         "OpSource GLSL 430 %fname \"" + makeLongUTF8String(65530) + "ccc\"")); // word count: 65535
4204         cases.push_back(CaseParameter("utf8_source_code",                                               "%fname = OpString \"filename\"\n"
4205                                                                                                                                                         "OpSource GLSL 430 %fname \"\xE2\x98\x82\xE2\x98\x85\"")); // umbrella & black star symbol
4206         cases.push_back(CaseParameter("normal_sourcecontinued",                                 "%fname = OpString \"filename\"\n"
4207                                                                                                                                                         "OpSource GLSL 430 %fname \"#version 430\nvo\"\n"
4208                                                                                                                                                         "OpSourceContinued \"id main() {}\""));
4209         cases.push_back(CaseParameter("empty_sourcecontinued",                                  "%fname = OpString \"filename\"\n"
4210                                                                                                                                                         "OpSource GLSL 430 %fname \"#version 430\nvoid main() {}\"\n"
4211                                                                                                                                                         "OpSourceContinued \"\""));
4212         cases.push_back(CaseParameter("long_sourcecontinued",                                   "%fname = OpString \"filename\"\n"
4213                                                                                                                                                         "OpSource GLSL 430 %fname \"#version 430\nvoid main() {}\"\n"
4214                                                                                                                                                         "OpSourceContinued \"" + makeLongUTF8String(65533) + "ccc\"")); // word count: 65535
4215         cases.push_back(CaseParameter("utf8_sourcecontinued",                                   "%fname = OpString \"filename\"\n"
4216                                                                                                                                                         "OpSource GLSL 430 %fname \"#version 430\nvoid main() {}\"\n"
4217                                                                                                                                                         "OpSourceContinued \"\xE2\x98\x8E\xE2\x9A\x91\"")); // white telephone & black flag symbol
4218         cases.push_back(CaseParameter("multi_sourcecontinued",                                  "%fname = OpString \"filename\"\n"
4219                                                                                                                                                         "OpSource GLSL 430 %fname \"#version 430\n\"\n"
4220                                                                                                                                                         "OpSourceContinued \"void\"\n"
4221                                                                                                                                                         "OpSourceContinued \"main()\"\n"
4222                                                                                                                                                         "OpSourceContinued \"{}\""));
4223         cases.push_back(CaseParameter("empty_source_before_sourcecontinued",    "%fname = OpString \"filename\"\n"
4224                                                                                                                                                         "OpSource GLSL 430 %fname \"\"\n"
4225                                                                                                                                                         "OpSourceContinued \"#version 430\nvoid main() {}\""));
4226
4227         fillRandomScalars(rnd, 1.f, 100.f, &positiveFloats[0], numElements);
4228
4229         for (size_t ndx = 0; ndx < numElements; ++ndx)
4230                 negativeFloats[ndx] = -positiveFloats[ndx];
4231
4232         for (size_t caseNdx = 0; caseNdx < cases.size(); ++caseNdx)
4233         {
4234                 map<string, string>             specializations;
4235                 ComputeShaderSpec               spec;
4236
4237                 specializations["SOURCE"] = cases[caseNdx].param;
4238                 spec.assembly = shaderTemplate.specialize(specializations);
4239                 spec.inputs.push_back(BufferSp(new Float32Buffer(positiveFloats)));
4240                 spec.outputs.push_back(BufferSp(new Float32Buffer(negativeFloats)));
4241                 spec.numWorkGroups = IVec3(numElements, 1, 1);
4242
4243                 group->addChild(new SpvAsmComputeShaderCase(testCtx, cases[caseNdx].name, cases[caseNdx].name, spec));
4244         }
4245
4246         return group.release();
4247 }
4248
4249 tcu::TestCaseGroup* createOpSourceExtensionGroup (tcu::TestContext& testCtx)
4250 {
4251         de::MovePtr<tcu::TestCaseGroup> group                   (new tcu::TestCaseGroup(testCtx, "opsourceextension", "Tests the OpSource instruction"));
4252         vector<CaseParameter>                   cases;
4253         de::Random                                              rnd                             (deStringHash(group->getName()));
4254         const int                                               numElements             = 100;
4255         vector<float>                                   inputFloats             (numElements, 0);
4256         vector<float>                                   outputFloats    (numElements, 0);
4257         const StringTemplate                    shaderTemplate  (
4258                 string(getComputeAsmShaderPreamble()) +
4259
4260                 "OpSourceExtension \"${EXTENSION}\"\n"
4261
4262                 "OpName %main           \"main\"\n"
4263                 "OpName %id             \"gl_GlobalInvocationID\"\n"
4264
4265                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
4266
4267                 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) + string(getComputeAsmInputOutputBuffer()) +
4268
4269                 "%id        = OpVariable %uvec3ptr Input\n"
4270                 "%zero      = OpConstant %i32 0\n"
4271
4272                 "%main      = OpFunction %void None %voidf\n"
4273                 "%label     = OpLabel\n"
4274                 "%idval     = OpLoad %uvec3 %id\n"
4275                 "%x         = OpCompositeExtract %u32 %idval 0\n"
4276                 "%inloc     = OpAccessChain %f32ptr %indata %zero %x\n"
4277                 "%inval     = OpLoad %f32 %inloc\n"
4278                 "%neg       = OpFNegate %f32 %inval\n"
4279                 "%outloc    = OpAccessChain %f32ptr %outdata %zero %x\n"
4280                 "             OpStore %outloc %neg\n"
4281                 "             OpReturn\n"
4282                 "             OpFunctionEnd\n");
4283
4284         cases.push_back(CaseParameter("empty_extension",        ""));
4285         cases.push_back(CaseParameter("real_extension",         "GL_ARB_texture_rectangle"));
4286         cases.push_back(CaseParameter("fake_extension",         "GL_ARB_im_the_ultimate_extension"));
4287         cases.push_back(CaseParameter("utf8_extension",         "GL_ARB_\xE2\x98\x82\xE2\x98\x85"));
4288         cases.push_back(CaseParameter("long_extension",         makeLongUTF8String(65533) + "ccc")); // word count: 65535
4289
4290         fillRandomScalars(rnd, -200.f, 200.f, &inputFloats[0], numElements);
4291
4292         for (size_t ndx = 0; ndx < numElements; ++ndx)
4293                 outputFloats[ndx] = -inputFloats[ndx];
4294
4295         for (size_t caseNdx = 0; caseNdx < cases.size(); ++caseNdx)
4296         {
4297                 map<string, string>             specializations;
4298                 ComputeShaderSpec               spec;
4299
4300                 specializations["EXTENSION"] = cases[caseNdx].param;
4301                 spec.assembly = shaderTemplate.specialize(specializations);
4302                 spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
4303                 spec.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
4304                 spec.numWorkGroups = IVec3(numElements, 1, 1);
4305
4306                 group->addChild(new SpvAsmComputeShaderCase(testCtx, cases[caseNdx].name, cases[caseNdx].name, spec));
4307         }
4308
4309         return group.release();
4310 }
4311
4312 // Checks that a compute shader can generate a constant null value of various types, without exercising a computation on it.
4313 tcu::TestCaseGroup* createOpConstantNullGroup (tcu::TestContext& testCtx)
4314 {
4315         de::MovePtr<tcu::TestCaseGroup> group                   (new tcu::TestCaseGroup(testCtx, "opconstantnull", "Tests the OpConstantNull instruction"));
4316         vector<CaseParameter>                   cases;
4317         de::Random                                              rnd                             (deStringHash(group->getName()));
4318         const int                                               numElements             = 100;
4319         vector<float>                                   positiveFloats  (numElements, 0);
4320         vector<float>                                   negativeFloats  (numElements, 0);
4321         const StringTemplate                    shaderTemplate  (
4322                 string(getComputeAsmShaderPreamble()) +
4323
4324                 "OpSource GLSL 430\n"
4325                 "OpName %main           \"main\"\n"
4326                 "OpName %id             \"gl_GlobalInvocationID\"\n"
4327
4328                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
4329
4330                 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) +
4331                 "%uvec2     = OpTypeVector %u32 2\n"
4332                 "%bvec3     = OpTypeVector %bool 3\n"
4333                 "%fvec4     = OpTypeVector %f32 4\n"
4334                 "%fmat33    = OpTypeMatrix %fvec3 3\n"
4335                 "%const100  = OpConstant %u32 100\n"
4336                 "%uarr100   = OpTypeArray %i32 %const100\n"
4337                 "%struct    = OpTypeStruct %f32 %i32 %u32\n"
4338                 "%pointer   = OpTypePointer Function %i32\n"
4339                 + string(getComputeAsmInputOutputBuffer()) +
4340
4341                 "%null      = OpConstantNull ${TYPE}\n"
4342
4343                 "%id        = OpVariable %uvec3ptr Input\n"
4344                 "%zero      = OpConstant %i32 0\n"
4345
4346                 "%main      = OpFunction %void None %voidf\n"
4347                 "%label     = OpLabel\n"
4348                 "%idval     = OpLoad %uvec3 %id\n"
4349                 "%x         = OpCompositeExtract %u32 %idval 0\n"
4350                 "%inloc     = OpAccessChain %f32ptr %indata %zero %x\n"
4351                 "%inval     = OpLoad %f32 %inloc\n"
4352                 "%neg       = OpFNegate %f32 %inval\n"
4353                 "%outloc    = OpAccessChain %f32ptr %outdata %zero %x\n"
4354                 "             OpStore %outloc %neg\n"
4355                 "             OpReturn\n"
4356                 "             OpFunctionEnd\n");
4357
4358         cases.push_back(CaseParameter("bool",                   "%bool"));
4359         cases.push_back(CaseParameter("sint32",                 "%i32"));
4360         cases.push_back(CaseParameter("uint32",                 "%u32"));
4361         cases.push_back(CaseParameter("float32",                "%f32"));
4362         cases.push_back(CaseParameter("vec4float32",    "%fvec4"));
4363         cases.push_back(CaseParameter("vec3bool",               "%bvec3"));
4364         cases.push_back(CaseParameter("vec2uint32",             "%uvec2"));
4365         cases.push_back(CaseParameter("matrix",                 "%fmat33"));
4366         cases.push_back(CaseParameter("array",                  "%uarr100"));
4367         cases.push_back(CaseParameter("struct",                 "%struct"));
4368         cases.push_back(CaseParameter("pointer",                "%pointer"));
4369
4370         fillRandomScalars(rnd, 1.f, 100.f, &positiveFloats[0], numElements);
4371
4372         for (size_t ndx = 0; ndx < numElements; ++ndx)
4373                 negativeFloats[ndx] = -positiveFloats[ndx];
4374
4375         for (size_t caseNdx = 0; caseNdx < cases.size(); ++caseNdx)
4376         {
4377                 map<string, string>             specializations;
4378                 ComputeShaderSpec               spec;
4379
4380                 specializations["TYPE"] = cases[caseNdx].param;
4381                 spec.assembly = shaderTemplate.specialize(specializations);
4382                 spec.inputs.push_back(BufferSp(new Float32Buffer(positiveFloats)));
4383                 spec.outputs.push_back(BufferSp(new Float32Buffer(negativeFloats)));
4384                 spec.numWorkGroups = IVec3(numElements, 1, 1);
4385
4386                 group->addChild(new SpvAsmComputeShaderCase(testCtx, cases[caseNdx].name, cases[caseNdx].name, spec));
4387         }
4388
4389         return group.release();
4390 }
4391
4392 // Checks that a compute shader can generate a constant composite value of various types, without exercising a computation on it.
4393 tcu::TestCaseGroup* createOpConstantCompositeGroup (tcu::TestContext& testCtx)
4394 {
4395         de::MovePtr<tcu::TestCaseGroup> group                   (new tcu::TestCaseGroup(testCtx, "opconstantcomposite", "Tests the OpConstantComposite instruction"));
4396         vector<CaseParameter>                   cases;
4397         de::Random                                              rnd                             (deStringHash(group->getName()));
4398         const int                                               numElements             = 100;
4399         vector<float>                                   positiveFloats  (numElements, 0);
4400         vector<float>                                   negativeFloats  (numElements, 0);
4401         const StringTemplate                    shaderTemplate  (
4402                 string(getComputeAsmShaderPreamble()) +
4403
4404                 "OpSource GLSL 430\n"
4405                 "OpName %main           \"main\"\n"
4406                 "OpName %id             \"gl_GlobalInvocationID\"\n"
4407
4408                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
4409
4410                 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) + string(getComputeAsmInputOutputBuffer()) +
4411
4412                 "%id        = OpVariable %uvec3ptr Input\n"
4413                 "%zero      = OpConstant %i32 0\n"
4414
4415                 "${CONSTANT}\n"
4416
4417                 "%main      = OpFunction %void None %voidf\n"
4418                 "%label     = OpLabel\n"
4419                 "%idval     = OpLoad %uvec3 %id\n"
4420                 "%x         = OpCompositeExtract %u32 %idval 0\n"
4421                 "%inloc     = OpAccessChain %f32ptr %indata %zero %x\n"
4422                 "%inval     = OpLoad %f32 %inloc\n"
4423                 "%neg       = OpFNegate %f32 %inval\n"
4424                 "%outloc    = OpAccessChain %f32ptr %outdata %zero %x\n"
4425                 "             OpStore %outloc %neg\n"
4426                 "             OpReturn\n"
4427                 "             OpFunctionEnd\n");
4428
4429         cases.push_back(CaseParameter("vector",                 "%five = OpConstant %u32 5\n"
4430                                                                                                         "%const = OpConstantComposite %uvec3 %five %zero %five"));
4431         cases.push_back(CaseParameter("matrix",                 "%m3fvec3 = OpTypeMatrix %fvec3 3\n"
4432                                                                                                         "%ten = OpConstant %f32 10.\n"
4433                                                                                                         "%fzero = OpConstant %f32 0.\n"
4434                                                                                                         "%vec = OpConstantComposite %fvec3 %ten %fzero %ten\n"
4435                                                                                                         "%mat = OpConstantComposite %m3fvec3 %vec %vec %vec"));
4436         cases.push_back(CaseParameter("struct",                 "%m2vec3 = OpTypeMatrix %fvec3 2\n"
4437                                                                                                         "%struct = OpTypeStruct %i32 %f32 %fvec3 %m2vec3\n"
4438                                                                                                         "%fzero = OpConstant %f32 0.\n"
4439                                                                                                         "%one = OpConstant %f32 1.\n"
4440                                                                                                         "%point5 = OpConstant %f32 0.5\n"
4441                                                                                                         "%vec = OpConstantComposite %fvec3 %one %one %fzero\n"
4442                                                                                                         "%mat = OpConstantComposite %m2vec3 %vec %vec\n"
4443                                                                                                         "%const = OpConstantComposite %struct %zero %point5 %vec %mat"));
4444         cases.push_back(CaseParameter("nested_struct",  "%st1 = OpTypeStruct %u32 %f32\n"
4445                                                                                                         "%st2 = OpTypeStruct %i32 %i32\n"
4446                                                                                                         "%struct = OpTypeStruct %st1 %st2\n"
4447                                                                                                         "%point5 = OpConstant %f32 0.5\n"
4448                                                                                                         "%one = OpConstant %u32 1\n"
4449                                                                                                         "%ten = OpConstant %i32 10\n"
4450                                                                                                         "%st1val = OpConstantComposite %st1 %one %point5\n"
4451                                                                                                         "%st2val = OpConstantComposite %st2 %ten %ten\n"
4452                                                                                                         "%const = OpConstantComposite %struct %st1val %st2val"));
4453
4454         fillRandomScalars(rnd, 1.f, 100.f, &positiveFloats[0], numElements);
4455
4456         for (size_t ndx = 0; ndx < numElements; ++ndx)
4457                 negativeFloats[ndx] = -positiveFloats[ndx];
4458
4459         for (size_t caseNdx = 0; caseNdx < cases.size(); ++caseNdx)
4460         {
4461                 map<string, string>             specializations;
4462                 ComputeShaderSpec               spec;
4463
4464                 specializations["CONSTANT"] = cases[caseNdx].param;
4465                 spec.assembly = shaderTemplate.specialize(specializations);
4466                 spec.inputs.push_back(BufferSp(new Float32Buffer(positiveFloats)));
4467                 spec.outputs.push_back(BufferSp(new Float32Buffer(negativeFloats)));
4468                 spec.numWorkGroups = IVec3(numElements, 1, 1);
4469
4470                 group->addChild(new SpvAsmComputeShaderCase(testCtx, cases[caseNdx].name, cases[caseNdx].name, spec));
4471         }
4472
4473         return group.release();
4474 }
4475
4476 // Creates a floating point number with the given exponent, and significand
4477 // bits set. It can only create normalized numbers. Only the least significant
4478 // 24 bits of the significand will be examined. The final bit of the
4479 // significand will also be ignored. This allows alignment to be written
4480 // similarly to C99 hex-floats.
4481 // For example if you wanted to write 0x1.7f34p-12 you would call
4482 // constructNormalizedFloat(-12, 0x7f3400)
4483 float constructNormalizedFloat (deInt32 exponent, deUint32 significand)
4484 {
4485         float f = 1.0f;
4486
4487         for (deInt32 idx = 0; idx < 23; ++idx)
4488         {
4489                 f += ((significand & 0x800000) == 0) ? 0.f : std::ldexp(1.0f, -(idx + 1));
4490                 significand <<= 1;
4491         }
4492
4493         return std::ldexp(f, exponent);
4494 }
4495
4496 // Compare instruction for the OpQuantizeF16 compute exact case.
4497 // Returns true if the output is what is expected from the test case.
4498 bool compareOpQuantizeF16ComputeExactCase (const std::vector<Resource>&, const vector<AllocationSp>& outputAllocs, const std::vector<Resource>& expectedOutputs, TestLog&)
4499 {
4500         if (outputAllocs.size() != 1)
4501                 return false;
4502
4503         // Only size is needed because we cannot compare Nans.
4504         size_t byteSize = expectedOutputs[0].getByteSize();
4505
4506         const float*    outputAsFloat   = static_cast<const float*>(outputAllocs[0]->getHostPtr());
4507
4508         if (byteSize != 4*sizeof(float)) {
4509                 return false;
4510         }
4511
4512         if (*outputAsFloat != constructNormalizedFloat(8, 0x304000) &&
4513                 *outputAsFloat != constructNormalizedFloat(8, 0x300000)) {
4514                 return false;
4515         }
4516         outputAsFloat++;
4517
4518         if (*outputAsFloat != -constructNormalizedFloat(-7, 0x600000) &&
4519                 *outputAsFloat != -constructNormalizedFloat(-7, 0x604000)) {
4520                 return false;
4521         }
4522         outputAsFloat++;
4523
4524         if (*outputAsFloat != constructNormalizedFloat(2, 0x01C000) &&
4525                 *outputAsFloat != constructNormalizedFloat(2, 0x020000)) {
4526                 return false;
4527         }
4528         outputAsFloat++;
4529
4530         if (*outputAsFloat != constructNormalizedFloat(1, 0xFFC000) &&
4531                 *outputAsFloat != constructNormalizedFloat(2, 0x000000)) {
4532                 return false;
4533         }
4534
4535         return true;
4536 }
4537
4538 // Checks that every output from a test-case is a float NaN.
4539 bool compareNan (const std::vector<Resource>&, const vector<AllocationSp>& outputAllocs, const std::vector<Resource>& expectedOutputs, TestLog&)
4540 {
4541         if (outputAllocs.size() != 1)
4542                 return false;
4543
4544         // Only size is needed because we cannot compare Nans.
4545         size_t byteSize = expectedOutputs[0].getByteSize();
4546
4547         const float* const      output_as_float = static_cast<const float*>(outputAllocs[0]->getHostPtr());
4548
4549         for (size_t idx = 0; idx < byteSize / sizeof(float); ++idx)
4550         {
4551                 if (!deFloatIsNaN(output_as_float[idx]))
4552                 {
4553                         return false;
4554                 }
4555         }
4556
4557         return true;
4558 }
4559
4560 // Checks that a compute shader can generate a constant composite value of various types, without exercising a computation on it.
4561 tcu::TestCaseGroup* createOpQuantizeToF16Group (tcu::TestContext& testCtx)
4562 {
4563         de::MovePtr<tcu::TestCaseGroup> group                   (new tcu::TestCaseGroup(testCtx, "opquantize", "Tests the OpQuantizeToF16 instruction"));
4564
4565         const std::string shader (
4566                 string(getComputeAsmShaderPreamble()) +
4567
4568                 "OpSource GLSL 430\n"
4569                 "OpName %main           \"main\"\n"
4570                 "OpName %id             \"gl_GlobalInvocationID\"\n"
4571
4572                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
4573
4574                 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) + string(getComputeAsmInputOutputBuffer()) +
4575
4576                 "%id        = OpVariable %uvec3ptr Input\n"
4577                 "%zero      = OpConstant %i32 0\n"
4578
4579                 "%main      = OpFunction %void None %voidf\n"
4580                 "%label     = OpLabel\n"
4581                 "%idval     = OpLoad %uvec3 %id\n"
4582                 "%x         = OpCompositeExtract %u32 %idval 0\n"
4583                 "%inloc     = OpAccessChain %f32ptr %indata %zero %x\n"
4584                 "%inval     = OpLoad %f32 %inloc\n"
4585                 "%quant     = OpQuantizeToF16 %f32 %inval\n"
4586                 "%outloc    = OpAccessChain %f32ptr %outdata %zero %x\n"
4587                 "             OpStore %outloc %quant\n"
4588                 "             OpReturn\n"
4589                 "             OpFunctionEnd\n");
4590
4591         {
4592                 ComputeShaderSpec       spec;
4593                 const deUint32          numElements             = 100;
4594                 vector<float>           infinities;
4595                 vector<float>           results;
4596
4597                 infinities.reserve(numElements);
4598                 results.reserve(numElements);
4599
4600                 for (size_t idx = 0; idx < numElements; ++idx)
4601                 {
4602                         switch(idx % 4)
4603                         {
4604                                 case 0:
4605                                         infinities.push_back(std::numeric_limits<float>::infinity());
4606                                         results.push_back(std::numeric_limits<float>::infinity());
4607                                         break;
4608                                 case 1:
4609                                         infinities.push_back(-std::numeric_limits<float>::infinity());
4610                                         results.push_back(-std::numeric_limits<float>::infinity());
4611                                         break;
4612                                 case 2:
4613                                         infinities.push_back(std::ldexp(1.0f, 16));
4614                                         results.push_back(std::numeric_limits<float>::infinity());
4615                                         break;
4616                                 case 3:
4617                                         infinities.push_back(std::ldexp(-1.0f, 32));
4618                                         results.push_back(-std::numeric_limits<float>::infinity());
4619                                         break;
4620                         }
4621                 }
4622
4623                 spec.assembly = shader;
4624                 spec.inputs.push_back(BufferSp(new Float32Buffer(infinities)));
4625                 spec.outputs.push_back(BufferSp(new Float32Buffer(results)));
4626                 spec.numWorkGroups = IVec3(numElements, 1, 1);
4627
4628                 group->addChild(new SpvAsmComputeShaderCase(
4629                         testCtx, "infinities", "Check that infinities propagated and created", spec));
4630         }
4631
4632         {
4633                 ComputeShaderSpec       spec;
4634                 vector<float>           nans;
4635                 const deUint32          numElements             = 100;
4636
4637                 nans.reserve(numElements);
4638
4639                 for (size_t idx = 0; idx < numElements; ++idx)
4640                 {
4641                         if (idx % 2 == 0)
4642                         {
4643                                 nans.push_back(std::numeric_limits<float>::quiet_NaN());
4644                         }
4645                         else
4646                         {
4647                                 nans.push_back(-std::numeric_limits<float>::quiet_NaN());
4648                         }
4649                 }
4650
4651                 spec.assembly = shader;
4652                 spec.inputs.push_back(BufferSp(new Float32Buffer(nans)));
4653                 spec.outputs.push_back(BufferSp(new Float32Buffer(nans)));
4654                 spec.numWorkGroups = IVec3(numElements, 1, 1);
4655                 spec.verifyIO = &compareNan;
4656
4657                 group->addChild(new SpvAsmComputeShaderCase(
4658                         testCtx, "propagated_nans", "Check that nans are propagated", spec));
4659         }
4660
4661         {
4662                 ComputeShaderSpec       spec;
4663                 vector<float>           small;
4664                 vector<float>           zeros;
4665                 const deUint32          numElements             = 100;
4666
4667                 small.reserve(numElements);
4668                 zeros.reserve(numElements);
4669
4670                 for (size_t idx = 0; idx < numElements; ++idx)
4671                 {
4672                         switch(idx % 6)
4673                         {
4674                                 case 0:
4675                                         small.push_back(0.f);
4676                                         zeros.push_back(0.f);
4677                                         break;
4678                                 case 1:
4679                                         small.push_back(-0.f);
4680                                         zeros.push_back(-0.f);
4681                                         break;
4682                                 case 2:
4683                                         small.push_back(std::ldexp(1.0f, -16));
4684                                         zeros.push_back(0.f);
4685                                         break;
4686                                 case 3:
4687                                         small.push_back(std::ldexp(-1.0f, -32));
4688                                         zeros.push_back(-0.f);
4689                                         break;
4690                                 case 4:
4691                                         small.push_back(std::ldexp(1.0f, -127));
4692                                         zeros.push_back(0.f);
4693                                         break;
4694                                 case 5:
4695                                         small.push_back(-std::ldexp(1.0f, -128));
4696                                         zeros.push_back(-0.f);
4697                                         break;
4698                         }
4699                 }
4700
4701                 spec.assembly = shader;
4702                 spec.inputs.push_back(BufferSp(new Float32Buffer(small)));
4703                 spec.outputs.push_back(BufferSp(new Float32Buffer(zeros)));
4704                 spec.numWorkGroups = IVec3(numElements, 1, 1);
4705
4706                 group->addChild(new SpvAsmComputeShaderCase(
4707                         testCtx, "flush_to_zero", "Check that values are zeroed correctly", spec));
4708         }
4709
4710         {
4711                 ComputeShaderSpec       spec;
4712                 vector<float>           exact;
4713                 const deUint32          numElements             = 200;
4714
4715                 exact.reserve(numElements);
4716
4717                 for (size_t idx = 0; idx < numElements; ++idx)
4718                         exact.push_back(static_cast<float>(static_cast<int>(idx) - 100));
4719
4720                 spec.assembly = shader;
4721                 spec.inputs.push_back(BufferSp(new Float32Buffer(exact)));
4722                 spec.outputs.push_back(BufferSp(new Float32Buffer(exact)));
4723                 spec.numWorkGroups = IVec3(numElements, 1, 1);
4724
4725                 group->addChild(new SpvAsmComputeShaderCase(
4726                         testCtx, "exact", "Check that values exactly preserved where appropriate", spec));
4727         }
4728
4729         {
4730                 ComputeShaderSpec       spec;
4731                 vector<float>           inputs;
4732                 const deUint32          numElements             = 4;
4733
4734                 inputs.push_back(constructNormalizedFloat(8,    0x300300));
4735                 inputs.push_back(-constructNormalizedFloat(-7,  0x600800));
4736                 inputs.push_back(constructNormalizedFloat(2,    0x01E000));
4737                 inputs.push_back(constructNormalizedFloat(1,    0xFFE000));
4738
4739                 spec.assembly = shader;
4740                 spec.verifyIO = &compareOpQuantizeF16ComputeExactCase;
4741                 spec.inputs.push_back(BufferSp(new Float32Buffer(inputs)));
4742                 spec.outputs.push_back(BufferSp(new Float32Buffer(inputs)));
4743                 spec.numWorkGroups = IVec3(numElements, 1, 1);
4744
4745                 group->addChild(new SpvAsmComputeShaderCase(
4746                         testCtx, "rounded", "Check that are rounded when needed", spec));
4747         }
4748
4749         return group.release();
4750 }
4751
4752 tcu::TestCaseGroup* createSpecConstantOpQuantizeToF16Group (tcu::TestContext& testCtx)
4753 {
4754         de::MovePtr<tcu::TestCaseGroup> group                   (new tcu::TestCaseGroup(testCtx, "opspecconstantop_opquantize", "Tests the OpQuantizeToF16 opcode for the OpSpecConstantOp instruction"));
4755
4756         const std::string shader (
4757                 string(getComputeAsmShaderPreamble()) +
4758
4759                 "OpName %main           \"main\"\n"
4760                 "OpName %id             \"gl_GlobalInvocationID\"\n"
4761
4762                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
4763
4764                 "OpDecorate %sc_0  SpecId 0\n"
4765                 "OpDecorate %sc_1  SpecId 1\n"
4766                 "OpDecorate %sc_2  SpecId 2\n"
4767                 "OpDecorate %sc_3  SpecId 3\n"
4768                 "OpDecorate %sc_4  SpecId 4\n"
4769                 "OpDecorate %sc_5  SpecId 5\n"
4770
4771                 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) + string(getComputeAsmInputOutputBuffer()) +
4772
4773                 "%id        = OpVariable %uvec3ptr Input\n"
4774                 "%zero      = OpConstant %i32 0\n"
4775                 "%c_u32_6   = OpConstant %u32 6\n"
4776
4777                 "%sc_0      = OpSpecConstant %f32 0.\n"
4778                 "%sc_1      = OpSpecConstant %f32 0.\n"
4779                 "%sc_2      = OpSpecConstant %f32 0.\n"
4780                 "%sc_3      = OpSpecConstant %f32 0.\n"
4781                 "%sc_4      = OpSpecConstant %f32 0.\n"
4782                 "%sc_5      = OpSpecConstant %f32 0.\n"
4783
4784                 "%sc_0_quant = OpSpecConstantOp %f32 QuantizeToF16 %sc_0\n"
4785                 "%sc_1_quant = OpSpecConstantOp %f32 QuantizeToF16 %sc_1\n"
4786                 "%sc_2_quant = OpSpecConstantOp %f32 QuantizeToF16 %sc_2\n"
4787                 "%sc_3_quant = OpSpecConstantOp %f32 QuantizeToF16 %sc_3\n"
4788                 "%sc_4_quant = OpSpecConstantOp %f32 QuantizeToF16 %sc_4\n"
4789                 "%sc_5_quant = OpSpecConstantOp %f32 QuantizeToF16 %sc_5\n"
4790
4791                 "%main      = OpFunction %void None %voidf\n"
4792                 "%label     = OpLabel\n"
4793                 "%idval     = OpLoad %uvec3 %id\n"
4794                 "%x         = OpCompositeExtract %u32 %idval 0\n"
4795                 "%outloc    = OpAccessChain %f32ptr %outdata %zero %x\n"
4796                 "%selector  = OpUMod %u32 %x %c_u32_6\n"
4797                 "            OpSelectionMerge %exit None\n"
4798                 "            OpSwitch %selector %exit 0 %case0 1 %case1 2 %case2 3 %case3 4 %case4 5 %case5\n"
4799
4800                 "%case0     = OpLabel\n"
4801                 "             OpStore %outloc %sc_0_quant\n"
4802                 "             OpBranch %exit\n"
4803
4804                 "%case1     = OpLabel\n"
4805                 "             OpStore %outloc %sc_1_quant\n"
4806                 "             OpBranch %exit\n"
4807
4808                 "%case2     = OpLabel\n"
4809                 "             OpStore %outloc %sc_2_quant\n"
4810                 "             OpBranch %exit\n"
4811
4812                 "%case3     = OpLabel\n"
4813                 "             OpStore %outloc %sc_3_quant\n"
4814                 "             OpBranch %exit\n"
4815
4816                 "%case4     = OpLabel\n"
4817                 "             OpStore %outloc %sc_4_quant\n"
4818                 "             OpBranch %exit\n"
4819
4820                 "%case5     = OpLabel\n"
4821                 "             OpStore %outloc %sc_5_quant\n"
4822                 "             OpBranch %exit\n"
4823
4824                 "%exit      = OpLabel\n"
4825                 "             OpReturn\n"
4826
4827                 "             OpFunctionEnd\n");
4828
4829         {
4830                 ComputeShaderSpec       spec;
4831                 const deUint8           numCases        = 4;
4832                 vector<float>           inputs          (numCases, 0.f);
4833                 vector<float>           outputs;
4834
4835                 spec.assembly           = shader;
4836                 spec.numWorkGroups      = IVec3(numCases, 1, 1);
4837
4838                 spec.specConstants.append<deInt32>(bitwiseCast<deUint32>(std::numeric_limits<float>::infinity()));
4839                 spec.specConstants.append<deInt32>(bitwiseCast<deUint32>(-std::numeric_limits<float>::infinity()));
4840                 spec.specConstants.append<deInt32>(bitwiseCast<deUint32>(std::ldexp(1.0f, 16)));
4841                 spec.specConstants.append<deInt32>(bitwiseCast<deUint32>(std::ldexp(-1.0f, 32)));
4842
4843                 outputs.push_back(std::numeric_limits<float>::infinity());
4844                 outputs.push_back(-std::numeric_limits<float>::infinity());
4845                 outputs.push_back(std::numeric_limits<float>::infinity());
4846                 outputs.push_back(-std::numeric_limits<float>::infinity());
4847
4848                 spec.inputs.push_back(BufferSp(new Float32Buffer(inputs)));
4849                 spec.outputs.push_back(BufferSp(new Float32Buffer(outputs)));
4850
4851                 group->addChild(new SpvAsmComputeShaderCase(
4852                         testCtx, "infinities", "Check that infinities propagated and created", spec));
4853         }
4854
4855         {
4856                 ComputeShaderSpec       spec;
4857                 const deUint8           numCases        = 2;
4858                 vector<float>           inputs          (numCases, 0.f);
4859                 vector<float>           outputs;
4860
4861                 spec.assembly           = shader;
4862                 spec.numWorkGroups      = IVec3(numCases, 1, 1);
4863                 spec.verifyIO           = &compareNan;
4864
4865                 outputs.push_back(std::numeric_limits<float>::quiet_NaN());
4866                 outputs.push_back(-std::numeric_limits<float>::quiet_NaN());
4867
4868                 for (deUint8 idx = 0; idx < numCases; ++idx)
4869                         spec.specConstants.append<deInt32>(bitwiseCast<deUint32>(outputs[idx]));
4870
4871                 spec.inputs.push_back(BufferSp(new Float32Buffer(inputs)));
4872                 spec.outputs.push_back(BufferSp(new Float32Buffer(outputs)));
4873
4874                 group->addChild(new SpvAsmComputeShaderCase(
4875                         testCtx, "propagated_nans", "Check that nans are propagated", spec));
4876         }
4877
4878         {
4879                 ComputeShaderSpec       spec;
4880                 const deUint8           numCases        = 6;
4881                 vector<float>           inputs          (numCases, 0.f);
4882                 vector<float>           outputs;
4883
4884                 spec.assembly           = shader;
4885                 spec.numWorkGroups      = IVec3(numCases, 1, 1);
4886
4887                 spec.specConstants.append<deInt32>(bitwiseCast<deUint32>(0.f));
4888                 spec.specConstants.append<deInt32>(bitwiseCast<deUint32>(-0.f));
4889                 spec.specConstants.append<deInt32>(bitwiseCast<deUint32>(std::ldexp(1.0f, -16)));
4890                 spec.specConstants.append<deInt32>(bitwiseCast<deUint32>(std::ldexp(-1.0f, -32)));
4891                 spec.specConstants.append<deInt32>(bitwiseCast<deUint32>(std::ldexp(1.0f, -127)));
4892                 spec.specConstants.append<deInt32>(bitwiseCast<deUint32>(-std::ldexp(1.0f, -128)));
4893
4894                 outputs.push_back(0.f);
4895                 outputs.push_back(-0.f);
4896                 outputs.push_back(0.f);
4897                 outputs.push_back(-0.f);
4898                 outputs.push_back(0.f);
4899                 outputs.push_back(-0.f);
4900
4901                 spec.inputs.push_back(BufferSp(new Float32Buffer(inputs)));
4902                 spec.outputs.push_back(BufferSp(new Float32Buffer(outputs)));
4903
4904                 group->addChild(new SpvAsmComputeShaderCase(
4905                         testCtx, "flush_to_zero", "Check that values are zeroed correctly", spec));
4906         }
4907
4908         {
4909                 ComputeShaderSpec       spec;
4910                 const deUint8           numCases        = 6;
4911                 vector<float>           inputs          (numCases, 0.f);
4912                 vector<float>           outputs;
4913
4914                 spec.assembly           = shader;
4915                 spec.numWorkGroups      = IVec3(numCases, 1, 1);
4916
4917                 for (deUint8 idx = 0; idx < 6; ++idx)
4918                 {
4919                         const float f = static_cast<float>(idx * 10 - 30) / 4.f;
4920                         spec.specConstants.append<deInt32>(bitwiseCast<deUint32>(f));
4921                         outputs.push_back(f);
4922                 }
4923
4924                 spec.inputs.push_back(BufferSp(new Float32Buffer(inputs)));
4925                 spec.outputs.push_back(BufferSp(new Float32Buffer(outputs)));
4926
4927                 group->addChild(new SpvAsmComputeShaderCase(
4928                         testCtx, "exact", "Check that values exactly preserved where appropriate", spec));
4929         }
4930
4931         {
4932                 ComputeShaderSpec       spec;
4933                 const deUint8           numCases        = 4;
4934                 vector<float>           inputs          (numCases, 0.f);
4935                 vector<float>           outputs;
4936
4937                 spec.assembly           = shader;
4938                 spec.numWorkGroups      = IVec3(numCases, 1, 1);
4939                 spec.verifyIO           = &compareOpQuantizeF16ComputeExactCase;
4940
4941                 outputs.push_back(constructNormalizedFloat(8, 0x300300));
4942                 outputs.push_back(-constructNormalizedFloat(-7, 0x600800));
4943                 outputs.push_back(constructNormalizedFloat(2, 0x01E000));
4944                 outputs.push_back(constructNormalizedFloat(1, 0xFFE000));
4945
4946                 for (deUint8 idx = 0; idx < numCases; ++idx)
4947                         spec.specConstants.append<deInt32>(bitwiseCast<deUint32>(outputs[idx]));
4948
4949                 spec.inputs.push_back(BufferSp(new Float32Buffer(inputs)));
4950                 spec.outputs.push_back(BufferSp(new Float32Buffer(outputs)));
4951
4952                 group->addChild(new SpvAsmComputeShaderCase(
4953                         testCtx, "rounded", "Check that are rounded when needed", spec));
4954         }
4955
4956         return group.release();
4957 }
4958
4959 // Checks that constant null/composite values can be used in computation.
4960 tcu::TestCaseGroup* createOpConstantUsageGroup (tcu::TestContext& testCtx)
4961 {
4962         de::MovePtr<tcu::TestCaseGroup> group                   (new tcu::TestCaseGroup(testCtx, "opconstantnullcomposite", "Spotcheck the OpConstantNull & OpConstantComposite instruction"));
4963         ComputeShaderSpec                               spec;
4964         de::Random                                              rnd                             (deStringHash(group->getName()));
4965         const int                                               numElements             = 100;
4966         vector<float>                                   positiveFloats  (numElements, 0);
4967         vector<float>                                   negativeFloats  (numElements, 0);
4968
4969         fillRandomScalars(rnd, 1.f, 100.f, &positiveFloats[0], numElements);
4970
4971         for (size_t ndx = 0; ndx < numElements; ++ndx)
4972                 negativeFloats[ndx] = -positiveFloats[ndx];
4973
4974         spec.assembly =
4975                 "OpCapability Shader\n"
4976                 "%std450 = OpExtInstImport \"GLSL.std.450\"\n"
4977                 "OpMemoryModel Logical GLSL450\n"
4978                 "OpEntryPoint GLCompute %main \"main\" %id\n"
4979                 "OpExecutionMode %main LocalSize 1 1 1\n"
4980
4981                 "OpSource GLSL 430\n"
4982                 "OpName %main           \"main\"\n"
4983                 "OpName %id             \"gl_GlobalInvocationID\"\n"
4984
4985                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
4986
4987                 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) +
4988
4989                 "%fmat      = OpTypeMatrix %fvec3 3\n"
4990                 "%ten       = OpConstant %u32 10\n"
4991                 "%f32arr10  = OpTypeArray %f32 %ten\n"
4992                 "%fst       = OpTypeStruct %f32 %f32\n"
4993
4994                 + string(getComputeAsmInputOutputBuffer()) +
4995
4996                 "%id        = OpVariable %uvec3ptr Input\n"
4997                 "%zero      = OpConstant %i32 0\n"
4998
4999                 // Create a bunch of null values
5000                 "%unull     = OpConstantNull %u32\n"
5001                 "%fnull     = OpConstantNull %f32\n"
5002                 "%vnull     = OpConstantNull %fvec3\n"
5003                 "%mnull     = OpConstantNull %fmat\n"
5004                 "%anull     = OpConstantNull %f32arr10\n"
5005                 "%snull     = OpConstantComposite %fst %fnull %fnull\n"
5006
5007                 "%main      = OpFunction %void None %voidf\n"
5008                 "%label     = OpLabel\n"
5009                 "%idval     = OpLoad %uvec3 %id\n"
5010                 "%x         = OpCompositeExtract %u32 %idval 0\n"
5011                 "%inloc     = OpAccessChain %f32ptr %indata %zero %x\n"
5012                 "%inval     = OpLoad %f32 %inloc\n"
5013                 "%neg       = OpFNegate %f32 %inval\n"
5014
5015                 // Get the abs() of (a certain element of) those null values
5016                 "%unull_cov = OpConvertUToF %f32 %unull\n"
5017                 "%unull_abs = OpExtInst %f32 %std450 FAbs %unull_cov\n"
5018                 "%fnull_abs = OpExtInst %f32 %std450 FAbs %fnull\n"
5019                 "%vnull_0   = OpCompositeExtract %f32 %vnull 0\n"
5020                 "%vnull_abs = OpExtInst %f32 %std450 FAbs %vnull_0\n"
5021                 "%mnull_12  = OpCompositeExtract %f32 %mnull 1 2\n"
5022                 "%mnull_abs = OpExtInst %f32 %std450 FAbs %mnull_12\n"
5023                 "%anull_3   = OpCompositeExtract %f32 %anull 3\n"
5024                 "%anull_abs = OpExtInst %f32 %std450 FAbs %anull_3\n"
5025                 "%snull_1   = OpCompositeExtract %f32 %snull 1\n"
5026                 "%snull_abs = OpExtInst %f32 %std450 FAbs %snull_1\n"
5027
5028                 // Add them all
5029                 "%add1      = OpFAdd %f32 %neg  %unull_abs\n"
5030                 "%add2      = OpFAdd %f32 %add1 %fnull_abs\n"
5031                 "%add3      = OpFAdd %f32 %add2 %vnull_abs\n"
5032                 "%add4      = OpFAdd %f32 %add3 %mnull_abs\n"
5033                 "%add5      = OpFAdd %f32 %add4 %anull_abs\n"
5034                 "%final     = OpFAdd %f32 %add5 %snull_abs\n"
5035
5036                 "%outloc    = OpAccessChain %f32ptr %outdata %zero %x\n"
5037                 "             OpStore %outloc %final\n" // write to output
5038                 "             OpReturn\n"
5039                 "             OpFunctionEnd\n";
5040         spec.inputs.push_back(BufferSp(new Float32Buffer(positiveFloats)));
5041         spec.outputs.push_back(BufferSp(new Float32Buffer(negativeFloats)));
5042         spec.numWorkGroups = IVec3(numElements, 1, 1);
5043
5044         group->addChild(new SpvAsmComputeShaderCase(testCtx, "spotcheck", "Check that values constructed via OpConstantNull & OpConstantComposite can be used", spec));
5045
5046         return group.release();
5047 }
5048
5049 // Assembly code used for testing loop control is based on GLSL source code:
5050 // #version 430
5051 //
5052 // layout(std140, set = 0, binding = 0) readonly buffer Input {
5053 //   float elements[];
5054 // } input_data;
5055 // layout(std140, set = 0, binding = 1) writeonly buffer Output {
5056 //   float elements[];
5057 // } output_data;
5058 //
5059 // void main() {
5060 //   uint x = gl_GlobalInvocationID.x;
5061 //   output_data.elements[x] = input_data.elements[x];
5062 //   for (uint i = 0; i < 4; ++i)
5063 //     output_data.elements[x] += 1.f;
5064 // }
5065 tcu::TestCaseGroup* createLoopControlGroup (tcu::TestContext& testCtx)
5066 {
5067         de::MovePtr<tcu::TestCaseGroup> group                   (new tcu::TestCaseGroup(testCtx, "loop_control", "Tests loop control cases"));
5068         vector<CaseParameter>                   cases;
5069         de::Random                                              rnd                             (deStringHash(group->getName()));
5070         const int                                               numElements             = 100;
5071         vector<float>                                   inputFloats             (numElements, 0);
5072         vector<float>                                   outputFloats    (numElements, 0);
5073         const StringTemplate                    shaderTemplate  (
5074                 string(getComputeAsmShaderPreamble()) +
5075
5076                 "OpSource GLSL 430\n"
5077                 "OpName %main \"main\"\n"
5078                 "OpName %id \"gl_GlobalInvocationID\"\n"
5079
5080                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
5081
5082                 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) + string(getComputeAsmInputOutputBuffer()) +
5083
5084                 "%u32ptr      = OpTypePointer Function %u32\n"
5085
5086                 "%id          = OpVariable %uvec3ptr Input\n"
5087                 "%zero        = OpConstant %i32 0\n"
5088                 "%uzero       = OpConstant %u32 0\n"
5089                 "%one         = OpConstant %i32 1\n"
5090                 "%constf1     = OpConstant %f32 1.0\n"
5091                 "%four        = OpConstant %u32 4\n"
5092
5093                 "%main        = OpFunction %void None %voidf\n"
5094                 "%entry       = OpLabel\n"
5095                 "%i           = OpVariable %u32ptr Function\n"
5096                 "               OpStore %i %uzero\n"
5097
5098                 "%idval       = OpLoad %uvec3 %id\n"
5099                 "%x           = OpCompositeExtract %u32 %idval 0\n"
5100                 "%inloc       = OpAccessChain %f32ptr %indata %zero %x\n"
5101                 "%inval       = OpLoad %f32 %inloc\n"
5102                 "%outloc      = OpAccessChain %f32ptr %outdata %zero %x\n"
5103                 "               OpStore %outloc %inval\n"
5104                 "               OpBranch %loop_entry\n"
5105
5106                 "%loop_entry  = OpLabel\n"
5107                 "%i_val       = OpLoad %u32 %i\n"
5108                 "%cmp_lt      = OpULessThan %bool %i_val %four\n"
5109                 "               OpLoopMerge %loop_merge %loop_body ${CONTROL}\n"
5110                 "               OpBranchConditional %cmp_lt %loop_body %loop_merge\n"
5111                 "%loop_body   = OpLabel\n"
5112                 "%outval      = OpLoad %f32 %outloc\n"
5113                 "%addf1       = OpFAdd %f32 %outval %constf1\n"
5114                 "               OpStore %outloc %addf1\n"
5115                 "%new_i       = OpIAdd %u32 %i_val %one\n"
5116                 "               OpStore %i %new_i\n"
5117                 "               OpBranch %loop_entry\n"
5118                 "%loop_merge  = OpLabel\n"
5119                 "               OpReturn\n"
5120                 "               OpFunctionEnd\n");
5121
5122         cases.push_back(CaseParameter("none",                           "None"));
5123         cases.push_back(CaseParameter("unroll",                         "Unroll"));
5124         cases.push_back(CaseParameter("dont_unroll",            "DontUnroll"));
5125
5126         fillRandomScalars(rnd, -100.f, 100.f, &inputFloats[0], numElements);
5127
5128         for (size_t ndx = 0; ndx < numElements; ++ndx)
5129                 outputFloats[ndx] = inputFloats[ndx] + 4.f;
5130
5131         for (size_t caseNdx = 0; caseNdx < cases.size(); ++caseNdx)
5132         {
5133                 map<string, string>             specializations;
5134                 ComputeShaderSpec               spec;
5135
5136                 specializations["CONTROL"] = cases[caseNdx].param;
5137                 spec.assembly = shaderTemplate.specialize(specializations);
5138                 spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
5139                 spec.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
5140                 spec.numWorkGroups = IVec3(numElements, 1, 1);
5141
5142                 group->addChild(new SpvAsmComputeShaderCase(testCtx, cases[caseNdx].name, cases[caseNdx].name, spec));
5143         }
5144
5145         group->addChild(new SpvAsmLoopControlDependencyLengthCase(testCtx, "dependency_length", "dependency_length"));
5146         group->addChild(new SpvAsmLoopControlDependencyInfiniteCase(testCtx, "dependency_infinite", "dependency_infinite"));
5147
5148         return group.release();
5149 }
5150
5151 // Assembly code used for testing selection control is based on GLSL source code:
5152 // #version 430
5153 //
5154 // layout(std140, set = 0, binding = 0) readonly buffer Input {
5155 //   float elements[];
5156 // } input_data;
5157 // layout(std140, set = 0, binding = 1) writeonly buffer Output {
5158 //   float elements[];
5159 // } output_data;
5160 //
5161 // void main() {
5162 //   uint x = gl_GlobalInvocationID.x;
5163 //   float val = input_data.elements[x];
5164 //   if (val > 10.f)
5165 //     output_data.elements[x] = val + 1.f;
5166 //   else
5167 //     output_data.elements[x] = val - 1.f;
5168 // }
5169 tcu::TestCaseGroup* createSelectionControlGroup (tcu::TestContext& testCtx)
5170 {
5171         de::MovePtr<tcu::TestCaseGroup> group                   (new tcu::TestCaseGroup(testCtx, "selection_control", "Tests selection control cases"));
5172         vector<CaseParameter>                   cases;
5173         de::Random                                              rnd                             (deStringHash(group->getName()));
5174         const int                                               numElements             = 100;
5175         vector<float>                                   inputFloats             (numElements, 0);
5176         vector<float>                                   outputFloats    (numElements, 0);
5177         const StringTemplate                    shaderTemplate  (
5178                 string(getComputeAsmShaderPreamble()) +
5179
5180                 "OpSource GLSL 430\n"
5181                 "OpName %main \"main\"\n"
5182                 "OpName %id \"gl_GlobalInvocationID\"\n"
5183
5184                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
5185
5186                 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) + string(getComputeAsmInputOutputBuffer()) +
5187
5188                 "%id       = OpVariable %uvec3ptr Input\n"
5189                 "%zero     = OpConstant %i32 0\n"
5190                 "%constf1  = OpConstant %f32 1.0\n"
5191                 "%constf10 = OpConstant %f32 10.0\n"
5192
5193                 "%main     = OpFunction %void None %voidf\n"
5194                 "%entry    = OpLabel\n"
5195                 "%idval    = OpLoad %uvec3 %id\n"
5196                 "%x        = OpCompositeExtract %u32 %idval 0\n"
5197                 "%inloc    = OpAccessChain %f32ptr %indata %zero %x\n"
5198                 "%inval    = OpLoad %f32 %inloc\n"
5199                 "%outloc   = OpAccessChain %f32ptr %outdata %zero %x\n"
5200                 "%cmp_gt   = OpFOrdGreaterThan %bool %inval %constf10\n"
5201
5202                 "            OpSelectionMerge %if_end ${CONTROL}\n"
5203                 "            OpBranchConditional %cmp_gt %if_true %if_false\n"
5204                 "%if_true  = OpLabel\n"
5205                 "%addf1    = OpFAdd %f32 %inval %constf1\n"
5206                 "            OpStore %outloc %addf1\n"
5207                 "            OpBranch %if_end\n"
5208                 "%if_false = OpLabel\n"
5209                 "%subf1    = OpFSub %f32 %inval %constf1\n"
5210                 "            OpStore %outloc %subf1\n"
5211                 "            OpBranch %if_end\n"
5212                 "%if_end   = OpLabel\n"
5213                 "            OpReturn\n"
5214                 "            OpFunctionEnd\n");
5215
5216         cases.push_back(CaseParameter("none",                                   "None"));
5217         cases.push_back(CaseParameter("flatten",                                "Flatten"));
5218         cases.push_back(CaseParameter("dont_flatten",                   "DontFlatten"));
5219         cases.push_back(CaseParameter("flatten_dont_flatten",   "DontFlatten|Flatten"));
5220
5221         fillRandomScalars(rnd, -100.f, 100.f, &inputFloats[0], numElements);
5222
5223         // CPU might not use the same rounding mode as the GPU. Use whole numbers to avoid rounding differences.
5224         floorAll(inputFloats);
5225
5226         for (size_t ndx = 0; ndx < numElements; ++ndx)
5227                 outputFloats[ndx] = inputFloats[ndx] + (inputFloats[ndx] > 10.f ? 1.f : -1.f);
5228
5229         for (size_t caseNdx = 0; caseNdx < cases.size(); ++caseNdx)
5230         {
5231                 map<string, string>             specializations;
5232                 ComputeShaderSpec               spec;
5233
5234                 specializations["CONTROL"] = cases[caseNdx].param;
5235                 spec.assembly = shaderTemplate.specialize(specializations);
5236                 spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
5237                 spec.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
5238                 spec.numWorkGroups = IVec3(numElements, 1, 1);
5239
5240                 group->addChild(new SpvAsmComputeShaderCase(testCtx, cases[caseNdx].name, cases[caseNdx].name, spec));
5241         }
5242
5243         return group.release();
5244 }
5245
5246 void getOpNameAbuseCases (vector<CaseParameter> &abuseCases)
5247 {
5248         // Generate a long name.
5249         std::string longname;
5250         longname.resize(65535, 'k'); // max string literal, spir-v 2.17
5251
5252         // Some bad names, abusing utf-8 encoding. This may also cause problems
5253         // with the logs.
5254         // 1. Various illegal code points in utf-8
5255         std::string utf8illegal =
5256                 "Illegal bytes in UTF-8: "
5257                 "\xc0 \xc1 \xf5 \xf6 \xf7 \xf8 \xf9 \xfa \xfb \xfc \xfd \xfe \xff"
5258                 "illegal surrogates: \xed\xad\xbf \xed\xbe\x80";
5259
5260         // 2. Zero encoded as overlong, not exactly legal but often supported to differentiate from terminating zero
5261         std::string utf8nul = "UTF-8 encoded nul \xC0\x80 (should not end name)";
5262
5263         // 3. Some overlong encodings
5264         std::string utf8overlong =
5265                 "UTF-8 overlong \xF0\x82\x82\xAC \xfc\x83\xbf\xbf\xbf\xbf \xf8\x87\xbf\xbf\xbf "
5266                 "\xf0\x8f\xbf\xbf";
5267
5268         // 4. Internet "zalgo" meme "bleeding text"
5269         std::string utf8zalgo =
5270                 "\x56\xcc\xb5\xcc\x85\xcc\x94\xcc\x88\xcd\x8a\xcc\x91\xcc\x88\xcd\x91\xcc\x83\xcd\x82"
5271                 "\xcc\x83\xcd\x90\xcc\x8a\xcc\x92\xcc\x92\xcd\x8b\xcc\x94\xcd\x9d\xcc\x98\xcc\xab\xcc"
5272                 "\xae\xcc\xa9\xcc\xad\xcc\x97\xcc\xb0\x75\xcc\xb6\xcc\xbe\xcc\x80\xcc\x82\xcc\x84\xcd"
5273                 "\x84\xcc\x90\xcd\x86\xcc\x9a\xcd\x84\xcc\x9b\xcd\x86\xcd\x92\xcc\x9a\xcd\x99\xcd\x99"
5274                 "\xcc\xbb\xcc\x98\xcd\x8e\xcd\x88\xcd\x9a\xcc\xa6\xcc\x9c\xcc\xab\xcc\x99\xcd\x94\xcd"
5275                 "\x99\xcd\x95\xcc\xa5\xcc\xab\xcd\x89\x6c\xcc\xb8\xcc\x8e\xcc\x8b\xcc\x8b\xcc\x9a\xcc"
5276                 "\x8e\xcd\x9d\xcc\x80\xcc\xa1\xcc\xad\xcd\x9c\xcc\xba\xcc\x96\xcc\xb3\xcc\xa2\xcd\x8e"
5277                 "\xcc\xa2\xcd\x96\x6b\xcc\xb8\xcc\x84\xcd\x81\xcc\xbf\xcc\x8d\xcc\x89\xcc\x85\xcc\x92"
5278                 "\xcc\x84\xcc\x90\xcd\x81\xcc\x93\xcd\x90\xcd\x92\xcd\x9d\xcc\x84\xcd\x98\xcd\x9d\xcd"
5279                 "\xa0\xcd\x91\xcc\x94\xcc\xb9\xcd\x93\xcc\xa5\xcd\x87\xcc\xad\xcc\xa7\xcd\x96\xcd\x99"
5280                 "\xcc\x9d\xcc\xbc\xcd\x96\xcd\x93\xcc\x9d\xcc\x99\xcc\xa8\xcc\xb1\xcd\x85\xcc\xba\xcc"
5281                 "\xa7\x61\xcc\xb8\xcc\x8e\xcc\x81\xcd\x90\xcd\x84\xcd\x8c\xcc\x8c\xcc\x85\xcd\x86\xcc"
5282                 "\x84\xcd\x84\xcc\x90\xcc\x84\xcc\x8d\xcd\x99\xcd\x8d\xcc\xb0\xcc\xa3\xcc\xa6\xcd\x89"
5283                 "\xcd\x8d\xcd\x87\xcc\x98\xcd\x8d\xcc\xa4\xcd\x9a\xcd\x8e\xcc\xab\xcc\xb9\xcc\xac\xcc"
5284                 "\xa2\xcd\x87\xcc\xa0\xcc\xb3\xcd\x89\xcc\xb9\xcc\xa7\xcc\xa6\xcd\x89\xcd\x95\x6e\xcc"
5285                 "\xb8\xcd\x8a\xcc\x8a\xcd\x82\xcc\x9b\xcd\x81\xcd\x90\xcc\x85\xcc\x9b\xcd\x80\xcd\x91"
5286                 "\xcd\x9b\xcc\x81\xcd\x81\xcc\x9a\xcc\xb3\xcd\x9c\xcc\x9e\xcc\x9d\xcd\x99\xcc\xa2\xcd"
5287                 "\x93\xcd\x96\xcc\x97\xff";
5288
5289         // General name abuses
5290         abuseCases.push_back(CaseParameter("_has_very_long_name", longname));
5291         abuseCases.push_back(CaseParameter("_utf8_illegal", utf8illegal));
5292         abuseCases.push_back(CaseParameter("_utf8_nul", utf8nul));
5293         abuseCases.push_back(CaseParameter("_utf8_overlong", utf8overlong));
5294         abuseCases.push_back(CaseParameter("_utf8_zalgo", utf8zalgo));
5295
5296         // GL keywords
5297         abuseCases.push_back(CaseParameter("_is_gl_Position", "gl_Position"));
5298         abuseCases.push_back(CaseParameter("_is_gl_InstanceID", "gl_InstanceID"));
5299         abuseCases.push_back(CaseParameter("_is_gl_PrimitiveID", "gl_PrimitiveID"));
5300         abuseCases.push_back(CaseParameter("_is_gl_TessCoord", "gl_TessCoord"));
5301         abuseCases.push_back(CaseParameter("_is_gl_PerVertex", "gl_PerVertex"));
5302         abuseCases.push_back(CaseParameter("_is_gl_InvocationID", "gl_InvocationID"));
5303         abuseCases.push_back(CaseParameter("_is_gl_PointSize", "gl_PointSize"));
5304         abuseCases.push_back(CaseParameter("_is_gl_PointCoord", "gl_PointCoord"));
5305         abuseCases.push_back(CaseParameter("_is_gl_Layer", "gl_Layer"));
5306         abuseCases.push_back(CaseParameter("_is_gl_FragDepth", "gl_FragDepth"));
5307         abuseCases.push_back(CaseParameter("_is_gl_NumWorkGroups", "gl_NumWorkGroups"));
5308         abuseCases.push_back(CaseParameter("_is_gl_WorkGroupID", "gl_WorkGroupID"));
5309         abuseCases.push_back(CaseParameter("_is_gl_LocalInvocationID", "gl_LocalInvocationID"));
5310         abuseCases.push_back(CaseParameter("_is_gl_GlobalInvocationID", "gl_GlobalInvocationID"));
5311         abuseCases.push_back(CaseParameter("_is_gl_MaxVertexAttribs", "gl_MaxVertexAttribs"));
5312         abuseCases.push_back(CaseParameter("_is_gl_MaxViewports", "gl_MaxViewports"));
5313         abuseCases.push_back(CaseParameter("_is_gl_MaxComputeWorkGroupCount", "gl_MaxComputeWorkGroupCount"));
5314         abuseCases.push_back(CaseParameter("_is_mat3", "mat3"));
5315         abuseCases.push_back(CaseParameter("_is_volatile", "volatile"));
5316         abuseCases.push_back(CaseParameter("_is_inout", "inout"));
5317         abuseCases.push_back(CaseParameter("_is_isampler3d", "isampler3d"));
5318 }
5319
5320 tcu::TestCaseGroup* createOpNameGroup (tcu::TestContext& testCtx)
5321 {
5322         de::MovePtr<tcu::TestCaseGroup> group                   (new tcu::TestCaseGroup(testCtx, "opname", "Tests OpName cases"));
5323         de::MovePtr<tcu::TestCaseGroup> entryMainGroup  (new tcu::TestCaseGroup(testCtx, "entry_main", "OpName tests with entry main"));
5324         de::MovePtr<tcu::TestCaseGroup> entryNotGroup   (new tcu::TestCaseGroup(testCtx, "entry_rdc", "OpName tests with entry rdc"));
5325         de::MovePtr<tcu::TestCaseGroup> abuseGroup              (new tcu::TestCaseGroup(testCtx, "abuse", "OpName abuse tests"));
5326         vector<CaseParameter>                   cases;
5327         vector<CaseParameter>                   abuseCases;
5328         vector<string>                                  testFunc;
5329         de::Random                                              rnd                             (deStringHash(group->getName()));
5330         const int                                               numElements             = 128;
5331         vector<float>                                   inputFloats             (numElements, 0);
5332         vector<float>                                   outputFloats    (numElements, 0);
5333
5334         getOpNameAbuseCases(abuseCases);
5335
5336         fillRandomScalars(rnd, -100.0f, 100.0f, &inputFloats[0], numElements);
5337
5338         for(size_t ndx = 0; ndx < numElements; ++ndx)
5339                 outputFloats[ndx] = -inputFloats[ndx];
5340
5341         const string commonShaderHeader =
5342                 "OpCapability Shader\n"
5343                 "OpMemoryModel Logical GLSL450\n"
5344                 "OpEntryPoint GLCompute %main \"main\" %id\n"
5345                 "OpExecutionMode %main LocalSize 1 1 1\n";
5346
5347         const string commonShaderFooter =
5348                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
5349
5350                 + string(getComputeAsmInputOutputBufferTraits())
5351                 + string(getComputeAsmCommonTypes())
5352                 + string(getComputeAsmInputOutputBuffer()) +
5353
5354                 "%id        = OpVariable %uvec3ptr Input\n"
5355                 "%zero      = OpConstant %i32 0\n"
5356
5357                 "%func      = OpFunction %void None %voidf\n"
5358                 "%5         = OpLabel\n"
5359                 "             OpReturn\n"
5360                 "             OpFunctionEnd\n"
5361
5362                 "%main      = OpFunction %void None %voidf\n"
5363                 "%entry     = OpLabel\n"
5364                 "%7         = OpFunctionCall %void %func\n"
5365
5366                 "%idval     = OpLoad %uvec3 %id\n"
5367                 "%x         = OpCompositeExtract %u32 %idval 0\n"
5368
5369                 "%inloc     = OpAccessChain %f32ptr %indata %zero %x\n"
5370                 "%inval     = OpLoad %f32 %inloc\n"
5371                 "%neg       = OpFNegate %f32 %inval\n"
5372                 "%outloc    = OpAccessChain %f32ptr %outdata %zero %x\n"
5373                 "             OpStore %outloc %neg\n"
5374
5375                 "             OpReturn\n"
5376                 "             OpFunctionEnd\n";
5377
5378         const StringTemplate shaderTemplate (
5379                 "OpCapability Shader\n"
5380                 "OpMemoryModel Logical GLSL450\n"
5381                 "OpEntryPoint GLCompute %main \"${ENTRY}\" %id\n"
5382                 "OpExecutionMode %main LocalSize 1 1 1\n"
5383                 "OpName %${ID} \"${NAME}\"\n" +
5384                 commonShaderFooter);
5385
5386         const std::string multipleNames =
5387                 commonShaderHeader +
5388                 "OpName %main \"to_be\"\n"
5389                 "OpName %id   \"or_not\"\n"
5390                 "OpName %main \"to_be\"\n"
5391                 "OpName %main \"makes_no\"\n"
5392                 "OpName %func \"difference\"\n"
5393                 "OpName %5    \"to_me\"\n" +
5394                 commonShaderFooter;
5395
5396         {
5397                 ComputeShaderSpec       spec;
5398
5399                 spec.assembly           = multipleNames;
5400                 spec.numWorkGroups      = IVec3(numElements, 1, 1);
5401                 spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
5402                 spec.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
5403
5404                 abuseGroup->addChild(new SpvAsmComputeShaderCase(testCtx, "main_has_multiple_names", "multiple_names", spec));
5405         }
5406
5407         const std::string everythingNamed =
5408                 commonShaderHeader +
5409                 "OpName %main   \"name1\"\n"
5410                 "OpName %id     \"name2\"\n"
5411                 "OpName %zero   \"name3\"\n"
5412                 "OpName %entry  \"name4\"\n"
5413                 "OpName %func   \"name5\"\n"
5414                 "OpName %5      \"name6\"\n"
5415                 "OpName %7      \"name7\"\n"
5416                 "OpName %idval  \"name8\"\n"
5417                 "OpName %inloc  \"name9\"\n"
5418                 "OpName %inval  \"name10\"\n"
5419                 "OpName %neg    \"name11\"\n"
5420                 "OpName %outloc \"name12\"\n"+
5421                 commonShaderFooter;
5422         {
5423                 ComputeShaderSpec       spec;
5424
5425                 spec.assembly           = everythingNamed;
5426                 spec.numWorkGroups      = IVec3(numElements, 1, 1);
5427                 spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
5428                 spec.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
5429
5430                 abuseGroup->addChild(new SpvAsmComputeShaderCase(testCtx, "everything_named", "everything_named", spec));
5431         }
5432
5433         const std::string everythingNamedTheSame =
5434                 commonShaderHeader +
5435                 "OpName %main   \"the_same\"\n"
5436                 "OpName %id     \"the_same\"\n"
5437                 "OpName %zero   \"the_same\"\n"
5438                 "OpName %entry  \"the_same\"\n"
5439                 "OpName %func   \"the_same\"\n"
5440                 "OpName %5      \"the_same\"\n"
5441                 "OpName %7      \"the_same\"\n"
5442                 "OpName %idval  \"the_same\"\n"
5443                 "OpName %inloc  \"the_same\"\n"
5444                 "OpName %inval  \"the_same\"\n"
5445                 "OpName %neg    \"the_same\"\n"
5446                 "OpName %outloc \"the_same\"\n"+
5447                 commonShaderFooter;
5448         {
5449                 ComputeShaderSpec       spec;
5450
5451                 spec.assembly           = everythingNamedTheSame;
5452                 spec.numWorkGroups      = IVec3(numElements, 1, 1);
5453                 spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
5454                 spec.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
5455
5456                 abuseGroup->addChild(new SpvAsmComputeShaderCase(testCtx, "everything_named_the_same", "everything_named_the_same", spec));
5457         }
5458
5459         // main_is_...
5460         for (size_t ndx = 0; ndx < abuseCases.size(); ++ndx)
5461         {
5462                 map<string, string>     specializations;
5463                 ComputeShaderSpec       spec;
5464
5465                 specializations["ENTRY"]        = "main";
5466                 specializations["ID"]           = "main";
5467                 specializations["NAME"]         = abuseCases[ndx].param;
5468                 spec.assembly                           = shaderTemplate.specialize(specializations);
5469                 spec.numWorkGroups                      = IVec3(numElements, 1, 1);
5470                 spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
5471                 spec.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
5472
5473                 abuseGroup->addChild(new SpvAsmComputeShaderCase(testCtx, (std::string("main") + abuseCases[ndx].name).c_str(), abuseCases[ndx].name, spec));
5474         }
5475
5476         // x_is_....
5477         for (size_t ndx = 0; ndx < abuseCases.size(); ++ndx)
5478         {
5479                 map<string, string>     specializations;
5480                 ComputeShaderSpec       spec;
5481
5482                 specializations["ENTRY"]        = "main";
5483                 specializations["ID"]           = "x";
5484                 specializations["NAME"]         = abuseCases[ndx].param;
5485                 spec.assembly                           = shaderTemplate.specialize(specializations);
5486                 spec.numWorkGroups                      = IVec3(numElements, 1, 1);
5487                 spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
5488                 spec.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
5489
5490                 abuseGroup->addChild(new SpvAsmComputeShaderCase(testCtx, (std::string("x") + abuseCases[ndx].name).c_str(), abuseCases[ndx].name, spec));
5491         }
5492
5493         cases.push_back(CaseParameter("_is_main", "main"));
5494         cases.push_back(CaseParameter("_is_not_main", "not_main"));
5495         testFunc.push_back("main");
5496         testFunc.push_back("func");
5497
5498         for(size_t fNdx = 0; fNdx < testFunc.size(); ++fNdx)
5499         {
5500                 for(size_t ndx = 0; ndx < cases.size(); ++ndx)
5501                 {
5502                         map<string, string>     specializations;
5503                         ComputeShaderSpec       spec;
5504
5505                         specializations["ENTRY"]        = "main";
5506                         specializations["ID"]           = testFunc[fNdx];
5507                         specializations["NAME"]         = cases[ndx].param;
5508                         spec.assembly                           = shaderTemplate.specialize(specializations);
5509                         spec.numWorkGroups                      = IVec3(numElements, 1, 1);
5510                         spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
5511                         spec.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
5512
5513                         entryMainGroup->addChild(new SpvAsmComputeShaderCase(testCtx, (testFunc[fNdx] + cases[ndx].name).c_str(), cases[ndx].name, spec));
5514                 }
5515         }
5516
5517         cases.push_back(CaseParameter("_is_entry", "rdc"));
5518
5519         for(size_t fNdx = 0; fNdx < testFunc.size(); ++fNdx)
5520         {
5521                 for(size_t ndx = 0; ndx < cases.size(); ++ndx)
5522                 {
5523                         map<string, string>     specializations;
5524                         ComputeShaderSpec       spec;
5525
5526                         specializations["ENTRY"]        = "rdc";
5527                         specializations["ID"]           = testFunc[fNdx];
5528                         specializations["NAME"]         = cases[ndx].param;
5529                         spec.assembly                           = shaderTemplate.specialize(specializations);
5530                         spec.numWorkGroups                      = IVec3(numElements, 1, 1);
5531                         spec.entryPoint                         = "rdc";
5532                         spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
5533                         spec.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
5534
5535                         entryNotGroup->addChild(new SpvAsmComputeShaderCase(testCtx, (testFunc[fNdx] + cases[ndx].name).c_str(), cases[ndx].name, spec));
5536                 }
5537         }
5538
5539         group->addChild(entryMainGroup.release());
5540         group->addChild(entryNotGroup.release());
5541         group->addChild(abuseGroup.release());
5542
5543         return group.release();
5544 }
5545
5546 tcu::TestCaseGroup* createOpMemberNameGroup (tcu::TestContext& testCtx)
5547 {
5548         de::MovePtr<tcu::TestCaseGroup> group(new tcu::TestCaseGroup(testCtx, "opmembername", "Tests OpMemberName cases"));
5549         de::MovePtr<tcu::TestCaseGroup> abuseGroup(new tcu::TestCaseGroup(testCtx, "abuse", "OpMemberName abuse tests"));
5550         vector<CaseParameter>                   abuseCases;
5551         vector<string>                                  testFunc;
5552         de::Random                                              rnd(deStringHash(group->getName()));
5553         const int                                               numElements = 128;
5554         vector<float>                                   inputFloats(numElements, 0);
5555         vector<float>                                   outputFloats(numElements, 0);
5556
5557         getOpNameAbuseCases(abuseCases);
5558
5559         fillRandomScalars(rnd, -100.0f, 100.0f, &inputFloats[0], numElements);
5560
5561         for (size_t ndx = 0; ndx < numElements; ++ndx)
5562                 outputFloats[ndx] = -inputFloats[ndx];
5563
5564         const string commonShaderHeader =
5565                 "OpCapability Shader\n"
5566                 "OpMemoryModel Logical GLSL450\n"
5567                 "OpEntryPoint GLCompute %main \"main\" %id\n"
5568                 "OpExecutionMode %main LocalSize 1 1 1\n";
5569
5570         const string commonShaderFooter =
5571                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
5572
5573                 + string(getComputeAsmInputOutputBufferTraits())
5574                 + string(getComputeAsmCommonTypes())
5575                 + string(getComputeAsmInputOutputBuffer()) +
5576
5577                 "%u3str     = OpTypeStruct %u32 %u32 %u32\n"
5578
5579                 "%id        = OpVariable %uvec3ptr Input\n"
5580                 "%zero      = OpConstant %i32 0\n"
5581
5582                 "%main      = OpFunction %void None %voidf\n"
5583                 "%entry     = OpLabel\n"
5584
5585                 "%idval     = OpLoad %uvec3 %id\n"
5586                 "%x0        = OpCompositeExtract %u32 %idval 0\n"
5587
5588                 "%idstr     = OpCompositeConstruct %u3str %x0 %x0 %x0\n"
5589                 "%x         = OpCompositeExtract %u32 %idstr 0\n"
5590
5591                 "%inloc     = OpAccessChain %f32ptr %indata %zero %x\n"
5592                 "%inval     = OpLoad %f32 %inloc\n"
5593                 "%neg       = OpFNegate %f32 %inval\n"
5594                 "%outloc    = OpAccessChain %f32ptr %outdata %zero %x\n"
5595                 "             OpStore %outloc %neg\n"
5596
5597                 "             OpReturn\n"
5598                 "             OpFunctionEnd\n";
5599
5600         const StringTemplate shaderTemplate(
5601                 commonShaderHeader +
5602                 "OpMemberName %u3str 0 \"${NAME}\"\n" +
5603                 commonShaderFooter);
5604
5605         const std::string multipleNames =
5606                 commonShaderHeader +
5607                 "OpMemberName %u3str 0 \"to_be\"\n"
5608                 "OpMemberName %u3str 1 \"or_not\"\n"
5609                 "OpMemberName %u3str 0 \"to_be\"\n"
5610                 "OpMemberName %u3str 2 \"makes_no\"\n"
5611                 "OpMemberName %u3str 0 \"difference\"\n"
5612                 "OpMemberName %u3str 0 \"to_me\"\n" +
5613                 commonShaderFooter;
5614         {
5615                 ComputeShaderSpec       spec;
5616
5617                 spec.assembly = multipleNames;
5618                 spec.numWorkGroups = IVec3(numElements, 1, 1);
5619                 spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
5620                 spec.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
5621
5622                 abuseGroup->addChild(new SpvAsmComputeShaderCase(testCtx, "u3str_x_has_multiple_names", "multiple_names", spec));
5623         }
5624
5625         const std::string everythingNamedTheSame =
5626                 commonShaderHeader +
5627                 "OpMemberName %u3str 0 \"the_same\"\n"
5628                 "OpMemberName %u3str 1 \"the_same\"\n"
5629                 "OpMemberName %u3str 2 \"the_same\"\n" +
5630                 commonShaderFooter;
5631
5632         {
5633                 ComputeShaderSpec       spec;
5634
5635                 spec.assembly = everythingNamedTheSame;
5636                 spec.numWorkGroups = IVec3(numElements, 1, 1);
5637                 spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
5638                 spec.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
5639
5640                 abuseGroup->addChild(new SpvAsmComputeShaderCase(testCtx, "everything_named_the_same", "everything_named_the_same", spec));
5641         }
5642
5643         // u3str_x_is_....
5644         for (size_t ndx = 0; ndx < abuseCases.size(); ++ndx)
5645         {
5646                 map<string, string>     specializations;
5647                 ComputeShaderSpec       spec;
5648
5649                 specializations["NAME"] = abuseCases[ndx].param;
5650                 spec.assembly = shaderTemplate.specialize(specializations);
5651                 spec.numWorkGroups = IVec3(numElements, 1, 1);
5652                 spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
5653                 spec.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
5654
5655                 abuseGroup->addChild(new SpvAsmComputeShaderCase(testCtx, (std::string("u3str_x") + abuseCases[ndx].name).c_str(), abuseCases[ndx].name, spec));
5656         }
5657
5658         group->addChild(abuseGroup.release());
5659
5660         return group.release();
5661 }
5662
5663 // Assembly code used for testing function control is based on GLSL source code:
5664 //
5665 // #version 430
5666 //
5667 // layout(std140, set = 0, binding = 0) readonly buffer Input {
5668 //   float elements[];
5669 // } input_data;
5670 // layout(std140, set = 0, binding = 1) writeonly buffer Output {
5671 //   float elements[];
5672 // } output_data;
5673 //
5674 // float const10() { return 10.f; }
5675 //
5676 // void main() {
5677 //   uint x = gl_GlobalInvocationID.x;
5678 //   output_data.elements[x] = input_data.elements[x] + const10();
5679 // }
5680 tcu::TestCaseGroup* createFunctionControlGroup (tcu::TestContext& testCtx)
5681 {
5682         de::MovePtr<tcu::TestCaseGroup> group                   (new tcu::TestCaseGroup(testCtx, "function_control", "Tests function control cases"));
5683         vector<CaseParameter>                   cases;
5684         de::Random                                              rnd                             (deStringHash(group->getName()));
5685         const int                                               numElements             = 100;
5686         vector<float>                                   inputFloats             (numElements, 0);
5687         vector<float>                                   outputFloats    (numElements, 0);
5688         const StringTemplate                    shaderTemplate  (
5689                 string(getComputeAsmShaderPreamble()) +
5690
5691                 "OpSource GLSL 430\n"
5692                 "OpName %main \"main\"\n"
5693                 "OpName %func_const10 \"const10(\"\n"
5694                 "OpName %id \"gl_GlobalInvocationID\"\n"
5695
5696                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
5697
5698                 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) + string(getComputeAsmInputOutputBuffer()) +
5699
5700                 "%f32f = OpTypeFunction %f32\n"
5701                 "%id = OpVariable %uvec3ptr Input\n"
5702                 "%zero = OpConstant %i32 0\n"
5703                 "%constf10 = OpConstant %f32 10.0\n"
5704
5705                 "%main         = OpFunction %void None %voidf\n"
5706                 "%entry        = OpLabel\n"
5707                 "%idval        = OpLoad %uvec3 %id\n"
5708                 "%x            = OpCompositeExtract %u32 %idval 0\n"
5709                 "%inloc        = OpAccessChain %f32ptr %indata %zero %x\n"
5710                 "%inval        = OpLoad %f32 %inloc\n"
5711                 "%ret_10       = OpFunctionCall %f32 %func_const10\n"
5712                 "%fadd         = OpFAdd %f32 %inval %ret_10\n"
5713                 "%outloc       = OpAccessChain %f32ptr %outdata %zero %x\n"
5714                 "                OpStore %outloc %fadd\n"
5715                 "                OpReturn\n"
5716                 "                OpFunctionEnd\n"
5717
5718                 "%func_const10 = OpFunction %f32 ${CONTROL} %f32f\n"
5719                 "%label        = OpLabel\n"
5720                 "                OpReturnValue %constf10\n"
5721                 "                OpFunctionEnd\n");
5722
5723         cases.push_back(CaseParameter("none",                                           "None"));
5724         cases.push_back(CaseParameter("inline",                                         "Inline"));
5725         cases.push_back(CaseParameter("dont_inline",                            "DontInline"));
5726         cases.push_back(CaseParameter("pure",                                           "Pure"));
5727         cases.push_back(CaseParameter("const",                                          "Const"));
5728         cases.push_back(CaseParameter("inline_pure",                            "Inline|Pure"));
5729         cases.push_back(CaseParameter("const_dont_inline",                      "Const|DontInline"));
5730         cases.push_back(CaseParameter("inline_dont_inline",                     "Inline|DontInline"));
5731         cases.push_back(CaseParameter("pure_inline_dont_inline",        "Pure|Inline|DontInline"));
5732
5733         fillRandomScalars(rnd, -100.f, 100.f, &inputFloats[0], numElements);
5734
5735         // CPU might not use the same rounding mode as the GPU. Use whole numbers to avoid rounding differences.
5736         floorAll(inputFloats);
5737
5738         for (size_t ndx = 0; ndx < numElements; ++ndx)
5739                 outputFloats[ndx] = inputFloats[ndx] + 10.f;
5740
5741         for (size_t caseNdx = 0; caseNdx < cases.size(); ++caseNdx)
5742         {
5743                 map<string, string>             specializations;
5744                 ComputeShaderSpec               spec;
5745
5746                 specializations["CONTROL"] = cases[caseNdx].param;
5747                 spec.assembly = shaderTemplate.specialize(specializations);
5748                 spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
5749                 spec.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
5750                 spec.numWorkGroups = IVec3(numElements, 1, 1);
5751
5752                 group->addChild(new SpvAsmComputeShaderCase(testCtx, cases[caseNdx].name, cases[caseNdx].name, spec));
5753         }
5754
5755         return group.release();
5756 }
5757
5758 tcu::TestCaseGroup* createMemoryAccessGroup (tcu::TestContext& testCtx)
5759 {
5760         de::MovePtr<tcu::TestCaseGroup> group                   (new tcu::TestCaseGroup(testCtx, "memory_access", "Tests memory access cases"));
5761         vector<CaseParameter>                   cases;
5762         de::Random                                              rnd                             (deStringHash(group->getName()));
5763         const int                                               numElements             = 100;
5764         vector<float>                                   inputFloats             (numElements, 0);
5765         vector<float>                                   outputFloats    (numElements, 0);
5766         const StringTemplate                    shaderTemplate  (
5767                 string(getComputeAsmShaderPreamble()) +
5768
5769                 "OpSource GLSL 430\n"
5770                 "OpName %main           \"main\"\n"
5771                 "OpName %id             \"gl_GlobalInvocationID\"\n"
5772
5773                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
5774
5775                 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) + string(getComputeAsmInputOutputBuffer()) +
5776
5777                 "%f32ptr_f  = OpTypePointer Function %f32\n"
5778
5779                 "%id        = OpVariable %uvec3ptr Input\n"
5780                 "%zero      = OpConstant %i32 0\n"
5781                 "%four      = OpConstant %i32 4\n"
5782
5783                 "%main      = OpFunction %void None %voidf\n"
5784                 "%label     = OpLabel\n"
5785                 "%copy      = OpVariable %f32ptr_f Function\n"
5786                 "%idval     = OpLoad %uvec3 %id ${ACCESS}\n"
5787                 "%x         = OpCompositeExtract %u32 %idval 0\n"
5788                 "%inloc     = OpAccessChain %f32ptr %indata  %zero %x\n"
5789                 "%outloc    = OpAccessChain %f32ptr %outdata %zero %x\n"
5790                 "             OpCopyMemory %copy %inloc ${ACCESS}\n"
5791                 "%val1      = OpLoad %f32 %copy\n"
5792                 "%val2      = OpLoad %f32 %inloc\n"
5793                 "%add       = OpFAdd %f32 %val1 %val2\n"
5794                 "             OpStore %outloc %add ${ACCESS}\n"
5795                 "             OpReturn\n"
5796                 "             OpFunctionEnd\n");
5797
5798         cases.push_back(CaseParameter("null",                                   ""));
5799         cases.push_back(CaseParameter("none",                                   "None"));
5800         cases.push_back(CaseParameter("volatile",                               "Volatile"));
5801         cases.push_back(CaseParameter("aligned",                                "Aligned 4"));
5802         cases.push_back(CaseParameter("nontemporal",                    "Nontemporal"));
5803         cases.push_back(CaseParameter("aligned_nontemporal",    "Aligned|Nontemporal 4"));
5804         cases.push_back(CaseParameter("aligned_volatile",               "Volatile|Aligned 4"));
5805
5806         fillRandomScalars(rnd, -100.f, 100.f, &inputFloats[0], numElements);
5807
5808         for (size_t ndx = 0; ndx < numElements; ++ndx)
5809                 outputFloats[ndx] = inputFloats[ndx] + inputFloats[ndx];
5810
5811         for (size_t caseNdx = 0; caseNdx < cases.size(); ++caseNdx)
5812         {
5813                 map<string, string>             specializations;
5814                 ComputeShaderSpec               spec;
5815
5816                 specializations["ACCESS"] = cases[caseNdx].param;
5817                 spec.assembly = shaderTemplate.specialize(specializations);
5818                 spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
5819                 spec.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
5820                 spec.numWorkGroups = IVec3(numElements, 1, 1);
5821
5822                 group->addChild(new SpvAsmComputeShaderCase(testCtx, cases[caseNdx].name, cases[caseNdx].name, spec));
5823         }
5824
5825         return group.release();
5826 }
5827
5828 // Checks that we can get undefined values for various types, without exercising a computation with it.
5829 tcu::TestCaseGroup* createOpUndefGroup (tcu::TestContext& testCtx)
5830 {
5831         de::MovePtr<tcu::TestCaseGroup> group                   (new tcu::TestCaseGroup(testCtx, "opundef", "Tests the OpUndef instruction"));
5832         vector<CaseParameter>                   cases;
5833         de::Random                                              rnd                             (deStringHash(group->getName()));
5834         const int                                               numElements             = 100;
5835         vector<float>                                   positiveFloats  (numElements, 0);
5836         vector<float>                                   negativeFloats  (numElements, 0);
5837         const StringTemplate                    shaderTemplate  (
5838                 string(getComputeAsmShaderPreamble()) +
5839
5840                 "OpSource GLSL 430\n"
5841                 "OpName %main           \"main\"\n"
5842                 "OpName %id             \"gl_GlobalInvocationID\"\n"
5843
5844                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
5845
5846                 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) +
5847                 "%uvec2     = OpTypeVector %u32 2\n"
5848                 "%fvec4     = OpTypeVector %f32 4\n"
5849                 "%fmat33    = OpTypeMatrix %fvec3 3\n"
5850                 "%image     = OpTypeImage %f32 2D 0 0 0 1 Unknown\n"
5851                 "%sampler   = OpTypeSampler\n"
5852                 "%simage    = OpTypeSampledImage %image\n"
5853                 "%const100  = OpConstant %u32 100\n"
5854                 "%uarr100   = OpTypeArray %i32 %const100\n"
5855                 "%struct    = OpTypeStruct %f32 %i32 %u32\n"
5856                 "%pointer   = OpTypePointer Function %i32\n"
5857                 + string(getComputeAsmInputOutputBuffer()) +
5858
5859                 "%id        = OpVariable %uvec3ptr Input\n"
5860                 "%zero      = OpConstant %i32 0\n"
5861
5862                 "%main      = OpFunction %void None %voidf\n"
5863                 "%label     = OpLabel\n"
5864
5865                 "%undef     = OpUndef ${TYPE}\n"
5866
5867                 "%idval     = OpLoad %uvec3 %id\n"
5868                 "%x         = OpCompositeExtract %u32 %idval 0\n"
5869
5870                 "%inloc     = OpAccessChain %f32ptr %indata %zero %x\n"
5871                 "%inval     = OpLoad %f32 %inloc\n"
5872                 "%neg       = OpFNegate %f32 %inval\n"
5873                 "%outloc    = OpAccessChain %f32ptr %outdata %zero %x\n"
5874                 "             OpStore %outloc %neg\n"
5875                 "             OpReturn\n"
5876                 "             OpFunctionEnd\n");
5877
5878         cases.push_back(CaseParameter("bool",                   "%bool"));
5879         cases.push_back(CaseParameter("sint32",                 "%i32"));
5880         cases.push_back(CaseParameter("uint32",                 "%u32"));
5881         cases.push_back(CaseParameter("float32",                "%f32"));
5882         cases.push_back(CaseParameter("vec4float32",    "%fvec4"));
5883         cases.push_back(CaseParameter("vec2uint32",             "%uvec2"));
5884         cases.push_back(CaseParameter("matrix",                 "%fmat33"));
5885         cases.push_back(CaseParameter("image",                  "%image"));
5886         cases.push_back(CaseParameter("sampler",                "%sampler"));
5887         cases.push_back(CaseParameter("sampledimage",   "%simage"));
5888         cases.push_back(CaseParameter("array",                  "%uarr100"));
5889         cases.push_back(CaseParameter("runtimearray",   "%f32arr"));
5890         cases.push_back(CaseParameter("struct",                 "%struct"));
5891         cases.push_back(CaseParameter("pointer",                "%pointer"));
5892
5893         fillRandomScalars(rnd, 1.f, 100.f, &positiveFloats[0], numElements);
5894
5895         for (size_t ndx = 0; ndx < numElements; ++ndx)
5896                 negativeFloats[ndx] = -positiveFloats[ndx];
5897
5898         for (size_t caseNdx = 0; caseNdx < cases.size(); ++caseNdx)
5899         {
5900                 map<string, string>             specializations;
5901                 ComputeShaderSpec               spec;
5902
5903                 specializations["TYPE"] = cases[caseNdx].param;
5904                 spec.assembly = shaderTemplate.specialize(specializations);
5905                 spec.inputs.push_back(BufferSp(new Float32Buffer(positiveFloats)));
5906                 spec.outputs.push_back(BufferSp(new Float32Buffer(negativeFloats)));
5907                 spec.numWorkGroups = IVec3(numElements, 1, 1);
5908
5909                 group->addChild(new SpvAsmComputeShaderCase(testCtx, cases[caseNdx].name, cases[caseNdx].name, spec));
5910         }
5911
5912                 return group.release();
5913 }
5914
5915 // Checks that a compute shader can generate a constant composite value of various types, without exercising a computation on it.
