1 /*-------------------------------------------------------------------------
2 * Vulkan Conformance Tests
3 * ------------------------
5 * Copyright (c) 2015 Google Inc.
6 * Copyright (c) 2016 The Khronos Group Inc.
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
12 * http://www.apache.org/licenses/LICENSE-2.0
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.
22 * \brief SPIR-V Assembly Tests for Instructions (special opcode/operand)
23 *//*--------------------------------------------------------------------*/
25 #include "vktSpvAsmInstructionTests.hpp"
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"
38 #include "vkDeviceUtil.hpp"
39 #include "vkMemUtil.hpp"
40 #include "vkPlatform.hpp"
41 #include "vkPrograms.hpp"
42 #include "vkQueryUtil.hpp"
44 #include "vkRefUtil.hpp"
45 #include "vkStrUtil.hpp"
46 #include "vkTypeUtil.hpp"
48 #include "deStringUtil.hpp"
49 #include "deUniquePtr.hpp"
51 #include "tcuStringTemplate.hpp"
53 #include "vktSpvAsmCrossStageInterfaceTests.hpp"
54 #include "vktSpvAsm8bitStorageTests.hpp"
55 #include "vktSpvAsm16bitStorageTests.hpp"
56 #include "vktSpvAsmUboMatrixPaddingTests.hpp"
57 #include "vktSpvAsmConditionalBranchTests.hpp"
58 #include "vktSpvAsmIndexingTests.hpp"
59 #include "vktSpvAsmImageSamplerTests.hpp"
60 #include "vktSpvAsmComputeShaderCase.hpp"
61 #include "vktSpvAsmComputeShaderTestUtil.hpp"
62 #include "vktSpvAsmFloatControlsTests.hpp"
63 #include "vktSpvAsmGraphicsShaderTestUtil.hpp"
64 #include "vktSpvAsmVariablePointersTests.hpp"
65 #include "vktSpvAsmVariableInitTests.hpp"
66 #include "vktSpvAsmPointerParameterTests.hpp"
67 #include "vktSpvAsmSpirvVersionTests.hpp"
68 #include "vktTestCaseUtil.hpp"
69 #include "vktSpvAsmLoopDepLenTests.hpp"
70 #include "vktSpvAsmLoopDepInfTests.hpp"
71 #include "vktSpvAsmCompositeInsertTests.hpp"
72 #include "vktSpvAsmVaryingNameTests.hpp"
73 #include "vktSpvAsmWorkgroupMemoryTests.hpp"
74 #include "vktSpvAsmSignedIntCompareTests.hpp"
86 namespace SpirVAssembly
100 using tcu::TestStatus;
103 using tcu::StringTemplate;
106 const bool TEST_WITH_NAN = true;
107 const bool TEST_WITHOUT_NAN = false;
110 static void fillRandomScalars (de::Random& rnd, T minValue, T maxValue, void* dst, int numValues, int offset = 0)
112 T* const typedPtr = (T*)dst;
113 for (int ndx = 0; ndx < numValues; ndx++)
114 typedPtr[offset + ndx] = randomScalar<T>(rnd, minValue, maxValue);
117 // Filter is a function that returns true if a value should pass, false otherwise.
118 template<typename T, typename FilterT>
119 static void fillRandomScalars (de::Random& rnd, T minValue, T maxValue, void* dst, int numValues, FilterT filter, int offset = 0)
121 T* const typedPtr = (T*)dst;
123 for (int ndx = 0; ndx < numValues; ndx++)
126 value = randomScalar<T>(rnd, minValue, maxValue);
127 while (!filter(value));
129 typedPtr[offset + ndx] = value;
133 // Gets a 64-bit integer with a more logarithmic distribution
134 deInt64 randomInt64LogDistributed (de::Random& rnd)
136 deInt64 val = rnd.getUint64();
137 val &= (1ull << rnd.getInt(1, 63)) - 1;
143 static void fillRandomInt64sLogDistributed (de::Random& rnd, vector<deInt64>& dst, int numValues)
145 for (int ndx = 0; ndx < numValues; ndx++)
146 dst[ndx] = randomInt64LogDistributed(rnd);
149 template<typename FilterT>
150 static void fillRandomInt64sLogDistributed (de::Random& rnd, vector<deInt64>& dst, int numValues, FilterT filter)
152 for (int ndx = 0; ndx < numValues; ndx++)
156 value = randomInt64LogDistributed(rnd);
157 } while (!filter(value));
162 inline bool filterNonNegative (const deInt64 value)
167 inline bool filterPositive (const deInt64 value)
172 inline bool filterNotZero (const deInt64 value)
177 static void floorAll (vector<float>& values)
179 for (size_t i = 0; i < values.size(); i++)
180 values[i] = deFloatFloor(values[i]);
183 static void floorAll (vector<Vec4>& values)
185 for (size_t i = 0; i < values.size(); i++)
186 values[i] = floor(values[i]);
194 CaseParameter (const char* case_, const string& param_) : name(case_), param(param_) {}
197 // Assembly code used for testing LocalSize, OpNop, OpConstant{Null|Composite}, Op[No]Line, OpSource[Continued], OpSourceExtension, OpUndef is based on GLSL source code:
201 // layout(std140, set = 0, binding = 0) readonly buffer Input {
204 // layout(std140, set = 0, binding = 1) writeonly buffer Output {
208 // layout (local_size_x = 1, local_size_y = 1, local_size_z = 1) in;
211 // uint x = gl_GlobalInvocationID.x;
212 // output_data.elements[x] = -input_data.elements[x];
215 static string getAsmForLocalSizeTest(bool useLiteralLocalSize, bool useSpecConstantWorkgroupSize, IVec3 workGroupSize, deUint32 ndx)
217 std::ostringstream out;
218 out << getComputeAsmShaderPreambleWithoutLocalSize();
220 if (useLiteralLocalSize)
222 out << "OpExecutionMode %main LocalSize "
223 << workGroupSize.x() << " " << workGroupSize.y() << " " << workGroupSize.z() << "\n";
226 out << "OpSource GLSL 430\n"
227 "OpName %main \"main\"\n"
228 "OpName %id \"gl_GlobalInvocationID\"\n"
229 "OpDecorate %id BuiltIn GlobalInvocationId\n";
231 if (useSpecConstantWorkgroupSize)
233 out << "OpDecorate %spec_0 SpecId 100\n"
234 << "OpDecorate %spec_1 SpecId 101\n"
235 << "OpDecorate %spec_2 SpecId 102\n"
236 << "OpDecorate %gl_WorkGroupSize BuiltIn WorkgroupSize\n";
239 out << getComputeAsmInputOutputBufferTraits()
240 << getComputeAsmCommonTypes()
241 << getComputeAsmInputOutputBuffer()
242 << "%id = OpVariable %uvec3ptr Input\n"
243 << "%zero = OpConstant %i32 0 \n";
245 if (useSpecConstantWorkgroupSize)
247 out << "%spec_0 = OpSpecConstant %u32 "<< workGroupSize.x() << "\n"
248 << "%spec_1 = OpSpecConstant %u32 "<< workGroupSize.y() << "\n"
249 << "%spec_2 = OpSpecConstant %u32 "<< workGroupSize.z() << "\n"
250 << "%gl_WorkGroupSize = OpSpecConstantComposite %uvec3 %spec_0 %spec_1 %spec_2\n";
253 out << "%main = OpFunction %void None %voidf\n"
254 << "%label = OpLabel\n"
255 << "%idval = OpLoad %uvec3 %id\n"
256 << "%ndx = OpCompositeExtract %u32 %idval " << ndx << "\n"
258 "%inloc = OpAccessChain %f32ptr %indata %zero %ndx\n"
259 "%inval = OpLoad %f32 %inloc\n"
260 "%neg = OpFNegate %f32 %inval\n"
261 "%outloc = OpAccessChain %f32ptr %outdata %zero %ndx\n"
262 " OpStore %outloc %neg\n"
268 tcu::TestCaseGroup* createLocalSizeGroup (tcu::TestContext& testCtx)
270 de::MovePtr<tcu::TestCaseGroup> group (new tcu::TestCaseGroup(testCtx, "localsize", ""));
271 ComputeShaderSpec spec;
272 de::Random rnd (deStringHash(group->getName()));
273 const deUint32 numElements = 64u;
274 vector<float> positiveFloats (numElements, 0);
275 vector<float> negativeFloats (numElements, 0);
277 fillRandomScalars(rnd, 1.f, 100.f, &positiveFloats[0], numElements);
279 for (size_t ndx = 0; ndx < numElements; ++ndx)
280 negativeFloats[ndx] = -positiveFloats[ndx];
282 spec.inputs.push_back(BufferSp(new Float32Buffer(positiveFloats)));
283 spec.outputs.push_back(BufferSp(new Float32Buffer(negativeFloats)));
285 spec.numWorkGroups = IVec3(numElements, 1, 1);
287 spec.assembly = getAsmForLocalSizeTest(true, false, IVec3(1, 1, 1), 0u);
288 group->addChild(new SpvAsmComputeShaderCase(testCtx, "literal_localsize", "", spec));
290 spec.assembly = getAsmForLocalSizeTest(true, true, IVec3(1, 1, 1), 0u);
291 group->addChild(new SpvAsmComputeShaderCase(testCtx, "literal_and_specid_localsize", "", spec));
293 spec.assembly = getAsmForLocalSizeTest(false, true, IVec3(1, 1, 1), 0u);
294 group->addChild(new SpvAsmComputeShaderCase(testCtx, "specid_localsize", "", spec));
296 spec.numWorkGroups = IVec3(1, 1, 1);
298 spec.assembly = getAsmForLocalSizeTest(true, false, IVec3(numElements, 1, 1), 0u);
299 group->addChild(new SpvAsmComputeShaderCase(testCtx, "literal_localsize_x", "", spec));
301 spec.assembly = getAsmForLocalSizeTest(true, true, IVec3(numElements, 1, 1), 0u);
302 group->addChild(new SpvAsmComputeShaderCase(testCtx, "literal_and_specid_localsize_x", "", spec));
304 spec.assembly = getAsmForLocalSizeTest(false, true, IVec3(numElements, 1, 1), 0u);
305 group->addChild(new SpvAsmComputeShaderCase(testCtx, "specid_localsize_x", "", spec));
307 spec.assembly = getAsmForLocalSizeTest(true, false, IVec3(1, numElements, 1), 1u);
308 group->addChild(new SpvAsmComputeShaderCase(testCtx, "literal_localsize_y", "", spec));
310 spec.assembly = getAsmForLocalSizeTest(true, true, IVec3(1, numElements, 1), 1u);
311 group->addChild(new SpvAsmComputeShaderCase(testCtx, "literal_and_specid_localsize_y", "", spec));
313 spec.assembly = getAsmForLocalSizeTest(false, true, IVec3(1, numElements, 1), 1u);
314 group->addChild(new SpvAsmComputeShaderCase(testCtx, "specid_localsize_y", "", spec));
316 spec.assembly = getAsmForLocalSizeTest(true, false, IVec3(1, 1, numElements), 2u);
317 group->addChild(new SpvAsmComputeShaderCase(testCtx, "literal_localsize_z", "", spec));
319 spec.assembly = getAsmForLocalSizeTest(true, true, IVec3(1, 1, numElements), 2u);
320 group->addChild(new SpvAsmComputeShaderCase(testCtx, "literal_and_specid_localsize_z", "", spec));
322 spec.assembly = getAsmForLocalSizeTest(false, true, IVec3(1, 1, numElements), 2u);
323 group->addChild(new SpvAsmComputeShaderCase(testCtx, "specid_localsize_z", "", spec));
325 return group.release();
328 tcu::TestCaseGroup* createOpNopGroup (tcu::TestContext& testCtx)
330 de::MovePtr<tcu::TestCaseGroup> group (new tcu::TestCaseGroup(testCtx, "opnop", "Test the OpNop instruction"));
331 ComputeShaderSpec spec;
332 de::Random rnd (deStringHash(group->getName()));
333 const int numElements = 100;
334 vector<float> positiveFloats (numElements, 0);
335 vector<float> negativeFloats (numElements, 0);
337 fillRandomScalars(rnd, 1.f, 100.f, &positiveFloats[0], numElements);
339 for (size_t ndx = 0; ndx < numElements; ++ndx)
340 negativeFloats[ndx] = -positiveFloats[ndx];
343 string(getComputeAsmShaderPreamble()) +
345 "OpSource GLSL 430\n"
346 "OpName %main \"main\"\n"
347 "OpName %id \"gl_GlobalInvocationID\"\n"
349 "OpDecorate %id BuiltIn GlobalInvocationId\n"
351 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes())
353 + string(getComputeAsmInputOutputBuffer()) +
355 "%id = OpVariable %uvec3ptr Input\n"
356 "%zero = OpConstant %i32 0\n"
358 "%main = OpFunction %void None %voidf\n"
360 "%idval = OpLoad %uvec3 %id\n"
361 "%x = OpCompositeExtract %u32 %idval 0\n"
363 " OpNop\n" // Inside a function body
365 "%inloc = OpAccessChain %f32ptr %indata %zero %x\n"
366 "%inval = OpLoad %f32 %inloc\n"
367 "%neg = OpFNegate %f32 %inval\n"
368 "%outloc = OpAccessChain %f32ptr %outdata %zero %x\n"
369 " OpStore %outloc %neg\n"
372 spec.inputs.push_back(BufferSp(new Float32Buffer(positiveFloats)));
373 spec.outputs.push_back(BufferSp(new Float32Buffer(negativeFloats)));
374 spec.numWorkGroups = IVec3(numElements, 1, 1);
376 group->addChild(new SpvAsmComputeShaderCase(testCtx, "all", "OpNop appearing at different places", spec));
378 return group.release();
381 template<bool nanSupported>
382 bool compareFUnord (const std::vector<Resource>& inputs, const vector<AllocationSp>& outputAllocs, const std::vector<Resource>& expectedOutputs, TestLog& log)
384 if (outputAllocs.size() != 1)
387 vector<deUint8> input1Bytes;
388 vector<deUint8> input2Bytes;
389 vector<deUint8> expectedBytes;
391 inputs[0].getBytes(input1Bytes);
392 inputs[1].getBytes(input2Bytes);
393 expectedOutputs[0].getBytes(expectedBytes);
395 const deInt32* const expectedOutputAsInt = reinterpret_cast<const deInt32*>(&expectedBytes.front());
396 const deInt32* const outputAsInt = static_cast<const deInt32*>(outputAllocs[0]->getHostPtr());
397 const float* const input1AsFloat = reinterpret_cast<const float*>(&input1Bytes.front());
398 const float* const input2AsFloat = reinterpret_cast<const float*>(&input2Bytes.front());
399 bool returnValue = true;
401 for (size_t idx = 0; idx < expectedBytes.size() / sizeof(deInt32); ++idx)
403 if (!nanSupported && (tcu::Float32(input1AsFloat[idx]).isNaN() || tcu::Float32(input2AsFloat[idx]).isNaN()))
406 if (outputAsInt[idx] != expectedOutputAsInt[idx])
408 log << TestLog::Message << "ERROR: Sub-case failed. inputs: " << input1AsFloat[idx] << "," << input2AsFloat[idx] << " output: " << outputAsInt[idx]<< " expected output: " << expectedOutputAsInt[idx] << TestLog::EndMessage;
415 typedef VkBool32 (*compareFuncType) (float, float);
421 compareFuncType compareFunc;
423 OpFUnordCase (const char* _name, const char* _opCode, compareFuncType _compareFunc)
426 , compareFunc (_compareFunc) {}
429 #define ADD_OPFUNORD_CASE(NAME, OPCODE, OPERATOR) \
431 struct compare_##NAME { static VkBool32 compare(float x, float y) { return (x OPERATOR y) ? VK_TRUE : VK_FALSE; } }; \
432 cases.push_back(OpFUnordCase(#NAME, OPCODE, compare_##NAME::compare)); \
433 } while (deGetFalse())
435 tcu::TestCaseGroup* createOpFUnordGroup (tcu::TestContext& testCtx, const bool testWithNan)
437 const string nan = testWithNan ? "_nan" : "";
438 const string groupName = "opfunord" + nan;
439 de::MovePtr<tcu::TestCaseGroup> group (new tcu::TestCaseGroup(testCtx, groupName.c_str(), "Test the OpFUnord* opcodes"));
440 de::Random rnd (deStringHash(group->getName()));
441 const int numElements = 100;
442 vector<OpFUnordCase> cases;
443 string extensions = testWithNan ? "OpExtension \"SPV_KHR_float_controls\"\n" : "";
444 string capabilities = testWithNan ? "OpCapability SignedZeroInfNanPreserve\n" : "";
445 string exeModes = testWithNan ? "OpExecutionMode %main SignedZeroInfNanPreserve 32\n" : "";
446 const StringTemplate shaderTemplate (
447 string(getComputeAsmShaderPreamble(capabilities, extensions, exeModes)) +
448 "OpSource GLSL 430\n"
449 "OpName %main \"main\"\n"
450 "OpName %id \"gl_GlobalInvocationID\"\n"
452 "OpDecorate %id BuiltIn GlobalInvocationId\n"
454 "OpDecorate %buf BufferBlock\n"
455 "OpDecorate %buf2 BufferBlock\n"
456 "OpDecorate %indata1 DescriptorSet 0\n"
457 "OpDecorate %indata1 Binding 0\n"
458 "OpDecorate %indata2 DescriptorSet 0\n"
459 "OpDecorate %indata2 Binding 1\n"
460 "OpDecorate %outdata DescriptorSet 0\n"
461 "OpDecorate %outdata Binding 2\n"
462 "OpDecorate %f32arr ArrayStride 4\n"
463 "OpDecorate %i32arr ArrayStride 4\n"
464 "OpMemberDecorate %buf 0 Offset 0\n"
465 "OpMemberDecorate %buf2 0 Offset 0\n"
467 + string(getComputeAsmCommonTypes()) +
469 "%buf = OpTypeStruct %f32arr\n"
470 "%bufptr = OpTypePointer Uniform %buf\n"
471 "%indata1 = OpVariable %bufptr Uniform\n"
472 "%indata2 = OpVariable %bufptr Uniform\n"
474 "%buf2 = OpTypeStruct %i32arr\n"
475 "%buf2ptr = OpTypePointer Uniform %buf2\n"
476 "%outdata = OpVariable %buf2ptr Uniform\n"
478 "%id = OpVariable %uvec3ptr Input\n"
479 "%zero = OpConstant %i32 0\n"
480 "%consti1 = OpConstant %i32 1\n"
481 "%constf1 = OpConstant %f32 1.0\n"
483 "%main = OpFunction %void None %voidf\n"
485 "%idval = OpLoad %uvec3 %id\n"
486 "%x = OpCompositeExtract %u32 %idval 0\n"
488 "%inloc1 = OpAccessChain %f32ptr %indata1 %zero %x\n"
489 "%inval1 = OpLoad %f32 %inloc1\n"
490 "%inloc2 = OpAccessChain %f32ptr %indata2 %zero %x\n"
491 "%inval2 = OpLoad %f32 %inloc2\n"
492 "%outloc = OpAccessChain %i32ptr %outdata %zero %x\n"
494 "%result = ${OPCODE} %bool %inval1 %inval2\n"
495 "%int_res = OpSelect %i32 %result %consti1 %zero\n"
496 " OpStore %outloc %int_res\n"
501 ADD_OPFUNORD_CASE(equal, "OpFUnordEqual", ==);
502 ADD_OPFUNORD_CASE(less, "OpFUnordLessThan", <);
503 ADD_OPFUNORD_CASE(lessequal, "OpFUnordLessThanEqual", <=);
504 ADD_OPFUNORD_CASE(greater, "OpFUnordGreaterThan", >);
505 ADD_OPFUNORD_CASE(greaterequal, "OpFUnordGreaterThanEqual", >=);
506 ADD_OPFUNORD_CASE(notequal, "OpFUnordNotEqual", !=);
508 for (size_t caseNdx = 0; caseNdx < cases.size(); ++caseNdx)
510 map<string, string> specializations;
511 ComputeShaderSpec spec;
512 const float NaN = std::numeric_limits<float>::quiet_NaN();
513 vector<float> inputFloats1 (numElements, 0);
514 vector<float> inputFloats2 (numElements, 0);
515 vector<deInt32> expectedInts (numElements, 0);
517 specializations["OPCODE"] = cases[caseNdx].opCode;
518 spec.assembly = shaderTemplate.specialize(specializations);
520 fillRandomScalars(rnd, 1.f, 100.f, &inputFloats1[0], numElements);
521 for (size_t ndx = 0; ndx < numElements; ++ndx)
525 case 0: inputFloats2[ndx] = inputFloats1[ndx] + 1.0f; break;
526 case 1: inputFloats2[ndx] = inputFloats1[ndx] - 1.0f; break;
527 case 2: inputFloats2[ndx] = inputFloats1[ndx]; break;
528 case 3: inputFloats2[ndx] = NaN; break;
529 case 4: inputFloats2[ndx] = inputFloats1[ndx]; inputFloats1[ndx] = NaN; break;
530 case 5: inputFloats2[ndx] = NaN; inputFloats1[ndx] = NaN; break;
532 expectedInts[ndx] = tcu::Float32(inputFloats1[ndx]).isNaN() || tcu::Float32(inputFloats2[ndx]).isNaN() || cases[caseNdx].compareFunc(inputFloats1[ndx], inputFloats2[ndx]);
535 spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats1)));
536 spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats2)));
537 spec.outputs.push_back(BufferSp(new Int32Buffer(expectedInts)));
538 spec.numWorkGroups = IVec3(numElements, 1, 1);
539 spec.verifyIO = testWithNan ? &compareFUnord<true> : &compareFUnord<false>;
543 spec.extensions.push_back("VK_KHR_shader_float_controls");
544 spec.requestedVulkanFeatures.floatControlsProperties.shaderSignedZeroInfNanPreserveFloat32 = DE_TRUE;
547 group->addChild(new SpvAsmComputeShaderCase(testCtx, cases[caseNdx].name, cases[caseNdx].name, spec));
550 return group.release();
556 const char* assembly;
557 const char* retValAssembly;
558 OpAtomicType opAtomic;
559 deInt32 numOutputElements;
561 OpAtomicCase(const char* _name, const char* _assembly, const char* _retValAssembly, OpAtomicType _opAtomic, deInt32 _numOutputElements)
563 , assembly (_assembly)
564 , retValAssembly (_retValAssembly)
565 , opAtomic (_opAtomic)
566 , numOutputElements (_numOutputElements) {}
569 tcu::TestCaseGroup* createOpAtomicGroup (tcu::TestContext& testCtx, bool useStorageBuffer, int numElements = 65535, bool verifyReturnValues = false)
571 std::string groupName ("opatomic");
572 if (useStorageBuffer)
573 groupName += "_storage_buffer";
574 if (verifyReturnValues)
575 groupName += "_return_values";
576 de::MovePtr<tcu::TestCaseGroup> group (new tcu::TestCaseGroup(testCtx, groupName.c_str(), "Test the OpAtomic* opcodes"));
577 vector<OpAtomicCase> cases;
579 const StringTemplate shaderTemplate (
581 string("OpCapability Shader\n") +
582 (useStorageBuffer ? "OpExtension \"SPV_KHR_storage_buffer_storage_class\"\n" : "") +
583 "OpMemoryModel Logical GLSL450\n"
584 "OpEntryPoint GLCompute %main \"main\" %id\n"
585 "OpExecutionMode %main LocalSize 1 1 1\n" +
587 "OpSource GLSL 430\n"
588 "OpName %main \"main\"\n"
589 "OpName %id \"gl_GlobalInvocationID\"\n"
591 "OpDecorate %id BuiltIn GlobalInvocationId\n"
593 "OpDecorate %buf ${BLOCK_DECORATION}\n"
594 "OpDecorate %indata DescriptorSet 0\n"
595 "OpDecorate %indata Binding 0\n"
596 "OpDecorate %i32arr ArrayStride 4\n"
597 "OpMemberDecorate %buf 0 Offset 0\n"
599 "OpDecorate %sumbuf ${BLOCK_DECORATION}\n"
600 "OpDecorate %sum DescriptorSet 0\n"
601 "OpDecorate %sum Binding 1\n"
602 "OpMemberDecorate %sumbuf 0 Coherent\n"
603 "OpMemberDecorate %sumbuf 0 Offset 0\n"
605 "${RETVAL_BUF_DECORATE}"
607 + getComputeAsmCommonTypes("${BLOCK_POINTER_TYPE}") +
609 "%buf = OpTypeStruct %i32arr\n"
610 "%bufptr = OpTypePointer ${BLOCK_POINTER_TYPE} %buf\n"
611 "%indata = OpVariable %bufptr ${BLOCK_POINTER_TYPE}\n"
613 "%sumbuf = OpTypeStruct %i32arr\n"
614 "%sumbufptr = OpTypePointer ${BLOCK_POINTER_TYPE} %sumbuf\n"
615 "%sum = OpVariable %sumbufptr ${BLOCK_POINTER_TYPE}\n"
619 "%id = OpVariable %uvec3ptr Input\n"
620 "%minusone = OpConstant %i32 -1\n"
621 "%zero = OpConstant %i32 0\n"
622 "%one = OpConstant %u32 1\n"
623 "%two = OpConstant %i32 2\n"
625 "%main = OpFunction %void None %voidf\n"
627 "%idval = OpLoad %uvec3 %id\n"
628 "%x = OpCompositeExtract %u32 %idval 0\n"
630 "%inloc = OpAccessChain %i32ptr %indata %zero %x\n"
631 "%inval = OpLoad %i32 %inloc\n"
633 "%outloc = OpAccessChain %i32ptr %sum %zero ${INDEX}\n"
640 #define ADD_OPATOMIC_CASE(NAME, ASSEMBLY, RETVAL_ASSEMBLY, OPATOMIC, NUM_OUTPUT_ELEMENTS) \
642 DE_ASSERT((NUM_OUTPUT_ELEMENTS) == 1 || (NUM_OUTPUT_ELEMENTS) == numElements); \
643 cases.push_back(OpAtomicCase(#NAME, ASSEMBLY, RETVAL_ASSEMBLY, OPATOMIC, NUM_OUTPUT_ELEMENTS)); \
644 } while (deGetFalse())
645 #define ADD_OPATOMIC_CASE_1(NAME, ASSEMBLY, RETVAL_ASSEMBLY, OPATOMIC) ADD_OPATOMIC_CASE(NAME, ASSEMBLY, RETVAL_ASSEMBLY, OPATOMIC, 1)
646 #define ADD_OPATOMIC_CASE_N(NAME, ASSEMBLY, RETVAL_ASSEMBLY, OPATOMIC) ADD_OPATOMIC_CASE(NAME, ASSEMBLY, RETVAL_ASSEMBLY, OPATOMIC, numElements)
648 ADD_OPATOMIC_CASE_1(iadd, "%retv = OpAtomicIAdd %i32 %outloc %one %zero %inval\n",
649 " OpStore %retloc %retv\n", OPATOMIC_IADD );
650 ADD_OPATOMIC_CASE_1(isub, "%retv = OpAtomicISub %i32 %outloc %one %zero %inval\n",
651 " OpStore %retloc %retv\n", OPATOMIC_ISUB );
652 ADD_OPATOMIC_CASE_1(iinc, "%retv = OpAtomicIIncrement %i32 %outloc %one %zero\n",
653 " OpStore %retloc %retv\n", OPATOMIC_IINC );
654 ADD_OPATOMIC_CASE_1(idec, "%retv = OpAtomicIDecrement %i32 %outloc %one %zero\n",
655 " OpStore %retloc %retv\n", OPATOMIC_IDEC );
656 if (!verifyReturnValues)
658 ADD_OPATOMIC_CASE_N(load, "%inval2 = OpAtomicLoad %i32 %inloc %one %zero\n"
659 " OpStore %outloc %inval2\n", "", OPATOMIC_LOAD );
660 ADD_OPATOMIC_CASE_N(store, " OpAtomicStore %outloc %one %zero %inval\n", "", OPATOMIC_STORE );
663 ADD_OPATOMIC_CASE_N(compex, "%even = OpSMod %i32 %inval %two\n"
664 " OpStore %outloc %even\n"
665 "%retv = OpAtomicCompareExchange %i32 %outloc %one %zero %zero %minusone %zero\n",
666 " OpStore %retloc %retv\n", OPATOMIC_COMPEX );
669 #undef ADD_OPATOMIC_CASE
670 #undef ADD_OPATOMIC_CASE_1
671 #undef ADD_OPATOMIC_CASE_N
673 for (size_t caseNdx = 0; caseNdx < cases.size(); ++caseNdx)
675 map<string, string> specializations;
676 ComputeShaderSpec spec;
677 vector<deInt32> inputInts (numElements, 0);
678 vector<deInt32> expected (cases[caseNdx].numOutputElements, -1);
680 specializations["INDEX"] = (cases[caseNdx].numOutputElements == 1) ? "%zero" : "%x";
681 specializations["INSTRUCTION"] = cases[caseNdx].assembly;
682 specializations["BLOCK_DECORATION"] = useStorageBuffer ? "Block" : "BufferBlock";
683 specializations["BLOCK_POINTER_TYPE"] = useStorageBuffer ? "StorageBuffer" : "Uniform";
685 if (verifyReturnValues)
687 const StringTemplate blockDecoration (
689 "OpDecorate %retbuf ${BLOCK_DECORATION}\n"
690 "OpDecorate %ret DescriptorSet 0\n"
691 "OpDecorate %ret Binding 2\n"
692 "OpMemberDecorate %retbuf 0 Offset 0\n\n");
694 const StringTemplate blockDeclaration (
696 "%retbuf = OpTypeStruct %i32arr\n"
697 "%retbufptr = OpTypePointer ${BLOCK_POINTER_TYPE} %retbuf\n"
698 "%ret = OpVariable %retbufptr ${BLOCK_POINTER_TYPE}\n\n");
700 specializations["RETVAL_ASSEMBLY"] =
701 "%retloc = OpAccessChain %i32ptr %ret %zero %x\n"
702 + std::string(cases[caseNdx].retValAssembly);
704 specializations["RETVAL_BUF_DECORATE"] = blockDecoration.specialize(specializations);
705 specializations["RETVAL_BUF_DECL"] = blockDeclaration.specialize(specializations);
709 specializations["RETVAL_ASSEMBLY"] = "";
710 specializations["RETVAL_BUF_DECORATE"] = "";
711 specializations["RETVAL_BUF_DECL"] = "";
714 spec.assembly = shaderTemplate.specialize(specializations);
716 if (useStorageBuffer)
717 spec.extensions.push_back("VK_KHR_storage_buffer_storage_class");
719 spec.inputs.push_back(BufferSp(new OpAtomicBuffer(numElements, cases[caseNdx].numOutputElements, cases[caseNdx].opAtomic, BUFFERTYPE_INPUT)));
720 spec.outputs.push_back(BufferSp(new OpAtomicBuffer(numElements, cases[caseNdx].numOutputElements, cases[caseNdx].opAtomic, BUFFERTYPE_EXPECTED)));
721 if (verifyReturnValues)
722 spec.outputs.push_back(BufferSp(new OpAtomicBuffer(numElements, cases[caseNdx].numOutputElements, cases[caseNdx].opAtomic, BUFFERTYPE_ATOMIC_RET)));
723 spec.numWorkGroups = IVec3(numElements, 1, 1);
725 if (verifyReturnValues)
727 switch (cases[caseNdx].opAtomic)
730 spec.verifyIO = OpAtomicBuffer::compareWithRetvals<OPATOMIC_IADD>;
733 spec.verifyIO = OpAtomicBuffer::compareWithRetvals<OPATOMIC_ISUB>;
736 spec.verifyIO = OpAtomicBuffer::compareWithRetvals<OPATOMIC_IINC>;
739 spec.verifyIO = OpAtomicBuffer::compareWithRetvals<OPATOMIC_IDEC>;
741 case OPATOMIC_COMPEX:
742 spec.verifyIO = OpAtomicBuffer::compareWithRetvals<OPATOMIC_COMPEX>;
745 DE_FATAL("Unsupported OpAtomic type for return value verification");
748 group->addChild(new SpvAsmComputeShaderCase(testCtx, cases[caseNdx].name, cases[caseNdx].name, spec));
751 return group.release();
754 tcu::TestCaseGroup* createOpLineGroup (tcu::TestContext& testCtx)
756 de::MovePtr<tcu::TestCaseGroup> group (new tcu::TestCaseGroup(testCtx, "opline", "Test the OpLine instruction"));
757 ComputeShaderSpec spec;
758 de::Random rnd (deStringHash(group->getName()));
759 const int numElements = 100;
760 vector<float> positiveFloats (numElements, 0);
761 vector<float> negativeFloats (numElements, 0);
763 fillRandomScalars(rnd, 1.f, 100.f, &positiveFloats[0], numElements);
765 for (size_t ndx = 0; ndx < numElements; ++ndx)
766 negativeFloats[ndx] = -positiveFloats[ndx];
769 string(getComputeAsmShaderPreamble()) +
771 "%fname1 = OpString \"negateInputs.comp\"\n"
772 "%fname2 = OpString \"negateInputs\"\n"
774 "OpSource GLSL 430\n"
775 "OpName %main \"main\"\n"
776 "OpName %id \"gl_GlobalInvocationID\"\n"
778 "OpDecorate %id BuiltIn GlobalInvocationId\n"
780 + string(getComputeAsmInputOutputBufferTraits()) +
782 "OpLine %fname1 0 0\n" // At the earliest possible position
784 + string(getComputeAsmCommonTypes()) + string(getComputeAsmInputOutputBuffer()) +
786 "OpLine %fname1 0 1\n" // Multiple OpLines in sequence
787 "OpLine %fname2 1 0\n" // Different filenames
788 "OpLine %fname1 1000 100000\n"
790 "%id = OpVariable %uvec3ptr Input\n"
791 "%zero = OpConstant %i32 0\n"
793 "OpLine %fname1 1 1\n" // Before a function
795 "%main = OpFunction %void None %voidf\n"
798 "OpLine %fname1 1 1\n" // In a function
800 "%idval = OpLoad %uvec3 %id\n"
801 "%x = OpCompositeExtract %u32 %idval 0\n"
802 "%inloc = OpAccessChain %f32ptr %indata %zero %x\n"
803 "%inval = OpLoad %f32 %inloc\n"
804 "%neg = OpFNegate %f32 %inval\n"
805 "%outloc = OpAccessChain %f32ptr %outdata %zero %x\n"
806 " OpStore %outloc %neg\n"
809 spec.inputs.push_back(BufferSp(new Float32Buffer(positiveFloats)));
810 spec.outputs.push_back(BufferSp(new Float32Buffer(negativeFloats)));
811 spec.numWorkGroups = IVec3(numElements, 1, 1);
813 group->addChild(new SpvAsmComputeShaderCase(testCtx, "all", "OpLine appearing at different places", spec));
815 return group.release();
818 bool veryfiBinaryShader (const ProgramBinary& binary)
820 const size_t paternCount = 3u;
821 bool paternsCheck[paternCount] =
825 const string patersns[paternCount] =
831 size_t paternNdx = 0u;
833 for (size_t ndx = 0u; ndx < binary.getSize(); ++ndx)
835 if (false == paternsCheck[paternNdx] &&
836 patersns[paternNdx][0] == static_cast<char>(binary.getBinary()[ndx]) &&
837 deMemoryEqual((const char*)&binary.getBinary()[ndx], &patersns[paternNdx][0], patersns[paternNdx].length()))
839 paternsCheck[paternNdx]= true;
841 if (paternNdx == paternCount)
846 for (size_t ndx = 0u; ndx < paternCount; ++ndx)
848 if (!paternsCheck[ndx])
855 tcu::TestCaseGroup* createOpModuleProcessedGroup (tcu::TestContext& testCtx)
857 de::MovePtr<tcu::TestCaseGroup> group (new tcu::TestCaseGroup(testCtx, "opmoduleprocessed", "Test the OpModuleProcessed instruction"));
858 ComputeShaderSpec spec;
859 de::Random rnd (deStringHash(group->getName()));
860 const int numElements = 10;
861 vector<float> positiveFloats (numElements, 0);
862 vector<float> negativeFloats (numElements, 0);
864 fillRandomScalars(rnd, 1.f, 100.f, &positiveFloats[0], numElements);
866 for (size_t ndx = 0; ndx < numElements; ++ndx)
867 negativeFloats[ndx] = -positiveFloats[ndx];
870 string(getComputeAsmShaderPreamble()) +
871 "%fname = OpString \"negateInputs.comp\"\n"
873 "OpSource GLSL 430\n"
874 "OpName %main \"main\"\n"
875 "OpName %id \"gl_GlobalInvocationID\"\n"
876 "OpModuleProcessed \"VULKAN CTS\"\n" //OpModuleProcessed;
877 "OpModuleProcessed \"Negative values\"\n"
878 "OpModuleProcessed \"Date: 2017/09/21\"\n"
879 "OpDecorate %id BuiltIn GlobalInvocationId\n"
881 + string(getComputeAsmInputOutputBufferTraits())
883 + string(getComputeAsmCommonTypes()) + string(getComputeAsmInputOutputBuffer()) +
885 "OpLine %fname 0 1\n"
887 "OpLine %fname 1000 1\n"
889 "%id = OpVariable %uvec3ptr Input\n"
890 "%zero = OpConstant %i32 0\n"
891 "%main = OpFunction %void None %voidf\n"
894 "%idval = OpLoad %uvec3 %id\n"
895 "%x = OpCompositeExtract %u32 %idval 0\n"
897 "%inloc = OpAccessChain %f32ptr %indata %zero %x\n"
898 "%inval = OpLoad %f32 %inloc\n"
899 "%neg = OpFNegate %f32 %inval\n"
900 "%outloc = OpAccessChain %f32ptr %outdata %zero %x\n"
901 " OpStore %outloc %neg\n"
904 spec.inputs.push_back(BufferSp(new Float32Buffer(positiveFloats)));
905 spec.outputs.push_back(BufferSp(new Float32Buffer(negativeFloats)));
906 spec.numWorkGroups = IVec3(numElements, 1, 1);
907 spec.verifyBinary = veryfiBinaryShader;
908 spec.spirvVersion = SPIRV_VERSION_1_3;
910 group->addChild(new SpvAsmComputeShaderCase(testCtx, "all", "OpModuleProcessed Tests", spec));
912 return group.release();
915 tcu::TestCaseGroup* createOpNoLineGroup (tcu::TestContext& testCtx)
917 de::MovePtr<tcu::TestCaseGroup> group (new tcu::TestCaseGroup(testCtx, "opnoline", "Test the OpNoLine instruction"));
918 ComputeShaderSpec spec;
919 de::Random rnd (deStringHash(group->getName()));
920 const int numElements = 100;
921 vector<float> positiveFloats (numElements, 0);
922 vector<float> negativeFloats (numElements, 0);
924 fillRandomScalars(rnd, 1.f, 100.f, &positiveFloats[0], numElements);
926 for (size_t ndx = 0; ndx < numElements; ++ndx)
927 negativeFloats[ndx] = -positiveFloats[ndx];
930 string(getComputeAsmShaderPreamble()) +
932 "%fname = OpString \"negateInputs.comp\"\n"
934 "OpSource GLSL 430\n"
935 "OpName %main \"main\"\n"
936 "OpName %id \"gl_GlobalInvocationID\"\n"
938 "OpDecorate %id BuiltIn GlobalInvocationId\n"
940 + string(getComputeAsmInputOutputBufferTraits()) +
942 "OpNoLine\n" // At the earliest possible position, without preceding OpLine
944 + string(getComputeAsmCommonTypes()) + string(getComputeAsmInputOutputBuffer()) +
946 "OpLine %fname 0 1\n"
947 "OpNoLine\n" // Immediately following a preceding OpLine
949 "OpLine %fname 1000 1\n"
951 "%id = OpVariable %uvec3ptr Input\n"
952 "%zero = OpConstant %i32 0\n"
954 "OpNoLine\n" // Contents after the previous OpLine
956 "%main = OpFunction %void None %voidf\n"
958 "%idval = OpLoad %uvec3 %id\n"
959 "%x = OpCompositeExtract %u32 %idval 0\n"
961 "OpNoLine\n" // Multiple OpNoLine
965 "%inloc = OpAccessChain %f32ptr %indata %zero %x\n"
966 "%inval = OpLoad %f32 %inloc\n"
967 "%neg = OpFNegate %f32 %inval\n"
968 "%outloc = OpAccessChain %f32ptr %outdata %zero %x\n"
969 " OpStore %outloc %neg\n"
972 spec.inputs.push_back(BufferSp(new Float32Buffer(positiveFloats)));
973 spec.outputs.push_back(BufferSp(new Float32Buffer(negativeFloats)));
974 spec.numWorkGroups = IVec3(numElements, 1, 1);
976 group->addChild(new SpvAsmComputeShaderCase(testCtx, "all", "OpNoLine appearing at different places", spec));
978 return group.release();
981 // Compare instruction for the contraction compute case.
982 // Returns true if the output is what is expected from the test case.
983 bool compareNoContractCase(const std::vector<Resource>&, const vector<AllocationSp>& outputAllocs, const std::vector<Resource>& expectedOutputs, TestLog&)
985 if (outputAllocs.size() != 1)
988 // Only size is needed because we are not comparing the exact values.
989 size_t byteSize = expectedOutputs[0].getByteSize();
991 const float* outputAsFloat = static_cast<const float*>(outputAllocs[0]->getHostPtr());
993 for(size_t i = 0; i < byteSize / sizeof(float); ++i) {
994 if (outputAsFloat[i] != 0.f &&
995 outputAsFloat[i] != -ldexp(1, -24)) {
1003 tcu::TestCaseGroup* createNoContractionGroup (tcu::TestContext& testCtx)
1005 de::MovePtr<tcu::TestCaseGroup> group (new tcu::TestCaseGroup(testCtx, "nocontraction", "Test the NoContraction decoration"));
1006 vector<CaseParameter> cases;
1007 const int numElements = 100;
1008 vector<float> inputFloats1 (numElements, 0);
1009 vector<float> inputFloats2 (numElements, 0);
1010 vector<float> outputFloats (numElements, 0);
1011 const StringTemplate shaderTemplate (
1012 string(getComputeAsmShaderPreamble()) +
1014 "OpName %main \"main\"\n"
1015 "OpName %id \"gl_GlobalInvocationID\"\n"
1017 "OpDecorate %id BuiltIn GlobalInvocationId\n"
1021 "OpDecorate %buf BufferBlock\n"
1022 "OpDecorate %indata1 DescriptorSet 0\n"
1023 "OpDecorate %indata1 Binding 0\n"
1024 "OpDecorate %indata2 DescriptorSet 0\n"
1025 "OpDecorate %indata2 Binding 1\n"
1026 "OpDecorate %outdata DescriptorSet 0\n"
1027 "OpDecorate %outdata Binding 2\n"
1028 "OpDecorate %f32arr ArrayStride 4\n"
1029 "OpMemberDecorate %buf 0 Offset 0\n"
1031 + string(getComputeAsmCommonTypes()) +
1033 "%buf = OpTypeStruct %f32arr\n"
1034 "%bufptr = OpTypePointer Uniform %buf\n"
1035 "%indata1 = OpVariable %bufptr Uniform\n"
1036 "%indata2 = OpVariable %bufptr Uniform\n"
1037 "%outdata = OpVariable %bufptr Uniform\n"
1039 "%id = OpVariable %uvec3ptr Input\n"
1040 "%zero = OpConstant %i32 0\n"
1041 "%c_f_m1 = OpConstant %f32 -1.\n"
1043 "%main = OpFunction %void None %voidf\n"
1044 "%label = OpLabel\n"
1045 "%idval = OpLoad %uvec3 %id\n"
1046 "%x = OpCompositeExtract %u32 %idval 0\n"
1047 "%inloc1 = OpAccessChain %f32ptr %indata1 %zero %x\n"
1048 "%inval1 = OpLoad %f32 %inloc1\n"
1049 "%inloc2 = OpAccessChain %f32ptr %indata2 %zero %x\n"
1050 "%inval2 = OpLoad %f32 %inloc2\n"
1051 "%mul = OpFMul %f32 %inval1 %inval2\n"
1052 "%add = OpFAdd %f32 %mul %c_f_m1\n"
1053 "%outloc = OpAccessChain %f32ptr %outdata %zero %x\n"
1054 " OpStore %outloc %add\n"
1056 " OpFunctionEnd\n");
1058 cases.push_back(CaseParameter("multiplication", "OpDecorate %mul NoContraction"));
1059 cases.push_back(CaseParameter("addition", "OpDecorate %add NoContraction"));
1060 cases.push_back(CaseParameter("both", "OpDecorate %mul NoContraction\nOpDecorate %add NoContraction"));
1062 for (size_t ndx = 0; ndx < numElements; ++ndx)
1064 inputFloats1[ndx] = 1.f + std::ldexp(1.f, -23); // 1 + 2^-23.
1065 inputFloats2[ndx] = 1.f - std::ldexp(1.f, -23); // 1 - 2^-23.
1066 // Result for (1 + 2^-23) * (1 - 2^-23) - 1. With NoContraction, the multiplication will be
1067 // conducted separately and the result is rounded to 1, or 0x1.fffffcp-1
1068 // So the final result will be 0.f or 0x1p-24.
1069 // If the operation is combined into a precise fused multiply-add, then the result would be
1070 // 2^-46 (0xa8800000).
1071 outputFloats[ndx] = 0.f;
1074 for (size_t caseNdx = 0; caseNdx < cases.size(); ++caseNdx)
1076 map<string, string> specializations;
1077 ComputeShaderSpec spec;
1079 specializations["DECORATION"] = cases[caseNdx].param;
1080 spec.assembly = shaderTemplate.specialize(specializations);
1081 spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats1)));
1082 spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats2)));
1083 spec.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
1084 spec.numWorkGroups = IVec3(numElements, 1, 1);
1085 // Check against the two possible answers based on rounding mode.
1086 spec.verifyIO = &compareNoContractCase;
1088 group->addChild(new SpvAsmComputeShaderCase(testCtx, cases[caseNdx].name, cases[caseNdx].name, spec));
1090 return group.release();
1093 bool compareFRem(const std::vector<Resource>&, const vector<AllocationSp>& outputAllocs, const std::vector<Resource>& expectedOutputs, TestLog&)
1095 if (outputAllocs.size() != 1)
1098 vector<deUint8> expectedBytes;
1099 expectedOutputs[0].getBytes(expectedBytes);
1101 const float* expectedOutputAsFloat = reinterpret_cast<const float*>(&expectedBytes.front());
1102 const float* outputAsFloat = static_cast<const float*>(outputAllocs[0]->getHostPtr());
1104 for (size_t idx = 0; idx < expectedBytes.size() / sizeof(float); ++idx)
1106 const float f0 = expectedOutputAsFloat[idx];
1107 const float f1 = outputAsFloat[idx];
1108 // \todo relative error needs to be fairly high because FRem may be implemented as
1109 // (roughly) frac(a/b)*b, so LSB errors can be magnified. But this should be fine for now.
1110 if (deFloatAbs((f1 - f0) / f0) > 0.02)
1117 tcu::TestCaseGroup* createOpFRemGroup (tcu::TestContext& testCtx)
1119 de::MovePtr<tcu::TestCaseGroup> group (new tcu::TestCaseGroup(testCtx, "opfrem", "Test the OpFRem instruction"));
1120 ComputeShaderSpec spec;
1121 de::Random rnd (deStringHash(group->getName()));
1122 const int numElements = 200;
1123 vector<float> inputFloats1 (numElements, 0);
1124 vector<float> inputFloats2 (numElements, 0);
1125 vector<float> outputFloats (numElements, 0);
1127 fillRandomScalars(rnd, -10000.f, 10000.f, &inputFloats1[0], numElements);
1128 fillRandomScalars(rnd, -100.f, 100.f, &inputFloats2[0], numElements);
1130 for (size_t ndx = 0; ndx < numElements; ++ndx)
1132 // Guard against divisors near zero.
1133 if (std::fabs(inputFloats2[ndx]) < 1e-3)
1134 inputFloats2[ndx] = 8.f;
1136 // The return value of std::fmod() has the same sign as its first operand, which is how OpFRem spec'd.
1137 outputFloats[ndx] = std::fmod(inputFloats1[ndx], inputFloats2[ndx]);
1141 string(getComputeAsmShaderPreamble()) +
1143 "OpName %main \"main\"\n"
1144 "OpName %id \"gl_GlobalInvocationID\"\n"
1146 "OpDecorate %id BuiltIn GlobalInvocationId\n"
1148 "OpDecorate %buf BufferBlock\n"
1149 "OpDecorate %indata1 DescriptorSet 0\n"
1150 "OpDecorate %indata1 Binding 0\n"
1151 "OpDecorate %indata2 DescriptorSet 0\n"
1152 "OpDecorate %indata2 Binding 1\n"
1153 "OpDecorate %outdata DescriptorSet 0\n"
1154 "OpDecorate %outdata Binding 2\n"
1155 "OpDecorate %f32arr ArrayStride 4\n"
1156 "OpMemberDecorate %buf 0 Offset 0\n"
1158 + string(getComputeAsmCommonTypes()) +
1160 "%buf = OpTypeStruct %f32arr\n"
1161 "%bufptr = OpTypePointer Uniform %buf\n"
1162 "%indata1 = OpVariable %bufptr Uniform\n"
1163 "%indata2 = OpVariable %bufptr Uniform\n"
1164 "%outdata = OpVariable %bufptr Uniform\n"
1166 "%id = OpVariable %uvec3ptr Input\n"
1167 "%zero = OpConstant %i32 0\n"
1169 "%main = OpFunction %void None %voidf\n"
1170 "%label = OpLabel\n"
1171 "%idval = OpLoad %uvec3 %id\n"
1172 "%x = OpCompositeExtract %u32 %idval 0\n"
1173 "%inloc1 = OpAccessChain %f32ptr %indata1 %zero %x\n"
1174 "%inval1 = OpLoad %f32 %inloc1\n"
1175 "%inloc2 = OpAccessChain %f32ptr %indata2 %zero %x\n"
1176 "%inval2 = OpLoad %f32 %inloc2\n"
1177 "%rem = OpFRem %f32 %inval1 %inval2\n"
1178 "%outloc = OpAccessChain %f32ptr %outdata %zero %x\n"
1179 " OpStore %outloc %rem\n"
1183 spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats1)));
1184 spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats2)));
1185 spec.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
1186 spec.numWorkGroups = IVec3(numElements, 1, 1);
1187 spec.verifyIO = &compareFRem;
1189 group->addChild(new SpvAsmComputeShaderCase(testCtx, "all", "", spec));
1191 return group.release();
1194 bool compareNMin (const std::vector<Resource>&, const vector<AllocationSp>& outputAllocs, const std::vector<Resource>& expectedOutputs, TestLog&)
1196 if (outputAllocs.size() != 1)
1199 const BufferSp& expectedOutput (expectedOutputs[0].getBuffer());
1200 std::vector<deUint8> data;
1201 expectedOutput->getBytes(data);
1203 const float* const expectedOutputAsFloat = reinterpret_cast<const float*>(&data.front());
1204 const float* const outputAsFloat = static_cast<const float*>(outputAllocs[0]->getHostPtr());
1206 for (size_t idx = 0; idx < expectedOutput->getByteSize() / sizeof(float); ++idx)
1208 const float f0 = expectedOutputAsFloat[idx];
1209 const float f1 = outputAsFloat[idx];
1211 // For NMin, we accept NaN as output if both inputs were NaN.
1212 // Otherwise the NaN is the wrong choise, as on architectures that
1213 // do not handle NaN, those are huge values.
1214 if (!(tcu::Float32(f1).isNaN() && tcu::Float32(f0).isNaN()) && deFloatAbs(f1 - f0) > 0.00001f)
1221 tcu::TestCaseGroup* createOpNMinGroup (tcu::TestContext& testCtx)
1223 de::MovePtr<tcu::TestCaseGroup> group (new tcu::TestCaseGroup(testCtx, "opnmin", "Test the OpNMin instruction"));
1224 ComputeShaderSpec spec;
1225 de::Random rnd (deStringHash(group->getName()));
1226 const int numElements = 200;
1227 vector<float> inputFloats1 (numElements, 0);
1228 vector<float> inputFloats2 (numElements, 0);
1229 vector<float> outputFloats (numElements, 0);
1231 fillRandomScalars(rnd, -10000.f, 10000.f, &inputFloats1[0], numElements);
1232 fillRandomScalars(rnd, -10000.f, 10000.f, &inputFloats2[0], numElements);
1234 // Make the first case a full-NAN case.
1235 inputFloats1[0] = TCU_NAN;
1236 inputFloats2[0] = TCU_NAN;
1238 for (size_t ndx = 0; ndx < numElements; ++ndx)
1240 // By default, pick the smallest
1241 outputFloats[ndx] = std::min(inputFloats1[ndx], inputFloats2[ndx]);
1243 // Make half of the cases NaN cases
1246 // Alternate between the NaN operand
1249 outputFloats[ndx] = inputFloats2[ndx];
1250 inputFloats1[ndx] = TCU_NAN;
1254 outputFloats[ndx] = inputFloats1[ndx];
1255 inputFloats2[ndx] = TCU_NAN;
1261 "OpCapability Shader\n"
1262 "%std450 = OpExtInstImport \"GLSL.std.450\"\n"
1263 "OpMemoryModel Logical GLSL450\n"
1264 "OpEntryPoint GLCompute %main \"main\" %id\n"
1265 "OpExecutionMode %main LocalSize 1 1 1\n"
1267 "OpName %main \"main\"\n"
1268 "OpName %id \"gl_GlobalInvocationID\"\n"
1270 "OpDecorate %id BuiltIn GlobalInvocationId\n"
1272 "OpDecorate %buf BufferBlock\n"
1273 "OpDecorate %indata1 DescriptorSet 0\n"
1274 "OpDecorate %indata1 Binding 0\n"
1275 "OpDecorate %indata2 DescriptorSet 0\n"
1276 "OpDecorate %indata2 Binding 1\n"
1277 "OpDecorate %outdata DescriptorSet 0\n"
1278 "OpDecorate %outdata Binding 2\n"
1279 "OpDecorate %f32arr ArrayStride 4\n"
1280 "OpMemberDecorate %buf 0 Offset 0\n"
1282 + string(getComputeAsmCommonTypes()) +
1284 "%buf = OpTypeStruct %f32arr\n"
1285 "%bufptr = OpTypePointer Uniform %buf\n"
1286 "%indata1 = OpVariable %bufptr Uniform\n"
1287 "%indata2 = OpVariable %bufptr Uniform\n"
1288 "%outdata = OpVariable %bufptr Uniform\n"
1290 "%id = OpVariable %uvec3ptr Input\n"
1291 "%zero = OpConstant %i32 0\n"
1293 "%main = OpFunction %void None %voidf\n"
1294 "%label = OpLabel\n"
1295 "%idval = OpLoad %uvec3 %id\n"
1296 "%x = OpCompositeExtract %u32 %idval 0\n"
1297 "%inloc1 = OpAccessChain %f32ptr %indata1 %zero %x\n"
1298 "%inval1 = OpLoad %f32 %inloc1\n"
1299 "%inloc2 = OpAccessChain %f32ptr %indata2 %zero %x\n"
1300 "%inval2 = OpLoad %f32 %inloc2\n"
1301 "%rem = OpExtInst %f32 %std450 NMin %inval1 %inval2\n"
1302 "%outloc = OpAccessChain %f32ptr %outdata %zero %x\n"
1303 " OpStore %outloc %rem\n"
1307 spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats1)));
1308 spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats2)));
1309 spec.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
1310 spec.numWorkGroups = IVec3(numElements, 1, 1);
1311 spec.verifyIO = &compareNMin;
1313 group->addChild(new SpvAsmComputeShaderCase(testCtx, "all", "", spec));
1315 return group.release();
1318 bool compareNMax (const std::vector<Resource>&, const vector<AllocationSp>& outputAllocs, const std::vector<Resource>& expectedOutputs, TestLog&)
1320 if (outputAllocs.size() != 1)
1323 const BufferSp& expectedOutput = expectedOutputs[0].getBuffer();
1324 std::vector<deUint8> data;
1325 expectedOutput->getBytes(data);
1327 const float* const expectedOutputAsFloat = reinterpret_cast<const float*>(&data.front());
1328 const float* const outputAsFloat = static_cast<const float*>(outputAllocs[0]->getHostPtr());
1330 for (size_t idx = 0; idx < expectedOutput->getByteSize() / sizeof(float); ++idx)
1332 const float f0 = expectedOutputAsFloat[idx];
1333 const float f1 = outputAsFloat[idx];
1335 // For NMax, NaN is considered acceptable result, since in
1336 // architectures that do not handle NaNs, those are huge values.
1337 if (!tcu::Float32(f1).isNaN() && deFloatAbs(f1 - f0) > 0.00001f)
1344 tcu::TestCaseGroup* createOpNMaxGroup (tcu::TestContext& testCtx)
1346 de::MovePtr<tcu::TestCaseGroup> group(new tcu::TestCaseGroup(testCtx, "opnmax", "Test the OpNMax instruction"));
1347 ComputeShaderSpec spec;
1348 de::Random rnd (deStringHash(group->getName()));
1349 const int numElements = 200;
1350 vector<float> inputFloats1 (numElements, 0);
1351 vector<float> inputFloats2 (numElements, 0);
1352 vector<float> outputFloats (numElements, 0);
1354 fillRandomScalars(rnd, -10000.f, 10000.f, &inputFloats1[0], numElements);
1355 fillRandomScalars(rnd, -10000.f, 10000.f, &inputFloats2[0], numElements);
1357 // Make the first case a full-NAN case.
1358 inputFloats1[0] = TCU_NAN;
1359 inputFloats2[0] = TCU_NAN;
1361 for (size_t ndx = 0; ndx < numElements; ++ndx)
1363 // By default, pick the biggest
1364 outputFloats[ndx] = std::max(inputFloats1[ndx], inputFloats2[ndx]);
1366 // Make half of the cases NaN cases
1369 // Alternate between the NaN operand
1372 outputFloats[ndx] = inputFloats2[ndx];
1373 inputFloats1[ndx] = TCU_NAN;
1377 outputFloats[ndx] = inputFloats1[ndx];
1378 inputFloats2[ndx] = TCU_NAN;
1384 "OpCapability Shader\n"
1385 "%std450 = OpExtInstImport \"GLSL.std.450\"\n"
1386 "OpMemoryModel Logical GLSL450\n"
1387 "OpEntryPoint GLCompute %main \"main\" %id\n"
1388 "OpExecutionMode %main LocalSize 1 1 1\n"
1390 "OpName %main \"main\"\n"
1391 "OpName %id \"gl_GlobalInvocationID\"\n"
1393 "OpDecorate %id BuiltIn GlobalInvocationId\n"
1395 "OpDecorate %buf BufferBlock\n"
1396 "OpDecorate %indata1 DescriptorSet 0\n"
1397 "OpDecorate %indata1 Binding 0\n"
1398 "OpDecorate %indata2 DescriptorSet 0\n"
1399 "OpDecorate %indata2 Binding 1\n"
1400 "OpDecorate %outdata DescriptorSet 0\n"
1401 "OpDecorate %outdata Binding 2\n"
1402 "OpDecorate %f32arr ArrayStride 4\n"
1403 "OpMemberDecorate %buf 0 Offset 0\n"
1405 + string(getComputeAsmCommonTypes()) +
1407 "%buf = OpTypeStruct %f32arr\n"
1408 "%bufptr = OpTypePointer Uniform %buf\n"
1409 "%indata1 = OpVariable %bufptr Uniform\n"
1410 "%indata2 = OpVariable %bufptr Uniform\n"
1411 "%outdata = OpVariable %bufptr Uniform\n"
1413 "%id = OpVariable %uvec3ptr Input\n"
1414 "%zero = OpConstant %i32 0\n"
1416 "%main = OpFunction %void None %voidf\n"
1417 "%label = OpLabel\n"
1418 "%idval = OpLoad %uvec3 %id\n"
1419 "%x = OpCompositeExtract %u32 %idval 0\n"
1420 "%inloc1 = OpAccessChain %f32ptr %indata1 %zero %x\n"
1421 "%inval1 = OpLoad %f32 %inloc1\n"
1422 "%inloc2 = OpAccessChain %f32ptr %indata2 %zero %x\n"
1423 "%inval2 = OpLoad %f32 %inloc2\n"
1424 "%rem = OpExtInst %f32 %std450 NMax %inval1 %inval2\n"
1425 "%outloc = OpAccessChain %f32ptr %outdata %zero %x\n"
1426 " OpStore %outloc %rem\n"
1430 spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats1)));
1431 spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats2)));
1432 spec.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
1433 spec.numWorkGroups = IVec3(numElements, 1, 1);
1434 spec.verifyIO = &compareNMax;
1436 group->addChild(new SpvAsmComputeShaderCase(testCtx, "all", "", spec));
1438 return group.release();
1441 bool compareNClamp (const std::vector<Resource>&, const vector<AllocationSp>& outputAllocs, const std::vector<Resource>& expectedOutputs, TestLog&)
1443 if (outputAllocs.size() != 1)
1446 const BufferSp& expectedOutput = expectedOutputs[0].getBuffer();
1447 std::vector<deUint8> data;
1448 expectedOutput->getBytes(data);
1450 const float* const expectedOutputAsFloat = reinterpret_cast<const float*>(&data.front());
1451 const float* const outputAsFloat = static_cast<const float*>(outputAllocs[0]->getHostPtr());
1453 for (size_t idx = 0; idx < expectedOutput->getByteSize() / sizeof(float) / 2; ++idx)
1455 const float e0 = expectedOutputAsFloat[idx * 2];
1456 const float e1 = expectedOutputAsFloat[idx * 2 + 1];
1457 const float res = outputAsFloat[idx];
1459 // For NClamp, we have two possible outcomes based on
1460 // whether NaNs are handled or not.
1461 // If either min or max value is NaN, the result is undefined,
1462 // so this test doesn't stress those. If the clamped value is
1463 // NaN, and NaNs are handled, the result is min; if NaNs are not
1464 // handled, they are big values that result in max.
1465 // If all three parameters are NaN, the result should be NaN.
1466 if (!((tcu::Float32(e0).isNaN() && tcu::Float32(res).isNaN()) ||
1467 (deFloatAbs(e0 - res) < 0.00001f) ||
1468 (deFloatAbs(e1 - res) < 0.00001f)))
1475 tcu::TestCaseGroup* createOpNClampGroup (tcu::TestContext& testCtx)
1477 de::MovePtr<tcu::TestCaseGroup> group (new tcu::TestCaseGroup(testCtx, "opnclamp", "Test the OpNClamp instruction"));
1478 ComputeShaderSpec spec;
1479 de::Random rnd (deStringHash(group->getName()));
1480 const int numElements = 200;
1481 vector<float> inputFloats1 (numElements, 0);
1482 vector<float> inputFloats2 (numElements, 0);
1483 vector<float> inputFloats3 (numElements, 0);
1484 vector<float> outputFloats (numElements * 2, 0);
1486 fillRandomScalars(rnd, -10000.f, 10000.f, &inputFloats1[0], numElements);
1487 fillRandomScalars(rnd, -10000.f, 10000.f, &inputFloats2[0], numElements);
1488 fillRandomScalars(rnd, -10000.f, 10000.f, &inputFloats3[0], numElements);
1490 for (size_t ndx = 0; ndx < numElements; ++ndx)
1492 // Results are only defined if max value is bigger than min value.
1493 if (inputFloats2[ndx] > inputFloats3[ndx])
1495 float t = inputFloats2[ndx];
1496 inputFloats2[ndx] = inputFloats3[ndx];
1497 inputFloats3[ndx] = t;
1500 // By default, do the clamp, setting both possible answers
1501 float defaultRes = std::min(std::max(inputFloats1[ndx], inputFloats2[ndx]), inputFloats3[ndx]);
1503 float maxResA = std::max(inputFloats1[ndx], inputFloats2[ndx]);
1504 float maxResB = maxResA;
1506 // Alternate between the NaN cases
1509 inputFloats1[ndx] = TCU_NAN;
1510 // If NaN is handled, the result should be same as the clamp minimum.
1511 // If NaN is not handled, the result should clamp to the clamp maximum.
1512 maxResA = inputFloats2[ndx];
1513 maxResB = inputFloats3[ndx];
1517 // Not a NaN case - only one legal result.
1518 maxResA = defaultRes;
1519 maxResB = defaultRes;
1522 outputFloats[ndx * 2] = maxResA;
1523 outputFloats[ndx * 2 + 1] = maxResB;
1526 // Make the first case a full-NAN case.
1527 inputFloats1[0] = TCU_NAN;
1528 inputFloats2[0] = TCU_NAN;
1529 inputFloats3[0] = TCU_NAN;
1530 outputFloats[0] = TCU_NAN;
1531 outputFloats[1] = TCU_NAN;
1534 "OpCapability Shader\n"
1535 "%std450 = OpExtInstImport \"GLSL.std.450\"\n"
1536 "OpMemoryModel Logical GLSL450\n"
1537 "OpEntryPoint GLCompute %main \"main\" %id\n"
1538 "OpExecutionMode %main LocalSize 1 1 1\n"
1540 "OpName %main \"main\"\n"
1541 "OpName %id \"gl_GlobalInvocationID\"\n"
1543 "OpDecorate %id BuiltIn GlobalInvocationId\n"
1545 "OpDecorate %buf BufferBlock\n"
1546 "OpDecorate %indata1 DescriptorSet 0\n"
1547 "OpDecorate %indata1 Binding 0\n"
1548 "OpDecorate %indata2 DescriptorSet 0\n"
1549 "OpDecorate %indata2 Binding 1\n"
1550 "OpDecorate %indata3 DescriptorSet 0\n"
1551 "OpDecorate %indata3 Binding 2\n"
1552 "OpDecorate %outdata DescriptorSet 0\n"
1553 "OpDecorate %outdata Binding 3\n"
1554 "OpDecorate %f32arr ArrayStride 4\n"
1555 "OpMemberDecorate %buf 0 Offset 0\n"
1557 + string(getComputeAsmCommonTypes()) +
1559 "%buf = OpTypeStruct %f32arr\n"
1560 "%bufptr = OpTypePointer Uniform %buf\n"
1561 "%indata1 = OpVariable %bufptr Uniform\n"
1562 "%indata2 = OpVariable %bufptr Uniform\n"
1563 "%indata3 = OpVariable %bufptr Uniform\n"
1564 "%outdata = OpVariable %bufptr Uniform\n"
1566 "%id = OpVariable %uvec3ptr Input\n"
1567 "%zero = OpConstant %i32 0\n"
1569 "%main = OpFunction %void None %voidf\n"
1570 "%label = OpLabel\n"
1571 "%idval = OpLoad %uvec3 %id\n"
1572 "%x = OpCompositeExtract %u32 %idval 0\n"
1573 "%inloc1 = OpAccessChain %f32ptr %indata1 %zero %x\n"
1574 "%inval1 = OpLoad %f32 %inloc1\n"
1575 "%inloc2 = OpAccessChain %f32ptr %indata2 %zero %x\n"
1576 "%inval2 = OpLoad %f32 %inloc2\n"
1577 "%inloc3 = OpAccessChain %f32ptr %indata3 %zero %x\n"
1578 "%inval3 = OpLoad %f32 %inloc3\n"
1579 "%rem = OpExtInst %f32 %std450 NClamp %inval1 %inval2 %inval3\n"
1580 "%outloc = OpAccessChain %f32ptr %outdata %zero %x\n"
1581 " OpStore %outloc %rem\n"
1585 spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats1)));
1586 spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats2)));
1587 spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats3)));
1588 spec.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
1589 spec.numWorkGroups = IVec3(numElements, 1, 1);
1590 spec.verifyIO = &compareNClamp;
1592 group->addChild(new SpvAsmComputeShaderCase(testCtx, "all", "", spec));
1594 return group.release();
1597 tcu::TestCaseGroup* createOpSRemComputeGroup (tcu::TestContext& testCtx, qpTestResult negFailResult)
1599 de::MovePtr<tcu::TestCaseGroup> group (new tcu::TestCaseGroup(testCtx, "opsrem", "Test the OpSRem instruction"));
1600 de::Random rnd (deStringHash(group->getName()));
1601 const int numElements = 200;
1603 const struct CaseParams
1606 const char* failMessage; // customized status message
1607 qpTestResult failResult; // override status on failure
1608 int op1Min, op1Max; // operand ranges
1612 { "positive", "Output doesn't match with expected", QP_TEST_RESULT_FAIL, 0, 65536, 0, 100 },
1613 { "all", "Inconsistent results, but within specification", negFailResult, -65536, 65536, -100, 100 }, // see below
1615 // If either operand is negative the result is undefined. Some implementations may still return correct values.
1617 for (int caseNdx = 0; caseNdx < DE_LENGTH_OF_ARRAY(cases); ++caseNdx)
1619 const CaseParams& params = cases[caseNdx];
1620 ComputeShaderSpec spec;
1621 vector<deInt32> inputInts1 (numElements, 0);
1622 vector<deInt32> inputInts2 (numElements, 0);
1623 vector<deInt32> outputInts (numElements, 0);
1625 fillRandomScalars(rnd, params.op1Min, params.op1Max, &inputInts1[0], numElements);
1626 fillRandomScalars(rnd, params.op2Min, params.op2Max, &inputInts2[0], numElements, filterNotZero);
1628 for (int ndx = 0; ndx < numElements; ++ndx)
1630 // The return value of std::fmod() has the same sign as its first operand, which is how OpFRem spec'd.
1631 outputInts[ndx] = inputInts1[ndx] % inputInts2[ndx];
1635 string(getComputeAsmShaderPreamble()) +
1637 "OpName %main \"main\"\n"
1638 "OpName %id \"gl_GlobalInvocationID\"\n"
1640 "OpDecorate %id BuiltIn GlobalInvocationId\n"
1642 "OpDecorate %buf BufferBlock\n"
1643 "OpDecorate %indata1 DescriptorSet 0\n"
1644 "OpDecorate %indata1 Binding 0\n"
1645 "OpDecorate %indata2 DescriptorSet 0\n"
1646 "OpDecorate %indata2 Binding 1\n"
1647 "OpDecorate %outdata DescriptorSet 0\n"
1648 "OpDecorate %outdata Binding 2\n"
1649 "OpDecorate %i32arr ArrayStride 4\n"
1650 "OpMemberDecorate %buf 0 Offset 0\n"
1652 + string(getComputeAsmCommonTypes()) +
1654 "%buf = OpTypeStruct %i32arr\n"
1655 "%bufptr = OpTypePointer Uniform %buf\n"
1656 "%indata1 = OpVariable %bufptr Uniform\n"
1657 "%indata2 = OpVariable %bufptr Uniform\n"
1658 "%outdata = OpVariable %bufptr Uniform\n"
1660 "%id = OpVariable %uvec3ptr Input\n"
1661 "%zero = OpConstant %i32 0\n"
1663 "%main = OpFunction %void None %voidf\n"
1664 "%label = OpLabel\n"
1665 "%idval = OpLoad %uvec3 %id\n"
1666 "%x = OpCompositeExtract %u32 %idval 0\n"
1667 "%inloc1 = OpAccessChain %i32ptr %indata1 %zero %x\n"
1668 "%inval1 = OpLoad %i32 %inloc1\n"
1669 "%inloc2 = OpAccessChain %i32ptr %indata2 %zero %x\n"
1670 "%inval2 = OpLoad %i32 %inloc2\n"
1671 "%rem = OpSRem %i32 %inval1 %inval2\n"
1672 "%outloc = OpAccessChain %i32ptr %outdata %zero %x\n"
1673 " OpStore %outloc %rem\n"
1677 spec.inputs.push_back (BufferSp(new Int32Buffer(inputInts1)));
1678 spec.inputs.push_back (BufferSp(new Int32Buffer(inputInts2)));
1679 spec.outputs.push_back (BufferSp(new Int32Buffer(outputInts)));
1680 spec.numWorkGroups = IVec3(numElements, 1, 1);
1681 spec.failResult = params.failResult;
1682 spec.failMessage = params.failMessage;
1684 group->addChild(new SpvAsmComputeShaderCase(testCtx, params.name, "", spec));
1687 return group.release();
1690 tcu::TestCaseGroup* createOpSRemComputeGroup64 (tcu::TestContext& testCtx, qpTestResult negFailResult)
1692 de::MovePtr<tcu::TestCaseGroup> group (new tcu::TestCaseGroup(testCtx, "opsrem64", "Test the 64-bit OpSRem instruction"));
1693 de::Random rnd (deStringHash(group->getName()));
1694 const int numElements = 200;
1696 const struct CaseParams
1699 const char* failMessage; // customized status message
1700 qpTestResult failResult; // override status on failure
1704 { "positive", "Output doesn't match with expected", QP_TEST_RESULT_FAIL, true },
1705 { "all", "Inconsistent results, but within specification", negFailResult, false }, // see below
1707 // If either operand is negative the result is undefined. Some implementations may still return correct values.
1709 for (int caseNdx = 0; caseNdx < DE_LENGTH_OF_ARRAY(cases); ++caseNdx)
1711 const CaseParams& params = cases[caseNdx];
1712 ComputeShaderSpec spec;
1713 vector<deInt64> inputInts1 (numElements, 0);
1714 vector<deInt64> inputInts2 (numElements, 0);
1715 vector<deInt64> outputInts (numElements, 0);
1717 if (params.positive)
1719 fillRandomInt64sLogDistributed(rnd, inputInts1, numElements, filterNonNegative);
1720 fillRandomInt64sLogDistributed(rnd, inputInts2, numElements, filterPositive);
1724 fillRandomInt64sLogDistributed(rnd, inputInts1, numElements);
1725 fillRandomInt64sLogDistributed(rnd, inputInts2, numElements, filterNotZero);
1728 for (int ndx = 0; ndx < numElements; ++ndx)
1730 // The return value of std::fmod() has the same sign as its first operand, which is how OpFRem spec'd.
1731 outputInts[ndx] = inputInts1[ndx] % inputInts2[ndx];
1735 "OpCapability Int64\n"
1737 + string(getComputeAsmShaderPreamble()) +
1739 "OpName %main \"main\"\n"
1740 "OpName %id \"gl_GlobalInvocationID\"\n"
1742 "OpDecorate %id BuiltIn GlobalInvocationId\n"
1744 "OpDecorate %buf BufferBlock\n"
1745 "OpDecorate %indata1 DescriptorSet 0\n"
1746 "OpDecorate %indata1 Binding 0\n"
1747 "OpDecorate %indata2 DescriptorSet 0\n"
1748 "OpDecorate %indata2 Binding 1\n"
1749 "OpDecorate %outdata DescriptorSet 0\n"
1750 "OpDecorate %outdata Binding 2\n"
1751 "OpDecorate %i64arr ArrayStride 8\n"
1752 "OpMemberDecorate %buf 0 Offset 0\n"
1754 + string(getComputeAsmCommonTypes())
1755 + string(getComputeAsmCommonInt64Types()) +
1757 "%buf = OpTypeStruct %i64arr\n"
1758 "%bufptr = OpTypePointer Uniform %buf\n"
1759 "%indata1 = OpVariable %bufptr Uniform\n"
1760 "%indata2 = OpVariable %bufptr Uniform\n"
1761 "%outdata = OpVariable %bufptr Uniform\n"
1763 "%id = OpVariable %uvec3ptr Input\n"
1764 "%zero = OpConstant %i64 0\n"
1766 "%main = OpFunction %void None %voidf\n"
1767 "%label = OpLabel\n"
1768 "%idval = OpLoad %uvec3 %id\n"
1769 "%x = OpCompositeExtract %u32 %idval 0\n"
1770 "%inloc1 = OpAccessChain %i64ptr %indata1 %zero %x\n"
1771 "%inval1 = OpLoad %i64 %inloc1\n"
1772 "%inloc2 = OpAccessChain %i64ptr %indata2 %zero %x\n"
1773 "%inval2 = OpLoad %i64 %inloc2\n"
1774 "%rem = OpSRem %i64 %inval1 %inval2\n"
1775 "%outloc = OpAccessChain %i64ptr %outdata %zero %x\n"
1776 " OpStore %outloc %rem\n"
1780 spec.inputs.push_back (BufferSp(new Int64Buffer(inputInts1)));
1781 spec.inputs.push_back (BufferSp(new Int64Buffer(inputInts2)));
1782 spec.outputs.push_back (BufferSp(new Int64Buffer(outputInts)));
1783 spec.numWorkGroups = IVec3(numElements, 1, 1);
1784 spec.failResult = params.failResult;
1785 spec.failMessage = params.failMessage;
1787 spec.requestedVulkanFeatures.coreFeatures.shaderInt64 = VK_TRUE;
1789 group->addChild(new SpvAsmComputeShaderCase(testCtx, params.name, "", spec));
1792 return group.release();
1795 tcu::TestCaseGroup* createOpSModComputeGroup (tcu::TestContext& testCtx, qpTestResult negFailResult)
1797 de::MovePtr<tcu::TestCaseGroup> group (new tcu::TestCaseGroup(testCtx, "opsmod", "Test the OpSMod instruction"));
1798 de::Random rnd (deStringHash(group->getName()));
1799 const int numElements = 200;
1801 const struct CaseParams
1804 const char* failMessage; // customized status message
1805 qpTestResult failResult; // override status on failure
1806 int op1Min, op1Max; // operand ranges
1810 { "positive", "Output doesn't match with expected", QP_TEST_RESULT_FAIL, 0, 65536, 0, 100 },
1811 { "all", "Inconsistent results, but within specification", negFailResult, -65536, 65536, -100, 100 }, // see below
1813 // If either operand is negative the result is undefined. Some implementations may still return correct values.
1815 for (int caseNdx = 0; caseNdx < DE_LENGTH_OF_ARRAY(cases); ++caseNdx)
1817 const CaseParams& params = cases[caseNdx];
1819 ComputeShaderSpec spec;
1820 vector<deInt32> inputInts1 (numElements, 0);
1821 vector<deInt32> inputInts2 (numElements, 0);
1822 vector<deInt32> outputInts (numElements, 0);
1824 fillRandomScalars(rnd, params.op1Min, params.op1Max, &inputInts1[0], numElements);
1825 fillRandomScalars(rnd, params.op2Min, params.op2Max, &inputInts2[0], numElements, filterNotZero);
1827 for (int ndx = 0; ndx < numElements; ++ndx)
1829 deInt32 rem = inputInts1[ndx] % inputInts2[ndx];
1832 outputInts[ndx] = 0;
1834 else if ((inputInts1[ndx] >= 0) == (inputInts2[ndx] >= 0))
1836 // They have the same sign
1837 outputInts[ndx] = rem;
1841 // They have opposite sign. The remainder operation takes the
1842 // sign inputInts1[ndx] but OpSMod is supposed to take ths sign
1843 // of inputInts2[ndx]. Adding inputInts2[ndx] will ensure that
1844 // the result has the correct sign and that it is still
1845 // congruent to inputInts1[ndx] modulo inputInts2[ndx]
1847 // See also http://mathforum.org/library/drmath/view/52343.html
1848 outputInts[ndx] = rem + inputInts2[ndx];
1853 string(getComputeAsmShaderPreamble()) +
1855 "OpName %main \"main\"\n"
1856 "OpName %id \"gl_GlobalInvocationID\"\n"
1858 "OpDecorate %id BuiltIn GlobalInvocationId\n"
1860 "OpDecorate %buf BufferBlock\n"
1861 "OpDecorate %indata1 DescriptorSet 0\n"
1862 "OpDecorate %indata1 Binding 0\n"
1863 "OpDecorate %indata2 DescriptorSet 0\n"
1864 "OpDecorate %indata2 Binding 1\n"
1865 "OpDecorate %outdata DescriptorSet 0\n"
1866 "OpDecorate %outdata Binding 2\n"
1867 "OpDecorate %i32arr ArrayStride 4\n"
1868 "OpMemberDecorate %buf 0 Offset 0\n"
1870 + string(getComputeAsmCommonTypes()) +
1872 "%buf = OpTypeStruct %i32arr\n"
1873 "%bufptr = OpTypePointer Uniform %buf\n"
1874 "%indata1 = OpVariable %bufptr Uniform\n"
1875 "%indata2 = OpVariable %bufptr Uniform\n"
1876 "%outdata = OpVariable %bufptr Uniform\n"
1878 "%id = OpVariable %uvec3ptr Input\n"
1879 "%zero = OpConstant %i32 0\n"
1881 "%main = OpFunction %void None %voidf\n"
1882 "%label = OpLabel\n"
1883 "%idval = OpLoad %uvec3 %id\n"
1884 "%x = OpCompositeExtract %u32 %idval 0\n"
1885 "%inloc1 = OpAccessChain %i32ptr %indata1 %zero %x\n"
1886 "%inval1 = OpLoad %i32 %inloc1\n"
1887 "%inloc2 = OpAccessChain %i32ptr %indata2 %zero %x\n"
1888 "%inval2 = OpLoad %i32 %inloc2\n"
1889 "%rem = OpSMod %i32 %inval1 %inval2\n"
1890 "%outloc = OpAccessChain %i32ptr %outdata %zero %x\n"
1891 " OpStore %outloc %rem\n"
1895 spec.inputs.push_back (BufferSp(new Int32Buffer(inputInts1)));
1896 spec.inputs.push_back (BufferSp(new Int32Buffer(inputInts2)));
1897 spec.outputs.push_back (BufferSp(new Int32Buffer(outputInts)));
1898 spec.numWorkGroups = IVec3(numElements, 1, 1);
1899 spec.failResult = params.failResult;
1900 spec.failMessage = params.failMessage;
1902 group->addChild(new SpvAsmComputeShaderCase(testCtx, params.name, "", spec));
1905 return group.release();
1908 tcu::TestCaseGroup* createOpSModComputeGroup64 (tcu::TestContext& testCtx, qpTestResult negFailResult)
1910 de::MovePtr<tcu::TestCaseGroup> group (new tcu::TestCaseGroup(testCtx, "opsmod64", "Test the OpSMod instruction"));
1911 de::Random rnd (deStringHash(group->getName()));
1912 const int numElements = 200;
1914 const struct CaseParams
1917 const char* failMessage; // customized status message
1918 qpTestResult failResult; // override status on failure
1922 { "positive", "Output doesn't match with expected", QP_TEST_RESULT_FAIL, true },
1923 { "all", "Inconsistent results, but within specification", negFailResult, false }, // see below
1925 // If either operand is negative the result is undefined. Some implementations may still return correct values.
1927 for (int caseNdx = 0; caseNdx < DE_LENGTH_OF_ARRAY(cases); ++caseNdx)
1929 const CaseParams& params = cases[caseNdx];
1931 ComputeShaderSpec spec;
1932 vector<deInt64> inputInts1 (numElements, 0);
1933 vector<deInt64> inputInts2 (numElements, 0);
1934 vector<deInt64> outputInts (numElements, 0);
1937 if (params.positive)
1939 fillRandomInt64sLogDistributed(rnd, inputInts1, numElements, filterNonNegative);
1940 fillRandomInt64sLogDistributed(rnd, inputInts2, numElements, filterPositive);
1944 fillRandomInt64sLogDistributed(rnd, inputInts1, numElements);
1945 fillRandomInt64sLogDistributed(rnd, inputInts2, numElements, filterNotZero);
1948 for (int ndx = 0; ndx < numElements; ++ndx)
1950 deInt64 rem = inputInts1[ndx] % inputInts2[ndx];
1953 outputInts[ndx] = 0;
1955 else if ((inputInts1[ndx] >= 0) == (inputInts2[ndx] >= 0))
1957 // They have the same sign
1958 outputInts[ndx] = rem;
1962 // They have opposite sign. The remainder operation takes the
1963 // sign inputInts1[ndx] but OpSMod is supposed to take ths sign
1964 // of inputInts2[ndx]. Adding inputInts2[ndx] will ensure that
1965 // the result has the correct sign and that it is still
1966 // congruent to inputInts1[ndx] modulo inputInts2[ndx]
1968 // See also http://mathforum.org/library/drmath/view/52343.html
1969 outputInts[ndx] = rem + inputInts2[ndx];
1974 "OpCapability Int64\n"
1976 + string(getComputeAsmShaderPreamble()) +
1978 "OpName %main \"main\"\n"
1979 "OpName %id \"gl_GlobalInvocationID\"\n"
1981 "OpDecorate %id BuiltIn GlobalInvocationId\n"
1983 "OpDecorate %buf BufferBlock\n"
1984 "OpDecorate %indata1 DescriptorSet 0\n"
1985 "OpDecorate %indata1 Binding 0\n"
1986 "OpDecorate %indata2 DescriptorSet 0\n"
1987 "OpDecorate %indata2 Binding 1\n"
1988 "OpDecorate %outdata DescriptorSet 0\n"
1989 "OpDecorate %outdata Binding 2\n"
1990 "OpDecorate %i64arr ArrayStride 8\n"
1991 "OpMemberDecorate %buf 0 Offset 0\n"
1993 + string(getComputeAsmCommonTypes())
1994 + string(getComputeAsmCommonInt64Types()) +
1996 "%buf = OpTypeStruct %i64arr\n"
1997 "%bufptr = OpTypePointer Uniform %buf\n"
1998 "%indata1 = OpVariable %bufptr Uniform\n"
1999 "%indata2 = OpVariable %bufptr Uniform\n"
2000 "%outdata = OpVariable %bufptr Uniform\n"
2002 "%id = OpVariable %uvec3ptr Input\n"
2003 "%zero = OpConstant %i64 0\n"
2005 "%main = OpFunction %void None %voidf\n"
2006 "%label = OpLabel\n"
2007 "%idval = OpLoad %uvec3 %id\n"
2008 "%x = OpCompositeExtract %u32 %idval 0\n"
2009 "%inloc1 = OpAccessChain %i64ptr %indata1 %zero %x\n"
2010 "%inval1 = OpLoad %i64 %inloc1\n"
2011 "%inloc2 = OpAccessChain %i64ptr %indata2 %zero %x\n"
2012 "%inval2 = OpLoad %i64 %inloc2\n"
2013 "%rem = OpSMod %i64 %inval1 %inval2\n"
2014 "%outloc = OpAccessChain %i64ptr %outdata %zero %x\n"
2015 " OpStore %outloc %rem\n"
2019 spec.inputs.push_back (BufferSp(new Int64Buffer(inputInts1)));
2020 spec.inputs.push_back (BufferSp(new Int64Buffer(inputInts2)));
2021 spec.outputs.push_back (BufferSp(new Int64Buffer(outputInts)));
2022 spec.numWorkGroups = IVec3(numElements, 1, 1);
2023 spec.failResult = params.failResult;
2024 spec.failMessage = params.failMessage;
2026 spec.requestedVulkanFeatures.coreFeatures.shaderInt64 = VK_TRUE;
2028 group->addChild(new SpvAsmComputeShaderCase(testCtx, params.name, "", spec));
2031 return group.release();
2034 // Copy contents in the input buffer to the output buffer.
2035 tcu::TestCaseGroup* createOpCopyMemoryGroup (tcu::TestContext& testCtx)
2037 de::MovePtr<tcu::TestCaseGroup> group (new tcu::TestCaseGroup(testCtx, "opcopymemory", "Test the OpCopyMemory instruction"));
2038 de::Random rnd (deStringHash(group->getName()));
2039 const int numElements = 100;
2041 // The following case adds vec4(0., 0.5, 1.5, 2.5) to each of the elements in the input buffer and writes output to the output buffer.
2042 ComputeShaderSpec spec1;
2043 vector<Vec4> inputFloats1 (numElements);
2044 vector<Vec4> outputFloats1 (numElements);
2046 fillRandomScalars(rnd, -200.f, 200.f, &inputFloats1[0], numElements * 4);
2048 // CPU might not use the same rounding mode as the GPU. Use whole numbers to avoid rounding differences.