5916 tcu::TestCaseGroup* createFloat16OpConstantCompositeGroup (tcu::TestContext& testCtx)
5917 {
5918         de::MovePtr<tcu::TestCaseGroup> group                   (new tcu::TestCaseGroup(testCtx, "opconstantcomposite", "Tests the OpConstantComposite instruction"));
5919         vector<CaseParameter>                   cases;
5920         de::Random                                              rnd                             (deStringHash(group->getName()));
5921         const int                                               numElements             = 100;
5922         vector<float>                                   positiveFloats  (numElements, 0);
5923         vector<float>                                   negativeFloats  (numElements, 0);
5924         const StringTemplate                    shaderTemplate  (
5925                 "OpCapability Shader\n"
5926                 "OpCapability Float16\n"
5927                 "OpMemoryModel Logical GLSL450\n"
5928                 "OpEntryPoint GLCompute %main \"main\" %id\n"
5929                 "OpExecutionMode %main LocalSize 1 1 1\n"
5930                 "OpSource GLSL 430\n"
5931                 "OpName %main           \"main\"\n"
5932                 "OpName %id             \"gl_GlobalInvocationID\"\n"
5933
5934                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
5935
5936                 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) + string(getComputeAsmInputOutputBuffer()) +
5937
5938                 "%id        = OpVariable %uvec3ptr Input\n"
5939                 "%zero      = OpConstant %i32 0\n"
5940                 "%f16       = OpTypeFloat 16\n"
5941                 "%c_f16_0   = OpConstant %f16 0.0\n"
5942                 "%c_f16_0_5 = OpConstant %f16 0.5\n"
5943                 "%c_f16_1   = OpConstant %f16 1.0\n"
5944                 "%v2f16     = OpTypeVector %f16 2\n"
5945                 "%v3f16     = OpTypeVector %f16 3\n"
5946                 "%v4f16     = OpTypeVector %f16 4\n"
5947
5948                 "${CONSTANT}\n"
5949
5950                 "%main      = OpFunction %void None %voidf\n"
5951                 "%label     = OpLabel\n"
5952                 "%idval     = OpLoad %uvec3 %id\n"
5953                 "%x         = OpCompositeExtract %u32 %idval 0\n"
5954                 "%inloc     = OpAccessChain %f32ptr %indata %zero %x\n"
5955                 "%inval     = OpLoad %f32 %inloc\n"
5956                 "%neg       = OpFNegate %f32 %inval\n"
5957                 "%outloc    = OpAccessChain %f32ptr %outdata %zero %x\n"
5958                 "             OpStore %outloc %neg\n"
5959                 "             OpReturn\n"
5960                 "             OpFunctionEnd\n");
5961
5962
5963         cases.push_back(CaseParameter("vector",                 "%const = OpConstantComposite %v3f16 %c_f16_0 %c_f16_0_5 %c_f16_1\n"));
5964         cases.push_back(CaseParameter("matrix",                 "%m3v3f16 = OpTypeMatrix %v3f16 3\n"
5965                                                                                                         "%vec = OpConstantComposite %v3f16 %c_f16_0 %c_f16_0_5 %c_f16_1\n"
5966                                                                                                         "%mat = OpConstantComposite %m3v3f16 %vec %vec %vec"));
5967         cases.push_back(CaseParameter("struct",                 "%m2v3f16 = OpTypeMatrix %v3f16 2\n"
5968                                                                                                         "%struct = OpTypeStruct %i32 %f16 %v3f16 %m2v3f16\n"
5969                                                                                                         "%vec = OpConstantComposite %v3f16 %c_f16_0 %c_f16_0_5 %c_f16_1\n"
5970                                                                                                         "%mat = OpConstantComposite %m2v3f16 %vec %vec\n"
5971                                                                                                         "%const = OpConstantComposite %struct %zero %c_f16_0_5 %vec %mat\n"));
5972         cases.push_back(CaseParameter("nested_struct",  "%st1 = OpTypeStruct %i32 %f16\n"
5973                                                                                                         "%st2 = OpTypeStruct %i32 %i32\n"
5974                                                                                                         "%struct = OpTypeStruct %st1 %st2\n"
5975                                                                                                         "%st1val = OpConstantComposite %st1 %zero %c_f16_0_5\n"
5976                                                                                                         "%st2val = OpConstantComposite %st2 %zero %zero\n"
5977                                                                                                         "%const = OpConstantComposite %struct %st1val %st2val"));
5978
5979         fillRandomScalars(rnd, 1.f, 100.f, &positiveFloats[0], numElements);
5980
5981         for (size_t ndx = 0; ndx < numElements; ++ndx)
5982                 negativeFloats[ndx] = -positiveFloats[ndx];
5983
5984         for (size_t caseNdx = 0; caseNdx < cases.size(); ++caseNdx)
5985         {
5986                 map<string, string>             specializations;
5987                 ComputeShaderSpec               spec;
5988
5989                 specializations["CONSTANT"] = cases[caseNdx].param;
5990                 spec.assembly = shaderTemplate.specialize(specializations);
5991                 spec.inputs.push_back(BufferSp(new Float32Buffer(positiveFloats)));
5992                 spec.outputs.push_back(BufferSp(new Float32Buffer(negativeFloats)));
5993                 spec.numWorkGroups = IVec3(numElements, 1, 1);
5994
5995                 spec.extensions.push_back("VK_KHR_16bit_storage");
5996                 spec.extensions.push_back("VK_KHR_shader_float16_int8");
5997
5998                 spec.requestedVulkanFeatures.ext16BitStorage = EXT16BITSTORAGEFEATURES_UNIFORM_BUFFER_BLOCK;
5999                 spec.requestedVulkanFeatures.extFloat16Int8 = EXTFLOAT16INT8FEATURES_FLOAT16;
6000
6001                 group->addChild(new SpvAsmComputeShaderCase(testCtx, cases[caseNdx].name, cases[caseNdx].name, spec));
6002         }
6003
6004         return group.release();
6005 }
6006
6007 const vector<deFloat16> squarize(const vector<deFloat16>& inData, const deUint32 argNo)
6008 {
6009         const size_t            inDataLength    = inData.size();
6010         vector<deFloat16>       result;
6011
6012         result.reserve(inDataLength * inDataLength);
6013
6014         if (argNo == 0)
6015         {
6016                 for (size_t numIdx = 0; numIdx < inDataLength; ++numIdx)
6017                         result.insert(result.end(), inData.begin(), inData.end());
6018         }
6019
6020         if (argNo == 1)
6021         {
6022                 for (size_t numIdx = 0; numIdx < inDataLength; ++numIdx)
6023                 {
6024                         const vector<deFloat16> tmp(inDataLength, inData[numIdx]);
6025
6026                         result.insert(result.end(), tmp.begin(), tmp.end());
6027                 }
6028         }
6029
6030         return result;
6031 }
6032
6033 const vector<deFloat16> squarizeVector(const vector<deFloat16>& inData, const deUint32 argNo)
6034 {
6035         vector<deFloat16>       vec;
6036         vector<deFloat16>       result;
6037
6038         // Create vectors. vec will contain each possible pair from inData
6039         {
6040                 const size_t    inDataLength    = inData.size();
6041
6042                 DE_ASSERT(inDataLength <= 64);
6043
6044                 vec.reserve(2 * inDataLength * inDataLength);
6045
6046                 for (size_t numIdxX = 0; numIdxX < inDataLength; ++numIdxX)
6047                 for (size_t numIdxY = 0; numIdxY < inDataLength; ++numIdxY)
6048                 {
6049                         vec.push_back(inData[numIdxX]);
6050                         vec.push_back(inData[numIdxY]);
6051                 }
6052         }
6053
6054         // Create vector pairs. result will contain each possible pair from vec
6055         {
6056                 const size_t    coordsPerVector = 2;
6057                 const size_t    vectorsCount    = vec.size() / coordsPerVector;
6058
6059                 result.reserve(coordsPerVector * vectorsCount * vectorsCount);
6060
6061                 if (argNo == 0)
6062                 {
6063                         for (size_t numIdxX = 0; numIdxX < vectorsCount; ++numIdxX)
6064                         for (size_t numIdxY = 0; numIdxY < vectorsCount; ++numIdxY)
6065                         {
6066                                 for (size_t coordNdx = 0; coordNdx < coordsPerVector; ++coordNdx)
6067                                         result.push_back(vec[coordsPerVector * numIdxY + coordNdx]);
6068                         }
6069                 }
6070
6071                 if (argNo == 1)
6072                 {
6073                         for (size_t numIdxX = 0; numIdxX < vectorsCount; ++numIdxX)
6074                         for (size_t numIdxY = 0; numIdxY < vectorsCount; ++numIdxY)
6075                         {
6076                                 for (size_t coordNdx = 0; coordNdx < coordsPerVector; ++coordNdx)
6077                                         result.push_back(vec[coordsPerVector * numIdxX + coordNdx]);
6078                         }
6079                 }
6080         }
6081
6082         return result;
6083 }
6084
6085 struct fp16isNan                        { bool operator()(const tcu::Float16 in1, const tcu::Float16)           { return in1.isNaN(); } };
6086 struct fp16isInf                        { bool operator()(const tcu::Float16 in1, const tcu::Float16)           { return in1.isInf(); } };
6087 struct fp16isEqual                      { bool operator()(const tcu::Float16 in1, const tcu::Float16 in2)       { return in1.asFloat() == in2.asFloat(); } };
6088 struct fp16isUnequal            { bool operator()(const tcu::Float16 in1, const tcu::Float16 in2)       { return in1.asFloat() != in2.asFloat(); } };
6089 struct fp16isLess                       { bool operator()(const tcu::Float16 in1, const tcu::Float16 in2)       { return in1.asFloat() <  in2.asFloat(); } };
6090 struct fp16isGreater            { bool operator()(const tcu::Float16 in1, const tcu::Float16 in2)       { return in1.asFloat() >  in2.asFloat(); } };
6091 struct fp16isLessOrEqual        { bool operator()(const tcu::Float16 in1, const tcu::Float16 in2)       { return in1.asFloat() <= in2.asFloat(); } };
6092 struct fp16isGreaterOrEqual     { bool operator()(const tcu::Float16 in1, const tcu::Float16 in2)       { return in1.asFloat() >= in2.asFloat(); } };
6093
6094 template <class TestedLogicalFunction, bool onlyTestFunc, bool unationModeAnd, bool nanSupported>
6095 bool compareFP16Logical (const std::vector<Resource>& inputs, const vector<AllocationSp>& outputAllocs, const std::vector<Resource>&, TestLog& log)
6096 {
6097         if (inputs.size() != 2 || outputAllocs.size() != 1)
6098                 return false;
6099
6100         vector<deUint8> input1Bytes;
6101         vector<deUint8> input2Bytes;
6102
6103         inputs[0].getBytes(input1Bytes);
6104         inputs[1].getBytes(input2Bytes);
6105
6106         const deUint32                  denormModesCount                        = 2;
6107         const deFloat16                 float16one                                      = tcu::Float16(1.0f).bits();
6108         const deFloat16                 float16zero                                     = tcu::Float16(0.0f).bits();
6109         const tcu::Float16              zero                                            = tcu::Float16::zero(1);
6110         const deFloat16* const  outputAsFP16                            = static_cast<deFloat16*>(outputAllocs[0]->getHostPtr());
6111         const deFloat16* const  input1AsFP16                            = reinterpret_cast<deFloat16* const>(&input1Bytes.front());
6112         const deFloat16* const  input2AsFP16                            = reinterpret_cast<deFloat16* const>(&input2Bytes.front());
6113         deUint32                                successfulRuns                          = denormModesCount;
6114         std::string                             results[denormModesCount];
6115         TestedLogicalFunction   testedLogicalFunction;
6116
6117         for (deUint32 denormMode = 0; denormMode < denormModesCount; denormMode++)
6118         {
6119                 const bool flushToZero = (denormMode == 1);
6120
6121                 for (size_t idx = 0; idx < input1Bytes.size() / sizeof(deFloat16); ++idx)
6122                 {
6123                         const tcu::Float16      f1pre                   = tcu::Float16(input1AsFP16[idx]);
6124                         const tcu::Float16      f2pre                   = tcu::Float16(input2AsFP16[idx]);
6125                         const tcu::Float16      f1                              = (flushToZero && f1pre.isDenorm()) ? zero : f1pre;
6126                         const tcu::Float16      f2                              = (flushToZero && f2pre.isDenorm()) ? zero : f2pre;
6127                         deFloat16                       expectedOutput  = float16zero;
6128
6129                         if (onlyTestFunc)
6130                         {
6131                                 if (testedLogicalFunction(f1, f2))
6132                                         expectedOutput = float16one;
6133                         }
6134                         else
6135                         {
6136                                 const bool      f1nan   = f1.isNaN();
6137                                 const bool      f2nan   = f2.isNaN();
6138
6139                                 // Skip NaN floats if not supported by implementation
6140                                 if (!nanSupported && (f1nan || f2nan))
6141                                         continue;
6142
6143                                 if (unationModeAnd)
6144                                 {
6145                                         const bool      ordered         = !f1nan && !f2nan;
6146
6147                                         if (ordered && testedLogicalFunction(f1, f2))
6148                                                 expectedOutput = float16one;
6149                                 }
6150                                 else
6151                                 {
6152                                         const bool      unordered       = f1nan || f2nan;
6153
6154                                         if (unordered || testedLogicalFunction(f1, f2))
6155                                                 expectedOutput = float16one;
6156                                 }
6157                         }
6158
6159                         if (outputAsFP16[idx] != expectedOutput)
6160                         {
6161                                 std::ostringstream str;
6162
6163                                 str << "ERROR: Sub-case #" << idx
6164                                         << " flushToZero:" << flushToZero
6165                                         << std::hex
6166                                         << " failed, inputs: 0x" << f1.bits()
6167                                         << ";0x" << f2.bits()
6168                                         << " output: 0x" << outputAsFP16[idx]
6169                                         << " expected output: 0x" << expectedOutput;
6170
6171                                 results[denormMode] = str.str();
6172
6173                                 successfulRuns--;
6174
6175                                 break;
6176                         }
6177                 }
6178         }
6179
6180         if (successfulRuns == 0)
6181                 for (deUint32 denormMode = 0; denormMode < denormModesCount; denormMode++)
6182                         log << TestLog::Message << results[denormMode] << TestLog::EndMessage;
6183
6184         return successfulRuns > 0;
6185 }
6186
6187 } // anonymous
6188
6189 tcu::TestCaseGroup* createOpSourceTests (tcu::TestContext& testCtx)
6190 {
6191         struct NameCodePair { string name, code; };
6192         RGBA                                                    defaultColors[4];
6193         de::MovePtr<tcu::TestCaseGroup> opSourceTests                   (new tcu::TestCaseGroup(testCtx, "opsource", "OpSource instruction"));
6194         const std::string                               opsourceGLSLWithFile    = "%opsrcfile = OpString \"foo.vert\"\nOpSource GLSL 450 %opsrcfile ";
6195         map<string, string>                             fragments                               = passthruFragments();
6196         const NameCodePair                              tests[]                                 =
6197         {
6198                 {"unknown", "OpSource Unknown 321"},
6199                 {"essl", "OpSource ESSL 310"},
6200                 {"glsl", "OpSource GLSL 450"},
6201                 {"opencl_cpp", "OpSource OpenCL_CPP 120"},
6202                 {"opencl_c", "OpSource OpenCL_C 120"},
6203                 {"multiple", "OpSource GLSL 450\nOpSource GLSL 450"},
6204                 {"file", opsourceGLSLWithFile},
6205                 {"source", opsourceGLSLWithFile + "\"void main(){}\""},
6206                 // Longest possible source string: SPIR-V limits instructions to 65535
6207                 // words, of which the first 4 are opsourceGLSLWithFile; the rest will
6208                 // contain 65530 UTF8 characters (one word each) plus one last word
6209                 // containing 3 ASCII characters and \0.
6210                 {"longsource", opsourceGLSLWithFile + '"' + makeLongUTF8String(65530) + "ccc" + '"'}
6211         };
6212
6213         getDefaultColors(defaultColors);
6214         for (size_t testNdx = 0; testNdx < sizeof(tests) / sizeof(NameCodePair); ++testNdx)
6215         {
6216                 fragments["debug"] = tests[testNdx].code;
6217                 createTestsForAllStages(tests[testNdx].name, defaultColors, defaultColors, fragments, opSourceTests.get());
6218         }
6219
6220         return opSourceTests.release();
6221 }
6222
6223 tcu::TestCaseGroup* createOpSourceContinuedTests (tcu::TestContext& testCtx)
6224 {
6225         struct NameCodePair { string name, code; };
6226         RGBA                                                            defaultColors[4];
6227         de::MovePtr<tcu::TestCaseGroup>         opSourceTests           (new tcu::TestCaseGroup(testCtx, "opsourcecontinued", "OpSourceContinued instruction"));
6228         map<string, string>                                     fragments                       = passthruFragments();
6229         const std::string                                       opsource                        = "%opsrcfile = OpString \"foo.vert\"\nOpSource GLSL 450 %opsrcfile \"void main(){}\"\n";
6230         const NameCodePair                                      tests[]                         =
6231         {
6232                 {"empty", opsource + "OpSourceContinued \"\""},
6233                 {"short", opsource + "OpSourceContinued \"abcde\""},
6234                 {"multiple", opsource + "OpSourceContinued \"abcde\"\nOpSourceContinued \"fghij\""},
6235                 // Longest possible source string: SPIR-V limits instructions to 65535
6236                 // words, of which the first one is OpSourceContinued/length; the rest
6237                 // will contain 65533 UTF8 characters (one word each) plus one last word
6238                 // containing 3 ASCII characters and \0.
6239                 {"long", opsource + "OpSourceContinued \"" + makeLongUTF8String(65533) + "ccc\""}
6240         };
6241
6242         getDefaultColors(defaultColors);
6243         for (size_t testNdx = 0; testNdx < sizeof(tests) / sizeof(NameCodePair); ++testNdx)
6244         {
6245                 fragments["debug"] = tests[testNdx].code;
6246                 createTestsForAllStages(tests[testNdx].name, defaultColors, defaultColors, fragments, opSourceTests.get());
6247         }
6248
6249         return opSourceTests.release();
6250 }
6251 tcu::TestCaseGroup* createOpNoLineTests(tcu::TestContext& testCtx)
6252 {
6253         RGBA                                                             defaultColors[4];
6254         de::MovePtr<tcu::TestCaseGroup>          opLineTests             (new tcu::TestCaseGroup(testCtx, "opnoline", "OpNoLine instruction"));
6255         map<string, string>                                      fragments;
6256         getDefaultColors(defaultColors);
6257         fragments["debug"]                      =
6258                 "%name = OpString \"name\"\n";
6259
6260         fragments["pre_main"]   =
6261                 "OpNoLine\n"
6262                 "OpNoLine\n"
6263                 "OpLine %name 1 1\n"
6264                 "OpNoLine\n"
6265                 "OpLine %name 1 1\n"
6266                 "OpLine %name 1 1\n"
6267                 "%second_function = OpFunction %v4f32 None %v4f32_v4f32_function\n"
6268                 "OpNoLine\n"
6269                 "OpLine %name 1 1\n"
6270                 "OpNoLine\n"
6271                 "OpLine %name 1 1\n"
6272                 "OpLine %name 1 1\n"
6273                 "%second_param1 = OpFunctionParameter %v4f32\n"
6274                 "OpNoLine\n"
6275                 "OpNoLine\n"
6276                 "%label_secondfunction = OpLabel\n"
6277                 "OpNoLine\n"
6278                 "OpReturnValue %second_param1\n"
6279                 "OpFunctionEnd\n"
6280                 "OpNoLine\n"
6281                 "OpNoLine\n";
6282
6283         fragments["testfun"]            =
6284                 // A %test_code function that returns its argument unchanged.
6285                 "OpNoLine\n"
6286                 "OpNoLine\n"
6287                 "OpLine %name 1 1\n"
6288                 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
6289                 "OpNoLine\n"
6290                 "%param1 = OpFunctionParameter %v4f32\n"
6291                 "OpNoLine\n"
6292                 "OpNoLine\n"
6293                 "%label_testfun = OpLabel\n"
6294                 "OpNoLine\n"
6295                 "%val1 = OpFunctionCall %v4f32 %second_function %param1\n"
6296                 "OpReturnValue %val1\n"
6297                 "OpFunctionEnd\n"
6298                 "OpLine %name 1 1\n"
6299                 "OpNoLine\n";
6300
6301         createTestsForAllStages("opnoline", defaultColors, defaultColors, fragments, opLineTests.get());
6302
6303         return opLineTests.release();
6304 }
6305
6306 tcu::TestCaseGroup* createOpModuleProcessedTests(tcu::TestContext& testCtx)
6307 {
6308         RGBA                                                            defaultColors[4];
6309         de::MovePtr<tcu::TestCaseGroup>         opModuleProcessedTests                  (new tcu::TestCaseGroup(testCtx, "opmoduleprocessed", "OpModuleProcessed instruction"));
6310         map<string, string>                                     fragments;
6311         std::vector<std::string>                        noExtensions;
6312         GraphicsResources                                       resources;
6313
6314         getDefaultColors(defaultColors);
6315         resources.verifyBinary = veryfiBinaryShader;
6316         resources.spirvVersion = SPIRV_VERSION_1_3;
6317
6318         fragments["moduleprocessed"]                                                    =
6319                 "OpModuleProcessed \"VULKAN CTS\"\n"
6320                 "OpModuleProcessed \"Negative values\"\n"
6321                 "OpModuleProcessed \"Date: 2017/09/21\"\n";
6322
6323         fragments["pre_main"]   =
6324                 "%second_function = OpFunction %v4f32 None %v4f32_v4f32_function\n"
6325                 "%second_param1 = OpFunctionParameter %v4f32\n"
6326                 "%label_secondfunction = OpLabel\n"
6327                 "OpReturnValue %second_param1\n"
6328                 "OpFunctionEnd\n";
6329
6330         fragments["testfun"]            =
6331                 // A %test_code function that returns its argument unchanged.
6332                 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
6333                 "%param1 = OpFunctionParameter %v4f32\n"
6334                 "%label_testfun = OpLabel\n"
6335                 "%val1 = OpFunctionCall %v4f32 %second_function %param1\n"
6336                 "OpReturnValue %val1\n"
6337                 "OpFunctionEnd\n";
6338
6339         createTestsForAllStages ("opmoduleprocessed", defaultColors, defaultColors, fragments, resources, noExtensions, opModuleProcessedTests.get());
6340
6341         return opModuleProcessedTests.release();
6342 }
6343
6344
6345 tcu::TestCaseGroup* createOpLineTests(tcu::TestContext& testCtx)
6346 {
6347         RGBA                                                                                                    defaultColors[4];
6348         de::MovePtr<tcu::TestCaseGroup>                                                 opLineTests                     (new tcu::TestCaseGroup(testCtx, "opline", "OpLine instruction"));
6349         map<string, string>                                                                             fragments;
6350         std::vector<std::pair<std::string, std::string> >               problemStrings;
6351
6352         problemStrings.push_back(std::make_pair<std::string, std::string>("empty_name", ""));
6353         problemStrings.push_back(std::make_pair<std::string, std::string>("short_name", "short_name"));
6354         problemStrings.push_back(std::make_pair<std::string, std::string>("long_name", makeLongUTF8String(65530) + "ccc"));
6355         getDefaultColors(defaultColors);
6356
6357         fragments["debug"]                      =
6358                 "%other_name = OpString \"other_name\"\n";
6359
6360         fragments["pre_main"]   =
6361                 "OpLine %file_name 32 0\n"
6362                 "OpLine %file_name 32 32\n"
6363                 "OpLine %file_name 32 40\n"
6364                 "OpLine %other_name 32 40\n"
6365                 "OpLine %other_name 0 100\n"
6366                 "OpLine %other_name 0 4294967295\n"
6367                 "OpLine %other_name 4294967295 0\n"
6368                 "OpLine %other_name 32 40\n"
6369                 "OpLine %file_name 0 0\n"
6370                 "%second_function = OpFunction %v4f32 None %v4f32_v4f32_function\n"
6371                 "OpLine %file_name 1 0\n"
6372                 "%second_param1 = OpFunctionParameter %v4f32\n"
6373                 "OpLine %file_name 1 3\n"
6374                 "OpLine %file_name 1 2\n"
6375                 "%label_secondfunction = OpLabel\n"
6376                 "OpLine %file_name 0 2\n"
6377                 "OpReturnValue %second_param1\n"
6378                 "OpFunctionEnd\n"
6379                 "OpLine %file_name 0 2\n"
6380                 "OpLine %file_name 0 2\n";
6381
6382         fragments["testfun"]            =
6383                 // A %test_code function that returns its argument unchanged.
6384                 "OpLine %file_name 1 0\n"
6385                 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
6386                 "OpLine %file_name 16 330\n"
6387                 "%param1 = OpFunctionParameter %v4f32\n"
6388                 "OpLine %file_name 14 442\n"
6389                 "%label_testfun = OpLabel\n"
6390                 "OpLine %file_name 11 1024\n"
6391                 "%val1 = OpFunctionCall %v4f32 %second_function %param1\n"
6392                 "OpLine %file_name 2 97\n"
6393                 "OpReturnValue %val1\n"
6394                 "OpFunctionEnd\n"
6395                 "OpLine %file_name 5 32\n";
6396
6397         for (size_t i = 0; i < problemStrings.size(); ++i)
6398         {
6399                 map<string, string> testFragments = fragments;
6400                 testFragments["debug"] += "%file_name = OpString \"" + problemStrings[i].second + "\"\n";
6401                 createTestsForAllStages(string("opline") + "_" + problemStrings[i].first, defaultColors, defaultColors, testFragments, opLineTests.get());
6402         }
6403
6404         return opLineTests.release();
6405 }
6406
6407 tcu::TestCaseGroup* createOpConstantNullTests(tcu::TestContext& testCtx)
6408 {
6409         de::MovePtr<tcu::TestCaseGroup> opConstantNullTests             (new tcu::TestCaseGroup(testCtx, "opconstantnull", "OpConstantNull instruction"));
6410         RGBA                                                    colors[4];
6411
6412
6413         const char                                              functionStart[] =
6414                 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
6415                 "%param1 = OpFunctionParameter %v4f32\n"
6416                 "%lbl    = OpLabel\n";
6417
6418         const char                                              functionEnd[]   =
6419                 "OpReturnValue %transformed_param\n"
6420                 "OpFunctionEnd\n";
6421
6422         struct NameConstantsCode
6423         {
6424                 string name;
6425                 string constants;
6426                 string code;
6427         };
6428
6429         NameConstantsCode tests[] =
6430         {
6431                 {
6432                         "vec4",
6433                         "%cnull = OpConstantNull %v4f32\n",
6434                         "%transformed_param = OpFAdd %v4f32 %param1 %cnull\n"
6435                 },
6436                 {
6437                         "float",
6438                         "%cnull = OpConstantNull %f32\n",
6439                         "%vp = OpVariable %fp_v4f32 Function\n"
6440                         "%v  = OpLoad %v4f32 %vp\n"
6441                         "%v0 = OpVectorInsertDynamic %v4f32 %v %cnull %c_i32_0\n"
6442                         "%v1 = OpVectorInsertDynamic %v4f32 %v0 %cnull %c_i32_1\n"
6443                         "%v2 = OpVectorInsertDynamic %v4f32 %v1 %cnull %c_i32_2\n"
6444                         "%v3 = OpVectorInsertDynamic %v4f32 %v2 %cnull %c_i32_3\n"
6445                         "%transformed_param = OpFAdd %v4f32 %param1 %v3\n"
6446                 },
6447                 {
6448                         "bool",
6449                         "%cnull             = OpConstantNull %bool\n",
6450                         "%v                 = OpVariable %fp_v4f32 Function\n"
6451                         "                     OpStore %v %param1\n"
6452                         "                     OpSelectionMerge %false_label None\n"
6453                         "                     OpBranchConditional %cnull %true_label %false_label\n"
6454                         "%true_label        = OpLabel\n"
6455                         "                     OpStore %v %c_v4f32_0_5_0_5_0_5_0_5\n"
6456                         "                     OpBranch %false_label\n"
6457                         "%false_label       = OpLabel\n"
6458                         "%transformed_param = OpLoad %v4f32 %v\n"
6459                 },
6460                 {
6461                         "i32",
6462                         "%cnull             = OpConstantNull %i32\n",
6463                         "%v                 = OpVariable %fp_v4f32 Function %c_v4f32_0_5_0_5_0_5_0_5\n"
6464                         "%b                 = OpIEqual %bool %cnull %c_i32_0\n"
6465                         "                     OpSelectionMerge %false_label None\n"
6466                         "                     OpBranchConditional %b %true_label %false_label\n"
6467                         "%true_label        = OpLabel\n"
6468                         "                     OpStore %v %param1\n"
6469                         "                     OpBranch %false_label\n"
6470                         "%false_label       = OpLabel\n"
6471                         "%transformed_param = OpLoad %v4f32 %v\n"
6472                 },
6473                 {
6474                         "struct",
6475                         "%stype             = OpTypeStruct %f32 %v4f32\n"
6476                         "%fp_stype          = OpTypePointer Function %stype\n"
6477                         "%cnull             = OpConstantNull %stype\n",
6478                         "%v                 = OpVariable %fp_stype Function %cnull\n"
6479                         "%f                 = OpAccessChain %fp_v4f32 %v %c_i32_1\n"
6480                         "%f_val             = OpLoad %v4f32 %f\n"
6481                         "%transformed_param = OpFAdd %v4f32 %param1 %f_val\n"
6482                 },
6483                 {
6484                         "array",
6485                         "%a4_v4f32          = OpTypeArray %v4f32 %c_u32_4\n"
6486                         "%fp_a4_v4f32       = OpTypePointer Function %a4_v4f32\n"
6487                         "%cnull             = OpConstantNull %a4_v4f32\n",
6488                         "%v                 = OpVariable %fp_a4_v4f32 Function %cnull\n"
6489                         "%f                 = OpAccessChain %fp_v4f32 %v %c_u32_0\n"
6490                         "%f1                = OpAccessChain %fp_v4f32 %v %c_u32_1\n"
6491                         "%f2                = OpAccessChain %fp_v4f32 %v %c_u32_2\n"
6492                         "%f3                = OpAccessChain %fp_v4f32 %v %c_u32_3\n"
6493                         "%f_val             = OpLoad %v4f32 %f\n"
6494                         "%f1_val            = OpLoad %v4f32 %f1\n"
6495                         "%f2_val            = OpLoad %v4f32 %f2\n"
6496                         "%f3_val            = OpLoad %v4f32 %f3\n"
6497                         "%t0                = OpFAdd %v4f32 %param1 %f_val\n"
6498                         "%t1                = OpFAdd %v4f32 %t0 %f1_val\n"
6499                         "%t2                = OpFAdd %v4f32 %t1 %f2_val\n"
6500                         "%transformed_param = OpFAdd %v4f32 %t2 %f3_val\n"
6501                 },
6502                 {
6503                         "matrix",
6504                         "%mat4x4_f32        = OpTypeMatrix %v4f32 4\n"
6505                         "%cnull             = OpConstantNull %mat4x4_f32\n",
6506                         // Our null matrix * any vector should result in a zero vector.
6507                         "%v                 = OpVectorTimesMatrix %v4f32 %param1 %cnull\n"
6508                         "%transformed_param = OpFAdd %v4f32 %param1 %v\n"
6509                 }
6510         };
6511
6512         getHalfColorsFullAlpha(colors);
6513
6514         for (size_t testNdx = 0; testNdx < sizeof(tests) / sizeof(NameConstantsCode); ++testNdx)
6515         {
6516                 map<string, string> fragments;
6517                 fragments["pre_main"] = tests[testNdx].constants;
6518                 fragments["testfun"] = string(functionStart) + tests[testNdx].code + functionEnd;
6519                 createTestsForAllStages(tests[testNdx].name, colors, colors, fragments, opConstantNullTests.get());
6520         }
6521         return opConstantNullTests.release();
6522 }
6523 tcu::TestCaseGroup* createOpConstantCompositeTests(tcu::TestContext& testCtx)
6524 {
6525         de::MovePtr<tcu::TestCaseGroup> opConstantCompositeTests                (new tcu::TestCaseGroup(testCtx, "opconstantcomposite", "OpConstantComposite instruction"));
6526         RGBA                                                    inputColors[4];
6527         RGBA                                                    outputColors[4];
6528
6529
6530         const char                                              functionStart[]  =
6531                 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
6532                 "%param1 = OpFunctionParameter %v4f32\n"
6533                 "%lbl    = OpLabel\n";
6534
6535         const char                                              functionEnd[]           =
6536                 "OpReturnValue %transformed_param\n"
6537                 "OpFunctionEnd\n";
6538
6539         struct NameConstantsCode
6540         {
6541                 string name;
6542                 string constants;
6543                 string code;
6544         };
6545
6546         NameConstantsCode tests[] =
6547         {
6548                 {
6549                         "vec4",
6550
6551                         "%cval              = OpConstantComposite %v4f32 %c_f32_0_5 %c_f32_0_5 %c_f32_0_5 %c_f32_0\n",
6552                         "%transformed_param = OpFAdd %v4f32 %param1 %cval\n"
6553                 },
6554                 {
6555                         "struct",
6556
6557                         "%stype             = OpTypeStruct %v4f32 %f32\n"
6558                         "%fp_stype          = OpTypePointer Function %stype\n"
6559                         "%f32_n_1           = OpConstant %f32 -1.0\n"
6560                         "%f32_1_5           = OpConstant %f32 !0x3fc00000\n" // +1.5
6561                         "%cvec              = OpConstantComposite %v4f32 %f32_1_5 %f32_1_5 %f32_1_5 %c_f32_1\n"
6562                         "%cval              = OpConstantComposite %stype %cvec %f32_n_1\n",
6563
6564                         "%v                 = OpVariable %fp_stype Function %cval\n"
6565                         "%vec_ptr           = OpAccessChain %fp_v4f32 %v %c_u32_0\n"
6566                         "%f32_ptr           = OpAccessChain %fp_f32 %v %c_u32_1\n"
6567                         "%vec_val           = OpLoad %v4f32 %vec_ptr\n"
6568                         "%f32_val           = OpLoad %f32 %f32_ptr\n"
6569                         "%tmp1              = OpVectorTimesScalar %v4f32 %c_v4f32_1_1_1_1 %f32_val\n" // vec4(-1)
6570                         "%tmp2              = OpFAdd %v4f32 %tmp1 %param1\n" // param1 + vec4(-1)
6571                         "%transformed_param = OpFAdd %v4f32 %tmp2 %vec_val\n" // param1 + vec4(-1) + vec4(1.5, 1.5, 1.5, 1.0)
6572                 },
6573                 {
6574                         // [1|0|0|0.5] [x] = x + 0.5
6575                         // [0|1|0|0.5] [y] = y + 0.5
6576                         // [0|0|1|0.5] [z] = z + 0.5
6577                         // [0|0|0|1  ] [1] = 1
6578                         "matrix",
6579
6580                         "%mat4x4_f32          = OpTypeMatrix %v4f32 4\n"
6581                         "%v4f32_1_0_0_0       = OpConstantComposite %v4f32 %c_f32_1 %c_f32_0 %c_f32_0 %c_f32_0\n"
6582                         "%v4f32_0_1_0_0       = OpConstantComposite %v4f32 %c_f32_0 %c_f32_1 %c_f32_0 %c_f32_0\n"
6583                         "%v4f32_0_0_1_0       = OpConstantComposite %v4f32 %c_f32_0 %c_f32_0 %c_f32_1 %c_f32_0\n"
6584                         "%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"
6585                         "%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",
6586
6587                         "%transformed_param   = OpMatrixTimesVector %v4f32 %cval %param1\n"
6588                 },
6589                 {
6590                         "array",
6591
6592                         "%c_v4f32_1_1_1_0     = OpConstantComposite %v4f32 %c_f32_1 %c_f32_1 %c_f32_1 %c_f32_0\n"
6593                         "%fp_a4f32            = OpTypePointer Function %a4f32\n"
6594                         "%f32_n_1             = OpConstant %f32 -1.0\n"
6595                         "%f32_1_5             = OpConstant %f32 !0x3fc00000\n" // +1.5
6596                         "%carr                = OpConstantComposite %a4f32 %c_f32_0 %f32_n_1 %f32_1_5 %c_f32_0\n",
6597
6598                         "%v                   = OpVariable %fp_a4f32 Function %carr\n"
6599                         "%f                   = OpAccessChain %fp_f32 %v %c_u32_0\n"
6600                         "%f1                  = OpAccessChain %fp_f32 %v %c_u32_1\n"
6601                         "%f2                  = OpAccessChain %fp_f32 %v %c_u32_2\n"
6602                         "%f3                  = OpAccessChain %fp_f32 %v %c_u32_3\n"
6603                         "%f_val               = OpLoad %f32 %f\n"
6604                         "%f1_val              = OpLoad %f32 %f1\n"
6605                         "%f2_val              = OpLoad %f32 %f2\n"
6606                         "%f3_val              = OpLoad %f32 %f3\n"
6607                         "%ftot1               = OpFAdd %f32 %f_val %f1_val\n"
6608                         "%ftot2               = OpFAdd %f32 %ftot1 %f2_val\n"
6609                         "%ftot3               = OpFAdd %f32 %ftot2 %f3_val\n"  // 0 - 1 + 1.5 + 0
6610                         "%add_vec             = OpVectorTimesScalar %v4f32 %c_v4f32_1_1_1_0 %ftot3\n"
6611                         "%transformed_param   = OpFAdd %v4f32 %param1 %add_vec\n"
6612                 },
6613                 {
6614                         //
6615                         // [
6616                         //   {
6617                         //      0.0,
6618                         //      [ 1.0, 1.0, 1.0, 1.0]
6619                         //   },
6620                         //   {
6621                         //      1.0,
6622                         //      [ 0.0, 0.5, 0.0, 0.0]
6623                         //   }, //     ^^^
6624                         //   {
6625                         //      0.0,
6626                         //      [ 1.0, 1.0, 1.0, 1.0]
6627                         //   }
6628                         // ]
6629                         "array_of_struct_of_array",
6630
6631                         "%c_v4f32_1_1_1_0     = OpConstantComposite %v4f32 %c_f32_1 %c_f32_1 %c_f32_1 %c_f32_0\n"
6632                         "%fp_a4f32            = OpTypePointer Function %a4f32\n"
6633                         "%stype               = OpTypeStruct %f32 %a4f32\n"
6634                         "%a3stype             = OpTypeArray %stype %c_u32_3\n"
6635                         "%fp_a3stype          = OpTypePointer Function %a3stype\n"
6636                         "%ca4f32_0            = OpConstantComposite %a4f32 %c_f32_0 %c_f32_0_5 %c_f32_0 %c_f32_0\n"
6637                         "%ca4f32_1            = OpConstantComposite %a4f32 %c_f32_1 %c_f32_1 %c_f32_1 %c_f32_1\n"
6638                         "%cstype1             = OpConstantComposite %stype %c_f32_0 %ca4f32_1\n"
6639                         "%cstype2             = OpConstantComposite %stype %c_f32_1 %ca4f32_0\n"
6640                         "%carr                = OpConstantComposite %a3stype %cstype1 %cstype2 %cstype1",
6641
6642                         "%v                   = OpVariable %fp_a3stype Function %carr\n"
6643                         "%f                   = OpAccessChain %fp_f32 %v %c_u32_1 %c_u32_1 %c_u32_1\n"
6644                         "%f_l                 = OpLoad %f32 %f\n"
6645                         "%add_vec             = OpVectorTimesScalar %v4f32 %c_v4f32_1_1_1_0 %f_l\n"
6646                         "%transformed_param   = OpFAdd %v4f32 %param1 %add_vec\n"
6647                 }
6648         };
6649
6650         getHalfColorsFullAlpha(inputColors);
6651         outputColors[0] = RGBA(255, 255, 255, 255);
6652         outputColors[1] = RGBA(255, 127, 127, 255);
6653         outputColors[2] = RGBA(127, 255, 127, 255);
6654         outputColors[3] = RGBA(127, 127, 255, 255);
6655
6656         for (size_t testNdx = 0; testNdx < sizeof(tests) / sizeof(NameConstantsCode); ++testNdx)
6657         {
6658                 map<string, string> fragments;
6659                 fragments["pre_main"] = tests[testNdx].constants;
6660                 fragments["testfun"] = string(functionStart) + tests[testNdx].code + functionEnd;
6661                 createTestsForAllStages(tests[testNdx].name, inputColors, outputColors, fragments, opConstantCompositeTests.get());
6662         }
6663         return opConstantCompositeTests.release();
6664 }
6665
6666 tcu::TestCaseGroup* createSelectionBlockOrderTests(tcu::TestContext& testCtx)
6667 {
6668         de::MovePtr<tcu::TestCaseGroup> group                           (new tcu::TestCaseGroup(testCtx, "selection_block_order", "Out-of-order blocks for selection"));
6669         RGBA                                                    inputColors[4];
6670         RGBA                                                    outputColors[4];
6671         map<string, string>                             fragments;
6672
6673         // vec4 test_code(vec4 param) {
6674         //   vec4 result = param;
6675         //   for (int i = 0; i < 4; ++i) {
6676         //     if (i == 0) result[i] = 0.;
6677         //     else        result[i] = 1. - result[i];
6678         //   }
6679         //   return result;
6680         // }
6681         const char                                              function[]                      =
6682                 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
6683                 "%param1    = OpFunctionParameter %v4f32\n"
6684                 "%lbl       = OpLabel\n"
6685                 "%iptr      = OpVariable %fp_i32 Function\n"
6686                 "%result    = OpVariable %fp_v4f32 Function\n"
6687                 "             OpStore %iptr %c_i32_0\n"
6688                 "             OpStore %result %param1\n"
6689                 "             OpBranch %loop\n"
6690
6691                 // Loop entry block.
6692                 "%loop      = OpLabel\n"
6693                 "%ival      = OpLoad %i32 %iptr\n"
6694                 "%lt_4      = OpSLessThan %bool %ival %c_i32_4\n"
6695                 "             OpLoopMerge %exit %if_entry None\n"
6696                 "             OpBranchConditional %lt_4 %if_entry %exit\n"
6697
6698                 // Merge block for loop.
6699                 "%exit      = OpLabel\n"
6700                 "%ret       = OpLoad %v4f32 %result\n"
6701                 "             OpReturnValue %ret\n"
6702
6703                 // If-statement entry block.
6704                 "%if_entry  = OpLabel\n"
6705                 "%loc       = OpAccessChain %fp_f32 %result %ival\n"
6706                 "%eq_0      = OpIEqual %bool %ival %c_i32_0\n"
6707                 "             OpSelectionMerge %if_exit None\n"
6708                 "             OpBranchConditional %eq_0 %if_true %if_false\n"
6709
6710                 // False branch for if-statement.
6711                 "%if_false  = OpLabel\n"
6712                 "%val       = OpLoad %f32 %loc\n"
6713                 "%sub       = OpFSub %f32 %c_f32_1 %val\n"
6714                 "             OpStore %loc %sub\n"
6715                 "             OpBranch %if_exit\n"
6716
6717                 // Merge block for if-statement.
6718                 "%if_exit   = OpLabel\n"
6719                 "%ival_next = OpIAdd %i32 %ival %c_i32_1\n"
6720                 "             OpStore %iptr %ival_next\n"
6721                 "             OpBranch %loop\n"
6722
6723                 // True branch for if-statement.
6724                 "%if_true   = OpLabel\n"
6725                 "             OpStore %loc %c_f32_0\n"
6726                 "             OpBranch %if_exit\n"
6727
6728                 "             OpFunctionEnd\n";
6729
6730         fragments["testfun"]    = function;
6731
6732         inputColors[0]                  = RGBA(127, 127, 127, 0);
6733         inputColors[1]                  = RGBA(127, 0,   0,   0);
6734         inputColors[2]                  = RGBA(0,   127, 0,   0);
6735         inputColors[3]                  = RGBA(0,   0,   127, 0);
6736
6737         outputColors[0]                 = RGBA(0, 128, 128, 255);
6738         outputColors[1]                 = RGBA(0, 255, 255, 255);
6739         outputColors[2]                 = RGBA(0, 128, 255, 255);
6740         outputColors[3]                 = RGBA(0, 255, 128, 255);
6741
6742         createTestsForAllStages("out_of_order", inputColors, outputColors, fragments, group.get());
6743
6744         return group.release();
6745 }
6746
6747 tcu::TestCaseGroup* createSwitchBlockOrderTests(tcu::TestContext& testCtx)
6748 {
6749         de::MovePtr<tcu::TestCaseGroup> group                           (new tcu::TestCaseGroup(testCtx, "switch_block_order", "Out-of-order blocks for switch"));
6750         RGBA                                                    inputColors[4];
6751         RGBA                                                    outputColors[4];
6752         map<string, string>                             fragments;
6753
6754         const char                                              typesAndConstants[]     =
6755                 "%c_f32_p2  = OpConstant %f32 0.2\n"
6756                 "%c_f32_p4  = OpConstant %f32 0.4\n"
6757                 "%c_f32_p6  = OpConstant %f32 0.6\n"
6758                 "%c_f32_p8  = OpConstant %f32 0.8\n";
6759
6760         // vec4 test_code(vec4 param) {
6761         //   vec4 result = param;
6762         //   for (int i = 0; i < 4; ++i) {
6763         //     switch (i) {
6764         //       case 0: result[i] += .2; break;
6765         //       case 1: result[i] += .6; break;
6766         //       case 2: result[i] += .4; break;
6767         //       case 3: result[i] += .8; break;
6768         //       default: break; // unreachable
6769         //     }
6770         //   }
6771         //   return result;
6772         // }
6773         const char                                              function[]                      =
6774                 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
6775                 "%param1    = OpFunctionParameter %v4f32\n"
6776                 "%lbl       = OpLabel\n"
6777                 "%iptr      = OpVariable %fp_i32 Function\n"
6778                 "%result    = OpVariable %fp_v4f32 Function\n"
6779                 "             OpStore %iptr %c_i32_0\n"
6780                 "             OpStore %result %param1\n"
6781                 "             OpBranch %loop\n"
6782
6783                 // Loop entry block.
6784                 "%loop      = OpLabel\n"
6785                 "%ival      = OpLoad %i32 %iptr\n"
6786                 "%lt_4      = OpSLessThan %bool %ival %c_i32_4\n"
6787                 "             OpLoopMerge %exit %switch_exit None\n"
6788                 "             OpBranchConditional %lt_4 %switch_entry %exit\n"
6789
6790                 // Merge block for loop.
6791                 "%exit      = OpLabel\n"
6792                 "%ret       = OpLoad %v4f32 %result\n"
6793                 "             OpReturnValue %ret\n"
6794
6795                 // Switch-statement entry block.
6796                 "%switch_entry   = OpLabel\n"
6797                 "%loc            = OpAccessChain %fp_f32 %result %ival\n"
6798                 "%val            = OpLoad %f32 %loc\n"
6799                 "                  OpSelectionMerge %switch_exit None\n"
6800                 "                  OpSwitch %ival %switch_default 0 %case0 1 %case1 2 %case2 3 %case3\n"
6801
6802                 "%case2          = OpLabel\n"
6803                 "%addp4          = OpFAdd %f32 %val %c_f32_p4\n"
6804                 "                  OpStore %loc %addp4\n"
6805                 "                  OpBranch %switch_exit\n"
6806
6807                 "%switch_default = OpLabel\n"
6808                 "                  OpUnreachable\n"
6809
6810                 "%case3          = OpLabel\n"
6811                 "%addp8          = OpFAdd %f32 %val %c_f32_p8\n"
6812                 "                  OpStore %loc %addp8\n"
6813                 "                  OpBranch %switch_exit\n"
6814
6815                 "%case0          = OpLabel\n"
6816                 "%addp2          = OpFAdd %f32 %val %c_f32_p2\n"
6817                 "                  OpStore %loc %addp2\n"
6818                 "                  OpBranch %switch_exit\n"
6819
6820                 // Merge block for switch-statement.
6821                 "%switch_exit    = OpLabel\n"
6822                 "%ival_next      = OpIAdd %i32 %ival %c_i32_1\n"
6823                 "                  OpStore %iptr %ival_next\n"
6824                 "                  OpBranch %loop\n"
6825
6826                 "%case1          = OpLabel\n"
6827                 "%addp6          = OpFAdd %f32 %val %c_f32_p6\n"
6828                 "                  OpStore %loc %addp6\n"
6829                 "                  OpBranch %switch_exit\n"
6830
6831                 "                  OpFunctionEnd\n";
6832
6833         fragments["pre_main"]   = typesAndConstants;
6834         fragments["testfun"]    = function;
6835
6836         inputColors[0]                  = RGBA(127, 27,  127, 51);
6837         inputColors[1]                  = RGBA(127, 0,   0,   51);
6838         inputColors[2]                  = RGBA(0,   27,  0,   51);
6839         inputColors[3]                  = RGBA(0,   0,   127, 51);
6840
6841         outputColors[0]                 = RGBA(178, 180, 229, 255);
6842         outputColors[1]                 = RGBA(178, 153, 102, 255);
6843         outputColors[2]                 = RGBA(51,  180, 102, 255);
6844         outputColors[3]                 = RGBA(51,  153, 229, 255);
6845
6846         createTestsForAllStages("out_of_order", inputColors, outputColors, fragments, group.get());
6847
6848         return group.release();
6849 }
6850
6851 tcu::TestCaseGroup* createDecorationGroupTests(tcu::TestContext& testCtx)
6852 {
6853         de::MovePtr<tcu::TestCaseGroup> group                           (new tcu::TestCaseGroup(testCtx, "decoration_group", "Decoration group tests"));
6854         RGBA                                                    inputColors[4];
6855         RGBA                                                    outputColors[4];
6856         map<string, string>                             fragments;
6857
6858         const char                                              decorations[]           =
6859                 "OpDecorate %array_group         ArrayStride 4\n"
6860                 "OpDecorate %struct_member_group Offset 0\n"
6861                 "%array_group         = OpDecorationGroup\n"
6862                 "%struct_member_group = OpDecorationGroup\n"
6863
6864                 "OpDecorate %group1 RelaxedPrecision\n"
6865                 "OpDecorate %group3 RelaxedPrecision\n"
6866                 "OpDecorate %group3 Invariant\n"
6867                 "OpDecorate %group3 Restrict\n"
6868                 "%group0 = OpDecorationGroup\n"
6869                 "%group1 = OpDecorationGroup\n"
6870                 "%group3 = OpDecorationGroup\n";
6871
6872         const char                                              typesAndConstants[]     =
6873                 "%a3f32     = OpTypeArray %f32 %c_u32_3\n"
6874                 "%struct1   = OpTypeStruct %a3f32\n"
6875                 "%struct2   = OpTypeStruct %a3f32\n"
6876                 "%fp_struct1 = OpTypePointer Function %struct1\n"
6877                 "%fp_struct2 = OpTypePointer Function %struct2\n"
6878                 "%c_f32_2    = OpConstant %f32 2.\n"
6879                 "%c_f32_n2   = OpConstant %f32 -2.\n"
6880
6881                 "%c_a3f32_1 = OpConstantComposite %a3f32 %c_f32_1 %c_f32_2 %c_f32_1\n"
6882                 "%c_a3f32_2 = OpConstantComposite %a3f32 %c_f32_n1 %c_f32_n2 %c_f32_n1\n"
6883                 "%c_struct1 = OpConstantComposite %struct1 %c_a3f32_1\n"
6884                 "%c_struct2 = OpConstantComposite %struct2 %c_a3f32_2\n";
6885
6886         const char                                              function[]                      =
6887                 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
6888                 "%param     = OpFunctionParameter %v4f32\n"
6889                 "%entry     = OpLabel\n"
6890                 "%result    = OpVariable %fp_v4f32 Function\n"
6891                 "%v_struct1 = OpVariable %fp_struct1 Function\n"
6892                 "%v_struct2 = OpVariable %fp_struct2 Function\n"
6893                 "             OpStore %result %param\n"
6894                 "             OpStore %v_struct1 %c_struct1\n"
6895                 "             OpStore %v_struct2 %c_struct2\n"
6896                 "%ptr1      = OpAccessChain %fp_f32 %v_struct1 %c_i32_0 %c_i32_2\n"
6897                 "%val1      = OpLoad %f32 %ptr1\n"
6898                 "%ptr2      = OpAccessChain %fp_f32 %v_struct2 %c_i32_0 %c_i32_2\n"
6899                 "%val2      = OpLoad %f32 %ptr2\n"
6900                 "%addvalues = OpFAdd %f32 %val1 %val2\n"
6901                 "%ptr       = OpAccessChain %fp_f32 %result %c_i32_1\n"
6902                 "%val       = OpLoad %f32 %ptr\n"
6903                 "%addresult = OpFAdd %f32 %addvalues %val\n"
6904                 "             OpStore %ptr %addresult\n"
6905                 "%ret       = OpLoad %v4f32 %result\n"
6906                 "             OpReturnValue %ret\n"
6907                 "             OpFunctionEnd\n";
6908
6909         struct CaseNameDecoration
6910         {
6911                 string name;
6912                 string decoration;
6913         };
6914
6915         CaseNameDecoration tests[] =
6916         {
6917                 {
6918                         "same_decoration_group_on_multiple_types",
6919                         "OpGroupMemberDecorate %struct_member_group %struct1 0 %struct2 0\n"
6920                 },
6921                 {
6922                         "empty_decoration_group",
6923                         "OpGroupDecorate %group0      %a3f32\n"
6924                         "OpGroupDecorate %group0      %result\n"
6925                 },
6926                 {
6927                         "one_element_decoration_group",
6928                         "OpGroupDecorate %array_group %a3f32\n"
6929                 },
6930                 {
6931                         "multiple_elements_decoration_group",
6932                         "OpGroupDecorate %group3      %v_struct1\n"
6933                 },
6934                 {
6935                         "multiple_decoration_groups_on_same_variable",
6936                         "OpGroupDecorate %group0      %v_struct2\n"
6937                         "OpGroupDecorate %group1      %v_struct2\n"
6938                         "OpGroupDecorate %group3      %v_struct2\n"
6939                 },
6940                 {
6941                         "same_decoration_group_multiple_times",
6942                         "OpGroupDecorate %group1      %addvalues\n"
6943                         "OpGroupDecorate %group1      %addvalues\n"
6944                         "OpGroupDecorate %group1      %addvalues\n"
6945                 },
6946
6947         };
6948
6949         getHalfColorsFullAlpha(inputColors);
6950         getHalfColorsFullAlpha(outputColors);
6951
6952         for (size_t idx = 0; idx < (sizeof(tests) / sizeof(tests[0])); ++idx)
6953         {
6954                 fragments["decoration"] = decorations + tests[idx].decoration;
6955                 fragments["pre_main"]   = typesAndConstants;
6956                 fragments["testfun"]    = function;
6957
6958                 createTestsForAllStages(tests[idx].name, inputColors, outputColors, fragments, group.get());
6959         }
6960
6961         return group.release();
6962 }
6963
6964 struct SpecConstantTwoIntGraphicsCase
6965 {
6966         const char*             caseName;
6967         const char*             scDefinition0;
6968         const char*             scDefinition1;
6969         const char*             scResultType;
6970         const char*             scOperation;
6971         deInt32                 scActualValue0;
6972         deInt32                 scActualValue1;
6973         const char*             resultOperation;
6974         RGBA                    expectedColors[4];
6975         deInt32                 scActualValueLength;
6976
6977                                         SpecConstantTwoIntGraphicsCase (const char*             name,
6978                                                                                                         const char*             definition0,
6979                                                                                                         const char*             definition1,
6980                                                                                                         const char*             resultType,
6981                                                                                                         const char*             operation,
6982                                                                                                         const deInt32   value0,
6983                                                                                                         const deInt32   value1,
6984                                                                                                         const char*             resultOp,
6985                                                                                                         const RGBA              (&output)[4],
6986                                                                                                         const deInt32   valueLength = sizeof(deInt32))
6987                                                 : caseName                              (name)
6988                                                 , scDefinition0                 (definition0)
6989                                                 , scDefinition1                 (definition1)
6990                                                 , scResultType                  (resultType)
6991                                                 , scOperation                   (operation)
6992                                                 , scActualValue0                (value0)
6993                                                 , scActualValue1                (value1)
6994                                                 , resultOperation               (resultOp)
6995                                                 , scActualValueLength   (valueLength)
6996         {
6997                 expectedColors[0] = output[0];
6998                 expectedColors[1] = output[1];
6999                 expectedColors[2] = output[2];
7000                 expectedColors[3] = output[3];
7001         }
7002 };
7003
7004 tcu::TestCaseGroup* createSpecConstantTests (tcu::TestContext& testCtx)
7005 {
7006         de::MovePtr<tcu::TestCaseGroup> group                           (new tcu::TestCaseGroup(testCtx, "opspecconstantop", "Test the OpSpecConstantOp instruction"));
7007         vector<SpecConstantTwoIntGraphicsCase>  cases;
7008         RGBA                                                    inputColors[4];
7009         RGBA                                                    outputColors0[4];
7010         RGBA                                                    outputColors1[4];
7011         RGBA                                                    outputColors2[4];
7012
7013         const deInt32                                   m1AsFloat16                     = 0xbc00; // -1(fp16) == 1 01111 0000000000 == 1011 1100 0000 0000
7014
7015         const char      decorations1[]                  =
7016                 "OpDecorate %sc_0  SpecId 0\n"
7017                 "OpDecorate %sc_1  SpecId 1\n";
7018
7019         const char      typesAndConstants1[]    =
7020                 "${OPTYPE_DEFINITIONS:opt}"
7021                 "%sc_0      = OpSpecConstant${SC_DEF0}\n"
7022                 "%sc_1      = OpSpecConstant${SC_DEF1}\n"
7023                 "%sc_op     = OpSpecConstantOp ${SC_RESULT_TYPE} ${SC_OP}\n";
7024
7025         const char      function1[]                             =
7026                 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
7027                 "%param     = OpFunctionParameter %v4f32\n"
7028                 "%label     = OpLabel\n"
7029                 "%result    = OpVariable %fp_v4f32 Function\n"
7030                 "${TYPE_CONVERT:opt}"
7031                 "             OpStore %result %param\n"
7032                 "%gen       = ${GEN_RESULT}\n"
7033                 "%index     = OpIAdd %i32 %gen %c_i32_1\n"
7034                 "%loc       = OpAccessChain %fp_f32 %result %index\n"
7035                 "%val       = OpLoad %f32 %loc\n"
7036                 "%add       = OpFAdd %f32 %val %c_f32_0_5\n"
7037                 "             OpStore %loc %add\n"
7038                 "%ret       = OpLoad %v4f32 %result\n"
7039                 "             OpReturnValue %ret\n"
7040                 "             OpFunctionEnd\n";
7041
7042         inputColors[0] = RGBA(127, 127, 127, 255);
7043         inputColors[1] = RGBA(127, 0,   0,   255);
7044         inputColors[2] = RGBA(0,   127, 0,   255);
7045         inputColors[3] = RGBA(0,   0,   127, 255);
7046
7047         // Derived from inputColors[x] by adding 128 to inputColors[x][0].
7048         outputColors0[0] = RGBA(255, 127, 127, 255);
7049         outputColors0[1] = RGBA(255, 0,   0,   255);
7050         outputColors0[2] = RGBA(128, 127, 0,   255);
7051         outputColors0[3] = RGBA(128, 0,   127, 255);
7052
7053         // Derived from inputColors[x] by adding 128 to inputColors[x][1].
7054         outputColors1[0] = RGBA(127, 255, 127, 255);
7055         outputColors1[1] = RGBA(127, 128, 0,   255);
7056         outputColors1[2] = RGBA(0,   255, 0,   255);
7057         outputColors1[3] = RGBA(0,   128, 127, 255);
7058
7059         // Derived from inputColors[x] by adding 128 to inputColors[x][2].
7060         outputColors2[0] = RGBA(127, 127, 255, 255);
7061         outputColors2[1] = RGBA(127, 0,   128, 255);
7062         outputColors2[2] = RGBA(0,   127, 128, 255);
7063         outputColors2[3] = RGBA(0,   0,   255, 255);
7064
7065         const char addZeroToSc[]                = "OpIAdd %i32 %c_i32_0 %sc_op";
7066         const char addZeroToSc32[]              = "OpIAdd %i32 %c_i32_0 %sc_op32";
7067         const char selectTrueUsingSc[]  = "OpSelect %i32 %sc_op %c_i32_1 %c_i32_0";
7068         const char selectFalseUsingSc[] = "OpSelect %i32 %sc_op %c_i32_0 %c_i32_1";
7069
7070         cases.push_back(SpecConstantTwoIntGraphicsCase("iadd",                                  " %i32 0",              " %i32 0",              "%i32",         "IAdd                 %sc_0 %sc_1",                             19,             -20,    addZeroToSc,            outputColors0));
7071         cases.push_back(SpecConstantTwoIntGraphicsCase("isub",                                  " %i32 0",              " %i32 0",              "%i32",         "ISub                 %sc_0 %sc_1",                             19,             20,             addZeroToSc,            outputColors0));
7072         cases.push_back(SpecConstantTwoIntGraphicsCase("imul",                                  " %i32 0",              " %i32 0",              "%i32",         "IMul                 %sc_0 %sc_1",                             -1,             -1,             addZeroToSc,            outputColors2));
7073         cases.push_back(SpecConstantTwoIntGraphicsCase("sdiv",                                  " %i32 0",              " %i32 0",              "%i32",         "SDiv                 %sc_0 %sc_1",                             -126,   126,    addZeroToSc,            outputColors0));
7074         cases.push_back(SpecConstantTwoIntGraphicsCase("udiv",                                  " %i32 0",              " %i32 0",              "%i32",         "UDiv                 %sc_0 %sc_1",                             126,    126,    addZeroToSc,            outputColors2));
7075         cases.push_back(SpecConstantTwoIntGraphicsCase("srem",                                  " %i32 0",              " %i32 0",              "%i32",         "SRem                 %sc_0 %sc_1",                             3,              2,              addZeroToSc,            outputColors2));
7076         cases.push_back(SpecConstantTwoIntGraphicsCase("smod",                                  " %i32 0",              " %i32 0",              "%i32",         "SMod                 %sc_0 %sc_1",                             3,              2,              addZeroToSc,            outputColors2));
7077         cases.push_back(SpecConstantTwoIntGraphicsCase("umod",                                  " %i32 0",              " %i32 0",              "%i32",         "UMod                 %sc_0 %sc_1",                             1001,   500,    addZeroToSc,            outputColors2));
7078         cases.push_back(SpecConstantTwoIntGraphicsCase("bitwiseand",                    " %i32 0",              " %i32 0",              "%i32",         "BitwiseAnd           %sc_0 %sc_1",                             0x33,   0x0d,   addZeroToSc,            outputColors2));
7079         cases.push_back(SpecConstantTwoIntGraphicsCase("bitwiseor",                             " %i32 0",              " %i32 0",              "%i32",         "BitwiseOr            %sc_0 %sc_1",                             0,              1,              addZeroToSc,            outputColors2));
7080         cases.push_back(SpecConstantTwoIntGraphicsCase("bitwisexor",                    " %i32 0",              " %i32 0",              "%i32",         "BitwiseXor           %sc_0 %sc_1",                             0x2e,   0x2f,   addZeroToSc,            outputColors2));
7081         cases.push_back(SpecConstantTwoIntGraphicsCase("shiftrightlogical",             " %i32 0",              " %i32 0",              "%i32",         "ShiftRightLogical    %sc_0 %sc_1",                             2,              1,              addZeroToSc,            outputColors2));
7082         cases.push_back(SpecConstantTwoIntGraphicsCase("shiftrightarithmetic",  " %i32 0",              " %i32 0",              "%i32",         "ShiftRightArithmetic %sc_0 %sc_1",                             -4,             2,              addZeroToSc,            outputColors0));
7083         cases.push_back(SpecConstantTwoIntGraphicsCase("shiftleftlogical",              " %i32 0",              " %i32 0",              "%i32",         "ShiftLeftLogical     %sc_0 %sc_1",                             1,              0,              addZeroToSc,            outputColors2));
7084         cases.push_back(SpecConstantTwoIntGraphicsCase("slessthan",                             " %i32 0",              " %i32 0",              "%bool",        "SLessThan            %sc_0 %sc_1",                             -20,    -10,    selectTrueUsingSc,      outputColors2));
7085         cases.push_back(SpecConstantTwoIntGraphicsCase("ulessthan",                             " %i32 0",              " %i32 0",              "%bool",        "ULessThan            %sc_0 %sc_1",                             10,             20,             selectTrueUsingSc,      outputColors2));
7086         cases.push_back(SpecConstantTwoIntGraphicsCase("sgreaterthan",                  " %i32 0",              " %i32 0",              "%bool",        "SGreaterThan         %sc_0 %sc_1",                             -1000,  50,             selectFalseUsingSc,     outputColors2));
7087         cases.push_back(SpecConstantTwoIntGraphicsCase("ugreaterthan",                  " %i32 0",              " %i32 0",              "%bool",        "UGreaterThan         %sc_0 %sc_1",                             10,             5,              selectTrueUsingSc,      outputColors2));
7088         cases.push_back(SpecConstantTwoIntGraphicsCase("slessthanequal",                " %i32 0",              " %i32 0",              "%bool",        "SLessThanEqual       %sc_0 %sc_1",                             -10,    -10,    selectTrueUsingSc,      outputColors2));
7089         cases.push_back(SpecConstantTwoIntGraphicsCase("ulessthanequal",                " %i32 0",              " %i32 0",              "%bool",        "ULessThanEqual       %sc_0 %sc_1",                             50,             100,    selectTrueUsingSc,      outputColors2));
7090         cases.push_back(SpecConstantTwoIntGraphicsCase("sgreaterthanequal",             " %i32 0",              " %i32 0",              "%bool",        "SGreaterThanEqual    %sc_0 %sc_1",                             -1000,  50,             selectFalseUsingSc,     outputColors2));
7091         cases.push_back(SpecConstantTwoIntGraphicsCase("ugreaterthanequal",             " %i32 0",              " %i32 0",              "%bool",        "UGreaterThanEqual    %sc_0 %sc_1",                             10,             10,             selectTrueUsingSc,      outputColors2));
7092         cases.push_back(SpecConstantTwoIntGraphicsCase("iequal",                                " %i32 0",              " %i32 0",              "%bool",        "IEqual               %sc_0 %sc_1",                             42,             24,             selectFalseUsingSc,     outputColors2));
7093         cases.push_back(SpecConstantTwoIntGraphicsCase("inotequal",                             " %i32 0",              " %i32 0",              "%bool",        "INotEqual            %sc_0 %sc_1",                             42,             24,             selectTrueUsingSc,      outputColors2));
7094         cases.push_back(SpecConstantTwoIntGraphicsCase("logicaland",                    "True %bool",   "True %bool",   "%bool",        "LogicalAnd           %sc_0 %sc_1",                             0,              1,              selectFalseUsingSc,     outputColors2));
7095         cases.push_back(SpecConstantTwoIntGraphicsCase("logicalor",                             "False %bool",  "False %bool",  "%bool",        "LogicalOr            %sc_0 %sc_1",                             1,              0,              selectTrueUsingSc,      outputColors2));
7096         cases.push_back(SpecConstantTwoIntGraphicsCase("logicalequal",                  "True %bool",   "True %bool",   "%bool",        "LogicalEqual         %sc_0 %sc_1",                             0,              1,              selectFalseUsingSc,     outputColors2));
7097         cases.push_back(SpecConstantTwoIntGraphicsCase("logicalnotequal",               "False %bool",  "False %bool",  "%bool",        "LogicalNotEqual      %sc_0 %sc_1",                             1,              0,              selectTrueUsingSc,      outputColors2));
7098         cases.push_back(SpecConstantTwoIntGraphicsCase("snegate",                               " %i32 0",              " %i32 0",              "%i32",         "SNegate              %sc_0",                                   -1,             0,              addZeroToSc,            outputColors2));
7099         cases.push_back(SpecConstantTwoIntGraphicsCase("not",                                   " %i32 0",              " %i32 0",              "%i32",         "Not                  %sc_0",                                   -2,             0,              addZeroToSc,            outputColors2));
7100         cases.push_back(SpecConstantTwoIntGraphicsCase("logicalnot",                    "False %bool",  "False %bool",  "%bool",        "LogicalNot           %sc_0",                                   1,              0,              selectFalseUsingSc,     outputColors2));
7101         cases.push_back(SpecConstantTwoIntGraphicsCase("select",                                "False %bool",  " %i32 0",              "%i32",         "Select               %sc_0 %sc_1 %c_i32_0",    1,              1,              addZeroToSc,            outputColors2));
7102         cases.push_back(SpecConstantTwoIntGraphicsCase("sconvert",                              " %i32 0",              " %i32 0",              "%i16",         "SConvert             %sc_0",                                   -1,             0,              addZeroToSc32,          outputColors0));
7103         // -1082130432 stored as 32-bit two's complement is the binary representation of -1 as IEEE-754 Float
7104         cases.push_back(SpecConstantTwoIntGraphicsCase("fconvert",                              " %f32 0",              " %f32 0",              "%f64",         "FConvert             %sc_0",                                   -1082130432, 0, addZeroToSc32,          outputColors0));
7105         cases.push_back(SpecConstantTwoIntGraphicsCase("fconvert16",                    " %f16 0",              " %f16 0",              "%f32",         "FConvert             %sc_0",                                   m1AsFloat16, 0, addZeroToSc32,          outputColors0, sizeof(deFloat16)));
7106         // \todo[2015-12-1 antiagainst] OpQuantizeToF16
7107
7108         for (size_t caseNdx = 0; caseNdx < cases.size(); ++caseNdx)
7109         {
7110                 map<string, string>                     specializations;
7111                 map<string, string>                     fragments;
7112                 SpecConstants                           specConstants;
7113                 PushConstants                           noPushConstants;
7114                 GraphicsResources                       noResources;
7115                 GraphicsInterfaces                      noInterfaces;
7116                 vector<string>                          extensions;
7117                 VulkanFeatures                          requiredFeatures;
7118
7119                 // Special SPIR-V code for SConvert-case
7120                 if (strcmp(cases[caseNdx].caseName, "sconvert") == 0)
7121                 {
7122                         requiredFeatures.coreFeatures.shaderInt16 = VK_TRUE;
7123                         fragments["capability"]                                 = "OpCapability Int16\n";                                       // Adds 16-bit integer capability
7124                         specializations["OPTYPE_DEFINITIONS"]   = "%i16 = OpTypeInt 16 1\n";                            // Adds 16-bit integer type
7125                         specializations["TYPE_CONVERT"]                 = "%sc_op32 = OpSConvert %i32 %sc_op\n";        // Converts 16-bit integer to 32-bit integer
7126                 }
7127
7128                 // Special SPIR-V code for FConvert-case
7129                 if (strcmp(cases[caseNdx].caseName, "fconvert") == 0)
7130                 {
7131                         requiredFeatures.coreFeatures.shaderFloat64 = VK_TRUE;
7132                         fragments["capability"]                                 = "OpCapability Float64\n";                                     // Adds 64-bit float capability
7133                         specializations["OPTYPE_DEFINITIONS"]   = "%f64 = OpTypeFloat 64\n";                            // Adds 64-bit float type
7134                         specializations["TYPE_CONVERT"]                 = "%sc_op32 = OpConvertFToS %i32 %sc_op\n";     // Converts 64-bit float to 32-bit integer
7135                 }
7136
7137                 // Special SPIR-V code for FConvert-case for 16-bit floats
7138                 if (strcmp(cases[caseNdx].caseName, "fconvert16") == 0)
7139                 {
7140                         extensions.push_back("VK_KHR_shader_float16_int8");
7141                         requiredFeatures.extFloat16Int8 = EXTFLOAT16INT8FEATURES_FLOAT16;
7142                         fragments["capability"]                                 = "OpCapability Float16\n";                                     // Adds 16-bit float capability
7143                         specializations["OPTYPE_DEFINITIONS"]   = "%f16 = OpTypeFloat 16\n";                            // Adds 16-bit float type
7144                         specializations["TYPE_CONVERT"]                 = "%sc_op32 = OpConvertFToS %i32 %sc_op\n";     // Converts 16-bit float to 32-bit integer
7145                 }
7146
7147                 specializations["SC_DEF0"]                      = cases[caseNdx].scDefinition0;
7148                 specializations["SC_DEF1"]                      = cases[caseNdx].scDefinition1;
7149                 specializations["SC_RESULT_TYPE"]       = cases[caseNdx].scResultType;
7150                 specializations["SC_OP"]                        = cases[caseNdx].scOperation;
7151                 specializations["GEN_RESULT"]           = cases[caseNdx].resultOperation;
7152
7153                 fragments["decoration"]                         = tcu::StringTemplate(decorations1).specialize(specializations);
7154                 fragments["pre_main"]                           = tcu::StringTemplate(typesAndConstants1).specialize(specializations);
7155                 fragments["testfun"]                            = tcu::StringTemplate(function1).specialize(specializations);
7156
7157                 specConstants.append(&cases[caseNdx].scActualValue0, cases[caseNdx].scActualValueLength);
7158                 specConstants.append(&cases[caseNdx].scActualValue1, cases[caseNdx].scActualValueLength);
7159
7160                 createTestsForAllStages(
7161                         cases[caseNdx].caseName, inputColors, cases[caseNdx].expectedColors, fragments, specConstants,
7162                         noPushConstants, noResources, noInterfaces, extensions, requiredFeatures, group.get());
7163         }
7164
7165         const char      decorations2[]                  =
7166                 "OpDecorate %sc_0  SpecId 0\n"
7167                 "OpDecorate %sc_1  SpecId 1\n"
7168                 "OpDecorate %sc_2  SpecId 2\n";
7169
7170         const char      typesAndConstants2[]    =
7171                 "%vec3_0      = OpConstantComposite %v3i32 %c_i32_0 %c_i32_0 %c_i32_0\n"
7172                 "%vec3_undef  = OpUndef %v3i32\n"
7173
7174                 "%sc_0        = OpSpecConstant %i32 0\n"
7175                 "%sc_1        = OpSpecConstant %i32 0\n"
7176                 "%sc_2        = OpSpecConstant %i32 0\n"
7177                 "%sc_vec3_0   = OpSpecConstantOp %v3i32 CompositeInsert  %sc_0        %vec3_0      0\n"                                                 // (sc_0, 0,    0)
7178                 "%sc_vec3_1   = OpSpecConstantOp %v3i32 CompositeInsert  %sc_1        %vec3_0      1\n"                                                 // (0,    sc_1, 0)
7179                 "%sc_vec3_2   = OpSpecConstantOp %v3i32 CompositeInsert  %sc_2        %vec3_0      2\n"                                                 // (0,    0,    sc_2)
7180                 "%sc_vec3_0_s = OpSpecConstantOp %v3i32 VectorShuffle    %sc_vec3_0   %vec3_undef  0          0xFFFFFFFF 2\n"   // (sc_0, ???,  0)
7181                 "%sc_vec3_1_s = OpSpecConstantOp %v3i32 VectorShuffle    %sc_vec3_1   %vec3_undef  0xFFFFFFFF 1          0\n"   // (???,  sc_1, 0)
7182                 "%sc_vec3_2_s = OpSpecConstantOp %v3i32 VectorShuffle    %vec3_undef  %sc_vec3_2   5          0xFFFFFFFF 5\n"   // (sc_2, ???,  sc_2)
7183                 "%sc_vec3_01  = OpSpecConstantOp %v3i32 VectorShuffle    %sc_vec3_0_s %sc_vec3_1_s 1 0 4\n"                                             // (0,    sc_0, sc_1)
7184                 "%sc_vec3_012 = OpSpecConstantOp %v3i32 VectorShuffle    %sc_vec3_01  %sc_vec3_2_s 5 1 2\n"                                             // (sc_2, sc_0, sc_1)
7185                 "%sc_ext_0    = OpSpecConstantOp %i32   CompositeExtract %sc_vec3_012              0\n"                                                 // sc_2
7186                 "%sc_ext_1    = OpSpecConstantOp %i32   CompositeExtract %sc_vec3_012              1\n"                                                 // sc_0
7187                 "%sc_ext_2    = OpSpecConstantOp %i32   CompositeExtract %sc_vec3_012              2\n"                                                 // sc_1
7188                 "%sc_sub      = OpSpecConstantOp %i32   ISub             %sc_ext_0    %sc_ext_1\n"                                                              // (sc_2 - sc_0)
7189                 "%sc_final    = OpSpecConstantOp %i32   IMul             %sc_sub      %sc_ext_2\n";                                                             // (sc_2 - sc_0) * sc_1
7190
7191         const char      function2[]                             =
7192                 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
7193                 "%param     = OpFunctionParameter %v4f32\n"
7194                 "%label     = OpLabel\n"
7195                 "%result    = OpVariable %fp_v4f32 Function\n"
7196                 "             OpStore %result %param\n"
7197                 "%loc       = OpAccessChain %fp_f32 %result %sc_final\n"
7198                 "%val       = OpLoad %f32 %loc\n"
7199                 "%add       = OpFAdd %f32 %val %c_f32_0_5\n"
7200                 "             OpStore %loc %add\n"
7201                 "%ret       = OpLoad %v4f32 %result\n"
7202                 "             OpReturnValue %ret\n"
7203                 "             OpFunctionEnd\n";
7204
7205         map<string, string>     fragments;
7206         SpecConstants           specConstants;
7207
7208         fragments["decoration"] = decorations2;
7209         fragments["pre_main"]   = typesAndConstants2;
7210         fragments["testfun"]    = function2;
7211
7212         specConstants.append<deInt32>(56789);
7213         specConstants.append<deInt32>(-2);
7214         specConstants.append<deInt32>(56788);
7215
7216         createTestsForAllStages("vector_related", inputColors, outputColors2, fragments, specConstants, group.get());
7217
7218         return group.release();
7219 }
7220
7221 tcu::TestCaseGroup* createOpPhiTests(tcu::TestContext& testCtx)
7222 {
7223         de::MovePtr<tcu::TestCaseGroup> group                           (new tcu::TestCaseGroup(testCtx, "opphi", "Test the OpPhi instruction"));
7224         RGBA                                                    inputColors[4];
7225         RGBA                                                    outputColors1[4];
7226         RGBA                                                    outputColors2[4];
7227         RGBA                                                    outputColors3[4];
7228         RGBA                                                    outputColors4[4];
7229         map<string, string>                             fragments1;
7230         map<string, string>                             fragments2;
7231         map<string, string>                             fragments3;
7232         map<string, string>                             fragments4;
7233         std::vector<std::string>                extensions4;
7234         GraphicsResources                               resources4;
7235         VulkanFeatures                                  vulkanFeatures4;
7236
7237         const char      typesAndConstants1[]    =
7238                 "%c_f32_p2  = OpConstant %f32 0.2\n"
7239                 "%c_f32_p4  = OpConstant %f32 0.4\n"
7240                 "%c_f32_p5  = OpConstant %f32 0.5\n"
7241                 "%c_f32_p8  = OpConstant %f32 0.8\n";
7242
7243         // vec4 test_code(vec4 param) {
7244         //   vec4 result = param;
7245         //   for (int i = 0; i < 4; ++i) {
7246         //     float operand;
7247         //     switch (i) {
7248         //       case 0: operand = .2; break;
7249         //       case 1: operand = .5; break;
7250         //       case 2: operand = .4; break;
7251         //       case 3: operand = .0; break;
7252         //       default: break; // unreachable
7253         //     }
7254         //     result[i] += operand;
7255         //   }
7256         //   return result;
7257         // }
7258         const char      function1[]                             =
7259                 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
7260                 "%param1    = OpFunctionParameter %v4f32\n"
7261                 "%lbl       = OpLabel\n"
7262                 "%iptr      = OpVariable %fp_i32 Function\n"
7263                 "%result    = OpVariable %fp_v4f32 Function\n"
7264                 "             OpStore %iptr %c_i32_0\n"
7265                 "             OpStore %result %param1\n"
7266                 "             OpBranch %loop\n"
7267
7268                 "%loop      = OpLabel\n"
7269                 "%ival      = OpLoad %i32 %iptr\n"
7270                 "%lt_4      = OpSLessThan %bool %ival %c_i32_4\n"
7271                 "             OpLoopMerge %exit %phi None\n"
7272                 "             OpBranchConditional %lt_4 %entry %exit\n"
7273
7274                 "%entry     = OpLabel\n"
7275                 "%loc       = OpAccessChain %fp_f32 %result %ival\n"
7276                 "%val       = OpLoad %f32 %loc\n"
7277                 "             OpSelectionMerge %phi None\n"
7278                 "             OpSwitch %ival %default 0 %case0 1 %case1 2 %case2 3 %case3\n"
7279
7280                 "%case0     = OpLabel\n"
7281                 "             OpBranch %phi\n"
7282                 "%case1     = OpLabel\n"
7283                 "             OpBranch %phi\n"
7284                 "%case2     = OpLabel\n"
7285                 "             OpBranch %phi\n"
7286                 "%case3     = OpLabel\n"
7287                 "             OpBranch %phi\n"
7288
7289                 "%default   = OpLabel\n"
7290                 "             OpUnreachable\n"
7291
7292                 "%phi       = OpLabel\n"
7293                 "%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
7294                 "%add       = OpFAdd %f32 %val %operand\n"
7295                 "             OpStore %loc %add\n"
7296                 "%ival_next = OpIAdd %i32 %ival %c_i32_1\n"
7297                 "             OpStore %iptr %ival_next\n"
7298                 "             OpBranch %loop\n"
7299
7300                 "%exit      = OpLabel\n"
7301                 "%ret       = OpLoad %v4f32 %result\n"
7302                 "             OpReturnValue %ret\n"
7303
7304                 "             OpFunctionEnd\n";
7305
7306         fragments1["pre_main"]  = typesAndConstants1;
7307         fragments1["testfun"]   = function1;
7308
7309         getHalfColorsFullAlpha(inputColors);
7310
7311         outputColors1[0]                = RGBA(178, 255, 229, 255);
7312         outputColors1[1]                = RGBA(178, 127, 102, 255);
7313         outputColors1[2]                = RGBA(51,  255, 102, 255);
7314         outputColors1[3]                = RGBA(51,  127, 229, 255);
7315
7316         createTestsForAllStages("out_of_order", inputColors, outputColors1, fragments1, group.get());
7317
7318         const char      typesAndConstants2[]    =
7319                 "%c_f32_p2  = OpConstant %f32 0.2\n";
7320
7321         // Add .4 to the second element of the given parameter.
7322         const char      function2[]                             =
7323                 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
7324                 "%param     = OpFunctionParameter %v4f32\n"
7325                 "%entry     = OpLabel\n"
7326                 "%result    = OpVariable %fp_v4f32 Function\n"
7327                 "             OpStore %result %param\n"
7328                 "%loc       = OpAccessChain %fp_f32 %result %c_i32_1\n"
7329                 "%val       = OpLoad %f32 %loc\n"
7330                 "             OpBranch %phi\n"
7331
7332                 "%phi        = OpLabel\n"
7333                 "%step       = OpPhi %i32 %c_i32_0  %entry %step_next  %phi\n"
7334                 "%accum      = OpPhi %f32 %val      %entry %accum_next %phi\n"
7335                 "%step_next  = OpIAdd %i32 %step  %c_i32_1\n"
7336                 "%accum_next = OpFAdd %f32 %accum %c_f32_p2\n"
7337                 "%still_loop = OpSLessThan %bool %step %c_i32_2\n"
7338                 "              OpLoopMerge %exit %phi None\n"
7339                 "              OpBranchConditional %still_loop %phi %exit\n"
7340
7341                 "%exit       = OpLabel\n"
7342                 "              OpStore %loc %accum\n"
7343                 "%ret        = OpLoad %v4f32 %result\n"
7344                 "              OpReturnValue %ret\n"
7345
7346                 "              OpFunctionEnd\n";
7347
7348         fragments2["pre_main"]  = typesAndConstants2;
7349         fragments2["testfun"]   = function2;
7350
7351         outputColors2[0]                        = RGBA(127, 229, 127, 255);
7352         outputColors2[1]                        = RGBA(127, 102, 0,   255);
7353         outputColors2[2]                        = RGBA(0,   229, 0,   255);
7354         outputColors2[3]                        = RGBA(0,   102, 127, 255);
7355
7356         createTestsForAllStages("induction", inputColors, outputColors2, fragments2, group.get());
7357
7358         const char      typesAndConstants3[]    =
7359                 "%true      = OpConstantTrue %bool\n"
7360                 "%false     = OpConstantFalse %bool\n"
7361                 "%c_f32_p2  = OpConstant %f32 0.2\n";
7362
7363         // Swap the second and the third element of the given parameter.
7364         const char      function3[]                             =
7365                 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
7366                 "%param     = OpFunctionParameter %v4f32\n"
7367                 "%entry     = OpLabel\n"
7368                 "%result    = OpVariable %fp_v4f32 Function\n"
7369                 "             OpStore %result %param\n"
7370                 "%a_loc     = OpAccessChain %fp_f32 %result %c_i32_1\n"
7371                 "%a_init    = OpLoad %f32 %a_loc\n"
7372                 "%b_loc     = OpAccessChain %fp_f32 %result %c_i32_2\n"
7373                 "%b_init    = OpLoad %f32 %b_loc\n"
7374                 "             OpBranch %phi\n"
7375
7376                 "%phi        = OpLabel\n"
7377                 "%still_loop = OpPhi %bool %true   %entry %false  %phi\n"
7378                 "%a_next     = OpPhi %f32  %a_init %entry %b_next %phi\n"
7379                 "%b_next     = OpPhi %f32  %b_init %entry %a_next %phi\n"
7380                 "              OpLoopMerge %exit %phi None\n"
7381                 "              OpBranchConditional %still_loop %phi %exit\n"
7382
7383                 "%exit       = OpLabel\n"
7384                 "              OpStore %a_loc %a_next\n"
7385                 "              OpStore %b_loc %b_next\n"
7386                 "%ret        = OpLoad %v4f32 %result\n"
7387                 "              OpReturnValue %ret\n"
7388
7389                 "              OpFunctionEnd\n";
7390
7391         fragments3["pre_main"]  = typesAndConstants3;
7392         fragments3["testfun"]   = function3;
7393
7394         outputColors3[0]                        = RGBA(127, 127, 127, 255);
7395         outputColors3[1]                        = RGBA(127, 0,   0,   255);
7396         outputColors3[2]                        = RGBA(0,   0,   127, 255);
7397         outputColors3[3]                        = RGBA(0,   127, 0,   255);
7398
7399         createTestsForAllStages("swap", inputColors, outputColors3, fragments3, group.get());
7400
7401         const char      typesAndConstants4[]    =
7402                 "%f16        = OpTypeFloat 16\n"
7403                 "%v4f16      = OpTypeVector %f16 4\n"
7404                 "%fp_f16     = OpTypePointer Function %f16\n"
7405                 "%fp_v4f16   = OpTypePointer Function %v4f16\n"
7406                 "%true       = OpConstantTrue %bool\n"
7407                 "%false      = OpConstantFalse %bool\n"
7408                 "%c_f32_p2   = OpConstant %f32 0.2\n";
7409
7410         // Swap the second and the third element of the given parameter.
7411         const char      function4[]                             =
7412                 "%test_code  = OpFunction %v4f32 None %v4f32_v4f32_function\n"
7413                 "%param      = OpFunctionParameter %v4f32\n"
7414                 "%entry      = OpLabel\n"
7415                 "%result     = OpVariable %fp_v4f16 Function\n"
7416                 "%param16    = OpFConvert %v4f16 %param\n"
7417                 "              OpStore %result %param16\n"
7418                 "%a_loc      = OpAccessChain %fp_f16 %result %c_i32_1\n"
7419                 "%a_init     = OpLoad %f16 %a_loc\n"
7420                 "%b_loc      = OpAccessChain %fp_f16 %result %c_i32_2\n"
7421                 "%b_init     = OpLoad %f16 %b_loc\n"
7422                 "              OpBranch %phi\n"
7423
7424                 "%phi        = OpLabel\n"
7425                 "%still_loop = OpPhi %bool %true   %entry %false  %phi\n"
7426                 "%a_next     = OpPhi %f16  %a_init %entry %b_next %phi\n"
7427                 "%b_next     = OpPhi %f16  %b_init %entry %a_next %phi\n"
7428                 "              OpLoopMerge %exit %phi None\n"
7429                 "              OpBranchConditional %still_loop %phi %exit\n"
7430
7431                 "%exit       = OpLabel\n"
7432                 "              OpStore %a_loc %a_next\n"
7433                 "              OpStore %b_loc %b_next\n"
7434                 "%ret16      = OpLoad %v4f16 %result\n"
7435                 "%ret        = OpFConvert %v4f32 %ret16\n"
7436                 "              OpReturnValue %ret\n"
7437
7438                 "              OpFunctionEnd\n";
7439
7440         fragments4["pre_main"]          = typesAndConstants4;
7441         fragments4["testfun"]           = function4;
7442         fragments4["capability"]        = "OpCapability StorageUniformBufferBlock16\n";
7443         fragments4["extension"]         = "OpExtension \"SPV_KHR_16bit_storage\"";
7444
7445         extensions4.push_back("VK_KHR_16bit_storage");
7446         extensions4.push_back("VK_KHR_shader_float16_int8");
7447
7448         vulkanFeatures4.ext16BitStorage = EXT16BITSTORAGEFEATURES_UNIFORM_BUFFER_BLOCK;
7449         vulkanFeatures4.extFloat16Int8  = EXTFLOAT16INT8FEATURES_FLOAT16;
7450
7451         outputColors4[0]                        = RGBA(127, 127, 127, 255);
7452         outputColors4[1]                        = RGBA(127, 0,   0,   255);
7453         outputColors4[2]                        = RGBA(0,   0,   127, 255);
7454         outputColors4[3]                        = RGBA(0,   127, 0,   255);
7455
7456         createTestsForAllStages("swap16", inputColors, outputColors4, fragments4, resources4, extensions4, group.get(), vulkanFeatures4);
7457
7458         return group.release();
7459 }
7460
7461 tcu::TestCaseGroup* createNoContractionTests(tcu::TestContext& testCtx)
7462 {
7463         de::MovePtr<tcu::TestCaseGroup> group                   (new tcu::TestCaseGroup(testCtx, "nocontraction", "Test the NoContraction decoration"));
7464         RGBA                                                    inputColors[4];
7465         RGBA                                                    outputColors[4];
7466
7467         // With NoContraction, (1 + 2^-23) * (1 - 2^-23) - 1 should be conducted as a multiplication and an addition separately.
7468         // For the multiplication, the result is 1 - 2^-46, which is out of the precision range for 32-bit float. (32-bit float
7469         // only have 23-bit fraction.) So it will be rounded to 1. Or 0x1.fffffc. Then the final result is 0 or -0x1p-24.
7470         // On the contrary, the result will be 2^-46, which is a normalized number perfectly representable as 32-bit float.
7471         const char                                              constantsAndTypes[]      =
7472                 "%c_vec4_0       = OpConstantComposite %v4f32 %c_f32_0 %c_f32_0 %c_f32_0 %c_f32_1\n"
7473                 "%c_vec4_1       = OpConstantComposite %v4f32 %c_f32_1 %c_f32_1 %c_f32_1 %c_f32_1\n"
7474                 "%c_f32_1pl2_23  = OpConstant %f32 0x1.000002p+0\n" // 1 + 2^-23
7475                 "%c_f32_1mi2_23  = OpConstant %f32 0x1.fffffcp-1\n" // 1 - 2^-23
7476                 "%c_f32_n1pn24   = OpConstant %f32 -0x1p-24\n";
7477
7478         const char                                              function[]       =
7479                 "%test_code      = OpFunction %v4f32 None %v4f32_v4f32_function\n"
7480                 "%param          = OpFunctionParameter %v4f32\n"
7481                 "%label          = OpLabel\n"
7482                 "%var1           = OpVariable %fp_f32 Function %c_f32_1pl2_23\n"
7483                 "%var2           = OpVariable %fp_f32 Function\n"
7484                 "%red            = OpCompositeExtract %f32 %param 0\n"
7485                 "%plus_red       = OpFAdd %f32 %c_f32_1mi2_23 %red\n"
7486                 "                  OpStore %var2 %plus_red\n"
7487                 "%val1           = OpLoad %f32 %var1\n"
7488                 "%val2           = OpLoad %f32 %var2\n"
7489                 "%mul            = OpFMul %f32 %val1 %val2\n"
7490                 "%add            = OpFAdd %f32 %mul %c_f32_n1\n"
7491                 "%is0            = OpFOrdEqual %bool %add %c_f32_0\n"
7492                 "%isn1n24         = OpFOrdEqual %bool %add %c_f32_n1pn24\n"
7493                 "%success        = OpLogicalOr %bool %is0 %isn1n24\n"
7494                 "%v4success      = OpCompositeConstruct %v4bool %success %success %success %success\n"
7495                 "%ret            = OpSelect %v4f32 %v4success %c_vec4_0 %c_vec4_1\n"
7496                 "                  OpReturnValue %ret\n"
7497                 "                  OpFunctionEnd\n";
7498
7499         struct CaseNameDecoration
7500         {
7501                 string name;
7502                 string decoration;
7503         };
7504
7505
7506         CaseNameDecoration tests[] = {
7507                 {"multiplication",      "OpDecorate %mul NoContraction"},
7508                 {"addition",            "OpDecorate %add NoContraction"},
7509                 {"both",                        "OpDecorate %mul NoContraction\nOpDecorate %add NoContraction"},
7510         };
7511
7512         getHalfColorsFullAlpha(inputColors);
7513
7514         for (deUint8 idx = 0; idx < 4; ++idx)
7515         {
7516                 inputColors[idx].setRed(0);
7517                 outputColors[idx] = RGBA(0, 0, 0, 255);
7518         }
7519
7520         for (size_t testNdx = 0; testNdx < sizeof(tests) / sizeof(CaseNameDecoration); ++testNdx)
7521         {
7522                 map<string, string> fragments;
7523
7524                 fragments["decoration"] = tests[testNdx].decoration;
7525                 fragments["pre_main"] = constantsAndTypes;
7526                 fragments["testfun"] = function;
7527
7528                 createTestsForAllStages(tests[testNdx].name, inputColors, outputColors, fragments, group.get());
7529         }
7530
7531         return group.release();
7532 }
7533
7534 tcu::TestCaseGroup* createMemoryAccessTests(tcu::TestContext& testCtx)
7535 {
7536         de::MovePtr<tcu::TestCaseGroup> memoryAccessTests (new tcu::TestCaseGroup(testCtx, "opmemoryaccess", "Memory Semantics"));
7537         RGBA                                                    colors[4];
7538
7539         const char                                              constantsAndTypes[]      =
7540                 "%c_a2f32_1         = OpConstantComposite %a2f32 %c_f32_1 %c_f32_1\n"
7541                 "%fp_a2f32          = OpTypePointer Function %a2f32\n"
7542                 "%stype             = OpTypeStruct  %v4f32 %a2f32 %f32\n"
7543                 "%fp_stype          = OpTypePointer Function %stype\n";
7544
7545         const char                                              function[]       =
7546                 "%test_code         = OpFunction %v4f32 None %v4f32_v4f32_function\n"
7547                 "%param1            = OpFunctionParameter %v4f32\n"
7548                 "%lbl               = OpLabel\n"
7549                 "%v1                = OpVariable %fp_v4f32 Function\n"
7550                 "%v2                = OpVariable %fp_a2f32 Function\n"
7551                 "%v3                = OpVariable %fp_f32 Function\n"
7552                 "%v                 = OpVariable %fp_stype Function\n"
7553                 "%vv                = OpVariable %fp_stype Function\n"
7554                 "%vvv               = OpVariable %fp_f32 Function\n"
7555
7556                 "                     OpStore %v1 %c_v4f32_1_1_1_1\n"
7557                 "                     OpStore %v2 %c_a2f32_1\n"
7558                 "                     OpStore %v3 %c_f32_1\n"
7559
7560                 "%p_v4f32          = OpAccessChain %fp_v4f32 %v %c_u32_0\n"
7561                 "%p_a2f32          = OpAccessChain %fp_a2f32 %v %c_u32_1\n"
7562                 "%p_f32            = OpAccessChain %fp_f32 %v %c_u32_2\n"
7563                 "%v1_v             = OpLoad %v4f32 %v1 ${access_type}\n"
7564                 "%v2_v             = OpLoad %a2f32 %v2 ${access_type}\n"
7565                 "%v3_v             = OpLoad %f32 %v3 ${access_type}\n"
7566
7567                 "                    OpStore %p_v4f32 %v1_v ${access_type}\n"
7568                 "                    OpStore %p_a2f32 %v2_v ${access_type}\n"
7569                 "                    OpStore %p_f32 %v3_v ${access_type}\n"
7570
7571                 "                    OpCopyMemory %vv %v ${access_type}\n"
7572                 "                    OpCopyMemory %vvv %p_f32 ${access_type}\n"
7573
7574                 "%p_f32_2          = OpAccessChain %fp_f32 %vv %c_u32_2\n"
7575                 "%v_f32_2          = OpLoad %f32 %p_f32_2\n"
7576                 "%v_f32_3          = OpLoad %f32 %vvv\n"
7577
7578                 "%ret1             = OpVectorTimesScalar %v4f32 %param1 %v_f32_2\n"
7579                 "%ret2             = OpVectorTimesScalar %v4f32 %ret1 %v_f32_3\n"
7580                 "                    OpReturnValue %ret2\n"
7581                 "                    OpFunctionEnd\n";
7582
7583         struct NameMemoryAccess
7584         {
7585                 string name;
7586                 string accessType;
7587         };
7588
7589
7590         NameMemoryAccess tests[] =
7591         {
7592                 { "none", "" },
7593                 { "volatile", "Volatile" },
7594                 { "aligned",  "Aligned 1" },
7595                 { "volatile_aligned",  "Volatile|Aligned 1" },
7596                 { "nontemporal_aligned",  "Nontemporal|Aligned 1" },
7597                 { "volatile_nontemporal",  "Volatile|Nontemporal" },
7598                 { "volatile_nontermporal_aligned",  "Volatile|Nontemporal|Aligned 1" },
7599         };
7600
7601         getHalfColorsFullAlpha(colors);
7602
7603         for (size_t testNdx = 0; testNdx < sizeof(tests) / sizeof(NameMemoryAccess); ++testNdx)
7604         {
7605                 map<string, string> fragments;
7606                 map<string, string> memoryAccess;
7607                 memoryAccess["access_type"] = tests[testNdx].accessType;
7608
7609                 fragments["pre_main"] = constantsAndTypes;
7610                 fragments["testfun"] = tcu::StringTemplate(function).specialize(memoryAccess);
7611                 createTestsForAllStages(tests[testNdx].name, colors, colors, fragments, memoryAccessTests.get());
7612         }
7613         return memoryAccessTests.release();
7614 }
7615 tcu::TestCaseGroup* createOpUndefTests(tcu::TestContext& testCtx)
7616 {
7617         de::MovePtr<tcu::TestCaseGroup>         opUndefTests             (new tcu::TestCaseGroup(testCtx, "opundef", "Test OpUndef"));
7618         RGBA                                                            defaultColors[4];
7619         map<string, string>                                     fragments;
7620         getDefaultColors(defaultColors);
7621
7622         // First, simple cases that don't do anything with the OpUndef result.
7623         struct NameCodePair { string name, decl, type; };
7624         const NameCodePair tests[] =
7625         {
7626                 {"bool", "", "%bool"},
7627                 {"vec2uint32", "", "%v2u32"},
7628                 {"image", "%type = OpTypeImage %f32 2D 0 0 0 1 Unknown", "%type"},
7629                 {"sampler", "%type = OpTypeSampler", "%type"},
7630                 {"sampledimage", "%img = OpTypeImage %f32 2D 0 0 0 1 Unknown\n" "%type = OpTypeSampledImage %img", "%type"},
7631                 {"pointer", "", "%fp_i32"},
7632                 {"runtimearray", "%type = OpTypeRuntimeArray %f32", "%type"},
7633                 {"array", "%c_u32_100 = OpConstant %u32 100\n" "%type = OpTypeArray %i32 %c_u32_100", "%type"},
7634                 {"struct", "%type = OpTypeStruct %f32 %i32 %u32", "%type"}};
7635         for (size_t testNdx = 0; testNdx < sizeof(tests) / sizeof(NameCodePair); ++testNdx)
7636         {
7637                 fragments["undef_type"] = tests[testNdx].type;
7638                 fragments["testfun"] = StringTemplate(
7639                         "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
7640                         "%param1 = OpFunctionParameter %v4f32\n"
7641                         "%label_testfun = OpLabel\n"
7642                         "%undef = OpUndef ${undef_type}\n"
7643                         "OpReturnValue %param1\n"
7644                         "OpFunctionEnd\n").specialize(fragments);
7645                 fragments["pre_main"] = tests[testNdx].decl;
7646                 createTestsForAllStages(tests[testNdx].name, defaultColors, defaultColors, fragments, opUndefTests.get());
7647         }
7648         fragments.clear();
7649
7650         fragments["testfun"] =
7651                 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
7652                 "%param1 = OpFunctionParameter %v4f32\n"
7653                 "%label_testfun = OpLabel\n"
7654                 "%undef = OpUndef %f32\n"
7655                 "%zero = OpFMul %f32 %undef %c_f32_0\n"
7656                 "%is_nan = OpIsNan %bool %zero\n" //OpUndef may result in NaN which may turn %zero into Nan.
7657                 "%actually_zero = OpSelect %f32 %is_nan %c_f32_0 %zero\n"
7658                 "%a = OpVectorExtractDynamic %f32 %param1 %c_i32_0\n"
7659                 "%b = OpFAdd %f32 %a %actually_zero\n"
7660                 "%ret = OpVectorInsertDynamic %v4f32 %param1 %b %c_i32_0\n"
7661                 "OpReturnValue %ret\n"
7662                 "OpFunctionEnd\n";
7663
7664         createTestsForAllStages("float32", defaultColors, defaultColors, fragments, opUndefTests.get());
7665
7666         fragments["testfun"] =
7667                 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
7668                 "%param1 = OpFunctionParameter %v4f32\n"
7669                 "%label_testfun = OpLabel\n"
7670                 "%undef = OpUndef %i32\n"
7671                 "%zero = OpIMul %i32 %undef %c_i32_0\n"
7672                 "%a = OpVectorExtractDynamic %f32 %param1 %zero\n"
7673                 "%ret = OpVectorInsertDynamic %v4f32 %param1 %a %c_i32_0\n"
7674                 "OpReturnValue %ret\n"
7675                 "OpFunctionEnd\n";
7676
7677         createTestsForAllStages("sint32", defaultColors, defaultColors, fragments, opUndefTests.get());
7678
7679         fragments["testfun"] =
7680                 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
7681                 "%param1 = OpFunctionParameter %v4f32\n"
7682                 "%label_testfun = OpLabel\n"
7683                 "%undef = OpUndef %u32\n"
7684                 "%zero = OpIMul %u32 %undef %c_i32_0\n"
7685                 "%a = OpVectorExtractDynamic %f32 %param1 %zero\n"
7686                 "%ret = OpVectorInsertDynamic %v4f32 %param1 %a %c_i32_0\n"
7687                 "OpReturnValue %ret\n"
7688                 "OpFunctionEnd\n";
7689
7690         createTestsForAllStages("uint32", defaultColors, defaultColors, fragments, opUndefTests.get());
7691
7692         fragments["testfun"] =
7693                 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
7694                 "%param1 = OpFunctionParameter %v4f32\n"
7695                 "%label_testfun = OpLabel\n"
7696                 "%undef = OpUndef %v4f32\n"
7697                 "%vzero = OpVectorTimesScalar %v4f32 %undef %c_f32_0\n"
7698                 "%zero_0 = OpVectorExtractDynamic %f32 %vzero %c_i32_0\n"
7699                 "%zero_1 = OpVectorExtractDynamic %f32 %vzero %c_i32_1\n"
7700                 "%zero_2 = OpVectorExtractDynamic %f32 %vzero %c_i32_2\n"
7701                 "%zero_3 = OpVectorExtractDynamic %f32 %vzero %c_i32_3\n"
7702                 "%is_nan_0 = OpIsNan %bool %zero_0\n"
7703                 "%is_nan_1 = OpIsNan %bool %zero_1\n"
7704                 "%is_nan_2 = OpIsNan %bool %zero_2\n"
7705                 "%is_nan_3 = OpIsNan %bool %zero_3\n"
7706                 "%actually_zero_0 = OpSelect %f32 %is_nan_0 %c_f32_0 %zero_0\n"
7707                 "%actually_zero_1 = OpSelect %f32 %is_nan_1 %c_f32_0 %zero_1\n"
7708                 "%actually_zero_2 = OpSelect %f32 %is_nan_2 %c_f32_0 %zero_2\n"
7709                 "%actually_zero_3 = OpSelect %f32 %is_nan_3 %c_f32_0 %zero_3\n"
7710                 "%param1_0 = OpVectorExtractDynamic %f32 %param1 %c_i32_0\n"
7711                 "%param1_1 = OpVectorExtractDynamic %f32 %param1 %c_i32_1\n"
7712                 "%param1_2 = OpVectorExtractDynamic %f32 %param1 %c_i32_2\n"
7713                 "%param1_3 = OpVectorExtractDynamic %f32 %param1 %c_i32_3\n"
7714                 "%sum_0 = OpFAdd %f32 %param1_0 %actually_zero_0\n"
7715                 "%sum_1 = OpFAdd %f32 %param1_1 %actually_zero_1\n"
7716                 "%sum_2 = OpFAdd %f32 %param1_2 %actually_zero_2\n"
7717                 "%sum_3 = OpFAdd %f32 %param1_3 %actually_zero_3\n"
7718                 "%ret3 = OpVectorInsertDynamic %v4f32 %param1 %sum_3 %c_i32_3\n"
7719                 "%ret2 = OpVectorInsertDynamic %v4f32 %ret3 %sum_2 %c_i32_2\n"
7720                 "%ret1 = OpVectorInsertDynamic %v4f32 %ret2 %sum_1 %c_i32_1\n"
7721                 "%ret = OpVectorInsertDynamic %v4f32 %ret1 %sum_0 %c_i32_0\n"
7722                 "OpReturnValue %ret\n"
7723                 "OpFunctionEnd\n";
7724
7725         createTestsForAllStages("vec4float32", defaultColors, defaultColors, fragments, opUndefTests.get());
7726
7727         fragments["pre_main"] =
7728                 "%m2x2f32 = OpTypeMatrix %v2f32 2\n";
7729         fragments["testfun"] =
7730                 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
7731                 "%param1 = OpFunctionParameter %v4f32\n"
7732                 "%label_testfun = OpLabel\n"
7733                 "%undef = OpUndef %m2x2f32\n"
7734                 "%mzero = OpMatrixTimesScalar %m2x2f32 %undef %c_f32_0\n"
7735                 "%zero_0 = OpCompositeExtract %f32 %mzero 0 0\n"
7736                 "%zero_1 = OpCompositeExtract %f32 %mzero 0 1\n"
7737                 "%zero_2 = OpCompositeExtract %f32 %mzero 1 0\n"
7738                 "%zero_3 = OpCompositeExtract %f32 %mzero 1 1\n"
7739                 "%is_nan_0 = OpIsNan %bool %zero_0\n"
7740                 "%is_nan_1 = OpIsNan %bool %zero_1\n"
7741                 "%is_nan_2 = OpIsNan %bool %zero_2\n"
7742                 "%is_nan_3 = OpIsNan %bool %zero_3\n"
7743                 "%actually_zero_0 = OpSelect %f32 %is_nan_0 %c_f32_0 %zero_0\n"
7744                 "%actually_zero_1 = OpSelect %f32 %is_nan_1 %c_f32_0 %zero_1\n"
7745                 "%actually_zero_2 = OpSelect %f32 %is_nan_2 %c_f32_0 %zero_2\n"
7746                 "%actually_zero_3 = OpSelect %f32 %is_nan_3 %c_f32_0 %zero_3\n"
7747                 "%param1_0 = OpVectorExtractDynamic %f32 %param1 %c_i32_0\n"
7748                 "%param1_1 = OpVectorExtractDynamic %f32 %param1 %c_i32_1\n"
7749                 "%param1_2 = OpVectorExtractDynamic %f32 %param1 %c_i32_2\n"
7750                 "%param1_3 = OpVectorExtractDynamic %f32 %param1 %c_i32_3\n"
7751                 "%sum_0 = OpFAdd %f32 %param1_0 %actually_zero_0\n"
7752                 "%sum_1 = OpFAdd %f32 %param1_1 %actually_zero_1\n"
7753                 "%sum_2 = OpFAdd %f32 %param1_2 %actually_zero_2\n"
7754                 "%sum_3 = OpFAdd %f32 %param1_3 %actually_zero_3\n"
7755                 "%ret3 = OpVectorInsertDynamic %v4f32 %param1 %sum_3 %c_i32_3\n"
7756                 "%ret2 = OpVectorInsertDynamic %v4f32 %ret3 %sum_2 %c_i32_2\n"
7757                 "%ret1 = OpVectorInsertDynamic %v4f32 %ret2 %sum_1 %c_i32_1\n"
7758                 "%ret = OpVectorInsertDynamic %v4f32 %ret1 %sum_0 %c_i32_0\n"
7759                 "OpReturnValue %ret\n"
7760                 "OpFunctionEnd\n";
7761
7762         createTestsForAllStages("matrix", defaultColors, defaultColors, fragments, opUndefTests.get());
7763
7764         return opUndefTests.release();
7765 }
7766
7767 void createOpQuantizeSingleOptionTests(tcu::TestCaseGroup* testCtx)
7768 {
7769         const RGBA              inputColors[4]          =
7770         {
7771                 RGBA(0,         0,              0,              255),
7772                 RGBA(0,         0,              255,    255),
7773                 RGBA(0,         255,    0,              255),
7774                 RGBA(0,         255,    255,    255)
7775         };
7776
7777         const RGBA              expectedColors[4]       =
7778         {
7779                 RGBA(255,        0,              0,              255),
7780                 RGBA(255,        0,              0,              255),
7781                 RGBA(255,        0,              0,              255),
7782                 RGBA(255,        0,              0,              255)
7783         };
7784
7785         const struct SingleFP16Possibility
7786         {
7787                 const char* name;
7788                 const char* constant;  // Value to assign to %test_constant.
7789                 float           valueAsFloat;
7790                 const char* condition; // Must assign to %cond an expression that evaluates to true after %c = OpQuantizeToF16(%test_constant + 0).
7791         }                               tests[]                         =
7792         {
7793                 {
7794                         "negative",
7795                         "-0x1.3p1\n",
7796                         -constructNormalizedFloat(1, 0x300000),
7797                         "%cond = OpFOrdEqual %bool %c %test_constant\n"
7798                 }, // -19
7799                 {
7800                         "positive",
7801                         "0x1.0p7\n",
7802                         constructNormalizedFloat(7, 0x000000),
7803                         "%cond = OpFOrdEqual %bool %c %test_constant\n"
7804                 },  // +128
7805                 // SPIR-V requires that OpQuantizeToF16 flushes
7806                 // any numbers that would end up denormalized in F16 to zero.
7807                 {
7808                         "denorm",
7809                         "0x0.0006p-126\n",
7810                         std::ldexp(1.5f, -140),
7811                         "%cond = OpFOrdEqual %bool %c %c_f32_0\n"
7812                 },  // denorm
7813                 {
7814                         "negative_denorm",
7815                         "-0x0.0006p-126\n",
7816                         -std::ldexp(1.5f, -140),
7817                         "%cond = OpFOrdEqual %bool %c %c_f32_0\n"
7818                 }, // -denorm
7819                 {
7820                         "too_small",
7821                         "0x1.0p-16\n",
7822                         std::ldexp(1.0f, -16),
7823                         "%cond = OpFOrdEqual %bool %c %c_f32_0\n"
7824                 },     // too small positive
7825                 {
7826                         "negative_too_small",
7827                         "-0x1.0p-32\n",
7828                         -std::ldexp(1.0f, -32),
7829                         "%cond = OpFOrdEqual %bool %c %c_f32_0\n"
7830                 },      // too small negative
7831                 {
7832                         "negative_inf",
7833                         "-0x1.0p128\n",
7834                         -std::ldexp(1.0f, 128),
7835
7836                         "%gz = OpFOrdLessThan %bool %c %c_f32_0\n"
7837                         "%inf = OpIsInf %bool %c\n"
7838                         "%cond = OpLogicalAnd %bool %gz %inf\n"
7839                 },     // -inf to -inf
7840                 {
7841                         "inf",
7842                         "0x1.0p128\n",
7843                         std::ldexp(1.0f, 128),
7844
7845                         "%gz = OpFOrdGreaterThan %bool %c %c_f32_0\n"
7846                         "%inf = OpIsInf %bool %c\n"
7847                         "%cond = OpLogicalAnd %bool %gz %inf\n"
7848                 },     // +inf to +inf
7849                 {
7850                         "round_to_negative_inf",
7851                         "-0x1.0p32\n",
7852                         -std::ldexp(1.0f, 32),
7853
7854                         "%gz = OpFOrdLessThan %bool %c %c_f32_0\n"
7855                         "%inf = OpIsInf %bool %c\n"
7856                         "%cond = OpLogicalAnd %bool %gz %inf\n"
7857                 },     // round to -inf
7858                 {
7859                         "round_to_inf",
7860                         "0x1.0p16\n",
7861                         std::ldexp(1.0f, 16),
7862
7863                         "%gz = OpFOrdGreaterThan %bool %c %c_f32_0\n"
7864                         "%inf = OpIsInf %bool %c\n"
7865                         "%cond = OpLogicalAnd %bool %gz %inf\n"
7866                 },     // round to +inf
7867                 {
7868                         "nan",
7869                         "0x1.1p128\n",
7870                         std::numeric_limits<float>::quiet_NaN(),
7871
7872                         // Test for any NaN value, as NaNs are not preserved
7873                         "%direct_quant = OpQuantizeToF16 %f32 %test_constant\n"
7874                         "%cond = OpIsNan %bool %direct_quant\n"
7875                 }, // nan
7876                 {
7877                         "negative_nan",
7878                         "-0x1.0001p128\n",
7879                         std::numeric_limits<float>::quiet_NaN(),
7880
7881                         // Test for any NaN value, as NaNs are not preserved
7882                         "%direct_quant = OpQuantizeToF16 %f32 %test_constant\n"
7883                         "%cond = OpIsNan %bool %direct_quant\n"
7884                 } // -nan
7885         };
7886         const char*             constants                       =
7887                 "%test_constant = OpConstant %f32 ";  // The value will be test.constant.
7888
7889         StringTemplate  function                        (
7890                 "%test_code     = OpFunction %v4f32 None %v4f32_v4f32_function\n"
7891                 "%param1        = OpFunctionParameter %v4f32\n"
7892                 "%label_testfun = OpLabel\n"
7893                 "%a             = OpVectorExtractDynamic %f32 %param1 %c_i32_0\n"
7894                 "%b             = OpFAdd %f32 %test_constant %a\n"
7895                 "%c             = OpQuantizeToF16 %f32 %b\n"
7896                 "${condition}\n"
7897                 "%v4cond        = OpCompositeConstruct %v4bool %cond %cond %cond %cond\n"
7898                 "%retval        = OpSelect %v4f32 %v4cond %c_v4f32_1_0_0_1 %param1\n"
7899                 "                 OpReturnValue %retval\n"
7900                 "OpFunctionEnd\n"
7901         );
7902
7903         const char*             specDecorations         = "OpDecorate %test_constant SpecId 0\n";
7904         const char*             specConstants           =
7905                         "%test_constant = OpSpecConstant %f32 0.\n"
7906                         "%c             = OpSpecConstantOp %f32 QuantizeToF16 %test_constant\n";
7907
7908         StringTemplate  specConstantFunction(
7909                 "%test_code     = OpFunction %v4f32 None %v4f32_v4f32_function\n"
7910                 "%param1        = OpFunctionParameter %v4f32\n"
7911                 "%label_testfun = OpLabel\n"
7912                 "${condition}\n"
7913                 "%v4cond        = OpCompositeConstruct %v4bool %cond %cond %cond %cond\n"
7914                 "%retval        = OpSelect %v4f32 %v4cond %c_v4f32_1_0_0_1 %param1\n"
7915                 "                 OpReturnValue %retval\n"
7916                 "OpFunctionEnd\n"
7917         );
7918
7919         for (size_t idx = 0; idx < (sizeof(tests)/sizeof(tests[0])); ++idx)
7920         {
7921                 map<string, string>                                                             codeSpecialization;
7922                 map<string, string>                                                             fragments;
7923                 codeSpecialization["condition"]                                 = tests[idx].condition;
7924                 fragments["testfun"]                                                    = function.specialize(codeSpecialization);
7925                 fragments["pre_main"]                                                   = string(constants) + tests[idx].constant + "\n";
7926                 createTestsForAllStages(tests[idx].name, inputColors, expectedColors, fragments, testCtx);
7927         }
7928
7929         for (size_t idx = 0; idx < (sizeof(tests)/sizeof(tests[0])); ++idx)
7930         {
7931                 map<string, string>                                                             codeSpecialization;
7932                 map<string, string>                                                             fragments;
7933                 SpecConstants                                                                   passConstants;
7934
7935                 codeSpecialization["condition"]                                 = tests[idx].condition;
7936                 fragments["testfun"]                                                    = specConstantFunction.specialize(codeSpecialization);
7937                 fragments["decoration"]                                                 = specDecorations;
7938                 fragments["pre_main"]                                                   = specConstants;
7939
7940                 passConstants.append<float>(tests[idx].valueAsFloat);
7941
7942                 createTestsForAllStages(string("spec_const_") + tests[idx].name, inputColors, expectedColors, fragments, passConstants, testCtx);
7943         }
7944 }
7945
7946 void createOpQuantizeTwoPossibilityTests(tcu::TestCaseGroup* testCtx)
7947 {
7948         RGBA inputColors[4] =  {
7949                 RGBA(0,         0,              0,              255),
7950                 RGBA(0,         0,              255,    255),
7951                 RGBA(0,         255,    0,              255),
7952                 RGBA(0,         255,    255,    255)
7953         };
7954
7955         RGBA expectedColors[4] =
7956         {
7957                 RGBA(255,        0,              0,              255),
7958                 RGBA(255,        0,              0,              255),
7959                 RGBA(255,        0,              0,              255),
7960                 RGBA(255,        0,              0,              255)
7961         };
7962
7963         struct DualFP16Possibility
7964         {
7965                 const char* name;
7966                 const char* input;
7967                 float           inputAsFloat;
7968                 const char* possibleOutput1;
7969                 const char* possibleOutput2;
7970         } tests[] = {
7971                 {
7972                         "positive_round_up_or_round_down",
7973                         "0x1.3003p8",
7974                         constructNormalizedFloat(8, 0x300300),
7975                         "0x1.304p8",
7976                         "0x1.3p8"
7977                 },
7978                 {
7979                         "negative_round_up_or_round_down",
7980                         "-0x1.6008p-7",
7981                         -constructNormalizedFloat(-7, 0x600800),
7982                         "-0x1.6p-7",
7983                         "-0x1.604p-7"
7984                 },
7985                 {
7986                         "carry_bit",
7987                         "0x1.01ep2",
7988                         constructNormalizedFloat(2, 0x01e000),
7989                         "0x1.01cp2",
7990                         "0x1.02p2"
7991                 },
7992                 {
7993                         "carry_to_exponent",
7994                         "0x1.ffep1",
7995                         constructNormalizedFloat(1, 0xffe000),
7996                         "0x1.ffcp1",
7997                         "0x1.0p2"
7998                 },
7999         };
8000         StringTemplate constants (
8001                 "%input_const = OpConstant %f32 ${input}\n"
8002                 "%possible_solution1 = OpConstant %f32 ${output1}\n"
8003                 "%possible_solution2 = OpConstant %f32 ${output2}\n"
8004                 );
8005
8006         StringTemplate specConstants (
8007                 "%input_const = OpSpecConstant %f32 0.\n"
8008                 "%possible_solution1 = OpConstant %f32 ${output1}\n"
8009                 "%possible_solution2 = OpConstant %f32 ${output2}\n"
8010         );
8011
8012         const char* specDecorations = "OpDecorate %input_const  SpecId 0\n";
8013
8014         const char* function  =
8015                 "%test_code     = OpFunction %v4f32 None %v4f32_v4f32_function\n"
8016                 "%param1        = OpFunctionParameter %v4f32\n"
8017                 "%label_testfun = OpLabel\n"
8018                 "%a             = OpVectorExtractDynamic %f32 %param1 %c_i32_0\n"
8019                 // For the purposes of this test we assume that 0.f will always get
8020                 // faithfully passed through the pipeline stages.
8021                 "%b             = OpFAdd %f32 %input_const %a\n"
8022                 "%c             = OpQuantizeToF16 %f32 %b\n"
8023                 "%eq_1          = OpFOrdEqual %bool %c %possible_solution1\n"
8024                 "%eq_2          = OpFOrdEqual %bool %c %possible_solution2\n"
8025                 "%cond          = OpLogicalOr %bool %eq_1 %eq_2\n"
8026                 "%v4cond        = OpCompositeConstruct %v4bool %cond %cond %cond %cond\n"
8027                 "%retval        = OpSelect %v4f32 %v4cond %c_v4f32_1_0_0_1 %param1"
8028                 "                 OpReturnValue %retval\n"
8029                 "OpFunctionEnd\n";
8030
8031         for(size_t idx = 0; idx < (sizeof(tests)/sizeof(tests[0])); ++idx) {
8032                 map<string, string>                                                                     fragments;
8033                 map<string, string>                                                                     constantSpecialization;
8034
8035                 constantSpecialization["input"]                                         = tests[idx].input;
8036                 constantSpecialization["output1"]                                       = tests[idx].possibleOutput1;
8037                 constantSpecialization["output2"]                                       = tests[idx].possibleOutput2;
8038                 fragments["testfun"]                                                            = function;
8039                 fragments["pre_main"]                                                           = constants.specialize(constantSpecialization);
8040                 createTestsForAllStages(tests[idx].name, inputColors, expectedColors, fragments, testCtx);
8041         }
8042
8043         for(size_t idx = 0; idx < (sizeof(tests)/sizeof(tests[0])); ++idx) {
8044                 map<string, string>                                                                     fragments;
8045                 map<string, string>                                                                     constantSpecialization;
8046                 SpecConstants                                                                           passConstants;
8047
8048                 constantSpecialization["output1"]                                       = tests[idx].possibleOutput1;
8049                 constantSpecialization["output2"]                                       = tests[idx].possibleOutput2;
8050                 fragments["testfun"]                                                            = function;
8051                 fragments["decoration"]                                                         = specDecorations;
8052                 fragments["pre_main"]                                                           = specConstants.specialize(constantSpecialization);
8053
8054                 passConstants.append<float>(tests[idx].inputAsFloat);
8055
8056                 createTestsForAllStages(string("spec_const_") + tests[idx].name, inputColors, expectedColors, fragments, passConstants, testCtx);
8057         }
8058 }
8059
8060 tcu::TestCaseGroup* createOpQuantizeTests(tcu::TestContext& testCtx)
8061 {
8062         de::MovePtr<tcu::TestCaseGroup> opQuantizeTests (new tcu::TestCaseGroup(testCtx, "opquantize", "Test OpQuantizeToF16"));
8063         createOpQuantizeSingleOptionTests(opQuantizeTests.get());
8064         createOpQuantizeTwoPossibilityTests(opQuantizeTests.get());
8065         return opQuantizeTests.release();
8066 }
8067
8068 struct ShaderPermutation
8069 {
8070         deUint8 vertexPermutation;
8071         deUint8 geometryPermutation;
8072         deUint8 tesscPermutation;
8073         deUint8 tessePermutation;
8074         deUint8 fragmentPermutation;
8075 };
8076
8077 ShaderPermutation getShaderPermutation(deUint8 inputValue)
8078 {
8079         ShaderPermutation       permutation =
8080         {
8081                 static_cast<deUint8>(inputValue & 0x10? 1u: 0u),
8082                 static_cast<deUint8>(inputValue & 0x08? 1u: 0u),
8083                 static_cast<deUint8>(inputValue & 0x04? 1u: 0u),
8084                 static_cast<deUint8>(inputValue & 0x02? 1u: 0u),
8085                 static_cast<deUint8>(inputValue & 0x01? 1u: 0u)
8086         };
8087         return permutation;
8088 }
8089
8090 tcu::TestCaseGroup* createModuleTests(tcu::TestContext& testCtx)
8091 {
8092         RGBA                                                            defaultColors[4];
8093         RGBA                                                            invertedColors[4];
8094         de::MovePtr<tcu::TestCaseGroup>         moduleTests                     (new tcu::TestCaseGroup(testCtx, "module", "Multiple entry points into shaders"));
8095
8096         getDefaultColors(defaultColors);
8097         getInvertedDefaultColors(invertedColors);
8098
8099         // Combined module tests
8100         {
8101                 // Shader stages: vertex and fragment
8102                 {
8103                         const ShaderElement combinedPipeline[]  =
8104                         {
8105                                 ShaderElement("module", "main", VK_SHADER_STAGE_VERTEX_BIT),
8106                                 ShaderElement("module", "main", VK_SHADER_STAGE_FRAGMENT_BIT)
8107                         };
8108
8109                         addFunctionCaseWithPrograms<InstanceContext>(
8110                                 moduleTests.get(), "same_module", "", createCombinedModule, runAndVerifyDefaultPipeline,
8111                                 createInstanceContext(combinedPipeline, map<string, string>()));
8112                 }
8113
8114                 // Shader stages: vertex, geometry and fragment
8115                 {
8116                         const ShaderElement combinedPipeline[]  =
8117                         {
8118                                 ShaderElement("module", "main", VK_SHADER_STAGE_VERTEX_BIT),
8119                                 ShaderElement("module", "main", VK_SHADER_STAGE_GEOMETRY_BIT),
8120                                 ShaderElement("module", "main", VK_SHADER_STAGE_FRAGMENT_BIT)
8121                         };
8122
8123                         addFunctionCaseWithPrograms<InstanceContext>(
8124                                 moduleTests.get(), "same_module_geom", "", createCombinedModule, runAndVerifyDefaultPipeline,
8125                                 createInstanceContext(combinedPipeline, map<string, string>()));
8126                 }
8127
8128                 // Shader stages: vertex, tessellation control, tessellation evaluation and fragment
8129                 {
8130                         const ShaderElement combinedPipeline[]  =
8131                         {
8132                                 ShaderElement("module", "main", VK_SHADER_STAGE_VERTEX_BIT),
8133                                 ShaderElement("module", "main", VK_SHADER_STAGE_TESSELLATION_CONTROL_BIT),
8134                                 ShaderElement("module", "main", VK_SHADER_STAGE_TESSELLATION_EVALUATION_BIT),
8135                                 ShaderElement("module", "main", VK_SHADER_STAGE_FRAGMENT_BIT)
8136                         };
8137
8138                         addFunctionCaseWithPrograms<InstanceContext>(
8139                                 moduleTests.get(), "same_module_tessc_tesse", "", createCombinedModule, runAndVerifyDefaultPipeline,
8140                                 createInstanceContext(combinedPipeline, map<string, string>()));
8141                 }
8142
8143                 // Shader stages: vertex, tessellation control, tessellation evaluation, geometry and fragment
8144                 {
8145                         const ShaderElement combinedPipeline[]  =
8146                         {
8147                                 ShaderElement("module", "main", VK_SHADER_STAGE_VERTEX_BIT),
8148                                 ShaderElement("module", "main", VK_SHADER_STAGE_TESSELLATION_CONTROL_BIT),
8149                                 ShaderElement("module", "main", VK_SHADER_STAGE_TESSELLATION_EVALUATION_BIT),
8150                                 ShaderElement("module", "main", VK_SHADER_STAGE_GEOMETRY_BIT),
8151                                 ShaderElement("module", "main", VK_SHADER_STAGE_FRAGMENT_BIT)
8152                         };
8153
8154                         addFunctionCaseWithPrograms<InstanceContext>(
8155                                 moduleTests.get(), "same_module_tessc_tesse_geom", "", createCombinedModule, runAndVerifyDefaultPipeline,
8156                                 createInstanceContext(combinedPipeline, map<string, string>()));
8157                 }
8158         }
8159
8160         const char* numbers[] =
8161         {
8162                 "1", "2"
8163         };
8164
8165         for (deInt8 idx = 0; idx < 32; ++idx)
8166         {
8167                 ShaderPermutation                       permutation             = getShaderPermutation(idx);
8168                 string                                          name                    = string("vert") + numbers[permutation.vertexPermutation] + "_geom" + numbers[permutation.geometryPermutation] + "_tessc" + numbers[permutation.tesscPermutation] + "_tesse" + numbers[permutation.tessePermutation] + "_frag" + numbers[permutation.fragmentPermutation];
8169                 const ShaderElement                     pipeline[]              =
8170                 {
8171                         ShaderElement("vert",   string("vert") +        numbers[permutation.vertexPermutation],         VK_SHADER_STAGE_VERTEX_BIT),
8172                         ShaderElement("geom",   string("geom") +        numbers[permutation.geometryPermutation],       VK_SHADER_STAGE_GEOMETRY_BIT),
8173                         ShaderElement("tessc",  string("tessc") +       numbers[permutation.tesscPermutation],          VK_SHADER_STAGE_TESSELLATION_CONTROL_BIT),
8174                         ShaderElement("tesse",  string("tesse") +       numbers[permutation.tessePermutation],          VK_SHADER_STAGE_TESSELLATION_EVALUATION_BIT),
8175                         ShaderElement("frag",   string("frag") +        numbers[permutation.fragmentPermutation],       VK_SHADER_STAGE_FRAGMENT_BIT)
8176                 };
8177
8178                 // If there are an even number of swaps, then it should be no-op.
8179                 // If there are an odd number, the color should be flipped.
8180                 if ((permutation.vertexPermutation + permutation.geometryPermutation + permutation.tesscPermutation + permutation.tessePermutation + permutation.fragmentPermutation) % 2 == 0)
8181                 {
8182                         addFunctionCaseWithPrograms<InstanceContext>(
8183                                         moduleTests.get(), name, "", createMultipleEntries, runAndVerifyDefaultPipeline,
8184                                         createInstanceContext(pipeline, defaultColors, defaultColors, map<string, string>()));
8185                 }
8186                 else
8187                 {
8188                         addFunctionCaseWithPrograms<InstanceContext>(
8189                                         moduleTests.get(), name, "", createMultipleEntries, runAndVerifyDefaultPipeline,
8190                                         createInstanceContext(pipeline, defaultColors, invertedColors, map<string, string>()));
8191                 }
8192         }
8193         return moduleTests.release();
8194 }
8195
8196 tcu::TestCaseGroup* createLoopTests(tcu::TestContext& testCtx)
8197 {
8198         de::MovePtr<tcu::TestCaseGroup> testGroup(new tcu::TestCaseGroup(testCtx, "loop", "Looping control flow"));
8199         RGBA defaultColors[4];
8200         getDefaultColors(defaultColors);
8201         map<string, string> fragments;
8202         fragments["pre_main"] =
8203                 "%c_f32_5 = OpConstant %f32 5.\n";
8204
8205         // A loop with a single block. The Continue Target is the loop block
8206         // itself. In SPIR-V terms, the "loop construct" contains no blocks at all
8207         // -- the "continue construct" forms the entire loop.
8208         fragments["testfun"] =
8209                 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
8210                 "%param1 = OpFunctionParameter %v4f32\n"
8211
8212                 "%entry = OpLabel\n"
8213                 "%val0 = OpVectorExtractDynamic %f32 %param1 %c_i32_0\n"
8214                 "OpBranch %loop\n"
8215
8216                 ";adds and subtracts 1.0 to %val in alternate iterations\n"
8217                 "%loop = OpLabel\n"
8218                 "%count = OpPhi %i32 %c_i32_4 %entry %count__ %loop\n"
8219                 "%delta = OpPhi %f32 %c_f32_1 %entry %minus_delta %loop\n"
8220                 "%val1 = OpPhi %f32 %val0 %entry %val %loop\n"
8221                 "%val = OpFAdd %f32 %val1 %delta\n"
8222                 "%minus_delta = OpFSub %f32 %c_f32_0 %delta\n"
8223                 "%count__ = OpISub %i32 %count %c_i32_1\n"
8224                 "%again = OpSGreaterThan %bool %count__ %c_i32_0\n"
8225                 "OpLoopMerge %exit %loop None\n"
8226                 "OpBranchConditional %again %loop %exit\n"
8227
8228                 "%exit = OpLabel\n"
8229                 "%result = OpVectorInsertDynamic %v4f32 %param1 %val %c_i32_0\n"
8230                 "OpReturnValue %result\n"
8231
8232                 "OpFunctionEnd\n";
8233
8234         createTestsForAllStages("single_block", defaultColors, defaultColors, fragments, testGroup.get());
8235
8236         // Body comprised of multiple basic blocks.
8237         const StringTemplate multiBlock(
8238                 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
8239                 "%param1 = OpFunctionParameter %v4f32\n"
8240
8241                 "%entry = OpLabel\n"
8242                 "%val0 = OpVectorExtractDynamic %f32 %param1 %c_i32_0\n"
8243                 "OpBranch %loop\n"
8244
8245                 ";adds and subtracts 1.0 to %val in alternate iterations\n"
8246                 "%loop = OpLabel\n"
8247                 "%count = OpPhi %i32 %c_i32_4 %entry %count__ %gather\n"
8248                 "%delta = OpPhi %f32 %c_f32_1 %entry %delta_next %gather\n"
8249                 "%val1 = OpPhi %f32 %val0 %entry %val %gather\n"
8250                 // There are several possibilities for the Continue Target below.  Each
8251                 // will be specialized into a separate test case.
8252                 "OpLoopMerge %exit ${continue_target} None\n"
8253                 "OpBranch %if\n"
8254
8255                 "%if = OpLabel\n"
8256                 ";delta_next = (delta > 0) ? -1 : 1;\n"
8257                 "%gt0 = OpFOrdGreaterThan %bool %delta %c_f32_0\n"
8258                 "OpSelectionMerge %gather DontFlatten\n"
8259                 "OpBranchConditional %gt0 %even %odd ;tells us if %count is even or odd\n"
8260
8261                 "%odd = OpLabel\n"
8262                 "OpBranch %gather\n"
8263
8264                 "%even = OpLabel\n"
8265                 "OpBranch %gather\n"
8266
8267                 "%gather = OpLabel\n"
8268                 "%delta_next = OpPhi %f32 %c_f32_n1 %even %c_f32_1 %odd\n"
8269                 "%val = OpFAdd %f32 %val1 %delta\n"
8270                 "%count__ = OpISub %i32 %count %c_i32_1\n"
8271                 "%again = OpSGreaterThan %bool %count__ %c_i32_0\n"
8272                 "OpBranchConditional %again %loop %exit\n"
8273
8274                 "%exit = OpLabel\n"
8275                 "%result = OpVectorInsertDynamic %v4f32 %param1 %val %c_i32_0\n"
8276                 "OpReturnValue %result\n"
8277
8278                 "OpFunctionEnd\n");
8279
8280         map<string, string> continue_target;
8281
8282         // The Continue Target is the loop block itself.
8283         continue_target["continue_target"] = "%loop";
8284         fragments["testfun"] = multiBlock.specialize(continue_target);
8285         createTestsForAllStages("multi_block_continue_construct", defaultColors, defaultColors, fragments, testGroup.get());
8286
8287         // The Continue Target is at the end of the loop.
8288         continue_target["continue_target"] = "%gather";
8289         fragments["testfun"] = multiBlock.specialize(continue_target);
8290         createTestsForAllStages("multi_block_loop_construct", defaultColors, defaultColors, fragments, testGroup.get());
8291
8292         // A loop with continue statement.
8293         fragments["testfun"] =
8294                 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
8295                 "%param1 = OpFunctionParameter %v4f32\n"
8296
8297                 "%entry = OpLabel\n"
8298                 "%val0 = OpVectorExtractDynamic %f32 %param1 %c_i32_0\n"
8299                 "OpBranch %loop\n"
8300
8301                 ";adds 4, 3, and 1 to %val0 (skips 2)\n"
8302                 "%loop = OpLabel\n"
8303                 "%count = OpPhi %i32 %c_i32_4 %entry %count__ %continue\n"
8304                 "%val1 = OpPhi %f32 %val0 %entry %val %continue\n"
8305                 "OpLoopMerge %exit %continue None\n"
8306                 "OpBranch %if\n"
8307
8308                 "%if = OpLabel\n"
8309                 ";skip if %count==2\n"
8310                 "%eq2 = OpIEqual %bool %count %c_i32_2\n"
8311                 "OpSelectionMerge %continue DontFlatten\n"
8312                 "OpBranchConditional %eq2 %continue %body\n"
8313
8314                 "%body = OpLabel\n"
8315                 "%fcount = OpConvertSToF %f32 %count\n"
8316                 "%val2 = OpFAdd %f32 %val1 %fcount\n"
8317                 "OpBranch %continue\n"
8318
8319                 "%continue = OpLabel\n"
8320                 "%val = OpPhi %f32 %val2 %body %val1 %if\n"
8321                 "%count__ = OpISub %i32 %count %c_i32_1\n"
8322                 "%again = OpSGreaterThan %bool %count__ %c_i32_0\n"
8323                 "OpBranchConditional %again %loop %exit\n"
8324
8325                 "%exit = OpLabel\n"
8326                 "%same = OpFSub %f32 %val %c_f32_8\n"
8327                 "%result = OpVectorInsertDynamic %v4f32 %param1 %same %c_i32_0\n"
8328                 "OpReturnValue %result\n"
8329                 "OpFunctionEnd\n";
8330         createTestsForAllStages("continue", defaultColors, defaultColors, fragments, testGroup.get());
8331
8332         // A loop with break.
8333         fragments["testfun"] =
8334                 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
8335                 "%param1 = OpFunctionParameter %v4f32\n"
8336
8337                 "%entry = OpLabel\n"
8338                 ";param1 components are between 0 and 1, so dot product is 4 or less\n"
8339                 "%dot = OpDot %f32 %param1 %param1\n"
8340                 "%div = OpFDiv %f32 %dot %c_f32_5\n"
8341                 "%zero = OpConvertFToU %u32 %div\n"
8342                 "%two = OpIAdd %i32 %zero %c_i32_2\n"
8343                 "%val0 = OpVectorExtractDynamic %f32 %param1 %c_i32_0\n"
8344                 "OpBranch %loop\n"
8345
8346                 ";adds 4 and 3 to %val0 (exits early)\n"
8347                 "%loop = OpLabel\n"
8348                 "%count = OpPhi %i32 %c_i32_4 %entry %count__ %continue\n"
8349                 "%val1 = OpPhi %f32 %val0 %entry %val2 %continue\n"
8350                 "OpLoopMerge %exit %continue None\n"
8351                 "OpBranch %if\n"
8352
8353                 "%if = OpLabel\n"
8354                 ";end loop if %count==%two\n"
8355                 "%above2 = OpSGreaterThan %bool %count %two\n"
8356                 "OpSelectionMerge %continue DontFlatten\n"
8357                 "OpBranchConditional %above2 %body %exit\n"
8358
8359                 "%body = OpLabel\n"
8360                 "%fcount = OpConvertSToF %f32 %count\n"
8361                 "%val2 = OpFAdd %f32 %val1 %fcount\n"
8362                 "OpBranch %continue\n"
8363
8364                 "%continue = OpLabel\n"
8365                 "%count__ = OpISub %i32 %count %c_i32_1\n"
8366                 "%again = OpSGreaterThan %bool %count__ %c_i32_0\n"
8367                 "OpBranchConditional %again %loop %exit\n"
8368
8369                 "%exit = OpLabel\n"
8370                 "%val_post = OpPhi %f32 %val2 %continue %val1 %if\n"
8371                 "%same = OpFSub %f32 %val_post %c_f32_7\n"
8372                 "%result = OpVectorInsertDynamic %v4f32 %param1 %same %c_i32_0\n"
8373                 "OpReturnValue %result\n"
8374                 "OpFunctionEnd\n";
8375         createTestsForAllStages("break", defaultColors, defaultColors, fragments, testGroup.get());
8376
8377         // A loop with return.
8378         fragments["testfun"] =
8379                 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
8380                 "%param1 = OpFunctionParameter %v4f32\n"
8381
8382                 "%entry = OpLabel\n"
8383                 ";param1 components are between 0 and 1, so dot product is 4 or less\n"
8384                 "%dot = OpDot %f32 %param1 %param1\n"
8385                 "%div = OpFDiv %f32 %dot %c_f32_5\n"
8386                 "%zero = OpConvertFToU %u32 %div\n"
8387                 "%two = OpIAdd %i32 %zero %c_i32_2\n"
8388                 "%val0 = OpVectorExtractDynamic %f32 %param1 %c_i32_0\n"
8389                 "OpBranch %loop\n"
8390
8391                 ";returns early without modifying %param1\n"
8392                 "%loop = OpLabel\n"
8393                 "%count = OpPhi %i32 %c_i32_4 %entry %count__ %continue\n"
8394                 "%val1 = OpPhi %f32 %val0 %entry %val2 %continue\n"
8395                 "OpLoopMerge %exit %continue None\n"
8396                 "OpBranch %if\n"
8397
8398                 "%if = OpLabel\n"
8399                 ";return if %count==%two\n"
8400                 "%above2 = OpSGreaterThan %bool %count %two\n"
8401                 "OpSelectionMerge %continue DontFlatten\n"
8402                 "OpBranchConditional %above2 %body %early_exit\n"
8403
8404                 "%early_exit = OpLabel\n"
8405                 "OpReturnValue %param1\n"
8406
8407                 "%body = OpLabel\n"
8408                 "%fcount = OpConvertSToF %f32 %count\n"
8409                 "%val2 = OpFAdd %f32 %val1 %fcount\n"
8410                 "OpBranch %continue\n"
8411
8412                 "%continue = OpLabel\n"
8413                 "%count__ = OpISub %i32 %count %c_i32_1\n"
8414                 "%again = OpSGreaterThan %bool %count__ %c_i32_0\n"
8415                 "OpBranchConditional %again %loop %exit\n"
8416
8417                 "%exit = OpLabel\n"
8418                 ";should never get here, so return an incorrect result\n"
8419                 "%result = OpVectorInsertDynamic %v4f32 %param1 %val2 %c_i32_0\n"
8420                 "OpReturnValue %result\n"
8421                 "OpFunctionEnd\n";
8422         createTestsForAllStages("return", defaultColors, defaultColors, fragments, testGroup.get());
8423
8424         // Continue inside a switch block to break to enclosing loop's merge block.
8425         // Matches roughly the following GLSL code:
8426         // for (; keep_going; keep_going = false)
8427         // {
8428         //     switch (int(param1.x))
8429         //     {
8430         //         case 0: continue;
8431         //         case 1: continue;
8432         //         default: continue;
8433         //     }
8434         //     dead code: modify return value to invalid result.
8435         // }
8436         fragments["pre_main"] =
8437                 "%fp_bool = OpTypePointer Function %bool\n"
8438                 "%true = OpConstantTrue %bool\n"
8439                 "%false = OpConstantFalse %bool\n";
8440
8441         fragments["testfun"] =
8442                 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
8443                 "%param1 = OpFunctionParameter %v4f32\n"
8444
8445                 "%entry = OpLabel\n"
8446                 "%keep_going = OpVariable %fp_bool Function\n"
8447                 "%val_ptr = OpVariable %fp_f32 Function\n"
8448                 "%param1_x = OpCompositeExtract %f32 %param1 0\n"
8449                 "OpStore %keep_going %true\n"
8450                 "OpBranch %forloop_begin\n"
8451
8452                 "%forloop_begin = OpLabel\n"
8453                 "OpLoopMerge %forloop_merge %forloop_continue None\n"
8454                 "OpBranch %forloop\n"
8455
8456                 "%forloop = OpLabel\n"
8457                 "%for_condition = OpLoad %bool %keep_going\n"
8458                 "OpBranchConditional %for_condition %forloop_body %forloop_merge\n"
8459
8460                 "%forloop_body = OpLabel\n"
8461                 "OpStore %val_ptr %param1_x\n"
8462                 "%param1_x_int = OpConvertFToS %i32 %param1_x\n"
8463
8464                 "OpSelectionMerge %switch_merge None\n"
8465                 "OpSwitch %param1_x_int %default 0 %case_0 1 %case_1\n"
8466                 "%case_0 = OpLabel\n"
8467                 "OpBranch %forloop_continue\n"
8468                 "%case_1 = OpLabel\n"
8469                 "OpBranch %forloop_continue\n"
8470                 "%default = OpLabel\n"
8471                 "OpBranch %forloop_continue\n"
8472                 "%switch_merge = OpLabel\n"
8473                 ";should never get here, so change the return value to invalid result\n"
8474                 "OpStore %val_ptr %c_f32_1\n"
8475                 "OpBranch %forloop_continue\n"
8476
8477                 "%forloop_continue = OpLabel\n"
8478                 "OpStore %keep_going %false\n"
8479                 "OpBranch %forloop_begin\n"
8480                 "%forloop_merge = OpLabel\n"
8481
8482                 "%val = OpLoad %f32 %val_ptr\n"
8483                 "%result = OpVectorInsertDynamic %v4f32 %param1 %val %c_i32_0\n"
8484                 "OpReturnValue %result\n"
8485                 "OpFunctionEnd\n";
8486         createTestsForAllStages("switch_continue", defaultColors, defaultColors, fragments, testGroup.get());
8487
8488         return testGroup.release();
8489 }
8490
8491 // A collection of tests putting OpControlBarrier in places GLSL forbids but SPIR-V allows.
8492 tcu::TestCaseGroup* createBarrierTests(tcu::TestContext& testCtx)
8493 {
8494         de::MovePtr<tcu::TestCaseGroup> testGroup(new tcu::TestCaseGroup(testCtx, "barrier", "OpControlBarrier"));
8495         map<string, string> fragments;
8496
8497         // A barrier inside a function body.
8498         fragments["pre_main"] =
8499                 "%Workgroup = OpConstant %i32 2\n"
8500                 "%WorkgroupAcquireRelease = OpConstant %i32 0x108\n";
8501         fragments["testfun"] =
8502                 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
8503                 "%param1 = OpFunctionParameter %v4f32\n"
8504                 "%label_testfun = OpLabel\n"
8505                 "OpControlBarrier %Workgroup %Workgroup %WorkgroupAcquireRelease\n"
8506                 "OpReturnValue %param1\n"
8507                 "OpFunctionEnd\n";
8508         addTessCtrlTest(testGroup.get(), "in_function", fragments);
8509
8510         // Common setup code for the following tests.
8511         fragments["pre_main"] =
8512                 "%Workgroup = OpConstant %i32 2\n"
8513                 "%WorkgroupAcquireRelease = OpConstant %i32 0x108\n"
8514                 "%c_f32_5 = OpConstant %f32 5.\n";
8515         const string setupPercentZero =  // Begins %test_code function with code that sets %zero to 0u but cannot be optimized away.
8516                 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
8517                 "%param1 = OpFunctionParameter %v4f32\n"
8518                 "%entry = OpLabel\n"
8519                 ";param1 components are between 0 and 1, so dot product is 4 or less\n"
8520                 "%dot = OpDot %f32 %param1 %param1\n"
8521                 "%div = OpFDiv %f32 %dot %c_f32_5\n"
8522                 "%zero = OpConvertFToU %u32 %div\n";
8523
8524         // Barriers inside OpSwitch branches.
8525         fragments["testfun"] =
8526                 setupPercentZero +
8527                 "OpSelectionMerge %switch_exit None\n"
8528                 "OpSwitch %zero %switch_default 0 %case0 1 %case1 ;should always go to %case0\n"
8529
8530                 "%case1 = OpLabel\n"
8531                 ";This barrier should never be executed, but its presence makes test failure more likely when there's a bug.\n"
8532                 "OpControlBarrier %Workgroup %Workgroup %WorkgroupAcquireRelease\n"
8533                 "%wrong_branch_alert1 = OpVectorInsertDynamic %v4f32 %param1 %c_f32_0_5 %c_i32_0\n"
8534                 "OpBranch %switch_exit\n"
8535
8536                 "%switch_default = OpLabel\n"
8537                 "%wrong_branch_alert2 = OpVectorInsertDynamic %v4f32 %param1 %c_f32_0_5 %c_i32_0\n"
8538                 ";This barrier should never be executed, but its presence makes test failure more likely when there's a bug.\n"
8539                 "OpControlBarrier %Workgroup %Workgroup %WorkgroupAcquireRelease\n"
8540                 "OpBranch %switch_exit\n"
8541
8542                 "%case0 = OpLabel\n"
8543                 "OpControlBarrier %Workgroup %Workgroup %WorkgroupAcquireRelease\n"
8544                 "OpBranch %switch_exit\n"
8545
8546                 "%switch_exit = OpLabel\n"
8547                 "%ret = OpPhi %v4f32 %param1 %case0 %wrong_branch_alert1 %case1 %wrong_branch_alert2 %switch_default\n"
8548                 "OpReturnValue %ret\n"
8549                 "OpFunctionEnd\n";
8550         addTessCtrlTest(testGroup.get(), "in_switch", fragments);
8551
8552         // Barriers inside if-then-else.
8553         fragments["testfun"] =
8554                 setupPercentZero +
8555                 "%eq0 = OpIEqual %bool %zero %c_u32_0\n"
8556                 "OpSelectionMerge %exit DontFlatten\n"
8557                 "OpBranchConditional %eq0 %then %else\n"
8558
8559                 "%else = OpLabel\n"
8560                 ";This barrier should never be executed, but its presence makes test failure more likely when there's a bug.\n"
8561                 "OpControlBarrier %Workgroup %Workgroup %WorkgroupAcquireRelease\n"
8562                 "%wrong_branch_alert = OpVectorInsertDynamic %v4f32 %param1 %c_f32_0_5 %c_i32_0\n"
8563                 "OpBranch %exit\n"
8564
8565                 "%then = OpLabel\n"
8566                 "OpControlBarrier %Workgroup %Workgroup %WorkgroupAcquireRelease\n"
8567                 "OpBranch %exit\n"
8568                 "%exit = OpLabel\n"
8569                 "%ret = OpPhi %v4f32 %param1 %then %wrong_branch_alert %else\n"
8570                 "OpReturnValue %ret\n"
8571                 "OpFunctionEnd\n";
8572         addTessCtrlTest(testGroup.get(), "in_if", fragments);
8573
8574         // A barrier after control-flow reconvergence, tempting the compiler to attempt something like this:
8575         // http://lists.llvm.org/pipermail/llvm-dev/2009-October/026317.html.
8576         fragments["testfun"] =
8577                 setupPercentZero +
8578                 "%thread_id = OpLoad %i32 %BP_gl_InvocationID\n"
8579                 "%thread0 = OpIEqual %bool %thread_id %c_i32_0\n"
8580                 "OpSelectionMerge %exit DontFlatten\n"
8581                 "OpBranchConditional %thread0 %then %else\n"
8582
8583                 "%else = OpLabel\n"
8584                 "%val0 = OpVectorExtractDynamic %f32 %param1 %c_i32_0\n"
8585                 "OpBranch %exit\n"
8586
8587                 "%then = OpLabel\n"
8588                 "%val1 = OpVectorExtractDynamic %f32 %param1 %zero\n"
8589                 "OpBranch %exit\n"
8590
8591                 "%exit = OpLabel\n"
8592                 "%val = OpPhi %f32 %val0 %else %val1 %then\n"
8593                 "OpControlBarrier %Workgroup %Workgroup %WorkgroupAcquireRelease\n"
8594                 "%ret = OpVectorInsertDynamic %v4f32 %param1 %val %zero\n"
8595                 "OpReturnValue %ret\n"
8596                 "OpFunctionEnd\n";
8597         addTessCtrlTest(testGroup.get(), "after_divergent_if", fragments);
8598
8599         // A barrier inside a loop.
8600         fragments["pre_main"] =
8601                 "%Workgroup = OpConstant %i32 2\n"
8602                 "%WorkgroupAcquireRelease = OpConstant %i32 0x108\n"
8603                 "%c_f32_10 = OpConstant %f32 10.\n";
8604         fragments["testfun"] =
8605                 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
8606                 "%param1 = OpFunctionParameter %v4f32\n"
8607                 "%entry = OpLabel\n"
8608                 "%val0 = OpVectorExtractDynamic %f32 %param1 %c_i32_0\n"
8609                 "OpBranch %loop\n"
8610
8611                 ";adds 4, 3, 2, and 1 to %val0\n"
8612                 "%loop = OpLabel\n"
8613                 "%count = OpPhi %i32 %c_i32_4 %entry %count__ %loop\n"
8614                 "%val1 = OpPhi %f32 %val0 %entry %val %loop\n"
8615                 "OpControlBarrier %Workgroup %Workgroup %WorkgroupAcquireRelease\n"
8616                 "%fcount = OpConvertSToF %f32 %count\n"
8617                 "%val = OpFAdd %f32 %val1 %fcount\n"
8618                 "%count__ = OpISub %i32 %count %c_i32_1\n"
8619                 "%again = OpSGreaterThan %bool %count__ %c_i32_0\n"
8620                 "OpLoopMerge %exit %loop None\n"
8621                 "OpBranchConditional %again %loop %exit\n"
8622
8623                 "%exit = OpLabel\n"
8624                 "%same = OpFSub %f32 %val %c_f32_10\n"
8625                 "%ret = OpVectorInsertDynamic %v4f32 %param1 %same %c_i32_0\n"
8626                 "OpReturnValue %ret\n"
8627                 "OpFunctionEnd\n";
8628         addTessCtrlTest(testGroup.get(), "in_loop", fragments);
8629
8630         return testGroup.release();
8631 }
8632
8633 // Test for the OpFRem instruction.
8634 tcu::TestCaseGroup* createFRemTests(tcu::TestContext& testCtx)
8635 {
8636         de::MovePtr<tcu::TestCaseGroup>         testGroup(new tcu::TestCaseGroup(testCtx, "frem", "OpFRem"));
8637         map<string, string>                                     fragments;
8638         RGBA                                                            inputColors[4];
8639         RGBA                                                            outputColors[4];
8640
8641         fragments["pre_main"]                            =
8642                 "%c_f32_3 = OpConstant %f32 3.0\n"
8643                 "%c_f32_n3 = OpConstant %f32 -3.0\n"
8644                 "%c_f32_4 = OpConstant %f32 4.0\n"
8645                 "%c_f32_p75 = OpConstant %f32 0.75\n"
8646                 "%c_v4f32_p75_p75_p75_p75 = OpConstantComposite %v4f32 %c_f32_p75 %c_f32_p75 %c_f32_p75 %c_f32_p75 \n"
8647                 "%c_v4f32_4_4_4_4 = OpConstantComposite %v4f32 %c_f32_4 %c_f32_4 %c_f32_4 %c_f32_4\n"
8648                 "%c_v4f32_3_n3_3_n3 = OpConstantComposite %v4f32 %c_f32_3 %c_f32_n3 %c_f32_3 %c_f32_n3\n";
8649
8650         // The test does the following.
8651         // vec4 result = (param1 * 8.0) - 4.0;
8652         // return (frem(result.x,3) + 0.75, frem(result.y, -3) + 0.75, 0, 1)
8653         fragments["testfun"]                             =
8654                 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
8655                 "%param1 = OpFunctionParameter %v4f32\n"
8656                 "%label_testfun = OpLabel\n"
8657                 "%v_times_8 = OpVectorTimesScalar %v4f32 %param1 %c_f32_8\n"
8658                 "%minus_4 = OpFSub %v4f32 %v_times_8 %c_v4f32_4_4_4_4\n"
8659                 "%frem = OpFRem %v4f32 %minus_4 %c_v4f32_3_n3_3_n3\n"
8660                 "%added = OpFAdd %v4f32 %frem %c_v4f32_p75_p75_p75_p75\n"
8661                 "%xyz_1 = OpVectorInsertDynamic %v4f32 %added %c_f32_1 %c_i32_3\n"
8662                 "%xy_0_1 = OpVectorInsertDynamic %v4f32 %xyz_1 %c_f32_0 %c_i32_2\n"
8663                 "OpReturnValue %xy_0_1\n"
8664                 "OpFunctionEnd\n";
8665
8666
8667         inputColors[0]          = RGBA(16,      16,             0, 255);
8668         inputColors[1]          = RGBA(232, 232,        0, 255);
8669         inputColors[2]          = RGBA(232, 16,         0, 255);
8670         inputColors[3]          = RGBA(16,      232,    0, 255);
8671
8672         outputColors[0]         = RGBA(64,      64,             0, 255);
8673         outputColors[1]         = RGBA(255, 255,        0, 255);
8674         outputColors[2]         = RGBA(255, 64,         0, 255);
8675         outputColors[3]         = RGBA(64,      255,    0, 255);
8676
8677         createTestsForAllStages("frem", inputColors, outputColors, fragments, testGroup.get());
8678         return testGroup.release();
8679 }
8680
8681 // Test for the OpSRem instruction.
8682 tcu::TestCaseGroup* createOpSRemGraphicsTests(tcu::TestContext& testCtx, qpTestResult negFailResult)
8683 {
8684         de::MovePtr<tcu::TestCaseGroup>         testGroup(new tcu::TestCaseGroup(testCtx, "srem", "OpSRem"));
8685         map<string, string>                                     fragments;
8686
8687         fragments["pre_main"]                            =
8688                 "%c_f32_255 = OpConstant %f32 255.0\n"
8689                 "%c_i32_128 = OpConstant %i32 128\n"
8690                 "%c_i32_255 = OpConstant %i32 255\n"
8691                 "%c_v4f32_255 = OpConstantComposite %v4f32 %c_f32_255 %c_f32_255 %c_f32_255 %c_f32_255 \n"
8692                 "%c_v4f32_0_5 = OpConstantComposite %v4f32 %c_f32_0_5 %c_f32_0_5 %c_f32_0_5 %c_f32_0_5 \n"
8693                 "%c_v4i32_128 = OpConstantComposite %v4i32 %c_i32_128 %c_i32_128 %c_i32_128 %c_i32_128 \n";
8694
8695         // The test does the following.
8696         // ivec4 ints = int(param1 * 255.0 + 0.5) - 128;
8697         // ivec4 result = ivec4(srem(ints.x, ints.y), srem(ints.y, ints.z), srem(ints.z, ints.x), 255);
8698         // return float(result + 128) / 255.0;
8699         fragments["testfun"]                             =
8700                 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
8701                 "%param1 = OpFunctionParameter %v4f32\n"
8702                 "%label_testfun = OpLabel\n"
8703                 "%div255 = OpFMul %v4f32 %param1 %c_v4f32_255\n"
8704                 "%add0_5 = OpFAdd %v4f32 %div255 %c_v4f32_0_5\n"
8705                 "%uints_in = OpConvertFToS %v4i32 %add0_5\n"
8706                 "%ints_in = OpISub %v4i32 %uints_in %c_v4i32_128\n"
8707                 "%x_in = OpCompositeExtract %i32 %ints_in 0\n"
8708                 "%y_in = OpCompositeExtract %i32 %ints_in 1\n"
8709                 "%z_in = OpCompositeExtract %i32 %ints_in 2\n"
8710                 "%x_out = OpSRem %i32 %x_in %y_in\n"
8711                 "%y_out = OpSRem %i32 %y_in %z_in\n"
8712                 "%z_out = OpSRem %i32 %z_in %x_in\n"
8713                 "%ints_out = OpCompositeConstruct %v4i32 %x_out %y_out %z_out %c_i32_255\n"
8714                 "%ints_offset = OpIAdd %v4i32 %ints_out %c_v4i32_128\n"
8715                 "%f_ints_offset = OpConvertSToF %v4f32 %ints_offset\n"
8716                 "%float_out = OpFDiv %v4f32 %f_ints_offset %c_v4f32_255\n"
8717                 "OpReturnValue %float_out\n"
8718                 "OpFunctionEnd\n";
8719
8720         const struct CaseParams
8721         {
8722                 const char*             name;
8723                 const char*             failMessageTemplate;    // customized status message
8724                 qpTestResult    failResult;                             // override status on failure
8725                 int                             operands[4][3];                 // four (x, y, z) vectors of operands
8726                 int                             results[4][3];                  // four (x, y, z) vectors of results
8727         } cases[] =
8728         {
8729                 {
8730                         "positive",
8731                         "${reason}",
8732                         QP_TEST_RESULT_FAIL,
8733                         { { 5, 12, 17 }, { 5, 5, 7 }, { 75, 8, 81 }, { 25, 60, 100 } },                 // operands
8734                         { { 5, 12,  2 }, { 0, 5, 2 }, {  3, 8,  6 }, { 25, 60,   0 } },                 // results
8735                 },
8736                 {
8737                         "all",
8738                         "Inconsistent results, but within specification: ${reason}",
8739                         negFailResult,                                                                                                                  // negative operands, not required by the spec
8740                         { { 5, 12, -17 }, { -5, -5, 7 }, { 75, 8, -81 }, { 25, -60, 100 } },    // operands
8741                         { { 5, 12,  -2 }, {  0, -5, 2 }, {  3, 8,  -6 }, { 25, -60,   0 } },    // results
8742                 },
8743         };
8744         // If either operand is negative the result is undefined. Some implementations may still return correct values.
8745
8746         for (int caseNdx = 0; caseNdx < DE_LENGTH_OF_ARRAY(cases); ++caseNdx)
8747         {
8748                 const CaseParams&       params                  = cases[caseNdx];
8749                 RGBA                            inputColors[4];
8750                 RGBA                            outputColors[4];
8751
8752                 for (int i = 0; i < 4; ++i)
8753                 {
8754                         inputColors [i] = RGBA(params.operands[i][0] + 128, params.operands[i][1] + 128, params.operands[i][2] + 128, 255);
8755                         outputColors[i] = RGBA(params.results [i][0] + 128, params.results [i][1] + 128, params.results [i][2] + 128, 255);
8756                 }
8757
8758                 createTestsForAllStages(params.name, inputColors, outputColors, fragments, testGroup.get(), params.failResult, params.failMessageTemplate);
8759         }
8760
8761         return testGroup.release();
8762 }
8763
8764 // Test for the OpSMod instruction.
8765 tcu::TestCaseGroup* createOpSModGraphicsTests(tcu::TestContext& testCtx, qpTestResult negFailResult)
8766 {
8767         de::MovePtr<tcu::TestCaseGroup>         testGroup(new tcu::TestCaseGroup(testCtx, "smod", "OpSMod"));
8768         map<string, string>                                     fragments;
8769
8770         fragments["pre_main"]                            =
8771                 "%c_f32_255 = OpConstant %f32 255.0\n"
8772                 "%c_i32_128 = OpConstant %i32 128\n"
8773                 "%c_i32_255 = OpConstant %i32 255\n"
8774                 "%c_v4f32_255 = OpConstantComposite %v4f32 %c_f32_255 %c_f32_255 %c_f32_255 %c_f32_255 \n"
8775                 "%c_v4f32_0_5 = OpConstantComposite %v4f32 %c_f32_0_5 %c_f32_0_5 %c_f32_0_5 %c_f32_0_5 \n"
8776                 "%c_v4i32_128 = OpConstantComposite %v4i32 %c_i32_128 %c_i32_128 %c_i32_128 %c_i32_128 \n";
8777
8778         // The test does the following.
8779         // ivec4 ints = int(param1 * 255.0 + 0.5) - 128;
8780         // ivec4 result = ivec4(smod(ints.x, ints.y), smod(ints.y, ints.z), smod(ints.z, ints.x), 255);
8781         // return float(result + 128) / 255.0;
8782         fragments["testfun"]                             =
8783                 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
8784                 "%param1 = OpFunctionParameter %v4f32\n"
8785                 "%label_testfun = OpLabel\n"
8786                 "%div255 = OpFMul %v4f32 %param1 %c_v4f32_255\n"
8787                 "%add0_5 = OpFAdd %v4f32 %div255 %c_v4f32_0_5\n"
8788                 "%uints_in = OpConvertFToS %v4i32 %add0_5\n"
8789                 "%ints_in = OpISub %v4i32 %uints_in %c_v4i32_128\n"
8790                 "%x_in = OpCompositeExtract %i32 %ints_in 0\n"
8791                 "%y_in = OpCompositeExtract %i32 %ints_in 1\n"
8792                 "%z_in = OpCompositeExtract %i32 %ints_in 2\n"
8793                 "%x_out = OpSMod %i32 %x_in %y_in\n"
8794                 "%y_out = OpSMod %i32 %y_in %z_in\n"
8795                 "%z_out = OpSMod %i32 %z_in %x_in\n"
8796                 "%ints_out = OpCompositeConstruct %v4i32 %x_out %y_out %z_out %c_i32_255\n"
8797                 "%ints_offset = OpIAdd %v4i32 %ints_out %c_v4i32_128\n"
8798                 "%f_ints_offset = OpConvertSToF %v4f32 %ints_offset\n"
8799                 "%float_out = OpFDiv %v4f32 %f_ints_offset %c_v4f32_255\n"
8800                 "OpReturnValue %float_out\n"
8801                 "OpFunctionEnd\n";
8802
8803         const struct CaseParams
8804         {
8805                 const char*             name;
8806                 const char*             failMessageTemplate;    // customized status message
8807                 qpTestResult    failResult;                             // override status on failure
8808                 int                             operands[4][3];                 // four (x, y, z) vectors of operands
8809                 int                             results[4][3];                  // four (x, y, z) vectors of results
8810         } cases[] =
8811         {
8812                 {
8813                         "positive",
8814                         "${reason}",
8815                         QP_TEST_RESULT_FAIL,
8816                         { { 5, 12, 17 }, { 5, 5, 7 }, { 75, 8, 81 }, { 25, 60, 100 } },                         // operands
8817                         { { 5, 12,  2 }, { 0, 5, 2 }, {  3, 8,  6 }, { 25, 60,   0 } },                         // results
8818                 },
8819                 {
8820                         "all",
8821                         "Inconsistent results, but within specification: ${reason}",
8822                         negFailResult,                                                                                                                          // negative operands, not required by the spec
8823                         { { 5, 12, -17 }, { -5, -5,  7 }, { 75,   8, -81 }, {  25, -60, 100 } },        // operands
8824                         { { 5, -5,   3 }, {  0,  2, -3 }, {  3, -73,  69 }, { -35,  40,   0 } },        // results
8825                 },
8826         };
8827         // If either operand is negative the result is undefined. Some implementations may still return correct values.
8828
8829         for (int caseNdx = 0; caseNdx < DE_LENGTH_OF_ARRAY(cases); ++caseNdx)
8830         {
8831                 const CaseParams&       params                  = cases[caseNdx];
8832                 RGBA                            inputColors[4];
8833                 RGBA                            outputColors[4];
8834
8835                 for (int i = 0; i < 4; ++i)
8836                 {
8837                         inputColors [i] = RGBA(params.operands[i][0] + 128, params.operands[i][1] + 128, params.operands[i][2] + 128, 255);
8838                         outputColors[i] = RGBA(params.results [i][0] + 128, params.results [i][1] + 128, params.results [i][2] + 128, 255);
8839                 }
8840
8841                 createTestsForAllStages(params.name, inputColors, outputColors, fragments, testGroup.get(), params.failResult, params.failMessageTemplate);
8842         }
8843         return testGroup.release();
8844 }
8845
8846 enum ConversionDataType
8847 {
8848         DATA_TYPE_SIGNED_8,
8849         DATA_TYPE_SIGNED_16,
8850         DATA_TYPE_SIGNED_32,
8851         DATA_TYPE_SIGNED_64,
8852         DATA_TYPE_UNSIGNED_8,
8853         DATA_TYPE_UNSIGNED_16,
8854         DATA_TYPE_UNSIGNED_32,
8855         DATA_TYPE_UNSIGNED_64,
8856         DATA_TYPE_FLOAT_16,
8857         DATA_TYPE_FLOAT_32,
8858         DATA_TYPE_FLOAT_64,
8859         DATA_TYPE_VEC2_SIGNED_16,
8860         DATA_TYPE_VEC2_SIGNED_32
8861 };
8862
8863 const string getBitWidthStr (ConversionDataType type)
8864 {
8865         switch (type)
8866         {
8867                 case DATA_TYPE_SIGNED_8:
8868                 case DATA_TYPE_UNSIGNED_8:
8869                         return "8";
8870
8871                 case DATA_TYPE_SIGNED_16:
8872                 case DATA_TYPE_UNSIGNED_16:
8873                 case DATA_TYPE_FLOAT_16:
8874                         return "16";
8875
8876                 case DATA_TYPE_SIGNED_32:
8877                 case DATA_TYPE_UNSIGNED_32:
8878                 case DATA_TYPE_FLOAT_32:
8879                 case DATA_TYPE_VEC2_SIGNED_16:
8880                         return "32";
8881
8882                 case DATA_TYPE_SIGNED_64:
8883                 case DATA_TYPE_UNSIGNED_64:
8884                 case DATA_TYPE_FLOAT_64:
8885                 case DATA_TYPE_VEC2_SIGNED_32:
8886                         return "64";
8887
8888                 default:
8889                         DE_ASSERT(false);
8890         }
8891         return "";
8892 }
8893
8894 const string getByteWidthStr (ConversionDataType type)
8895 {
8896         switch (type)
8897         {
8898                 case DATA_TYPE_SIGNED_8:
8899                 case DATA_TYPE_UNSIGNED_8:
8900                         return "1";
8901
8902                 case DATA_TYPE_SIGNED_16:
8903                 case DATA_TYPE_UNSIGNED_16:
8904                 case DATA_TYPE_FLOAT_16:
8905                         return "2";
8906
8907                 case DATA_TYPE_SIGNED_32:
8908                 case DATA_TYPE_UNSIGNED_32:
8909                 case DATA_TYPE_FLOAT_32:
8910                 case DATA_TYPE_VEC2_SIGNED_16:
8911                         return "4";
8912
8913                 case DATA_TYPE_SIGNED_64:
8914                 case DATA_TYPE_UNSIGNED_64:
8915                 case DATA_TYPE_FLOAT_64:
8916                 case DATA_TYPE_VEC2_SIGNED_32:
8917                         return "8";
8918
8919                 default:
8920                         DE_ASSERT(false);
8921         }
8922         return "";
8923 }
8924
8925 bool isSigned (ConversionDataType type)
8926 {
8927         switch (type)
8928         {
8929                 case DATA_TYPE_SIGNED_8:
8930                 case DATA_TYPE_SIGNED_16:
8931                 case DATA_TYPE_SIGNED_32:
8932                 case DATA_TYPE_SIGNED_64:
8933                 case DATA_TYPE_FLOAT_16:
8934                 case DATA_TYPE_FLOAT_32:
8935                 case DATA_TYPE_FLOAT_64:
8936                 case DATA_TYPE_VEC2_SIGNED_16:
8937                 case DATA_TYPE_VEC2_SIGNED_32:
8938                         return true;
8939
8940                 case DATA_TYPE_UNSIGNED_8:
8941                 case DATA_TYPE_UNSIGNED_16:
8942                 case DATA_TYPE_UNSIGNED_32:
8943                 case DATA_TYPE_UNSIGNED_64:
8944                         return false;
8945
8946                 default:
8947                         DE_ASSERT(false);
8948         }
8949         return false;
8950 }
8951
8952 bool isInt (ConversionDataType type)
8953 {
8954         switch (type)
8955         {
8956                 case DATA_TYPE_SIGNED_8:
8957                 case DATA_TYPE_SIGNED_16:
8958                 case DATA_TYPE_SIGNED_32:
8959                 case DATA_TYPE_SIGNED_64:
8960                 case DATA_TYPE_UNSIGNED_8:
8961                 case DATA_TYPE_UNSIGNED_16:
8962                 case DATA_TYPE_UNSIGNED_32:
8963                 case DATA_TYPE_UNSIGNED_64:
8964                         return true;
8965
8966                 case DATA_TYPE_FLOAT_16:
8967                 case DATA_TYPE_FLOAT_32:
8968                 case DATA_TYPE_FLOAT_64:
8969                 case DATA_TYPE_VEC2_SIGNED_16:
8970                 case DATA_TYPE_VEC2_SIGNED_32:
8971                         return false;
8972
8973                 default:
8974                         DE_ASSERT(false);
8975         }
8976         return false;
8977 }
8978
8979 bool isFloat (ConversionDataType type)
8980 {
8981         switch (type)
8982         {
8983                 case DATA_TYPE_SIGNED_8:
8984                 case DATA_TYPE_SIGNED_16:
8985                 case DATA_TYPE_SIGNED_32:
8986                 case DATA_TYPE_SIGNED_64:
8987                 case DATA_TYPE_UNSIGNED_8:
8988                 case DATA_TYPE_UNSIGNED_16:
8989                 case DATA_TYPE_UNSIGNED_32:
8990                 case DATA_TYPE_UNSIGNED_64:
8991                 case DATA_TYPE_VEC2_SIGNED_16:
8992                 case DATA_TYPE_VEC2_SIGNED_32:
8993                         return false;
8994
8995                 case DATA_TYPE_FLOAT_16:
8996                 case DATA_TYPE_FLOAT_32:
8997                 case DATA_TYPE_FLOAT_64:
8998                         return true;
8999
9000                 default:
9001                         DE_ASSERT(false);
9002         }
9003         return false;
9004 }
9005
9006 const string getTypeName (ConversionDataType type)
9007 {
9008         string prefix = isSigned(type) ? "" : "u";
9009
9010         if              (isInt(type))                                           return prefix + "int"   + getBitWidthStr(type);
9011         else if (isFloat(type))                                         return prefix + "float" + getBitWidthStr(type);
9012         else if (type == DATA_TYPE_VEC2_SIGNED_16)      return "i16vec2";
9013         else if (type == DATA_TYPE_VEC2_SIGNED_32)      return "i32vec2";
9014         else                                                                            DE_ASSERT(false);
9015
9016         return "";
9017 }
9018
9019 const string getTestName (ConversionDataType from, ConversionDataType to, const char* suffix)
9020 {
9021         const string fullSuffix(suffix == DE_NULL ? "" : string("_") + string(suffix));
9022
9023         return getTypeName(from) + "_to_" + getTypeName(to) + fullSuffix;
9024 }
9025
9026 const string getAsmTypeName (ConversionDataType type)
9027 {
9028         string prefix;
9029
9030         if              (isInt(type))                                           prefix = isSigned(type) ? "i" : "u";
9031         else if (isFloat(type))                                         prefix = "f";
9032         else if (type == DATA_TYPE_VEC2_SIGNED_16)      return "i16vec2";
9033         else if (type == DATA_TYPE_VEC2_SIGNED_32)      return "v2i32";
9034         else                                                                            DE_ASSERT(false);
9035
9036         return prefix + getBitWidthStr(type);
9037 }
9038
9039 template<typename T>
9040 BufferSp getSpecializedBuffer (deInt64 number)
9041 {
9042         return BufferSp(new Buffer<T>(vector<T>(1, (T)number)));
9043 }
9044
9045 BufferSp getBuffer (ConversionDataType type, deInt64 number)
9046 {
9047         switch (type)
9048         {
9049                 case DATA_TYPE_SIGNED_8:                return getSpecializedBuffer<deInt8>(number);
9050                 case DATA_TYPE_SIGNED_16:               return getSpecializedBuffer<deInt16>(number);
9051                 case DATA_TYPE_SIGNED_32:               return getSpecializedBuffer<deInt32>(number);
9052                 case DATA_TYPE_SIGNED_64:               return getSpecializedBuffer<deInt64>(number);
9053                 case DATA_TYPE_UNSIGNED_8:              return getSpecializedBuffer<deUint8>(number);
9054                 case DATA_TYPE_UNSIGNED_16:             return getSpecializedBuffer<deUint16>(number);
9055                 case DATA_TYPE_UNSIGNED_32:             return getSpecializedBuffer<deUint32>(number);
9056                 case DATA_TYPE_UNSIGNED_64:             return getSpecializedBuffer<deUint64>(number);
9057                 case DATA_TYPE_FLOAT_16:                return getSpecializedBuffer<deUint16>(number);
9058                 case DATA_TYPE_FLOAT_32:                return getSpecializedBuffer<deUint32>(number);
9059                 case DATA_TYPE_FLOAT_64:                return getSpecializedBuffer<deUint64>(number);
9060                 case DATA_TYPE_VEC2_SIGNED_16:  return getSpecializedBuffer<deUint32>(number);
9061                 case DATA_TYPE_VEC2_SIGNED_32:  return getSpecializedBuffer<deUint64>(number);
9062
9063                 default:                                                TCU_THROW(InternalError, "Unimplemented type passed");
9064         }
9065 }
9066
9067 bool usesInt8 (ConversionDataType from, ConversionDataType to)
9068 {
9069         return (from == DATA_TYPE_SIGNED_8 || to == DATA_TYPE_SIGNED_8 ||
9070                         from == DATA_TYPE_UNSIGNED_8 || to == DATA_TYPE_UNSIGNED_8);
9071 }
9072
9073 bool usesInt16 (ConversionDataType from, ConversionDataType to)
9074 {
9075         return (from == DATA_TYPE_SIGNED_16 || to == DATA_TYPE_SIGNED_16 ||
9076                         from == DATA_TYPE_UNSIGNED_16 || to == DATA_TYPE_UNSIGNED_16 ||
9077                         from == DATA_TYPE_VEC2_SIGNED_16 || to == DATA_TYPE_VEC2_SIGNED_16);
9078 }
9079
9080 bool usesInt32 (ConversionDataType from, ConversionDataType to)
9081 {
9082         return (from == DATA_TYPE_SIGNED_32 || to == DATA_TYPE_SIGNED_32 ||
9083                         from == DATA_TYPE_UNSIGNED_32 || to == DATA_TYPE_UNSIGNED_32 ||
9084                         from == DATA_TYPE_VEC2_SIGNED_32|| to == DATA_TYPE_VEC2_SIGNED_32);
9085 }
9086
9087 bool usesInt64 (ConversionDataType from, ConversionDataType to)
9088 {
9089         return (from == DATA_TYPE_SIGNED_64 || to == DATA_TYPE_SIGNED_64 ||
9090                         from == DATA_TYPE_UNSIGNED_64 || to == DATA_TYPE_UNSIGNED_64);
9091 }
9092
9093 bool usesFloat16 (ConversionDataType from, ConversionDataType to)
9094 {
9095         return (from == DATA_TYPE_FLOAT_16 || to == DATA_TYPE_FLOAT_16);
9096 }
9097
9098 bool usesFloat32 (ConversionDataType from, ConversionDataType to)
9099 {
9100         return (from == DATA_TYPE_FLOAT_32 || to == DATA_TYPE_FLOAT_32);
9101 }
9102
9103 bool usesFloat64 (ConversionDataType from, ConversionDataType to)
9104 {
9105         return (from == DATA_TYPE_FLOAT_64 || to == DATA_TYPE_FLOAT_64);
9106 }
9107
9108 void getVulkanFeaturesAndExtensions (ConversionDataType from, ConversionDataType to, VulkanFeatures& vulkanFeatures, vector<string>& extensions)
9109 {
9110         if (usesInt16(from, to) && !usesInt32(from, to))
9111                 vulkanFeatures.coreFeatures.shaderInt16 = DE_TRUE;
9112
9113         if (usesInt64(from, to))
9114                 vulkanFeatures.coreFeatures.shaderInt64 = DE_TRUE;
9115
9116         if (usesFloat64(from, to))
9117                 vulkanFeatures.coreFeatures.shaderFloat64 = DE_TRUE;
9118
9119         if (usesInt16(from, to) || usesFloat16(from, to))
9120         {
9121                 extensions.push_back("VK_KHR_16bit_storage");
9122                 vulkanFeatures.ext16BitStorage |= EXT16BITSTORAGEFEATURES_UNIFORM_BUFFER_BLOCK;
9123         }
9124
9125         if (usesFloat16(from, to) || usesInt8(from, to))
9126         {
9127                 extensions.push_back("VK_KHR_shader_float16_int8");
9128
9129                 if (usesFloat16(from, to))
9130                 {
9131                         vulkanFeatures.extFloat16Int8 |= EXTFLOAT16INT8FEATURES_FLOAT16;
9132                 }
9133
9134                 if (usesInt8(from, to))
9135                 {
9136                         vulkanFeatures.extFloat16Int8 |= EXTFLOAT16INT8FEATURES_INT8;
9137
9138                         extensions.push_back("VK_KHR_8bit_storage");
9139                         vulkanFeatures.ext8BitStorage |= EXT8BITSTORAGEFEATURES_STORAGE_BUFFER;
9140                 }
9141         }
9142 }
9143
9144 struct ConvertCase
9145 {
9146         ConvertCase (const string& instruction, ConversionDataType from, ConversionDataType to, deInt64 number, bool separateOutput = false, deInt64 outputNumber = 0, const char* suffix = DE_NULL)
9147         : m_fromType            (from)
9148         , m_toType                      (to)
9149         , m_name                        (getTestName(from, to, suffix))
9150         , m_inputBuffer         (getBuffer(from, number))
9151         {
9152                 string caps;
9153                 string decl;
9154                 string exts;
9155
9156                 m_asmTypes["inputType"]         = getAsmTypeName(from);
9157                 m_asmTypes["outputType"]        = getAsmTypeName(to);
9158
9159                 if (separateOutput)
9160                         m_outputBuffer = getBuffer(to, outputNumber);
9161                 else
9162                         m_outputBuffer = getBuffer(to, number);
9163
9164                 if (usesInt8(from, to))
9165                 {
9166                         bool requiresInt8Capability = true;
9167                         if (instruction == "OpUConvert" || instruction == "OpSConvert")
9168                         {
9169                                 // Conversions between 8 and 32 bit are provided by SPV_KHR_8bit_storage. The rest requires explicit Int8
9170                                 if (usesInt32(from, to))
9171                                         requiresInt8Capability = false;
9172                         }
9173
9174                         caps += "OpCapability StorageBuffer8BitAccess\n";
9175                         if (requiresInt8Capability)
9176                                 caps += "OpCapability Int8\n";
9177
9178                         decl += "%i8         = OpTypeInt 8 1\n"
9179                                         "%u8         = OpTypeInt 8 0\n";
9180                         exts += "OpExtension \"SPV_KHR_8bit_storage\"\n";
9181                 }
9182
9183                 if (usesInt16(from, to))
9184                 {
9185                         bool requiresInt16Capability = true;
9186
9187                         if (instruction == "OpUConvert" || instruction == "OpSConvert" || instruction == "OpFConvert")
9188                         {
9189                                 // Conversions between 16 and 32 bit are provided by SPV_KHR_16bit_storage. The rest requires explicit Int16
9190                                 if (usesInt32(from, to) || usesFloat32(from, to))
9191                                         requiresInt16Capability = false;
9192                         }
9193
9194                         decl += "%i16        = OpTypeInt 16 1\n"
9195                                         "%u16        = OpTypeInt 16 0\n"
9196                                         "%i16vec2    = OpTypeVector %i16 2\n";
9197
9198                         // Conversions between 16 and 32 bit are provided by SPV_KHR_16bit_storage. The rest requires explicit Int16
9199                         if (requiresInt16Capability)
9200                                 caps += "OpCapability Int16\n";
9201                 }
9202
9203                 if (usesFloat16(from, to))
9204                 {
9205                         decl += "%f16        = OpTypeFloat 16\n";
9206
9207                         // Conversions between 16 and 32 bit are provided by SPV_KHR_16bit_storage. The rest requires explicit Float16
9208                         if (!(usesInt32(from, to) || usesFloat32(from, to)))
9209                                 caps += "OpCapability Float16\n";
9210                 }
9211
9212                 if (usesInt16(from, to) || usesFloat16(from, to))
9213                 {
9214                         caps += "OpCapability StorageUniformBufferBlock16\n";
9215                         exts += "OpExtension \"SPV_KHR_16bit_storage\"\n";
9216                 }
9217
9218                 if (usesInt64(from, to))
9219                 {
9220                         caps += "OpCapability Int64\n";
9221                         decl += "%i64        = OpTypeInt 64 1\n"
9222                                         "%u64        = OpTypeInt 64 0\n";
9223                 }
9224
9225                 if (usesFloat64(from, to))
9226                 {
9227                         caps += "OpCapability Float64\n";
9228                         decl += "%f64        = OpTypeFloat 64\n";
9229                 }
9230
9231                 m_asmTypes["datatype_capabilities"]             = caps;
9232                 m_asmTypes["datatype_additional_decl"]  = decl;
9233                 m_asmTypes["datatype_extensions"]               = exts;
9234         }
9235
9236         ConversionDataType              m_fromType;
9237         ConversionDataType              m_toType;
9238         string                                  m_name;
9239         map<string, string>             m_asmTypes;
9240         BufferSp                                m_inputBuffer;
9241         BufferSp                                m_outputBuffer;
9242 };
9243
9244 const string getConvertCaseShaderStr (const string& instruction, const ConvertCase& convertCase)
9245 {
9246         map<string, string> params = convertCase.m_asmTypes;
9247
9248         params["instruction"]   = instruction;
9249         params["inDecorator"]   = getByteWidthStr(convertCase.m_fromType);
9250         params["outDecorator"]  = getByteWidthStr(convertCase.m_toType);
9251
9252         const StringTemplate shader (
9253                 "OpCapability Shader\n"
9254                 "${datatype_capabilities}"
9255                 "${datatype_extensions:opt}"
9256                 "OpMemoryModel Logical GLSL450\n"
9257                 "OpEntryPoint GLCompute %main \"main\"\n"
9258                 "OpExecutionMode %main LocalSize 1 1 1\n"
9259                 "OpSource GLSL 430\n"
9260                 "OpName %main           \"main\"\n"
9261                 // Decorators
9262                 "OpDecorate %indata DescriptorSet 0\n"
9263                 "OpDecorate %indata Binding 0\n"
9264                 "OpDecorate %outdata DescriptorSet 0\n"
9265                 "OpDecorate %outdata Binding 1\n"
9266                 "OpDecorate %in_buf BufferBlock\n"
9267                 "OpDecorate %out_buf BufferBlock\n"
9268                 "OpMemberDecorate %in_buf 0 Offset 0\n"
9269                 "OpMemberDecorate %out_buf 0 Offset 0\n"
9270                 // Base types
9271                 "%void       = OpTypeVoid\n"
9272                 "%voidf      = OpTypeFunction %void\n"
9273                 "%u32        = OpTypeInt 32 0\n"
9274                 "%i32        = OpTypeInt 32 1\n"
9275                 "%f32        = OpTypeFloat 32\n"
9276                 "%v2i32      = OpTypeVector %i32 2\n"
9277                 "${datatype_additional_decl}"
9278                 "%uvec3      = OpTypeVector %u32 3\n"
9279                 // Derived types
9280                 "%in_ptr     = OpTypePointer Uniform %${inputType}\n"
9281                 "%out_ptr    = OpTypePointer Uniform %${outputType}\n"
9282                 "%in_buf     = OpTypeStruct %${inputType}\n"
9283                 "%out_buf    = OpTypeStruct %${outputType}\n"
9284                 "%in_bufptr  = OpTypePointer Uniform %in_buf\n"
9285                 "%out_bufptr = OpTypePointer Uniform %out_buf\n"
9286                 "%indata     = OpVariable %in_bufptr Uniform\n"
9287                 "%outdata    = OpVariable %out_bufptr Uniform\n"
9288                 // Constants
9289                 "%zero       = OpConstant %i32 0\n"
9290                 // Main function
9291                 "%main       = OpFunction %void None %voidf\n"
9292                 "%label      = OpLabel\n"
9293                 "%inloc      = OpAccessChain %in_ptr %indata %zero\n"
9294                 "%outloc     = OpAccessChain %out_ptr %outdata %zero\n"
9295                 "%inval      = OpLoad %${inputType} %inloc\n"
9296                 "%conv       = ${instruction} %${outputType} %inval\n"
9297                 "              OpStore %outloc %conv\n"
9298                 "              OpReturn\n"
9299                 "              OpFunctionEnd\n"
9300         );
9301
9302         return shader.specialize(params);
9303 }
9304
9305 void createConvertCases (vector<ConvertCase>& testCases, const string& instruction)
9306 {
9307         if (instruction == "OpUConvert")
9308         {
9309                 // Convert unsigned int to unsigned int
9310                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_8,           DATA_TYPE_UNSIGNED_16,          42));
9311                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_8,           DATA_TYPE_UNSIGNED_32,          73));
9312                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_8,           DATA_TYPE_UNSIGNED_64,          121));
9313
9314                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_16,          DATA_TYPE_UNSIGNED_8,           33));
9315                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_16,          DATA_TYPE_UNSIGNED_32,          60653));
9316                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_16,          DATA_TYPE_UNSIGNED_64,          17991));
9317
9318                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_32,          DATA_TYPE_UNSIGNED_64,          904256275));
9319                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_32,          DATA_TYPE_UNSIGNED_16,          6275));
9320                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_32,          DATA_TYPE_UNSIGNED_8,           17));
9321
9322                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_64,          DATA_TYPE_UNSIGNED_32,          701256243));
9323                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_64,          DATA_TYPE_UNSIGNED_16,          4741));
9324                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_64,          DATA_TYPE_UNSIGNED_8,           65));
9325         }
9326         else if (instruction == "OpSConvert")
9327         {
9328                 // Sign extension int->int
9329                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_8,                     DATA_TYPE_SIGNED_16,            -30));
9330                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_8,                     DATA_TYPE_SIGNED_32,            55));
9331                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_8,                     DATA_TYPE_SIGNED_64,            -3));
9332                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_16,            DATA_TYPE_SIGNED_32,            14669));
9333                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_16,            DATA_TYPE_SIGNED_64,            -3341));
9334                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_32,            DATA_TYPE_SIGNED_64,            973610259));
9335
9336                 // Truncate for int->int
9337                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_16,            DATA_TYPE_SIGNED_8,                     81));
9338                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_32,            DATA_TYPE_SIGNED_8,                     -93));
9339                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_64,            DATA_TYPE_SIGNED_8,                     3182748172687672ll,                                     true,   56));
9340                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_32,            DATA_TYPE_SIGNED_16,            12382));
9341                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_64,            DATA_TYPE_SIGNED_32,            -972812359));
9342                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_64,            DATA_TYPE_SIGNED_16,            -1067742499291926803ll,                         true,   -4371));
9343
9344                 // Sign extension for int->uint
9345                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_8,                     DATA_TYPE_UNSIGNED_16,          56));
9346                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_8,                     DATA_TYPE_UNSIGNED_32,          -47,                                                            true,   4294967249u));
9347                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_8,                     DATA_TYPE_UNSIGNED_64,          -5,                                                                     true,   18446744073709551611ull));
9348                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_16,            DATA_TYPE_UNSIGNED_32,          14669));
9349                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_16,            DATA_TYPE_UNSIGNED_64,          -3341,                                                          true,   18446744073709548275ull));
9350                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_32,            DATA_TYPE_UNSIGNED_64,          973610259));
9351
9352                 // Truncate for int->uint
9353                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_16,            DATA_TYPE_UNSIGNED_8,           -25711,                                                         true,   145));
9354                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_32,            DATA_TYPE_UNSIGNED_8,           103));
9355                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_64,            DATA_TYPE_UNSIGNED_8,           -1067742499291926803ll,                         true,   61165));
9356                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_32,            DATA_TYPE_UNSIGNED_16,          12382));
9357                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_64,            DATA_TYPE_UNSIGNED_32,          -972812359,                                                     true,   3322154937u));
9358                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_64,            DATA_TYPE_UNSIGNED_16,          -1067742499291926803ll,                         true,   61165));
9359
9360                 // Sign extension for uint->int
9361                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_8,           DATA_TYPE_SIGNED_16,            71));
9362                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_8,           DATA_TYPE_SIGNED_32,            201,                                                            true,   -55));
9363                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_8,           DATA_TYPE_SIGNED_64,            188,                                                            true,   -68));
9364                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_16,          DATA_TYPE_SIGNED_32,            14669));
9365                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_16,          DATA_TYPE_SIGNED_64,            62195,                                                          true,   -3341));
9366                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_32,          DATA_TYPE_SIGNED_64,            973610259));
9367
9368                 // Truncate for uint->int
9369                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_16,          DATA_TYPE_SIGNED_8,                     67));
9370                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_32,          DATA_TYPE_SIGNED_8,                     133,                                                            true,   -123));
9371                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_64,          DATA_TYPE_SIGNED_8,                     836927654193256494ull,                          true,   46));
9372                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_32,          DATA_TYPE_SIGNED_16,            12382));
9373                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_64,          DATA_TYPE_SIGNED_32,            18446744072736739257ull,                        true,   -972812359));
9374                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_64,          DATA_TYPE_SIGNED_16,            17379001574417624813ull,                        true,   -4371));
9375
9376                 // Convert i16vec2 to i32vec2 and vice versa
9377                 // Unsigned values are used here to represent negative signed values and to allow defined shifting behaviour.