2049 floorAll(inputFloats1);
2051 for (size_t ndx = 0; ndx < numElements; ++ndx)
2052 outputFloats1[ndx] = inputFloats1[ndx] + Vec4(0.f, 0.5f, 1.5f, 2.5f);
2055 string(getComputeAsmShaderPreamble()) +
2057 "OpName %main \"main\"\n"
2058 "OpName %id \"gl_GlobalInvocationID\"\n"
2060 "OpDecorate %id BuiltIn GlobalInvocationId\n"
2061 "OpDecorate %vec4arr ArrayStride 16\n"
2063 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) +
2065 "%vec4 = OpTypeVector %f32 4\n"
2066 "%vec4ptr_u = OpTypePointer Uniform %vec4\n"
2067 "%vec4ptr_f = OpTypePointer Function %vec4\n"
2068 "%vec4arr = OpTypeRuntimeArray %vec4\n"
2069 "%buf = OpTypeStruct %vec4arr\n"
2070 "%bufptr = OpTypePointer Uniform %buf\n"
2071 "%indata = OpVariable %bufptr Uniform\n"
2072 "%outdata = OpVariable %bufptr Uniform\n"
2074 "%id = OpVariable %uvec3ptr Input\n"
2075 "%zero = OpConstant %i32 0\n"
2076 "%c_f_0 = OpConstant %f32 0.\n"
2077 "%c_f_0_5 = OpConstant %f32 0.5\n"
2078 "%c_f_1_5 = OpConstant %f32 1.5\n"
2079 "%c_f_2_5 = OpConstant %f32 2.5\n"
2080 "%c_vec4 = OpConstantComposite %vec4 %c_f_0 %c_f_0_5 %c_f_1_5 %c_f_2_5\n"
2082 "%main = OpFunction %void None %voidf\n"
2083 "%label = OpLabel\n"
2084 "%v_vec4 = OpVariable %vec4ptr_f Function\n"
2085 "%idval = OpLoad %uvec3 %id\n"
2086 "%x = OpCompositeExtract %u32 %idval 0\n"
2087 "%inloc = OpAccessChain %vec4ptr_u %indata %zero %x\n"
2088 "%outloc = OpAccessChain %vec4ptr_u %outdata %zero %x\n"
2089 " OpCopyMemory %v_vec4 %inloc\n"
2090 "%v_vec4_val = OpLoad %vec4 %v_vec4\n"
2091 "%add = OpFAdd %vec4 %v_vec4_val %c_vec4\n"
2092 " OpStore %outloc %add\n"
2096 spec1.inputs.push_back(BufferSp(new Vec4Buffer(inputFloats1)));
2097 spec1.outputs.push_back(BufferSp(new Vec4Buffer(outputFloats1)));
2098 spec1.numWorkGroups = IVec3(numElements, 1, 1);
2100 group->addChild(new SpvAsmComputeShaderCase(testCtx, "vector", "OpCopyMemory elements of vector type", spec1));
2102 // The following case copies a float[100] variable from the input buffer to the output buffer.
2103 ComputeShaderSpec spec2;
2104 vector<float> inputFloats2 (numElements);
2105 vector<float> outputFloats2 (numElements);
2107 fillRandomScalars(rnd, -200.f, 200.f, &inputFloats2[0], numElements);
2109 for (size_t ndx = 0; ndx < numElements; ++ndx)
2110 outputFloats2[ndx] = inputFloats2[ndx];
2113 string(getComputeAsmShaderPreamble()) +
2115 "OpName %main \"main\"\n"
2116 "OpName %id \"gl_GlobalInvocationID\"\n"
2118 "OpDecorate %id BuiltIn GlobalInvocationId\n"
2119 "OpDecorate %f32arr100 ArrayStride 4\n"
2121 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) +
2123 "%hundred = OpConstant %u32 100\n"
2124 "%f32arr100 = OpTypeArray %f32 %hundred\n"
2125 "%f32arr100ptr_f = OpTypePointer Function %f32arr100\n"
2126 "%f32arr100ptr_u = OpTypePointer Uniform %f32arr100\n"
2127 "%buf = OpTypeStruct %f32arr100\n"
2128 "%bufptr = OpTypePointer Uniform %buf\n"
2129 "%indata = OpVariable %bufptr Uniform\n"
2130 "%outdata = OpVariable %bufptr Uniform\n"
2132 "%id = OpVariable %uvec3ptr Input\n"
2133 "%zero = OpConstant %i32 0\n"
2135 "%main = OpFunction %void None %voidf\n"
2136 "%label = OpLabel\n"
2137 "%var = OpVariable %f32arr100ptr_f Function\n"
2138 "%inarr = OpAccessChain %f32arr100ptr_u %indata %zero\n"
2139 "%outarr = OpAccessChain %f32arr100ptr_u %outdata %zero\n"
2140 " OpCopyMemory %var %inarr\n"
2141 " OpCopyMemory %outarr %var\n"
2145 spec2.inputs.push_back(BufferSp(new Float32Buffer(inputFloats2)));
2146 spec2.outputs.push_back(BufferSp(new Float32Buffer(outputFloats2)));
2147 spec2.numWorkGroups = IVec3(1, 1, 1);
2149 group->addChild(new SpvAsmComputeShaderCase(testCtx, "array", "OpCopyMemory elements of array type", spec2));
2151 // The following case copies a struct{vec4, vec4, vec4, vec4} variable from the input buffer to the output buffer.
2152 ComputeShaderSpec spec3;
2153 vector<float> inputFloats3 (16);
2154 vector<float> outputFloats3 (16);
2156 fillRandomScalars(rnd, -200.f, 200.f, &inputFloats3[0], 16);
2158 for (size_t ndx = 0; ndx < 16; ++ndx)
2159 outputFloats3[ndx] = inputFloats3[ndx];
2162 string(getComputeAsmShaderPreamble()) +
2164 "OpName %main \"main\"\n"
2165 "OpName %id \"gl_GlobalInvocationID\"\n"
2167 "OpDecorate %id BuiltIn GlobalInvocationId\n"
2168 //"OpMemberDecorate %buf 0 Offset 0\n" - exists in getComputeAsmInputOutputBufferTraits
2169 "OpMemberDecorate %buf 1 Offset 16\n"
2170 "OpMemberDecorate %buf 2 Offset 32\n"
2171 "OpMemberDecorate %buf 3 Offset 48\n"
2173 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) +
2175 "%vec4 = OpTypeVector %f32 4\n"
2176 "%buf = OpTypeStruct %vec4 %vec4 %vec4 %vec4\n"
2177 "%bufptr = OpTypePointer Uniform %buf\n"
2178 "%indata = OpVariable %bufptr Uniform\n"
2179 "%outdata = OpVariable %bufptr Uniform\n"
2180 "%vec4stptr = OpTypePointer Function %buf\n"
2182 "%id = OpVariable %uvec3ptr Input\n"
2183 "%zero = OpConstant %i32 0\n"
2185 "%main = OpFunction %void None %voidf\n"
2186 "%label = OpLabel\n"
2187 "%var = OpVariable %vec4stptr Function\n"
2188 " OpCopyMemory %var %indata\n"
2189 " OpCopyMemory %outdata %var\n"
2193 spec3.inputs.push_back(BufferSp(new Float32Buffer(inputFloats3)));
2194 spec3.outputs.push_back(BufferSp(new Float32Buffer(outputFloats3)));
2195 spec3.numWorkGroups = IVec3(1, 1, 1);
2197 group->addChild(new SpvAsmComputeShaderCase(testCtx, "struct", "OpCopyMemory elements of struct type", spec3));
2199 // The following case negates multiple float variables from the input buffer and stores the results to the output buffer.
2200 ComputeShaderSpec spec4;
2201 vector<float> inputFloats4 (numElements);
2202 vector<float> outputFloats4 (numElements);
2204 fillRandomScalars(rnd, -200.f, 200.f, &inputFloats4[0], numElements);
2206 for (size_t ndx = 0; ndx < numElements; ++ndx)
2207 outputFloats4[ndx] = -inputFloats4[ndx];
2210 string(getComputeAsmShaderPreamble()) +
2212 "OpName %main \"main\"\n"
2213 "OpName %id \"gl_GlobalInvocationID\"\n"
2215 "OpDecorate %id BuiltIn GlobalInvocationId\n"
2217 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) + string(getComputeAsmInputOutputBuffer()) +
2219 "%f32ptr_f = OpTypePointer Function %f32\n"
2220 "%id = OpVariable %uvec3ptr Input\n"
2221 "%zero = OpConstant %i32 0\n"
2223 "%main = OpFunction %void None %voidf\n"
2224 "%label = OpLabel\n"
2225 "%var = OpVariable %f32ptr_f Function\n"
2226 "%idval = OpLoad %uvec3 %id\n"
2227 "%x = OpCompositeExtract %u32 %idval 0\n"
2228 "%inloc = OpAccessChain %f32ptr %indata %zero %x\n"
2229 "%outloc = OpAccessChain %f32ptr %outdata %zero %x\n"
2230 " OpCopyMemory %var %inloc\n"
2231 "%val = OpLoad %f32 %var\n"
2232 "%neg = OpFNegate %f32 %val\n"
2233 " OpStore %outloc %neg\n"
2237 spec4.inputs.push_back(BufferSp(new Float32Buffer(inputFloats4)));
2238 spec4.outputs.push_back(BufferSp(new Float32Buffer(outputFloats4)));
2239 spec4.numWorkGroups = IVec3(numElements, 1, 1);
2241 group->addChild(new SpvAsmComputeShaderCase(testCtx, "float", "OpCopyMemory elements of float type", spec4));
2243 return group.release();
2246 tcu::TestCaseGroup* createOpCopyObjectGroup (tcu::TestContext& testCtx)
2248 de::MovePtr<tcu::TestCaseGroup> group (new tcu::TestCaseGroup(testCtx, "opcopyobject", "Test the OpCopyObject instruction"));
2249 ComputeShaderSpec spec;
2250 de::Random rnd (deStringHash(group->getName()));
2251 const int numElements = 100;
2252 vector<float> inputFloats (numElements, 0);
2253 vector<float> outputFloats (numElements, 0);
2255 fillRandomScalars(rnd, -200.f, 200.f, &inputFloats[0], numElements);
2257 // CPU might not use the same rounding mode as the GPU. Use whole numbers to avoid rounding differences.
2258 floorAll(inputFloats);
2260 for (size_t ndx = 0; ndx < numElements; ++ndx)
2261 outputFloats[ndx] = inputFloats[ndx] + 7.5f;
2264 string(getComputeAsmShaderPreamble()) +
2266 "OpName %main \"main\"\n"
2267 "OpName %id \"gl_GlobalInvocationID\"\n"
2269 "OpDecorate %id BuiltIn GlobalInvocationId\n"
2271 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) +
2273 "%fmat = OpTypeMatrix %fvec3 3\n"
2274 "%three = OpConstant %u32 3\n"
2275 "%farr = OpTypeArray %f32 %three\n"
2276 "%fst = OpTypeStruct %f32 %f32\n"
2278 + string(getComputeAsmInputOutputBuffer()) +
2280 "%id = OpVariable %uvec3ptr Input\n"
2281 "%zero = OpConstant %i32 0\n"
2282 "%c_f = OpConstant %f32 1.5\n"
2283 "%c_fvec3 = OpConstantComposite %fvec3 %c_f %c_f %c_f\n"
2284 "%c_fmat = OpConstantComposite %fmat %c_fvec3 %c_fvec3 %c_fvec3\n"
2285 "%c_farr = OpConstantComposite %farr %c_f %c_f %c_f\n"
2286 "%c_fst = OpConstantComposite %fst %c_f %c_f\n"
2288 "%main = OpFunction %void None %voidf\n"
2289 "%label = OpLabel\n"
2290 "%c_f_copy = OpCopyObject %f32 %c_f\n"
2291 "%c_fvec3_copy = OpCopyObject %fvec3 %c_fvec3\n"
2292 "%c_fmat_copy = OpCopyObject %fmat %c_fmat\n"
2293 "%c_farr_copy = OpCopyObject %farr %c_farr\n"
2294 "%c_fst_copy = OpCopyObject %fst %c_fst\n"
2295 "%fvec3_elem = OpCompositeExtract %f32 %c_fvec3_copy 0\n"
2296 "%fmat_elem = OpCompositeExtract %f32 %c_fmat_copy 1 2\n"
2297 "%farr_elem = OpCompositeExtract %f32 %c_farr_copy 2\n"
2298 "%fst_elem = OpCompositeExtract %f32 %c_fst_copy 1\n"
2299 // Add up. 1.5 * 5 = 7.5.
2300 "%add1 = OpFAdd %f32 %c_f_copy %fvec3_elem\n"
2301 "%add2 = OpFAdd %f32 %add1 %fmat_elem\n"
2302 "%add3 = OpFAdd %f32 %add2 %farr_elem\n"
2303 "%add4 = OpFAdd %f32 %add3 %fst_elem\n"
2305 "%idval = OpLoad %uvec3 %id\n"
2306 "%x = OpCompositeExtract %u32 %idval 0\n"
2307 "%inloc = OpAccessChain %f32ptr %indata %zero %x\n"
2308 "%outloc = OpAccessChain %f32ptr %outdata %zero %x\n"
2309 "%inval = OpLoad %f32 %inloc\n"
2310 "%add = OpFAdd %f32 %add4 %inval\n"
2311 " OpStore %outloc %add\n"
2314 spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
2315 spec.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
2316 spec.numWorkGroups = IVec3(numElements, 1, 1);
2318 group->addChild(new SpvAsmComputeShaderCase(testCtx, "spotcheck", "OpCopyObject on different types", spec));
2320 return group.release();
2322 // Assembly code used for testing OpUnreachable is based on GLSL source code:
2326 // layout(std140, set = 0, binding = 0) readonly buffer Input {
2327 // float elements[];
2329 // layout(std140, set = 0, binding = 1) writeonly buffer Output {
2330 // float elements[];
2333 // void not_called_func() {
2334 // // place OpUnreachable here
2337 // uint modulo4(uint val) {
2338 // switch (val % uint(4)) {
2339 // case 0: return 3;
2340 // case 1: return 2;
2341 // case 2: return 1;
2342 // case 3: return 0;
2343 // default: return 100; // place OpUnreachable here
2349 // // place OpUnreachable here
2353 // uint x = gl_GlobalInvocationID.x;
2354 // if (const5() > modulo4(1000)) {
2355 // output_data.elements[x] = -input_data.elements[x];
2357 // // place OpUnreachable here
2358 // output_data.elements[x] = input_data.elements[x];
2362 tcu::TestCaseGroup* createOpUnreachableGroup (tcu::TestContext& testCtx)
2364 de::MovePtr<tcu::TestCaseGroup> group (new tcu::TestCaseGroup(testCtx, "opunreachable", "Test the OpUnreachable instruction"));
2365 ComputeShaderSpec spec;
2366 de::Random rnd (deStringHash(group->getName()));
2367 const int numElements = 100;
2368 vector<float> positiveFloats (numElements, 0);
2369 vector<float> negativeFloats (numElements, 0);
2371 fillRandomScalars(rnd, 1.f, 100.f, &positiveFloats[0], numElements);
2373 for (size_t ndx = 0; ndx < numElements; ++ndx)
2374 negativeFloats[ndx] = -positiveFloats[ndx];
2377 string(getComputeAsmShaderPreamble()) +
2379 "OpSource GLSL 430\n"
2380 "OpName %main \"main\"\n"
2381 "OpName %func_not_called_func \"not_called_func(\"\n"
2382 "OpName %func_modulo4 \"modulo4(u1;\"\n"
2383 "OpName %func_const5 \"const5(\"\n"
2384 "OpName %id \"gl_GlobalInvocationID\"\n"
2386 "OpDecorate %id BuiltIn GlobalInvocationId\n"
2388 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) +
2390 "%u32ptr = OpTypePointer Function %u32\n"
2391 "%uintfuint = OpTypeFunction %u32 %u32ptr\n"
2392 "%unitf = OpTypeFunction %u32\n"
2394 "%id = OpVariable %uvec3ptr Input\n"
2395 "%zero = OpConstant %u32 0\n"
2396 "%one = OpConstant %u32 1\n"
2397 "%two = OpConstant %u32 2\n"
2398 "%three = OpConstant %u32 3\n"
2399 "%four = OpConstant %u32 4\n"
2400 "%five = OpConstant %u32 5\n"
2401 "%hundred = OpConstant %u32 100\n"
2402 "%thousand = OpConstant %u32 1000\n"
2404 + string(getComputeAsmInputOutputBuffer()) +
2407 "%main = OpFunction %void None %voidf\n"
2408 "%main_entry = OpLabel\n"
2409 "%v_thousand = OpVariable %u32ptr Function %thousand\n"
2410 "%idval = OpLoad %uvec3 %id\n"
2411 "%x = OpCompositeExtract %u32 %idval 0\n"
2412 "%inloc = OpAccessChain %f32ptr %indata %zero %x\n"
2413 "%inval = OpLoad %f32 %inloc\n"
2414 "%outloc = OpAccessChain %f32ptr %outdata %zero %x\n"
2415 "%ret_const5 = OpFunctionCall %u32 %func_const5\n"
2416 "%ret_modulo4 = OpFunctionCall %u32 %func_modulo4 %v_thousand\n"
2417 "%cmp_gt = OpUGreaterThan %bool %ret_const5 %ret_modulo4\n"
2418 " OpSelectionMerge %if_end None\n"
2419 " OpBranchConditional %cmp_gt %if_true %if_false\n"
2420 "%if_true = OpLabel\n"
2421 "%negate = OpFNegate %f32 %inval\n"
2422 " OpStore %outloc %negate\n"
2423 " OpBranch %if_end\n"
2424 "%if_false = OpLabel\n"
2425 " OpUnreachable\n" // Unreachable else branch for if statement
2426 "%if_end = OpLabel\n"
2430 // not_called_function()
2431 "%func_not_called_func = OpFunction %void None %voidf\n"
2432 "%not_called_func_entry = OpLabel\n"
2433 " OpUnreachable\n" // Unreachable entry block in not called static function
2437 "%func_modulo4 = OpFunction %u32 None %uintfuint\n"
2438 "%valptr = OpFunctionParameter %u32ptr\n"
2439 "%modulo4_entry = OpLabel\n"
2440 "%val = OpLoad %u32 %valptr\n"
2441 "%modulo = OpUMod %u32 %val %four\n"
2442 " OpSelectionMerge %switch_merge None\n"
2443 " OpSwitch %modulo %default 0 %case0 1 %case1 2 %case2 3 %case3\n"
2444 "%case0 = OpLabel\n"
2445 " OpReturnValue %three\n"
2446 "%case1 = OpLabel\n"
2447 " OpReturnValue %two\n"
2448 "%case2 = OpLabel\n"
2449 " OpReturnValue %one\n"
2450 "%case3 = OpLabel\n"
2451 " OpReturnValue %zero\n"
2452 "%default = OpLabel\n"
2453 " OpUnreachable\n" // Unreachable default case for switch statement
2454 "%switch_merge = OpLabel\n"
2455 " OpUnreachable\n" // Unreachable merge block for switch statement
2459 "%func_const5 = OpFunction %u32 None %unitf\n"
2460 "%const5_entry = OpLabel\n"
2461 " OpReturnValue %five\n"
2462 "%unreachable = OpLabel\n"
2463 " OpUnreachable\n" // Unreachable block in function
2465 spec.inputs.push_back(BufferSp(new Float32Buffer(positiveFloats)));
2466 spec.outputs.push_back(BufferSp(new Float32Buffer(negativeFloats)));
2467 spec.numWorkGroups = IVec3(numElements, 1, 1);
2469 group->addChild(new SpvAsmComputeShaderCase(testCtx, "all", "OpUnreachable appearing at different places", spec));
2471 return group.release();
2474 // Assembly code used for testing decoration group is based on GLSL source code:
2478 // layout(std140, set = 0, binding = 0) readonly buffer Input0 {
2479 // float elements[];
2481 // layout(std140, set = 0, binding = 1) readonly buffer Input1 {
2482 // float elements[];
2484 // layout(std140, set = 0, binding = 2) readonly buffer Input2 {
2485 // float elements[];
2487 // layout(std140, set = 0, binding = 3) readonly buffer Input3 {
2488 // float elements[];
2490 // layout(std140, set = 0, binding = 4) readonly buffer Input4 {
2491 // float elements[];
2493 // layout(std140, set = 0, binding = 5) writeonly buffer Output {
2494 // float elements[];
2498 // uint x = gl_GlobalInvocationID.x;
2499 // output_data.elements[x] = input_data0.elements[x] + input_data1.elements[x] + input_data2.elements[x] + input_data3.elements[x] + input_data4.elements[x];
2501 tcu::TestCaseGroup* createDecorationGroupGroup (tcu::TestContext& testCtx)
2503 de::MovePtr<tcu::TestCaseGroup> group (new tcu::TestCaseGroup(testCtx, "decoration_group", "Test the OpDecorationGroup & OpGroupDecorate instruction"));
2504 ComputeShaderSpec spec;
2505 de::Random rnd (deStringHash(group->getName()));
2506 const int numElements = 100;
2507 vector<float> inputFloats0 (numElements, 0);
2508 vector<float> inputFloats1 (numElements, 0);
2509 vector<float> inputFloats2 (numElements, 0);
2510 vector<float> inputFloats3 (numElements, 0);
2511 vector<float> inputFloats4 (numElements, 0);
2512 vector<float> outputFloats (numElements, 0);
2514 fillRandomScalars(rnd, -300.f, 300.f, &inputFloats0[0], numElements);
2515 fillRandomScalars(rnd, -300.f, 300.f, &inputFloats1[0], numElements);
2516 fillRandomScalars(rnd, -300.f, 300.f, &inputFloats2[0], numElements);
2517 fillRandomScalars(rnd, -300.f, 300.f, &inputFloats3[0], numElements);
2518 fillRandomScalars(rnd, -300.f, 300.f, &inputFloats4[0], numElements);
2520 // CPU might not use the same rounding mode as the GPU. Use whole numbers to avoid rounding differences.
2521 floorAll(inputFloats0);
2522 floorAll(inputFloats1);
2523 floorAll(inputFloats2);
2524 floorAll(inputFloats3);
2525 floorAll(inputFloats4);
2527 for (size_t ndx = 0; ndx < numElements; ++ndx)
2528 outputFloats[ndx] = inputFloats0[ndx] + inputFloats1[ndx] + inputFloats2[ndx] + inputFloats3[ndx] + inputFloats4[ndx];
2531 string(getComputeAsmShaderPreamble()) +
2533 "OpSource GLSL 430\n"
2534 "OpName %main \"main\"\n"
2535 "OpName %id \"gl_GlobalInvocationID\"\n"
2537 // Not using group decoration on variable.
2538 "OpDecorate %id BuiltIn GlobalInvocationId\n"
2539 // Not using group decoration on type.
2540 "OpDecorate %f32arr ArrayStride 4\n"
2542 "OpDecorate %groups BufferBlock\n"
2543 "OpDecorate %groupm Offset 0\n"
2544 "%groups = OpDecorationGroup\n"
2545 "%groupm = OpDecorationGroup\n"
2547 // Group decoration on multiple structs.
2548 "OpGroupDecorate %groups %outbuf %inbuf0 %inbuf1 %inbuf2 %inbuf3 %inbuf4\n"
2549 // Group decoration on multiple struct members.
2550 "OpGroupMemberDecorate %groupm %outbuf 0 %inbuf0 0 %inbuf1 0 %inbuf2 0 %inbuf3 0 %inbuf4 0\n"
2552 "OpDecorate %group1 DescriptorSet 0\n"
2553 "OpDecorate %group3 DescriptorSet 0\n"
2554 "OpDecorate %group3 NonWritable\n"
2555 "OpDecorate %group3 Restrict\n"
2556 "%group0 = OpDecorationGroup\n"
2557 "%group1 = OpDecorationGroup\n"
2558 "%group3 = OpDecorationGroup\n"
2560 // Applying the same decoration group multiple times.
2561 "OpGroupDecorate %group1 %outdata\n"
2562 "OpGroupDecorate %group1 %outdata\n"
2563 "OpGroupDecorate %group1 %outdata\n"
2564 "OpDecorate %outdata DescriptorSet 0\n"
2565 "OpDecorate %outdata Binding 5\n"
2566 // Applying decoration group containing nothing.
2567 "OpGroupDecorate %group0 %indata0\n"
2568 "OpDecorate %indata0 DescriptorSet 0\n"
2569 "OpDecorate %indata0 Binding 0\n"
2570 // Applying decoration group containing one decoration.
2571 "OpGroupDecorate %group1 %indata1\n"
2572 "OpDecorate %indata1 Binding 1\n"
2573 // Applying decoration group containing multiple decorations.
2574 "OpGroupDecorate %group3 %indata2 %indata3\n"
2575 "OpDecorate %indata2 Binding 2\n"
2576 "OpDecorate %indata3 Binding 3\n"
2577 // Applying multiple decoration groups (with overlapping).
2578 "OpGroupDecorate %group0 %indata4\n"
2579 "OpGroupDecorate %group1 %indata4\n"
2580 "OpGroupDecorate %group3 %indata4\n"
2581 "OpDecorate %indata4 Binding 4\n"
2583 + string(getComputeAsmCommonTypes()) +
2585 "%id = OpVariable %uvec3ptr Input\n"
2586 "%zero = OpConstant %i32 0\n"
2588 "%outbuf = OpTypeStruct %f32arr\n"
2589 "%outbufptr = OpTypePointer Uniform %outbuf\n"
2590 "%outdata = OpVariable %outbufptr Uniform\n"
2591 "%inbuf0 = OpTypeStruct %f32arr\n"
2592 "%inbuf0ptr = OpTypePointer Uniform %inbuf0\n"
2593 "%indata0 = OpVariable %inbuf0ptr Uniform\n"
2594 "%inbuf1 = OpTypeStruct %f32arr\n"
2595 "%inbuf1ptr = OpTypePointer Uniform %inbuf1\n"
2596 "%indata1 = OpVariable %inbuf1ptr Uniform\n"
2597 "%inbuf2 = OpTypeStruct %f32arr\n"
2598 "%inbuf2ptr = OpTypePointer Uniform %inbuf2\n"
2599 "%indata2 = OpVariable %inbuf2ptr Uniform\n"
2600 "%inbuf3 = OpTypeStruct %f32arr\n"
2601 "%inbuf3ptr = OpTypePointer Uniform %inbuf3\n"
2602 "%indata3 = OpVariable %inbuf3ptr Uniform\n"
2603 "%inbuf4 = OpTypeStruct %f32arr\n"
2604 "%inbufptr = OpTypePointer Uniform %inbuf4\n"
2605 "%indata4 = OpVariable %inbufptr Uniform\n"
2607 "%main = OpFunction %void None %voidf\n"
2608 "%label = OpLabel\n"
2609 "%idval = OpLoad %uvec3 %id\n"
2610 "%x = OpCompositeExtract %u32 %idval 0\n"
2611 "%inloc0 = OpAccessChain %f32ptr %indata0 %zero %x\n"
2612 "%inloc1 = OpAccessChain %f32ptr %indata1 %zero %x\n"
2613 "%inloc2 = OpAccessChain %f32ptr %indata2 %zero %x\n"
2614 "%inloc3 = OpAccessChain %f32ptr %indata3 %zero %x\n"
2615 "%inloc4 = OpAccessChain %f32ptr %indata4 %zero %x\n"
2616 "%outloc = OpAccessChain %f32ptr %outdata %zero %x\n"
2617 "%inval0 = OpLoad %f32 %inloc0\n"
2618 "%inval1 = OpLoad %f32 %inloc1\n"
2619 "%inval2 = OpLoad %f32 %inloc2\n"
2620 "%inval3 = OpLoad %f32 %inloc3\n"
2621 "%inval4 = OpLoad %f32 %inloc4\n"
2622 "%add0 = OpFAdd %f32 %inval0 %inval1\n"
2623 "%add1 = OpFAdd %f32 %add0 %inval2\n"
2624 "%add2 = OpFAdd %f32 %add1 %inval3\n"
2625 "%add = OpFAdd %f32 %add2 %inval4\n"
2626 " OpStore %outloc %add\n"
2629 spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats0)));
2630 spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats1)));
2631 spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats2)));
2632 spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats3)));
2633 spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats4)));
2634 spec.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
2635 spec.numWorkGroups = IVec3(numElements, 1, 1);
2637 group->addChild(new SpvAsmComputeShaderCase(testCtx, "all", "decoration group cases", spec));
2639 return group.release();
2642 struct SpecConstantTwoIntCase
2644 const char* caseName;
2645 const char* scDefinition0;
2646 const char* scDefinition1;
2647 const char* scResultType;
2648 const char* scOperation;
2649 deInt32 scActualValue0;
2650 deInt32 scActualValue1;
2651 const char* resultOperation;
2652 vector<deInt32> expectedOutput;
2653 deInt32 scActualValueLength;
2655 SpecConstantTwoIntCase (const char* name,
2656 const char* definition0,
2657 const char* definition1,
2658 const char* resultType,
2659 const char* operation,
2662 const char* resultOp,
2663 const vector<deInt32>& output,
2664 const deInt32 valueLength = sizeof(deInt32))
2666 , scDefinition0 (definition0)
2667 , scDefinition1 (definition1)
2668 , scResultType (resultType)
2669 , scOperation (operation)
2670 , scActualValue0 (value0)
2671 , scActualValue1 (value1)
2672 , resultOperation (resultOp)
2673 , expectedOutput (output)
2674 , scActualValueLength (valueLength)
2678 tcu::TestCaseGroup* createSpecConstantGroup (tcu::TestContext& testCtx)
2680 de::MovePtr<tcu::TestCaseGroup> group (new tcu::TestCaseGroup(testCtx, "opspecconstantop", "Test the OpSpecConstantOp instruction"));
2681 vector<SpecConstantTwoIntCase> cases;
2682 de::Random rnd (deStringHash(group->getName()));
2683 const int numElements = 100;
2684 const deInt32 p1AsFloat16 = 0x3c00; // +1(fp16) == 0 01111 0000000000 == 0011 1100 0000 0000
2685 vector<deInt32> inputInts (numElements, 0);
2686 vector<deInt32> outputInts1 (numElements, 0);
2687 vector<deInt32> outputInts2 (numElements, 0);
2688 vector<deInt32> outputInts3 (numElements, 0);
2689 vector<deInt32> outputInts4 (numElements, 0);
2690 const StringTemplate shaderTemplate (
2691 "${CAPABILITIES:opt}"
2692 + string(getComputeAsmShaderPreamble()) +
2694 "OpName %main \"main\"\n"
2695 "OpName %id \"gl_GlobalInvocationID\"\n"
2697 "OpDecorate %id BuiltIn GlobalInvocationId\n"
2698 "OpDecorate %sc_0 SpecId 0\n"
2699 "OpDecorate %sc_1 SpecId 1\n"
2700 "OpDecorate %i32arr ArrayStride 4\n"
2702 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) +
2704 "${OPTYPE_DEFINITIONS:opt}"
2705 "%buf = OpTypeStruct %i32arr\n"
2706 "%bufptr = OpTypePointer Uniform %buf\n"
2707 "%indata = OpVariable %bufptr Uniform\n"
2708 "%outdata = OpVariable %bufptr Uniform\n"
2710 "%id = OpVariable %uvec3ptr Input\n"
2711 "%zero = OpConstant %i32 0\n"
2713 "%sc_0 = OpSpecConstant${SC_DEF0}\n"
2714 "%sc_1 = OpSpecConstant${SC_DEF1}\n"
2715 "%sc_final = OpSpecConstantOp ${SC_RESULT_TYPE} ${SC_OP}\n"
2717 "%main = OpFunction %void None %voidf\n"
2718 "%label = OpLabel\n"
2719 "${TYPE_CONVERT:opt}"
2720 "%idval = OpLoad %uvec3 %id\n"
2721 "%x = OpCompositeExtract %u32 %idval 0\n"
2722 "%inloc = OpAccessChain %i32ptr %indata %zero %x\n"
2723 "%inval = OpLoad %i32 %inloc\n"
2724 "%final = ${GEN_RESULT}\n"
2725 "%outloc = OpAccessChain %i32ptr %outdata %zero %x\n"
2726 " OpStore %outloc %final\n"
2728 " OpFunctionEnd\n");
2730 fillRandomScalars(rnd, -65536, 65536, &inputInts[0], numElements);
2732 for (size_t ndx = 0; ndx < numElements; ++ndx)
2734 outputInts1[ndx] = inputInts[ndx] + 42;
2735 outputInts2[ndx] = inputInts[ndx];
2736 outputInts3[ndx] = inputInts[ndx] - 11200;
2737 outputInts4[ndx] = inputInts[ndx] + 1;
2740 const char addScToInput[] = "OpIAdd %i32 %inval %sc_final";
2741 const char addSc32ToInput[] = "OpIAdd %i32 %inval %sc_final32";
2742 const char selectTrueUsingSc[] = "OpSelect %i32 %sc_final %inval %zero";
2743 const char selectFalseUsingSc[] = "OpSelect %i32 %sc_final %zero %inval";
2745 cases.push_back(SpecConstantTwoIntCase("iadd", " %i32 0", " %i32 0", "%i32", "IAdd %sc_0 %sc_1", 62, -20, addScToInput, outputInts1));
2746 cases.push_back(SpecConstantTwoIntCase("isub", " %i32 0", " %i32 0", "%i32", "ISub %sc_0 %sc_1", 100, 58, addScToInput, outputInts1));
2747 cases.push_back(SpecConstantTwoIntCase("imul", " %i32 0", " %i32 0", "%i32", "IMul %sc_0 %sc_1", -2, -21, addScToInput, outputInts1));
2748 cases.push_back(SpecConstantTwoIntCase("sdiv", " %i32 0", " %i32 0", "%i32", "SDiv %sc_0 %sc_1", -126, -3, addScToInput, outputInts1));
2749 cases.push_back(SpecConstantTwoIntCase("udiv", " %i32 0", " %i32 0", "%i32", "UDiv %sc_0 %sc_1", 126, 3, addScToInput, outputInts1));
2750 cases.push_back(SpecConstantTwoIntCase("srem", " %i32 0", " %i32 0", "%i32", "SRem %sc_0 %sc_1", 7, 3, addScToInput, outputInts4));
2751 cases.push_back(SpecConstantTwoIntCase("smod", " %i32 0", " %i32 0", "%i32", "SMod %sc_0 %sc_1", 7, 3, addScToInput, outputInts4));
2752 cases.push_back(SpecConstantTwoIntCase("umod", " %i32 0", " %i32 0", "%i32", "UMod %sc_0 %sc_1", 342, 50, addScToInput, outputInts1));
2753 cases.push_back(SpecConstantTwoIntCase("bitwiseand", " %i32 0", " %i32 0", "%i32", "BitwiseAnd %sc_0 %sc_1", 42, 63, addScToInput, outputInts1));
2754 cases.push_back(SpecConstantTwoIntCase("bitwiseor", " %i32 0", " %i32 0", "%i32", "BitwiseOr %sc_0 %sc_1", 34, 8, addScToInput, outputInts1));
2755 cases.push_back(SpecConstantTwoIntCase("bitwisexor", " %i32 0", " %i32 0", "%i32", "BitwiseXor %sc_0 %sc_1", 18, 56, addScToInput, outputInts1));
2756 cases.push_back(SpecConstantTwoIntCase("shiftrightlogical", " %i32 0", " %i32 0", "%i32", "ShiftRightLogical %sc_0 %sc_1", 168, 2, addScToInput, outputInts1));
2757 cases.push_back(SpecConstantTwoIntCase("shiftrightarithmetic", " %i32 0", " %i32 0", "%i32", "ShiftRightArithmetic %sc_0 %sc_1", 168, 2, addScToInput, outputInts1));
2758 cases.push_back(SpecConstantTwoIntCase("shiftleftlogical", " %i32 0", " %i32 0", "%i32", "ShiftLeftLogical %sc_0 %sc_1", 21, 1, addScToInput, outputInts1));
2759 cases.push_back(SpecConstantTwoIntCase("slessthan", " %i32 0", " %i32 0", "%bool", "SLessThan %sc_0 %sc_1", -20, -10, selectTrueUsingSc, outputInts2));
2760 cases.push_back(SpecConstantTwoIntCase("ulessthan", " %i32 0", " %i32 0", "%bool", "ULessThan %sc_0 %sc_1", 10, 20, selectTrueUsingSc, outputInts2));
2761 cases.push_back(SpecConstantTwoIntCase("sgreaterthan", " %i32 0", " %i32 0", "%bool", "SGreaterThan %sc_0 %sc_1", -1000, 50, selectFalseUsingSc, outputInts2));
2762 cases.push_back(SpecConstantTwoIntCase("ugreaterthan", " %i32 0", " %i32 0", "%bool", "UGreaterThan %sc_0 %sc_1", 10, 5, selectTrueUsingSc, outputInts2));
2763 cases.push_back(SpecConstantTwoIntCase("slessthanequal", " %i32 0", " %i32 0", "%bool", "SLessThanEqual %sc_0 %sc_1", -10, -10, selectTrueUsingSc, outputInts2));
2764 cases.push_back(SpecConstantTwoIntCase("ulessthanequal", " %i32 0", " %i32 0", "%bool", "ULessThanEqual %sc_0 %sc_1", 50, 100, selectTrueUsingSc, outputInts2));
2765 cases.push_back(SpecConstantTwoIntCase("sgreaterthanequal", " %i32 0", " %i32 0", "%bool", "SGreaterThanEqual %sc_0 %sc_1", -1000, 50, selectFalseUsingSc, outputInts2));
2766 cases.push_back(SpecConstantTwoIntCase("ugreaterthanequal", " %i32 0", " %i32 0", "%bool", "UGreaterThanEqual %sc_0 %sc_1", 10, 10, selectTrueUsingSc, outputInts2));
2767 cases.push_back(SpecConstantTwoIntCase("iequal", " %i32 0", " %i32 0", "%bool", "IEqual %sc_0 %sc_1", 42, 24, selectFalseUsingSc, outputInts2));
2768 cases.push_back(SpecConstantTwoIntCase("inotequal", " %i32 0", " %i32 0", "%bool", "INotEqual %sc_0 %sc_1", 42, 24, selectTrueUsingSc, outputInts2));
2769 cases.push_back(SpecConstantTwoIntCase("logicaland", "True %bool", "True %bool", "%bool", "LogicalAnd %sc_0 %sc_1", 0, 1, selectFalseUsingSc, outputInts2));
2770 cases.push_back(SpecConstantTwoIntCase("logicalor", "False %bool", "False %bool", "%bool", "LogicalOr %sc_0 %sc_1", 1, 0, selectTrueUsingSc, outputInts2));
2771 cases.push_back(SpecConstantTwoIntCase("logicalequal", "True %bool", "True %bool", "%bool", "LogicalEqual %sc_0 %sc_1", 0, 1, selectFalseUsingSc, outputInts2));
2772 cases.push_back(SpecConstantTwoIntCase("logicalnotequal", "False %bool", "False %bool", "%bool", "LogicalNotEqual %sc_0 %sc_1", 1, 0, selectTrueUsingSc, outputInts2));
2773 cases.push_back(SpecConstantTwoIntCase("snegate", " %i32 0", " %i32 0", "%i32", "SNegate %sc_0", -42, 0, addScToInput, outputInts1));
2774 cases.push_back(SpecConstantTwoIntCase("not", " %i32 0", " %i32 0", "%i32", "Not %sc_0", -43, 0, addScToInput, outputInts1));
2775 cases.push_back(SpecConstantTwoIntCase("logicalnot", "False %bool", "False %bool", "%bool", "LogicalNot %sc_0", 1, 0, selectFalseUsingSc, outputInts2));
2776 cases.push_back(SpecConstantTwoIntCase("select", "False %bool", " %i32 0", "%i32", "Select %sc_0 %sc_1 %zero", 1, 42, addScToInput, outputInts1));
2777 cases.push_back(SpecConstantTwoIntCase("sconvert", " %i32 0", " %i32 0", "%i16", "SConvert %sc_0", -11200, 0, addSc32ToInput, outputInts3));
2778 // -969998336 stored as 32-bit two's complement is the binary representation of -11200 as IEEE-754 Float
2779 cases.push_back(SpecConstantTwoIntCase("fconvert", " %f32 0", " %f32 0", "%f64", "FConvert %sc_0", -969998336, 0, addSc32ToInput, outputInts3));
2780 cases.push_back(SpecConstantTwoIntCase("fconvert16", " %f16 0", " %f16 0", "%f32", "FConvert %sc_0", p1AsFloat16, 0, addSc32ToInput, outputInts4, sizeof(deFloat16)));
2782 for (size_t caseNdx = 0; caseNdx < cases.size(); ++caseNdx)
2784 map<string, string> specializations;
2785 ComputeShaderSpec spec;
2787 specializations["SC_DEF0"] = cases[caseNdx].scDefinition0;
2788 specializations["SC_DEF1"] = cases[caseNdx].scDefinition1;
2789 specializations["SC_RESULT_TYPE"] = cases[caseNdx].scResultType;
2790 specializations["SC_OP"] = cases[caseNdx].scOperation;
2791 specializations["GEN_RESULT"] = cases[caseNdx].resultOperation;
2793 // Special SPIR-V code for SConvert-case
2794 if (strcmp(cases[caseNdx].caseName, "sconvert") == 0)
2796 spec.requestedVulkanFeatures.coreFeatures.shaderInt16 = VK_TRUE;
2797 specializations["CAPABILITIES"] = "OpCapability Int16\n"; // Adds 16-bit integer capability
2798 specializations["OPTYPE_DEFINITIONS"] = "%i16 = OpTypeInt 16 1\n"; // Adds 16-bit integer type
2799 specializations["TYPE_CONVERT"] = "%sc_final32 = OpSConvert %i32 %sc_final\n"; // Converts 16-bit integer to 32-bit integer
2802 // Special SPIR-V code for FConvert-case
2803 if (strcmp(cases[caseNdx].caseName, "fconvert") == 0)
2805 spec.requestedVulkanFeatures.coreFeatures.shaderFloat64 = VK_TRUE;
2806 specializations["CAPABILITIES"] = "OpCapability Float64\n"; // Adds 64-bit float capability
2807 specializations["OPTYPE_DEFINITIONS"] = "%f64 = OpTypeFloat 64\n"; // Adds 64-bit float type
2808 specializations["TYPE_CONVERT"] = "%sc_final32 = OpConvertFToS %i32 %sc_final\n"; // Converts 64-bit float to 32-bit integer
2811 // Special SPIR-V code for FConvert-case for 16-bit floats
2812 if (strcmp(cases[caseNdx].caseName, "fconvert16") == 0)
2814 spec.extensions.push_back("VK_KHR_shader_float16_int8");
2815 spec.requestedVulkanFeatures.extFloat16Int8 = EXTFLOAT16INT8FEATURES_FLOAT16;
2816 specializations["CAPABILITIES"] = "OpCapability Float16\n"; // Adds 16-bit float capability
2817 specializations["OPTYPE_DEFINITIONS"] = "%f16 = OpTypeFloat 16\n"; // Adds 16-bit float type
2818 specializations["TYPE_CONVERT"] = "%sc_final32 = OpConvertFToS %i32 %sc_final\n"; // Converts 16-bit float to 32-bit integer
2821 spec.assembly = shaderTemplate.specialize(specializations);
2822 spec.inputs.push_back(BufferSp(new Int32Buffer(inputInts)));
2823 spec.outputs.push_back(BufferSp(new Int32Buffer(cases[caseNdx].expectedOutput)));
2824 spec.numWorkGroups = IVec3(numElements, 1, 1);
2825 spec.specConstants.append(&cases[caseNdx].scActualValue0, cases[caseNdx].scActualValueLength);
2826 spec.specConstants.append(&cases[caseNdx].scActualValue1, cases[caseNdx].scActualValueLength);
2828 group->addChild(new SpvAsmComputeShaderCase(testCtx, cases[caseNdx].caseName, cases[caseNdx].caseName, spec));
2831 ComputeShaderSpec spec;
2834 string(getComputeAsmShaderPreamble()) +
2836 "OpName %main \"main\"\n"
2837 "OpName %id \"gl_GlobalInvocationID\"\n"
2839 "OpDecorate %id BuiltIn GlobalInvocationId\n"
2840 "OpDecorate %sc_0 SpecId 0\n"
2841 "OpDecorate %sc_1 SpecId 1\n"
2842 "OpDecorate %sc_2 SpecId 2\n"
2843 "OpDecorate %i32arr ArrayStride 4\n"
2845 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) +
2847 "%ivec3 = OpTypeVector %i32 3\n"
2848 "%buf = OpTypeStruct %i32arr\n"
2849 "%bufptr = OpTypePointer Uniform %buf\n"
2850 "%indata = OpVariable %bufptr Uniform\n"
2851 "%outdata = OpVariable %bufptr Uniform\n"
2853 "%id = OpVariable %uvec3ptr Input\n"
2854 "%zero = OpConstant %i32 0\n"
2855 "%ivec3_0 = OpConstantComposite %ivec3 %zero %zero %zero\n"
2856 "%vec3_undef = OpUndef %ivec3\n"
2858 "%sc_0 = OpSpecConstant %i32 0\n"
2859 "%sc_1 = OpSpecConstant %i32 0\n"
2860 "%sc_2 = OpSpecConstant %i32 0\n"
2861 "%sc_vec3_0 = OpSpecConstantOp %ivec3 CompositeInsert %sc_0 %ivec3_0 0\n" // (sc_0, 0, 0)
2862 "%sc_vec3_1 = OpSpecConstantOp %ivec3 CompositeInsert %sc_1 %ivec3_0 1\n" // (0, sc_1, 0)
2863 "%sc_vec3_2 = OpSpecConstantOp %ivec3 CompositeInsert %sc_2 %ivec3_0 2\n" // (0, 0, sc_2)
2864 "%sc_vec3_0_s = OpSpecConstantOp %ivec3 VectorShuffle %sc_vec3_0 %vec3_undef 0 0xFFFFFFFF 2\n" // (sc_0, ???, 0)
2865 "%sc_vec3_1_s = OpSpecConstantOp %ivec3 VectorShuffle %sc_vec3_1 %vec3_undef 0xFFFFFFFF 1 0\n" // (???, sc_1, 0)
2866 "%sc_vec3_2_s = OpSpecConstantOp %ivec3 VectorShuffle %vec3_undef %sc_vec3_2 5 0xFFFFFFFF 5\n" // (sc_2, ???, sc_2)
2867 "%sc_vec3_01 = OpSpecConstantOp %ivec3 VectorShuffle %sc_vec3_0_s %sc_vec3_1_s 1 0 4\n" // (0, sc_0, sc_1)
2868 "%sc_vec3_012 = OpSpecConstantOp %ivec3 VectorShuffle %sc_vec3_01 %sc_vec3_2_s 5 1 2\n" // (sc_2, sc_0, sc_1)
2869 "%sc_ext_0 = OpSpecConstantOp %i32 CompositeExtract %sc_vec3_012 0\n" // sc_2
2870 "%sc_ext_1 = OpSpecConstantOp %i32 CompositeExtract %sc_vec3_012 1\n" // sc_0
2871 "%sc_ext_2 = OpSpecConstantOp %i32 CompositeExtract %sc_vec3_012 2\n" // sc_1
2872 "%sc_sub = OpSpecConstantOp %i32 ISub %sc_ext_0 %sc_ext_1\n" // (sc_2 - sc_0)
2873 "%sc_final = OpSpecConstantOp %i32 IMul %sc_sub %sc_ext_2\n" // (sc_2 - sc_0) * sc_1
2875 "%main = OpFunction %void None %voidf\n"
2876 "%label = OpLabel\n"
2877 "%idval = OpLoad %uvec3 %id\n"
2878 "%x = OpCompositeExtract %u32 %idval 0\n"
2879 "%inloc = OpAccessChain %i32ptr %indata %zero %x\n"
2880 "%inval = OpLoad %i32 %inloc\n"
2881 "%final = OpIAdd %i32 %inval %sc_final\n"
2882 "%outloc = OpAccessChain %i32ptr %outdata %zero %x\n"
2883 " OpStore %outloc %final\n"
2886 spec.inputs.push_back(BufferSp(new Int32Buffer(inputInts)));
2887 spec.outputs.push_back(BufferSp(new Int32Buffer(outputInts3)));
2888 spec.numWorkGroups = IVec3(numElements, 1, 1);
2889 spec.specConstants.append<deInt32>(123);
2890 spec.specConstants.append<deInt32>(56);
2891 spec.specConstants.append<deInt32>(-77);
2893 group->addChild(new SpvAsmComputeShaderCase(testCtx, "vector_related", "VectorShuffle, CompositeExtract, & CompositeInsert", spec));
2895 return group.release();
2898 void createOpPhiVartypeTests (de::MovePtr<tcu::TestCaseGroup>& group, tcu::TestContext& testCtx)
2900 ComputeShaderSpec specInt;
2901 ComputeShaderSpec specFloat;
2902 ComputeShaderSpec specFloat16;
2903 ComputeShaderSpec specVec3;
2904 ComputeShaderSpec specMat4;
2905 ComputeShaderSpec specArray;
2906 ComputeShaderSpec specStruct;
2907 de::Random rnd (deStringHash(group->getName()));
2908 const int numElements = 100;
2909 vector<float> inputFloats (numElements, 0);
2910 vector<float> outputFloats (numElements, 0);
2911 vector<deFloat16> inputFloats16 (numElements, 0);
2912 vector<deFloat16> outputFloats16 (numElements, 0);
2914 fillRandomScalars(rnd, -300.f, 300.f, &inputFloats[0], numElements);
2916 // CPU might not use the same rounding mode as the GPU. Use whole numbers to avoid rounding differences.
2917 floorAll(inputFloats);
2919 for (size_t ndx = 0; ndx < numElements; ++ndx)
2921 // Just check if the value is positive or not
2922 outputFloats[ndx] = (inputFloats[ndx] > 0) ? 1.0f : -1.0f;
2925 for (size_t ndx = 0; ndx < numElements; ++ndx)
2927 inputFloats16[ndx] = tcu::Float16(inputFloats[ndx]).bits();
2928 outputFloats16[ndx] = tcu::Float16(outputFloats[ndx]).bits();
2931 // All of the tests are of the form:
2935 // if (inputdata > 0)
2942 specFloat.assembly =
2943 string(getComputeAsmShaderPreamble()) +
2945 "OpSource GLSL 430\n"
2946 "OpName %main \"main\"\n"
2947 "OpName %id \"gl_GlobalInvocationID\"\n"
2949 "OpDecorate %id BuiltIn GlobalInvocationId\n"
2951 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) + string(getComputeAsmInputOutputBuffer()) +
2953 "%id = OpVariable %uvec3ptr Input\n"
2954 "%zero = OpConstant %i32 0\n"
2955 "%float_0 = OpConstant %f32 0.0\n"
2956 "%float_1 = OpConstant %f32 1.0\n"
2957 "%float_n1 = OpConstant %f32 -1.0\n"
2959 "%main = OpFunction %void None %voidf\n"
2960 "%entry = OpLabel\n"
2961 "%idval = OpLoad %uvec3 %id\n"
2962 "%x = OpCompositeExtract %u32 %idval 0\n"
2963 "%inloc = OpAccessChain %f32ptr %indata %zero %x\n"
2964 "%inval = OpLoad %f32 %inloc\n"
2966 "%comp = OpFOrdGreaterThan %bool %inval %float_0\n"
2967 " OpSelectionMerge %cm None\n"
2968 " OpBranchConditional %comp %tb %fb\n"
2974 "%res = OpPhi %f32 %float_1 %tb %float_n1 %fb\n"
2976 "%outloc = OpAccessChain %f32ptr %outdata %zero %x\n"
2977 " OpStore %outloc %res\n"
2981 specFloat.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
2982 specFloat.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
2983 specFloat.numWorkGroups = IVec3(numElements, 1, 1);
2985 specFloat16.assembly =
2986 "OpCapability Shader\n"
2987 "OpCapability StorageUniformBufferBlock16\n"
2988 "OpExtension \"SPV_KHR_16bit_storage\"\n"
2989 "OpMemoryModel Logical GLSL450\n"
2990 "OpEntryPoint GLCompute %main \"main\" %id\n"
2991 "OpExecutionMode %main LocalSize 1 1 1\n"
2993 "OpSource GLSL 430\n"
2994 "OpName %main \"main\"\n"
2995 "OpName %id \"gl_GlobalInvocationID\"\n"
2997 "OpDecorate %id BuiltIn GlobalInvocationId\n"
2999 "OpDecorate %buf BufferBlock\n"
3000 "OpDecorate %indata DescriptorSet 0\n"
3001 "OpDecorate %indata Binding 0\n"
3002 "OpDecorate %outdata DescriptorSet 0\n"
3003 "OpDecorate %outdata Binding 1\n"
3004 "OpDecorate %f16arr ArrayStride 2\n"
3005 "OpMemberDecorate %buf 0 Offset 0\n"
3007 "%f16 = OpTypeFloat 16\n"
3008 "%f16ptr = OpTypePointer Uniform %f16\n"
3009 "%f16arr = OpTypeRuntimeArray %f16\n"
3011 + string(getComputeAsmCommonTypes()) +
3013 "%buf = OpTypeStruct %f16arr\n"
3014 "%bufptr = OpTypePointer Uniform %buf\n"
3015 "%indata = OpVariable %bufptr Uniform\n"
3016 "%outdata = OpVariable %bufptr Uniform\n"
3018 "%id = OpVariable %uvec3ptr Input\n"
3019 "%zero = OpConstant %i32 0\n"
3020 "%float_0 = OpConstant %f16 0.0\n"
3021 "%float_1 = OpConstant %f16 1.0\n"
3022 "%float_n1 = OpConstant %f16 -1.0\n"
3024 "%main = OpFunction %void None %voidf\n"
3025 "%entry = OpLabel\n"
3026 "%idval = OpLoad %uvec3 %id\n"
3027 "%x = OpCompositeExtract %u32 %idval 0\n"
3028 "%inloc = OpAccessChain %f16ptr %indata %zero %x\n"
3029 "%inval = OpLoad %f16 %inloc\n"
3031 "%comp = OpFOrdGreaterThan %bool %inval %float_0\n"
3032 " OpSelectionMerge %cm None\n"
3033 " OpBranchConditional %comp %tb %fb\n"
3039 "%res = OpPhi %f16 %float_1 %tb %float_n1 %fb\n"
3041 "%outloc = OpAccessChain %f16ptr %outdata %zero %x\n"
3042 " OpStore %outloc %res\n"
3046 specFloat16.inputs.push_back(BufferSp(new Float16Buffer(inputFloats16)));
3047 specFloat16.outputs.push_back(BufferSp(new Float16Buffer(outputFloats16)));
3048 specFloat16.numWorkGroups = IVec3(numElements, 1, 1);
3049 specFloat16.requestedVulkanFeatures.ext16BitStorage = EXT16BITSTORAGEFEATURES_UNIFORM_BUFFER_BLOCK;
3050 specFloat16.requestedVulkanFeatures.extFloat16Int8 = EXTFLOAT16INT8FEATURES_FLOAT16;
3053 string(getComputeAsmShaderPreamble()) +
3055 "OpSource GLSL 430\n"
3056 "OpName %main \"main\"\n"
3057 "OpName %id \"gl_GlobalInvocationID\"\n"
3059 "OpDecorate %id BuiltIn GlobalInvocationId\n"
3061 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) + string(getComputeAsmInputOutputBuffer()) +
3063 "%id = OpVariable %uvec3ptr Input\n"
3064 "%v4f32 = OpTypeVector %f32 4\n"
3065 "%mat4v4f32 = OpTypeMatrix %v4f32 4\n"
3066 "%zero = OpConstant %i32 0\n"
3067 "%float_0 = OpConstant %f32 0.0\n"
3068 "%float_1 = OpConstant %f32 1.0\n"
3069 "%float_n1 = OpConstant %f32 -1.0\n"
3070 "%m11 = OpConstantComposite %v4f32 %float_1 %float_0 %float_0 %float_0\n"
3071 "%m12 = OpConstantComposite %v4f32 %float_0 %float_1 %float_0 %float_0\n"
3072 "%m13 = OpConstantComposite %v4f32 %float_0 %float_0 %float_1 %float_0\n"
3073 "%m14 = OpConstantComposite %v4f32 %float_0 %float_0 %float_0 %float_1\n"
3074 "%m1 = OpConstantComposite %mat4v4f32 %m11 %m12 %m13 %m14\n"
3075 "%m21 = OpConstantComposite %v4f32 %float_n1 %float_0 %float_0 %float_0\n"
3076 "%m22 = OpConstantComposite %v4f32 %float_0 %float_n1 %float_0 %float_0\n"
3077 "%m23 = OpConstantComposite %v4f32 %float_0 %float_0 %float_n1 %float_0\n"
3078 "%m24 = OpConstantComposite %v4f32 %float_0 %float_0 %float_0 %float_n1\n"
3079 "%m2 = OpConstantComposite %mat4v4f32 %m21 %m22 %m23 %m24\n"
3081 "%main = OpFunction %void None %voidf\n"
3082 "%entry = OpLabel\n"
3083 "%idval = OpLoad %uvec3 %id\n"
3084 "%x = OpCompositeExtract %u32 %idval 0\n"
3085 "%inloc = OpAccessChain %f32ptr %indata %zero %x\n"
3086 "%inval = OpLoad %f32 %inloc\n"
3088 "%comp = OpFOrdGreaterThan %bool %inval %float_0\n"
3089 " OpSelectionMerge %cm None\n"
3090 " OpBranchConditional %comp %tb %fb\n"
3096 "%mres = OpPhi %mat4v4f32 %m1 %tb %m2 %fb\n"
3097 "%res = OpCompositeExtract %f32 %mres 2 2\n"
3099 "%outloc = OpAccessChain %f32ptr %outdata %zero %x\n"
3100 " OpStore %outloc %res\n"
3104 specMat4.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
3105 specMat4.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
3106 specMat4.numWorkGroups = IVec3(numElements, 1, 1);
3109 string(getComputeAsmShaderPreamble()) +
3111 "OpSource GLSL 430\n"
3112 "OpName %main \"main\"\n"
3113 "OpName %id \"gl_GlobalInvocationID\"\n"
3115 "OpDecorate %id BuiltIn GlobalInvocationId\n"
3117 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) + string(getComputeAsmInputOutputBuffer()) +
3119 "%id = OpVariable %uvec3ptr Input\n"
3120 "%zero = OpConstant %i32 0\n"
3121 "%float_0 = OpConstant %f32 0.0\n"
3122 "%float_1 = OpConstant %f32 1.0\n"
3123 "%float_n1 = OpConstant %f32 -1.0\n"
3124 "%v1 = OpConstantComposite %fvec3 %float_1 %float_1 %float_1\n"
3125 "%v2 = OpConstantComposite %fvec3 %float_n1 %float_n1 %float_n1\n"
3127 "%main = OpFunction %void None %voidf\n"
3128 "%entry = OpLabel\n"
3129 "%idval = OpLoad %uvec3 %id\n"
3130 "%x = OpCompositeExtract %u32 %idval 0\n"
3131 "%inloc = OpAccessChain %f32ptr %indata %zero %x\n"
3132 "%inval = OpLoad %f32 %inloc\n"
3134 "%comp = OpFOrdGreaterThan %bool %inval %float_0\n"
3135 " OpSelectionMerge %cm None\n"
3136 " OpBranchConditional %comp %tb %fb\n"
3142 "%vres = OpPhi %fvec3 %v1 %tb %v2 %fb\n"
3143 "%res = OpCompositeExtract %f32 %vres 2\n"
3145 "%outloc = OpAccessChain %f32ptr %outdata %zero %x\n"
3146 " OpStore %outloc %res\n"
3150 specVec3.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
3151 specVec3.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
3152 specVec3.numWorkGroups = IVec3(numElements, 1, 1);
3155 string(getComputeAsmShaderPreamble()) +
3157 "OpSource GLSL 430\n"
3158 "OpName %main \"main\"\n"
3159 "OpName %id \"gl_GlobalInvocationID\"\n"
3161 "OpDecorate %id BuiltIn GlobalInvocationId\n"
3163 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) + string(getComputeAsmInputOutputBuffer()) +
3165 "%id = OpVariable %uvec3ptr Input\n"
3166 "%zero = OpConstant %i32 0\n"
3167 "%float_0 = OpConstant %f32 0.0\n"
3168 "%i1 = OpConstant %i32 1\n"
3169 "%i2 = OpConstant %i32 -1\n"
3171 "%main = OpFunction %void None %voidf\n"
3172 "%entry = OpLabel\n"
3173 "%idval = OpLoad %uvec3 %id\n"
3174 "%x = OpCompositeExtract %u32 %idval 0\n"
3175 "%inloc = OpAccessChain %f32ptr %indata %zero %x\n"
3176 "%inval = OpLoad %f32 %inloc\n"
3178 "%comp = OpFOrdGreaterThan %bool %inval %float_0\n"
3179 " OpSelectionMerge %cm None\n"
3180 " OpBranchConditional %comp %tb %fb\n"
3186 "%ires = OpPhi %i32 %i1 %tb %i2 %fb\n"
3187 "%res = OpConvertSToF %f32 %ires\n"
3189 "%outloc = OpAccessChain %f32ptr %outdata %zero %x\n"
3190 " OpStore %outloc %res\n"
3194 specInt.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
3195 specInt.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
3196 specInt.numWorkGroups = IVec3(numElements, 1, 1);
3198 specArray.assembly =
3199 string(getComputeAsmShaderPreamble()) +
3201 "OpSource GLSL 430\n"
3202 "OpName %main \"main\"\n"
3203 "OpName %id \"gl_GlobalInvocationID\"\n"
3205 "OpDecorate %id BuiltIn GlobalInvocationId\n"
3207 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) + string(getComputeAsmInputOutputBuffer()) +
3209 "%id = OpVariable %uvec3ptr Input\n"
3210 "%zero = OpConstant %i32 0\n"
3211 "%u7 = OpConstant %u32 7\n"
3212 "%float_0 = OpConstant %f32 0.0\n"
3213 "%float_1 = OpConstant %f32 1.0\n"
3214 "%float_n1 = OpConstant %f32 -1.0\n"
3215 "%f32a7 = OpTypeArray %f32 %u7\n"
3216 "%a1 = OpConstantComposite %f32a7 %float_1 %float_1 %float_1 %float_1 %float_1 %float_1 %float_1\n"
3217 "%a2 = OpConstantComposite %f32a7 %float_n1 %float_n1 %float_n1 %float_n1 %float_n1 %float_n1 %float_n1\n"
3218 "%main = OpFunction %void None %voidf\n"
3219 "%entry = OpLabel\n"
3220 "%idval = OpLoad %uvec3 %id\n"
3221 "%x = OpCompositeExtract %u32 %idval 0\n"
3222 "%inloc = OpAccessChain %f32ptr %indata %zero %x\n"
3223 "%inval = OpLoad %f32 %inloc\n"
3225 "%comp = OpFOrdGreaterThan %bool %inval %float_0\n"
3226 " OpSelectionMerge %cm None\n"
3227 " OpBranchConditional %comp %tb %fb\n"
3233 "%ares = OpPhi %f32a7 %a1 %tb %a2 %fb\n"
3234 "%res = OpCompositeExtract %f32 %ares 5\n"
3236 "%outloc = OpAccessChain %f32ptr %outdata %zero %x\n"
3237 " OpStore %outloc %res\n"
3241 specArray.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
3242 specArray.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
3243 specArray.numWorkGroups = IVec3(numElements, 1, 1);
3245 specStruct.assembly =
3246 string(getComputeAsmShaderPreamble()) +
3248 "OpSource GLSL 430\n"
3249 "OpName %main \"main\"\n"
3250 "OpName %id \"gl_GlobalInvocationID\"\n"
3252 "OpDecorate %id BuiltIn GlobalInvocationId\n"
3254 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) + string(getComputeAsmInputOutputBuffer()) +
3256 "%id = OpVariable %uvec3ptr Input\n"
3257 "%zero = OpConstant %i32 0\n"
3258 "%float_0 = OpConstant %f32 0.0\n"
3259 "%float_1 = OpConstant %f32 1.0\n"
3260 "%float_n1 = OpConstant %f32 -1.0\n"
3262 "%v2f32 = OpTypeVector %f32 2\n"
3263 "%Data2 = OpTypeStruct %f32 %v2f32\n"
3264 "%Data = OpTypeStruct %Data2 %f32\n"
3266 "%in1a = OpConstantComposite %v2f32 %float_1 %float_1\n"
3267 "%in1b = OpConstantComposite %Data2 %float_1 %in1a\n"
3268 "%s1 = OpConstantComposite %Data %in1b %float_1\n"
3269 "%in2a = OpConstantComposite %v2f32 %float_n1 %float_n1\n"
3270 "%in2b = OpConstantComposite %Data2 %float_n1 %in2a\n"
3271 "%s2 = OpConstantComposite %Data %in2b %float_n1\n"
3273 "%main = OpFunction %void None %voidf\n"
3274 "%entry = OpLabel\n"
3275 "%idval = OpLoad %uvec3 %id\n"
3276 "%x = OpCompositeExtract %u32 %idval 0\n"
3277 "%inloc = OpAccessChain %f32ptr %indata %zero %x\n"
3278 "%inval = OpLoad %f32 %inloc\n"
3280 "%comp = OpFOrdGreaterThan %bool %inval %float_0\n"
3281 " OpSelectionMerge %cm None\n"
3282 " OpBranchConditional %comp %tb %fb\n"
3288 "%sres = OpPhi %Data %s1 %tb %s2 %fb\n"
3289 "%res = OpCompositeExtract %f32 %sres 0 0\n"
3291 "%outloc = OpAccessChain %f32ptr %outdata %zero %x\n"
3292 " OpStore %outloc %res\n"
3296 specStruct.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
3297 specStruct.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
3298 specStruct.numWorkGroups = IVec3(numElements, 1, 1);
3300 group->addChild(new SpvAsmComputeShaderCase(testCtx, "vartype_int", "OpPhi with int variables", specInt));
3301 group->addChild(new SpvAsmComputeShaderCase(testCtx, "vartype_float", "OpPhi with float variables", specFloat));
3302 group->addChild(new SpvAsmComputeShaderCase(testCtx, "vartype_float16", "OpPhi with 16bit float variables", specFloat16));
3303 group->addChild(new SpvAsmComputeShaderCase(testCtx, "vartype_vec3", "OpPhi with vec3 variables", specVec3));
3304 group->addChild(new SpvAsmComputeShaderCase(testCtx, "vartype_mat4", "OpPhi with mat4 variables", specMat4));
3305 group->addChild(new SpvAsmComputeShaderCase(testCtx, "vartype_array", "OpPhi with array variables", specArray));
3306 group->addChild(new SpvAsmComputeShaderCase(testCtx, "vartype_struct", "OpPhi with struct variables", specStruct));
3309 string generateConstantDefinitions (int count)
3311 std::ostringstream r;
3312 for (int i = 0; i < count; i++)
3313 r << "%cf" << (i * 10 + 5) << " = OpConstant %f32 " <<(i * 10 + 5) << ".0\n";
3318 string generateSwitchCases (int count)
3320 std::ostringstream r;
3321 for (int i = 0; i < count; i++)
3322 r << " " << i << " %case" << i;
3327 string generateSwitchTargets (int count)
3329 std::ostringstream r;
3330 for (int i = 0; i < count; i++)
3331 r << "%case" << i << " = OpLabel\n OpBranch %phi\n";
3336 string generateOpPhiParams (int count)
3338 std::ostringstream r;
3339 for (int i = 0; i < count; i++)
3340 r << " %cf" << (i * 10 + 5) << " %case" << i;
3345 string generateIntWidth (int value)
3347 std::ostringstream r;
3352 // Expand input string by injecting "ABC" between the input
3353 // string characters. The acc/add/treshold parameters are used
3354 // to skip some of the injections to make the result less
3355 // uniform (and a lot shorter).
3356 string expandOpPhiCase5 (const string& s, int &acc, int add, int treshold)
3358 std::ostringstream res;
3359 const char* p = s.c_str();
3375 // Calculate expected result based on the code string
3376 float calcOpPhiCase5 (float val, const string& s)
3378 const char* p = s.c_str();
3381 const float tv[8] = { 0.5f, 1.5f, 3.5f, 7.5f, 15.5f, 31.5f, 63.5f, 127.5f };
3382 const float v = deFloatAbs(val);
3387 for (int i = 7; i >= 0; --i)
3388 x[i] = std::fmod((float)v, (float)(2 << i));
3389 for (int i = 7; i >= 0; --i)
3390 b[i] = x[i] > tv[i];
3397 if (skip == 0 && b[depth])
3408 if (b[depth] || skip)
3422 // In the code string, the letters represent the following:
3425 // if (certain bit is set)
3436 // AABCBC leads to if(){r++;if(){r++;}else{}}else{}
3437 // ABABCC leads to if(){r++;}else{if(){r++;}else{}}
3438 // ABCABC leads to if(){r++;}else{}if(){r++;}else{}
3440 // Code generation gets a bit complicated due to the else-branches,
3441 // which do not generate new values. Thus, the generator needs to
3442 // keep track of the previous variable change seen by the else
3444 string generateOpPhiCase5 (const string& s)
3446 std::stack<int> idStack;
3447 std::stack<std::string> value;
3448 std::stack<std::string> valueLabel;
3449 std::stack<std::string> mergeLeft;
3450 std::stack<std::string> mergeRight;
3451 std::ostringstream res;
3452 const char* p = s.c_str();
3458 value.push("%f32_0");
3459 valueLabel.push("%f32_0 %entry");
3467 idStack.push(currId);
3468 res << "\tOpSelectionMerge %m" << currId << " None\n";
3469 res << "\tOpBranchConditional %b" << depth << " %t" << currId << " %f" << currId << "\n";
3470 res << "%t" << currId << " = OpLabel\n";
3471 res << "%rt" << currId << " = OpFAdd %f32 " << value.top() << " %f32_1\n";
3472 std::ostringstream tag;
3473 tag << "%rt" << currId;
3474 value.push(tag.str());
3475 tag << " %t" << currId;
3476 valueLabel.push(tag.str());
3481 mergeLeft.push(valueLabel.top());
3484 res << "\tOpBranch %m" << currId << "\n";
3485 res << "%f" << currId << " = OpLabel\n";
3486 std::ostringstream tag;
3487 tag << value.top() << " %f" << currId;
3489 valueLabel.push(tag.str());
3494 mergeRight.push(valueLabel.top());
3495 res << "\tOpBranch %m" << currId << "\n";
3496 res << "%m" << currId << " = OpLabel\n";
3498 res << "%res"; // last result goes to %res
3500 res << "%rm" << currId;
3501 res << " = OpPhi %f32 " << mergeLeft.top() << " " << mergeRight.top() << "\n";
3502 std::ostringstream tag;
3503 tag << "%rm" << currId;
3505 value.push(tag.str());
3506 tag << " %m" << currId;
3508 valueLabel.push(tag.str());
3513 currId = idStack.top();
3521 tcu::TestCaseGroup* createOpPhiGroup (tcu::TestContext& testCtx)
3523 de::MovePtr<tcu::TestCaseGroup> group (new tcu::TestCaseGroup(testCtx, "opphi", "Test the OpPhi instruction"));
3524 ComputeShaderSpec spec1;
3525 ComputeShaderSpec spec2;
3526 ComputeShaderSpec spec3;
3527 ComputeShaderSpec spec4;
3528 ComputeShaderSpec spec5;
3529 de::Random rnd (deStringHash(group->getName()));
3530 const int numElements = 100;
3531 vector<float> inputFloats (numElements, 0);
3532 vector<float> outputFloats1 (numElements, 0);
3533 vector<float> outputFloats2 (numElements, 0);
3534 vector<float> outputFloats3 (numElements, 0);
3535 vector<float> outputFloats4 (numElements, 0);
3536 vector<float> outputFloats5 (numElements, 0);
3537 std::string codestring = "ABC";
3538 const int test4Width = 1024;
3540 // Build case 5 code string. Each iteration makes the hierarchy more complicated.
3541 // 9 iterations with (7, 24) parameters makes the hierarchy 8 deep with about 1500 lines of
3543 for (int i = 0, acc = 0; i < 9; i++)
3544 codestring = expandOpPhiCase5(codestring, acc, 7, 24);
3546 fillRandomScalars(rnd, -300.f, 300.f, &inputFloats[0], numElements);
3548 // CPU might not use the same rounding mode as the GPU. Use whole numbers to avoid rounding differences.
3549 floorAll(inputFloats);
3551 for (size_t ndx = 0; ndx < numElements; ++ndx)
3555 case 0: outputFloats1[ndx] = inputFloats[ndx] + 5.5f; break;
3556 case 1: outputFloats1[ndx] = inputFloats[ndx] + 20.5f; break;
3557 case 2: outputFloats1[ndx] = inputFloats[ndx] + 1.75f; break;
3560 outputFloats2[ndx] = inputFloats[ndx] + 6.5f * 3;
3561 outputFloats3[ndx] = 8.5f - inputFloats[ndx];
3563 int index4 = (int)deFloor(deAbs((float)ndx * inputFloats[ndx]));
3564 outputFloats4[ndx] = (float)(index4 % test4Width) * 10.0f + 5.0f;
3566 outputFloats5[ndx] = calcOpPhiCase5(inputFloats[ndx], codestring);
3570 string(getComputeAsmShaderPreamble()) +
3572 "OpSource GLSL 430\n"
3573 "OpName %main \"main\"\n"
3574 "OpName %id \"gl_GlobalInvocationID\"\n"
3576 "OpDecorate %id BuiltIn GlobalInvocationId\n"
3578 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) + string(getComputeAsmInputOutputBuffer()) +
3580 "%id = OpVariable %uvec3ptr Input\n"
3581 "%zero = OpConstant %i32 0\n"
3582 "%three = OpConstant %u32 3\n"
3583 "%constf5p5 = OpConstant %f32 5.5\n"
3584 "%constf20p5 = OpConstant %f32 20.5\n"
3585 "%constf1p75 = OpConstant %f32 1.75\n"
3586 "%constf8p5 = OpConstant %f32 8.5\n"
3587 "%constf6p5 = OpConstant %f32 6.5\n"
3589 "%main = OpFunction %void None %voidf\n"
3590 "%entry = OpLabel\n"
3591 "%idval = OpLoad %uvec3 %id\n"
3592 "%x = OpCompositeExtract %u32 %idval 0\n"
3593 "%selector = OpUMod %u32 %x %three\n"
3594 " OpSelectionMerge %phi None\n"
3595 " OpSwitch %selector %default 0 %case0 1 %case1 2 %case2\n"
3597 // Case 1 before OpPhi.
3598 "%case1 = OpLabel\n"
3601 "%default = OpLabel\n"
3605 "%operand = OpPhi %f32 %constf1p75 %case2 %constf20p5 %case1 %constf5p5 %case0\n" // not in the order of blocks
3606 "%inloc = OpAccessChain %f32ptr %indata %zero %x\n"
3607 "%inval = OpLoad %f32 %inloc\n"
3608 "%add = OpFAdd %f32 %inval %operand\n"
3609 "%outloc = OpAccessChain %f32ptr %outdata %zero %x\n"
3610 " OpStore %outloc %add\n"
3613 // Case 0 after OpPhi.
3614 "%case0 = OpLabel\n"
3618 // Case 2 after OpPhi.
3619 "%case2 = OpLabel\n"
3623 spec1.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
3624 spec1.outputs.push_back(BufferSp(new Float32Buffer(outputFloats1)));
3625 spec1.numWorkGroups = IVec3(numElements, 1, 1);
3627 group->addChild(new SpvAsmComputeShaderCase(testCtx, "block", "out-of-order and unreachable blocks for OpPhi", spec1));
3630 string(getComputeAsmShaderPreamble()) +
3632 "OpName %main \"main\"\n"
3633 "OpName %id \"gl_GlobalInvocationID\"\n"
3635 "OpDecorate %id BuiltIn GlobalInvocationId\n"
3637 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) + string(getComputeAsmInputOutputBuffer()) +
3639 "%id = OpVariable %uvec3ptr Input\n"
3640 "%zero = OpConstant %i32 0\n"
3641 "%one = OpConstant %i32 1\n"
3642 "%three = OpConstant %i32 3\n"
3643 "%constf6p5 = OpConstant %f32 6.5\n"
3645 "%main = OpFunction %void None %voidf\n"
3646 "%entry = OpLabel\n"
3647 "%idval = OpLoad %uvec3 %id\n"
3648 "%x = OpCompositeExtract %u32 %idval 0\n"
3649 "%inloc = OpAccessChain %f32ptr %indata %zero %x\n"
3650 "%outloc = OpAccessChain %f32ptr %outdata %zero %x\n"
3651 "%inval = OpLoad %f32 %inloc\n"
3655 "%step = OpPhi %i32 %zero %entry %step_next %phi\n"
3656 "%accum = OpPhi %f32 %inval %entry %accum_next %phi\n"
3657 "%step_next = OpIAdd %i32 %step %one\n"
3658 "%accum_next = OpFAdd %f32 %accum %constf6p5\n"
3659 "%still_loop = OpSLessThan %bool %step %three\n"
3660 " OpLoopMerge %exit %phi None\n"
3661 " OpBranchConditional %still_loop %phi %exit\n"
3664 " OpStore %outloc %accum\n"
3667 spec2.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
3668 spec2.outputs.push_back(BufferSp(new Float32Buffer(outputFloats2)));
3669 spec2.numWorkGroups = IVec3(numElements, 1, 1);
3671 group->addChild(new SpvAsmComputeShaderCase(testCtx, "induction", "The usual way induction variables are handled in LLVM IR", spec2));
3674 string(getComputeAsmShaderPreamble()) +
3676 "OpName %main \"main\"\n"
3677 "OpName %id \"gl_GlobalInvocationID\"\n"
3679 "OpDecorate %id BuiltIn GlobalInvocationId\n"
3681 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) + string(getComputeAsmInputOutputBuffer()) +
3683 "%f32ptr_f = OpTypePointer Function %f32\n"
3684 "%id = OpVariable %uvec3ptr Input\n"
3685 "%true = OpConstantTrue %bool\n"
3686 "%false = OpConstantFalse %bool\n"
3687 "%zero = OpConstant %i32 0\n"
3688 "%constf8p5 = OpConstant %f32 8.5\n"
3690 "%main = OpFunction %void None %voidf\n"
3691 "%entry = OpLabel\n"
3692 "%b = OpVariable %f32ptr_f Function %constf8p5\n"
3693 "%idval = OpLoad %uvec3 %id\n"
3694 "%x = OpCompositeExtract %u32 %idval 0\n"
3695 "%inloc = OpAccessChain %f32ptr %indata %zero %x\n"
3696 "%outloc = OpAccessChain %f32ptr %outdata %zero %x\n"
3697 "%a_init = OpLoad %f32 %inloc\n"
3698 "%b_init = OpLoad %f32 %b\n"
3702 "%still_loop = OpPhi %bool %true %entry %false %phi\n"
3703 "%a_next = OpPhi %f32 %a_init %entry %b_next %phi\n"
3704 "%b_next = OpPhi %f32 %b_init %entry %a_next %phi\n"
3705 " OpLoopMerge %exit %phi None\n"
3706 " OpBranchConditional %still_loop %phi %exit\n"
3709 "%sub = OpFSub %f32 %a_next %b_next\n"
3710 " OpStore %outloc %sub\n"
3713 spec3.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
3714 spec3.outputs.push_back(BufferSp(new Float32Buffer(outputFloats3)));
3715 spec3.numWorkGroups = IVec3(numElements, 1, 1);
3717 group->addChild(new SpvAsmComputeShaderCase(testCtx, "swap", "Swap the values of two variables using OpPhi", spec3));
3720 "OpCapability Shader\n"
3721 "%ext = OpExtInstImport \"GLSL.std.450\"\n"
3722 "OpMemoryModel Logical GLSL450\n"
3723 "OpEntryPoint GLCompute %main \"main\" %id\n"
3724 "OpExecutionMode %main LocalSize 1 1 1\n"
3726 "OpSource GLSL 430\n"
3727 "OpName %main \"main\"\n"
3728 "OpName %id \"gl_GlobalInvocationID\"\n"
3730 "OpDecorate %id BuiltIn GlobalInvocationId\n"
3732 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) + string(getComputeAsmInputOutputBuffer()) +
3734 "%id = OpVariable %uvec3ptr Input\n"
3735 "%zero = OpConstant %i32 0\n"
3736 "%cimod = OpConstant %u32 " + generateIntWidth(test4Width) + "\n"
3738 + generateConstantDefinitions(test4Width) +
3740 "%main = OpFunction %void None %voidf\n"
3741 "%entry = OpLabel\n"
3742 "%idval = OpLoad %uvec3 %id\n"
3743 "%x = OpCompositeExtract %u32 %idval 0\n"
3744 "%inloc = OpAccessChain %f32ptr %indata %zero %x\n"
3745 "%inval = OpLoad %f32 %inloc\n"
3746 "%xf = OpConvertUToF %f32 %x\n"
3747 "%xm = OpFMul %f32 %xf %inval\n"
3748 "%xa = OpExtInst %f32 %ext FAbs %xm\n"
3749 "%xi = OpConvertFToU %u32 %xa\n"
3750 "%selector = OpUMod %u32 %xi %cimod\n"
3751 " OpSelectionMerge %phi None\n"
3752 " OpSwitch %selector %default "
3754 + generateSwitchCases(test4Width) +
3756 "%default = OpLabel\n"
3759 + generateSwitchTargets(test4Width) +
3762 "%result = OpPhi %f32"
3764 + generateOpPhiParams(test4Width) +
3766 "%outloc = OpAccessChain %f32ptr %outdata %zero %x\n"
3767 " OpStore %outloc %result\n"
3771 spec4.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
3772 spec4.outputs.push_back(BufferSp(new Float32Buffer(outputFloats4)));
3773 spec4.numWorkGroups = IVec3(numElements, 1, 1);
3775 group->addChild(new SpvAsmComputeShaderCase(testCtx, "wide", "OpPhi with a lot of parameters", spec4));
3778 "OpCapability Shader\n"
3779 "%ext = OpExtInstImport \"GLSL.std.450\"\n"
3780 "OpMemoryModel Logical GLSL450\n"
3781 "OpEntryPoint GLCompute %main \"main\" %id\n"
3782 "OpExecutionMode %main LocalSize 1 1 1\n"
3783 "%code = OpString \"" + codestring + "\"\n"
3785 "OpSource GLSL 430\n"
3786 "OpName %main \"main\"\n"
3787 "OpName %id \"gl_GlobalInvocationID\"\n"
3789 "OpDecorate %id BuiltIn GlobalInvocationId\n"
3791 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) + string(getComputeAsmInputOutputBuffer()) +
3793 "%id = OpVariable %uvec3ptr Input\n"
3794 "%zero = OpConstant %i32 0\n"
3795 "%f32_0 = OpConstant %f32 0.0\n"
3796 "%f32_0_5 = OpConstant %f32 0.5\n"
3797 "%f32_1 = OpConstant %f32 1.0\n"
3798 "%f32_1_5 = OpConstant %f32 1.5\n"
3799 "%f32_2 = OpConstant %f32 2.0\n"
3800 "%f32_3_5 = OpConstant %f32 3.5\n"
3801 "%f32_4 = OpConstant %f32 4.0\n"
3802 "%f32_7_5 = OpConstant %f32 7.5\n"
3803 "%f32_8 = OpConstant %f32 8.0\n"
3804 "%f32_15_5 = OpConstant %f32 15.5\n"
3805 "%f32_16 = OpConstant %f32 16.0\n"
3806 "%f32_31_5 = OpConstant %f32 31.5\n"
3807 "%f32_32 = OpConstant %f32 32.0\n"
3808 "%f32_63_5 = OpConstant %f32 63.5\n"
3809 "%f32_64 = OpConstant %f32 64.0\n"
3810 "%f32_127_5 = OpConstant %f32 127.5\n"
3811 "%f32_128 = OpConstant %f32 128.0\n"
3812 "%f32_256 = OpConstant %f32 256.0\n"
3814 "%main = OpFunction %void None %voidf\n"
3815 "%entry = OpLabel\n"
3816 "%idval = OpLoad %uvec3 %id\n"
3817 "%x = OpCompositeExtract %u32 %idval 0\n"
3818 "%inloc = OpAccessChain %f32ptr %indata %zero %x\n"
3819 "%inval = OpLoad %f32 %inloc\n"
3821 "%xabs = OpExtInst %f32 %ext FAbs %inval\n"
3822 "%x8 = OpFMod %f32 %xabs %f32_256\n"
3823 "%x7 = OpFMod %f32 %xabs %f32_128\n"
3824 "%x6 = OpFMod %f32 %xabs %f32_64\n"
3825 "%x5 = OpFMod %f32 %xabs %f32_32\n"
3826 "%x4 = OpFMod %f32 %xabs %f32_16\n"
3827 "%x3 = OpFMod %f32 %xabs %f32_8\n"
3828 "%x2 = OpFMod %f32 %xabs %f32_4\n"
3829 "%x1 = OpFMod %f32 %xabs %f32_2\n"
3831 "%b7 = OpFOrdGreaterThanEqual %bool %x8 %f32_127_5\n"
3832 "%b6 = OpFOrdGreaterThanEqual %bool %x7 %f32_63_5\n"
3833 "%b5 = OpFOrdGreaterThanEqual %bool %x6 %f32_31_5\n"
3834 "%b4 = OpFOrdGreaterThanEqual %bool %x5 %f32_15_5\n"
3835 "%b3 = OpFOrdGreaterThanEqual %bool %x4 %f32_7_5\n"
3836 "%b2 = OpFOrdGreaterThanEqual %bool %x3 %f32_3_5\n"
3837 "%b1 = OpFOrdGreaterThanEqual %bool %x2 %f32_1_5\n"
3838 "%b0 = OpFOrdGreaterThanEqual %bool %x1 %f32_0_5\n"
3840 + generateOpPhiCase5(codestring) +
3842 "%outloc = OpAccessChain %f32ptr %outdata %zero %x\n"
3843 " OpStore %outloc %res\n"
3847 spec5.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
3848 spec5.outputs.push_back(BufferSp(new Float32Buffer(outputFloats5)));
3849 spec5.numWorkGroups = IVec3(numElements, 1, 1);
3851 group->addChild(new SpvAsmComputeShaderCase(testCtx, "nested", "Stress OpPhi with a lot of nesting", spec5));
3853 createOpPhiVartypeTests(group, testCtx);
3855 return group.release();
3858 // Assembly code used for testing block order is based on GLSL source code:
3862 // layout(std140, set = 0, binding = 0) readonly buffer Input {
3863 // float elements[];
3865 // layout(std140, set = 0, binding = 1) writeonly buffer Output {
3866 // float elements[];
3870 // uint x = gl_GlobalInvocationID.x;
3871 // output_data.elements[x] = input_data.elements[x];
3872 // if (x > uint(50)) {
3873 // switch (x % uint(3)) {
3874 // case 0: output_data.elements[x] += 1.5f; break;
3875 // case 1: output_data.elements[x] += 42.f; break;
3876 // case 2: output_data.elements[x] -= 27.f; break;
3880 // output_data.elements[x] = -input_data.elements[x];
3883 tcu::TestCaseGroup* createBlockOrderGroup (tcu::TestContext& testCtx)
3885 de::MovePtr<tcu::TestCaseGroup> group (new tcu::TestCaseGroup(testCtx, "block_order", "Test block orders"));
3886 ComputeShaderSpec spec;
3887 de::Random rnd (deStringHash(group->getName()));
3888 const int numElements = 100;
3889 vector<float> inputFloats (numElements, 0);
3890 vector<float> outputFloats (numElements, 0);
3892 fillRandomScalars(rnd, -100.f, 100.f, &inputFloats[0], numElements);
3894 // CPU might not use the same rounding mode as the GPU. Use whole numbers to avoid rounding differences.