9378                 // The actual signed value -32123 is used here as uint16 value 33413 and uint32 value 4294935173
9379                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_VEC2_SIGNED_16,       DATA_TYPE_VEC2_SIGNED_32,       (33413u << 16)                  | 27593,        true,   (4294935173ull << 32)   | 27593));
9380                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_VEC2_SIGNED_32,       DATA_TYPE_VEC2_SIGNED_16,       (4294935173ull << 32)   | 27593,        true,   (33413u << 16)                  | 27593));
9381         }
9382         else if (instruction == "OpFConvert")
9383         {
9384                 // All hexadecimal values below represent 1234.0 as 16/32/64-bit IEEE 754 float
9385                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_32,                     DATA_TYPE_FLOAT_64,                     0x449a4000,                                                     true,   0x4093480000000000));
9386                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_64,                     DATA_TYPE_FLOAT_32,                     0x4093480000000000,                                     true,   0x449a4000));
9387
9388                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_32,                     DATA_TYPE_FLOAT_16,                     0x449a4000,                                                     true,   0x64D2));
9389                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_16,                     DATA_TYPE_FLOAT_32,                     0x64D2,                                                         true,   0x449a4000));
9390
9391                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_16,                     DATA_TYPE_FLOAT_64,                     0x64D2,                                                         true,   0x4093480000000000));
9392                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_64,                     DATA_TYPE_FLOAT_16,                     0x4093480000000000,                                     true,   0x64D2));
9393         }
9394         else if (instruction == "OpConvertFToU")
9395         {
9396                 // Normal numbers from uint8 range
9397                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_16,                     DATA_TYPE_UNSIGNED_8,           0x5020,                                                         true,   33,                                                                     "33"));
9398                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_32,                     DATA_TYPE_UNSIGNED_8,           0x42280000,                                                     true,   42,                                                                     "42"));
9399                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_64,                     DATA_TYPE_UNSIGNED_8,           0x4067800000000000ull,                          true,   188,                                                            "188"));
9400
9401                 // Maximum uint8 value
9402                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_16,                     DATA_TYPE_UNSIGNED_8,           0x5BF8,                                                         true,   255,                                                            "max"));
9403                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_32,                     DATA_TYPE_UNSIGNED_8,           0x437F0000,                                                     true,   255,                                                            "max"));
9404                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_64,                     DATA_TYPE_UNSIGNED_8,           0x406FE00000000000ull,                          true,   255,                                                            "max"));
9405
9406                 // +0
9407                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_16,                     DATA_TYPE_UNSIGNED_8,           0x0000,                                                         true,   0,                                                                      "p0"));
9408                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_32,                     DATA_TYPE_UNSIGNED_8,           0x00000000,                                                     true,   0,                                                                      "p0"));
9409                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_64,                     DATA_TYPE_UNSIGNED_8,           0x0000000000000000ull,                          true,   0,                                                                      "p0"));
9410
9411                 // -0
9412                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_16,                     DATA_TYPE_UNSIGNED_8,           0x8000,                                                         true,   0,                                                                      "m0"));
9413                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_32,                     DATA_TYPE_UNSIGNED_8,           0x80000000,                                                     true,   0,                                                                      "m0"));
9414                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_64,                     DATA_TYPE_UNSIGNED_8,           0x8000000000000000ull,                          true,   0,                                                                      "m0"));
9415
9416                 // All hexadecimal values below represent 1234.0 as 16/32/64-bit IEEE 754 float
9417                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_16,                     DATA_TYPE_UNSIGNED_16,          0x64D2,                                                         true,   1234,                                                           "1234"));
9418                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_16,                     DATA_TYPE_UNSIGNED_32,          0x64D2,                                                         true,   1234,                                                           "1234"));
9419                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_16,                     DATA_TYPE_UNSIGNED_64,          0x64D2,                                                         true,   1234,                                                           "1234"));
9420
9421                 // 0x7BFF = 0111 1011 1111 1111 = 0 11110 1111111111 = 65504
9422                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_16,                     DATA_TYPE_UNSIGNED_16,          0x7BFF,                                                         true,   65504,                                                          "max"));
9423                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_16,                     DATA_TYPE_UNSIGNED_32,          0x7BFF,                                                         true,   65504,                                                          "max"));
9424                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_16,                     DATA_TYPE_UNSIGNED_64,          0x7BFF,                                                         true,   65504,                                                          "max"));
9425
9426                 // +0
9427                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_16,                     DATA_TYPE_UNSIGNED_32,          0x0000,                                                         true,   0,                                                                      "p0"));
9428                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_16,                     DATA_TYPE_UNSIGNED_16,          0x0000,                                                         true,   0,                                                                      "p0"));
9429                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_16,                     DATA_TYPE_UNSIGNED_64,          0x0000,                                                         true,   0,                                                                      "p0"));
9430
9431                 // -0
9432                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_16,                     DATA_TYPE_UNSIGNED_16,          0x8000,                                                         true,   0,                                                                      "m0"));
9433                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_16,                     DATA_TYPE_UNSIGNED_32,          0x8000,                                                         true,   0,                                                                      "m0"));
9434                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_16,                     DATA_TYPE_UNSIGNED_64,          0x8000,                                                         true,   0,                                                                      "m0"));
9435
9436                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_32,                     DATA_TYPE_UNSIGNED_16,          0x449a4000,                                                     true,   1234));
9437                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_32,                     DATA_TYPE_UNSIGNED_32,          0x449a4000,                                                     true,   1234));
9438                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_32,                     DATA_TYPE_UNSIGNED_64,          0x449a4000,                                                     true,   1234));
9439                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_64,                     DATA_TYPE_UNSIGNED_16,          0x4093480000000000,                                     true,   1234));
9440                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_64,                     DATA_TYPE_UNSIGNED_32,          0x4093480000000000,                                     true,   1234));
9441                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_64,                     DATA_TYPE_UNSIGNED_64,          0x4093480000000000,                                     true,   1234));
9442         }
9443         else if (instruction == "OpConvertUToF")
9444         {
9445                 // Normal numbers from uint8 range
9446                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_8,           DATA_TYPE_FLOAT_16,                     116,                                                            true,   0x5740,                                                         "116"));
9447                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_8,           DATA_TYPE_FLOAT_32,                     232,                                                            true,   0x43680000,                                                     "232"));
9448                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_8,           DATA_TYPE_FLOAT_64,                     164,                                                            true,   0x4064800000000000ull,                          "164"));
9449
9450                 // Maximum uint8 value
9451                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_8,           DATA_TYPE_FLOAT_16,                     255,                                                            true,   0x5BF8,                                                         "max"));
9452                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_8,           DATA_TYPE_FLOAT_32,                     255,                                                            true,   0x437F0000,                                                     "max"));
9453                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_8,           DATA_TYPE_FLOAT_64,                     255,                                                            true,   0x406FE00000000000ull,                          "max"));
9454
9455                 // All hexadecimal values below represent 1234.0 as 32/64-bit IEEE 754 float
9456                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_16,          DATA_TYPE_FLOAT_16,                     1234,                                                           true,   0x64D2,                                                         "1234"));
9457                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_32,          DATA_TYPE_FLOAT_16,                     1234,                                                           true,   0x64D2,                                                         "1234"));
9458                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_64,          DATA_TYPE_FLOAT_16,                     1234,                                                           true,   0x64D2,                                                         "1234"));
9459
9460                 // 0x7BFF = 0111 1011 1111 1111 = 0 11110 1111111111 = 65504
9461                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_16,          DATA_TYPE_FLOAT_16,                     65504,                                                          true,   0x7BFF,                                                         "max"));
9462                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_32,          DATA_TYPE_FLOAT_16,                     65504,                                                          true,   0x7BFF,                                                         "max"));
9463                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_64,          DATA_TYPE_FLOAT_16,                     65504,                                                          true,   0x7BFF,                                                         "max"));
9464
9465                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_16,          DATA_TYPE_FLOAT_32,                     1234,                                                           true,   0x449a4000));
9466                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_16,          DATA_TYPE_FLOAT_64,                     1234,                                                           true,   0x4093480000000000));
9467                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_32,          DATA_TYPE_FLOAT_32,                     1234,                                                           true,   0x449a4000));
9468                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_32,          DATA_TYPE_FLOAT_64,                     1234,                                                           true,   0x4093480000000000));
9469                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_64,          DATA_TYPE_FLOAT_32,                     1234,                                                           true,   0x449a4000));
9470                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_UNSIGNED_64,          DATA_TYPE_FLOAT_64,                     1234,                                                           true,   0x4093480000000000));
9471         }
9472         else if (instruction == "OpConvertFToS")
9473         {
9474                 // Normal numbers from int8 range
9475                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_16,                     DATA_TYPE_SIGNED_8,                     0xC980,                                                         true,   -11,                                                            "m11"));
9476                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_32,                     DATA_TYPE_SIGNED_8,                     0xC2140000,                                                     true,   -37,                                                            "m37"));
9477                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_64,                     DATA_TYPE_SIGNED_8,                     0xC050800000000000ull,                          true,   -66,                                                            "m66"));
9478
9479                 // Minimum int8 value
9480                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_16,                     DATA_TYPE_SIGNED_8,                     0xD800,                                                         true,   -128,                                                           "min"));
9481                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_32,                     DATA_TYPE_SIGNED_8,                     0xC3000000,                                                     true,   -128,                                                           "min"));
9482                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_64,                     DATA_TYPE_SIGNED_8,                     0xC060000000000000ull,                          true,   -128,                                                           "min"));
9483
9484                 // Maximum int8 value
9485                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_16,                     DATA_TYPE_SIGNED_8,                     0x57F0,                                                         true,   127,                                                            "max"));
9486                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_32,                     DATA_TYPE_SIGNED_8,                     0x42FE0000,                                                     true,   127,                                                            "max"));
9487                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_64,                     DATA_TYPE_SIGNED_8,                     0x405FC00000000000ull,                          true,   127,                                                            "max"));
9488
9489                 // +0
9490                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_16,                     DATA_TYPE_SIGNED_8,                     0x0000,                                                         true,   0,                                                                      "p0"));
9491                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_32,                     DATA_TYPE_SIGNED_8,                     0x00000000,                                                     true,   0,                                                                      "p0"));
9492                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_64,                     DATA_TYPE_SIGNED_8,                     0x0000000000000000ull,                          true,   0,                                                                      "p0"));
9493
9494                 // -0
9495                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_16,                     DATA_TYPE_SIGNED_8,                     0x8000,                                                         true,   0,                                                                      "m0"));
9496                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_32,                     DATA_TYPE_SIGNED_8,                     0x80000000,                                                     true,   0,                                                                      "m0"));
9497                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_64,                     DATA_TYPE_SIGNED_8,                     0x8000000000000000ull,                          true,   0,                                                                      "m0"));
9498
9499                 // All hexadecimal values below represent -1234.0 as 32/64-bit IEEE 754 float
9500                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_16,                     DATA_TYPE_SIGNED_16,            0xE4D2,                                                         true,   -1234,                                                          "m1234"));
9501                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_16,                     DATA_TYPE_SIGNED_32,            0xE4D2,                                                         true,   -1234,                                                          "m1234"));
9502                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_16,                     DATA_TYPE_SIGNED_64,            0xE4D2,                                                         true,   -1234,                                                          "m1234"));
9503
9504                 // 0xF800 = 1111 1000 0000 0000 = 1 11110 0000000000 = -32768
9505                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_16,                     DATA_TYPE_SIGNED_16,            0xF800,                                                         true,   -32768,                                                         "min"));
9506                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_16,                     DATA_TYPE_SIGNED_32,            0xF800,                                                         true,   -32768,                                                         "min"));
9507                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_16,                     DATA_TYPE_SIGNED_64,            0xF800,                                                         true,   -32768,                                                         "min"));
9508
9509                 // 0x77FF = 0111 0111 1111 1111 = 0 11101 1111111111 = 32752
9510                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_16,                     DATA_TYPE_SIGNED_16,            0x77FF,                                                         true,   32752,                                                          "max"));
9511                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_16,                     DATA_TYPE_SIGNED_32,            0x77FF,                                                         true,   32752,                                                          "max"));
9512                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_16,                     DATA_TYPE_SIGNED_64,            0x77FF,                                                         true,   32752,                                                          "max"));
9513
9514                 // +0
9515                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_16,                     DATA_TYPE_SIGNED_16,            0x0000,                                                         true,   0,                                                                      "p0"));
9516                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_16,                     DATA_TYPE_SIGNED_32,            0x0000,                                                         true,   0,                                                                      "p0"));
9517                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_16,                     DATA_TYPE_SIGNED_64,            0x0000,                                                         true,   0,                                                                      "p0"));
9518
9519                 // -0
9520                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_16,                     DATA_TYPE_SIGNED_16,            0x8000,                                                         true,   0,                                                                      "m0"));
9521                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_16,                     DATA_TYPE_SIGNED_32,            0x8000,                                                         true,   0,                                                                      "m0"));
9522                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_16,                     DATA_TYPE_SIGNED_64,            0x8000,                                                         true,   0,                                                                      "m0"));
9523
9524                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_32,                     DATA_TYPE_SIGNED_16,            0xc49a4000,                                                     true,   -1234));
9525                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_32,                     DATA_TYPE_SIGNED_32,            0xc49a4000,                                                     true,   -1234));
9526                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_32,                     DATA_TYPE_SIGNED_64,            0xc49a4000,                                                     true,   -1234));
9527                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_64,                     DATA_TYPE_SIGNED_16,            0xc093480000000000,                                     true,   -1234));
9528                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_64,                     DATA_TYPE_SIGNED_32,            0xc093480000000000,                                     true,   -1234));
9529                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_64,                     DATA_TYPE_SIGNED_64,            0xc093480000000000,                                     true,   -1234));
9530                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_32,                     DATA_TYPE_SIGNED_16,            0x453b9000,                                                     true,    3001,                                                          "p3001"));
9531                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_FLOAT_32,                     DATA_TYPE_SIGNED_16,            0xc53b9000,                                                     true,   -3001,                                                          "m3001"));
9532         }
9533         else if (instruction == "OpConvertSToF")
9534         {
9535                 // Normal numbers from int8 range
9536                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_8,                     DATA_TYPE_FLOAT_16,                     -12,                                                            true,   0xCA00,                                                         "m21"));
9537                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_8,                     DATA_TYPE_FLOAT_32,                     -21,                                                            true,   0xC1A80000,                                                     "m21"));
9538                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_8,                     DATA_TYPE_FLOAT_64,                     -99,                                                            true,   0xC058C00000000000ull,                          "m99"));
9539
9540                 // Minimum int8 value
9541                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_8,                     DATA_TYPE_FLOAT_16,                     -128,                                                           true,   0xD800,                                                         "min"));
9542                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_8,                     DATA_TYPE_FLOAT_32,                     -128,                                                           true,   0xC3000000,                                                     "min"));
9543                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_8,                     DATA_TYPE_FLOAT_64,                     -128,                                                           true,   0xC060000000000000ull,                          "min"));
9544
9545                 // Maximum int8 value
9546                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_8,                     DATA_TYPE_FLOAT_16,                     127,                                                            true,   0x57F0,                                                         "max"));
9547                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_8,                     DATA_TYPE_FLOAT_32,                     127,                                                            true,   0x42FE0000,                                                     "max"));
9548                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_8,                     DATA_TYPE_FLOAT_64,                     127,                                                            true,   0x405FC00000000000ull,                          "max"));
9549
9550                 // All hexadecimal values below represent 1234.0 as 32/64-bit IEEE 754 float
9551                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_16,            DATA_TYPE_FLOAT_16,                     -1234,                                                          true,   0xE4D2,                                                         "m1234"));
9552                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_32,            DATA_TYPE_FLOAT_16,                     -1234,                                                          true,   0xE4D2,                                                         "m1234"));
9553                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_64,            DATA_TYPE_FLOAT_16,                     -1234,                                                          true,   0xE4D2,                                                         "m1234"));
9554
9555                 // 0xF800 = 1111 1000 0000 0000 = 1 11110 0000000000 = -32768
9556                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_16,            DATA_TYPE_FLOAT_16,                     -32768,                                                         true,   0xF800,                                                         "min"));
9557                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_32,            DATA_TYPE_FLOAT_16,                     -32768,                                                         true,   0xF800,                                                         "min"));
9558                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_64,            DATA_TYPE_FLOAT_16,                     -32768,                                                         true,   0xF800,                                                         "min"));
9559
9560                 // 0x77FF = 0111 0111 1111 1111 = 0 11101 1111111111 = 32752
9561                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_16,            DATA_TYPE_FLOAT_16,                     32752,                                                          true,   0x77FF,                                                         "max"));
9562                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_32,            DATA_TYPE_FLOAT_16,                     32752,                                                          true,   0x77FF,                                                         "max"));
9563                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_64,            DATA_TYPE_FLOAT_16,                     32752,                                                          true,   0x77FF,                                                         "max"));
9564
9565                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_16,            DATA_TYPE_FLOAT_32,                     -1234,                                                          true,   0xc49a4000));
9566                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_16,            DATA_TYPE_FLOAT_64,                     -1234,                                                          true,   0xc093480000000000));
9567                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_32,            DATA_TYPE_FLOAT_32,                     -1234,                                                          true,   0xc49a4000));
9568                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_32,            DATA_TYPE_FLOAT_64,                     -1234,                                                          true,   0xc093480000000000));
9569                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_64,            DATA_TYPE_FLOAT_32,                     -1234,                                                          true,   0xc49a4000));
9570                 testCases.push_back(ConvertCase(instruction,    DATA_TYPE_SIGNED_64,            DATA_TYPE_FLOAT_64,                     -1234,                                                          true,   0xc093480000000000));
9571         }
9572         else
9573                 DE_FATAL("Unknown instruction");
9574 }
9575
9576 const map<string, string> getConvertCaseFragments (string instruction, const ConvertCase& convertCase)
9577 {
9578         map<string, string> params = convertCase.m_asmTypes;
9579         map<string, string> fragments;
9580
9581         params["instruction"] = instruction;
9582         params["inDecorator"] = getByteWidthStr(convertCase.m_fromType);
9583
9584         const StringTemplate decoration (
9585                 "      OpDecorate %SSBOi DescriptorSet 0\n"
9586                 "      OpDecorate %SSBOo DescriptorSet 0\n"
9587                 "      OpDecorate %SSBOi Binding 0\n"
9588                 "      OpDecorate %SSBOo Binding 1\n"
9589                 "      OpDecorate %s_SSBOi Block\n"
9590                 "      OpDecorate %s_SSBOo Block\n"
9591                 "OpMemberDecorate %s_SSBOi 0 Offset 0\n"
9592                 "OpMemberDecorate %s_SSBOo 0 Offset 0\n");
9593
9594         const StringTemplate pre_main (
9595                 "${datatype_additional_decl:opt}"
9596                 "    %ptr_in = OpTypePointer StorageBuffer %${inputType}\n"
9597                 "   %ptr_out = OpTypePointer StorageBuffer %${outputType}\n"
9598                 "   %s_SSBOi = OpTypeStruct %${inputType}\n"
9599                 "   %s_SSBOo = OpTypeStruct %${outputType}\n"
9600                 " %ptr_SSBOi = OpTypePointer StorageBuffer %s_SSBOi\n"
9601                 " %ptr_SSBOo = OpTypePointer StorageBuffer %s_SSBOo\n"
9602                 "     %SSBOi = OpVariable %ptr_SSBOi StorageBuffer\n"
9603                 "     %SSBOo = OpVariable %ptr_SSBOo StorageBuffer\n");
9604
9605         const StringTemplate testfun (
9606                 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
9607                 "%param     = OpFunctionParameter %v4f32\n"
9608                 "%label     = OpLabel\n"
9609                 "%iLoc      = OpAccessChain %ptr_in %SSBOi %c_u32_0\n"
9610                 "%oLoc      = OpAccessChain %ptr_out %SSBOo %c_u32_0\n"
9611                 "%valIn     = OpLoad %${inputType} %iLoc\n"
9612                 "%valOut    = ${instruction} %${outputType} %valIn\n"
9613                 "             OpStore %oLoc %valOut\n"
9614                 "             OpReturnValue %param\n"
9615                 "             OpFunctionEnd\n");
9616
9617         params["datatype_extensions"] =
9618                 params["datatype_extensions"] +
9619                 "OpExtension \"SPV_KHR_storage_buffer_storage_class\"\n";
9620
9621         fragments["capability"] = params["datatype_capabilities"];
9622         fragments["extension"]  = params["datatype_extensions"];
9623         fragments["decoration"] = decoration.specialize(params);
9624         fragments["pre_main"]   = pre_main.specialize(params);
9625         fragments["testfun"]    = testfun.specialize(params);
9626
9627         return fragments;
9628 }
9629
9630 // Test for OpSConvert, OpUConvert, OpFConvert and OpConvert* in compute shaders
9631 tcu::TestCaseGroup* createConvertComputeTests (tcu::TestContext& testCtx, const string& instruction, const string& name)
9632 {
9633         de::MovePtr<tcu::TestCaseGroup>         group(new tcu::TestCaseGroup(testCtx, name.c_str(), instruction.c_str()));
9634         vector<ConvertCase>                                     testCases;
9635         createConvertCases(testCases, instruction);
9636
9637         for (vector<ConvertCase>::const_iterator test = testCases.begin(); test != testCases.end(); ++test)
9638         {
9639                 ComputeShaderSpec spec;
9640                 spec.assembly                   = getConvertCaseShaderStr(instruction, *test);
9641                 spec.numWorkGroups              = IVec3(1, 1, 1);
9642                 spec.inputs.push_back   (test->m_inputBuffer);
9643                 spec.outputs.push_back  (test->m_outputBuffer);
9644
9645                 getVulkanFeaturesAndExtensions(test->m_fromType, test->m_toType, spec.requestedVulkanFeatures, spec.extensions);
9646
9647                 group->addChild(new SpvAsmComputeShaderCase(testCtx, test->m_name.c_str(), "", spec));
9648         }
9649         return group.release();
9650 }
9651
9652 // Test for OpSConvert, OpUConvert, OpFConvert and OpConvert* in graphics shaders
9653 tcu::TestCaseGroup* createConvertGraphicsTests (tcu::TestContext& testCtx, const string& instruction, const string& name)
9654 {
9655         de::MovePtr<tcu::TestCaseGroup>         group(new tcu::TestCaseGroup(testCtx, name.c_str(), instruction.c_str()));
9656         vector<ConvertCase>                                     testCases;
9657         createConvertCases(testCases, instruction);
9658
9659         for (vector<ConvertCase>::const_iterator test = testCases.begin(); test != testCases.end(); ++test)
9660         {
9661                 map<string, string>     fragments               = getConvertCaseFragments(instruction, *test);
9662                 VulkanFeatures          vulkanFeatures;
9663                 GraphicsResources       resources;
9664                 vector<string>          extensions;
9665                 SpecConstants           noSpecConstants;
9666                 PushConstants           noPushConstants;
9667                 GraphicsInterfaces      noInterfaces;
9668                 tcu::RGBA                       defaultColors[4];
9669
9670                 getDefaultColors                        (defaultColors);
9671                 resources.inputs.push_back      (Resource(test->m_inputBuffer, VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
9672                 resources.outputs.push_back     (Resource(test->m_outputBuffer, VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
9673                 extensions.push_back            ("VK_KHR_storage_buffer_storage_class");
9674
9675                 getVulkanFeaturesAndExtensions(test->m_fromType, test->m_toType, vulkanFeatures, extensions);
9676
9677                 createTestsForAllStages(
9678                         test->m_name, defaultColors, defaultColors, fragments, noSpecConstants,
9679                         noPushConstants, resources, noInterfaces, extensions, vulkanFeatures, group.get());
9680         }
9681         return group.release();
9682 }
9683
9684 // Constant-Creation Instructions: OpConstant, OpConstantComposite
9685 tcu::TestCaseGroup* createOpConstantFloat16Tests(tcu::TestContext& testCtx)
9686 {
9687         de::MovePtr<tcu::TestCaseGroup> opConstantCompositeTests                (new tcu::TestCaseGroup(testCtx, "opconstant", "OpConstant and OpConstantComposite instruction"));
9688         RGBA                                                    inputColors[4];
9689         RGBA                                                    outputColors[4];
9690         vector<string>                                  extensions;
9691         GraphicsResources                               resources;
9692         VulkanFeatures                                  features;
9693
9694         const char                                              functionStart[]  =
9695                 "%test_code             = OpFunction %v4f32 None %v4f32_v4f32_function\n"
9696                 "%param1                = OpFunctionParameter %v4f32\n"
9697                 "%lbl                   = OpLabel\n";
9698
9699         const char                                              functionEnd[]           =
9700                 "%transformed_param_32  = OpFConvert %v4f32 %transformed_param\n"
9701                 "                         OpReturnValue %transformed_param_32\n"
9702                 "                         OpFunctionEnd\n";
9703
9704         struct NameConstantsCode
9705         {
9706                 string name;
9707                 string constants;
9708                 string code;
9709         };
9710
9711 #define FLOAT_16_COMMON_TYPES_AND_CONSTS \
9712                         "%f16                  = OpTypeFloat 16\n"                                                 \
9713                         "%c_f16_0              = OpConstant %f16 0.0\n"                                            \
9714                         "%c_f16_0_5            = OpConstant %f16 0.5\n"                                            \
9715                         "%c_f16_1              = OpConstant %f16 1.0\n"                                            \
9716                         "%v4f16                = OpTypeVector %f16 4\n"                                            \
9717                         "%fp_f16               = OpTypePointer Function %f16\n"                                    \
9718                         "%fp_v4f16             = OpTypePointer Function %v4f16\n"                                  \
9719                         "%c_v4f16_1_1_1_1      = OpConstantComposite %v4f16 %c_f16_1 %c_f16_1 %c_f16_1 %c_f16_1\n" \
9720                         "%a4f16                = OpTypeArray %f16 %c_u32_4\n"                                      \
9721
9722         NameConstantsCode                               tests[] =
9723         {
9724                 {
9725                         "vec4",
9726
9727                         FLOAT_16_COMMON_TYPES_AND_CONSTS
9728                         "%cval                 = OpConstantComposite %v4f16 %c_f16_0_5 %c_f16_0_5 %c_f16_0_5 %c_f16_0\n",
9729                         "%param1_16            = OpFConvert %v4f16 %param1\n"
9730                         "%transformed_param    = OpFAdd %v4f16 %param1_16 %cval\n"
9731                 },
9732                 {
9733                         "struct",
9734
9735                         FLOAT_16_COMMON_TYPES_AND_CONSTS
9736                         "%stype                = OpTypeStruct %v4f16 %f16\n"
9737                         "%fp_stype             = OpTypePointer Function %stype\n"
9738                         "%f16_n_1              = OpConstant %f16 -1.0\n"
9739                         "%f16_1_5              = OpConstant %f16 !0x3e00\n" // +1.5
9740                         "%cvec                 = OpConstantComposite %v4f16 %f16_1_5 %f16_1_5 %f16_1_5 %c_f16_1\n"
9741                         "%cval                 = OpConstantComposite %stype %cvec %f16_n_1\n",
9742
9743                         "%v                    = OpVariable %fp_stype Function %cval\n"
9744                         "%vec_ptr              = OpAccessChain %fp_v4f16 %v %c_u32_0\n"
9745                         "%f16_ptr              = OpAccessChain %fp_f16 %v %c_u32_1\n"
9746                         "%vec_val              = OpLoad %v4f16 %vec_ptr\n"
9747                         "%f16_val              = OpLoad %f16 %f16_ptr\n"
9748                         "%tmp1                 = OpVectorTimesScalar %v4f16 %c_v4f16_1_1_1_1 %f16_val\n" // vec4(-1)
9749                         "%param1_16            = OpFConvert %v4f16 %param1\n"
9750                         "%tmp2                 = OpFAdd %v4f16 %tmp1 %param1_16\n" // param1 + vec4(-1)
9751                         "%transformed_param    = OpFAdd %v4f16 %tmp2 %vec_val\n" // param1 + vec4(-1) + vec4(1.5, 1.5, 1.5, 1.0)
9752                 },
9753                 {
9754                         // [1|0|0|0.5] [x] = x + 0.5
9755                         // [0|1|0|0.5] [y] = y + 0.5
9756                         // [0|0|1|0.5] [z] = z + 0.5
9757                         // [0|0|0|1  ] [1] = 1
9758                         "matrix",
9759
9760                         FLOAT_16_COMMON_TYPES_AND_CONSTS
9761                         "%mat4x4_f16           = OpTypeMatrix %v4f16 4\n"
9762                         "%v4f16_1_0_0_0        = OpConstantComposite %v4f16 %c_f16_1 %c_f16_0 %c_f16_0 %c_f16_0\n"
9763                         "%v4f16_0_1_0_0        = OpConstantComposite %v4f16 %c_f16_0 %c_f16_1 %c_f16_0 %c_f16_0\n"
9764                         "%v4f16_0_0_1_0        = OpConstantComposite %v4f16 %c_f16_0 %c_f16_0 %c_f16_1 %c_f16_0\n"
9765                         "%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"
9766                         "%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",
9767
9768                         "%param1_16            = OpFConvert %v4f16 %param1\n"
9769                         "%transformed_param    = OpMatrixTimesVector %v4f16 %cval %param1_16\n"
9770                 },
9771                 {
9772                         "array",
9773
9774                         FLOAT_16_COMMON_TYPES_AND_CONSTS
9775                         "%c_v4f16_1_1_1_0      = OpConstantComposite %v4f16 %c_f16_1 %c_f16_1 %c_f16_1 %c_f16_0\n"
9776                         "%fp_a4f16             = OpTypePointer Function %a4f16\n"
9777                         "%f16_n_1              = OpConstant %f16 -1.0\n"
9778                         "%f16_1_5              = OpConstant %f16 !0x3e00\n" // +1.5
9779                         "%carr                 = OpConstantComposite %a4f16 %c_f16_0 %f16_n_1 %f16_1_5 %c_f16_0\n",
9780
9781                         "%v                    = OpVariable %fp_a4f16 Function %carr\n"
9782                         "%f                    = OpAccessChain %fp_f16 %v %c_u32_0\n"
9783                         "%f1                   = OpAccessChain %fp_f16 %v %c_u32_1\n"
9784                         "%f2                   = OpAccessChain %fp_f16 %v %c_u32_2\n"
9785                         "%f3                   = OpAccessChain %fp_f16 %v %c_u32_3\n"
9786                         "%f_val                = OpLoad %f16 %f\n"
9787                         "%f1_val               = OpLoad %f16 %f1\n"
9788                         "%f2_val               = OpLoad %f16 %f2\n"
9789                         "%f3_val               = OpLoad %f16 %f3\n"
9790                         "%ftot1                = OpFAdd %f16 %f_val %f1_val\n"
9791                         "%ftot2                = OpFAdd %f16 %ftot1 %f2_val\n"
9792                         "%ftot3                = OpFAdd %f16 %ftot2 %f3_val\n"  // 0 - 1 + 1.5 + 0
9793                         "%add_vec              = OpVectorTimesScalar %v4f16 %c_v4f16_1_1_1_0 %ftot3\n"
9794                         "%param1_16            = OpFConvert %v4f16 %param1\n"
9795                         "%transformed_param    = OpFAdd %v4f16 %param1_16 %add_vec\n"
9796                 },
9797                 {
9798                         //
9799                         // [
9800                         //   {
9801                         //      0.0,
9802                         //      [ 1.0, 1.0, 1.0, 1.0]
9803                         //   },
9804                         //   {
9805                         //      1.0,
9806                         //      [ 0.0, 0.5, 0.0, 0.0]
9807                         //   }, //     ^^^
9808                         //   {
9809                         //      0.0,
9810                         //      [ 1.0, 1.0, 1.0, 1.0]
9811                         //   }
9812                         // ]
9813                         "array_of_struct_of_array",
9814
9815                         FLOAT_16_COMMON_TYPES_AND_CONSTS
9816                         "%c_v4f16_1_1_1_0      = OpConstantComposite %v4f16 %c_f16_1 %c_f16_1 %c_f16_1 %c_f16_0\n"
9817                         "%fp_a4f16             = OpTypePointer Function %a4f16\n"
9818                         "%stype                = OpTypeStruct %f16 %a4f16\n"
9819                         "%a3stype              = OpTypeArray %stype %c_u32_3\n"
9820                         "%fp_a3stype           = OpTypePointer Function %a3stype\n"
9821                         "%ca4f16_0             = OpConstantComposite %a4f16 %c_f16_0 %c_f16_0_5 %c_f16_0 %c_f16_0\n"
9822                         "%ca4f16_1             = OpConstantComposite %a4f16 %c_f16_1 %c_f16_1 %c_f16_1 %c_f16_1\n"
9823                         "%cstype1              = OpConstantComposite %stype %c_f16_0 %ca4f16_1\n"
9824                         "%cstype2              = OpConstantComposite %stype %c_f16_1 %ca4f16_0\n"
9825                         "%carr                 = OpConstantComposite %a3stype %cstype1 %cstype2 %cstype1",
9826
9827                         "%v                    = OpVariable %fp_a3stype Function %carr\n"
9828                         "%f                    = OpAccessChain %fp_f16 %v %c_u32_1 %c_u32_1 %c_u32_1\n"
9829                         "%f_l                  = OpLoad %f16 %f\n"
9830                         "%add_vec              = OpVectorTimesScalar %v4f16 %c_v4f16_1_1_1_0 %f_l\n"
9831                         "%param1_16            = OpFConvert %v4f16 %param1\n"
9832                         "%transformed_param    = OpFAdd %v4f16 %param1_16 %add_vec\n"
9833                 }
9834         };
9835
9836         getHalfColorsFullAlpha(inputColors);
9837         outputColors[0] = RGBA(255, 255, 255, 255);
9838         outputColors[1] = RGBA(255, 127, 127, 255);
9839         outputColors[2] = RGBA(127, 255, 127, 255);
9840         outputColors[3] = RGBA(127, 127, 255, 255);
9841
9842         extensions.push_back("VK_KHR_16bit_storage");
9843         extensions.push_back("VK_KHR_shader_float16_int8");
9844         features.extFloat16Int8 = EXTFLOAT16INT8FEATURES_FLOAT16;
9845
9846         for (size_t testNdx = 0; testNdx < sizeof(tests) / sizeof(NameConstantsCode); ++testNdx)
9847         {
9848                 map<string, string> fragments;
9849
9850                 fragments["extension"]  = "OpExtension \"SPV_KHR_16bit_storage\"";
9851                 fragments["capability"] = "OpCapability Float16\n";
9852                 fragments["pre_main"]   = tests[testNdx].constants;
9853                 fragments["testfun"]    = string(functionStart) + tests[testNdx].code + functionEnd;
9854
9855                 createTestsForAllStages(tests[testNdx].name, inputColors, outputColors, fragments, resources, extensions, opConstantCompositeTests.get(), features);
9856         }
9857         return opConstantCompositeTests.release();
9858 }
9859
9860 template<typename T>
9861 void finalizeTestsCreation (T&                                                  specResource,
9862                                                         const map<string, string>&      fragments,
9863                                                         tcu::TestContext&                       testCtx,
9864                                                         tcu::TestCaseGroup&                     testGroup,
9865                                                         const std::string&                      testName,
9866                                                         const VulkanFeatures&           vulkanFeatures,
9867                                                         const vector<string>&           extensions,
9868                                                         const IVec3&                            numWorkGroups);
9869
9870 template<>
9871 void finalizeTestsCreation (GraphicsResources&                  specResource,
9872                                                         const map<string, string>&      fragments,
9873                                                         tcu::TestContext&                       ,
9874                                                         tcu::TestCaseGroup&                     testGroup,
9875                                                         const std::string&                      testName,
9876                                                         const VulkanFeatures&           vulkanFeatures,
9877                                                         const vector<string>&           extensions,
9878                                                         const IVec3&                            )
9879 {
9880         RGBA defaultColors[4];
9881         getDefaultColors(defaultColors);
9882
9883         createTestsForAllStages(testName, defaultColors, defaultColors, fragments, specResource, extensions, &testGroup, vulkanFeatures);
9884 }
9885
9886 template<>
9887 void finalizeTestsCreation (ComputeShaderSpec&                  specResource,
9888                                                         const map<string, string>&      fragments,
9889                                                         tcu::TestContext&                       testCtx,
9890                                                         tcu::TestCaseGroup&                     testGroup,
9891                                                         const std::string&                      testName,
9892                                                         const VulkanFeatures&           vulkanFeatures,
9893                                                         const vector<string>&           extensions,
9894                                                         const IVec3&                            numWorkGroups)
9895 {
9896         specResource.numWorkGroups = numWorkGroups;
9897         specResource.requestedVulkanFeatures = vulkanFeatures;
9898         specResource.extensions = extensions;
9899
9900         specResource.assembly = makeComputeShaderAssembly(fragments);
9901
9902         testGroup.addChild(new SpvAsmComputeShaderCase(testCtx, testName.c_str(), "", specResource));
9903 }
9904
9905 template<class SpecResource>
9906 tcu::TestCaseGroup* createFloat16LogicalSet (tcu::TestContext& testCtx, const bool nanSupported)
9907 {
9908         const string                                            nan                                     = nanSupported ? "_nan" : "";
9909         const string                                            groupName                       = "logical" + nan;
9910         de::MovePtr<tcu::TestCaseGroup>         testGroup                       (new tcu::TestCaseGroup(testCtx, groupName.c_str(), "Float 16 logical tests"));
9911
9912         de::Random                                                      rnd                                     (deStringHash(testGroup->getName()));
9913         const string                                            spvCapabilities         = string("OpCapability StorageUniformBufferBlock16\n") + (nanSupported ? "OpCapability SignedZeroInfNanPreserve\n" : "");
9914         const string                                            spvExtensions           = string("OpExtension \"SPV_KHR_16bit_storage\"\n") + (nanSupported ? "OpExtension \"SPV_KHR_float_controls\"\n" : "");
9915         const string                                            spvExecutionMode        = nanSupported ? "OpExecutionMode %BP_main SignedZeroInfNanPreserve 16\n" : "";
9916         const deUint32                                          numDataPoints           = 16;
9917         const vector<deFloat16>                         float16Data                     = getFloat16s(rnd, numDataPoints);
9918         const vector<deFloat16>                         float16Data1            = squarize(float16Data, 0);
9919         const vector<deFloat16>                         float16Data2            = squarize(float16Data, 1);
9920         const vector<deFloat16>                         float16DataVec1         = squarizeVector(float16Data, 0);
9921         const vector<deFloat16>                         float16DataVec2         = squarizeVector(float16Data, 1);
9922         const vector<deFloat16>                         float16OutDummy         (float16Data1.size(), 0);
9923         const vector<deFloat16>                         float16OutVecDummy      (float16DataVec1.size(), 0);
9924
9925         struct TestOp
9926         {
9927                 const char*             opCode;
9928                 VerifyIOFunc    verifyFuncNan;
9929                 VerifyIOFunc    verifyFuncNonNan;
9930                 const deUint32  argCount;
9931         };
9932
9933         const TestOp    testOps[]       =
9934         {
9935                 { "OpIsNan"                                             ,       compareFP16Logical<fp16isNan,                           true,  false, true>,    compareFP16Logical<fp16isNan,                           true,  false, false>,   1       },
9936                 { "OpIsInf"                                             ,       compareFP16Logical<fp16isInf,                           true,  false, true>,    compareFP16Logical<fp16isInf,                           true,  false, false>,   1       },
9937                 { "OpFOrdEqual"                                 ,       compareFP16Logical<fp16isEqual,                         false, true,  true>,    compareFP16Logical<fp16isEqual,                         false, true,  false>,   2       },
9938                 { "OpFUnordEqual"                               ,       compareFP16Logical<fp16isEqual,                         false, false, true>,    compareFP16Logical<fp16isEqual,                         false, false, false>,   2       },
9939                 { "OpFOrdNotEqual"                              ,       compareFP16Logical<fp16isUnequal,                       false, true,  true>,    compareFP16Logical<fp16isUnequal,                       false, true,  false>,   2       },
9940                 { "OpFUnordNotEqual"                    ,       compareFP16Logical<fp16isUnequal,                       false, false, true>,    compareFP16Logical<fp16isUnequal,                       false, false, false>,   2       },
9941                 { "OpFOrdLessThan"                              ,       compareFP16Logical<fp16isLess,                          false, true,  true>,    compareFP16Logical<fp16isLess,                          false, true,  false>,   2       },
9942                 { "OpFUnordLessThan"                    ,       compareFP16Logical<fp16isLess,                          false, false, true>,    compareFP16Logical<fp16isLess,                          false, false, false>,   2       },
9943                 { "OpFOrdGreaterThan"                   ,       compareFP16Logical<fp16isGreater,                       false, true,  true>,    compareFP16Logical<fp16isGreater,                       false, true,  false>,   2       },
9944                 { "OpFUnordGreaterThan"                 ,       compareFP16Logical<fp16isGreater,                       false, false, true>,    compareFP16Logical<fp16isGreater,                       false, false, false>,   2       },
9945                 { "OpFOrdLessThanEqual"                 ,       compareFP16Logical<fp16isLessOrEqual,           false, true,  true>,    compareFP16Logical<fp16isLessOrEqual,           false, true,  false>,   2       },
9946                 { "OpFUnordLessThanEqual"               ,       compareFP16Logical<fp16isLessOrEqual,           false, false, true>,    compareFP16Logical<fp16isLessOrEqual,           false, false, false>,   2       },
9947                 { "OpFOrdGreaterThanEqual"              ,       compareFP16Logical<fp16isGreaterOrEqual,        false, true,  true>,    compareFP16Logical<fp16isGreaterOrEqual,        false, true,  false>,   2       },
9948                 { "OpFUnordGreaterThanEqual"    ,       compareFP16Logical<fp16isGreaterOrEqual,        false, false, true>,    compareFP16Logical<fp16isGreaterOrEqual,        false, false, false>,   2       },
9949         };
9950
9951         { // scalar cases
9952                 const StringTemplate preMain
9953                 (
9954                         "%c_i32_ndp = OpConstant %i32 ${num_data_points}\n"
9955                         "      %f16 = OpTypeFloat 16\n"
9956                         "  %c_f16_0 = OpConstant %f16 0.0\n"
9957                         "  %c_f16_1 = OpConstant %f16 1.0\n"
9958                         "   %up_f16 = OpTypePointer Uniform %f16\n"
9959                         "   %ra_f16 = OpTypeArray %f16 %c_i32_ndp\n"
9960                         "   %SSBO16 = OpTypeStruct %ra_f16\n"
9961                         "%up_SSBO16 = OpTypePointer Uniform %SSBO16\n"
9962                         "%ssbo_src0 = OpVariable %up_SSBO16 Uniform\n"
9963                         "%ssbo_src1 = OpVariable %up_SSBO16 Uniform\n"
9964                         " %ssbo_dst = OpVariable %up_SSBO16 Uniform\n"
9965                 );
9966
9967                 const StringTemplate decoration
9968                 (
9969                         "OpDecorate %ra_f16 ArrayStride 2\n"
9970                         "OpMemberDecorate %SSBO16 0 Offset 0\n"
9971                         "OpDecorate %SSBO16 BufferBlock\n"
9972                         "OpDecorate %ssbo_src0 DescriptorSet 0\n"
9973                         "OpDecorate %ssbo_src0 Binding 0\n"
9974                         "OpDecorate %ssbo_src1 DescriptorSet 0\n"
9975                         "OpDecorate %ssbo_src1 Binding 1\n"
9976                         "OpDecorate %ssbo_dst DescriptorSet 0\n"
9977                         "OpDecorate %ssbo_dst Binding 2\n"
9978                 );
9979
9980                 const StringTemplate testFun
9981                 (
9982                         "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
9983                         "    %param = OpFunctionParameter %v4f32\n"
9984
9985                         "    %entry = OpLabel\n"
9986                         "        %i = OpVariable %fp_i32 Function\n"
9987                         "             OpStore %i %c_i32_0\n"
9988                         "             OpBranch %loop\n"
9989
9990                         "     %loop = OpLabel\n"
9991                         "    %i_cmp = OpLoad %i32 %i\n"
9992                         "       %lt = OpSLessThan %bool %i_cmp %c_i32_ndp\n"
9993                         "             OpLoopMerge %merge %next None\n"
9994                         "             OpBranchConditional %lt %write %merge\n"
9995
9996                         "    %write = OpLabel\n"
9997                         "      %ndx = OpLoad %i32 %i\n"
9998
9999                         "     %src0 = OpAccessChain %up_f16 %ssbo_src0 %c_i32_0 %ndx\n"
10000                         " %val_src0 = OpLoad %f16 %src0\n"
10001
10002                         "${op_arg1_calc}"
10003
10004                         " %val_bdst = ${op_code} %bool %val_src0 ${op_arg1}\n"
10005                         "  %val_dst = OpSelect %f16 %val_bdst %c_f16_1 %c_f16_0\n"
10006                         "      %dst = OpAccessChain %up_f16 %ssbo_dst %c_i32_0 %ndx\n"
10007                         "             OpStore %dst %val_dst\n"
10008                         "             OpBranch %next\n"
10009
10010                         "     %next = OpLabel\n"
10011                         "    %i_cur = OpLoad %i32 %i\n"
10012                         "    %i_new = OpIAdd %i32 %i_cur %c_i32_1\n"
10013                         "             OpStore %i %i_new\n"
10014                         "             OpBranch %loop\n"
10015
10016                         "    %merge = OpLabel\n"
10017                         "             OpReturnValue %param\n"
10018
10019                         "             OpFunctionEnd\n"
10020                 );
10021
10022                 const StringTemplate arg1Calc
10023                 (
10024                         "     %src1 = OpAccessChain %up_f16 %ssbo_src1 %c_i32_0 %ndx\n"
10025                         " %val_src1 = OpLoad %f16 %src1\n"
10026                 );
10027
10028                 for (deUint32 testOpsIdx = 0; testOpsIdx < DE_LENGTH_OF_ARRAY(testOps); ++testOpsIdx)
10029                 {
10030                         const size_t            iterations              = float16Data1.size();
10031                         const TestOp&           testOp                  = testOps[testOpsIdx];
10032                         const string            testName                = de::toLower(string(testOp.opCode)) + "_scalar";
10033                         SpecResource            specResource;
10034                         map<string, string>     specs;
10035                         VulkanFeatures          features;
10036                         map<string, string>     fragments;
10037                         vector<string>          extensions;
10038
10039                         specs["num_data_points"]        = de::toString(iterations);
10040                         specs["op_code"]                        = testOp.opCode;
10041                         specs["op_arg1"]                        = (testOp.argCount == 1) ? "" : "%val_src1";
10042                         specs["op_arg1_calc"]           = (testOp.argCount == 1) ? "" : arg1Calc.specialize(specs);
10043
10044                         fragments["extension"]          = spvExtensions;
10045                         fragments["capability"]         = spvCapabilities;
10046                         fragments["execution_mode"]     = spvExecutionMode;
10047                         fragments["decoration"]         = decoration.specialize(specs);
10048                         fragments["pre_main"]           = preMain.specialize(specs);
10049                         fragments["testfun"]            = testFun.specialize(specs);
10050
10051                         specResource.inputs.push_back(Resource(BufferSp(new Float16Buffer(float16Data1)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
10052                         specResource.inputs.push_back(Resource(BufferSp(new Float16Buffer(float16Data2)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
10053                         specResource.outputs.push_back(Resource(BufferSp(new Float16Buffer(float16OutDummy)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
10054                         specResource.verifyIO = nanSupported ? testOp.verifyFuncNan : testOp.verifyFuncNonNan;
10055
10056                         extensions.push_back("VK_KHR_16bit_storage");
10057                         extensions.push_back("VK_KHR_shader_float16_int8");
10058
10059                         if (nanSupported)
10060                         {
10061                                 extensions.push_back("VK_KHR_shader_float_controls");
10062
10063                                 features.floatControlsProperties.shaderSignedZeroInfNanPreserveFloat16 = DE_TRUE;
10064                         }
10065
10066                         features.ext16BitStorage = EXT16BITSTORAGEFEATURES_UNIFORM_BUFFER_BLOCK;
10067                         features.extFloat16Int8 = EXTFLOAT16INT8FEATURES_FLOAT16;
10068
10069                         finalizeTestsCreation(specResource, fragments, testCtx, *testGroup.get(), testName, features, extensions, IVec3(1, 1, 1));
10070                 }
10071         }
10072         { // vector cases
10073                 const StringTemplate preMain
10074                 (
10075                         "  %c_i32_ndp = OpConstant %i32 ${num_data_points}\n"
10076                         "     %v2bool = OpTypeVector %bool 2\n"
10077                         "        %f16 = OpTypeFloat 16\n"
10078                         "    %c_f16_0 = OpConstant %f16 0.0\n"
10079                         "    %c_f16_1 = OpConstant %f16 1.0\n"
10080                         "      %v2f16 = OpTypeVector %f16 2\n"
10081                         "%c_v2f16_0_0 = OpConstantComposite %v2f16 %c_f16_0 %c_f16_0\n"
10082                         "%c_v2f16_1_1 = OpConstantComposite %v2f16 %c_f16_1 %c_f16_1\n"
10083                         "   %up_v2f16 = OpTypePointer Uniform %v2f16\n"
10084                         "   %ra_v2f16 = OpTypeArray %v2f16 %c_i32_ndp\n"
10085                         "     %SSBO16 = OpTypeStruct %ra_v2f16\n"
10086                         "  %up_SSBO16 = OpTypePointer Uniform %SSBO16\n"
10087                         "  %ssbo_src0 = OpVariable %up_SSBO16 Uniform\n"
10088                         "  %ssbo_src1 = OpVariable %up_SSBO16 Uniform\n"
10089                         "   %ssbo_dst = OpVariable %up_SSBO16 Uniform\n"
10090                 );
10091
10092                 const StringTemplate decoration
10093                 (
10094                         "OpDecorate %ra_v2f16 ArrayStride 4\n"
10095                         "OpMemberDecorate %SSBO16 0 Offset 0\n"
10096                         "OpDecorate %SSBO16 BufferBlock\n"
10097                         "OpDecorate %ssbo_src0 DescriptorSet 0\n"
10098                         "OpDecorate %ssbo_src0 Binding 0\n"
10099                         "OpDecorate %ssbo_src1 DescriptorSet 0\n"
10100                         "OpDecorate %ssbo_src1 Binding 1\n"
10101                         "OpDecorate %ssbo_dst DescriptorSet 0\n"
10102                         "OpDecorate %ssbo_dst Binding 2\n"
10103                 );
10104
10105                 const StringTemplate testFun
10106                 (
10107                         "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
10108                         "    %param = OpFunctionParameter %v4f32\n"
10109
10110                         "    %entry = OpLabel\n"
10111                         "        %i = OpVariable %fp_i32 Function\n"
10112                         "             OpStore %i %c_i32_0\n"
10113                         "             OpBranch %loop\n"
10114
10115                         "     %loop = OpLabel\n"
10116                         "    %i_cmp = OpLoad %i32 %i\n"
10117                         "       %lt = OpSLessThan %bool %i_cmp %c_i32_ndp\n"
10118                         "             OpLoopMerge %merge %next None\n"
10119                         "             OpBranchConditional %lt %write %merge\n"
10120
10121                         "    %write = OpLabel\n"
10122                         "      %ndx = OpLoad %i32 %i\n"
10123
10124                         "     %src0 = OpAccessChain %up_v2f16 %ssbo_src0 %c_i32_0 %ndx\n"
10125                         " %val_src0 = OpLoad %v2f16 %src0\n"
10126
10127                         "${op_arg1_calc}"
10128
10129                         " %val_bdst = ${op_code} %v2bool %val_src0 ${op_arg1}\n"
10130                         "  %val_dst = OpSelect %v2f16 %val_bdst %c_v2f16_1_1 %c_v2f16_0_0\n"
10131                         "      %dst = OpAccessChain %up_v2f16 %ssbo_dst %c_i32_0 %ndx\n"
10132                         "             OpStore %dst %val_dst\n"
10133                         "             OpBranch %next\n"
10134
10135                         "     %next = OpLabel\n"
10136                         "    %i_cur = OpLoad %i32 %i\n"
10137                         "    %i_new = OpIAdd %i32 %i_cur %c_i32_1\n"
10138                         "             OpStore %i %i_new\n"
10139                         "             OpBranch %loop\n"
10140
10141                         "    %merge = OpLabel\n"
10142                         "             OpReturnValue %param\n"
10143
10144                         "             OpFunctionEnd\n"
10145                 );
10146
10147                 const StringTemplate arg1Calc
10148                 (
10149                         "     %src1 = OpAccessChain %up_v2f16 %ssbo_src1 %c_i32_0 %ndx\n"
10150                         " %val_src1 = OpLoad %v2f16 %src1\n"
10151                 );
10152
10153                 for (deUint32 testOpsIdx = 0; testOpsIdx < DE_LENGTH_OF_ARRAY(testOps); ++testOpsIdx)
10154                 {
10155                         const deUint32          itemsPerVec     = 2;
10156                         const size_t            iterations      = float16DataVec1.size() / itemsPerVec;
10157                         const TestOp&           testOp          = testOps[testOpsIdx];
10158                         const string            testName        = de::toLower(string(testOp.opCode)) + "_vector";
10159                         SpecResource            specResource;
10160                         map<string, string>     specs;
10161                         vector<string>          extensions;
10162                         VulkanFeatures          features;
10163                         map<string, string>     fragments;
10164
10165                         specs["num_data_points"]        = de::toString(iterations);
10166                         specs["op_code"]                        = testOp.opCode;
10167                         specs["op_arg1"]                        = (testOp.argCount == 1) ? "" : "%val_src1";
10168                         specs["op_arg1_calc"]           = (testOp.argCount == 1) ? "" : arg1Calc.specialize(specs);
10169
10170                         fragments["extension"]          = spvExtensions;
10171                         fragments["capability"]         = spvCapabilities;
10172                         fragments["execution_mode"]     = spvExecutionMode;
10173                         fragments["decoration"]         = decoration.specialize(specs);
10174                         fragments["pre_main"]           = preMain.specialize(specs);
10175                         fragments["testfun"]            = testFun.specialize(specs);
10176
10177                         specResource.inputs.push_back(Resource(BufferSp(new Float16Buffer(float16DataVec1)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
10178                         specResource.inputs.push_back(Resource(BufferSp(new Float16Buffer(float16DataVec2)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
10179                         specResource.outputs.push_back(Resource(BufferSp(new Float16Buffer(float16OutVecDummy)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
10180                         specResource.verifyIO = nanSupported ? testOp.verifyFuncNan : testOp.verifyFuncNonNan;
10181
10182                         extensions.push_back("VK_KHR_16bit_storage");
10183                         extensions.push_back("VK_KHR_shader_float16_int8");
10184
10185                         if (nanSupported)
10186                         {
10187                                 extensions.push_back("VK_KHR_shader_float_controls");
10188
10189                                 features.floatControlsProperties.shaderSignedZeroInfNanPreserveFloat16 = DE_TRUE;
10190                         }
10191
10192                         features.ext16BitStorage = EXT16BITSTORAGEFEATURES_UNIFORM_BUFFER_BLOCK;
10193                         features.extFloat16Int8 = EXTFLOAT16INT8FEATURES_FLOAT16;
10194
10195                         finalizeTestsCreation(specResource, fragments, testCtx, *testGroup.get(), testName, features, extensions, IVec3(1, 1, 1));
10196                 }
10197         }
10198
10199         return testGroup.release();
10200 }
10201
10202 bool compareFP16FunctionSetFunc (const std::vector<Resource>& inputs, const vector<AllocationSp>& outputAllocs, const std::vector<Resource>&, TestLog& log)
10203 {
10204         if (inputs.size() != 1 || outputAllocs.size() != 1)
10205                 return false;
10206
10207         vector<deUint8> input1Bytes;
10208
10209         inputs[0].getBytes(input1Bytes);
10210
10211         const deUint16* const   input1AsFP16    = (const deUint16*)&input1Bytes[0];
10212         const deUint16* const   outputAsFP16    = (const deUint16*)outputAllocs[0]->getHostPtr();
10213         std::string                             error;
10214
10215         for (size_t idx = 0; idx < input1Bytes.size() / sizeof(deUint16); ++idx)
10216         {
10217                 if (!compare16BitFloat(input1AsFP16[idx], outputAsFP16[idx], error))
10218                 {
10219                         log << TestLog::Message << error << TestLog::EndMessage;
10220
10221                         return false;
10222                 }
10223         }
10224
10225         return true;
10226 }
10227
10228 template<class SpecResource>
10229 tcu::TestCaseGroup* createFloat16FuncSet (tcu::TestContext& testCtx)
10230 {
10231         de::MovePtr<tcu::TestCaseGroup>         testGroup                       (new tcu::TestCaseGroup(testCtx, "function", "Float 16 function call related tests"));
10232
10233         de::Random                                                      rnd                                     (deStringHash(testGroup->getName()));
10234         const StringTemplate                            capabilities            ("OpCapability ${cap}\n");
10235         const deUint32                                          numDataPoints           = 256;
10236         const vector<deFloat16>                         float16InputData        = getFloat16s(rnd, numDataPoints);
10237         const vector<deFloat16>                         float16OutputDummy      (float16InputData.size(), 0);
10238         map<string, string>                                     fragments;
10239
10240         struct TestType
10241         {
10242                 const deUint32  typeComponents;
10243                 const char*             typeName;
10244                 const char*             typeDecls;
10245         };
10246
10247         const TestType  testTypes[]     =
10248         {
10249                 {
10250                         1,
10251                         "f16",
10252                         ""
10253                 },
10254                 {
10255                         2,
10256                         "v2f16",
10257                         "      %v2f16 = OpTypeVector %f16 2\n"
10258                         "  %c_v2f16_0 = OpConstantComposite %v2f16 %c_f16_0 %c_f16_0\n"
10259                 },
10260                 {
10261                         4,
10262                         "v4f16",
10263                         "      %v4f16 = OpTypeVector %f16 4\n"
10264                         "  %c_v4f16_0 = OpConstantComposite %v4f16 %c_f16_0 %c_f16_0 %c_f16_0 %c_f16_0\n"
10265                 },
10266         };
10267
10268         const StringTemplate preMain
10269         (
10270                 "  %c_i32_ndp = OpConstant %i32 ${num_data_points}\n"
10271                 "     %v2bool = OpTypeVector %bool 2\n"
10272                 "        %f16 = OpTypeFloat 16\n"
10273                 "    %c_f16_0 = OpConstant %f16 0.0\n"
10274
10275                 "${type_decls}"
10276
10277                 "  %${tt}_fun = OpTypeFunction %${tt} %${tt}\n"
10278                 "   %up_${tt} = OpTypePointer Uniform %${tt}\n"
10279                 "   %ra_${tt} = OpTypeArray %${tt} %c_i32_ndp\n"
10280                 "     %SSBO16 = OpTypeStruct %ra_${tt}\n"
10281                 "  %up_SSBO16 = OpTypePointer Uniform %SSBO16\n"
10282                 "   %ssbo_src = OpVariable %up_SSBO16 Uniform\n"
10283                 "   %ssbo_dst = OpVariable %up_SSBO16 Uniform\n"
10284         );
10285
10286         const StringTemplate decoration
10287         (
10288                 "OpDecorate %ra_${tt} ArrayStride ${tt_stride}\n"
10289                 "OpMemberDecorate %SSBO16 0 Offset 0\n"
10290                 "OpDecorate %SSBO16 BufferBlock\n"
10291                 "OpDecorate %ssbo_src DescriptorSet 0\n"
10292                 "OpDecorate %ssbo_src Binding 0\n"
10293                 "OpDecorate %ssbo_dst DescriptorSet 0\n"
10294                 "OpDecorate %ssbo_dst Binding 1\n"
10295         );
10296
10297         const StringTemplate testFun
10298         (
10299                 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
10300                 "    %param = OpFunctionParameter %v4f32\n"
10301                 "    %entry = OpLabel\n"
10302
10303                 "        %i = OpVariable %fp_i32 Function\n"
10304                 "             OpStore %i %c_i32_0\n"
10305                 "             OpBranch %loop\n"
10306
10307                 "     %loop = OpLabel\n"
10308                 "    %i_cmp = OpLoad %i32 %i\n"
10309                 "       %lt = OpSLessThan %bool %i_cmp %c_i32_ndp\n"
10310                 "             OpLoopMerge %merge %next None\n"
10311                 "             OpBranchConditional %lt %write %merge\n"
10312
10313                 "    %write = OpLabel\n"
10314                 "      %ndx = OpLoad %i32 %i\n"
10315
10316                 "      %src = OpAccessChain %up_${tt} %ssbo_src %c_i32_0 %ndx\n"
10317                 "  %val_src = OpLoad %${tt} %src\n"
10318
10319                 "  %val_dst = OpFunctionCall %${tt} %pass_fun %val_src\n"
10320                 "      %dst = OpAccessChain %up_${tt} %ssbo_dst %c_i32_0 %ndx\n"
10321                 "             OpStore %dst %val_dst\n"
10322                 "             OpBranch %next\n"
10323
10324                 "     %next = OpLabel\n"
10325                 "    %i_cur = OpLoad %i32 %i\n"
10326                 "    %i_new = OpIAdd %i32 %i_cur %c_i32_1\n"
10327                 "             OpStore %i %i_new\n"
10328                 "             OpBranch %loop\n"
10329
10330                 "    %merge = OpLabel\n"
10331                 "             OpReturnValue %param\n"
10332
10333                 "             OpFunctionEnd\n"
10334
10335                 " %pass_fun = OpFunction %${tt} None %${tt}_fun\n"
10336                 "   %param0 = OpFunctionParameter %${tt}\n"
10337                 " %entry_pf = OpLabel\n"
10338                 "     %res0 = OpFAdd %${tt} %param0 %c_${tt}_0\n"
10339                 "             OpReturnValue %res0\n"
10340                 "             OpFunctionEnd\n"
10341         );
10342
10343         for (deUint32 testTypeIdx = 0; testTypeIdx < DE_LENGTH_OF_ARRAY(testTypes); ++testTypeIdx)
10344         {
10345                 const TestType&         testType                = testTypes[testTypeIdx];
10346                 const string            testName                = testType.typeName;
10347                 const deUint32          itemsPerType    = testType.typeComponents;
10348                 const size_t            iterations              = float16InputData.size() / itemsPerType;
10349                 const size_t            typeStride              = itemsPerType * sizeof(deFloat16);
10350                 SpecResource            specResource;
10351                 map<string, string>     specs;
10352                 VulkanFeatures          features;
10353                 vector<string>          extensions;
10354
10355                 specs["cap"]                            = "StorageUniformBufferBlock16";
10356                 specs["num_data_points"]        = de::toString(iterations);
10357                 specs["tt"]                                     = testType.typeName;
10358                 specs["tt_stride"]                      = de::toString(typeStride);
10359                 specs["type_decls"]                     = testType.typeDecls;
10360
10361                 fragments["extension"]          = "OpExtension \"SPV_KHR_16bit_storage\"";
10362                 fragments["capability"]         = capabilities.specialize(specs);
10363                 fragments["decoration"]         = decoration.specialize(specs);
10364                 fragments["pre_main"]           = preMain.specialize(specs);
10365                 fragments["testfun"]            = testFun.specialize(specs);
10366
10367                 specResource.inputs.push_back(Resource(BufferSp(new Float16Buffer(float16InputData)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
10368                 specResource.outputs.push_back(Resource(BufferSp(new Float16Buffer(float16OutputDummy)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
10369                 specResource.verifyIO = compareFP16FunctionSetFunc;
10370
10371                 extensions.push_back("VK_KHR_16bit_storage");
10372                 extensions.push_back("VK_KHR_shader_float16_int8");
10373
10374                 features.ext16BitStorage = EXT16BITSTORAGEFEATURES_UNIFORM_BUFFER_BLOCK;
10375                 features.extFloat16Int8 = EXTFLOAT16INT8FEATURES_FLOAT16;
10376
10377                 finalizeTestsCreation(specResource, fragments, testCtx, *testGroup.get(), testName, features, extensions, IVec3(1, 1, 1));
10378         }
10379
10380         return testGroup.release();
10381 }
10382
10383 struct getV_    { deUint32 inline operator()(deUint32 v) const  { return v;        } getV_(){} };
10384 struct getV0    { deUint32 inline operator()(deUint32 v) const  { return v & (~1); } getV0(){} };
10385 struct getV1    { deUint32 inline operator()(deUint32 v) const  { return v | ( 1); } getV1(){} };
10386
10387 template<deUint32 R, deUint32 N>
10388 inline static deUint32 getOffset(deUint32 x, deUint32 y, deUint32 n)
10389 {
10390         return N * ((R * y) + x) + n;
10391 }
10392
10393 template<deUint32 R, deUint32 N, class X0, class X1, class Y0, class Y1>
10394 struct getFDelta
10395 {
10396         float operator() (const deFloat16* data, deUint32 x, deUint32 y, deUint32 n, deUint32 flavor) const
10397         {
10398                 DE_STATIC_ASSERT(R%2 == 0);
10399                 DE_ASSERT(flavor == 0);
10400                 DE_UNREF(flavor);
10401
10402                 const X0                        x0;
10403                 const X1                        x1;
10404                 const Y0                        y0;
10405                 const Y1                        y1;
10406                 const deFloat16         v0      = data[getOffset<R, N>(x0(x), y0(y), n)];
10407                 const deFloat16         v1      = data[getOffset<R, N>(x1(x), y1(y), n)];
10408                 const tcu::Float16      f0      = tcu::Float16(v0);
10409                 const tcu::Float16      f1      = tcu::Float16(v1);
10410                 const float                     d0      = f0.asFloat();
10411                 const float                     d1      = f1.asFloat();
10412                 const float                     d       = d1 - d0;
10413
10414                 return d;
10415         }
10416
10417         getFDelta(){}
10418 };
10419
10420 template<deUint32 F, class Class0, class Class1>
10421 struct getFOneOf
10422 {
10423         float operator() (const deFloat16* data, deUint32 x, deUint32 y, deUint32 n, deUint32 flavor) const
10424         {
10425                 DE_ASSERT(flavor < F);
10426
10427                 if (flavor == 0)
10428                 {
10429                         Class0 c;
10430
10431                         return c(data, x, y, n, flavor);
10432                 }
10433                 else
10434                 {
10435                         Class1 c;
10436
10437                         return c(data, x, y, n, flavor - 1);
10438                 }
10439         }
10440
10441         getFOneOf(){}
10442 };
10443
10444 template<class FineX0, class FineX1, class FineY0, class FineY1>
10445 struct calcWidthOf4
10446 {
10447         float operator() (const deFloat16* data, deUint32 x, deUint32 y, deUint32 n, deUint32 flavor) const
10448         {
10449                 DE_ASSERT(flavor < 4);
10450
10451                 const deUint32                                          flavorX = (flavor & 1) == 0 ? 0 : 1;
10452                 const deUint32                                          flavorY = (flavor & 2) == 0 ? 0 : 1;
10453                 const getFOneOf<2, FineX0, FineX1>      cx;
10454                 const getFOneOf<2, FineY0, FineY1>      cy;
10455                 float                                                           v               = 0;
10456
10457                 v += fabsf(cx(data, x, y, n, flavorX));
10458                 v += fabsf(cy(data, x, y, n, flavorY));
10459
10460                 return v;
10461         }
10462
10463         calcWidthOf4(){}
10464 };
10465
10466 template<deUint32 R, deUint32 N, class Derivative>
10467 bool compareDerivativeWithFlavor (const deFloat16* inputAsFP16, const deFloat16* outputAsFP16, deUint32 flavor, std::string& error)
10468 {
10469         const deUint32          numDataPointsByAxis     = R;
10470         const Derivative        derivativeFunc;
10471
10472         for (deUint32 y = 0; y < numDataPointsByAxis; ++y)
10473         for (deUint32 x = 0; x < numDataPointsByAxis; ++x)
10474         for (deUint32 n = 0; n < N; ++n)
10475         {
10476                 const float             expectedFloat   = derivativeFunc(inputAsFP16, x, y, n, flavor);
10477                 deFloat16               expected                = deFloat32To16Round(expectedFloat, DE_ROUNDINGMODE_TO_NEAREST_EVEN);
10478                 const deFloat16 output                  = outputAsFP16[getOffset<R, N>(x, y, n)];
10479
10480                 bool                    reportError             = !compare16BitFloat(expected, output, error);
10481
10482                 if (reportError)
10483                 {
10484                         expected        = deFloat32To16Round(expectedFloat, DE_ROUNDINGMODE_TO_ZERO);
10485                         reportError     = !compare16BitFloat(expected, output, error);
10486                 }
10487
10488                 if (reportError)
10489                 {
10490                         error = "subcase at " + de::toString(x) + "," + de::toString(y) + "," + de::toString(n) + ": " + error;
10491
10492                         return false;
10493                 }
10494         }
10495
10496         return true;
10497 }
10498
10499 template<deUint32 R, deUint32 N, deUint32 FLAVOUR_COUNT, class Derivative>
10500 bool compareDerivative (const std::vector<Resource>& inputs, const vector<AllocationSp>& outputAllocs, const std::vector<Resource>&, TestLog& log)
10501 {
10502         if (inputs.size() != 1 || outputAllocs.size() != 1)
10503                 return false;
10504
10505         deUint32                        successfulRuns                  = FLAVOUR_COUNT;
10506         std::string                     results[FLAVOUR_COUNT];
10507         vector<deUint8>         inputBytes;
10508
10509         inputs[0].getBytes(inputBytes);
10510
10511         const deFloat16*        inputAsFP16             = reinterpret_cast<deFloat16* const>(&inputBytes.front());
10512         const deFloat16*        outputAsFP16    = static_cast<deFloat16*>(outputAllocs[0]->getHostPtr());
10513
10514         DE_ASSERT(inputBytes.size() ==  R * R * N * sizeof(deFloat16));
10515
10516         for (deUint32 flavor = 0; flavor < FLAVOUR_COUNT; ++flavor)
10517                 if (compareDerivativeWithFlavor<R, N, Derivative> (inputAsFP16, outputAsFP16, flavor, results[flavor]))
10518                 {
10519                         break;
10520                 }
10521                 else
10522                 {
10523                         successfulRuns--;
10524                 }
10525
10526         if (successfulRuns == 0)
10527                 for (deUint32 flavor = 0; flavor < FLAVOUR_COUNT; flavor++)
10528                         log << TestLog::Message << "At flavor #" << flavor << " " << results[flavor] << TestLog::EndMessage;
10529
10530         return successfulRuns > 0;
10531 }
10532
10533 template<deUint32 R, deUint32 N>
10534 tcu::TestCaseGroup* createDerivativeTests (tcu::TestContext& testCtx)
10535 {
10536         typedef getFDelta<R, N, getV0, getV1, getV_, getV_> getFDxFine;
10537         typedef getFDelta<R, N, getV_, getV_, getV0, getV1> getFDyFine;
10538
10539         typedef getFDelta<R, N, getV0, getV1, getV0, getV0> getFdxCoarse0;
10540         typedef getFDelta<R, N, getV0, getV1, getV1, getV1> getFdxCoarse1;
10541         typedef getFDelta<R, N, getV0, getV0, getV0, getV1> getFdyCoarse0;
10542         typedef getFDelta<R, N, getV1, getV1, getV0, getV1> getFdyCoarse1;
10543         typedef getFOneOf<2, getFdxCoarse0, getFdxCoarse1> getFDxCoarse;
10544         typedef getFOneOf<2, getFdyCoarse0, getFdyCoarse1> getFDyCoarse;
10545
10546         typedef calcWidthOf4<getFDxFine, getFDxFine, getFDyFine, getFDyFine> getFWidthFine;
10547         typedef calcWidthOf4<getFdxCoarse0, getFdxCoarse1, getFdyCoarse0, getFdyCoarse1> getFWidthCoarse;
10548
10549         typedef getFOneOf<3, getFDxFine, getFDxCoarse> getFDx;
10550         typedef getFOneOf<3, getFDyFine, getFDyCoarse> getFDy;
10551         typedef getFOneOf<5, getFWidthFine, getFWidthCoarse> getFWidth;
10552
10553         const std::string                                       testGroupName           (std::string("derivative_") + de::toString(N));
10554         de::MovePtr<tcu::TestCaseGroup>         testGroup                       (new tcu::TestCaseGroup(testCtx, testGroupName.c_str(), "Derivative instruction tests"));
10555
10556         de::Random                                                      rnd                                     (deStringHash(testGroup->getName()));
10557         const deUint32                                          numDataPointsByAxis     = R;
10558         const deUint32                                          numDataPoints           = N * numDataPointsByAxis * numDataPointsByAxis;
10559         vector<deFloat16>                                       float16InputX;
10560         vector<deFloat16>                                       float16InputY;
10561         vector<deFloat16>                                       float16InputW;
10562         vector<deFloat16>                                       float16OutputDummy      (numDataPoints, 0);
10563         RGBA                                                            defaultColors[4];
10564
10565         getDefaultColors(defaultColors);
10566
10567         float16InputX.reserve(numDataPoints);
10568         for (deUint32 y = 0; y < numDataPointsByAxis; ++y)
10569         for (deUint32 x = 0; x < numDataPointsByAxis; ++x)
10570         for (deUint32 n = 0; n < N; ++n)
10571         {
10572                 const float arg = static_cast<float>(2 * DE_PI) * static_cast<float>(x * (n + 1)) / static_cast<float>(1 * numDataPointsByAxis);
10573
10574                 if (y%2 == 0)
10575                         float16InputX.push_back(tcu::Float16(sin(arg)).bits());
10576                 else
10577                         float16InputX.push_back(tcu::Float16(cos(arg)).bits());
10578         }
10579
10580         float16InputY.reserve(numDataPoints);
10581         for (deUint32 y = 0; y < numDataPointsByAxis; ++y)
10582         for (deUint32 x = 0; x < numDataPointsByAxis; ++x)
10583         for (deUint32 n = 0; n < N; ++n)
10584         {
10585                 const float arg = static_cast<float>(2 * DE_PI) * static_cast<float>(y * (n + 1)) / static_cast<float>(1 * numDataPointsByAxis);
10586
10587                 if (x%2 == 0)
10588                         float16InputY.push_back(tcu::Float16(sin(arg)).bits());
10589                 else
10590                         float16InputY.push_back(tcu::Float16(cos(arg)).bits());
10591         }
10592
10593         const deFloat16 testNumbers[]   =
10594         {
10595                 tcu::Float16( 2.0  ).bits(),
10596                 tcu::Float16( 4.0  ).bits(),
10597                 tcu::Float16( 8.0  ).bits(),
10598                 tcu::Float16( 16.0 ).bits(),
10599                 tcu::Float16( 32.0 ).bits(),
10600                 tcu::Float16( 64.0 ).bits(),
10601                 tcu::Float16( 128.0).bits(),
10602                 tcu::Float16( 256.0).bits(),
10603                 tcu::Float16( 512.0).bits(),
10604                 tcu::Float16(-2.0  ).bits(),
10605                 tcu::Float16(-4.0  ).bits(),
10606                 tcu::Float16(-8.0  ).bits(),
10607                 tcu::Float16(-16.0 ).bits(),
10608                 tcu::Float16(-32.0 ).bits(),
10609                 tcu::Float16(-64.0 ).bits(),
10610                 tcu::Float16(-128.0).bits(),
10611                 tcu::Float16(-256.0).bits(),
10612                 tcu::Float16(-512.0).bits(),
10613         };
10614
10615         float16InputW.reserve(numDataPoints);
10616         for (deUint32 y = 0; y < numDataPointsByAxis; ++y)
10617         for (deUint32 x = 0; x < numDataPointsByAxis; ++x)
10618         for (deUint32 n = 0; n < N; ++n)
10619                 float16InputW.push_back(testNumbers[rnd.getInt(0, DE_LENGTH_OF_ARRAY(testNumbers) - 1)]);
10620
10621         struct TestOp
10622         {
10623                 const char*                     opCode;
10624                 vector<deFloat16>&      inputData;
10625                 VerifyIOFunc            verifyFunc;
10626         };
10627
10628         const TestOp    testOps[]       =
10629         {
10630                 { "OpDPdxFine"          ,       float16InputX   ,       compareDerivative<R, N, 1, getFDxFine           >       },
10631                 { "OpDPdyFine"          ,       float16InputY   ,       compareDerivative<R, N, 1, getFDyFine           >       },