3895 floorAll(inputFloats);
3897 for (size_t ndx = 0; ndx <= 50; ++ndx)
3898 outputFloats[ndx] = -inputFloats[ndx];
3900 for (size_t ndx = 51; ndx < numElements; ++ndx)
3904 case 0: outputFloats[ndx] = inputFloats[ndx] + 1.5f; break;
3905 case 1: outputFloats[ndx] = inputFloats[ndx] + 42.f; break;
3906 case 2: outputFloats[ndx] = inputFloats[ndx] - 27.f; break;
3912 string(getComputeAsmShaderPreamble()) +
3914 "OpSource GLSL 430\n"
3915 "OpName %main \"main\"\n"
3916 "OpName %id \"gl_GlobalInvocationID\"\n"
3918 "OpDecorate %id BuiltIn GlobalInvocationId\n"
3920 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) +
3922 "%u32ptr = OpTypePointer Function %u32\n"
3923 "%u32ptr_input = OpTypePointer Input %u32\n"
3925 + string(getComputeAsmInputOutputBuffer()) +
3927 "%id = OpVariable %uvec3ptr Input\n"
3928 "%zero = OpConstant %i32 0\n"
3929 "%const3 = OpConstant %u32 3\n"
3930 "%const50 = OpConstant %u32 50\n"
3931 "%constf1p5 = OpConstant %f32 1.5\n"
3932 "%constf27 = OpConstant %f32 27.0\n"
3933 "%constf42 = OpConstant %f32 42.0\n"
3935 "%main = OpFunction %void None %voidf\n"
3938 "%entry = OpLabel\n"
3940 // Create a temporary variable to hold the value of gl_GlobalInvocationID.x.
3941 "%xvar = OpVariable %u32ptr Function\n"
3942 "%xptr = OpAccessChain %u32ptr_input %id %zero\n"
3943 "%x = OpLoad %u32 %xptr\n"
3944 " OpStore %xvar %x\n"
3946 "%cmp = OpUGreaterThan %bool %x %const50\n"
3947 " OpSelectionMerge %if_merge None\n"
3948 " OpBranchConditional %cmp %if_true %if_false\n"
3950 // False branch for if-statement: placed in the middle of switch cases and before true branch.
3951 "%if_false = OpLabel\n"
3952 "%x_f = OpLoad %u32 %xvar\n"
3953 "%inloc_f = OpAccessChain %f32ptr %indata %zero %x_f\n"
3954 "%inval_f = OpLoad %f32 %inloc_f\n"
3955 "%negate = OpFNegate %f32 %inval_f\n"
3956 "%outloc_f = OpAccessChain %f32ptr %outdata %zero %x_f\n"
3957 " OpStore %outloc_f %negate\n"
3958 " OpBranch %if_merge\n"
3960 // Merge block for if-statement: placed in the middle of true and false branch.
3961 "%if_merge = OpLabel\n"
3964 // True branch for if-statement: placed in the middle of swtich cases and after the false branch.
3965 "%if_true = OpLabel\n"
3966 "%xval_t = OpLoad %u32 %xvar\n"
3967 "%mod = OpUMod %u32 %xval_t %const3\n"
3968 " OpSelectionMerge %switch_merge None\n"
3969 " OpSwitch %mod %default 0 %case0 1 %case1 2 %case2\n"
3971 // Merge block for switch-statement: placed before the case
3972 // bodies. But it must follow OpSwitch which dominates it.
3973 "%switch_merge = OpLabel\n"
3974 " OpBranch %if_merge\n"
3976 // Case 1 for switch-statement: placed before case 0.
3977 // It must follow the OpSwitch that dominates it.
3978 "%case1 = OpLabel\n"
3979 "%x_1 = OpLoad %u32 %xvar\n"
3980 "%inloc_1 = OpAccessChain %f32ptr %indata %zero %x_1\n"
3981 "%inval_1 = OpLoad %f32 %inloc_1\n"
3982 "%addf42 = OpFAdd %f32 %inval_1 %constf42\n"
3983 "%outloc_1 = OpAccessChain %f32ptr %outdata %zero %x_1\n"
3984 " OpStore %outloc_1 %addf42\n"
3985 " OpBranch %switch_merge\n"
3987 // Case 2 for switch-statement.
3988 "%case2 = OpLabel\n"
3989 "%x_2 = OpLoad %u32 %xvar\n"
3990 "%inloc_2 = OpAccessChain %f32ptr %indata %zero %x_2\n"
3991 "%inval_2 = OpLoad %f32 %inloc_2\n"
3992 "%subf27 = OpFSub %f32 %inval_2 %constf27\n"
3993 "%outloc_2 = OpAccessChain %f32ptr %outdata %zero %x_2\n"
3994 " OpStore %outloc_2 %subf27\n"
3995 " OpBranch %switch_merge\n"
3997 // Default case for switch-statement: placed in the middle of normal cases.
3998 "%default = OpLabel\n"
3999 " OpBranch %switch_merge\n"
4001 // Case 0 for switch-statement: out of order.
4002 "%case0 = OpLabel\n"
4003 "%x_0 = OpLoad %u32 %xvar\n"
4004 "%inloc_0 = OpAccessChain %f32ptr %indata %zero %x_0\n"
4005 "%inval_0 = OpLoad %f32 %inloc_0\n"
4006 "%addf1p5 = OpFAdd %f32 %inval_0 %constf1p5\n"
4007 "%outloc_0 = OpAccessChain %f32ptr %outdata %zero %x_0\n"
4008 " OpStore %outloc_0 %addf1p5\n"
4009 " OpBranch %switch_merge\n"
4012 spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
4013 spec.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
4014 spec.numWorkGroups = IVec3(numElements, 1, 1);
4016 group->addChild(new SpvAsmComputeShaderCase(testCtx, "all", "various out-of-order blocks", spec));
4018 return group.release();
4021 tcu::TestCaseGroup* createMultipleShaderGroup (tcu::TestContext& testCtx)
4023 de::MovePtr<tcu::TestCaseGroup> group (new tcu::TestCaseGroup(testCtx, "multiple_shaders", "Test multiple shaders in the same module"));
4024 ComputeShaderSpec spec1;
4025 ComputeShaderSpec spec2;
4026 de::Random rnd (deStringHash(group->getName()));
4027 const int numElements = 100;
4028 vector<float> inputFloats (numElements, 0);
4029 vector<float> outputFloats1 (numElements, 0);
4030 vector<float> outputFloats2 (numElements, 0);
4031 fillRandomScalars(rnd, -500.f, 500.f, &inputFloats[0], numElements);
4033 for (size_t ndx = 0; ndx < numElements; ++ndx)
4035 outputFloats1[ndx] = inputFloats[ndx] + inputFloats[ndx];
4036 outputFloats2[ndx] = -inputFloats[ndx];
4039 const string assembly(
4040 "OpCapability Shader\n"
4041 "OpMemoryModel Logical GLSL450\n"
4042 "OpEntryPoint GLCompute %comp_main1 \"entrypoint1\" %id\n"
4043 "OpEntryPoint GLCompute %comp_main2 \"entrypoint2\" %id\n"
4044 // A module cannot have two OpEntryPoint instructions with the same Execution Model and the same Name string.
4045 "OpEntryPoint Vertex %vert_main \"entrypoint2\" %vert_builtins %vertexIndex %instanceIndex\n"
4046 "OpExecutionMode %comp_main1 LocalSize 1 1 1\n"
4047 "OpExecutionMode %comp_main2 LocalSize 1 1 1\n"
4049 "OpName %comp_main1 \"entrypoint1\"\n"
4050 "OpName %comp_main2 \"entrypoint2\"\n"
4051 "OpName %vert_main \"entrypoint2\"\n"
4052 "OpName %id \"gl_GlobalInvocationID\"\n"
4053 "OpName %vert_builtin_st \"gl_PerVertex\"\n"
4054 "OpName %vertexIndex \"gl_VertexIndex\"\n"
4055 "OpName %instanceIndex \"gl_InstanceIndex\"\n"
4056 "OpMemberName %vert_builtin_st 0 \"gl_Position\"\n"
4057 "OpMemberName %vert_builtin_st 1 \"gl_PointSize\"\n"
4058 "OpMemberName %vert_builtin_st 2 \"gl_ClipDistance\"\n"
4060 "OpDecorate %id BuiltIn GlobalInvocationId\n"
4061 "OpDecorate %vertexIndex BuiltIn VertexIndex\n"
4062 "OpDecorate %instanceIndex BuiltIn InstanceIndex\n"
4063 "OpDecorate %vert_builtin_st Block\n"
4064 "OpMemberDecorate %vert_builtin_st 0 BuiltIn Position\n"
4065 "OpMemberDecorate %vert_builtin_st 1 BuiltIn PointSize\n"
4066 "OpMemberDecorate %vert_builtin_st 2 BuiltIn ClipDistance\n"
4068 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) + string(getComputeAsmInputOutputBuffer()) +
4070 "%zero = OpConstant %i32 0\n"
4071 "%one = OpConstant %u32 1\n"
4072 "%c_f32_1 = OpConstant %f32 1\n"
4074 "%i32inputptr = OpTypePointer Input %i32\n"
4075 "%vec4 = OpTypeVector %f32 4\n"
4076 "%vec4ptr = OpTypePointer Output %vec4\n"
4077 "%f32arr1 = OpTypeArray %f32 %one\n"
4078 "%vert_builtin_st = OpTypeStruct %vec4 %f32 %f32arr1\n"
4079 "%vert_builtin_st_ptr = OpTypePointer Output %vert_builtin_st\n"
4080 "%vert_builtins = OpVariable %vert_builtin_st_ptr Output\n"
4082 "%id = OpVariable %uvec3ptr Input\n"
4083 "%vertexIndex = OpVariable %i32inputptr Input\n"
4084 "%instanceIndex = OpVariable %i32inputptr Input\n"
4085 "%c_vec4_1 = OpConstantComposite %vec4 %c_f32_1 %c_f32_1 %c_f32_1 %c_f32_1\n"
4087 // gl_Position = vec4(1.);
4088 "%vert_main = OpFunction %void None %voidf\n"
4089 "%vert_entry = OpLabel\n"
4090 "%position = OpAccessChain %vec4ptr %vert_builtins %zero\n"
4091 " OpStore %position %c_vec4_1\n"
4096 "%comp_main1 = OpFunction %void None %voidf\n"
4097 "%comp1_entry = OpLabel\n"
4098 "%idval1 = OpLoad %uvec3 %id\n"
4099 "%x1 = OpCompositeExtract %u32 %idval1 0\n"
4100 "%inloc1 = OpAccessChain %f32ptr %indata %zero %x1\n"
4101 "%inval1 = OpLoad %f32 %inloc1\n"
4102 "%add = OpFAdd %f32 %inval1 %inval1\n"
4103 "%outloc1 = OpAccessChain %f32ptr %outdata %zero %x1\n"
4104 " OpStore %outloc1 %add\n"
4109 "%comp_main2 = OpFunction %void None %voidf\n"
4110 "%comp2_entry = OpLabel\n"
4111 "%idval2 = OpLoad %uvec3 %id\n"
4112 "%x2 = OpCompositeExtract %u32 %idval2 0\n"
4113 "%inloc2 = OpAccessChain %f32ptr %indata %zero %x2\n"
4114 "%inval2 = OpLoad %f32 %inloc2\n"
4115 "%neg = OpFNegate %f32 %inval2\n"
4116 "%outloc2 = OpAccessChain %f32ptr %outdata %zero %x2\n"
4117 " OpStore %outloc2 %neg\n"
4119 " OpFunctionEnd\n");
4121 spec1.assembly = assembly;
4122 spec1.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
4123 spec1.outputs.push_back(BufferSp(new Float32Buffer(outputFloats1)));
4124 spec1.numWorkGroups = IVec3(numElements, 1, 1);
4125 spec1.entryPoint = "entrypoint1";
4127 spec2.assembly = assembly;
4128 spec2.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
4129 spec2.outputs.push_back(BufferSp(new Float32Buffer(outputFloats2)));
4130 spec2.numWorkGroups = IVec3(numElements, 1, 1);
4131 spec2.entryPoint = "entrypoint2";
4133 group->addChild(new SpvAsmComputeShaderCase(testCtx, "shader1", "multiple shaders in the same module", spec1));
4134 group->addChild(new SpvAsmComputeShaderCase(testCtx, "shader2", "multiple shaders in the same module", spec2));
4136 return group.release();
4139 inline std::string makeLongUTF8String (size_t num4ByteChars)
4141 // An example of a longest valid UTF-8 character. Be explicit about the
4142 // character type because Microsoft compilers can otherwise interpret the
4143 // character string as being over wide (16-bit) characters. Ideally, we
4144 // would just use a C++11 UTF-8 string literal, but we want to support older
4145 // Microsoft compilers.
4146 const std::basic_string<char> earthAfrica("\xF0\x9F\x8C\x8D");
4147 std::string longString;
4148 longString.reserve(num4ByteChars * 4);
4149 for (size_t count = 0; count < num4ByteChars; count++)
4151 longString += earthAfrica;
4156 tcu::TestCaseGroup* createOpSourceGroup (tcu::TestContext& testCtx)
4158 de::MovePtr<tcu::TestCaseGroup> group (new tcu::TestCaseGroup(testCtx, "opsource", "Tests the OpSource & OpSourceContinued instruction"));
4159 vector<CaseParameter> cases;
4160 de::Random rnd (deStringHash(group->getName()));
4161 const int numElements = 100;
4162 vector<float> positiveFloats (numElements, 0);
4163 vector<float> negativeFloats (numElements, 0);
4164 const StringTemplate shaderTemplate (
4165 "OpCapability Shader\n"
4166 "OpMemoryModel Logical GLSL450\n"
4168 "OpEntryPoint GLCompute %main \"main\" %id\n"
4169 "OpExecutionMode %main LocalSize 1 1 1\n"
4173 "OpName %main \"main\"\n"
4174 "OpName %id \"gl_GlobalInvocationID\"\n"
4176 "OpDecorate %id BuiltIn GlobalInvocationId\n"
4178 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) + string(getComputeAsmInputOutputBuffer()) +
4180 "%id = OpVariable %uvec3ptr Input\n"
4181 "%zero = OpConstant %i32 0\n"
4183 "%main = OpFunction %void None %voidf\n"
4184 "%label = OpLabel\n"
4185 "%idval = OpLoad %uvec3 %id\n"
4186 "%x = OpCompositeExtract %u32 %idval 0\n"
4187 "%inloc = OpAccessChain %f32ptr %indata %zero %x\n"
4188 "%inval = OpLoad %f32 %inloc\n"
4189 "%neg = OpFNegate %f32 %inval\n"
4190 "%outloc = OpAccessChain %f32ptr %outdata %zero %x\n"
4191 " OpStore %outloc %neg\n"
4193 " OpFunctionEnd\n");
4195 cases.push_back(CaseParameter("unknown_source", "OpSource Unknown 0"));
4196 cases.push_back(CaseParameter("wrong_source", "OpSource OpenCL_C 210"));
4197 cases.push_back(CaseParameter("normal_filename", "%fname = OpString \"filename\"\n"
4198 "OpSource GLSL 430 %fname"));
4199 cases.push_back(CaseParameter("empty_filename", "%fname = OpString \"\"\n"
4200 "OpSource GLSL 430 %fname"));
4201 cases.push_back(CaseParameter("normal_source_code", "%fname = OpString \"filename\"\n"
4202 "OpSource GLSL 430 %fname \"#version 430\nvoid main() {}\""));
4203 cases.push_back(CaseParameter("empty_source_code", "%fname = OpString \"filename\"\n"
4204 "OpSource GLSL 430 %fname \"\""));
4205 cases.push_back(CaseParameter("long_source_code", "%fname = OpString \"filename\"\n"
4206 "OpSource GLSL 430 %fname \"" + makeLongUTF8String(65530) + "ccc\"")); // word count: 65535
4207 cases.push_back(CaseParameter("utf8_source_code", "%fname = OpString \"filename\"\n"
4208 "OpSource GLSL 430 %fname \"\xE2\x98\x82\xE2\x98\x85\"")); // umbrella & black star symbol
4209 cases.push_back(CaseParameter("normal_sourcecontinued", "%fname = OpString \"filename\"\n"
4210 "OpSource GLSL 430 %fname \"#version 430\nvo\"\n"
4211 "OpSourceContinued \"id main() {}\""));
4212 cases.push_back(CaseParameter("empty_sourcecontinued", "%fname = OpString \"filename\"\n"
4213 "OpSource GLSL 430 %fname \"#version 430\nvoid main() {}\"\n"
4214 "OpSourceContinued \"\""));
4215 cases.push_back(CaseParameter("long_sourcecontinued", "%fname = OpString \"filename\"\n"
4216 "OpSource GLSL 430 %fname \"#version 430\nvoid main() {}\"\n"
4217 "OpSourceContinued \"" + makeLongUTF8String(65533) + "ccc\"")); // word count: 65535
4218 cases.push_back(CaseParameter("utf8_sourcecontinued", "%fname = OpString \"filename\"\n"
4219 "OpSource GLSL 430 %fname \"#version 430\nvoid main() {}\"\n"
4220 "OpSourceContinued \"\xE2\x98\x8E\xE2\x9A\x91\"")); // white telephone & black flag symbol
4221 cases.push_back(CaseParameter("multi_sourcecontinued", "%fname = OpString \"filename\"\n"
4222 "OpSource GLSL 430 %fname \"#version 430\n\"\n"
4223 "OpSourceContinued \"void\"\n"
4224 "OpSourceContinued \"main()\"\n"
4225 "OpSourceContinued \"{}\""));
4226 cases.push_back(CaseParameter("empty_source_before_sourcecontinued", "%fname = OpString \"filename\"\n"
4227 "OpSource GLSL 430 %fname \"\"\n"
4228 "OpSourceContinued \"#version 430\nvoid main() {}\""));
4230 fillRandomScalars(rnd, 1.f, 100.f, &positiveFloats[0], numElements);
4232 for (size_t ndx = 0; ndx < numElements; ++ndx)
4233 negativeFloats[ndx] = -positiveFloats[ndx];
4235 for (size_t caseNdx = 0; caseNdx < cases.size(); ++caseNdx)
4237 map<string, string> specializations;
4238 ComputeShaderSpec spec;
4240 specializations["SOURCE"] = cases[caseNdx].param;
4241 spec.assembly = shaderTemplate.specialize(specializations);
4242 spec.inputs.push_back(BufferSp(new Float32Buffer(positiveFloats)));
4243 spec.outputs.push_back(BufferSp(new Float32Buffer(negativeFloats)));
4244 spec.numWorkGroups = IVec3(numElements, 1, 1);
4246 group->addChild(new SpvAsmComputeShaderCase(testCtx, cases[caseNdx].name, cases[caseNdx].name, spec));
4249 return group.release();
4252 tcu::TestCaseGroup* createOpSourceExtensionGroup (tcu::TestContext& testCtx)
4254 de::MovePtr<tcu::TestCaseGroup> group (new tcu::TestCaseGroup(testCtx, "opsourceextension", "Tests the OpSource instruction"));
4255 vector<CaseParameter> cases;
4256 de::Random rnd (deStringHash(group->getName()));
4257 const int numElements = 100;
4258 vector<float> inputFloats (numElements, 0);
4259 vector<float> outputFloats (numElements, 0);
4260 const StringTemplate shaderTemplate (
4261 string(getComputeAsmShaderPreamble()) +
4263 "OpSourceExtension \"${EXTENSION}\"\n"
4265 "OpName %main \"main\"\n"
4266 "OpName %id \"gl_GlobalInvocationID\"\n"
4268 "OpDecorate %id BuiltIn GlobalInvocationId\n"
4270 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) + string(getComputeAsmInputOutputBuffer()) +
4272 "%id = OpVariable %uvec3ptr Input\n"
4273 "%zero = OpConstant %i32 0\n"
4275 "%main = OpFunction %void None %voidf\n"
4276 "%label = OpLabel\n"
4277 "%idval = OpLoad %uvec3 %id\n"
4278 "%x = OpCompositeExtract %u32 %idval 0\n"
4279 "%inloc = OpAccessChain %f32ptr %indata %zero %x\n"
4280 "%inval = OpLoad %f32 %inloc\n"
4281 "%neg = OpFNegate %f32 %inval\n"
4282 "%outloc = OpAccessChain %f32ptr %outdata %zero %x\n"
4283 " OpStore %outloc %neg\n"
4285 " OpFunctionEnd\n");
4287 cases.push_back(CaseParameter("empty_extension", ""));
4288 cases.push_back(CaseParameter("real_extension", "GL_ARB_texture_rectangle"));
4289 cases.push_back(CaseParameter("fake_extension", "GL_ARB_im_the_ultimate_extension"));
4290 cases.push_back(CaseParameter("utf8_extension", "GL_ARB_\xE2\x98\x82\xE2\x98\x85"));
4291 cases.push_back(CaseParameter("long_extension", makeLongUTF8String(65533) + "ccc")); // word count: 65535
4293 fillRandomScalars(rnd, -200.f, 200.f, &inputFloats[0], numElements);
4295 for (size_t ndx = 0; ndx < numElements; ++ndx)
4296 outputFloats[ndx] = -inputFloats[ndx];
4298 for (size_t caseNdx = 0; caseNdx < cases.size(); ++caseNdx)
4300 map<string, string> specializations;
4301 ComputeShaderSpec spec;
4303 specializations["EXTENSION"] = cases[caseNdx].param;
4304 spec.assembly = shaderTemplate.specialize(specializations);
4305 spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
4306 spec.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
4307 spec.numWorkGroups = IVec3(numElements, 1, 1);
4309 group->addChild(new SpvAsmComputeShaderCase(testCtx, cases[caseNdx].name, cases[caseNdx].name, spec));
4312 return group.release();
4315 // Checks that a compute shader can generate a constant null value of various types, without exercising a computation on it.
4316 tcu::TestCaseGroup* createOpConstantNullGroup (tcu::TestContext& testCtx)
4318 de::MovePtr<tcu::TestCaseGroup> group (new tcu::TestCaseGroup(testCtx, "opconstantnull", "Tests the OpConstantNull instruction"));
4319 vector<CaseParameter> cases;
4320 de::Random rnd (deStringHash(group->getName()));
4321 const int numElements = 100;
4322 vector<float> positiveFloats (numElements, 0);
4323 vector<float> negativeFloats (numElements, 0);
4324 const StringTemplate shaderTemplate (
4325 string(getComputeAsmShaderPreamble()) +
4327 "OpSource GLSL 430\n"
4328 "OpName %main \"main\"\n"
4329 "OpName %id \"gl_GlobalInvocationID\"\n"
4331 "OpDecorate %id BuiltIn GlobalInvocationId\n"
4333 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) +
4334 "%uvec2 = OpTypeVector %u32 2\n"
4335 "%bvec3 = OpTypeVector %bool 3\n"
4336 "%fvec4 = OpTypeVector %f32 4\n"
4337 "%fmat33 = OpTypeMatrix %fvec3 3\n"
4338 "%const100 = OpConstant %u32 100\n"
4339 "%uarr100 = OpTypeArray %i32 %const100\n"
4340 "%struct = OpTypeStruct %f32 %i32 %u32\n"
4341 "%pointer = OpTypePointer Function %i32\n"
4342 + string(getComputeAsmInputOutputBuffer()) +
4344 "%null = OpConstantNull ${TYPE}\n"
4346 "%id = OpVariable %uvec3ptr Input\n"
4347 "%zero = OpConstant %i32 0\n"
4349 "%main = OpFunction %void None %voidf\n"
4350 "%label = OpLabel\n"
4351 "%idval = OpLoad %uvec3 %id\n"
4352 "%x = OpCompositeExtract %u32 %idval 0\n"
4353 "%inloc = OpAccessChain %f32ptr %indata %zero %x\n"
4354 "%inval = OpLoad %f32 %inloc\n"
4355 "%neg = OpFNegate %f32 %inval\n"
4356 "%outloc = OpAccessChain %f32ptr %outdata %zero %x\n"
4357 " OpStore %outloc %neg\n"
4359 " OpFunctionEnd\n");
4361 cases.push_back(CaseParameter("bool", "%bool"));
4362 cases.push_back(CaseParameter("sint32", "%i32"));
4363 cases.push_back(CaseParameter("uint32", "%u32"));
4364 cases.push_back(CaseParameter("float32", "%f32"));
4365 cases.push_back(CaseParameter("vec4float32", "%fvec4"));
4366 cases.push_back(CaseParameter("vec3bool", "%bvec3"));
4367 cases.push_back(CaseParameter("vec2uint32", "%uvec2"));
4368 cases.push_back(CaseParameter("matrix", "%fmat33"));
4369 cases.push_back(CaseParameter("array", "%uarr100"));
4370 cases.push_back(CaseParameter("struct", "%struct"));
4371 cases.push_back(CaseParameter("pointer", "%pointer"));
4373 fillRandomScalars(rnd, 1.f, 100.f, &positiveFloats[0], numElements);
4375 for (size_t ndx = 0; ndx < numElements; ++ndx)
4376 negativeFloats[ndx] = -positiveFloats[ndx];
4378 for (size_t caseNdx = 0; caseNdx < cases.size(); ++caseNdx)
4380 map<string, string> specializations;
4381 ComputeShaderSpec spec;
4383 specializations["TYPE"] = cases[caseNdx].param;
4384 spec.assembly = shaderTemplate.specialize(specializations);
4385 spec.inputs.push_back(BufferSp(new Float32Buffer(positiveFloats)));
4386 spec.outputs.push_back(BufferSp(new Float32Buffer(negativeFloats)));
4387 spec.numWorkGroups = IVec3(numElements, 1, 1);
4389 group->addChild(new SpvAsmComputeShaderCase(testCtx, cases[caseNdx].name, cases[caseNdx].name, spec));
4392 return group.release();
4395 // Checks that a compute shader can generate a constant composite value of various types, without exercising a computation on it.
4396 tcu::TestCaseGroup* createOpConstantCompositeGroup (tcu::TestContext& testCtx)
4398 de::MovePtr<tcu::TestCaseGroup> group (new tcu::TestCaseGroup(testCtx, "opconstantcomposite", "Tests the OpConstantComposite instruction"));
4399 vector<CaseParameter> cases;
4400 de::Random rnd (deStringHash(group->getName()));
4401 const int numElements = 100;
4402 vector<float> positiveFloats (numElements, 0);
4403 vector<float> negativeFloats (numElements, 0);
4404 const StringTemplate shaderTemplate (
4405 string(getComputeAsmShaderPreamble()) +
4407 "OpSource GLSL 430\n"
4408 "OpName %main \"main\"\n"
4409 "OpName %id \"gl_GlobalInvocationID\"\n"
4411 "OpDecorate %id BuiltIn GlobalInvocationId\n"
4413 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) + string(getComputeAsmInputOutputBuffer()) +
4415 "%id = OpVariable %uvec3ptr Input\n"
4416 "%zero = OpConstant %i32 0\n"
4420 "%main = OpFunction %void None %voidf\n"
4421 "%label = OpLabel\n"
4422 "%idval = OpLoad %uvec3 %id\n"
4423 "%x = OpCompositeExtract %u32 %idval 0\n"
4424 "%inloc = OpAccessChain %f32ptr %indata %zero %x\n"
4425 "%inval = OpLoad %f32 %inloc\n"
4426 "%neg = OpFNegate %f32 %inval\n"
4427 "%outloc = OpAccessChain %f32ptr %outdata %zero %x\n"
4428 " OpStore %outloc %neg\n"
4430 " OpFunctionEnd\n");
4432 cases.push_back(CaseParameter("vector", "%five = OpConstant %u32 5\n"
4433 "%const = OpConstantComposite %uvec3 %five %zero %five"));
4434 cases.push_back(CaseParameter("matrix", "%m3fvec3 = OpTypeMatrix %fvec3 3\n"
4435 "%ten = OpConstant %f32 10.\n"
4436 "%fzero = OpConstant %f32 0.\n"
4437 "%vec = OpConstantComposite %fvec3 %ten %fzero %ten\n"
4438 "%mat = OpConstantComposite %m3fvec3 %vec %vec %vec"));
4439 cases.push_back(CaseParameter("struct", "%m2vec3 = OpTypeMatrix %fvec3 2\n"
4440 "%struct = OpTypeStruct %i32 %f32 %fvec3 %m2vec3\n"
4441 "%fzero = OpConstant %f32 0.\n"
4442 "%one = OpConstant %f32 1.\n"
4443 "%point5 = OpConstant %f32 0.5\n"
4444 "%vec = OpConstantComposite %fvec3 %one %one %fzero\n"
4445 "%mat = OpConstantComposite %m2vec3 %vec %vec\n"
4446 "%const = OpConstantComposite %struct %zero %point5 %vec %mat"));
4447 cases.push_back(CaseParameter("nested_struct", "%st1 = OpTypeStruct %u32 %f32\n"
4448 "%st2 = OpTypeStruct %i32 %i32\n"
4449 "%struct = OpTypeStruct %st1 %st2\n"
4450 "%point5 = OpConstant %f32 0.5\n"
4451 "%one = OpConstant %u32 1\n"
4452 "%ten = OpConstant %i32 10\n"
4453 "%st1val = OpConstantComposite %st1 %one %point5\n"
4454 "%st2val = OpConstantComposite %st2 %ten %ten\n"
4455 "%const = OpConstantComposite %struct %st1val %st2val"));
4457 fillRandomScalars(rnd, 1.f, 100.f, &positiveFloats[0], numElements);
4459 for (size_t ndx = 0; ndx < numElements; ++ndx)
4460 negativeFloats[ndx] = -positiveFloats[ndx];
4462 for (size_t caseNdx = 0; caseNdx < cases.size(); ++caseNdx)
4464 map<string, string> specializations;
4465 ComputeShaderSpec spec;
4467 specializations["CONSTANT"] = cases[caseNdx].param;
4468 spec.assembly = shaderTemplate.specialize(specializations);
4469 spec.inputs.push_back(BufferSp(new Float32Buffer(positiveFloats)));
4470 spec.outputs.push_back(BufferSp(new Float32Buffer(negativeFloats)));
4471 spec.numWorkGroups = IVec3(numElements, 1, 1);
4473 group->addChild(new SpvAsmComputeShaderCase(testCtx, cases[caseNdx].name, cases[caseNdx].name, spec));
4476 return group.release();
4479 // Creates a floating point number with the given exponent, and significand
4480 // bits set. It can only create normalized numbers. Only the least significant
4481 // 24 bits of the significand will be examined. The final bit of the
4482 // significand will also be ignored. This allows alignment to be written
4483 // similarly to C99 hex-floats.
4484 // For example if you wanted to write 0x1.7f34p-12 you would call
4485 // constructNormalizedFloat(-12, 0x7f3400)
4486 float constructNormalizedFloat (deInt32 exponent, deUint32 significand)
4490 for (deInt32 idx = 0; idx < 23; ++idx)
4492 f += ((significand & 0x800000) == 0) ? 0.f : std::ldexp(1.0f, -(idx + 1));
4496 return std::ldexp(f, exponent);
4499 // Compare instruction for the OpQuantizeF16 compute exact case.
4500 // Returns true if the output is what is expected from the test case.
4501 bool compareOpQuantizeF16ComputeExactCase (const std::vector<Resource>&, const vector<AllocationSp>& outputAllocs, const std::vector<Resource>& expectedOutputs, TestLog&)
4503 if (outputAllocs.size() != 1)
4506 // Only size is needed because we cannot compare Nans.
4507 size_t byteSize = expectedOutputs[0].getByteSize();
4509 const float* outputAsFloat = static_cast<const float*>(outputAllocs[0]->getHostPtr());
4511 if (byteSize != 4*sizeof(float)) {
4515 if (*outputAsFloat != constructNormalizedFloat(8, 0x304000) &&
4516 *outputAsFloat != constructNormalizedFloat(8, 0x300000)) {
4521 if (*outputAsFloat != -constructNormalizedFloat(-7, 0x600000) &&
4522 *outputAsFloat != -constructNormalizedFloat(-7, 0x604000)) {
4527 if (*outputAsFloat != constructNormalizedFloat(2, 0x01C000) &&
4528 *outputAsFloat != constructNormalizedFloat(2, 0x020000)) {
4533 if (*outputAsFloat != constructNormalizedFloat(1, 0xFFC000) &&
4534 *outputAsFloat != constructNormalizedFloat(2, 0x000000)) {
4541 // Checks that every output from a test-case is a float NaN.
4542 bool compareNan (const std::vector<Resource>&, const vector<AllocationSp>& outputAllocs, const std::vector<Resource>& expectedOutputs, TestLog&)
4544 if (outputAllocs.size() != 1)
4547 // Only size is needed because we cannot compare Nans.
4548 size_t byteSize = expectedOutputs[0].getByteSize();
4550 const float* const output_as_float = static_cast<const float*>(outputAllocs[0]->getHostPtr());
4552 for (size_t idx = 0; idx < byteSize / sizeof(float); ++idx)
4554 if (!deFloatIsNaN(output_as_float[idx]))
4563 // Checks that a compute shader can generate a constant composite value of various types, without exercising a computation on it.
4564 tcu::TestCaseGroup* createOpQuantizeToF16Group (tcu::TestContext& testCtx)
4566 de::MovePtr<tcu::TestCaseGroup> group (new tcu::TestCaseGroup(testCtx, "opquantize", "Tests the OpQuantizeToF16 instruction"));
4568 const std::string shader (
4569 string(getComputeAsmShaderPreamble()) +
4571 "OpSource GLSL 430\n"
4572 "OpName %main \"main\"\n"
4573 "OpName %id \"gl_GlobalInvocationID\"\n"
4575 "OpDecorate %id BuiltIn GlobalInvocationId\n"
4577 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) + string(getComputeAsmInputOutputBuffer()) +
4579 "%id = OpVariable %uvec3ptr Input\n"
4580 "%zero = OpConstant %i32 0\n"
4582 "%main = OpFunction %void None %voidf\n"
4583 "%label = OpLabel\n"
4584 "%idval = OpLoad %uvec3 %id\n"
4585 "%x = OpCompositeExtract %u32 %idval 0\n"
4586 "%inloc = OpAccessChain %f32ptr %indata %zero %x\n"
4587 "%inval = OpLoad %f32 %inloc\n"
4588 "%quant = OpQuantizeToF16 %f32 %inval\n"
4589 "%outloc = OpAccessChain %f32ptr %outdata %zero %x\n"
4590 " OpStore %outloc %quant\n"
4592 " OpFunctionEnd\n");
4595 ComputeShaderSpec spec;
4596 const deUint32 numElements = 100;
4597 vector<float> infinities;
4598 vector<float> results;
4600 infinities.reserve(numElements);
4601 results.reserve(numElements);
4603 for (size_t idx = 0; idx < numElements; ++idx)
4608 infinities.push_back(std::numeric_limits<float>::infinity());
4609 results.push_back(std::numeric_limits<float>::infinity());
4612 infinities.push_back(-std::numeric_limits<float>::infinity());
4613 results.push_back(-std::numeric_limits<float>::infinity());
4616 infinities.push_back(std::ldexp(1.0f, 16));
4617 results.push_back(std::numeric_limits<float>::infinity());
4620 infinities.push_back(std::ldexp(-1.0f, 32));
4621 results.push_back(-std::numeric_limits<float>::infinity());
4626 spec.assembly = shader;
4627 spec.inputs.push_back(BufferSp(new Float32Buffer(infinities)));
4628 spec.outputs.push_back(BufferSp(new Float32Buffer(results)));
4629 spec.numWorkGroups = IVec3(numElements, 1, 1);
4631 group->addChild(new SpvAsmComputeShaderCase(
4632 testCtx, "infinities", "Check that infinities propagated and created", spec));
4636 ComputeShaderSpec spec;
4638 const deUint32 numElements = 100;
4640 nans.reserve(numElements);
4642 for (size_t idx = 0; idx < numElements; ++idx)
4646 nans.push_back(std::numeric_limits<float>::quiet_NaN());
4650 nans.push_back(-std::numeric_limits<float>::quiet_NaN());
4654 spec.assembly = shader;
4655 spec.inputs.push_back(BufferSp(new Float32Buffer(nans)));
4656 spec.outputs.push_back(BufferSp(new Float32Buffer(nans)));
4657 spec.numWorkGroups = IVec3(numElements, 1, 1);
4658 spec.verifyIO = &compareNan;
4660 group->addChild(new SpvAsmComputeShaderCase(
4661 testCtx, "propagated_nans", "Check that nans are propagated", spec));
4665 ComputeShaderSpec spec;
4666 vector<float> small;
4667 vector<float> zeros;
4668 const deUint32 numElements = 100;
4670 small.reserve(numElements);
4671 zeros.reserve(numElements);
4673 for (size_t idx = 0; idx < numElements; ++idx)
4678 small.push_back(0.f);
4679 zeros.push_back(0.f);
4682 small.push_back(-0.f);
4683 zeros.push_back(-0.f);
4686 small.push_back(std::ldexp(1.0f, -16));
4687 zeros.push_back(0.f);
4690 small.push_back(std::ldexp(-1.0f, -32));
4691 zeros.push_back(-0.f);
4694 small.push_back(std::ldexp(1.0f, -127));
4695 zeros.push_back(0.f);
4698 small.push_back(-std::ldexp(1.0f, -128));
4699 zeros.push_back(-0.f);
4704 spec.assembly = shader;
4705 spec.inputs.push_back(BufferSp(new Float32Buffer(small)));
4706 spec.outputs.push_back(BufferSp(new Float32Buffer(zeros)));
4707 spec.numWorkGroups = IVec3(numElements, 1, 1);
4709 group->addChild(new SpvAsmComputeShaderCase(
4710 testCtx, "flush_to_zero", "Check that values are zeroed correctly", spec));
4714 ComputeShaderSpec spec;
4715 vector<float> exact;
4716 const deUint32 numElements = 200;
4718 exact.reserve(numElements);
4720 for (size_t idx = 0; idx < numElements; ++idx)
4721 exact.push_back(static_cast<float>(static_cast<int>(idx) - 100));
4723 spec.assembly = shader;
4724 spec.inputs.push_back(BufferSp(new Float32Buffer(exact)));
4725 spec.outputs.push_back(BufferSp(new Float32Buffer(exact)));
4726 spec.numWorkGroups = IVec3(numElements, 1, 1);
4728 group->addChild(new SpvAsmComputeShaderCase(
4729 testCtx, "exact", "Check that values exactly preserved where appropriate", spec));
4733 ComputeShaderSpec spec;
4734 vector<float> inputs;
4735 const deUint32 numElements = 4;
4737 inputs.push_back(constructNormalizedFloat(8, 0x300300));
4738 inputs.push_back(-constructNormalizedFloat(-7, 0x600800));
4739 inputs.push_back(constructNormalizedFloat(2, 0x01E000));
4740 inputs.push_back(constructNormalizedFloat(1, 0xFFE000));
4742 spec.assembly = shader;
4743 spec.verifyIO = &compareOpQuantizeF16ComputeExactCase;
4744 spec.inputs.push_back(BufferSp(new Float32Buffer(inputs)));
4745 spec.outputs.push_back(BufferSp(new Float32Buffer(inputs)));
4746 spec.numWorkGroups = IVec3(numElements, 1, 1);
4748 group->addChild(new SpvAsmComputeShaderCase(
4749 testCtx, "rounded", "Check that are rounded when needed", spec));
4752 return group.release();
4755 tcu::TestCaseGroup* createSpecConstantOpQuantizeToF16Group (tcu::TestContext& testCtx)
4757 de::MovePtr<tcu::TestCaseGroup> group (new tcu::TestCaseGroup(testCtx, "opspecconstantop_opquantize", "Tests the OpQuantizeToF16 opcode for the OpSpecConstantOp instruction"));
4759 const std::string shader (
4760 string(getComputeAsmShaderPreamble()) +
4762 "OpName %main \"main\"\n"
4763 "OpName %id \"gl_GlobalInvocationID\"\n"
4765 "OpDecorate %id BuiltIn GlobalInvocationId\n"
4767 "OpDecorate %sc_0 SpecId 0\n"
4768 "OpDecorate %sc_1 SpecId 1\n"
4769 "OpDecorate %sc_2 SpecId 2\n"
4770 "OpDecorate %sc_3 SpecId 3\n"
4771 "OpDecorate %sc_4 SpecId 4\n"
4772 "OpDecorate %sc_5 SpecId 5\n"
4774 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) + string(getComputeAsmInputOutputBuffer()) +
4776 "%id = OpVariable %uvec3ptr Input\n"
4777 "%zero = OpConstant %i32 0\n"
4778 "%c_u32_6 = OpConstant %u32 6\n"
4780 "%sc_0 = OpSpecConstant %f32 0.\n"
4781 "%sc_1 = OpSpecConstant %f32 0.\n"
4782 "%sc_2 = OpSpecConstant %f32 0.\n"
4783 "%sc_3 = OpSpecConstant %f32 0.\n"
4784 "%sc_4 = OpSpecConstant %f32 0.\n"
4785 "%sc_5 = OpSpecConstant %f32 0.\n"
4787 "%sc_0_quant = OpSpecConstantOp %f32 QuantizeToF16 %sc_0\n"
4788 "%sc_1_quant = OpSpecConstantOp %f32 QuantizeToF16 %sc_1\n"
4789 "%sc_2_quant = OpSpecConstantOp %f32 QuantizeToF16 %sc_2\n"
4790 "%sc_3_quant = OpSpecConstantOp %f32 QuantizeToF16 %sc_3\n"
4791 "%sc_4_quant = OpSpecConstantOp %f32 QuantizeToF16 %sc_4\n"
4792 "%sc_5_quant = OpSpecConstantOp %f32 QuantizeToF16 %sc_5\n"
4794 "%main = OpFunction %void None %voidf\n"
4795 "%label = OpLabel\n"
4796 "%idval = OpLoad %uvec3 %id\n"
4797 "%x = OpCompositeExtract %u32 %idval 0\n"
4798 "%outloc = OpAccessChain %f32ptr %outdata %zero %x\n"
4799 "%selector = OpUMod %u32 %x %c_u32_6\n"
4800 " OpSelectionMerge %exit None\n"
4801 " OpSwitch %selector %exit 0 %case0 1 %case1 2 %case2 3 %case3 4 %case4 5 %case5\n"
4803 "%case0 = OpLabel\n"
4804 " OpStore %outloc %sc_0_quant\n"
4807 "%case1 = OpLabel\n"
4808 " OpStore %outloc %sc_1_quant\n"
4811 "%case2 = OpLabel\n"
4812 " OpStore %outloc %sc_2_quant\n"
4815 "%case3 = OpLabel\n"
4816 " OpStore %outloc %sc_3_quant\n"
4819 "%case4 = OpLabel\n"
4820 " OpStore %outloc %sc_4_quant\n"
4823 "%case5 = OpLabel\n"
4824 " OpStore %outloc %sc_5_quant\n"
4830 " OpFunctionEnd\n");
4833 ComputeShaderSpec spec;
4834 const deUint8 numCases = 4;
4835 vector<float> inputs (numCases, 0.f);
4836 vector<float> outputs;
4838 spec.assembly = shader;
4839 spec.numWorkGroups = IVec3(numCases, 1, 1);
4841 spec.specConstants.append<deInt32>(bitwiseCast<deUint32>(std::numeric_limits<float>::infinity()));
4842 spec.specConstants.append<deInt32>(bitwiseCast<deUint32>(-std::numeric_limits<float>::infinity()));
4843 spec.specConstants.append<deInt32>(bitwiseCast<deUint32>(std::ldexp(1.0f, 16)));
4844 spec.specConstants.append<deInt32>(bitwiseCast<deUint32>(std::ldexp(-1.0f, 32)));
4846 outputs.push_back(std::numeric_limits<float>::infinity());
4847 outputs.push_back(-std::numeric_limits<float>::infinity());
4848 outputs.push_back(std::numeric_limits<float>::infinity());
4849 outputs.push_back(-std::numeric_limits<float>::infinity());
4851 spec.inputs.push_back(BufferSp(new Float32Buffer(inputs)));
4852 spec.outputs.push_back(BufferSp(new Float32Buffer(outputs)));
4854 group->addChild(new SpvAsmComputeShaderCase(
4855 testCtx, "infinities", "Check that infinities propagated and created", spec));
4859 ComputeShaderSpec spec;
4860 const deUint8 numCases = 2;
4861 vector<float> inputs (numCases, 0.f);
4862 vector<float> outputs;
4864 spec.assembly = shader;
4865 spec.numWorkGroups = IVec3(numCases, 1, 1);
4866 spec.verifyIO = &compareNan;
4868 outputs.push_back(std::numeric_limits<float>::quiet_NaN());
4869 outputs.push_back(-std::numeric_limits<float>::quiet_NaN());
4871 for (deUint8 idx = 0; idx < numCases; ++idx)
4872 spec.specConstants.append<deInt32>(bitwiseCast<deUint32>(outputs[idx]));
4874 spec.inputs.push_back(BufferSp(new Float32Buffer(inputs)));
4875 spec.outputs.push_back(BufferSp(new Float32Buffer(outputs)));
4877 group->addChild(new SpvAsmComputeShaderCase(
4878 testCtx, "propagated_nans", "Check that nans are propagated", spec));
4882 ComputeShaderSpec spec;
4883 const deUint8 numCases = 6;
4884 vector<float> inputs (numCases, 0.f);
4885 vector<float> outputs;
4887 spec.assembly = shader;
4888 spec.numWorkGroups = IVec3(numCases, 1, 1);
4890 spec.specConstants.append<deInt32>(bitwiseCast<deUint32>(0.f));
4891 spec.specConstants.append<deInt32>(bitwiseCast<deUint32>(-0.f));
4892 spec.specConstants.append<deInt32>(bitwiseCast<deUint32>(std::ldexp(1.0f, -16)));
4893 spec.specConstants.append<deInt32>(bitwiseCast<deUint32>(std::ldexp(-1.0f, -32)));
4894 spec.specConstants.append<deInt32>(bitwiseCast<deUint32>(std::ldexp(1.0f, -127)));
4895 spec.specConstants.append<deInt32>(bitwiseCast<deUint32>(-std::ldexp(1.0f, -128)));
4897 outputs.push_back(0.f);
4898 outputs.push_back(-0.f);
4899 outputs.push_back(0.f);
4900 outputs.push_back(-0.f);
4901 outputs.push_back(0.f);
4902 outputs.push_back(-0.f);
4904 spec.inputs.push_back(BufferSp(new Float32Buffer(inputs)));
4905 spec.outputs.push_back(BufferSp(new Float32Buffer(outputs)));
4907 group->addChild(new SpvAsmComputeShaderCase(
4908 testCtx, "flush_to_zero", "Check that values are zeroed correctly", spec));
4912 ComputeShaderSpec spec;
4913 const deUint8 numCases = 6;
4914 vector<float> inputs (numCases, 0.f);
4915 vector<float> outputs;
4917 spec.assembly = shader;
4918 spec.numWorkGroups = IVec3(numCases, 1, 1);
4920 for (deUint8 idx = 0; idx < 6; ++idx)
4922 const float f = static_cast<float>(idx * 10 - 30) / 4.f;
4923 spec.specConstants.append<deInt32>(bitwiseCast<deUint32>(f));
4924 outputs.push_back(f);
4927 spec.inputs.push_back(BufferSp(new Float32Buffer(inputs)));
4928 spec.outputs.push_back(BufferSp(new Float32Buffer(outputs)));
4930 group->addChild(new SpvAsmComputeShaderCase(
4931 testCtx, "exact", "Check that values exactly preserved where appropriate", spec));
4935 ComputeShaderSpec spec;
4936 const deUint8 numCases = 4;
4937 vector<float> inputs (numCases, 0.f);
4938 vector<float> outputs;
4940 spec.assembly = shader;
4941 spec.numWorkGroups = IVec3(numCases, 1, 1);
4942 spec.verifyIO = &compareOpQuantizeF16ComputeExactCase;
4944 outputs.push_back(constructNormalizedFloat(8, 0x300300));
4945 outputs.push_back(-constructNormalizedFloat(-7, 0x600800));
4946 outputs.push_back(constructNormalizedFloat(2, 0x01E000));
4947 outputs.push_back(constructNormalizedFloat(1, 0xFFE000));
4949 for (deUint8 idx = 0; idx < numCases; ++idx)
4950 spec.specConstants.append<deInt32>(bitwiseCast<deUint32>(outputs[idx]));
4952 spec.inputs.push_back(BufferSp(new Float32Buffer(inputs)));
4953 spec.outputs.push_back(BufferSp(new Float32Buffer(outputs)));
4955 group->addChild(new SpvAsmComputeShaderCase(
4956 testCtx, "rounded", "Check that are rounded when needed", spec));
4959 return group.release();
4962 // Checks that constant null/composite values can be used in computation.
4963 tcu::TestCaseGroup* createOpConstantUsageGroup (tcu::TestContext& testCtx)
4965 de::MovePtr<tcu::TestCaseGroup> group (new tcu::TestCaseGroup(testCtx, "opconstantnullcomposite", "Spotcheck the OpConstantNull & OpConstantComposite instruction"));
4966 ComputeShaderSpec spec;
4967 de::Random rnd (deStringHash(group->getName()));
4968 const int numElements = 100;
4969 vector<float> positiveFloats (numElements, 0);
4970 vector<float> negativeFloats (numElements, 0);
4972 fillRandomScalars(rnd, 1.f, 100.f, &positiveFloats[0], numElements);
4974 for (size_t ndx = 0; ndx < numElements; ++ndx)
4975 negativeFloats[ndx] = -positiveFloats[ndx];
4978 "OpCapability Shader\n"
4979 "%std450 = OpExtInstImport \"GLSL.std.450\"\n"
4980 "OpMemoryModel Logical GLSL450\n"
4981 "OpEntryPoint GLCompute %main \"main\" %id\n"
4982 "OpExecutionMode %main LocalSize 1 1 1\n"
4984 "OpSource GLSL 430\n"
4985 "OpName %main \"main\"\n"
4986 "OpName %id \"gl_GlobalInvocationID\"\n"
4988 "OpDecorate %id BuiltIn GlobalInvocationId\n"
4990 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) +
4992 "%fmat = OpTypeMatrix %fvec3 3\n"
4993 "%ten = OpConstant %u32 10\n"
4994 "%f32arr10 = OpTypeArray %f32 %ten\n"
4995 "%fst = OpTypeStruct %f32 %f32\n"
4997 + string(getComputeAsmInputOutputBuffer()) +
4999 "%id = OpVariable %uvec3ptr Input\n"
5000 "%zero = OpConstant %i32 0\n"
5002 // Create a bunch of null values
5003 "%unull = OpConstantNull %u32\n"
5004 "%fnull = OpConstantNull %f32\n"
5005 "%vnull = OpConstantNull %fvec3\n"
5006 "%mnull = OpConstantNull %fmat\n"
5007 "%anull = OpConstantNull %f32arr10\n"
5008 "%snull = OpConstantComposite %fst %fnull %fnull\n"
5010 "%main = OpFunction %void None %voidf\n"
5011 "%label = OpLabel\n"
5012 "%idval = OpLoad %uvec3 %id\n"
5013 "%x = OpCompositeExtract %u32 %idval 0\n"
5014 "%inloc = OpAccessChain %f32ptr %indata %zero %x\n"
5015 "%inval = OpLoad %f32 %inloc\n"
5016 "%neg = OpFNegate %f32 %inval\n"
5018 // Get the abs() of (a certain element of) those null values
5019 "%unull_cov = OpConvertUToF %f32 %unull\n"
5020 "%unull_abs = OpExtInst %f32 %std450 FAbs %unull_cov\n"
5021 "%fnull_abs = OpExtInst %f32 %std450 FAbs %fnull\n"
5022 "%vnull_0 = OpCompositeExtract %f32 %vnull 0\n"
5023 "%vnull_abs = OpExtInst %f32 %std450 FAbs %vnull_0\n"
5024 "%mnull_12 = OpCompositeExtract %f32 %mnull 1 2\n"
5025 "%mnull_abs = OpExtInst %f32 %std450 FAbs %mnull_12\n"
5026 "%anull_3 = OpCompositeExtract %f32 %anull 3\n"
5027 "%anull_abs = OpExtInst %f32 %std450 FAbs %anull_3\n"
5028 "%snull_1 = OpCompositeExtract %f32 %snull 1\n"
5029 "%snull_abs = OpExtInst %f32 %std450 FAbs %snull_1\n"
5032 "%add1 = OpFAdd %f32 %neg %unull_abs\n"
5033 "%add2 = OpFAdd %f32 %add1 %fnull_abs\n"
5034 "%add3 = OpFAdd %f32 %add2 %vnull_abs\n"
5035 "%add4 = OpFAdd %f32 %add3 %mnull_abs\n"
5036 "%add5 = OpFAdd %f32 %add4 %anull_abs\n"
5037 "%final = OpFAdd %f32 %add5 %snull_abs\n"
5039 "%outloc = OpAccessChain %f32ptr %outdata %zero %x\n"
5040 " OpStore %outloc %final\n" // write to output
5043 spec.inputs.push_back(BufferSp(new Float32Buffer(positiveFloats)));
5044 spec.outputs.push_back(BufferSp(new Float32Buffer(negativeFloats)));
5045 spec.numWorkGroups = IVec3(numElements, 1, 1);
5047 group->addChild(new SpvAsmComputeShaderCase(testCtx, "spotcheck", "Check that values constructed via OpConstantNull & OpConstantComposite can be used", spec));
5049 return group.release();
5052 // Assembly code used for testing loop control is based on GLSL source code:
5055 // layout(std140, set = 0, binding = 0) readonly buffer Input {
5056 // float elements[];
5058 // layout(std140, set = 0, binding = 1) writeonly buffer Output {
5059 // float elements[];
5063 // uint x = gl_GlobalInvocationID.x;
5064 // output_data.elements[x] = input_data.elements[x];
5065 // for (uint i = 0; i < 4; ++i)
5066 // output_data.elements[x] += 1.f;
5068 tcu::TestCaseGroup* createLoopControlGroup (tcu::TestContext& testCtx)
5070 de::MovePtr<tcu::TestCaseGroup> group (new tcu::TestCaseGroup(testCtx, "loop_control", "Tests loop control cases"));
5071 vector<CaseParameter> cases;
5072 de::Random rnd (deStringHash(group->getName()));
5073 const int numElements = 100;
5074 vector<float> inputFloats (numElements, 0);
5075 vector<float> outputFloats (numElements, 0);
5076 const StringTemplate shaderTemplate (
5077 string(getComputeAsmShaderPreamble()) +
5079 "OpSource GLSL 430\n"
5080 "OpName %main \"main\"\n"
5081 "OpName %id \"gl_GlobalInvocationID\"\n"
5083 "OpDecorate %id BuiltIn GlobalInvocationId\n"
5085 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) + string(getComputeAsmInputOutputBuffer()) +
5087 "%u32ptr = OpTypePointer Function %u32\n"
5089 "%id = OpVariable %uvec3ptr Input\n"
5090 "%zero = OpConstant %i32 0\n"
5091 "%uzero = OpConstant %u32 0\n"
5092 "%one = OpConstant %i32 1\n"
5093 "%constf1 = OpConstant %f32 1.0\n"
5094 "%four = OpConstant %u32 4\n"
5096 "%main = OpFunction %void None %voidf\n"
5097 "%entry = OpLabel\n"
5098 "%i = OpVariable %u32ptr Function\n"
5099 " OpStore %i %uzero\n"
5101 "%idval = OpLoad %uvec3 %id\n"
5102 "%x = OpCompositeExtract %u32 %idval 0\n"
5103 "%inloc = OpAccessChain %f32ptr %indata %zero %x\n"
5104 "%inval = OpLoad %f32 %inloc\n"
5105 "%outloc = OpAccessChain %f32ptr %outdata %zero %x\n"
5106 " OpStore %outloc %inval\n"
5107 " OpBranch %loop_entry\n"
5109 "%loop_entry = OpLabel\n"
5110 "%i_val = OpLoad %u32 %i\n"
5111 "%cmp_lt = OpULessThan %bool %i_val %four\n"
5112 " OpLoopMerge %loop_merge %loop_body ${CONTROL}\n"
5113 " OpBranchConditional %cmp_lt %loop_body %loop_merge\n"
5114 "%loop_body = OpLabel\n"
5115 "%outval = OpLoad %f32 %outloc\n"
5116 "%addf1 = OpFAdd %f32 %outval %constf1\n"
5117 " OpStore %outloc %addf1\n"
5118 "%new_i = OpIAdd %u32 %i_val %one\n"
5119 " OpStore %i %new_i\n"
5120 " OpBranch %loop_entry\n"
5121 "%loop_merge = OpLabel\n"
5123 " OpFunctionEnd\n");
5125 cases.push_back(CaseParameter("none", "None"));
5126 cases.push_back(CaseParameter("unroll", "Unroll"));
5127 cases.push_back(CaseParameter("dont_unroll", "DontUnroll"));
5128 cases.push_back(CaseParameter("unroll_dont_unroll", "Unroll|DontUnroll"));
5130 fillRandomScalars(rnd, -100.f, 100.f, &inputFloats[0], numElements);
5132 for (size_t ndx = 0; ndx < numElements; ++ndx)
5133 outputFloats[ndx] = inputFloats[ndx] + 4.f;
5135 for (size_t caseNdx = 0; caseNdx < cases.size(); ++caseNdx)
5137 map<string, string> specializations;
5138 ComputeShaderSpec spec;
5140 specializations["CONTROL"] = cases[caseNdx].param;
5141 spec.assembly = shaderTemplate.specialize(specializations);
5142 spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
5143 spec.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
5144 spec.numWorkGroups = IVec3(numElements, 1, 1);
5146 group->addChild(new SpvAsmComputeShaderCase(testCtx, cases[caseNdx].name, cases[caseNdx].name, spec));
5149 group->addChild(new SpvAsmLoopControlDependencyLengthCase(testCtx, "dependency_length", "dependency_length"));
5150 group->addChild(new SpvAsmLoopControlDependencyInfiniteCase(testCtx, "dependency_infinite", "dependency_infinite"));
5152 return group.release();
5155 // Assembly code used for testing selection control is based on GLSL source code:
5158 // layout(std140, set = 0, binding = 0) readonly buffer Input {
5159 // float elements[];
5161 // layout(std140, set = 0, binding = 1) writeonly buffer Output {
5162 // float elements[];
5166 // uint x = gl_GlobalInvocationID.x;
5167 // float val = input_data.elements[x];
5169 // output_data.elements[x] = val + 1.f;
5171 // output_data.elements[x] = val - 1.f;
5173 tcu::TestCaseGroup* createSelectionControlGroup (tcu::TestContext& testCtx)
5175 de::MovePtr<tcu::TestCaseGroup> group (new tcu::TestCaseGroup(testCtx, "selection_control", "Tests selection control cases"));
5176 vector<CaseParameter> cases;
5177 de::Random rnd (deStringHash(group->getName()));
5178 const int numElements = 100;
5179 vector<float> inputFloats (numElements, 0);
5180 vector<float> outputFloats (numElements, 0);
5181 const StringTemplate shaderTemplate (
5182 string(getComputeAsmShaderPreamble()) +
5184 "OpSource GLSL 430\n"
5185 "OpName %main \"main\"\n"
5186 "OpName %id \"gl_GlobalInvocationID\"\n"
5188 "OpDecorate %id BuiltIn GlobalInvocationId\n"
5190 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) + string(getComputeAsmInputOutputBuffer()) +
5192 "%id = OpVariable %uvec3ptr Input\n"
5193 "%zero = OpConstant %i32 0\n"
5194 "%constf1 = OpConstant %f32 1.0\n"
5195 "%constf10 = OpConstant %f32 10.0\n"
5197 "%main = OpFunction %void None %voidf\n"
5198 "%entry = OpLabel\n"
5199 "%idval = OpLoad %uvec3 %id\n"
5200 "%x = OpCompositeExtract %u32 %idval 0\n"
5201 "%inloc = OpAccessChain %f32ptr %indata %zero %x\n"
5202 "%inval = OpLoad %f32 %inloc\n"
5203 "%outloc = OpAccessChain %f32ptr %outdata %zero %x\n"
5204 "%cmp_gt = OpFOrdGreaterThan %bool %inval %constf10\n"
5206 " OpSelectionMerge %if_end ${CONTROL}\n"
5207 " OpBranchConditional %cmp_gt %if_true %if_false\n"
5208 "%if_true = OpLabel\n"
5209 "%addf1 = OpFAdd %f32 %inval %constf1\n"
5210 " OpStore %outloc %addf1\n"
5211 " OpBranch %if_end\n"
5212 "%if_false = OpLabel\n"
5213 "%subf1 = OpFSub %f32 %inval %constf1\n"
5214 " OpStore %outloc %subf1\n"
5215 " OpBranch %if_end\n"
5216 "%if_end = OpLabel\n"
5218 " OpFunctionEnd\n");
5220 cases.push_back(CaseParameter("none", "None"));
5221 cases.push_back(CaseParameter("flatten", "Flatten"));
5222 cases.push_back(CaseParameter("dont_flatten", "DontFlatten"));
5223 cases.push_back(CaseParameter("flatten_dont_flatten", "DontFlatten|Flatten"));
5225 fillRandomScalars(rnd, -100.f, 100.f, &inputFloats[0], numElements);
5227 // CPU might not use the same rounding mode as the GPU. Use whole numbers to avoid rounding differences.
5228 floorAll(inputFloats);
5230 for (size_t ndx = 0; ndx < numElements; ++ndx)
5231 outputFloats[ndx] = inputFloats[ndx] + (inputFloats[ndx] > 10.f ? 1.f : -1.f);
5233 for (size_t caseNdx = 0; caseNdx < cases.size(); ++caseNdx)
5235 map<string, string> specializations;
5236 ComputeShaderSpec spec;
5238 specializations["CONTROL"] = cases[caseNdx].param;
5239 spec.assembly = shaderTemplate.specialize(specializations);
5240 spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
5241 spec.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
5242 spec.numWorkGroups = IVec3(numElements, 1, 1);
5244 group->addChild(new SpvAsmComputeShaderCase(testCtx, cases[caseNdx].name, cases[caseNdx].name, spec));
5247 return group.release();
5250 void getOpNameAbuseCases (vector<CaseParameter> &abuseCases)
5252 // Generate a long name.
5253 std::string longname;
5254 longname.resize(65535, 'k'); // max string literal, spir-v 2.17
5256 // Some bad names, abusing utf-8 encoding. This may also cause problems
5258 // 1. Various illegal code points in utf-8
5259 std::string utf8illegal =
5260 "Illegal bytes in UTF-8: "
5261 "\xc0 \xc1 \xf5 \xf6 \xf7 \xf8 \xf9 \xfa \xfb \xfc \xfd \xfe \xff"
5262 "illegal surrogates: \xed\xad\xbf \xed\xbe\x80";
5264 // 2. Zero encoded as overlong, not exactly legal but often supported to differentiate from terminating zero
5265 std::string utf8nul = "UTF-8 encoded nul \xC0\x80 (should not end name)";
5267 // 3. Some overlong encodings
5268 std::string utf8overlong =
5269 "UTF-8 overlong \xF0\x82\x82\xAC \xfc\x83\xbf\xbf\xbf\xbf \xf8\x87\xbf\xbf\xbf "
5272 // 4. Internet "zalgo" meme "bleeding text"
5273 std::string utf8zalgo =
5274 "\x56\xcc\xb5\xcc\x85\xcc\x94\xcc\x88\xcd\x8a\xcc\x91\xcc\x88\xcd\x91\xcc\x83\xcd\x82"
5275 "\xcc\x83\xcd\x90\xcc\x8a\xcc\x92\xcc\x92\xcd\x8b\xcc\x94\xcd\x9d\xcc\x98\xcc\xab\xcc"
5276 "\xae\xcc\xa9\xcc\xad\xcc\x97\xcc\xb0\x75\xcc\xb6\xcc\xbe\xcc\x80\xcc\x82\xcc\x84\xcd"
5277 "\x84\xcc\x90\xcd\x86\xcc\x9a\xcd\x84\xcc\x9b\xcd\x86\xcd\x92\xcc\x9a\xcd\x99\xcd\x99"
5278 "\xcc\xbb\xcc\x98\xcd\x8e\xcd\x88\xcd\x9a\xcc\xa6\xcc\x9c\xcc\xab\xcc\x99\xcd\x94\xcd"
5279 "\x99\xcd\x95\xcc\xa5\xcc\xab\xcd\x89\x6c\xcc\xb8\xcc\x8e\xcc\x8b\xcc\x8b\xcc\x9a\xcc"
5280 "\x8e\xcd\x9d\xcc\x80\xcc\xa1\xcc\xad\xcd\x9c\xcc\xba\xcc\x96\xcc\xb3\xcc\xa2\xcd\x8e"
5281 "\xcc\xa2\xcd\x96\x6b\xcc\xb8\xcc\x84\xcd\x81\xcc\xbf\xcc\x8d\xcc\x89\xcc\x85\xcc\x92"
5282 "\xcc\x84\xcc\x90\xcd\x81\xcc\x93\xcd\x90\xcd\x92\xcd\x9d\xcc\x84\xcd\x98\xcd\x9d\xcd"
5283 "\xa0\xcd\x91\xcc\x94\xcc\xb9\xcd\x93\xcc\xa5\xcd\x87\xcc\xad\xcc\xa7\xcd\x96\xcd\x99"
5284 "\xcc\x9d\xcc\xbc\xcd\x96\xcd\x93\xcc\x9d\xcc\x99\xcc\xa8\xcc\xb1\xcd\x85\xcc\xba\xcc"
5285 "\xa7\x61\xcc\xb8\xcc\x8e\xcc\x81\xcd\x90\xcd\x84\xcd\x8c\xcc\x8c\xcc\x85\xcd\x86\xcc"
5286 "\x84\xcd\x84\xcc\x90\xcc\x84\xcc\x8d\xcd\x99\xcd\x8d\xcc\xb0\xcc\xa3\xcc\xa6\xcd\x89"
5287 "\xcd\x8d\xcd\x87\xcc\x98\xcd\x8d\xcc\xa4\xcd\x9a\xcd\x8e\xcc\xab\xcc\xb9\xcc\xac\xcc"
5288 "\xa2\xcd\x87\xcc\xa0\xcc\xb3\xcd\x89\xcc\xb9\xcc\xa7\xcc\xa6\xcd\x89\xcd\x95\x6e\xcc"
5289 "\xb8\xcd\x8a\xcc\x8a\xcd\x82\xcc\x9b\xcd\x81\xcd\x90\xcc\x85\xcc\x9b\xcd\x80\xcd\x91"
5290 "\xcd\x9b\xcc\x81\xcd\x81\xcc\x9a\xcc\xb3\xcd\x9c\xcc\x9e\xcc\x9d\xcd\x99\xcc\xa2\xcd"
5291 "\x93\xcd\x96\xcc\x97\xff";
5293 // General name abuses
5294 abuseCases.push_back(CaseParameter("_has_very_long_name", longname));
5295 abuseCases.push_back(CaseParameter("_utf8_illegal", utf8illegal));
5296 abuseCases.push_back(CaseParameter("_utf8_nul", utf8nul));
5297 abuseCases.push_back(CaseParameter("_utf8_overlong", utf8overlong));
5298 abuseCases.push_back(CaseParameter("_utf8_zalgo", utf8zalgo));
5301 abuseCases.push_back(CaseParameter("_is_gl_Position", "gl_Position"));
5302 abuseCases.push_back(CaseParameter("_is_gl_InstanceID", "gl_InstanceID"));
5303 abuseCases.push_back(CaseParameter("_is_gl_PrimitiveID", "gl_PrimitiveID"));
5304 abuseCases.push_back(CaseParameter("_is_gl_TessCoord", "gl_TessCoord"));
5305 abuseCases.push_back(CaseParameter("_is_gl_PerVertex", "gl_PerVertex"));
5306 abuseCases.push_back(CaseParameter("_is_gl_InvocationID", "gl_InvocationID"));
5307 abuseCases.push_back(CaseParameter("_is_gl_PointSize", "gl_PointSize"));
5308 abuseCases.push_back(CaseParameter("_is_gl_PointCoord", "gl_PointCoord"));
5309 abuseCases.push_back(CaseParameter("_is_gl_Layer", "gl_Layer"));
5310 abuseCases.push_back(CaseParameter("_is_gl_FragDepth", "gl_FragDepth"));
5311 abuseCases.push_back(CaseParameter("_is_gl_NumWorkGroups", "gl_NumWorkGroups"));
5312 abuseCases.push_back(CaseParameter("_is_gl_WorkGroupID", "gl_WorkGroupID"));
5313 abuseCases.push_back(CaseParameter("_is_gl_LocalInvocationID", "gl_LocalInvocationID"));
5314 abuseCases.push_back(CaseParameter("_is_gl_GlobalInvocationID", "gl_GlobalInvocationID"));
5315 abuseCases.push_back(CaseParameter("_is_gl_MaxVertexAttribs", "gl_MaxVertexAttribs"));
5316 abuseCases.push_back(CaseParameter("_is_gl_MaxViewports", "gl_MaxViewports"));
5317 abuseCases.push_back(CaseParameter("_is_gl_MaxComputeWorkGroupCount", "gl_MaxComputeWorkGroupCount"));
5318 abuseCases.push_back(CaseParameter("_is_mat3", "mat3"));
5319 abuseCases.push_back(CaseParameter("_is_volatile", "volatile"));
5320 abuseCases.push_back(CaseParameter("_is_inout", "inout"));
5321 abuseCases.push_back(CaseParameter("_is_isampler3d", "isampler3d"));
5324 tcu::TestCaseGroup* createOpNameGroup (tcu::TestContext& testCtx)
5326 de::MovePtr<tcu::TestCaseGroup> group (new tcu::TestCaseGroup(testCtx, "opname", "Tests OpName cases"));
5327 de::MovePtr<tcu::TestCaseGroup> entryMainGroup (new tcu::TestCaseGroup(testCtx, "entry_main", "OpName tests with entry main"));
5328 de::MovePtr<tcu::TestCaseGroup> entryNotGroup (new tcu::TestCaseGroup(testCtx, "entry_rdc", "OpName tests with entry rdc"));
5329 de::MovePtr<tcu::TestCaseGroup> abuseGroup (new tcu::TestCaseGroup(testCtx, "abuse", "OpName abuse tests"));
5330 vector<CaseParameter> cases;
5331 vector<CaseParameter> abuseCases;
5332 vector<string> testFunc;
5333 de::Random rnd (deStringHash(group->getName()));
5334 const int numElements = 128;
5335 vector<float> inputFloats (numElements, 0);
5336 vector<float> outputFloats (numElements, 0);
5338 getOpNameAbuseCases(abuseCases);
5340 fillRandomScalars(rnd, -100.0f, 100.0f, &inputFloats[0], numElements);
5342 for(size_t ndx = 0; ndx < numElements; ++ndx)
5343 outputFloats[ndx] = -inputFloats[ndx];
5345 const string commonShaderHeader =
5346 "OpCapability Shader\n"
5347 "OpMemoryModel Logical GLSL450\n"
5348 "OpEntryPoint GLCompute %main \"main\" %id\n"
5349 "OpExecutionMode %main LocalSize 1 1 1\n";
5351 const string commonShaderFooter =
5352 "OpDecorate %id BuiltIn GlobalInvocationId\n"
5354 + string(getComputeAsmInputOutputBufferTraits())
5355 + string(getComputeAsmCommonTypes())
5356 + string(getComputeAsmInputOutputBuffer()) +
5358 "%id = OpVariable %uvec3ptr Input\n"
5359 "%zero = OpConstant %i32 0\n"
5361 "%func = OpFunction %void None %voidf\n"
5366 "%main = OpFunction %void None %voidf\n"
5367 "%entry = OpLabel\n"
5368 "%7 = OpFunctionCall %void %func\n"
5370 "%idval = OpLoad %uvec3 %id\n"
5371 "%x = OpCompositeExtract %u32 %idval 0\n"
5373 "%inloc = OpAccessChain %f32ptr %indata %zero %x\n"
5374 "%inval = OpLoad %f32 %inloc\n"
5375 "%neg = OpFNegate %f32 %inval\n"
5376 "%outloc = OpAccessChain %f32ptr %outdata %zero %x\n"
5377 " OpStore %outloc %neg\n"
5382 const StringTemplate shaderTemplate (
5383 "OpCapability Shader\n"
5384 "OpMemoryModel Logical GLSL450\n"
5385 "OpEntryPoint GLCompute %main \"${ENTRY}\" %id\n"
5386 "OpExecutionMode %main LocalSize 1 1 1\n"
5387 "OpName %${ID} \"${NAME}\"\n" +
5388 commonShaderFooter);
5390 const std::string multipleNames =
5391 commonShaderHeader +
5392 "OpName %main \"to_be\"\n"
5393 "OpName %id \"or_not\"\n"
5394 "OpName %main \"to_be\"\n"
5395 "OpName %main \"makes_no\"\n"
5396 "OpName %func \"difference\"\n"
5397 "OpName %5 \"to_me\"\n" +
5401 ComputeShaderSpec spec;
5403 spec.assembly = multipleNames;
5404 spec.numWorkGroups = IVec3(numElements, 1, 1);
5405 spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
5406 spec.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
5408 abuseGroup->addChild(new SpvAsmComputeShaderCase(testCtx, "main_has_multiple_names", "multiple_names", spec));
5411 const std::string everythingNamed =
5412 commonShaderHeader +
5413 "OpName %main \"name1\"\n"
5414 "OpName %id \"name2\"\n"
5415 "OpName %zero \"name3\"\n"
5416 "OpName %entry \"name4\"\n"
5417 "OpName %func \"name5\"\n"
5418 "OpName %5 \"name6\"\n"
5419 "OpName %7 \"name7\"\n"
5420 "OpName %idval \"name8\"\n"
5421 "OpName %inloc \"name9\"\n"
5422 "OpName %inval \"name10\"\n"
5423 "OpName %neg \"name11\"\n"
5424 "OpName %outloc \"name12\"\n"+
5427 ComputeShaderSpec spec;
5429 spec.assembly = everythingNamed;
5430 spec.numWorkGroups = IVec3(numElements, 1, 1);
5431 spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
5432 spec.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
5434 abuseGroup->addChild(new SpvAsmComputeShaderCase(testCtx, "everything_named", "everything_named", spec));
5437 const std::string everythingNamedTheSame =
5438 commonShaderHeader +
5439 "OpName %main \"the_same\"\n"
5440 "OpName %id \"the_same\"\n"
5441 "OpName %zero \"the_same\"\n"
5442 "OpName %entry \"the_same\"\n"
5443 "OpName %func \"the_same\"\n"
5444 "OpName %5 \"the_same\"\n"
5445 "OpName %7 \"the_same\"\n"
5446 "OpName %idval \"the_same\"\n"
5447 "OpName %inloc \"the_same\"\n"
5448 "OpName %inval \"the_same\"\n"
5449 "OpName %neg \"the_same\"\n"
5450 "OpName %outloc \"the_same\"\n"+
5453 ComputeShaderSpec spec;
5455 spec.assembly = everythingNamedTheSame;
5456 spec.numWorkGroups = IVec3(numElements, 1, 1);
5457 spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
5458 spec.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
5460 abuseGroup->addChild(new SpvAsmComputeShaderCase(testCtx, "everything_named_the_same", "everything_named_the_same", spec));
5464 for (size_t ndx = 0; ndx < abuseCases.size(); ++ndx)
5466 map<string, string> specializations;
5467 ComputeShaderSpec spec;
5469 specializations["ENTRY"] = "main";
5470 specializations["ID"] = "main";
5471 specializations["NAME"] = abuseCases[ndx].param;
5472 spec.assembly = shaderTemplate.specialize(specializations);
5473 spec.numWorkGroups = IVec3(numElements, 1, 1);
5474 spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
5475 spec.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
5477 abuseGroup->addChild(new SpvAsmComputeShaderCase(testCtx, (std::string("main") + abuseCases[ndx].name).c_str(), abuseCases[ndx].name, spec));
5481 for (size_t ndx = 0; ndx < abuseCases.size(); ++ndx)
5483 map<string, string> specializations;
5484 ComputeShaderSpec spec;
5486 specializations["ENTRY"] = "main";
5487 specializations["ID"] = "x";
5488 specializations["NAME"] = abuseCases[ndx].param;
5489 spec.assembly = shaderTemplate.specialize(specializations);
5490 spec.numWorkGroups = IVec3(numElements, 1, 1);
5491 spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
5492 spec.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
5494 abuseGroup->addChild(new SpvAsmComputeShaderCase(testCtx, (std::string("x") + abuseCases[ndx].name).c_str(), abuseCases[ndx].name, spec));
5497 cases.push_back(CaseParameter("_is_main", "main"));
5498 cases.push_back(CaseParameter("_is_not_main", "not_main"));
5499 testFunc.push_back("main");
5500 testFunc.push_back("func");
5502 for(size_t fNdx = 0; fNdx < testFunc.size(); ++fNdx)
5504 for(size_t ndx = 0; ndx < cases.size(); ++ndx)
5506 map<string, string> specializations;
5507 ComputeShaderSpec spec;
5509 specializations["ENTRY"] = "main";
5510 specializations["ID"] = testFunc[fNdx];
5511 specializations["NAME"] = cases[ndx].param;
5512 spec.assembly = shaderTemplate.specialize(specializations);
5513 spec.numWorkGroups = IVec3(numElements, 1, 1);
5514 spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
5515 spec.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
5517 entryMainGroup->addChild(new SpvAsmComputeShaderCase(testCtx, (testFunc[fNdx] + cases[ndx].name).c_str(), cases[ndx].name, spec));
5521 cases.push_back(CaseParameter("_is_entry", "rdc"));
5523 for(size_t fNdx = 0; fNdx < testFunc.size(); ++fNdx)
5525 for(size_t ndx = 0; ndx < cases.size(); ++ndx)
5527 map<string, string> specializations;
5528 ComputeShaderSpec spec;
5530 specializations["ENTRY"] = "rdc";
5531 specializations["ID"] = testFunc[fNdx];
5532 specializations["NAME"] = cases[ndx].param;
5533 spec.assembly = shaderTemplate.specialize(specializations);
5534 spec.numWorkGroups = IVec3(numElements, 1, 1);
5535 spec.entryPoint = "rdc";
5536 spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
5537 spec.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
5539 entryNotGroup->addChild(new SpvAsmComputeShaderCase(testCtx, (testFunc[fNdx] + cases[ndx].name).c_str(), cases[ndx].name, spec));
5543 group->addChild(entryMainGroup.release());
5544 group->addChild(entryNotGroup.release());
5545 group->addChild(abuseGroup.release());
5547 return group.release();
5550 tcu::TestCaseGroup* createOpMemberNameGroup (tcu::TestContext& testCtx)
5552 de::MovePtr<tcu::TestCaseGroup> group(new tcu::TestCaseGroup(testCtx, "opmembername", "Tests OpMemberName cases"));
5553 de::MovePtr<tcu::TestCaseGroup> abuseGroup(new tcu::TestCaseGroup(testCtx, "abuse", "OpMemberName abuse tests"));
5554 vector<CaseParameter> abuseCases;
5555 vector<string> testFunc;
5556 de::Random rnd(deStringHash(group->getName()));
5557 const int numElements = 128;
5558 vector<float> inputFloats(numElements, 0);
5559 vector<float> outputFloats(numElements, 0);
5561 getOpNameAbuseCases(abuseCases);
5563 fillRandomScalars(rnd, -100.0f, 100.0f, &inputFloats[0], numElements);
5565 for (size_t ndx = 0; ndx < numElements; ++ndx)
5566 outputFloats[ndx] = -inputFloats[ndx];
5568 const string commonShaderHeader =
5569 "OpCapability Shader\n"
5570 "OpMemoryModel Logical GLSL450\n"
5571 "OpEntryPoint GLCompute %main \"main\" %id\n"
5572 "OpExecutionMode %main LocalSize 1 1 1\n";
5574 const string commonShaderFooter =
5575 "OpDecorate %id BuiltIn GlobalInvocationId\n"
5577 + string(getComputeAsmInputOutputBufferTraits())
5578 + string(getComputeAsmCommonTypes())
5579 + string(getComputeAsmInputOutputBuffer()) +
5581 "%u3str = OpTypeStruct %u32 %u32 %u32\n"
5583 "%id = OpVariable %uvec3ptr Input\n"
5584 "%zero = OpConstant %i32 0\n"
5586 "%main = OpFunction %void None %voidf\n"
5587 "%entry = OpLabel\n"
5589 "%idval = OpLoad %uvec3 %id\n"
5590 "%x0 = OpCompositeExtract %u32 %idval 0\n"
5592 "%idstr = OpCompositeConstruct %u3str %x0 %x0 %x0\n"
5593 "%x = OpCompositeExtract %u32 %idstr 0\n"
5595 "%inloc = OpAccessChain %f32ptr %indata %zero %x\n"
5596 "%inval = OpLoad %f32 %inloc\n"
5597 "%neg = OpFNegate %f32 %inval\n"
5598 "%outloc = OpAccessChain %f32ptr %outdata %zero %x\n"
5599 " OpStore %outloc %neg\n"
5604 const StringTemplate shaderTemplate(
5605 commonShaderHeader +
5606 "OpMemberName %u3str 0 \"${NAME}\"\n" +
5607 commonShaderFooter);
5609 const std::string multipleNames =
5610 commonShaderHeader +
5611 "OpMemberName %u3str 0 \"to_be\"\n"
5612 "OpMemberName %u3str 1 \"or_not\"\n"
5613 "OpMemberName %u3str 0 \"to_be\"\n"
5614 "OpMemberName %u3str 2 \"makes_no\"\n"
5615 "OpMemberName %u3str 0 \"difference\"\n"
5616 "OpMemberName %u3str 0 \"to_me\"\n" +
5619 ComputeShaderSpec spec;
5621 spec.assembly = multipleNames;
5622 spec.numWorkGroups = IVec3(numElements, 1, 1);
5623 spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
5624 spec.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
5626 abuseGroup->addChild(new SpvAsmComputeShaderCase(testCtx, "u3str_x_has_multiple_names", "multiple_names", spec));
5629 const std::string everythingNamedTheSame =
5630 commonShaderHeader +
5631 "OpMemberName %u3str 0 \"the_same\"\n"
5632 "OpMemberName %u3str 1 \"the_same\"\n"
5633 "OpMemberName %u3str 2 \"the_same\"\n" +
5637 ComputeShaderSpec spec;
5639 spec.assembly = everythingNamedTheSame;
5640 spec.numWorkGroups = IVec3(numElements, 1, 1);
5641 spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
5642 spec.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
5644 abuseGroup->addChild(new SpvAsmComputeShaderCase(testCtx, "everything_named_the_same", "everything_named_the_same", spec));
5648 for (size_t ndx = 0; ndx < abuseCases.size(); ++ndx)
5650 map<string, string> specializations;
5651 ComputeShaderSpec spec;
5653 specializations["NAME"] = abuseCases[ndx].param;
5654 spec.assembly = shaderTemplate.specialize(specializations);
5655 spec.numWorkGroups = IVec3(numElements, 1, 1);
5656 spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
5657 spec.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
5659 abuseGroup->addChild(new SpvAsmComputeShaderCase(testCtx, (std::string("u3str_x") + abuseCases[ndx].name).c_str(), abuseCases[ndx].name, spec));
5662 group->addChild(abuseGroup.release());
5664 return group.release();
5667 // Assembly code used for testing function control is based on GLSL source code:
5671 // layout(std140, set = 0, binding = 0) readonly buffer Input {
5672 // float elements[];
5674 // layout(std140, set = 0, binding = 1) writeonly buffer Output {
5675 // float elements[];
5678 // float const10() { return 10.f; }
5681 // uint x = gl_GlobalInvocationID.x;
5682 // output_data.elements[x] = input_data.elements[x] + const10();
5684 tcu::TestCaseGroup* createFunctionControlGroup (tcu::TestContext& testCtx)
5686 de::MovePtr<tcu::TestCaseGroup> group (new tcu::TestCaseGroup(testCtx, "function_control", "Tests function control cases"));
5687 vector<CaseParameter> cases;
5688 de::Random rnd (deStringHash(group->getName()));
5689 const int numElements = 100;
5690 vector<float> inputFloats (numElements, 0);
5691 vector<float> outputFloats (numElements, 0);
5692 const StringTemplate shaderTemplate (
5693 string(getComputeAsmShaderPreamble()) +
5695 "OpSource GLSL 430\n"
5696 "OpName %main \"main\"\n"
5697 "OpName %func_const10 \"const10(\"\n"
5698 "OpName %id \"gl_GlobalInvocationID\"\n"
5700 "OpDecorate %id BuiltIn GlobalInvocationId\n"
5702 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) + string(getComputeAsmInputOutputBuffer()) +
5704 "%f32f = OpTypeFunction %f32\n"
5705 "%id = OpVariable %uvec3ptr Input\n"
5706 "%zero = OpConstant %i32 0\n"
5707 "%constf10 = OpConstant %f32 10.0\n"
5709 "%main = OpFunction %void None %voidf\n"
5710 "%entry = OpLabel\n"
5711 "%idval = OpLoad %uvec3 %id\n"
5712 "%x = OpCompositeExtract %u32 %idval 0\n"
5713 "%inloc = OpAccessChain %f32ptr %indata %zero %x\n"
5714 "%inval = OpLoad %f32 %inloc\n"
5715 "%ret_10 = OpFunctionCall %f32 %func_const10\n"
5716 "%fadd = OpFAdd %f32 %inval %ret_10\n"
5717 "%outloc = OpAccessChain %f32ptr %outdata %zero %x\n"
5718 " OpStore %outloc %fadd\n"
5722 "%func_const10 = OpFunction %f32 ${CONTROL} %f32f\n"
5723 "%label = OpLabel\n"
5724 " OpReturnValue %constf10\n"
5725 " OpFunctionEnd\n");
5727 cases.push_back(CaseParameter("none", "None"));
5728 cases.push_back(CaseParameter("inline", "Inline"));
5729 cases.push_back(CaseParameter("dont_inline", "DontInline"));
5730 cases.push_back(CaseParameter("pure", "Pure"));
5731 cases.push_back(CaseParameter("const", "Const"));
5732 cases.push_back(CaseParameter("inline_pure", "Inline|Pure"));
5733 cases.push_back(CaseParameter("const_dont_inline", "Const|DontInline"));
5734 cases.push_back(CaseParameter("inline_dont_inline", "Inline|DontInline"));
5735 cases.push_back(CaseParameter("pure_inline_dont_inline", "Pure|Inline|DontInline"));
5737 fillRandomScalars(rnd, -100.f, 100.f, &inputFloats[0], numElements);
5739 // CPU might not use the same rounding mode as the GPU. Use whole numbers to avoid rounding differences.