10632                 { "OpFwidthFine"        ,       float16InputW   ,       compareDerivative<R, N, 1, getFWidthFine        >       },
10633                 { "OpDPdxCoarse"        ,       float16InputX   ,       compareDerivative<R, N, 3, getFDx                       >       },
10634                 { "OpDPdyCoarse"        ,       float16InputY   ,       compareDerivative<R, N, 3, getFDy                       >       },
10635                 { "OpFwidthCoarse"      ,       float16InputW   ,       compareDerivative<R, N, 5, getFWidth            >       },
10636                 { "OpDPdx"                      ,       float16InputX   ,       compareDerivative<R, N, 3, getFDx                       >       },
10637                 { "OpDPdy"                      ,       float16InputY   ,       compareDerivative<R, N, 3, getFDy                       >       },
10638                 { "OpFwidth"            ,       float16InputW   ,       compareDerivative<R, N, 5, getFWidth            >       },
10639         };
10640
10641         struct TestType
10642         {
10643                 const deUint32  typeComponents;
10644                 const char*             typeName;
10645                 const char*             typeDecls;
10646         };
10647
10648         const TestType  testTypes[]     =
10649         {
10650                 {
10651                         1,
10652                         "f16",
10653                         ""
10654                 },
10655                 {
10656                         2,
10657                         "v2f16",
10658                         "      %v2f16 = OpTypeVector %f16 2\n"
10659                 },
10660                 {
10661                         4,
10662                         "v4f16",
10663                         "      %v4f16 = OpTypeVector %f16 4\n"
10664                 },
10665         };
10666
10667         const deUint32  testTypeNdx     = (N == 1) ? 0
10668                                                                 : (N == 2) ? 1
10669                                                                 : (N == 4) ? 2
10670                                                                 : DE_LENGTH_OF_ARRAY(testTypes);
10671         const TestType& testType        =       testTypes[testTypeNdx];
10672
10673         DE_ASSERT(testTypeNdx < DE_LENGTH_OF_ARRAY(testTypes));
10674         DE_ASSERT(testType.typeComponents == N);
10675
10676         const StringTemplate preMain
10677         (
10678                 "%c_i32_ndp = OpConstant %i32 ${num_data_points}\n"
10679                 " %c_u32_xw = OpConstant %u32 ${items_by_x}\n"
10680                 "      %f16 = OpTypeFloat 16\n"
10681                 "${type_decls}"
10682                 " %up_${tt} = OpTypePointer Uniform %${tt}\n"
10683                 " %ra_${tt} = OpTypeArray %${tt} %c_i32_ndp\n"
10684                 "   %SSBO16 = OpTypeStruct %ra_${tt}\n"
10685                 "%up_SSBO16 = OpTypePointer Uniform %SSBO16\n"
10686                 " %ssbo_src = OpVariable %up_SSBO16 Uniform\n"
10687                 " %ssbo_dst = OpVariable %up_SSBO16 Uniform\n"
10688         );
10689
10690         const StringTemplate decoration
10691         (
10692                 "OpDecorate %ra_${tt} ArrayStride ${tt_stride}\n"
10693                 "OpMemberDecorate %SSBO16 0 Offset 0\n"
10694                 "OpDecorate %SSBO16 BufferBlock\n"
10695                 "OpDecorate %ssbo_src DescriptorSet 0\n"
10696                 "OpDecorate %ssbo_src Binding 0\n"
10697                 "OpDecorate %ssbo_dst DescriptorSet 0\n"
10698                 "OpDecorate %ssbo_dst Binding 1\n"
10699         );
10700
10701         const StringTemplate testFun
10702         (
10703                 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
10704                 "    %param = OpFunctionParameter %v4f32\n"
10705                 "    %entry = OpLabel\n"
10706
10707                 "  %loc_x_c = OpAccessChain %ip_f32 %BP_gl_FragCoord %c_i32_0\n"
10708                 "  %loc_y_c = OpAccessChain %ip_f32 %BP_gl_FragCoord %c_i32_1\n"
10709                 "      %x_c = OpLoad %f32 %loc_x_c\n"
10710                 "      %y_c = OpLoad %f32 %loc_y_c\n"
10711                 "    %x_idx = OpConvertFToU %u32 %x_c\n"
10712                 "    %y_idx = OpConvertFToU %u32 %y_c\n"
10713                 "    %ndx_y = OpIMul %u32 %y_idx %c_u32_xw\n"
10714                 "      %ndx = OpIAdd %u32 %ndx_y %x_idx\n"
10715
10716                 "      %src = OpAccessChain %up_${tt} %ssbo_src %c_i32_0 %ndx\n"
10717                 "  %val_src = OpLoad %${tt} %src\n"
10718                 "  %val_dst = ${op_code} %${tt} %val_src\n"
10719                 "      %dst = OpAccessChain %up_${tt} %ssbo_dst %c_i32_0 %ndx\n"
10720                 "             OpStore %dst %val_dst\n"
10721                 "             OpBranch %merge\n"
10722
10723                 "    %merge = OpLabel\n"
10724                 "             OpReturnValue %param\n"
10725
10726                 "             OpFunctionEnd\n"
10727         );
10728
10729         for (deUint32 testOpsIdx = 0; testOpsIdx < DE_LENGTH_OF_ARRAY(testOps); ++testOpsIdx)
10730         {
10731                 const TestOp&           testOp                  = testOps[testOpsIdx];
10732                 const string            testName                = de::toLower(string(testOp.opCode));
10733                 const size_t            typeStride              = N * sizeof(deFloat16);
10734                 GraphicsResources       specResource;
10735                 map<string, string>     specs;
10736                 VulkanFeatures          features;
10737                 vector<string>          extensions;
10738                 map<string, string>     fragments;
10739                 SpecConstants           noSpecConstants;
10740                 PushConstants           noPushConstants;
10741                 GraphicsInterfaces      noInterfaces;
10742
10743                 specs["op_code"]                        = testOp.opCode;
10744                 specs["num_data_points"]        = de::toString(testOp.inputData.size() / N);
10745                 specs["items_by_x"]                     = de::toString(numDataPointsByAxis);
10746                 specs["tt"]                                     = testType.typeName;
10747                 specs["tt_stride"]                      = de::toString(typeStride);
10748                 specs["type_decls"]                     = testType.typeDecls;
10749
10750                 fragments["extension"]          = "OpExtension \"SPV_KHR_16bit_storage\"";
10751                 fragments["capability"]         = "OpCapability DerivativeControl\nOpCapability StorageUniformBufferBlock16\n";
10752                 fragments["decoration"]         = decoration.specialize(specs);
10753                 fragments["pre_main"]           = preMain.specialize(specs);
10754                 fragments["testfun"]            = testFun.specialize(specs);
10755
10756                 specResource.inputs.push_back(Resource(BufferSp(new Float16Buffer(testOp.inputData)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
10757                 specResource.outputs.push_back(Resource(BufferSp(new Float16Buffer(float16OutputDummy)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
10758                 specResource.verifyIO = testOp.verifyFunc;
10759
10760                 extensions.push_back("VK_KHR_16bit_storage");
10761                 extensions.push_back("VK_KHR_shader_float16_int8");
10762
10763                 features.extFloat16Int8         = EXTFLOAT16INT8FEATURES_FLOAT16;
10764                 features.ext16BitStorage        = EXT16BITSTORAGEFEATURES_UNIFORM_BUFFER_BLOCK;
10765
10766                 createTestForStage(VK_SHADER_STAGE_FRAGMENT_BIT, testName.c_str(), defaultColors, defaultColors, fragments, noSpecConstants,
10767                                                         noPushConstants, specResource, noInterfaces, extensions, features, testGroup.get(), QP_TEST_RESULT_FAIL, string(), true);
10768         }
10769
10770         return testGroup.release();
10771 }
10772
10773 bool compareFP16VectorExtractFunc (const std::vector<Resource>& inputs, const vector<AllocationSp>& outputAllocs, const std::vector<Resource>&, TestLog& log)
10774 {
10775         if (inputs.size() != 2 || outputAllocs.size() != 1)
10776                 return false;
10777
10778         vector<deUint8> input1Bytes;
10779         vector<deUint8> input2Bytes;
10780
10781         inputs[0].getBytes(input1Bytes);
10782         inputs[1].getBytes(input2Bytes);
10783
10784         DE_ASSERT(input1Bytes.size() > 0);
10785         DE_ASSERT(input2Bytes.size() > 0);
10786         DE_ASSERT(input2Bytes.size() % sizeof(deUint32) == 0);
10787
10788         const size_t                    iterations              = input2Bytes.size() / sizeof(deUint32);
10789         const size_t                    components              = input1Bytes.size() / (sizeof(deFloat16) * iterations);
10790         const deFloat16* const  input1AsFP16    = (const deFloat16*)&input1Bytes[0];
10791         const deUint32* const   inputIndices    = (const deUint32*)&input2Bytes[0];
10792         const deFloat16* const  outputAsFP16    = (const deFloat16*)outputAllocs[0]->getHostPtr();
10793         std::string                             error;
10794
10795         DE_ASSERT(components == 2 || components == 4);
10796         DE_ASSERT(input1Bytes.size() == iterations * components * sizeof(deFloat16));
10797
10798         for (size_t idx = 0; idx < iterations; ++idx)
10799         {
10800                 const deUint32  componentNdx    = inputIndices[idx];
10801
10802                 DE_ASSERT(componentNdx < components);
10803
10804                 const deFloat16 expected                = input1AsFP16[components * idx + componentNdx];
10805
10806                 if (!compare16BitFloat(expected, outputAsFP16[idx], error))
10807                 {
10808                         log << TestLog::Message << "At " << idx << error << TestLog::EndMessage;
10809
10810                         return false;
10811                 }
10812         }
10813
10814         return true;
10815 }
10816
10817 template<class SpecResource>
10818 tcu::TestCaseGroup* createFloat16VectorExtractSet (tcu::TestContext& testCtx)
10819 {
10820         de::MovePtr<tcu::TestCaseGroup>         testGroup                       (new tcu::TestCaseGroup(testCtx, "opvectorextractdynamic", "OpVectorExtractDynamic tests"));
10821
10822         de::Random                                                      rnd                                     (deStringHash(testGroup->getName()));
10823         const deUint32                                          numDataPoints           = 256;
10824         const vector<deFloat16>                         float16InputData        = getFloat16s(rnd, numDataPoints);
10825         const vector<deFloat16>                         float16OutputDummy      (float16InputData.size(), 0);
10826
10827         struct TestType
10828         {
10829                 const deUint32  typeComponents;
10830                 const size_t    typeStride;
10831                 const char*             typeName;
10832                 const char*             typeDecls;
10833         };
10834
10835         const TestType  testTypes[]     =
10836         {
10837                 {
10838                         2,
10839                         2 * sizeof(deFloat16),
10840                         "v2f16",
10841                         "      %v2f16 = OpTypeVector %f16 2\n"
10842                 },
10843                 {
10844                         3,
10845                         4 * sizeof(deFloat16),
10846                         "v3f16",
10847                         "      %v3f16 = OpTypeVector %f16 3\n"
10848                 },
10849                 {
10850                         4,
10851                         4 * sizeof(deFloat16),
10852                         "v4f16",
10853                         "      %v4f16 = OpTypeVector %f16 4\n"
10854                 },
10855         };
10856
10857         const StringTemplate preMain
10858         (
10859                 "  %c_i32_ndp = OpConstant %i32 ${num_data_points}\n"
10860                 "        %f16 = OpTypeFloat 16\n"
10861
10862                 "${type_decl}"
10863
10864                 "   %up_${tt} = OpTypePointer Uniform %${tt}\n"
10865                 "   %ra_${tt} = OpTypeArray %${tt} %c_i32_ndp\n"
10866                 "   %SSBO_SRC = OpTypeStruct %ra_${tt}\n"
10867                 "%up_SSBO_SRC = OpTypePointer Uniform %SSBO_SRC\n"
10868
10869                 "     %up_u32 = OpTypePointer Uniform %u32\n"
10870                 "     %ra_u32 = OpTypeArray %u32 %c_i32_ndp\n"
10871                 "   %SSBO_IDX = OpTypeStruct %ra_u32\n"
10872                 "%up_SSBO_IDX = OpTypePointer Uniform %SSBO_IDX\n"
10873
10874                 "     %up_f16 = OpTypePointer Uniform %f16\n"
10875                 "     %ra_f16 = OpTypeArray %f16 %c_i32_ndp\n"
10876                 "   %SSBO_DST = OpTypeStruct %ra_f16\n"
10877                 "%up_SSBO_DST = OpTypePointer Uniform %SSBO_DST\n"
10878
10879                 "   %ssbo_src = OpVariable %up_SSBO_SRC Uniform\n"
10880                 "   %ssbo_idx = OpVariable %up_SSBO_IDX Uniform\n"
10881                 "   %ssbo_dst = OpVariable %up_SSBO_DST Uniform\n"
10882         );
10883
10884         const StringTemplate decoration
10885         (
10886                 "OpDecorate %ra_${tt} ArrayStride ${tt_stride}\n"
10887                 "OpMemberDecorate %SSBO_SRC 0 Offset 0\n"
10888                 "OpDecorate %SSBO_SRC BufferBlock\n"
10889                 "OpDecorate %ssbo_src DescriptorSet 0\n"
10890                 "OpDecorate %ssbo_src Binding 0\n"
10891
10892                 "OpDecorate %ra_u32 ArrayStride 4\n"
10893                 "OpMemberDecorate %SSBO_IDX 0 Offset 0\n"
10894                 "OpDecorate %SSBO_IDX BufferBlock\n"
10895                 "OpDecorate %ssbo_idx DescriptorSet 0\n"
10896                 "OpDecorate %ssbo_idx Binding 1\n"
10897
10898                 "OpDecorate %ra_f16 ArrayStride 2\n"
10899                 "OpMemberDecorate %SSBO_DST 0 Offset 0\n"
10900                 "OpDecorate %SSBO_DST BufferBlock\n"
10901                 "OpDecorate %ssbo_dst DescriptorSet 0\n"
10902                 "OpDecorate %ssbo_dst Binding 2\n"
10903         );
10904
10905         const StringTemplate testFun
10906         (
10907                 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
10908                 "    %param = OpFunctionParameter %v4f32\n"
10909                 "    %entry = OpLabel\n"
10910
10911                 "        %i = OpVariable %fp_i32 Function\n"
10912                 "             OpStore %i %c_i32_0\n"
10913
10914                 " %will_run = OpFunctionCall %bool %isUniqueIdZero\n"
10915                 "             OpSelectionMerge %end_if None\n"
10916                 "             OpBranchConditional %will_run %run_test %end_if\n"
10917
10918                 " %run_test = OpLabel\n"
10919                 "             OpBranch %loop\n"
10920
10921                 "     %loop = OpLabel\n"
10922                 "    %i_cmp = OpLoad %i32 %i\n"
10923                 "       %lt = OpSLessThan %bool %i_cmp %c_i32_ndp\n"
10924                 "             OpLoopMerge %merge %next None\n"
10925                 "             OpBranchConditional %lt %write %merge\n"
10926
10927                 "    %write = OpLabel\n"
10928                 "      %ndx = OpLoad %i32 %i\n"
10929
10930                 "      %src = OpAccessChain %up_${tt} %ssbo_src %c_i32_0 %ndx\n"
10931                 "  %val_src = OpLoad %${tt} %src\n"
10932
10933                 "  %src_idx = OpAccessChain %up_u32 %ssbo_idx %c_i32_0 %ndx\n"
10934                 "  %val_idx = OpLoad %u32 %src_idx\n"
10935
10936                 "  %val_dst = OpVectorExtractDynamic %f16 %val_src %val_idx\n"
10937                 "      %dst = OpAccessChain %up_f16 %ssbo_dst %c_i32_0 %ndx\n"
10938
10939                 "             OpStore %dst %val_dst\n"
10940                 "             OpBranch %next\n"
10941
10942                 "     %next = OpLabel\n"
10943                 "    %i_cur = OpLoad %i32 %i\n"
10944                 "    %i_new = OpIAdd %i32 %i_cur %c_i32_1\n"
10945                 "             OpStore %i %i_new\n"
10946                 "             OpBranch %loop\n"
10947
10948                 "    %merge = OpLabel\n"
10949                 "             OpBranch %end_if\n"
10950                 "   %end_if = OpLabel\n"
10951                 "             OpReturnValue %param\n"
10952
10953                 "             OpFunctionEnd\n"
10954         );
10955
10956         for (deUint32 testTypeIdx = 0; testTypeIdx < DE_LENGTH_OF_ARRAY(testTypes); ++testTypeIdx)
10957         {
10958                 const TestType&         testType                = testTypes[testTypeIdx];
10959                 const string            testName                = testType.typeName;
10960                 const size_t            itemsPerType    = testType.typeStride / sizeof(deFloat16);
10961                 const size_t            iterations              = float16InputData.size() / itemsPerType;
10962                 SpecResource            specResource;
10963                 map<string, string>     specs;
10964                 VulkanFeatures          features;
10965                 vector<deUint32>        inputDataNdx;
10966                 map<string, string>     fragments;
10967                 vector<string>          extensions;
10968
10969                 for (deUint32 ndx = 0; ndx < iterations; ++ndx)
10970                         inputDataNdx.push_back(rnd.getUint32() % testType.typeComponents);
10971
10972                 specs["num_data_points"]        = de::toString(iterations);
10973                 specs["tt"]                                     = testType.typeName;
10974                 specs["tt_stride"]                      = de::toString(testType.typeStride);
10975                 specs["type_decl"]                      = testType.typeDecls;
10976
10977                 fragments["extension"]          = "OpExtension \"SPV_KHR_16bit_storage\"";
10978                 fragments["capability"]         = "OpCapability StorageUniformBufferBlock16\n";
10979                 fragments["decoration"]         = decoration.specialize(specs);
10980                 fragments["pre_main"]           = preMain.specialize(specs);
10981                 fragments["testfun"]            = testFun.specialize(specs);
10982
10983                 specResource.inputs.push_back(Resource(BufferSp(new Float16Buffer(float16InputData)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
10984                 specResource.inputs.push_back(Resource(BufferSp(new Uint32Buffer(inputDataNdx)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
10985                 specResource.outputs.push_back(Resource(BufferSp(new Float16Buffer(float16OutputDummy)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
10986                 specResource.verifyIO = compareFP16VectorExtractFunc;
10987
10988                 extensions.push_back("VK_KHR_16bit_storage");
10989                 extensions.push_back("VK_KHR_shader_float16_int8");
10990
10991                 features.extFloat16Int8         = EXTFLOAT16INT8FEATURES_FLOAT16;
10992                 features.ext16BitStorage        = EXT16BITSTORAGEFEATURES_UNIFORM_BUFFER_BLOCK;
10993
10994                 finalizeTestsCreation(specResource, fragments, testCtx, *testGroup.get(), testName, features, extensions, IVec3(1, 1, 1));
10995         }
10996
10997         return testGroup.release();
10998 }
10999
11000 template<deUint32 COMPONENTS_COUNT, deUint32 REPLACEMENT>
11001 bool compareFP16VectorInsertFunc (const std::vector<Resource>& inputs, const vector<AllocationSp>& outputAllocs, const std::vector<Resource>&, TestLog& log)
11002 {
11003         if (inputs.size() != 2 || outputAllocs.size() != 1)
11004                 return false;
11005
11006         vector<deUint8> input1Bytes;
11007         vector<deUint8> input2Bytes;
11008
11009         inputs[0].getBytes(input1Bytes);
11010         inputs[1].getBytes(input2Bytes);
11011
11012         DE_ASSERT(input1Bytes.size() > 0);
11013         DE_ASSERT(input2Bytes.size() > 0);
11014         DE_ASSERT(input2Bytes.size() % sizeof(deUint32) == 0);
11015
11016         const size_t                    iterations                      = input2Bytes.size() / sizeof(deUint32);
11017         const size_t                    componentsStride        = input1Bytes.size() / (sizeof(deFloat16) * iterations);
11018         const deFloat16* const  input1AsFP16            = (const deFloat16*)&input1Bytes[0];
11019         const deUint32* const   inputIndices            = (const deUint32*)&input2Bytes[0];
11020         const deFloat16* const  outputAsFP16            = (const deFloat16*)outputAllocs[0]->getHostPtr();
11021         const deFloat16                 magic                           = tcu::Float16(float(REPLACEMENT)).bits();
11022         std::string                             error;
11023
11024         DE_ASSERT(componentsStride == 2 || componentsStride == 4);
11025         DE_ASSERT(input1Bytes.size() == iterations * componentsStride * sizeof(deFloat16));
11026
11027         for (size_t idx = 0; idx < iterations; ++idx)
11028         {
11029                 const deFloat16*        inputVec                = &input1AsFP16[componentsStride * idx];
11030                 const deFloat16*        outputVec               = &outputAsFP16[componentsStride * idx];
11031                 const deUint32          replacedCompNdx = inputIndices[idx];
11032
11033                 DE_ASSERT(replacedCompNdx < COMPONENTS_COUNT);
11034
11035                 for (size_t compNdx = 0; compNdx < COMPONENTS_COUNT; ++compNdx)
11036                 {
11037                         const deFloat16 expected        = (compNdx == replacedCompNdx) ? magic : inputVec[compNdx];
11038
11039                         if (!compare16BitFloat(expected, outputVec[compNdx], error))
11040                         {
11041                                 log << TestLog::Message << "At " << idx << "[" << compNdx << "]: " << error << TestLog::EndMessage;
11042
11043                                 return false;
11044                         }
11045                 }
11046         }
11047
11048         return true;
11049 }
11050
11051 template<class SpecResource>
11052 tcu::TestCaseGroup* createFloat16VectorInsertSet (tcu::TestContext& testCtx)
11053 {
11054         de::MovePtr<tcu::TestCaseGroup>         testGroup                       (new tcu::TestCaseGroup(testCtx, "opvectorinsertdynamic", "OpVectorInsertDynamic tests"));
11055
11056         de::Random                                                      rnd                                     (deStringHash(testGroup->getName()));
11057         const deUint32                                          replacement                     = 42;
11058         const deUint32                                          numDataPoints           = 256;
11059         const vector<deFloat16>                         float16InputData        = getFloat16s(rnd, numDataPoints);
11060         const vector<deFloat16>                         float16OutputDummy      (float16InputData.size(), 0);
11061
11062         struct TestType
11063         {
11064                 const deUint32  typeComponents;
11065                 const size_t    typeStride;
11066                 const char*             typeName;
11067                 const char*             typeDecls;
11068                 VerifyIOFunc    verifyIOFunc;
11069         };
11070
11071         const TestType  testTypes[]     =
11072         {
11073                 {
11074                         2,
11075                         2 * sizeof(deFloat16),
11076                         "v2f16",
11077                         "      %v2f16 = OpTypeVector %f16 2\n",
11078                         compareFP16VectorInsertFunc<2, replacement>
11079                 },
11080                 {
11081                         3,
11082                         4 * sizeof(deFloat16),
11083                         "v3f16",
11084                         "      %v3f16 = OpTypeVector %f16 3\n",
11085                         compareFP16VectorInsertFunc<3, replacement>
11086                 },
11087                 {
11088                         4,
11089                         4 * sizeof(deFloat16),
11090                         "v4f16",
11091                         "      %v4f16 = OpTypeVector %f16 4\n",
11092                         compareFP16VectorInsertFunc<4, replacement>
11093                 },
11094         };
11095
11096         const StringTemplate preMain
11097         (
11098                 "  %c_i32_ndp = OpConstant %i32 ${num_data_points}\n"
11099                 "        %f16 = OpTypeFloat 16\n"
11100                 "  %c_f16_ins = OpConstant %f16 ${replacement}\n"
11101
11102                 "${type_decl}"
11103
11104                 "   %up_${tt} = OpTypePointer Uniform %${tt}\n"
11105                 "   %ra_${tt} = OpTypeArray %${tt} %c_i32_ndp\n"
11106                 "   %SSBO_SRC = OpTypeStruct %ra_${tt}\n"
11107                 "%up_SSBO_SRC = OpTypePointer Uniform %SSBO_SRC\n"
11108
11109                 "     %up_u32 = OpTypePointer Uniform %u32\n"
11110                 "     %ra_u32 = OpTypeArray %u32 %c_i32_ndp\n"
11111                 "   %SSBO_IDX = OpTypeStruct %ra_u32\n"
11112                 "%up_SSBO_IDX = OpTypePointer Uniform %SSBO_IDX\n"
11113
11114                 "   %SSBO_DST = OpTypeStruct %ra_${tt}\n"
11115                 "%up_SSBO_DST = OpTypePointer Uniform %SSBO_DST\n"
11116
11117                 "   %ssbo_src = OpVariable %up_SSBO_SRC Uniform\n"
11118                 "   %ssbo_idx = OpVariable %up_SSBO_IDX Uniform\n"
11119                 "   %ssbo_dst = OpVariable %up_SSBO_DST Uniform\n"
11120         );
11121
11122         const StringTemplate decoration
11123         (
11124                 "OpDecorate %ra_${tt} ArrayStride ${tt_stride}\n"
11125                 "OpMemberDecorate %SSBO_SRC 0 Offset 0\n"
11126                 "OpDecorate %SSBO_SRC BufferBlock\n"
11127                 "OpDecorate %ssbo_src DescriptorSet 0\n"
11128                 "OpDecorate %ssbo_src Binding 0\n"
11129
11130                 "OpDecorate %ra_u32 ArrayStride 4\n"
11131                 "OpMemberDecorate %SSBO_IDX 0 Offset 0\n"
11132                 "OpDecorate %SSBO_IDX BufferBlock\n"
11133                 "OpDecorate %ssbo_idx DescriptorSet 0\n"
11134                 "OpDecorate %ssbo_idx Binding 1\n"
11135
11136                 "OpMemberDecorate %SSBO_DST 0 Offset 0\n"
11137                 "OpDecorate %SSBO_DST BufferBlock\n"
11138                 "OpDecorate %ssbo_dst DescriptorSet 0\n"
11139                 "OpDecorate %ssbo_dst Binding 2\n"
11140         );
11141
11142         const StringTemplate testFun
11143         (
11144                 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
11145                 "    %param = OpFunctionParameter %v4f32\n"
11146                 "    %entry = OpLabel\n"
11147
11148                 "        %i = OpVariable %fp_i32 Function\n"
11149                 "             OpStore %i %c_i32_0\n"
11150
11151                 " %will_run = OpFunctionCall %bool %isUniqueIdZero\n"
11152                 "             OpSelectionMerge %end_if None\n"
11153                 "             OpBranchConditional %will_run %run_test %end_if\n"
11154
11155                 " %run_test = OpLabel\n"
11156                 "             OpBranch %loop\n"
11157
11158                 "     %loop = OpLabel\n"
11159                 "    %i_cmp = OpLoad %i32 %i\n"
11160                 "       %lt = OpSLessThan %bool %i_cmp %c_i32_ndp\n"
11161                 "             OpLoopMerge %merge %next None\n"
11162                 "             OpBranchConditional %lt %write %merge\n"
11163
11164                 "    %write = OpLabel\n"
11165                 "      %ndx = OpLoad %i32 %i\n"
11166
11167                 "      %src = OpAccessChain %up_${tt} %ssbo_src %c_i32_0 %ndx\n"
11168                 "  %val_src = OpLoad %${tt} %src\n"
11169
11170                 "  %src_idx = OpAccessChain %up_u32 %ssbo_idx %c_i32_0 %ndx\n"
11171                 "  %val_idx = OpLoad %u32 %src_idx\n"
11172
11173                 "  %val_dst = OpVectorInsertDynamic %${tt} %val_src %c_f16_ins %val_idx\n"
11174                 "      %dst = OpAccessChain %up_${tt} %ssbo_dst %c_i32_0 %ndx\n"
11175
11176                 "             OpStore %dst %val_dst\n"
11177                 "             OpBranch %next\n"
11178
11179                 "     %next = OpLabel\n"
11180                 "    %i_cur = OpLoad %i32 %i\n"
11181                 "    %i_new = OpIAdd %i32 %i_cur %c_i32_1\n"
11182                 "             OpStore %i %i_new\n"
11183                 "             OpBranch %loop\n"
11184
11185                 "    %merge = OpLabel\n"
11186                 "             OpBranch %end_if\n"
11187                 "   %end_if = OpLabel\n"
11188                 "             OpReturnValue %param\n"
11189
11190                 "             OpFunctionEnd\n"
11191         );
11192
11193         for (deUint32 testTypeIdx = 0; testTypeIdx < DE_LENGTH_OF_ARRAY(testTypes); ++testTypeIdx)
11194         {
11195                 const TestType&         testType                = testTypes[testTypeIdx];
11196                 const string            testName                = testType.typeName;
11197                 const size_t            itemsPerType    = testType.typeStride / sizeof(deFloat16);
11198                 const size_t            iterations              = float16InputData.size() / itemsPerType;
11199                 SpecResource            specResource;
11200                 map<string, string>     specs;
11201                 VulkanFeatures          features;
11202                 vector<deUint32>        inputDataNdx;
11203                 map<string, string>     fragments;
11204                 vector<string>          extensions;
11205
11206                 for (deUint32 ndx = 0; ndx < iterations; ++ndx)
11207                         inputDataNdx.push_back(rnd.getUint32() % testType.typeComponents);
11208
11209                 specs["num_data_points"]        = de::toString(iterations);
11210                 specs["tt"]                                     = testType.typeName;
11211                 specs["tt_stride"]                      = de::toString(testType.typeStride);
11212                 specs["type_decl"]                      = testType.typeDecls;
11213                 specs["replacement"]            = de::toString(replacement);
11214
11215                 fragments["extension"]          = "OpExtension \"SPV_KHR_16bit_storage\"";
11216                 fragments["capability"]         = "OpCapability StorageUniformBufferBlock16\n";
11217                 fragments["decoration"]         = decoration.specialize(specs);
11218                 fragments["pre_main"]           = preMain.specialize(specs);
11219                 fragments["testfun"]            = testFun.specialize(specs);
11220
11221                 specResource.inputs.push_back(Resource(BufferSp(new Float16Buffer(float16InputData)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
11222                 specResource.inputs.push_back(Resource(BufferSp(new Uint32Buffer(inputDataNdx)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
11223                 specResource.outputs.push_back(Resource(BufferSp(new Float16Buffer(float16OutputDummy)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
11224                 specResource.verifyIO = testType.verifyIOFunc;
11225
11226                 extensions.push_back("VK_KHR_16bit_storage");
11227                 extensions.push_back("VK_KHR_shader_float16_int8");
11228
11229                 features.extFloat16Int8         = EXTFLOAT16INT8FEATURES_FLOAT16;
11230                 features.ext16BitStorage        = EXT16BITSTORAGEFEATURES_UNIFORM_BUFFER_BLOCK;
11231
11232                 finalizeTestsCreation(specResource, fragments, testCtx, *testGroup.get(), testName, features, extensions, IVec3(1, 1, 1));
11233         }
11234
11235         return testGroup.release();
11236 }
11237
11238 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)
11239 {
11240         const size_t    compNdxCount    = (vec1Len + vec2Len + 1);
11241         const size_t    compNdxLimited  = iteration % (compNdxCount * compNdxCount);
11242         size_t                  comp;
11243
11244         switch (componentNdx)
11245         {
11246                 case 0: comp = compNdxLimited / compNdxCount; break;
11247                 case 1: comp = compNdxLimited % compNdxCount; break;
11248                 case 2: comp = 0; break;
11249                 case 3: comp = 1; break;
11250                 default: TCU_THROW(InternalError, "Impossible");
11251         }
11252
11253         if (comp >= vec1Len + vec2Len)
11254         {
11255                 validate = false;
11256                 return 0;
11257         }
11258         else
11259         {
11260                 validate = true;
11261                 return (comp < vec1Len) ? input1Vec[comp] : input2Vec[comp - vec1Len];
11262         }
11263 }
11264
11265 template<deUint32 DST_COMPONENTS_COUNT, deUint32 SRC0_COMPONENTS_COUNT, deUint32 SRC1_COMPONENTS_COUNT>
11266 bool compareFP16VectorShuffleFunc (const std::vector<Resource>& inputs, const vector<AllocationSp>& outputAllocs, const std::vector<Resource>&, TestLog& log)
11267 {
11268         DE_STATIC_ASSERT(DST_COMPONENTS_COUNT == 2 || DST_COMPONENTS_COUNT == 3 || DST_COMPONENTS_COUNT == 4);
11269         DE_STATIC_ASSERT(SRC0_COMPONENTS_COUNT == 2 || SRC0_COMPONENTS_COUNT == 3 || SRC0_COMPONENTS_COUNT == 4);
11270         DE_STATIC_ASSERT(SRC1_COMPONENTS_COUNT == 2 || SRC1_COMPONENTS_COUNT == 3 || SRC1_COMPONENTS_COUNT == 4);
11271
11272         if (inputs.size() != 2 || outputAllocs.size() != 1)
11273                 return false;
11274
11275         vector<deUint8> input1Bytes;
11276         vector<deUint8> input2Bytes;
11277
11278         inputs[0].getBytes(input1Bytes);
11279         inputs[1].getBytes(input2Bytes);
11280
11281         DE_ASSERT(input1Bytes.size() > 0);
11282         DE_ASSERT(input2Bytes.size() > 0);
11283         DE_ASSERT(input2Bytes.size() % sizeof(deFloat16) == 0);
11284
11285         const size_t                    componentsStrideDst             = (DST_COMPONENTS_COUNT == 3) ? 4 : DST_COMPONENTS_COUNT;
11286         const size_t                    componentsStrideSrc0    = (SRC0_COMPONENTS_COUNT == 3) ? 4 : SRC0_COMPONENTS_COUNT;
11287         const size_t                    componentsStrideSrc1    = (SRC1_COMPONENTS_COUNT == 3) ? 4 : SRC1_COMPONENTS_COUNT;
11288         const size_t                    iterations                              = input1Bytes.size() / (componentsStrideSrc0 * sizeof(deFloat16));
11289         const deFloat16* const  input1AsFP16                    = (const deFloat16*)&input1Bytes[0];
11290         const deFloat16* const  input2AsFP16                    = (const deFloat16*)&input2Bytes[0];
11291         const deFloat16* const  outputAsFP16                    = (const deFloat16*)outputAllocs[0]->getHostPtr();
11292         std::string                             error;
11293
11294         DE_ASSERT(input1Bytes.size() == iterations * componentsStrideSrc0 * sizeof(deFloat16));
11295         DE_ASSERT(input2Bytes.size() == iterations * componentsStrideSrc1 * sizeof(deFloat16));
11296
11297         for (size_t idx = 0; idx < iterations; ++idx)
11298         {
11299                 const deFloat16*        input1Vec       = &input1AsFP16[componentsStrideSrc0 * idx];
11300                 const deFloat16*        input2Vec       = &input2AsFP16[componentsStrideSrc1 * idx];
11301                 const deFloat16*        outputVec       = &outputAsFP16[componentsStrideDst * idx];
11302
11303                 for (size_t compNdx = 0; compNdx < DST_COMPONENTS_COUNT; ++compNdx)
11304                 {
11305                         bool            validate        = true;
11306                         deFloat16       expected        = getShuffledComponent(idx, compNdx, input1Vec, input2Vec, SRC0_COMPONENTS_COUNT, SRC1_COMPONENTS_COUNT, validate);
11307
11308                         if (validate && !compare16BitFloat(expected, outputVec[compNdx], error))
11309                         {
11310                                 log << TestLog::Message << "At " << idx << "[" << compNdx << "]: " << error << TestLog::EndMessage;
11311
11312                                 return false;
11313                         }
11314                 }
11315         }
11316
11317         return true;
11318 }
11319
11320 VerifyIOFunc getFloat16VectorShuffleVerifyIOFunc (deUint32 dstComponentsCount, deUint32 src0ComponentsCount, deUint32 src1ComponentsCount)
11321 {
11322         DE_ASSERT(dstComponentsCount <= 4);
11323         DE_ASSERT(src0ComponentsCount <= 4);
11324         DE_ASSERT(src1ComponentsCount <= 4);
11325         deUint32 funcCode = 100 * dstComponentsCount + 10 * src0ComponentsCount + src1ComponentsCount;
11326
11327         switch (funcCode)
11328         {
11329                 case 222:return compareFP16VectorShuffleFunc<2, 2, 2>;
11330                 case 223:return compareFP16VectorShuffleFunc<2, 2, 3>;
11331                 case 224:return compareFP16VectorShuffleFunc<2, 2, 4>;
11332                 case 232:return compareFP16VectorShuffleFunc<2, 3, 2>;
11333                 case 233:return compareFP16VectorShuffleFunc<2, 3, 3>;
11334                 case 234:return compareFP16VectorShuffleFunc<2, 3, 4>;
11335                 case 242:return compareFP16VectorShuffleFunc<2, 4, 2>;
11336                 case 243:return compareFP16VectorShuffleFunc<2, 4, 3>;
11337                 case 244:return compareFP16VectorShuffleFunc<2, 4, 4>;
11338                 case 322:return compareFP16VectorShuffleFunc<3, 2, 2>;
11339                 case 323:return compareFP16VectorShuffleFunc<3, 2, 3>;
11340                 case 324:return compareFP16VectorShuffleFunc<3, 2, 4>;
11341                 case 332:return compareFP16VectorShuffleFunc<3, 3, 2>;
11342                 case 333:return compareFP16VectorShuffleFunc<3, 3, 3>;
11343                 case 334:return compareFP16VectorShuffleFunc<3, 3, 4>;
11344                 case 342:return compareFP16VectorShuffleFunc<3, 4, 2>;
11345                 case 343:return compareFP16VectorShuffleFunc<3, 4, 3>;
11346                 case 344:return compareFP16VectorShuffleFunc<3, 4, 4>;
11347                 case 422:return compareFP16VectorShuffleFunc<4, 2, 2>;
11348                 case 423:return compareFP16VectorShuffleFunc<4, 2, 3>;
11349                 case 424:return compareFP16VectorShuffleFunc<4, 2, 4>;
11350                 case 432:return compareFP16VectorShuffleFunc<4, 3, 2>;
11351                 case 433:return compareFP16VectorShuffleFunc<4, 3, 3>;
11352                 case 434:return compareFP16VectorShuffleFunc<4, 3, 4>;
11353                 case 442:return compareFP16VectorShuffleFunc<4, 4, 2>;
11354                 case 443:return compareFP16VectorShuffleFunc<4, 4, 3>;
11355                 case 444:return compareFP16VectorShuffleFunc<4, 4, 4>;
11356                 default: TCU_THROW(InternalError, "Invalid number of components specified.");
11357         }
11358 }
11359
11360 template<class SpecResource>
11361 tcu::TestCaseGroup* createFloat16VectorShuffleSet (tcu::TestContext& testCtx)
11362 {
11363         de::MovePtr<tcu::TestCaseGroup>         testGroup                       (new tcu::TestCaseGroup(testCtx, "opvectorshuffle", "OpVectorShuffle tests"));
11364         const int                                                       testSpecificSeed        = deStringHash(testGroup->getName());
11365         const int                                                       seed                            = testCtx.getCommandLine().getBaseSeed() ^ testSpecificSeed;
11366         de::Random                                                      rnd                                     (seed);
11367         const deUint32                                          numDataPoints           = 128;
11368         map<string, string>                                     fragments;
11369
11370         struct TestType
11371         {
11372                 const deUint32  typeComponents;
11373                 const char*             typeName;
11374         };
11375
11376         const TestType  testTypes[]     =
11377         {
11378                 {
11379                         2,
11380                         "v2f16",
11381                 },
11382                 {
11383                         3,
11384                         "v3f16",
11385                 },
11386                 {
11387                         4,
11388                         "v4f16",
11389                 },
11390         };
11391
11392         const StringTemplate preMain
11393         (
11394                 "    %c_i32_ndp = OpConstant %i32 ${num_data_points}\n"
11395                 "     %c_i32_cc = OpConstant %i32 ${case_count}\n"
11396                 "          %f16 = OpTypeFloat 16\n"
11397                 "        %v2f16 = OpTypeVector %f16 2\n"
11398                 "        %v3f16 = OpTypeVector %f16 3\n"
11399                 "        %v4f16 = OpTypeVector %f16 4\n"
11400
11401                 "     %up_v2f16 = OpTypePointer Uniform %v2f16\n"
11402                 "     %ra_v2f16 = OpTypeArray %v2f16 %c_i32_ndp\n"
11403                 "   %SSBO_v2f16 = OpTypeStruct %ra_v2f16\n"
11404                 "%up_SSBO_v2f16 = OpTypePointer Uniform %SSBO_v2f16\n"
11405
11406                 "     %up_v3f16 = OpTypePointer Uniform %v3f16\n"
11407                 "     %ra_v3f16 = OpTypeArray %v3f16 %c_i32_ndp\n"
11408                 "   %SSBO_v3f16 = OpTypeStruct %ra_v3f16\n"
11409                 "%up_SSBO_v3f16 = OpTypePointer Uniform %SSBO_v3f16\n"
11410
11411                 "     %up_v4f16 = OpTypePointer Uniform %v4f16\n"
11412                 "     %ra_v4f16 = OpTypeArray %v4f16 %c_i32_ndp\n"
11413                 "   %SSBO_v4f16 = OpTypeStruct %ra_v4f16\n"
11414                 "%up_SSBO_v4f16 = OpTypePointer Uniform %SSBO_v4f16\n"
11415
11416                 "        %fun_t = OpTypeFunction %${tt_dst} %${tt_src0} %${tt_src1} %i32\n"
11417
11418                 "    %ssbo_src0 = OpVariable %up_SSBO_${tt_src0} Uniform\n"
11419                 "    %ssbo_src1 = OpVariable %up_SSBO_${tt_src1} Uniform\n"
11420                 "     %ssbo_dst = OpVariable %up_SSBO_${tt_dst} Uniform\n"
11421         );
11422
11423         const StringTemplate decoration
11424         (
11425                 "OpDecorate %ra_v2f16 ArrayStride 4\n"
11426                 "OpDecorate %ra_v3f16 ArrayStride 8\n"
11427                 "OpDecorate %ra_v4f16 ArrayStride 8\n"
11428
11429                 "OpMemberDecorate %SSBO_v2f16 0 Offset 0\n"
11430                 "OpDecorate %SSBO_v2f16 BufferBlock\n"
11431
11432                 "OpMemberDecorate %SSBO_v3f16 0 Offset 0\n"
11433                 "OpDecorate %SSBO_v3f16 BufferBlock\n"
11434
11435                 "OpMemberDecorate %SSBO_v4f16 0 Offset 0\n"
11436                 "OpDecorate %SSBO_v4f16 BufferBlock\n"
11437
11438                 "OpDecorate %ssbo_src0 DescriptorSet 0\n"
11439                 "OpDecorate %ssbo_src0 Binding 0\n"
11440                 "OpDecorate %ssbo_src1 DescriptorSet 0\n"
11441                 "OpDecorate %ssbo_src1 Binding 1\n"
11442                 "OpDecorate %ssbo_dst DescriptorSet 0\n"
11443                 "OpDecorate %ssbo_dst Binding 2\n"
11444         );
11445
11446         const StringTemplate testFun
11447         (
11448                 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
11449                 "    %param = OpFunctionParameter %v4f32\n"
11450                 "    %entry = OpLabel\n"
11451
11452                 "        %i = OpVariable %fp_i32 Function\n"
11453                 "             OpStore %i %c_i32_0\n"
11454
11455                 " %will_run = OpFunctionCall %bool %isUniqueIdZero\n"
11456                 "             OpSelectionMerge %end_if None\n"
11457                 "             OpBranchConditional %will_run %run_test %end_if\n"
11458
11459                 " %run_test = OpLabel\n"
11460                 "             OpBranch %loop\n"
11461
11462                 "     %loop = OpLabel\n"
11463                 "    %i_cmp = OpLoad %i32 %i\n"
11464                 "       %lt = OpSLessThan %bool %i_cmp %c_i32_ndp\n"
11465                 "             OpLoopMerge %merge %next None\n"
11466                 "             OpBranchConditional %lt %write %merge\n"
11467
11468                 "    %write = OpLabel\n"
11469                 "      %ndx = OpLoad %i32 %i\n"
11470                 "     %src0 = OpAccessChain %up_${tt_src0} %ssbo_src0 %c_i32_0 %ndx\n"
11471                 " %val_src0 = OpLoad %${tt_src0} %src0\n"
11472                 "     %src1 = OpAccessChain %up_${tt_src1} %ssbo_src1 %c_i32_0 %ndx\n"
11473                 " %val_src1 = OpLoad %${tt_src1} %src1\n"
11474                 "  %val_dst = OpFunctionCall %${tt_dst} %sw_fun %val_src0 %val_src1 %ndx\n"
11475                 "      %dst = OpAccessChain %up_${tt_dst} %ssbo_dst %c_i32_0 %ndx\n"
11476                 "             OpStore %dst %val_dst\n"
11477                 "             OpBranch %next\n"
11478
11479                 "     %next = OpLabel\n"
11480                 "    %i_cur = OpLoad %i32 %i\n"
11481                 "    %i_new = OpIAdd %i32 %i_cur %c_i32_1\n"
11482                 "             OpStore %i %i_new\n"
11483                 "             OpBranch %loop\n"
11484
11485                 "    %merge = OpLabel\n"
11486                 "             OpBranch %end_if\n"
11487                 "   %end_if = OpLabel\n"
11488                 "             OpReturnValue %param\n"
11489                 "             OpFunctionEnd\n"
11490                 "\n"
11491
11492                 "   %sw_fun = OpFunction %${tt_dst} None %fun_t\n"
11493                 "%sw_param0 = OpFunctionParameter %${tt_src0}\n"
11494                 "%sw_param1 = OpFunctionParameter %${tt_src1}\n"
11495                 "%sw_paramn = OpFunctionParameter %i32\n"
11496                 " %sw_entry = OpLabel\n"
11497                 "   %modulo = OpSMod %i32 %sw_paramn %c_i32_cc\n"
11498                 "             OpSelectionMerge %switch_e None\n"
11499                 "             OpSwitch %modulo %default ${case_list}\n"
11500                 "${case_bodies}"
11501                 "%default   = OpLabel\n"
11502                 "             OpUnreachable\n" // Unreachable default case for switch statement
11503                 "%switch_e  = OpLabel\n"
11504                 "             OpUnreachable\n" // Unreachable merge block for switch statement
11505                 "             OpFunctionEnd\n"
11506         );
11507
11508         const StringTemplate testCaseBody
11509         (
11510                 "%case_${case_ndx}    = OpLabel\n"
11511                 "%val_dst_${case_ndx} = OpVectorShuffle %${tt_dst} %sw_param0 %sw_param1 ${shuffle}\n"
11512                 "             OpReturnValue %val_dst_${case_ndx}\n"
11513         );
11514
11515         for (deUint32 dstTypeIdx = 0; dstTypeIdx < DE_LENGTH_OF_ARRAY(testTypes); ++dstTypeIdx)
11516         {
11517                 const TestType& dstType                 = testTypes[dstTypeIdx];
11518
11519                 for (deUint32 comp0Idx = 0; comp0Idx < DE_LENGTH_OF_ARRAY(testTypes); ++comp0Idx)
11520                 {
11521                         const TestType& src0Type        = testTypes[comp0Idx];
11522
11523                         for (deUint32 comp1Idx = 0; comp1Idx < DE_LENGTH_OF_ARRAY(testTypes); ++comp1Idx)
11524                         {
11525                                 const TestType&                 src1Type                        = testTypes[comp1Idx];
11526                                 const deUint32                  input0Stride            = (src0Type.typeComponents == 3) ? 4 : src0Type.typeComponents;
11527                                 const deUint32                  input1Stride            = (src1Type.typeComponents == 3) ? 4 : src1Type.typeComponents;
11528                                 const deUint32                  outputStride            = (dstType.typeComponents == 3) ? 4 : dstType.typeComponents;
11529                                 const vector<deFloat16> float16Input0Data       = getFloat16s(rnd, input0Stride * numDataPoints);
11530                                 const vector<deFloat16> float16Input1Data       = getFloat16s(rnd, input1Stride * numDataPoints);
11531                                 const vector<deFloat16> float16OutputDummy      (outputStride * numDataPoints, 0);
11532                                 const string                    testName                        = de::toString(dstType.typeComponents) + de::toString(src0Type.typeComponents) + de::toString(src1Type.typeComponents);
11533                                 deUint32                                caseCount                       = 0;
11534                                 SpecResource                    specResource;
11535                                 map<string, string>             specs;
11536                                 vector<string>                  extensions;
11537                                 VulkanFeatures                  features;
11538                                 string                                  caseBodies;
11539                                 string                                  caseList;
11540
11541                                 // Generate case
11542                                 {
11543                                         vector<string>  componentList;
11544
11545                                         // Generate component possible indices for OpVectorShuffle for components 0 and 1 in output vector
11546                                         {
11547                                                 deUint32                caseNo          = 0;
11548
11549                                                 for (deUint32 comp0IdxLocal = 0; comp0IdxLocal < src0Type.typeComponents; ++comp0IdxLocal)
11550                                                         componentList.push_back(de::toString(caseNo++));
11551                                                 for (deUint32 comp1IdxLocal = 0; comp1IdxLocal < src1Type.typeComponents; ++comp1IdxLocal)
11552                                                         componentList.push_back(de::toString(caseNo++));
11553                                                 componentList.push_back("0xFFFFFFFF");
11554                                         }
11555
11556                                         for (deUint32 comp0IdxLocal = 0; comp0IdxLocal < componentList.size(); ++comp0IdxLocal)
11557                                         {
11558                                                 for (deUint32 comp1IdxLocal = 0; comp1IdxLocal < componentList.size(); ++comp1IdxLocal)
11559                                                 {
11560                                                         map<string, string>     specCase;
11561                                                         string                          shuffle         = componentList[comp0IdxLocal] + " " + componentList[comp1IdxLocal];
11562
11563                                                         for (deUint32 compIdx = 2; compIdx < dstType.typeComponents; ++compIdx)
11564                                                                 shuffle += " " + de::toString(compIdx - 2);
11565
11566                                                         specCase["case_ndx"]    = de::toString(caseCount);
11567                                                         specCase["shuffle"]             = shuffle;
11568                                                         specCase["tt_dst"]              = dstType.typeName;
11569
11570                                                         caseBodies      += testCaseBody.specialize(specCase);
11571                                                         caseList        += de::toString(caseCount) + " %case_" + de::toString(caseCount) + " ";
11572
11573                                                         caseCount++;
11574                                                 }
11575                                         }
11576                                 }
11577
11578                                 specs["num_data_points"]        = de::toString(numDataPoints);
11579                                 specs["tt_dst"]                         = dstType.typeName;
11580                                 specs["tt_src0"]                        = src0Type.typeName;
11581                                 specs["tt_src1"]                        = src1Type.typeName;
11582                                 specs["case_bodies"]            = caseBodies;
11583                                 specs["case_list"]                      = caseList;
11584                                 specs["case_count"]                     = de::toString(caseCount);
11585
11586                                 fragments["extension"]          = "OpExtension \"SPV_KHR_16bit_storage\"";
11587                                 fragments["capability"]         = "OpCapability StorageUniformBufferBlock16\n";
11588                                 fragments["decoration"]         = decoration.specialize(specs);
11589                                 fragments["pre_main"]           = preMain.specialize(specs);
11590                                 fragments["testfun"]            = testFun.specialize(specs);
11591
11592                                 specResource.inputs.push_back(Resource(BufferSp(new Float16Buffer(float16Input0Data)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
11593                                 specResource.inputs.push_back(Resource(BufferSp(new Float16Buffer(float16Input1Data)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
11594                                 specResource.outputs.push_back(Resource(BufferSp(new Float16Buffer(float16OutputDummy)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
11595                                 specResource.verifyIO = getFloat16VectorShuffleVerifyIOFunc(dstType.typeComponents, src0Type.typeComponents, src1Type.typeComponents);
11596
11597                                 extensions.push_back("VK_KHR_16bit_storage");
11598                                 extensions.push_back("VK_KHR_shader_float16_int8");
11599
11600                                 features.extFloat16Int8         = EXTFLOAT16INT8FEATURES_FLOAT16;
11601                                 features.ext16BitStorage        = EXT16BITSTORAGEFEATURES_UNIFORM_BUFFER_BLOCK;
11602
11603                                 finalizeTestsCreation(specResource, fragments, testCtx, *testGroup.get(), testName, features, extensions, IVec3(1, 1, 1));
11604                         }
11605                 }
11606         }
11607
11608         return testGroup.release();
11609 }
11610
11611 bool compareFP16CompositeFunc (const std::vector<Resource>& inputs, const vector<AllocationSp>& outputAllocs, const std::vector<Resource>&, TestLog& log)
11612 {
11613         if (inputs.size() != 1 || outputAllocs.size() != 1)
11614                 return false;
11615
11616         vector<deUint8> input1Bytes;
11617
11618         inputs[0].getBytes(input1Bytes);
11619
11620         DE_ASSERT(input1Bytes.size() > 0);
11621         DE_ASSERT(input1Bytes.size() % sizeof(deFloat16) == 0);
11622
11623         const size_t                    iterations              = input1Bytes.size() / sizeof(deFloat16);
11624         const deFloat16* const  input1AsFP16    = (const deFloat16*)&input1Bytes[0];
11625         const deFloat16* const  outputAsFP16    = (const deFloat16*)outputAllocs[0]->getHostPtr();
11626         const deFloat16                 exceptionValue  = tcu::Float16(-1.0).bits();
11627         std::string                             error;
11628
11629         for (size_t idx = 0; idx < iterations; ++idx)
11630         {
11631                 if (input1AsFP16[idx] == exceptionValue)
11632                         continue;
11633
11634                 if (!compare16BitFloat(input1AsFP16[idx], outputAsFP16[idx], error))
11635                 {
11636                         log << TestLog::Message << "At " << idx << ":" << error << TestLog::EndMessage;
11637
11638                         return false;
11639                 }
11640         }
11641
11642         return true;
11643 }
11644
11645 template<class SpecResource>
11646 tcu::TestCaseGroup* createFloat16CompositeConstructSet (tcu::TestContext& testCtx)
11647 {
11648         de::MovePtr<tcu::TestCaseGroup>         testGroup                               (new tcu::TestCaseGroup(testCtx, "opcompositeconstruct", "OpCompositeConstruct tests"));
11649         const deUint32                                          numElements                             = 8;
11650         const string                                            testName                                = "struct";
11651         const deUint32                                          structItemsCount                = 88;
11652         const deUint32                                          exceptionIndices[]              = { 1, 7, 15, 17, 25, 33, 51, 55, 59, 63, 67, 71, 84, 85, 86, 87 };
11653         const deFloat16                                         exceptionValue                  = tcu::Float16(-1.0).bits();
11654         const deUint32                                          fieldModifier                   = 2;
11655         const deUint32                                          fieldModifiedMulIndex   = 60;
11656         const deUint32                                          fieldModifiedAddIndex   = 66;
11657
11658         const StringTemplate preMain
11659         (
11660                 "    %c_i32_ndp = OpConstant %i32 ${num_elements}\n"
11661                 "          %f16 = OpTypeFloat 16\n"
11662                 "        %v2f16 = OpTypeVector %f16 2\n"
11663                 "        %v3f16 = OpTypeVector %f16 3\n"
11664                 "        %v4f16 = OpTypeVector %f16 4\n"
11665                 "    %c_f16_mod = OpConstant %f16 ${field_modifier}\n"
11666
11667                 "${consts}"
11668
11669                 "      %c_u32_5 = OpConstant %u32 5\n"
11670
11671                 " %f16arr3      = OpTypeArray %f16 %c_u32_3\n"
11672                 " %v2f16arr3    = OpTypeArray %v2f16 %c_u32_3\n"
11673                 " %v2f16arr5    = OpTypeArray %v2f16 %c_u32_5\n"
11674                 " %v3f16arr5    = OpTypeArray %v3f16 %c_u32_5\n"
11675                 " %v4f16arr3    = OpTypeArray %v4f16 %c_u32_3\n"
11676                 " %struct16     = OpTypeStruct %f16 %v2f16arr3\n"
11677                 " %struct16arr3 = OpTypeArray %struct16 %c_u32_3\n"
11678                 " %st_test      = OpTypeStruct %f16 %v2f16 %v3f16 %v4f16 %f16arr3 %struct16arr3 %v2f16arr5 %f16 %v3f16arr5 %v4f16arr3\n"
11679
11680                 "        %up_st = OpTypePointer Uniform %st_test\n"
11681                 "        %ra_st = OpTypeArray %st_test %c_i32_ndp\n"
11682                 "      %SSBO_st = OpTypeStruct %ra_st\n"
11683                 "   %up_SSBO_st = OpTypePointer Uniform %SSBO_st\n"
11684
11685                 "     %ssbo_dst = OpVariable %up_SSBO_st Uniform\n"
11686         );
11687
11688         const StringTemplate decoration
11689         (
11690                 "OpDecorate %SSBO_st BufferBlock\n"
11691                 "OpDecorate %ra_st ArrayStride ${struct_item_size}\n"
11692                 "OpDecorate %ssbo_dst DescriptorSet 0\n"
11693                 "OpDecorate %ssbo_dst Binding 1\n"
11694
11695                 "OpMemberDecorate %SSBO_st 0 Offset 0\n"
11696
11697                 "OpDecorate %v2f16arr3 ArrayStride 4\n"
11698                 "OpMemberDecorate %struct16 0 Offset 0\n"
11699                 "OpMemberDecorate %struct16 1 Offset 4\n"
11700                 "OpDecorate %struct16arr3 ArrayStride 16\n"
11701                 "OpDecorate %f16arr3 ArrayStride 2\n"
11702                 "OpDecorate %v2f16arr5 ArrayStride 4\n"
11703                 "OpDecorate %v3f16arr5 ArrayStride 8\n"
11704                 "OpDecorate %v4f16arr3 ArrayStride 8\n"
11705
11706                 "OpMemberDecorate %st_test 0 Offset 0\n"
11707                 "OpMemberDecorate %st_test 1 Offset 4\n"
11708                 "OpMemberDecorate %st_test 2 Offset 8\n"
11709                 "OpMemberDecorate %st_test 3 Offset 16\n"
11710                 "OpMemberDecorate %st_test 4 Offset 24\n"
11711                 "OpMemberDecorate %st_test 5 Offset 32\n"
11712                 "OpMemberDecorate %st_test 6 Offset 80\n"
11713                 "OpMemberDecorate %st_test 7 Offset 100\n"
11714                 "OpMemberDecorate %st_test 8 Offset 104\n"
11715                 "OpMemberDecorate %st_test 9 Offset 144\n"
11716         );
11717
11718         const StringTemplate testFun
11719         (
11720                 " %test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
11721                 "     %param = OpFunctionParameter %v4f32\n"
11722                 "     %entry = OpLabel\n"
11723
11724                 "         %i = OpVariable %fp_i32 Function\n"
11725                 "              OpStore %i %c_i32_0\n"
11726
11727                 "  %will_run = OpFunctionCall %bool %isUniqueIdZero\n"
11728                 "              OpSelectionMerge %end_if None\n"
11729                 "              OpBranchConditional %will_run %run_test %end_if\n"
11730
11731                 "  %run_test = OpLabel\n"
11732                 "              OpBranch %loop\n"
11733
11734                 "      %loop = OpLabel\n"
11735                 "     %i_cmp = OpLoad %i32 %i\n"
11736                 "        %lt = OpSLessThan %bool %i_cmp %c_i32_ndp\n"
11737                 "              OpLoopMerge %merge %next None\n"
11738                 "              OpBranchConditional %lt %write %merge\n"
11739
11740                 "     %write = OpLabel\n"
11741                 "       %ndx = OpLoad %i32 %i\n"
11742
11743                 "      %fld1 = OpCompositeConstruct %v2f16 %c_f16_2 %c_f16_3\n"
11744                 "      %fld2 = OpCompositeConstruct %v3f16 %c_f16_4 %c_f16_5 %c_f16_6\n"
11745                 "      %fld3 = OpCompositeConstruct %v4f16 %c_f16_8 %c_f16_9 %c_f16_10 %c_f16_11\n"
11746
11747                 "      %fld4 = OpCompositeConstruct %f16arr3 %c_f16_12 %c_f16_13 %c_f16_14\n"
11748
11749                 "%fld5_0_1_0 = OpCompositeConstruct %v2f16 %c_f16_18 %c_f16_19\n"
11750                 "%fld5_0_1_1 = OpCompositeConstruct %v2f16 %c_f16_20 %c_f16_21\n"
11751                 "%fld5_0_1_2 = OpCompositeConstruct %v2f16 %c_f16_22 %c_f16_23\n"
11752                 "  %fld5_0_1 = OpCompositeConstruct %v2f16arr3 %fld5_0_1_0 %fld5_0_1_1 %fld5_0_1_2\n"
11753                 "    %fld5_0 = OpCompositeConstruct %struct16 %c_f16_16 %fld5_0_1\n"
11754
11755                 "%fld5_1_1_0 = OpCompositeConstruct %v2f16 %c_f16_26 %c_f16_27\n"
11756                 "%fld5_1_1_1 = OpCompositeConstruct %v2f16 %c_f16_28 %c_f16_29\n"
11757                 "%fld5_1_1_2 = OpCompositeConstruct %v2f16 %c_f16_30 %c_f16_31\n"
11758                 "  %fld5_1_1 = OpCompositeConstruct %v2f16arr3 %fld5_1_1_0 %fld5_1_1_1 %fld5_1_1_2\n"
11759                 "    %fld5_1 = OpCompositeConstruct %struct16 %c_f16_24 %fld5_1_1\n"
11760
11761                 "%fld5_2_1_0 = OpCompositeConstruct %v2f16 %c_f16_34 %c_f16_35\n"
11762                 "%fld5_2_1_1 = OpCompositeConstruct %v2f16 %c_f16_36 %c_f16_37\n"
11763                 "%fld5_2_1_2 = OpCompositeConstruct %v2f16 %c_f16_38 %c_f16_39\n"
11764                 "  %fld5_2_1 = OpCompositeConstruct %v2f16arr3 %fld5_2_1_0 %fld5_2_1_1 %fld5_2_1_2\n"
11765                 "    %fld5_2 = OpCompositeConstruct %struct16 %c_f16_32 %fld5_2_1\n"
11766
11767                 "      %fld5 = OpCompositeConstruct %struct16arr3 %fld5_0 %fld5_1 %fld5_2\n"
11768
11769                 "    %fld6_0 = OpCompositeConstruct %v2f16 %c_f16_40 %c_f16_41\n"
11770                 "    %fld6_1 = OpCompositeConstruct %v2f16 %c_f16_42 %c_f16_43\n"
11771                 "    %fld6_2 = OpCompositeConstruct %v2f16 %c_f16_44 %c_f16_45\n"
11772                 "    %fld6_3 = OpCompositeConstruct %v2f16 %c_f16_46 %c_f16_47\n"
11773                 "    %fld6_4 = OpCompositeConstruct %v2f16 %c_f16_48 %c_f16_49\n"
11774                 "      %fld6 = OpCompositeConstruct %v2f16arr5 %fld6_0 %fld6_1 %fld6_2 %fld6_3 %fld6_4\n"
11775
11776                 "      %fndx = OpConvertSToF %f16 %ndx\n"
11777                 "  %fld8_2a0 = OpFMul %f16 %fndx %c_f16_mod\n"
11778                 "  %fld8_3b1 = OpFAdd %f16 %fndx %c_f16_mod\n"
11779
11780                 "   %fld8_2a = OpCompositeConstruct %v2f16 %fld8_2a0 %c_f16_61\n"
11781                 "   %fld8_3b = OpCompositeConstruct %v2f16 %c_f16_65 %fld8_3b1\n"
11782                 "    %fld8_0 = OpCompositeConstruct %v3f16 %c_f16_52 %c_f16_53 %c_f16_54\n"
11783                 "    %fld8_1 = OpCompositeConstruct %v3f16 %c_f16_56 %c_f16_57 %c_f16_58\n"
11784                 "    %fld8_2 = OpCompositeConstruct %v3f16 %fld8_2a %c_f16_62\n"
11785                 "    %fld8_3 = OpCompositeConstruct %v3f16 %c_f16_64 %fld8_3b\n"
11786                 "    %fld8_4 = OpCompositeConstruct %v3f16 %c_f16_68 %c_f16_69 %c_f16_70\n"
11787                 "      %fld8 = OpCompositeConstruct %v3f16arr5 %fld8_0 %fld8_1 %fld8_2 %fld8_3 %fld8_4\n"
11788
11789                 "    %fld9_0 = OpCompositeConstruct %v4f16 %c_f16_72 %c_f16_73 %c_f16_74 %c_f16_75\n"
11790                 "    %fld9_1 = OpCompositeConstruct %v4f16 %c_f16_76 %c_f16_77 %c_f16_78 %c_f16_79\n"
11791                 "    %fld9_2 = OpCompositeConstruct %v4f16 %c_f16_80 %c_f16_81 %c_f16_82 %c_f16_83\n"
11792                 "      %fld9 = OpCompositeConstruct %v4f16arr3 %fld9_0 %fld9_1 %fld9_2\n"
11793
11794                 "    %st_val = OpCompositeConstruct %st_test %c_f16_0 %fld1 %fld2 %fld3 %fld4 %fld5 %fld6 %c_f16_50 %fld8 %fld9\n"
11795                 "       %dst = OpAccessChain %up_st %ssbo_dst %c_i32_0 %ndx\n"
11796                 "              OpStore %dst %st_val\n"
11797
11798                 "              OpBranch %next\n"
11799
11800                 "      %next = OpLabel\n"
11801                 "     %i_cur = OpLoad %i32 %i\n"
11802                 "     %i_new = OpIAdd %i32 %i_cur %c_i32_1\n"
11803                 "              OpStore %i %i_new\n"
11804                 "              OpBranch %loop\n"
11805
11806                 "     %merge = OpLabel\n"
11807                 "              OpBranch %end_if\n"
11808                 "    %end_if = OpLabel\n"
11809                 "              OpReturnValue %param\n"
11810                 "              OpFunctionEnd\n"
11811         );
11812
11813         {
11814                 SpecResource            specResource;
11815                 map<string, string>     specs;
11816                 VulkanFeatures          features;
11817                 map<string, string>     fragments;
11818                 vector<string>          extensions;
11819                 vector<deFloat16>       expectedOutput;
11820                 string                          consts;
11821
11822                 for (deUint32 elementNdx = 0; elementNdx < numElements; ++elementNdx)
11823                 {
11824                         vector<deFloat16>       expectedIterationOutput;
11825
11826                         for (deUint32 structItemNdx = 0; structItemNdx < structItemsCount; ++structItemNdx)
11827                                 expectedIterationOutput.push_back(tcu::Float16(float(structItemNdx)).bits());
11828
11829                         for (deUint32 structItemNdx = 0; structItemNdx < DE_LENGTH_OF_ARRAY(exceptionIndices); ++structItemNdx)
11830                                 expectedIterationOutput[exceptionIndices[structItemNdx]] = exceptionValue;
11831
11832                         expectedIterationOutput[fieldModifiedMulIndex] = tcu::Float16(float(elementNdx * fieldModifier)).bits();
11833                         expectedIterationOutput[fieldModifiedAddIndex] = tcu::Float16(float(elementNdx + fieldModifier)).bits();
11834
11835                         expectedOutput.insert(expectedOutput.end(), expectedIterationOutput.begin(), expectedIterationOutput.end());
11836                 }
11837
11838                 for (deUint32 i = 0; i < structItemsCount; ++i)
11839                         consts += "     %c_f16_" + de::toString(i) + " = OpConstant %f16 "  + de::toString(i) + "\n";
11840
11841                 specs["num_elements"]           = de::toString(numElements);
11842                 specs["struct_item_size"]       = de::toString(structItemsCount * sizeof(deFloat16));
11843                 specs["field_modifier"]         = de::toString(fieldModifier);
11844                 specs["consts"]                         = consts;
11845
11846                 fragments["extension"]          = "OpExtension \"SPV_KHR_16bit_storage\"";
11847                 fragments["capability"]         = "OpCapability StorageUniformBufferBlock16\n";
11848                 fragments["decoration"]         = decoration.specialize(specs);
11849                 fragments["pre_main"]           = preMain.specialize(specs);
11850                 fragments["testfun"]            = testFun.specialize(specs);
11851
11852                 specResource.inputs.push_back(Resource(BufferSp(new Float16Buffer(expectedOutput)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
11853                 specResource.outputs.push_back(Resource(BufferSp(new Float16Buffer(expectedOutput)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
11854                 specResource.verifyIO = compareFP16CompositeFunc;
11855
11856                 extensions.push_back("VK_KHR_16bit_storage");
11857                 extensions.push_back("VK_KHR_shader_float16_int8");
11858
11859                 features.extFloat16Int8         = EXTFLOAT16INT8FEATURES_FLOAT16;
11860                 features.ext16BitStorage        = EXT16BITSTORAGEFEATURES_UNIFORM_BUFFER_BLOCK;
11861
11862                 finalizeTestsCreation(specResource, fragments, testCtx, *testGroup.get(), testName, features, extensions, IVec3(1, 1, 1));
11863         }
11864
11865         return testGroup.release();
11866 }
11867
11868 template<class SpecResource>
11869 tcu::TestCaseGroup* createFloat16CompositeInsertExtractSet (tcu::TestContext& testCtx, const char* op)
11870 {
11871         de::MovePtr<tcu::TestCaseGroup>         testGroup               (new tcu::TestCaseGroup(testCtx, de::toLower(op).c_str(), op));
11872         const deFloat16                                         exceptionValue  = tcu::Float16(-1.0).bits();
11873         const string                                            opName                  (op);
11874         const deUint32                                          opIndex                 = (opName == "OpCompositeInsert") ? 0
11875                                                                                                                 : (opName == "OpCompositeExtract") ? 1
11876                                                                                                                 : -1;
11877
11878         const StringTemplate preMain
11879         (
11880                 "   %c_i32_ndp = OpConstant %i32 ${num_elements}\n"
11881                 "         %f16 = OpTypeFloat 16\n"
11882                 "       %v2f16 = OpTypeVector %f16 2\n"
11883                 "       %v3f16 = OpTypeVector %f16 3\n"
11884                 "       %v4f16 = OpTypeVector %f16 4\n"
11885                 "    %c_f16_na = OpConstant %f16 -1.0\n"
11886                 "     %c_u32_5 = OpConstant %u32 5\n"
11887
11888                 "%f16arr3      = OpTypeArray %f16 %c_u32_3\n"
11889                 "%v2f16arr3    = OpTypeArray %v2f16 %c_u32_3\n"
11890                 "%v2f16arr5    = OpTypeArray %v2f16 %c_u32_5\n"
11891                 "%v3f16arr5    = OpTypeArray %v3f16 %c_u32_5\n"
11892                 "%v4f16arr3    = OpTypeArray %v4f16 %c_u32_3\n"
11893                 "%struct16     = OpTypeStruct %f16 %v2f16arr3\n"
11894                 "%struct16arr3 = OpTypeArray %struct16 %c_u32_3\n"
11895                 "%st_test      = OpTypeStruct %${field_type}\n"
11896
11897                 "      %up_f16 = OpTypePointer Uniform %f16\n"
11898                 "       %up_st = OpTypePointer Uniform %st_test\n"
11899                 "      %ra_f16 = OpTypeArray %f16 %c_i32_ndp\n"
11900                 "       %ra_st = OpTypeArray %st_test %c_i32_1\n"
11901
11902                 "${op_premain_decls}"
11903
11904                 " %up_SSBO_src = OpTypePointer Uniform %SSBO_src\n"
11905                 " %up_SSBO_dst = OpTypePointer Uniform %SSBO_dst\n"
11906
11907                 "    %ssbo_src = OpVariable %up_SSBO_src Uniform\n"
11908                 "    %ssbo_dst = OpVariable %up_SSBO_dst Uniform\n"
11909         );
11910
11911         const StringTemplate decoration
11912         (
11913                 "OpDecorate %SSBO_src BufferBlock\n"
11914                 "OpDecorate %SSBO_dst BufferBlock\n"
11915                 "OpDecorate %ra_f16 ArrayStride 2\n"
11916                 "OpDecorate %ra_st ArrayStride ${struct_item_size}\n"
11917                 "OpDecorate %ssbo_src DescriptorSet 0\n"
11918                 "OpDecorate %ssbo_src Binding 0\n"
11919                 "OpDecorate %ssbo_dst DescriptorSet 0\n"
11920                 "OpDecorate %ssbo_dst Binding 1\n"
11921
11922                 "OpMemberDecorate %SSBO_src 0 Offset 0\n"
11923                 "OpMemberDecorate %SSBO_dst 0 Offset 0\n"
11924
11925                 "OpDecorate %v2f16arr3 ArrayStride 4\n"
11926                 "OpMemberDecorate %struct16 0 Offset 0\n"
11927                 "OpMemberDecorate %struct16 1 Offset 4\n"
11928                 "OpDecorate %struct16arr3 ArrayStride 16\n"
11929                 "OpDecorate %f16arr3 ArrayStride 2\n"
11930                 "OpDecorate %v2f16arr5 ArrayStride 4\n"
11931                 "OpDecorate %v3f16arr5 ArrayStride 8\n"
11932                 "OpDecorate %v4f16arr3 ArrayStride 8\n"
11933
11934                 "OpMemberDecorate %st_test 0 Offset 0\n"
11935         );
11936
11937         const StringTemplate testFun
11938         (
11939                 " %test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
11940                 "     %param = OpFunctionParameter %v4f32\n"
11941                 "     %entry = OpLabel\n"
11942
11943                 "         %i = OpVariable %fp_i32 Function\n"
11944                 "              OpStore %i %c_i32_0\n"
11945
11946                 "  %will_run = OpFunctionCall %bool %isUniqueIdZero\n"
11947                 "              OpSelectionMerge %end_if None\n"
11948                 "              OpBranchConditional %will_run %run_test %end_if\n"
11949
11950                 "  %run_test = OpLabel\n"
11951                 "              OpBranch %loop\n"
11952
11953                 "      %loop = OpLabel\n"
11954                 "     %i_cmp = OpLoad %i32 %i\n"
11955                 "        %lt = OpSLessThan %bool %i_cmp %c_i32_ndp\n"
11956                 "              OpLoopMerge %merge %next None\n"
11957                 "              OpBranchConditional %lt %write %merge\n"
11958
11959                 "     %write = OpLabel\n"
11960                 "       %ndx = OpLoad %i32 %i\n"
11961
11962                 "${op_sw_fun_call}"
11963
11964                 "              OpStore %dst %val_dst\n"
11965                 "              OpBranch %next\n"
11966
11967                 "      %next = OpLabel\n"
11968                 "     %i_cur = OpLoad %i32 %i\n"
11969                 "     %i_new = OpIAdd %i32 %i_cur %c_i32_1\n"
11970                 "              OpStore %i %i_new\n"
11971                 "              OpBranch %loop\n"
11972
11973                 "     %merge = OpLabel\n"
11974                 "              OpBranch %end_if\n"
11975                 "    %end_if = OpLabel\n"
11976                 "              OpReturnValue %param\n"
11977                 "              OpFunctionEnd\n"
11978
11979                 "${op_sw_fun_header}"
11980                 " %sw_param = OpFunctionParameter %st_test\n"
11981                 "%sw_paramn = OpFunctionParameter %i32\n"
11982                 " %sw_entry = OpLabel\n"
11983                 "             OpSelectionMerge %switch_e None\n"
11984                 "             OpSwitch %sw_paramn %default ${case_list}\n"
11985
11986                 "${case_bodies}"
11987
11988                 "%default   = OpLabel\n"
11989                 "             OpReturnValue ${op_case_default_value}\n"
11990                 "%switch_e  = OpLabel\n"
11991                 "             OpUnreachable\n" // Unreachable merge block for switch statement
11992                 "             OpFunctionEnd\n"
11993         );
11994
11995         const StringTemplate testCaseBody
11996         (
11997                 "%case_${case_ndx}    = OpLabel\n"
11998                 "%val_ret_${case_ndx} = ${op_name} ${op_args_part} ${access_path}\n"
11999                 "             OpReturnValue %val_ret_${case_ndx}\n"
12000         );
12001
12002         struct OpParts
12003         {
12004                 const char*     premainDecls;
12005                 const char*     swFunCall;
12006                 const char*     swFunHeader;
12007                 const char*     caseDefaultValue;
12008                 const char*     argsPartial;
12009         };
12010
12011         OpParts                                                         opPartsArray[]                  =
12012         {
12013                 // OpCompositeInsert
12014                 {
12015                         "       %fun_t = OpTypeFunction %st_test %f16 %st_test %i32\n"
12016                         "    %SSBO_src = OpTypeStruct %ra_f16\n"
12017                         "    %SSBO_dst = OpTypeStruct %ra_st\n",
12018
12019                         "       %src = OpAccessChain %up_f16 %ssbo_src %c_i32_0 %ndx\n"
12020                         "       %dst = OpAccessChain %up_st %ssbo_dst %c_i32_0 %c_i32_0\n"
12021                         "   %val_new = OpLoad %f16 %src\n"
12022                         "   %val_old = OpLoad %st_test %dst\n"
12023                         "   %val_dst = OpFunctionCall %st_test %sw_fun %val_new %val_old %ndx\n",
12024
12025                         "   %sw_fun = OpFunction %st_test None %fun_t\n"
12026                         "%sw_paramv = OpFunctionParameter %f16\n",
12027
12028                         "%sw_param",
12029
12030                         "%st_test %sw_paramv %sw_param",
12031                 },
12032                 // OpCompositeExtract
12033                 {
12034                         "       %fun_t = OpTypeFunction %f16 %st_test %i32\n"
12035                         "    %SSBO_src = OpTypeStruct %ra_st\n"
12036                         "    %SSBO_dst = OpTypeStruct %ra_f16\n",
12037
12038                         "       %src = OpAccessChain %up_st %ssbo_src %c_i32_0 %c_i32_0\n"
12039                         "       %dst = OpAccessChain %up_f16 %ssbo_dst %c_i32_0 %ndx\n"
12040                         "   %val_src = OpLoad %st_test %src\n"
12041                         "   %val_dst = OpFunctionCall %f16 %sw_fun %val_src %ndx\n",
12042
12043                         "   %sw_fun = OpFunction %f16 None %fun_t\n",
12044
12045                         "%c_f16_na",
12046
12047                         "%f16 %sw_param",
12048                 },
12049         };
12050
12051         DE_ASSERT(opIndex >= 0 && opIndex < DE_LENGTH_OF_ARRAY(opPartsArray));
12052
12053         const char*     accessPathF16[] =
12054         {
12055                 "0",                    // %f16
12056                 DE_NULL,
12057         };
12058         const char*     accessPathV2F16[] =
12059         {
12060                 "0 0",                  // %v2f16
12061                 "0 1",
12062         };
12063         const char*     accessPathV3F16[] =
12064         {
12065                 "0 0",                  // %v3f16
12066                 "0 1",
12067                 "0 2",
12068                 DE_NULL,
12069         };
12070         const char*     accessPathV4F16[] =
12071         {
12072                 "0 0",                  // %v4f16"
12073                 "0 1",
12074                 "0 2",
12075                 "0 3",
12076         };
12077         const char*     accessPathF16Arr3[] =
12078         {
12079                 "0 0",                  // %f16arr3
12080                 "0 1",
12081                 "0 2",
12082                 DE_NULL,
12083         };
12084         const char*     accessPathStruct16Arr3[] =
12085         {
12086                 "0 0 0",                // %struct16arr3
12087                 DE_NULL,
12088                 "0 0 1 0 0",
12089                 "0 0 1 0 1",
12090                 "0 0 1 1 0",
12091                 "0 0 1 1 1",
12092                 "0 0 1 2 0",
12093                 "0 0 1 2 1",
12094                 "0 1 0",
12095                 DE_NULL,
12096                 "0 1 1 0 0",
12097                 "0 1 1 0 1",
12098                 "0 1 1 1 0",
12099                 "0 1 1 1 1",
12100                 "0 1 1 2 0",
12101                 "0 1 1 2 1",
12102                 "0 2 0",
12103                 DE_NULL,
12104                 "0 2 1 0 0",
12105                 "0 2 1 0 1",
12106                 "0 2 1 1 0",
12107                 "0 2 1 1 1",
12108                 "0 2 1 2 0",
12109                 "0 2 1 2 1",
12110         };
12111         const char*     accessPathV2F16Arr5[] =
12112         {
12113                 "0 0 0",                // %v2f16arr5
12114                 "0 0 1",
12115                 "0 1 0",
12116                 "0 1 1",
12117                 "0 2 0",
12118                 "0 2 1",
12119                 "0 3 0",
12120                 "0 3 1",
12121                 "0 4 0",
12122                 "0 4 1",
12123         };
12124         const char*     accessPathV3F16Arr5[] =
12125         {
12126                 "0 0 0",                // %v3f16arr5
12127                 "0 0 1",
12128                 "0 0 2",
12129                 DE_NULL,
12130                 "0 1 0",
12131                 "0 1 1",
12132                 "0 1 2",
12133                 DE_NULL,
12134                 "0 2 0",
12135                 "0 2 1",
12136                 "0 2 2",
12137                 DE_NULL,
12138                 "0 3 0",
12139                 "0 3 1",
12140                 "0 3 2",
12141                 DE_NULL,
12142                 "0 4 0",
12143                 "0 4 1",
12144                 "0 4 2",
12145                 DE_NULL,
12146         };
12147         const char*     accessPathV4F16Arr3[] =
12148         {
12149                 "0 0 0",                // %v4f16arr3
12150                 "0 0 1",
12151                 "0 0 2",
12152                 "0 0 3",
12153                 "0 1 0",
12154                 "0 1 1",
12155                 "0 1 2",
12156                 "0 1 3",
12157                 "0 2 0",
12158                 "0 2 1",
12159                 "0 2 2",
12160                 "0 2 3",
12161                 DE_NULL,
12162                 DE_NULL,
12163                 DE_NULL,
12164                 DE_NULL,
12165         };
12166
12167         struct TypeTestParameters
12168         {
12169                 const char*             name;
12170                 size_t                  accessPathLength;
12171                 const char**    accessPath;
12172         };
12173
12174         const TypeTestParameters typeTestParameters[] =
12175         {
12176                 {       "f16",                  DE_LENGTH_OF_ARRAY(accessPathF16),                      accessPathF16                   },
12177                 {       "v2f16",                DE_LENGTH_OF_ARRAY(accessPathV2F16),            accessPathV2F16                 },
12178                 {       "v3f16",                DE_LENGTH_OF_ARRAY(accessPathV3F16),            accessPathV3F16                 },
12179                 {       "v4f16",                DE_LENGTH_OF_ARRAY(accessPathV4F16),            accessPathV4F16                 },
12180                 {       "f16arr3",              DE_LENGTH_OF_ARRAY(accessPathF16Arr3),          accessPathF16Arr3               },
12181                 {       "v2f16arr5",    DE_LENGTH_OF_ARRAY(accessPathV2F16Arr5),        accessPathV2F16Arr5             },
12182                 {       "v3f16arr5",    DE_LENGTH_OF_ARRAY(accessPathV3F16Arr5),        accessPathV3F16Arr5             },
12183                 {       "v4f16arr3",    DE_LENGTH_OF_ARRAY(accessPathV4F16Arr3),        accessPathV4F16Arr3             },
12184                 {       "struct16arr3", DE_LENGTH_OF_ARRAY(accessPathStruct16Arr3),     accessPathStruct16Arr3  },
12185         };
12186
12187         for (size_t typeTestNdx = 0; typeTestNdx < DE_LENGTH_OF_ARRAY(typeTestParameters); ++typeTestNdx)
12188         {
12189                 const OpParts           opParts                         = opPartsArray[opIndex];
12190                 const string            testName                        = typeTestParameters[typeTestNdx].name;
12191                 const size_t            structItemsCount        = typeTestParameters[typeTestNdx].accessPathLength;
12192                 const char**            accessPath                      = typeTestParameters[typeTestNdx].accessPath;
12193                 SpecResource            specResource;
12194                 map<string, string>     specs;
12195                 VulkanFeatures          features;
12196                 map<string, string>     fragments;
12197                 vector<string>          extensions;
12198                 vector<deFloat16>       inputFP16;
12199                 vector<deFloat16>       dummyFP16Output;
12200
12201                 // Generate values for input
12202                 inputFP16.reserve(structItemsCount);
12203                 for (deUint32 structItemNdx = 0; structItemNdx < structItemsCount; ++structItemNdx)
12204                         inputFP16.push_back((accessPath[structItemNdx] == DE_NULL) ? exceptionValue : tcu::Float16(float(structItemNdx)).bits());
12205
12206                 dummyFP16Output.resize(structItemsCount);
12207
12208                 // Generate cases for OpSwitch
12209                 {
12210                         string  caseBodies;
12211                         string  caseList;
12212
12213                         for (deUint32 caseNdx = 0; caseNdx < structItemsCount; ++caseNdx)
12214                                 if (accessPath[caseNdx] != DE_NULL)
12215                                 {
12216                                         map<string, string>     specCase;
12217
12218                                         specCase["case_ndx"]            = de::toString(caseNdx);
12219                                         specCase["access_path"]         = accessPath[caseNdx];
12220                                         specCase["op_args_part"]        = opParts.argsPartial;
12221                                         specCase["op_name"]                     = opName;
12222
12223                                         caseBodies      += testCaseBody.specialize(specCase);
12224                                         caseList        += de::toString(caseNdx) + " %case_" + de::toString(caseNdx) + " ";
12225                                 }
12226
12227                         specs["case_bodies"]    = caseBodies;
12228                         specs["case_list"]              = caseList;
12229                 }
12230
12231                 specs["num_elements"]                   = de::toString(structItemsCount);
12232                 specs["field_type"]                             = typeTestParameters[typeTestNdx].name;
12233                 specs["struct_item_size"]               = de::toString(structItemsCount * sizeof(deFloat16));
12234                 specs["op_premain_decls"]               = opParts.premainDecls;
12235                 specs["op_sw_fun_call"]                 = opParts.swFunCall;
12236                 specs["op_sw_fun_header"]               = opParts.swFunHeader;
12237                 specs["op_case_default_value"]  = opParts.caseDefaultValue;
12238
12239                 fragments["extension"]          = "OpExtension \"SPV_KHR_16bit_storage\"";
12240                 fragments["capability"]         = "OpCapability StorageUniformBufferBlock16\n";
12241                 fragments["decoration"]         = decoration.specialize(specs);
12242                 fragments["pre_main"]           = preMain.specialize(specs);
12243                 fragments["testfun"]            = testFun.specialize(specs);
12244
12245                 specResource.inputs.push_back(Resource(BufferSp(new Float16Buffer(inputFP16)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
12246                 specResource.outputs.push_back(Resource(BufferSp(new Float16Buffer(dummyFP16Output)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
12247                 specResource.verifyIO = compareFP16CompositeFunc;
12248
12249                 extensions.push_back("VK_KHR_16bit_storage");
12250                 extensions.push_back("VK_KHR_shader_float16_int8");
12251
12252                 features.extFloat16Int8         = EXTFLOAT16INT8FEATURES_FLOAT16;
12253                 features.ext16BitStorage        = EXT16BITSTORAGEFEATURES_UNIFORM_BUFFER_BLOCK;
12254
12255                 finalizeTestsCreation(specResource, fragments, testCtx, *testGroup.get(), testName, features, extensions, IVec3(1, 1, 1));
12256         }
12257
12258         return testGroup.release();
12259 }
12260
12261 struct fp16PerComponent
12262 {
12263         fp16PerComponent()
12264                 : flavor(0)
12265                 , floatFormat16 (-14, 15, 10, true)
12266                 , outCompCount(0)
12267                 , argCompCount(3, 0)
12268         {
12269         }
12270
12271         bool                    callOncePerComponent    ()                                                                      { return true; }
12272         deUint32                getComponentValidity    ()                                                                      { return static_cast<deUint32>(-1); }
12273
12274         virtual double  getULPs                                 (vector<const deFloat16*>&)                     { return 1.0; }
12275         virtual double  getMin                                  (double value, double ulps)                     { return value - floatFormat16.ulp(deAbs(value), ulps); }
12276         virtual double  getMax                                  (double value, double ulps)                     { return value + floatFormat16.ulp(deAbs(value), ulps); }
12277
12278         virtual size_t  getFlavorCount                  ()                                                                      { return flavorNames.empty() ? 1 : flavorNames.size(); }
12279         virtual void    setFlavor                               (size_t flavorNo)                                       { DE_ASSERT(flavorNo < getFlavorCount()); flavor = flavorNo; }
12280         virtual size_t  getFlavor                               ()                                                                      { return flavor; }
12281         virtual string  getCurrentFlavorName    ()                                                                      { return flavorNames.empty() ? string("") : flavorNames[getFlavor()]; }
12282
12283         virtual void    setOutCompCount                 (size_t compCount)                                      { outCompCount = compCount; }
12284         virtual size_t  getOutCompCount                 ()                                                                      { return outCompCount; }
12285
12286         virtual void    setArgCompCount                 (size_t argNo, size_t compCount)        { argCompCount[argNo] = compCount; }
12287         virtual size_t  getArgCompCount                 (size_t argNo)                                          { return argCompCount[argNo]; }
12288
12289 protected:
12290         size_t                          flavor;
12291         tcu::FloatFormat        floatFormat16;
12292         size_t                          outCompCount;
12293         vector<size_t>          argCompCount;
12294         vector<string>          flavorNames;
12295 };
12296
12297 struct fp16OpFNegate : public fp16PerComponent
12298 {
12299         template <class fp16type>
12300         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
12301         {
12302                 const fp16type  x               (*in[0]);
12303                 const double    d               (x.asDouble());
12304                 const double    result  (0.0 - d);
12305
12306                 out[0] = fp16type(result).bits();
12307                 min[0] = getMin(result, getULPs(in));
12308                 max[0] = getMax(result, getULPs(in));
12309
12310                 return true;
12311         }
12312 };
12313
12314 struct fp16Round : public fp16PerComponent
12315 {
12316         fp16Round() : fp16PerComponent()
12317         {
12318                 flavorNames.push_back("Floor(x+0.5)");
12319                 flavorNames.push_back("Floor(x-0.5)");
12320                 flavorNames.push_back("RoundEven");
12321         }
12322
12323         template<class fp16type>
12324         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
12325         {
12326                 const fp16type  x               (*in[0]);
12327                 const double    d               (x.asDouble());
12328                 double                  result  (0.0);
12329
12330                 switch (flavor)
12331                 {
12332                         case 0:         result = deRound(d);            break;
12333                         case 1:         result = deFloor(d - 0.5);      break;
12334                         case 2:         result = deRoundEven(d);        break;
12335                         default:        TCU_THROW(InternalError, "Invalid flavor specified");
12336                 }
12337
12338                 out[0] = fp16type(result).bits();
12339                 min[0] = getMin(result, getULPs(in));
12340                 max[0] = getMax(result, getULPs(in));
12341
12342                 return true;
12343         }
12344 };
12345
12346 struct fp16RoundEven : public fp16PerComponent
12347 {
12348         template<class fp16type>
12349         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
12350         {
12351                 const fp16type  x               (*in[0]);
12352                 const double    d               (x.asDouble());
12353                 const double    result  (deRoundEven(d));
12354
12355                 out[0] = fp16type(result).bits();
12356                 min[0] = getMin(result, getULPs(in));
12357                 max[0] = getMax(result, getULPs(in));
12358
12359                 return true;
12360         }
12361 };
12362
12363 struct fp16Trunc : public fp16PerComponent
12364 {
12365         template<class fp16type>
12366         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
12367         {
12368                 const fp16type  x               (*in[0]);
12369                 const double    d               (x.asDouble());
12370                 const double    result  (deTrunc(d));
12371
12372                 out[0] = fp16type(result).bits();
12373                 min[0] = getMin(result, getULPs(in));
12374                 max[0] = getMax(result, getULPs(in));
12375
12376                 return true;
12377         }
12378 };
12379
12380 struct fp16FAbs : public fp16PerComponent
12381 {
12382         template<class fp16type>
12383         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
12384         {
12385                 const fp16type  x               (*in[0]);
12386                 const double    d               (x.asDouble());
12387                 const double    result  (deAbs(d));
12388
12389                 out[0] = fp16type(result).bits();
12390                 min[0] = getMin(result, getULPs(in));