5740 floorAll(inputFloats);
5742 for (size_t ndx = 0; ndx < numElements; ++ndx)
5743 outputFloats[ndx] = inputFloats[ndx] + 10.f;
5745 for (size_t caseNdx = 0; caseNdx < cases.size(); ++caseNdx)
5747 map<string, string> specializations;
5748 ComputeShaderSpec spec;
5750 specializations["CONTROL"] = cases[caseNdx].param;
5751 spec.assembly = shaderTemplate.specialize(specializations);
5752 spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
5753 spec.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
5754 spec.numWorkGroups = IVec3(numElements, 1, 1);
5756 group->addChild(new SpvAsmComputeShaderCase(testCtx, cases[caseNdx].name, cases[caseNdx].name, spec));
5759 return group.release();
5762 tcu::TestCaseGroup* createMemoryAccessGroup (tcu::TestContext& testCtx)
5764 de::MovePtr<tcu::TestCaseGroup> group (new tcu::TestCaseGroup(testCtx, "memory_access", "Tests memory access cases"));
5765 vector<CaseParameter> cases;
5766 de::Random rnd (deStringHash(group->getName()));
5767 const int numElements = 100;
5768 vector<float> inputFloats (numElements, 0);
5769 vector<float> outputFloats (numElements, 0);
5770 const StringTemplate shaderTemplate (
5771 string(getComputeAsmShaderPreamble()) +
5773 "OpSource GLSL 430\n"
5774 "OpName %main \"main\"\n"
5775 "OpName %id \"gl_GlobalInvocationID\"\n"
5777 "OpDecorate %id BuiltIn GlobalInvocationId\n"
5779 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) + string(getComputeAsmInputOutputBuffer()) +
5781 "%f32ptr_f = OpTypePointer Function %f32\n"
5783 "%id = OpVariable %uvec3ptr Input\n"
5784 "%zero = OpConstant %i32 0\n"
5785 "%four = OpConstant %i32 4\n"
5787 "%main = OpFunction %void None %voidf\n"
5788 "%label = OpLabel\n"
5789 "%copy = OpVariable %f32ptr_f Function\n"
5790 "%idval = OpLoad %uvec3 %id ${ACCESS}\n"
5791 "%x = OpCompositeExtract %u32 %idval 0\n"
5792 "%inloc = OpAccessChain %f32ptr %indata %zero %x\n"
5793 "%outloc = OpAccessChain %f32ptr %outdata %zero %x\n"
5794 " OpCopyMemory %copy %inloc ${ACCESS}\n"
5795 "%val1 = OpLoad %f32 %copy\n"
5796 "%val2 = OpLoad %f32 %inloc\n"
5797 "%add = OpFAdd %f32 %val1 %val2\n"
5798 " OpStore %outloc %add ${ACCESS}\n"
5800 " OpFunctionEnd\n");
5802 cases.push_back(CaseParameter("null", ""));
5803 cases.push_back(CaseParameter("none", "None"));
5804 cases.push_back(CaseParameter("volatile", "Volatile"));
5805 cases.push_back(CaseParameter("aligned", "Aligned 4"));
5806 cases.push_back(CaseParameter("nontemporal", "Nontemporal"));
5807 cases.push_back(CaseParameter("aligned_nontemporal", "Aligned|Nontemporal 4"));
5808 cases.push_back(CaseParameter("aligned_volatile", "Volatile|Aligned 4"));
5810 fillRandomScalars(rnd, -100.f, 100.f, &inputFloats[0], numElements);
5812 for (size_t ndx = 0; ndx < numElements; ++ndx)
5813 outputFloats[ndx] = inputFloats[ndx] + inputFloats[ndx];
5815 for (size_t caseNdx = 0; caseNdx < cases.size(); ++caseNdx)
5817 map<string, string> specializations;
5818 ComputeShaderSpec spec;
5820 specializations["ACCESS"] = cases[caseNdx].param;
5821 spec.assembly = shaderTemplate.specialize(specializations);
5822 spec.inputs.push_back(BufferSp(new Float32Buffer(inputFloats)));
5823 spec.outputs.push_back(BufferSp(new Float32Buffer(outputFloats)));
5824 spec.numWorkGroups = IVec3(numElements, 1, 1);
5826 group->addChild(new SpvAsmComputeShaderCase(testCtx, cases[caseNdx].name, cases[caseNdx].name, spec));
5829 return group.release();
5832 // Checks that we can get undefined values for various types, without exercising a computation with it.
5833 tcu::TestCaseGroup* createOpUndefGroup (tcu::TestContext& testCtx)
5835 de::MovePtr<tcu::TestCaseGroup> group (new tcu::TestCaseGroup(testCtx, "opundef", "Tests the OpUndef instruction"));
5836 vector<CaseParameter> cases;
5837 de::Random rnd (deStringHash(group->getName()));
5838 const int numElements = 100;
5839 vector<float> positiveFloats (numElements, 0);
5840 vector<float> negativeFloats (numElements, 0);
5841 const StringTemplate shaderTemplate (
5842 string(getComputeAsmShaderPreamble()) +
5844 "OpSource GLSL 430\n"
5845 "OpName %main \"main\"\n"
5846 "OpName %id \"gl_GlobalInvocationID\"\n"
5848 "OpDecorate %id BuiltIn GlobalInvocationId\n"
5850 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) +
5851 "%uvec2 = OpTypeVector %u32 2\n"
5852 "%fvec4 = OpTypeVector %f32 4\n"
5853 "%fmat33 = OpTypeMatrix %fvec3 3\n"
5854 "%image = OpTypeImage %f32 2D 0 0 0 1 Unknown\n"
5855 "%sampler = OpTypeSampler\n"
5856 "%simage = OpTypeSampledImage %image\n"
5857 "%const100 = OpConstant %u32 100\n"
5858 "%uarr100 = OpTypeArray %i32 %const100\n"
5859 "%struct = OpTypeStruct %f32 %i32 %u32\n"
5860 "%pointer = OpTypePointer Function %i32\n"
5861 + string(getComputeAsmInputOutputBuffer()) +
5863 "%id = OpVariable %uvec3ptr Input\n"
5864 "%zero = OpConstant %i32 0\n"
5866 "%main = OpFunction %void None %voidf\n"
5867 "%label = OpLabel\n"
5869 "%undef = OpUndef ${TYPE}\n"
5871 "%idval = OpLoad %uvec3 %id\n"
5872 "%x = OpCompositeExtract %u32 %idval 0\n"
5874 "%inloc = OpAccessChain %f32ptr %indata %zero %x\n"
5875 "%inval = OpLoad %f32 %inloc\n"
5876 "%neg = OpFNegate %f32 %inval\n"
5877 "%outloc = OpAccessChain %f32ptr %outdata %zero %x\n"
5878 " OpStore %outloc %neg\n"
5880 " OpFunctionEnd\n");
5882 cases.push_back(CaseParameter("bool", "%bool"));
5883 cases.push_back(CaseParameter("sint32", "%i32"));
5884 cases.push_back(CaseParameter("uint32", "%u32"));
5885 cases.push_back(CaseParameter("float32", "%f32"));
5886 cases.push_back(CaseParameter("vec4float32", "%fvec4"));
5887 cases.push_back(CaseParameter("vec2uint32", "%uvec2"));
5888 cases.push_back(CaseParameter("matrix", "%fmat33"));
5889 cases.push_back(CaseParameter("image", "%image"));
5890 cases.push_back(CaseParameter("sampler", "%sampler"));
5891 cases.push_back(CaseParameter("sampledimage", "%simage"));
5892 cases.push_back(CaseParameter("array", "%uarr100"));
5893 cases.push_back(CaseParameter("runtimearray", "%f32arr"));
5894 cases.push_back(CaseParameter("struct", "%struct"));
5895 cases.push_back(CaseParameter("pointer", "%pointer"));
5897 fillRandomScalars(rnd, 1.f, 100.f, &positiveFloats[0], numElements);
5899 for (size_t ndx = 0; ndx < numElements; ++ndx)
5900 negativeFloats[ndx] = -positiveFloats[ndx];
5902 for (size_t caseNdx = 0; caseNdx < cases.size(); ++caseNdx)
5904 map<string, string> specializations;
5905 ComputeShaderSpec spec;
5907 specializations["TYPE"] = cases[caseNdx].param;
5908 spec.assembly = shaderTemplate.specialize(specializations);
5909 spec.inputs.push_back(BufferSp(new Float32Buffer(positiveFloats)));
5910 spec.outputs.push_back(BufferSp(new Float32Buffer(negativeFloats)));
5911 spec.numWorkGroups = IVec3(numElements, 1, 1);
5913 group->addChild(new SpvAsmComputeShaderCase(testCtx, cases[caseNdx].name, cases[caseNdx].name, spec));
5916 return group.release();
5919 // Checks that a compute shader can generate a constant composite value of various types, without exercising a computation on it.
5920 tcu::TestCaseGroup* createFloat16OpConstantCompositeGroup (tcu::TestContext& testCtx)
5922 de::MovePtr<tcu::TestCaseGroup> group (new tcu::TestCaseGroup(testCtx, "opconstantcomposite", "Tests the OpConstantComposite instruction"));
5923 vector<CaseParameter> cases;
5924 de::Random rnd (deStringHash(group->getName()));
5925 const int numElements = 100;
5926 vector<float> positiveFloats (numElements, 0);
5927 vector<float> negativeFloats (numElements, 0);
5928 const StringTemplate shaderTemplate (
5929 "OpCapability Shader\n"
5930 "OpCapability Float16\n"
5931 "OpMemoryModel Logical GLSL450\n"
5932 "OpEntryPoint GLCompute %main \"main\" %id\n"
5933 "OpExecutionMode %main LocalSize 1 1 1\n"
5934 "OpSource GLSL 430\n"
5935 "OpName %main \"main\"\n"
5936 "OpName %id \"gl_GlobalInvocationID\"\n"
5938 "OpDecorate %id BuiltIn GlobalInvocationId\n"
5940 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) + string(getComputeAsmInputOutputBuffer()) +
5942 "%id = OpVariable %uvec3ptr Input\n"
5943 "%zero = OpConstant %i32 0\n"
5944 "%f16 = OpTypeFloat 16\n"
5945 "%c_f16_0 = OpConstant %f16 0.0\n"
5946 "%c_f16_0_5 = OpConstant %f16 0.5\n"
5947 "%c_f16_1 = OpConstant %f16 1.0\n"
5948 "%v2f16 = OpTypeVector %f16 2\n"
5949 "%v3f16 = OpTypeVector %f16 3\n"
5950 "%v4f16 = OpTypeVector %f16 4\n"
5954 "%main = OpFunction %void None %voidf\n"
5955 "%label = OpLabel\n"
5956 "%idval = OpLoad %uvec3 %id\n"
5957 "%x = OpCompositeExtract %u32 %idval 0\n"
5958 "%inloc = OpAccessChain %f32ptr %indata %zero %x\n"
5959 "%inval = OpLoad %f32 %inloc\n"
5960 "%neg = OpFNegate %f32 %inval\n"
5961 "%outloc = OpAccessChain %f32ptr %outdata %zero %x\n"
5962 " OpStore %outloc %neg\n"
5964 " OpFunctionEnd\n");
5967 cases.push_back(CaseParameter("vector", "%const = OpConstantComposite %v3f16 %c_f16_0 %c_f16_0_5 %c_f16_1\n"));
5968 cases.push_back(CaseParameter("matrix", "%m3v3f16 = OpTypeMatrix %v3f16 3\n"
5969 "%vec = OpConstantComposite %v3f16 %c_f16_0 %c_f16_0_5 %c_f16_1\n"
5970 "%mat = OpConstantComposite %m3v3f16 %vec %vec %vec"));
5971 cases.push_back(CaseParameter("struct", "%m2v3f16 = OpTypeMatrix %v3f16 2\n"
5972 "%struct = OpTypeStruct %i32 %f16 %v3f16 %m2v3f16\n"
5973 "%vec = OpConstantComposite %v3f16 %c_f16_0 %c_f16_0_5 %c_f16_1\n"
5974 "%mat = OpConstantComposite %m2v3f16 %vec %vec\n"
5975 "%const = OpConstantComposite %struct %zero %c_f16_0_5 %vec %mat\n"));
5976 cases.push_back(CaseParameter("nested_struct", "%st1 = OpTypeStruct %i32 %f16\n"
5977 "%st2 = OpTypeStruct %i32 %i32\n"
5978 "%struct = OpTypeStruct %st1 %st2\n"
5979 "%st1val = OpConstantComposite %st1 %zero %c_f16_0_5\n"
5980 "%st2val = OpConstantComposite %st2 %zero %zero\n"
5981 "%const = OpConstantComposite %struct %st1val %st2val"));
5983 fillRandomScalars(rnd, 1.f, 100.f, &positiveFloats[0], numElements);
5985 for (size_t ndx = 0; ndx < numElements; ++ndx)
5986 negativeFloats[ndx] = -positiveFloats[ndx];
5988 for (size_t caseNdx = 0; caseNdx < cases.size(); ++caseNdx)
5990 map<string, string> specializations;
5991 ComputeShaderSpec spec;
5993 specializations["CONSTANT"] = cases[caseNdx].param;
5994 spec.assembly = shaderTemplate.specialize(specializations);
5995 spec.inputs.push_back(BufferSp(new Float32Buffer(positiveFloats)));
5996 spec.outputs.push_back(BufferSp(new Float32Buffer(negativeFloats)));
5997 spec.numWorkGroups = IVec3(numElements, 1, 1);
5999 spec.extensions.push_back("VK_KHR_16bit_storage");
6000 spec.extensions.push_back("VK_KHR_shader_float16_int8");
6002 spec.requestedVulkanFeatures.ext16BitStorage = EXT16BITSTORAGEFEATURES_UNIFORM_BUFFER_BLOCK;
6003 spec.requestedVulkanFeatures.extFloat16Int8 = EXTFLOAT16INT8FEATURES_FLOAT16;
6005 group->addChild(new SpvAsmComputeShaderCase(testCtx, cases[caseNdx].name, cases[caseNdx].name, spec));
6008 return group.release();
6011 const vector<deFloat16> squarize(const vector<deFloat16>& inData, const deUint32 argNo)
6013 const size_t inDataLength = inData.size();
6014 vector<deFloat16> result;
6016 result.reserve(inDataLength * inDataLength);
6020 for (size_t numIdx = 0; numIdx < inDataLength; ++numIdx)
6021 result.insert(result.end(), inData.begin(), inData.end());
6026 for (size_t numIdx = 0; numIdx < inDataLength; ++numIdx)
6028 const vector<deFloat16> tmp(inDataLength, inData[numIdx]);
6030 result.insert(result.end(), tmp.begin(), tmp.end());
6037 const vector<deFloat16> squarizeVector(const vector<deFloat16>& inData, const deUint32 argNo)
6039 vector<deFloat16> vec;
6040 vector<deFloat16> result;
6042 // Create vectors. vec will contain each possible pair from inData
6044 const size_t inDataLength = inData.size();
6046 DE_ASSERT(inDataLength <= 64);
6048 vec.reserve(2 * inDataLength * inDataLength);
6050 for (size_t numIdxX = 0; numIdxX < inDataLength; ++numIdxX)
6051 for (size_t numIdxY = 0; numIdxY < inDataLength; ++numIdxY)
6053 vec.push_back(inData[numIdxX]);
6054 vec.push_back(inData[numIdxY]);
6058 // Create vector pairs. result will contain each possible pair from vec
6060 const size_t coordsPerVector = 2;
6061 const size_t vectorsCount = vec.size() / coordsPerVector;
6063 result.reserve(coordsPerVector * vectorsCount * vectorsCount);
6067 for (size_t numIdxX = 0; numIdxX < vectorsCount; ++numIdxX)
6068 for (size_t numIdxY = 0; numIdxY < vectorsCount; ++numIdxY)
6070 for (size_t coordNdx = 0; coordNdx < coordsPerVector; ++coordNdx)
6071 result.push_back(vec[coordsPerVector * numIdxY + coordNdx]);
6077 for (size_t numIdxX = 0; numIdxX < vectorsCount; ++numIdxX)
6078 for (size_t numIdxY = 0; numIdxY < vectorsCount; ++numIdxY)
6080 for (size_t coordNdx = 0; coordNdx < coordsPerVector; ++coordNdx)
6081 result.push_back(vec[coordsPerVector * numIdxX + coordNdx]);
6089 struct fp16isNan { bool operator()(const tcu::Float16 in1, const tcu::Float16) { return in1.isNaN(); } };
6090 struct fp16isInf { bool operator()(const tcu::Float16 in1, const tcu::Float16) { return in1.isInf(); } };
6091 struct fp16isEqual { bool operator()(const tcu::Float16 in1, const tcu::Float16 in2) { return in1.asFloat() == in2.asFloat(); } };
6092 struct fp16isUnequal { bool operator()(const tcu::Float16 in1, const tcu::Float16 in2) { return in1.asFloat() != in2.asFloat(); } };
6093 struct fp16isLess { bool operator()(const tcu::Float16 in1, const tcu::Float16 in2) { return in1.asFloat() < in2.asFloat(); } };
6094 struct fp16isGreater { bool operator()(const tcu::Float16 in1, const tcu::Float16 in2) { return in1.asFloat() > in2.asFloat(); } };
6095 struct fp16isLessOrEqual { bool operator()(const tcu::Float16 in1, const tcu::Float16 in2) { return in1.asFloat() <= in2.asFloat(); } };
6096 struct fp16isGreaterOrEqual { bool operator()(const tcu::Float16 in1, const tcu::Float16 in2) { return in1.asFloat() >= in2.asFloat(); } };
6098 template <class TestedLogicalFunction, bool onlyTestFunc, bool unationModeAnd, bool nanSupported>
6099 bool compareFP16Logical (const std::vector<Resource>& inputs, const vector<AllocationSp>& outputAllocs, const std::vector<Resource>&, TestLog& log)
6101 if (inputs.size() != 2 || outputAllocs.size() != 1)
6104 vector<deUint8> input1Bytes;
6105 vector<deUint8> input2Bytes;
6107 inputs[0].getBytes(input1Bytes);
6108 inputs[1].getBytes(input2Bytes);
6110 const deUint32 denormModesCount = 2;
6111 const deFloat16 float16one = tcu::Float16(1.0f).bits();
6112 const deFloat16 float16zero = tcu::Float16(0.0f).bits();
6113 const tcu::Float16 zero = tcu::Float16::zero(1);
6114 const deFloat16* const outputAsFP16 = static_cast<deFloat16*>(outputAllocs[0]->getHostPtr());
6115 const deFloat16* const input1AsFP16 = reinterpret_cast<deFloat16* const>(&input1Bytes.front());
6116 const deFloat16* const input2AsFP16 = reinterpret_cast<deFloat16* const>(&input2Bytes.front());
6117 deUint32 successfulRuns = denormModesCount;
6118 std::string results[denormModesCount];
6119 TestedLogicalFunction testedLogicalFunction;
6121 for (deUint32 denormMode = 0; denormMode < denormModesCount; denormMode++)
6123 const bool flushToZero = (denormMode == 1);
6125 for (size_t idx = 0; idx < input1Bytes.size() / sizeof(deFloat16); ++idx)
6127 const tcu::Float16 f1pre = tcu::Float16(input1AsFP16[idx]);
6128 const tcu::Float16 f2pre = tcu::Float16(input2AsFP16[idx]);
6129 const tcu::Float16 f1 = (flushToZero && f1pre.isDenorm()) ? zero : f1pre;
6130 const tcu::Float16 f2 = (flushToZero && f2pre.isDenorm()) ? zero : f2pre;
6131 deFloat16 expectedOutput = float16zero;
6135 if (testedLogicalFunction(f1, f2))
6136 expectedOutput = float16one;
6140 const bool f1nan = f1.isNaN();
6141 const bool f2nan = f2.isNaN();
6143 // Skip NaN floats if not supported by implementation
6144 if (!nanSupported && (f1nan || f2nan))
6149 const bool ordered = !f1nan && !f2nan;
6151 if (ordered && testedLogicalFunction(f1, f2))
6152 expectedOutput = float16one;
6156 const bool unordered = f1nan || f2nan;
6158 if (unordered || testedLogicalFunction(f1, f2))
6159 expectedOutput = float16one;
6163 if (outputAsFP16[idx] != expectedOutput)
6165 std::ostringstream str;
6167 str << "ERROR: Sub-case #" << idx
6168 << " flushToZero:" << flushToZero
6170 << " failed, inputs: 0x" << f1.bits()
6171 << ";0x" << f2.bits()
6172 << " output: 0x" << outputAsFP16[idx]
6173 << " expected output: 0x" << expectedOutput;
6175 results[denormMode] = str.str();
6184 if (successfulRuns == 0)
6185 for (deUint32 denormMode = 0; denormMode < denormModesCount; denormMode++)
6186 log << TestLog::Message << results[denormMode] << TestLog::EndMessage;
6188 return successfulRuns > 0;
6193 tcu::TestCaseGroup* createOpSourceTests (tcu::TestContext& testCtx)
6195 struct NameCodePair { string name, code; };
6196 RGBA defaultColors[4];
6197 de::MovePtr<tcu::TestCaseGroup> opSourceTests (new tcu::TestCaseGroup(testCtx, "opsource", "OpSource instruction"));
6198 const std::string opsourceGLSLWithFile = "%opsrcfile = OpString \"foo.vert\"\nOpSource GLSL 450 %opsrcfile ";
6199 map<string, string> fragments = passthruFragments();
6200 const NameCodePair tests[] =
6202 {"unknown", "OpSource Unknown 321"},
6203 {"essl", "OpSource ESSL 310"},
6204 {"glsl", "OpSource GLSL 450"},
6205 {"opencl_cpp", "OpSource OpenCL_CPP 120"},
6206 {"opencl_c", "OpSource OpenCL_C 120"},
6207 {"multiple", "OpSource GLSL 450\nOpSource GLSL 450"},
6208 {"file", opsourceGLSLWithFile},
6209 {"source", opsourceGLSLWithFile + "\"void main(){}\""},
6210 // Longest possible source string: SPIR-V limits instructions to 65535
6211 // words, of which the first 4 are opsourceGLSLWithFile; the rest will
6212 // contain 65530 UTF8 characters (one word each) plus one last word
6213 // containing 3 ASCII characters and \0.
6214 {"longsource", opsourceGLSLWithFile + '"' + makeLongUTF8String(65530) + "ccc" + '"'}
6217 getDefaultColors(defaultColors);
6218 for (size_t testNdx = 0; testNdx < sizeof(tests) / sizeof(NameCodePair); ++testNdx)
6220 fragments["debug"] = tests[testNdx].code;
6221 createTestsForAllStages(tests[testNdx].name, defaultColors, defaultColors, fragments, opSourceTests.get());
6224 return opSourceTests.release();
6227 tcu::TestCaseGroup* createOpSourceContinuedTests (tcu::TestContext& testCtx)
6229 struct NameCodePair { string name, code; };
6230 RGBA defaultColors[4];
6231 de::MovePtr<tcu::TestCaseGroup> opSourceTests (new tcu::TestCaseGroup(testCtx, "opsourcecontinued", "OpSourceContinued instruction"));
6232 map<string, string> fragments = passthruFragments();
6233 const std::string opsource = "%opsrcfile = OpString \"foo.vert\"\nOpSource GLSL 450 %opsrcfile \"void main(){}\"\n";
6234 const NameCodePair tests[] =
6236 {"empty", opsource + "OpSourceContinued \"\""},
6237 {"short", opsource + "OpSourceContinued \"abcde\""},
6238 {"multiple", opsource + "OpSourceContinued \"abcde\"\nOpSourceContinued \"fghij\""},
6239 // Longest possible source string: SPIR-V limits instructions to 65535
6240 // words, of which the first one is OpSourceContinued/length; the rest
6241 // will contain 65533 UTF8 characters (one word each) plus one last word
6242 // containing 3 ASCII characters and \0.
6243 {"long", opsource + "OpSourceContinued \"" + makeLongUTF8String(65533) + "ccc\""}
6246 getDefaultColors(defaultColors);
6247 for (size_t testNdx = 0; testNdx < sizeof(tests) / sizeof(NameCodePair); ++testNdx)
6249 fragments["debug"] = tests[testNdx].code;
6250 createTestsForAllStages(tests[testNdx].name, defaultColors, defaultColors, fragments, opSourceTests.get());
6253 return opSourceTests.release();
6255 tcu::TestCaseGroup* createOpNoLineTests(tcu::TestContext& testCtx)
6257 RGBA defaultColors[4];
6258 de::MovePtr<tcu::TestCaseGroup> opLineTests (new tcu::TestCaseGroup(testCtx, "opnoline", "OpNoLine instruction"));
6259 map<string, string> fragments;
6260 getDefaultColors(defaultColors);
6261 fragments["debug"] =
6262 "%name = OpString \"name\"\n";
6264 fragments["pre_main"] =
6267 "OpLine %name 1 1\n"
6269 "OpLine %name 1 1\n"
6270 "OpLine %name 1 1\n"
6271 "%second_function = OpFunction %v4f32 None %v4f32_v4f32_function\n"
6273 "OpLine %name 1 1\n"
6275 "OpLine %name 1 1\n"
6276 "OpLine %name 1 1\n"
6277 "%second_param1 = OpFunctionParameter %v4f32\n"
6280 "%label_secondfunction = OpLabel\n"
6282 "OpReturnValue %second_param1\n"
6287 fragments["testfun"] =
6288 // A %test_code function that returns its argument unchanged.
6291 "OpLine %name 1 1\n"
6292 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
6294 "%param1 = OpFunctionParameter %v4f32\n"
6297 "%label_testfun = OpLabel\n"
6299 "%val1 = OpFunctionCall %v4f32 %second_function %param1\n"
6300 "OpReturnValue %val1\n"
6302 "OpLine %name 1 1\n"
6305 createTestsForAllStages("opnoline", defaultColors, defaultColors, fragments, opLineTests.get());
6307 return opLineTests.release();
6310 tcu::TestCaseGroup* createOpModuleProcessedTests(tcu::TestContext& testCtx)
6312 RGBA defaultColors[4];
6313 de::MovePtr<tcu::TestCaseGroup> opModuleProcessedTests (new tcu::TestCaseGroup(testCtx, "opmoduleprocessed", "OpModuleProcessed instruction"));
6314 map<string, string> fragments;
6315 std::vector<std::string> noExtensions;
6316 GraphicsResources resources;
6318 getDefaultColors(defaultColors);
6319 resources.verifyBinary = veryfiBinaryShader;
6320 resources.spirvVersion = SPIRV_VERSION_1_3;
6322 fragments["moduleprocessed"] =
6323 "OpModuleProcessed \"VULKAN CTS\"\n"
6324 "OpModuleProcessed \"Negative values\"\n"
6325 "OpModuleProcessed \"Date: 2017/09/21\"\n";
6327 fragments["pre_main"] =
6328 "%second_function = OpFunction %v4f32 None %v4f32_v4f32_function\n"
6329 "%second_param1 = OpFunctionParameter %v4f32\n"
6330 "%label_secondfunction = OpLabel\n"
6331 "OpReturnValue %second_param1\n"
6334 fragments["testfun"] =
6335 // A %test_code function that returns its argument unchanged.
6336 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
6337 "%param1 = OpFunctionParameter %v4f32\n"
6338 "%label_testfun = OpLabel\n"
6339 "%val1 = OpFunctionCall %v4f32 %second_function %param1\n"
6340 "OpReturnValue %val1\n"
6343 createTestsForAllStages ("opmoduleprocessed", defaultColors, defaultColors, fragments, resources, noExtensions, opModuleProcessedTests.get());
6345 return opModuleProcessedTests.release();
6349 tcu::TestCaseGroup* createOpLineTests(tcu::TestContext& testCtx)
6351 RGBA defaultColors[4];
6352 de::MovePtr<tcu::TestCaseGroup> opLineTests (new tcu::TestCaseGroup(testCtx, "opline", "OpLine instruction"));
6353 map<string, string> fragments;
6354 std::vector<std::pair<std::string, std::string> > problemStrings;
6356 problemStrings.push_back(std::make_pair<std::string, std::string>("empty_name", ""));
6357 problemStrings.push_back(std::make_pair<std::string, std::string>("short_name", "short_name"));
6358 problemStrings.push_back(std::make_pair<std::string, std::string>("long_name", makeLongUTF8String(65530) + "ccc"));
6359 getDefaultColors(defaultColors);
6361 fragments["debug"] =
6362 "%other_name = OpString \"other_name\"\n";
6364 fragments["pre_main"] =
6365 "OpLine %file_name 32 0\n"
6366 "OpLine %file_name 32 32\n"
6367 "OpLine %file_name 32 40\n"
6368 "OpLine %other_name 32 40\n"
6369 "OpLine %other_name 0 100\n"
6370 "OpLine %other_name 0 4294967295\n"
6371 "OpLine %other_name 4294967295 0\n"
6372 "OpLine %other_name 32 40\n"
6373 "OpLine %file_name 0 0\n"
6374 "%second_function = OpFunction %v4f32 None %v4f32_v4f32_function\n"
6375 "OpLine %file_name 1 0\n"
6376 "%second_param1 = OpFunctionParameter %v4f32\n"
6377 "OpLine %file_name 1 3\n"
6378 "OpLine %file_name 1 2\n"
6379 "%label_secondfunction = OpLabel\n"
6380 "OpLine %file_name 0 2\n"
6381 "OpReturnValue %second_param1\n"
6383 "OpLine %file_name 0 2\n"
6384 "OpLine %file_name 0 2\n";
6386 fragments["testfun"] =
6387 // A %test_code function that returns its argument unchanged.
6388 "OpLine %file_name 1 0\n"
6389 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
6390 "OpLine %file_name 16 330\n"
6391 "%param1 = OpFunctionParameter %v4f32\n"
6392 "OpLine %file_name 14 442\n"
6393 "%label_testfun = OpLabel\n"
6394 "OpLine %file_name 11 1024\n"
6395 "%val1 = OpFunctionCall %v4f32 %second_function %param1\n"
6396 "OpLine %file_name 2 97\n"
6397 "OpReturnValue %val1\n"
6399 "OpLine %file_name 5 32\n";
6401 for (size_t i = 0; i < problemStrings.size(); ++i)
6403 map<string, string> testFragments = fragments;
6404 testFragments["debug"] += "%file_name = OpString \"" + problemStrings[i].second + "\"\n";
6405 createTestsForAllStages(string("opline") + "_" + problemStrings[i].first, defaultColors, defaultColors, testFragments, opLineTests.get());
6408 return opLineTests.release();
6411 tcu::TestCaseGroup* createOpConstantNullTests(tcu::TestContext& testCtx)
6413 de::MovePtr<tcu::TestCaseGroup> opConstantNullTests (new tcu::TestCaseGroup(testCtx, "opconstantnull", "OpConstantNull instruction"));
6417 const char functionStart[] =
6418 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
6419 "%param1 = OpFunctionParameter %v4f32\n"
6422 const char functionEnd[] =
6423 "OpReturnValue %transformed_param\n"
6426 struct NameConstantsCode
6433 NameConstantsCode tests[] =
6437 "%cnull = OpConstantNull %v4f32\n",
6438 "%transformed_param = OpFAdd %v4f32 %param1 %cnull\n"
6442 "%cnull = OpConstantNull %f32\n",
6443 "%vp = OpVariable %fp_v4f32 Function\n"
6444 "%v = OpLoad %v4f32 %vp\n"
6445 "%v0 = OpVectorInsertDynamic %v4f32 %v %cnull %c_i32_0\n"
6446 "%v1 = OpVectorInsertDynamic %v4f32 %v0 %cnull %c_i32_1\n"
6447 "%v2 = OpVectorInsertDynamic %v4f32 %v1 %cnull %c_i32_2\n"
6448 "%v3 = OpVectorInsertDynamic %v4f32 %v2 %cnull %c_i32_3\n"
6449 "%transformed_param = OpFAdd %v4f32 %param1 %v3\n"
6453 "%cnull = OpConstantNull %bool\n",
6454 "%v = OpVariable %fp_v4f32 Function\n"
6455 " OpStore %v %param1\n"
6456 " OpSelectionMerge %false_label None\n"
6457 " OpBranchConditional %cnull %true_label %false_label\n"
6458 "%true_label = OpLabel\n"
6459 " OpStore %v %c_v4f32_0_5_0_5_0_5_0_5\n"
6460 " OpBranch %false_label\n"
6461 "%false_label = OpLabel\n"
6462 "%transformed_param = OpLoad %v4f32 %v\n"
6466 "%cnull = OpConstantNull %i32\n",
6467 "%v = OpVariable %fp_v4f32 Function %c_v4f32_0_5_0_5_0_5_0_5\n"
6468 "%b = OpIEqual %bool %cnull %c_i32_0\n"
6469 " OpSelectionMerge %false_label None\n"
6470 " OpBranchConditional %b %true_label %false_label\n"
6471 "%true_label = OpLabel\n"
6472 " OpStore %v %param1\n"
6473 " OpBranch %false_label\n"
6474 "%false_label = OpLabel\n"
6475 "%transformed_param = OpLoad %v4f32 %v\n"
6479 "%stype = OpTypeStruct %f32 %v4f32\n"
6480 "%fp_stype = OpTypePointer Function %stype\n"
6481 "%cnull = OpConstantNull %stype\n",
6482 "%v = OpVariable %fp_stype Function %cnull\n"
6483 "%f = OpAccessChain %fp_v4f32 %v %c_i32_1\n"
6484 "%f_val = OpLoad %v4f32 %f\n"
6485 "%transformed_param = OpFAdd %v4f32 %param1 %f_val\n"
6489 "%a4_v4f32 = OpTypeArray %v4f32 %c_u32_4\n"
6490 "%fp_a4_v4f32 = OpTypePointer Function %a4_v4f32\n"
6491 "%cnull = OpConstantNull %a4_v4f32\n",
6492 "%v = OpVariable %fp_a4_v4f32 Function %cnull\n"
6493 "%f = OpAccessChain %fp_v4f32 %v %c_u32_0\n"
6494 "%f1 = OpAccessChain %fp_v4f32 %v %c_u32_1\n"
6495 "%f2 = OpAccessChain %fp_v4f32 %v %c_u32_2\n"
6496 "%f3 = OpAccessChain %fp_v4f32 %v %c_u32_3\n"
6497 "%f_val = OpLoad %v4f32 %f\n"
6498 "%f1_val = OpLoad %v4f32 %f1\n"
6499 "%f2_val = OpLoad %v4f32 %f2\n"
6500 "%f3_val = OpLoad %v4f32 %f3\n"
6501 "%t0 = OpFAdd %v4f32 %param1 %f_val\n"
6502 "%t1 = OpFAdd %v4f32 %t0 %f1_val\n"
6503 "%t2 = OpFAdd %v4f32 %t1 %f2_val\n"
6504 "%transformed_param = OpFAdd %v4f32 %t2 %f3_val\n"
6508 "%mat4x4_f32 = OpTypeMatrix %v4f32 4\n"
6509 "%cnull = OpConstantNull %mat4x4_f32\n",
6510 // Our null matrix * any vector should result in a zero vector.
6511 "%v = OpVectorTimesMatrix %v4f32 %param1 %cnull\n"
6512 "%transformed_param = OpFAdd %v4f32 %param1 %v\n"
6516 getHalfColorsFullAlpha(colors);
6518 for (size_t testNdx = 0; testNdx < sizeof(tests) / sizeof(NameConstantsCode); ++testNdx)
6520 map<string, string> fragments;
6521 fragments["pre_main"] = tests[testNdx].constants;
6522 fragments["testfun"] = string(functionStart) + tests[testNdx].code + functionEnd;
6523 createTestsForAllStages(tests[testNdx].name, colors, colors, fragments, opConstantNullTests.get());
6525 return opConstantNullTests.release();
6527 tcu::TestCaseGroup* createOpConstantCompositeTests(tcu::TestContext& testCtx)
6529 de::MovePtr<tcu::TestCaseGroup> opConstantCompositeTests (new tcu::TestCaseGroup(testCtx, "opconstantcomposite", "OpConstantComposite instruction"));
6530 RGBA inputColors[4];
6531 RGBA outputColors[4];
6534 const char functionStart[] =
6535 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
6536 "%param1 = OpFunctionParameter %v4f32\n"
6539 const char functionEnd[] =
6540 "OpReturnValue %transformed_param\n"
6543 struct NameConstantsCode
6550 NameConstantsCode tests[] =
6555 "%cval = OpConstantComposite %v4f32 %c_f32_0_5 %c_f32_0_5 %c_f32_0_5 %c_f32_0\n",
6556 "%transformed_param = OpFAdd %v4f32 %param1 %cval\n"
6561 "%stype = OpTypeStruct %v4f32 %f32\n"
6562 "%fp_stype = OpTypePointer Function %stype\n"
6563 "%f32_n_1 = OpConstant %f32 -1.0\n"
6564 "%f32_1_5 = OpConstant %f32 !0x3fc00000\n" // +1.5
6565 "%cvec = OpConstantComposite %v4f32 %f32_1_5 %f32_1_5 %f32_1_5 %c_f32_1\n"
6566 "%cval = OpConstantComposite %stype %cvec %f32_n_1\n",
6568 "%v = OpVariable %fp_stype Function %cval\n"
6569 "%vec_ptr = OpAccessChain %fp_v4f32 %v %c_u32_0\n"
6570 "%f32_ptr = OpAccessChain %fp_f32 %v %c_u32_1\n"
6571 "%vec_val = OpLoad %v4f32 %vec_ptr\n"
6572 "%f32_val = OpLoad %f32 %f32_ptr\n"
6573 "%tmp1 = OpVectorTimesScalar %v4f32 %c_v4f32_1_1_1_1 %f32_val\n" // vec4(-1)
6574 "%tmp2 = OpFAdd %v4f32 %tmp1 %param1\n" // param1 + vec4(-1)
6575 "%transformed_param = OpFAdd %v4f32 %tmp2 %vec_val\n" // param1 + vec4(-1) + vec4(1.5, 1.5, 1.5, 1.0)
6578 // [1|0|0|0.5] [x] = x + 0.5
6579 // [0|1|0|0.5] [y] = y + 0.5
6580 // [0|0|1|0.5] [z] = z + 0.5
6581 // [0|0|0|1 ] [1] = 1
6584 "%mat4x4_f32 = OpTypeMatrix %v4f32 4\n"
6585 "%v4f32_1_0_0_0 = OpConstantComposite %v4f32 %c_f32_1 %c_f32_0 %c_f32_0 %c_f32_0\n"
6586 "%v4f32_0_1_0_0 = OpConstantComposite %v4f32 %c_f32_0 %c_f32_1 %c_f32_0 %c_f32_0\n"
6587 "%v4f32_0_0_1_0 = OpConstantComposite %v4f32 %c_f32_0 %c_f32_0 %c_f32_1 %c_f32_0\n"
6588 "%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"
6589 "%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",
6591 "%transformed_param = OpMatrixTimesVector %v4f32 %cval %param1\n"
6596 "%c_v4f32_1_1_1_0 = OpConstantComposite %v4f32 %c_f32_1 %c_f32_1 %c_f32_1 %c_f32_0\n"
6597 "%fp_a4f32 = OpTypePointer Function %a4f32\n"
6598 "%f32_n_1 = OpConstant %f32 -1.0\n"
6599 "%f32_1_5 = OpConstant %f32 !0x3fc00000\n" // +1.5
6600 "%carr = OpConstantComposite %a4f32 %c_f32_0 %f32_n_1 %f32_1_5 %c_f32_0\n",
6602 "%v = OpVariable %fp_a4f32 Function %carr\n"
6603 "%f = OpAccessChain %fp_f32 %v %c_u32_0\n"
6604 "%f1 = OpAccessChain %fp_f32 %v %c_u32_1\n"
6605 "%f2 = OpAccessChain %fp_f32 %v %c_u32_2\n"
6606 "%f3 = OpAccessChain %fp_f32 %v %c_u32_3\n"
6607 "%f_val = OpLoad %f32 %f\n"
6608 "%f1_val = OpLoad %f32 %f1\n"
6609 "%f2_val = OpLoad %f32 %f2\n"
6610 "%f3_val = OpLoad %f32 %f3\n"
6611 "%ftot1 = OpFAdd %f32 %f_val %f1_val\n"
6612 "%ftot2 = OpFAdd %f32 %ftot1 %f2_val\n"
6613 "%ftot3 = OpFAdd %f32 %ftot2 %f3_val\n" // 0 - 1 + 1.5 + 0
6614 "%add_vec = OpVectorTimesScalar %v4f32 %c_v4f32_1_1_1_0 %ftot3\n"
6615 "%transformed_param = OpFAdd %v4f32 %param1 %add_vec\n"
6622 // [ 1.0, 1.0, 1.0, 1.0]
6626 // [ 0.0, 0.5, 0.0, 0.0]
6630 // [ 1.0, 1.0, 1.0, 1.0]
6633 "array_of_struct_of_array",
6635 "%c_v4f32_1_1_1_0 = OpConstantComposite %v4f32 %c_f32_1 %c_f32_1 %c_f32_1 %c_f32_0\n"
6636 "%fp_a4f32 = OpTypePointer Function %a4f32\n"
6637 "%stype = OpTypeStruct %f32 %a4f32\n"
6638 "%a3stype = OpTypeArray %stype %c_u32_3\n"
6639 "%fp_a3stype = OpTypePointer Function %a3stype\n"
6640 "%ca4f32_0 = OpConstantComposite %a4f32 %c_f32_0 %c_f32_0_5 %c_f32_0 %c_f32_0\n"
6641 "%ca4f32_1 = OpConstantComposite %a4f32 %c_f32_1 %c_f32_1 %c_f32_1 %c_f32_1\n"
6642 "%cstype1 = OpConstantComposite %stype %c_f32_0 %ca4f32_1\n"
6643 "%cstype2 = OpConstantComposite %stype %c_f32_1 %ca4f32_0\n"
6644 "%carr = OpConstantComposite %a3stype %cstype1 %cstype2 %cstype1",
6646 "%v = OpVariable %fp_a3stype Function %carr\n"
6647 "%f = OpAccessChain %fp_f32 %v %c_u32_1 %c_u32_1 %c_u32_1\n"
6648 "%f_l = OpLoad %f32 %f\n"
6649 "%add_vec = OpVectorTimesScalar %v4f32 %c_v4f32_1_1_1_0 %f_l\n"
6650 "%transformed_param = OpFAdd %v4f32 %param1 %add_vec\n"
6654 getHalfColorsFullAlpha(inputColors);
6655 outputColors[0] = RGBA(255, 255, 255, 255);
6656 outputColors[1] = RGBA(255, 127, 127, 255);
6657 outputColors[2] = RGBA(127, 255, 127, 255);
6658 outputColors[3] = RGBA(127, 127, 255, 255);
6660 for (size_t testNdx = 0; testNdx < sizeof(tests) / sizeof(NameConstantsCode); ++testNdx)
6662 map<string, string> fragments;
6663 fragments["pre_main"] = tests[testNdx].constants;
6664 fragments["testfun"] = string(functionStart) + tests[testNdx].code + functionEnd;
6665 createTestsForAllStages(tests[testNdx].name, inputColors, outputColors, fragments, opConstantCompositeTests.get());
6667 return opConstantCompositeTests.release();
6670 tcu::TestCaseGroup* createSelectionBlockOrderTests(tcu::TestContext& testCtx)
6672 de::MovePtr<tcu::TestCaseGroup> group (new tcu::TestCaseGroup(testCtx, "selection_block_order", "Out-of-order blocks for selection"));
6673 RGBA inputColors[4];
6674 RGBA outputColors[4];
6675 map<string, string> fragments;
6677 // vec4 test_code(vec4 param) {
6678 // vec4 result = param;
6679 // for (int i = 0; i < 4; ++i) {
6680 // if (i == 0) result[i] = 0.;
6681 // else result[i] = 1. - result[i];
6685 const char function[] =
6686 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
6687 "%param1 = OpFunctionParameter %v4f32\n"
6689 "%iptr = OpVariable %fp_i32 Function\n"
6690 "%result = OpVariable %fp_v4f32 Function\n"
6691 " OpStore %iptr %c_i32_0\n"
6692 " OpStore %result %param1\n"
6695 // Loop entry block.
6697 "%ival = OpLoad %i32 %iptr\n"
6698 "%lt_4 = OpSLessThan %bool %ival %c_i32_4\n"
6699 " OpLoopMerge %exit %if_entry None\n"
6700 " OpBranchConditional %lt_4 %if_entry %exit\n"
6702 // Merge block for loop.
6704 "%ret = OpLoad %v4f32 %result\n"
6705 " OpReturnValue %ret\n"
6707 // If-statement entry block.
6708 "%if_entry = OpLabel\n"
6709 "%loc = OpAccessChain %fp_f32 %result %ival\n"
6710 "%eq_0 = OpIEqual %bool %ival %c_i32_0\n"
6711 " OpSelectionMerge %if_exit None\n"
6712 " OpBranchConditional %eq_0 %if_true %if_false\n"
6714 // False branch for if-statement.
6715 "%if_false = OpLabel\n"
6716 "%val = OpLoad %f32 %loc\n"
6717 "%sub = OpFSub %f32 %c_f32_1 %val\n"
6718 " OpStore %loc %sub\n"
6719 " OpBranch %if_exit\n"
6721 // Merge block for if-statement.
6722 "%if_exit = OpLabel\n"
6723 "%ival_next = OpIAdd %i32 %ival %c_i32_1\n"
6724 " OpStore %iptr %ival_next\n"
6727 // True branch for if-statement.
6728 "%if_true = OpLabel\n"
6729 " OpStore %loc %c_f32_0\n"
6730 " OpBranch %if_exit\n"
6734 fragments["testfun"] = function;
6736 inputColors[0] = RGBA(127, 127, 127, 0);
6737 inputColors[1] = RGBA(127, 0, 0, 0);
6738 inputColors[2] = RGBA(0, 127, 0, 0);
6739 inputColors[3] = RGBA(0, 0, 127, 0);
6741 outputColors[0] = RGBA(0, 128, 128, 255);
6742 outputColors[1] = RGBA(0, 255, 255, 255);
6743 outputColors[2] = RGBA(0, 128, 255, 255);
6744 outputColors[3] = RGBA(0, 255, 128, 255);
6746 createTestsForAllStages("out_of_order", inputColors, outputColors, fragments, group.get());
6748 return group.release();
6751 tcu::TestCaseGroup* createSwitchBlockOrderTests(tcu::TestContext& testCtx)
6753 de::MovePtr<tcu::TestCaseGroup> group (new tcu::TestCaseGroup(testCtx, "switch_block_order", "Out-of-order blocks for switch"));
6754 RGBA inputColors[4];
6755 RGBA outputColors[4];
6756 map<string, string> fragments;
6758 const char typesAndConstants[] =
6759 "%c_f32_p2 = OpConstant %f32 0.2\n"
6760 "%c_f32_p4 = OpConstant %f32 0.4\n"
6761 "%c_f32_p6 = OpConstant %f32 0.6\n"
6762 "%c_f32_p8 = OpConstant %f32 0.8\n";
6764 // vec4 test_code(vec4 param) {
6765 // vec4 result = param;
6766 // for (int i = 0; i < 4; ++i) {
6768 // case 0: result[i] += .2; break;
6769 // case 1: result[i] += .6; break;
6770 // case 2: result[i] += .4; break;
6771 // case 3: result[i] += .8; break;
6772 // default: break; // unreachable
6777 const char function[] =
6778 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
6779 "%param1 = OpFunctionParameter %v4f32\n"
6781 "%iptr = OpVariable %fp_i32 Function\n"
6782 "%result = OpVariable %fp_v4f32 Function\n"
6783 " OpStore %iptr %c_i32_0\n"
6784 " OpStore %result %param1\n"
6787 // Loop entry block.
6789 "%ival = OpLoad %i32 %iptr\n"
6790 "%lt_4 = OpSLessThan %bool %ival %c_i32_4\n"
6791 " OpLoopMerge %exit %switch_exit None\n"
6792 " OpBranchConditional %lt_4 %switch_entry %exit\n"
6794 // Merge block for loop.
6796 "%ret = OpLoad %v4f32 %result\n"
6797 " OpReturnValue %ret\n"
6799 // Switch-statement entry block.
6800 "%switch_entry = OpLabel\n"
6801 "%loc = OpAccessChain %fp_f32 %result %ival\n"
6802 "%val = OpLoad %f32 %loc\n"
6803 " OpSelectionMerge %switch_exit None\n"
6804 " OpSwitch %ival %switch_default 0 %case0 1 %case1 2 %case2 3 %case3\n"
6806 "%case2 = OpLabel\n"
6807 "%addp4 = OpFAdd %f32 %val %c_f32_p4\n"
6808 " OpStore %loc %addp4\n"
6809 " OpBranch %switch_exit\n"
6811 "%switch_default = OpLabel\n"
6814 "%case3 = OpLabel\n"
6815 "%addp8 = OpFAdd %f32 %val %c_f32_p8\n"
6816 " OpStore %loc %addp8\n"
6817 " OpBranch %switch_exit\n"
6819 "%case0 = OpLabel\n"
6820 "%addp2 = OpFAdd %f32 %val %c_f32_p2\n"
6821 " OpStore %loc %addp2\n"
6822 " OpBranch %switch_exit\n"
6824 // Merge block for switch-statement.
6825 "%switch_exit = OpLabel\n"
6826 "%ival_next = OpIAdd %i32 %ival %c_i32_1\n"
6827 " OpStore %iptr %ival_next\n"
6830 "%case1 = OpLabel\n"
6831 "%addp6 = OpFAdd %f32 %val %c_f32_p6\n"
6832 " OpStore %loc %addp6\n"
6833 " OpBranch %switch_exit\n"
6837 fragments["pre_main"] = typesAndConstants;
6838 fragments["testfun"] = function;
6840 inputColors[0] = RGBA(127, 27, 127, 51);
6841 inputColors[1] = RGBA(127, 0, 0, 51);
6842 inputColors[2] = RGBA(0, 27, 0, 51);
6843 inputColors[3] = RGBA(0, 0, 127, 51);
6845 outputColors[0] = RGBA(178, 180, 229, 255);
6846 outputColors[1] = RGBA(178, 153, 102, 255);
6847 outputColors[2] = RGBA(51, 180, 102, 255);
6848 outputColors[3] = RGBA(51, 153, 229, 255);
6850 createTestsForAllStages("out_of_order", inputColors, outputColors, fragments, group.get());
6852 return group.release();
6855 tcu::TestCaseGroup* createDecorationGroupTests(tcu::TestContext& testCtx)
6857 de::MovePtr<tcu::TestCaseGroup> group (new tcu::TestCaseGroup(testCtx, "decoration_group", "Decoration group tests"));
6858 RGBA inputColors[4];
6859 RGBA outputColors[4];
6860 map<string, string> fragments;
6862 const char decorations[] =
6863 "OpDecorate %array_group ArrayStride 4\n"
6864 "OpDecorate %struct_member_group Offset 0\n"
6865 "%array_group = OpDecorationGroup\n"
6866 "%struct_member_group = OpDecorationGroup\n"
6868 "OpDecorate %group1 RelaxedPrecision\n"
6869 "OpDecorate %group3 RelaxedPrecision\n"
6870 "OpDecorate %group3 Invariant\n"
6871 "OpDecorate %group3 Restrict\n"
6872 "%group0 = OpDecorationGroup\n"
6873 "%group1 = OpDecorationGroup\n"
6874 "%group3 = OpDecorationGroup\n";
6876 const char typesAndConstants[] =
6877 "%a3f32 = OpTypeArray %f32 %c_u32_3\n"
6878 "%struct1 = OpTypeStruct %a3f32\n"
6879 "%struct2 = OpTypeStruct %a3f32\n"
6880 "%fp_struct1 = OpTypePointer Function %struct1\n"
6881 "%fp_struct2 = OpTypePointer Function %struct2\n"
6882 "%c_f32_2 = OpConstant %f32 2.\n"
6883 "%c_f32_n2 = OpConstant %f32 -2.\n"
6885 "%c_a3f32_1 = OpConstantComposite %a3f32 %c_f32_1 %c_f32_2 %c_f32_1\n"
6886 "%c_a3f32_2 = OpConstantComposite %a3f32 %c_f32_n1 %c_f32_n2 %c_f32_n1\n"
6887 "%c_struct1 = OpConstantComposite %struct1 %c_a3f32_1\n"
6888 "%c_struct2 = OpConstantComposite %struct2 %c_a3f32_2\n";
6890 const char function[] =
6891 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
6892 "%param = OpFunctionParameter %v4f32\n"
6893 "%entry = OpLabel\n"
6894 "%result = OpVariable %fp_v4f32 Function\n"
6895 "%v_struct1 = OpVariable %fp_struct1 Function\n"
6896 "%v_struct2 = OpVariable %fp_struct2 Function\n"
6897 " OpStore %result %param\n"
6898 " OpStore %v_struct1 %c_struct1\n"
6899 " OpStore %v_struct2 %c_struct2\n"
6900 "%ptr1 = OpAccessChain %fp_f32 %v_struct1 %c_i32_0 %c_i32_2\n"
6901 "%val1 = OpLoad %f32 %ptr1\n"
6902 "%ptr2 = OpAccessChain %fp_f32 %v_struct2 %c_i32_0 %c_i32_2\n"
6903 "%val2 = OpLoad %f32 %ptr2\n"
6904 "%addvalues = OpFAdd %f32 %val1 %val2\n"
6905 "%ptr = OpAccessChain %fp_f32 %result %c_i32_1\n"
6906 "%val = OpLoad %f32 %ptr\n"
6907 "%addresult = OpFAdd %f32 %addvalues %val\n"
6908 " OpStore %ptr %addresult\n"
6909 "%ret = OpLoad %v4f32 %result\n"
6910 " OpReturnValue %ret\n"
6913 struct CaseNameDecoration
6919 CaseNameDecoration tests[] =
6922 "same_decoration_group_on_multiple_types",
6923 "OpGroupMemberDecorate %struct_member_group %struct1 0 %struct2 0\n"
6926 "empty_decoration_group",
6927 "OpGroupDecorate %group0 %a3f32\n"
6928 "OpGroupDecorate %group0 %result\n"
6931 "one_element_decoration_group",
6932 "OpGroupDecorate %array_group %a3f32\n"
6935 "multiple_elements_decoration_group",
6936 "OpGroupDecorate %group3 %v_struct1\n"
6939 "multiple_decoration_groups_on_same_variable",
6940 "OpGroupDecorate %group0 %v_struct2\n"
6941 "OpGroupDecorate %group1 %v_struct2\n"
6942 "OpGroupDecorate %group3 %v_struct2\n"
6945 "same_decoration_group_multiple_times",
6946 "OpGroupDecorate %group1 %addvalues\n"
6947 "OpGroupDecorate %group1 %addvalues\n"
6948 "OpGroupDecorate %group1 %addvalues\n"
6953 getHalfColorsFullAlpha(inputColors);
6954 getHalfColorsFullAlpha(outputColors);
6956 for (size_t idx = 0; idx < (sizeof(tests) / sizeof(tests[0])); ++idx)
6958 fragments["decoration"] = decorations + tests[idx].decoration;
6959 fragments["pre_main"] = typesAndConstants;
6960 fragments["testfun"] = function;
6962 createTestsForAllStages(tests[idx].name, inputColors, outputColors, fragments, group.get());
6965 return group.release();
6968 struct SpecConstantTwoIntGraphicsCase
6970 const char* caseName;
6971 const char* scDefinition0;
6972 const char* scDefinition1;
6973 const char* scResultType;
6974 const char* scOperation;
6975 deInt32 scActualValue0;
6976 deInt32 scActualValue1;
6977 const char* resultOperation;
6978 RGBA expectedColors[4];
6979 deInt32 scActualValueLength;
6981 SpecConstantTwoIntGraphicsCase (const char* name,
6982 const char* definition0,
6983 const char* definition1,
6984 const char* resultType,
6985 const char* operation,
6986 const deInt32 value0,
6987 const deInt32 value1,
6988 const char* resultOp,
6989 const RGBA (&output)[4],
6990 const deInt32 valueLength = sizeof(deInt32))
6992 , scDefinition0 (definition0)
6993 , scDefinition1 (definition1)
6994 , scResultType (resultType)
6995 , scOperation (operation)
6996 , scActualValue0 (value0)
6997 , scActualValue1 (value1)
6998 , resultOperation (resultOp)
6999 , scActualValueLength (valueLength)
7001 expectedColors[0] = output[0];
7002 expectedColors[1] = output[1];
7003 expectedColors[2] = output[2];
7004 expectedColors[3] = output[3];
7008 tcu::TestCaseGroup* createSpecConstantTests (tcu::TestContext& testCtx)
7010 de::MovePtr<tcu::TestCaseGroup> group (new tcu::TestCaseGroup(testCtx, "opspecconstantop", "Test the OpSpecConstantOp instruction"));
7011 vector<SpecConstantTwoIntGraphicsCase> cases;
7012 RGBA inputColors[4];
7013 RGBA outputColors0[4];
7014 RGBA outputColors1[4];
7015 RGBA outputColors2[4];
7017 const deInt32 m1AsFloat16 = 0xbc00; // -1(fp16) == 1 01111 0000000000 == 1011 1100 0000 0000
7019 const char decorations1[] =
7020 "OpDecorate %sc_0 SpecId 0\n"
7021 "OpDecorate %sc_1 SpecId 1\n";
7023 const char typesAndConstants1[] =
7024 "${OPTYPE_DEFINITIONS:opt}"
7025 "%sc_0 = OpSpecConstant${SC_DEF0}\n"
7026 "%sc_1 = OpSpecConstant${SC_DEF1}\n"
7027 "%sc_op = OpSpecConstantOp ${SC_RESULT_TYPE} ${SC_OP}\n";
7029 const char function1[] =
7030 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
7031 "%param = OpFunctionParameter %v4f32\n"
7032 "%label = OpLabel\n"
7033 "%result = OpVariable %fp_v4f32 Function\n"
7034 "${TYPE_CONVERT:opt}"
7035 " OpStore %result %param\n"
7036 "%gen = ${GEN_RESULT}\n"
7037 "%index = OpIAdd %i32 %gen %c_i32_1\n"
7038 "%loc = OpAccessChain %fp_f32 %result %index\n"
7039 "%val = OpLoad %f32 %loc\n"
7040 "%add = OpFAdd %f32 %val %c_f32_0_5\n"
7041 " OpStore %loc %add\n"
7042 "%ret = OpLoad %v4f32 %result\n"
7043 " OpReturnValue %ret\n"
7046 inputColors[0] = RGBA(127, 127, 127, 255);
7047 inputColors[1] = RGBA(127, 0, 0, 255);
7048 inputColors[2] = RGBA(0, 127, 0, 255);
7049 inputColors[3] = RGBA(0, 0, 127, 255);
7051 // Derived from inputColors[x] by adding 128 to inputColors[x][0].
7052 outputColors0[0] = RGBA(255, 127, 127, 255);
7053 outputColors0[1] = RGBA(255, 0, 0, 255);
7054 outputColors0[2] = RGBA(128, 127, 0, 255);
7055 outputColors0[3] = RGBA(128, 0, 127, 255);
7057 // Derived from inputColors[x] by adding 128 to inputColors[x][1].
7058 outputColors1[0] = RGBA(127, 255, 127, 255);
7059 outputColors1[1] = RGBA(127, 128, 0, 255);
7060 outputColors1[2] = RGBA(0, 255, 0, 255);
7061 outputColors1[3] = RGBA(0, 128, 127, 255);
7063 // Derived from inputColors[x] by adding 128 to inputColors[x][2].
7064 outputColors2[0] = RGBA(127, 127, 255, 255);
7065 outputColors2[1] = RGBA(127, 0, 128, 255);
7066 outputColors2[2] = RGBA(0, 127, 128, 255);
7067 outputColors2[3] = RGBA(0, 0, 255, 255);
7069 const char addZeroToSc[] = "OpIAdd %i32 %c_i32_0 %sc_op";
7070 const char addZeroToSc32[] = "OpIAdd %i32 %c_i32_0 %sc_op32";
7071 const char selectTrueUsingSc[] = "OpSelect %i32 %sc_op %c_i32_1 %c_i32_0";
7072 const char selectFalseUsingSc[] = "OpSelect %i32 %sc_op %c_i32_0 %c_i32_1";
7074 cases.push_back(SpecConstantTwoIntGraphicsCase("iadd", " %i32 0", " %i32 0", "%i32", "IAdd %sc_0 %sc_1", 19, -20, addZeroToSc, outputColors0));
7075 cases.push_back(SpecConstantTwoIntGraphicsCase("isub", " %i32 0", " %i32 0", "%i32", "ISub %sc_0 %sc_1", 19, 20, addZeroToSc, outputColors0));
7076 cases.push_back(SpecConstantTwoIntGraphicsCase("imul", " %i32 0", " %i32 0", "%i32", "IMul %sc_0 %sc_1", -1, -1, addZeroToSc, outputColors2));
7077 cases.push_back(SpecConstantTwoIntGraphicsCase("sdiv", " %i32 0", " %i32 0", "%i32", "SDiv %sc_0 %sc_1", -126, 126, addZeroToSc, outputColors0));
7078 cases.push_back(SpecConstantTwoIntGraphicsCase("udiv", " %i32 0", " %i32 0", "%i32", "UDiv %sc_0 %sc_1", 126, 126, addZeroToSc, outputColors2));
7079 cases.push_back(SpecConstantTwoIntGraphicsCase("srem", " %i32 0", " %i32 0", "%i32", "SRem %sc_0 %sc_1", 3, 2, addZeroToSc, outputColors2));
7080 cases.push_back(SpecConstantTwoIntGraphicsCase("smod", " %i32 0", " %i32 0", "%i32", "SMod %sc_0 %sc_1", 3, 2, addZeroToSc, outputColors2));
7081 cases.push_back(SpecConstantTwoIntGraphicsCase("umod", " %i32 0", " %i32 0", "%i32", "UMod %sc_0 %sc_1", 1001, 500, addZeroToSc, outputColors2));
7082 cases.push_back(SpecConstantTwoIntGraphicsCase("bitwiseand", " %i32 0", " %i32 0", "%i32", "BitwiseAnd %sc_0 %sc_1", 0x33, 0x0d, addZeroToSc, outputColors2));
7083 cases.push_back(SpecConstantTwoIntGraphicsCase("bitwiseor", " %i32 0", " %i32 0", "%i32", "BitwiseOr %sc_0 %sc_1", 0, 1, addZeroToSc, outputColors2));
7084 cases.push_back(SpecConstantTwoIntGraphicsCase("bitwisexor", " %i32 0", " %i32 0", "%i32", "BitwiseXor %sc_0 %sc_1", 0x2e, 0x2f, addZeroToSc, outputColors2));
7085 cases.push_back(SpecConstantTwoIntGraphicsCase("shiftrightlogical", " %i32 0", " %i32 0", "%i32", "ShiftRightLogical %sc_0 %sc_1", 2, 1, addZeroToSc, outputColors2));
7086 cases.push_back(SpecConstantTwoIntGraphicsCase("shiftrightarithmetic", " %i32 0", " %i32 0", "%i32", "ShiftRightArithmetic %sc_0 %sc_1", -4, 2, addZeroToSc, outputColors0));
7087 cases.push_back(SpecConstantTwoIntGraphicsCase("shiftleftlogical", " %i32 0", " %i32 0", "%i32", "ShiftLeftLogical %sc_0 %sc_1", 1, 0, addZeroToSc, outputColors2));
7088 cases.push_back(SpecConstantTwoIntGraphicsCase("slessthan", " %i32 0", " %i32 0", "%bool", "SLessThan %sc_0 %sc_1", -20, -10, selectTrueUsingSc, outputColors2));
7089 cases.push_back(SpecConstantTwoIntGraphicsCase("ulessthan", " %i32 0", " %i32 0", "%bool", "ULessThan %sc_0 %sc_1", 10, 20, selectTrueUsingSc, outputColors2));
7090 cases.push_back(SpecConstantTwoIntGraphicsCase("sgreaterthan", " %i32 0", " %i32 0", "%bool", "SGreaterThan %sc_0 %sc_1", -1000, 50, selectFalseUsingSc, outputColors2));
7091 cases.push_back(SpecConstantTwoIntGraphicsCase("ugreaterthan", " %i32 0", " %i32 0", "%bool", "UGreaterThan %sc_0 %sc_1", 10, 5, selectTrueUsingSc, outputColors2));
7092 cases.push_back(SpecConstantTwoIntGraphicsCase("slessthanequal", " %i32 0", " %i32 0", "%bool", "SLessThanEqual %sc_0 %sc_1", -10, -10, selectTrueUsingSc, outputColors2));
7093 cases.push_back(SpecConstantTwoIntGraphicsCase("ulessthanequal", " %i32 0", " %i32 0", "%bool", "ULessThanEqual %sc_0 %sc_1", 50, 100, selectTrueUsingSc, outputColors2));
7094 cases.push_back(SpecConstantTwoIntGraphicsCase("sgreaterthanequal", " %i32 0", " %i32 0", "%bool", "SGreaterThanEqual %sc_0 %sc_1", -1000, 50, selectFalseUsingSc, outputColors2));
7095 cases.push_back(SpecConstantTwoIntGraphicsCase("ugreaterthanequal", " %i32 0", " %i32 0", "%bool", "UGreaterThanEqual %sc_0 %sc_1", 10, 10, selectTrueUsingSc, outputColors2));
7096 cases.push_back(SpecConstantTwoIntGraphicsCase("iequal", " %i32 0", " %i32 0", "%bool", "IEqual %sc_0 %sc_1", 42, 24, selectFalseUsingSc, outputColors2));
7097 cases.push_back(SpecConstantTwoIntGraphicsCase("inotequal", " %i32 0", " %i32 0", "%bool", "INotEqual %sc_0 %sc_1", 42, 24, selectTrueUsingSc, outputColors2));
7098 cases.push_back(SpecConstantTwoIntGraphicsCase("logicaland", "True %bool", "True %bool", "%bool", "LogicalAnd %sc_0 %sc_1", 0, 1, selectFalseUsingSc, outputColors2));
7099 cases.push_back(SpecConstantTwoIntGraphicsCase("logicalor", "False %bool", "False %bool", "%bool", "LogicalOr %sc_0 %sc_1", 1, 0, selectTrueUsingSc, outputColors2));
7100 cases.push_back(SpecConstantTwoIntGraphicsCase("logicalequal", "True %bool", "True %bool", "%bool", "LogicalEqual %sc_0 %sc_1", 0, 1, selectFalseUsingSc, outputColors2));
7101 cases.push_back(SpecConstantTwoIntGraphicsCase("logicalnotequal", "False %bool", "False %bool", "%bool", "LogicalNotEqual %sc_0 %sc_1", 1, 0, selectTrueUsingSc, outputColors2));
7102 cases.push_back(SpecConstantTwoIntGraphicsCase("snegate", " %i32 0", " %i32 0", "%i32", "SNegate %sc_0", -1, 0, addZeroToSc, outputColors2));
7103 cases.push_back(SpecConstantTwoIntGraphicsCase("not", " %i32 0", " %i32 0", "%i32", "Not %sc_0", -2, 0, addZeroToSc, outputColors2));
7104 cases.push_back(SpecConstantTwoIntGraphicsCase("logicalnot", "False %bool", "False %bool", "%bool", "LogicalNot %sc_0", 1, 0, selectFalseUsingSc, outputColors2));
7105 cases.push_back(SpecConstantTwoIntGraphicsCase("select", "False %bool", " %i32 0", "%i32", "Select %sc_0 %sc_1 %c_i32_0", 1, 1, addZeroToSc, outputColors2));
7106 cases.push_back(SpecConstantTwoIntGraphicsCase("sconvert", " %i32 0", " %i32 0", "%i16", "SConvert %sc_0", -1, 0, addZeroToSc32, outputColors0));
7107 // -1082130432 stored as 32-bit two's complement is the binary representation of -1 as IEEE-754 Float
7108 cases.push_back(SpecConstantTwoIntGraphicsCase("fconvert", " %f32 0", " %f32 0", "%f64", "FConvert %sc_0", -1082130432, 0, addZeroToSc32, outputColors0));
7109 cases.push_back(SpecConstantTwoIntGraphicsCase("fconvert16", " %f16 0", " %f16 0", "%f32", "FConvert %sc_0", m1AsFloat16, 0, addZeroToSc32, outputColors0, sizeof(deFloat16)));
7110 // \todo[2015-12-1 antiagainst] OpQuantizeToF16
7112 for (size_t caseNdx = 0; caseNdx < cases.size(); ++caseNdx)
7114 map<string, string> specializations;
7115 map<string, string> fragments;
7116 SpecConstants specConstants;
7117 PushConstants noPushConstants;
7118 GraphicsResources noResources;
7119 GraphicsInterfaces noInterfaces;
7120 vector<string> extensions;
7121 VulkanFeatures requiredFeatures;
7123 // Special SPIR-V code for SConvert-case
7124 if (strcmp(cases[caseNdx].caseName, "sconvert") == 0)
7126 requiredFeatures.coreFeatures.shaderInt16 = VK_TRUE;
7127 fragments["capability"] = "OpCapability Int16\n"; // Adds 16-bit integer capability
7128 specializations["OPTYPE_DEFINITIONS"] = "%i16 = OpTypeInt 16 1\n"; // Adds 16-bit integer type
7129 specializations["TYPE_CONVERT"] = "%sc_op32 = OpSConvert %i32 %sc_op\n"; // Converts 16-bit integer to 32-bit integer
7132 // Special SPIR-V code for FConvert-case
7133 if (strcmp(cases[caseNdx].caseName, "fconvert") == 0)
7135 requiredFeatures.coreFeatures.shaderFloat64 = VK_TRUE;
7136 fragments["capability"] = "OpCapability Float64\n"; // Adds 64-bit float capability
7137 specializations["OPTYPE_DEFINITIONS"] = "%f64 = OpTypeFloat 64\n"; // Adds 64-bit float type
7138 specializations["TYPE_CONVERT"] = "%sc_op32 = OpConvertFToS %i32 %sc_op\n"; // Converts 64-bit float to 32-bit integer
7141 // Special SPIR-V code for FConvert-case for 16-bit floats
7142 if (strcmp(cases[caseNdx].caseName, "fconvert16") == 0)
7144 extensions.push_back("VK_KHR_shader_float16_int8");
7145 requiredFeatures.extFloat16Int8 = EXTFLOAT16INT8FEATURES_FLOAT16;
7146 fragments["capability"] = "OpCapability Float16\n"; // Adds 16-bit float capability
7147 specializations["OPTYPE_DEFINITIONS"] = "%f16 = OpTypeFloat 16\n"; // Adds 16-bit float type
7148 specializations["TYPE_CONVERT"] = "%sc_op32 = OpConvertFToS %i32 %sc_op\n"; // Converts 16-bit float to 32-bit integer
7151 specializations["SC_DEF0"] = cases[caseNdx].scDefinition0;
7152 specializations["SC_DEF1"] = cases[caseNdx].scDefinition1;
7153 specializations["SC_RESULT_TYPE"] = cases[caseNdx].scResultType;
7154 specializations["SC_OP"] = cases[caseNdx].scOperation;
7155 specializations["GEN_RESULT"] = cases[caseNdx].resultOperation;
7157 fragments["decoration"] = tcu::StringTemplate(decorations1).specialize(specializations);
7158 fragments["pre_main"] = tcu::StringTemplate(typesAndConstants1).specialize(specializations);
7159 fragments["testfun"] = tcu::StringTemplate(function1).specialize(specializations);
7161 specConstants.append(&cases[caseNdx].scActualValue0, cases[caseNdx].scActualValueLength);
7162 specConstants.append(&cases[caseNdx].scActualValue1, cases[caseNdx].scActualValueLength);
7164 createTestsForAllStages(
7165 cases[caseNdx].caseName, inputColors, cases[caseNdx].expectedColors, fragments, specConstants,
7166 noPushConstants, noResources, noInterfaces, extensions, requiredFeatures, group.get());
7169 const char decorations2[] =
7170 "OpDecorate %sc_0 SpecId 0\n"
7171 "OpDecorate %sc_1 SpecId 1\n"
7172 "OpDecorate %sc_2 SpecId 2\n";
7174 const char typesAndConstants2[] =
7175 "%vec3_0 = OpConstantComposite %v3i32 %c_i32_0 %c_i32_0 %c_i32_0\n"
7176 "%vec3_undef = OpUndef %v3i32\n"
7178 "%sc_0 = OpSpecConstant %i32 0\n"
7179 "%sc_1 = OpSpecConstant %i32 0\n"
7180 "%sc_2 = OpSpecConstant %i32 0\n"
7181 "%sc_vec3_0 = OpSpecConstantOp %v3i32 CompositeInsert %sc_0 %vec3_0 0\n" // (sc_0, 0, 0)
7182 "%sc_vec3_1 = OpSpecConstantOp %v3i32 CompositeInsert %sc_1 %vec3_0 1\n" // (0, sc_1, 0)
7183 "%sc_vec3_2 = OpSpecConstantOp %v3i32 CompositeInsert %sc_2 %vec3_0 2\n" // (0, 0, sc_2)
7184 "%sc_vec3_0_s = OpSpecConstantOp %v3i32 VectorShuffle %sc_vec3_0 %vec3_undef 0 0xFFFFFFFF 2\n" // (sc_0, ???, 0)
7185 "%sc_vec3_1_s = OpSpecConstantOp %v3i32 VectorShuffle %sc_vec3_1 %vec3_undef 0xFFFFFFFF 1 0\n" // (???, sc_1, 0)
7186 "%sc_vec3_2_s = OpSpecConstantOp %v3i32 VectorShuffle %vec3_undef %sc_vec3_2 5 0xFFFFFFFF 5\n" // (sc_2, ???, sc_2)
7187 "%sc_vec3_01 = OpSpecConstantOp %v3i32 VectorShuffle %sc_vec3_0_s %sc_vec3_1_s 1 0 4\n" // (0, sc_0, sc_1)
7188 "%sc_vec3_012 = OpSpecConstantOp %v3i32 VectorShuffle %sc_vec3_01 %sc_vec3_2_s 5 1 2\n" // (sc_2, sc_0, sc_1)
7189 "%sc_ext_0 = OpSpecConstantOp %i32 CompositeExtract %sc_vec3_012 0\n" // sc_2
7190 "%sc_ext_1 = OpSpecConstantOp %i32 CompositeExtract %sc_vec3_012 1\n" // sc_0
7191 "%sc_ext_2 = OpSpecConstantOp %i32 CompositeExtract %sc_vec3_012 2\n" // sc_1
7192 "%sc_sub = OpSpecConstantOp %i32 ISub %sc_ext_0 %sc_ext_1\n" // (sc_2 - sc_0)
7193 "%sc_final = OpSpecConstantOp %i32 IMul %sc_sub %sc_ext_2\n"; // (sc_2 - sc_0) * sc_1
7195 const char function2[] =
7196 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
7197 "%param = OpFunctionParameter %v4f32\n"
7198 "%label = OpLabel\n"
7199 "%result = OpVariable %fp_v4f32 Function\n"
7200 " OpStore %result %param\n"
7201 "%loc = OpAccessChain %fp_f32 %result %sc_final\n"
7202 "%val = OpLoad %f32 %loc\n"
7203 "%add = OpFAdd %f32 %val %c_f32_0_5\n"
7204 " OpStore %loc %add\n"
7205 "%ret = OpLoad %v4f32 %result\n"
7206 " OpReturnValue %ret\n"
7209 map<string, string> fragments;
7210 SpecConstants specConstants;
7212 fragments["decoration"] = decorations2;
7213 fragments["pre_main"] = typesAndConstants2;
7214 fragments["testfun"] = function2;
7216 specConstants.append<deInt32>(56789);
7217 specConstants.append<deInt32>(-2);
7218 specConstants.append<deInt32>(56788);
7220 createTestsForAllStages("vector_related", inputColors, outputColors2, fragments, specConstants, group.get());
7222 return group.release();
7225 tcu::TestCaseGroup* createOpPhiTests(tcu::TestContext& testCtx)
7227 de::MovePtr<tcu::TestCaseGroup> group (new tcu::TestCaseGroup(testCtx, "opphi", "Test the OpPhi instruction"));
7228 RGBA inputColors[4];
7229 RGBA outputColors1[4];
7230 RGBA outputColors2[4];
7231 RGBA outputColors3[4];
7232 RGBA outputColors4[4];
7233 map<string, string> fragments1;
7234 map<string, string> fragments2;
7235 map<string, string> fragments3;
7236 map<string, string> fragments4;
7237 std::vector<std::string> extensions4;
7238 GraphicsResources resources4;
7239 VulkanFeatures vulkanFeatures4;
7241 const char typesAndConstants1[] =
7242 "%c_f32_p2 = OpConstant %f32 0.2\n"
7243 "%c_f32_p4 = OpConstant %f32 0.4\n"
7244 "%c_f32_p5 = OpConstant %f32 0.5\n"
7245 "%c_f32_p8 = OpConstant %f32 0.8\n";
7247 // vec4 test_code(vec4 param) {
7248 // vec4 result = param;
7249 // for (int i = 0; i < 4; ++i) {
7252 // case 0: operand = .2; break;
7253 // case 1: operand = .5; break;
7254 // case 2: operand = .4; break;
7255 // case 3: operand = .0; break;
7256 // default: break; // unreachable
7258 // result[i] += operand;
7262 const char function1[] =
7263 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
7264 "%param1 = OpFunctionParameter %v4f32\n"
7266 "%iptr = OpVariable %fp_i32 Function\n"
7267 "%result = OpVariable %fp_v4f32 Function\n"
7268 " OpStore %iptr %c_i32_0\n"
7269 " OpStore %result %param1\n"
7273 "%ival = OpLoad %i32 %iptr\n"
7274 "%lt_4 = OpSLessThan %bool %ival %c_i32_4\n"
7275 " OpLoopMerge %exit %phi None\n"
7276 " OpBranchConditional %lt_4 %entry %exit\n"
7278 "%entry = OpLabel\n"
7279 "%loc = OpAccessChain %fp_f32 %result %ival\n"
7280 "%val = OpLoad %f32 %loc\n"
7281 " OpSelectionMerge %phi None\n"
7282 " OpSwitch %ival %default 0 %case0 1 %case1 2 %case2 3 %case3\n"
7284 "%case0 = OpLabel\n"
7286 "%case1 = OpLabel\n"
7288 "%case2 = OpLabel\n"
7290 "%case3 = OpLabel\n"
7293 "%default = OpLabel\n"
7297 "%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
7298 "%add = OpFAdd %f32 %val %operand\n"
7299 " OpStore %loc %add\n"
7300 "%ival_next = OpIAdd %i32 %ival %c_i32_1\n"
7301 " OpStore %iptr %ival_next\n"
7305 "%ret = OpLoad %v4f32 %result\n"
7306 " OpReturnValue %ret\n"
7310 fragments1["pre_main"] = typesAndConstants1;
7311 fragments1["testfun"] = function1;
7313 getHalfColorsFullAlpha(inputColors);
7315 outputColors1[0] = RGBA(178, 255, 229, 255);
7316 outputColors1[1] = RGBA(178, 127, 102, 255);
7317 outputColors1[2] = RGBA(51, 255, 102, 255);
7318 outputColors1[3] = RGBA(51, 127, 229, 255);
7320 createTestsForAllStages("out_of_order", inputColors, outputColors1, fragments1, group.get());
7322 const char typesAndConstants2[] =
7323 "%c_f32_p2 = OpConstant %f32 0.2\n";
7325 // Add .4 to the second element of the given parameter.
7326 const char function2[] =
7327 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
7328 "%param = OpFunctionParameter %v4f32\n"
7329 "%entry = OpLabel\n"
7330 "%result = OpVariable %fp_v4f32 Function\n"
7331 " OpStore %result %param\n"
7332 "%loc = OpAccessChain %fp_f32 %result %c_i32_1\n"
7333 "%val = OpLoad %f32 %loc\n"
7337 "%step = OpPhi %i32 %c_i32_0 %entry %step_next %phi\n"
7338 "%accum = OpPhi %f32 %val %entry %accum_next %phi\n"
7339 "%step_next = OpIAdd %i32 %step %c_i32_1\n"
7340 "%accum_next = OpFAdd %f32 %accum %c_f32_p2\n"
7341 "%still_loop = OpSLessThan %bool %step %c_i32_2\n"
7342 " OpLoopMerge %exit %phi None\n"
7343 " OpBranchConditional %still_loop %phi %exit\n"
7346 " OpStore %loc %accum\n"
7347 "%ret = OpLoad %v4f32 %result\n"
7348 " OpReturnValue %ret\n"
7352 fragments2["pre_main"] = typesAndConstants2;
7353 fragments2["testfun"] = function2;
7355 outputColors2[0] = RGBA(127, 229, 127, 255);
7356 outputColors2[1] = RGBA(127, 102, 0, 255);
7357 outputColors2[2] = RGBA(0, 229, 0, 255);
7358 outputColors2[3] = RGBA(0, 102, 127, 255);
7360 createTestsForAllStages("induction", inputColors, outputColors2, fragments2, group.get());
7362 const char typesAndConstants3[] =
7363 "%true = OpConstantTrue %bool\n"
7364 "%false = OpConstantFalse %bool\n"
7365 "%c_f32_p2 = OpConstant %f32 0.2\n";
7367 // Swap the second and the third element of the given parameter.
7368 const char function3[] =
7369 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
7370 "%param = OpFunctionParameter %v4f32\n"
7371 "%entry = OpLabel\n"
7372 "%result = OpVariable %fp_v4f32 Function\n"
7373 " OpStore %result %param\n"
7374 "%a_loc = OpAccessChain %fp_f32 %result %c_i32_1\n"
7375 "%a_init = OpLoad %f32 %a_loc\n"
7376 "%b_loc = OpAccessChain %fp_f32 %result %c_i32_2\n"
7377 "%b_init = OpLoad %f32 %b_loc\n"
7381 "%still_loop = OpPhi %bool %true %entry %false %phi\n"
7382 "%a_next = OpPhi %f32 %a_init %entry %b_next %phi\n"
7383 "%b_next = OpPhi %f32 %b_init %entry %a_next %phi\n"
7384 " OpLoopMerge %exit %phi None\n"
7385 " OpBranchConditional %still_loop %phi %exit\n"
7388 " OpStore %a_loc %a_next\n"
7389 " OpStore %b_loc %b_next\n"
7390 "%ret = OpLoad %v4f32 %result\n"
7391 " OpReturnValue %ret\n"
7395 fragments3["pre_main"] = typesAndConstants3;
7396 fragments3["testfun"] = function3;
7398 outputColors3[0] = RGBA(127, 127, 127, 255);
7399 outputColors3[1] = RGBA(127, 0, 0, 255);
7400 outputColors3[2] = RGBA(0, 0, 127, 255);
7401 outputColors3[3] = RGBA(0, 127, 0, 255);
7403 createTestsForAllStages("swap", inputColors, outputColors3, fragments3, group.get());
7405 const char typesAndConstants4[] =
7406 "%f16 = OpTypeFloat 16\n"
7407 "%v4f16 = OpTypeVector %f16 4\n"
7408 "%fp_f16 = OpTypePointer Function %f16\n"
7409 "%fp_v4f16 = OpTypePointer Function %v4f16\n"
7410 "%true = OpConstantTrue %bool\n"
7411 "%false = OpConstantFalse %bool\n"
7412 "%c_f32_p2 = OpConstant %f32 0.2\n";
7414 // Swap the second and the third element of the given parameter.
7415 const char function4[] =
7416 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
7417 "%param = OpFunctionParameter %v4f32\n"
7418 "%entry = OpLabel\n"
7419 "%result = OpVariable %fp_v4f16 Function\n"
7420 "%param16 = OpFConvert %v4f16 %param\n"
7421 " OpStore %result %param16\n"
7422 "%a_loc = OpAccessChain %fp_f16 %result %c_i32_1\n"
7423 "%a_init = OpLoad %f16 %a_loc\n"
7424 "%b_loc = OpAccessChain %fp_f16 %result %c_i32_2\n"
7425 "%b_init = OpLoad %f16 %b_loc\n"
7429 "%still_loop = OpPhi %bool %true %entry %false %phi\n"
7430 "%a_next = OpPhi %f16 %a_init %entry %b_next %phi\n"
7431 "%b_next = OpPhi %f16 %b_init %entry %a_next %phi\n"
7432 " OpLoopMerge %exit %phi None\n"
7433 " OpBranchConditional %still_loop %phi %exit\n"
7436 " OpStore %a_loc %a_next\n"
7437 " OpStore %b_loc %b_next\n"
7438 "%ret16 = OpLoad %v4f16 %result\n"
7439 "%ret = OpFConvert %v4f32 %ret16\n"
7440 " OpReturnValue %ret\n"
7444 fragments4["pre_main"] = typesAndConstants4;
7445 fragments4["testfun"] = function4;
7446 fragments4["capability"] = "OpCapability StorageUniformBufferBlock16\n";
7447 fragments4["extension"] = "OpExtension \"SPV_KHR_16bit_storage\"";
7449 extensions4.push_back("VK_KHR_16bit_storage");
7450 extensions4.push_back("VK_KHR_shader_float16_int8");
7452 vulkanFeatures4.ext16BitStorage = EXT16BITSTORAGEFEATURES_UNIFORM_BUFFER_BLOCK;
7453 vulkanFeatures4.extFloat16Int8 = EXTFLOAT16INT8FEATURES_FLOAT16;
7455 outputColors4[0] = RGBA(127, 127, 127, 255);
7456 outputColors4[1] = RGBA(127, 0, 0, 255);
7457 outputColors4[2] = RGBA(0, 0, 127, 255);
7458 outputColors4[3] = RGBA(0, 127, 0, 255);
7460 createTestsForAllStages("swap16", inputColors, outputColors4, fragments4, resources4, extensions4, group.get(), vulkanFeatures4);
7462 return group.release();
7465 tcu::TestCaseGroup* createNoContractionTests(tcu::TestContext& testCtx)
7467 de::MovePtr<tcu::TestCaseGroup> group (new tcu::TestCaseGroup(testCtx, "nocontraction", "Test the NoContraction decoration"));
7468 RGBA inputColors[4];
7469 RGBA outputColors[4];
7471 // With NoContraction, (1 + 2^-23) * (1 - 2^-23) - 1 should be conducted as a multiplication and an addition separately.
7472 // For the multiplication, the result is 1 - 2^-46, which is out of the precision range for 32-bit float. (32-bit float
7473 // only have 23-bit fraction.) So it will be rounded to 1. Or 0x1.fffffc. Then the final result is 0 or -0x1p-24.
7474 // On the contrary, the result will be 2^-46, which is a normalized number perfectly representable as 32-bit float.