12391                 max[0] = getMax(result, getULPs(in));
12392
12393                 return true;
12394         }
12395 };
12396
12397 struct fp16FSign : public fp16PerComponent
12398 {
12399         template<class fp16type>
12400         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
12401         {
12402                 const fp16type  x               (*in[0]);
12403                 const double    d               (x.asDouble());
12404                 const double    result  (deSign(d));
12405
12406                 if (x.isNaN())
12407                         return false;
12408
12409                 out[0] = fp16type(result).bits();
12410                 min[0] = getMin(result, getULPs(in));
12411                 max[0] = getMax(result, getULPs(in));
12412
12413                 return true;
12414         }
12415 };
12416
12417 struct fp16Floor : public fp16PerComponent
12418 {
12419         template<class fp16type>
12420         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
12421         {
12422                 const fp16type  x               (*in[0]);
12423                 const double    d               (x.asDouble());
12424                 const double    result  (deFloor(d));
12425
12426                 out[0] = fp16type(result).bits();
12427                 min[0] = getMin(result, getULPs(in));
12428                 max[0] = getMax(result, getULPs(in));
12429
12430                 return true;
12431         }
12432 };
12433
12434 struct fp16Ceil : public fp16PerComponent
12435 {
12436         template<class fp16type>
12437         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
12438         {
12439                 const fp16type  x               (*in[0]);
12440                 const double    d               (x.asDouble());
12441                 const double    result  (deCeil(d));
12442
12443                 out[0] = fp16type(result).bits();
12444                 min[0] = getMin(result, getULPs(in));
12445                 max[0] = getMax(result, getULPs(in));
12446
12447                 return true;
12448         }
12449 };
12450
12451 struct fp16Fract : public fp16PerComponent
12452 {
12453         template<class fp16type>
12454         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
12455         {
12456                 const fp16type  x               (*in[0]);
12457                 const double    d               (x.asDouble());
12458                 const double    result  (deFrac(d));
12459
12460                 out[0] = fp16type(result).bits();
12461                 min[0] = getMin(result, getULPs(in));
12462                 max[0] = getMax(result, getULPs(in));
12463
12464                 return true;
12465         }
12466 };
12467
12468 struct fp16Radians : public fp16PerComponent
12469 {
12470         virtual double getULPs (vector<const deFloat16*>& in)
12471         {
12472                 DE_UNREF(in);
12473
12474                 return 2.5;
12475         }
12476
12477         template<class fp16type>
12478         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
12479         {
12480                 const fp16type  x               (*in[0]);
12481                 const float             d               (x.asFloat());
12482                 const float             result  (deFloatRadians(d));
12483
12484                 out[0] = fp16type(result).bits();
12485                 min[0] = getMin(result, getULPs(in));
12486                 max[0] = getMax(result, getULPs(in));
12487
12488                 return true;
12489         }
12490 };
12491
12492 struct fp16Degrees : public fp16PerComponent
12493 {
12494         virtual double getULPs (vector<const deFloat16*>& in)
12495         {
12496                 DE_UNREF(in);
12497
12498                 return 2.5;
12499         }
12500
12501         template<class fp16type>
12502         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
12503         {
12504                 const fp16type  x               (*in[0]);
12505                 const float             d               (x.asFloat());
12506                 const float             result  (deFloatDegrees(d));
12507
12508                 out[0] = fp16type(result).bits();
12509                 min[0] = getMin(result, getULPs(in));
12510                 max[0] = getMax(result, getULPs(in));
12511
12512                 return true;
12513         }
12514 };
12515
12516 struct fp16Sin : public fp16PerComponent
12517 {
12518         template<class fp16type>
12519         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
12520         {
12521                 const fp16type  x                       (*in[0]);
12522                 const double    d                       (x.asDouble());
12523                 const double    result          (deSin(d));
12524                 const double    unspecUlp       (16.0);
12525                 const double    err                     (de::inRange(d, -DE_PI_DOUBLE, DE_PI_DOUBLE) ? deLdExp(1.0, -7) : floatFormat16.ulp(deAbs(result), unspecUlp));
12526
12527                 if (!de::inRange(d, -DE_PI_DOUBLE, DE_PI_DOUBLE))
12528                         return false;
12529
12530                 out[0] = fp16type(result).bits();
12531                 min[0] = result - err;
12532                 max[0] = result + err;
12533
12534                 return true;
12535         }
12536 };
12537
12538 struct fp16Cos : public fp16PerComponent
12539 {
12540         template<class fp16type>
12541         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
12542         {
12543                 const fp16type  x                       (*in[0]);
12544                 const double    d                       (x.asDouble());
12545                 const double    result          (deCos(d));
12546                 const double    unspecUlp       (16.0);
12547                 const double    err                     (de::inRange(d, -DE_PI_DOUBLE, DE_PI_DOUBLE) ? deLdExp(1.0, -7) : floatFormat16.ulp(deAbs(result), unspecUlp));
12548
12549                 if (!de::inRange(d, -DE_PI_DOUBLE, DE_PI_DOUBLE))
12550                         return false;
12551
12552                 out[0] = fp16type(result).bits();
12553                 min[0] = result - err;
12554                 max[0] = result + err;
12555
12556                 return true;
12557         }
12558 };
12559
12560 struct fp16Tan : public fp16PerComponent
12561 {
12562         template<class fp16type>
12563         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
12564         {
12565                 const fp16type  x               (*in[0]);
12566                 const double    d               (x.asDouble());
12567                 const double    result  (deTan(d));
12568
12569                 if (!de::inRange(d, -DE_PI_DOUBLE, DE_PI_DOUBLE))
12570                         return false;
12571
12572                 out[0] = fp16type(result).bits();
12573                 {
12574                         const double    err                     = deLdExp(1.0, -7);
12575                         const double    s1                      = deSin(d) + err;
12576                         const double    s2                      = deSin(d) - err;
12577                         const double    c1                      = deCos(d) + err;
12578                         const double    c2                      = deCos(d) - err;
12579                         const double    edgeVals[]      = {s1/c1, s1/c2, s2/c1, s2/c2};
12580                         double                  edgeLeft        = out[0];
12581                         double                  edgeRight       = out[0];
12582
12583                         if (deSign(c1 * c2) < 0.0)
12584                         {
12585                                 edgeLeft        = -std::numeric_limits<double>::infinity();
12586                                 edgeRight       = +std::numeric_limits<double>::infinity();
12587                         }
12588                         else
12589                         {
12590                                 edgeLeft        = *std::min_element(&edgeVals[0], &edgeVals[DE_LENGTH_OF_ARRAY(edgeVals)]);
12591                                 edgeRight       = *std::max_element(&edgeVals[0], &edgeVals[DE_LENGTH_OF_ARRAY(edgeVals)]);
12592                         }
12593
12594                         min[0] = edgeLeft;
12595                         max[0] = edgeRight;
12596                 }
12597
12598                 return true;
12599         }
12600 };
12601
12602 struct fp16Asin : public fp16PerComponent
12603 {
12604         template<class fp16type>
12605         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
12606         {
12607                 const fp16type  x               (*in[0]);
12608                 const double    d               (x.asDouble());
12609                 const double    result  (deAsin(d));
12610                 const double    error   (deAtan2(d, sqrt(1.0 - d * d)));
12611
12612                 if (!x.isNaN() && deAbs(d) > 1.0)
12613                         return false;
12614
12615                 out[0] = fp16type(result).bits();
12616                 min[0] = result - floatFormat16.ulp(deAbs(error), 2 * 5.0); // This is not a precision test. Value is not from spec
12617                 max[0] = result + floatFormat16.ulp(deAbs(error), 2 * 5.0); // This is not a precision test. Value is not from spec
12618
12619                 return true;
12620         }
12621 };
12622
12623 struct fp16Acos : public fp16PerComponent
12624 {
12625         template<class fp16type>
12626         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
12627         {
12628                 const fp16type  x               (*in[0]);
12629                 const double    d               (x.asDouble());
12630                 const double    result  (deAcos(d));
12631                 const double    error   (deAtan2(sqrt(1.0 - d * d), d));
12632
12633                 if (!x.isNaN() && deAbs(d) > 1.0)
12634                         return false;
12635
12636                 out[0] = fp16type(result).bits();
12637                 min[0] = result - floatFormat16.ulp(deAbs(error), 2 * 5.0); // This is not a precision test. Value is not from spec
12638                 max[0] = result + floatFormat16.ulp(deAbs(error), 2 * 5.0); // This is not a precision test. Value is not from spec
12639
12640                 return true;
12641         }
12642 };
12643
12644 struct fp16Atan : public fp16PerComponent
12645 {
12646         virtual double getULPs(vector<const deFloat16*>& in)
12647         {
12648                 DE_UNREF(in);
12649
12650                 return 2 * 5.0; // This is not a precision test. Value is not from spec
12651         }
12652
12653         template<class fp16type>
12654         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
12655         {
12656                 const fp16type  x               (*in[0]);
12657                 const double    d               (x.asDouble());
12658                 const double    result  (deAtanOver(d));
12659
12660                 out[0] = fp16type(result).bits();
12661                 min[0] = getMin(result, getULPs(in));
12662                 max[0] = getMax(result, getULPs(in));
12663
12664                 return true;
12665         }
12666 };
12667
12668 struct fp16Sinh : public fp16PerComponent
12669 {
12670         fp16Sinh() : fp16PerComponent()
12671         {
12672                 flavorNames.push_back("Double");
12673                 flavorNames.push_back("ExpFP16");
12674         }
12675
12676         template<class fp16type>
12677         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
12678         {
12679                 const fp16type  x               (*in[0]);
12680                 const double    d               (x.asDouble());
12681                 const double    ulps    (64 * (1.0 + 2 * deAbs(d))); // This is not a precision test. Value is not from spec
12682                 double                  result  (0.0);
12683                 double                  error   (0.0);
12684
12685                 if (getFlavor() == 0)
12686                 {
12687                         result  = deSinh(d);
12688                         error   = floatFormat16.ulp(deAbs(result), ulps);
12689                 }
12690                 else if (getFlavor() == 1)
12691                 {
12692                         const fp16type  epx     (deExp(d));
12693                         const fp16type  enx     (deExp(-d));
12694                         const fp16type  esx     (epx.asDouble() - enx.asDouble());
12695                         const fp16type  sx2     (esx.asDouble() / 2.0);
12696
12697                         result  = sx2.asDouble();
12698                         error   = deAbs(floatFormat16.ulp(epx.asDouble(), ulps)) + deAbs(floatFormat16.ulp(enx.asDouble(), ulps));
12699                 }
12700                 else
12701                 {
12702                         TCU_THROW(InternalError, "Unknown flavor");
12703                 }
12704
12705                 out[0] = fp16type(result).bits();
12706                 min[0] = result - error;
12707                 max[0] = result + error;
12708
12709                 return true;
12710         }
12711 };
12712
12713 struct fp16Cosh : public fp16PerComponent
12714 {
12715         fp16Cosh() : fp16PerComponent()
12716         {
12717                 flavorNames.push_back("Double");
12718                 flavorNames.push_back("ExpFP16");
12719         }
12720
12721         template<class fp16type>
12722         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
12723         {
12724                 const fp16type  x               (*in[0]);
12725                 const double    d               (x.asDouble());
12726                 const double    ulps    (64 * (1.0 + 2 * deAbs(d))); // This is not a precision test. Value is not from spec
12727                 double                  result  (0.0);
12728
12729                 if (getFlavor() == 0)
12730                 {
12731                         result = deCosh(d);
12732                 }
12733                 else if (getFlavor() == 1)
12734                 {
12735                         const fp16type  epx     (deExp(d));
12736                         const fp16type  enx     (deExp(-d));
12737                         const fp16type  esx     (epx.asDouble() + enx.asDouble());
12738                         const fp16type  sx2     (esx.asDouble() / 2.0);
12739
12740                         result = sx2.asDouble();
12741                 }
12742                 else
12743                 {
12744                         TCU_THROW(InternalError, "Unknown flavor");
12745                 }
12746
12747                 out[0] = fp16type(result).bits();
12748                 min[0] = result - floatFormat16.ulp(deAbs(result), ulps);
12749                 max[0] = result + floatFormat16.ulp(deAbs(result), ulps);
12750
12751                 return true;
12752         }
12753 };
12754
12755 struct fp16Tanh : public fp16PerComponent
12756 {
12757         fp16Tanh() : fp16PerComponent()
12758         {
12759                 flavorNames.push_back("Tanh");
12760                 flavorNames.push_back("SinhCosh");
12761                 flavorNames.push_back("SinhCoshFP16");
12762                 flavorNames.push_back("PolyFP16");
12763         }
12764
12765         virtual double getULPs (vector<const deFloat16*>& in)
12766         {
12767                 const tcu::Float16      x       (*in[0]);
12768                 const double            d       (x.asDouble());
12769
12770                 return 2 * (1.0 + 2 * deAbs(d)); // This is not a precision test. Value is not from spec
12771         }
12772
12773         template<class fp16type>
12774         inline double calcPoly (const fp16type& espx, const fp16type& esnx, const fp16type& ecpx, const fp16type& ecnx)
12775         {
12776                 const fp16type  esx     (espx.asDouble() - esnx.asDouble());
12777                 const fp16type  sx2     (esx.asDouble() / 2.0);
12778                 const fp16type  ecx     (ecpx.asDouble() + ecnx.asDouble());
12779                 const fp16type  cx2     (ecx.asDouble() / 2.0);
12780                 const fp16type  tg      (sx2.asDouble() / cx2.asDouble());
12781                 const double    rez     (tg.asDouble());
12782
12783                 return rez;
12784         }
12785
12786         template<class fp16type>
12787         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
12788         {
12789                 const fp16type  x               (*in[0]);
12790                 const double    d               (x.asDouble());
12791                 double                  result  (0.0);
12792
12793                 if (getFlavor() == 0)
12794                 {
12795                         result  = deTanh(d);
12796                         min[0]  = getMin(result, getULPs(in));
12797                         max[0]  = getMax(result, getULPs(in));
12798                 }
12799                 else if (getFlavor() == 1)
12800                 {
12801                         result  = deSinh(d) / deCosh(d);
12802                         min[0]  = getMin(result, getULPs(in));
12803                         max[0]  = getMax(result, getULPs(in));
12804                 }
12805                 else if (getFlavor() == 2)
12806                 {
12807                         const fp16type  s       (deSinh(d));
12808                         const fp16type  c       (deCosh(d));
12809
12810                         result  = s.asDouble() / c.asDouble();
12811                         min[0]  = getMin(result, getULPs(in));
12812                         max[0]  = getMax(result, getULPs(in));
12813                 }
12814                 else if (getFlavor() == 3)
12815                 {
12816                         const double    ulps    (getULPs(in));
12817                         const double    epxm    (deExp( d));
12818                         const double    enxm    (deExp(-d));
12819                         const double    epxmerr = floatFormat16.ulp(epxm, ulps);
12820                         const double    enxmerr = floatFormat16.ulp(enxm, ulps);
12821                         const fp16type  epx[]   = { fp16type(epxm - epxmerr), fp16type(epxm + epxmerr) };
12822                         const fp16type  enx[]   = { fp16type(enxm - enxmerr), fp16type(enxm + enxmerr) };
12823                         const fp16type  epxm16  (epxm);
12824                         const fp16type  enxm16  (enxm);
12825                         vector<double>  tgs;
12826
12827                         for (size_t spNdx = 0; spNdx < DE_LENGTH_OF_ARRAY(epx); ++spNdx)
12828                         for (size_t snNdx = 0; snNdx < DE_LENGTH_OF_ARRAY(enx); ++snNdx)
12829                         for (size_t cpNdx = 0; cpNdx < DE_LENGTH_OF_ARRAY(epx); ++cpNdx)
12830                         for (size_t cnNdx = 0; cnNdx < DE_LENGTH_OF_ARRAY(enx); ++cnNdx)
12831                         {
12832                                 const double tgh = calcPoly(epx[spNdx], enx[snNdx], epx[cpNdx], enx[cnNdx]);
12833
12834                                 tgs.push_back(tgh);
12835                         }
12836
12837                         result = calcPoly(epxm16, enxm16, epxm16, enxm16);
12838                         min[0] = *std::min_element(tgs.begin(), tgs.end());
12839                         max[0] = *std::max_element(tgs.begin(), tgs.end());
12840                 }
12841                 else
12842                 {
12843                         TCU_THROW(InternalError, "Unknown flavor");
12844                 }
12845
12846                 out[0] = fp16type(result).bits();
12847
12848                 return true;
12849         }
12850 };
12851
12852 struct fp16Asinh : public fp16PerComponent
12853 {
12854         fp16Asinh() : fp16PerComponent()
12855         {
12856                 flavorNames.push_back("Double");
12857                 flavorNames.push_back("PolyFP16Wiki");
12858                 flavorNames.push_back("PolyFP16Abs");
12859         }
12860
12861         virtual double getULPs (vector<const deFloat16*>& in)
12862         {
12863                 DE_UNREF(in);
12864
12865                 return 256.0; // This is not a precision test. Value is not from spec
12866         }
12867
12868         template<class fp16type>
12869         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
12870         {
12871                 const fp16type  x               (*in[0]);
12872                 const double    d               (x.asDouble());
12873                 double                  result  (0.0);
12874
12875                 if (getFlavor() == 0)
12876                 {
12877                         result = deAsinh(d);
12878                 }
12879                 else if (getFlavor() == 1)
12880                 {
12881                         const fp16type  x2              (d * d);
12882                         const fp16type  x2p1    (x2.asDouble() + 1.0);
12883                         const fp16type  sq              (deSqrt(x2p1.asDouble()));
12884                         const fp16type  sxsq    (d + sq.asDouble());
12885                         const fp16type  lsxsq   (deLog(sxsq.asDouble()));
12886
12887                         if (lsxsq.isInf())
12888                                 return false;
12889
12890                         result = lsxsq.asDouble();
12891                 }
12892                 else if (getFlavor() == 2)
12893                 {
12894                         const fp16type  x2              (d * d);
12895                         const fp16type  x2p1    (x2.asDouble() + 1.0);
12896                         const fp16type  sq              (deSqrt(x2p1.asDouble()));
12897                         const fp16type  sxsq    (deAbs(d) + sq.asDouble());
12898                         const fp16type  lsxsq   (deLog(sxsq.asDouble()));
12899
12900                         result = deSign(d) * lsxsq.asDouble();
12901                 }
12902                 else
12903                 {
12904                         TCU_THROW(InternalError, "Unknown flavor");
12905                 }
12906
12907                 out[0] = fp16type(result).bits();
12908                 min[0] = getMin(result, getULPs(in));
12909                 max[0] = getMax(result, getULPs(in));
12910
12911                 return true;
12912         }
12913 };
12914
12915 struct fp16Acosh : public fp16PerComponent
12916 {
12917         fp16Acosh() : fp16PerComponent()
12918         {
12919                 flavorNames.push_back("Double");
12920                 flavorNames.push_back("PolyFP16");
12921         }
12922
12923         virtual double getULPs (vector<const deFloat16*>& in)
12924         {
12925                 DE_UNREF(in);
12926
12927                 return 16.0; // This is not a precision test. Value is not from spec
12928         }
12929
12930         template<class fp16type>
12931         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
12932         {
12933                 const fp16type  x               (*in[0]);
12934                 const double    d               (x.asDouble());
12935                 double                  result  (0.0);
12936
12937                 if (!x.isNaN() && d < 1.0)
12938                         return false;
12939
12940                 if (getFlavor() == 0)
12941                 {
12942                         result = deAcosh(d);
12943                 }
12944                 else if (getFlavor() == 1)
12945                 {
12946                         const fp16type  x2              (d * d);
12947                         const fp16type  x2m1    (x2.asDouble() - 1.0);
12948                         const fp16type  sq              (deSqrt(x2m1.asDouble()));
12949                         const fp16type  sxsq    (d + sq.asDouble());
12950                         const fp16type  lsxsq   (deLog(sxsq.asDouble()));
12951
12952                         result = lsxsq.asDouble();
12953                 }
12954                 else
12955                 {
12956                         TCU_THROW(InternalError, "Unknown flavor");
12957                 }
12958
12959                 out[0] = fp16type(result).bits();
12960                 min[0] = getMin(result, getULPs(in));
12961                 max[0] = getMax(result, getULPs(in));
12962
12963                 return true;
12964         }
12965 };
12966
12967 struct fp16Atanh : public fp16PerComponent
12968 {
12969         fp16Atanh() : fp16PerComponent()
12970         {
12971                 flavorNames.push_back("Double");
12972                 flavorNames.push_back("PolyFP16");
12973         }
12974
12975         template<class fp16type>
12976         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
12977         {
12978                 const fp16type  x               (*in[0]);
12979                 const double    d               (x.asDouble());
12980                 double                  result  (0.0);
12981
12982                 if (deAbs(d) >= 1.0)
12983                         return false;
12984
12985                 if (getFlavor() == 0)
12986                 {
12987                         const double    ulps    (16.0); // This is not a precision test. Value is not from spec
12988
12989                         result = deAtanh(d);
12990                         min[0] = getMin(result, ulps);
12991                         max[0] = getMax(result, ulps);
12992                 }
12993                 else if (getFlavor() == 1)
12994                 {
12995                         const fp16type  x1a             (1.0 + d);
12996                         const fp16type  x1b             (1.0 - d);
12997                         const fp16type  x1d             (x1a.asDouble() / x1b.asDouble());
12998                         const fp16type  lx1d    (deLog(x1d.asDouble()));
12999                         const fp16type  lx1d2   (0.5 * lx1d.asDouble());
13000                         const double    error   (2 * (de::inRange(deAbs(x1d.asDouble()), 0.5, 2.0) ? deLdExp(2.0, -7) : floatFormat16.ulp(deAbs(x1d.asDouble()), 3.0)));
13001
13002                         result = lx1d2.asDouble();
13003                         min[0] = result - error;
13004                         max[0] = result + error;
13005                 }
13006                 else
13007                 {
13008                         TCU_THROW(InternalError, "Unknown flavor");
13009                 }
13010
13011                 out[0] = fp16type(result).bits();
13012
13013                 return true;
13014         }
13015 };
13016
13017 struct fp16Exp : public fp16PerComponent
13018 {
13019         template<class fp16type>
13020         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
13021         {
13022                 const fp16type  x               (*in[0]);
13023                 const double    d               (x.asDouble());
13024                 const double    ulps    (10.0 * (1.0 + 2.0 * deAbs(d)));
13025                 const double    result  (deExp(d));
13026
13027                 out[0] = fp16type(result).bits();
13028                 min[0] = getMin(result, ulps);
13029                 max[0] = getMax(result, ulps);
13030
13031                 return true;
13032         }
13033 };
13034
13035 struct fp16Log : public fp16PerComponent
13036 {
13037         template<class fp16type>
13038         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
13039         {
13040                 const fp16type  x               (*in[0]);
13041                 const double    d               (x.asDouble());
13042                 const double    result  (deLog(d));
13043                 const double    error   (de::inRange(deAbs(d), 0.5, 2.0) ? deLdExp(2.0, -7) : floatFormat16.ulp(deAbs(result), 3.0));
13044
13045                 if (d <= 0.0)
13046                         return false;
13047
13048                 out[0] = fp16type(result).bits();
13049                 min[0] = result - error;
13050                 max[0] = result + error;
13051
13052                 return true;
13053         }
13054 };
13055
13056 struct fp16Exp2 : public fp16PerComponent
13057 {
13058         template<class fp16type>
13059         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
13060         {
13061                 const fp16type  x               (*in[0]);
13062                 const double    d               (x.asDouble());
13063                 const double    result  (deExp2(d));
13064                 const double    ulps    (1.0 + 2.0 * deAbs(fp16type(in[0][0]).asDouble()));
13065
13066                 out[0] = fp16type(result).bits();
13067                 min[0] = getMin(result, ulps);
13068                 max[0] = getMax(result, ulps);
13069
13070                 return true;
13071         }
13072 };
13073
13074 struct fp16Log2 : public fp16PerComponent
13075 {
13076         template<class fp16type>
13077         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
13078         {
13079                 const fp16type  x               (*in[0]);
13080                 const double    d               (x.asDouble());
13081                 const double    result  (deLog2(d));
13082                 const double    error   (de::inRange(deAbs(d), 0.5, 2.0) ? deLdExp(2.0, -7) : floatFormat16.ulp(deAbs(result), 3.0));
13083
13084                 if (d <= 0.0)
13085                         return false;
13086
13087                 out[0] = fp16type(result).bits();
13088                 min[0] = result - error;
13089                 max[0] = result + error;
13090
13091                 return true;
13092         }
13093 };
13094
13095 struct fp16Sqrt : public fp16PerComponent
13096 {
13097         virtual double getULPs (vector<const deFloat16*>& in)
13098         {
13099                 DE_UNREF(in);
13100
13101                 return 6.0;
13102         }
13103
13104         template<class fp16type>
13105         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
13106         {
13107                 const fp16type  x               (*in[0]);
13108                 const double    d               (x.asDouble());
13109                 const double    result  (deSqrt(d));
13110
13111                 if (!x.isNaN() && d < 0.0)
13112                         return false;
13113
13114                 out[0] = fp16type(result).bits();
13115                 min[0] = getMin(result, getULPs(in));
13116                 max[0] = getMax(result, getULPs(in));
13117
13118                 return true;
13119         }
13120 };
13121
13122 struct fp16InverseSqrt : public fp16PerComponent
13123 {
13124         virtual double getULPs (vector<const deFloat16*>& in)
13125         {
13126                 DE_UNREF(in);
13127
13128                 return 2.0;
13129         }
13130
13131         template<class fp16type>
13132         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
13133         {
13134                 const fp16type  x               (*in[0]);
13135                 const double    d               (x.asDouble());
13136                 const double    result  (1.0/deSqrt(d));
13137
13138                 if (!x.isNaN() && d <= 0.0)
13139                         return false;
13140
13141                 out[0] = fp16type(result).bits();
13142                 min[0] = getMin(result, getULPs(in));
13143                 max[0] = getMax(result, getULPs(in));
13144
13145                 return true;
13146         }
13147 };
13148
13149 struct fp16ModfFrac : public fp16PerComponent
13150 {
13151         template<class fp16type>
13152         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
13153         {
13154                 const fp16type  x               (*in[0]);
13155                 const double    d               (x.asDouble());
13156                 double                  i               (0.0);
13157                 const double    result  (deModf(d, &i));
13158
13159                 if (x.isInf() || x.isNaN())
13160                         return false;
13161
13162                 out[0] = fp16type(result).bits();
13163                 min[0] = getMin(result, getULPs(in));
13164                 max[0] = getMax(result, getULPs(in));
13165
13166                 return true;
13167         }
13168 };
13169
13170 struct fp16ModfInt : public fp16PerComponent
13171 {
13172         template<class fp16type>
13173         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
13174         {
13175                 const fp16type  x               (*in[0]);
13176                 const double    d               (x.asDouble());
13177                 double                  i               (0.0);
13178                 const double    dummy   (deModf(d, &i));
13179                 const double    result  (i);
13180
13181                 DE_UNREF(dummy);
13182
13183                 if (x.isInf() || x.isNaN())
13184                         return false;
13185
13186                 out[0] = fp16type(result).bits();
13187                 min[0] = getMin(result, getULPs(in));
13188                 max[0] = getMax(result, getULPs(in));
13189
13190                 return true;
13191         }
13192 };
13193
13194 struct fp16FrexpS : public fp16PerComponent
13195 {
13196         template<class fp16type>
13197         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
13198         {
13199                 const fp16type  x               (*in[0]);
13200                 const double    d               (x.asDouble());
13201                 int                             e               (0);
13202                 const double    result  (deFrExp(d, &e));
13203
13204                 if (x.isNaN() || x.isInf())
13205                         return false;
13206
13207                 out[0] = fp16type(result).bits();
13208                 min[0] = getMin(result, getULPs(in));
13209                 max[0] = getMax(result, getULPs(in));
13210
13211                 return true;
13212         }
13213 };
13214
13215 struct fp16FrexpE : public fp16PerComponent
13216 {
13217         template<class fp16type>
13218         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
13219         {
13220                 const fp16type  x               (*in[0]);
13221                 const double    d               (x.asDouble());
13222                 int                             e               (0);
13223                 const double    dummy   (deFrExp(d, &e));
13224                 const double    result  (static_cast<double>(e));
13225
13226                 DE_UNREF(dummy);
13227
13228                 if (x.isNaN() || x.isInf())
13229                         return false;
13230
13231                 out[0] = fp16type(result).bits();
13232                 min[0] = getMin(result, getULPs(in));
13233                 max[0] = getMax(result, getULPs(in));
13234
13235                 return true;
13236         }
13237 };
13238
13239 struct fp16OpFAdd : public fp16PerComponent
13240 {
13241         template<class fp16type>
13242         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
13243         {
13244                 const fp16type  x               (*in[0]);
13245                 const fp16type  y               (*in[1]);
13246                 const double    xd              (x.asDouble());
13247                 const double    yd              (y.asDouble());
13248                 const double    result  (xd + yd);
13249
13250                 out[0] = fp16type(result).bits();
13251                 min[0] = getMin(result, getULPs(in));
13252                 max[0] = getMax(result, getULPs(in));
13253
13254                 return true;
13255         }
13256 };
13257
13258 struct fp16OpFSub : public fp16PerComponent
13259 {
13260         template<class fp16type>
13261         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
13262         {
13263                 const fp16type  x               (*in[0]);
13264                 const fp16type  y               (*in[1]);
13265                 const double    xd              (x.asDouble());
13266                 const double    yd              (y.asDouble());
13267                 const double    result  (xd - yd);
13268
13269                 out[0] = fp16type(result).bits();
13270                 min[0] = getMin(result, getULPs(in));
13271                 max[0] = getMax(result, getULPs(in));
13272
13273                 return true;
13274         }
13275 };
13276
13277 struct fp16OpFMul : public fp16PerComponent
13278 {
13279         template<class fp16type>
13280         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
13281         {
13282                 const fp16type  x               (*in[0]);
13283                 const fp16type  y               (*in[1]);
13284                 const double    xd              (x.asDouble());
13285                 const double    yd              (y.asDouble());
13286                 const double    result  (xd * yd);
13287
13288                 out[0] = fp16type(result).bits();
13289                 min[0] = getMin(result, getULPs(in));
13290                 max[0] = getMax(result, getULPs(in));
13291
13292                 return true;
13293         }
13294 };
13295
13296 struct fp16OpFDiv : public fp16PerComponent
13297 {
13298         fp16OpFDiv() : fp16PerComponent()
13299         {
13300                 flavorNames.push_back("DirectDiv");
13301                 flavorNames.push_back("InverseDiv");
13302         }
13303
13304         template<class fp16type>
13305         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
13306         {
13307                 const fp16type  x                       (*in[0]);
13308                 const fp16type  y                       (*in[1]);
13309                 const double    xd                      (x.asDouble());
13310                 const double    yd                      (y.asDouble());
13311                 const double    unspecUlp       (16.0);
13312                 const double    ulpCnt          (de::inRange(deAbs(yd), deLdExp(1, -14), deLdExp(1, 14)) ? 2.5 : unspecUlp);
13313                 double                  result          (0.0);
13314
13315                 if (y.isZero())
13316                         return false;
13317
13318                 if (getFlavor() == 0)
13319                 {
13320                         result = (xd / yd);
13321                 }
13322                 else if (getFlavor() == 1)
13323                 {
13324                         const double    invyd   (1.0 / yd);
13325                         const fp16type  invy    (invyd);
13326
13327                         result = (xd * invy.asDouble());
13328                 }
13329                 else
13330                 {
13331                         TCU_THROW(InternalError, "Unknown flavor");
13332                 }
13333
13334                 out[0] = fp16type(result).bits();
13335                 min[0] = getMin(result, ulpCnt);
13336                 max[0] = getMax(result, ulpCnt);
13337
13338                 return true;
13339         }
13340 };
13341
13342 struct fp16Atan2 : public fp16PerComponent
13343 {
13344         fp16Atan2() : fp16PerComponent()
13345         {
13346                 flavorNames.push_back("DoubleCalc");
13347                 flavorNames.push_back("DoubleCalc_PI");
13348         }
13349
13350         virtual double getULPs(vector<const deFloat16*>& in)
13351         {
13352                 DE_UNREF(in);
13353
13354                 return 2 * 5.0; // This is not a precision test. Value is not from spec
13355         }
13356
13357         template<class fp16type>
13358         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
13359         {
13360                 const fp16type  x               (*in[0]);
13361                 const fp16type  y               (*in[1]);
13362                 const double    xd              (x.asDouble());
13363                 const double    yd              (y.asDouble());
13364                 double                  result  (0.0);
13365
13366                 if (x.isZero() && y.isZero())
13367                         return false;
13368
13369                 if (getFlavor() == 0)
13370                 {
13371                         result  = deAtan2(xd, yd);
13372                 }
13373                 else if (getFlavor() == 1)
13374                 {
13375                         const double    ulps    (2.0 * 5.0); // This is not a precision test. Value is not from spec
13376                         const double    eps             (floatFormat16.ulp(DE_PI_DOUBLE, ulps));
13377
13378                         result  = deAtan2(xd, yd);
13379
13380                         if (de::inRange(deAbs(result), DE_PI_DOUBLE - eps, DE_PI_DOUBLE + eps))
13381                                 result  = -result;
13382                 }
13383                 else
13384                 {
13385                         TCU_THROW(InternalError, "Unknown flavor");
13386                 }
13387
13388                 out[0] = fp16type(result).bits();
13389                 min[0] = getMin(result, getULPs(in));
13390                 max[0] = getMax(result, getULPs(in));
13391
13392                 return true;
13393         }
13394 };
13395
13396 struct fp16Pow : public fp16PerComponent
13397 {
13398         fp16Pow() : fp16PerComponent()
13399         {
13400                 flavorNames.push_back("Pow");
13401                 flavorNames.push_back("PowLog2");
13402                 flavorNames.push_back("PowLog2FP16");
13403         }
13404
13405         template<class fp16type>
13406         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
13407         {
13408                 const fp16type  x               (*in[0]);
13409                 const fp16type  y               (*in[1]);
13410                 const double    xd              (x.asDouble());
13411                 const double    yd              (y.asDouble());
13412                 const double    logxeps (de::inRange(deAbs(xd), 0.5, 2.0) ? deLdExp(1.0, -7) : floatFormat16.ulp(deLog2(xd), 3.0));
13413                 const double    ulps1   (1.0 + 4.0 * deAbs(yd * (deLog2(xd) - logxeps)));
13414                 const double    ulps2   (1.0 + 4.0 * deAbs(yd * (deLog2(xd) + logxeps)));
13415                 const double    ulps    (deMax(deAbs(ulps1), deAbs(ulps2)));
13416                 double                  result  (0.0);
13417
13418                 if (xd < 0.0)
13419                         return false;
13420
13421                 if (x.isZero() && yd <= 0.0)
13422                         return false;
13423
13424                 if (getFlavor() == 0)
13425                 {
13426                         result = dePow(xd, yd);
13427                 }
13428                 else if (getFlavor() == 1)
13429                 {
13430                         const double    l2d     (deLog2(xd));
13431                         const double    e2d     (deExp2(yd * l2d));
13432
13433                         result = e2d;
13434                 }
13435                 else if (getFlavor() == 2)
13436                 {
13437                         const double    l2d     (deLog2(xd));
13438                         const fp16type  l2      (l2d);
13439                         const double    e2d     (deExp2(yd * l2.asDouble()));
13440                         const fp16type  e2      (e2d);
13441
13442                         result = e2.asDouble();
13443                 }
13444                 else
13445                 {
13446                         TCU_THROW(InternalError, "Unknown flavor");
13447                 }
13448
13449                 out[0] = fp16type(result).bits();
13450                 min[0] = getMin(result, ulps);
13451                 max[0] = getMax(result, ulps);
13452
13453                 return true;
13454         }
13455 };
13456
13457 struct fp16FMin : public fp16PerComponent
13458 {
13459         template<class fp16type>
13460         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
13461         {
13462                 const fp16type  x               (*in[0]);
13463                 const fp16type  y               (*in[1]);
13464                 const double    xd              (x.asDouble());
13465                 const double    yd              (y.asDouble());
13466                 const double    result  (deMin(xd, yd));
13467
13468                 if (x.isNaN() || y.isNaN())
13469                         return false;
13470
13471                 out[0] = fp16type(result).bits();
13472                 min[0] = getMin(result, getULPs(in));
13473                 max[0] = getMax(result, getULPs(in));
13474
13475                 return true;
13476         }
13477 };
13478
13479 struct fp16FMax : public fp16PerComponent
13480 {
13481         template<class fp16type>
13482         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
13483         {
13484                 const fp16type  x               (*in[0]);
13485                 const fp16type  y               (*in[1]);
13486                 const double    xd              (x.asDouble());
13487                 const double    yd              (y.asDouble());
13488                 const double    result  (deMax(xd, yd));
13489
13490                 if (x.isNaN() || y.isNaN())
13491                         return false;
13492
13493                 out[0] = fp16type(result).bits();
13494                 min[0] = getMin(result, getULPs(in));
13495                 max[0] = getMax(result, getULPs(in));
13496
13497                 return true;
13498         }
13499 };
13500
13501 struct fp16Step : public fp16PerComponent
13502 {
13503         template<class fp16type>
13504         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
13505         {
13506                 const fp16type  edge    (*in[0]);
13507                 const fp16type  x               (*in[1]);
13508                 const double    edged   (edge.asDouble());
13509                 const double    xd              (x.asDouble());
13510                 const double    result  (deStep(edged, xd));
13511
13512                 out[0] = fp16type(result).bits();
13513                 min[0] = getMin(result, getULPs(in));
13514                 max[0] = getMax(result, getULPs(in));
13515
13516                 return true;
13517         }
13518 };
13519
13520 struct fp16Ldexp : public fp16PerComponent
13521 {
13522         template<class fp16type>
13523         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
13524         {
13525                 const fp16type  x               (*in[0]);
13526                 const fp16type  y               (*in[1]);
13527                 const double    xd              (x.asDouble());
13528                 const int               yd              (static_cast<int>(deTrunc(y.asDouble())));
13529                 const double    result  (deLdExp(xd, yd));
13530
13531                 if (y.isNaN() || y.isInf() || y.isDenorm() || yd < -14 || yd > 15)
13532                         return false;
13533
13534                 // Spec: "If this product is too large to be represented in the floating-point type, the result is undefined."
13535                 if (fp16type(result).isInf())
13536                         return false;
13537
13538                 out[0] = fp16type(result).bits();
13539                 min[0] = getMin(result, getULPs(in));
13540                 max[0] = getMax(result, getULPs(in));
13541
13542                 return true;
13543         }
13544 };
13545
13546 struct fp16FClamp : public fp16PerComponent
13547 {
13548         template<class fp16type>
13549         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
13550         {
13551                 const fp16type  x               (*in[0]);
13552                 const fp16type  minVal  (*in[1]);
13553                 const fp16type  maxVal  (*in[2]);
13554                 const double    xd              (x.asDouble());
13555                 const double    minVald (minVal.asDouble());
13556                 const double    maxVald (maxVal.asDouble());
13557                 const double    result  (deClamp(xd, minVald, maxVald));
13558
13559                 if (minVal.isNaN() || maxVal.isNaN() || minVald > maxVald)
13560                         return false;
13561
13562                 out[0] = fp16type(result).bits();
13563                 min[0] = getMin(result, getULPs(in));
13564                 max[0] = getMax(result, getULPs(in));
13565
13566                 return true;
13567         }
13568 };
13569
13570 struct fp16FMix : public fp16PerComponent
13571 {
13572         fp16FMix() : fp16PerComponent()
13573         {
13574                 flavorNames.push_back("DoubleCalc");
13575                 flavorNames.push_back("EmulatingFP16");
13576                 flavorNames.push_back("EmulatingFP16YminusX");
13577         }
13578
13579         template<class fp16type>
13580         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
13581         {
13582                 const fp16type  x               (*in[0]);
13583                 const fp16type  y               (*in[1]);
13584                 const fp16type  a               (*in[2]);
13585                 const double    ulps    (8.0); // This is not a precision test. Value is not from spec
13586                 double                  result  (0.0);
13587
13588                 if (getFlavor() == 0)
13589                 {
13590                         const double    xd              (x.asDouble());
13591                         const double    yd              (y.asDouble());
13592                         const double    ad              (a.asDouble());
13593                         const double    xeps    (floatFormat16.ulp(deAbs(xd * (1.0 - ad)), ulps));
13594                         const double    yeps    (floatFormat16.ulp(deAbs(yd * ad), ulps));
13595                         const double    eps             (xeps + yeps);
13596
13597                         result = deMix(xd, yd, ad);
13598                         min[0] = result - eps;
13599                         max[0] = result + eps;
13600                 }
13601                 else if (getFlavor() == 1)
13602                 {
13603                         const double    xd              (x.asDouble());
13604                         const double    yd              (y.asDouble());
13605                         const double    ad              (a.asDouble());
13606                         const fp16type  am              (1.0 - ad);
13607                         const double    amd             (am.asDouble());
13608                         const fp16type  xam             (xd * amd);
13609                         const double    xamd    (xam.asDouble());
13610                         const fp16type  ya              (yd * ad);
13611                         const double    yad             (ya.asDouble());
13612                         const double    xeps    (floatFormat16.ulp(deAbs(xd * (1.0 - ad)), ulps));
13613                         const double    yeps    (floatFormat16.ulp(deAbs(yd * ad), ulps));
13614                         const double    eps             (xeps + yeps);
13615
13616                         result = xamd + yad;
13617                         min[0] = result - eps;
13618                         max[0] = result + eps;
13619                 }
13620                 else if (getFlavor() == 2)
13621                 {
13622                         const double    xd              (x.asDouble());
13623                         const double    yd              (y.asDouble());
13624                         const double    ad              (a.asDouble());
13625                         const fp16type  ymx             (yd - xd);
13626                         const double    ymxd    (ymx.asDouble());
13627                         const fp16type  ymxa    (ymxd * ad);
13628                         const double    ymxad   (ymxa.asDouble());
13629                         const double    xeps    (floatFormat16.ulp(deAbs(xd * (1.0 - ad)), ulps));
13630                         const double    yeps    (floatFormat16.ulp(deAbs(yd * ad), ulps));
13631                         const double    eps             (xeps + yeps);
13632
13633                         result = xd + ymxad;
13634                         min[0] = result - eps;
13635                         max[0] = result + eps;
13636                 }
13637                 else
13638                 {
13639                         TCU_THROW(InternalError, "Unknown flavor");
13640                 }
13641
13642                 out[0] = fp16type(result).bits();
13643
13644                 return true;
13645         }
13646 };
13647
13648 struct fp16SmoothStep : public fp16PerComponent
13649 {
13650         fp16SmoothStep() : fp16PerComponent()
13651         {
13652                 flavorNames.push_back("FloatCalc");
13653                 flavorNames.push_back("EmulatingFP16");
13654                 flavorNames.push_back("EmulatingFP16WClamp");
13655         }
13656
13657         virtual double getULPs(vector<const deFloat16*>& in)
13658         {
13659                 DE_UNREF(in);
13660
13661                 return 4.0; // This is not a precision test. Value is not from spec
13662         }
13663
13664         template<class fp16type>
13665         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
13666         {
13667                 const fp16type  edge0   (*in[0]);
13668                 const fp16type  edge1   (*in[1]);
13669                 const fp16type  x               (*in[2]);
13670                 double                  result  (0.0);
13671
13672                 if (edge0.isNaN() || edge1.isNaN() || x.isNaN() || edge0.asDouble() >= edge1.asDouble())
13673                         return false;
13674
13675                 if (edge0.isInf() || edge1.isInf() || x.isInf())
13676                         return false;
13677
13678                 if (getFlavor() == 0)
13679                 {
13680                         const float     edge0d  (edge0.asFloat());
13681                         const float     edge1d  (edge1.asFloat());
13682                         const float     xd              (x.asFloat());
13683                         const float     sstep   (deFloatSmoothStep(edge0d, edge1d, xd));
13684
13685                         result = sstep;
13686                 }
13687                 else if (getFlavor() == 1)
13688                 {
13689                         const double    edge0d  (edge0.asDouble());
13690                         const double    edge1d  (edge1.asDouble());
13691                         const double    xd              (x.asDouble());
13692
13693                         if (xd <= edge0d)
13694                                 result = 0.0;
13695                         else if (xd >= edge1d)
13696                                 result = 1.0;
13697                         else
13698                         {
13699                                 const fp16type  a       (xd - edge0d);
13700                                 const fp16type  b       (edge1d - edge0d);
13701                                 const fp16type  t       (a.asDouble() / b.asDouble());
13702                                 const fp16type  t2      (2.0 * t.asDouble());
13703                                 const fp16type  t3      (3.0 - t2.asDouble());
13704                                 const fp16type  t4      (t.asDouble() * t3.asDouble());
13705                                 const fp16type  t5      (t.asDouble() * t4.asDouble());
13706
13707                                 result = t5.asDouble();
13708                         }
13709                 }
13710                 else if (getFlavor() == 2)
13711                 {
13712                         const double    edge0d  (edge0.asDouble());
13713                         const double    edge1d  (edge1.asDouble());
13714                         const double    xd              (x.asDouble());
13715                         const fp16type  a       (xd - edge0d);
13716                         const fp16type  b       (edge1d - edge0d);
13717                         const fp16type  bi      (1.0 / b.asDouble());
13718                         const fp16type  t0      (a.asDouble() * bi.asDouble());
13719                         const double    tc      (deClamp(t0.asDouble(), 0.0, 1.0));
13720                         const fp16type  t       (tc);
13721                         const fp16type  t2      (2.0 * t.asDouble());
13722                         const fp16type  t3      (3.0 - t2.asDouble());
13723                         const fp16type  t4      (t.asDouble() * t3.asDouble());
13724                         const fp16type  t5      (t.asDouble() * t4.asDouble());
13725
13726                         result = t5.asDouble();
13727                 }
13728                 else
13729                 {
13730                         TCU_THROW(InternalError, "Unknown flavor");
13731                 }
13732
13733                 out[0] = fp16type(result).bits();
13734                 min[0] = getMin(result, getULPs(in));
13735                 max[0] = getMax(result, getULPs(in));
13736
13737                 return true;
13738         }
13739 };
13740
13741 struct fp16Fma : public fp16PerComponent
13742 {
13743         fp16Fma()
13744         {
13745                 flavorNames.push_back("DoubleCalc");
13746                 flavorNames.push_back("EmulatingFP16");
13747         }
13748
13749         virtual double getULPs(vector<const deFloat16*>& in)
13750         {
13751                 DE_UNREF(in);
13752
13753                 return 16.0;
13754         }
13755
13756         template<class fp16type>
13757         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
13758         {
13759                 DE_ASSERT(in.size() == 3);
13760                 DE_ASSERT(getArgCompCount(0) == getOutCompCount());
13761                 DE_ASSERT(getArgCompCount(1) == getOutCompCount());
13762                 DE_ASSERT(getArgCompCount(2) == getOutCompCount());
13763                 DE_ASSERT(getOutCompCount() > 0);
13764
13765                 const fp16type  a               (*in[0]);
13766                 const fp16type  b               (*in[1]);
13767                 const fp16type  c               (*in[2]);
13768                 double                  result  (0.0);
13769
13770                 if (getFlavor() == 0)
13771                 {
13772                         const double    ad      (a.asDouble());
13773                         const double    bd      (b.asDouble());
13774                         const double    cd      (c.asDouble());
13775
13776                         result  = deMadd(ad, bd, cd);
13777                 }
13778                 else if (getFlavor() == 1)
13779                 {
13780                         const double    ad      (a.asDouble());
13781                         const double    bd      (b.asDouble());
13782                         const double    cd      (c.asDouble());
13783                         const fp16type  ab      (ad * bd);
13784                         const fp16type  r       (ab.asDouble() + cd);
13785
13786                         result  = r.asDouble();
13787                 }
13788                 else
13789                 {
13790                         TCU_THROW(InternalError, "Unknown flavor");
13791                 }
13792
13793                 out[0] = fp16type(result).bits();
13794                 min[0] = getMin(result, getULPs(in));
13795                 max[0] = getMax(result, getULPs(in));
13796
13797                 return true;
13798         }
13799 };
13800
13801
13802 struct fp16AllComponents : public fp16PerComponent
13803 {
13804         bool            callOncePerComponent    ()      { return false; }
13805 };
13806
13807 struct fp16Length : public fp16AllComponents
13808 {
13809         fp16Length() : fp16AllComponents()
13810         {
13811                 flavorNames.push_back("EmulatingFP16");
13812                 flavorNames.push_back("DoubleCalc");
13813         }
13814
13815         virtual double getULPs(vector<const deFloat16*>& in)
13816         {
13817                 DE_UNREF(in);
13818
13819                 return 4.0;
13820         }
13821
13822         template<class fp16type>
13823         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
13824         {
13825                 DE_ASSERT(getOutCompCount() == 1);
13826                 DE_ASSERT(in.size() == 1);
13827
13828                 double  result  (0.0);
13829
13830                 if (getFlavor() == 0)
13831                 {
13832                         fp16type        r       (0.0);
13833
13834                         for (size_t componentNdx = 0; componentNdx < getArgCompCount(0); ++componentNdx)
13835                         {
13836                                 const fp16type  x       (in[0][componentNdx]);
13837                                 const fp16type  q       (x.asDouble() * x.asDouble());
13838
13839                                 r = fp16type(r.asDouble() + q.asDouble());
13840                         }
13841
13842                         result = deSqrt(r.asDouble());
13843
13844                         out[0] = fp16type(result).bits();
13845                 }
13846                 else if (getFlavor() == 1)
13847                 {
13848                         double  r       (0.0);
13849
13850                         for (size_t componentNdx = 0; componentNdx < getArgCompCount(0); ++componentNdx)
13851                         {
13852                                 const fp16type  x       (in[0][componentNdx]);
13853                                 const double    q       (x.asDouble() * x.asDouble());
13854
13855                                 r += q;
13856                         }
13857
13858                         result = deSqrt(r);
13859
13860                         out[0] = fp16type(result).bits();
13861                 }
13862                 else
13863                 {
13864                         TCU_THROW(InternalError, "Unknown flavor");
13865                 }
13866
13867                 min[0] = getMin(result, getULPs(in));
13868                 max[0] = getMax(result, getULPs(in));
13869
13870                 return true;
13871         }
13872 };
13873
13874 struct fp16Distance : public fp16AllComponents
13875 {
13876         fp16Distance() : fp16AllComponents()
13877         {
13878                 flavorNames.push_back("EmulatingFP16");
13879                 flavorNames.push_back("DoubleCalc");
13880         }
13881
13882         virtual double getULPs(vector<const deFloat16*>& in)
13883         {
13884                 DE_UNREF(in);
13885
13886                 return 4.0;
13887         }
13888
13889         template<class fp16type>
13890         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
13891         {
13892                 DE_ASSERT(getOutCompCount() == 1);
13893                 DE_ASSERT(in.size() == 2);
13894                 DE_ASSERT(getArgCompCount(0) == getArgCompCount(1));
13895
13896                 double  result  (0.0);
13897
13898                 if (getFlavor() == 0)
13899                 {
13900                         fp16type        r       (0.0);
13901
13902                         for (size_t componentNdx = 0; componentNdx < getArgCompCount(0); ++componentNdx)
13903                         {
13904                                 const fp16type  x       (in[0][componentNdx]);
13905                                 const fp16type  y       (in[1][componentNdx]);
13906                                 const fp16type  d       (x.asDouble() - y.asDouble());
13907                                 const fp16type  q       (d.asDouble() * d.asDouble());
13908
13909                                 r = fp16type(r.asDouble() + q.asDouble());
13910                         }
13911
13912                         result = deSqrt(r.asDouble());
13913                 }
13914                 else if (getFlavor() == 1)
13915                 {
13916                         double  r       (0.0);
13917
13918                         for (size_t componentNdx = 0; componentNdx < getArgCompCount(0); ++componentNdx)
13919                         {
13920                                 const fp16type  x       (in[0][componentNdx]);
13921                                 const fp16type  y       (in[1][componentNdx]);
13922                                 const double    d       (x.asDouble() - y.asDouble());
13923                                 const double    q       (d * d);
13924
13925                                 r += q;
13926                         }
13927
13928                         result = deSqrt(r);
13929                 }
13930                 else
13931                 {
13932                         TCU_THROW(InternalError, "Unknown flavor");
13933                 }
13934
13935                 out[0] = fp16type(result).bits();
13936                 min[0] = getMin(result, getULPs(in));
13937                 max[0] = getMax(result, getULPs(in));
13938
13939                 return true;
13940         }
13941 };
13942
13943 struct fp16Cross : public fp16AllComponents
13944 {
13945         fp16Cross() : fp16AllComponents()
13946         {
13947                 flavorNames.push_back("EmulatingFP16");
13948                 flavorNames.push_back("DoubleCalc");
13949         }
13950
13951         virtual double getULPs(vector<const deFloat16*>& in)
13952         {
13953                 DE_UNREF(in);
13954
13955                 return 4.0;
13956         }
13957
13958         template<class fp16type>
13959         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
13960         {
13961                 DE_ASSERT(getOutCompCount() == 3);
13962                 DE_ASSERT(in.size() == 2);
13963                 DE_ASSERT(getArgCompCount(0) == 3);
13964                 DE_ASSERT(getArgCompCount(1) == 3);
13965
13966                 if (getFlavor() == 0)
13967                 {
13968                         const fp16type  x0              (in[0][0]);
13969                         const fp16type  x1              (in[0][1]);
13970                         const fp16type  x2              (in[0][2]);
13971                         const fp16type  y0              (in[1][0]);
13972                         const fp16type  y1              (in[1][1]);
13973                         const fp16type  y2              (in[1][2]);
13974                         const fp16type  x1y2    (x1.asDouble() * y2.asDouble());
13975                         const fp16type  y1x2    (y1.asDouble() * x2.asDouble());
13976                         const fp16type  x2y0    (x2.asDouble() * y0.asDouble());
13977                         const fp16type  y2x0    (y2.asDouble() * x0.asDouble());
13978                         const fp16type  x0y1    (x0.asDouble() * y1.asDouble());
13979                         const fp16type  y0x1    (y0.asDouble() * x1.asDouble());
13980
13981                         out[0] = fp16type(x1y2.asDouble() - y1x2.asDouble()).bits();
13982                         out[1] = fp16type(x2y0.asDouble() - y2x0.asDouble()).bits();
13983                         out[2] = fp16type(x0y1.asDouble() - y0x1.asDouble()).bits();
13984                 }
13985                 else if (getFlavor() == 1)
13986                 {
13987                         const fp16type  x0              (in[0][0]);
13988                         const fp16type  x1              (in[0][1]);
13989                         const fp16type  x2              (in[0][2]);
13990                         const fp16type  y0              (in[1][0]);
13991                         const fp16type  y1              (in[1][1]);
13992                         const fp16type  y2              (in[1][2]);
13993                         const double    x1y2    (x1.asDouble() * y2.asDouble());
13994                         const double    y1x2    (y1.asDouble() * x2.asDouble());
13995                         const double    x2y0    (x2.asDouble() * y0.asDouble());
13996                         const double    y2x0    (y2.asDouble() * x0.asDouble());
13997                         const double    x0y1    (x0.asDouble() * y1.asDouble());
13998                         const double    y0x1    (y0.asDouble() * x1.asDouble());
13999
14000                         out[0] = fp16type(x1y2 - y1x2).bits();
14001                         out[1] = fp16type(x2y0 - y2x0).bits();
14002                         out[2] = fp16type(x0y1 - y0x1).bits();
14003                 }
14004                 else
14005                 {
14006                         TCU_THROW(InternalError, "Unknown flavor");
14007                 }
14008
14009                 for (size_t ndx = 0; ndx < getOutCompCount(); ++ndx)
14010                         min[ndx] = getMin(fp16type(out[ndx]).asDouble(), getULPs(in));
14011                 for (size_t ndx = 0; ndx < getOutCompCount(); ++ndx)
14012                         max[ndx] = getMax(fp16type(out[ndx]).asDouble(), getULPs(in));
14013
14014                 return true;
14015         }
14016 };
14017
14018 struct fp16Normalize : public fp16AllComponents
14019 {
14020         fp16Normalize() : fp16AllComponents()
14021         {
14022                 flavorNames.push_back("EmulatingFP16");
14023                 flavorNames.push_back("DoubleCalc");
14024
14025                 // flavorNames will be extended later
14026         }
14027
14028         virtual void    setArgCompCount                 (size_t argNo, size_t compCount)
14029         {
14030                 DE_ASSERT(argCompCount[argNo] == 0); // Once only
14031
14032                 if (argNo == 0 && argCompCount[argNo] == 0)
14033                 {
14034                         const size_t            maxPermutationsCount    = 24u; // Equal to 4!
14035                         std::vector<int>        indices;
14036
14037                         for (size_t componentNdx = 0; componentNdx < compCount; ++componentNdx)
14038                                 indices.push_back(static_cast<int>(componentNdx));
14039
14040                         m_permutations.reserve(maxPermutationsCount);
14041
14042                         permutationsFlavorStart = flavorNames.size();
14043
14044                         do
14045                         {
14046                                 tcu::UVec4      permutation;
14047                                 std::string     name            = "Permutted_";
14048
14049                                 for (size_t componentNdx = 0; componentNdx < compCount; ++componentNdx)
14050                                 {
14051                                         permutation[static_cast<int>(componentNdx)] = indices[componentNdx];
14052                                         name += de::toString(indices[componentNdx]);
14053                                 }
14054
14055                                 m_permutations.push_back(permutation);
14056                                 flavorNames.push_back(name);
14057
14058                         } while(std::next_permutation(indices.begin(), indices.end()));
14059
14060                         permutationsFlavorEnd = flavorNames.size();
14061                 }
14062
14063                 fp16AllComponents::setArgCompCount(argNo, compCount);
14064         }
14065         virtual double getULPs(vector<const deFloat16*>& in)
14066         {
14067                 DE_UNREF(in);
14068
14069                 return 8.0;
14070         }
14071
14072         template<class fp16type>
14073         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
14074         {
14075                 DE_ASSERT(in.size() == 1);
14076                 DE_ASSERT(getArgCompCount(0) == getOutCompCount());
14077
14078                 if (getFlavor() == 0)
14079                 {
14080                         fp16type        r(0.0);
14081
14082                         for (size_t componentNdx = 0; componentNdx < getArgCompCount(0); ++componentNdx)
14083                         {
14084                                 const fp16type  x       (in[0][componentNdx]);
14085                                 const fp16type  q       (x.asDouble() * x.asDouble());
14086
14087                                 r = fp16type(r.asDouble() + q.asDouble());
14088                         }
14089
14090                         r = fp16type(deSqrt(r.asDouble()));
14091
14092                         if (r.isZero())
14093                                 return false;
14094
14095                         for (size_t componentNdx = 0; componentNdx < getArgCompCount(0); ++componentNdx)
14096                         {
14097                                 const fp16type  x       (in[0][componentNdx]);
14098
14099                                 out[componentNdx] = fp16type(x.asDouble() / r.asDouble()).bits();
14100                         }
14101                 }
14102                 else if (getFlavor() == 1)
14103                 {
14104                         double  r(0.0);
14105
14106                         for (size_t componentNdx = 0; componentNdx < getArgCompCount(0); ++componentNdx)
14107                         {
14108                                 const fp16type  x       (in[0][componentNdx]);
14109                                 const double    q       (x.asDouble() * x.asDouble());
14110
14111                                 r += q;
14112                         }
14113
14114                         r = deSqrt(r);
14115
14116                         if (r == 0)
14117                                 return false;
14118
14119                         for (size_t componentNdx = 0; componentNdx < getArgCompCount(0); ++componentNdx)
14120                         {
14121                                 const fp16type  x       (in[0][componentNdx]);
14122
14123                                 out[componentNdx] = fp16type(x.asDouble() / r).bits();
14124                         }
14125                 }
14126                 else if (de::inBounds<size_t>(getFlavor(), permutationsFlavorStart, permutationsFlavorEnd))
14127                 {
14128                         const int                       compCount               (static_cast<int>(getArgCompCount(0)));
14129                         const size_t            permutationNdx  (getFlavor() - permutationsFlavorStart);
14130                         const tcu::UVec4&       permutation             (m_permutations[permutationNdx]);
14131                         fp16type                        r                               (0.0);
14132
14133                         for (int permComponentNdx = 0; permComponentNdx < compCount; ++permComponentNdx)
14134                         {
14135                                 const size_t    componentNdx    (permutation[permComponentNdx]);
14136                                 const fp16type  x                               (in[0][componentNdx]);
14137                                 const fp16type  q                               (x.asDouble() * x.asDouble());
14138
14139                                 r = fp16type(r.asDouble() + q.asDouble());
14140                         }
14141
14142                         r = fp16type(deSqrt(r.asDouble()));
14143
14144                         if (r.isZero())
14145                                 return false;
14146
14147                         for (int permComponentNdx = 0; permComponentNdx < compCount; ++permComponentNdx)
14148                         {
14149                                 const size_t    componentNdx    (permutation[permComponentNdx]);
14150                                 const fp16type  x                               (in[0][componentNdx]);
14151
14152                                 out[componentNdx] = fp16type(x.asDouble() / r.asDouble()).bits();
14153                         }
14154                 }
14155                 else
14156                 {
14157                         TCU_THROW(InternalError, "Unknown flavor");
14158                 }
14159
14160                 for (size_t ndx = 0; ndx < getOutCompCount(); ++ndx)
14161                         min[ndx] = getMin(fp16type(out[ndx]).asDouble(), getULPs(in));
14162                 for (size_t ndx = 0; ndx < getOutCompCount(); ++ndx)
14163                         max[ndx] = getMax(fp16type(out[ndx]).asDouble(), getULPs(in));
14164
14165                 return true;
14166         }
14167
14168 private:
14169         std::vector<tcu::UVec4> m_permutations;
14170         size_t                                  permutationsFlavorStart;
14171         size_t                                  permutationsFlavorEnd;
14172 };
14173
14174 struct fp16FaceForward : public fp16AllComponents
14175 {
14176         virtual double getULPs(vector<const deFloat16*>& in)
14177         {
14178                 DE_UNREF(in);
14179
14180                 return 4.0;
14181         }
14182
14183         template<class fp16type>
14184         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
14185         {
14186                 DE_ASSERT(in.size() == 3);
14187                 DE_ASSERT(getArgCompCount(0) == getOutCompCount());
14188                 DE_ASSERT(getArgCompCount(1) == getOutCompCount());
14189                 DE_ASSERT(getArgCompCount(2) == getOutCompCount());
14190
14191                 fp16type        dp(0.0);
14192
14193                 for (size_t componentNdx = 0; componentNdx < getArgCompCount(1); ++componentNdx)
14194                 {
14195                         const fp16type  x       (in[1][componentNdx]);
14196                         const fp16type  y       (in[2][componentNdx]);
14197                         const double    xd      (x.asDouble());
14198                         const double    yd      (y.asDouble());
14199                         const fp16type  q       (xd * yd);
14200
14201                         dp = fp16type(dp.asDouble() + q.asDouble());
14202                 }
14203
14204                 if (dp.isNaN() || dp.isZero())
14205                         return false;
14206
14207                 for (size_t componentNdx = 0; componentNdx < getOutCompCount(); ++componentNdx)
14208                 {
14209                         const fp16type  n       (in[0][componentNdx]);
14210
14211                         out[componentNdx] = (dp.signBit() == 1) ? n.bits() : fp16type(-n.asDouble()).bits();
14212                 }
14213
14214                 for (size_t ndx = 0; ndx < getOutCompCount(); ++ndx)
14215                         min[ndx] = getMin(fp16type(out[ndx]).asDouble(), getULPs(in));
14216                 for (size_t ndx = 0; ndx < getOutCompCount(); ++ndx)
14217                         max[ndx] = getMax(fp16type(out[ndx]).asDouble(), getULPs(in));
14218
14219                 return true;
14220         }
14221 };
14222
14223 struct fp16Reflect : public fp16AllComponents
14224 {
14225         fp16Reflect() : fp16AllComponents()
14226         {
14227                 flavorNames.push_back("EmulatingFP16");
14228                 flavorNames.push_back("EmulatingFP16+KeepZeroSign");
14229                 flavorNames.push_back("FloatCalc");
14230                 flavorNames.push_back("FloatCalc+KeepZeroSign");
14231                 flavorNames.push_back("EmulatingFP16+2Nfirst");
14232                 flavorNames.push_back("EmulatingFP16+2Ifirst");
14233         }
14234
14235         virtual double getULPs(vector<const deFloat16*>& in)
14236         {
14237                 DE_UNREF(in);
14238
14239                 return 256.0; // This is not a precision test. Value is not from spec
14240         }
14241
14242         template<class fp16type>
14243         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
14244         {
14245                 DE_ASSERT(in.size() == 2);
14246                 DE_ASSERT(getArgCompCount(0) == getOutCompCount());
14247                 DE_ASSERT(getArgCompCount(1) == getOutCompCount());
14248
14249                 if (getFlavor() < 4)
14250                 {
14251                         const bool      keepZeroSign    ((flavor & 1) != 0 ? true : false);
14252                         const bool      floatCalc               ((flavor & 2) != 0 ? true : false);
14253
14254                         if (floatCalc)
14255                         {
14256                                 float   dp(0.0f);
14257
14258                                 for (size_t componentNdx = 0; componentNdx < getArgCompCount(1); ++componentNdx)
14259                                 {
14260                                         const fp16type  i       (in[0][componentNdx]);
14261                                         const fp16type  n       (in[1][componentNdx]);
14262                                         const float             id      (i.asFloat());
14263                                         const float             nd      (n.asFloat());
14264                                         const float             qd      (id * nd);
14265
14266                                         if (keepZeroSign)
14267                                                 dp = (componentNdx == 0) ? qd : dp + qd;
14268                                         else
14269                                                 dp = dp + qd;
14270                                 }
14271
14272                                 for (size_t componentNdx = 0; componentNdx < getOutCompCount(); ++componentNdx)
14273                                 {
14274                                         const fp16type  i               (in[0][componentNdx]);
14275                                         const fp16type  n               (in[1][componentNdx]);
14276                                         const float             dpnd    (dp * n.asFloat());
14277                                         const float             dpn2d   (2.0f * dpnd);
14278                                         const float             idpn2d  (i.asFloat() - dpn2d);
14279                                         const fp16type  result  (idpn2d);
14280
14281                                         out[componentNdx] = result.bits();
14282                                 }
14283                         }
14284                         else
14285                         {
14286                                 fp16type        dp(0.0);
14287
14288                                 for (size_t componentNdx = 0; componentNdx < getArgCompCount(1); ++componentNdx)
14289                                 {
14290                                         const fp16type  i       (in[0][componentNdx]);
14291                                         const fp16type  n       (in[1][componentNdx]);
14292                                         const double    id      (i.asDouble());
14293                                         const double    nd      (n.asDouble());
14294                                         const fp16type  q       (id * nd);
14295
14296                                         if (keepZeroSign)
14297                                                 dp = (componentNdx == 0) ? q : fp16type(dp.asDouble() + q.asDouble());
14298                                         else
14299                                                 dp = fp16type(dp.asDouble() + q.asDouble());
14300                                 }
14301
14302                                 if (dp.isNaN())
14303                                         return false;
14304
14305                                 for (size_t componentNdx = 0; componentNdx < getOutCompCount(); ++componentNdx)
14306                                 {
14307                                         const fp16type  i               (in[0][componentNdx]);
14308                                         const fp16type  n               (in[1][componentNdx]);
14309                                         const fp16type  dpn             (dp.asDouble() * n.asDouble());
14310                                         const fp16type  dpn2    (2 * dpn.asDouble());
14311                                         const fp16type  idpn2   (i.asDouble() - dpn2.asDouble());
14312
14313                                         out[componentNdx] = idpn2.bits();
14314                                 }
14315                         }
14316                 }
14317                 else if (getFlavor() == 4)
14318                 {
14319                         fp16type        dp(0.0);
14320
14321                         for (size_t componentNdx = 0; componentNdx < getArgCompCount(1); ++componentNdx)
14322                         {
14323                                 const fp16type  i       (in[0][componentNdx]);
14324                                 const fp16type  n       (in[1][componentNdx]);
14325                                 const double    id      (i.asDouble());
14326                                 const double    nd      (n.asDouble());
14327                                 const fp16type  q       (id * nd);
14328
14329                                 dp = fp16type(dp.asDouble() + q.asDouble());
14330                         }
14331
14332                         if (dp.isNaN())
14333                                 return false;
14334
14335                         for (size_t componentNdx = 0; componentNdx < getOutCompCount(); ++componentNdx)
14336                         {
14337                                 const fp16type  i               (in[0][componentNdx]);
14338                                 const fp16type  n               (in[1][componentNdx]);
14339                                 const fp16type  n2              (2 * n.asDouble());
14340                                 const fp16type  dpn2    (dp.asDouble() * n2.asDouble());
14341                                 const fp16type  idpn2   (i.asDouble() - dpn2.asDouble());
14342
14343                                 out[componentNdx] = idpn2.bits();
14344                         }
14345                 }
14346                 else if (getFlavor() == 5)
14347                 {
14348                         fp16type        dp2(0.0);
14349
14350                         for (size_t componentNdx = 0; componentNdx < getArgCompCount(1); ++componentNdx)
14351                         {
14352                                 const fp16type  i       (in[0][componentNdx]);
14353                                 const fp16type  n       (in[1][componentNdx]);
14354                                 const fp16type  i2      (2.0 * i.asDouble());
14355                                 const double    i2d     (i2.asDouble());
14356                                 const double    nd      (n.asDouble());
14357                                 const fp16type  q       (i2d * nd);
14358
14359                                 dp2 = fp16type(dp2.asDouble() + q.asDouble());
14360                         }
14361
14362                         if (dp2.isNaN())
14363                                 return false;
14364
14365                         for (size_t componentNdx = 0; componentNdx < getOutCompCount(); ++componentNdx)
14366                         {
14367                                 const fp16type  i               (in[0][componentNdx]);
14368                                 const fp16type  n               (in[1][componentNdx]);
14369                                 const fp16type  dpn2    (dp2.asDouble() * n.asDouble());
14370                                 const fp16type  idpn2   (i.asDouble() - dpn2.asDouble());
14371
14372                                 out[componentNdx] = idpn2.bits();
14373                         }
14374                 }
14375                 else
14376                 {
14377                         TCU_THROW(InternalError, "Unknown flavor");
14378                 }
14379
14380                 for (size_t ndx = 0; ndx < getOutCompCount(); ++ndx)
14381                         min[ndx] = getMin(fp16type(out[ndx]).asDouble(), getULPs(in));
14382                 for (size_t ndx = 0; ndx < getOutCompCount(); ++ndx)
14383                         max[ndx] = getMax(fp16type(out[ndx]).asDouble(), getULPs(in));
14384
14385                 return true;
14386         }
14387 };
14388
14389 struct fp16Refract : public fp16AllComponents
14390 {
14391         fp16Refract() : fp16AllComponents()
14392         {
14393                 flavorNames.push_back("EmulatingFP16");
14394                 flavorNames.push_back("EmulatingFP16+KeepZeroSign");
14395                 flavorNames.push_back("FloatCalc");
14396                 flavorNames.push_back("FloatCalc+KeepZeroSign");
14397         }
14398
14399         virtual double getULPs(vector<const deFloat16*>& in)
14400         {
14401                 DE_UNREF(in);
14402
14403                 return 8192.0; // This is not a precision test. Value is not from spec
14404         }
14405
14406         template<class fp16type>
14407         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
14408         {
14409                 DE_ASSERT(in.size() == 3);
14410                 DE_ASSERT(getArgCompCount(0) == getOutCompCount());
14411                 DE_ASSERT(getArgCompCount(1) == getOutCompCount());
14412                 DE_ASSERT(getArgCompCount(2) == 1);
14413
14414                 const bool              keepZeroSign    ((flavor & 1) != 0 ? true : false);
14415                 const bool              doubleCalc              ((flavor & 2) != 0 ? true : false);
14416                 const fp16type  eta                             (*in[2]);
14417
14418                 if (doubleCalc)
14419                 {
14420                         double  dp      (0.0);
14421
14422                         for (size_t componentNdx = 0; componentNdx < getArgCompCount(1); ++componentNdx)
14423                         {
14424                                 const fp16type  i       (in[0][componentNdx]);
14425                                 const fp16type  n       (in[1][componentNdx]);
14426                                 const double    id      (i.asDouble());
14427                                 const double    nd      (n.asDouble());
14428                                 const double    qd      (id * nd);
14429
14430                                 if (keepZeroSign)
14431                                         dp = (componentNdx == 0) ? qd : dp + qd;
14432                                 else
14433                                         dp = dp + qd;
14434                         }
14435
14436                         const double    eta2    (eta.asDouble() * eta.asDouble());
14437                         const double    dp2             (dp * dp);
14438                         const double    dp1             (1.0 - dp2);
14439                         const double    dpe             (eta2 * dp1);
14440                         const double    k               (1.0 - dpe);
14441
14442                         if (k < 0.0)
14443                         {
14444                                 const fp16type  zero    (0.0);
14445
14446                                 for (size_t componentNdx = 0; componentNdx < getOutCompCount(); ++componentNdx)
14447                                         out[componentNdx] = zero.bits();
14448                         }
14449                         else
14450                         {
14451                                 const double    sk      (deSqrt(k));
14452
14453                                 for (size_t componentNdx = 0; componentNdx < getOutCompCount(); ++componentNdx)
14454                                 {
14455                                         const fp16type  i               (in[0][componentNdx]);
14456                                         const fp16type  n               (in[1][componentNdx]);
14457                                         const double    etai    (i.asDouble() * eta.asDouble());
14458                                         const double    etadp   (eta.asDouble() * dp);
14459                                         const double    etadpk  (etadp + sk);
14460                                         const double    etadpkn (etadpk * n.asDouble());
14461                                         const double    full    (etai - etadpkn);
14462                                         const fp16type  result  (full);
14463
14464                                         if (result.isInf())
14465                                                 return false;
14466
14467                                         out[componentNdx] = result.bits();
14468                                 }
14469                         }
14470                 }
14471                 else
14472                 {
14473                         fp16type        dp      (0.0);
14474
14475                         for (size_t componentNdx = 0; componentNdx < getArgCompCount(1); ++componentNdx)
14476                         {
14477                                 const fp16type  i       (in[0][componentNdx]);
14478                                 const fp16type  n       (in[1][componentNdx]);
14479                                 const double    id      (i.asDouble());
14480                                 const double    nd      (n.asDouble());
14481                                 const fp16type  q       (id * nd);
14482
14483                                 if (keepZeroSign)
14484                                         dp = (componentNdx == 0) ? q : fp16type(dp.asDouble() + q.asDouble());
14485                                 else
14486                                         dp = fp16type(dp.asDouble() + q.asDouble());
14487                         }
14488
14489                         if (dp.isNaN())
14490                                 return false;
14491
14492                         const fp16type  eta2(eta.asDouble() * eta.asDouble());
14493                         const fp16type  dp2     (dp.asDouble() * dp.asDouble());
14494                         const fp16type  dp1     (1.0 - dp2.asDouble());
14495                         const fp16type  dpe     (eta2.asDouble() * dp1.asDouble());
14496                         const fp16type  k       (1.0 - dpe.asDouble());
14497
14498                         if (k.asDouble() < 0.0)
14499                         {
14500                                 const fp16type  zero    (0.0);
14501
14502                                 for (size_t componentNdx = 0; componentNdx < getOutCompCount(); ++componentNdx)
14503                                         out[componentNdx] = zero.bits();
14504                         }
14505                         else
14506                         {
14507                                 const fp16type  sk      (deSqrt(k.asDouble()));
14508
14509                                 for (size_t componentNdx = 0; componentNdx < getOutCompCount(); ++componentNdx)
14510                                 {
14511                                         const fp16type  i               (in[0][componentNdx]);
14512                                         const fp16type  n               (in[1][componentNdx]);
14513                                         const fp16type  etai    (i.asDouble() * eta.asDouble());
14514                                         const fp16type  etadp   (eta.asDouble() * dp.asDouble());
14515                                         const fp16type  etadpk  (etadp.asDouble() + sk.asDouble());
14516                                         const fp16type  etadpkn (etadpk.asDouble() * n.asDouble());
14517                                         const fp16type  full    (etai.asDouble() - etadpkn.asDouble());
14518
14519                                         if (full.isNaN() || full.isInf())
14520                                                 return false;
14521
14522                                         out[componentNdx] = full.bits();
14523                                 }
14524                         }
14525                 }
14526
14527                 for (size_t ndx = 0; ndx < getOutCompCount(); ++ndx)
14528                         min[ndx] = getMin(fp16type(out[ndx]).asDouble(), getULPs(in));
14529                 for (size_t ndx = 0; ndx < getOutCompCount(); ++ndx)
14530                         max[ndx] = getMax(fp16type(out[ndx]).asDouble(), getULPs(in));
14531
14532                 return true;
14533         }
14534 };
14535
14536 struct fp16Dot : public fp16AllComponents
14537 {
14538         fp16Dot() : fp16AllComponents()
14539         {
14540                 flavorNames.push_back("EmulatingFP16");
14541                 flavorNames.push_back("FloatCalc");
14542                 flavorNames.push_back("DoubleCalc");
14543
14544                 // flavorNames will be extended later
14545         }
14546
14547         virtual void    setArgCompCount                 (size_t argNo, size_t compCount)
14548         {
14549                 DE_ASSERT(argCompCount[argNo] == 0); // Once only
14550
14551                 if (argNo == 0 && argCompCount[argNo] == 0)
14552                 {
14553                         const size_t            maxPermutationsCount    = 24u; // Equal to 4!