7475 const char constantsAndTypes[] =
7476 "%c_vec4_0 = OpConstantComposite %v4f32 %c_f32_0 %c_f32_0 %c_f32_0 %c_f32_1\n"
7477 "%c_vec4_1 = OpConstantComposite %v4f32 %c_f32_1 %c_f32_1 %c_f32_1 %c_f32_1\n"
7478 "%c_f32_1pl2_23 = OpConstant %f32 0x1.000002p+0\n" // 1 + 2^-23
7479 "%c_f32_1mi2_23 = OpConstant %f32 0x1.fffffcp-1\n" // 1 - 2^-23
7480 "%c_f32_n1pn24 = OpConstant %f32 -0x1p-24\n";
7482 const char function[] =
7483 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
7484 "%param = OpFunctionParameter %v4f32\n"
7485 "%label = OpLabel\n"
7486 "%var1 = OpVariable %fp_f32 Function %c_f32_1pl2_23\n"
7487 "%var2 = OpVariable %fp_f32 Function\n"
7488 "%red = OpCompositeExtract %f32 %param 0\n"
7489 "%plus_red = OpFAdd %f32 %c_f32_1mi2_23 %red\n"
7490 " OpStore %var2 %plus_red\n"
7491 "%val1 = OpLoad %f32 %var1\n"
7492 "%val2 = OpLoad %f32 %var2\n"
7493 "%mul = OpFMul %f32 %val1 %val2\n"
7494 "%add = OpFAdd %f32 %mul %c_f32_n1\n"
7495 "%is0 = OpFOrdEqual %bool %add %c_f32_0\n"
7496 "%isn1n24 = OpFOrdEqual %bool %add %c_f32_n1pn24\n"
7497 "%success = OpLogicalOr %bool %is0 %isn1n24\n"
7498 "%v4success = OpCompositeConstruct %v4bool %success %success %success %success\n"
7499 "%ret = OpSelect %v4f32 %v4success %c_vec4_0 %c_vec4_1\n"
7500 " OpReturnValue %ret\n"
7503 struct CaseNameDecoration
7510 CaseNameDecoration tests[] = {
7511 {"multiplication", "OpDecorate %mul NoContraction"},
7512 {"addition", "OpDecorate %add NoContraction"},
7513 {"both", "OpDecorate %mul NoContraction\nOpDecorate %add NoContraction"},
7516 getHalfColorsFullAlpha(inputColors);
7518 for (deUint8 idx = 0; idx < 4; ++idx)
7520 inputColors[idx].setRed(0);
7521 outputColors[idx] = RGBA(0, 0, 0, 255);
7524 for (size_t testNdx = 0; testNdx < sizeof(tests) / sizeof(CaseNameDecoration); ++testNdx)
7526 map<string, string> fragments;
7528 fragments["decoration"] = tests[testNdx].decoration;
7529 fragments["pre_main"] = constantsAndTypes;
7530 fragments["testfun"] = function;
7532 createTestsForAllStages(tests[testNdx].name, inputColors, outputColors, fragments, group.get());
7535 return group.release();
7538 tcu::TestCaseGroup* createMemoryAccessTests(tcu::TestContext& testCtx)
7540 de::MovePtr<tcu::TestCaseGroup> memoryAccessTests (new tcu::TestCaseGroup(testCtx, "opmemoryaccess", "Memory Semantics"));
7543 const char constantsAndTypes[] =
7544 "%c_a2f32_1 = OpConstantComposite %a2f32 %c_f32_1 %c_f32_1\n"
7545 "%fp_a2f32 = OpTypePointer Function %a2f32\n"
7546 "%stype = OpTypeStruct %v4f32 %a2f32 %f32\n"
7547 "%fp_stype = OpTypePointer Function %stype\n";
7549 const char function[] =
7550 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
7551 "%param1 = OpFunctionParameter %v4f32\n"
7553 "%v1 = OpVariable %fp_v4f32 Function\n"
7554 "%v2 = OpVariable %fp_a2f32 Function\n"
7555 "%v3 = OpVariable %fp_f32 Function\n"
7556 "%v = OpVariable %fp_stype Function\n"
7557 "%vv = OpVariable %fp_stype Function\n"
7558 "%vvv = OpVariable %fp_f32 Function\n"
7560 " OpStore %v1 %c_v4f32_1_1_1_1\n"
7561 " OpStore %v2 %c_a2f32_1\n"
7562 " OpStore %v3 %c_f32_1\n"
7564 "%p_v4f32 = OpAccessChain %fp_v4f32 %v %c_u32_0\n"
7565 "%p_a2f32 = OpAccessChain %fp_a2f32 %v %c_u32_1\n"
7566 "%p_f32 = OpAccessChain %fp_f32 %v %c_u32_2\n"
7567 "%v1_v = OpLoad %v4f32 %v1 ${access_type}\n"
7568 "%v2_v = OpLoad %a2f32 %v2 ${access_type}\n"
7569 "%v3_v = OpLoad %f32 %v3 ${access_type}\n"
7571 " OpStore %p_v4f32 %v1_v ${access_type}\n"
7572 " OpStore %p_a2f32 %v2_v ${access_type}\n"
7573 " OpStore %p_f32 %v3_v ${access_type}\n"
7575 " OpCopyMemory %vv %v ${access_type}\n"
7576 " OpCopyMemory %vvv %p_f32 ${access_type}\n"
7578 "%p_f32_2 = OpAccessChain %fp_f32 %vv %c_u32_2\n"
7579 "%v_f32_2 = OpLoad %f32 %p_f32_2\n"
7580 "%v_f32_3 = OpLoad %f32 %vvv\n"
7582 "%ret1 = OpVectorTimesScalar %v4f32 %param1 %v_f32_2\n"
7583 "%ret2 = OpVectorTimesScalar %v4f32 %ret1 %v_f32_3\n"
7584 " OpReturnValue %ret2\n"
7587 struct NameMemoryAccess
7594 NameMemoryAccess tests[] =
7597 { "volatile", "Volatile" },
7598 { "aligned", "Aligned 1" },
7599 { "volatile_aligned", "Volatile|Aligned 1" },
7600 { "nontemporal_aligned", "Nontemporal|Aligned 1" },
7601 { "volatile_nontemporal", "Volatile|Nontemporal" },
7602 { "volatile_nontermporal_aligned", "Volatile|Nontemporal|Aligned 1" },
7605 getHalfColorsFullAlpha(colors);
7607 for (size_t testNdx = 0; testNdx < sizeof(tests) / sizeof(NameMemoryAccess); ++testNdx)
7609 map<string, string> fragments;
7610 map<string, string> memoryAccess;
7611 memoryAccess["access_type"] = tests[testNdx].accessType;
7613 fragments["pre_main"] = constantsAndTypes;
7614 fragments["testfun"] = tcu::StringTemplate(function).specialize(memoryAccess);
7615 createTestsForAllStages(tests[testNdx].name, colors, colors, fragments, memoryAccessTests.get());
7617 return memoryAccessTests.release();
7619 tcu::TestCaseGroup* createOpUndefTests(tcu::TestContext& testCtx)
7621 de::MovePtr<tcu::TestCaseGroup> opUndefTests (new tcu::TestCaseGroup(testCtx, "opundef", "Test OpUndef"));
7622 RGBA defaultColors[4];
7623 map<string, string> fragments;
7624 getDefaultColors(defaultColors);
7626 // First, simple cases that don't do anything with the OpUndef result.
7627 struct NameCodePair { string name, decl, type; };
7628 const NameCodePair tests[] =
7630 {"bool", "", "%bool"},
7631 {"vec2uint32", "", "%v2u32"},
7632 {"image", "%type = OpTypeImage %f32 2D 0 0 0 1 Unknown", "%type"},
7633 {"sampler", "%type = OpTypeSampler", "%type"},
7634 {"sampledimage", "%img = OpTypeImage %f32 2D 0 0 0 1 Unknown\n" "%type = OpTypeSampledImage %img", "%type"},
7635 {"pointer", "", "%fp_i32"},
7636 {"runtimearray", "%type = OpTypeRuntimeArray %f32", "%type"},
7637 {"array", "%c_u32_100 = OpConstant %u32 100\n" "%type = OpTypeArray %i32 %c_u32_100", "%type"},
7638 {"struct", "%type = OpTypeStruct %f32 %i32 %u32", "%type"}};
7639 for (size_t testNdx = 0; testNdx < sizeof(tests) / sizeof(NameCodePair); ++testNdx)
7641 fragments["undef_type"] = tests[testNdx].type;
7642 fragments["testfun"] = StringTemplate(
7643 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
7644 "%param1 = OpFunctionParameter %v4f32\n"
7645 "%label_testfun = OpLabel\n"
7646 "%undef = OpUndef ${undef_type}\n"
7647 "OpReturnValue %param1\n"
7648 "OpFunctionEnd\n").specialize(fragments);
7649 fragments["pre_main"] = tests[testNdx].decl;
7650 createTestsForAllStages(tests[testNdx].name, defaultColors, defaultColors, fragments, opUndefTests.get());
7654 fragments["testfun"] =
7655 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
7656 "%param1 = OpFunctionParameter %v4f32\n"
7657 "%label_testfun = OpLabel\n"
7658 "%undef = OpUndef %f32\n"
7659 "%zero = OpFMul %f32 %undef %c_f32_0\n"
7660 "%is_nan = OpIsNan %bool %zero\n" //OpUndef may result in NaN which may turn %zero into Nan.
7661 "%actually_zero = OpSelect %f32 %is_nan %c_f32_0 %zero\n"
7662 "%a = OpVectorExtractDynamic %f32 %param1 %c_i32_0\n"
7663 "%b = OpFAdd %f32 %a %actually_zero\n"
7664 "%ret = OpVectorInsertDynamic %v4f32 %param1 %b %c_i32_0\n"
7665 "OpReturnValue %ret\n"
7668 createTestsForAllStages("float32", defaultColors, defaultColors, fragments, opUndefTests.get());
7670 fragments["testfun"] =
7671 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
7672 "%param1 = OpFunctionParameter %v4f32\n"
7673 "%label_testfun = OpLabel\n"
7674 "%undef = OpUndef %i32\n"
7675 "%zero = OpIMul %i32 %undef %c_i32_0\n"
7676 "%a = OpVectorExtractDynamic %f32 %param1 %zero\n"
7677 "%ret = OpVectorInsertDynamic %v4f32 %param1 %a %c_i32_0\n"
7678 "OpReturnValue %ret\n"
7681 createTestsForAllStages("sint32", defaultColors, defaultColors, fragments, opUndefTests.get());
7683 fragments["testfun"] =
7684 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
7685 "%param1 = OpFunctionParameter %v4f32\n"
7686 "%label_testfun = OpLabel\n"
7687 "%undef = OpUndef %u32\n"
7688 "%zero = OpIMul %u32 %undef %c_i32_0\n"
7689 "%a = OpVectorExtractDynamic %f32 %param1 %zero\n"
7690 "%ret = OpVectorInsertDynamic %v4f32 %param1 %a %c_i32_0\n"
7691 "OpReturnValue %ret\n"
7694 createTestsForAllStages("uint32", defaultColors, defaultColors, fragments, opUndefTests.get());
7696 fragments["testfun"] =
7697 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
7698 "%param1 = OpFunctionParameter %v4f32\n"
7699 "%label_testfun = OpLabel\n"
7700 "%undef = OpUndef %v4f32\n"
7701 "%vzero = OpVectorTimesScalar %v4f32 %undef %c_f32_0\n"
7702 "%zero_0 = OpVectorExtractDynamic %f32 %vzero %c_i32_0\n"
7703 "%zero_1 = OpVectorExtractDynamic %f32 %vzero %c_i32_1\n"
7704 "%zero_2 = OpVectorExtractDynamic %f32 %vzero %c_i32_2\n"
7705 "%zero_3 = OpVectorExtractDynamic %f32 %vzero %c_i32_3\n"
7706 "%is_nan_0 = OpIsNan %bool %zero_0\n"
7707 "%is_nan_1 = OpIsNan %bool %zero_1\n"
7708 "%is_nan_2 = OpIsNan %bool %zero_2\n"
7709 "%is_nan_3 = OpIsNan %bool %zero_3\n"
7710 "%actually_zero_0 = OpSelect %f32 %is_nan_0 %c_f32_0 %zero_0\n"
7711 "%actually_zero_1 = OpSelect %f32 %is_nan_1 %c_f32_0 %zero_1\n"
7712 "%actually_zero_2 = OpSelect %f32 %is_nan_2 %c_f32_0 %zero_2\n"
7713 "%actually_zero_3 = OpSelect %f32 %is_nan_3 %c_f32_0 %zero_3\n"
7714 "%param1_0 = OpVectorExtractDynamic %f32 %param1 %c_i32_0\n"
7715 "%param1_1 = OpVectorExtractDynamic %f32 %param1 %c_i32_1\n"
7716 "%param1_2 = OpVectorExtractDynamic %f32 %param1 %c_i32_2\n"
7717 "%param1_3 = OpVectorExtractDynamic %f32 %param1 %c_i32_3\n"
7718 "%sum_0 = OpFAdd %f32 %param1_0 %actually_zero_0\n"
7719 "%sum_1 = OpFAdd %f32 %param1_1 %actually_zero_1\n"
7720 "%sum_2 = OpFAdd %f32 %param1_2 %actually_zero_2\n"
7721 "%sum_3 = OpFAdd %f32 %param1_3 %actually_zero_3\n"
7722 "%ret3 = OpVectorInsertDynamic %v4f32 %param1 %sum_3 %c_i32_3\n"
7723 "%ret2 = OpVectorInsertDynamic %v4f32 %ret3 %sum_2 %c_i32_2\n"
7724 "%ret1 = OpVectorInsertDynamic %v4f32 %ret2 %sum_1 %c_i32_1\n"
7725 "%ret = OpVectorInsertDynamic %v4f32 %ret1 %sum_0 %c_i32_0\n"
7726 "OpReturnValue %ret\n"
7729 createTestsForAllStages("vec4float32", defaultColors, defaultColors, fragments, opUndefTests.get());
7731 fragments["pre_main"] =
7732 "%m2x2f32 = OpTypeMatrix %v2f32 2\n";
7733 fragments["testfun"] =
7734 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
7735 "%param1 = OpFunctionParameter %v4f32\n"
7736 "%label_testfun = OpLabel\n"
7737 "%undef = OpUndef %m2x2f32\n"
7738 "%mzero = OpMatrixTimesScalar %m2x2f32 %undef %c_f32_0\n"
7739 "%zero_0 = OpCompositeExtract %f32 %mzero 0 0\n"
7740 "%zero_1 = OpCompositeExtract %f32 %mzero 0 1\n"
7741 "%zero_2 = OpCompositeExtract %f32 %mzero 1 0\n"
7742 "%zero_3 = OpCompositeExtract %f32 %mzero 1 1\n"
7743 "%is_nan_0 = OpIsNan %bool %zero_0\n"
7744 "%is_nan_1 = OpIsNan %bool %zero_1\n"
7745 "%is_nan_2 = OpIsNan %bool %zero_2\n"
7746 "%is_nan_3 = OpIsNan %bool %zero_3\n"
7747 "%actually_zero_0 = OpSelect %f32 %is_nan_0 %c_f32_0 %zero_0\n"
7748 "%actually_zero_1 = OpSelect %f32 %is_nan_1 %c_f32_0 %zero_1\n"
7749 "%actually_zero_2 = OpSelect %f32 %is_nan_2 %c_f32_0 %zero_2\n"
7750 "%actually_zero_3 = OpSelect %f32 %is_nan_3 %c_f32_0 %zero_3\n"
7751 "%param1_0 = OpVectorExtractDynamic %f32 %param1 %c_i32_0\n"
7752 "%param1_1 = OpVectorExtractDynamic %f32 %param1 %c_i32_1\n"
7753 "%param1_2 = OpVectorExtractDynamic %f32 %param1 %c_i32_2\n"
7754 "%param1_3 = OpVectorExtractDynamic %f32 %param1 %c_i32_3\n"
7755 "%sum_0 = OpFAdd %f32 %param1_0 %actually_zero_0\n"
7756 "%sum_1 = OpFAdd %f32 %param1_1 %actually_zero_1\n"
7757 "%sum_2 = OpFAdd %f32 %param1_2 %actually_zero_2\n"
7758 "%sum_3 = OpFAdd %f32 %param1_3 %actually_zero_3\n"
7759 "%ret3 = OpVectorInsertDynamic %v4f32 %param1 %sum_3 %c_i32_3\n"
7760 "%ret2 = OpVectorInsertDynamic %v4f32 %ret3 %sum_2 %c_i32_2\n"
7761 "%ret1 = OpVectorInsertDynamic %v4f32 %ret2 %sum_1 %c_i32_1\n"
7762 "%ret = OpVectorInsertDynamic %v4f32 %ret1 %sum_0 %c_i32_0\n"
7763 "OpReturnValue %ret\n"
7766 createTestsForAllStages("matrix", defaultColors, defaultColors, fragments, opUndefTests.get());
7768 return opUndefTests.release();
7771 void createOpQuantizeSingleOptionTests(tcu::TestCaseGroup* testCtx)
7773 const RGBA inputColors[4] =
7776 RGBA(0, 0, 255, 255),
7777 RGBA(0, 255, 0, 255),
7778 RGBA(0, 255, 255, 255)
7781 const RGBA expectedColors[4] =
7783 RGBA(255, 0, 0, 255),
7784 RGBA(255, 0, 0, 255),
7785 RGBA(255, 0, 0, 255),
7786 RGBA(255, 0, 0, 255)
7789 const struct SingleFP16Possibility
7792 const char* constant; // Value to assign to %test_constant.
7794 const char* condition; // Must assign to %cond an expression that evaluates to true after %c = OpQuantizeToF16(%test_constant + 0).
7800 -constructNormalizedFloat(1, 0x300000),
7801 "%cond = OpFOrdEqual %bool %c %test_constant\n"
7806 constructNormalizedFloat(7, 0x000000),
7807 "%cond = OpFOrdEqual %bool %c %test_constant\n"
7809 // SPIR-V requires that OpQuantizeToF16 flushes
7810 // any numbers that would end up denormalized in F16 to zero.
7814 std::ldexp(1.5f, -140),
7815 "%cond = OpFOrdEqual %bool %c %c_f32_0\n"
7820 -std::ldexp(1.5f, -140),
7821 "%cond = OpFOrdEqual %bool %c %c_f32_0\n"
7826 std::ldexp(1.0f, -16),
7827 "%cond = OpFOrdEqual %bool %c %c_f32_0\n"
7828 }, // too small positive
7830 "negative_too_small",
7832 -std::ldexp(1.0f, -32),
7833 "%cond = OpFOrdEqual %bool %c %c_f32_0\n"
7834 }, // too small negative
7838 -std::ldexp(1.0f, 128),
7840 "%gz = OpFOrdLessThan %bool %c %c_f32_0\n"
7841 "%inf = OpIsInf %bool %c\n"
7842 "%cond = OpLogicalAnd %bool %gz %inf\n"
7847 std::ldexp(1.0f, 128),
7849 "%gz = OpFOrdGreaterThan %bool %c %c_f32_0\n"
7850 "%inf = OpIsInf %bool %c\n"
7851 "%cond = OpLogicalAnd %bool %gz %inf\n"
7854 "round_to_negative_inf",
7856 -std::ldexp(1.0f, 32),
7858 "%gz = OpFOrdLessThan %bool %c %c_f32_0\n"
7859 "%inf = OpIsInf %bool %c\n"
7860 "%cond = OpLogicalAnd %bool %gz %inf\n"
7865 std::ldexp(1.0f, 16),
7867 "%gz = OpFOrdGreaterThan %bool %c %c_f32_0\n"
7868 "%inf = OpIsInf %bool %c\n"
7869 "%cond = OpLogicalAnd %bool %gz %inf\n"
7874 std::numeric_limits<float>::quiet_NaN(),
7876 // Test for any NaN value, as NaNs are not preserved
7877 "%direct_quant = OpQuantizeToF16 %f32 %test_constant\n"
7878 "%cond = OpIsNan %bool %direct_quant\n"
7883 std::numeric_limits<float>::quiet_NaN(),
7885 // Test for any NaN value, as NaNs are not preserved
7886 "%direct_quant = OpQuantizeToF16 %f32 %test_constant\n"
7887 "%cond = OpIsNan %bool %direct_quant\n"
7890 const char* constants =
7891 "%test_constant = OpConstant %f32 "; // The value will be test.constant.
7893 StringTemplate function (
7894 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
7895 "%param1 = OpFunctionParameter %v4f32\n"
7896 "%label_testfun = OpLabel\n"
7897 "%a = OpVectorExtractDynamic %f32 %param1 %c_i32_0\n"
7898 "%b = OpFAdd %f32 %test_constant %a\n"
7899 "%c = OpQuantizeToF16 %f32 %b\n"
7901 "%v4cond = OpCompositeConstruct %v4bool %cond %cond %cond %cond\n"
7902 "%retval = OpSelect %v4f32 %v4cond %c_v4f32_1_0_0_1 %param1\n"
7903 " OpReturnValue %retval\n"
7907 const char* specDecorations = "OpDecorate %test_constant SpecId 0\n";
7908 const char* specConstants =
7909 "%test_constant = OpSpecConstant %f32 0.\n"
7910 "%c = OpSpecConstantOp %f32 QuantizeToF16 %test_constant\n";
7912 StringTemplate specConstantFunction(
7913 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
7914 "%param1 = OpFunctionParameter %v4f32\n"
7915 "%label_testfun = OpLabel\n"
7917 "%v4cond = OpCompositeConstruct %v4bool %cond %cond %cond %cond\n"
7918 "%retval = OpSelect %v4f32 %v4cond %c_v4f32_1_0_0_1 %param1\n"
7919 " OpReturnValue %retval\n"
7923 for (size_t idx = 0; idx < (sizeof(tests)/sizeof(tests[0])); ++idx)
7925 map<string, string> codeSpecialization;
7926 map<string, string> fragments;
7927 codeSpecialization["condition"] = tests[idx].condition;
7928 fragments["testfun"] = function.specialize(codeSpecialization);
7929 fragments["pre_main"] = string(constants) + tests[idx].constant + "\n";
7930 createTestsForAllStages(tests[idx].name, inputColors, expectedColors, fragments, testCtx);
7933 for (size_t idx = 0; idx < (sizeof(tests)/sizeof(tests[0])); ++idx)
7935 map<string, string> codeSpecialization;
7936 map<string, string> fragments;
7937 SpecConstants passConstants;
7939 codeSpecialization["condition"] = tests[idx].condition;
7940 fragments["testfun"] = specConstantFunction.specialize(codeSpecialization);
7941 fragments["decoration"] = specDecorations;
7942 fragments["pre_main"] = specConstants;
7944 passConstants.append<float>(tests[idx].valueAsFloat);
7946 createTestsForAllStages(string("spec_const_") + tests[idx].name, inputColors, expectedColors, fragments, passConstants, testCtx);
7950 void createOpQuantizeTwoPossibilityTests(tcu::TestCaseGroup* testCtx)
7952 RGBA inputColors[4] = {
7954 RGBA(0, 0, 255, 255),
7955 RGBA(0, 255, 0, 255),
7956 RGBA(0, 255, 255, 255)
7959 RGBA expectedColors[4] =
7961 RGBA(255, 0, 0, 255),
7962 RGBA(255, 0, 0, 255),
7963 RGBA(255, 0, 0, 255),
7964 RGBA(255, 0, 0, 255)
7967 struct DualFP16Possibility
7972 const char* possibleOutput1;
7973 const char* possibleOutput2;
7976 "positive_round_up_or_round_down",
7978 constructNormalizedFloat(8, 0x300300),
7983 "negative_round_up_or_round_down",
7985 -constructNormalizedFloat(-7, 0x600800),
7992 constructNormalizedFloat(2, 0x01e000),
7997 "carry_to_exponent",
7999 constructNormalizedFloat(1, 0xffe000),
8004 StringTemplate constants (
8005 "%input_const = OpConstant %f32 ${input}\n"
8006 "%possible_solution1 = OpConstant %f32 ${output1}\n"
8007 "%possible_solution2 = OpConstant %f32 ${output2}\n"
8010 StringTemplate specConstants (
8011 "%input_const = OpSpecConstant %f32 0.\n"
8012 "%possible_solution1 = OpConstant %f32 ${output1}\n"
8013 "%possible_solution2 = OpConstant %f32 ${output2}\n"
8016 const char* specDecorations = "OpDecorate %input_const SpecId 0\n";
8018 const char* function =
8019 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
8020 "%param1 = OpFunctionParameter %v4f32\n"
8021 "%label_testfun = OpLabel\n"
8022 "%a = OpVectorExtractDynamic %f32 %param1 %c_i32_0\n"
8023 // For the purposes of this test we assume that 0.f will always get
8024 // faithfully passed through the pipeline stages.
8025 "%b = OpFAdd %f32 %input_const %a\n"
8026 "%c = OpQuantizeToF16 %f32 %b\n"
8027 "%eq_1 = OpFOrdEqual %bool %c %possible_solution1\n"
8028 "%eq_2 = OpFOrdEqual %bool %c %possible_solution2\n"
8029 "%cond = OpLogicalOr %bool %eq_1 %eq_2\n"
8030 "%v4cond = OpCompositeConstruct %v4bool %cond %cond %cond %cond\n"
8031 "%retval = OpSelect %v4f32 %v4cond %c_v4f32_1_0_0_1 %param1"
8032 " OpReturnValue %retval\n"
8035 for(size_t idx = 0; idx < (sizeof(tests)/sizeof(tests[0])); ++idx) {
8036 map<string, string> fragments;
8037 map<string, string> constantSpecialization;
8039 constantSpecialization["input"] = tests[idx].input;
8040 constantSpecialization["output1"] = tests[idx].possibleOutput1;
8041 constantSpecialization["output2"] = tests[idx].possibleOutput2;
8042 fragments["testfun"] = function;
8043 fragments["pre_main"] = constants.specialize(constantSpecialization);
8044 createTestsForAllStages(tests[idx].name, inputColors, expectedColors, fragments, testCtx);
8047 for(size_t idx = 0; idx < (sizeof(tests)/sizeof(tests[0])); ++idx) {
8048 map<string, string> fragments;
8049 map<string, string> constantSpecialization;
8050 SpecConstants passConstants;
8052 constantSpecialization["output1"] = tests[idx].possibleOutput1;
8053 constantSpecialization["output2"] = tests[idx].possibleOutput2;
8054 fragments["testfun"] = function;
8055 fragments["decoration"] = specDecorations;
8056 fragments["pre_main"] = specConstants.specialize(constantSpecialization);
8058 passConstants.append<float>(tests[idx].inputAsFloat);
8060 createTestsForAllStages(string("spec_const_") + tests[idx].name, inputColors, expectedColors, fragments, passConstants, testCtx);
8064 tcu::TestCaseGroup* createOpQuantizeTests(tcu::TestContext& testCtx)
8066 de::MovePtr<tcu::TestCaseGroup> opQuantizeTests (new tcu::TestCaseGroup(testCtx, "opquantize", "Test OpQuantizeToF16"));
8067 createOpQuantizeSingleOptionTests(opQuantizeTests.get());
8068 createOpQuantizeTwoPossibilityTests(opQuantizeTests.get());
8069 return opQuantizeTests.release();
8072 struct ShaderPermutation
8074 deUint8 vertexPermutation;
8075 deUint8 geometryPermutation;
8076 deUint8 tesscPermutation;
8077 deUint8 tessePermutation;
8078 deUint8 fragmentPermutation;
8081 ShaderPermutation getShaderPermutation(deUint8 inputValue)
8083 ShaderPermutation permutation =
8085 static_cast<deUint8>(inputValue & 0x10? 1u: 0u),
8086 static_cast<deUint8>(inputValue & 0x08? 1u: 0u),
8087 static_cast<deUint8>(inputValue & 0x04? 1u: 0u),
8088 static_cast<deUint8>(inputValue & 0x02? 1u: 0u),
8089 static_cast<deUint8>(inputValue & 0x01? 1u: 0u)
8094 tcu::TestCaseGroup* createModuleTests(tcu::TestContext& testCtx)
8096 RGBA defaultColors[4];
8097 RGBA invertedColors[4];
8098 de::MovePtr<tcu::TestCaseGroup> moduleTests (new tcu::TestCaseGroup(testCtx, "module", "Multiple entry points into shaders"));
8100 getDefaultColors(defaultColors);
8101 getInvertedDefaultColors(invertedColors);
8103 // Combined module tests
8105 // Shader stages: vertex and fragment
8107 const ShaderElement combinedPipeline[] =
8109 ShaderElement("module", "main", VK_SHADER_STAGE_VERTEX_BIT),
8110 ShaderElement("module", "main", VK_SHADER_STAGE_FRAGMENT_BIT)
8113 addFunctionCaseWithPrograms<InstanceContext>(
8114 moduleTests.get(), "same_module", "", createCombinedModule, runAndVerifyDefaultPipeline,
8115 createInstanceContext(combinedPipeline, map<string, string>()));
8118 // Shader stages: vertex, geometry and fragment
8120 const ShaderElement combinedPipeline[] =
8122 ShaderElement("module", "main", VK_SHADER_STAGE_VERTEX_BIT),
8123 ShaderElement("module", "main", VK_SHADER_STAGE_GEOMETRY_BIT),
8124 ShaderElement("module", "main", VK_SHADER_STAGE_FRAGMENT_BIT)
8127 addFunctionCaseWithPrograms<InstanceContext>(
8128 moduleTests.get(), "same_module_geom", "", createCombinedModule, runAndVerifyDefaultPipeline,
8129 createInstanceContext(combinedPipeline, map<string, string>()));
8132 // Shader stages: vertex, tessellation control, tessellation evaluation and fragment
8134 const ShaderElement combinedPipeline[] =
8136 ShaderElement("module", "main", VK_SHADER_STAGE_VERTEX_BIT),
8137 ShaderElement("module", "main", VK_SHADER_STAGE_TESSELLATION_CONTROL_BIT),
8138 ShaderElement("module", "main", VK_SHADER_STAGE_TESSELLATION_EVALUATION_BIT),
8139 ShaderElement("module", "main", VK_SHADER_STAGE_FRAGMENT_BIT)
8142 addFunctionCaseWithPrograms<InstanceContext>(
8143 moduleTests.get(), "same_module_tessc_tesse", "", createCombinedModule, runAndVerifyDefaultPipeline,
8144 createInstanceContext(combinedPipeline, map<string, string>()));
8147 // Shader stages: vertex, tessellation control, tessellation evaluation, geometry and fragment
8149 const ShaderElement combinedPipeline[] =
8151 ShaderElement("module", "main", VK_SHADER_STAGE_VERTEX_BIT),
8152 ShaderElement("module", "main", VK_SHADER_STAGE_TESSELLATION_CONTROL_BIT),
8153 ShaderElement("module", "main", VK_SHADER_STAGE_TESSELLATION_EVALUATION_BIT),
8154 ShaderElement("module", "main", VK_SHADER_STAGE_GEOMETRY_BIT),
8155 ShaderElement("module", "main", VK_SHADER_STAGE_FRAGMENT_BIT)
8158 addFunctionCaseWithPrograms<InstanceContext>(
8159 moduleTests.get(), "same_module_tessc_tesse_geom", "", createCombinedModule, runAndVerifyDefaultPipeline,
8160 createInstanceContext(combinedPipeline, map<string, string>()));
8164 const char* numbers[] =
8169 for (deInt8 idx = 0; idx < 32; ++idx)
8171 ShaderPermutation permutation = getShaderPermutation(idx);
8172 string name = string("vert") + numbers[permutation.vertexPermutation] + "_geom" + numbers[permutation.geometryPermutation] + "_tessc" + numbers[permutation.tesscPermutation] + "_tesse" + numbers[permutation.tessePermutation] + "_frag" + numbers[permutation.fragmentPermutation];
8173 const ShaderElement pipeline[] =
8175 ShaderElement("vert", string("vert") + numbers[permutation.vertexPermutation], VK_SHADER_STAGE_VERTEX_BIT),
8176 ShaderElement("geom", string("geom") + numbers[permutation.geometryPermutation], VK_SHADER_STAGE_GEOMETRY_BIT),
8177 ShaderElement("tessc", string("tessc") + numbers[permutation.tesscPermutation], VK_SHADER_STAGE_TESSELLATION_CONTROL_BIT),
8178 ShaderElement("tesse", string("tesse") + numbers[permutation.tessePermutation], VK_SHADER_STAGE_TESSELLATION_EVALUATION_BIT),
8179 ShaderElement("frag", string("frag") + numbers[permutation.fragmentPermutation], VK_SHADER_STAGE_FRAGMENT_BIT)
8182 // If there are an even number of swaps, then it should be no-op.
8183 // If there are an odd number, the color should be flipped.
8184 if ((permutation.vertexPermutation + permutation.geometryPermutation + permutation.tesscPermutation + permutation.tessePermutation + permutation.fragmentPermutation) % 2 == 0)
8186 addFunctionCaseWithPrograms<InstanceContext>(
8187 moduleTests.get(), name, "", createMultipleEntries, runAndVerifyDefaultPipeline,
8188 createInstanceContext(pipeline, defaultColors, defaultColors, map<string, string>()));
8192 addFunctionCaseWithPrograms<InstanceContext>(
8193 moduleTests.get(), name, "", createMultipleEntries, runAndVerifyDefaultPipeline,
8194 createInstanceContext(pipeline, defaultColors, invertedColors, map<string, string>()));
8197 return moduleTests.release();
8200 tcu::TestCaseGroup* createLoopTests(tcu::TestContext& testCtx)
8202 de::MovePtr<tcu::TestCaseGroup> testGroup(new tcu::TestCaseGroup(testCtx, "loop", "Looping control flow"));
8203 RGBA defaultColors[4];
8204 getDefaultColors(defaultColors);
8205 map<string, string> fragments;
8206 fragments["pre_main"] =
8207 "%c_f32_5 = OpConstant %f32 5.\n";
8209 // A loop with a single block. The Continue Target is the loop block
8210 // itself. In SPIR-V terms, the "loop construct" contains no blocks at all
8211 // -- the "continue construct" forms the entire loop.
8212 fragments["testfun"] =
8213 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
8214 "%param1 = OpFunctionParameter %v4f32\n"
8216 "%entry = OpLabel\n"
8217 "%val0 = OpVectorExtractDynamic %f32 %param1 %c_i32_0\n"
8220 ";adds and subtracts 1.0 to %val in alternate iterations\n"
8222 "%count = OpPhi %i32 %c_i32_4 %entry %count__ %loop\n"
8223 "%delta = OpPhi %f32 %c_f32_1 %entry %minus_delta %loop\n"
8224 "%val1 = OpPhi %f32 %val0 %entry %val %loop\n"
8225 "%val = OpFAdd %f32 %val1 %delta\n"
8226 "%minus_delta = OpFSub %f32 %c_f32_0 %delta\n"
8227 "%count__ = OpISub %i32 %count %c_i32_1\n"
8228 "%again = OpSGreaterThan %bool %count__ %c_i32_0\n"
8229 "OpLoopMerge %exit %loop None\n"
8230 "OpBranchConditional %again %loop %exit\n"
8233 "%result = OpVectorInsertDynamic %v4f32 %param1 %val %c_i32_0\n"
8234 "OpReturnValue %result\n"
8238 createTestsForAllStages("single_block", defaultColors, defaultColors, fragments, testGroup.get());
8240 // Body comprised of multiple basic blocks.
8241 const StringTemplate multiBlock(
8242 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
8243 "%param1 = OpFunctionParameter %v4f32\n"
8245 "%entry = OpLabel\n"
8246 "%val0 = OpVectorExtractDynamic %f32 %param1 %c_i32_0\n"
8249 ";adds and subtracts 1.0 to %val in alternate iterations\n"
8251 "%count = OpPhi %i32 %c_i32_4 %entry %count__ %gather\n"
8252 "%delta = OpPhi %f32 %c_f32_1 %entry %delta_next %gather\n"
8253 "%val1 = OpPhi %f32 %val0 %entry %val %gather\n"
8254 // There are several possibilities for the Continue Target below. Each
8255 // will be specialized into a separate test case.
8256 "OpLoopMerge %exit ${continue_target} None\n"
8260 ";delta_next = (delta > 0) ? -1 : 1;\n"
8261 "%gt0 = OpFOrdGreaterThan %bool %delta %c_f32_0\n"
8262 "OpSelectionMerge %gather DontFlatten\n"
8263 "OpBranchConditional %gt0 %even %odd ;tells us if %count is even or odd\n"
8266 "OpBranch %gather\n"
8269 "OpBranch %gather\n"
8271 "%gather = OpLabel\n"
8272 "%delta_next = OpPhi %f32 %c_f32_n1 %even %c_f32_1 %odd\n"
8273 "%val = OpFAdd %f32 %val1 %delta\n"
8274 "%count__ = OpISub %i32 %count %c_i32_1\n"
8275 "%again = OpSGreaterThan %bool %count__ %c_i32_0\n"
8276 "OpBranchConditional %again %loop %exit\n"
8279 "%result = OpVectorInsertDynamic %v4f32 %param1 %val %c_i32_0\n"
8280 "OpReturnValue %result\n"
8284 map<string, string> continue_target;
8286 // The Continue Target is the loop block itself.
8287 continue_target["continue_target"] = "%loop";
8288 fragments["testfun"] = multiBlock.specialize(continue_target);
8289 createTestsForAllStages("multi_block_continue_construct", defaultColors, defaultColors, fragments, testGroup.get());
8291 // The Continue Target is at the end of the loop.
8292 continue_target["continue_target"] = "%gather";
8293 fragments["testfun"] = multiBlock.specialize(continue_target);
8294 createTestsForAllStages("multi_block_loop_construct", defaultColors, defaultColors, fragments, testGroup.get());
8296 // A loop with continue statement.
8297 fragments["testfun"] =
8298 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
8299 "%param1 = OpFunctionParameter %v4f32\n"
8301 "%entry = OpLabel\n"
8302 "%val0 = OpVectorExtractDynamic %f32 %param1 %c_i32_0\n"
8305 ";adds 4, 3, and 1 to %val0 (skips 2)\n"
8307 "%count = OpPhi %i32 %c_i32_4 %entry %count__ %continue\n"
8308 "%val1 = OpPhi %f32 %val0 %entry %val %continue\n"
8309 "OpLoopMerge %exit %continue None\n"
8313 ";skip if %count==2\n"
8314 "%eq2 = OpIEqual %bool %count %c_i32_2\n"
8315 "OpSelectionMerge %continue DontFlatten\n"
8316 "OpBranchConditional %eq2 %continue %body\n"
8319 "%fcount = OpConvertSToF %f32 %count\n"
8320 "%val2 = OpFAdd %f32 %val1 %fcount\n"
8321 "OpBranch %continue\n"
8323 "%continue = OpLabel\n"
8324 "%val = OpPhi %f32 %val2 %body %val1 %if\n"
8325 "%count__ = OpISub %i32 %count %c_i32_1\n"
8326 "%again = OpSGreaterThan %bool %count__ %c_i32_0\n"
8327 "OpBranchConditional %again %loop %exit\n"
8330 "%same = OpFSub %f32 %val %c_f32_8\n"
8331 "%result = OpVectorInsertDynamic %v4f32 %param1 %same %c_i32_0\n"
8332 "OpReturnValue %result\n"
8334 createTestsForAllStages("continue", defaultColors, defaultColors, fragments, testGroup.get());
8336 // A loop with break.
8337 fragments["testfun"] =
8338 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
8339 "%param1 = OpFunctionParameter %v4f32\n"
8341 "%entry = OpLabel\n"
8342 ";param1 components are between 0 and 1, so dot product is 4 or less\n"
8343 "%dot = OpDot %f32 %param1 %param1\n"
8344 "%div = OpFDiv %f32 %dot %c_f32_5\n"
8345 "%zero = OpConvertFToU %u32 %div\n"
8346 "%two = OpIAdd %i32 %zero %c_i32_2\n"
8347 "%val0 = OpVectorExtractDynamic %f32 %param1 %c_i32_0\n"
8350 ";adds 4 and 3 to %val0 (exits early)\n"
8352 "%count = OpPhi %i32 %c_i32_4 %entry %count__ %continue\n"
8353 "%val1 = OpPhi %f32 %val0 %entry %val2 %continue\n"
8354 "OpLoopMerge %exit %continue None\n"
8358 ";end loop if %count==%two\n"
8359 "%above2 = OpSGreaterThan %bool %count %two\n"
8360 "OpSelectionMerge %continue DontFlatten\n"
8361 "OpBranchConditional %above2 %body %exit\n"
8364 "%fcount = OpConvertSToF %f32 %count\n"
8365 "%val2 = OpFAdd %f32 %val1 %fcount\n"
8366 "OpBranch %continue\n"
8368 "%continue = OpLabel\n"
8369 "%count__ = OpISub %i32 %count %c_i32_1\n"
8370 "%again = OpSGreaterThan %bool %count__ %c_i32_0\n"
8371 "OpBranchConditional %again %loop %exit\n"
8374 "%val_post = OpPhi %f32 %val2 %continue %val1 %if\n"
8375 "%same = OpFSub %f32 %val_post %c_f32_7\n"
8376 "%result = OpVectorInsertDynamic %v4f32 %param1 %same %c_i32_0\n"
8377 "OpReturnValue %result\n"
8379 createTestsForAllStages("break", defaultColors, defaultColors, fragments, testGroup.get());
8381 // A loop with return.
8382 fragments["testfun"] =
8383 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
8384 "%param1 = OpFunctionParameter %v4f32\n"
8386 "%entry = OpLabel\n"
8387 ";param1 components are between 0 and 1, so dot product is 4 or less\n"
8388 "%dot = OpDot %f32 %param1 %param1\n"
8389 "%div = OpFDiv %f32 %dot %c_f32_5\n"
8390 "%zero = OpConvertFToU %u32 %div\n"
8391 "%two = OpIAdd %i32 %zero %c_i32_2\n"
8392 "%val0 = OpVectorExtractDynamic %f32 %param1 %c_i32_0\n"
8395 ";returns early without modifying %param1\n"
8397 "%count = OpPhi %i32 %c_i32_4 %entry %count__ %continue\n"
8398 "%val1 = OpPhi %f32 %val0 %entry %val2 %continue\n"
8399 "OpLoopMerge %exit %continue None\n"
8403 ";return if %count==%two\n"
8404 "%above2 = OpSGreaterThan %bool %count %two\n"
8405 "OpSelectionMerge %continue DontFlatten\n"
8406 "OpBranchConditional %above2 %body %early_exit\n"
8408 "%early_exit = OpLabel\n"
8409 "OpReturnValue %param1\n"
8412 "%fcount = OpConvertSToF %f32 %count\n"
8413 "%val2 = OpFAdd %f32 %val1 %fcount\n"
8414 "OpBranch %continue\n"
8416 "%continue = OpLabel\n"
8417 "%count__ = OpISub %i32 %count %c_i32_1\n"
8418 "%again = OpSGreaterThan %bool %count__ %c_i32_0\n"
8419 "OpBranchConditional %again %loop %exit\n"
8422 ";should never get here, so return an incorrect result\n"
8423 "%result = OpVectorInsertDynamic %v4f32 %param1 %val2 %c_i32_0\n"
8424 "OpReturnValue %result\n"
8426 createTestsForAllStages("return", defaultColors, defaultColors, fragments, testGroup.get());
8428 // Continue inside a switch block to break to enclosing loop's merge block.
8429 // Matches roughly the following GLSL code:
8430 // for (; keep_going; keep_going = false)
8432 // switch (int(param1.x))
8434 // case 0: continue;
8435 // case 1: continue;
8436 // default: continue;
8438 // dead code: modify return value to invalid result.
8440 fragments["pre_main"] =
8441 "%fp_bool = OpTypePointer Function %bool\n"
8442 "%true = OpConstantTrue %bool\n"
8443 "%false = OpConstantFalse %bool\n";
8445 fragments["testfun"] =
8446 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
8447 "%param1 = OpFunctionParameter %v4f32\n"
8449 "%entry = OpLabel\n"
8450 "%keep_going = OpVariable %fp_bool Function\n"
8451 "%val_ptr = OpVariable %fp_f32 Function\n"
8452 "%param1_x = OpCompositeExtract %f32 %param1 0\n"
8453 "OpStore %keep_going %true\n"
8454 "OpBranch %forloop_begin\n"
8456 "%forloop_begin = OpLabel\n"
8457 "OpLoopMerge %forloop_merge %forloop_continue None\n"
8458 "OpBranch %forloop\n"
8460 "%forloop = OpLabel\n"
8461 "%for_condition = OpLoad %bool %keep_going\n"
8462 "OpBranchConditional %for_condition %forloop_body %forloop_merge\n"
8464 "%forloop_body = OpLabel\n"
8465 "OpStore %val_ptr %param1_x\n"
8466 "%param1_x_int = OpConvertFToS %i32 %param1_x\n"
8468 "OpSelectionMerge %switch_merge None\n"
8469 "OpSwitch %param1_x_int %default 0 %case_0 1 %case_1\n"
8470 "%case_0 = OpLabel\n"
8471 "OpBranch %forloop_continue\n"
8472 "%case_1 = OpLabel\n"
8473 "OpBranch %forloop_continue\n"
8474 "%default = OpLabel\n"
8475 "OpBranch %forloop_continue\n"
8476 "%switch_merge = OpLabel\n"
8477 ";should never get here, so change the return value to invalid result\n"
8478 "OpStore %val_ptr %c_f32_1\n"
8479 "OpBranch %forloop_continue\n"
8481 "%forloop_continue = OpLabel\n"
8482 "OpStore %keep_going %false\n"
8483 "OpBranch %forloop_begin\n"
8484 "%forloop_merge = OpLabel\n"
8486 "%val = OpLoad %f32 %val_ptr\n"
8487 "%result = OpVectorInsertDynamic %v4f32 %param1 %val %c_i32_0\n"
8488 "OpReturnValue %result\n"
8490 createTestsForAllStages("switch_continue", defaultColors, defaultColors, fragments, testGroup.get());
8492 return testGroup.release();
8495 // A collection of tests putting OpControlBarrier in places GLSL forbids but SPIR-V allows.
8496 tcu::TestCaseGroup* createBarrierTests(tcu::TestContext& testCtx)
8498 de::MovePtr<tcu::TestCaseGroup> testGroup(new tcu::TestCaseGroup(testCtx, "barrier", "OpControlBarrier"));
8499 map<string, string> fragments;
8501 // A barrier inside a function body.
8502 fragments["pre_main"] =
8503 "%Workgroup = OpConstant %i32 2\n"
8504 "%WorkgroupAcquireRelease = OpConstant %i32 0x108\n";
8505 fragments["testfun"] =
8506 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
8507 "%param1 = OpFunctionParameter %v4f32\n"
8508 "%label_testfun = OpLabel\n"
8509 "OpControlBarrier %Workgroup %Workgroup %WorkgroupAcquireRelease\n"
8510 "OpReturnValue %param1\n"
8512 addTessCtrlTest(testGroup.get(), "in_function", fragments);
8514 // Common setup code for the following tests.
8515 fragments["pre_main"] =
8516 "%Workgroup = OpConstant %i32 2\n"
8517 "%WorkgroupAcquireRelease = OpConstant %i32 0x108\n"
8518 "%c_f32_5 = OpConstant %f32 5.\n";
8519 const string setupPercentZero = // Begins %test_code function with code that sets %zero to 0u but cannot be optimized away.
8520 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
8521 "%param1 = OpFunctionParameter %v4f32\n"
8522 "%entry = OpLabel\n"
8523 ";param1 components are between 0 and 1, so dot product is 4 or less\n"
8524 "%dot = OpDot %f32 %param1 %param1\n"
8525 "%div = OpFDiv %f32 %dot %c_f32_5\n"
8526 "%zero = OpConvertFToU %u32 %div\n";
8528 // Barriers inside OpSwitch branches.
8529 fragments["testfun"] =
8531 "OpSelectionMerge %switch_exit None\n"
8532 "OpSwitch %zero %switch_default 0 %case0 1 %case1 ;should always go to %case0\n"
8534 "%case1 = OpLabel\n"
8535 ";This barrier should never be executed, but its presence makes test failure more likely when there's a bug.\n"
8536 "OpControlBarrier %Workgroup %Workgroup %WorkgroupAcquireRelease\n"
8537 "%wrong_branch_alert1 = OpVectorInsertDynamic %v4f32 %param1 %c_f32_0_5 %c_i32_0\n"
8538 "OpBranch %switch_exit\n"
8540 "%switch_default = OpLabel\n"
8541 "%wrong_branch_alert2 = OpVectorInsertDynamic %v4f32 %param1 %c_f32_0_5 %c_i32_0\n"
8542 ";This barrier should never be executed, but its presence makes test failure more likely when there's a bug.\n"
8543 "OpControlBarrier %Workgroup %Workgroup %WorkgroupAcquireRelease\n"
8544 "OpBranch %switch_exit\n"
8546 "%case0 = OpLabel\n"
8547 "OpControlBarrier %Workgroup %Workgroup %WorkgroupAcquireRelease\n"
8548 "OpBranch %switch_exit\n"
8550 "%switch_exit = OpLabel\n"
8551 "%ret = OpPhi %v4f32 %param1 %case0 %wrong_branch_alert1 %case1 %wrong_branch_alert2 %switch_default\n"
8552 "OpReturnValue %ret\n"
8554 addTessCtrlTest(testGroup.get(), "in_switch", fragments);
8556 // Barriers inside if-then-else.
8557 fragments["testfun"] =
8559 "%eq0 = OpIEqual %bool %zero %c_u32_0\n"
8560 "OpSelectionMerge %exit DontFlatten\n"
8561 "OpBranchConditional %eq0 %then %else\n"
8564 ";This barrier should never be executed, but its presence makes test failure more likely when there's a bug.\n"
8565 "OpControlBarrier %Workgroup %Workgroup %WorkgroupAcquireRelease\n"
8566 "%wrong_branch_alert = OpVectorInsertDynamic %v4f32 %param1 %c_f32_0_5 %c_i32_0\n"
8570 "OpControlBarrier %Workgroup %Workgroup %WorkgroupAcquireRelease\n"
8573 "%ret = OpPhi %v4f32 %param1 %then %wrong_branch_alert %else\n"
8574 "OpReturnValue %ret\n"
8576 addTessCtrlTest(testGroup.get(), "in_if", fragments);
8578 // A barrier after control-flow reconvergence, tempting the compiler to attempt something like this:
8579 // http://lists.llvm.org/pipermail/llvm-dev/2009-October/026317.html.
8580 fragments["testfun"] =
8582 "%thread_id = OpLoad %i32 %BP_gl_InvocationID\n"
8583 "%thread0 = OpIEqual %bool %thread_id %c_i32_0\n"
8584 "OpSelectionMerge %exit DontFlatten\n"
8585 "OpBranchConditional %thread0 %then %else\n"
8588 "%val0 = OpVectorExtractDynamic %f32 %param1 %c_i32_0\n"
8592 "%val1 = OpVectorExtractDynamic %f32 %param1 %zero\n"
8596 "%val = OpPhi %f32 %val0 %else %val1 %then\n"
8597 "OpControlBarrier %Workgroup %Workgroup %WorkgroupAcquireRelease\n"
8598 "%ret = OpVectorInsertDynamic %v4f32 %param1 %val %zero\n"
8599 "OpReturnValue %ret\n"
8601 addTessCtrlTest(testGroup.get(), "after_divergent_if", fragments);
8603 // A barrier inside a loop.
8604 fragments["pre_main"] =
8605 "%Workgroup = OpConstant %i32 2\n"
8606 "%WorkgroupAcquireRelease = OpConstant %i32 0x108\n"
8607 "%c_f32_10 = OpConstant %f32 10.\n";
8608 fragments["testfun"] =
8609 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
8610 "%param1 = OpFunctionParameter %v4f32\n"
8611 "%entry = OpLabel\n"
8612 "%val0 = OpVectorExtractDynamic %f32 %param1 %c_i32_0\n"
8615 ";adds 4, 3, 2, and 1 to %val0\n"
8617 "%count = OpPhi %i32 %c_i32_4 %entry %count__ %loop\n"
8618 "%val1 = OpPhi %f32 %val0 %entry %val %loop\n"
8619 "OpControlBarrier %Workgroup %Workgroup %WorkgroupAcquireRelease\n"
8620 "%fcount = OpConvertSToF %f32 %count\n"
8621 "%val = OpFAdd %f32 %val1 %fcount\n"
8622 "%count__ = OpISub %i32 %count %c_i32_1\n"
8623 "%again = OpSGreaterThan %bool %count__ %c_i32_0\n"
8624 "OpLoopMerge %exit %loop None\n"
8625 "OpBranchConditional %again %loop %exit\n"
8628 "%same = OpFSub %f32 %val %c_f32_10\n"
8629 "%ret = OpVectorInsertDynamic %v4f32 %param1 %same %c_i32_0\n"
8630 "OpReturnValue %ret\n"
8632 addTessCtrlTest(testGroup.get(), "in_loop", fragments);
8634 return testGroup.release();
8637 // Test for the OpFRem instruction.
8638 tcu::TestCaseGroup* createFRemTests(tcu::TestContext& testCtx)
8640 de::MovePtr<tcu::TestCaseGroup> testGroup(new tcu::TestCaseGroup(testCtx, "frem", "OpFRem"));
8641 map<string, string> fragments;
8642 RGBA inputColors[4];
8643 RGBA outputColors[4];
8645 fragments["pre_main"] =
8646 "%c_f32_3 = OpConstant %f32 3.0\n"
8647 "%c_f32_n3 = OpConstant %f32 -3.0\n"
8648 "%c_f32_4 = OpConstant %f32 4.0\n"
8649 "%c_f32_p75 = OpConstant %f32 0.75\n"
8650 "%c_v4f32_p75_p75_p75_p75 = OpConstantComposite %v4f32 %c_f32_p75 %c_f32_p75 %c_f32_p75 %c_f32_p75 \n"
8651 "%c_v4f32_4_4_4_4 = OpConstantComposite %v4f32 %c_f32_4 %c_f32_4 %c_f32_4 %c_f32_4\n"
8652 "%c_v4f32_3_n3_3_n3 = OpConstantComposite %v4f32 %c_f32_3 %c_f32_n3 %c_f32_3 %c_f32_n3\n";
8654 // The test does the following.
8655 // vec4 result = (param1 * 8.0) - 4.0;
8656 // return (frem(result.x,3) + 0.75, frem(result.y, -3) + 0.75, 0, 1)
8657 fragments["testfun"] =
8658 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
8659 "%param1 = OpFunctionParameter %v4f32\n"
8660 "%label_testfun = OpLabel\n"
8661 "%v_times_8 = OpVectorTimesScalar %v4f32 %param1 %c_f32_8\n"
8662 "%minus_4 = OpFSub %v4f32 %v_times_8 %c_v4f32_4_4_4_4\n"
8663 "%frem = OpFRem %v4f32 %minus_4 %c_v4f32_3_n3_3_n3\n"
8664 "%added = OpFAdd %v4f32 %frem %c_v4f32_p75_p75_p75_p75\n"
8665 "%xyz_1 = OpVectorInsertDynamic %v4f32 %added %c_f32_1 %c_i32_3\n"
8666 "%xy_0_1 = OpVectorInsertDynamic %v4f32 %xyz_1 %c_f32_0 %c_i32_2\n"
8667 "OpReturnValue %xy_0_1\n"
8671 inputColors[0] = RGBA(16, 16, 0, 255);
8672 inputColors[1] = RGBA(232, 232, 0, 255);
8673 inputColors[2] = RGBA(232, 16, 0, 255);
8674 inputColors[3] = RGBA(16, 232, 0, 255);
8676 outputColors[0] = RGBA(64, 64, 0, 255);
8677 outputColors[1] = RGBA(255, 255, 0, 255);
8678 outputColors[2] = RGBA(255, 64, 0, 255);
8679 outputColors[3] = RGBA(64, 255, 0, 255);
8681 createTestsForAllStages("frem", inputColors, outputColors, fragments, testGroup.get());
8682 return testGroup.release();
8685 // Test for the OpSRem instruction.
8686 tcu::TestCaseGroup* createOpSRemGraphicsTests(tcu::TestContext& testCtx, qpTestResult negFailResult)
8688 de::MovePtr<tcu::TestCaseGroup> testGroup(new tcu::TestCaseGroup(testCtx, "srem", "OpSRem"));
8689 map<string, string> fragments;
8691 fragments["pre_main"] =
8692 "%c_f32_255 = OpConstant %f32 255.0\n"
8693 "%c_i32_128 = OpConstant %i32 128\n"
8694 "%c_i32_255 = OpConstant %i32 255\n"
8695 "%c_v4f32_255 = OpConstantComposite %v4f32 %c_f32_255 %c_f32_255 %c_f32_255 %c_f32_255 \n"
8696 "%c_v4f32_0_5 = OpConstantComposite %v4f32 %c_f32_0_5 %c_f32_0_5 %c_f32_0_5 %c_f32_0_5 \n"
8697 "%c_v4i32_128 = OpConstantComposite %v4i32 %c_i32_128 %c_i32_128 %c_i32_128 %c_i32_128 \n";
8699 // The test does the following.
8700 // ivec4 ints = int(param1 * 255.0 + 0.5) - 128;
8701 // ivec4 result = ivec4(srem(ints.x, ints.y), srem(ints.y, ints.z), srem(ints.z, ints.x), 255);
8702 // return float(result + 128) / 255.0;
8703 fragments["testfun"] =
8704 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
8705 "%param1 = OpFunctionParameter %v4f32\n"
8706 "%label_testfun = OpLabel\n"
8707 "%div255 = OpFMul %v4f32 %param1 %c_v4f32_255\n"
8708 "%add0_5 = OpFAdd %v4f32 %div255 %c_v4f32_0_5\n"
8709 "%uints_in = OpConvertFToS %v4i32 %add0_5\n"
8710 "%ints_in = OpISub %v4i32 %uints_in %c_v4i32_128\n"
8711 "%x_in = OpCompositeExtract %i32 %ints_in 0\n"
8712 "%y_in = OpCompositeExtract %i32 %ints_in 1\n"
8713 "%z_in = OpCompositeExtract %i32 %ints_in 2\n"
8714 "%x_out = OpSRem %i32 %x_in %y_in\n"
8715 "%y_out = OpSRem %i32 %y_in %z_in\n"
8716 "%z_out = OpSRem %i32 %z_in %x_in\n"
8717 "%ints_out = OpCompositeConstruct %v4i32 %x_out %y_out %z_out %c_i32_255\n"
8718 "%ints_offset = OpIAdd %v4i32 %ints_out %c_v4i32_128\n"
8719 "%f_ints_offset = OpConvertSToF %v4f32 %ints_offset\n"
8720 "%float_out = OpFDiv %v4f32 %f_ints_offset %c_v4f32_255\n"
8721 "OpReturnValue %float_out\n"
8724 const struct CaseParams
8727 const char* failMessageTemplate; // customized status message
8728 qpTestResult failResult; // override status on failure
8729 int operands[4][3]; // four (x, y, z) vectors of operands
8730 int results[4][3]; // four (x, y, z) vectors of results
8736 QP_TEST_RESULT_FAIL,
8737 { { 5, 12, 17 }, { 5, 5, 7 }, { 75, 8, 81 }, { 25, 60, 100 } }, // operands
8738 { { 5, 12, 2 }, { 0, 5, 2 }, { 3, 8, 6 }, { 25, 60, 0 } }, // results
8742 "Inconsistent results, but within specification: ${reason}",
8743 negFailResult, // negative operands, not required by the spec
8744 { { 5, 12, -17 }, { -5, -5, 7 }, { 75, 8, -81 }, { 25, -60, 100 } }, // operands
8745 { { 5, 12, -2 }, { 0, -5, 2 }, { 3, 8, -6 }, { 25, -60, 0 } }, // results
8748 // If either operand is negative the result is undefined. Some implementations may still return correct values.
8750 for (int caseNdx = 0; caseNdx < DE_LENGTH_OF_ARRAY(cases); ++caseNdx)
8752 const CaseParams& params = cases[caseNdx];
8753 RGBA inputColors[4];
8754 RGBA outputColors[4];
8756 for (int i = 0; i < 4; ++i)
8758 inputColors [i] = RGBA(params.operands[i][0] + 128, params.operands[i][1] + 128, params.operands[i][2] + 128, 255);
8759 outputColors[i] = RGBA(params.results [i][0] + 128, params.results [i][1] + 128, params.results [i][2] + 128, 255);
8762 createTestsForAllStages(params.name, inputColors, outputColors, fragments, testGroup.get(), params.failResult, params.failMessageTemplate);
8765 return testGroup.release();
8768 // Test for the OpSMod instruction.
8769 tcu::TestCaseGroup* createOpSModGraphicsTests(tcu::TestContext& testCtx, qpTestResult negFailResult)
8771 de::MovePtr<tcu::TestCaseGroup> testGroup(new tcu::TestCaseGroup(testCtx, "smod", "OpSMod"));
8772 map<string, string> fragments;
8774 fragments["pre_main"] =
8775 "%c_f32_255 = OpConstant %f32 255.0\n"
8776 "%c_i32_128 = OpConstant %i32 128\n"
8777 "%c_i32_255 = OpConstant %i32 255\n"
8778 "%c_v4f32_255 = OpConstantComposite %v4f32 %c_f32_255 %c_f32_255 %c_f32_255 %c_f32_255 \n"
8779 "%c_v4f32_0_5 = OpConstantComposite %v4f32 %c_f32_0_5 %c_f32_0_5 %c_f32_0_5 %c_f32_0_5 \n"
8780 "%c_v4i32_128 = OpConstantComposite %v4i32 %c_i32_128 %c_i32_128 %c_i32_128 %c_i32_128 \n";
8782 // The test does the following.
8783 // ivec4 ints = int(param1 * 255.0 + 0.5) - 128;
8784 // ivec4 result = ivec4(smod(ints.x, ints.y), smod(ints.y, ints.z), smod(ints.z, ints.x), 255);
8785 // return float(result + 128) / 255.0;
8786 fragments["testfun"] =
8787 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
8788 "%param1 = OpFunctionParameter %v4f32\n"
8789 "%label_testfun = OpLabel\n"
8790 "%div255 = OpFMul %v4f32 %param1 %c_v4f32_255\n"
8791 "%add0_5 = OpFAdd %v4f32 %div255 %c_v4f32_0_5\n"
8792 "%uints_in = OpConvertFToS %v4i32 %add0_5\n"
8793 "%ints_in = OpISub %v4i32 %uints_in %c_v4i32_128\n"
8794 "%x_in = OpCompositeExtract %i32 %ints_in 0\n"
8795 "%y_in = OpCompositeExtract %i32 %ints_in 1\n"
8796 "%z_in = OpCompositeExtract %i32 %ints_in 2\n"
8797 "%x_out = OpSMod %i32 %x_in %y_in\n"
8798 "%y_out = OpSMod %i32 %y_in %z_in\n"
8799 "%z_out = OpSMod %i32 %z_in %x_in\n"
8800 "%ints_out = OpCompositeConstruct %v4i32 %x_out %y_out %z_out %c_i32_255\n"
8801 "%ints_offset = OpIAdd %v4i32 %ints_out %c_v4i32_128\n"
8802 "%f_ints_offset = OpConvertSToF %v4f32 %ints_offset\n"
8803 "%float_out = OpFDiv %v4f32 %f_ints_offset %c_v4f32_255\n"
8804 "OpReturnValue %float_out\n"
8807 const struct CaseParams
8810 const char* failMessageTemplate; // customized status message
8811 qpTestResult failResult; // override status on failure
8812 int operands[4][3]; // four (x, y, z) vectors of operands
8813 int results[4][3]; // four (x, y, z) vectors of results
8819 QP_TEST_RESULT_FAIL,
8820 { { 5, 12, 17 }, { 5, 5, 7 }, { 75, 8, 81 }, { 25, 60, 100 } }, // operands
8821 { { 5, 12, 2 }, { 0, 5, 2 }, { 3, 8, 6 }, { 25, 60, 0 } }, // results
8825 "Inconsistent results, but within specification: ${reason}",
8826 negFailResult, // negative operands, not required by the spec
8827 { { 5, 12, -17 }, { -5, -5, 7 }, { 75, 8, -81 }, { 25, -60, 100 } }, // operands
8828 { { 5, -5, 3 }, { 0, 2, -3 }, { 3, -73, 69 }, { -35, 40, 0 } }, // results
8831 // If either operand is negative the result is undefined. Some implementations may still return correct values.
8833 for (int caseNdx = 0; caseNdx < DE_LENGTH_OF_ARRAY(cases); ++caseNdx)
8835 const CaseParams& params = cases[caseNdx];
8836 RGBA inputColors[4];
8837 RGBA outputColors[4];
8839 for (int i = 0; i < 4; ++i)
8841 inputColors [i] = RGBA(params.operands[i][0] + 128, params.operands[i][1] + 128, params.operands[i][2] + 128, 255);
8842 outputColors[i] = RGBA(params.results [i][0] + 128, params.results [i][1] + 128, params.results [i][2] + 128, 255);
8845 createTestsForAllStages(params.name, inputColors, outputColors, fragments, testGroup.get(), params.failResult, params.failMessageTemplate);
8847 return testGroup.release();
8850 enum ConversionDataType
8853 DATA_TYPE_SIGNED_16,
8854 DATA_TYPE_SIGNED_32,
8855 DATA_TYPE_SIGNED_64,
8856 DATA_TYPE_UNSIGNED_8,
8857 DATA_TYPE_UNSIGNED_16,
8858 DATA_TYPE_UNSIGNED_32,
8859 DATA_TYPE_UNSIGNED_64,
8863 DATA_TYPE_VEC2_SIGNED_16,
8864 DATA_TYPE_VEC2_SIGNED_32
8867 const string getBitWidthStr (ConversionDataType type)
8871 case DATA_TYPE_SIGNED_8:
8872 case DATA_TYPE_UNSIGNED_8:
8875 case DATA_TYPE_SIGNED_16:
8876 case DATA_TYPE_UNSIGNED_16:
8877 case DATA_TYPE_FLOAT_16:
8880 case DATA_TYPE_SIGNED_32:
8881 case DATA_TYPE_UNSIGNED_32:
8882 case DATA_TYPE_FLOAT_32:
8883 case DATA_TYPE_VEC2_SIGNED_16:
8886 case DATA_TYPE_SIGNED_64:
8887 case DATA_TYPE_UNSIGNED_64:
8888 case DATA_TYPE_FLOAT_64:
8889 case DATA_TYPE_VEC2_SIGNED_32:
8898 const string getByteWidthStr (ConversionDataType type)
8902 case DATA_TYPE_SIGNED_8:
8903 case DATA_TYPE_UNSIGNED_8:
8906 case DATA_TYPE_SIGNED_16:
8907 case DATA_TYPE_UNSIGNED_16:
8908 case DATA_TYPE_FLOAT_16:
8911 case DATA_TYPE_SIGNED_32:
8912 case DATA_TYPE_UNSIGNED_32:
8913 case DATA_TYPE_FLOAT_32:
8914 case DATA_TYPE_VEC2_SIGNED_16:
8917 case DATA_TYPE_SIGNED_64:
8918 case DATA_TYPE_UNSIGNED_64:
8919 case DATA_TYPE_FLOAT_64:
8920 case DATA_TYPE_VEC2_SIGNED_32:
8929 bool isSigned (ConversionDataType type)
8933 case DATA_TYPE_SIGNED_8:
8934 case DATA_TYPE_SIGNED_16:
8935 case DATA_TYPE_SIGNED_32:
8936 case DATA_TYPE_SIGNED_64:
8937 case DATA_TYPE_FLOAT_16:
8938 case DATA_TYPE_FLOAT_32:
8939 case DATA_TYPE_FLOAT_64:
8940 case DATA_TYPE_VEC2_SIGNED_16:
8941 case DATA_TYPE_VEC2_SIGNED_32:
8944 case DATA_TYPE_UNSIGNED_8:
8945 case DATA_TYPE_UNSIGNED_16:
8946 case DATA_TYPE_UNSIGNED_32:
8947 case DATA_TYPE_UNSIGNED_64:
8956 bool isInt (ConversionDataType type)
8960 case DATA_TYPE_SIGNED_8:
8961 case DATA_TYPE_SIGNED_16:
8962 case DATA_TYPE_SIGNED_32:
8963 case DATA_TYPE_SIGNED_64:
8964 case DATA_TYPE_UNSIGNED_8:
8965 case DATA_TYPE_UNSIGNED_16:
8966 case DATA_TYPE_UNSIGNED_32:
8967 case DATA_TYPE_UNSIGNED_64:
8970 case DATA_TYPE_FLOAT_16:
8971 case DATA_TYPE_FLOAT_32:
8972 case DATA_TYPE_FLOAT_64:
8973 case DATA_TYPE_VEC2_SIGNED_16:
8974 case DATA_TYPE_VEC2_SIGNED_32:
8983 bool isFloat (ConversionDataType type)
8987 case DATA_TYPE_SIGNED_8:
8988 case DATA_TYPE_SIGNED_16:
8989 case DATA_TYPE_SIGNED_32:
8990 case DATA_TYPE_SIGNED_64:
8991 case DATA_TYPE_UNSIGNED_8:
8992 case DATA_TYPE_UNSIGNED_16:
8993 case DATA_TYPE_UNSIGNED_32:
8994 case DATA_TYPE_UNSIGNED_64:
8995 case DATA_TYPE_VEC2_SIGNED_16:
8996 case DATA_TYPE_VEC2_SIGNED_32:
8999 case DATA_TYPE_FLOAT_16:
9000 case DATA_TYPE_FLOAT_32:
9001 case DATA_TYPE_FLOAT_64:
9010 const string getTypeName (ConversionDataType type)
9012 string prefix = isSigned(type) ? "" : "u";
9014 if (isInt(type)) return prefix + "int" + getBitWidthStr(type);
9015 else if (isFloat(type)) return prefix + "float" + getBitWidthStr(type);
9016 else if (type == DATA_TYPE_VEC2_SIGNED_16) return "i16vec2";
9017 else if (type == DATA_TYPE_VEC2_SIGNED_32) return "i32vec2";
9018 else DE_ASSERT(false);
9023 const string getTestName (ConversionDataType from, ConversionDataType to, const char* suffix)
9025 const string fullSuffix(suffix == DE_NULL ? "" : string("_") + string(suffix));
9027 return getTypeName(from) + "_to_" + getTypeName(to) + fullSuffix;
9030 const string getAsmTypeName (ConversionDataType type)
9034 if (isInt(type)) prefix = isSigned(type) ? "i" : "u";
9035 else if (isFloat(type)) prefix = "f";
9036 else if (type == DATA_TYPE_VEC2_SIGNED_16) return "i16vec2";
9037 else if (type == DATA_TYPE_VEC2_SIGNED_32) return "v2i32";
9038 else DE_ASSERT(false);
9040 return prefix + getBitWidthStr(type);
9043 template<typename T>
9044 BufferSp getSpecializedBuffer (deInt64 number)
9046 return BufferSp(new Buffer<T>(vector<T>(1, (T)number)));
9049 BufferSp getBuffer (ConversionDataType type, deInt64 number)
9053 case DATA_TYPE_SIGNED_8: return getSpecializedBuffer<deInt8>(number);
9054 case DATA_TYPE_SIGNED_16: return getSpecializedBuffer<deInt16>(number);
9055 case DATA_TYPE_SIGNED_32: return getSpecializedBuffer<deInt32>(number);
9056 case DATA_TYPE_SIGNED_64: return getSpecializedBuffer<deInt64>(number);
9057 case DATA_TYPE_UNSIGNED_8: return getSpecializedBuffer<deUint8>(number);
9058 case DATA_TYPE_UNSIGNED_16: return getSpecializedBuffer<deUint16>(number);
9059 case DATA_TYPE_UNSIGNED_32: return getSpecializedBuffer<deUint32>(number);
9060 case DATA_TYPE_UNSIGNED_64: return getSpecializedBuffer<deUint64>(number);
9061 case DATA_TYPE_FLOAT_16: return getSpecializedBuffer<deUint16>(number);
9062 case DATA_TYPE_FLOAT_32: return getSpecializedBuffer<deUint32>(number);
9063 case DATA_TYPE_FLOAT_64: return getSpecializedBuffer<deUint64>(number);
9064 case DATA_TYPE_VEC2_SIGNED_16: return getSpecializedBuffer<deUint32>(number);
9065 case DATA_TYPE_VEC2_SIGNED_32: return getSpecializedBuffer<deUint64>(number);
9067 default: TCU_THROW(InternalError, "Unimplemented type passed");
9071 bool usesInt8 (ConversionDataType from, ConversionDataType to)
9073 return (from == DATA_TYPE_SIGNED_8 || to == DATA_TYPE_SIGNED_8 ||
9074 from == DATA_TYPE_UNSIGNED_8 || to == DATA_TYPE_UNSIGNED_8);
9077 bool usesInt16 (ConversionDataType from, ConversionDataType to)
9079 return (from == DATA_TYPE_SIGNED_16 || to == DATA_TYPE_SIGNED_16 ||
9080 from == DATA_TYPE_UNSIGNED_16 || to == DATA_TYPE_UNSIGNED_16 ||
9081 from == DATA_TYPE_VEC2_SIGNED_16 || to == DATA_TYPE_VEC2_SIGNED_16);
9084 bool usesInt32 (ConversionDataType from, ConversionDataType to)
9086 return (from == DATA_TYPE_SIGNED_32 || to == DATA_TYPE_SIGNED_32 ||
9087 from == DATA_TYPE_UNSIGNED_32 || to == DATA_TYPE_UNSIGNED_32 ||
9088 from == DATA_TYPE_VEC2_SIGNED_32|| to == DATA_TYPE_VEC2_SIGNED_32);
9091 bool usesInt64 (ConversionDataType from, ConversionDataType to)
9093 return (from == DATA_TYPE_SIGNED_64 || to == DATA_TYPE_SIGNED_64 ||
9094 from == DATA_TYPE_UNSIGNED_64 || to == DATA_TYPE_UNSIGNED_64);
9097 bool usesFloat16 (ConversionDataType from, ConversionDataType to)
9099 return (from == DATA_TYPE_FLOAT_16 || to == DATA_TYPE_FLOAT_16);
9102 bool usesFloat32 (ConversionDataType from, ConversionDataType to)
9104 return (from == DATA_TYPE_FLOAT_32 || to == DATA_TYPE_FLOAT_32);
9107 bool usesFloat64 (ConversionDataType from, ConversionDataType to)
9109 return (from == DATA_TYPE_FLOAT_64 || to == DATA_TYPE_FLOAT_64);
9112 void getVulkanFeaturesAndExtensions (ConversionDataType from, ConversionDataType to, VulkanFeatures& vulkanFeatures, vector<string>& extensions)
9114 if (usesInt16(from, to) && !usesInt32(from, to))
9115 vulkanFeatures.coreFeatures.shaderInt16 = DE_TRUE;
9117 if (usesInt64(from, to))
9118 vulkanFeatures.coreFeatures.shaderInt64 = DE_TRUE;
9120 if (usesFloat64(from, to))
9121 vulkanFeatures.coreFeatures.shaderFloat64 = DE_TRUE;
9123 if (usesInt16(from, to) || usesFloat16(from, to))
9125 extensions.push_back("VK_KHR_16bit_storage");
9126 vulkanFeatures.ext16BitStorage |= EXT16BITSTORAGEFEATURES_UNIFORM_BUFFER_BLOCK;
9129 if (usesFloat16(from, to) || usesInt8(from, to))
9131 extensions.push_back("VK_KHR_shader_float16_int8");
9133 if (usesFloat16(from, to))
9135 vulkanFeatures.extFloat16Int8 |= EXTFLOAT16INT8FEATURES_FLOAT16;
9138 if (usesInt8(from, to))
9140 vulkanFeatures.extFloat16Int8 |= EXTFLOAT16INT8FEATURES_INT8;
9142 extensions.push_back("VK_KHR_8bit_storage");
9143 vulkanFeatures.ext8BitStorage |= EXT8BITSTORAGEFEATURES_STORAGE_BUFFER;
9150 ConvertCase (const string& instruction, ConversionDataType from, ConversionDataType to, deInt64 number, bool separateOutput = false, deInt64 outputNumber = 0, const char* suffix = DE_NULL)
9153 , m_name (getTestName(from, to, suffix))
9154 , m_inputBuffer (getBuffer(from, number))
9160 m_asmTypes["inputType"] = getAsmTypeName(from);
9161 m_asmTypes["outputType"] = getAsmTypeName(to);
9164 m_outputBuffer = getBuffer(to, outputNumber);
9166 m_outputBuffer = getBuffer(to, number);
9168 if (usesInt8(from, to))
9170 bool requiresInt8Capability = true;
9171 if (instruction == "OpUConvert" || instruction == "OpSConvert")
9173 // Conversions between 8 and 32 bit are provided by SPV_KHR_8bit_storage. The rest requires explicit Int8
9174 if (usesInt32(from, to))
9175 requiresInt8Capability = false;
9178 caps += "OpCapability StorageBuffer8BitAccess\n";
9179 if (requiresInt8Capability)
9180 caps += "OpCapability Int8\n";
9182 decl += "%i8 = OpTypeInt 8 1\n"
9183 "%u8 = OpTypeInt 8 0\n";
9184 exts += "OpExtension \"SPV_KHR_8bit_storage\"\n";
9187 if (usesInt16(from, to))
9189 bool requiresInt16Capability = true;
9191 if (instruction == "OpUConvert" || instruction == "OpSConvert" || instruction == "OpFConvert")
9193 // Conversions between 16 and 32 bit are provided by SPV_KHR_16bit_storage. The rest requires explicit Int16
9194 if (usesInt32(from, to) || usesFloat32(from, to))
9195 requiresInt16Capability = false;
9198 decl += "%i16 = OpTypeInt 16 1\n"
9199 "%u16 = OpTypeInt 16 0\n"
9200 "%i16vec2 = OpTypeVector %i16 2\n";
9202 // Conversions between 16 and 32 bit are provided by SPV_KHR_16bit_storage. The rest requires explicit Int16
9203 if (requiresInt16Capability)
9204 caps += "OpCapability Int16\n";
9207 if (usesFloat16(from, to))
9209 decl += "%f16 = OpTypeFloat 16\n";
9211 // Conversions between 16 and 32 bit are provided by SPV_KHR_16bit_storage. The rest requires explicit Float16
9212 if (!(usesInt32(from, to) || usesFloat32(from, to)))
9213 caps += "OpCapability Float16\n";
9216 if (usesInt16(from, to) || usesFloat16(from, to))
9218 caps += "OpCapability StorageUniformBufferBlock16\n";
9219 exts += "OpExtension \"SPV_KHR_16bit_storage\"\n";
9222 if (usesInt64(from, to))
9224 caps += "OpCapability Int64\n";
9225 decl += "%i64 = OpTypeInt 64 1\n"
9226 "%u64 = OpTypeInt 64 0\n";
9229 if (usesFloat64(from, to))
9231 caps += "OpCapability Float64\n";
9232 decl += "%f64 = OpTypeFloat 64\n";
9235 m_asmTypes["datatype_capabilities"] = caps;
9236 m_asmTypes["datatype_additional_decl"] = decl;
9237 m_asmTypes["datatype_extensions"] = exts;
9240 ConversionDataType m_fromType;
9241 ConversionDataType m_toType;
9243 map<string, string> m_asmTypes;
9244 BufferSp m_inputBuffer;
9245 BufferSp m_outputBuffer;
9248 const string getConvertCaseShaderStr (const string& instruction, const ConvertCase& convertCase)
9250 map<string, string> params = convertCase.m_asmTypes;
9252 params["instruction"] = instruction;
9253 params["inDecorator"] = getByteWidthStr(convertCase.m_fromType);
9254 params["outDecorator"] = getByteWidthStr(convertCase.m_toType);
9256 const StringTemplate shader (
9257 "OpCapability Shader\n"
9258 "${datatype_capabilities}"
9259 "${datatype_extensions:opt}"
9260 "OpMemoryModel Logical GLSL450\n"
9261 "OpEntryPoint GLCompute %main \"main\"\n"
9262 "OpExecutionMode %main LocalSize 1 1 1\n"
9263 "OpSource GLSL 430\n"
9264 "OpName %main \"main\"\n"
9266 "OpDecorate %indata DescriptorSet 0\n"
9267 "OpDecorate %indata Binding 0\n"
9268 "OpDecorate %outdata DescriptorSet 0\n"
9269 "OpDecorate %outdata Binding 1\n"
9270 "OpDecorate %in_buf BufferBlock\n"
9271 "OpDecorate %out_buf BufferBlock\n"
9272 "OpMemberDecorate %in_buf 0 Offset 0\n"
9273 "OpMemberDecorate %out_buf 0 Offset 0\n"
9275 "%void = OpTypeVoid\n"
9276 "%voidf = OpTypeFunction %void\n"
9277 "%u32 = OpTypeInt 32 0\n"
9278 "%i32 = OpTypeInt 32 1\n"
9279 "%f32 = OpTypeFloat 32\n"
9280 "%v2i32 = OpTypeVector %i32 2\n"
9281 "${datatype_additional_decl}"
9282 "%uvec3 = OpTypeVector %u32 3\n"
9284 "%in_ptr = OpTypePointer Uniform %${inputType}\n"
9285 "%out_ptr = OpTypePointer Uniform %${outputType}\n"
9286 "%in_buf = OpTypeStruct %${inputType}\n"
9287 "%out_buf = OpTypeStruct %${outputType}\n"
9288 "%in_bufptr = OpTypePointer Uniform %in_buf\n"
9289 "%out_bufptr = OpTypePointer Uniform %out_buf\n"
9290 "%indata = OpVariable %in_bufptr Uniform\n"
9291 "%outdata = OpVariable %out_bufptr Uniform\n"
9293 "%zero = OpConstant %i32 0\n"
9295 "%main = OpFunction %void None %voidf\n"
9296 "%label = OpLabel\n"
9297 "%inloc = OpAccessChain %in_ptr %indata %zero\n"
9298 "%outloc = OpAccessChain %out_ptr %outdata %zero\n"
9299 "%inval = OpLoad %${inputType} %inloc\n"
9300 "%conv = ${instruction} %${outputType} %inval\n"
9301 " OpStore %outloc %conv\n"
9306 return shader.specialize(params);
9309 void createConvertCases (vector<ConvertCase>& testCases, const string& instruction)
9311 if (instruction == "OpUConvert")
9313 // Convert unsigned int to unsigned int
9314 testCases.push_back(ConvertCase(instruction, DATA_TYPE_UNSIGNED_8, DATA_TYPE_UNSIGNED_16, 42));
9315 testCases.push_back(ConvertCase(instruction, DATA_TYPE_UNSIGNED_8, DATA_TYPE_UNSIGNED_32, 73));
9316 testCases.push_back(ConvertCase(instruction, DATA_TYPE_UNSIGNED_8, DATA_TYPE_UNSIGNED_64, 121));
9318 testCases.push_back(ConvertCase(instruction, DATA_TYPE_UNSIGNED_16, DATA_TYPE_UNSIGNED_8, 33));
9319 testCases.push_back(ConvertCase(instruction, DATA_TYPE_UNSIGNED_16, DATA_TYPE_UNSIGNED_32, 60653));
9320 testCases.push_back(ConvertCase(instruction, DATA_TYPE_UNSIGNED_16, DATA_TYPE_UNSIGNED_64, 17991));
9322 testCases.push_back(ConvertCase(instruction, DATA_TYPE_UNSIGNED_32, DATA_TYPE_UNSIGNED_64, 904256275));
9323 testCases.push_back(ConvertCase(instruction, DATA_TYPE_UNSIGNED_32, DATA_TYPE_UNSIGNED_16, 6275));
9324 testCases.push_back(ConvertCase(instruction, DATA_TYPE_UNSIGNED_32, DATA_TYPE_UNSIGNED_8, 17));
9326 testCases.push_back(ConvertCase(instruction, DATA_TYPE_UNSIGNED_64, DATA_TYPE_UNSIGNED_32, 701256243));
9327 testCases.push_back(ConvertCase(instruction, DATA_TYPE_UNSIGNED_64, DATA_TYPE_UNSIGNED_16, 4741));
9328 testCases.push_back(ConvertCase(instruction, DATA_TYPE_UNSIGNED_64, DATA_TYPE_UNSIGNED_8, 65));
9330 else if (instruction == "OpSConvert")
9332 // Sign extension int->int
9333 testCases.push_back(ConvertCase(instruction, DATA_TYPE_SIGNED_8, DATA_TYPE_SIGNED_16, -30));
9334 testCases.push_back(ConvertCase(instruction, DATA_TYPE_SIGNED_8, DATA_TYPE_SIGNED_32, 55));
9335 testCases.push_back(ConvertCase(instruction, DATA_TYPE_SIGNED_8, DATA_TYPE_SIGNED_64, -3));
9336 testCases.push_back(ConvertCase(instruction, DATA_TYPE_SIGNED_16, DATA_TYPE_SIGNED_32, 14669));
9337 testCases.push_back(ConvertCase(instruction, DATA_TYPE_SIGNED_16, DATA_TYPE_SIGNED_64, -3341));
9338 testCases.push_back(ConvertCase(instruction, DATA_TYPE_SIGNED_32, DATA_TYPE_SIGNED_64, 973610259));
9340 // Truncate for int->int
9341 testCases.push_back(ConvertCase(instruction, DATA_TYPE_SIGNED_16, DATA_TYPE_SIGNED_8, 81));
9342 testCases.push_back(ConvertCase(instruction, DATA_TYPE_SIGNED_32, DATA_TYPE_SIGNED_8, -93));
9343 testCases.push_back(ConvertCase(instruction, DATA_TYPE_SIGNED_64, DATA_TYPE_SIGNED_8, 3182748172687672ll, true, 56));
9344 testCases.push_back(ConvertCase(instruction, DATA_TYPE_SIGNED_32, DATA_TYPE_SIGNED_16, 12382));
9345 testCases.push_back(ConvertCase(instruction, DATA_TYPE_SIGNED_64, DATA_TYPE_SIGNED_32, -972812359));
9346 testCases.push_back(ConvertCase(instruction, DATA_TYPE_SIGNED_64, DATA_TYPE_SIGNED_16, -1067742499291926803ll, true, -4371));
9348 // Sign extension for int->uint
9349 testCases.push_back(ConvertCase(instruction, DATA_TYPE_SIGNED_8, DATA_TYPE_UNSIGNED_16, 56));
9350 testCases.push_back(ConvertCase(instruction, DATA_TYPE_SIGNED_8, DATA_TYPE_UNSIGNED_32, -47, true, 4294967249u));
9351 testCases.push_back(ConvertCase(instruction, DATA_TYPE_SIGNED_8, DATA_TYPE_UNSIGNED_64, -5, true, 18446744073709551611ull));
9352 testCases.push_back(ConvertCase(instruction, DATA_TYPE_SIGNED_16, DATA_TYPE_UNSIGNED_32, 14669));
9353 testCases.push_back(ConvertCase(instruction, DATA_TYPE_SIGNED_16, DATA_TYPE_UNSIGNED_64, -3341, true, 18446744073709548275ull));
9354 testCases.push_back(ConvertCase(instruction, DATA_TYPE_SIGNED_32, DATA_TYPE_UNSIGNED_64, 973610259));
9356 // Truncate for int->uint
9357 testCases.push_back(ConvertCase(instruction, DATA_TYPE_SIGNED_16, DATA_TYPE_UNSIGNED_8, -25711, true, 145));
9358 testCases.push_back(ConvertCase(instruction, DATA_TYPE_SIGNED_32, DATA_TYPE_UNSIGNED_8, 103));
9359 testCases.push_back(ConvertCase(instruction, DATA_TYPE_SIGNED_64, DATA_TYPE_UNSIGNED_8, -1067742499291926803ll, true, 61165));
9360 testCases.push_back(ConvertCase(instruction, DATA_TYPE_SIGNED_32, DATA_TYPE_UNSIGNED_16, 12382));
9361 testCases.push_back(ConvertCase(instruction, DATA_TYPE_SIGNED_64, DATA_TYPE_UNSIGNED_32, -972812359, true, 3322154937u));
9362 testCases.push_back(ConvertCase(instruction, DATA_TYPE_SIGNED_64, DATA_TYPE_UNSIGNED_16, -1067742499291926803ll, true, 61165));
9364 // Sign extension for uint->int
9365 testCases.push_back(ConvertCase(instruction, DATA_TYPE_UNSIGNED_8, DATA_TYPE_SIGNED_16, 71));
9366 testCases.push_back(ConvertCase(instruction, DATA_TYPE_UNSIGNED_8, DATA_TYPE_SIGNED_32, 201, true, -55));
9367 testCases.push_back(ConvertCase(instruction, DATA_TYPE_UNSIGNED_8, DATA_TYPE_SIGNED_64, 188, true, -68));
9368 testCases.push_back(ConvertCase(instruction, DATA_TYPE_UNSIGNED_16, DATA_TYPE_SIGNED_32, 14669));
9369 testCases.push_back(ConvertCase(instruction, DATA_TYPE_UNSIGNED_16, DATA_TYPE_SIGNED_64, 62195, true, -3341));
9370 testCases.push_back(ConvertCase(instruction, DATA_TYPE_UNSIGNED_32, DATA_TYPE_SIGNED_64, 973610259));
9372 // Truncate for uint->int
9373 testCases.push_back(ConvertCase(instruction, DATA_TYPE_UNSIGNED_16, DATA_TYPE_SIGNED_8, 67));
9374 testCases.push_back(ConvertCase(instruction, DATA_TYPE_UNSIGNED_32, DATA_TYPE_SIGNED_8, 133, true, -123));
9375 testCases.push_back(ConvertCase(instruction, DATA_TYPE_UNSIGNED_64, DATA_TYPE_SIGNED_8, 836927654193256494ull, true, 46));
9376 testCases.push_back(ConvertCase(instruction, DATA_TYPE_UNSIGNED_32, DATA_TYPE_SIGNED_16, 12382));
9377 testCases.push_back(ConvertCase(instruction, DATA_TYPE_UNSIGNED_64, DATA_TYPE_SIGNED_32, 18446744072736739257ull, true, -972812359));
9378 testCases.push_back(ConvertCase(instruction, DATA_TYPE_UNSIGNED_64, DATA_TYPE_SIGNED_16, 17379001574417624813ull, true, -4371));
9380 // Convert i16vec2 to i32vec2 and vice versa
9381 // Unsigned values are used here to represent negative signed values and to allow defined shifting behaviour.
9382 // The actual signed value -32123 is used here as uint16 value 33413 and uint32 value 4294935173
9383 testCases.push_back(ConvertCase(instruction, DATA_TYPE_VEC2_SIGNED_16, DATA_TYPE_VEC2_SIGNED_32, (33413u << 16) | 27593, true, (4294935173ull << 32) | 27593));
9384 testCases.push_back(ConvertCase(instruction, DATA_TYPE_VEC2_SIGNED_32, DATA_TYPE_VEC2_SIGNED_16, (4294935173ull << 32) | 27593, true, (33413u << 16) | 27593));
9386 else if (instruction == "OpFConvert")
9388 // All hexadecimal values below represent 1234.0 as 16/32/64-bit IEEE 754 float
9389 testCases.push_back(ConvertCase(instruction, DATA_TYPE_FLOAT_32, DATA_TYPE_FLOAT_64, 0x449a4000, true, 0x4093480000000000));
9390 testCases.push_back(ConvertCase(instruction, DATA_TYPE_FLOAT_64, DATA_TYPE_FLOAT_32, 0x4093480000000000, true, 0x449a4000));
9392 testCases.push_back(ConvertCase(instruction, DATA_TYPE_FLOAT_32, DATA_TYPE_FLOAT_16, 0x449a4000, true, 0x64D2));
9393 testCases.push_back(ConvertCase(instruction, DATA_TYPE_FLOAT_16, DATA_TYPE_FLOAT_32, 0x64D2, true, 0x449a4000));
9395 testCases.push_back(ConvertCase(instruction, DATA_TYPE_FLOAT_16, DATA_TYPE_FLOAT_64, 0x64D2, true, 0x4093480000000000));
9396 testCases.push_back(ConvertCase(instruction, DATA_TYPE_FLOAT_64, DATA_TYPE_FLOAT_16, 0x4093480000000000, true, 0x64D2));
9398 else if (instruction == "OpConvertFToU")
9400 // Normal numbers from uint8 range
9401 testCases.push_back(ConvertCase(instruction, DATA_TYPE_FLOAT_16, DATA_TYPE_UNSIGNED_8, 0x5020, true, 33, "33"));
9402 testCases.push_back(ConvertCase(instruction, DATA_TYPE_FLOAT_32, DATA_TYPE_UNSIGNED_8, 0x42280000, true, 42, "42"));
9403 testCases.push_back(ConvertCase(instruction, DATA_TYPE_FLOAT_64, DATA_TYPE_UNSIGNED_8, 0x4067800000000000ull, true, 188, "188"));
9405 // Maximum uint8 value
9406 testCases.push_back(ConvertCase(instruction, DATA_TYPE_FLOAT_16, DATA_TYPE_UNSIGNED_8, 0x5BF8, true, 255, "max"));
9407 testCases.push_back(ConvertCase(instruction, DATA_TYPE_FLOAT_32, DATA_TYPE_UNSIGNED_8, 0x437F0000, true, 255, "max"));
9408 testCases.push_back(ConvertCase(instruction, DATA_TYPE_FLOAT_64, DATA_TYPE_UNSIGNED_8, 0x406FE00000000000ull, true, 255, "max"));
9411 testCases.push_back(ConvertCase(instruction, DATA_TYPE_FLOAT_16, DATA_TYPE_UNSIGNED_8, 0x0000, true, 0, "p0"));
9412 testCases.push_back(ConvertCase(instruction, DATA_TYPE_FLOAT_32, DATA_TYPE_UNSIGNED_8, 0x00000000, true, 0, "p0"));
9413 testCases.push_back(ConvertCase(instruction, DATA_TYPE_FLOAT_64, DATA_TYPE_UNSIGNED_8, 0x0000000000000000ull, true, 0, "p0"));
9416 testCases.push_back(ConvertCase(instruction, DATA_TYPE_FLOAT_16, DATA_TYPE_UNSIGNED_8, 0x8000, true, 0, "m0"));
9417 testCases.push_back(ConvertCase(instruction, DATA_TYPE_FLOAT_32, DATA_TYPE_UNSIGNED_8, 0x80000000, true, 0, "m0"));
9418 testCases.push_back(ConvertCase(instruction, DATA_TYPE_FLOAT_64, DATA_TYPE_UNSIGNED_8, 0x8000000000000000ull, true, 0, "m0"));
9420 // All hexadecimal values below represent 1234.0 as 16/32/64-bit IEEE 754 float
9421 testCases.push_back(ConvertCase(instruction, DATA_TYPE_FLOAT_16, DATA_TYPE_UNSIGNED_16, 0x64D2, true, 1234, "1234"));
9422 testCases.push_back(ConvertCase(instruction, DATA_TYPE_FLOAT_16, DATA_TYPE_UNSIGNED_32, 0x64D2, true, 1234, "1234"));
9423 testCases.push_back(ConvertCase(instruction, DATA_TYPE_FLOAT_16, DATA_TYPE_UNSIGNED_64, 0x64D2, true, 1234, "1234"));
9425 // 0x7BFF = 0111 1011 1111 1111 = 0 11110 1111111111 = 65504
9426 testCases.push_back(ConvertCase(instruction, DATA_TYPE_FLOAT_16, DATA_TYPE_UNSIGNED_16, 0x7BFF, true, 65504, "max"));
9427 testCases.push_back(ConvertCase(instruction, DATA_TYPE_FLOAT_16, DATA_TYPE_UNSIGNED_32, 0x7BFF, true, 65504, "max"));
9428 testCases.push_back(ConvertCase(instruction, DATA_TYPE_FLOAT_16, DATA_TYPE_UNSIGNED_64, 0x7BFF, true, 65504, "max"));
9431 testCases.push_back(ConvertCase(instruction, DATA_TYPE_FLOAT_16, DATA_TYPE_UNSIGNED_32, 0x0000, true, 0, "p0"));
9432 testCases.push_back(ConvertCase(instruction, DATA_TYPE_FLOAT_16, DATA_TYPE_UNSIGNED_16, 0x0000, true, 0, "p0"));
9433 testCases.push_back(ConvertCase(instruction, DATA_TYPE_FLOAT_16, DATA_TYPE_UNSIGNED_64, 0x0000, true, 0, "p0"));
9436 testCases.push_back(ConvertCase(instruction, DATA_TYPE_FLOAT_16, DATA_TYPE_UNSIGNED_16, 0x8000, true, 0, "m0"));
9437 testCases.push_back(ConvertCase(instruction, DATA_TYPE_FLOAT_16, DATA_TYPE_UNSIGNED_32, 0x8000, true, 0, "m0"));
9438 testCases.push_back(ConvertCase(instruction, DATA_TYPE_FLOAT_16, DATA_TYPE_UNSIGNED_64, 0x8000, true, 0, "m0"));
9440 testCases.push_back(ConvertCase(instruction, DATA_TYPE_FLOAT_32, DATA_TYPE_UNSIGNED_16, 0x449a4000, true, 1234));
9441 testCases.push_back(ConvertCase(instruction, DATA_TYPE_FLOAT_32, DATA_TYPE_UNSIGNED_32, 0x449a4000, true, 1234));
9442 testCases.push_back(ConvertCase(instruction, DATA_TYPE_FLOAT_32, DATA_TYPE_UNSIGNED_64, 0x449a4000, true, 1234));
9443 testCases.push_back(ConvertCase(instruction, DATA_TYPE_FLOAT_64, DATA_TYPE_UNSIGNED_16, 0x4093480000000000, true, 1234));
9444 testCases.push_back(ConvertCase(instruction, DATA_TYPE_FLOAT_64, DATA_TYPE_UNSIGNED_32, 0x4093480000000000, true, 1234));
9445 testCases.push_back(ConvertCase(instruction, DATA_TYPE_FLOAT_64, DATA_TYPE_UNSIGNED_64, 0x4093480000000000, true, 1234));
9447 else if (instruction == "OpConvertUToF")
9449 // Normal numbers from uint8 range
9450 testCases.push_back(ConvertCase(instruction, DATA_TYPE_UNSIGNED_8, DATA_TYPE_FLOAT_16, 116, true, 0x5740, "116"));
9451 testCases.push_back(ConvertCase(instruction, DATA_TYPE_UNSIGNED_8, DATA_TYPE_FLOAT_32, 232, true, 0x43680000, "232"));
9452 testCases.push_back(ConvertCase(instruction, DATA_TYPE_UNSIGNED_8, DATA_TYPE_FLOAT_64, 164, true, 0x4064800000000000ull, "164"));
9454 // Maximum uint8 value
9455 testCases.push_back(ConvertCase(instruction, DATA_TYPE_UNSIGNED_8, DATA_TYPE_FLOAT_16, 255, true, 0x5BF8, "max"));
9456 testCases.push_back(ConvertCase(instruction, DATA_TYPE_UNSIGNED_8, DATA_TYPE_FLOAT_32, 255, true, 0x437F0000, "max"));
9457 testCases.push_back(ConvertCase(instruction, DATA_TYPE_UNSIGNED_8, DATA_TYPE_FLOAT_64, 255, true, 0x406FE00000000000ull, "max"));
9459 // All hexadecimal values below represent 1234.0 as 32/64-bit IEEE 754 float
9460 testCases.push_back(ConvertCase(instruction, DATA_TYPE_UNSIGNED_16, DATA_TYPE_FLOAT_16, 1234, true, 0x64D2, "1234"));
9461 testCases.push_back(ConvertCase(instruction, DATA_TYPE_UNSIGNED_32, DATA_TYPE_FLOAT_16, 1234, true, 0x64D2, "1234"));
9462 testCases.push_back(ConvertCase(instruction, DATA_TYPE_UNSIGNED_64, DATA_TYPE_FLOAT_16, 1234, true, 0x64D2, "1234"));
9464 // 0x7BFF = 0111 1011 1111 1111 = 0 11110 1111111111 = 65504
9465 testCases.push_back(ConvertCase(instruction, DATA_TYPE_UNSIGNED_16, DATA_TYPE_FLOAT_16, 65504, true, 0x7BFF, "max"));
9466 testCases.push_back(ConvertCase(instruction, DATA_TYPE_UNSIGNED_32, DATA_TYPE_FLOAT_16, 65504, true, 0x7BFF, "max"));
9467 testCases.push_back(ConvertCase(instruction, DATA_TYPE_UNSIGNED_64, DATA_TYPE_FLOAT_16, 65504, true, 0x7BFF, "max"));
9469 testCases.push_back(ConvertCase(instruction, DATA_TYPE_UNSIGNED_16, DATA_TYPE_FLOAT_32, 1234, true, 0x449a4000));
9470 testCases.push_back(ConvertCase(instruction, DATA_TYPE_UNSIGNED_16, DATA_TYPE_FLOAT_64, 1234, true, 0x4093480000000000));
9471 testCases.push_back(ConvertCase(instruction, DATA_TYPE_UNSIGNED_32, DATA_TYPE_FLOAT_32, 1234, true, 0x449a4000));
9472 testCases.push_back(ConvertCase(instruction, DATA_TYPE_UNSIGNED_32, DATA_TYPE_FLOAT_64, 1234, true, 0x4093480000000000));
9473 testCases.push_back(ConvertCase(instruction, DATA_TYPE_UNSIGNED_64, DATA_TYPE_FLOAT_32, 1234, true, 0x449a4000));
9474 testCases.push_back(ConvertCase(instruction, DATA_TYPE_UNSIGNED_64, DATA_TYPE_FLOAT_64, 1234, true, 0x4093480000000000));
9476 else if (instruction == "OpConvertFToS")
9478 // Normal numbers from int8 range
9479 testCases.push_back(ConvertCase(instruction, DATA_TYPE_FLOAT_16, DATA_TYPE_SIGNED_8, 0xC980, true, -11, "m11"));
9480 testCases.push_back(ConvertCase(instruction, DATA_TYPE_FLOAT_32, DATA_TYPE_SIGNED_8, 0xC2140000, true, -37, "m37"));
9481 testCases.push_back(ConvertCase(instruction, DATA_TYPE_FLOAT_64, DATA_TYPE_SIGNED_8, 0xC050800000000000ull, true, -66, "m66"));
9483 // Minimum int8 value
9484 testCases.push_back(ConvertCase(instruction, DATA_TYPE_FLOAT_16, DATA_TYPE_SIGNED_8, 0xD800, true, -128, "min"));
9485 testCases.push_back(ConvertCase(instruction, DATA_TYPE_FLOAT_32, DATA_TYPE_SIGNED_8, 0xC3000000, true, -128, "min"));
9486 testCases.push_back(ConvertCase(instruction, DATA_TYPE_FLOAT_64, DATA_TYPE_SIGNED_8, 0xC060000000000000ull, true, -128, "min"));
9488 // Maximum int8 value
9489 testCases.push_back(ConvertCase(instruction, DATA_TYPE_FLOAT_16, DATA_TYPE_SIGNED_8, 0x57F0, true, 127, "max"));
9490 testCases.push_back(ConvertCase(instruction, DATA_TYPE_FLOAT_32, DATA_TYPE_SIGNED_8, 0x42FE0000, true, 127, "max"));
9491 testCases.push_back(ConvertCase(instruction, DATA_TYPE_FLOAT_64, DATA_TYPE_SIGNED_8, 0x405FC00000000000ull, true, 127, "max"));
9494 testCases.push_back(ConvertCase(instruction, DATA_TYPE_FLOAT_16, DATA_TYPE_SIGNED_8, 0x0000, true, 0, "p0"));
9495 testCases.push_back(ConvertCase(instruction, DATA_TYPE_FLOAT_32, DATA_TYPE_SIGNED_8, 0x00000000, true, 0, "p0"));
9496 testCases.push_back(ConvertCase(instruction, DATA_TYPE_FLOAT_64, DATA_TYPE_SIGNED_8, 0x0000000000000000ull, true, 0, "p0"));
9499 testCases.push_back(ConvertCase(instruction, DATA_TYPE_FLOAT_16, DATA_TYPE_SIGNED_8, 0x8000, true, 0, "m0"));
9500 testCases.push_back(ConvertCase(instruction, DATA_TYPE_FLOAT_32, DATA_TYPE_SIGNED_8, 0x80000000, true, 0, "m0"));
9501 testCases.push_back(ConvertCase(instruction, DATA_TYPE_FLOAT_64, DATA_TYPE_SIGNED_8, 0x8000000000000000ull, true, 0, "m0"));
9503 // All hexadecimal values below represent -1234.0 as 32/64-bit IEEE 754 float
9504 testCases.push_back(ConvertCase(instruction, DATA_TYPE_FLOAT_16, DATA_TYPE_SIGNED_16, 0xE4D2, true, -1234, "m1234"));
9505 testCases.push_back(ConvertCase(instruction, DATA_TYPE_FLOAT_16, DATA_TYPE_SIGNED_32, 0xE4D2, true, -1234, "m1234"));
9506 testCases.push_back(ConvertCase(instruction, DATA_TYPE_FLOAT_16, DATA_TYPE_SIGNED_64, 0xE4D2, true, -1234, "m1234"));
9508 // 0xF800 = 1111 1000 0000 0000 = 1 11110 0000000000 = -32768
9509 testCases.push_back(ConvertCase(instruction, DATA_TYPE_FLOAT_16, DATA_TYPE_SIGNED_16, 0xF800, true, -32768, "min"));
9510 testCases.push_back(ConvertCase(instruction, DATA_TYPE_FLOAT_16, DATA_TYPE_SIGNED_32, 0xF800, true, -32768, "min"));
9511 testCases.push_back(ConvertCase(instruction, DATA_TYPE_FLOAT_16, DATA_TYPE_SIGNED_64, 0xF800, true, -32768, "min"));
9513 // 0x77FF = 0111 0111 1111 1111 = 0 11101 1111111111 = 32752
9514 testCases.push_back(ConvertCase(instruction, DATA_TYPE_FLOAT_16, DATA_TYPE_SIGNED_16, 0x77FF, true, 32752, "max"));
9515 testCases.push_back(ConvertCase(instruction, DATA_TYPE_FLOAT_16, DATA_TYPE_SIGNED_32, 0x77FF, true, 32752, "max"));
9516 testCases.push_back(ConvertCase(instruction, DATA_TYPE_FLOAT_16, DATA_TYPE_SIGNED_64, 0x77FF, true, 32752, "max"));
9519 testCases.push_back(ConvertCase(instruction, DATA_TYPE_FLOAT_16, DATA_TYPE_SIGNED_16, 0x0000, true, 0, "p0"));
9520 testCases.push_back(ConvertCase(instruction, DATA_TYPE_FLOAT_16, DATA_TYPE_SIGNED_32, 0x0000, true, 0, "p0"));
9521 testCases.push_back(ConvertCase(instruction, DATA_TYPE_FLOAT_16, DATA_TYPE_SIGNED_64, 0x0000, true, 0, "p0"));
9524 testCases.push_back(ConvertCase(instruction, DATA_TYPE_FLOAT_16, DATA_TYPE_SIGNED_16, 0x8000, true, 0, "m0"));
9525 testCases.push_back(ConvertCase(instruction, DATA_TYPE_FLOAT_16, DATA_TYPE_SIGNED_32, 0x8000, true, 0, "m0"));
9526 testCases.push_back(ConvertCase(instruction, DATA_TYPE_FLOAT_16, DATA_TYPE_SIGNED_64, 0x8000, true, 0, "m0"));
9528 testCases.push_back(ConvertCase(instruction, DATA_TYPE_FLOAT_32, DATA_TYPE_SIGNED_16, 0xc49a4000, true, -1234));
9529 testCases.push_back(ConvertCase(instruction, DATA_TYPE_FLOAT_32, DATA_TYPE_SIGNED_32, 0xc49a4000, true, -1234));
9530 testCases.push_back(ConvertCase(instruction, DATA_TYPE_FLOAT_32, DATA_TYPE_SIGNED_64, 0xc49a4000, true, -1234));
9531 testCases.push_back(ConvertCase(instruction, DATA_TYPE_FLOAT_64, DATA_TYPE_SIGNED_16, 0xc093480000000000, true, -1234));
9532 testCases.push_back(ConvertCase(instruction, DATA_TYPE_FLOAT_64, DATA_TYPE_SIGNED_32, 0xc093480000000000, true, -1234));
9533 testCases.push_back(ConvertCase(instruction, DATA_TYPE_FLOAT_64, DATA_TYPE_SIGNED_64, 0xc093480000000000, true, -1234));
9534 testCases.push_back(ConvertCase(instruction, DATA_TYPE_FLOAT_32, DATA_TYPE_SIGNED_16, 0x453b9000, true, 3001, "p3001"));
9535 testCases.push_back(ConvertCase(instruction, DATA_TYPE_FLOAT_32, DATA_TYPE_SIGNED_16, 0xc53b9000, true, -3001, "m3001"));
9537 else if (instruction == "OpConvertSToF")
9539 // Normal numbers from int8 range
9540 testCases.push_back(ConvertCase(instruction, DATA_TYPE_SIGNED_8, DATA_TYPE_FLOAT_16, -12, true, 0xCA00, "m21"));
9541 testCases.push_back(ConvertCase(instruction, DATA_TYPE_SIGNED_8, DATA_TYPE_FLOAT_32, -21, true, 0xC1A80000, "m21"));
9542 testCases.push_back(ConvertCase(instruction, DATA_TYPE_SIGNED_8, DATA_TYPE_FLOAT_64, -99, true, 0xC058C00000000000ull, "m99"));
9544 // Minimum int8 value
9545 testCases.push_back(ConvertCase(instruction, DATA_TYPE_SIGNED_8, DATA_TYPE_FLOAT_16, -128, true, 0xD800, "min"));
9546 testCases.push_back(ConvertCase(instruction, DATA_TYPE_SIGNED_8, DATA_TYPE_FLOAT_32, -128, true, 0xC3000000, "min"));
9547 testCases.push_back(ConvertCase(instruction, DATA_TYPE_SIGNED_8, DATA_TYPE_FLOAT_64, -128, true, 0xC060000000000000ull, "min"));
9549 // Maximum int8 value
9550 testCases.push_back(ConvertCase(instruction, DATA_TYPE_SIGNED_8, DATA_TYPE_FLOAT_16, 127, true, 0x57F0, "max"));
9551 testCases.push_back(ConvertCase(instruction, DATA_TYPE_SIGNED_8, DATA_TYPE_FLOAT_32, 127, true, 0x42FE0000, "max"));
9552 testCases.push_back(ConvertCase(instruction, DATA_TYPE_SIGNED_8, DATA_TYPE_FLOAT_64, 127, true, 0x405FC00000000000ull, "max"));
9554 // All hexadecimal values below represent 1234.0 as 32/64-bit IEEE 754 float
9555 testCases.push_back(ConvertCase(instruction, DATA_TYPE_SIGNED_16, DATA_TYPE_FLOAT_16, -1234, true, 0xE4D2, "m1234"));
9556 testCases.push_back(ConvertCase(instruction, DATA_TYPE_SIGNED_32, DATA_TYPE_FLOAT_16, -1234, true, 0xE4D2, "m1234"));
9557 testCases.push_back(ConvertCase(instruction, DATA_TYPE_SIGNED_64, DATA_TYPE_FLOAT_16, -1234, true, 0xE4D2, "m1234"));
9559 // 0xF800 = 1111 1000 0000 0000 = 1 11110 0000000000 = -32768
9560 testCases.push_back(ConvertCase(instruction, DATA_TYPE_SIGNED_16, DATA_TYPE_FLOAT_16, -32768, true, 0xF800, "min"));
9561 testCases.push_back(ConvertCase(instruction, DATA_TYPE_SIGNED_32, DATA_TYPE_FLOAT_16, -32768, true, 0xF800, "min"));
9562 testCases.push_back(ConvertCase(instruction, DATA_TYPE_SIGNED_64, DATA_TYPE_FLOAT_16, -32768, true, 0xF800, "min"));
9564 // 0x77FF = 0111 0111 1111 1111 = 0 11101 1111111111 = 32752
9565 testCases.push_back(ConvertCase(instruction, DATA_TYPE_SIGNED_16, DATA_TYPE_FLOAT_16, 32752, true, 0x77FF, "max"));
9566 testCases.push_back(ConvertCase(instruction, DATA_TYPE_SIGNED_32, DATA_TYPE_FLOAT_16, 32752, true, 0x77FF, "max"));
9567 testCases.push_back(ConvertCase(instruction, DATA_TYPE_SIGNED_64, DATA_TYPE_FLOAT_16, 32752, true, 0x77FF, "max"));
9569 testCases.push_back(ConvertCase(instruction, DATA_TYPE_SIGNED_16, DATA_TYPE_FLOAT_32, -1234, true, 0xc49a4000));
9570 testCases.push_back(ConvertCase(instruction, DATA_TYPE_SIGNED_16, DATA_TYPE_FLOAT_64, -1234, true, 0xc093480000000000));
9571 testCases.push_back(ConvertCase(instruction, DATA_TYPE_SIGNED_32, DATA_TYPE_FLOAT_32, -1234, true, 0xc49a4000));
9572 testCases.push_back(ConvertCase(instruction, DATA_TYPE_SIGNED_32, DATA_TYPE_FLOAT_64, -1234, true, 0xc093480000000000));
9573 testCases.push_back(ConvertCase(instruction, DATA_TYPE_SIGNED_64, DATA_TYPE_FLOAT_32, -1234, true, 0xc49a4000));
9574 testCases.push_back(ConvertCase(instruction, DATA_TYPE_SIGNED_64, DATA_TYPE_FLOAT_64, -1234, true, 0xc093480000000000));
9577 DE_FATAL("Unknown instruction");
9580 const map<string, string> getConvertCaseFragments (string instruction, const ConvertCase& convertCase)
9582 map<string, string> params = convertCase.m_asmTypes;
9583 map<string, string> fragments;
9585 params["instruction"] = instruction;
9586 params["inDecorator"] = getByteWidthStr(convertCase.m_fromType);
9588 const StringTemplate decoration (
9589 " OpDecorate %SSBOi DescriptorSet 0\n"
9590 " OpDecorate %SSBOo DescriptorSet 0\n"
9591 " OpDecorate %SSBOi Binding 0\n"
9592 " OpDecorate %SSBOo Binding 1\n"
9593 " OpDecorate %s_SSBOi Block\n"
9594 " OpDecorate %s_SSBOo Block\n"
9595 "OpMemberDecorate %s_SSBOi 0 Offset 0\n"
9596 "OpMemberDecorate %s_SSBOo 0 Offset 0\n");
9598 const StringTemplate pre_main (
9599 "${datatype_additional_decl:opt}"
9600 " %ptr_in = OpTypePointer StorageBuffer %${inputType}\n"
9601 " %ptr_out = OpTypePointer StorageBuffer %${outputType}\n"
9602 " %s_SSBOi = OpTypeStruct %${inputType}\n"
9603 " %s_SSBOo = OpTypeStruct %${outputType}\n"
9604 " %ptr_SSBOi = OpTypePointer StorageBuffer %s_SSBOi\n"
9605 " %ptr_SSBOo = OpTypePointer StorageBuffer %s_SSBOo\n"
9606 " %SSBOi = OpVariable %ptr_SSBOi StorageBuffer\n"
9607 " %SSBOo = OpVariable %ptr_SSBOo StorageBuffer\n");
9609 const StringTemplate testfun (
9610 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
9611 "%param = OpFunctionParameter %v4f32\n"
9612 "%label = OpLabel\n"
9613 "%iLoc = OpAccessChain %ptr_in %SSBOi %c_u32_0\n"
9614 "%oLoc = OpAccessChain %ptr_out %SSBOo %c_u32_0\n"
9615 "%valIn = OpLoad %${inputType} %iLoc\n"
9616 "%valOut = ${instruction} %${outputType} %valIn\n"
9617 " OpStore %oLoc %valOut\n"
9618 " OpReturnValue %param\n"
9619 " OpFunctionEnd\n");
9621 params["datatype_extensions"] =
9622 params["datatype_extensions"] +
9623 "OpExtension \"SPV_KHR_storage_buffer_storage_class\"\n";
9625 fragments["capability"] = params["datatype_capabilities"];
9626 fragments["extension"] = params["datatype_extensions"];
9627 fragments["decoration"] = decoration.specialize(params);
9628 fragments["pre_main"] = pre_main.specialize(params);
9629 fragments["testfun"] = testfun.specialize(params);
9634 // Test for OpSConvert, OpUConvert, OpFConvert and OpConvert* in compute shaders
9635 tcu::TestCaseGroup* createConvertComputeTests (tcu::TestContext& testCtx, const string& instruction, const string& name)
9637 de::MovePtr<tcu::TestCaseGroup> group(new tcu::TestCaseGroup(testCtx, name.c_str(), instruction.c_str()));
9638 vector<ConvertCase> testCases;
9639 createConvertCases(testCases, instruction);
9641 for (vector<ConvertCase>::const_iterator test = testCases.begin(); test != testCases.end(); ++test)
9643 ComputeShaderSpec spec;
9644 spec.assembly = getConvertCaseShaderStr(instruction, *test);
9645 spec.numWorkGroups = IVec3(1, 1, 1);
9646 spec.inputs.push_back (test->m_inputBuffer);
9647 spec.outputs.push_back (test->m_outputBuffer);
9649 getVulkanFeaturesAndExtensions(test->m_fromType, test->m_toType, spec.requestedVulkanFeatures, spec.extensions);
9651 group->addChild(new SpvAsmComputeShaderCase(testCtx, test->m_name.c_str(), "", spec));
9653 return group.release();
9656 // Test for OpSConvert, OpUConvert, OpFConvert and OpConvert* in graphics shaders
9657 tcu::TestCaseGroup* createConvertGraphicsTests (tcu::TestContext& testCtx, const string& instruction, const string& name)
9659 de::MovePtr<tcu::TestCaseGroup> group(new tcu::TestCaseGroup(testCtx, name.c_str(), instruction.c_str()));
9660 vector<ConvertCase> testCases;
9661 createConvertCases(testCases, instruction);
9663 for (vector<ConvertCase>::const_iterator test = testCases.begin(); test != testCases.end(); ++test)
9665 map<string, string> fragments = getConvertCaseFragments(instruction, *test);
9666 VulkanFeatures vulkanFeatures;
9667 GraphicsResources resources;
9668 vector<string> extensions;
9669 SpecConstants noSpecConstants;
9670 PushConstants noPushConstants;
9671 GraphicsInterfaces noInterfaces;
9672 tcu::RGBA defaultColors[4];
9674 getDefaultColors (defaultColors);
9675 resources.inputs.push_back (Resource(test->m_inputBuffer, VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
9676 resources.outputs.push_back (Resource(test->m_outputBuffer, VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
9677 extensions.push_back ("VK_KHR_storage_buffer_storage_class");
9679 getVulkanFeaturesAndExtensions(test->m_fromType, test->m_toType, vulkanFeatures, extensions);
9681 vulkanFeatures.coreFeatures.vertexPipelineStoresAndAtomics = true;
9682 vulkanFeatures.coreFeatures.fragmentStoresAndAtomics = true;
9684 createTestsForAllStages(
9685 test->m_name, defaultColors, defaultColors, fragments, noSpecConstants,
9686 noPushConstants, resources, noInterfaces, extensions, vulkanFeatures, group.get());
9688 return group.release();
9691 // Constant-Creation Instructions: OpConstant, OpConstantComposite
9692 tcu::TestCaseGroup* createOpConstantFloat16Tests(tcu::TestContext& testCtx)
9694 de::MovePtr<tcu::TestCaseGroup> opConstantCompositeTests (new tcu::TestCaseGroup(testCtx, "opconstant", "OpConstant and OpConstantComposite instruction"));
9695 RGBA inputColors[4];
9696 RGBA outputColors[4];
9697 vector<string> extensions;
9698 GraphicsResources resources;
9699 VulkanFeatures features;
9701 const char functionStart[] =
9702 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
9703 "%param1 = OpFunctionParameter %v4f32\n"
9706 const char functionEnd[] =
9707 "%transformed_param_32 = OpFConvert %v4f32 %transformed_param\n"
9708 " OpReturnValue %transformed_param_32\n"
9711 struct NameConstantsCode
9718 #define FLOAT_16_COMMON_TYPES_AND_CONSTS \
9719 "%f16 = OpTypeFloat 16\n" \
9720 "%c_f16_0 = OpConstant %f16 0.0\n" \
9721 "%c_f16_0_5 = OpConstant %f16 0.5\n" \
9722 "%c_f16_1 = OpConstant %f16 1.0\n" \
9723 "%v4f16 = OpTypeVector %f16 4\n" \
9724 "%fp_f16 = OpTypePointer Function %f16\n" \
9725 "%fp_v4f16 = OpTypePointer Function %v4f16\n" \
9726 "%c_v4f16_1_1_1_1 = OpConstantComposite %v4f16 %c_f16_1 %c_f16_1 %c_f16_1 %c_f16_1\n" \
9727 "%a4f16 = OpTypeArray %f16 %c_u32_4\n" \
9729 NameConstantsCode tests[] =
9734 FLOAT_16_COMMON_TYPES_AND_CONSTS
9735 "%cval = OpConstantComposite %v4f16 %c_f16_0_5 %c_f16_0_5 %c_f16_0_5 %c_f16_0\n",
9736 "%param1_16 = OpFConvert %v4f16 %param1\n"
9737 "%transformed_param = OpFAdd %v4f16 %param1_16 %cval\n"
9742 FLOAT_16_COMMON_TYPES_AND_CONSTS
9743 "%stype = OpTypeStruct %v4f16 %f16\n"
9744 "%fp_stype = OpTypePointer Function %stype\n"
9745 "%f16_n_1 = OpConstant %f16 -1.0\n"
9746 "%f16_1_5 = OpConstant %f16 !0x3e00\n" // +1.5
9747 "%cvec = OpConstantComposite %v4f16 %f16_1_5 %f16_1_5 %f16_1_5 %c_f16_1\n"
9748 "%cval = OpConstantComposite %stype %cvec %f16_n_1\n",
9750 "%v = OpVariable %fp_stype Function %cval\n"
9751 "%vec_ptr = OpAccessChain %fp_v4f16 %v %c_u32_0\n"
9752 "%f16_ptr = OpAccessChain %fp_f16 %v %c_u32_1\n"
9753 "%vec_val = OpLoad %v4f16 %vec_ptr\n"
9754 "%f16_val = OpLoad %f16 %f16_ptr\n"
9755 "%tmp1 = OpVectorTimesScalar %v4f16 %c_v4f16_1_1_1_1 %f16_val\n" // vec4(-1)
9756 "%param1_16 = OpFConvert %v4f16 %param1\n"
9757 "%tmp2 = OpFAdd %v4f16 %tmp1 %param1_16\n" // param1 + vec4(-1)
9758 "%transformed_param = OpFAdd %v4f16 %tmp2 %vec_val\n" // param1 + vec4(-1) + vec4(1.5, 1.5, 1.5, 1.0)
9761 // [1|0|0|0.5] [x] = x + 0.5
9762 // [0|1|0|0.5] [y] = y + 0.5
9763 // [0|0|1|0.5] [z] = z + 0.5
9764 // [0|0|0|1 ] [1] = 1
9767 FLOAT_16_COMMON_TYPES_AND_CONSTS
9768 "%mat4x4_f16 = OpTypeMatrix %v4f16 4\n"
9769 "%v4f16_1_0_0_0 = OpConstantComposite %v4f16 %c_f16_1 %c_f16_0 %c_f16_0 %c_f16_0\n"
9770 "%v4f16_0_1_0_0 = OpConstantComposite %v4f16 %c_f16_0 %c_f16_1 %c_f16_0 %c_f16_0\n"
9771 "%v4f16_0_0_1_0 = OpConstantComposite %v4f16 %c_f16_0 %c_f16_0 %c_f16_1 %c_f16_0\n"
9772 "%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"
9773 "%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",
9775 "%param1_16 = OpFConvert %v4f16 %param1\n"
9776 "%transformed_param = OpMatrixTimesVector %v4f16 %cval %param1_16\n"
9781 FLOAT_16_COMMON_TYPES_AND_CONSTS
9782 "%c_v4f16_1_1_1_0 = OpConstantComposite %v4f16 %c_f16_1 %c_f16_1 %c_f16_1 %c_f16_0\n"
9783 "%fp_a4f16 = OpTypePointer Function %a4f16\n"
9784 "%f16_n_1 = OpConstant %f16 -1.0\n"
9785 "%f16_1_5 = OpConstant %f16 !0x3e00\n" // +1.5
9786 "%carr = OpConstantComposite %a4f16 %c_f16_0 %f16_n_1 %f16_1_5 %c_f16_0\n",
9788 "%v = OpVariable %fp_a4f16 Function %carr\n"
9789 "%f = OpAccessChain %fp_f16 %v %c_u32_0\n"
9790 "%f1 = OpAccessChain %fp_f16 %v %c_u32_1\n"
9791 "%f2 = OpAccessChain %fp_f16 %v %c_u32_2\n"
9792 "%f3 = OpAccessChain %fp_f16 %v %c_u32_3\n"
9793 "%f_val = OpLoad %f16 %f\n"
9794 "%f1_val = OpLoad %f16 %f1\n"
9795 "%f2_val = OpLoad %f16 %f2\n"
9796 "%f3_val = OpLoad %f16 %f3\n"
9797 "%ftot1 = OpFAdd %f16 %f_val %f1_val\n"
9798 "%ftot2 = OpFAdd %f16 %ftot1 %f2_val\n"
9799 "%ftot3 = OpFAdd %f16 %ftot2 %f3_val\n" // 0 - 1 + 1.5 + 0
9800 "%add_vec = OpVectorTimesScalar %v4f16 %c_v4f16_1_1_1_0 %ftot3\n"
9801 "%param1_16 = OpFConvert %v4f16 %param1\n"
9802 "%transformed_param = OpFAdd %v4f16 %param1_16 %add_vec\n"
9809 // [ 1.0, 1.0, 1.0, 1.0]
9813 // [ 0.0, 0.5, 0.0, 0.0]
9817 // [ 1.0, 1.0, 1.0, 1.0]
9820 "array_of_struct_of_array",
9822 FLOAT_16_COMMON_TYPES_AND_CONSTS
9823 "%c_v4f16_1_1_1_0 = OpConstantComposite %v4f16 %c_f16_1 %c_f16_1 %c_f16_1 %c_f16_0\n"
9824 "%fp_a4f16 = OpTypePointer Function %a4f16\n"
9825 "%stype = OpTypeStruct %f16 %a4f16\n"
9826 "%a3stype = OpTypeArray %stype %c_u32_3\n"
9827 "%fp_a3stype = OpTypePointer Function %a3stype\n"
9828 "%ca4f16_0 = OpConstantComposite %a4f16 %c_f16_0 %c_f16_0_5 %c_f16_0 %c_f16_0\n"
9829 "%ca4f16_1 = OpConstantComposite %a4f16 %c_f16_1 %c_f16_1 %c_f16_1 %c_f16_1\n"
9830 "%cstype1 = OpConstantComposite %stype %c_f16_0 %ca4f16_1\n"
9831 "%cstype2 = OpConstantComposite %stype %c_f16_1 %ca4f16_0\n"
9832 "%carr = OpConstantComposite %a3stype %cstype1 %cstype2 %cstype1",
9834 "%v = OpVariable %fp_a3stype Function %carr\n"
9835 "%f = OpAccessChain %fp_f16 %v %c_u32_1 %c_u32_1 %c_u32_1\n"
9836 "%f_l = OpLoad %f16 %f\n"
9837 "%add_vec = OpVectorTimesScalar %v4f16 %c_v4f16_1_1_1_0 %f_l\n"
9838 "%param1_16 = OpFConvert %v4f16 %param1\n"
9839 "%transformed_param = OpFAdd %v4f16 %param1_16 %add_vec\n"
9843 getHalfColorsFullAlpha(inputColors);
9844 outputColors[0] = RGBA(255, 255, 255, 255);
9845 outputColors[1] = RGBA(255, 127, 127, 255);
9846 outputColors[2] = RGBA(127, 255, 127, 255);
9847 outputColors[3] = RGBA(127, 127, 255, 255);
9849 extensions.push_back("VK_KHR_16bit_storage");
9850 extensions.push_back("VK_KHR_shader_float16_int8");
9851 features.extFloat16Int8 = EXTFLOAT16INT8FEATURES_FLOAT16;
9853 for (size_t testNdx = 0; testNdx < sizeof(tests) / sizeof(NameConstantsCode); ++testNdx)
9855 map<string, string> fragments;
9857 fragments["extension"] = "OpExtension \"SPV_KHR_16bit_storage\"";
9858 fragments["capability"] = "OpCapability Float16\n";
9859 fragments["pre_main"] = tests[testNdx].constants;
9860 fragments["testfun"] = string(functionStart) + tests[testNdx].code + functionEnd;
9862 createTestsForAllStages(tests[testNdx].name, inputColors, outputColors, fragments, resources, extensions, opConstantCompositeTests.get(), features);
9864 return opConstantCompositeTests.release();
9867 template<typename T>
9868 void finalizeTestsCreation (T& specResource,
9869 const map<string, string>& fragments,
9870 tcu::TestContext& testCtx,
9871 tcu::TestCaseGroup& testGroup,
9872 const std::string& testName,
9873 const VulkanFeatures& vulkanFeatures,
9874 const vector<string>& extensions,
9875 const IVec3& numWorkGroups);
9878 void finalizeTestsCreation (GraphicsResources& specResource,
9879 const map<string, string>& fragments,
9881 tcu::TestCaseGroup& testGroup,
9882 const std::string& testName,
9883 const VulkanFeatures& vulkanFeatures,
9884 const vector<string>& extensions,
9887 RGBA defaultColors[4];
9888 getDefaultColors(defaultColors);
9890 createTestsForAllStages(testName, defaultColors, defaultColors, fragments, specResource, extensions, &testGroup, vulkanFeatures);
9894 void finalizeTestsCreation (ComputeShaderSpec& specResource,
9895 const map<string, string>& fragments,
9896 tcu::TestContext& testCtx,
9897 tcu::TestCaseGroup& testGroup,
9898 const std::string& testName,
9899 const VulkanFeatures& vulkanFeatures,
9900 const vector<string>& extensions,
9901 const IVec3& numWorkGroups)
9903 specResource.numWorkGroups = numWorkGroups;
9904 specResource.requestedVulkanFeatures = vulkanFeatures;
9905 specResource.extensions = extensions;
9907 specResource.assembly = makeComputeShaderAssembly(fragments);
9909 testGroup.addChild(new SpvAsmComputeShaderCase(testCtx, testName.c_str(), "", specResource));
9912 template<class SpecResource>
9913 tcu::TestCaseGroup* createFloat16LogicalSet (tcu::TestContext& testCtx, const bool nanSupported)
9915 const string nan = nanSupported ? "_nan" : "";
9916 const string groupName = "logical" + nan;
9917 de::MovePtr<tcu::TestCaseGroup> testGroup (new tcu::TestCaseGroup(testCtx, groupName.c_str(), "Float 16 logical tests"));
9919 de::Random rnd (deStringHash(testGroup->getName()));
9920 const string spvCapabilities = string("OpCapability StorageUniformBufferBlock16\n") + (nanSupported ? "OpCapability SignedZeroInfNanPreserve\n" : "");
9921 const string spvExtensions = string("OpExtension \"SPV_KHR_16bit_storage\"\n") + (nanSupported ? "OpExtension \"SPV_KHR_float_controls\"\n" : "");
9922 const string spvExecutionMode = nanSupported ? "OpExecutionMode %BP_main SignedZeroInfNanPreserve 16\n" : "";
9923 const deUint32 numDataPoints = 16;
9924 const vector<deFloat16> float16Data = getFloat16s(rnd, numDataPoints);
9925 const vector<deFloat16> float16Data1 = squarize(float16Data, 0);
9926 const vector<deFloat16> float16Data2 = squarize(float16Data, 1);
9927 const vector<deFloat16> float16DataVec1 = squarizeVector(float16Data, 0);
9928 const vector<deFloat16> float16DataVec2 = squarizeVector(float16Data, 1);
9929 const vector<deFloat16> float16OutDummy (float16Data1.size(), 0);
9930 const vector<deFloat16> float16OutVecDummy (float16DataVec1.size(), 0);
9935 VerifyIOFunc verifyFuncNan;
9936 VerifyIOFunc verifyFuncNonNan;
9937 const deUint32 argCount;
9940 const TestOp testOps[] =
9942 { "OpIsNan" , compareFP16Logical<fp16isNan, true, false, true>, compareFP16Logical<fp16isNan, true, false, false>, 1 },
9943 { "OpIsInf" , compareFP16Logical<fp16isInf, true, false, true>, compareFP16Logical<fp16isInf, true, false, false>, 1 },
9944 { "OpFOrdEqual" , compareFP16Logical<fp16isEqual, false, true, true>, compareFP16Logical<fp16isEqual, false, true, false>, 2 },
9945 { "OpFUnordEqual" , compareFP16Logical<fp16isEqual, false, false, true>, compareFP16Logical<fp16isEqual, false, false, false>, 2 },
9946 { "OpFOrdNotEqual" , compareFP16Logical<fp16isUnequal, false, true, true>, compareFP16Logical<fp16isUnequal, false, true, false>, 2 },
9947 { "OpFUnordNotEqual" , compareFP16Logical<fp16isUnequal, false, false, true>, compareFP16Logical<fp16isUnequal, false, false, false>, 2 },
9948 { "OpFOrdLessThan" , compareFP16Logical<fp16isLess, false, true, true>, compareFP16Logical<fp16isLess, false, true, false>, 2 },
9949 { "OpFUnordLessThan" , compareFP16Logical<fp16isLess, false, false, true>, compareFP16Logical<fp16isLess, false, false, false>, 2 },
9950 { "OpFOrdGreaterThan" , compareFP16Logical<fp16isGreater, false, true, true>, compareFP16Logical<fp16isGreater, false, true, false>, 2 },
9951 { "OpFUnordGreaterThan" , compareFP16Logical<fp16isGreater, false, false, true>, compareFP16Logical<fp16isGreater, false, false, false>, 2 },
9952 { "OpFOrdLessThanEqual" , compareFP16Logical<fp16isLessOrEqual, false, true, true>, compareFP16Logical<fp16isLessOrEqual, false, true, false>, 2 },
9953 { "OpFUnordLessThanEqual" , compareFP16Logical<fp16isLessOrEqual, false, false, true>, compareFP16Logical<fp16isLessOrEqual, false, false, false>, 2 },
9954 { "OpFOrdGreaterThanEqual" , compareFP16Logical<fp16isGreaterOrEqual, false, true, true>, compareFP16Logical<fp16isGreaterOrEqual, false, true, false>, 2 },
9955 { "OpFUnordGreaterThanEqual" , compareFP16Logical<fp16isGreaterOrEqual, false, false, true>, compareFP16Logical<fp16isGreaterOrEqual, false, false, false>, 2 },
9959 const StringTemplate preMain
9961 "%c_i32_ndp = OpConstant %i32 ${num_data_points}\n"
9962 " %f16 = OpTypeFloat 16\n"
9963 " %c_f16_0 = OpConstant %f16 0.0\n"
9964 " %c_f16_1 = OpConstant %f16 1.0\n"
9965 " %up_f16 = OpTypePointer Uniform %f16\n"
9966 " %ra_f16 = OpTypeArray %f16 %c_i32_ndp\n"
9967 " %SSBO16 = OpTypeStruct %ra_f16\n"
9968 "%up_SSBO16 = OpTypePointer Uniform %SSBO16\n"
9969 "%ssbo_src0 = OpVariable %up_SSBO16 Uniform\n"
9970 "%ssbo_src1 = OpVariable %up_SSBO16 Uniform\n"
9971 " %ssbo_dst = OpVariable %up_SSBO16 Uniform\n"
9974 const StringTemplate decoration
9976 "OpDecorate %ra_f16 ArrayStride 2\n"
9977 "OpMemberDecorate %SSBO16 0 Offset 0\n"
9978 "OpDecorate %SSBO16 BufferBlock\n"
9979 "OpDecorate %ssbo_src0 DescriptorSet 0\n"
9980 "OpDecorate %ssbo_src0 Binding 0\n"
9981 "OpDecorate %ssbo_src1 DescriptorSet 0\n"
9982 "OpDecorate %ssbo_src1 Binding 1\n"
9983 "OpDecorate %ssbo_dst DescriptorSet 0\n"
9984 "OpDecorate %ssbo_dst Binding 2\n"
9987 const StringTemplate testFun
9989 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
9990 " %param = OpFunctionParameter %v4f32\n"
9992 " %entry = OpLabel\n"
9993 " %i = OpVariable %fp_i32 Function\n"
9994 " OpStore %i %c_i32_0\n"
9997 " %loop = OpLabel\n"
9998 " %i_cmp = OpLoad %i32 %i\n"
9999 " %lt = OpSLessThan %bool %i_cmp %c_i32_ndp\n"
10000 " OpLoopMerge %merge %next None\n"
10001 " OpBranchConditional %lt %write %merge\n"
10003 " %write = OpLabel\n"
10004 " %ndx = OpLoad %i32 %i\n"
10006 " %src0 = OpAccessChain %up_f16 %ssbo_src0 %c_i32_0 %ndx\n"
10007 " %val_src0 = OpLoad %f16 %src0\n"
10011 " %val_bdst = ${op_code} %bool %val_src0 ${op_arg1}\n"
10012 " %val_dst = OpSelect %f16 %val_bdst %c_f16_1 %c_f16_0\n"
10013 " %dst = OpAccessChain %up_f16 %ssbo_dst %c_i32_0 %ndx\n"
10014 " OpStore %dst %val_dst\n"
10015 " OpBranch %next\n"
10017 " %next = OpLabel\n"
10018 " %i_cur = OpLoad %i32 %i\n"
10019 " %i_new = OpIAdd %i32 %i_cur %c_i32_1\n"
10020 " OpStore %i %i_new\n"
10021 " OpBranch %loop\n"
10023 " %merge = OpLabel\n"
10024 " OpReturnValue %param\n"
10029 const StringTemplate arg1Calc
10031 " %src1 = OpAccessChain %up_f16 %ssbo_src1 %c_i32_0 %ndx\n"
10032 " %val_src1 = OpLoad %f16 %src1\n"
10035 for (deUint32 testOpsIdx = 0; testOpsIdx < DE_LENGTH_OF_ARRAY(testOps); ++testOpsIdx)
10037 const size_t iterations = float16Data1.size();
10038 const TestOp& testOp = testOps[testOpsIdx];
10039 const string testName = de::toLower(string(testOp.opCode)) + "_scalar";
10040 SpecResource specResource;
10041 map<string, string> specs;
10042 VulkanFeatures features;
10043 map<string, string> fragments;
10044 vector<string> extensions;
10046 specs["num_data_points"] = de::toString(iterations);
10047 specs["op_code"] = testOp.opCode;
10048 specs["op_arg1"] = (testOp.argCount == 1) ? "" : "%val_src1";
10049 specs["op_arg1_calc"] = (testOp.argCount == 1) ? "" : arg1Calc.specialize(specs);
10051 fragments["extension"] = spvExtensions;
10052 fragments["capability"] = spvCapabilities;
10053 fragments["execution_mode"] = spvExecutionMode;
10054 fragments["decoration"] = decoration.specialize(specs);
10055 fragments["pre_main"] = preMain.specialize(specs);
10056 fragments["testfun"] = testFun.specialize(specs);
10058 specResource.inputs.push_back(Resource(BufferSp(new Float16Buffer(float16Data1)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
10059 specResource.inputs.push_back(Resource(BufferSp(new Float16Buffer(float16Data2)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
10060 specResource.outputs.push_back(Resource(BufferSp(new Float16Buffer(float16OutDummy)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
10061 specResource.verifyIO = nanSupported ? testOp.verifyFuncNan : testOp.verifyFuncNonNan;
10063 extensions.push_back("VK_KHR_16bit_storage");
10064 extensions.push_back("VK_KHR_shader_float16_int8");
10068 extensions.push_back("VK_KHR_shader_float_controls");
10070 features.floatControlsProperties.shaderSignedZeroInfNanPreserveFloat16 = DE_TRUE;
10073 features.ext16BitStorage = EXT16BITSTORAGEFEATURES_UNIFORM_BUFFER_BLOCK;
10074 features.extFloat16Int8 = EXTFLOAT16INT8FEATURES_FLOAT16;
10076 finalizeTestsCreation(specResource, fragments, testCtx, *testGroup.get(), testName, features, extensions, IVec3(1, 1, 1));
10080 const StringTemplate preMain
10082 " %c_i32_ndp = OpConstant %i32 ${num_data_points}\n"
10083 " %v2bool = OpTypeVector %bool 2\n"
10084 " %f16 = OpTypeFloat 16\n"
10085 " %c_f16_0 = OpConstant %f16 0.0\n"
10086 " %c_f16_1 = OpConstant %f16 1.0\n"
10087 " %v2f16 = OpTypeVector %f16 2\n"
10088 "%c_v2f16_0_0 = OpConstantComposite %v2f16 %c_f16_0 %c_f16_0\n"
10089 "%c_v2f16_1_1 = OpConstantComposite %v2f16 %c_f16_1 %c_f16_1\n"
10090 " %up_v2f16 = OpTypePointer Uniform %v2f16\n"
10091 " %ra_v2f16 = OpTypeArray %v2f16 %c_i32_ndp\n"
10092 " %SSBO16 = OpTypeStruct %ra_v2f16\n"
10093 " %up_SSBO16 = OpTypePointer Uniform %SSBO16\n"
10094 " %ssbo_src0 = OpVariable %up_SSBO16 Uniform\n"
10095 " %ssbo_src1 = OpVariable %up_SSBO16 Uniform\n"
10096 " %ssbo_dst = OpVariable %up_SSBO16 Uniform\n"
10099 const StringTemplate decoration
10101 "OpDecorate %ra_v2f16 ArrayStride 4\n"
10102 "OpMemberDecorate %SSBO16 0 Offset 0\n"
10103 "OpDecorate %SSBO16 BufferBlock\n"
10104 "OpDecorate %ssbo_src0 DescriptorSet 0\n"
10105 "OpDecorate %ssbo_src0 Binding 0\n"
10106 "OpDecorate %ssbo_src1 DescriptorSet 0\n"
10107 "OpDecorate %ssbo_src1 Binding 1\n"
10108 "OpDecorate %ssbo_dst DescriptorSet 0\n"
10109 "OpDecorate %ssbo_dst Binding 2\n"
10112 const StringTemplate testFun
10114 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
10115 " %param = OpFunctionParameter %v4f32\n"
10117 " %entry = OpLabel\n"
10118 " %i = OpVariable %fp_i32 Function\n"
10119 " OpStore %i %c_i32_0\n"
10120 " OpBranch %loop\n"
10122 " %loop = OpLabel\n"
10123 " %i_cmp = OpLoad %i32 %i\n"
10124 " %lt = OpSLessThan %bool %i_cmp %c_i32_ndp\n"
10125 " OpLoopMerge %merge %next None\n"
10126 " OpBranchConditional %lt %write %merge\n"
10128 " %write = OpLabel\n"
10129 " %ndx = OpLoad %i32 %i\n"
10131 " %src0 = OpAccessChain %up_v2f16 %ssbo_src0 %c_i32_0 %ndx\n"
10132 " %val_src0 = OpLoad %v2f16 %src0\n"
10136 " %val_bdst = ${op_code} %v2bool %val_src0 ${op_arg1}\n"
10137 " %val_dst = OpSelect %v2f16 %val_bdst %c_v2f16_1_1 %c_v2f16_0_0\n"
10138 " %dst = OpAccessChain %up_v2f16 %ssbo_dst %c_i32_0 %ndx\n"
10139 " OpStore %dst %val_dst\n"
10140 " OpBranch %next\n"
10142 " %next = OpLabel\n"
10143 " %i_cur = OpLoad %i32 %i\n"
10144 " %i_new = OpIAdd %i32 %i_cur %c_i32_1\n"
10145 " OpStore %i %i_new\n"
10146 " OpBranch %loop\n"
10148 " %merge = OpLabel\n"
10149 " OpReturnValue %param\n"
10154 const StringTemplate arg1Calc
10156 " %src1 = OpAccessChain %up_v2f16 %ssbo_src1 %c_i32_0 %ndx\n"
10157 " %val_src1 = OpLoad %v2f16 %src1\n"
10160 for (deUint32 testOpsIdx = 0; testOpsIdx < DE_LENGTH_OF_ARRAY(testOps); ++testOpsIdx)
10162 const deUint32 itemsPerVec = 2;
10163 const size_t iterations = float16DataVec1.size() / itemsPerVec;
10164 const TestOp& testOp = testOps[testOpsIdx];
10165 const string testName = de::toLower(string(testOp.opCode)) + "_vector";
10166 SpecResource specResource;
10167 map<string, string> specs;
10168 vector<string> extensions;
10169 VulkanFeatures features;
10170 map<string, string> fragments;
10172 specs["num_data_points"] = de::toString(iterations);
10173 specs["op_code"] = testOp.opCode;
10174 specs["op_arg1"] = (testOp.argCount == 1) ? "" : "%val_src1";
10175 specs["op_arg1_calc"] = (testOp.argCount == 1) ? "" : arg1Calc.specialize(specs);
10177 fragments["extension"] = spvExtensions;
10178 fragments["capability"] = spvCapabilities;
10179 fragments["execution_mode"] = spvExecutionMode;
10180 fragments["decoration"] = decoration.specialize(specs);
10181 fragments["pre_main"] = preMain.specialize(specs);
10182 fragments["testfun"] = testFun.specialize(specs);
10184 specResource.inputs.push_back(Resource(BufferSp(new Float16Buffer(float16DataVec1)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
10185 specResource.inputs.push_back(Resource(BufferSp(new Float16Buffer(float16DataVec2)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
10186 specResource.outputs.push_back(Resource(BufferSp(new Float16Buffer(float16OutVecDummy)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
10187 specResource.verifyIO = nanSupported ? testOp.verifyFuncNan : testOp.verifyFuncNonNan;
10189 extensions.push_back("VK_KHR_16bit_storage");
10190 extensions.push_back("VK_KHR_shader_float16_int8");
10194 extensions.push_back("VK_KHR_shader_float_controls");
10196 features.floatControlsProperties.shaderSignedZeroInfNanPreserveFloat16 = DE_TRUE;
10199 features.ext16BitStorage = EXT16BITSTORAGEFEATURES_UNIFORM_BUFFER_BLOCK;
10200 features.extFloat16Int8 = EXTFLOAT16INT8FEATURES_FLOAT16;
10202 finalizeTestsCreation(specResource, fragments, testCtx, *testGroup.get(), testName, features, extensions, IVec3(1, 1, 1));
10206 return testGroup.release();
10209 bool compareFP16FunctionSetFunc (const std::vector<Resource>& inputs, const vector<AllocationSp>& outputAllocs, const std::vector<Resource>&, TestLog& log)
10211 if (inputs.size() != 1 || outputAllocs.size() != 1)
10214 vector<deUint8> input1Bytes;
10216 inputs[0].getBytes(input1Bytes);
10218 const deUint16* const input1AsFP16 = (const deUint16*)&input1Bytes[0];
10219 const deUint16* const outputAsFP16 = (const deUint16*)outputAllocs[0]->getHostPtr();
10222 for (size_t idx = 0; idx < input1Bytes.size() / sizeof(deUint16); ++idx)
10224 if (!compare16BitFloat(input1AsFP16[idx], outputAsFP16[idx], error))
10226 log << TestLog::Message << error << TestLog::EndMessage;
10235 template<class SpecResource>
10236 tcu::TestCaseGroup* createFloat16FuncSet (tcu::TestContext& testCtx)
10238 de::MovePtr<tcu::TestCaseGroup> testGroup (new tcu::TestCaseGroup(testCtx, "function", "Float 16 function call related tests"));
10240 de::Random rnd (deStringHash(testGroup->getName()));
10241 const StringTemplate capabilities ("OpCapability ${cap}\n");
10242 const deUint32 numDataPoints = 256;
10243 const vector<deFloat16> float16InputData = getFloat16s(rnd, numDataPoints);
10244 const vector<deFloat16> float16OutputDummy (float16InputData.size(), 0);
10245 map<string, string> fragments;
10249 const deUint32 typeComponents;
10250 const char* typeName;
10251 const char* typeDecls;
10254 const TestType testTypes[] =
10264 " %v2f16 = OpTypeVector %f16 2\n"
10265 " %c_v2f16_0 = OpConstantComposite %v2f16 %c_f16_0 %c_f16_0\n"
10270 " %v4f16 = OpTypeVector %f16 4\n"
10271 " %c_v4f16_0 = OpConstantComposite %v4f16 %c_f16_0 %c_f16_0 %c_f16_0 %c_f16_0\n"
10275 const StringTemplate preMain
10277 " %c_i32_ndp = OpConstant %i32 ${num_data_points}\n"
10278 " %v2bool = OpTypeVector %bool 2\n"
10279 " %f16 = OpTypeFloat 16\n"
10280 " %c_f16_0 = OpConstant %f16 0.0\n"
10284 " %${tt}_fun = OpTypeFunction %${tt} %${tt}\n"
10285 " %up_${tt} = OpTypePointer Uniform %${tt}\n"
10286 " %ra_${tt} = OpTypeArray %${tt} %c_i32_ndp\n"
10287 " %SSBO16 = OpTypeStruct %ra_${tt}\n"
10288 " %up_SSBO16 = OpTypePointer Uniform %SSBO16\n"
10289 " %ssbo_src = OpVariable %up_SSBO16 Uniform\n"
10290 " %ssbo_dst = OpVariable %up_SSBO16 Uniform\n"
10293 const StringTemplate decoration
10295 "OpDecorate %ra_${tt} ArrayStride ${tt_stride}\n"
10296 "OpMemberDecorate %SSBO16 0 Offset 0\n"
10297 "OpDecorate %SSBO16 BufferBlock\n"
10298 "OpDecorate %ssbo_src DescriptorSet 0\n"
10299 "OpDecorate %ssbo_src Binding 0\n"
10300 "OpDecorate %ssbo_dst DescriptorSet 0\n"
10301 "OpDecorate %ssbo_dst Binding 1\n"
10304 const StringTemplate testFun
10306 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
10307 " %param = OpFunctionParameter %v4f32\n"
10308 " %entry = OpLabel\n"
10310 " %i = OpVariable %fp_i32 Function\n"
10311 " OpStore %i %c_i32_0\n"
10312 " OpBranch %loop\n"
10314 " %loop = OpLabel\n"
10315 " %i_cmp = OpLoad %i32 %i\n"
10316 " %lt = OpSLessThan %bool %i_cmp %c_i32_ndp\n"
10317 " OpLoopMerge %merge %next None\n"
10318 " OpBranchConditional %lt %write %merge\n"
10320 " %write = OpLabel\n"
10321 " %ndx = OpLoad %i32 %i\n"
10323 " %src = OpAccessChain %up_${tt} %ssbo_src %c_i32_0 %ndx\n"
10324 " %val_src = OpLoad %${tt} %src\n"
10326 " %val_dst = OpFunctionCall %${tt} %pass_fun %val_src\n"
10327 " %dst = OpAccessChain %up_${tt} %ssbo_dst %c_i32_0 %ndx\n"
10328 " OpStore %dst %val_dst\n"
10329 " OpBranch %next\n"
10331 " %next = OpLabel\n"
10332 " %i_cur = OpLoad %i32 %i\n"
10333 " %i_new = OpIAdd %i32 %i_cur %c_i32_1\n"
10334 " OpStore %i %i_new\n"
10335 " OpBranch %loop\n"
10337 " %merge = OpLabel\n"
10338 " OpReturnValue %param\n"
10342 " %pass_fun = OpFunction %${tt} None %${tt}_fun\n"
10343 " %param0 = OpFunctionParameter %${tt}\n"
10344 " %entry_pf = OpLabel\n"
10345 " %res0 = OpFAdd %${tt} %param0 %c_${tt}_0\n"
10346 " OpReturnValue %res0\n"
10350 for (deUint32 testTypeIdx = 0; testTypeIdx < DE_LENGTH_OF_ARRAY(testTypes); ++testTypeIdx)
10352 const TestType& testType = testTypes[testTypeIdx];
10353 const string testName = testType.typeName;
10354 const deUint32 itemsPerType = testType.typeComponents;
10355 const size_t iterations = float16InputData.size() / itemsPerType;
10356 const size_t typeStride = itemsPerType * sizeof(deFloat16);
10357 SpecResource specResource;
10358 map<string, string> specs;
10359 VulkanFeatures features;
10360 vector<string> extensions;
10362 specs["cap"] = "StorageUniformBufferBlock16";
10363 specs["num_data_points"] = de::toString(iterations);
10364 specs["tt"] = testType.typeName;
10365 specs["tt_stride"] = de::toString(typeStride);
10366 specs["type_decls"] = testType.typeDecls;
10368 fragments["extension"] = "OpExtension \"SPV_KHR_16bit_storage\"";
10369 fragments["capability"] = capabilities.specialize(specs);
10370 fragments["decoration"] = decoration.specialize(specs);
10371 fragments["pre_main"] = preMain.specialize(specs);
10372 fragments["testfun"] = testFun.specialize(specs);
10374 specResource.inputs.push_back(Resource(BufferSp(new Float16Buffer(float16InputData)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
10375 specResource.outputs.push_back(Resource(BufferSp(new Float16Buffer(float16OutputDummy)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
10376 specResource.verifyIO = compareFP16FunctionSetFunc;
10378 extensions.push_back("VK_KHR_16bit_storage");
10379 extensions.push_back("VK_KHR_shader_float16_int8");
10381 features.ext16BitStorage = EXT16BITSTORAGEFEATURES_UNIFORM_BUFFER_BLOCK;
10382 features.extFloat16Int8 = EXTFLOAT16INT8FEATURES_FLOAT16;
10384 finalizeTestsCreation(specResource, fragments, testCtx, *testGroup.get(), testName, features, extensions, IVec3(1, 1, 1));
10387 return testGroup.release();
10390 struct getV_ { deUint32 inline operator()(deUint32 v) const { return v; } getV_(){} };
10391 struct getV0 { deUint32 inline operator()(deUint32 v) const { return v & (~1); } getV0(){} };
10392 struct getV1 { deUint32 inline operator()(deUint32 v) const { return v | ( 1); } getV1(){} };
10394 template<deUint32 R, deUint32 N>
10395 inline static deUint32 getOffset(deUint32 x, deUint32 y, deUint32 n)
10397 return N * ((R * y) + x) + n;
10400 template<deUint32 R, deUint32 N, class X0, class X1, class Y0, class Y1>
10403 float operator() (const deFloat16* data, deUint32 x, deUint32 y, deUint32 n, deUint32 flavor) const
10405 DE_STATIC_ASSERT(R%2 == 0);
10406 DE_ASSERT(flavor == 0);
10413 const deFloat16 v0 = data[getOffset<R, N>(x0(x), y0(y), n)];
10414 const deFloat16 v1 = data[getOffset<R, N>(x1(x), y1(y), n)];
10415 const tcu::Float16 f0 = tcu::Float16(v0);
10416 const tcu::Float16 f1 = tcu::Float16(v1);
10417 const float d0 = f0.asFloat();
10418 const float d1 = f1.asFloat();
10419 const float d = d1 - d0;
10427 template<deUint32 F, class Class0, class Class1>
10430 float operator() (const deFloat16* data, deUint32 x, deUint32 y, deUint32 n, deUint32 flavor) const
10432 DE_ASSERT(flavor < F);
10438 return c(data, x, y, n, flavor);
10444 return c(data, x, y, n, flavor - 1);
10451 template<class FineX0, class FineX1, class FineY0, class FineY1>
10452 struct calcWidthOf4
10454 float operator() (const deFloat16* data, deUint32 x, deUint32 y, deUint32 n, deUint32 flavor) const
10456 DE_ASSERT(flavor < 4);
10458 const deUint32 flavorX = (flavor & 1) == 0 ? 0 : 1;
10459 const deUint32 flavorY = (flavor & 2) == 0 ? 0 : 1;
10460 const getFOneOf<2, FineX0, FineX1> cx;
10461 const getFOneOf<2, FineY0, FineY1> cy;
10464 v += fabsf(cx(data, x, y, n, flavorX));
10465 v += fabsf(cy(data, x, y, n, flavorY));
10473 template<deUint32 R, deUint32 N, class Derivative>
10474 bool compareDerivativeWithFlavor (const deFloat16* inputAsFP16, const deFloat16* outputAsFP16, deUint32 flavor, std::string& error)
10476 const deUint32 numDataPointsByAxis = R;
10477 const Derivative derivativeFunc;
10479 for (deUint32 y = 0; y < numDataPointsByAxis; ++y)
10480 for (deUint32 x = 0; x < numDataPointsByAxis; ++x)
10481 for (deUint32 n = 0; n < N; ++n)
10483 const float expectedFloat = derivativeFunc(inputAsFP16, x, y, n, flavor);
10484 deFloat16 expected = deFloat32To16Round(expectedFloat, DE_ROUNDINGMODE_TO_NEAREST_EVEN);
10485 const deFloat16 output = outputAsFP16[getOffset<R, N>(x, y, n)];
10487 bool reportError = !compare16BitFloat(expected, output, error);
10491 expected = deFloat32To16Round(expectedFloat, DE_ROUNDINGMODE_TO_ZERO);
10492 reportError = !compare16BitFloat(expected, output, error);
10497 error = "subcase at " + de::toString(x) + "," + de::toString(y) + "," + de::toString(n) + ": " + error;
10506 template<deUint32 R, deUint32 N, deUint32 FLAVOUR_COUNT, class Derivative>
10507 bool compareDerivative (const std::vector<Resource>& inputs, const vector<AllocationSp>& outputAllocs, const std::vector<Resource>&, TestLog& log)
10509 if (inputs.size() != 1 || outputAllocs.size() != 1)
10512 deUint32 successfulRuns = FLAVOUR_COUNT;
10513 std::string results[FLAVOUR_COUNT];
10514 vector<deUint8> inputBytes;
10516 inputs[0].getBytes(inputBytes);
10518 const deFloat16* inputAsFP16 = reinterpret_cast<deFloat16* const>(&inputBytes.front());
10519 const deFloat16* outputAsFP16 = static_cast<deFloat16*>(outputAllocs[0]->getHostPtr());
10521 DE_ASSERT(inputBytes.size() == R * R * N * sizeof(deFloat16));
10523 for (deUint32 flavor = 0; flavor < FLAVOUR_COUNT; ++flavor)
10524 if (compareDerivativeWithFlavor<R, N, Derivative> (inputAsFP16, outputAsFP16, flavor, results[flavor]))
10533 if (successfulRuns == 0)
10534 for (deUint32 flavor = 0; flavor < FLAVOUR_COUNT; flavor++)
10535 log << TestLog::Message << "At flavor #" << flavor << " " << results[flavor] << TestLog::EndMessage;
10537 return successfulRuns > 0;
10540 template<deUint32 R, deUint32 N>
10541 tcu::TestCaseGroup* createDerivativeTests (tcu::TestContext& testCtx)
10543 typedef getFDelta<R, N, getV0, getV1, getV_, getV_> getFDxFine;
10544 typedef getFDelta<R, N, getV_, getV_, getV0, getV1> getFDyFine;
10546 typedef getFDelta<R, N, getV0, getV1, getV0, getV0> getFdxCoarse0;
10547 typedef getFDelta<R, N, getV0, getV1, getV1, getV1> getFdxCoarse1;
10548 typedef getFDelta<R, N, getV0, getV0, getV0, getV1> getFdyCoarse0;
10549 typedef getFDelta<R, N, getV1, getV1, getV0, getV1> getFdyCoarse1;
10550 typedef getFOneOf<2, getFdxCoarse0, getFdxCoarse1> getFDxCoarse;
10551 typedef getFOneOf<2, getFdyCoarse0, getFdyCoarse1> getFDyCoarse;
10553 typedef calcWidthOf4<getFDxFine, getFDxFine, getFDyFine, getFDyFine> getFWidthFine;
10554 typedef calcWidthOf4<getFdxCoarse0, getFdxCoarse1, getFdyCoarse0, getFdyCoarse1> getFWidthCoarse;
10556 typedef getFOneOf<3, getFDxFine, getFDxCoarse> getFDx;
10557 typedef getFOneOf<3, getFDyFine, getFDyCoarse> getFDy;
10558 typedef getFOneOf<5, getFWidthFine, getFWidthCoarse> getFWidth;
10560 const std::string testGroupName (std::string("derivative_") + de::toString(N));
10561 de::MovePtr<tcu::TestCaseGroup> testGroup (new tcu::TestCaseGroup(testCtx, testGroupName.c_str(), "Derivative instruction tests"));
10563 de::Random rnd (deStringHash(testGroup->getName()));
10564 const deUint32 numDataPointsByAxis = R;
10565 const deUint32 numDataPoints = N * numDataPointsByAxis * numDataPointsByAxis;
10566 vector<deFloat16> float16InputX;
10567 vector<deFloat16> float16InputY;
10568 vector<deFloat16> float16InputW;
10569 vector<deFloat16> float16OutputDummy (numDataPoints, 0);
10570 RGBA defaultColors[4];
10572 getDefaultColors(defaultColors);
10574 float16InputX.reserve(numDataPoints);
10575 for (deUint32 y = 0; y < numDataPointsByAxis; ++y)
10576 for (deUint32 x = 0; x < numDataPointsByAxis; ++x)
10577 for (deUint32 n = 0; n < N; ++n)
10579 const float arg = static_cast<float>(2 * DE_PI) * static_cast<float>(x * (n + 1)) / static_cast<float>(1 * numDataPointsByAxis);
10582 float16InputX.push_back(tcu::Float16(sin(arg)).bits());
10584 float16InputX.push_back(tcu::Float16(cos(arg)).bits());
10587 float16InputY.reserve(numDataPoints);
10588 for (deUint32 y = 0; y < numDataPointsByAxis; ++y)
10589 for (deUint32 x = 0; x < numDataPointsByAxis; ++x)
10590 for (deUint32 n = 0; n < N; ++n)
10592 const float arg = static_cast<float>(2 * DE_PI) * static_cast<float>(y * (n + 1)) / static_cast<float>(1 * numDataPointsByAxis);
10595 float16InputY.push_back(tcu::Float16(sin(arg)).bits());
10597 float16InputY.push_back(tcu::Float16(cos(arg)).bits());
10600 const deFloat16 testNumbers[] =
10602 tcu::Float16( 2.0 ).bits(),
10603 tcu::Float16( 4.0 ).bits(),
10604 tcu::Float16( 8.0 ).bits(),
10605 tcu::Float16( 16.0 ).bits(),
10606 tcu::Float16( 32.0 ).bits(),
10607 tcu::Float16( 64.0 ).bits(),
10608 tcu::Float16( 128.0).bits(),
10609 tcu::Float16( 256.0).bits(),
10610 tcu::Float16( 512.0).bits(),
10611 tcu::Float16(-2.0 ).bits(),
10612 tcu::Float16(-4.0 ).bits(),
10613 tcu::Float16(-8.0 ).bits(),
10614 tcu::Float16(-16.0 ).bits(),
10615 tcu::Float16(-32.0 ).bits(),
10616 tcu::Float16(-64.0 ).bits(),
10617 tcu::Float16(-128.0).bits(),
10618 tcu::Float16(-256.0).bits(),
10619 tcu::Float16(-512.0).bits(),
10622 float16InputW.reserve(numDataPoints);
10623 for (deUint32 y = 0; y < numDataPointsByAxis; ++y)
10624 for (deUint32 x = 0; x < numDataPointsByAxis; ++x)
10625 for (deUint32 n = 0; n < N; ++n)
10626 float16InputW.push_back(testNumbers[rnd.getInt(0, DE_LENGTH_OF_ARRAY(testNumbers) - 1)]);
10630 const char* opCode;
10631 vector<deFloat16>& inputData;
10632 VerifyIOFunc verifyFunc;
10635 const TestOp testOps[] =
10637 { "OpDPdxFine" , float16InputX , compareDerivative<R, N, 1, getFDxFine > },
10638 { "OpDPdyFine" , float16InputY , compareDerivative<R, N, 1, getFDyFine > },
10639 { "OpFwidthFine" , float16InputW , compareDerivative<R, N, 1, getFWidthFine > },
10640 { "OpDPdxCoarse" , float16InputX , compareDerivative<R, N, 3, getFDx > },
10641 { "OpDPdyCoarse" , float16InputY , compareDerivative<R, N, 3, getFDy > },
10642 { "OpFwidthCoarse" , float16InputW , compareDerivative<R, N, 5, getFWidth > },
10643 { "OpDPdx" , float16InputX , compareDerivative<R, N, 3, getFDx > },
10644 { "OpDPdy" , float16InputY , compareDerivative<R, N, 3, getFDy > },
10645 { "OpFwidth" , float16InputW , compareDerivative<R, N, 5, getFWidth > },
10650 const deUint32 typeComponents;
10651 const char* typeName;
10652 const char* typeDecls;
10655 const TestType testTypes[] =
10665 " %v2f16 = OpTypeVector %f16 2\n"
10670 " %v4f16 = OpTypeVector %f16 4\n"
10674 const deUint32 testTypeNdx = (N == 1) ? 0
10677 : DE_LENGTH_OF_ARRAY(testTypes);
10678 const TestType& testType = testTypes[testTypeNdx];
10680 DE_ASSERT(testTypeNdx < DE_LENGTH_OF_ARRAY(testTypes));
10681 DE_ASSERT(testType.typeComponents == N);
10683 const StringTemplate preMain
10685 "%c_i32_ndp = OpConstant %i32 ${num_data_points}\n"
10686 " %c_u32_xw = OpConstant %u32 ${items_by_x}\n"
10687 " %f16 = OpTypeFloat 16\n"
10689 " %up_${tt} = OpTypePointer Uniform %${tt}\n"
10690 " %ra_${tt} = OpTypeArray %${tt} %c_i32_ndp\n"
10691 " %SSBO16 = OpTypeStruct %ra_${tt}\n"
10692 "%up_SSBO16 = OpTypePointer Uniform %SSBO16\n"
10693 " %ssbo_src = OpVariable %up_SSBO16 Uniform\n"
10694 " %ssbo_dst = OpVariable %up_SSBO16 Uniform\n"
10697 const StringTemplate decoration
10699 "OpDecorate %ra_${tt} ArrayStride ${tt_stride}\n"
10700 "OpMemberDecorate %SSBO16 0 Offset 0\n"
10701 "OpDecorate %SSBO16 BufferBlock\n"
10702 "OpDecorate %ssbo_src DescriptorSet 0\n"
10703 "OpDecorate %ssbo_src Binding 0\n"
10704 "OpDecorate %ssbo_dst DescriptorSet 0\n"
10705 "OpDecorate %ssbo_dst Binding 1\n"
10708 const StringTemplate testFun
10710 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
10711 " %param = OpFunctionParameter %v4f32\n"
10712 " %entry = OpLabel\n"
10714 " %loc_x_c = OpAccessChain %ip_f32 %BP_gl_FragCoord %c_i32_0\n"
10715 " %loc_y_c = OpAccessChain %ip_f32 %BP_gl_FragCoord %c_i32_1\n"
10716 " %x_c = OpLoad %f32 %loc_x_c\n"
10717 " %y_c = OpLoad %f32 %loc_y_c\n"
10718 " %x_idx = OpConvertFToU %u32 %x_c\n"
10719 " %y_idx = OpConvertFToU %u32 %y_c\n"
10720 " %ndx_y = OpIMul %u32 %y_idx %c_u32_xw\n"
10721 " %ndx = OpIAdd %u32 %ndx_y %x_idx\n"
10723 " %src = OpAccessChain %up_${tt} %ssbo_src %c_i32_0 %ndx\n"
10724 " %val_src = OpLoad %${tt} %src\n"
10725 " %val_dst = ${op_code} %${tt} %val_src\n"
10726 " %dst = OpAccessChain %up_${tt} %ssbo_dst %c_i32_0 %ndx\n"
10727 " OpStore %dst %val_dst\n"
10728 " OpBranch %merge\n"
10730 " %merge = OpLabel\n"
10731 " OpReturnValue %param\n"
10736 for (deUint32 testOpsIdx = 0; testOpsIdx < DE_LENGTH_OF_ARRAY(testOps); ++testOpsIdx)
10738 const TestOp& testOp = testOps[testOpsIdx];
10739 const string testName = de::toLower(string(testOp.opCode));
10740 const size_t typeStride = N * sizeof(deFloat16);
10741 GraphicsResources specResource;
10742 map<string, string> specs;
10743 VulkanFeatures features;
10744 vector<string> extensions;
10745 map<string, string> fragments;
10746 SpecConstants noSpecConstants;
10747 PushConstants noPushConstants;
10748 GraphicsInterfaces noInterfaces;
10750 specs["op_code"] = testOp.opCode;
10751 specs["num_data_points"] = de::toString(testOp.inputData.size() / N);
10752 specs["items_by_x"] = de::toString(numDataPointsByAxis);
10753 specs["tt"] = testType.typeName;
10754 specs["tt_stride"] = de::toString(typeStride);
10755 specs["type_decls"] = testType.typeDecls;
10757 fragments["extension"] = "OpExtension \"SPV_KHR_16bit_storage\"";
10758 fragments["capability"] = "OpCapability DerivativeControl\nOpCapability StorageUniformBufferBlock16\n";
10759 fragments["decoration"] = decoration.specialize(specs);
10760 fragments["pre_main"] = preMain.specialize(specs);
10761 fragments["testfun"] = testFun.specialize(specs);
10763 specResource.inputs.push_back(Resource(BufferSp(new Float16Buffer(testOp.inputData)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
10764 specResource.outputs.push_back(Resource(BufferSp(new Float16Buffer(float16OutputDummy)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
10765 specResource.verifyIO = testOp.verifyFunc;
10767 extensions.push_back("VK_KHR_16bit_storage");
10768 extensions.push_back("VK_KHR_shader_float16_int8");
10770 features.extFloat16Int8 = EXTFLOAT16INT8FEATURES_FLOAT16;
10771 features.ext16BitStorage = EXT16BITSTORAGEFEATURES_UNIFORM_BUFFER_BLOCK;
10773 createTestForStage(VK_SHADER_STAGE_FRAGMENT_BIT, testName.c_str(), defaultColors, defaultColors, fragments, noSpecConstants,
10774 noPushConstants, specResource, noInterfaces, extensions, features, testGroup.get(), QP_TEST_RESULT_FAIL, string(), true);
10777 return testGroup.release();
10780 bool compareFP16VectorExtractFunc (const std::vector<Resource>& inputs, const vector<AllocationSp>& outputAllocs, const std::vector<Resource>&, TestLog& log)
10782 if (inputs.size() != 2 || outputAllocs.size() != 1)
10785 vector<deUint8> input1Bytes;
10786 vector<deUint8> input2Bytes;
10788 inputs[0].getBytes(input1Bytes);
10789 inputs[1].getBytes(input2Bytes);
10791 DE_ASSERT(input1Bytes.size() > 0);
10792 DE_ASSERT(input2Bytes.size() > 0);
10793 DE_ASSERT(input2Bytes.size() % sizeof(deUint32) == 0);
10795 const size_t iterations = input2Bytes.size() / sizeof(deUint32);
10796 const size_t components = input1Bytes.size() / (sizeof(deFloat16) * iterations);
10797 const deFloat16* const input1AsFP16 = (const deFloat16*)&input1Bytes[0];
10798 const deUint32* const inputIndices = (const deUint32*)&input2Bytes[0];
10799 const deFloat16* const outputAsFP16 = (const deFloat16*)outputAllocs[0]->getHostPtr();
10802 DE_ASSERT(components == 2 || components == 4);
10803 DE_ASSERT(input1Bytes.size() == iterations * components * sizeof(deFloat16));
10805 for (size_t idx = 0; idx < iterations; ++idx)
10807 const deUint32 componentNdx = inputIndices[idx];
10809 DE_ASSERT(componentNdx < components);
10811 const deFloat16 expected = input1AsFP16[components * idx + componentNdx];
10813 if (!compare16BitFloat(expected, outputAsFP16[idx], error))
10815 log << TestLog::Message << "At " << idx << error << TestLog::EndMessage;
10824 template<class SpecResource>
10825 tcu::TestCaseGroup* createFloat16VectorExtractSet (tcu::TestContext& testCtx)
10827 de::MovePtr<tcu::TestCaseGroup> testGroup (new tcu::TestCaseGroup(testCtx, "opvectorextractdynamic", "OpVectorExtractDynamic tests"));
10829 de::Random rnd (deStringHash(testGroup->getName()));
10830 const deUint32 numDataPoints = 256;
10831 const vector<deFloat16> float16InputData = getFloat16s(rnd, numDataPoints);
10832 const vector<deFloat16> float16OutputDummy (float16InputData.size(), 0);
10836 const deUint32 typeComponents;
10837 const size_t typeStride;
10838 const char* typeName;
10839 const char* typeDecls;
10842 const TestType testTypes[] =
10846 2 * sizeof(deFloat16),
10848 " %v2f16 = OpTypeVector %f16 2\n"
10852 4 * sizeof(deFloat16),
10854 " %v3f16 = OpTypeVector %f16 3\n"
10858 4 * sizeof(deFloat16),
10860 " %v4f16 = OpTypeVector %f16 4\n"
10864 const StringTemplate preMain
10866 " %c_i32_ndp = OpConstant %i32 ${num_data_points}\n"
10867 " %f16 = OpTypeFloat 16\n"
10871 " %up_${tt} = OpTypePointer Uniform %${tt}\n"
10872 " %ra_${tt} = OpTypeArray %${tt} %c_i32_ndp\n"
10873 " %SSBO_SRC = OpTypeStruct %ra_${tt}\n"
10874 "%up_SSBO_SRC = OpTypePointer Uniform %SSBO_SRC\n"
10876 " %up_u32 = OpTypePointer Uniform %u32\n"
10877 " %ra_u32 = OpTypeArray %u32 %c_i32_ndp\n"
10878 " %SSBO_IDX = OpTypeStruct %ra_u32\n"
10879 "%up_SSBO_IDX = OpTypePointer Uniform %SSBO_IDX\n"
10881 " %up_f16 = OpTypePointer Uniform %f16\n"
10882 " %ra_f16 = OpTypeArray %f16 %c_i32_ndp\n"
10883 " %SSBO_DST = OpTypeStruct %ra_f16\n"
10884 "%up_SSBO_DST = OpTypePointer Uniform %SSBO_DST\n"
10886 " %ssbo_src = OpVariable %up_SSBO_SRC Uniform\n"
10887 " %ssbo_idx = OpVariable %up_SSBO_IDX Uniform\n"
10888 " %ssbo_dst = OpVariable %up_SSBO_DST Uniform\n"
10891 const StringTemplate decoration
10893 "OpDecorate %ra_${tt} ArrayStride ${tt_stride}\n"
10894 "OpMemberDecorate %SSBO_SRC 0 Offset 0\n"
10895 "OpDecorate %SSBO_SRC BufferBlock\n"
10896 "OpDecorate %ssbo_src DescriptorSet 0\n"
10897 "OpDecorate %ssbo_src Binding 0\n"
10899 "OpDecorate %ra_u32 ArrayStride 4\n"
10900 "OpMemberDecorate %SSBO_IDX 0 Offset 0\n"
10901 "OpDecorate %SSBO_IDX BufferBlock\n"
10902 "OpDecorate %ssbo_idx DescriptorSet 0\n"
10903 "OpDecorate %ssbo_idx Binding 1\n"
10905 "OpDecorate %ra_f16 ArrayStride 2\n"
10906 "OpMemberDecorate %SSBO_DST 0 Offset 0\n"
10907 "OpDecorate %SSBO_DST BufferBlock\n"
10908 "OpDecorate %ssbo_dst DescriptorSet 0\n"
10909 "OpDecorate %ssbo_dst Binding 2\n"
10912 const StringTemplate testFun
10914 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
10915 " %param = OpFunctionParameter %v4f32\n"
10916 " %entry = OpLabel\n"
10918 " %i = OpVariable %fp_i32 Function\n"
10919 " OpStore %i %c_i32_0\n"
10921 " %will_run = OpFunctionCall %bool %isUniqueIdZero\n"
10922 " OpSelectionMerge %end_if None\n"
10923 " OpBranchConditional %will_run %run_test %end_if\n"
10925 " %run_test = OpLabel\n"
10926 " OpBranch %loop\n"
10928 " %loop = OpLabel\n"
10929 " %i_cmp = OpLoad %i32 %i\n"
10930 " %lt = OpSLessThan %bool %i_cmp %c_i32_ndp\n"
10931 " OpLoopMerge %merge %next None\n"
10932 " OpBranchConditional %lt %write %merge\n"
10934 " %write = OpLabel\n"
10935 " %ndx = OpLoad %i32 %i\n"
10937 " %src = OpAccessChain %up_${tt} %ssbo_src %c_i32_0 %ndx\n"
10938 " %val_src = OpLoad %${tt} %src\n"
10940 " %src_idx = OpAccessChain %up_u32 %ssbo_idx %c_i32_0 %ndx\n"
10941 " %val_idx = OpLoad %u32 %src_idx\n"
10943 " %val_dst = OpVectorExtractDynamic %f16 %val_src %val_idx\n"
10944 " %dst = OpAccessChain %up_f16 %ssbo_dst %c_i32_0 %ndx\n"
10946 " OpStore %dst %val_dst\n"
10947 " OpBranch %next\n"
10949 " %next = OpLabel\n"
10950 " %i_cur = OpLoad %i32 %i\n"
10951 " %i_new = OpIAdd %i32 %i_cur %c_i32_1\n"
10952 " OpStore %i %i_new\n"
10953 " OpBranch %loop\n"
10955 " %merge = OpLabel\n"
10956 " OpBranch %end_if\n"
10957 " %end_if = OpLabel\n"
10958 " OpReturnValue %param\n"
10963 for (deUint32 testTypeIdx = 0; testTypeIdx < DE_LENGTH_OF_ARRAY(testTypes); ++testTypeIdx)
10965 const TestType& testType = testTypes[testTypeIdx];
10966 const string testName = testType.typeName;
10967 const size_t itemsPerType = testType.typeStride / sizeof(deFloat16);
10968 const size_t iterations = float16InputData.size() / itemsPerType;
10969 SpecResource specResource;
10970 map<string, string> specs;
10971 VulkanFeatures features;
10972 vector<deUint32> inputDataNdx;
10973 map<string, string> fragments;
10974 vector<string> extensions;
10976 for (deUint32 ndx = 0; ndx < iterations; ++ndx)
10977 inputDataNdx.push_back(rnd.getUint32() % testType.typeComponents);
10979 specs["num_data_points"] = de::toString(iterations);
10980 specs["tt"] = testType.typeName;
10981 specs["tt_stride"] = de::toString(testType.typeStride);
10982 specs["type_decl"] = testType.typeDecls;
10984 fragments["extension"] = "OpExtension \"SPV_KHR_16bit_storage\"";
10985 fragments["capability"] = "OpCapability StorageUniformBufferBlock16\n";
10986 fragments["decoration"] = decoration.specialize(specs);
10987 fragments["pre_main"] = preMain.specialize(specs);
10988 fragments["testfun"] = testFun.specialize(specs);
10990 specResource.inputs.push_back(Resource(BufferSp(new Float16Buffer(float16InputData)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
10991 specResource.inputs.push_back(Resource(BufferSp(new Uint32Buffer(inputDataNdx)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
10992 specResource.outputs.push_back(Resource(BufferSp(new Float16Buffer(float16OutputDummy)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
10993 specResource.verifyIO = compareFP16VectorExtractFunc;
10995 extensions.push_back("VK_KHR_16bit_storage");
10996 extensions.push_back("VK_KHR_shader_float16_int8");
10998 features.extFloat16Int8 = EXTFLOAT16INT8FEATURES_FLOAT16;
10999 features.ext16BitStorage = EXT16BITSTORAGEFEATURES_UNIFORM_BUFFER_BLOCK;
11001 finalizeTestsCreation(specResource, fragments, testCtx, *testGroup.get(), testName, features, extensions, IVec3(1, 1, 1));
11004 return testGroup.release();
11007 template<deUint32 COMPONENTS_COUNT, deUint32 REPLACEMENT>
11008 bool compareFP16VectorInsertFunc (const std::vector<Resource>& inputs, const vector<AllocationSp>& outputAllocs, const std::vector<Resource>&, TestLog& log)
11010 if (inputs.size() != 2 || outputAllocs.size() != 1)
11013 vector<deUint8> input1Bytes;
11014 vector<deUint8> input2Bytes;
11016 inputs[0].getBytes(input1Bytes);
11017 inputs[1].getBytes(input2Bytes);
11019 DE_ASSERT(input1Bytes.size() > 0);
11020 DE_ASSERT(input2Bytes.size() > 0);
11021 DE_ASSERT(input2Bytes.size() % sizeof(deUint32) == 0);
11023 const size_t iterations = input2Bytes.size() / sizeof(deUint32);
11024 const size_t componentsStride = input1Bytes.size() / (sizeof(deFloat16) * iterations);
11025 const deFloat16* const input1AsFP16 = (const deFloat16*)&input1Bytes[0];
11026 const deUint32* const inputIndices = (const deUint32*)&input2Bytes[0];
11027 const deFloat16* const outputAsFP16 = (const deFloat16*)outputAllocs[0]->getHostPtr();
11028 const deFloat16 magic = tcu::Float16(float(REPLACEMENT)).bits();
11031 DE_ASSERT(componentsStride == 2 || componentsStride == 4);
11032 DE_ASSERT(input1Bytes.size() == iterations * componentsStride * sizeof(deFloat16));
11034 for (size_t idx = 0; idx < iterations; ++idx)
11036 const deFloat16* inputVec = &input1AsFP16[componentsStride * idx];
11037 const deFloat16* outputVec = &outputAsFP16[componentsStride * idx];
11038 const deUint32 replacedCompNdx = inputIndices[idx];
11040 DE_ASSERT(replacedCompNdx < COMPONENTS_COUNT);
11042 for (size_t compNdx = 0; compNdx < COMPONENTS_COUNT; ++compNdx)
11044 const deFloat16 expected = (compNdx == replacedCompNdx) ? magic : inputVec[compNdx];
11046 if (!compare16BitFloat(expected, outputVec[compNdx], error))
11048 log << TestLog::Message << "At " << idx << "[" << compNdx << "]: " << error << TestLog::EndMessage;
11058 template<class SpecResource>
11059 tcu::TestCaseGroup* createFloat16VectorInsertSet (tcu::TestContext& testCtx)
11061 de::MovePtr<tcu::TestCaseGroup> testGroup (new tcu::TestCaseGroup(testCtx, "opvectorinsertdynamic", "OpVectorInsertDynamic tests"));
11063 de::Random rnd (deStringHash(testGroup->getName()));
11064 const deUint32 replacement = 42;
11065 const deUint32 numDataPoints = 256;
11066 const vector<deFloat16> float16InputData = getFloat16s(rnd, numDataPoints);
11067 const vector<deFloat16> float16OutputDummy (float16InputData.size(), 0);
11071 const deUint32 typeComponents;
11072 const size_t typeStride;
11073 const char* typeName;
11074 const char* typeDecls;
11075 VerifyIOFunc verifyIOFunc;
11078 const TestType testTypes[] =
11082 2 * sizeof(deFloat16),
11084 " %v2f16 = OpTypeVector %f16 2\n",
11085 compareFP16VectorInsertFunc<2, replacement>
11089 4 * sizeof(deFloat16),
11091 " %v3f16 = OpTypeVector %f16 3\n",
11092 compareFP16VectorInsertFunc<3, replacement>
11096 4 * sizeof(deFloat16),
11098 " %v4f16 = OpTypeVector %f16 4\n",