14554                         std::vector<int>        indices;
14555
14556                         for (size_t componentNdx = 0; componentNdx < compCount; ++componentNdx)
14557                                 indices.push_back(static_cast<int>(componentNdx));
14558
14559                         m_permutations.reserve(maxPermutationsCount);
14560
14561                         permutationsFlavorStart = flavorNames.size();
14562
14563                         do
14564                         {
14565                                 tcu::UVec4      permutation;
14566                                 std::string     name            = "Permutted_";
14567
14568                                 for (size_t componentNdx = 0; componentNdx < compCount; ++componentNdx)
14569                                 {
14570                                         permutation[static_cast<int>(componentNdx)] = indices[componentNdx];
14571                                         name += de::toString(indices[componentNdx]);
14572                                 }
14573
14574                                 m_permutations.push_back(permutation);
14575                                 flavorNames.push_back(name);
14576
14577                         } while(std::next_permutation(indices.begin(), indices.end()));
14578
14579                         permutationsFlavorEnd = flavorNames.size();
14580                 }
14581
14582                 fp16AllComponents::setArgCompCount(argNo, compCount);
14583         }
14584
14585         virtual double  getULPs(vector<const deFloat16*>& in)
14586         {
14587                 DE_UNREF(in);
14588
14589                 return 16.0; // This is not a precision test. Value is not from spec
14590         }
14591
14592         template<class fp16type>
14593         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
14594         {
14595                 DE_ASSERT(in.size() == 2);
14596                 DE_ASSERT(getArgCompCount(0) == getArgCompCount(1));
14597                 DE_ASSERT(getOutCompCount() == 1);
14598
14599                 double  result  (0.0);
14600                 double  eps             (0.0);
14601
14602                 if (getFlavor() == 0)
14603                 {
14604                         fp16type        dp      (0.0);
14605
14606                         for (size_t componentNdx = 0; componentNdx < getArgCompCount(1); ++componentNdx)
14607                         {
14608                                 const fp16type  x       (in[0][componentNdx]);
14609                                 const fp16type  y       (in[1][componentNdx]);
14610                                 const fp16type  q       (x.asDouble() * y.asDouble());
14611
14612                                 dp = fp16type(dp.asDouble() + q.asDouble());
14613                                 eps += floatFormat16.ulp(q.asDouble(), 2.0);
14614                         }
14615
14616                         result = dp.asDouble();
14617                 }
14618                 else if (getFlavor() == 1)
14619                 {
14620                         float   dp      (0.0);
14621
14622                         for (size_t componentNdx = 0; componentNdx < getArgCompCount(1); ++componentNdx)
14623                         {
14624                                 const fp16type  x       (in[0][componentNdx]);
14625                                 const fp16type  y       (in[1][componentNdx]);
14626                                 const float             q       (x.asFloat() * y.asFloat());
14627
14628                                 dp += q;
14629                                 eps += floatFormat16.ulp(static_cast<double>(q), 2.0);
14630                         }
14631
14632                         result = dp;
14633                 }
14634                 else if (getFlavor() == 2)
14635                 {
14636                         double  dp      (0.0);
14637
14638                         for (size_t componentNdx = 0; componentNdx < getArgCompCount(1); ++componentNdx)
14639                         {
14640                                 const fp16type  x       (in[0][componentNdx]);
14641                                 const fp16type  y       (in[1][componentNdx]);
14642                                 const double    q       (x.asDouble() * y.asDouble());
14643
14644                                 dp += q;
14645                                 eps += floatFormat16.ulp(q, 2.0);
14646                         }
14647
14648                         result = dp;
14649                 }
14650                 else if (de::inBounds<size_t>(getFlavor(), permutationsFlavorStart, permutationsFlavorEnd))
14651                 {
14652                         const int                       compCount               (static_cast<int>(getArgCompCount(1)));
14653                         const size_t            permutationNdx  (getFlavor() - permutationsFlavorStart);
14654                         const tcu::UVec4&       permutation             (m_permutations[permutationNdx]);
14655                         fp16type                        dp                              (0.0);
14656
14657                         for (int permComponentNdx = 0; permComponentNdx < compCount; ++permComponentNdx)
14658                         {
14659                                 const size_t            componentNdx    (permutation[permComponentNdx]);
14660                                 const fp16type          x                               (in[0][componentNdx]);
14661                                 const fp16type          y                               (in[1][componentNdx]);
14662                                 const fp16type          q                               (x.asDouble() * y.asDouble());
14663
14664                                 dp = fp16type(dp.asDouble() + q.asDouble());
14665                                 eps += floatFormat16.ulp(q.asDouble(), 2.0);
14666                         }
14667
14668                         result = dp.asDouble();
14669                 }
14670                 else
14671                 {
14672                         TCU_THROW(InternalError, "Unknown flavor");
14673                 }
14674
14675                 out[0] = fp16type(result).bits();
14676                 min[0] = result - eps;
14677                 max[0] = result + eps;
14678
14679                 return true;
14680         }
14681
14682 private:
14683         std::vector<tcu::UVec4> m_permutations;
14684         size_t                                  permutationsFlavorStart;
14685         size_t                                  permutationsFlavorEnd;
14686 };
14687
14688 struct fp16VectorTimesScalar : public fp16AllComponents
14689 {
14690         virtual double getULPs(vector<const deFloat16*>& in)
14691         {
14692                 DE_UNREF(in);
14693
14694                 return 2.0;
14695         }
14696
14697         template<class fp16type>
14698         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
14699         {
14700                 DE_ASSERT(in.size() == 2);
14701                 DE_ASSERT(getArgCompCount(0) == getOutCompCount());
14702                 DE_ASSERT(getArgCompCount(1) == 1);
14703
14704                 fp16type        s       (*in[1]);
14705
14706                 for (size_t componentNdx = 0; componentNdx < getArgCompCount(0); ++componentNdx)
14707                 {
14708                         const fp16type  x          (in[0][componentNdx]);
14709                         const double    result (s.asDouble() * x.asDouble());
14710                         const fp16type  m          (result);
14711
14712                         out[componentNdx] = m.bits();
14713                         min[componentNdx] = getMin(result, getULPs(in));
14714                         max[componentNdx] = getMax(result, getULPs(in));
14715                 }
14716
14717                 return true;
14718         }
14719 };
14720
14721 struct fp16MatrixBase : public fp16AllComponents
14722 {
14723         deUint32                getComponentValidity                    ()
14724         {
14725                 return static_cast<deUint32>(-1);
14726         }
14727
14728         inline size_t   getNdx                                                  (const size_t rowCount, const size_t col, const size_t row)
14729         {
14730                 const size_t minComponentCount  = 0;
14731                 const size_t maxComponentCount  = 3;
14732                 const size_t alignedRowsCount   = (rowCount == 3) ? 4 : rowCount;
14733
14734                 DE_ASSERT(de::inRange(rowCount, minComponentCount + 1, maxComponentCount + 1));
14735                 DE_ASSERT(de::inRange(col, minComponentCount, maxComponentCount));
14736                 DE_ASSERT(de::inBounds(row, minComponentCount, rowCount));
14737                 DE_UNREF(minComponentCount);
14738                 DE_UNREF(maxComponentCount);
14739
14740                 return col * alignedRowsCount + row;
14741         }
14742
14743         deUint32                getComponentMatrixValidityMask  (size_t cols, size_t rows)
14744         {
14745                 deUint32        result  = 0u;
14746
14747                 for (size_t rowNdx = 0; rowNdx < rows; ++rowNdx)
14748                         for (size_t colNdx = 0; colNdx < cols; ++colNdx)
14749                         {
14750                                 const size_t bitNdx = getNdx(rows, colNdx, rowNdx);
14751
14752                                 DE_ASSERT(bitNdx < sizeof(result) * 8);
14753
14754                                 result |= (1<<bitNdx);
14755                         }
14756
14757                 return result;
14758         }
14759 };
14760
14761 template<size_t cols, size_t rows>
14762 struct fp16Transpose : public fp16MatrixBase
14763 {
14764         virtual double getULPs(vector<const deFloat16*>& in)
14765         {
14766                 DE_UNREF(in);
14767
14768                 return 1.0;
14769         }
14770
14771         deUint32        getComponentValidity    ()
14772         {
14773                 return getComponentMatrixValidityMask(rows, cols);
14774         }
14775
14776         template<class fp16type>
14777         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
14778         {
14779                 DE_ASSERT(in.size() == 1);
14780
14781                 const size_t            alignedCols     = (cols == 3) ? 4 : cols;
14782                 const size_t            alignedRows     = (rows == 3) ? 4 : rows;
14783                 vector<deFloat16>       output          (alignedCols * alignedRows, 0);
14784
14785                 DE_ASSERT(output.size() == alignedCols * alignedRows);
14786
14787                 for (size_t rowNdx = 0; rowNdx < rows; ++rowNdx)
14788                         for (size_t colNdx = 0; colNdx < cols; ++colNdx)
14789                                 output[rowNdx * alignedCols + colNdx] = in[0][colNdx * alignedRows + rowNdx];
14790
14791                 deMemcpy(out, &output[0], sizeof(deFloat16) * output.size());
14792                 deMemcpy(min, &output[0], sizeof(deFloat16) * output.size());
14793                 deMemcpy(max, &output[0], sizeof(deFloat16) * output.size());
14794
14795                 return true;
14796         }
14797 };
14798
14799 template<size_t cols, size_t rows>
14800 struct fp16MatrixTimesScalar : public fp16MatrixBase
14801 {
14802         virtual double getULPs(vector<const deFloat16*>& in)
14803         {
14804                 DE_UNREF(in);
14805
14806                 return 4.0;
14807         }
14808
14809         deUint32        getComponentValidity    ()
14810         {
14811                 return getComponentMatrixValidityMask(cols, rows);
14812         }
14813
14814         template<class fp16type>
14815         bool calc(vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
14816         {
14817                 DE_ASSERT(in.size() == 2);
14818                 DE_ASSERT(getArgCompCount(1) == 1);
14819
14820                 const fp16type  y                       (in[1][0]);
14821                 const float             scalar          (y.asFloat());
14822                 const size_t    alignedCols     = (cols == 3) ? 4 : cols;
14823                 const size_t    alignedRows     = (rows == 3) ? 4 : rows;
14824
14825                 DE_ASSERT(getArgCompCount(0) == alignedCols * alignedRows);
14826                 DE_ASSERT(getOutCompCount() == alignedCols * alignedRows);
14827                 DE_UNREF(alignedCols);
14828
14829                 for (size_t rowNdx = 0; rowNdx < rows; ++rowNdx)
14830                         for (size_t colNdx = 0; colNdx < cols; ++colNdx)
14831                         {
14832                                 const size_t    ndx     (colNdx * alignedRows + rowNdx);
14833                                 const fp16type  x       (in[0][ndx]);
14834                                 const double    result  (scalar * x.asFloat());
14835
14836                                 out[ndx] = fp16type(result).bits();
14837                                 min[ndx] = getMin(result, getULPs(in));
14838                                 max[ndx] = getMax(result, getULPs(in));
14839                         }
14840
14841                 return true;
14842         }
14843 };
14844
14845 template<size_t cols, size_t rows>
14846 struct fp16VectorTimesMatrix : public fp16MatrixBase
14847 {
14848         fp16VectorTimesMatrix() : fp16MatrixBase()
14849         {
14850                 flavorNames.push_back("EmulatingFP16");
14851                 flavorNames.push_back("FloatCalc");
14852         }
14853
14854         virtual double getULPs (vector<const deFloat16*>& in)
14855         {
14856                 DE_UNREF(in);
14857
14858                 return (8.0 * cols);
14859         }
14860
14861         deUint32 getComponentValidity ()
14862         {
14863                 return getComponentMatrixValidityMask(cols, 1);
14864         }
14865
14866         template<class fp16type>
14867         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
14868         {
14869                 DE_ASSERT(in.size() == 2);
14870
14871                 const size_t    alignedCols     = (cols == 3) ? 4 : cols;
14872                 const size_t    alignedRows     = (rows == 3) ? 4 : rows;
14873
14874                 DE_ASSERT(getOutCompCount() == cols);
14875                 DE_ASSERT(getArgCompCount(0) == rows);
14876                 DE_ASSERT(getArgCompCount(1) == alignedCols * alignedRows);
14877                 DE_UNREF(alignedCols);
14878
14879                 if (getFlavor() == 0)
14880                 {
14881                         for (size_t colNdx = 0; colNdx < cols; ++colNdx)
14882                         {
14883                                 fp16type        s       (fp16type::zero(1));
14884
14885                                 for (size_t rowNdx = 0; rowNdx < rows; ++rowNdx)
14886                                 {
14887                                         const fp16type  v       (in[0][rowNdx]);
14888                                         const float             vf      (v.asFloat());
14889                                         const size_t    ndx     (colNdx * alignedRows + rowNdx);
14890                                         const fp16type  x       (in[1][ndx]);
14891                                         const float             xf      (x.asFloat());
14892                                         const fp16type  m       (vf * xf);
14893
14894                                         s = fp16type(s.asFloat() + m.asFloat());
14895                                 }
14896
14897                                 out[colNdx] = s.bits();
14898                                 min[colNdx] = getMin(s.asDouble(), getULPs(in));
14899                                 max[colNdx] = getMax(s.asDouble(), getULPs(in));
14900                         }
14901                 }
14902                 else if (getFlavor() == 1)
14903                 {
14904                         for (size_t colNdx = 0; colNdx < cols; ++colNdx)
14905                         {
14906                                 float   s       (0.0f);
14907
14908                                 for (size_t rowNdx = 0; rowNdx < rows; ++rowNdx)
14909                                 {
14910                                         const fp16type  v       (in[0][rowNdx]);
14911                                         const float             vf      (v.asFloat());
14912                                         const size_t    ndx     (colNdx * alignedRows + rowNdx);
14913                                         const fp16type  x       (in[1][ndx]);
14914                                         const float             xf      (x.asFloat());
14915                                         const float             m       (vf * xf);
14916
14917                                         s += m;
14918                                 }
14919
14920                                 out[colNdx] = fp16type(s).bits();
14921                                 min[colNdx] = getMin(static_cast<double>(s), getULPs(in));
14922                                 max[colNdx] = getMax(static_cast<double>(s), getULPs(in));
14923                         }
14924                 }
14925                 else
14926                 {
14927                         TCU_THROW(InternalError, "Unknown flavor");
14928                 }
14929
14930                 return true;
14931         }
14932 };
14933
14934 template<size_t cols, size_t rows>
14935 struct fp16MatrixTimesVector : public fp16MatrixBase
14936 {
14937         fp16MatrixTimesVector() : fp16MatrixBase()
14938         {
14939                 flavorNames.push_back("EmulatingFP16");
14940                 flavorNames.push_back("FloatCalc");
14941         }
14942
14943         virtual double getULPs (vector<const deFloat16*>& in)
14944         {
14945                 DE_UNREF(in);
14946
14947                 return (8.0 * rows);
14948         }
14949
14950         deUint32 getComponentValidity ()
14951         {
14952                 return getComponentMatrixValidityMask(rows, 1);
14953         }
14954
14955         template<class fp16type>
14956         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
14957         {
14958                 DE_ASSERT(in.size() == 2);
14959
14960                 const size_t    alignedCols     = (cols == 3) ? 4 : cols;
14961                 const size_t    alignedRows     = (rows == 3) ? 4 : rows;
14962
14963                 DE_ASSERT(getOutCompCount() == rows);
14964                 DE_ASSERT(getArgCompCount(0) == alignedCols * alignedRows);
14965                 DE_ASSERT(getArgCompCount(1) == cols);
14966                 DE_UNREF(alignedCols);
14967
14968                 if (getFlavor() == 0)
14969                 {
14970                         for (size_t rowNdx = 0; rowNdx < rows; ++rowNdx)
14971                         {
14972                                 fp16type        s       (fp16type::zero(1));
14973
14974                                 for (size_t colNdx = 0; colNdx < cols; ++colNdx)
14975                                 {
14976                                         const size_t    ndx     (colNdx * alignedRows + rowNdx);
14977                                         const fp16type  x       (in[0][ndx]);
14978                                         const float             xf      (x.asFloat());
14979                                         const fp16type  v       (in[1][colNdx]);
14980                                         const float             vf      (v.asFloat());
14981                                         const fp16type  m       (vf * xf);
14982
14983                                         s = fp16type(s.asFloat() + m.asFloat());
14984                                 }
14985
14986                                 out[rowNdx] = s.bits();
14987                                 min[rowNdx] = getMin(s.asDouble(), getULPs(in));
14988                                 max[rowNdx] = getMax(s.asDouble(), getULPs(in));
14989                         }
14990                 }
14991                 else if (getFlavor() == 1)
14992                 {
14993                         for (size_t rowNdx = 0; rowNdx < rows; ++rowNdx)
14994                         {
14995                                 float   s       (0.0f);
14996
14997                                 for (size_t colNdx = 0; colNdx < cols; ++colNdx)
14998                                 {
14999                                         const size_t    ndx     (colNdx * alignedRows + rowNdx);
15000                                         const fp16type  x       (in[0][ndx]);
15001                                         const float             xf      (x.asFloat());
15002                                         const fp16type  v       (in[1][colNdx]);
15003                                         const float             vf      (v.asFloat());
15004                                         const float             m       (vf * xf);
15005
15006                                         s += m;
15007                                 }
15008
15009                                 out[rowNdx] = fp16type(s).bits();
15010                                 min[rowNdx] = getMin(static_cast<double>(s), getULPs(in));
15011                                 max[rowNdx] = getMax(static_cast<double>(s), getULPs(in));
15012                         }
15013                 }
15014                 else
15015                 {
15016                         TCU_THROW(InternalError, "Unknown flavor");
15017                 }
15018
15019                 return true;
15020         }
15021 };
15022
15023 template<size_t colsL, size_t rowsL, size_t colsR, size_t rowsR>
15024 struct fp16MatrixTimesMatrix : public fp16MatrixBase
15025 {
15026         fp16MatrixTimesMatrix() : fp16MatrixBase()
15027         {
15028                 flavorNames.push_back("EmulatingFP16");
15029                 flavorNames.push_back("FloatCalc");
15030         }
15031
15032         virtual double getULPs (vector<const deFloat16*>& in)
15033         {
15034                 DE_UNREF(in);
15035
15036                 return 32.0;
15037         }
15038
15039         deUint32 getComponentValidity ()
15040         {
15041                 return getComponentMatrixValidityMask(colsR, rowsL);
15042         }
15043
15044         template<class fp16type>
15045         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
15046         {
15047                 DE_STATIC_ASSERT(colsL == rowsR);
15048
15049                 DE_ASSERT(in.size() == 2);
15050
15051                 const size_t    alignedColsL    = (colsL == 3) ? 4 : colsL;
15052                 const size_t    alignedRowsL    = (rowsL == 3) ? 4 : rowsL;
15053                 const size_t    alignedColsR    = (colsR == 3) ? 4 : colsR;
15054                 const size_t    alignedRowsR    = (rowsR == 3) ? 4 : rowsR;
15055
15056                 DE_ASSERT(getOutCompCount() == alignedColsR * alignedRowsL);
15057                 DE_ASSERT(getArgCompCount(0) == alignedColsL * alignedRowsL);
15058                 DE_ASSERT(getArgCompCount(1) == alignedColsR * alignedRowsR);
15059                 DE_UNREF(alignedColsL);
15060                 DE_UNREF(alignedColsR);
15061
15062                 if (getFlavor() == 0)
15063                 {
15064                         for (size_t rowNdx = 0; rowNdx < rowsL; ++rowNdx)
15065                         {
15066                                 for (size_t colNdx = 0; colNdx < colsR; ++colNdx)
15067                                 {
15068                                         const size_t    ndx     (colNdx * alignedRowsL + rowNdx);
15069                                         fp16type                s       (fp16type::zero(1));
15070
15071                                         for (size_t commonNdx = 0; commonNdx < colsL; ++commonNdx)
15072                                         {
15073                                                 const size_t    ndxl    (commonNdx * alignedRowsL + rowNdx);
15074                                                 const fp16type  l               (in[0][ndxl]);
15075                                                 const float             lf              (l.asFloat());
15076                                                 const size_t    ndxr    (colNdx * alignedRowsR + commonNdx);
15077                                                 const fp16type  r               (in[1][ndxr]);
15078                                                 const float             rf              (r.asFloat());
15079                                                 const fp16type  m               (lf * rf);
15080
15081                                                 s = fp16type(s.asFloat() + m.asFloat());
15082                                         }
15083
15084                                         out[ndx] = s.bits();
15085                                         min[ndx] = getMin(s.asDouble(), getULPs(in));
15086                                         max[ndx] = getMax(s.asDouble(), getULPs(in));
15087                                 }
15088                         }
15089                 }
15090                 else if (getFlavor() == 1)
15091                 {
15092                         for (size_t rowNdx = 0; rowNdx < rowsL; ++rowNdx)
15093                         {
15094                                 for (size_t colNdx = 0; colNdx < colsR; ++colNdx)
15095                                 {
15096                                         const size_t    ndx     (colNdx * alignedRowsL + rowNdx);
15097                                         float                   s       (0.0f);
15098
15099                                         for (size_t commonNdx = 0; commonNdx < colsL; ++commonNdx)
15100                                         {
15101                                                 const size_t    ndxl    (commonNdx * alignedRowsL + rowNdx);
15102                                                 const fp16type  l               (in[0][ndxl]);
15103                                                 const float             lf              (l.asFloat());
15104                                                 const size_t    ndxr    (colNdx * alignedRowsR + commonNdx);
15105                                                 const fp16type  r               (in[1][ndxr]);
15106                                                 const float             rf              (r.asFloat());
15107                                                 const float             m               (lf * rf);
15108
15109                                                 s += m;
15110                                         }
15111
15112                                         out[ndx] = fp16type(s).bits();
15113                                         min[ndx] = getMin(static_cast<double>(s), getULPs(in));
15114                                         max[ndx] = getMax(static_cast<double>(s), getULPs(in));
15115                                 }
15116                         }
15117                 }
15118                 else
15119                 {
15120                         TCU_THROW(InternalError, "Unknown flavor");
15121                 }
15122
15123                 return true;
15124         }
15125 };
15126
15127 template<size_t cols, size_t rows>
15128 struct fp16OuterProduct : public fp16MatrixBase
15129 {
15130         virtual double getULPs (vector<const deFloat16*>& in)
15131         {
15132                 DE_UNREF(in);
15133
15134                 return 2.0;
15135         }
15136
15137         deUint32 getComponentValidity ()
15138         {
15139                 return getComponentMatrixValidityMask(cols, rows);
15140         }
15141
15142         template<class fp16type>
15143         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
15144         {
15145                 DE_ASSERT(in.size() == 2);
15146
15147                 const size_t    alignedCols     = (cols == 3) ? 4 : cols;
15148                 const size_t    alignedRows     = (rows == 3) ? 4 : rows;
15149
15150                 DE_ASSERT(getArgCompCount(0) == rows);
15151                 DE_ASSERT(getArgCompCount(1) == cols);
15152                 DE_ASSERT(getOutCompCount() == alignedCols * alignedRows);
15153                 DE_UNREF(alignedCols);
15154
15155                 for (size_t rowNdx = 0; rowNdx < rows; ++rowNdx)
15156                 {
15157                         for (size_t colNdx = 0; colNdx < cols; ++colNdx)
15158                         {
15159                                 const size_t    ndx     (colNdx * alignedRows + rowNdx);
15160                                 const fp16type  x       (in[0][rowNdx]);
15161                                 const float             xf      (x.asFloat());
15162                                 const fp16type  y       (in[1][colNdx]);
15163                                 const float             yf      (y.asFloat());
15164                                 const fp16type  m       (xf * yf);
15165
15166                                 out[ndx] = m.bits();
15167                                 min[ndx] = getMin(m.asDouble(), getULPs(in));
15168                                 max[ndx] = getMax(m.asDouble(), getULPs(in));
15169                         }
15170                 }
15171
15172                 return true;
15173         }
15174 };
15175
15176 template<size_t size>
15177 struct fp16Determinant;
15178
15179 template<>
15180 struct fp16Determinant<2> : public fp16MatrixBase
15181 {
15182         virtual double getULPs (vector<const deFloat16*>& in)
15183         {
15184                 DE_UNREF(in);
15185
15186                 return 128.0; // This is not a precision test. Value is not from spec
15187         }
15188
15189         deUint32 getComponentValidity ()
15190         {
15191                 return 1;
15192         }
15193
15194         template<class fp16type>
15195         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
15196         {
15197                 const size_t    cols            = 2;
15198                 const size_t    rows            = 2;
15199                 const size_t    alignedCols     = (cols == 3) ? 4 : cols;
15200                 const size_t    alignedRows     = (rows == 3) ? 4 : rows;
15201
15202                 DE_ASSERT(in.size() == 1);
15203                 DE_ASSERT(getOutCompCount() == 1);
15204                 DE_ASSERT(getArgCompCount(0) == alignedRows * alignedCols);
15205                 DE_UNREF(alignedCols);
15206                 DE_UNREF(alignedRows);
15207
15208                 // [ a b ]
15209                 // [ c d ]
15210                 const float             a               (fp16type(in[0][getNdx(rows, 0, 0)]).asFloat());
15211                 const float             b               (fp16type(in[0][getNdx(rows, 1, 0)]).asFloat());
15212                 const float             c               (fp16type(in[0][getNdx(rows, 0, 1)]).asFloat());
15213                 const float             d               (fp16type(in[0][getNdx(rows, 1, 1)]).asFloat());
15214                 const float             ad              (a * d);
15215                 const fp16type  adf16   (ad);
15216                 const float             bc              (b * c);
15217                 const fp16type  bcf16   (bc);
15218                 const float             r               (adf16.asFloat() - bcf16.asFloat());
15219                 const fp16type  rf16    (r);
15220
15221                 out[0] = rf16.bits();
15222                 min[0] = getMin(r, getULPs(in));
15223                 max[0] = getMax(r, getULPs(in));
15224
15225                 return true;
15226         }
15227 };
15228
15229 template<>
15230 struct fp16Determinant<3> : public fp16MatrixBase
15231 {
15232         virtual double getULPs (vector<const deFloat16*>& in)
15233         {
15234                 DE_UNREF(in);
15235
15236                 return 128.0; // This is not a precision test. Value is not from spec
15237         }
15238
15239         deUint32 getComponentValidity ()
15240         {
15241                 return 1;
15242         }
15243
15244         template<class fp16type>
15245         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
15246         {
15247                 const size_t    cols            = 3;
15248                 const size_t    rows            = 3;
15249                 const size_t    alignedCols     = (cols == 3) ? 4 : cols;
15250                 const size_t    alignedRows     = (rows == 3) ? 4 : rows;
15251
15252                 DE_ASSERT(in.size() == 1);
15253                 DE_ASSERT(getOutCompCount() == 1);
15254                 DE_ASSERT(getArgCompCount(0) == alignedRows * alignedCols);
15255                 DE_UNREF(alignedCols);
15256                 DE_UNREF(alignedRows);
15257
15258                 // [ a b c ]
15259                 // [ d e f ]
15260                 // [ g h i ]
15261                 const float             a               (fp16type(in[0][getNdx(rows, 0, 0)]).asFloat());
15262                 const float             b               (fp16type(in[0][getNdx(rows, 1, 0)]).asFloat());
15263                 const float             c               (fp16type(in[0][getNdx(rows, 2, 0)]).asFloat());
15264                 const float             d               (fp16type(in[0][getNdx(rows, 0, 1)]).asFloat());
15265                 const float             e               (fp16type(in[0][getNdx(rows, 1, 1)]).asFloat());
15266                 const float             f               (fp16type(in[0][getNdx(rows, 2, 1)]).asFloat());
15267                 const float             g               (fp16type(in[0][getNdx(rows, 0, 2)]).asFloat());
15268                 const float             h               (fp16type(in[0][getNdx(rows, 1, 2)]).asFloat());
15269                 const float             i               (fp16type(in[0][getNdx(rows, 2, 2)]).asFloat());
15270                 const fp16type  aei             (a * e * i);
15271                 const fp16type  bfg             (b * f * g);
15272                 const fp16type  cdh             (c * d * h);
15273                 const fp16type  ceg             (c * e * g);
15274                 const fp16type  bdi             (b * d * i);
15275                 const fp16type  afh             (a * f * h);
15276                 const float             r               (aei.asFloat() + bfg.asFloat() + cdh.asFloat() - ceg.asFloat() - bdi.asFloat() - afh.asFloat());
15277                 const fp16type  rf16    (r);
15278
15279                 out[0] = rf16.bits();
15280                 min[0] = getMin(r, getULPs(in));
15281                 max[0] = getMax(r, getULPs(in));
15282
15283                 return true;
15284         }
15285 };
15286
15287 template<>
15288 struct fp16Determinant<4> : public fp16MatrixBase
15289 {
15290         virtual double getULPs (vector<const deFloat16*>& in)
15291         {
15292                 DE_UNREF(in);
15293
15294                 return 128.0; // This is not a precision test. Value is not from spec
15295         }
15296
15297         deUint32 getComponentValidity ()
15298         {
15299                 return 1;
15300         }
15301
15302         template<class fp16type>
15303         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
15304         {
15305                 const size_t    rows            = 4;
15306                 const size_t    cols            = 4;
15307                 const size_t    alignedCols     = (cols == 3) ? 4 : cols;
15308                 const size_t    alignedRows     = (rows == 3) ? 4 : rows;
15309
15310                 DE_ASSERT(in.size() == 1);
15311                 DE_ASSERT(getOutCompCount() == 1);
15312                 DE_ASSERT(getArgCompCount(0) == alignedRows * alignedCols);
15313                 DE_UNREF(alignedCols);
15314                 DE_UNREF(alignedRows);
15315
15316                 // [ a b c d ]
15317                 // [ e f g h ]
15318                 // [ i j k l ]
15319                 // [ m n o p ]
15320                 const float             a               (fp16type(in[0][getNdx(rows, 0, 0)]).asFloat());
15321                 const float             b               (fp16type(in[0][getNdx(rows, 1, 0)]).asFloat());
15322                 const float             c               (fp16type(in[0][getNdx(rows, 2, 0)]).asFloat());
15323                 const float             d               (fp16type(in[0][getNdx(rows, 3, 0)]).asFloat());
15324                 const float             e               (fp16type(in[0][getNdx(rows, 0, 1)]).asFloat());
15325                 const float             f               (fp16type(in[0][getNdx(rows, 1, 1)]).asFloat());
15326                 const float             g               (fp16type(in[0][getNdx(rows, 2, 1)]).asFloat());
15327                 const float             h               (fp16type(in[0][getNdx(rows, 3, 1)]).asFloat());
15328                 const float             i               (fp16type(in[0][getNdx(rows, 0, 2)]).asFloat());
15329                 const float             j               (fp16type(in[0][getNdx(rows, 1, 2)]).asFloat());
15330                 const float             k               (fp16type(in[0][getNdx(rows, 2, 2)]).asFloat());
15331                 const float             l               (fp16type(in[0][getNdx(rows, 3, 2)]).asFloat());
15332                 const float             m               (fp16type(in[0][getNdx(rows, 0, 3)]).asFloat());
15333                 const float             n               (fp16type(in[0][getNdx(rows, 1, 3)]).asFloat());
15334                 const float             o               (fp16type(in[0][getNdx(rows, 2, 3)]).asFloat());
15335                 const float             p               (fp16type(in[0][getNdx(rows, 3, 3)]).asFloat());
15336
15337                 // [ f g h ]
15338                 // [ j k l ]
15339                 // [ n o p ]
15340                 const fp16type  fkp             (f * k * p);
15341                 const fp16type  gln             (g * l * n);
15342                 const fp16type  hjo             (h * j * o);
15343                 const fp16type  hkn             (h * k * n);
15344                 const fp16type  gjp             (g * j * p);
15345                 const fp16type  flo             (f * l * o);
15346                 const fp16type  detA    (a * (fkp.asFloat() + gln.asFloat() + hjo.asFloat() - hkn.asFloat() - gjp.asFloat() - flo.asFloat()));
15347
15348                 // [ e g h ]
15349                 // [ i k l ]
15350                 // [ m o p ]
15351                 const fp16type  ekp             (e * k * p);
15352                 const fp16type  glm             (g * l * m);
15353                 const fp16type  hio             (h * i * o);
15354                 const fp16type  hkm             (h * k * m);
15355                 const fp16type  gip             (g * i * p);
15356                 const fp16type  elo             (e * l * o);
15357                 const fp16type  detB    (b * (ekp.asFloat() + glm.asFloat() + hio.asFloat() - hkm.asFloat() - gip.asFloat() - elo.asFloat()));
15358
15359                 // [ e f h ]
15360                 // [ i j l ]
15361                 // [ m n p ]
15362                 const fp16type  ejp             (e * j * p);
15363                 const fp16type  flm             (f * l * m);
15364                 const fp16type  hin             (h * i * n);
15365                 const fp16type  hjm             (h * j * m);
15366                 const fp16type  fip             (f * i * p);
15367                 const fp16type  eln             (e * l * n);
15368                 const fp16type  detC    (c * (ejp.asFloat() + flm.asFloat() + hin.asFloat() - hjm.asFloat() - fip.asFloat() - eln.asFloat()));
15369
15370                 // [ e f g ]
15371                 // [ i j k ]
15372                 // [ m n o ]
15373                 const fp16type  ejo             (e * j * o);
15374                 const fp16type  fkm             (f * k * m);
15375                 const fp16type  gin             (g * i * n);
15376                 const fp16type  gjm             (g * j * m);
15377                 const fp16type  fio             (f * i * o);
15378                 const fp16type  ekn             (e * k * n);
15379                 const fp16type  detD    (d * (ejo.asFloat() + fkm.asFloat() + gin.asFloat() - gjm.asFloat() - fio.asFloat() - ekn.asFloat()));
15380
15381                 const float             r               (detA.asFloat() - detB.asFloat() + detC.asFloat() - detD.asFloat());
15382                 const fp16type  rf16    (r);
15383
15384                 out[0] = rf16.bits();
15385                 min[0] = getMin(r, getULPs(in));
15386                 max[0] = getMax(r, getULPs(in));
15387
15388                 return true;
15389         }
15390 };
15391
15392 template<size_t size>
15393 struct fp16Inverse;
15394
15395 template<>
15396 struct fp16Inverse<2> : public fp16MatrixBase
15397 {
15398         virtual double getULPs (vector<const deFloat16*>& in)
15399         {
15400                 DE_UNREF(in);
15401
15402                 return 128.0; // This is not a precision test. Value is not from spec
15403         }
15404
15405         deUint32 getComponentValidity ()
15406         {
15407                 return getComponentMatrixValidityMask(2, 2);
15408         }
15409
15410         template<class fp16type>
15411         bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
15412         {
15413                 const size_t    cols            = 2;
15414                 const size_t    rows            = 2;
15415                 const size_t    alignedCols     = (cols == 3) ? 4 : cols;
15416                 const size_t    alignedRows     = (rows == 3) ? 4 : rows;
15417
15418                 DE_ASSERT(in.size() == 1);
15419                 DE_ASSERT(getOutCompCount() == alignedRows * alignedCols);
15420                 DE_ASSERT(getArgCompCount(0) == alignedRows * alignedCols);
15421                 DE_UNREF(alignedCols);
15422
15423                 // [ a b ]
15424                 // [ c d ]
15425                 const float             a               (fp16type(in[0][getNdx(rows, 0, 0)]).asFloat());
15426                 const float             b               (fp16type(in[0][getNdx(rows, 1, 0)]).asFloat());
15427                 const float             c               (fp16type(in[0][getNdx(rows, 0, 1)]).asFloat());
15428                 const float             d               (fp16type(in[0][getNdx(rows, 1, 1)]).asFloat());
15429                 const float             ad              (a * d);
15430                 const fp16type  adf16   (ad);
15431                 const float             bc              (b * c);
15432                 const fp16type  bcf16   (bc);
15433                 const float             det             (adf16.asFloat() - bcf16.asFloat());
15434                 const fp16type  det16   (det);
15435
15436                 out[0] = fp16type( d / det16.asFloat()).bits();
15437                 out[1] = fp16type(-c / det16.asFloat()).bits();
15438                 out[2] = fp16type(-b / det16.asFloat()).bits();
15439                 out[3] = fp16type( a / det16.asFloat()).bits();
15440
15441                 for (size_t rowNdx = 0; rowNdx < rows; ++rowNdx)
15442                         for (size_t colNdx = 0; colNdx < cols; ++colNdx)
15443                         {
15444                                 const size_t    ndx     (colNdx * alignedRows + rowNdx);
15445                                 const fp16type  s       (out[ndx]);
15446
15447                                 min[ndx] = getMin(s.asDouble(), getULPs(in));
15448                                 max[ndx] = getMax(s.asDouble(), getULPs(in));
15449                         }
15450
15451                 return true;
15452         }
15453 };
15454
15455 inline std::string fp16ToString(deFloat16 val)
15456 {
15457         return tcu::toHex<4>(val).toString() + " (" + de::floatToString(tcu::Float16(val).asFloat(), 10) + ")";
15458 }
15459
15460 template <size_t RES_COMPONENTS, size_t ARG0_COMPONENTS, size_t ARG1_COMPONENTS, size_t ARG2_COMPONENTS, class TestedArithmeticFunction>
15461 bool compareFP16ArithmeticFunc (const std::vector<Resource>& inputs, const vector<AllocationSp>& outputAllocs, const std::vector<Resource>& expectedOutputs, TestLog& log)
15462 {
15463         if (inputs.size() < 1 || inputs.size() > 3 || outputAllocs.size() != 1 || expectedOutputs.size() != 1)
15464                 return false;
15465
15466         const size_t    resultStep                      = (RES_COMPONENTS == 3) ? 4 : RES_COMPONENTS;
15467         const size_t    iterationsCount         = expectedOutputs[0].getByteSize() / (sizeof(deFloat16) * resultStep);
15468         const size_t    inputsSteps[3]          =
15469         {
15470                 (ARG0_COMPONENTS == 3) ? 4 : ARG0_COMPONENTS,
15471                 (ARG1_COMPONENTS == 3) ? 4 : ARG1_COMPONENTS,
15472                 (ARG2_COMPONENTS == 3) ? 4 : ARG2_COMPONENTS,
15473         };
15474
15475         DE_ASSERT(expectedOutputs[0].getByteSize() > 0);
15476         DE_ASSERT(expectedOutputs[0].getByteSize() == sizeof(deFloat16) * iterationsCount * resultStep);
15477
15478         for (size_t inputNdx = 0; inputNdx < inputs.size(); ++inputNdx)
15479         {
15480                 DE_ASSERT(inputs[inputNdx].getByteSize() > 0);
15481                 DE_ASSERT(inputs[inputNdx].getByteSize() == sizeof(deFloat16) * iterationsCount * inputsSteps[inputNdx]);
15482         }
15483
15484         const deFloat16* const          outputAsFP16                                    = (const deFloat16*)outputAllocs[0]->getHostPtr();
15485         TestedArithmeticFunction        func;
15486
15487         func.setOutCompCount(RES_COMPONENTS);
15488         func.setArgCompCount(0, ARG0_COMPONENTS);
15489         func.setArgCompCount(1, ARG1_COMPONENTS);
15490         func.setArgCompCount(2, ARG2_COMPONENTS);
15491
15492         const bool                                      callOncePerComponent                    = func.callOncePerComponent();
15493         const deUint32                          componentValidityMask                   = func.getComponentValidity();
15494         const size_t                            denormModesCount                                = 2;
15495         const char*                                     denormModes[denormModesCount]   = { "keep denormal numbers", "flush to zero" };
15496         const size_t                            successfulRunsPerComponent              = denormModesCount * func.getFlavorCount();
15497         bool                                            success                                                 = true;
15498         size_t                                          validatedCount                                  = 0;
15499
15500         vector<deUint8> inputBytes[3];
15501
15502         for (size_t inputNdx = 0; inputNdx < inputs.size(); ++inputNdx)
15503                 inputs[inputNdx].getBytes(inputBytes[inputNdx]);
15504
15505         const deFloat16* const                  inputsAsFP16[3]                 =
15506         {
15507                 inputs.size() >= 1 ? (const deFloat16*)&inputBytes[0][0] : DE_NULL,
15508                 inputs.size() >= 2 ? (const deFloat16*)&inputBytes[1][0] : DE_NULL,
15509                 inputs.size() >= 3 ? (const deFloat16*)&inputBytes[2][0] : DE_NULL,
15510         };
15511
15512         for (size_t idx = 0; idx < iterationsCount; ++idx)
15513         {
15514                 std::vector<size_t>                     successfulRuns          (RES_COMPONENTS, successfulRunsPerComponent);
15515                 std::vector<std::string>        errors                          (RES_COMPONENTS);
15516                 bool                                            iterationValidated      (true);
15517
15518                 for (size_t denormNdx = 0; denormNdx < 2; ++denormNdx)
15519                 {
15520                         for (size_t flavorNdx = 0; flavorNdx < func.getFlavorCount(); ++flavorNdx)
15521                         {
15522                                 func.setFlavor(flavorNdx);
15523
15524                                 const deFloat16*                        iterationOutputFP16             = &outputAsFP16[idx * resultStep];
15525                                 vector<deFloat16>                       iterationCalculatedFP16 (resultStep, 0);
15526                                 vector<double>                          iterationEdgeMin                (resultStep, 0.0);
15527                                 vector<double>                          iterationEdgeMax                (resultStep, 0.0);
15528                                 vector<const deFloat16*>        arguments;
15529
15530                                 for (size_t componentNdx = 0; componentNdx < RES_COMPONENTS; ++componentNdx)
15531                                 {
15532                                         std::string     error;
15533                                         bool            reportError = false;
15534
15535                                         if (callOncePerComponent || componentNdx == 0)
15536                                         {
15537                                                 bool funcCallResult;
15538
15539                                                 arguments.clear();
15540
15541                                                 for (size_t inputNdx = 0; inputNdx < inputs.size(); ++inputNdx)
15542                                                         arguments.push_back(&inputsAsFP16[inputNdx][idx * inputsSteps[inputNdx] + componentNdx]);
15543
15544                                                 if (denormNdx == 0)
15545                                                         funcCallResult = func.template calc<tcu::Float16>(arguments, &iterationCalculatedFP16[componentNdx], &iterationEdgeMin[componentNdx], &iterationEdgeMax[componentNdx]);
15546                                                 else
15547                                                         funcCallResult = func.template calc<tcu::Float16Denormless>(arguments, &iterationCalculatedFP16[componentNdx], &iterationEdgeMin[componentNdx], &iterationEdgeMax[componentNdx]);
15548
15549                                                 if (!funcCallResult)
15550                                                 {
15551                                                         iterationValidated = false;
15552
15553                                                         if (callOncePerComponent)
15554                                                                 continue;
15555                                                         else
15556                                                                 break;
15557                                                 }
15558                                         }
15559
15560                                         if ((componentValidityMask != 0) && (componentValidityMask & (1<<componentNdx)) == 0)
15561                                                 continue;
15562
15563                                         reportError = !compare16BitFloat(iterationCalculatedFP16[componentNdx], iterationOutputFP16[componentNdx], error);
15564
15565                                         if (reportError)
15566                                         {
15567                                                 tcu::Float16 expected   (iterationCalculatedFP16[componentNdx]);
15568                                                 tcu::Float16 outputted  (iterationOutputFP16[componentNdx]);
15569
15570                                                 if (reportError && expected.isNaN())
15571                                                         reportError = false;
15572
15573                                                 if (reportError && !expected.isNaN() && !outputted.isNaN())
15574                                                 {
15575                                                         if (reportError && !expected.isInf() && !outputted.isInf())
15576                                                         {
15577                                                                 // Ignore rounding
15578                                                                 if (expected.bits() == outputted.bits() + 1 || expected.bits() + 1 == outputted.bits())
15579                                                                         reportError = false;
15580                                                         }
15581
15582                                                         if (reportError && expected.isInf())
15583                                                         {
15584                                                                 // RTZ rounding mode returns +/-65504 instead of Inf on overflow
15585                                                                 if (expected.sign() == 1 && outputted.bits() == 0x7bff && iterationEdgeMin[componentNdx] <= std::numeric_limits<double>::max())
15586                                                                         reportError = false;
15587                                                                 else if (expected.sign() == -1 && outputted.bits() == 0xfbff && iterationEdgeMax[componentNdx] >= -std::numeric_limits<double>::max())
15588                                                                         reportError = false;
15589                                                         }
15590
15591                                                         if (reportError)
15592                                                         {
15593                                                                 const double    outputtedDouble = outputted.asDouble();
15594
15595                                                                 DE_ASSERT(iterationEdgeMin[componentNdx] <= iterationEdgeMax[componentNdx]);
15596
15597                                                                 if (de::inRange(outputtedDouble, iterationEdgeMin[componentNdx], iterationEdgeMax[componentNdx]))
15598                                                                         reportError = false;
15599                                                         }
15600                                                 }
15601
15602                                                 if (reportError)
15603                                                 {
15604                                                         const size_t            inputsComps[3]  =
15605                                                         {
15606                                                                 ARG0_COMPONENTS,
15607                                                                 ARG1_COMPONENTS,
15608                                                                 ARG2_COMPONENTS,
15609                                                         };
15610                                                         string                          inputsValues    ("Inputs:");
15611                                                         string                          flavorName              (func.getFlavorCount() == 1 ? "" : string(" flavor ") + de::toString(flavorNdx) + " (" + func.getCurrentFlavorName() + ")");
15612                                                         std::stringstream       errStream;
15613
15614                                                         for (size_t inputNdx = 0; inputNdx < inputs.size(); ++inputNdx)
15615                                                         {
15616                                                                 const size_t    inputCompsCount = inputsComps[inputNdx];
15617
15618                                                                 inputsValues += " [" + de::toString(inputNdx) + "]=(";
15619
15620                                                                 for (size_t compNdx = 0; compNdx < inputCompsCount; ++compNdx)
15621                                                                 {
15622                                                                         const deFloat16 inputComponentValue = inputsAsFP16[inputNdx][idx * inputsSteps[inputNdx] + compNdx];
15623
15624                                                                         inputsValues += fp16ToString(inputComponentValue) + ((compNdx + 1 == inputCompsCount) ? ")": ", ");
15625                                                                 }
15626                                                         }
15627
15628                                                         errStream       << "At"
15629                                                                                 << " iteration " << de::toString(idx)
15630                                                                                 << " component " << de::toString(componentNdx)
15631                                                                                 << " denormMode " << de::toString(denormNdx)
15632                                                                                 << " (" << denormModes[denormNdx] << ")"
15633                                                                                 << " " << flavorName
15634                                                                                 << " " << inputsValues
15635                                                                                 << " outputted:" + fp16ToString(iterationOutputFP16[componentNdx])
15636                                                                                 << " expected:" + fp16ToString(iterationCalculatedFP16[componentNdx])
15637                                                                                 << " or in range: [" << iterationEdgeMin[componentNdx] << ", " << iterationEdgeMax[componentNdx] << "]."
15638                                                                                 << " " << error << "."
15639                                                                                 << std::endl;
15640
15641                                                         errors[componentNdx] += errStream.str();
15642
15643                                                         successfulRuns[componentNdx]--;
15644                                                 }
15645                                         }
15646                                 }
15647                         }
15648                 }
15649
15650                 for (size_t componentNdx = 0; componentNdx < RES_COMPONENTS; ++componentNdx)
15651                 {
15652                         // Check if any component has total failure
15653                         if (successfulRuns[componentNdx] == 0)
15654                         {
15655                                 // Test failed in all denorm modes and all flavors for certain component: dump errors
15656                                 log << TestLog::Message << errors[componentNdx] << TestLog::EndMessage;
15657
15658                                 success = false;
15659                         }
15660                 }
15661
15662                 if (iterationValidated)
15663                         validatedCount++;
15664         }
15665
15666         if (validatedCount < 16)
15667                 TCU_THROW(InternalError, "Too few samples has been validated.");
15668
15669         return success;
15670 }
15671
15672 // IEEE-754 floating point numbers:
15673 // +--------+------+----------+-------------+
15674 // | binary | sign | exponent | significand |
15675 // +--------+------+----------+-------------+
15676 // | 16-bit |  1   |    5     |     10      |
15677 // +--------+------+----------+-------------+
15678 // | 32-bit |  1   |    8     |     23      |
15679 // +--------+------+----------+-------------+
15680 //
15681 // 16-bit floats:
15682 //
15683 // 0   000 00   00 0000 0001 (0x0001: 2e-24:         minimum positive denormalized)
15684 // 0   000 00   11 1111 1111 (0x03ff: 2e-14 - 2e-24: maximum positive denormalized)
15685 // 0   000 01   00 0000 0000 (0x0400: 2e-14:         minimum positive normalized)
15686 // 0   111 10   11 1111 1111 (0x7bff: 65504:         maximum positive normalized)
15687 //
15688 // 0   000 00   00 0000 0000 (0x0000: +0)
15689 // 0   111 11   00 0000 0000 (0x7c00: +Inf)
15690 // 0   000 00   11 1111 0000 (0x03f0: +Denorm)
15691 // 0   000 01   00 0000 0001 (0x0401: +Norm)
15692 // 0   111 11   00 0000 1111 (0x7c0f: +SNaN)
15693 // 0   111 11   11 1111 0000 (0x7ff0: +QNaN)
15694 // Generate and return 16-bit floats and their corresponding 32-bit values.
15695 //
15696 // The first 14 number pairs are manually picked, while the rest are randomly generated.
15697 // Expected count to be at least 14 (numPicks).
15698 vector<deFloat16> getFloat16a (de::Random& rnd, deUint32 count)
15699 {
15700         vector<deFloat16>       float16;
15701
15702         float16.reserve(count);
15703
15704         // Zero
15705         float16.push_back(deUint16(0x0000));
15706         float16.push_back(deUint16(0x8000));
15707         // Infinity
15708         float16.push_back(deUint16(0x7c00));
15709         float16.push_back(deUint16(0xfc00));
15710         // Normalized
15711         float16.push_back(deUint16(0x0401));
15712         float16.push_back(deUint16(0x8401));
15713         // Some normal number
15714         float16.push_back(deUint16(0x14cb));
15715         float16.push_back(deUint16(0x94cb));
15716         // Min/max positive normal
15717         float16.push_back(deUint16(0x0400));
15718         float16.push_back(deUint16(0x7bff));
15719         // Min/max negative normal
15720         float16.push_back(deUint16(0x8400));
15721         float16.push_back(deUint16(0xfbff));
15722         // PI
15723         float16.push_back(deUint16(0x4248)); // 3.140625
15724         float16.push_back(deUint16(0xb248)); // -3.140625
15725         // PI/2
15726         float16.push_back(deUint16(0x3e48)); // 1.5703125
15727         float16.push_back(deUint16(0xbe48)); // -1.5703125
15728         float16.push_back(deUint16(0x3c00)); // 1.0
15729         float16.push_back(deUint16(0x3800)); // 0.5
15730         // Some useful constants
15731         float16.push_back(tcu::Float16(-2.5f).bits());
15732         float16.push_back(tcu::Float16(-1.0f).bits());
15733         float16.push_back(tcu::Float16( 0.4f).bits());
15734         float16.push_back(tcu::Float16( 2.5f).bits());
15735
15736         const deUint32          numPicks        = static_cast<deUint32>(float16.size());
15737
15738         DE_ASSERT(count >= numPicks);
15739         count -= numPicks;
15740
15741         for (deUint32 numIdx = 0; numIdx < count; ++numIdx)
15742         {
15743                 int                     sign            = (rnd.getUint16() % 2 == 0) ? +1 : -1;
15744                 int                     exponent        = (rnd.getUint16() % 29) - 14 + 1;
15745                 deUint16        mantissa        = static_cast<deUint16>(2 * (rnd.getUint16() % 512));
15746
15747                 // Exclude power of -14 to avoid denorms
15748                 DE_ASSERT(de::inRange(exponent, -13, 15));
15749
15750                 float16.push_back(tcu::Float16::constructBits(sign, exponent, mantissa).bits());
15751         }
15752
15753         return float16;
15754 }
15755
15756 static inline vector<deFloat16> getInputData1 (deUint32 seed, size_t count, size_t argNo)
15757 {
15758         DE_UNREF(argNo);
15759
15760         de::Random      rnd(seed);
15761
15762         return getFloat16a(rnd, static_cast<deUint32>(count));
15763 }
15764
15765 static inline vector<deFloat16> getInputData2 (deUint32 seed, size_t count, size_t argNo)
15766 {
15767         de::Random      rnd             (seed);
15768         size_t          newCount = static_cast<size_t>(deSqrt(double(count)));
15769
15770         DE_ASSERT(newCount * newCount == count);
15771
15772         vector<deFloat16>       float16 = getFloat16a(rnd, static_cast<deUint32>(newCount));
15773
15774         return squarize(float16, static_cast<deUint32>(argNo));
15775 }
15776
15777 static inline vector<deFloat16> getInputData3 (deUint32 seed, size_t count, size_t argNo)
15778 {
15779         if (argNo == 0 || argNo == 1)
15780                 return getInputData2(seed, count, argNo);
15781         else
15782                 return getInputData1(seed<<argNo, count, argNo);
15783 }
15784
15785 vector<deFloat16> getInputData (deUint32 seed, size_t count, size_t compCount, size_t stride, size_t argCount, size_t argNo)
15786 {
15787         DE_UNREF(stride);
15788
15789         vector<deFloat16>       result;
15790
15791         switch (argCount)
15792         {
15793                 case 1:result = getInputData1(seed, count, argNo); break;
15794                 case 2:result = getInputData2(seed, count, argNo); break;
15795                 case 3:result = getInputData3(seed, count, argNo); break;
15796                 default: TCU_THROW(InternalError, "Invalid argument count specified");
15797         }
15798
15799         if (compCount == 3)
15800         {
15801                 const size_t            newCount = (3 * count) / 4;
15802                 vector<deFloat16>       newResult;
15803
15804                 newResult.reserve(result.size());
15805
15806                 for (size_t ndx = 0; ndx < newCount; ++ndx)
15807                 {
15808                         newResult.push_back(result[ndx]);
15809
15810                         if (ndx % 3 == 2)
15811                                 newResult.push_back(0);
15812                 }
15813
15814                 result = newResult;
15815         }
15816
15817         DE_ASSERT(result.size() == count);
15818
15819         return result;
15820 }
15821
15822 // Generator for functions requiring data in range [1, inf]
15823 vector<deFloat16> getInputDataAC (deUint32 seed, size_t count, size_t compCount, size_t stride, size_t argCount, size_t argNo)
15824 {
15825         vector<deFloat16>       result;
15826
15827         result = getInputData(seed, count, compCount, stride, argCount, argNo);
15828
15829         // Filter out values below 1.0 from upper half of numbers
15830         for (size_t idx = result.size() / 2; idx < result.size(); ++idx)
15831         {
15832                 const float f = tcu::Float16(result[idx]).asFloat();
15833
15834                 if (f < 1.0f)
15835                         result[idx] = tcu::Float16(1.0f - f).bits();
15836         }
15837
15838         return result;
15839 }
15840
15841 // Generator for functions requiring data in range [-1, 1]
15842 vector<deFloat16> getInputDataA (deUint32 seed, size_t count, size_t compCount, size_t stride, size_t argCount, size_t argNo)
15843 {
15844         vector<deFloat16>       result;
15845
15846         result = getInputData(seed, count, compCount, stride, argCount, argNo);
15847
15848         for (size_t idx = result.size() / 2; idx < result.size(); ++idx)
15849         {
15850                 const float f = tcu::Float16(result[idx]).asFloat();
15851
15852                 if (!de::inRange(f, -1.0f, 1.0f))
15853                         result[idx] = tcu::Float16(deFloatFrac(f)).bits();
15854         }
15855
15856         return result;
15857 }
15858
15859 // Generator for functions requiring data in range [-pi, pi]
15860 vector<deFloat16> getInputDataPI (deUint32 seed, size_t count, size_t compCount, size_t stride, size_t argCount, size_t argNo)
15861 {
15862         vector<deFloat16>       result;
15863
15864         result = getInputData(seed, count, compCount, stride, argCount, argNo);
15865
15866         for (size_t idx = result.size() / 2; idx < result.size(); ++idx)
15867         {
15868                 const float f = tcu::Float16(result[idx]).asFloat();
15869
15870                 if (!de::inRange(f, -DE_PI, DE_PI))
15871                         result[idx] = tcu::Float16(fmodf(f, DE_PI)).bits();
15872         }
15873
15874         return result;
15875 }
15876
15877 // Generator for functions requiring data in range [0, inf]
15878 vector<deFloat16> getInputDataP (deUint32 seed, size_t count, size_t compCount, size_t stride, size_t argCount, size_t argNo)
15879 {
15880         vector<deFloat16>       result;
15881
15882         result = getInputData(seed, count, compCount, stride, argCount, argNo);
15883
15884         if (argNo == 0)
15885         {
15886                 for (size_t idx = result.size() / 2; idx < result.size(); ++idx)
15887                         result[idx] &= static_cast<deFloat16>(~0x8000);
15888         }
15889
15890         return result;
15891 }
15892
15893 vector<deFloat16> getInputDataV (deUint32 seed, size_t count, size_t compCount, size_t stride, size_t argCount, size_t argNo)
15894 {
15895         DE_UNREF(stride);
15896         DE_UNREF(argCount);
15897
15898         vector<deFloat16>       result;
15899
15900         if (argNo == 0)
15901                 result = getInputData2(seed, count, argNo);
15902         else
15903         {
15904                 const size_t            alignedCount    = (compCount == 3) ? 4 : compCount;
15905                 const size_t            newCountX               = static_cast<size_t>(deSqrt(double(count * alignedCount)));
15906                 const size_t            newCountY               = count / newCountX;
15907                 de::Random                      rnd                             (seed);
15908                 vector<deFloat16>       float16                 = getFloat16a(rnd, static_cast<deUint32>(newCountX));
15909
15910                 DE_ASSERT(newCountX * newCountX == alignedCount * count);
15911
15912                 for (size_t numIdx = 0; numIdx < newCountX; ++numIdx)
15913                 {
15914                         const vector<deFloat16> tmp(newCountY, float16[numIdx]);
15915
15916                         result.insert(result.end(), tmp.begin(), tmp.end());
15917                 }
15918         }
15919
15920         DE_ASSERT(result.size() == count);
15921
15922         return result;
15923 }
15924
15925 vector<deFloat16> getInputDataM (deUint32 seed, size_t count, size_t compCount, size_t stride, size_t argCount, size_t argNo)
15926 {
15927         DE_UNREF(compCount);
15928         DE_UNREF(stride);
15929         DE_UNREF(argCount);
15930
15931         de::Random                      rnd             (seed << argNo);
15932         vector<deFloat16>       result;
15933
15934         result = getFloat16a(rnd, static_cast<deUint32>(count));
15935
15936         DE_ASSERT(result.size() == count);
15937
15938         return result;
15939 }
15940
15941 vector<deFloat16> getInputDataD (deUint32 seed, size_t count, size_t compCount, size_t stride, size_t argCount, size_t argNo)
15942 {
15943         DE_UNREF(compCount);
15944         DE_UNREF(argCount);
15945
15946         de::Random                      rnd             (seed << argNo);
15947         vector<deFloat16>       result;
15948
15949         for (deUint32 numIdx = 0; numIdx < count; ++numIdx)
15950         {
15951                 int num = (rnd.getUint16() % 16) - 8;
15952
15953                 result.push_back(tcu::Float16(float(num)).bits());
15954         }
15955
15956         result[0 * stride] = deUint16(0x7c00); // +Inf
15957         result[1 * stride] = deUint16(0xfc00); // -Inf
15958
15959         DE_ASSERT(result.size() == count);
15960
15961         return result;
15962 }
15963
15964 // Generator for smoothstep function
15965 vector<deFloat16> getInputDataSS (deUint32 seed, size_t count, size_t compCount, size_t stride, size_t argCount, size_t argNo)
15966 {
15967         vector<deFloat16>       result;
15968
15969         result = getInputDataD(seed, count, compCount, stride, argCount, argNo);
15970
15971         if (argNo == 0)
15972         {
15973                 for (size_t idx = result.size() / 2; idx < result.size(); ++idx)
15974                 {
15975                         const float f = tcu::Float16(result[idx]).asFloat();
15976
15977                         if (f > 4.0f)
15978                                 result[idx] = tcu::Float16(-f).bits();
15979                 }
15980         }
15981
15982         if (argNo == 1)
15983         {
15984                 for (size_t idx = result.size() / 2; idx < result.size(); ++idx)
15985                 {
15986                         const float f = tcu::Float16(result[idx]).asFloat();
15987
15988                         if (f < 4.0f)
15989                                 result[idx] = tcu::Float16(-f).bits();
15990                 }
15991         }
15992
15993         return result;
15994 }
15995
15996 // Generates normalized vectors for arguments 0 and 1
15997 vector<deFloat16> getInputDataN (deUint32 seed, size_t count, size_t compCount, size_t stride, size_t argCount, size_t argNo)
15998 {
15999         DE_UNREF(compCount);
16000         DE_UNREF(argCount);
16001
16002         de::Random                      rnd             (seed << argNo);
16003         vector<deFloat16>       result;
16004
16005         if (argNo == 0 || argNo == 1)
16006         {
16007                 // The input parameters for the incident vector I and the surface normal N must already be normalized
16008                 for (size_t numIdx = 0; numIdx < count; numIdx += stride)
16009                 {
16010                         vector <float>  unnormolized;
16011                         float                   sum                             = 0;
16012
16013                         for (size_t compIdx = 0; compIdx < compCount; ++compIdx)
16014                                 unnormolized.push_back(float((rnd.getUint16() % 16) - 8));
16015
16016                         for (size_t compIdx = 0; compIdx < compCount; ++compIdx)
16017                                 sum += unnormolized[compIdx] * unnormolized[compIdx];
16018
16019                         sum = deFloatSqrt(sum);
16020                         if (sum == 0.0f)
16021                                 unnormolized[0] = sum = 1.0f;
16022
16023                         for (size_t compIdx = 0; compIdx < compCount; ++compIdx)
16024                                 result.push_back(tcu::Float16(unnormolized[compIdx] / sum).bits());
16025
16026                         for (size_t compIdx = compCount; compIdx < stride; ++compIdx)
16027                                 result.push_back(0);
16028                 }
16029         }
16030         else
16031         {
16032                 // Input parameter eta
16033                 for (deUint32 numIdx = 0; numIdx < count; ++numIdx)
16034                 {
16035                         int num = (rnd.getUint16() % 16) - 8;
16036
16037                         result.push_back(tcu::Float16(float(num)).bits());
16038                 }
16039         }
16040
16041         DE_ASSERT(result.size() == count);
16042
16043         return result;
16044 }
16045
16046 // Data generator for complex matrix functions like determinant and inverse
16047 vector<deFloat16> getInputDataC (deUint32 seed, size_t count, size_t compCount, size_t stride, size_t argCount, size_t argNo)
16048 {
16049         DE_UNREF(compCount);
16050         DE_UNREF(stride);
16051         DE_UNREF(argCount);
16052
16053         de::Random                      rnd             (seed << argNo);
16054         vector<deFloat16>       result;
16055
16056         for (deUint32 numIdx = 0; numIdx < count; ++numIdx)
16057         {
16058                 int num = (rnd.getUint16() % 16) - 8;
16059
16060                 result.push_back(tcu::Float16(float(num)).bits());
16061         }
16062
16063         DE_ASSERT(result.size() == count);
16064
16065         return result;
16066 }
16067
16068 struct Math16TestType
16069 {
16070         const char*             typePrefix;
16071         const size_t    typeComponents;
16072         const size_t    typeArrayStride;
16073         const size_t    typeStructStride;
16074 };
16075
16076 enum Math16DataTypes
16077 {
16078         NONE    = 0,
16079         SCALAR  = 1,
16080         VEC2    = 2,
16081         VEC3    = 3,
16082         VEC4    = 4,
16083         MAT2X2,
16084         MAT2X3,
16085         MAT2X4,
16086         MAT3X2,
16087         MAT3X3,
16088         MAT3X4,
16089         MAT4X2,
16090         MAT4X3,
16091         MAT4X4,
16092         MATH16_TYPE_LAST
16093 };
16094
16095 struct Math16ArgFragments
16096 {
16097         const char*     bodies;
16098         const char*     variables;
16099         const char*     decorations;
16100         const char*     funcVariables;
16101 };
16102
16103 typedef vector<deFloat16> Math16GetInputData (deUint32 seed, size_t count, size_t compCount, size_t stride, size_t argCount, size_t argNo);
16104
16105 struct Math16TestFunc
16106 {
16107         const char*                                     funcName;
16108         const char*                                     funcSuffix;
16109         size_t                                          funcArgsCount;
16110         size_t                                          typeResult;
16111         size_t                                          typeArg0;
16112         size_t                                          typeArg1;
16113         size_t                                          typeArg2;
16114         Math16GetInputData*                     getInputDataFunc;
16115         VerifyIOFunc                            verifyFunc;
16116 };
16117
16118 template<class SpecResource>
16119 void createFloat16ArithmeticFuncTest (tcu::TestContext& testCtx, tcu::TestCaseGroup& testGroup, const size_t testTypeIdx, const Math16TestFunc& testFunc)
16120 {
16121         const int                                       testSpecificSeed                        = deStringHash(testGroup.getName());
16122         const int                                       seed                                            = testCtx.getCommandLine().getBaseSeed() ^ testSpecificSeed;
16123         const size_t                            numDataPointsByAxis                     = 32;
16124         const size_t                            numDataPoints                           = numDataPointsByAxis * numDataPointsByAxis;
16125         const char*                                     componentType                           = "f16";
16126         const Math16TestType            testTypes[MATH16_TYPE_LAST]     =
16127         {
16128                 { "",           0,       0,                                              0,                                             },
16129                 { "",           1,       1 * sizeof(deFloat16),  2 * sizeof(deFloat16)  },
16130                 { "v2",         2,       2 * sizeof(deFloat16),  2 * sizeof(deFloat16)  },
16131                 { "v3",         3,       4 * sizeof(deFloat16),  4 * sizeof(deFloat16)  },
16132                 { "v4",         4,       4 * sizeof(deFloat16),  4 * sizeof(deFloat16)  },
16133                 { "m2x2",       0,       4 * sizeof(deFloat16),  4 * sizeof(deFloat16)  },
16134                 { "m2x3",       0,       8 * sizeof(deFloat16),  8 * sizeof(deFloat16)  },
16135                 { "m2x4",       0,       8 * sizeof(deFloat16),  8 * sizeof(deFloat16)  },
16136                 { "m3x2",       0,       8 * sizeof(deFloat16),  8 * sizeof(deFloat16)  },
16137                 { "m3x3",       0,      16 * sizeof(deFloat16), 16 * sizeof(deFloat16)  },
16138                 { "m3x4",       0,      16 * sizeof(deFloat16), 16 * sizeof(deFloat16)  },
16139                 { "m4x2",       0,       8 * sizeof(deFloat16),  8 * sizeof(deFloat16)  },
16140                 { "m4x3",       0,      16 * sizeof(deFloat16), 16 * sizeof(deFloat16)  },
16141                 { "m4x4",       0,      16 * sizeof(deFloat16), 16 * sizeof(deFloat16)  },
16142         };
16143
16144         DE_ASSERT(testTypeIdx == testTypes[testTypeIdx].typeComponents);
16145
16146
16147         const StringTemplate preMain
16148         (
16149                 "     %c_i32_ndp  = OpConstant %i32 ${num_data_points}\n"
16150
16151                 "        %f16     = OpTypeFloat 16\n"
16152                 "        %v2f16   = OpTypeVector %f16 2\n"
16153                 "        %v3f16   = OpTypeVector %f16 3\n"
16154                 "        %v4f16   = OpTypeVector %f16 4\n"
16155                 "        %m2x2f16 = OpTypeMatrix %v2f16 2\n"
16156                 "        %m2x3f16 = OpTypeMatrix %v3f16 2\n"
16157                 "        %m2x4f16 = OpTypeMatrix %v4f16 2\n"
16158                 "        %m3x2f16 = OpTypeMatrix %v2f16 3\n"
16159                 "        %m3x3f16 = OpTypeMatrix %v3f16 3\n"
16160                 "        %m3x4f16 = OpTypeMatrix %v4f16 3\n"
16161                 "        %m4x2f16 = OpTypeMatrix %v2f16 4\n"
16162                 "        %m4x3f16 = OpTypeMatrix %v3f16 4\n"
16163                 "        %m4x4f16 = OpTypeMatrix %v4f16 4\n"
16164
16165                 "     %up_f16     = OpTypePointer Uniform %f16    \n"
16166                 "     %up_v2f16   = OpTypePointer Uniform %v2f16  \n"
16167                 "     %up_v3f16   = OpTypePointer Uniform %v3f16  \n"
16168                 "     %up_v4f16   = OpTypePointer Uniform %v4f16  \n"
16169                 "     %up_m2x2f16 = OpTypePointer Uniform %m2x2f16\n"
16170                 "     %up_m2x3f16 = OpTypePointer Uniform %m2x3f16\n"
16171                 "     %up_m2x4f16 = OpTypePointer Uniform %m2x4f16\n"
16172                 "     %up_m3x2f16 = OpTypePointer Uniform %m3x2f16\n"
16173                 "     %up_m3x3f16 = OpTypePointer Uniform %m3x3f16\n"
16174                 "     %up_m3x4f16 = OpTypePointer Uniform %m3x4f16\n"
16175                 "     %up_m4x2f16 = OpTypePointer Uniform %m4x2f16\n"
16176                 "     %up_m4x3f16 = OpTypePointer Uniform %m4x3f16\n"
16177                 "     %up_m4x4f16 = OpTypePointer Uniform %m4x4f16\n"
16178
16179                 "     %ra_f16     = OpTypeArray %f16     %c_i32_ndp\n"
16180                 "     %ra_v2f16   = OpTypeArray %v2f16   %c_i32_ndp\n"
16181                 "     %ra_v3f16   = OpTypeArray %v3f16   %c_i32_ndp\n"
16182                 "     %ra_v4f16   = OpTypeArray %v4f16   %c_i32_ndp\n"
16183                 "     %ra_m2x2f16 = OpTypeArray %m2x2f16 %c_i32_ndp\n"
16184                 "     %ra_m2x3f16 = OpTypeArray %m2x3f16 %c_i32_ndp\n"
16185                 "     %ra_m2x4f16 = OpTypeArray %m2x4f16 %c_i32_ndp\n"
16186                 "     %ra_m3x2f16 = OpTypeArray %m3x2f16 %c_i32_ndp\n"
16187                 "     %ra_m3x3f16 = OpTypeArray %m3x3f16 %c_i32_ndp\n"
16188                 "     %ra_m3x4f16 = OpTypeArray %m3x4f16 %c_i32_ndp\n"
16189                 "     %ra_m4x2f16 = OpTypeArray %m4x2f16 %c_i32_ndp\n"
16190                 "     %ra_m4x3f16 = OpTypeArray %m4x3f16 %c_i32_ndp\n"
16191                 "     %ra_m4x4f16 = OpTypeArray %m4x4f16 %c_i32_ndp\n"
16192
16193                 "   %SSBO_f16     = OpTypeStruct %ra_f16    \n"
16194                 "   %SSBO_v2f16   = OpTypeStruct %ra_v2f16  \n"
16195                 "   %SSBO_v3f16   = OpTypeStruct %ra_v3f16  \n"
16196                 "   %SSBO_v4f16   = OpTypeStruct %ra_v4f16  \n"
16197                 "   %SSBO_m2x2f16 = OpTypeStruct %ra_m2x2f16\n"
16198                 "   %SSBO_m2x3f16 = OpTypeStruct %ra_m2x3f16\n"
16199                 "   %SSBO_m2x4f16 = OpTypeStruct %ra_m2x4f16\n"
16200                 "   %SSBO_m3x2f16 = OpTypeStruct %ra_m3x2f16\n"
16201                 "   %SSBO_m3x3f16 = OpTypeStruct %ra_m3x3f16\n"
16202                 "   %SSBO_m3x4f16 = OpTypeStruct %ra_m3x4f16\n"
16203                 "   %SSBO_m4x2f16 = OpTypeStruct %ra_m4x2f16\n"
16204                 "   %SSBO_m4x3f16 = OpTypeStruct %ra_m4x3f16\n"
16205                 "   %SSBO_m4x4f16 = OpTypeStruct %ra_m4x4f16\n"
16206
16207                 "%up_SSBO_f16     = OpTypePointer Uniform %SSBO_f16    \n"
16208                 "%up_SSBO_v2f16   = OpTypePointer Uniform %SSBO_v2f16  \n"
16209                 "%up_SSBO_v3f16   = OpTypePointer Uniform %SSBO_v3f16  \n"
16210                 "%up_SSBO_v4f16   = OpTypePointer Uniform %SSBO_v4f16  \n"
16211                 "%up_SSBO_m2x2f16 = OpTypePointer Uniform %SSBO_m2x2f16\n"
16212                 "%up_SSBO_m2x3f16 = OpTypePointer Uniform %SSBO_m2x3f16\n"
16213                 "%up_SSBO_m2x4f16 = OpTypePointer Uniform %SSBO_m2x4f16\n"
16214                 "%up_SSBO_m3x2f16 = OpTypePointer Uniform %SSBO_m3x2f16\n"
16215                 "%up_SSBO_m3x3f16 = OpTypePointer Uniform %SSBO_m3x3f16\n"
16216                 "%up_SSBO_m3x4f16 = OpTypePointer Uniform %SSBO_m3x4f16\n"
16217                 "%up_SSBO_m4x2f16 = OpTypePointer Uniform %SSBO_m4x2f16\n"
16218                 "%up_SSBO_m4x3f16 = OpTypePointer Uniform %SSBO_m4x3f16\n"
16219                 "%up_SSBO_m4x4f16 = OpTypePointer Uniform %SSBO_m4x4f16\n"
16220
16221                 "       %fp_v2i32 = OpTypePointer Function %v2i32\n"
16222                 "       %fp_v3i32 = OpTypePointer Function %v3i32\n"
16223                 "       %fp_v4i32 = OpTypePointer Function %v4i32\n"
16224                 "${arg_vars}"
16225         );
16226
16227         const StringTemplate decoration
16228         (
16229                 "OpDecorate %ra_f16     ArrayStride 2 \n"
16230                 "OpDecorate %ra_v2f16   ArrayStride 4 \n"
16231                 "OpDecorate %ra_v3f16   ArrayStride 8 \n"
16232                 "OpDecorate %ra_v4f16   ArrayStride 8 \n"
16233                 "OpDecorate %ra_m2x2f16 ArrayStride 8 \n"
16234                 "OpDecorate %ra_m2x3f16 ArrayStride 16\n"
16235                 "OpDecorate %ra_m2x4f16 ArrayStride 16\n"
16236                 "OpDecorate %ra_m3x2f16 ArrayStride 16\n"
16237                 "OpDecorate %ra_m3x3f16 ArrayStride 32\n"
16238                 "OpDecorate %ra_m3x4f16 ArrayStride 32\n"
16239                 "OpDecorate %ra_m4x2f16 ArrayStride 16\n"
16240                 "OpDecorate %ra_m4x3f16 ArrayStride 32\n"
16241                 "OpDecorate %ra_m4x4f16 ArrayStride 32\n"
16242
16243                 "OpMemberDecorate %SSBO_f16     0 Offset 0\n"
16244                 "OpMemberDecorate %SSBO_v2f16   0 Offset 0\n"
16245                 "OpMemberDecorate %SSBO_v3f16   0 Offset 0\n"
16246                 "OpMemberDecorate %SSBO_v4f16   0 Offset 0\n"
16247                 "OpMemberDecorate %SSBO_m2x2f16 0 Offset 0\n"
16248                 "OpMemberDecorate %SSBO_m2x3f16 0 Offset 0\n"
16249                 "OpMemberDecorate %SSBO_m2x4f16 0 Offset 0\n"
16250                 "OpMemberDecorate %SSBO_m3x2f16 0 Offset 0\n"
16251                 "OpMemberDecorate %SSBO_m3x3f16 0 Offset 0\n"
16252                 "OpMemberDecorate %SSBO_m3x4f16 0 Offset 0\n"
16253                 "OpMemberDecorate %SSBO_m4x2f16 0 Offset 0\n"
16254                 "OpMemberDecorate %SSBO_m4x3f16 0 Offset 0\n"
16255                 "OpMemberDecorate %SSBO_m4x4f16 0 Offset 0\n"
16256
16257                 "OpDecorate %SSBO_f16     BufferBlock\n"
16258                 "OpDecorate %SSBO_v2f16   BufferBlock\n"
16259                 "OpDecorate %SSBO_v3f16   BufferBlock\n"
16260                 "OpDecorate %SSBO_v4f16   BufferBlock\n"
16261                 "OpDecorate %SSBO_m2x2f16 BufferBlock\n"
16262                 "OpDecorate %SSBO_m2x3f16 BufferBlock\n"
16263                 "OpDecorate %SSBO_m2x4f16 BufferBlock\n"
16264                 "OpDecorate %SSBO_m3x2f16 BufferBlock\n"
16265                 "OpDecorate %SSBO_m3x3f16 BufferBlock\n"
16266                 "OpDecorate %SSBO_m3x4f16 BufferBlock\n"
16267                 "OpDecorate %SSBO_m4x2f16 BufferBlock\n"
16268                 "OpDecorate %SSBO_m4x3f16 BufferBlock\n"
16269                 "OpDecorate %SSBO_m4x4f16 BufferBlock\n"
16270
16271                 "OpMemberDecorate %SSBO_m2x2f16 0 ColMajor\n"
16272                 "OpMemberDecorate %SSBO_m2x3f16 0 ColMajor\n"
16273                 "OpMemberDecorate %SSBO_m2x4f16 0 ColMajor\n"
16274                 "OpMemberDecorate %SSBO_m3x2f16 0 ColMajor\n"
16275                 "OpMemberDecorate %SSBO_m3x3f16 0 ColMajor\n"
16276                 "OpMemberDecorate %SSBO_m3x4f16 0 ColMajor\n"
16277                 "OpMemberDecorate %SSBO_m4x2f16 0 ColMajor\n"
16278                 "OpMemberDecorate %SSBO_m4x3f16 0 ColMajor\n"
16279                 "OpMemberDecorate %SSBO_m4x4f16 0 ColMajor\n"
16280
16281                 "OpMemberDecorate %SSBO_m2x2f16 0 MatrixStride 4\n"
16282                 "OpMemberDecorate %SSBO_m2x3f16 0 MatrixStride 8\n"
16283                 "OpMemberDecorate %SSBO_m2x4f16 0 MatrixStride 8\n"
16284                 "OpMemberDecorate %SSBO_m3x2f16 0 MatrixStride 4\n"
16285                 "OpMemberDecorate %SSBO_m3x3f16 0 MatrixStride 8\n"
16286                 "OpMemberDecorate %SSBO_m3x4f16 0 MatrixStride 8\n"
16287                 "OpMemberDecorate %SSBO_m4x2f16 0 MatrixStride 4\n"
16288                 "OpMemberDecorate %SSBO_m4x3f16 0 MatrixStride 8\n"
16289                 "OpMemberDecorate %SSBO_m4x4f16 0 MatrixStride 8\n"
16290
16291                 "${arg_decorations}"
16292         );
16293
16294         const StringTemplate testFun
16295         (
16296                 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
16297                 "    %param = OpFunctionParameter %v4f32\n"
16298                 "    %entry = OpLabel\n"
16299
16300                 "        %i = OpVariable %fp_i32 Function\n"
16301                 "${arg_infunc_vars}"
16302                 "             OpStore %i %c_i32_0\n"
16303                 "             OpBranch %loop\n"
16304
16305                 "     %loop = OpLabel\n"
16306                 "    %i_cmp = OpLoad %i32 %i\n"
16307                 "       %lt = OpSLessThan %bool %i_cmp %c_i32_ndp\n"
16308                 "             OpLoopMerge %merge %next None\n"
16309                 "             OpBranchConditional %lt %write %merge\n"
16310
16311                 "    %write = OpLabel\n"
16312                 "      %ndx = OpLoad %i32 %i\n"
16313
16314                 "${arg_func_call}"
16315
16316                 "             OpBranch %next\n"
16317
16318                 "     %next = OpLabel\n"
16319                 "    %i_cur = OpLoad %i32 %i\n"
16320                 "    %i_new = OpIAdd %i32 %i_cur %c_i32_1\n"
16321                 "             OpStore %i %i_new\n"
16322                 "             OpBranch %loop\n"
16323
16324                 "    %merge = OpLabel\n"
16325                 "             OpReturnValue %param\n"
16326                 "             OpFunctionEnd\n"
16327         );
16328
16329         const Math16ArgFragments        argFragment1    =
16330         {
16331                 "     %src0 = OpAccessChain %up_${t0} %ssbo_src0 %c_i32_0 %ndx\n"
16332                 " %val_src0 = OpLoad %${t0} %src0\n"
16333                 "  %val_dst = ${op} %${tr} ${ext_inst} %val_src0\n"
16334                 "      %dst = OpAccessChain %up_${tr} %ssbo_dst %c_i32_0 %ndx\n"
16335                 "             OpStore %dst %val_dst\n",
16336                 "",
16337                 "",
16338                 "",
16339         };
16340
16341         const Math16ArgFragments        argFragment2    =
16342         {
16343                 "     %src0 = OpAccessChain %up_${t0} %ssbo_src0 %c_i32_0 %ndx\n"
16344                 " %val_src0 = OpLoad %${t0} %src0\n"
16345                 "     %src1 = OpAccessChain %up_${t1} %ssbo_src1 %c_i32_0 %ndx\n"
16346                 " %val_src1 = OpLoad %${t1} %src1\n"
16347                 "  %val_dst = ${op} %${tr} ${ext_inst} %val_src0 %val_src1\n"
16348                 "      %dst = OpAccessChain %up_${tr} %ssbo_dst %c_i32_0 %ndx\n"
16349                 "             OpStore %dst %val_dst\n",
16350                 "",
16351                 "",
16352                 "",
16353         };
16354
16355         const Math16ArgFragments        argFragment3    =
16356         {
16357                 "     %src0 = OpAccessChain %up_${t0} %ssbo_src0 %c_i32_0 %ndx\n"
16358                 " %val_src0 = OpLoad %${t0} %src0\n"
16359                 "     %src1 = OpAccessChain %up_${t1} %ssbo_src1 %c_i32_0 %ndx\n"
16360                 " %val_src1 = OpLoad %${t1} %src1\n"
16361                 "     %src2 = OpAccessChain %up_${t2} %ssbo_src2 %c_i32_0 %ndx\n"
16362                 " %val_src2 = OpLoad %${t2} %src2\n"
16363                 "  %val_dst = ${op} %${tr} ${ext_inst} %val_src0 %val_src1 %val_src2\n"
16364                 "      %dst = OpAccessChain %up_${tr} %ssbo_dst %c_i32_0 %ndx\n"
16365                 "             OpStore %dst %val_dst\n",
16366                 "",
16367                 "",
16368                 "",
16369         };
16370
16371         const Math16ArgFragments        argFragmentLdExp        =
16372         {
16373                 "     %src0 = OpAccessChain %up_${t0} %ssbo_src0 %c_i32_0 %ndx\n"
16374                 " %val_src0 = OpLoad %${t0} %src0\n"
16375                 "     %src1 = OpAccessChain %up_${t1} %ssbo_src1 %c_i32_0 %ndx\n"
16376                 " %val_src1 = OpLoad %${t1} %src1\n"
16377                 "%val_src1i = OpConvertFToS %${dr}i32 %val_src1\n"
16378                 "  %val_dst = ${op} %${tr} ${ext_inst} %val_src0 %val_src1i\n"
16379                 "      %dst = OpAccessChain %up_${tr} %ssbo_dst %c_i32_0 %ndx\n"
16380                 "             OpStore %dst %val_dst\n",
16381
16382                 "",
16383
16384                 "",
16385
16386                 "",
16387         };
16388
16389         const Math16ArgFragments        argFragmentModfFrac     =
16390         {
16391                 "     %src0 = OpAccessChain %up_${t0} %ssbo_src0 %c_i32_0 %ndx\n"
16392                 " %val_src0 = OpLoad %${t0} %src0\n"
16393                 "  %val_dst = ${op} %${tr} ${ext_inst} %val_src0 %tmp\n"
16394                 "      %dst = OpAccessChain %up_${tr} %ssbo_dst %c_i32_0 %ndx\n"
16395                 "             OpStore %dst %val_dst\n",
16396
16397                 "   %fp_tmp = OpTypePointer Function %${tr}\n",
16398
16399                 "",
16400
16401                 "      %tmp = OpVariable %fp_tmp Function\n",
16402         };
16403
16404         const Math16ArgFragments        argFragmentModfInt      =
16405         {
16406                 "     %src0 = OpAccessChain %up_${t0} %ssbo_src0 %c_i32_0 %ndx\n"
16407                 " %val_src0 = OpLoad %${t0} %src0\n"
16408                 "%val_dummy = ${op} %${tr} ${ext_inst} %val_src0 %tmp\n"
16409                 "     %tmp0 = OpAccessChain %fp_tmp %tmp\n"
16410                 "  %val_dst = OpLoad %${tr} %tmp0\n"
16411                 "      %dst = OpAccessChain %up_${tr} %ssbo_dst %c_i32_0 %ndx\n"
16412                 "             OpStore %dst %val_dst\n",
16413
16414                 "   %fp_tmp = OpTypePointer Function %${tr}\n",
16415
16416                 "",
16417
16418                 "      %tmp = OpVariable %fp_tmp Function\n",
16419         };
16420
16421         const Math16ArgFragments        argFragmentModfStruct   =
16422         {
16423                 "     %src0 = OpAccessChain %up_${t0} %ssbo_src0 %c_i32_0 %ndx\n"
16424                 " %val_src0 = OpLoad %${t0} %src0\n"
16425                 "  %val_tmp = ${op} %st_tmp ${ext_inst} %val_src0\n"
16426                 "%tmp_ptr_s = OpAccessChain %fp_tmp %tmp\n"
16427                 "             OpStore %tmp_ptr_s %val_tmp\n"
16428                 "%tmp_ptr_l = OpAccessChain %fp_${tr} %tmp %c_${struct_member}\n"
16429                 "  %val_dst = OpLoad %${tr} %tmp_ptr_l\n"
16430                 "      %dst = OpAccessChain %up_${tr} %ssbo_dst %c_i32_0 %ndx\n"
16431                 "             OpStore %dst %val_dst\n",
16432
16433                 "  %fp_${tr} = OpTypePointer Function %${tr}\n"
16434                 "   %st_tmp = OpTypeStruct %${tr} %${tr}\n"
16435                 "   %fp_tmp = OpTypePointer Function %st_tmp\n"
16436                 "   %c_frac = OpConstant %i32 0\n"
16437                 "    %c_int = OpConstant %i32 1\n",
16438
16439                 "OpMemberDecorate %st_tmp 0 Offset 0\n"
16440                 "OpMemberDecorate %st_tmp 1 Offset ${struct_stride}\n",
16441
16442                 "      %tmp = OpVariable %fp_tmp Function\n",
16443         };
16444
16445         const Math16ArgFragments        argFragmentFrexpStructS =
16446         {
16447                 "     %src0 = OpAccessChain %up_${t0} %ssbo_src0 %c_i32_0 %ndx\n"
16448                 " %val_src0 = OpLoad %${t0} %src0\n"
16449                 "  %val_tmp = ${op} %st_tmp ${ext_inst} %val_src0\n"
16450                 "%tmp_ptr_s = OpAccessChain %fp_tmp %tmp\n"
16451                 "             OpStore %tmp_ptr_s %val_tmp\n"
16452                 "%tmp_ptr_l = OpAccessChain %fp_${tr} %tmp %c_i32_0\n"
16453                 "  %val_dst = OpLoad %${tr} %tmp_ptr_l\n"
16454                 "      %dst = OpAccessChain %up_${tr} %ssbo_dst %c_i32_0 %ndx\n"
16455                 "             OpStore %dst %val_dst\n",
16456
16457                 "  %fp_${tr} = OpTypePointer Function %${tr}\n"
16458                 "   %st_tmp = OpTypeStruct %${tr} %${dr}i32\n"
16459                 "   %fp_tmp = OpTypePointer Function %st_tmp\n",
16460
16461                 "OpMemberDecorate %st_tmp 0 Offset 0\n"
16462                 "OpMemberDecorate %st_tmp 1 Offset ${struct_stride}\n",
16463
16464                 "      %tmp = OpVariable %fp_tmp Function\n",
16465         };
16466
16467         const Math16ArgFragments        argFragmentFrexpStructE =
16468         {
16469                 "     %src0 = OpAccessChain %up_${t0} %ssbo_src0 %c_i32_0 %ndx\n"
16470                 " %val_src0 = OpLoad %${t0} %src0\n"
16471                 "  %val_tmp = ${op} %st_tmp ${ext_inst} %val_src0\n"
16472                 "%tmp_ptr_s = OpAccessChain %fp_tmp %tmp\n"
16473                 "             OpStore %tmp_ptr_s %val_tmp\n"
16474                 "%tmp_ptr_l = OpAccessChain %fp_${dr}i32 %tmp %c_i32_1\n"
16475                 "%val_dst_i = OpLoad %${dr}i32 %tmp_ptr_l\n"
16476                 "  %val_dst = OpConvertSToF %${tr} %val_dst_i\n"
16477                 "      %dst = OpAccessChain %up_${tr} %ssbo_dst %c_i32_0 %ndx\n"
16478                 "             OpStore %dst %val_dst\n",
16479
16480                 "   %st_tmp = OpTypeStruct %${tr} %${dr}i32\n"
16481                 "   %fp_tmp = OpTypePointer Function %st_tmp\n",
16482
16483                 "OpMemberDecorate %st_tmp 0 Offset 0\n"
16484                 "OpMemberDecorate %st_tmp 1 Offset ${struct_stride}\n",
16485
16486                 "      %tmp = OpVariable %fp_tmp Function\n",
16487         };
16488
16489         const Math16ArgFragments        argFragmentFrexpS               =
16490         {
16491                 "     %src0 = OpAccessChain %up_${t0} %ssbo_src0 %c_i32_0 %ndx\n"
16492                 " %val_src0 = OpLoad %${t0} %src0\n"
16493                 "  %out_exp = OpAccessChain %fp_${dr}i32 %tmp\n"
16494                 "  %val_dst = ${op} %${tr} ${ext_inst} %val_src0 %out_exp\n"
16495                 "      %dst = OpAccessChain %up_${tr} %ssbo_dst %c_i32_0 %ndx\n"
16496                 "             OpStore %dst %val_dst\n",
16497
16498                 "",
16499
16500                 "",
16501
16502                 "      %tmp = OpVariable %fp_${dr}i32 Function\n",
16503         };
16504
16505         const Math16ArgFragments        argFragmentFrexpE               =
16506         {
16507                 "     %src0 = OpAccessChain %up_${t0} %ssbo_src0 %c_i32_0 %ndx\n"
16508                 " %val_src0 = OpLoad %${t0} %src0\n"
16509                 "  %out_exp = OpAccessChain %fp_${dr}i32 %tmp\n"
16510                 "%val_dummy = ${op} %${tr} ${ext_inst} %val_src0 %out_exp\n"
16511                 "%val_dst_i = OpLoad %${dr}i32 %out_exp\n"
16512                 "  %val_dst = OpConvertSToF %${tr} %val_dst_i\n"
16513                 "      %dst = OpAccessChain %up_${tr} %ssbo_dst %c_i32_0 %ndx\n"
16514                 "             OpStore %dst %val_dst\n",
16515
16516                 "",
16517
16518                 "",
16519
16520                 "      %tmp = OpVariable %fp_${dr}i32 Function\n",
16521         };
16522
16523         const Math16TestType&           testType                                = testTypes[testTypeIdx];
16524         const string                            funcNameString                  = string(testFunc.funcName) + string(testFunc.funcSuffix);
16525         const string                            testName                                = de::toLower(funcNameString);
16526         const Math16ArgFragments*       argFragments                    = DE_NULL;
16527         const size_t                            typeStructStride                = testType.typeStructStride;
16528         const bool                                      extInst                                 = !(testFunc.funcName[0] == 'O' && testFunc.funcName[1] == 'p');
16529         const size_t                            numFloatsPerArg0Type    = testTypes[testFunc.typeArg0].typeArrayStride / sizeof(deFloat16);
16530         const size_t                            iterations                              = numDataPoints / numFloatsPerArg0Type;
16531         const size_t                            numFloatsPerResultType  = testTypes[testFunc.typeResult].typeArrayStride / sizeof(deFloat16);
16532         const vector<deFloat16>         float16DummyOutput              (iterations * numFloatsPerResultType, 0);
16533         VulkanFeatures                          features;
16534         SpecResource                            specResource;
16535         map<string, string>                     specs;
16536         map<string, string>                     fragments;
16537         vector<string>                          extensions;
16538         string                                          funcCall;
16539         string                                          funcVariables;
16540         string                                          variables;
16541         string                                          declarations;
16542         string                                          decorations;
16543
16544         switch (testFunc.funcArgsCount)
16545         {
16546                 case 1:
16547                 {
16548                         argFragments = &argFragment1;
16549
16550                         if (funcNameString == "ModfFrac")               argFragments = &argFragmentModfFrac;
16551                         if (funcNameString == "ModfInt")                argFragments = &argFragmentModfInt;
16552                         if (funcNameString == "ModfStructFrac") argFragments = &argFragmentModfStruct;
16553                         if (funcNameString == "ModfStructInt")  argFragments = &argFragmentModfStruct;
16554                         if (funcNameString == "FrexpS")                 argFragments = &argFragmentFrexpS;
16555                         if (funcNameString == "FrexpE")                 argFragments = &argFragmentFrexpE;
16556                         if (funcNameString == "FrexpStructS")   argFragments = &argFragmentFrexpStructS;
16557                         if (funcNameString == "FrexpStructE")   argFragments = &argFragmentFrexpStructE;
16558
16559                         break;
16560                 }
16561                 case 2:
16562                 {
16563                         argFragments = &argFragment2;
16564
16565                         if (funcNameString == "Ldexp")                  argFragments = &argFragmentLdExp;
16566
16567                         break;
16568                 }
16569                 case 3:
16570                 {
16571                         argFragments = &argFragment3;
16572
16573                         break;
16574                 }
16575                 default:
16576                 {
16577                         TCU_THROW(InternalError, "Invalid number of arguments");
16578                 }
16579         }
16580
16581         if (testFunc.funcArgsCount == 1)
16582         {
16583                 variables +=
16584                         " %ssbo_src0 = OpVariable %up_SSBO_${t0} Uniform\n"
16585                         "  %ssbo_dst = OpVariable %up_SSBO_${tr} Uniform\n";
16586
16587                 decorations +=
16588                         "OpDecorate %ssbo_src0 DescriptorSet 0\n"
16589                         "OpDecorate %ssbo_src0 Binding 0\n"
16590                         "OpDecorate %ssbo_dst DescriptorSet 0\n"
16591                         "OpDecorate %ssbo_dst Binding 1\n";
16592         }
16593         else if (testFunc.funcArgsCount == 2)
16594         {
16595                 variables +=
16596                         " %ssbo_src0 = OpVariable %up_SSBO_${t0} Uniform\n"
16597                         " %ssbo_src1 = OpVariable %up_SSBO_${t1} Uniform\n"
16598                         "  %ssbo_dst = OpVariable %up_SSBO_${tr} Uniform\n";
16599
16600                 decorations +=
16601                         "OpDecorate %ssbo_src0 DescriptorSet 0\n"
16602                         "OpDecorate %ssbo_src0 Binding 0\n"
16603                         "OpDecorate %ssbo_src1 DescriptorSet 0\n"
16604                         "OpDecorate %ssbo_src1 Binding 1\n"
16605                         "OpDecorate %ssbo_dst DescriptorSet 0\n"
16606                         "OpDecorate %ssbo_dst Binding 2\n";
16607         }
16608         else if (testFunc.funcArgsCount == 3)
16609         {
16610                 variables +=
16611                         " %ssbo_src0 = OpVariable %up_SSBO_${t0} Uniform\n"
16612                         " %ssbo_src1 = OpVariable %up_SSBO_${t1} Uniform\n"
16613                         " %ssbo_src2 = OpVariable %up_SSBO_${t2} Uniform\n"
16614                         "  %ssbo_dst = OpVariable %up_SSBO_${tr} Uniform\n";
16615
16616                 decorations +=
16617                         "OpDecorate %ssbo_src0 DescriptorSet 0\n"
16618                         "OpDecorate %ssbo_src0 Binding 0\n"
16619                         "OpDecorate %ssbo_src1 DescriptorSet 0\n"
16620                         "OpDecorate %ssbo_src1 Binding 1\n"
16621                         "OpDecorate %ssbo_src2 DescriptorSet 0\n"
16622                         "OpDecorate %ssbo_src2 Binding 2\n"
16623                         "OpDecorate %ssbo_dst DescriptorSet 0\n"
16624                         "OpDecorate %ssbo_dst Binding 3\n";
16625         }
16626         else
16627         {
16628                 TCU_THROW(InternalError, "Invalid number of function arguments");
16629         }
16630
16631         variables       += argFragments->variables;
16632         decorations     += argFragments->decorations;
16633
16634         specs["dr"]                                     = testTypes[testFunc.typeResult].typePrefix;
16635         specs["d0"]                                     = testTypes[testFunc.typeArg0].typePrefix;
16636         specs["d1"]                                     = testTypes[testFunc.typeArg1].typePrefix;
16637         specs["d2"]                                     = testTypes[testFunc.typeArg2].typePrefix;
16638         specs["tr"]                                     = string(testTypes[testFunc.typeResult].typePrefix) + componentType;
16639         specs["t0"]                                     = string(testTypes[testFunc.typeArg0].typePrefix) + componentType;
16640         specs["t1"]                                     = string(testTypes[testFunc.typeArg1].typePrefix) + componentType;
16641         specs["t2"]                                     = string(testTypes[testFunc.typeArg2].typePrefix) + componentType;
16642         specs["struct_stride"]          = de::toString(typeStructStride);
16643         specs["op"]                                     = extInst ? "OpExtInst" : testFunc.funcName;
16644         specs["ext_inst"]                       = extInst ? string("%ext_import ") + testFunc.funcName : "";
16645         specs["struct_member"]          = de::toLower(testFunc.funcSuffix);
16646
16647         variables                                       = StringTemplate(variables).specialize(specs);
16648         decorations                                     = StringTemplate(decorations).specialize(specs);
16649         funcVariables                           = StringTemplate(argFragments->funcVariables).specialize(specs);
16650         funcCall                                        = StringTemplate(argFragments->bodies).specialize(specs);
16651
16652         specs["num_data_points"]        = de::toString(iterations);
16653         specs["arg_vars"]                       = variables;
16654         specs["arg_decorations"]        = decorations;
16655         specs["arg_infunc_vars"]        = funcVariables;
16656         specs["arg_func_call"]          = funcCall;
16657
16658         fragments["extension"]          = "OpExtension \"SPV_KHR_16bit_storage\"\n%ext_import = OpExtInstImport \"GLSL.std.450\"";
16659         fragments["capability"]         = "OpCapability Matrix\nOpCapability StorageUniformBufferBlock16";
16660         fragments["decoration"]         = decoration.specialize(specs);
16661         fragments["pre_main"]           = preMain.specialize(specs);
16662         fragments["testfun"]            = testFun.specialize(specs);
16663
16664         for (size_t inputArgNdx = 0; inputArgNdx < testFunc.funcArgsCount; ++inputArgNdx)
16665         {
16666                 const size_t                    numFloatsPerItem        = (inputArgNdx == 0) ? testTypes[testFunc.typeArg0].typeArrayStride / sizeof(deFloat16)
16667                                                                                                         : (inputArgNdx == 1) ? testTypes[testFunc.typeArg1].typeArrayStride / sizeof(deFloat16)
16668                                                                                                         : (inputArgNdx == 2) ? testTypes[testFunc.typeArg2].typeArrayStride / sizeof(deFloat16)
16669                                                                                                         : -1;
16670                 const vector<deFloat16> inputData                       = testFunc.getInputDataFunc(seed, numFloatsPerItem * iterations, testTypeIdx, numFloatsPerItem, testFunc.funcArgsCount, inputArgNdx);
16671
16672                 specResource.inputs.push_back(Resource(BufferSp(new Float16Buffer(inputData)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
16673         }
16674
16675         specResource.outputs.push_back(Resource(BufferSp(new Float16Buffer(float16DummyOutput)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
16676         specResource.verifyIO = testFunc.verifyFunc;
16677
16678         extensions.push_back("VK_KHR_16bit_storage");
16679         extensions.push_back("VK_KHR_shader_float16_int8");
16680
16681         features.ext16BitStorage        = EXT16BITSTORAGEFEATURES_UNIFORM_BUFFER_BLOCK;
16682         features.extFloat16Int8         = EXTFLOAT16INT8FEATURES_FLOAT16;
16683
16684         finalizeTestsCreation(specResource, fragments, testCtx, testGroup, testName, features, extensions, IVec3(1, 1, 1));
16685 }
16686
16687 template<size_t C, class SpecResource>
16688 tcu::TestCaseGroup* createFloat16ArithmeticSet (tcu::TestContext& testCtx)
16689 {
16690         DE_STATIC_ASSERT(C >= 1 && C <= 4);
16691
16692         const std::string                               testGroupName   (string("arithmetic_") + de::toString(C));