11099 compareFP16VectorInsertFunc<4, replacement>
11103 const StringTemplate preMain
11105 " %c_i32_ndp = OpConstant %i32 ${num_data_points}\n"
11106 " %f16 = OpTypeFloat 16\n"
11107 " %c_f16_ins = OpConstant %f16 ${replacement}\n"
11111 " %up_${tt} = OpTypePointer Uniform %${tt}\n"
11112 " %ra_${tt} = OpTypeArray %${tt} %c_i32_ndp\n"
11113 " %SSBO_SRC = OpTypeStruct %ra_${tt}\n"
11114 "%up_SSBO_SRC = OpTypePointer Uniform %SSBO_SRC\n"
11116 " %up_u32 = OpTypePointer Uniform %u32\n"
11117 " %ra_u32 = OpTypeArray %u32 %c_i32_ndp\n"
11118 " %SSBO_IDX = OpTypeStruct %ra_u32\n"
11119 "%up_SSBO_IDX = OpTypePointer Uniform %SSBO_IDX\n"
11121 " %SSBO_DST = OpTypeStruct %ra_${tt}\n"
11122 "%up_SSBO_DST = OpTypePointer Uniform %SSBO_DST\n"
11124 " %ssbo_src = OpVariable %up_SSBO_SRC Uniform\n"
11125 " %ssbo_idx = OpVariable %up_SSBO_IDX Uniform\n"
11126 " %ssbo_dst = OpVariable %up_SSBO_DST Uniform\n"
11129 const StringTemplate decoration
11131 "OpDecorate %ra_${tt} ArrayStride ${tt_stride}\n"
11132 "OpMemberDecorate %SSBO_SRC 0 Offset 0\n"
11133 "OpDecorate %SSBO_SRC BufferBlock\n"
11134 "OpDecorate %ssbo_src DescriptorSet 0\n"
11135 "OpDecorate %ssbo_src Binding 0\n"
11137 "OpDecorate %ra_u32 ArrayStride 4\n"
11138 "OpMemberDecorate %SSBO_IDX 0 Offset 0\n"
11139 "OpDecorate %SSBO_IDX BufferBlock\n"
11140 "OpDecorate %ssbo_idx DescriptorSet 0\n"
11141 "OpDecorate %ssbo_idx Binding 1\n"
11143 "OpMemberDecorate %SSBO_DST 0 Offset 0\n"
11144 "OpDecorate %SSBO_DST BufferBlock\n"
11145 "OpDecorate %ssbo_dst DescriptorSet 0\n"
11146 "OpDecorate %ssbo_dst Binding 2\n"
11149 const StringTemplate testFun
11151 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
11152 " %param = OpFunctionParameter %v4f32\n"
11153 " %entry = OpLabel\n"
11155 " %i = OpVariable %fp_i32 Function\n"
11156 " OpStore %i %c_i32_0\n"
11158 " %will_run = OpFunctionCall %bool %isUniqueIdZero\n"
11159 " OpSelectionMerge %end_if None\n"
11160 " OpBranchConditional %will_run %run_test %end_if\n"
11162 " %run_test = OpLabel\n"
11163 " OpBranch %loop\n"
11165 " %loop = OpLabel\n"
11166 " %i_cmp = OpLoad %i32 %i\n"
11167 " %lt = OpSLessThan %bool %i_cmp %c_i32_ndp\n"
11168 " OpLoopMerge %merge %next None\n"
11169 " OpBranchConditional %lt %write %merge\n"
11171 " %write = OpLabel\n"
11172 " %ndx = OpLoad %i32 %i\n"
11174 " %src = OpAccessChain %up_${tt} %ssbo_src %c_i32_0 %ndx\n"
11175 " %val_src = OpLoad %${tt} %src\n"
11177 " %src_idx = OpAccessChain %up_u32 %ssbo_idx %c_i32_0 %ndx\n"
11178 " %val_idx = OpLoad %u32 %src_idx\n"
11180 " %val_dst = OpVectorInsertDynamic %${tt} %val_src %c_f16_ins %val_idx\n"
11181 " %dst = OpAccessChain %up_${tt} %ssbo_dst %c_i32_0 %ndx\n"
11183 " OpStore %dst %val_dst\n"
11184 " OpBranch %next\n"
11186 " %next = OpLabel\n"
11187 " %i_cur = OpLoad %i32 %i\n"
11188 " %i_new = OpIAdd %i32 %i_cur %c_i32_1\n"
11189 " OpStore %i %i_new\n"
11190 " OpBranch %loop\n"
11192 " %merge = OpLabel\n"
11193 " OpBranch %end_if\n"
11194 " %end_if = OpLabel\n"
11195 " OpReturnValue %param\n"
11200 for (deUint32 testTypeIdx = 0; testTypeIdx < DE_LENGTH_OF_ARRAY(testTypes); ++testTypeIdx)
11202 const TestType& testType = testTypes[testTypeIdx];
11203 const string testName = testType.typeName;
11204 const size_t itemsPerType = testType.typeStride / sizeof(deFloat16);
11205 const size_t iterations = float16InputData.size() / itemsPerType;
11206 SpecResource specResource;
11207 map<string, string> specs;
11208 VulkanFeatures features;
11209 vector<deUint32> inputDataNdx;
11210 map<string, string> fragments;
11211 vector<string> extensions;
11213 for (deUint32 ndx = 0; ndx < iterations; ++ndx)
11214 inputDataNdx.push_back(rnd.getUint32() % testType.typeComponents);
11216 specs["num_data_points"] = de::toString(iterations);
11217 specs["tt"] = testType.typeName;
11218 specs["tt_stride"] = de::toString(testType.typeStride);
11219 specs["type_decl"] = testType.typeDecls;
11220 specs["replacement"] = de::toString(replacement);
11222 fragments["extension"] = "OpExtension \"SPV_KHR_16bit_storage\"";
11223 fragments["capability"] = "OpCapability StorageUniformBufferBlock16\n";
11224 fragments["decoration"] = decoration.specialize(specs);
11225 fragments["pre_main"] = preMain.specialize(specs);
11226 fragments["testfun"] = testFun.specialize(specs);
11228 specResource.inputs.push_back(Resource(BufferSp(new Float16Buffer(float16InputData)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
11229 specResource.inputs.push_back(Resource(BufferSp(new Uint32Buffer(inputDataNdx)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
11230 specResource.outputs.push_back(Resource(BufferSp(new Float16Buffer(float16OutputDummy)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
11231 specResource.verifyIO = testType.verifyIOFunc;
11233 extensions.push_back("VK_KHR_16bit_storage");
11234 extensions.push_back("VK_KHR_shader_float16_int8");
11236 features.extFloat16Int8 = EXTFLOAT16INT8FEATURES_FLOAT16;
11237 features.ext16BitStorage = EXT16BITSTORAGEFEATURES_UNIFORM_BUFFER_BLOCK;
11239 finalizeTestsCreation(specResource, fragments, testCtx, *testGroup.get(), testName, features, extensions, IVec3(1, 1, 1));
11242 return testGroup.release();
11245 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)
11247 const size_t compNdxCount = (vec1Len + vec2Len + 1);
11248 const size_t compNdxLimited = iteration % (compNdxCount * compNdxCount);
11251 switch (componentNdx)
11253 case 0: comp = compNdxLimited / compNdxCount; break;
11254 case 1: comp = compNdxLimited % compNdxCount; break;
11255 case 2: comp = 0; break;
11256 case 3: comp = 1; break;
11257 default: TCU_THROW(InternalError, "Impossible");
11260 if (comp >= vec1Len + vec2Len)
11268 return (comp < vec1Len) ? input1Vec[comp] : input2Vec[comp - vec1Len];
11272 template<deUint32 DST_COMPONENTS_COUNT, deUint32 SRC0_COMPONENTS_COUNT, deUint32 SRC1_COMPONENTS_COUNT>
11273 bool compareFP16VectorShuffleFunc (const std::vector<Resource>& inputs, const vector<AllocationSp>& outputAllocs, const std::vector<Resource>&, TestLog& log)
11275 DE_STATIC_ASSERT(DST_COMPONENTS_COUNT == 2 || DST_COMPONENTS_COUNT == 3 || DST_COMPONENTS_COUNT == 4);
11276 DE_STATIC_ASSERT(SRC0_COMPONENTS_COUNT == 2 || SRC0_COMPONENTS_COUNT == 3 || SRC0_COMPONENTS_COUNT == 4);
11277 DE_STATIC_ASSERT(SRC1_COMPONENTS_COUNT == 2 || SRC1_COMPONENTS_COUNT == 3 || SRC1_COMPONENTS_COUNT == 4);
11279 if (inputs.size() != 2 || outputAllocs.size() != 1)
11282 vector<deUint8> input1Bytes;
11283 vector<deUint8> input2Bytes;
11285 inputs[0].getBytes(input1Bytes);
11286 inputs[1].getBytes(input2Bytes);
11288 DE_ASSERT(input1Bytes.size() > 0);
11289 DE_ASSERT(input2Bytes.size() > 0);
11290 DE_ASSERT(input2Bytes.size() % sizeof(deFloat16) == 0);
11292 const size_t componentsStrideDst = (DST_COMPONENTS_COUNT == 3) ? 4 : DST_COMPONENTS_COUNT;
11293 const size_t componentsStrideSrc0 = (SRC0_COMPONENTS_COUNT == 3) ? 4 : SRC0_COMPONENTS_COUNT;
11294 const size_t componentsStrideSrc1 = (SRC1_COMPONENTS_COUNT == 3) ? 4 : SRC1_COMPONENTS_COUNT;
11295 const size_t iterations = input1Bytes.size() / (componentsStrideSrc0 * sizeof(deFloat16));
11296 const deFloat16* const input1AsFP16 = (const deFloat16*)&input1Bytes[0];
11297 const deFloat16* const input2AsFP16 = (const deFloat16*)&input2Bytes[0];
11298 const deFloat16* const outputAsFP16 = (const deFloat16*)outputAllocs[0]->getHostPtr();
11301 DE_ASSERT(input1Bytes.size() == iterations * componentsStrideSrc0 * sizeof(deFloat16));
11302 DE_ASSERT(input2Bytes.size() == iterations * componentsStrideSrc1 * sizeof(deFloat16));
11304 for (size_t idx = 0; idx < iterations; ++idx)
11306 const deFloat16* input1Vec = &input1AsFP16[componentsStrideSrc0 * idx];
11307 const deFloat16* input2Vec = &input2AsFP16[componentsStrideSrc1 * idx];
11308 const deFloat16* outputVec = &outputAsFP16[componentsStrideDst * idx];
11310 for (size_t compNdx = 0; compNdx < DST_COMPONENTS_COUNT; ++compNdx)
11312 bool validate = true;
11313 deFloat16 expected = getShuffledComponent(idx, compNdx, input1Vec, input2Vec, SRC0_COMPONENTS_COUNT, SRC1_COMPONENTS_COUNT, validate);
11315 if (validate && !compare16BitFloat(expected, outputVec[compNdx], error))
11317 log << TestLog::Message << "At " << idx << "[" << compNdx << "]: " << error << TestLog::EndMessage;
11327 VerifyIOFunc getFloat16VectorShuffleVerifyIOFunc (deUint32 dstComponentsCount, deUint32 src0ComponentsCount, deUint32 src1ComponentsCount)
11329 DE_ASSERT(dstComponentsCount <= 4);
11330 DE_ASSERT(src0ComponentsCount <= 4);
11331 DE_ASSERT(src1ComponentsCount <= 4);
11332 deUint32 funcCode = 100 * dstComponentsCount + 10 * src0ComponentsCount + src1ComponentsCount;
11336 case 222:return compareFP16VectorShuffleFunc<2, 2, 2>;
11337 case 223:return compareFP16VectorShuffleFunc<2, 2, 3>;
11338 case 224:return compareFP16VectorShuffleFunc<2, 2, 4>;
11339 case 232:return compareFP16VectorShuffleFunc<2, 3, 2>;
11340 case 233:return compareFP16VectorShuffleFunc<2, 3, 3>;
11341 case 234:return compareFP16VectorShuffleFunc<2, 3, 4>;
11342 case 242:return compareFP16VectorShuffleFunc<2, 4, 2>;
11343 case 243:return compareFP16VectorShuffleFunc<2, 4, 3>;
11344 case 244:return compareFP16VectorShuffleFunc<2, 4, 4>;
11345 case 322:return compareFP16VectorShuffleFunc<3, 2, 2>;
11346 case 323:return compareFP16VectorShuffleFunc<3, 2, 3>;
11347 case 324:return compareFP16VectorShuffleFunc<3, 2, 4>;
11348 case 332:return compareFP16VectorShuffleFunc<3, 3, 2>;
11349 case 333:return compareFP16VectorShuffleFunc<3, 3, 3>;
11350 case 334:return compareFP16VectorShuffleFunc<3, 3, 4>;
11351 case 342:return compareFP16VectorShuffleFunc<3, 4, 2>;
11352 case 343:return compareFP16VectorShuffleFunc<3, 4, 3>;
11353 case 344:return compareFP16VectorShuffleFunc<3, 4, 4>;
11354 case 422:return compareFP16VectorShuffleFunc<4, 2, 2>;
11355 case 423:return compareFP16VectorShuffleFunc<4, 2, 3>;
11356 case 424:return compareFP16VectorShuffleFunc<4, 2, 4>;
11357 case 432:return compareFP16VectorShuffleFunc<4, 3, 2>;
11358 case 433:return compareFP16VectorShuffleFunc<4, 3, 3>;
11359 case 434:return compareFP16VectorShuffleFunc<4, 3, 4>;
11360 case 442:return compareFP16VectorShuffleFunc<4, 4, 2>;
11361 case 443:return compareFP16VectorShuffleFunc<4, 4, 3>;
11362 case 444:return compareFP16VectorShuffleFunc<4, 4, 4>;
11363 default: TCU_THROW(InternalError, "Invalid number of components specified.");
11367 template<class SpecResource>
11368 tcu::TestCaseGroup* createFloat16VectorShuffleSet (tcu::TestContext& testCtx)
11370 de::MovePtr<tcu::TestCaseGroup> testGroup (new tcu::TestCaseGroup(testCtx, "opvectorshuffle", "OpVectorShuffle tests"));
11371 const int testSpecificSeed = deStringHash(testGroup->getName());
11372 const int seed = testCtx.getCommandLine().getBaseSeed() ^ testSpecificSeed;
11373 de::Random rnd (seed);
11374 const deUint32 numDataPoints = 128;
11375 map<string, string> fragments;
11379 const deUint32 typeComponents;
11380 const char* typeName;
11383 const TestType testTypes[] =
11399 const StringTemplate preMain
11401 " %c_i32_ndp = OpConstant %i32 ${num_data_points}\n"
11402 " %c_i32_cc = OpConstant %i32 ${case_count}\n"
11403 " %f16 = OpTypeFloat 16\n"
11404 " %v2f16 = OpTypeVector %f16 2\n"
11405 " %v3f16 = OpTypeVector %f16 3\n"
11406 " %v4f16 = OpTypeVector %f16 4\n"
11408 " %up_v2f16 = OpTypePointer Uniform %v2f16\n"
11409 " %ra_v2f16 = OpTypeArray %v2f16 %c_i32_ndp\n"
11410 " %SSBO_v2f16 = OpTypeStruct %ra_v2f16\n"
11411 "%up_SSBO_v2f16 = OpTypePointer Uniform %SSBO_v2f16\n"
11413 " %up_v3f16 = OpTypePointer Uniform %v3f16\n"
11414 " %ra_v3f16 = OpTypeArray %v3f16 %c_i32_ndp\n"
11415 " %SSBO_v3f16 = OpTypeStruct %ra_v3f16\n"
11416 "%up_SSBO_v3f16 = OpTypePointer Uniform %SSBO_v3f16\n"
11418 " %up_v4f16 = OpTypePointer Uniform %v4f16\n"
11419 " %ra_v4f16 = OpTypeArray %v4f16 %c_i32_ndp\n"
11420 " %SSBO_v4f16 = OpTypeStruct %ra_v4f16\n"
11421 "%up_SSBO_v4f16 = OpTypePointer Uniform %SSBO_v4f16\n"
11423 " %fun_t = OpTypeFunction %${tt_dst} %${tt_src0} %${tt_src1} %i32\n"
11425 " %ssbo_src0 = OpVariable %up_SSBO_${tt_src0} Uniform\n"
11426 " %ssbo_src1 = OpVariable %up_SSBO_${tt_src1} Uniform\n"
11427 " %ssbo_dst = OpVariable %up_SSBO_${tt_dst} Uniform\n"
11430 const StringTemplate decoration
11432 "OpDecorate %ra_v2f16 ArrayStride 4\n"
11433 "OpDecorate %ra_v3f16 ArrayStride 8\n"
11434 "OpDecorate %ra_v4f16 ArrayStride 8\n"
11436 "OpMemberDecorate %SSBO_v2f16 0 Offset 0\n"
11437 "OpDecorate %SSBO_v2f16 BufferBlock\n"
11439 "OpMemberDecorate %SSBO_v3f16 0 Offset 0\n"
11440 "OpDecorate %SSBO_v3f16 BufferBlock\n"
11442 "OpMemberDecorate %SSBO_v4f16 0 Offset 0\n"
11443 "OpDecorate %SSBO_v4f16 BufferBlock\n"
11445 "OpDecorate %ssbo_src0 DescriptorSet 0\n"
11446 "OpDecorate %ssbo_src0 Binding 0\n"
11447 "OpDecorate %ssbo_src1 DescriptorSet 0\n"
11448 "OpDecorate %ssbo_src1 Binding 1\n"
11449 "OpDecorate %ssbo_dst DescriptorSet 0\n"
11450 "OpDecorate %ssbo_dst Binding 2\n"
11453 const StringTemplate testFun
11455 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
11456 " %param = OpFunctionParameter %v4f32\n"
11457 " %entry = OpLabel\n"
11459 " %i = OpVariable %fp_i32 Function\n"
11460 " OpStore %i %c_i32_0\n"
11462 " %will_run = OpFunctionCall %bool %isUniqueIdZero\n"
11463 " OpSelectionMerge %end_if None\n"
11464 " OpBranchConditional %will_run %run_test %end_if\n"
11466 " %run_test = OpLabel\n"
11467 " OpBranch %loop\n"
11469 " %loop = OpLabel\n"
11470 " %i_cmp = OpLoad %i32 %i\n"
11471 " %lt = OpSLessThan %bool %i_cmp %c_i32_ndp\n"
11472 " OpLoopMerge %merge %next None\n"
11473 " OpBranchConditional %lt %write %merge\n"
11475 " %write = OpLabel\n"
11476 " %ndx = OpLoad %i32 %i\n"
11477 " %src0 = OpAccessChain %up_${tt_src0} %ssbo_src0 %c_i32_0 %ndx\n"
11478 " %val_src0 = OpLoad %${tt_src0} %src0\n"
11479 " %src1 = OpAccessChain %up_${tt_src1} %ssbo_src1 %c_i32_0 %ndx\n"
11480 " %val_src1 = OpLoad %${tt_src1} %src1\n"
11481 " %val_dst = OpFunctionCall %${tt_dst} %sw_fun %val_src0 %val_src1 %ndx\n"
11482 " %dst = OpAccessChain %up_${tt_dst} %ssbo_dst %c_i32_0 %ndx\n"
11483 " OpStore %dst %val_dst\n"
11484 " OpBranch %next\n"
11486 " %next = OpLabel\n"
11487 " %i_cur = OpLoad %i32 %i\n"
11488 " %i_new = OpIAdd %i32 %i_cur %c_i32_1\n"
11489 " OpStore %i %i_new\n"
11490 " OpBranch %loop\n"
11492 " %merge = OpLabel\n"
11493 " OpBranch %end_if\n"
11494 " %end_if = OpLabel\n"
11495 " OpReturnValue %param\n"
11499 " %sw_fun = OpFunction %${tt_dst} None %fun_t\n"
11500 "%sw_param0 = OpFunctionParameter %${tt_src0}\n"
11501 "%sw_param1 = OpFunctionParameter %${tt_src1}\n"
11502 "%sw_paramn = OpFunctionParameter %i32\n"
11503 " %sw_entry = OpLabel\n"
11504 " %modulo = OpSMod %i32 %sw_paramn %c_i32_cc\n"
11505 " OpSelectionMerge %switch_e None\n"
11506 " OpSwitch %modulo %default ${case_list}\n"
11508 "%default = OpLabel\n"
11509 " OpUnreachable\n" // Unreachable default case for switch statement
11510 "%switch_e = OpLabel\n"
11511 " OpUnreachable\n" // Unreachable merge block for switch statement
11515 const StringTemplate testCaseBody
11517 "%case_${case_ndx} = OpLabel\n"
11518 "%val_dst_${case_ndx} = OpVectorShuffle %${tt_dst} %sw_param0 %sw_param1 ${shuffle}\n"
11519 " OpReturnValue %val_dst_${case_ndx}\n"
11522 for (deUint32 dstTypeIdx = 0; dstTypeIdx < DE_LENGTH_OF_ARRAY(testTypes); ++dstTypeIdx)
11524 const TestType& dstType = testTypes[dstTypeIdx];
11526 for (deUint32 comp0Idx = 0; comp0Idx < DE_LENGTH_OF_ARRAY(testTypes); ++comp0Idx)
11528 const TestType& src0Type = testTypes[comp0Idx];
11530 for (deUint32 comp1Idx = 0; comp1Idx < DE_LENGTH_OF_ARRAY(testTypes); ++comp1Idx)
11532 const TestType& src1Type = testTypes[comp1Idx];
11533 const deUint32 input0Stride = (src0Type.typeComponents == 3) ? 4 : src0Type.typeComponents;
11534 const deUint32 input1Stride = (src1Type.typeComponents == 3) ? 4 : src1Type.typeComponents;
11535 const deUint32 outputStride = (dstType.typeComponents == 3) ? 4 : dstType.typeComponents;
11536 const vector<deFloat16> float16Input0Data = getFloat16s(rnd, input0Stride * numDataPoints);
11537 const vector<deFloat16> float16Input1Data = getFloat16s(rnd, input1Stride * numDataPoints);
11538 const vector<deFloat16> float16OutputDummy (outputStride * numDataPoints, 0);
11539 const string testName = de::toString(dstType.typeComponents) + de::toString(src0Type.typeComponents) + de::toString(src1Type.typeComponents);
11540 deUint32 caseCount = 0;
11541 SpecResource specResource;
11542 map<string, string> specs;
11543 vector<string> extensions;
11544 VulkanFeatures features;
11550 vector<string> componentList;
11552 // Generate component possible indices for OpVectorShuffle for components 0 and 1 in output vector
11554 deUint32 caseNo = 0;
11556 for (deUint32 comp0IdxLocal = 0; comp0IdxLocal < src0Type.typeComponents; ++comp0IdxLocal)
11557 componentList.push_back(de::toString(caseNo++));
11558 for (deUint32 comp1IdxLocal = 0; comp1IdxLocal < src1Type.typeComponents; ++comp1IdxLocal)
11559 componentList.push_back(de::toString(caseNo++));
11560 componentList.push_back("0xFFFFFFFF");
11563 for (deUint32 comp0IdxLocal = 0; comp0IdxLocal < componentList.size(); ++comp0IdxLocal)
11565 for (deUint32 comp1IdxLocal = 0; comp1IdxLocal < componentList.size(); ++comp1IdxLocal)
11567 map<string, string> specCase;
11568 string shuffle = componentList[comp0IdxLocal] + " " + componentList[comp1IdxLocal];
11570 for (deUint32 compIdx = 2; compIdx < dstType.typeComponents; ++compIdx)
11571 shuffle += " " + de::toString(compIdx - 2);
11573 specCase["case_ndx"] = de::toString(caseCount);
11574 specCase["shuffle"] = shuffle;
11575 specCase["tt_dst"] = dstType.typeName;
11577 caseBodies += testCaseBody.specialize(specCase);
11578 caseList += de::toString(caseCount) + " %case_" + de::toString(caseCount) + " ";
11585 specs["num_data_points"] = de::toString(numDataPoints);
11586 specs["tt_dst"] = dstType.typeName;
11587 specs["tt_src0"] = src0Type.typeName;
11588 specs["tt_src1"] = src1Type.typeName;
11589 specs["case_bodies"] = caseBodies;
11590 specs["case_list"] = caseList;
11591 specs["case_count"] = de::toString(caseCount);
11593 fragments["extension"] = "OpExtension \"SPV_KHR_16bit_storage\"";
11594 fragments["capability"] = "OpCapability StorageUniformBufferBlock16\n";
11595 fragments["decoration"] = decoration.specialize(specs);
11596 fragments["pre_main"] = preMain.specialize(specs);
11597 fragments["testfun"] = testFun.specialize(specs);
11599 specResource.inputs.push_back(Resource(BufferSp(new Float16Buffer(float16Input0Data)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
11600 specResource.inputs.push_back(Resource(BufferSp(new Float16Buffer(float16Input1Data)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
11601 specResource.outputs.push_back(Resource(BufferSp(new Float16Buffer(float16OutputDummy)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
11602 specResource.verifyIO = getFloat16VectorShuffleVerifyIOFunc(dstType.typeComponents, src0Type.typeComponents, src1Type.typeComponents);
11604 extensions.push_back("VK_KHR_16bit_storage");
11605 extensions.push_back("VK_KHR_shader_float16_int8");
11607 features.extFloat16Int8 = EXTFLOAT16INT8FEATURES_FLOAT16;
11608 features.ext16BitStorage = EXT16BITSTORAGEFEATURES_UNIFORM_BUFFER_BLOCK;
11610 finalizeTestsCreation(specResource, fragments, testCtx, *testGroup.get(), testName, features, extensions, IVec3(1, 1, 1));
11615 return testGroup.release();
11618 bool compareFP16CompositeFunc (const std::vector<Resource>& inputs, const vector<AllocationSp>& outputAllocs, const std::vector<Resource>&, TestLog& log)
11620 if (inputs.size() != 1 || outputAllocs.size() != 1)
11623 vector<deUint8> input1Bytes;
11625 inputs[0].getBytes(input1Bytes);
11627 DE_ASSERT(input1Bytes.size() > 0);
11628 DE_ASSERT(input1Bytes.size() % sizeof(deFloat16) == 0);
11630 const size_t iterations = input1Bytes.size() / sizeof(deFloat16);
11631 const deFloat16* const input1AsFP16 = (const deFloat16*)&input1Bytes[0];
11632 const deFloat16* const outputAsFP16 = (const deFloat16*)outputAllocs[0]->getHostPtr();
11633 const deFloat16 exceptionValue = tcu::Float16(-1.0).bits();
11636 for (size_t idx = 0; idx < iterations; ++idx)
11638 if (input1AsFP16[idx] == exceptionValue)
11641 if (!compare16BitFloat(input1AsFP16[idx], outputAsFP16[idx], error))
11643 log << TestLog::Message << "At " << idx << ":" << error << TestLog::EndMessage;
11652 template<class SpecResource>
11653 tcu::TestCaseGroup* createFloat16CompositeConstructSet (tcu::TestContext& testCtx)
11655 de::MovePtr<tcu::TestCaseGroup> testGroup (new tcu::TestCaseGroup(testCtx, "opcompositeconstruct", "OpCompositeConstruct tests"));
11656 const deUint32 numElements = 8;
11657 const string testName = "struct";
11658 const deUint32 structItemsCount = 88;
11659 const deUint32 exceptionIndices[] = { 1, 7, 15, 17, 25, 33, 51, 55, 59, 63, 67, 71, 84, 85, 86, 87 };
11660 const deFloat16 exceptionValue = tcu::Float16(-1.0).bits();
11661 const deUint32 fieldModifier = 2;
11662 const deUint32 fieldModifiedMulIndex = 60;
11663 const deUint32 fieldModifiedAddIndex = 66;
11665 const StringTemplate preMain
11667 " %c_i32_ndp = OpConstant %i32 ${num_elements}\n"
11668 " %f16 = OpTypeFloat 16\n"
11669 " %v2f16 = OpTypeVector %f16 2\n"
11670 " %v3f16 = OpTypeVector %f16 3\n"
11671 " %v4f16 = OpTypeVector %f16 4\n"
11672 " %c_f16_mod = OpConstant %f16 ${field_modifier}\n"
11676 " %c_u32_5 = OpConstant %u32 5\n"
11678 " %f16arr3 = OpTypeArray %f16 %c_u32_3\n"
11679 " %v2f16arr3 = OpTypeArray %v2f16 %c_u32_3\n"
11680 " %v2f16arr5 = OpTypeArray %v2f16 %c_u32_5\n"
11681 " %v3f16arr5 = OpTypeArray %v3f16 %c_u32_5\n"
11682 " %v4f16arr3 = OpTypeArray %v4f16 %c_u32_3\n"
11683 " %struct16 = OpTypeStruct %f16 %v2f16arr3\n"
11684 " %struct16arr3 = OpTypeArray %struct16 %c_u32_3\n"
11685 " %st_test = OpTypeStruct %f16 %v2f16 %v3f16 %v4f16 %f16arr3 %struct16arr3 %v2f16arr5 %f16 %v3f16arr5 %v4f16arr3\n"
11687 " %up_st = OpTypePointer Uniform %st_test\n"
11688 " %ra_st = OpTypeArray %st_test %c_i32_ndp\n"
11689 " %SSBO_st = OpTypeStruct %ra_st\n"
11690 " %up_SSBO_st = OpTypePointer Uniform %SSBO_st\n"
11692 " %ssbo_dst = OpVariable %up_SSBO_st Uniform\n"
11695 const StringTemplate decoration
11697 "OpDecorate %SSBO_st BufferBlock\n"
11698 "OpDecorate %ra_st ArrayStride ${struct_item_size}\n"
11699 "OpDecorate %ssbo_dst DescriptorSet 0\n"
11700 "OpDecorate %ssbo_dst Binding 1\n"
11702 "OpMemberDecorate %SSBO_st 0 Offset 0\n"
11704 "OpDecorate %v2f16arr3 ArrayStride 4\n"
11705 "OpMemberDecorate %struct16 0 Offset 0\n"
11706 "OpMemberDecorate %struct16 1 Offset 4\n"
11707 "OpDecorate %struct16arr3 ArrayStride 16\n"
11708 "OpDecorate %f16arr3 ArrayStride 2\n"
11709 "OpDecorate %v2f16arr5 ArrayStride 4\n"
11710 "OpDecorate %v3f16arr5 ArrayStride 8\n"
11711 "OpDecorate %v4f16arr3 ArrayStride 8\n"
11713 "OpMemberDecorate %st_test 0 Offset 0\n"
11714 "OpMemberDecorate %st_test 1 Offset 4\n"
11715 "OpMemberDecorate %st_test 2 Offset 8\n"
11716 "OpMemberDecorate %st_test 3 Offset 16\n"
11717 "OpMemberDecorate %st_test 4 Offset 24\n"
11718 "OpMemberDecorate %st_test 5 Offset 32\n"
11719 "OpMemberDecorate %st_test 6 Offset 80\n"
11720 "OpMemberDecorate %st_test 7 Offset 100\n"
11721 "OpMemberDecorate %st_test 8 Offset 104\n"
11722 "OpMemberDecorate %st_test 9 Offset 144\n"
11725 const StringTemplate testFun
11727 " %test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
11728 " %param = OpFunctionParameter %v4f32\n"
11729 " %entry = OpLabel\n"
11731 " %i = OpVariable %fp_i32 Function\n"
11732 " OpStore %i %c_i32_0\n"
11734 " %will_run = OpFunctionCall %bool %isUniqueIdZero\n"
11735 " OpSelectionMerge %end_if None\n"
11736 " OpBranchConditional %will_run %run_test %end_if\n"
11738 " %run_test = OpLabel\n"
11739 " OpBranch %loop\n"
11741 " %loop = OpLabel\n"
11742 " %i_cmp = OpLoad %i32 %i\n"
11743 " %lt = OpSLessThan %bool %i_cmp %c_i32_ndp\n"
11744 " OpLoopMerge %merge %next None\n"
11745 " OpBranchConditional %lt %write %merge\n"
11747 " %write = OpLabel\n"
11748 " %ndx = OpLoad %i32 %i\n"
11750 " %fld1 = OpCompositeConstruct %v2f16 %c_f16_2 %c_f16_3\n"
11751 " %fld2 = OpCompositeConstruct %v3f16 %c_f16_4 %c_f16_5 %c_f16_6\n"
11752 " %fld3 = OpCompositeConstruct %v4f16 %c_f16_8 %c_f16_9 %c_f16_10 %c_f16_11\n"
11754 " %fld4 = OpCompositeConstruct %f16arr3 %c_f16_12 %c_f16_13 %c_f16_14\n"
11756 "%fld5_0_1_0 = OpCompositeConstruct %v2f16 %c_f16_18 %c_f16_19\n"
11757 "%fld5_0_1_1 = OpCompositeConstruct %v2f16 %c_f16_20 %c_f16_21\n"
11758 "%fld5_0_1_2 = OpCompositeConstruct %v2f16 %c_f16_22 %c_f16_23\n"
11759 " %fld5_0_1 = OpCompositeConstruct %v2f16arr3 %fld5_0_1_0 %fld5_0_1_1 %fld5_0_1_2\n"
11760 " %fld5_0 = OpCompositeConstruct %struct16 %c_f16_16 %fld5_0_1\n"
11762 "%fld5_1_1_0 = OpCompositeConstruct %v2f16 %c_f16_26 %c_f16_27\n"
11763 "%fld5_1_1_1 = OpCompositeConstruct %v2f16 %c_f16_28 %c_f16_29\n"
11764 "%fld5_1_1_2 = OpCompositeConstruct %v2f16 %c_f16_30 %c_f16_31\n"
11765 " %fld5_1_1 = OpCompositeConstruct %v2f16arr3 %fld5_1_1_0 %fld5_1_1_1 %fld5_1_1_2\n"
11766 " %fld5_1 = OpCompositeConstruct %struct16 %c_f16_24 %fld5_1_1\n"
11768 "%fld5_2_1_0 = OpCompositeConstruct %v2f16 %c_f16_34 %c_f16_35\n"
11769 "%fld5_2_1_1 = OpCompositeConstruct %v2f16 %c_f16_36 %c_f16_37\n"
11770 "%fld5_2_1_2 = OpCompositeConstruct %v2f16 %c_f16_38 %c_f16_39\n"
11771 " %fld5_2_1 = OpCompositeConstruct %v2f16arr3 %fld5_2_1_0 %fld5_2_1_1 %fld5_2_1_2\n"
11772 " %fld5_2 = OpCompositeConstruct %struct16 %c_f16_32 %fld5_2_1\n"
11774 " %fld5 = OpCompositeConstruct %struct16arr3 %fld5_0 %fld5_1 %fld5_2\n"
11776 " %fld6_0 = OpCompositeConstruct %v2f16 %c_f16_40 %c_f16_41\n"
11777 " %fld6_1 = OpCompositeConstruct %v2f16 %c_f16_42 %c_f16_43\n"
11778 " %fld6_2 = OpCompositeConstruct %v2f16 %c_f16_44 %c_f16_45\n"
11779 " %fld6_3 = OpCompositeConstruct %v2f16 %c_f16_46 %c_f16_47\n"
11780 " %fld6_4 = OpCompositeConstruct %v2f16 %c_f16_48 %c_f16_49\n"
11781 " %fld6 = OpCompositeConstruct %v2f16arr5 %fld6_0 %fld6_1 %fld6_2 %fld6_3 %fld6_4\n"
11783 " %fndx = OpConvertSToF %f16 %ndx\n"
11784 " %fld8_2a0 = OpFMul %f16 %fndx %c_f16_mod\n"
11785 " %fld8_3b1 = OpFAdd %f16 %fndx %c_f16_mod\n"
11787 " %fld8_2a = OpCompositeConstruct %v2f16 %fld8_2a0 %c_f16_61\n"
11788 " %fld8_3b = OpCompositeConstruct %v2f16 %c_f16_65 %fld8_3b1\n"
11789 " %fld8_0 = OpCompositeConstruct %v3f16 %c_f16_52 %c_f16_53 %c_f16_54\n"
11790 " %fld8_1 = OpCompositeConstruct %v3f16 %c_f16_56 %c_f16_57 %c_f16_58\n"
11791 " %fld8_2 = OpCompositeConstruct %v3f16 %fld8_2a %c_f16_62\n"
11792 " %fld8_3 = OpCompositeConstruct %v3f16 %c_f16_64 %fld8_3b\n"
11793 " %fld8_4 = OpCompositeConstruct %v3f16 %c_f16_68 %c_f16_69 %c_f16_70\n"
11794 " %fld8 = OpCompositeConstruct %v3f16arr5 %fld8_0 %fld8_1 %fld8_2 %fld8_3 %fld8_4\n"
11796 " %fld9_0 = OpCompositeConstruct %v4f16 %c_f16_72 %c_f16_73 %c_f16_74 %c_f16_75\n"
11797 " %fld9_1 = OpCompositeConstruct %v4f16 %c_f16_76 %c_f16_77 %c_f16_78 %c_f16_79\n"
11798 " %fld9_2 = OpCompositeConstruct %v4f16 %c_f16_80 %c_f16_81 %c_f16_82 %c_f16_83\n"
11799 " %fld9 = OpCompositeConstruct %v4f16arr3 %fld9_0 %fld9_1 %fld9_2\n"
11801 " %st_val = OpCompositeConstruct %st_test %c_f16_0 %fld1 %fld2 %fld3 %fld4 %fld5 %fld6 %c_f16_50 %fld8 %fld9\n"
11802 " %dst = OpAccessChain %up_st %ssbo_dst %c_i32_0 %ndx\n"
11803 " OpStore %dst %st_val\n"
11805 " OpBranch %next\n"
11807 " %next = OpLabel\n"
11808 " %i_cur = OpLoad %i32 %i\n"
11809 " %i_new = OpIAdd %i32 %i_cur %c_i32_1\n"
11810 " OpStore %i %i_new\n"
11811 " OpBranch %loop\n"
11813 " %merge = OpLabel\n"
11814 " OpBranch %end_if\n"
11815 " %end_if = OpLabel\n"
11816 " OpReturnValue %param\n"
11821 SpecResource specResource;
11822 map<string, string> specs;
11823 VulkanFeatures features;
11824 map<string, string> fragments;
11825 vector<string> extensions;
11826 vector<deFloat16> expectedOutput;
11829 for (deUint32 elementNdx = 0; elementNdx < numElements; ++elementNdx)
11831 vector<deFloat16> expectedIterationOutput;
11833 for (deUint32 structItemNdx = 0; structItemNdx < structItemsCount; ++structItemNdx)
11834 expectedIterationOutput.push_back(tcu::Float16(float(structItemNdx)).bits());
11836 for (deUint32 structItemNdx = 0; structItemNdx < DE_LENGTH_OF_ARRAY(exceptionIndices); ++structItemNdx)
11837 expectedIterationOutput[exceptionIndices[structItemNdx]] = exceptionValue;
11839 expectedIterationOutput[fieldModifiedMulIndex] = tcu::Float16(float(elementNdx * fieldModifier)).bits();
11840 expectedIterationOutput[fieldModifiedAddIndex] = tcu::Float16(float(elementNdx + fieldModifier)).bits();
11842 expectedOutput.insert(expectedOutput.end(), expectedIterationOutput.begin(), expectedIterationOutput.end());
11845 for (deUint32 i = 0; i < structItemsCount; ++i)
11846 consts += " %c_f16_" + de::toString(i) + " = OpConstant %f16 " + de::toString(i) + "\n";
11848 specs["num_elements"] = de::toString(numElements);
11849 specs["struct_item_size"] = de::toString(structItemsCount * sizeof(deFloat16));
11850 specs["field_modifier"] = de::toString(fieldModifier);
11851 specs["consts"] = consts;
11853 fragments["extension"] = "OpExtension \"SPV_KHR_16bit_storage\"";
11854 fragments["capability"] = "OpCapability StorageUniformBufferBlock16\n";
11855 fragments["decoration"] = decoration.specialize(specs);
11856 fragments["pre_main"] = preMain.specialize(specs);
11857 fragments["testfun"] = testFun.specialize(specs);
11859 specResource.inputs.push_back(Resource(BufferSp(new Float16Buffer(expectedOutput)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
11860 specResource.outputs.push_back(Resource(BufferSp(new Float16Buffer(expectedOutput)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
11861 specResource.verifyIO = compareFP16CompositeFunc;
11863 extensions.push_back("VK_KHR_16bit_storage");
11864 extensions.push_back("VK_KHR_shader_float16_int8");
11866 features.extFloat16Int8 = EXTFLOAT16INT8FEATURES_FLOAT16;
11867 features.ext16BitStorage = EXT16BITSTORAGEFEATURES_UNIFORM_BUFFER_BLOCK;
11869 finalizeTestsCreation(specResource, fragments, testCtx, *testGroup.get(), testName, features, extensions, IVec3(1, 1, 1));
11872 return testGroup.release();
11875 template<class SpecResource>
11876 tcu::TestCaseGroup* createFloat16CompositeInsertExtractSet (tcu::TestContext& testCtx, const char* op)
11878 de::MovePtr<tcu::TestCaseGroup> testGroup (new tcu::TestCaseGroup(testCtx, de::toLower(op).c_str(), op));
11879 const deFloat16 exceptionValue = tcu::Float16(-1.0).bits();
11880 const string opName (op);
11881 const deUint32 opIndex = (opName == "OpCompositeInsert") ? 0
11882 : (opName == "OpCompositeExtract") ? 1
11885 const StringTemplate preMain
11887 " %c_i32_ndp = OpConstant %i32 ${num_elements}\n"
11888 " %f16 = OpTypeFloat 16\n"
11889 " %v2f16 = OpTypeVector %f16 2\n"
11890 " %v3f16 = OpTypeVector %f16 3\n"
11891 " %v4f16 = OpTypeVector %f16 4\n"
11892 " %c_f16_na = OpConstant %f16 -1.0\n"
11893 " %c_u32_5 = OpConstant %u32 5\n"
11895 "%f16arr3 = OpTypeArray %f16 %c_u32_3\n"
11896 "%v2f16arr3 = OpTypeArray %v2f16 %c_u32_3\n"
11897 "%v2f16arr5 = OpTypeArray %v2f16 %c_u32_5\n"
11898 "%v3f16arr5 = OpTypeArray %v3f16 %c_u32_5\n"
11899 "%v4f16arr3 = OpTypeArray %v4f16 %c_u32_3\n"
11900 "%struct16 = OpTypeStruct %f16 %v2f16arr3\n"
11901 "%struct16arr3 = OpTypeArray %struct16 %c_u32_3\n"
11902 "%st_test = OpTypeStruct %${field_type}\n"
11904 " %up_f16 = OpTypePointer Uniform %f16\n"
11905 " %up_st = OpTypePointer Uniform %st_test\n"
11906 " %ra_f16 = OpTypeArray %f16 %c_i32_ndp\n"
11907 " %ra_st = OpTypeArray %st_test %c_i32_1\n"
11909 "${op_premain_decls}"
11911 " %up_SSBO_src = OpTypePointer Uniform %SSBO_src\n"
11912 " %up_SSBO_dst = OpTypePointer Uniform %SSBO_dst\n"
11914 " %ssbo_src = OpVariable %up_SSBO_src Uniform\n"
11915 " %ssbo_dst = OpVariable %up_SSBO_dst Uniform\n"
11918 const StringTemplate decoration
11920 "OpDecorate %SSBO_src BufferBlock\n"
11921 "OpDecorate %SSBO_dst BufferBlock\n"
11922 "OpDecorate %ra_f16 ArrayStride 2\n"
11923 "OpDecorate %ra_st ArrayStride ${struct_item_size}\n"
11924 "OpDecorate %ssbo_src DescriptorSet 0\n"
11925 "OpDecorate %ssbo_src Binding 0\n"
11926 "OpDecorate %ssbo_dst DescriptorSet 0\n"
11927 "OpDecorate %ssbo_dst Binding 1\n"
11929 "OpMemberDecorate %SSBO_src 0 Offset 0\n"
11930 "OpMemberDecorate %SSBO_dst 0 Offset 0\n"
11932 "OpDecorate %v2f16arr3 ArrayStride 4\n"
11933 "OpMemberDecorate %struct16 0 Offset 0\n"
11934 "OpMemberDecorate %struct16 1 Offset 4\n"
11935 "OpDecorate %struct16arr3 ArrayStride 16\n"
11936 "OpDecorate %f16arr3 ArrayStride 2\n"
11937 "OpDecorate %v2f16arr5 ArrayStride 4\n"
11938 "OpDecorate %v3f16arr5 ArrayStride 8\n"
11939 "OpDecorate %v4f16arr3 ArrayStride 8\n"
11941 "OpMemberDecorate %st_test 0 Offset 0\n"
11944 const StringTemplate testFun
11946 " %test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
11947 " %param = OpFunctionParameter %v4f32\n"
11948 " %entry = OpLabel\n"
11950 " %i = OpVariable %fp_i32 Function\n"
11951 " OpStore %i %c_i32_0\n"
11953 " %will_run = OpFunctionCall %bool %isUniqueIdZero\n"
11954 " OpSelectionMerge %end_if None\n"
11955 " OpBranchConditional %will_run %run_test %end_if\n"
11957 " %run_test = OpLabel\n"
11958 " OpBranch %loop\n"
11960 " %loop = OpLabel\n"
11961 " %i_cmp = OpLoad %i32 %i\n"
11962 " %lt = OpSLessThan %bool %i_cmp %c_i32_ndp\n"
11963 " OpLoopMerge %merge %next None\n"
11964 " OpBranchConditional %lt %write %merge\n"
11966 " %write = OpLabel\n"
11967 " %ndx = OpLoad %i32 %i\n"
11969 "${op_sw_fun_call}"
11971 " OpStore %dst %val_dst\n"
11972 " OpBranch %next\n"
11974 " %next = OpLabel\n"
11975 " %i_cur = OpLoad %i32 %i\n"
11976 " %i_new = OpIAdd %i32 %i_cur %c_i32_1\n"
11977 " OpStore %i %i_new\n"
11978 " OpBranch %loop\n"
11980 " %merge = OpLabel\n"
11981 " OpBranch %end_if\n"
11982 " %end_if = OpLabel\n"
11983 " OpReturnValue %param\n"
11986 "${op_sw_fun_header}"
11987 " %sw_param = OpFunctionParameter %st_test\n"
11988 "%sw_paramn = OpFunctionParameter %i32\n"
11989 " %sw_entry = OpLabel\n"
11990 " OpSelectionMerge %switch_e None\n"
11991 " OpSwitch %sw_paramn %default ${case_list}\n"
11995 "%default = OpLabel\n"
11996 " OpReturnValue ${op_case_default_value}\n"
11997 "%switch_e = OpLabel\n"
11998 " OpUnreachable\n" // Unreachable merge block for switch statement
12002 const StringTemplate testCaseBody
12004 "%case_${case_ndx} = OpLabel\n"
12005 "%val_ret_${case_ndx} = ${op_name} ${op_args_part} ${access_path}\n"
12006 " OpReturnValue %val_ret_${case_ndx}\n"
12011 const char* premainDecls;
12012 const char* swFunCall;
12013 const char* swFunHeader;
12014 const char* caseDefaultValue;
12015 const char* argsPartial;
12018 OpParts opPartsArray[] =
12020 // OpCompositeInsert
12022 " %fun_t = OpTypeFunction %st_test %f16 %st_test %i32\n"
12023 " %SSBO_src = OpTypeStruct %ra_f16\n"
12024 " %SSBO_dst = OpTypeStruct %ra_st\n",
12026 " %src = OpAccessChain %up_f16 %ssbo_src %c_i32_0 %ndx\n"
12027 " %dst = OpAccessChain %up_st %ssbo_dst %c_i32_0 %c_i32_0\n"
12028 " %val_new = OpLoad %f16 %src\n"
12029 " %val_old = OpLoad %st_test %dst\n"
12030 " %val_dst = OpFunctionCall %st_test %sw_fun %val_new %val_old %ndx\n",
12032 " %sw_fun = OpFunction %st_test None %fun_t\n"
12033 "%sw_paramv = OpFunctionParameter %f16\n",
12037 "%st_test %sw_paramv %sw_param",
12039 // OpCompositeExtract
12041 " %fun_t = OpTypeFunction %f16 %st_test %i32\n"
12042 " %SSBO_src = OpTypeStruct %ra_st\n"
12043 " %SSBO_dst = OpTypeStruct %ra_f16\n",
12045 " %src = OpAccessChain %up_st %ssbo_src %c_i32_0 %c_i32_0\n"
12046 " %dst = OpAccessChain %up_f16 %ssbo_dst %c_i32_0 %ndx\n"
12047 " %val_src = OpLoad %st_test %src\n"
12048 " %val_dst = OpFunctionCall %f16 %sw_fun %val_src %ndx\n",
12050 " %sw_fun = OpFunction %f16 None %fun_t\n",
12058 DE_ASSERT(opIndex >= 0 && opIndex < DE_LENGTH_OF_ARRAY(opPartsArray));
12060 const char* accessPathF16[] =
12065 const char* accessPathV2F16[] =
12070 const char* accessPathV3F16[] =
12077 const char* accessPathV4F16[] =
12084 const char* accessPathF16Arr3[] =
12091 const char* accessPathStruct16Arr3[] =
12093 "0 0 0", // %struct16arr3
12118 const char* accessPathV2F16Arr5[] =
12120 "0 0 0", // %v2f16arr5
12131 const char* accessPathV3F16Arr5[] =
12133 "0 0 0", // %v3f16arr5
12154 const char* accessPathV4F16Arr3[] =
12156 "0 0 0", // %v4f16arr3
12174 struct TypeTestParameters
12177 size_t accessPathLength;
12178 const char** accessPath;
12181 const TypeTestParameters typeTestParameters[] =
12183 { "f16", DE_LENGTH_OF_ARRAY(accessPathF16), accessPathF16 },
12184 { "v2f16", DE_LENGTH_OF_ARRAY(accessPathV2F16), accessPathV2F16 },
12185 { "v3f16", DE_LENGTH_OF_ARRAY(accessPathV3F16), accessPathV3F16 },
12186 { "v4f16", DE_LENGTH_OF_ARRAY(accessPathV4F16), accessPathV4F16 },
12187 { "f16arr3", DE_LENGTH_OF_ARRAY(accessPathF16Arr3), accessPathF16Arr3 },
12188 { "v2f16arr5", DE_LENGTH_OF_ARRAY(accessPathV2F16Arr5), accessPathV2F16Arr5 },
12189 { "v3f16arr5", DE_LENGTH_OF_ARRAY(accessPathV3F16Arr5), accessPathV3F16Arr5 },
12190 { "v4f16arr3", DE_LENGTH_OF_ARRAY(accessPathV4F16Arr3), accessPathV4F16Arr3 },
12191 { "struct16arr3", DE_LENGTH_OF_ARRAY(accessPathStruct16Arr3), accessPathStruct16Arr3 },
12194 for (size_t typeTestNdx = 0; typeTestNdx < DE_LENGTH_OF_ARRAY(typeTestParameters); ++typeTestNdx)
12196 const OpParts opParts = opPartsArray[opIndex];
12197 const string testName = typeTestParameters[typeTestNdx].name;
12198 const size_t structItemsCount = typeTestParameters[typeTestNdx].accessPathLength;
12199 const char** accessPath = typeTestParameters[typeTestNdx].accessPath;
12200 SpecResource specResource;
12201 map<string, string> specs;
12202 VulkanFeatures features;
12203 map<string, string> fragments;
12204 vector<string> extensions;
12205 vector<deFloat16> inputFP16;
12206 vector<deFloat16> dummyFP16Output;
12208 // Generate values for input
12209 inputFP16.reserve(structItemsCount);
12210 for (deUint32 structItemNdx = 0; structItemNdx < structItemsCount; ++structItemNdx)
12211 inputFP16.push_back((accessPath[structItemNdx] == DE_NULL) ? exceptionValue : tcu::Float16(float(structItemNdx)).bits());
12213 dummyFP16Output.resize(structItemsCount);
12215 // Generate cases for OpSwitch
12220 for (deUint32 caseNdx = 0; caseNdx < structItemsCount; ++caseNdx)
12221 if (accessPath[caseNdx] != DE_NULL)
12223 map<string, string> specCase;
12225 specCase["case_ndx"] = de::toString(caseNdx);
12226 specCase["access_path"] = accessPath[caseNdx];
12227 specCase["op_args_part"] = opParts.argsPartial;
12228 specCase["op_name"] = opName;
12230 caseBodies += testCaseBody.specialize(specCase);
12231 caseList += de::toString(caseNdx) + " %case_" + de::toString(caseNdx) + " ";
12234 specs["case_bodies"] = caseBodies;
12235 specs["case_list"] = caseList;
12238 specs["num_elements"] = de::toString(structItemsCount);
12239 specs["field_type"] = typeTestParameters[typeTestNdx].name;
12240 specs["struct_item_size"] = de::toString(structItemsCount * sizeof(deFloat16));
12241 specs["op_premain_decls"] = opParts.premainDecls;
12242 specs["op_sw_fun_call"] = opParts.swFunCall;
12243 specs["op_sw_fun_header"] = opParts.swFunHeader;
12244 specs["op_case_default_value"] = opParts.caseDefaultValue;
12246 fragments["extension"] = "OpExtension \"SPV_KHR_16bit_storage\"";
12247 fragments["capability"] = "OpCapability StorageUniformBufferBlock16\n";
12248 fragments["decoration"] = decoration.specialize(specs);
12249 fragments["pre_main"] = preMain.specialize(specs);
12250 fragments["testfun"] = testFun.specialize(specs);
12252 specResource.inputs.push_back(Resource(BufferSp(new Float16Buffer(inputFP16)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
12253 specResource.outputs.push_back(Resource(BufferSp(new Float16Buffer(dummyFP16Output)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
12254 specResource.verifyIO = compareFP16CompositeFunc;
12256 extensions.push_back("VK_KHR_16bit_storage");
12257 extensions.push_back("VK_KHR_shader_float16_int8");
12259 features.extFloat16Int8 = EXTFLOAT16INT8FEATURES_FLOAT16;
12260 features.ext16BitStorage = EXT16BITSTORAGEFEATURES_UNIFORM_BUFFER_BLOCK;
12262 finalizeTestsCreation(specResource, fragments, testCtx, *testGroup.get(), testName, features, extensions, IVec3(1, 1, 1));
12265 return testGroup.release();
12268 struct fp16PerComponent
12272 , floatFormat16 (-14, 15, 10, true)
12274 , argCompCount(3, 0)
12278 bool callOncePerComponent () { return true; }
12279 deUint32 getComponentValidity () { return static_cast<deUint32>(-1); }
12281 virtual double getULPs (vector<const deFloat16*>&) { return 1.0; }
12282 virtual double getMin (double value, double ulps) { return value - floatFormat16.ulp(deAbs(value), ulps); }
12283 virtual double getMax (double value, double ulps) { return value + floatFormat16.ulp(deAbs(value), ulps); }
12285 virtual size_t getFlavorCount () { return flavorNames.empty() ? 1 : flavorNames.size(); }
12286 virtual void setFlavor (size_t flavorNo) { DE_ASSERT(flavorNo < getFlavorCount()); flavor = flavorNo; }
12287 virtual size_t getFlavor () { return flavor; }
12288 virtual string getCurrentFlavorName () { return flavorNames.empty() ? string("") : flavorNames[getFlavor()]; }
12290 virtual void setOutCompCount (size_t compCount) { outCompCount = compCount; }
12291 virtual size_t getOutCompCount () { return outCompCount; }
12293 virtual void setArgCompCount (size_t argNo, size_t compCount) { argCompCount[argNo] = compCount; }
12294 virtual size_t getArgCompCount (size_t argNo) { return argCompCount[argNo]; }
12298 tcu::FloatFormat floatFormat16;
12299 size_t outCompCount;
12300 vector<size_t> argCompCount;
12301 vector<string> flavorNames;
12304 struct fp16OpFNegate : public fp16PerComponent
12306 template <class fp16type>
12307 bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
12309 const fp16type x (*in[0]);
12310 const double d (x.asDouble());
12311 const double result (0.0 - d);
12313 out[0] = fp16type(result).bits();
12314 min[0] = getMin(result, getULPs(in));
12315 max[0] = getMax(result, getULPs(in));
12321 struct fp16Round : public fp16PerComponent
12323 fp16Round() : fp16PerComponent()
12325 flavorNames.push_back("Floor(x+0.5)");
12326 flavorNames.push_back("Floor(x-0.5)");
12327 flavorNames.push_back("RoundEven");
12330 template<class fp16type>
12331 bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
12333 const fp16type x (*in[0]);
12334 const double d (x.asDouble());
12335 double result (0.0);
12339 case 0: result = deRound(d); break;
12340 case 1: result = deFloor(d - 0.5); break;
12341 case 2: result = deRoundEven(d); break;
12342 default: TCU_THROW(InternalError, "Invalid flavor specified");
12345 out[0] = fp16type(result).bits();
12346 min[0] = getMin(result, getULPs(in));
12347 max[0] = getMax(result, getULPs(in));
12353 struct fp16RoundEven : public fp16PerComponent
12355 template<class fp16type>
12356 bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
12358 const fp16type x (*in[0]);
12359 const double d (x.asDouble());
12360 const double result (deRoundEven(d));
12362 out[0] = fp16type(result).bits();
12363 min[0] = getMin(result, getULPs(in));
12364 max[0] = getMax(result, getULPs(in));
12370 struct fp16Trunc : public fp16PerComponent
12372 template<class fp16type>
12373 bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
12375 const fp16type x (*in[0]);
12376 const double d (x.asDouble());
12377 const double result (deTrunc(d));
12379 out[0] = fp16type(result).bits();
12380 min[0] = getMin(result, getULPs(in));
12381 max[0] = getMax(result, getULPs(in));
12387 struct fp16FAbs : public fp16PerComponent
12389 template<class fp16type>
12390 bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
12392 const fp16type x (*in[0]);
12393 const double d (x.asDouble());
12394 const double result (deAbs(d));
12396 out[0] = fp16type(result).bits();
12397 min[0] = getMin(result, getULPs(in));
12398 max[0] = getMax(result, getULPs(in));
12404 struct fp16FSign : public fp16PerComponent
12406 template<class fp16type>
12407 bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
12409 const fp16type x (*in[0]);
12410 const double d (x.asDouble());
12411 const double result (deSign(d));
12416 out[0] = fp16type(result).bits();
12417 min[0] = getMin(result, getULPs(in));
12418 max[0] = getMax(result, getULPs(in));
12424 struct fp16Floor : public fp16PerComponent
12426 template<class fp16type>
12427 bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
12429 const fp16type x (*in[0]);
12430 const double d (x.asDouble());
12431 const double result (deFloor(d));
12433 out[0] = fp16type(result).bits();
12434 min[0] = getMin(result, getULPs(in));
12435 max[0] = getMax(result, getULPs(in));
12441 struct fp16Ceil : public fp16PerComponent
12443 template<class fp16type>
12444 bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
12446 const fp16type x (*in[0]);
12447 const double d (x.asDouble());
12448 const double result (deCeil(d));
12450 out[0] = fp16type(result).bits();
12451 min[0] = getMin(result, getULPs(in));
12452 max[0] = getMax(result, getULPs(in));
12458 struct fp16Fract : public fp16PerComponent
12460 template<class fp16type>
12461 bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
12463 const fp16type x (*in[0]);
12464 const double d (x.asDouble());
12465 const double result (deFrac(d));
12467 out[0] = fp16type(result).bits();
12468 min[0] = getMin(result, getULPs(in));
12469 max[0] = getMax(result, getULPs(in));
12475 struct fp16Radians : public fp16PerComponent
12477 virtual double getULPs (vector<const deFloat16*>& in)
12484 template<class fp16type>
12485 bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
12487 const fp16type x (*in[0]);
12488 const float d (x.asFloat());
12489 const float result (deFloatRadians(d));
12491 out[0] = fp16type(result).bits();
12492 min[0] = getMin(result, getULPs(in));
12493 max[0] = getMax(result, getULPs(in));
12499 struct fp16Degrees : public fp16PerComponent
12501 virtual double getULPs (vector<const deFloat16*>& in)
12508 template<class fp16type>
12509 bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
12511 const fp16type x (*in[0]);
12512 const float d (x.asFloat());
12513 const float result (deFloatDegrees(d));
12515 out[0] = fp16type(result).bits();
12516 min[0] = getMin(result, getULPs(in));
12517 max[0] = getMax(result, getULPs(in));
12523 struct fp16Sin : public fp16PerComponent
12525 template<class fp16type>
12526 bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
12528 const fp16type x (*in[0]);
12529 const double d (x.asDouble());
12530 const double result (deSin(d));
12531 const double unspecUlp (16.0);
12532 const double err (de::inRange(d, -DE_PI_DOUBLE, DE_PI_DOUBLE) ? deLdExp(1.0, -7) : floatFormat16.ulp(deAbs(result), unspecUlp));
12534 if (!de::inRange(d, -DE_PI_DOUBLE, DE_PI_DOUBLE))
12537 out[0] = fp16type(result).bits();
12538 min[0] = result - err;
12539 max[0] = result + err;
12545 struct fp16Cos : public fp16PerComponent
12547 template<class fp16type>
12548 bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
12550 const fp16type x (*in[0]);
12551 const double d (x.asDouble());
12552 const double result (deCos(d));
12553 const double unspecUlp (16.0);
12554 const double err (de::inRange(d, -DE_PI_DOUBLE, DE_PI_DOUBLE) ? deLdExp(1.0, -7) : floatFormat16.ulp(deAbs(result), unspecUlp));
12556 if (!de::inRange(d, -DE_PI_DOUBLE, DE_PI_DOUBLE))
12559 out[0] = fp16type(result).bits();
12560 min[0] = result - err;
12561 max[0] = result + err;
12567 struct fp16Tan : public fp16PerComponent
12569 template<class fp16type>
12570 bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
12572 const fp16type x (*in[0]);
12573 const double d (x.asDouble());
12574 const double result (deTan(d));
12576 if (!de::inRange(d, -DE_PI_DOUBLE, DE_PI_DOUBLE))
12579 out[0] = fp16type(result).bits();
12581 const double err = deLdExp(1.0, -7);
12582 const double s1 = deSin(d) + err;
12583 const double s2 = deSin(d) - err;
12584 const double c1 = deCos(d) + err;
12585 const double c2 = deCos(d) - err;
12586 const double edgeVals[] = {s1/c1, s1/c2, s2/c1, s2/c2};
12587 double edgeLeft = out[0];
12588 double edgeRight = out[0];
12590 if (deSign(c1 * c2) < 0.0)
12592 edgeLeft = -std::numeric_limits<double>::infinity();
12593 edgeRight = +std::numeric_limits<double>::infinity();
12597 edgeLeft = *std::min_element(&edgeVals[0], &edgeVals[DE_LENGTH_OF_ARRAY(edgeVals)]);
12598 edgeRight = *std::max_element(&edgeVals[0], &edgeVals[DE_LENGTH_OF_ARRAY(edgeVals)]);
12602 max[0] = edgeRight;
12609 struct fp16Asin : public fp16PerComponent
12611 template<class fp16type>
12612 bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
12614 const fp16type x (*in[0]);
12615 const double d (x.asDouble());
12616 const double result (deAsin(d));
12617 const double error (deAtan2(d, sqrt(1.0 - d * d)));
12619 if (!x.isNaN() && deAbs(d) > 1.0)
12622 out[0] = fp16type(result).bits();
12623 min[0] = result - floatFormat16.ulp(deAbs(error), 2 * 5.0); // This is not a precision test. Value is not from spec
12624 max[0] = result + floatFormat16.ulp(deAbs(error), 2 * 5.0); // This is not a precision test. Value is not from spec
12630 struct fp16Acos : public fp16PerComponent
12632 template<class fp16type>
12633 bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
12635 const fp16type x (*in[0]);
12636 const double d (x.asDouble());
12637 const double result (deAcos(d));
12638 const double error (deAtan2(sqrt(1.0 - d * d), d));
12640 if (!x.isNaN() && deAbs(d) > 1.0)
12643 out[0] = fp16type(result).bits();
12644 min[0] = result - floatFormat16.ulp(deAbs(error), 2 * 5.0); // This is not a precision test. Value is not from spec
12645 max[0] = result + floatFormat16.ulp(deAbs(error), 2 * 5.0); // This is not a precision test. Value is not from spec
12651 struct fp16Atan : public fp16PerComponent
12653 virtual double getULPs(vector<const deFloat16*>& in)
12657 return 2 * 5.0; // This is not a precision test. Value is not from spec
12660 template<class fp16type>
12661 bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
12663 const fp16type x (*in[0]);
12664 const double d (x.asDouble());
12665 const double result (deAtanOver(d));
12667 out[0] = fp16type(result).bits();
12668 min[0] = getMin(result, getULPs(in));
12669 max[0] = getMax(result, getULPs(in));
12675 struct fp16Sinh : public fp16PerComponent
12677 fp16Sinh() : fp16PerComponent()
12679 flavorNames.push_back("Double");
12680 flavorNames.push_back("ExpFP16");
12683 template<class fp16type>
12684 bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
12686 const fp16type x (*in[0]);
12687 const double d (x.asDouble());
12688 const double ulps (64 * (1.0 + 2 * deAbs(d))); // This is not a precision test. Value is not from spec
12689 double result (0.0);
12690 double error (0.0);
12692 if (getFlavor() == 0)
12694 result = deSinh(d);
12695 error = floatFormat16.ulp(deAbs(result), ulps);
12697 else if (getFlavor() == 1)
12699 const fp16type epx (deExp(d));
12700 const fp16type enx (deExp(-d));
12701 const fp16type esx (epx.asDouble() - enx.asDouble());
12702 const fp16type sx2 (esx.asDouble() / 2.0);
12704 result = sx2.asDouble();
12705 error = deAbs(floatFormat16.ulp(epx.asDouble(), ulps)) + deAbs(floatFormat16.ulp(enx.asDouble(), ulps));
12709 TCU_THROW(InternalError, "Unknown flavor");
12712 out[0] = fp16type(result).bits();
12713 min[0] = result - error;
12714 max[0] = result + error;
12720 struct fp16Cosh : public fp16PerComponent
12722 fp16Cosh() : fp16PerComponent()
12724 flavorNames.push_back("Double");
12725 flavorNames.push_back("ExpFP16");
12728 template<class fp16type>
12729 bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
12731 const fp16type x (*in[0]);
12732 const double d (x.asDouble());
12733 const double ulps (64 * (1.0 + 2 * deAbs(d))); // This is not a precision test. Value is not from spec
12734 double result (0.0);
12736 if (getFlavor() == 0)
12738 result = deCosh(d);
12740 else if (getFlavor() == 1)
12742 const fp16type epx (deExp(d));
12743 const fp16type enx (deExp(-d));
12744 const fp16type esx (epx.asDouble() + enx.asDouble());
12745 const fp16type sx2 (esx.asDouble() / 2.0);
12747 result = sx2.asDouble();
12751 TCU_THROW(InternalError, "Unknown flavor");
12754 out[0] = fp16type(result).bits();
12755 min[0] = result - floatFormat16.ulp(deAbs(result), ulps);
12756 max[0] = result + floatFormat16.ulp(deAbs(result), ulps);
12762 struct fp16Tanh : public fp16PerComponent
12764 fp16Tanh() : fp16PerComponent()
12766 flavorNames.push_back("Tanh");
12767 flavorNames.push_back("SinhCosh");
12768 flavorNames.push_back("SinhCoshFP16");
12769 flavorNames.push_back("PolyFP16");
12772 virtual double getULPs (vector<const deFloat16*>& in)
12774 const tcu::Float16 x (*in[0]);
12775 const double d (x.asDouble());
12777 return 2 * (1.0 + 2 * deAbs(d)); // This is not a precision test. Value is not from spec
12780 template<class fp16type>
12781 inline double calcPoly (const fp16type& espx, const fp16type& esnx, const fp16type& ecpx, const fp16type& ecnx)
12783 const fp16type esx (espx.asDouble() - esnx.asDouble());
12784 const fp16type sx2 (esx.asDouble() / 2.0);
12785 const fp16type ecx (ecpx.asDouble() + ecnx.asDouble());
12786 const fp16type cx2 (ecx.asDouble() / 2.0);
12787 const fp16type tg (sx2.asDouble() / cx2.asDouble());
12788 const double rez (tg.asDouble());
12793 template<class fp16type>
12794 bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
12796 const fp16type x (*in[0]);
12797 const double d (x.asDouble());
12798 double result (0.0);
12800 if (getFlavor() == 0)
12802 result = deTanh(d);
12803 min[0] = getMin(result, getULPs(in));
12804 max[0] = getMax(result, getULPs(in));
12806 else if (getFlavor() == 1)
12808 result = deSinh(d) / deCosh(d);
12809 min[0] = getMin(result, getULPs(in));
12810 max[0] = getMax(result, getULPs(in));
12812 else if (getFlavor() == 2)
12814 const fp16type s (deSinh(d));
12815 const fp16type c (deCosh(d));
12817 result = s.asDouble() / c.asDouble();
12818 min[0] = getMin(result, getULPs(in));
12819 max[0] = getMax(result, getULPs(in));
12821 else if (getFlavor() == 3)
12823 const double ulps (getULPs(in));
12824 const double epxm (deExp( d));
12825 const double enxm (deExp(-d));
12826 const double epxmerr = floatFormat16.ulp(epxm, ulps);
12827 const double enxmerr = floatFormat16.ulp(enxm, ulps);
12828 const fp16type epx[] = { fp16type(epxm - epxmerr), fp16type(epxm + epxmerr) };
12829 const fp16type enx[] = { fp16type(enxm - enxmerr), fp16type(enxm + enxmerr) };
12830 const fp16type epxm16 (epxm);
12831 const fp16type enxm16 (enxm);
12832 vector<double> tgs;
12834 for (size_t spNdx = 0; spNdx < DE_LENGTH_OF_ARRAY(epx); ++spNdx)
12835 for (size_t snNdx = 0; snNdx < DE_LENGTH_OF_ARRAY(enx); ++snNdx)
12836 for (size_t cpNdx = 0; cpNdx < DE_LENGTH_OF_ARRAY(epx); ++cpNdx)
12837 for (size_t cnNdx = 0; cnNdx < DE_LENGTH_OF_ARRAY(enx); ++cnNdx)
12839 const double tgh = calcPoly(epx[spNdx], enx[snNdx], epx[cpNdx], enx[cnNdx]);
12841 tgs.push_back(tgh);
12844 result = calcPoly(epxm16, enxm16, epxm16, enxm16);
12845 min[0] = *std::min_element(tgs.begin(), tgs.end());
12846 max[0] = *std::max_element(tgs.begin(), tgs.end());
12850 TCU_THROW(InternalError, "Unknown flavor");
12853 out[0] = fp16type(result).bits();
12859 struct fp16Asinh : public fp16PerComponent
12861 fp16Asinh() : fp16PerComponent()
12863 flavorNames.push_back("Double");
12864 flavorNames.push_back("PolyFP16Wiki");
12865 flavorNames.push_back("PolyFP16Abs");
12868 virtual double getULPs (vector<const deFloat16*>& in)
12872 return 256.0; // This is not a precision test. Value is not from spec
12875 template<class fp16type>
12876 bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
12878 const fp16type x (*in[0]);
12879 const double d (x.asDouble());
12880 double result (0.0);
12882 if (getFlavor() == 0)
12884 result = deAsinh(d);
12886 else if (getFlavor() == 1)
12888 const fp16type x2 (d * d);
12889 const fp16type x2p1 (x2.asDouble() + 1.0);
12890 const fp16type sq (deSqrt(x2p1.asDouble()));
12891 const fp16type sxsq (d + sq.asDouble());
12892 const fp16type lsxsq (deLog(sxsq.asDouble()));
12897 result = lsxsq.asDouble();
12899 else if (getFlavor() == 2)
12901 const fp16type x2 (d * d);
12902 const fp16type x2p1 (x2.asDouble() + 1.0);
12903 const fp16type sq (deSqrt(x2p1.asDouble()));
12904 const fp16type sxsq (deAbs(d) + sq.asDouble());
12905 const fp16type lsxsq (deLog(sxsq.asDouble()));
12907 result = deSign(d) * lsxsq.asDouble();
12911 TCU_THROW(InternalError, "Unknown flavor");
12914 out[0] = fp16type(result).bits();
12915 min[0] = getMin(result, getULPs(in));
12916 max[0] = getMax(result, getULPs(in));
12922 struct fp16Acosh : public fp16PerComponent
12924 fp16Acosh() : fp16PerComponent()
12926 flavorNames.push_back("Double");
12927 flavorNames.push_back("PolyFP16");
12930 virtual double getULPs (vector<const deFloat16*>& in)
12934 return 16.0; // This is not a precision test. Value is not from spec
12937 template<class fp16type>
12938 bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
12940 const fp16type x (*in[0]);
12941 const double d (x.asDouble());
12942 double result (0.0);
12944 if (!x.isNaN() && d < 1.0)
12947 if (getFlavor() == 0)
12949 result = deAcosh(d);
12951 else if (getFlavor() == 1)
12953 const fp16type x2 (d * d);
12954 const fp16type x2m1 (x2.asDouble() - 1.0);
12955 const fp16type sq (deSqrt(x2m1.asDouble()));
12956 const fp16type sxsq (d + sq.asDouble());
12957 const fp16type lsxsq (deLog(sxsq.asDouble()));
12959 result = lsxsq.asDouble();
12963 TCU_THROW(InternalError, "Unknown flavor");
12966 out[0] = fp16type(result).bits();
12967 min[0] = getMin(result, getULPs(in));
12968 max[0] = getMax(result, getULPs(in));
12974 struct fp16Atanh : public fp16PerComponent
12976 fp16Atanh() : fp16PerComponent()
12978 flavorNames.push_back("Double");
12979 flavorNames.push_back("PolyFP16");
12982 template<class fp16type>
12983 bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
12985 const fp16type x (*in[0]);
12986 const double d (x.asDouble());
12987 double result (0.0);
12989 if (deAbs(d) >= 1.0)
12992 if (getFlavor() == 0)
12994 const double ulps (16.0); // This is not a precision test. Value is not from spec
12996 result = deAtanh(d);
12997 min[0] = getMin(result, ulps);
12998 max[0] = getMax(result, ulps);
13000 else if (getFlavor() == 1)
13002 const fp16type x1a (1.0 + d);
13003 const fp16type x1b (1.0 - d);
13004 const fp16type x1d (x1a.asDouble() / x1b.asDouble());
13005 const fp16type lx1d (deLog(x1d.asDouble()));
13006 const fp16type lx1d2 (0.5 * lx1d.asDouble());
13007 const double error (2 * (de::inRange(deAbs(x1d.asDouble()), 0.5, 2.0) ? deLdExp(2.0, -7) : floatFormat16.ulp(deAbs(x1d.asDouble()), 3.0)));
13009 result = lx1d2.asDouble();
13010 min[0] = result - error;
13011 max[0] = result + error;
13015 TCU_THROW(InternalError, "Unknown flavor");
13018 out[0] = fp16type(result).bits();
13024 struct fp16Exp : public fp16PerComponent
13026 template<class fp16type>
13027 bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
13029 const fp16type x (*in[0]);
13030 const double d (x.asDouble());
13031 const double ulps (10.0 * (1.0 + 2.0 * deAbs(d)));
13032 const double result (deExp(d));
13034 out[0] = fp16type(result).bits();
13035 min[0] = getMin(result, ulps);
13036 max[0] = getMax(result, ulps);
13042 struct fp16Log : public fp16PerComponent
13044 template<class fp16type>
13045 bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
13047 const fp16type x (*in[0]);
13048 const double d (x.asDouble());
13049 const double result (deLog(d));
13050 const double error (de::inRange(deAbs(d), 0.5, 2.0) ? deLdExp(2.0, -7) : floatFormat16.ulp(deAbs(result), 3.0));
13055 out[0] = fp16type(result).bits();
13056 min[0] = result - error;
13057 max[0] = result + error;
13063 struct fp16Exp2 : public fp16PerComponent
13065 template<class fp16type>
13066 bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
13068 const fp16type x (*in[0]);
13069 const double d (x.asDouble());
13070 const double result (deExp2(d));
13071 const double ulps (1.0 + 2.0 * deAbs(fp16type(in[0][0]).asDouble()));
13073 out[0] = fp16type(result).bits();
13074 min[0] = getMin(result, ulps);
13075 max[0] = getMax(result, ulps);
13081 struct fp16Log2 : public fp16PerComponent
13083 template<class fp16type>
13084 bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
13086 const fp16type x (*in[0]);
13087 const double d (x.asDouble());
13088 const double result (deLog2(d));
13089 const double error (de::inRange(deAbs(d), 0.5, 2.0) ? deLdExp(2.0, -7) : floatFormat16.ulp(deAbs(result), 3.0));
13094 out[0] = fp16type(result).bits();
13095 min[0] = result - error;
13096 max[0] = result + error;
13102 struct fp16Sqrt : public fp16PerComponent
13104 virtual double getULPs (vector<const deFloat16*>& in)
13111 template<class fp16type>
13112 bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
13114 const fp16type x (*in[0]);
13115 const double d (x.asDouble());
13116 const double result (deSqrt(d));
13118 if (!x.isNaN() && d < 0.0)
13121 out[0] = fp16type(result).bits();
13122 min[0] = getMin(result, getULPs(in));
13123 max[0] = getMax(result, getULPs(in));
13129 struct fp16InverseSqrt : public fp16PerComponent
13131 virtual double getULPs (vector<const deFloat16*>& in)
13138 template<class fp16type>
13139 bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
13141 const fp16type x (*in[0]);
13142 const double d (x.asDouble());
13143 const double result (1.0/deSqrt(d));
13145 if (!x.isNaN() && d <= 0.0)
13148 out[0] = fp16type(result).bits();
13149 min[0] = getMin(result, getULPs(in));
13150 max[0] = getMax(result, getULPs(in));
13156 struct fp16ModfFrac : public fp16PerComponent
13158 template<class fp16type>
13159 bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
13161 const fp16type x (*in[0]);
13162 const double d (x.asDouble());
13164 const double result (deModf(d, &i));
13166 if (x.isInf() || x.isNaN())
13169 out[0] = fp16type(result).bits();
13170 min[0] = getMin(result, getULPs(in));
13171 max[0] = getMax(result, getULPs(in));
13177 struct fp16ModfInt : public fp16PerComponent
13179 template<class fp16type>
13180 bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
13182 const fp16type x (*in[0]);
13183 const double d (x.asDouble());
13185 const double dummy (deModf(d, &i));
13186 const double result (i);
13190 if (x.isInf() || x.isNaN())
13193 out[0] = fp16type(result).bits();
13194 min[0] = getMin(result, getULPs(in));
13195 max[0] = getMax(result, getULPs(in));
13201 struct fp16FrexpS : public fp16PerComponent
13203 template<class fp16type>
13204 bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
13206 const fp16type x (*in[0]);
13207 const double d (x.asDouble());
13209 const double result (deFrExp(d, &e));
13211 if (x.isNaN() || x.isInf())
13214 out[0] = fp16type(result).bits();
13215 min[0] = getMin(result, getULPs(in));
13216 max[0] = getMax(result, getULPs(in));
13222 struct fp16FrexpE : public fp16PerComponent
13224 template<class fp16type>
13225 bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
13227 const fp16type x (*in[0]);
13228 const double d (x.asDouble());
13230 const double dummy (deFrExp(d, &e));
13231 const double result (static_cast<double>(e));
13235 if (x.isNaN() || x.isInf())
13238 out[0] = fp16type(result).bits();
13239 min[0] = getMin(result, getULPs(in));
13240 max[0] = getMax(result, getULPs(in));
13246 struct fp16OpFAdd : public fp16PerComponent
13248 template<class fp16type>
13249 bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
13251 const fp16type x (*in[0]);
13252 const fp16type y (*in[1]);
13253 const double xd (x.asDouble());
13254 const double yd (y.asDouble());
13255 const double result (xd + yd);
13257 out[0] = fp16type(result).bits();
13258 min[0] = getMin(result, getULPs(in));
13259 max[0] = getMax(result, getULPs(in));
13265 struct fp16OpFSub : public fp16PerComponent
13267 template<class fp16type>
13268 bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
13270 const fp16type x (*in[0]);
13271 const fp16type y (*in[1]);
13272 const double xd (x.asDouble());
13273 const double yd (y.asDouble());
13274 const double result (xd - yd);
13276 out[0] = fp16type(result).bits();
13277 min[0] = getMin(result, getULPs(in));
13278 max[0] = getMax(result, getULPs(in));
13284 struct fp16OpFMul : public fp16PerComponent
13286 template<class fp16type>
13287 bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
13289 const fp16type x (*in[0]);
13290 const fp16type y (*in[1]);
13291 const double xd (x.asDouble());
13292 const double yd (y.asDouble());
13293 const double result (xd * yd);
13295 out[0] = fp16type(result).bits();
13296 min[0] = getMin(result, getULPs(in));
13297 max[0] = getMax(result, getULPs(in));
13303 struct fp16OpFDiv : public fp16PerComponent
13305 fp16OpFDiv() : fp16PerComponent()
13307 flavorNames.push_back("DirectDiv");
13308 flavorNames.push_back("InverseDiv");
13311 template<class fp16type>
13312 bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
13314 const fp16type x (*in[0]);
13315 const fp16type y (*in[1]);
13316 const double xd (x.asDouble());
13317 const double yd (y.asDouble());
13318 const double unspecUlp (16.0);
13319 const double ulpCnt (de::inRange(deAbs(yd), deLdExp(1, -14), deLdExp(1, 14)) ? 2.5 : unspecUlp);
13320 double result (0.0);
13325 if (getFlavor() == 0)
13327 result = (xd / yd);
13329 else if (getFlavor() == 1)
13331 const double invyd (1.0 / yd);
13332 const fp16type invy (invyd);
13334 result = (xd * invy.asDouble());
13338 TCU_THROW(InternalError, "Unknown flavor");
13341 out[0] = fp16type(result).bits();
13342 min[0] = getMin(result, ulpCnt);
13343 max[0] = getMax(result, ulpCnt);
13349 struct fp16Atan2 : public fp16PerComponent
13351 fp16Atan2() : fp16PerComponent()
13353 flavorNames.push_back("DoubleCalc");
13354 flavorNames.push_back("DoubleCalc_PI");
13357 virtual double getULPs(vector<const deFloat16*>& in)
13361 return 2 * 5.0; // This is not a precision test. Value is not from spec
13364 template<class fp16type>
13365 bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
13367 const fp16type x (*in[0]);
13368 const fp16type y (*in[1]);
13369 const double xd (x.asDouble());
13370 const double yd (y.asDouble());
13371 double result (0.0);
13373 if (x.isZero() && y.isZero())
13376 if (getFlavor() == 0)
13378 result = deAtan2(xd, yd);
13380 else if (getFlavor() == 1)
13382 const double ulps (2.0 * 5.0); // This is not a precision test. Value is not from spec
13383 const double eps (floatFormat16.ulp(DE_PI_DOUBLE, ulps));
13385 result = deAtan2(xd, yd);
13387 if (de::inRange(deAbs(result), DE_PI_DOUBLE - eps, DE_PI_DOUBLE + eps))
13392 TCU_THROW(InternalError, "Unknown flavor");
13395 out[0] = fp16type(result).bits();
13396 min[0] = getMin(result, getULPs(in));
13397 max[0] = getMax(result, getULPs(in));
13403 struct fp16Pow : public fp16PerComponent
13405 fp16Pow() : fp16PerComponent()
13407 flavorNames.push_back("Pow");
13408 flavorNames.push_back("PowLog2");
13409 flavorNames.push_back("PowLog2FP16");
13412 template<class fp16type>
13413 bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
13415 const fp16type x (*in[0]);
13416 const fp16type y (*in[1]);
13417 const double xd (x.asDouble());
13418 const double yd (y.asDouble());
13419 const double logxeps (de::inRange(deAbs(xd), 0.5, 2.0) ? deLdExp(1.0, -7) : floatFormat16.ulp(deLog2(xd), 3.0));
13420 const double ulps1 (1.0 + 4.0 * deAbs(yd * (deLog2(xd) - logxeps)));
13421 const double ulps2 (1.0 + 4.0 * deAbs(yd * (deLog2(xd) + logxeps)));
13422 const double ulps (deMax(deAbs(ulps1), deAbs(ulps2)));
13423 double result (0.0);
13428 if (x.isZero() && yd <= 0.0)
13431 if (getFlavor() == 0)
13433 result = dePow(xd, yd);
13435 else if (getFlavor() == 1)
13437 const double l2d (deLog2(xd));
13438 const double e2d (deExp2(yd * l2d));
13442 else if (getFlavor() == 2)
13444 const double l2d (deLog2(xd));
13445 const fp16type l2 (l2d);
13446 const double e2d (deExp2(yd * l2.asDouble()));
13447 const fp16type e2 (e2d);
13449 result = e2.asDouble();
13453 TCU_THROW(InternalError, "Unknown flavor");
13456 out[0] = fp16type(result).bits();
13457 min[0] = getMin(result, ulps);
13458 max[0] = getMax(result, ulps);
13464 struct fp16FMin : public fp16PerComponent
13466 template<class fp16type>
13467 bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
13469 const fp16type x (*in[0]);
13470 const fp16type y (*in[1]);
13471 const double xd (x.asDouble());
13472 const double yd (y.asDouble());
13473 const double result (deMin(xd, yd));
13475 if (x.isNaN() || y.isNaN())
13478 out[0] = fp16type(result).bits();
13479 min[0] = getMin(result, getULPs(in));
13480 max[0] = getMax(result, getULPs(in));
13486 struct fp16FMax : public fp16PerComponent
13488 template<class fp16type>
13489 bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
13491 const fp16type x (*in[0]);
13492 const fp16type y (*in[1]);
13493 const double xd (x.asDouble());
13494 const double yd (y.asDouble());
13495 const double result (deMax(xd, yd));
13497 if (x.isNaN() || y.isNaN())
13500 out[0] = fp16type(result).bits();
13501 min[0] = getMin(result, getULPs(in));
13502 max[0] = getMax(result, getULPs(in));
13508 struct fp16Step : public fp16PerComponent
13510 template<class fp16type>
13511 bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
13513 const fp16type edge (*in[0]);
13514 const fp16type x (*in[1]);
13515 const double edged (edge.asDouble());
13516 const double xd (x.asDouble());
13517 const double result (deStep(edged, xd));
13519 out[0] = fp16type(result).bits();
13520 min[0] = getMin(result, getULPs(in));
13521 max[0] = getMax(result, getULPs(in));
13527 struct fp16Ldexp : public fp16PerComponent
13529 template<class fp16type>
13530 bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
13532 const fp16type x (*in[0]);
13533 const fp16type y (*in[1]);
13534 const double xd (x.asDouble());
13535 const int yd (static_cast<int>(deTrunc(y.asDouble())));
13536 const double result (deLdExp(xd, yd));
13538 if (y.isNaN() || y.isInf() || y.isDenorm() || yd < -14 || yd > 15)
13541 // Spec: "If this product is too large to be represented in the floating-point type, the result is undefined."