16693         de::MovePtr<tcu::TestCaseGroup> testGroup               (new tcu::TestCaseGroup(testCtx, testGroupName.c_str(), "Float 16 arithmetic and related tests"));
16694         const Math16TestFunc                    testFuncs[]             =
16695         {
16696                 {       "OpFNegate",                    "",                     1,      C,              C,              0,              0, &getInputData,       compareFP16ArithmeticFunc<  C,  C,  0,  0, fp16OpFNegate>                                       },
16697                 {       "Round",                                "",                     1,      C,              C,              0,              0, &getInputData,       compareFP16ArithmeticFunc<  C,  C,  0,  0, fp16Round>                                           },
16698                 {       "RoundEven",                    "",                     1,      C,              C,              0,              0, &getInputData,       compareFP16ArithmeticFunc<  C,  C,  0,  0, fp16RoundEven>                                       },
16699                 {       "Trunc",                                "",                     1,      C,              C,              0,              0, &getInputData,       compareFP16ArithmeticFunc<  C,  C,  0,  0, fp16Trunc>                                           },
16700                 {       "FAbs",                                 "",                     1,      C,              C,              0,              0, &getInputData,       compareFP16ArithmeticFunc<  C,  C,  0,  0, fp16FAbs>                                            },
16701                 {       "FSign",                                "",                     1,      C,              C,              0,              0, &getInputData,       compareFP16ArithmeticFunc<  C,  C,  0,  0, fp16FSign>                                           },
16702                 {       "Floor",                                "",                     1,      C,              C,              0,              0, &getInputData,       compareFP16ArithmeticFunc<  C,  C,  0,  0, fp16Floor>                                           },
16703                 {       "Ceil",                                 "",                     1,      C,              C,              0,              0, &getInputData,       compareFP16ArithmeticFunc<  C,  C,  0,  0, fp16Ceil>                                            },
16704                 {       "Fract",                                "",                     1,      C,              C,              0,              0, &getInputData,       compareFP16ArithmeticFunc<  C,  C,  0,  0, fp16Fract>                                           },
16705                 {       "Radians",                              "",                     1,      C,              C,              0,              0, &getInputData,       compareFP16ArithmeticFunc<  C,  C,  0,  0, fp16Radians>                                         },
16706                 {       "Degrees",                              "",                     1,      C,              C,              0,              0, &getInputData,       compareFP16ArithmeticFunc<  C,  C,  0,  0, fp16Degrees>                                         },
16707                 {       "Sin",                                  "",                     1,      C,              C,              0,              0, &getInputDataPI,     compareFP16ArithmeticFunc<  C,  C,  0,  0, fp16Sin>                                                     },
16708                 {       "Cos",                                  "",                     1,      C,              C,              0,              0, &getInputDataPI,     compareFP16ArithmeticFunc<  C,  C,  0,  0, fp16Cos>                                                     },
16709                 {       "Tan",                                  "",                     1,      C,              C,              0,              0, &getInputDataPI,     compareFP16ArithmeticFunc<  C,  C,  0,  0, fp16Tan>                                                     },
16710                 {       "Asin",                                 "",                     1,      C,              C,              0,              0, &getInputDataA,      compareFP16ArithmeticFunc<  C,  C,  0,  0, fp16Asin>                                            },
16711                 {       "Acos",                                 "",                     1,      C,              C,              0,              0, &getInputDataA,      compareFP16ArithmeticFunc<  C,  C,  0,  0, fp16Acos>                                            },
16712                 {       "Atan",                                 "",                     1,      C,              C,              0,              0, &getInputData,       compareFP16ArithmeticFunc<  C,  C,  0,  0, fp16Atan>                                            },
16713                 {       "Sinh",                                 "",                     1,      C,              C,              0,              0, &getInputData,       compareFP16ArithmeticFunc<  C,  C,  0,  0, fp16Sinh>                                            },
16714                 {       "Cosh",                                 "",                     1,      C,              C,              0,              0, &getInputData,       compareFP16ArithmeticFunc<  C,  C,  0,  0, fp16Cosh>                                            },
16715                 {       "Tanh",                                 "",                     1,      C,              C,              0,              0, &getInputData,       compareFP16ArithmeticFunc<  C,  C,  0,  0, fp16Tanh>                                            },
16716                 {       "Asinh",                                "",                     1,      C,              C,              0,              0, &getInputData,       compareFP16ArithmeticFunc<  C,  C,  0,  0, fp16Asinh>                                           },
16717                 {       "Acosh",                                "",                     1,      C,              C,              0,              0, &getInputDataAC,     compareFP16ArithmeticFunc<  C,  C,  0,  0, fp16Acosh>                                           },
16718                 {       "Atanh",                                "",                     1,      C,              C,              0,              0, &getInputDataA,      compareFP16ArithmeticFunc<  C,  C,  0,  0, fp16Atanh>                                           },
16719                 {       "Exp",                                  "",                     1,      C,              C,              0,              0, &getInputData,       compareFP16ArithmeticFunc<  C,  C,  0,  0, fp16Exp>                                                     },
16720                 {       "Log",                                  "",                     1,      C,              C,              0,              0, &getInputDataP,      compareFP16ArithmeticFunc<  C,  C,  0,  0, fp16Log>                                                     },
16721                 {       "Exp2",                                 "",                     1,      C,              C,              0,              0, &getInputData,       compareFP16ArithmeticFunc<  C,  C,  0,  0, fp16Exp2>                                            },
16722                 {       "Log2",                                 "",                     1,      C,              C,              0,              0, &getInputDataP,      compareFP16ArithmeticFunc<  C,  C,  0,  0, fp16Log2>                                            },
16723                 {       "Sqrt",                                 "",                     1,      C,              C,              0,              0, &getInputDataP,      compareFP16ArithmeticFunc<  C,  C,  0,  0, fp16Sqrt>                                            },
16724                 {       "InverseSqrt",                  "",                     1,      C,              C,              0,              0, &getInputDataP,      compareFP16ArithmeticFunc<  C,  C,  0,  0, fp16InverseSqrt>                                     },
16725                 {       "Modf",                                 "Frac",         1,      C,              C,              0,              0, &getInputData,       compareFP16ArithmeticFunc<  C,  C,  0,  0, fp16ModfFrac>                                        },
16726                 {       "Modf",                                 "Int",          1,      C,              C,              0,              0, &getInputData,       compareFP16ArithmeticFunc<  C,  C,  0,  0, fp16ModfInt>                                         },
16727                 {       "ModfStruct",                   "Frac",         1,      C,              C,              0,              0, &getInputData,       compareFP16ArithmeticFunc<  C,  C,  0,  0, fp16ModfFrac>                                        },
16728                 {       "ModfStruct",                   "Int",          1,      C,              C,              0,              0, &getInputData,       compareFP16ArithmeticFunc<  C,  C,  0,  0, fp16ModfInt>                                         },
16729                 {       "Frexp",                                "S",            1,      C,              C,              0,              0, &getInputData,       compareFP16ArithmeticFunc<  C,  C,  0,  0, fp16FrexpS>                                          },
16730                 {       "Frexp",                                "E",            1,      C,              C,              0,              0, &getInputData,       compareFP16ArithmeticFunc<  C,  C,  0,  0, fp16FrexpE>                                          },
16731                 {       "FrexpStruct",                  "S",            1,      C,              C,              0,              0, &getInputData,       compareFP16ArithmeticFunc<  C,  C,  0,  0, fp16FrexpS>                                          },
16732                 {       "FrexpStruct",                  "E",            1,      C,              C,              0,              0, &getInputData,       compareFP16ArithmeticFunc<  C,  C,  0,  0, fp16FrexpE>                                          },
16733                 {       "OpFAdd",                               "",                     2,      C,              C,              C,              0, &getInputData,       compareFP16ArithmeticFunc<  C,  C,  C,  0, fp16OpFAdd>                                          },
16734                 {       "OpFSub",                               "",                     2,      C,              C,              C,              0, &getInputData,       compareFP16ArithmeticFunc<  C,  C,  C,  0, fp16OpFSub>                                          },
16735                 {       "OpFMul",                               "",                     2,      C,              C,              C,              0, &getInputData,       compareFP16ArithmeticFunc<  C,  C,  C,  0, fp16OpFMul>                                          },
16736                 {       "OpFDiv",                               "",                     2,      C,              C,              C,              0, &getInputData,       compareFP16ArithmeticFunc<  C,  C,  C,  0, fp16OpFDiv>                                          },
16737                 {       "Atan2",                                "",                     2,      C,              C,              C,              0, &getInputData,       compareFP16ArithmeticFunc<  C,  C,  C,  0, fp16Atan2>                                           },
16738                 {       "Pow",                                  "",                     2,      C,              C,              C,              0, &getInputDataP,      compareFP16ArithmeticFunc<  C,  C,  C,  0, fp16Pow>                                                     },
16739                 {       "FMin",                                 "",                     2,      C,              C,              C,              0, &getInputData,       compareFP16ArithmeticFunc<  C,  C,  C,  0, fp16FMin>                                            },
16740                 {       "FMax",                                 "",                     2,      C,              C,              C,              0, &getInputData,       compareFP16ArithmeticFunc<  C,  C,  C,  0, fp16FMax>                                            },
16741                 {       "Step",                                 "",                     2,      C,              C,              C,              0, &getInputData,       compareFP16ArithmeticFunc<  C,  C,  C,  0, fp16Step>                                            },
16742                 {       "Ldexp",                                "",                     2,      C,              C,              C,              0, &getInputData,       compareFP16ArithmeticFunc<  C,  C,  C,  0, fp16Ldexp>                                           },
16743                 {       "FClamp",                               "",                     3,      C,              C,              C,              C, &getInputData,       compareFP16ArithmeticFunc<  C,  C,  C,  C, fp16FClamp>                                          },
16744                 {       "FMix",                                 "",                     3,      C,              C,              C,              C, &getInputDataD,      compareFP16ArithmeticFunc<  C,  C,  C,  C, fp16FMix>                                            },
16745                 {       "SmoothStep",                   "",                     3,      C,              C,              C,              C, &getInputDataSS,     compareFP16ArithmeticFunc<  C,  C,  C,  C, fp16SmoothStep>                                      },
16746                 {       "Fma",                                  "",                     3,      C,              C,              C,              C, &getInputData,       compareFP16ArithmeticFunc<  C,  C,  C,  C, fp16Fma>                                                     },
16747                 {       "Length",                               "",                     1,      1,              C,              0,              0, &getInputData,       compareFP16ArithmeticFunc<  1,  C,  0,  0, fp16Length>                                          },
16748                 {       "Distance",                             "",                     2,      1,              C,              C,              0, &getInputData,       compareFP16ArithmeticFunc<  1,  C,  C,  0, fp16Distance>                                        },
16749                 {       "Cross",                                "",                     2,      C,              C,              C,              0, &getInputDataD,      compareFP16ArithmeticFunc<  C,  C,  C,  0, fp16Cross>                                           },
16750                 {       "Normalize",                    "",                     1,      C,              C,              0,              0, &getInputData,       compareFP16ArithmeticFunc<  C,  C,  0,  0, fp16Normalize>                                       },
16751                 {       "FaceForward",                  "",                     3,      C,              C,              C,              C, &getInputDataD,      compareFP16ArithmeticFunc<  C,  C,  C,  C, fp16FaceForward>                                     },
16752                 {       "Reflect",                              "",                     2,      C,              C,              C,              0, &getInputDataD,      compareFP16ArithmeticFunc<  C,  C,  C,  0, fp16Reflect>                                         },
16753                 {       "Refract",                              "",                     3,      C,              C,              C,              1, &getInputDataN,      compareFP16ArithmeticFunc<  C,  C,  C,  1, fp16Refract>                                         },
16754                 {       "OpDot",                                "",                     2,      1,              C,              C,              0, &getInputDataD,      compareFP16ArithmeticFunc<  1,  C,  C,  0, fp16Dot>                                                     },
16755                 {       "OpVectorTimesScalar",  "",                     2,      C,              C,              1,              0, &getInputDataV,      compareFP16ArithmeticFunc<  C,  C,  1,  0, fp16VectorTimesScalar>                       },
16756         };
16757
16758         for (deUint32 testFuncIdx = 0; testFuncIdx < DE_LENGTH_OF_ARRAY(testFuncs); ++testFuncIdx)
16759         {
16760                 const Math16TestFunc&   testFunc                = testFuncs[testFuncIdx];
16761                 const string                    funcNameString  = testFunc.funcName;
16762
16763                 if ((C != 3) && funcNameString == "Cross")
16764                         continue;
16765
16766                 if ((C < 2) && funcNameString == "OpDot")
16767                         continue;
16768
16769                 if ((C < 2) && funcNameString == "OpVectorTimesScalar")
16770                         continue;
16771
16772                 createFloat16ArithmeticFuncTest<SpecResource>(testCtx, *testGroup.get(), C, testFunc);
16773         }
16774
16775         return testGroup.release();
16776 }
16777
16778 template<class SpecResource>
16779 tcu::TestCaseGroup* createFloat16ArithmeticSet (tcu::TestContext& testCtx)
16780 {
16781         const std::string                               testGroupName   ("arithmetic");
16782         de::MovePtr<tcu::TestCaseGroup> testGroup               (new tcu::TestCaseGroup(testCtx, testGroupName.c_str(), "Float 16 arithmetic and related tests"));
16783         const Math16TestFunc                    testFuncs[]             =
16784         {
16785                 {       "OpTranspose",                  "2x2",          1,      MAT2X2, MAT2X2, 0,              0, &getInputDataM,      compareFP16ArithmeticFunc<  4,  4,  0,  0, fp16Transpose<2,2> >                         },
16786                 {       "OpTranspose",                  "3x2",          1,      MAT2X3, MAT3X2, 0,              0, &getInputDataM,      compareFP16ArithmeticFunc<  8,  8,  0,  0, fp16Transpose<3,2> >                         },
16787                 {       "OpTranspose",                  "4x2",          1,      MAT2X4, MAT4X2, 0,              0, &getInputDataM,      compareFP16ArithmeticFunc<  8,  8,  0,  0, fp16Transpose<4,2> >                         },
16788                 {       "OpTranspose",                  "2x3",          1,      MAT3X2, MAT2X3, 0,              0, &getInputDataM,      compareFP16ArithmeticFunc<  8,  8,  0,  0, fp16Transpose<2,3> >                         },
16789                 {       "OpTranspose",                  "3x3",          1,      MAT3X3, MAT3X3, 0,              0, &getInputDataM,      compareFP16ArithmeticFunc< 16, 16,  0,  0, fp16Transpose<3,3> >                         },
16790                 {       "OpTranspose",                  "4x3",          1,      MAT3X4, MAT4X3, 0,              0, &getInputDataM,      compareFP16ArithmeticFunc< 16, 16,  0,  0, fp16Transpose<4,3> >                         },
16791                 {       "OpTranspose",                  "2x4",          1,      MAT4X2, MAT2X4, 0,              0, &getInputDataM,      compareFP16ArithmeticFunc<  8,  8,  0,  0, fp16Transpose<2,4> >                         },
16792                 {       "OpTranspose",                  "3x4",          1,      MAT4X3, MAT3X4, 0,              0, &getInputDataM,      compareFP16ArithmeticFunc< 16, 16,  0,  0, fp16Transpose<3,4> >                         },
16793                 {       "OpTranspose",                  "4x4",          1,      MAT4X4, MAT4X4, 0,              0, &getInputDataM,      compareFP16ArithmeticFunc< 16, 16,  0,  0, fp16Transpose<4,4> >                         },
16794                 {       "OpMatrixTimesScalar",  "2x2",          2,      MAT2X2, MAT2X2, 1,              0, &getInputDataD,      compareFP16ArithmeticFunc<  4,  4,  1,  0, fp16MatrixTimesScalar<2,2> >         },
16795                 {       "OpMatrixTimesScalar",  "2x3",          2,      MAT2X3, MAT2X3, 1,              0, &getInputDataD,      compareFP16ArithmeticFunc<  8,  8,  1,  0, fp16MatrixTimesScalar<2,3> >         },
16796                 {       "OpMatrixTimesScalar",  "2x4",          2,      MAT2X4, MAT2X4, 1,              0, &getInputDataD,      compareFP16ArithmeticFunc<  8,  8,  1,  0, fp16MatrixTimesScalar<2,4> >         },
16797                 {       "OpMatrixTimesScalar",  "3x2",          2,      MAT3X2, MAT3X2, 1,              0, &getInputDataD,      compareFP16ArithmeticFunc<  8,  8,  1,  0, fp16MatrixTimesScalar<3,2> >         },
16798                 {       "OpMatrixTimesScalar",  "3x3",          2,      MAT3X3, MAT3X3, 1,              0, &getInputDataD,      compareFP16ArithmeticFunc< 16, 16,  1,  0, fp16MatrixTimesScalar<3,3> >         },
16799                 {       "OpMatrixTimesScalar",  "3x4",          2,      MAT3X4, MAT3X4, 1,              0, &getInputDataD,      compareFP16ArithmeticFunc< 16, 16,  1,  0, fp16MatrixTimesScalar<3,4> >         },
16800                 {       "OpMatrixTimesScalar",  "4x2",          2,      MAT4X2, MAT4X2, 1,              0, &getInputDataD,      compareFP16ArithmeticFunc<  8,  8,  1,  0, fp16MatrixTimesScalar<4,2> >         },
16801                 {       "OpMatrixTimesScalar",  "4x3",          2,      MAT4X3, MAT4X3, 1,              0, &getInputDataD,      compareFP16ArithmeticFunc< 16, 16,  1,  0, fp16MatrixTimesScalar<4,3> >         },
16802                 {       "OpMatrixTimesScalar",  "4x4",          2,      MAT4X4, MAT4X4, 1,              0, &getInputDataD,      compareFP16ArithmeticFunc< 16, 16,  1,  0, fp16MatrixTimesScalar<4,4> >         },
16803                 {       "OpVectorTimesMatrix",  "2x2",          2,      VEC2,   VEC2,   MAT2X2, 0, &getInputDataD,      compareFP16ArithmeticFunc<  2,  2,  4,  0, fp16VectorTimesMatrix<2,2> >         },
16804                 {       "OpVectorTimesMatrix",  "2x3",          2,      VEC2,   VEC3,   MAT2X3, 0, &getInputDataD,      compareFP16ArithmeticFunc<  2,  3,  8,  0, fp16VectorTimesMatrix<2,3> >         },
16805                 {       "OpVectorTimesMatrix",  "2x4",          2,      VEC2,   VEC4,   MAT2X4, 0, &getInputDataD,      compareFP16ArithmeticFunc<  2,  4,  8,  0, fp16VectorTimesMatrix<2,4> >         },
16806                 {       "OpVectorTimesMatrix",  "3x2",          2,      VEC3,   VEC2,   MAT3X2, 0, &getInputDataD,      compareFP16ArithmeticFunc<  3,  2,  8,  0, fp16VectorTimesMatrix<3,2> >         },
16807                 {       "OpVectorTimesMatrix",  "3x3",          2,      VEC3,   VEC3,   MAT3X3, 0, &getInputDataD,      compareFP16ArithmeticFunc<  3,  3, 16,  0, fp16VectorTimesMatrix<3,3> >         },
16808                 {       "OpVectorTimesMatrix",  "3x4",          2,      VEC3,   VEC4,   MAT3X4, 0, &getInputDataD,      compareFP16ArithmeticFunc<  3,  4, 16,  0, fp16VectorTimesMatrix<3,4> >         },
16809                 {       "OpVectorTimesMatrix",  "4x2",          2,      VEC4,   VEC2,   MAT4X2, 0, &getInputDataD,      compareFP16ArithmeticFunc<  4,  2,  8,  0, fp16VectorTimesMatrix<4,2> >         },
16810                 {       "OpVectorTimesMatrix",  "4x3",          2,      VEC4,   VEC3,   MAT4X3, 0, &getInputDataD,      compareFP16ArithmeticFunc<  4,  3, 16,  0, fp16VectorTimesMatrix<4,3> >         },
16811                 {       "OpVectorTimesMatrix",  "4x4",          2,      VEC4,   VEC4,   MAT4X4, 0, &getInputDataD,      compareFP16ArithmeticFunc<  4,  4, 16,  0, fp16VectorTimesMatrix<4,4> >         },
16812                 {       "OpMatrixTimesVector",  "2x2",          2,      VEC2,   MAT2X2, VEC2,   0, &getInputDataD,      compareFP16ArithmeticFunc<  2,  4,  2,  0, fp16MatrixTimesVector<2,2> >         },
16813                 {       "OpMatrixTimesVector",  "2x3",          2,      VEC3,   MAT2X3, VEC2,   0, &getInputDataD,      compareFP16ArithmeticFunc<  3,  8,  2,  0, fp16MatrixTimesVector<2,3> >         },
16814                 {       "OpMatrixTimesVector",  "2x4",          2,      VEC4,   MAT2X4, VEC2,   0, &getInputDataD,      compareFP16ArithmeticFunc<  4,  8,  2,  0, fp16MatrixTimesVector<2,4> >         },
16815                 {       "OpMatrixTimesVector",  "3x2",          2,      VEC2,   MAT3X2, VEC3,   0, &getInputDataD,      compareFP16ArithmeticFunc<  2,  8,  3,  0, fp16MatrixTimesVector<3,2> >         },
16816                 {       "OpMatrixTimesVector",  "3x3",          2,      VEC3,   MAT3X3, VEC3,   0, &getInputDataD,      compareFP16ArithmeticFunc<  3, 16,  3,  0, fp16MatrixTimesVector<3,3> >         },
16817                 {       "OpMatrixTimesVector",  "3x4",          2,      VEC4,   MAT3X4, VEC3,   0, &getInputDataD,      compareFP16ArithmeticFunc<  4, 16,  3,  0, fp16MatrixTimesVector<3,4> >         },
16818                 {       "OpMatrixTimesVector",  "4x2",          2,      VEC2,   MAT4X2, VEC4,   0, &getInputDataD,      compareFP16ArithmeticFunc<  2,  8,  4,  0, fp16MatrixTimesVector<4,2> >         },
16819                 {       "OpMatrixTimesVector",  "4x3",          2,      VEC3,   MAT4X3, VEC4,   0, &getInputDataD,      compareFP16ArithmeticFunc<  3, 16,  4,  0, fp16MatrixTimesVector<4,3> >         },
16820                 {       "OpMatrixTimesVector",  "4x4",          2,      VEC4,   MAT4X4, VEC4,   0, &getInputDataD,      compareFP16ArithmeticFunc<  4, 16,  4,  0, fp16MatrixTimesVector<4,4> >         },
16821                 {       "OpMatrixTimesMatrix",  "2x2_2x2",      2,      MAT2X2, MAT2X2, MAT2X2, 0, &getInputDataD,      compareFP16ArithmeticFunc<  4,  4,  4,  0, fp16MatrixTimesMatrix<2,2,2,2> >     },
16822                 {       "OpMatrixTimesMatrix",  "2x2_3x2",      2,      MAT3X2, MAT2X2, MAT3X2, 0, &getInputDataD,      compareFP16ArithmeticFunc<  8,  4,  8,  0, fp16MatrixTimesMatrix<2,2,3,2> >     },
16823                 {       "OpMatrixTimesMatrix",  "2x2_4x2",      2,      MAT4X2, MAT2X2, MAT4X2, 0, &getInputDataD,      compareFP16ArithmeticFunc<  8,  4,  8,  0, fp16MatrixTimesMatrix<2,2,4,2> >     },
16824                 {       "OpMatrixTimesMatrix",  "2x3_2x2",      2,      MAT2X3, MAT2X3, MAT2X2, 0, &getInputDataD,      compareFP16ArithmeticFunc<  8,  8,  4,  0, fp16MatrixTimesMatrix<2,3,2,2> >     },
16825                 {       "OpMatrixTimesMatrix",  "2x3_3x2",      2,      MAT3X3, MAT2X3, MAT3X2, 0, &getInputDataD,      compareFP16ArithmeticFunc< 16,  8,  8,  0, fp16MatrixTimesMatrix<2,3,3,2> >     },
16826                 {       "OpMatrixTimesMatrix",  "2x3_4x2",      2,      MAT4X3, MAT2X3, MAT4X2, 0, &getInputDataD,      compareFP16ArithmeticFunc< 16,  8,  8,  0, fp16MatrixTimesMatrix<2,3,4,2> >     },
16827                 {       "OpMatrixTimesMatrix",  "2x4_2x2",      2,      MAT2X4, MAT2X4, MAT2X2, 0, &getInputDataD,      compareFP16ArithmeticFunc<  8,  8,  4,  0, fp16MatrixTimesMatrix<2,4,2,2> >     },
16828                 {       "OpMatrixTimesMatrix",  "2x4_3x2",      2,      MAT3X4, MAT2X4, MAT3X2, 0, &getInputDataD,      compareFP16ArithmeticFunc< 16,  8,  8,  0, fp16MatrixTimesMatrix<2,4,3,2> >     },
16829                 {       "OpMatrixTimesMatrix",  "2x4_4x2",      2,      MAT4X4, MAT2X4, MAT4X2, 0, &getInputDataD,      compareFP16ArithmeticFunc< 16,  8,  8,  0, fp16MatrixTimesMatrix<2,4,4,2> >     },
16830                 {       "OpMatrixTimesMatrix",  "3x2_2x3",      2,      MAT2X2, MAT3X2, MAT2X3, 0, &getInputDataD,      compareFP16ArithmeticFunc<  4,  8,  8,  0, fp16MatrixTimesMatrix<3,2,2,3> >     },
16831                 {       "OpMatrixTimesMatrix",  "3x2_3x3",      2,      MAT3X2, MAT3X2, MAT3X3, 0, &getInputDataD,      compareFP16ArithmeticFunc<  8,  8, 16,  0, fp16MatrixTimesMatrix<3,2,3,3> >     },
16832                 {       "OpMatrixTimesMatrix",  "3x2_4x3",      2,      MAT4X2, MAT3X2, MAT4X3, 0, &getInputDataD,      compareFP16ArithmeticFunc<  8,  8, 16,  0, fp16MatrixTimesMatrix<3,2,4,3> >     },
16833                 {       "OpMatrixTimesMatrix",  "3x3_2x3",      2,      MAT2X3, MAT3X3, MAT2X3, 0, &getInputDataD,      compareFP16ArithmeticFunc<  8, 16,  8,  0, fp16MatrixTimesMatrix<3,3,2,3> >     },
16834                 {       "OpMatrixTimesMatrix",  "3x3_3x3",      2,      MAT3X3, MAT3X3, MAT3X3, 0, &getInputDataD,      compareFP16ArithmeticFunc< 16, 16, 16,  0, fp16MatrixTimesMatrix<3,3,3,3> >     },
16835                 {       "OpMatrixTimesMatrix",  "3x3_4x3",      2,      MAT4X3, MAT3X3, MAT4X3, 0, &getInputDataD,      compareFP16ArithmeticFunc< 16, 16, 16,  0, fp16MatrixTimesMatrix<3,3,4,3> >     },
16836                 {       "OpMatrixTimesMatrix",  "3x4_2x3",      2,      MAT2X4, MAT3X4, MAT2X3, 0, &getInputDataD,      compareFP16ArithmeticFunc<  8, 16,  8,  0, fp16MatrixTimesMatrix<3,4,2,3> >     },
16837                 {       "OpMatrixTimesMatrix",  "3x4_3x3",      2,      MAT3X4, MAT3X4, MAT3X3, 0, &getInputDataD,      compareFP16ArithmeticFunc< 16, 16, 16,  0, fp16MatrixTimesMatrix<3,4,3,3> >     },
16838                 {       "OpMatrixTimesMatrix",  "3x4_4x3",      2,      MAT4X4, MAT3X4, MAT4X3, 0, &getInputDataD,      compareFP16ArithmeticFunc< 16, 16, 16,  0, fp16MatrixTimesMatrix<3,4,4,3> >     },
16839                 {       "OpMatrixTimesMatrix",  "4x2_2x4",      2,      MAT2X2, MAT4X2, MAT2X4, 0, &getInputDataD,      compareFP16ArithmeticFunc<  4,  8,  8,  0, fp16MatrixTimesMatrix<4,2,2,4> >     },
16840                 {       "OpMatrixTimesMatrix",  "4x2_3x4",      2,      MAT3X2, MAT4X2, MAT3X4, 0, &getInputDataD,      compareFP16ArithmeticFunc<  8,  8, 16,  0, fp16MatrixTimesMatrix<4,2,3,4> >     },
16841                 {       "OpMatrixTimesMatrix",  "4x2_4x4",      2,      MAT4X2, MAT4X2, MAT4X4, 0, &getInputDataD,      compareFP16ArithmeticFunc<  8,  8, 16,  0, fp16MatrixTimesMatrix<4,2,4,4> >     },
16842                 {       "OpMatrixTimesMatrix",  "4x3_2x4",      2,      MAT2X3, MAT4X3, MAT2X4, 0, &getInputDataD,      compareFP16ArithmeticFunc<  8, 16,  8,  0, fp16MatrixTimesMatrix<4,3,2,4> >     },
16843                 {       "OpMatrixTimesMatrix",  "4x3_3x4",      2,      MAT3X3, MAT4X3, MAT3X4, 0, &getInputDataD,      compareFP16ArithmeticFunc< 16, 16, 16,  0, fp16MatrixTimesMatrix<4,3,3,4> >     },
16844                 {       "OpMatrixTimesMatrix",  "4x3_4x4",      2,      MAT4X3, MAT4X3, MAT4X4, 0, &getInputDataD,      compareFP16ArithmeticFunc< 16, 16, 16,  0, fp16MatrixTimesMatrix<4,3,4,4> >     },
16845                 {       "OpMatrixTimesMatrix",  "4x4_2x4",      2,      MAT2X4, MAT4X4, MAT2X4, 0, &getInputDataD,      compareFP16ArithmeticFunc<  8, 16,  8,  0, fp16MatrixTimesMatrix<4,4,2,4> >     },
16846                 {       "OpMatrixTimesMatrix",  "4x4_3x4",      2,      MAT3X4, MAT4X4, MAT3X4, 0, &getInputDataD,      compareFP16ArithmeticFunc< 16, 16, 16,  0, fp16MatrixTimesMatrix<4,4,3,4> >     },
16847                 {       "OpMatrixTimesMatrix",  "4x4_4x4",      2,      MAT4X4, MAT4X4, MAT4X4, 0, &getInputDataD,      compareFP16ArithmeticFunc< 16, 16, 16,  0, fp16MatrixTimesMatrix<4,4,4,4> >     },
16848                 {       "OpOuterProduct",               "2x2",          2,      MAT2X2, VEC2,   VEC2,   0, &getInputDataD,      compareFP16ArithmeticFunc<  4,  2,  2,  0, fp16OuterProduct<2,2> >                      },
16849                 {       "OpOuterProduct",               "2x3",          2,      MAT2X3, VEC3,   VEC2,   0, &getInputDataD,      compareFP16ArithmeticFunc<  8,  3,  2,  0, fp16OuterProduct<2,3> >                      },
16850                 {       "OpOuterProduct",               "2x4",          2,      MAT2X4, VEC4,   VEC2,   0, &getInputDataD,      compareFP16ArithmeticFunc<  8,  4,  2,  0, fp16OuterProduct<2,4> >                      },
16851                 {       "OpOuterProduct",               "3x2",          2,      MAT3X2, VEC2,   VEC3,   0, &getInputDataD,      compareFP16ArithmeticFunc<  8,  2,  3,  0, fp16OuterProduct<3,2> >                      },
16852                 {       "OpOuterProduct",               "3x3",          2,      MAT3X3, VEC3,   VEC3,   0, &getInputDataD,      compareFP16ArithmeticFunc< 16,  3,  3,  0, fp16OuterProduct<3,3> >                      },
16853                 {       "OpOuterProduct",               "3x4",          2,      MAT3X4, VEC4,   VEC3,   0, &getInputDataD,      compareFP16ArithmeticFunc< 16,  4,  3,  0, fp16OuterProduct<3,4> >                      },
16854                 {       "OpOuterProduct",               "4x2",          2,      MAT4X2, VEC2,   VEC4,   0, &getInputDataD,      compareFP16ArithmeticFunc<  8,  2,  4,  0, fp16OuterProduct<4,2> >                      },
16855                 {       "OpOuterProduct",               "4x3",          2,      MAT4X3, VEC3,   VEC4,   0, &getInputDataD,      compareFP16ArithmeticFunc< 16,  3,  4,  0, fp16OuterProduct<4,3> >                      },
16856                 {       "OpOuterProduct",               "4x4",          2,      MAT4X4, VEC4,   VEC4,   0, &getInputDataD,      compareFP16ArithmeticFunc< 16,  4,  4,  0, fp16OuterProduct<4,4> >                      },
16857                 {       "Determinant",                  "2x2",          1,      SCALAR, MAT2X2, NONE,   0, &getInputDataC,      compareFP16ArithmeticFunc<  1,  4,  0,  0, fp16Determinant<2> >                         },
16858                 {       "Determinant",                  "3x3",          1,      SCALAR, MAT3X3, NONE,   0, &getInputDataC,      compareFP16ArithmeticFunc<  1, 16,  0,  0, fp16Determinant<3> >                         },
16859                 {       "Determinant",                  "4x4",          1,      SCALAR, MAT4X4, NONE,   0, &getInputDataC,      compareFP16ArithmeticFunc<  1, 16,  0,  0, fp16Determinant<4> >                         },
16860                 {       "MatrixInverse",                "2x2",          1,      MAT2X2, MAT2X2, NONE,   0, &getInputDataC,      compareFP16ArithmeticFunc<  4,  4,  0,  0, fp16Inverse<2> >                                     },
16861         };
16862
16863         for (deUint32 testFuncIdx = 0; testFuncIdx < DE_LENGTH_OF_ARRAY(testFuncs); ++testFuncIdx)
16864         {
16865                 const Math16TestFunc&   testFunc        = testFuncs[testFuncIdx];
16866
16867                 createFloat16ArithmeticFuncTest<SpecResource>(testCtx, *testGroup.get(), 0, testFunc);
16868         }
16869
16870         return testGroup.release();
16871 }
16872
16873 const string getNumberTypeName (const NumberType type)
16874 {
16875         if (type == NUMBERTYPE_INT32)
16876         {
16877                 return "int";
16878         }
16879         else if (type == NUMBERTYPE_UINT32)
16880         {
16881                 return "uint";
16882         }
16883         else if (type == NUMBERTYPE_FLOAT32)
16884         {
16885                 return "float";
16886         }
16887         else
16888         {
16889                 DE_ASSERT(false);
16890                 return "";
16891         }
16892 }
16893
16894 deInt32 getInt(de::Random& rnd)
16895 {
16896         return rnd.getInt(std::numeric_limits<int>::min(), std::numeric_limits<int>::max());
16897 }
16898
16899 const string repeatString (const string& str, int times)
16900 {
16901         string filler;
16902         for (int i = 0; i < times; ++i)
16903         {
16904                 filler += str;
16905         }
16906         return filler;
16907 }
16908
16909 const string getRandomConstantString (const NumberType type, de::Random& rnd)
16910 {
16911         if (type == NUMBERTYPE_INT32)
16912         {
16913                 return numberToString<deInt32>(getInt(rnd));
16914         }
16915         else if (type == NUMBERTYPE_UINT32)
16916         {
16917                 return numberToString<deUint32>(rnd.getUint32());
16918         }
16919         else if (type == NUMBERTYPE_FLOAT32)
16920         {
16921                 return numberToString<float>(rnd.getFloat());
16922         }
16923         else
16924         {
16925                 DE_ASSERT(false);
16926                 return "";
16927         }
16928 }
16929
16930 void createVectorCompositeCases (vector<map<string, string> >& testCases, de::Random& rnd, const NumberType type)
16931 {
16932         map<string, string> params;
16933
16934         // Vec2 to Vec4
16935         for (int width = 2; width <= 4; ++width)
16936         {
16937                 const string randomConst = numberToString(getInt(rnd));
16938                 const string widthStr = numberToString(width);
16939                 const string composite_type = "${customType}vec" + widthStr;
16940                 const int index = rnd.getInt(0, width-1);
16941
16942                 params["type"]                  = "vec";
16943                 params["name"]                  = params["type"] + "_" + widthStr;
16944                 params["compositeDecl"]         = composite_type + " = OpTypeVector ${customType} " + widthStr +"\n";
16945                 params["compositeType"]         = composite_type;
16946                 params["filler"]                = string("%filler    = OpConstant ${customType} ") + getRandomConstantString(type, rnd) + "\n";
16947                 params["compositeConstruct"]    = "%instance  = OpCompositeConstruct " + composite_type + repeatString(" %filler", width) + "\n";
16948                 params["indexes"]               = numberToString(index);
16949                 testCases.push_back(params);
16950         }
16951 }
16952
16953 void createArrayCompositeCases (vector<map<string, string> >& testCases, de::Random& rnd, const NumberType type)
16954 {
16955         const int limit = 10;
16956         map<string, string> params;
16957
16958         for (int width = 2; width <= limit; ++width)
16959         {
16960                 string randomConst = numberToString(getInt(rnd));
16961                 string widthStr = numberToString(width);
16962                 int index = rnd.getInt(0, width-1);
16963
16964                 params["type"]                  = "array";
16965                 params["name"]                  = params["type"] + "_" + widthStr;
16966                 params["compositeDecl"]         = string("%arraywidth = OpConstant %u32 " + widthStr + "\n")
16967                                                                                         +        "%composite = OpTypeArray ${customType} %arraywidth\n";
16968                 params["compositeType"]         = "%composite";
16969                 params["filler"]                = string("%filler    = OpConstant ${customType} ") + getRandomConstantString(type, rnd) + "\n";
16970                 params["compositeConstruct"]    = "%instance  = OpCompositeConstruct %composite" + repeatString(" %filler", width) + "\n";
16971                 params["indexes"]               = numberToString(index);
16972                 testCases.push_back(params);
16973         }
16974 }
16975
16976 void createStructCompositeCases (vector<map<string, string> >& testCases, de::Random& rnd, const NumberType type)
16977 {
16978         const int limit = 10;
16979         map<string, string> params;
16980
16981         for (int width = 2; width <= limit; ++width)
16982         {
16983                 string randomConst = numberToString(getInt(rnd));
16984                 int index = rnd.getInt(0, width-1);
16985
16986                 params["type"]                  = "struct";
16987                 params["name"]                  = params["type"] + "_" + numberToString(width);
16988                 params["compositeDecl"]         = "%composite = OpTypeStruct" + repeatString(" ${customType}", width) + "\n";
16989                 params["compositeType"]         = "%composite";
16990                 params["filler"]                = string("%filler    = OpConstant ${customType} ") + getRandomConstantString(type, rnd) + "\n";
16991                 params["compositeConstruct"]    = "%instance  = OpCompositeConstruct %composite" + repeatString(" %filler", width) + "\n";
16992                 params["indexes"]               = numberToString(index);
16993                 testCases.push_back(params);
16994         }
16995 }
16996
16997 void createMatrixCompositeCases (vector<map<string, string> >& testCases, de::Random& rnd, const NumberType type)
16998 {
16999         map<string, string> params;
17000
17001         // Vec2 to Vec4
17002         for (int width = 2; width <= 4; ++width)
17003         {
17004                 string widthStr = numberToString(width);
17005
17006                 for (int column = 2 ; column <= 4; ++column)
17007                 {
17008                         int index_0 = rnd.getInt(0, column-1);
17009                         int index_1 = rnd.getInt(0, width-1);
17010                         string columnStr = numberToString(column);
17011
17012                         params["type"]          = "matrix";
17013                         params["name"]          = params["type"] + "_" + widthStr + "x" + columnStr;
17014                         params["compositeDecl"] = string("%vectype   = OpTypeVector ${customType} " + widthStr + "\n")
17015                                                                                                 +        "%composite = OpTypeMatrix %vectype " + columnStr + "\n";
17016                         params["compositeType"] = "%composite";
17017
17018                         params["filler"]        = string("%filler    = OpConstant ${customType} ") + getRandomConstantString(type, rnd) + "\n"
17019                                                                                                 +        "%fillerVec = OpConstantComposite %vectype" + repeatString(" %filler", width) + "\n";
17020
17021                         params["compositeConstruct"]    = "%instance  = OpCompositeConstruct %composite" + repeatString(" %fillerVec", column) + "\n";
17022                         params["indexes"]       = numberToString(index_0) + " " + numberToString(index_1);
17023                         testCases.push_back(params);
17024                 }
17025         }
17026 }
17027
17028 void createCompositeCases (vector<map<string, string> >& testCases, de::Random& rnd, const NumberType type)
17029 {
17030         createVectorCompositeCases(testCases, rnd, type);
17031         createArrayCompositeCases(testCases, rnd, type);
17032         createStructCompositeCases(testCases, rnd, type);
17033         // Matrix only supports float types
17034         if (type == NUMBERTYPE_FLOAT32)
17035         {
17036                 createMatrixCompositeCases(testCases, rnd, type);
17037         }
17038 }
17039
17040 const string getAssemblyTypeDeclaration (const NumberType type)
17041 {
17042         switch (type)
17043         {
17044                 case NUMBERTYPE_INT32:          return "OpTypeInt 32 1";
17045                 case NUMBERTYPE_UINT32:         return "OpTypeInt 32 0";
17046                 case NUMBERTYPE_FLOAT32:        return "OpTypeFloat 32";
17047                 default:                        DE_ASSERT(false); return "";
17048         }
17049 }
17050
17051 const string getAssemblyTypeName (const NumberType type)
17052 {
17053         switch (type)
17054         {
17055                 case NUMBERTYPE_INT32:          return "%i32";
17056                 case NUMBERTYPE_UINT32:         return "%u32";
17057                 case NUMBERTYPE_FLOAT32:        return "%f32";
17058                 default:                        DE_ASSERT(false); return "";
17059         }
17060 }
17061
17062 const string specializeCompositeInsertShaderTemplate (const NumberType type, const map<string, string>& params)
17063 {
17064         map<string, string>     parameters(params);
17065
17066         const string customType = getAssemblyTypeName(type);
17067         map<string, string> substCustomType;
17068         substCustomType["customType"] = customType;
17069         parameters["compositeDecl"] = StringTemplate(parameters.at("compositeDecl")).specialize(substCustomType);
17070         parameters["compositeType"] = StringTemplate(parameters.at("compositeType")).specialize(substCustomType);
17071         parameters["compositeConstruct"] = StringTemplate(parameters.at("compositeConstruct")).specialize(substCustomType);
17072         parameters["filler"] = StringTemplate(parameters.at("filler")).specialize(substCustomType);
17073         parameters["customType"] = customType;
17074         parameters["compositeDecorator"] = (parameters["type"] == "array") ? "OpDecorate %composite ArrayStride 4\n" : "";
17075
17076         if (parameters.at("compositeType") != "%u32vec3")
17077         {
17078                 parameters["u32vec3Decl"] = "%u32vec3   = OpTypeVector %u32 3\n";
17079         }
17080
17081         return StringTemplate(
17082                 "OpCapability Shader\n"
17083                 "OpCapability Matrix\n"
17084                 "OpMemoryModel Logical GLSL450\n"
17085                 "OpEntryPoint GLCompute %main \"main\" %id\n"
17086                 "OpExecutionMode %main LocalSize 1 1 1\n"
17087
17088                 "OpSource GLSL 430\n"
17089                 "OpName %main           \"main\"\n"
17090                 "OpName %id             \"gl_GlobalInvocationID\"\n"
17091
17092                 // Decorators
17093                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
17094                 "OpDecorate %buf BufferBlock\n"
17095                 "OpDecorate %indata DescriptorSet 0\n"
17096                 "OpDecorate %indata Binding 0\n"
17097                 "OpDecorate %outdata DescriptorSet 0\n"
17098                 "OpDecorate %outdata Binding 1\n"
17099                 "OpDecorate %customarr ArrayStride 4\n"
17100                 "${compositeDecorator}"
17101                 "OpMemberDecorate %buf 0 Offset 0\n"
17102
17103                 // General types
17104                 "%void      = OpTypeVoid\n"
17105                 "%voidf     = OpTypeFunction %void\n"
17106                 "%u32       = OpTypeInt 32 0\n"
17107                 "%i32       = OpTypeInt 32 1\n"
17108                 "%f32       = OpTypeFloat 32\n"
17109
17110                 // Composite declaration
17111                 "${compositeDecl}"
17112
17113                 // Constants
17114                 "${filler}"
17115
17116                 "${u32vec3Decl:opt}"
17117                 "%uvec3ptr  = OpTypePointer Input %u32vec3\n"
17118
17119                 // Inherited from custom
17120                 "%customptr = OpTypePointer Uniform ${customType}\n"
17121                 "%customarr = OpTypeRuntimeArray ${customType}\n"
17122                 "%buf       = OpTypeStruct %customarr\n"
17123                 "%bufptr    = OpTypePointer Uniform %buf\n"
17124
17125                 "%indata    = OpVariable %bufptr Uniform\n"
17126                 "%outdata   = OpVariable %bufptr Uniform\n"
17127
17128                 "%id        = OpVariable %uvec3ptr Input\n"
17129                 "%zero      = OpConstant %i32 0\n"
17130
17131                 "%main      = OpFunction %void None %voidf\n"
17132                 "%label     = OpLabel\n"
17133                 "%idval     = OpLoad %u32vec3 %id\n"
17134                 "%x         = OpCompositeExtract %u32 %idval 0\n"
17135
17136                 "%inloc     = OpAccessChain %customptr %indata %zero %x\n"
17137                 "%outloc    = OpAccessChain %customptr %outdata %zero %x\n"
17138                 // Read the input value
17139                 "%inval     = OpLoad ${customType} %inloc\n"
17140                 // Create the composite and fill it
17141                 "${compositeConstruct}"
17142                 // Insert the input value to a place
17143                 "%instance2 = OpCompositeInsert ${compositeType} %inval %instance ${indexes}\n"
17144                 // Read back the value from the position
17145                 "%out_val   = OpCompositeExtract ${customType} %instance2 ${indexes}\n"
17146                 // Store it in the output position
17147                 "             OpStore %outloc %out_val\n"
17148                 "             OpReturn\n"
17149                 "             OpFunctionEnd\n"
17150         ).specialize(parameters);
17151 }
17152
17153 template<typename T>
17154 BufferSp createCompositeBuffer(T number)
17155 {
17156         return BufferSp(new Buffer<T>(vector<T>(1, number)));
17157 }
17158
17159 tcu::TestCaseGroup* createOpCompositeInsertGroup (tcu::TestContext& testCtx)
17160 {
17161         de::MovePtr<tcu::TestCaseGroup> group   (new tcu::TestCaseGroup(testCtx, "opcompositeinsert", "Test the OpCompositeInsert instruction"));
17162         de::Random                                              rnd             (deStringHash(group->getName()));
17163
17164         for (int type = NUMBERTYPE_INT32; type != NUMBERTYPE_END32; ++type)
17165         {
17166                 NumberType                                              numberType              = NumberType(type);
17167                 const string                                    typeName                = getNumberTypeName(numberType);
17168                 const string                                    description             = "Test the OpCompositeInsert instruction with " + typeName + "s";
17169                 de::MovePtr<tcu::TestCaseGroup> subGroup                (new tcu::TestCaseGroup(testCtx, typeName.c_str(), description.c_str()));
17170                 vector<map<string, string> >    testCases;
17171
17172                 createCompositeCases(testCases, rnd, numberType);
17173
17174                 for (vector<map<string, string> >::const_iterator test = testCases.begin(); test != testCases.end(); ++test)
17175                 {
17176                         ComputeShaderSpec       spec;
17177
17178                         spec.assembly = specializeCompositeInsertShaderTemplate(numberType, *test);
17179
17180                         switch (numberType)
17181                         {
17182                                 case NUMBERTYPE_INT32:
17183                                 {
17184                                         deInt32 number = getInt(rnd);
17185                                         spec.inputs.push_back(createCompositeBuffer<deInt32>(number));
17186                                         spec.outputs.push_back(createCompositeBuffer<deInt32>(number));
17187                                         break;
17188                                 }
17189                                 case NUMBERTYPE_UINT32:
17190                                 {
17191                                         deUint32 number = rnd.getUint32();
17192                                         spec.inputs.push_back(createCompositeBuffer<deUint32>(number));
17193                                         spec.outputs.push_back(createCompositeBuffer<deUint32>(number));
17194                                         break;
17195                                 }
17196                                 case NUMBERTYPE_FLOAT32:
17197                                 {
17198                                         float number = rnd.getFloat();
17199                                         spec.inputs.push_back(createCompositeBuffer<float>(number));
17200                                         spec.outputs.push_back(createCompositeBuffer<float>(number));
17201                                         break;
17202                                 }
17203                                 default:
17204                                         DE_ASSERT(false);
17205                         }
17206
17207                         spec.numWorkGroups = IVec3(1, 1, 1);
17208                         subGroup->addChild(new SpvAsmComputeShaderCase(testCtx, test->at("name").c_str(), "OpCompositeInsert test", spec));
17209                 }
17210                 group->addChild(subGroup.release());
17211         }
17212         return group.release();
17213 }
17214
17215 struct AssemblyStructInfo
17216 {
17217         AssemblyStructInfo (const deUint32 comp, const deUint32 idx)
17218         : components    (comp)
17219         , index                 (idx)
17220         {}
17221
17222         deUint32 components;
17223         deUint32 index;
17224 };
17225
17226 const string specializeInBoundsShaderTemplate (const NumberType type, const AssemblyStructInfo& structInfo, const map<string, string>& params)
17227 {
17228         // Create the full index string
17229         string                          fullIndex       = numberToString(structInfo.index) + " " + params.at("indexes");
17230         // Convert it to list of indexes
17231         vector<string>          indexes         = de::splitString(fullIndex, ' ');
17232
17233         map<string, string>     parameters      (params);
17234         parameters["structType"]        = repeatString(" ${compositeType}", structInfo.components);
17235         parameters["structConstruct"]   = repeatString(" %instance", structInfo.components);
17236         parameters["insertIndexes"]     = fullIndex;
17237
17238         // In matrix cases the last two index is the CompositeExtract indexes
17239         const deUint32 extractIndexes = (parameters["type"] == "matrix") ? 2 : 1;
17240
17241         // Construct the extractIndex
17242         for (vector<string>::const_iterator index = indexes.end() - extractIndexes; index != indexes.end(); ++index)
17243         {
17244                 parameters["extractIndexes"] += " " + *index;
17245         }
17246
17247         // Remove the last 1 or 2 element depends on matrix case or not
17248         indexes.erase(indexes.end() - extractIndexes, indexes.end());
17249
17250         deUint32 id = 0;
17251         // Generate AccessChain index expressions (except for the last one, because we use ptr to the composite)
17252         for (vector<string>::const_iterator index = indexes.begin(); index != indexes.end(); ++index)
17253         {
17254                 string indexId = "%index_" + numberToString(id++);
17255                 parameters["accessChainConstDeclaration"] += indexId + "   = OpConstant %u32 " + *index + "\n";
17256                 parameters["accessChainIndexes"] += " " + indexId;
17257         }
17258
17259         parameters["compositeDecorator"] = (parameters["type"] == "array") ? "OpDecorate %composite ArrayStride 4\n" : "";
17260
17261         const string customType = getAssemblyTypeName(type);
17262         map<string, string> substCustomType;
17263         substCustomType["customType"] = customType;
17264         parameters["compositeDecl"] = StringTemplate(parameters.at("compositeDecl")).specialize(substCustomType);
17265         parameters["compositeType"] = StringTemplate(parameters.at("compositeType")).specialize(substCustomType);
17266         parameters["compositeConstruct"] = StringTemplate(parameters.at("compositeConstruct")).specialize(substCustomType);
17267         parameters["filler"] = StringTemplate(parameters.at("filler")).specialize(substCustomType);
17268         parameters["customType"] = customType;
17269
17270         const string compositeType = parameters.at("compositeType");
17271         map<string, string> substCompositeType;
17272         substCompositeType["compositeType"] = compositeType;
17273         parameters["structType"] = StringTemplate(parameters.at("structType")).specialize(substCompositeType);
17274         if (compositeType != "%u32vec3")
17275         {
17276                 parameters["u32vec3Decl"] = "%u32vec3   = OpTypeVector %u32 3\n";
17277         }
17278
17279         return StringTemplate(
17280                 "OpCapability Shader\n"
17281                 "OpCapability Matrix\n"
17282                 "OpMemoryModel Logical GLSL450\n"
17283                 "OpEntryPoint GLCompute %main \"main\" %id\n"
17284                 "OpExecutionMode %main LocalSize 1 1 1\n"
17285
17286                 "OpSource GLSL 430\n"
17287                 "OpName %main           \"main\"\n"
17288                 "OpName %id             \"gl_GlobalInvocationID\"\n"
17289                 // Decorators
17290                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
17291                 "OpDecorate %buf BufferBlock\n"
17292                 "OpDecorate %indata DescriptorSet 0\n"
17293                 "OpDecorate %indata Binding 0\n"
17294                 "OpDecorate %outdata DescriptorSet 0\n"
17295                 "OpDecorate %outdata Binding 1\n"
17296                 "OpDecorate %customarr ArrayStride 4\n"
17297                 "${compositeDecorator}"
17298                 "OpMemberDecorate %buf 0 Offset 0\n"
17299                 // General types
17300                 "%void      = OpTypeVoid\n"
17301                 "%voidf     = OpTypeFunction %void\n"
17302                 "%i32       = OpTypeInt 32 1\n"
17303                 "%u32       = OpTypeInt 32 0\n"
17304                 "%f32       = OpTypeFloat 32\n"
17305                 // Custom types
17306                 "${compositeDecl}"
17307                 // %u32vec3 if not already declared in ${compositeDecl}
17308                 "${u32vec3Decl:opt}"
17309                 "%uvec3ptr  = OpTypePointer Input %u32vec3\n"
17310                 // Inherited from composite
17311                 "%composite_p = OpTypePointer Function ${compositeType}\n"
17312                 "%struct_t  = OpTypeStruct${structType}\n"
17313                 "%struct_p  = OpTypePointer Function %struct_t\n"
17314                 // Constants
17315                 "${filler}"
17316                 "${accessChainConstDeclaration}"
17317                 // Inherited from custom
17318                 "%customptr = OpTypePointer Uniform ${customType}\n"
17319                 "%customarr = OpTypeRuntimeArray ${customType}\n"
17320                 "%buf       = OpTypeStruct %customarr\n"
17321                 "%bufptr    = OpTypePointer Uniform %buf\n"
17322                 "%indata    = OpVariable %bufptr Uniform\n"
17323                 "%outdata   = OpVariable %bufptr Uniform\n"
17324
17325                 "%id        = OpVariable %uvec3ptr Input\n"
17326                 "%zero      = OpConstant %u32 0\n"
17327                 "%main      = OpFunction %void None %voidf\n"
17328                 "%label     = OpLabel\n"
17329                 "%struct_v  = OpVariable %struct_p Function\n"
17330                 "%idval     = OpLoad %u32vec3 %id\n"
17331                 "%x         = OpCompositeExtract %u32 %idval 0\n"
17332                 // Create the input/output type
17333                 "%inloc     = OpInBoundsAccessChain %customptr %indata %zero %x\n"
17334                 "%outloc    = OpInBoundsAccessChain %customptr %outdata %zero %x\n"
17335                 // Read the input value
17336                 "%inval     = OpLoad ${customType} %inloc\n"
17337                 // Create the composite and fill it
17338                 "${compositeConstruct}"
17339                 // Create the struct and fill it with the composite
17340                 "%struct    = OpCompositeConstruct %struct_t${structConstruct}\n"
17341                 // Insert the value
17342                 "%comp_obj  = OpCompositeInsert %struct_t %inval %struct ${insertIndexes}\n"
17343                 // Store the object
17344                 "             OpStore %struct_v %comp_obj\n"
17345                 // Get deepest possible composite pointer
17346                 "%inner_ptr = OpInBoundsAccessChain %composite_p %struct_v${accessChainIndexes}\n"
17347                 "%read_obj  = OpLoad ${compositeType} %inner_ptr\n"
17348                 // Read back the stored value
17349                 "%read_val  = OpCompositeExtract ${customType} %read_obj${extractIndexes}\n"
17350                 "             OpStore %outloc %read_val\n"
17351                 "             OpReturn\n"
17352                 "             OpFunctionEnd\n"
17353         ).specialize(parameters);
17354 }
17355
17356 tcu::TestCaseGroup* createOpInBoundsAccessChainGroup (tcu::TestContext& testCtx)
17357 {
17358         de::MovePtr<tcu::TestCaseGroup> group                   (new tcu::TestCaseGroup(testCtx, "opinboundsaccesschain", "Test the OpInBoundsAccessChain instruction"));
17359         de::Random                                              rnd                             (deStringHash(group->getName()));
17360
17361         for (int type = NUMBERTYPE_INT32; type != NUMBERTYPE_END32; ++type)
17362         {
17363                 NumberType                                              numberType      = NumberType(type);
17364                 const string                                    typeName        = getNumberTypeName(numberType);
17365                 const string                                    description     = "Test the OpInBoundsAccessChain instruction with " + typeName + "s";
17366                 de::MovePtr<tcu::TestCaseGroup> subGroup        (new tcu::TestCaseGroup(testCtx, typeName.c_str(), description.c_str()));
17367
17368                 vector<map<string, string> >    testCases;
17369                 createCompositeCases(testCases, rnd, numberType);
17370
17371                 for (vector<map<string, string> >::const_iterator test = testCases.begin(); test != testCases.end(); ++test)
17372                 {
17373                         ComputeShaderSpec       spec;
17374
17375                         // Number of components inside of a struct
17376                         deUint32 structComponents = rnd.getInt(2, 8);
17377                         // Component index value
17378                         deUint32 structIndex = rnd.getInt(0, structComponents - 1);
17379                         AssemblyStructInfo structInfo(structComponents, structIndex);
17380
17381                         spec.assembly = specializeInBoundsShaderTemplate(numberType, structInfo, *test);
17382
17383                         switch (numberType)
17384                         {
17385                                 case NUMBERTYPE_INT32:
17386                                 {
17387                                         deInt32 number = getInt(rnd);
17388                                         spec.inputs.push_back(createCompositeBuffer<deInt32>(number));
17389                                         spec.outputs.push_back(createCompositeBuffer<deInt32>(number));
17390                                         break;
17391                                 }
17392                                 case NUMBERTYPE_UINT32:
17393                                 {
17394                                         deUint32 number = rnd.getUint32();
17395                                         spec.inputs.push_back(createCompositeBuffer<deUint32>(number));
17396                                         spec.outputs.push_back(createCompositeBuffer<deUint32>(number));
17397                                         break;
17398                                 }
17399                                 case NUMBERTYPE_FLOAT32:
17400                                 {
17401                                         float number = rnd.getFloat();
17402                                         spec.inputs.push_back(createCompositeBuffer<float>(number));
17403                                         spec.outputs.push_back(createCompositeBuffer<float>(number));
17404                                         break;
17405                                 }
17406                                 default:
17407                                         DE_ASSERT(false);
17408                         }
17409                         spec.numWorkGroups = IVec3(1, 1, 1);
17410                         subGroup->addChild(new SpvAsmComputeShaderCase(testCtx, test->at("name").c_str(), "OpInBoundsAccessChain test", spec));
17411                 }
17412                 group->addChild(subGroup.release());
17413         }
17414         return group.release();
17415 }
17416
17417 // If the params missing, uninitialized case
17418 const string specializeDefaultOutputShaderTemplate (const NumberType type, const map<string, string>& params = map<string, string>())
17419 {
17420         map<string, string> parameters(params);
17421
17422         parameters["customType"]        = getAssemblyTypeName(type);
17423
17424         // Declare the const value, and use it in the initializer
17425         if (params.find("constValue") != params.end())
17426         {
17427                 parameters["variableInitializer"]       = " %const";
17428         }
17429         // Uninitialized case
17430         else
17431         {
17432                 parameters["commentDecl"]       = ";";
17433         }
17434
17435         return StringTemplate(
17436                 "OpCapability Shader\n"
17437                 "OpMemoryModel Logical GLSL450\n"
17438                 "OpEntryPoint GLCompute %main \"main\" %id\n"
17439                 "OpExecutionMode %main LocalSize 1 1 1\n"
17440                 "OpSource GLSL 430\n"
17441                 "OpName %main           \"main\"\n"
17442                 "OpName %id             \"gl_GlobalInvocationID\"\n"
17443                 // Decorators
17444                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
17445                 "OpDecorate %indata DescriptorSet 0\n"
17446                 "OpDecorate %indata Binding 0\n"
17447                 "OpDecorate %outdata DescriptorSet 0\n"
17448                 "OpDecorate %outdata Binding 1\n"
17449                 "OpDecorate %in_arr ArrayStride 4\n"
17450                 "OpDecorate %in_buf BufferBlock\n"
17451                 "OpMemberDecorate %in_buf 0 Offset 0\n"
17452                 // Base types
17453                 "%void       = OpTypeVoid\n"
17454                 "%voidf      = OpTypeFunction %void\n"
17455                 "%u32        = OpTypeInt 32 0\n"
17456                 "%i32        = OpTypeInt 32 1\n"
17457                 "%f32        = OpTypeFloat 32\n"
17458                 "%uvec3      = OpTypeVector %u32 3\n"
17459                 "%uvec3ptr   = OpTypePointer Input %uvec3\n"
17460                 "${commentDecl:opt}%const      = OpConstant ${customType} ${constValue:opt}\n"
17461                 // Derived types
17462                 "%in_ptr     = OpTypePointer Uniform ${customType}\n"
17463                 "%in_arr     = OpTypeRuntimeArray ${customType}\n"
17464                 "%in_buf     = OpTypeStruct %in_arr\n"
17465                 "%in_bufptr  = OpTypePointer Uniform %in_buf\n"
17466                 "%indata     = OpVariable %in_bufptr Uniform\n"
17467                 "%outdata    = OpVariable %in_bufptr Uniform\n"
17468                 "%id         = OpVariable %uvec3ptr Input\n"
17469                 "%var_ptr    = OpTypePointer Function ${customType}\n"
17470                 // Constants
17471                 "%zero       = OpConstant %i32 0\n"
17472                 // Main function
17473                 "%main       = OpFunction %void None %voidf\n"
17474                 "%label      = OpLabel\n"
17475                 "%out_var    = OpVariable %var_ptr Function${variableInitializer:opt}\n"
17476                 "%idval      = OpLoad %uvec3 %id\n"
17477                 "%x          = OpCompositeExtract %u32 %idval 0\n"
17478                 "%inloc      = OpAccessChain %in_ptr %indata %zero %x\n"
17479                 "%outloc     = OpAccessChain %in_ptr %outdata %zero %x\n"
17480
17481                 "%outval     = OpLoad ${customType} %out_var\n"
17482                 "              OpStore %outloc %outval\n"
17483                 "              OpReturn\n"
17484                 "              OpFunctionEnd\n"
17485         ).specialize(parameters);
17486 }
17487
17488 bool compareFloats (const std::vector<Resource>&, const vector<AllocationSp>& outputAllocs, const std::vector<Resource>& expectedOutputs, TestLog& log)
17489 {
17490         DE_ASSERT(outputAllocs.size() != 0);
17491         DE_ASSERT(outputAllocs.size() == expectedOutputs.size());
17492
17493         // Use custom epsilon because of the float->string conversion
17494         const float     epsilon = 0.00001f;
17495
17496         for (size_t outputNdx = 0; outputNdx < outputAllocs.size(); ++outputNdx)
17497         {
17498                 vector<deUint8> expectedBytes;
17499                 float                   expected;
17500                 float                   actual;
17501
17502                 expectedOutputs[outputNdx].getBytes(expectedBytes);
17503                 memcpy(&expected, &expectedBytes.front(), expectedBytes.size());
17504                 memcpy(&actual, outputAllocs[outputNdx]->getHostPtr(), expectedBytes.size());
17505
17506                 // Test with epsilon
17507                 if (fabs(expected - actual) > epsilon)
17508                 {
17509                         log << TestLog::Message << "Error: The actual and expected values not matching."
17510                                 << " Expected: " << expected << " Actual: " << actual << " Epsilon: " << epsilon << TestLog::EndMessage;
17511                         return false;
17512                 }
17513         }
17514         return true;
17515 }
17516
17517 // Checks if the driver crash with uninitialized cases
17518 bool passthruVerify (const std::vector<Resource>&, const vector<AllocationSp>& outputAllocs, const std::vector<Resource>& expectedOutputs, TestLog&)
17519 {
17520         DE_ASSERT(outputAllocs.size() != 0);
17521         DE_ASSERT(outputAllocs.size() == expectedOutputs.size());
17522
17523         // Copy and discard the result.
17524         for (size_t outputNdx = 0; outputNdx < outputAllocs.size(); ++outputNdx)
17525         {
17526                 vector<deUint8> expectedBytes;
17527                 expectedOutputs[outputNdx].getBytes(expectedBytes);
17528
17529                 const size_t    width                   = expectedBytes.size();
17530                 vector<char>    data                    (width);
17531
17532                 memcpy(&data[0], outputAllocs[outputNdx]->getHostPtr(), width);
17533         }
17534         return true;
17535 }
17536
17537 tcu::TestCaseGroup* createShaderDefaultOutputGroup (tcu::TestContext& testCtx)
17538 {
17539         de::MovePtr<tcu::TestCaseGroup> group   (new tcu::TestCaseGroup(testCtx, "shader_default_output", "Test shader default output."));
17540         de::Random                                              rnd             (deStringHash(group->getName()));
17541
17542         for (int type = NUMBERTYPE_INT32; type != NUMBERTYPE_END32; ++type)
17543         {
17544                 NumberType                                              numberType      = NumberType(type);
17545                 const string                                    typeName        = getNumberTypeName(numberType);
17546                 const string                                    description     = "Test the OpVariable initializer with " + typeName + ".";
17547                 de::MovePtr<tcu::TestCaseGroup> subGroup        (new tcu::TestCaseGroup(testCtx, typeName.c_str(), description.c_str()));
17548
17549                 // 2 similar subcases (initialized and uninitialized)
17550                 for (int subCase = 0; subCase < 2; ++subCase)
17551                 {
17552                         ComputeShaderSpec spec;
17553                         spec.numWorkGroups = IVec3(1, 1, 1);
17554
17555                         map<string, string>                             params;
17556
17557                         switch (numberType)
17558                         {
17559                                 case NUMBERTYPE_INT32:
17560                                 {
17561                                         deInt32 number = getInt(rnd);
17562                                         spec.inputs.push_back(createCompositeBuffer<deInt32>(number));
17563                                         spec.outputs.push_back(createCompositeBuffer<deInt32>(number));
17564                                         params["constValue"] = numberToString(number);
17565                                         break;
17566                                 }
17567                                 case NUMBERTYPE_UINT32:
17568                                 {
17569                                         deUint32 number = rnd.getUint32();
17570                                         spec.inputs.push_back(createCompositeBuffer<deUint32>(number));
17571                                         spec.outputs.push_back(createCompositeBuffer<deUint32>(number));
17572                                         params["constValue"] = numberToString(number);
17573                                         break;
17574                                 }
17575                                 case NUMBERTYPE_FLOAT32:
17576                                 {
17577                                         float number = rnd.getFloat();
17578                                         spec.inputs.push_back(createCompositeBuffer<float>(number));
17579                                         spec.outputs.push_back(createCompositeBuffer<float>(number));
17580                                         spec.verifyIO = &compareFloats;
17581                                         params["constValue"] = numberToString(number);
17582                                         break;
17583                                 }
17584                                 default:
17585                                         DE_ASSERT(false);
17586                         }
17587
17588                         // Initialized subcase
17589                         if (!subCase)
17590                         {
17591                                 spec.assembly = specializeDefaultOutputShaderTemplate(numberType, params);
17592                                 subGroup->addChild(new SpvAsmComputeShaderCase(testCtx, "initialized", "OpVariable initializer tests.", spec));
17593                         }
17594                         // Uninitialized subcase
17595                         else
17596                         {
17597                                 spec.assembly = specializeDefaultOutputShaderTemplate(numberType);
17598                                 spec.verifyIO = &passthruVerify;
17599                                 subGroup->addChild(new SpvAsmComputeShaderCase(testCtx, "uninitialized", "OpVariable initializer tests.", spec));
17600                         }
17601                 }
17602                 group->addChild(subGroup.release());
17603         }
17604         return group.release();
17605 }
17606
17607 tcu::TestCaseGroup* createOpNopTests (tcu::TestContext& testCtx)
17608 {
17609         de::MovePtr<tcu::TestCaseGroup> testGroup (new tcu::TestCaseGroup(testCtx, "opnop", "Test OpNop"));
17610         RGBA                                                    defaultColors[4];
17611         map<string, string>                             opNopFragments;
17612
17613         getDefaultColors(defaultColors);
17614
17615         opNopFragments["testfun"]               =
17616                 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
17617                 "%param1 = OpFunctionParameter %v4f32\n"
17618                 "%label_testfun = OpLabel\n"
17619                 "OpNop\n"
17620                 "OpNop\n"
17621                 "OpNop\n"
17622                 "OpNop\n"
17623                 "OpNop\n"
17624                 "OpNop\n"
17625                 "OpNop\n"
17626                 "OpNop\n"
17627                 "%a = OpVectorExtractDynamic %f32 %param1 %c_i32_0\n"
17628                 "%b = OpFAdd %f32 %a %a\n"
17629                 "OpNop\n"
17630                 "%c = OpFSub %f32 %b %a\n"
17631                 "%ret = OpVectorInsertDynamic %v4f32 %param1 %c %c_i32_0\n"
17632                 "OpNop\n"
17633                 "OpNop\n"
17634                 "OpReturnValue %ret\n"
17635                 "OpFunctionEnd\n";
17636
17637         createTestsForAllStages("opnop", defaultColors, defaultColors, opNopFragments, testGroup.get());
17638
17639         return testGroup.release();
17640 }
17641
17642 tcu::TestCaseGroup* createOpNameTests (tcu::TestContext& testCtx)
17643 {
17644         de::MovePtr<tcu::TestCaseGroup> testGroup       (new tcu::TestCaseGroup(testCtx, "opname","Test OpName"));
17645         RGBA                                                    defaultColors[4];
17646         map<string, string>                             opNameFragments;
17647
17648         getDefaultColors(defaultColors);
17649
17650         opNameFragments["testfun"] =
17651                 "%test_code  = OpFunction %v4f32 None %v4f32_v4f32_function\n"
17652                 "%param1     = OpFunctionParameter %v4f32\n"
17653                 "%label_func = OpLabel\n"
17654                 "%a          = OpVectorExtractDynamic %f32 %param1 %c_i32_0\n"
17655                 "%b          = OpFAdd %f32 %a %a\n"
17656                 "%c          = OpFSub %f32 %b %a\n"
17657                 "%ret        = OpVectorInsertDynamic %v4f32 %param1 %c %c_i32_0\n"
17658                 "OpReturnValue %ret\n"
17659                 "OpFunctionEnd\n";
17660
17661         opNameFragments["debug"] =
17662                 "OpName %BP_main \"not_main\"";
17663
17664         createTestsForAllStages("opname", defaultColors, defaultColors, opNameFragments, testGroup.get());
17665
17666         return testGroup.release();
17667 }
17668
17669 tcu::TestCaseGroup* createFloat16Tests (tcu::TestContext& testCtx)
17670 {
17671         de::MovePtr<tcu::TestCaseGroup>         testGroup                       (new tcu::TestCaseGroup(testCtx, "float16", "Float 16 tests"));
17672
17673         testGroup->addChild(createOpConstantFloat16Tests(testCtx));
17674         testGroup->addChild(createFloat16LogicalSet<GraphicsResources>(testCtx, TEST_WITH_NAN));
17675         testGroup->addChild(createFloat16LogicalSet<GraphicsResources>(testCtx, TEST_WITHOUT_NAN));
17676         testGroup->addChild(createFloat16FuncSet<GraphicsResources>(testCtx));
17677         testGroup->addChild(createDerivativeTests<256, 1>(testCtx));
17678         testGroup->addChild(createDerivativeTests<256, 2>(testCtx));
17679         testGroup->addChild(createDerivativeTests<256, 4>(testCtx));
17680         testGroup->addChild(createFloat16VectorExtractSet<GraphicsResources>(testCtx));
17681         testGroup->addChild(createFloat16VectorInsertSet<GraphicsResources>(testCtx));
17682         testGroup->addChild(createFloat16VectorShuffleSet<GraphicsResources>(testCtx));
17683         testGroup->addChild(createFloat16CompositeConstructSet<GraphicsResources>(testCtx));
17684         testGroup->addChild(createFloat16CompositeInsertExtractSet<GraphicsResources>(testCtx, "OpCompositeExtract"));
17685         testGroup->addChild(createFloat16CompositeInsertExtractSet<GraphicsResources>(testCtx, "OpCompositeInsert"));
17686         testGroup->addChild(createFloat16ArithmeticSet<GraphicsResources>(testCtx));
17687         testGroup->addChild(createFloat16ArithmeticSet<1, GraphicsResources>(testCtx));
17688         testGroup->addChild(createFloat16ArithmeticSet<2, GraphicsResources>(testCtx));
17689         testGroup->addChild(createFloat16ArithmeticSet<3, GraphicsResources>(testCtx));
17690         testGroup->addChild(createFloat16ArithmeticSet<4, GraphicsResources>(testCtx));
17691
17692         return testGroup.release();
17693 }
17694
17695 tcu::TestCaseGroup* createFloat16Group (tcu::TestContext& testCtx)
17696 {
17697         de::MovePtr<tcu::TestCaseGroup>         testGroup                       (new tcu::TestCaseGroup(testCtx, "float16", "Float 16 tests"));
17698
17699         testGroup->addChild(createFloat16OpConstantCompositeGroup(testCtx));
17700         testGroup->addChild(createFloat16LogicalSet<ComputeShaderSpec>(testCtx, TEST_WITH_NAN));
17701         testGroup->addChild(createFloat16LogicalSet<ComputeShaderSpec>(testCtx, TEST_WITHOUT_NAN));
17702         testGroup->addChild(createFloat16FuncSet<ComputeShaderSpec>(testCtx));
17703         testGroup->addChild(createFloat16VectorExtractSet<ComputeShaderSpec>(testCtx));
17704         testGroup->addChild(createFloat16VectorInsertSet<ComputeShaderSpec>(testCtx));
17705         testGroup->addChild(createFloat16VectorShuffleSet<ComputeShaderSpec>(testCtx));
17706         testGroup->addChild(createFloat16CompositeConstructSet<ComputeShaderSpec>(testCtx));
17707         testGroup->addChild(createFloat16CompositeInsertExtractSet<ComputeShaderSpec>(testCtx, "OpCompositeExtract"));
17708         testGroup->addChild(createFloat16CompositeInsertExtractSet<ComputeShaderSpec>(testCtx, "OpCompositeInsert"));
17709         testGroup->addChild(createFloat16ArithmeticSet<ComputeShaderSpec>(testCtx));
17710         testGroup->addChild(createFloat16ArithmeticSet<1, ComputeShaderSpec>(testCtx));
17711         testGroup->addChild(createFloat16ArithmeticSet<2, ComputeShaderSpec>(testCtx));
17712         testGroup->addChild(createFloat16ArithmeticSet<3, ComputeShaderSpec>(testCtx));
17713         testGroup->addChild(createFloat16ArithmeticSet<4, ComputeShaderSpec>(testCtx));
17714
17715         return testGroup.release();
17716 }
17717
17718 tcu::TestCaseGroup* createBoolMixedBitSizeGroup (tcu::TestContext& testCtx)
17719 {
17720         de::MovePtr<tcu::TestCaseGroup> group                   (new tcu::TestCaseGroup(testCtx, "mixed_bitsize", "Tests boolean operands produced from instructions of different bit-sizes"));
17721
17722         de::Random                                              rnd                             (deStringHash(group->getName()));
17723         const int               numElements             = 100;
17724         vector<float>   inputData               (numElements, 0);
17725         vector<float>   outputData              (numElements, 0);
17726         fillRandomScalars(rnd, 0.0f, 100.0f, &inputData[0], 100);
17727
17728         const StringTemplate                    shaderTemplate  (
17729                 "${CAPS}\n"
17730                 "OpMemoryModel Logical GLSL450\n"
17731                 "OpEntryPoint GLCompute %main \"main\" %id\n"
17732                 "OpExecutionMode %main LocalSize 1 1 1\n"
17733                 "OpSource GLSL 430\n"
17734                 "OpName %main           \"main\"\n"
17735                 "OpName %id             \"gl_GlobalInvocationID\"\n"
17736
17737                 "OpDecorate %id BuiltIn GlobalInvocationId\n"
17738
17739                 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) + string(getComputeAsmInputOutputBuffer()) +
17740
17741                 "%id        = OpVariable %uvec3ptr Input\n"
17742                 "${CONST}\n"
17743                 "%main      = OpFunction %void None %voidf\n"
17744                 "%label     = OpLabel\n"
17745                 "%idval     = OpLoad %uvec3 %id\n"
17746                 "%x         = OpCompositeExtract %u32 %idval 0\n"
17747                 "%inloc     = OpAccessChain %f32ptr %indata %c0i32 %x\n"
17748
17749                 "${TEST}\n"
17750
17751                 "%outloc    = OpAccessChain %f32ptr %outdata %c0i32 %x\n"
17752                 "             OpStore %outloc %res\n"
17753                 "             OpReturn\n"
17754                 "             OpFunctionEnd\n"
17755         );
17756
17757         // Each test case produces 4 boolean values, and we want each of these values
17758         // to come froma different combination of the available bit-sizes, so compute
17759         // all possible combinations here.
17760         vector<deUint32>        widths;
17761         widths.push_back(32);
17762         widths.push_back(16);
17763         widths.push_back(8);
17764
17765         vector<IVec4>   cases;
17766         for (size_t width0 = 0; width0 < widths.size(); width0++)
17767         {
17768                 for (size_t width1 = 0; width1 < widths.size(); width1++)
17769                 {
17770                         for (size_t width2 = 0; width2 < widths.size(); width2++)
17771                         {
17772                                 for (size_t width3 = 0; width3 < widths.size(); width3++)
17773                                 {
17774                                         cases.push_back(IVec4(widths[width0], widths[width1], widths[width2], widths[width3]));
17775                                 }
17776                         }
17777                 }
17778         }
17779
17780         for (size_t caseNdx = 0; caseNdx < cases.size(); caseNdx++)
17781         {
17782                 /// Skip cases where all bitsizes are the same, we are only interested in testing booleans produced from instructions with different native bit-sizes
17783                 if (cases[caseNdx][0] == cases[caseNdx][1] && cases[caseNdx][0] == cases[caseNdx][2] && cases[caseNdx][0] == cases[caseNdx][3])
17784                         continue;
17785
17786                 map<string, string>     specializations;
17787                 ComputeShaderSpec       spec;
17788
17789                 // Inject appropriate capabilities and reference constants depending
17790                 // on the bit-sizes required by this test case
17791                 bool hasFloat32 = cases[caseNdx][0] == 32 || cases[caseNdx][1] == 32 || cases[caseNdx][2] == 32 || cases[caseNdx][3] == 32;
17792                 bool hasFloat16 = cases[caseNdx][0] == 16 || cases[caseNdx][1] == 16 || cases[caseNdx][2] == 16 || cases[caseNdx][3] == 16;
17793                 bool hasInt8    = cases[caseNdx][0] == 8 || cases[caseNdx][1] == 8 || cases[caseNdx][2] == 8 || cases[caseNdx][3] == 8;
17794
17795                 string capsStr  = "OpCapability Shader\n";
17796                 string constStr =
17797                         "%c0i32     = OpConstant %i32 0\n"
17798                         "%c1f32     = OpConstant %f32 1.0\n"
17799                         "%c0f32     = OpConstant %f32 0.0\n";
17800
17801                 if (hasFloat32)
17802                 {
17803                         constStr        +=
17804                                 "%c10f32    = OpConstant %f32 10.0\n"
17805                                 "%c25f32    = OpConstant %f32 25.0\n"
17806                                 "%c50f32    = OpConstant %f32 50.0\n"
17807                                 "%c90f32    = OpConstant %f32 90.0\n";
17808                 }
17809
17810                 if (hasFloat16)
17811                 {
17812                         capsStr         += "OpCapability Float16\n";
17813                         constStr        +=
17814                                 "%f16       = OpTypeFloat 16\n"
17815                                 "%c10f16    = OpConstant %f16 10.0\n"
17816                                 "%c25f16    = OpConstant %f16 25.0\n"
17817                                 "%c50f16    = OpConstant %f16 50.0\n"
17818                                 "%c90f16    = OpConstant %f16 90.0\n";
17819                 }
17820
17821                 if (hasInt8)
17822                 {
17823                         capsStr         += "OpCapability Int8\n";
17824                         constStr        +=
17825                                 "%i8        = OpTypeInt 8 1\n"
17826                                 "%c10i8     = OpConstant %i8 10\n"
17827                                 "%c25i8     = OpConstant %i8 25\n"
17828                                 "%c50i8     = OpConstant %i8 50\n"
17829                                 "%c90i8     = OpConstant %i8 90\n";
17830                 }
17831
17832                 // Each invocation reads a different float32 value as input. Depending on
17833                 // the bit-sizes required by the particular test case, we also produce
17834                 // float16 and/or and int8 values by converting from the 32-bit float.
17835                 string testStr  = "";
17836                 testStr                 += "%inval32   = OpLoad %f32 %inloc\n";
17837                 if (hasFloat16)
17838                         testStr         += "%inval16   = OpFConvert %f16 %inval32\n";
17839                 if (hasInt8)
17840                         testStr         += "%inval8    = OpConvertFToS %i8 %inval32\n";
17841
17842                 // Because conversions from Float to Int round towards 0 we want our "greater" comparisons to be >=,
17843                 // that way a float32/float16 comparison such as 50.6f >= 50.0f will preserve its result
17844                 // when converted to int8, since FtoS(50.6f) results in 50. For "less" comparisons, it is the
17845                 // other way around, so in this case we want < instead of <=.
17846                 if (cases[caseNdx][0] == 32)
17847                         testStr         += "%cmp1      = OpFOrdGreaterThanEqual %bool %inval32 %c25f32\n";
17848                 else if (cases[caseNdx][0] == 16)
17849                         testStr         += "%cmp1      = OpFOrdGreaterThanEqual %bool %inval16 %c25f16\n";
17850                 else
17851                         testStr         += "%cmp1      = OpSGreaterThanEqual %bool %inval8 %c25i8\n";
17852
17853                 if (cases[caseNdx][1] == 32)
17854                         testStr         += "%cmp2      = OpFOrdLessThan %bool %inval32 %c50f32\n";
17855                 else if (cases[caseNdx][1] == 16)
17856                         testStr         += "%cmp2      = OpFOrdLessThan %bool %inval16 %c50f16\n";
17857                 else
17858                         testStr         += "%cmp2      = OpSLessThan %bool %inval8 %c50i8\n";
17859
17860                 if (cases[caseNdx][2] == 32)
17861                         testStr         += "%cmp3      = OpFOrdLessThan %bool %inval32 %c10f32\n";
17862                 else if (cases[caseNdx][2] == 16)
17863                         testStr         += "%cmp3      = OpFOrdLessThan %bool %inval16 %c10f16\n";
17864                 else
17865                         testStr         += "%cmp3      = OpSLessThan %bool %inval8 %c10i8\n";
17866
17867                 if (cases[caseNdx][3] == 32)
17868                         testStr         += "%cmp4      = OpFOrdGreaterThanEqual %bool %inval32 %c90f32\n";
17869                 else if (cases[caseNdx][3] == 16)
17870                         testStr         += "%cmp4      = OpFOrdGreaterThanEqual %bool %inval16 %c90f16\n";
17871                 else
17872                         testStr         += "%cmp4      = OpSGreaterThanEqual %bool %inval8 %c90i8\n";
17873
17874                 testStr                 += "%and1      = OpLogicalAnd %bool %cmp1 %cmp2\n";
17875                 testStr                 += "%or1       = OpLogicalOr %bool %cmp3 %cmp4\n";
17876                 testStr                 += "%or2       = OpLogicalOr %bool %and1 %or1\n";
17877                 testStr                 += "%not1      = OpLogicalNot %bool %or2\n";
17878                 testStr                 += "%res       = OpSelect %f32 %not1 %c1f32 %c0f32\n";
17879
17880                 specializations["CAPS"]         = capsStr;
17881                 specializations["CONST"]        = constStr;
17882                 specializations["TEST"]         = testStr;
17883
17884                 // Compute expected result by evaluating the boolean expression computed in the shader for each input value
17885                 for (size_t ndx = 0; ndx < numElements; ++ndx)
17886                         outputData[ndx] = !((inputData[ndx] >= 25.0f && inputData[ndx] < 50.0f) || (inputData[ndx] < 10.0f || inputData[ndx] >= 90.0f));
17887
17888                 spec.assembly = shaderTemplate.specialize(specializations);
17889                 spec.inputs.push_back(BufferSp(new Float32Buffer(inputData)));
17890                 spec.outputs.push_back(BufferSp(new Float32Buffer(outputData)));
17891                 spec.numWorkGroups = IVec3(numElements, 1, 1);
17892                 if (hasFloat16)
17893                         spec.requestedVulkanFeatures.extFloat16Int8 |= EXTFLOAT16INT8FEATURES_FLOAT16;
17894                 if (hasInt8)
17895                         spec.requestedVulkanFeatures.extFloat16Int8 |= EXTFLOAT16INT8FEATURES_INT8;
17896                 spec.extensions.push_back("VK_KHR_shader_float16_int8");
17897
17898                 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]);
17899                 group->addChild(new SpvAsmComputeShaderCase(testCtx, testName.c_str(), "", spec));
17900         }
17901
17902         return group.release();
17903 }
17904
17905 tcu::TestCaseGroup* createBoolGroup (tcu::TestContext& testCtx)
17906 {
17907         de::MovePtr<tcu::TestCaseGroup>         testGroup                       (new tcu::TestCaseGroup(testCtx, "bool", "Boolean tests"));
17908
17909         testGroup->addChild(createBoolMixedBitSizeGroup(testCtx));
17910
17911         return testGroup.release();
17912 }
17913
17914 tcu::TestCaseGroup* createOpNameAbuseTests (tcu::TestContext& testCtx)
17915 {
17916         de::MovePtr<tcu::TestCaseGroup> abuseGroup(new tcu::TestCaseGroup(testCtx, "opname_abuse", "OpName abuse tests"));
17917         vector<CaseParameter>                   abuseCases;
17918         RGBA                                                    defaultColors[4];
17919         map<string, string>                             opNameFragments;
17920
17921         getOpNameAbuseCases(abuseCases);
17922         getDefaultColors(defaultColors);
17923
17924         opNameFragments["testfun"] =
17925                 "%test_code  = OpFunction %v4f32 None %v4f32_v4f32_function\n"
17926                 "%param1     = OpFunctionParameter %v4f32\n"
17927                 "%label_func = OpLabel\n"
17928                 "%a          = OpVectorExtractDynamic %f32 %param1 %c_i32_0\n"
17929                 "%b          = OpFAdd %f32 %a %a\n"
17930                 "%c          = OpFSub %f32 %b %a\n"
17931                 "%ret        = OpVectorInsertDynamic %v4f32 %param1 %c %c_i32_0\n"
17932                 "OpReturnValue %ret\n"
17933                 "OpFunctionEnd\n";
17934
17935         for (unsigned int i = 0; i < abuseCases.size(); i++)
17936         {
17937                 string casename;
17938                 casename = string("main") + abuseCases[i].name;
17939
17940                 opNameFragments["debug"] =
17941                         "OpName %BP_main \"" + abuseCases[i].param + "\"";
17942
17943                 createTestsForAllStages(casename, defaultColors, defaultColors, opNameFragments, abuseGroup.get());
17944         }
17945
17946         for (unsigned int i = 0; i < abuseCases.size(); i++)
17947         {
17948                 string casename;
17949                 casename = string("b") + abuseCases[i].name;
17950
17951                 opNameFragments["debug"] =
17952                         "OpName %b \"" + abuseCases[i].param + "\"";
17953
17954                 createTestsForAllStages(casename, defaultColors, defaultColors, opNameFragments, abuseGroup.get());
17955         }
17956
17957         {
17958                 opNameFragments["debug"] =
17959                         "OpName %test_code \"name1\"\n"
17960                         "OpName %param1    \"name2\"\n"
17961                         "OpName %a         \"name3\"\n"
17962                         "OpName %b         \"name4\"\n"
17963                         "OpName %c         \"name5\"\n"
17964                         "OpName %ret       \"name6\"\n";
17965
17966                 createTestsForAllStages("everything_named", defaultColors, defaultColors, opNameFragments, abuseGroup.get());
17967         }
17968
17969         {
17970                 opNameFragments["debug"] =
17971                         "OpName %test_code \"the_same\"\n"
17972                         "OpName %param1    \"the_same\"\n"
17973                         "OpName %a         \"the_same\"\n"
17974                         "OpName %b         \"the_same\"\n"
17975                         "OpName %c         \"the_same\"\n"
17976                         "OpName %ret       \"the_same\"\n";
17977
17978                 createTestsForAllStages("everything_named_the_same", defaultColors, defaultColors, opNameFragments, abuseGroup.get());
17979         }
17980
17981         {
17982                 opNameFragments["debug"] =
17983                         "OpName %BP_main \"to_be\"\n"
17984                         "OpName %BP_main \"or_not\"\n"
17985                         "OpName %BP_main \"to_be\"\n";
17986
17987                 createTestsForAllStages("main_has_multiple_names", defaultColors, defaultColors, opNameFragments, abuseGroup.get());
17988         }
17989
17990         {
17991                 opNameFragments["debug"] =
17992                         "OpName %b \"to_be\"\n"
17993                         "OpName %b \"or_not\"\n"
17994                         "OpName %b \"to_be\"\n";
17995
17996                 createTestsForAllStages("b_has_multiple_names", defaultColors, defaultColors, opNameFragments, abuseGroup.get());
17997         }
17998
17999         return abuseGroup.release();
18000 }
18001
18002
18003 tcu::TestCaseGroup* createOpMemberNameAbuseTests (tcu::TestContext& testCtx)
18004 {
18005         de::MovePtr<tcu::TestCaseGroup> abuseGroup(new tcu::TestCaseGroup(testCtx, "opmembername_abuse", "OpName abuse tests"));
18006         vector<CaseParameter>                   abuseCases;
18007         RGBA                                                    defaultColors[4];
18008         map<string, string>                             opMemberNameFragments;
18009
18010         getOpNameAbuseCases(abuseCases);
18011         getDefaultColors(defaultColors);
18012
18013         opMemberNameFragments["pre_main"] =
18014                 "%f3str = OpTypeStruct %f32 %f32 %f32\n";
18015
18016         opMemberNameFragments["testfun"] =
18017                 "%test_code  = OpFunction %v4f32 None %v4f32_v4f32_function\n"
18018                 "%param1     = OpFunctionParameter %v4f32\n"
18019                 "%label_func = OpLabel\n"
18020                 "%a          = OpVectorExtractDynamic %f32 %param1 %c_i32_0\n"
18021                 "%b          = OpFAdd %f32 %a %a\n"
18022                 "%c          = OpFSub %f32 %b %a\n"
18023                 "%cstr       = OpCompositeConstruct %f3str %c %c %c\n"
18024                 "%d          = OpCompositeExtract %f32 %cstr 0\n"
18025                 "%ret        = OpVectorInsertDynamic %v4f32 %param1 %d %c_i32_0\n"
18026                 "OpReturnValue %ret\n"
18027                 "OpFunctionEnd\n";
18028
18029         for (unsigned int i = 0; i < abuseCases.size(); i++)
18030         {
18031                 string casename;
18032                 casename = string("f3str_x") + abuseCases[i].name;
18033
18034                 opMemberNameFragments["debug"] =
18035                         "OpMemberName %f3str 0 \"" + abuseCases[i].param + "\"";
18036
18037                 createTestsForAllStages(casename, defaultColors, defaultColors, opMemberNameFragments, abuseGroup.get());
18038         }
18039
18040         {
18041                 opMemberNameFragments["debug"] =
18042                         "OpMemberName %f3str 0 \"name1\"\n"
18043                         "OpMemberName %f3str 1 \"name2\"\n"
18044                         "OpMemberName %f3str 2 \"name3\"\n";
18045
18046                 createTestsForAllStages("everything_named", defaultColors, defaultColors, opMemberNameFragments, abuseGroup.get());
18047         }
18048
18049         {
18050                 opMemberNameFragments["debug"] =
18051                         "OpMemberName %f3str 0 \"the_same\"\n"
18052                         "OpMemberName %f3str 1 \"the_same\"\n"
18053                         "OpMemberName %f3str 2 \"the_same\"\n";
18054
18055                 createTestsForAllStages("everything_named_the_same", defaultColors, defaultColors, opMemberNameFragments, abuseGroup.get());
18056         }
18057
18058         {
18059                 opMemberNameFragments["debug"] =
18060                         "OpMemberName %f3str 0 \"to_be\"\n"
18061                         "OpMemberName %f3str 1 \"or_not\"\n"
18062                         "OpMemberName %f3str 0 \"to_be\"\n"
18063                         "OpMemberName %f3str 2 \"makes_no\"\n"
18064                         "OpMemberName %f3str 0 \"difference\"\n"
18065                         "OpMemberName %f3str 0 \"to_me\"\n";
18066
18067
18068                 createTestsForAllStages("f3str_x_has_multiple_names", defaultColors, defaultColors, opMemberNameFragments, abuseGroup.get());
18069         }
18070
18071         return abuseGroup.release();
18072 }
18073
18074 tcu::TestCaseGroup* createInstructionTests (tcu::TestContext& testCtx)
18075 {
18076         const bool testComputePipeline = true;
18077
18078         de::MovePtr<tcu::TestCaseGroup> instructionTests        (new tcu::TestCaseGroup(testCtx, "instruction", "Instructions with special opcodes/operands"));
18079         de::MovePtr<tcu::TestCaseGroup> computeTests            (new tcu::TestCaseGroup(testCtx, "compute", "Compute Instructions with special opcodes/operands"));
18080         de::MovePtr<tcu::TestCaseGroup> graphicsTests           (new tcu::TestCaseGroup(testCtx, "graphics", "Graphics Instructions with special opcodes/operands"));
18081
18082         computeTests->addChild(createSpivVersionCheckTests(testCtx, testComputePipeline));
18083         computeTests->addChild(createLocalSizeGroup(testCtx));
18084         computeTests->addChild(createOpNopGroup(testCtx));
18085         computeTests->addChild(createOpFUnordGroup(testCtx, TEST_WITHOUT_NAN));
18086         computeTests->addChild(createOpAtomicGroup(testCtx, false));
18087         computeTests->addChild(createOpAtomicGroup(testCtx, true));                                     // Using new StorageBuffer decoration
18088         computeTests->addChild(createOpAtomicGroup(testCtx, false, 1024, true));        // Return value validation
18089         computeTests->addChild(createOpLineGroup(testCtx));
18090         computeTests->addChild(createOpModuleProcessedGroup(testCtx));
18091         computeTests->addChild(createOpNoLineGroup(testCtx));
18092         computeTests->addChild(createOpConstantNullGroup(testCtx));
18093         computeTests->addChild(createOpConstantCompositeGroup(testCtx));
18094         computeTests->addChild(createOpConstantUsageGroup(testCtx));
18095         computeTests->addChild(createSpecConstantGroup(testCtx));
18096         computeTests->addChild(createOpSourceGroup(testCtx));
18097         computeTests->addChild(createOpSourceExtensionGroup(testCtx));
18098         computeTests->addChild(createDecorationGroupGroup(testCtx));
18099         computeTests->addChild(createOpPhiGroup(testCtx));
18100         computeTests->addChild(createLoopControlGroup(testCtx));
18101         computeTests->addChild(createFunctionControlGroup(testCtx));
18102         computeTests->addChild(createSelectionControlGroup(testCtx));
18103         computeTests->addChild(createBlockOrderGroup(testCtx));
18104         computeTests->addChild(createMultipleShaderGroup(testCtx));
18105         computeTests->addChild(createMemoryAccessGroup(testCtx));
18106         computeTests->addChild(createOpCopyMemoryGroup(testCtx));
18107         computeTests->addChild(createOpCopyObjectGroup(testCtx));
18108         computeTests->addChild(createNoContractionGroup(testCtx));
18109         computeTests->addChild(createOpUndefGroup(testCtx));
18110         computeTests->addChild(createOpUnreachableGroup(testCtx));
18111         computeTests->addChild(createOpQuantizeToF16Group(testCtx));
18112         computeTests->addChild(createOpFRemGroup(testCtx));
18113         computeTests->addChild(createOpSRemComputeGroup(testCtx, QP_TEST_RESULT_PASS));
18114         computeTests->addChild(createOpSRemComputeGroup64(testCtx, QP_TEST_RESULT_PASS));
18115         computeTests->addChild(createOpSModComputeGroup(testCtx, QP_TEST_RESULT_PASS));
18116         computeTests->addChild(createOpSModComputeGroup64(testCtx, QP_TEST_RESULT_PASS));
18117         computeTests->addChild(createConvertComputeTests(testCtx, "OpSConvert", "sconvert"));
18118         computeTests->addChild(createConvertComputeTests(testCtx, "OpUConvert", "uconvert"));
18119         computeTests->addChild(createConvertComputeTests(testCtx, "OpFConvert", "fconvert"));
18120         computeTests->addChild(createConvertComputeTests(testCtx, "OpConvertSToF", "convertstof"));
18121         computeTests->addChild(createConvertComputeTests(testCtx, "OpConvertFToS", "convertftos"));
18122         computeTests->addChild(createConvertComputeTests(testCtx, "OpConvertUToF", "convertutof"));
18123         computeTests->addChild(createConvertComputeTests(testCtx, "OpConvertFToU", "convertftou"));
18124         computeTests->addChild(createOpCompositeInsertGroup(testCtx));
18125         computeTests->addChild(createOpInBoundsAccessChainGroup(testCtx));
18126         computeTests->addChild(createShaderDefaultOutputGroup(testCtx));
18127         computeTests->addChild(createOpNMinGroup(testCtx));
18128         computeTests->addChild(createOpNMaxGroup(testCtx));
18129         computeTests->addChild(createOpNClampGroup(testCtx));
18130         {
18131                 de::MovePtr<tcu::TestCaseGroup> computeAndroidTests     (new tcu::TestCaseGroup(testCtx, "android", "Android CTS Tests"));
18132
18133                 computeAndroidTests->addChild(createOpSRemComputeGroup(testCtx, QP_TEST_RESULT_QUALITY_WARNING));
18134                 computeAndroidTests->addChild(createOpSModComputeGroup(testCtx, QP_TEST_RESULT_QUALITY_WARNING));
18135
18136                 computeTests->addChild(computeAndroidTests.release());
18137         }
18138
18139         computeTests->addChild(create8BitStorageComputeGroup(testCtx));
18140         computeTests->addChild(create16BitStorageComputeGroup(testCtx));
18141         computeTests->addChild(createFloatControlsComputeGroup(testCtx));
18142         computeTests->addChild(createUboMatrixPaddingComputeGroup(testCtx));
18143         computeTests->addChild(createCompositeInsertComputeGroup(testCtx));
18144         computeTests->addChild(createVariableInitComputeGroup(testCtx));
18145         computeTests->addChild(createConditionalBranchComputeGroup(testCtx));
18146         computeTests->addChild(createIndexingComputeGroup(testCtx));
18147         computeTests->addChild(createVariablePointersComputeGroup(testCtx));
18148         computeTests->addChild(createImageSamplerComputeGroup(testCtx));
18149         computeTests->addChild(createOpNameGroup(testCtx));
18150         computeTests->addChild(createOpMemberNameGroup(testCtx));
18151         computeTests->addChild(createPointerParameterComputeGroup(testCtx));
18152         computeTests->addChild(createFloat16Group(testCtx));
18153         computeTests->addChild(createBoolGroup(testCtx));
18154         computeTests->addChild(createWorkgroupMemoryComputeGroup(testCtx));
18155
18156         graphicsTests->addChild(createCrossStageInterfaceTests(testCtx));
18157         graphicsTests->addChild(createSpivVersionCheckTests(testCtx, !testComputePipeline));
18158         graphicsTests->addChild(createOpNopTests(testCtx));
18159         graphicsTests->addChild(createOpSourceTests(testCtx));
18160         graphicsTests->addChild(createOpSourceContinuedTests(testCtx));
18161         graphicsTests->addChild(createOpModuleProcessedTests(testCtx));
18162         graphicsTests->addChild(createOpLineTests(testCtx));
18163         graphicsTests->addChild(createOpNoLineTests(testCtx));
18164         graphicsTests->addChild(createOpConstantNullTests(testCtx));
18165         graphicsTests->addChild(createOpConstantCompositeTests(testCtx));
18166         graphicsTests->addChild(createMemoryAccessTests(testCtx));
18167         graphicsTests->addChild(createOpUndefTests(testCtx));
18168         graphicsTests->addChild(createSelectionBlockOrderTests(testCtx));
18169         graphicsTests->addChild(createModuleTests(testCtx));
18170         graphicsTests->addChild(createSwitchBlockOrderTests(testCtx));
18171         graphicsTests->addChild(createOpPhiTests(testCtx));
18172         graphicsTests->addChild(createNoContractionTests(testCtx));
18173         graphicsTests->addChild(createOpQuantizeTests(testCtx));
18174         graphicsTests->addChild(createLoopTests(testCtx));
18175         graphicsTests->addChild(createSpecConstantTests(testCtx));
18176         graphicsTests->addChild(createSpecConstantOpQuantizeToF16Group(testCtx));
18177         graphicsTests->addChild(createBarrierTests(testCtx));
18178         graphicsTests->addChild(createDecorationGroupTests(testCtx));
18179         graphicsTests->addChild(createFRemTests(testCtx));
18180         graphicsTests->addChild(createOpSRemGraphicsTests(testCtx, QP_TEST_RESULT_PASS));
18181         graphicsTests->addChild(createOpSModGraphicsTests(testCtx, QP_TEST_RESULT_PASS));
18182
18183         {
18184                 de::MovePtr<tcu::TestCaseGroup> graphicsAndroidTests    (new tcu::TestCaseGroup(testCtx, "android", "Android CTS Tests"));
18185
18186                 graphicsAndroidTests->addChild(createOpSRemGraphicsTests(testCtx, QP_TEST_RESULT_QUALITY_WARNING));
18187                 graphicsAndroidTests->addChild(createOpSModGraphicsTests(testCtx, QP_TEST_RESULT_QUALITY_WARNING));
18188
18189                 graphicsTests->addChild(graphicsAndroidTests.release());
18190         }
18191         graphicsTests->addChild(createOpNameTests(testCtx));
18192         graphicsTests->addChild(createOpNameAbuseTests(testCtx));
18193         graphicsTests->addChild(createOpMemberNameAbuseTests(testCtx));
18194
18195         graphicsTests->addChild(create8BitStorageGraphicsGroup(testCtx));
18196         graphicsTests->addChild(create16BitStorageGraphicsGroup(testCtx));
18197         graphicsTests->addChild(createFloatControlsGraphicsGroup(testCtx));
18198         graphicsTests->addChild(createUboMatrixPaddingGraphicsGroup(testCtx));
18199         graphicsTests->addChild(createCompositeInsertGraphicsGroup(testCtx));
18200         graphicsTests->addChild(createVariableInitGraphicsGroup(testCtx));
18201         graphicsTests->addChild(createConditionalBranchGraphicsGroup(testCtx));
18202         graphicsTests->addChild(createIndexingGraphicsGroup(testCtx));
18203         graphicsTests->addChild(createVariablePointersGraphicsGroup(testCtx));
18204         graphicsTests->addChild(createImageSamplerGraphicsGroup(testCtx));
18205         graphicsTests->addChild(createConvertGraphicsTests(testCtx, "OpSConvert", "sconvert"));
18206         graphicsTests->addChild(createConvertGraphicsTests(testCtx, "OpUConvert", "uconvert"));
18207         graphicsTests->addChild(createConvertGraphicsTests(testCtx, "OpFConvert", "fconvert"));
18208         graphicsTests->addChild(createConvertGraphicsTests(testCtx, "OpConvertSToF", "convertstof"));
18209         graphicsTests->addChild(createConvertGraphicsTests(testCtx, "OpConvertFToS", "convertftos"));
18210         graphicsTests->addChild(createConvertGraphicsTests(testCtx, "OpConvertUToF", "convertutof"));
18211         graphicsTests->addChild(createConvertGraphicsTests(testCtx, "OpConvertFToU", "convertftou"));
18212         graphicsTests->addChild(createPointerParameterGraphicsGroup(testCtx));
18213         graphicsTests->addChild(createVaryingNameGraphicsGroup(testCtx));
18214
18215         graphicsTests->addChild(createFloat16Tests(testCtx));
18216
18217         instructionTests->addChild(computeTests.release());
18218         instructionTests->addChild(graphicsTests.release());
18219
18220         return instructionTests.release();
18221 }
18222
18223 } // SpirVAssembly
18224 } // vkt