13542 if (fp16type(result).isInf())
13545 out[0] = fp16type(result).bits();
13546 min[0] = getMin(result, getULPs(in));
13547 max[0] = getMax(result, getULPs(in));
13553 struct fp16FClamp : public fp16PerComponent
13555 template<class fp16type>
13556 bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
13558 const fp16type x (*in[0]);
13559 const fp16type minVal (*in[1]);
13560 const fp16type maxVal (*in[2]);
13561 const double xd (x.asDouble());
13562 const double minVald (minVal.asDouble());
13563 const double maxVald (maxVal.asDouble());
13564 const double result (deClamp(xd, minVald, maxVald));
13566 if (minVal.isNaN() || maxVal.isNaN() || minVald > maxVald)
13569 out[0] = fp16type(result).bits();
13570 min[0] = getMin(result, getULPs(in));
13571 max[0] = getMax(result, getULPs(in));
13577 struct fp16FMix : public fp16PerComponent
13579 fp16FMix() : fp16PerComponent()
13581 flavorNames.push_back("DoubleCalc");
13582 flavorNames.push_back("EmulatingFP16");
13583 flavorNames.push_back("EmulatingFP16YminusX");
13586 template<class fp16type>
13587 bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
13589 const fp16type x (*in[0]);
13590 const fp16type y (*in[1]);
13591 const fp16type a (*in[2]);
13592 const double ulps (8.0); // This is not a precision test. Value is not from spec
13593 double result (0.0);
13595 if (getFlavor() == 0)
13597 const double xd (x.asDouble());
13598 const double yd (y.asDouble());
13599 const double ad (a.asDouble());
13600 const double xeps (floatFormat16.ulp(deAbs(xd * (1.0 - ad)), ulps));
13601 const double yeps (floatFormat16.ulp(deAbs(yd * ad), ulps));
13602 const double eps (xeps + yeps);
13604 result = deMix(xd, yd, ad);
13605 min[0] = result - eps;
13606 max[0] = result + eps;
13608 else if (getFlavor() == 1)
13610 const double xd (x.asDouble());
13611 const double yd (y.asDouble());
13612 const double ad (a.asDouble());
13613 const fp16type am (1.0 - ad);
13614 const double amd (am.asDouble());
13615 const fp16type xam (xd * amd);
13616 const double xamd (xam.asDouble());
13617 const fp16type ya (yd * ad);
13618 const double yad (ya.asDouble());
13619 const double xeps (floatFormat16.ulp(deAbs(xd * (1.0 - ad)), ulps));
13620 const double yeps (floatFormat16.ulp(deAbs(yd * ad), ulps));
13621 const double eps (xeps + yeps);
13623 result = xamd + yad;
13624 min[0] = result - eps;
13625 max[0] = result + eps;
13627 else if (getFlavor() == 2)
13629 const double xd (x.asDouble());
13630 const double yd (y.asDouble());
13631 const double ad (a.asDouble());
13632 const fp16type ymx (yd - xd);
13633 const double ymxd (ymx.asDouble());
13634 const fp16type ymxa (ymxd * ad);
13635 const double ymxad (ymxa.asDouble());
13636 const double xeps (floatFormat16.ulp(deAbs(xd * (1.0 - ad)), ulps));
13637 const double yeps (floatFormat16.ulp(deAbs(yd * ad), ulps));
13638 const double eps (xeps + yeps);
13640 result = xd + ymxad;
13641 min[0] = result - eps;
13642 max[0] = result + eps;
13646 TCU_THROW(InternalError, "Unknown flavor");
13649 out[0] = fp16type(result).bits();
13655 struct fp16SmoothStep : public fp16PerComponent
13657 fp16SmoothStep() : fp16PerComponent()
13659 flavorNames.push_back("FloatCalc");
13660 flavorNames.push_back("EmulatingFP16");
13661 flavorNames.push_back("EmulatingFP16WClamp");
13664 virtual double getULPs(vector<const deFloat16*>& in)
13668 return 4.0; // This is not a precision test. Value is not from spec
13671 template<class fp16type>
13672 bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
13674 const fp16type edge0 (*in[0]);
13675 const fp16type edge1 (*in[1]);
13676 const fp16type x (*in[2]);
13677 double result (0.0);
13679 if (edge0.isNaN() || edge1.isNaN() || x.isNaN() || edge0.asDouble() >= edge1.asDouble())
13682 if (edge0.isInf() || edge1.isInf() || x.isInf())
13685 if (getFlavor() == 0)
13687 const float edge0d (edge0.asFloat());
13688 const float edge1d (edge1.asFloat());
13689 const float xd (x.asFloat());
13690 const float sstep (deFloatSmoothStep(edge0d, edge1d, xd));
13694 else if (getFlavor() == 1)
13696 const double edge0d (edge0.asDouble());
13697 const double edge1d (edge1.asDouble());
13698 const double xd (x.asDouble());
13702 else if (xd >= edge1d)
13706 const fp16type a (xd - edge0d);
13707 const fp16type b (edge1d - edge0d);
13708 const fp16type t (a.asDouble() / b.asDouble());
13709 const fp16type t2 (2.0 * t.asDouble());
13710 const fp16type t3 (3.0 - t2.asDouble());
13711 const fp16type t4 (t.asDouble() * t3.asDouble());
13712 const fp16type t5 (t.asDouble() * t4.asDouble());
13714 result = t5.asDouble();
13717 else if (getFlavor() == 2)
13719 const double edge0d (edge0.asDouble());
13720 const double edge1d (edge1.asDouble());
13721 const double xd (x.asDouble());
13722 const fp16type a (xd - edge0d);
13723 const fp16type b (edge1d - edge0d);
13724 const fp16type bi (1.0 / b.asDouble());
13725 const fp16type t0 (a.asDouble() * bi.asDouble());
13726 const double tc (deClamp(t0.asDouble(), 0.0, 1.0));
13727 const fp16type t (tc);
13728 const fp16type t2 (2.0 * t.asDouble());
13729 const fp16type t3 (3.0 - t2.asDouble());
13730 const fp16type t4 (t.asDouble() * t3.asDouble());
13731 const fp16type t5 (t.asDouble() * t4.asDouble());
13733 result = t5.asDouble();
13737 TCU_THROW(InternalError, "Unknown flavor");
13740 out[0] = fp16type(result).bits();
13741 min[0] = getMin(result, getULPs(in));
13742 max[0] = getMax(result, getULPs(in));
13748 struct fp16Fma : public fp16PerComponent
13750 virtual double getULPs(vector<const deFloat16*>& in)
13757 template<class fp16type>
13758 bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
13760 DE_ASSERT(in.size() == 3);
13761 DE_ASSERT(getArgCompCount(0) == getOutCompCount());
13762 DE_ASSERT(getArgCompCount(1) == getOutCompCount());
13763 DE_ASSERT(getArgCompCount(2) == getOutCompCount());
13764 DE_ASSERT(getOutCompCount() > 0);
13766 const fp16type a (*in[0]);
13767 const fp16type b (*in[1]);
13768 const fp16type c (*in[2]);
13769 const double ad (a.asDouble());
13770 const double bd (b.asDouble());
13771 const double cd (c.asDouble());
13772 const double result (deMadd(ad, bd, cd));
13774 out[0] = fp16type(result).bits();
13775 min[0] = getMin(result, getULPs(in));
13776 max[0] = getMax(result, getULPs(in));
13783 struct fp16AllComponents : public fp16PerComponent
13785 bool callOncePerComponent () { return false; }
13788 struct fp16Length : public fp16AllComponents
13790 fp16Length() : fp16AllComponents()
13792 flavorNames.push_back("EmulatingFP16");
13793 flavorNames.push_back("DoubleCalc");
13796 virtual double getULPs(vector<const deFloat16*>& in)
13803 template<class fp16type>
13804 bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
13806 DE_ASSERT(getOutCompCount() == 1);
13807 DE_ASSERT(in.size() == 1);
13809 double result (0.0);
13811 if (getFlavor() == 0)
13815 for (size_t componentNdx = 0; componentNdx < getArgCompCount(0); ++componentNdx)
13817 const fp16type x (in[0][componentNdx]);
13818 const fp16type q (x.asDouble() * x.asDouble());
13820 r = fp16type(r.asDouble() + q.asDouble());
13823 result = deSqrt(r.asDouble());
13825 out[0] = fp16type(result).bits();
13827 else if (getFlavor() == 1)
13831 for (size_t componentNdx = 0; componentNdx < getArgCompCount(0); ++componentNdx)
13833 const fp16type x (in[0][componentNdx]);
13834 const double q (x.asDouble() * x.asDouble());
13839 result = deSqrt(r);
13841 out[0] = fp16type(result).bits();
13845 TCU_THROW(InternalError, "Unknown flavor");
13848 min[0] = getMin(result, getULPs(in));
13849 max[0] = getMax(result, getULPs(in));
13855 struct fp16Distance : public fp16AllComponents
13857 fp16Distance() : fp16AllComponents()
13859 flavorNames.push_back("EmulatingFP16");
13860 flavorNames.push_back("DoubleCalc");
13863 virtual double getULPs(vector<const deFloat16*>& in)
13870 template<class fp16type>
13871 bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
13873 DE_ASSERT(getOutCompCount() == 1);
13874 DE_ASSERT(in.size() == 2);
13875 DE_ASSERT(getArgCompCount(0) == getArgCompCount(1));
13877 double result (0.0);
13879 if (getFlavor() == 0)
13883 for (size_t componentNdx = 0; componentNdx < getArgCompCount(0); ++componentNdx)
13885 const fp16type x (in[0][componentNdx]);
13886 const fp16type y (in[1][componentNdx]);
13887 const fp16type d (x.asDouble() - y.asDouble());
13888 const fp16type q (d.asDouble() * d.asDouble());
13890 r = fp16type(r.asDouble() + q.asDouble());
13893 result = deSqrt(r.asDouble());
13895 else if (getFlavor() == 1)
13899 for (size_t componentNdx = 0; componentNdx < getArgCompCount(0); ++componentNdx)
13901 const fp16type x (in[0][componentNdx]);
13902 const fp16type y (in[1][componentNdx]);
13903 const double d (x.asDouble() - y.asDouble());
13904 const double q (d * d);
13909 result = deSqrt(r);
13913 TCU_THROW(InternalError, "Unknown flavor");
13916 out[0] = fp16type(result).bits();
13917 min[0] = getMin(result, getULPs(in));
13918 max[0] = getMax(result, getULPs(in));
13924 struct fp16Cross : public fp16AllComponents
13926 fp16Cross() : fp16AllComponents()
13928 flavorNames.push_back("EmulatingFP16");
13929 flavorNames.push_back("DoubleCalc");
13932 virtual double getULPs(vector<const deFloat16*>& in)
13939 template<class fp16type>
13940 bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
13942 DE_ASSERT(getOutCompCount() == 3);
13943 DE_ASSERT(in.size() == 2);
13944 DE_ASSERT(getArgCompCount(0) == 3);
13945 DE_ASSERT(getArgCompCount(1) == 3);
13947 if (getFlavor() == 0)
13949 const fp16type x0 (in[0][0]);
13950 const fp16type x1 (in[0][1]);
13951 const fp16type x2 (in[0][2]);
13952 const fp16type y0 (in[1][0]);
13953 const fp16type y1 (in[1][1]);
13954 const fp16type y2 (in[1][2]);
13955 const fp16type x1y2 (x1.asDouble() * y2.asDouble());
13956 const fp16type y1x2 (y1.asDouble() * x2.asDouble());
13957 const fp16type x2y0 (x2.asDouble() * y0.asDouble());
13958 const fp16type y2x0 (y2.asDouble() * x0.asDouble());
13959 const fp16type x0y1 (x0.asDouble() * y1.asDouble());
13960 const fp16type y0x1 (y0.asDouble() * x1.asDouble());
13962 out[0] = fp16type(x1y2.asDouble() - y1x2.asDouble()).bits();
13963 out[1] = fp16type(x2y0.asDouble() - y2x0.asDouble()).bits();
13964 out[2] = fp16type(x0y1.asDouble() - y0x1.asDouble()).bits();
13966 else if (getFlavor() == 1)
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 double x1y2 (x1.asDouble() * y2.asDouble());
13975 const double y1x2 (y1.asDouble() * x2.asDouble());
13976 const double x2y0 (x2.asDouble() * y0.asDouble());
13977 const double y2x0 (y2.asDouble() * x0.asDouble());
13978 const double x0y1 (x0.asDouble() * y1.asDouble());
13979 const double y0x1 (y0.asDouble() * x1.asDouble());
13981 out[0] = fp16type(x1y2 - y1x2).bits();
13982 out[1] = fp16type(x2y0 - y2x0).bits();
13983 out[2] = fp16type(x0y1 - y0x1).bits();
13987 TCU_THROW(InternalError, "Unknown flavor");
13990 for (size_t ndx = 0; ndx < getOutCompCount(); ++ndx)
13991 min[ndx] = getMin(fp16type(out[ndx]).asDouble(), getULPs(in));
13992 for (size_t ndx = 0; ndx < getOutCompCount(); ++ndx)
13993 max[ndx] = getMax(fp16type(out[ndx]).asDouble(), getULPs(in));
13999 struct fp16Normalize : public fp16AllComponents
14001 fp16Normalize() : fp16AllComponents()
14003 flavorNames.push_back("EmulatingFP16");
14004 flavorNames.push_back("DoubleCalc");
14006 // flavorNames will be extended later
14009 virtual void setArgCompCount (size_t argNo, size_t compCount)
14011 DE_ASSERT(argCompCount[argNo] == 0); // Once only
14013 if (argNo == 0 && argCompCount[argNo] == 0)
14015 const size_t maxPermutationsCount = 24u; // Equal to 4!
14016 std::vector<int> indices;
14018 for (size_t componentNdx = 0; componentNdx < compCount; ++componentNdx)
14019 indices.push_back(static_cast<int>(componentNdx));
14021 m_permutations.reserve(maxPermutationsCount);
14023 permutationsFlavorStart = flavorNames.size();
14027 tcu::UVec4 permutation;
14028 std::string name = "Permutted_";
14030 for (size_t componentNdx = 0; componentNdx < compCount; ++componentNdx)
14032 permutation[static_cast<int>(componentNdx)] = indices[componentNdx];
14033 name += de::toString(indices[componentNdx]);
14036 m_permutations.push_back(permutation);
14037 flavorNames.push_back(name);
14039 } while(std::next_permutation(indices.begin(), indices.end()));
14041 permutationsFlavorEnd = flavorNames.size();
14044 fp16AllComponents::setArgCompCount(argNo, compCount);
14046 virtual double getULPs(vector<const deFloat16*>& in)
14053 template<class fp16type>
14054 bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
14056 DE_ASSERT(in.size() == 1);
14057 DE_ASSERT(getArgCompCount(0) == getOutCompCount());
14059 if (getFlavor() == 0)
14063 for (size_t componentNdx = 0; componentNdx < getArgCompCount(0); ++componentNdx)
14065 const fp16type x (in[0][componentNdx]);
14066 const fp16type q (x.asDouble() * x.asDouble());
14068 r = fp16type(r.asDouble() + q.asDouble());
14071 r = fp16type(deSqrt(r.asDouble()));
14076 for (size_t componentNdx = 0; componentNdx < getArgCompCount(0); ++componentNdx)
14078 const fp16type x (in[0][componentNdx]);
14080 out[componentNdx] = fp16type(x.asDouble() / r.asDouble()).bits();
14083 else if (getFlavor() == 1)
14087 for (size_t componentNdx = 0; componentNdx < getArgCompCount(0); ++componentNdx)
14089 const fp16type x (in[0][componentNdx]);
14090 const double q (x.asDouble() * x.asDouble());
14100 for (size_t componentNdx = 0; componentNdx < getArgCompCount(0); ++componentNdx)
14102 const fp16type x (in[0][componentNdx]);
14104 out[componentNdx] = fp16type(x.asDouble() / r).bits();
14107 else if (de::inBounds<size_t>(getFlavor(), permutationsFlavorStart, permutationsFlavorEnd))
14109 const int compCount (static_cast<int>(getArgCompCount(0)));
14110 const size_t permutationNdx (getFlavor() - permutationsFlavorStart);
14111 const tcu::UVec4& permutation (m_permutations[permutationNdx]);
14114 for (int permComponentNdx = 0; permComponentNdx < compCount; ++permComponentNdx)
14116 const size_t componentNdx (permutation[permComponentNdx]);
14117 const fp16type x (in[0][componentNdx]);
14118 const fp16type q (x.asDouble() * x.asDouble());
14120 r = fp16type(r.asDouble() + q.asDouble());
14123 r = fp16type(deSqrt(r.asDouble()));
14128 for (int permComponentNdx = 0; permComponentNdx < compCount; ++permComponentNdx)
14130 const size_t componentNdx (permutation[permComponentNdx]);
14131 const fp16type x (in[0][componentNdx]);
14133 out[componentNdx] = fp16type(x.asDouble() / r.asDouble()).bits();
14138 TCU_THROW(InternalError, "Unknown flavor");
14141 for (size_t ndx = 0; ndx < getOutCompCount(); ++ndx)
14142 min[ndx] = getMin(fp16type(out[ndx]).asDouble(), getULPs(in));
14143 for (size_t ndx = 0; ndx < getOutCompCount(); ++ndx)
14144 max[ndx] = getMax(fp16type(out[ndx]).asDouble(), getULPs(in));
14150 std::vector<tcu::UVec4> m_permutations;
14151 size_t permutationsFlavorStart;
14152 size_t permutationsFlavorEnd;
14155 struct fp16FaceForward : public fp16AllComponents
14157 virtual double getULPs(vector<const deFloat16*>& in)
14164 template<class fp16type>
14165 bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
14167 DE_ASSERT(in.size() == 3);
14168 DE_ASSERT(getArgCompCount(0) == getOutCompCount());
14169 DE_ASSERT(getArgCompCount(1) == getOutCompCount());
14170 DE_ASSERT(getArgCompCount(2) == getOutCompCount());
14174 for (size_t componentNdx = 0; componentNdx < getArgCompCount(1); ++componentNdx)
14176 const fp16type x (in[1][componentNdx]);
14177 const fp16type y (in[2][componentNdx]);
14178 const double xd (x.asDouble());
14179 const double yd (y.asDouble());
14180 const fp16type q (xd * yd);
14182 dp = fp16type(dp.asDouble() + q.asDouble());
14185 if (dp.isNaN() || dp.isZero())
14188 for (size_t componentNdx = 0; componentNdx < getOutCompCount(); ++componentNdx)
14190 const fp16type n (in[0][componentNdx]);
14192 out[componentNdx] = (dp.signBit() == 1) ? n.bits() : fp16type(-n.asDouble()).bits();
14195 for (size_t ndx = 0; ndx < getOutCompCount(); ++ndx)
14196 min[ndx] = getMin(fp16type(out[ndx]).asDouble(), getULPs(in));
14197 for (size_t ndx = 0; ndx < getOutCompCount(); ++ndx)
14198 max[ndx] = getMax(fp16type(out[ndx]).asDouble(), getULPs(in));
14204 struct fp16Reflect : public fp16AllComponents
14206 fp16Reflect() : fp16AllComponents()
14208 flavorNames.push_back("EmulatingFP16");
14209 flavorNames.push_back("EmulatingFP16+KeepZeroSign");
14210 flavorNames.push_back("FloatCalc");
14211 flavorNames.push_back("FloatCalc+KeepZeroSign");
14212 flavorNames.push_back("EmulatingFP16+2Nfirst");
14213 flavorNames.push_back("EmulatingFP16+2Ifirst");
14216 virtual double getULPs(vector<const deFloat16*>& in)
14220 return 256.0; // This is not a precision test. Value is not from spec
14223 template<class fp16type>
14224 bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
14226 DE_ASSERT(in.size() == 2);
14227 DE_ASSERT(getArgCompCount(0) == getOutCompCount());
14228 DE_ASSERT(getArgCompCount(1) == getOutCompCount());
14230 if (getFlavor() < 4)
14232 const bool keepZeroSign ((flavor & 1) != 0 ? true : false);
14233 const bool floatCalc ((flavor & 2) != 0 ? true : false);
14239 for (size_t componentNdx = 0; componentNdx < getArgCompCount(1); ++componentNdx)
14241 const fp16type i (in[0][componentNdx]);
14242 const fp16type n (in[1][componentNdx]);
14243 const float id (i.asFloat());
14244 const float nd (n.asFloat());
14245 const float qd (id * nd);
14248 dp = (componentNdx == 0) ? qd : dp + qd;
14253 for (size_t componentNdx = 0; componentNdx < getOutCompCount(); ++componentNdx)
14255 const fp16type i (in[0][componentNdx]);
14256 const fp16type n (in[1][componentNdx]);
14257 const float dpnd (dp * n.asFloat());
14258 const float dpn2d (2.0f * dpnd);
14259 const float idpn2d (i.asFloat() - dpn2d);
14260 const fp16type result (idpn2d);
14262 out[componentNdx] = result.bits();
14269 for (size_t componentNdx = 0; componentNdx < getArgCompCount(1); ++componentNdx)
14271 const fp16type i (in[0][componentNdx]);
14272 const fp16type n (in[1][componentNdx]);
14273 const double id (i.asDouble());
14274 const double nd (n.asDouble());
14275 const fp16type q (id * nd);
14278 dp = (componentNdx == 0) ? q : fp16type(dp.asDouble() + q.asDouble());
14280 dp = fp16type(dp.asDouble() + q.asDouble());
14286 for (size_t componentNdx = 0; componentNdx < getOutCompCount(); ++componentNdx)
14288 const fp16type i (in[0][componentNdx]);
14289 const fp16type n (in[1][componentNdx]);
14290 const fp16type dpn (dp.asDouble() * n.asDouble());
14291 const fp16type dpn2 (2 * dpn.asDouble());
14292 const fp16type idpn2 (i.asDouble() - dpn2.asDouble());
14294 out[componentNdx] = idpn2.bits();
14298 else if (getFlavor() == 4)
14302 for (size_t componentNdx = 0; componentNdx < getArgCompCount(1); ++componentNdx)
14304 const fp16type i (in[0][componentNdx]);
14305 const fp16type n (in[1][componentNdx]);
14306 const double id (i.asDouble());
14307 const double nd (n.asDouble());
14308 const fp16type q (id * nd);
14310 dp = fp16type(dp.asDouble() + q.asDouble());
14316 for (size_t componentNdx = 0; componentNdx < getOutCompCount(); ++componentNdx)
14318 const fp16type i (in[0][componentNdx]);
14319 const fp16type n (in[1][componentNdx]);
14320 const fp16type n2 (2 * n.asDouble());
14321 const fp16type dpn2 (dp.asDouble() * n2.asDouble());
14322 const fp16type idpn2 (i.asDouble() - dpn2.asDouble());
14324 out[componentNdx] = idpn2.bits();
14327 else if (getFlavor() == 5)
14331 for (size_t componentNdx = 0; componentNdx < getArgCompCount(1); ++componentNdx)
14333 const fp16type i (in[0][componentNdx]);
14334 const fp16type n (in[1][componentNdx]);
14335 const fp16type i2 (2.0 * i.asDouble());
14336 const double i2d (i2.asDouble());
14337 const double nd (n.asDouble());
14338 const fp16type q (i2d * nd);
14340 dp2 = fp16type(dp2.asDouble() + q.asDouble());
14346 for (size_t componentNdx = 0; componentNdx < getOutCompCount(); ++componentNdx)
14348 const fp16type i (in[0][componentNdx]);
14349 const fp16type n (in[1][componentNdx]);
14350 const fp16type dpn2 (dp2.asDouble() * n.asDouble());
14351 const fp16type idpn2 (i.asDouble() - dpn2.asDouble());
14353 out[componentNdx] = idpn2.bits();
14358 TCU_THROW(InternalError, "Unknown flavor");
14361 for (size_t ndx = 0; ndx < getOutCompCount(); ++ndx)
14362 min[ndx] = getMin(fp16type(out[ndx]).asDouble(), getULPs(in));
14363 for (size_t ndx = 0; ndx < getOutCompCount(); ++ndx)
14364 max[ndx] = getMax(fp16type(out[ndx]).asDouble(), getULPs(in));
14370 struct fp16Refract : public fp16AllComponents
14372 fp16Refract() : fp16AllComponents()
14374 flavorNames.push_back("EmulatingFP16");
14375 flavorNames.push_back("EmulatingFP16+KeepZeroSign");
14376 flavorNames.push_back("FloatCalc");
14377 flavorNames.push_back("FloatCalc+KeepZeroSign");
14380 virtual double getULPs(vector<const deFloat16*>& in)
14384 return 8192.0; // This is not a precision test. Value is not from spec
14387 template<class fp16type>
14388 bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
14390 DE_ASSERT(in.size() == 3);
14391 DE_ASSERT(getArgCompCount(0) == getOutCompCount());
14392 DE_ASSERT(getArgCompCount(1) == getOutCompCount());
14393 DE_ASSERT(getArgCompCount(2) == 1);
14395 const bool keepZeroSign ((flavor & 1) != 0 ? true : false);
14396 const bool doubleCalc ((flavor & 2) != 0 ? true : false);
14397 const fp16type eta (*in[2]);
14403 for (size_t componentNdx = 0; componentNdx < getArgCompCount(1); ++componentNdx)
14405 const fp16type i (in[0][componentNdx]);
14406 const fp16type n (in[1][componentNdx]);
14407 const double id (i.asDouble());
14408 const double nd (n.asDouble());
14409 const double qd (id * nd);
14412 dp = (componentNdx == 0) ? qd : dp + qd;
14417 const double eta2 (eta.asDouble() * eta.asDouble());
14418 const double dp2 (dp * dp);
14419 const double dp1 (1.0 - dp2);
14420 const double dpe (eta2 * dp1);
14421 const double k (1.0 - dpe);
14425 const fp16type zero (0.0);
14427 for (size_t componentNdx = 0; componentNdx < getOutCompCount(); ++componentNdx)
14428 out[componentNdx] = zero.bits();
14432 const double sk (deSqrt(k));
14434 for (size_t componentNdx = 0; componentNdx < getOutCompCount(); ++componentNdx)
14436 const fp16type i (in[0][componentNdx]);
14437 const fp16type n (in[1][componentNdx]);
14438 const double etai (i.asDouble() * eta.asDouble());
14439 const double etadp (eta.asDouble() * dp);
14440 const double etadpk (etadp + sk);
14441 const double etadpkn (etadpk * n.asDouble());
14442 const double full (etai - etadpkn);
14443 const fp16type result (full);
14445 if (result.isInf())
14448 out[componentNdx] = result.bits();
14456 for (size_t componentNdx = 0; componentNdx < getArgCompCount(1); ++componentNdx)
14458 const fp16type i (in[0][componentNdx]);
14459 const fp16type n (in[1][componentNdx]);
14460 const double id (i.asDouble());
14461 const double nd (n.asDouble());
14462 const fp16type q (id * nd);
14465 dp = (componentNdx == 0) ? q : fp16type(dp.asDouble() + q.asDouble());
14467 dp = fp16type(dp.asDouble() + q.asDouble());
14473 const fp16type eta2(eta.asDouble() * eta.asDouble());
14474 const fp16type dp2 (dp.asDouble() * dp.asDouble());
14475 const fp16type dp1 (1.0 - dp2.asDouble());
14476 const fp16type dpe (eta2.asDouble() * dp1.asDouble());
14477 const fp16type k (1.0 - dpe.asDouble());
14479 if (k.asDouble() < 0.0)
14481 const fp16type zero (0.0);
14483 for (size_t componentNdx = 0; componentNdx < getOutCompCount(); ++componentNdx)
14484 out[componentNdx] = zero.bits();
14488 const fp16type sk (deSqrt(k.asDouble()));
14490 for (size_t componentNdx = 0; componentNdx < getOutCompCount(); ++componentNdx)
14492 const fp16type i (in[0][componentNdx]);
14493 const fp16type n (in[1][componentNdx]);
14494 const fp16type etai (i.asDouble() * eta.asDouble());
14495 const fp16type etadp (eta.asDouble() * dp.asDouble());
14496 const fp16type etadpk (etadp.asDouble() + sk.asDouble());
14497 const fp16type etadpkn (etadpk.asDouble() * n.asDouble());
14498 const fp16type full (etai.asDouble() - etadpkn.asDouble());
14500 if (full.isNaN() || full.isInf())
14503 out[componentNdx] = full.bits();
14508 for (size_t ndx = 0; ndx < getOutCompCount(); ++ndx)
14509 min[ndx] = getMin(fp16type(out[ndx]).asDouble(), getULPs(in));
14510 for (size_t ndx = 0; ndx < getOutCompCount(); ++ndx)
14511 max[ndx] = getMax(fp16type(out[ndx]).asDouble(), getULPs(in));
14517 struct fp16Dot : public fp16AllComponents
14519 fp16Dot() : fp16AllComponents()
14521 flavorNames.push_back("EmulatingFP16");
14522 flavorNames.push_back("FloatCalc");
14523 flavorNames.push_back("DoubleCalc");
14525 // flavorNames will be extended later
14528 virtual void setArgCompCount (size_t argNo, size_t compCount)
14530 DE_ASSERT(argCompCount[argNo] == 0); // Once only
14532 if (argNo == 0 && argCompCount[argNo] == 0)
14534 const size_t maxPermutationsCount = 24u; // Equal to 4!
14535 std::vector<int> indices;
14537 for (size_t componentNdx = 0; componentNdx < compCount; ++componentNdx)
14538 indices.push_back(static_cast<int>(componentNdx));
14540 m_permutations.reserve(maxPermutationsCount);
14542 permutationsFlavorStart = flavorNames.size();
14546 tcu::UVec4 permutation;
14547 std::string name = "Permutted_";
14549 for (size_t componentNdx = 0; componentNdx < compCount; ++componentNdx)
14551 permutation[static_cast<int>(componentNdx)] = indices[componentNdx];
14552 name += de::toString(indices[componentNdx]);
14555 m_permutations.push_back(permutation);
14556 flavorNames.push_back(name);
14558 } while(std::next_permutation(indices.begin(), indices.end()));
14560 permutationsFlavorEnd = flavorNames.size();
14563 fp16AllComponents::setArgCompCount(argNo, compCount);
14566 virtual double getULPs(vector<const deFloat16*>& in)
14570 return 16.0; // This is not a precision test. Value is not from spec
14573 template<class fp16type>
14574 bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
14576 DE_ASSERT(in.size() == 2);
14577 DE_ASSERT(getArgCompCount(0) == getArgCompCount(1));
14578 DE_ASSERT(getOutCompCount() == 1);
14580 double result (0.0);
14583 if (getFlavor() == 0)
14587 for (size_t componentNdx = 0; componentNdx < getArgCompCount(1); ++componentNdx)
14589 const fp16type x (in[0][componentNdx]);
14590 const fp16type y (in[1][componentNdx]);
14591 const fp16type q (x.asDouble() * y.asDouble());
14593 dp = fp16type(dp.asDouble() + q.asDouble());
14594 eps += floatFormat16.ulp(q.asDouble(), 2.0);
14597 result = dp.asDouble();
14599 else if (getFlavor() == 1)
14603 for (size_t componentNdx = 0; componentNdx < getArgCompCount(1); ++componentNdx)
14605 const fp16type x (in[0][componentNdx]);
14606 const fp16type y (in[1][componentNdx]);
14607 const float q (x.asFloat() * y.asFloat());
14610 eps += floatFormat16.ulp(static_cast<double>(q), 2.0);
14615 else if (getFlavor() == 2)
14619 for (size_t componentNdx = 0; componentNdx < getArgCompCount(1); ++componentNdx)
14621 const fp16type x (in[0][componentNdx]);
14622 const fp16type y (in[1][componentNdx]);
14623 const double q (x.asDouble() * y.asDouble());
14626 eps += floatFormat16.ulp(q, 2.0);
14631 else if (de::inBounds<size_t>(getFlavor(), permutationsFlavorStart, permutationsFlavorEnd))
14633 const int compCount (static_cast<int>(getArgCompCount(1)));
14634 const size_t permutationNdx (getFlavor() - permutationsFlavorStart);
14635 const tcu::UVec4& permutation (m_permutations[permutationNdx]);
14638 for (int permComponentNdx = 0; permComponentNdx < compCount; ++permComponentNdx)
14640 const size_t componentNdx (permutation[permComponentNdx]);
14641 const fp16type x (in[0][componentNdx]);
14642 const fp16type y (in[1][componentNdx]);
14643 const fp16type q (x.asDouble() * y.asDouble());
14645 dp = fp16type(dp.asDouble() + q.asDouble());
14646 eps += floatFormat16.ulp(q.asDouble(), 2.0);
14649 result = dp.asDouble();
14653 TCU_THROW(InternalError, "Unknown flavor");
14656 out[0] = fp16type(result).bits();
14657 min[0] = result - eps;
14658 max[0] = result + eps;
14664 std::vector<tcu::UVec4> m_permutations;
14665 size_t permutationsFlavorStart;
14666 size_t permutationsFlavorEnd;
14669 struct fp16VectorTimesScalar : public fp16AllComponents
14671 virtual double getULPs(vector<const deFloat16*>& in)
14678 template<class fp16type>
14679 bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
14681 DE_ASSERT(in.size() == 2);
14682 DE_ASSERT(getArgCompCount(0) == getOutCompCount());
14683 DE_ASSERT(getArgCompCount(1) == 1);
14685 fp16type s (*in[1]);
14687 for (size_t componentNdx = 0; componentNdx < getArgCompCount(0); ++componentNdx)
14689 const fp16type x (in[0][componentNdx]);
14690 const double result (s.asDouble() * x.asDouble());
14691 const fp16type m (result);
14693 out[componentNdx] = m.bits();
14694 min[componentNdx] = getMin(result, getULPs(in));
14695 max[componentNdx] = getMax(result, getULPs(in));
14702 struct fp16MatrixBase : public fp16AllComponents
14704 deUint32 getComponentValidity ()
14706 return static_cast<deUint32>(-1);
14709 inline size_t getNdx (const size_t rowCount, const size_t col, const size_t row)
14711 const size_t minComponentCount = 0;
14712 const size_t maxComponentCount = 3;
14713 const size_t alignedRowsCount = (rowCount == 3) ? 4 : rowCount;
14715 DE_ASSERT(de::inRange(rowCount, minComponentCount + 1, maxComponentCount + 1));
14716 DE_ASSERT(de::inRange(col, minComponentCount, maxComponentCount));
14717 DE_ASSERT(de::inBounds(row, minComponentCount, rowCount));
14718 DE_UNREF(minComponentCount);
14719 DE_UNREF(maxComponentCount);
14721 return col * alignedRowsCount + row;
14724 deUint32 getComponentMatrixValidityMask (size_t cols, size_t rows)
14726 deUint32 result = 0u;
14728 for (size_t rowNdx = 0; rowNdx < rows; ++rowNdx)
14729 for (size_t colNdx = 0; colNdx < cols; ++colNdx)
14731 const size_t bitNdx = getNdx(rows, colNdx, rowNdx);
14733 DE_ASSERT(bitNdx < sizeof(result) * 8);
14735 result |= (1<<bitNdx);
14742 template<size_t cols, size_t rows>
14743 struct fp16Transpose : public fp16MatrixBase
14745 virtual double getULPs(vector<const deFloat16*>& in)
14752 deUint32 getComponentValidity ()
14754 return getComponentMatrixValidityMask(rows, cols);
14757 template<class fp16type>
14758 bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
14760 DE_ASSERT(in.size() == 1);
14762 const size_t alignedCols = (cols == 3) ? 4 : cols;
14763 const size_t alignedRows = (rows == 3) ? 4 : rows;
14764 vector<deFloat16> output (alignedCols * alignedRows, 0);
14766 DE_ASSERT(output.size() == alignedCols * alignedRows);
14768 for (size_t rowNdx = 0; rowNdx < rows; ++rowNdx)
14769 for (size_t colNdx = 0; colNdx < cols; ++colNdx)
14770 output[rowNdx * alignedCols + colNdx] = in[0][colNdx * alignedRows + rowNdx];
14772 deMemcpy(out, &output[0], sizeof(deFloat16) * output.size());
14773 deMemcpy(min, &output[0], sizeof(deFloat16) * output.size());
14774 deMemcpy(max, &output[0], sizeof(deFloat16) * output.size());
14780 template<size_t cols, size_t rows>
14781 struct fp16MatrixTimesScalar : public fp16MatrixBase
14783 virtual double getULPs(vector<const deFloat16*>& in)
14790 deUint32 getComponentValidity ()
14792 return getComponentMatrixValidityMask(cols, rows);
14795 template<class fp16type>
14796 bool calc(vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
14798 DE_ASSERT(in.size() == 2);
14799 DE_ASSERT(getArgCompCount(1) == 1);
14801 const fp16type y (in[1][0]);
14802 const float scalar (y.asFloat());
14803 const size_t alignedCols = (cols == 3) ? 4 : cols;
14804 const size_t alignedRows = (rows == 3) ? 4 : rows;
14806 DE_ASSERT(getArgCompCount(0) == alignedCols * alignedRows);
14807 DE_ASSERT(getOutCompCount() == alignedCols * alignedRows);
14808 DE_UNREF(alignedCols);
14810 for (size_t rowNdx = 0; rowNdx < rows; ++rowNdx)
14811 for (size_t colNdx = 0; colNdx < cols; ++colNdx)
14813 const size_t ndx (colNdx * alignedRows + rowNdx);
14814 const fp16type x (in[0][ndx]);
14815 const double result (scalar * x.asFloat());
14817 out[ndx] = fp16type(result).bits();
14818 min[ndx] = getMin(result, getULPs(in));
14819 max[ndx] = getMax(result, getULPs(in));
14826 template<size_t cols, size_t rows>
14827 struct fp16VectorTimesMatrix : public fp16MatrixBase
14829 fp16VectorTimesMatrix() : fp16MatrixBase()
14831 flavorNames.push_back("EmulatingFP16");
14832 flavorNames.push_back("FloatCalc");
14835 virtual double getULPs (vector<const deFloat16*>& in)
14839 return (8.0 * cols);
14842 deUint32 getComponentValidity ()
14844 return getComponentMatrixValidityMask(cols, 1);
14847 template<class fp16type>
14848 bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
14850 DE_ASSERT(in.size() == 2);
14852 const size_t alignedCols = (cols == 3) ? 4 : cols;
14853 const size_t alignedRows = (rows == 3) ? 4 : rows;
14855 DE_ASSERT(getOutCompCount() == cols);
14856 DE_ASSERT(getArgCompCount(0) == rows);
14857 DE_ASSERT(getArgCompCount(1) == alignedCols * alignedRows);
14858 DE_UNREF(alignedCols);
14860 if (getFlavor() == 0)
14862 for (size_t colNdx = 0; colNdx < cols; ++colNdx)
14864 fp16type s (fp16type::zero(1));
14866 for (size_t rowNdx = 0; rowNdx < rows; ++rowNdx)
14868 const fp16type v (in[0][rowNdx]);
14869 const float vf (v.asFloat());
14870 const size_t ndx (colNdx * alignedRows + rowNdx);
14871 const fp16type x (in[1][ndx]);
14872 const float xf (x.asFloat());
14873 const fp16type m (vf * xf);
14875 s = fp16type(s.asFloat() + m.asFloat());
14878 out[colNdx] = s.bits();
14879 min[colNdx] = getMin(s.asDouble(), getULPs(in));
14880 max[colNdx] = getMax(s.asDouble(), getULPs(in));
14883 else if (getFlavor() == 1)
14885 for (size_t colNdx = 0; colNdx < cols; ++colNdx)
14889 for (size_t rowNdx = 0; rowNdx < rows; ++rowNdx)
14891 const fp16type v (in[0][rowNdx]);
14892 const float vf (v.asFloat());
14893 const size_t ndx (colNdx * alignedRows + rowNdx);
14894 const fp16type x (in[1][ndx]);
14895 const float xf (x.asFloat());
14896 const float m (vf * xf);
14901 out[colNdx] = fp16type(s).bits();
14902 min[colNdx] = getMin(static_cast<double>(s), getULPs(in));
14903 max[colNdx] = getMax(static_cast<double>(s), getULPs(in));
14908 TCU_THROW(InternalError, "Unknown flavor");
14915 template<size_t cols, size_t rows>
14916 struct fp16MatrixTimesVector : public fp16MatrixBase
14918 fp16MatrixTimesVector() : fp16MatrixBase()
14920 flavorNames.push_back("EmulatingFP16");
14921 flavorNames.push_back("FloatCalc");
14924 virtual double getULPs (vector<const deFloat16*>& in)
14928 return (8.0 * rows);
14931 deUint32 getComponentValidity ()
14933 return getComponentMatrixValidityMask(rows, 1);
14936 template<class fp16type>
14937 bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
14939 DE_ASSERT(in.size() == 2);
14941 const size_t alignedCols = (cols == 3) ? 4 : cols;
14942 const size_t alignedRows = (rows == 3) ? 4 : rows;
14944 DE_ASSERT(getOutCompCount() == rows);
14945 DE_ASSERT(getArgCompCount(0) == alignedCols * alignedRows);
14946 DE_ASSERT(getArgCompCount(1) == cols);
14947 DE_UNREF(alignedCols);
14949 if (getFlavor() == 0)
14951 for (size_t rowNdx = 0; rowNdx < rows; ++rowNdx)
14953 fp16type s (fp16type::zero(1));
14955 for (size_t colNdx = 0; colNdx < cols; ++colNdx)
14957 const size_t ndx (colNdx * alignedRows + rowNdx);
14958 const fp16type x (in[0][ndx]);
14959 const float xf (x.asFloat());
14960 const fp16type v (in[1][colNdx]);
14961 const float vf (v.asFloat());
14962 const fp16type m (vf * xf);
14964 s = fp16type(s.asFloat() + m.asFloat());
14967 out[rowNdx] = s.bits();
14968 min[rowNdx] = getMin(s.asDouble(), getULPs(in));
14969 max[rowNdx] = getMax(s.asDouble(), getULPs(in));
14972 else if (getFlavor() == 1)
14974 for (size_t rowNdx = 0; rowNdx < rows; ++rowNdx)
14978 for (size_t colNdx = 0; colNdx < cols; ++colNdx)
14980 const size_t ndx (colNdx * alignedRows + rowNdx);
14981 const fp16type x (in[0][ndx]);
14982 const float xf (x.asFloat());
14983 const fp16type v (in[1][colNdx]);
14984 const float vf (v.asFloat());
14985 const float m (vf * xf);
14990 out[rowNdx] = fp16type(s).bits();
14991 min[rowNdx] = getMin(static_cast<double>(s), getULPs(in));
14992 max[rowNdx] = getMax(static_cast<double>(s), getULPs(in));
14997 TCU_THROW(InternalError, "Unknown flavor");
15004 template<size_t colsL, size_t rowsL, size_t colsR, size_t rowsR>
15005 struct fp16MatrixTimesMatrix : public fp16MatrixBase
15007 fp16MatrixTimesMatrix() : fp16MatrixBase()
15009 flavorNames.push_back("EmulatingFP16");
15010 flavorNames.push_back("FloatCalc");
15013 virtual double getULPs (vector<const deFloat16*>& in)
15020 deUint32 getComponentValidity ()
15022 return getComponentMatrixValidityMask(colsR, rowsL);
15025 template<class fp16type>
15026 bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
15028 DE_STATIC_ASSERT(colsL == rowsR);
15030 DE_ASSERT(in.size() == 2);
15032 const size_t alignedColsL = (colsL == 3) ? 4 : colsL;
15033 const size_t alignedRowsL = (rowsL == 3) ? 4 : rowsL;
15034 const size_t alignedColsR = (colsR == 3) ? 4 : colsR;
15035 const size_t alignedRowsR = (rowsR == 3) ? 4 : rowsR;
15037 DE_ASSERT(getOutCompCount() == alignedColsR * alignedRowsL);
15038 DE_ASSERT(getArgCompCount(0) == alignedColsL * alignedRowsL);
15039 DE_ASSERT(getArgCompCount(1) == alignedColsR * alignedRowsR);
15040 DE_UNREF(alignedColsL);
15041 DE_UNREF(alignedColsR);
15043 if (getFlavor() == 0)
15045 for (size_t rowNdx = 0; rowNdx < rowsL; ++rowNdx)
15047 for (size_t colNdx = 0; colNdx < colsR; ++colNdx)
15049 const size_t ndx (colNdx * alignedRowsL + rowNdx);
15050 fp16type s (fp16type::zero(1));
15052 for (size_t commonNdx = 0; commonNdx < colsL; ++commonNdx)
15054 const size_t ndxl (commonNdx * alignedRowsL + rowNdx);
15055 const fp16type l (in[0][ndxl]);
15056 const float lf (l.asFloat());
15057 const size_t ndxr (colNdx * alignedRowsR + commonNdx);
15058 const fp16type r (in[1][ndxr]);
15059 const float rf (r.asFloat());
15060 const fp16type m (lf * rf);
15062 s = fp16type(s.asFloat() + m.asFloat());
15065 out[ndx] = s.bits();
15066 min[ndx] = getMin(s.asDouble(), getULPs(in));
15067 max[ndx] = getMax(s.asDouble(), getULPs(in));
15071 else if (getFlavor() == 1)
15073 for (size_t rowNdx = 0; rowNdx < rowsL; ++rowNdx)
15075 for (size_t colNdx = 0; colNdx < colsR; ++colNdx)
15077 const size_t ndx (colNdx * alignedRowsL + rowNdx);
15080 for (size_t commonNdx = 0; commonNdx < colsL; ++commonNdx)
15082 const size_t ndxl (commonNdx * alignedRowsL + rowNdx);
15083 const fp16type l (in[0][ndxl]);
15084 const float lf (l.asFloat());
15085 const size_t ndxr (colNdx * alignedRowsR + commonNdx);
15086 const fp16type r (in[1][ndxr]);
15087 const float rf (r.asFloat());
15088 const float m (lf * rf);
15093 out[ndx] = fp16type(s).bits();
15094 min[ndx] = getMin(static_cast<double>(s), getULPs(in));
15095 max[ndx] = getMax(static_cast<double>(s), getULPs(in));
15101 TCU_THROW(InternalError, "Unknown flavor");
15108 template<size_t cols, size_t rows>
15109 struct fp16OuterProduct : public fp16MatrixBase
15111 virtual double getULPs (vector<const deFloat16*>& in)
15118 deUint32 getComponentValidity ()
15120 return getComponentMatrixValidityMask(cols, rows);
15123 template<class fp16type>
15124 bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
15126 DE_ASSERT(in.size() == 2);
15128 const size_t alignedCols = (cols == 3) ? 4 : cols;
15129 const size_t alignedRows = (rows == 3) ? 4 : rows;
15131 DE_ASSERT(getArgCompCount(0) == rows);
15132 DE_ASSERT(getArgCompCount(1) == cols);
15133 DE_ASSERT(getOutCompCount() == alignedCols * alignedRows);
15134 DE_UNREF(alignedCols);
15136 for (size_t rowNdx = 0; rowNdx < rows; ++rowNdx)
15138 for (size_t colNdx = 0; colNdx < cols; ++colNdx)
15140 const size_t ndx (colNdx * alignedRows + rowNdx);
15141 const fp16type x (in[0][rowNdx]);
15142 const float xf (x.asFloat());
15143 const fp16type y (in[1][colNdx]);
15144 const float yf (y.asFloat());
15145 const fp16type m (xf * yf);
15147 out[ndx] = m.bits();
15148 min[ndx] = getMin(m.asDouble(), getULPs(in));
15149 max[ndx] = getMax(m.asDouble(), getULPs(in));
15157 template<size_t size>
15158 struct fp16Determinant;
15161 struct fp16Determinant<2> : public fp16MatrixBase
15163 virtual double getULPs (vector<const deFloat16*>& in)
15167 return 128.0; // This is not a precision test. Value is not from spec
15170 deUint32 getComponentValidity ()
15175 template<class fp16type>
15176 bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
15178 const size_t cols = 2;
15179 const size_t rows = 2;
15180 const size_t alignedCols = (cols == 3) ? 4 : cols;
15181 const size_t alignedRows = (rows == 3) ? 4 : rows;
15183 DE_ASSERT(in.size() == 1);
15184 DE_ASSERT(getOutCompCount() == 1);
15185 DE_ASSERT(getArgCompCount(0) == alignedRows * alignedCols);
15186 DE_UNREF(alignedCols);
15187 DE_UNREF(alignedRows);
15191 const float a (fp16type(in[0][getNdx(rows, 0, 0)]).asFloat());
15192 const float b (fp16type(in[0][getNdx(rows, 1, 0)]).asFloat());
15193 const float c (fp16type(in[0][getNdx(rows, 0, 1)]).asFloat());
15194 const float d (fp16type(in[0][getNdx(rows, 1, 1)]).asFloat());
15195 const float ad (a * d);
15196 const fp16type adf16 (ad);
15197 const float bc (b * c);
15198 const fp16type bcf16 (bc);
15199 const float r (adf16.asFloat() - bcf16.asFloat());
15200 const fp16type rf16 (r);
15202 out[0] = rf16.bits();
15203 min[0] = getMin(r, getULPs(in));
15204 max[0] = getMax(r, getULPs(in));
15211 struct fp16Determinant<3> : public fp16MatrixBase
15213 virtual double getULPs (vector<const deFloat16*>& in)
15217 return 128.0; // This is not a precision test. Value is not from spec
15220 deUint32 getComponentValidity ()
15225 template<class fp16type>
15226 bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
15228 const size_t cols = 3;
15229 const size_t rows = 3;
15230 const size_t alignedCols = (cols == 3) ? 4 : cols;
15231 const size_t alignedRows = (rows == 3) ? 4 : rows;
15233 DE_ASSERT(in.size() == 1);
15234 DE_ASSERT(getOutCompCount() == 1);
15235 DE_ASSERT(getArgCompCount(0) == alignedRows * alignedCols);
15236 DE_UNREF(alignedCols);
15237 DE_UNREF(alignedRows);
15242 const float a (fp16type(in[0][getNdx(rows, 0, 0)]).asFloat());
15243 const float b (fp16type(in[0][getNdx(rows, 1, 0)]).asFloat());
15244 const float c (fp16type(in[0][getNdx(rows, 2, 0)]).asFloat());
15245 const float d (fp16type(in[0][getNdx(rows, 0, 1)]).asFloat());
15246 const float e (fp16type(in[0][getNdx(rows, 1, 1)]).asFloat());
15247 const float f (fp16type(in[0][getNdx(rows, 2, 1)]).asFloat());
15248 const float g (fp16type(in[0][getNdx(rows, 0, 2)]).asFloat());
15249 const float h (fp16type(in[0][getNdx(rows, 1, 2)]).asFloat());
15250 const float i (fp16type(in[0][getNdx(rows, 2, 2)]).asFloat());
15251 const fp16type aei (a * e * i);
15252 const fp16type bfg (b * f * g);
15253 const fp16type cdh (c * d * h);
15254 const fp16type ceg (c * e * g);
15255 const fp16type bdi (b * d * i);
15256 const fp16type afh (a * f * h);
15257 const float r (aei.asFloat() + bfg.asFloat() + cdh.asFloat() - ceg.asFloat() - bdi.asFloat() - afh.asFloat());
15258 const fp16type rf16 (r);
15260 out[0] = rf16.bits();
15261 min[0] = getMin(r, getULPs(in));
15262 max[0] = getMax(r, getULPs(in));
15269 struct fp16Determinant<4> : public fp16MatrixBase
15271 virtual double getULPs (vector<const deFloat16*>& in)
15275 return 128.0; // This is not a precision test. Value is not from spec
15278 deUint32 getComponentValidity ()
15283 template<class fp16type>
15284 bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
15286 const size_t rows = 4;
15287 const size_t cols = 4;
15288 const size_t alignedCols = (cols == 3) ? 4 : cols;
15289 const size_t alignedRows = (rows == 3) ? 4 : rows;
15291 DE_ASSERT(in.size() == 1);
15292 DE_ASSERT(getOutCompCount() == 1);
15293 DE_ASSERT(getArgCompCount(0) == alignedRows * alignedCols);
15294 DE_UNREF(alignedCols);
15295 DE_UNREF(alignedRows);
15301 const float a (fp16type(in[0][getNdx(rows, 0, 0)]).asFloat());
15302 const float b (fp16type(in[0][getNdx(rows, 1, 0)]).asFloat());
15303 const float c (fp16type(in[0][getNdx(rows, 2, 0)]).asFloat());
15304 const float d (fp16type(in[0][getNdx(rows, 3, 0)]).asFloat());
15305 const float e (fp16type(in[0][getNdx(rows, 0, 1)]).asFloat());
15306 const float f (fp16type(in[0][getNdx(rows, 1, 1)]).asFloat());
15307 const float g (fp16type(in[0][getNdx(rows, 2, 1)]).asFloat());
15308 const float h (fp16type(in[0][getNdx(rows, 3, 1)]).asFloat());
15309 const float i (fp16type(in[0][getNdx(rows, 0, 2)]).asFloat());
15310 const float j (fp16type(in[0][getNdx(rows, 1, 2)]).asFloat());
15311 const float k (fp16type(in[0][getNdx(rows, 2, 2)]).asFloat());
15312 const float l (fp16type(in[0][getNdx(rows, 3, 2)]).asFloat());
15313 const float m (fp16type(in[0][getNdx(rows, 0, 3)]).asFloat());
15314 const float n (fp16type(in[0][getNdx(rows, 1, 3)]).asFloat());
15315 const float o (fp16type(in[0][getNdx(rows, 2, 3)]).asFloat());
15316 const float p (fp16type(in[0][getNdx(rows, 3, 3)]).asFloat());
15321 const fp16type fkp (f * k * p);
15322 const fp16type gln (g * l * n);
15323 const fp16type hjo (h * j * o);
15324 const fp16type hkn (h * k * n);
15325 const fp16type gjp (g * j * p);
15326 const fp16type flo (f * l * o);
15327 const fp16type detA (a * (fkp.asFloat() + gln.asFloat() + hjo.asFloat() - hkn.asFloat() - gjp.asFloat() - flo.asFloat()));
15332 const fp16type ekp (e * k * p);
15333 const fp16type glm (g * l * m);
15334 const fp16type hio (h * i * o);
15335 const fp16type hkm (h * k * m);
15336 const fp16type gip (g * i * p);
15337 const fp16type elo (e * l * o);
15338 const fp16type detB (b * (ekp.asFloat() + glm.asFloat() + hio.asFloat() - hkm.asFloat() - gip.asFloat() - elo.asFloat()));
15343 const fp16type ejp (e * j * p);
15344 const fp16type flm (f * l * m);
15345 const fp16type hin (h * i * n);
15346 const fp16type hjm (h * j * m);
15347 const fp16type fip (f * i * p);
15348 const fp16type eln (e * l * n);
15349 const fp16type detC (c * (ejp.asFloat() + flm.asFloat() + hin.asFloat() - hjm.asFloat() - fip.asFloat() - eln.asFloat()));
15354 const fp16type ejo (e * j * o);
15355 const fp16type fkm (f * k * m);
15356 const fp16type gin (g * i * n);
15357 const fp16type gjm (g * j * m);
15358 const fp16type fio (f * i * o);
15359 const fp16type ekn (e * k * n);
15360 const fp16type detD (d * (ejo.asFloat() + fkm.asFloat() + gin.asFloat() - gjm.asFloat() - fio.asFloat() - ekn.asFloat()));
15362 const float r (detA.asFloat() - detB.asFloat() + detC.asFloat() - detD.asFloat());
15363 const fp16type rf16 (r);
15365 out[0] = rf16.bits();
15366 min[0] = getMin(r, getULPs(in));
15367 max[0] = getMax(r, getULPs(in));
15373 template<size_t size>
15374 struct fp16Inverse;
15377 struct fp16Inverse<2> : public fp16MatrixBase
15379 virtual double getULPs (vector<const deFloat16*>& in)
15383 return 128.0; // This is not a precision test. Value is not from spec
15386 deUint32 getComponentValidity ()
15388 return getComponentMatrixValidityMask(2, 2);
15391 template<class fp16type>
15392 bool calc (vector<const deFloat16*>& in, deFloat16* out, double* min, double* max)
15394 const size_t cols = 2;
15395 const size_t rows = 2;
15396 const size_t alignedCols = (cols == 3) ? 4 : cols;
15397 const size_t alignedRows = (rows == 3) ? 4 : rows;
15399 DE_ASSERT(in.size() == 1);
15400 DE_ASSERT(getOutCompCount() == alignedRows * alignedCols);
15401 DE_ASSERT(getArgCompCount(0) == alignedRows * alignedCols);
15402 DE_UNREF(alignedCols);
15406 const float a (fp16type(in[0][getNdx(rows, 0, 0)]).asFloat());
15407 const float b (fp16type(in[0][getNdx(rows, 1, 0)]).asFloat());
15408 const float c (fp16type(in[0][getNdx(rows, 0, 1)]).asFloat());
15409 const float d (fp16type(in[0][getNdx(rows, 1, 1)]).asFloat());
15410 const float ad (a * d);
15411 const fp16type adf16 (ad);
15412 const float bc (b * c);
15413 const fp16type bcf16 (bc);
15414 const float det (adf16.asFloat() - bcf16.asFloat());
15415 const fp16type det16 (det);
15417 out[0] = fp16type( d / det16.asFloat()).bits();
15418 out[1] = fp16type(-c / det16.asFloat()).bits();
15419 out[2] = fp16type(-b / det16.asFloat()).bits();
15420 out[3] = fp16type( a / det16.asFloat()).bits();
15422 for (size_t rowNdx = 0; rowNdx < rows; ++rowNdx)
15423 for (size_t colNdx = 0; colNdx < cols; ++colNdx)
15425 const size_t ndx (colNdx * alignedRows + rowNdx);
15426 const fp16type s (out[ndx]);
15428 min[ndx] = getMin(s.asDouble(), getULPs(in));
15429 max[ndx] = getMax(s.asDouble(), getULPs(in));
15436 inline std::string fp16ToString(deFloat16 val)
15438 return tcu::toHex<4>(val).toString() + " (" + de::floatToString(tcu::Float16(val).asFloat(), 10) + ")";
15441 template <size_t RES_COMPONENTS, size_t ARG0_COMPONENTS, size_t ARG1_COMPONENTS, size_t ARG2_COMPONENTS, class TestedArithmeticFunction>
15442 bool compareFP16ArithmeticFunc (const std::vector<Resource>& inputs, const vector<AllocationSp>& outputAllocs, const std::vector<Resource>& expectedOutputs, TestLog& log)
15444 if (inputs.size() < 1 || inputs.size() > 3 || outputAllocs.size() != 1 || expectedOutputs.size() != 1)
15447 const size_t resultStep = (RES_COMPONENTS == 3) ? 4 : RES_COMPONENTS;
15448 const size_t iterationsCount = expectedOutputs[0].getByteSize() / (sizeof(deFloat16) * resultStep);
15449 const size_t inputsSteps[3] =
15451 (ARG0_COMPONENTS == 3) ? 4 : ARG0_COMPONENTS,
15452 (ARG1_COMPONENTS == 3) ? 4 : ARG1_COMPONENTS,
15453 (ARG2_COMPONENTS == 3) ? 4 : ARG2_COMPONENTS,
15456 DE_ASSERT(expectedOutputs[0].getByteSize() > 0);
15457 DE_ASSERT(expectedOutputs[0].getByteSize() == sizeof(deFloat16) * iterationsCount * resultStep);
15459 for (size_t inputNdx = 0; inputNdx < inputs.size(); ++inputNdx)
15461 DE_ASSERT(inputs[inputNdx].getByteSize() > 0);
15462 DE_ASSERT(inputs[inputNdx].getByteSize() == sizeof(deFloat16) * iterationsCount * inputsSteps[inputNdx]);
15465 const deFloat16* const outputAsFP16 = (const deFloat16*)outputAllocs[0]->getHostPtr();
15466 TestedArithmeticFunction func;
15468 func.setOutCompCount(RES_COMPONENTS);
15469 func.setArgCompCount(0, ARG0_COMPONENTS);
15470 func.setArgCompCount(1, ARG1_COMPONENTS);
15471 func.setArgCompCount(2, ARG2_COMPONENTS);
15473 const bool callOncePerComponent = func.callOncePerComponent();
15474 const deUint32 componentValidityMask = func.getComponentValidity();
15475 const size_t denormModesCount = 2;
15476 const char* denormModes[denormModesCount] = { "keep denormal numbers", "flush to zero" };
15477 const size_t successfulRunsPerComponent = denormModesCount * func.getFlavorCount();
15478 bool success = true;
15479 size_t validatedCount = 0;
15481 vector<deUint8> inputBytes[3];
15483 for (size_t inputNdx = 0; inputNdx < inputs.size(); ++inputNdx)
15484 inputs[inputNdx].getBytes(inputBytes[inputNdx]);
15486 const deFloat16* const inputsAsFP16[3] =
15488 inputs.size() >= 1 ? (const deFloat16*)&inputBytes[0][0] : DE_NULL,
15489 inputs.size() >= 2 ? (const deFloat16*)&inputBytes[1][0] : DE_NULL,
15490 inputs.size() >= 3 ? (const deFloat16*)&inputBytes[2][0] : DE_NULL,
15493 for (size_t idx = 0; idx < iterationsCount; ++idx)
15495 std::vector<size_t> successfulRuns (RES_COMPONENTS, successfulRunsPerComponent);
15496 std::vector<std::string> errors (RES_COMPONENTS);
15497 bool iterationValidated (true);
15499 for (size_t denormNdx = 0; denormNdx < 2; ++denormNdx)
15501 for (size_t flavorNdx = 0; flavorNdx < func.getFlavorCount(); ++flavorNdx)
15503 func.setFlavor(flavorNdx);
15505 const deFloat16* iterationOutputFP16 = &outputAsFP16[idx * resultStep];
15506 vector<deFloat16> iterationCalculatedFP16 (resultStep, 0);
15507 vector<double> iterationEdgeMin (resultStep, 0.0);
15508 vector<double> iterationEdgeMax (resultStep, 0.0);
15509 vector<const deFloat16*> arguments;
15511 for (size_t componentNdx = 0; componentNdx < RES_COMPONENTS; ++componentNdx)
15514 bool reportError = false;
15516 if (callOncePerComponent || componentNdx == 0)
15518 bool funcCallResult;
15522 for (size_t inputNdx = 0; inputNdx < inputs.size(); ++inputNdx)
15523 arguments.push_back(&inputsAsFP16[inputNdx][idx * inputsSteps[inputNdx] + componentNdx]);
15525 if (denormNdx == 0)
15526 funcCallResult = func.template calc<tcu::Float16>(arguments, &iterationCalculatedFP16[componentNdx], &iterationEdgeMin[componentNdx], &iterationEdgeMax[componentNdx]);
15528 funcCallResult = func.template calc<tcu::Float16Denormless>(arguments, &iterationCalculatedFP16[componentNdx], &iterationEdgeMin[componentNdx], &iterationEdgeMax[componentNdx]);
15530 if (!funcCallResult)
15532 iterationValidated = false;
15534 if (callOncePerComponent)
15541 if ((componentValidityMask != 0) && (componentValidityMask & (1<<componentNdx)) == 0)
15544 reportError = !compare16BitFloat(iterationCalculatedFP16[componentNdx], iterationOutputFP16[componentNdx], error);
15548 tcu::Float16 expected (iterationCalculatedFP16[componentNdx]);
15549 tcu::Float16 outputted (iterationOutputFP16[componentNdx]);
15551 if (reportError && expected.isNaN())
15552 reportError = false;
15554 if (reportError && !expected.isNaN() && !outputted.isNaN())
15556 if (reportError && !expected.isInf() && !outputted.isInf())
15559 if (expected.bits() == outputted.bits() + 1 || expected.bits() + 1 == outputted.bits())
15560 reportError = false;
15563 if (reportError && expected.isInf())
15565 // RTZ rounding mode returns +/-65504 instead of Inf on overflow
15566 if (expected.sign() == 1 && outputted.bits() == 0x7bff && iterationEdgeMin[componentNdx] <= std::numeric_limits<double>::max())
15567 reportError = false;
15568 else if (expected.sign() == -1 && outputted.bits() == 0xfbff && iterationEdgeMax[componentNdx] >= -std::numeric_limits<double>::max())
15569 reportError = false;
15574 const double outputtedDouble = outputted.asDouble();
15576 DE_ASSERT(iterationEdgeMin[componentNdx] <= iterationEdgeMax[componentNdx]);
15578 if (de::inRange(outputtedDouble, iterationEdgeMin[componentNdx], iterationEdgeMax[componentNdx]))
15579 reportError = false;
15585 const size_t inputsComps[3] =
15591 string inputsValues ("Inputs:");
15592 string flavorName (func.getFlavorCount() == 1 ? "" : string(" flavor ") + de::toString(flavorNdx) + " (" + func.getCurrentFlavorName() + ")");
15593 std::stringstream errStream;
15595 for (size_t inputNdx = 0; inputNdx < inputs.size(); ++inputNdx)
15597 const size_t inputCompsCount = inputsComps[inputNdx];
15599 inputsValues += " [" + de::toString(inputNdx) + "]=(";
15601 for (size_t compNdx = 0; compNdx < inputCompsCount; ++compNdx)
15603 const deFloat16 inputComponentValue = inputsAsFP16[inputNdx][idx * inputsSteps[inputNdx] + compNdx];
15605 inputsValues += fp16ToString(inputComponentValue) + ((compNdx + 1 == inputCompsCount) ? ")": ", ");
15610 << " iteration " << de::toString(idx)
15611 << " component " << de::toString(componentNdx)
15612 << " denormMode " << de::toString(denormNdx)
15613 << " (" << denormModes[denormNdx] << ")"
15614 << " " << flavorName
15615 << " " << inputsValues
15616 << " outputted:" + fp16ToString(iterationOutputFP16[componentNdx])
15617 << " expected:" + fp16ToString(iterationCalculatedFP16[componentNdx])
15618 << " or in range: [" << iterationEdgeMin[componentNdx] << ", " << iterationEdgeMax[componentNdx] << "]."
15619 << " " << error << "."
15622 errors[componentNdx] += errStream.str();
15624 successfulRuns[componentNdx]--;
15631 for (size_t componentNdx = 0; componentNdx < RES_COMPONENTS; ++componentNdx)
15633 // Check if any component has total failure
15634 if (successfulRuns[componentNdx] == 0)
15636 // Test failed in all denorm modes and all flavors for certain component: dump errors
15637 log << TestLog::Message << errors[componentNdx] << TestLog::EndMessage;
15643 if (iterationValidated)
15647 if (validatedCount < 16)
15648 TCU_THROW(InternalError, "Too few samples has been validated.");
15653 // IEEE-754 floating point numbers:
15654 // +--------+------+----------+-------------+
15655 // | binary | sign | exponent | significand |
15656 // +--------+------+----------+-------------+
15657 // | 16-bit | 1 | 5 | 10 |
15658 // +--------+------+----------+-------------+
15659 // | 32-bit | 1 | 8 | 23 |
15660 // +--------+------+----------+-------------+
15664 // 0 000 00 00 0000 0001 (0x0001: 2e-24: minimum positive denormalized)
15665 // 0 000 00 11 1111 1111 (0x03ff: 2e-14 - 2e-24: maximum positive denormalized)
15666 // 0 000 01 00 0000 0000 (0x0400: 2e-14: minimum positive normalized)
15667 // 0 111 10 11 1111 1111 (0x7bff: 65504: maximum positive normalized)
15669 // 0 000 00 00 0000 0000 (0x0000: +0)
15670 // 0 111 11 00 0000 0000 (0x7c00: +Inf)
15671 // 0 000 00 11 1111 0000 (0x03f0: +Denorm)
15672 // 0 000 01 00 0000 0001 (0x0401: +Norm)
15673 // 0 111 11 00 0000 1111 (0x7c0f: +SNaN)
15674 // 0 111 11 11 1111 0000 (0x7ff0: +QNaN)
15675 // Generate and return 16-bit floats and their corresponding 32-bit values.
15677 // The first 14 number pairs are manually picked, while the rest are randomly generated.
15678 // Expected count to be at least 14 (numPicks).
15679 vector<deFloat16> getFloat16a (de::Random& rnd, deUint32 count)
15681 vector<deFloat16> float16;
15683 float16.reserve(count);
15686 float16.push_back(deUint16(0x0000));
15687 float16.push_back(deUint16(0x8000));
15689 float16.push_back(deUint16(0x7c00));
15690 float16.push_back(deUint16(0xfc00));
15692 float16.push_back(deUint16(0x0401));
15693 float16.push_back(deUint16(0x8401));
15694 // Some normal number
15695 float16.push_back(deUint16(0x14cb));
15696 float16.push_back(deUint16(0x94cb));
15697 // Min/max positive normal
15698 float16.push_back(deUint16(0x0400));
15699 float16.push_back(deUint16(0x7bff));
15700 // Min/max negative normal
15701 float16.push_back(deUint16(0x8400));
15702 float16.push_back(deUint16(0xfbff));
15704 float16.push_back(deUint16(0x4248)); // 3.140625
15705 float16.push_back(deUint16(0xb248)); // -3.140625
15707 float16.push_back(deUint16(0x3e48)); // 1.5703125
15708 float16.push_back(deUint16(0xbe48)); // -1.5703125
15709 float16.push_back(deUint16(0x3c00)); // 1.0
15710 float16.push_back(deUint16(0x3800)); // 0.5
15711 // Some useful constants
15712 float16.push_back(tcu::Float16(-2.5f).bits());
15713 float16.push_back(tcu::Float16(-1.0f).bits());
15714 float16.push_back(tcu::Float16( 0.4f).bits());
15715 float16.push_back(tcu::Float16( 2.5f).bits());
15717 const deUint32 numPicks = static_cast<deUint32>(float16.size());
15719 DE_ASSERT(count >= numPicks);
15722 for (deUint32 numIdx = 0; numIdx < count; ++numIdx)
15724 int sign = (rnd.getUint16() % 2 == 0) ? +1 : -1;
15725 int exponent = (rnd.getUint16() % 29) - 14 + 1;
15726 deUint16 mantissa = static_cast<deUint16>(2 * (rnd.getUint16() % 512));
15728 // Exclude power of -14 to avoid denorms
15729 DE_ASSERT(de::inRange(exponent, -13, 15));
15731 float16.push_back(tcu::Float16::constructBits(sign, exponent, mantissa).bits());
15737 static inline vector<deFloat16> getInputData1 (deUint32 seed, size_t count, size_t argNo)
15741 de::Random rnd(seed);
15743 return getFloat16a(rnd, static_cast<deUint32>(count));
15746 static inline vector<deFloat16> getInputData2 (deUint32 seed, size_t count, size_t argNo)
15748 de::Random rnd (seed);
15749 size_t newCount = static_cast<size_t>(deSqrt(double(count)));
15751 DE_ASSERT(newCount * newCount == count);
15753 vector<deFloat16> float16 = getFloat16a(rnd, static_cast<deUint32>(newCount));
15755 return squarize(float16, static_cast<deUint32>(argNo));
15758 static inline vector<deFloat16> getInputData3 (deUint32 seed, size_t count, size_t argNo)
15760 if (argNo == 0 || argNo == 1)
15761 return getInputData2(seed, count, argNo);
15763 return getInputData1(seed<<argNo, count, argNo);
15766 vector<deFloat16> getInputData (deUint32 seed, size_t count, size_t compCount, size_t stride, size_t argCount, size_t argNo)
15770 vector<deFloat16> result;
15774 case 1:result = getInputData1(seed, count, argNo); break;
15775 case 2:result = getInputData2(seed, count, argNo); break;
15776 case 3:result = getInputData3(seed, count, argNo); break;
15777 default: TCU_THROW(InternalError, "Invalid argument count specified");
15780 if (compCount == 3)
15782 const size_t newCount = (3 * count) / 4;
15783 vector<deFloat16> newResult;
15785 newResult.reserve(result.size());
15787 for (size_t ndx = 0; ndx < newCount; ++ndx)
15789 newResult.push_back(result[ndx]);
15792 newResult.push_back(0);
15795 result = newResult;
15798 DE_ASSERT(result.size() == count);
15803 // Generator for functions requiring data in range [1, inf]
15804 vector<deFloat16> getInputDataAC (deUint32 seed, size_t count, size_t compCount, size_t stride, size_t argCount, size_t argNo)
15806 vector<deFloat16> result;
15808 result = getInputData(seed, count, compCount, stride, argCount, argNo);
15810 // Filter out values below 1.0 from upper half of numbers
15811 for (size_t idx = result.size() / 2; idx < result.size(); ++idx)
15813 const float f = tcu::Float16(result[idx]).asFloat();
15816 result[idx] = tcu::Float16(1.0f - f).bits();
15822 // Generator for functions requiring data in range [-1, 1]
15823 vector<deFloat16> getInputDataA (deUint32 seed, size_t count, size_t compCount, size_t stride, size_t argCount, size_t argNo)
15825 vector<deFloat16> result;
15827 result = getInputData(seed, count, compCount, stride, argCount, argNo);
15829 for (size_t idx = result.size() / 2; idx < result.size(); ++idx)
15831 const float f = tcu::Float16(result[idx]).asFloat();
15833 if (!de::inRange(f, -1.0f, 1.0f))
15834 result[idx] = tcu::Float16(deFloatFrac(f)).bits();
15840 // Generator for functions requiring data in range [-pi, pi]
15841 vector<deFloat16> getInputDataPI (deUint32 seed, size_t count, size_t compCount, size_t stride, size_t argCount, size_t argNo)
15843 vector<deFloat16> result;
15845 result = getInputData(seed, count, compCount, stride, argCount, argNo);
15847 for (size_t idx = result.size() / 2; idx < result.size(); ++idx)
15849 const float f = tcu::Float16(result[idx]).asFloat();
15851 if (!de::inRange(f, -DE_PI, DE_PI))
15852 result[idx] = tcu::Float16(fmodf(f, DE_PI)).bits();
15858 // Generator for functions requiring data in range [0, inf]
15859 vector<deFloat16> getInputDataP (deUint32 seed, size_t count, size_t compCount, size_t stride, size_t argCount, size_t argNo)
15861 vector<deFloat16> result;
15863 result = getInputData(seed, count, compCount, stride, argCount, argNo);
15867 for (size_t idx = result.size() / 2; idx < result.size(); ++idx)
15868 result[idx] &= static_cast<deFloat16>(~0x8000);
15874 vector<deFloat16> getInputDataV (deUint32 seed, size_t count, size_t compCount, size_t stride, size_t argCount, size_t argNo)
15877 DE_UNREF(argCount);
15879 vector<deFloat16> result;
15882 result = getInputData2(seed, count, argNo);
15885 const size_t alignedCount = (compCount == 3) ? 4 : compCount;
15886 const size_t newCountX = static_cast<size_t>(deSqrt(double(count * alignedCount)));
15887 const size_t newCountY = count / newCountX;
15888 de::Random rnd (seed);
15889 vector<deFloat16> float16 = getFloat16a(rnd, static_cast<deUint32>(newCountX));
15891 DE_ASSERT(newCountX * newCountX == alignedCount * count);
15893 for (size_t numIdx = 0; numIdx < newCountX; ++numIdx)
15895 const vector<deFloat16> tmp(newCountY, float16[numIdx]);
15897 result.insert(result.end(), tmp.begin(), tmp.end());
15901 DE_ASSERT(result.size() == count);
15906 vector<deFloat16> getInputDataM (deUint32 seed, size_t count, size_t compCount, size_t stride, size_t argCount, size_t argNo)
15908 DE_UNREF(compCount);
15910 DE_UNREF(argCount);
15912 de::Random rnd (seed << argNo);
15913 vector<deFloat16> result;
15915 result = getFloat16a(rnd, static_cast<deUint32>(count));
15917 DE_ASSERT(result.size() == count);
15922 vector<deFloat16> getInputDataD (deUint32 seed, size_t count, size_t compCount, size_t stride, size_t argCount, size_t argNo)
15924 DE_UNREF(compCount);
15925 DE_UNREF(argCount);
15927 de::Random rnd (seed << argNo);
15928 vector<deFloat16> result;
15930 for (deUint32 numIdx = 0; numIdx < count; ++numIdx)
15932 int num = (rnd.getUint16() % 16) - 8;
15934 result.push_back(tcu::Float16(float(num)).bits());
15937 result[0 * stride] = deUint16(0x7c00); // +Inf
15938 result[1 * stride] = deUint16(0xfc00); // -Inf
15940 DE_ASSERT(result.size() == count);
15945 // Generator for smoothstep function
15946 vector<deFloat16> getInputDataSS (deUint32 seed, size_t count, size_t compCount, size_t stride, size_t argCount, size_t argNo)
15948 vector<deFloat16> result;
15950 result = getInputDataD(seed, count, compCount, stride, argCount, argNo);
15954 for (size_t idx = result.size() / 2; idx < result.size(); ++idx)
15956 const float f = tcu::Float16(result[idx]).asFloat();
15959 result[idx] = tcu::Float16(-f).bits();
15965 for (size_t idx = result.size() / 2; idx < result.size(); ++idx)
15967 const float f = tcu::Float16(result[idx]).asFloat();
15970 result[idx] = tcu::Float16(-f).bits();
15977 // Generates normalized vectors for arguments 0 and 1
15978 vector<deFloat16> getInputDataN (deUint32 seed, size_t count, size_t compCount, size_t stride, size_t argCount, size_t argNo)
15980 DE_UNREF(compCount);
15981 DE_UNREF(argCount);
15983 de::Random rnd (seed << argNo);
15984 vector<deFloat16> result;
15986 if (argNo == 0 || argNo == 1)
15988 // The input parameters for the incident vector I and the surface normal N must already be normalized
15989 for (size_t numIdx = 0; numIdx < count; numIdx += stride)
15991 vector <float> unnormolized;
15994 for (size_t compIdx = 0; compIdx < compCount; ++compIdx)
15995 unnormolized.push_back(float((rnd.getUint16() % 16) - 8));
15997 for (size_t compIdx = 0; compIdx < compCount; ++compIdx)
15998 sum += unnormolized[compIdx] * unnormolized[compIdx];
16000 sum = deFloatSqrt(sum);
16002 unnormolized[0] = sum = 1.0f;
16004 for (size_t compIdx = 0; compIdx < compCount; ++compIdx)
16005 result.push_back(tcu::Float16(unnormolized[compIdx] / sum).bits());
16007 for (size_t compIdx = compCount; compIdx < stride; ++compIdx)
16008 result.push_back(0);
16013 // Input parameter eta
16014 for (deUint32 numIdx = 0; numIdx < count; ++numIdx)
16016 int num = (rnd.getUint16() % 16) - 8;
16018 result.push_back(tcu::Float16(float(num)).bits());
16022 DE_ASSERT(result.size() == count);
16027 // Data generator for complex matrix functions like determinant and inverse
16028 vector<deFloat16> getInputDataC (deUint32 seed, size_t count, size_t compCount, size_t stride, size_t argCount, size_t argNo)
16030 DE_UNREF(compCount);
16032 DE_UNREF(argCount);
16034 de::Random rnd (seed << argNo);
16035 vector<deFloat16> result;
16037 for (deUint32 numIdx = 0; numIdx < count; ++numIdx)
16039 int num = (rnd.getUint16() % 16) - 8;
16041 result.push_back(tcu::Float16(float(num)).bits());
16044 DE_ASSERT(result.size() == count);
16049 struct Math16TestType
16051 const char* typePrefix;
16052 const size_t typeComponents;
16053 const size_t typeArrayStride;
16054 const size_t typeStructStride;
16057 enum Math16DataTypes
16076 struct Math16ArgFragments
16078 const char* bodies;
16079 const char* variables;
16080 const char* decorations;
16081 const char* funcVariables;
16084 typedef vector<deFloat16> Math16GetInputData (deUint32 seed, size_t count, size_t compCount, size_t stride, size_t argCount, size_t argNo);
16086 struct Math16TestFunc
16088 const char* funcName;
16089 const char* funcSuffix;
16090 size_t funcArgsCount;
16095 Math16GetInputData* getInputDataFunc;
16096 VerifyIOFunc verifyFunc;
16099 template<class SpecResource>
16100 void createFloat16ArithmeticFuncTest (tcu::TestContext& testCtx, tcu::TestCaseGroup& testGroup, const size_t testTypeIdx, const Math16TestFunc& testFunc)
16102 const int testSpecificSeed = deStringHash(testGroup.getName());
16103 const int seed = testCtx.getCommandLine().getBaseSeed() ^ testSpecificSeed;
16104 const size_t numDataPointsByAxis = 32;
16105 const size_t numDataPoints = numDataPointsByAxis * numDataPointsByAxis;
16106 const char* componentType = "f16";
16107 const Math16TestType testTypes[MATH16_TYPE_LAST] =
16110 { "", 1, 1 * sizeof(deFloat16), 2 * sizeof(deFloat16) },
16111 { "v2", 2, 2 * sizeof(deFloat16), 2 * sizeof(deFloat16) },
16112 { "v3", 3, 4 * sizeof(deFloat16), 4 * sizeof(deFloat16) },
16113 { "v4", 4, 4 * sizeof(deFloat16), 4 * sizeof(deFloat16) },
16114 { "m2x2", 0, 4 * sizeof(deFloat16), 4 * sizeof(deFloat16) },
16115 { "m2x3", 0, 8 * sizeof(deFloat16), 8 * sizeof(deFloat16) },
16116 { "m2x4", 0, 8 * sizeof(deFloat16), 8 * sizeof(deFloat16) },
16117 { "m3x2", 0, 8 * sizeof(deFloat16), 8 * sizeof(deFloat16) },
16118 { "m3x3", 0, 16 * sizeof(deFloat16), 16 * sizeof(deFloat16) },
16119 { "m3x4", 0, 16 * sizeof(deFloat16), 16 * sizeof(deFloat16) },
16120 { "m4x2", 0, 8 * sizeof(deFloat16), 8 * sizeof(deFloat16) },
16121 { "m4x3", 0, 16 * sizeof(deFloat16), 16 * sizeof(deFloat16) },
16122 { "m4x4", 0, 16 * sizeof(deFloat16), 16 * sizeof(deFloat16) },
16125 DE_ASSERT(testTypeIdx == testTypes[testTypeIdx].typeComponents);
16128 const StringTemplate preMain
16130 " %c_i32_ndp = OpConstant %i32 ${num_data_points}\n"
16132 " %f16 = OpTypeFloat 16\n"
16133 " %v2f16 = OpTypeVector %f16 2\n"
16134 " %v3f16 = OpTypeVector %f16 3\n"
16135 " %v4f16 = OpTypeVector %f16 4\n"
16136 " %m2x2f16 = OpTypeMatrix %v2f16 2\n"
16137 " %m2x3f16 = OpTypeMatrix %v3f16 2\n"
16138 " %m2x4f16 = OpTypeMatrix %v4f16 2\n"
16139 " %m3x2f16 = OpTypeMatrix %v2f16 3\n"
16140 " %m3x3f16 = OpTypeMatrix %v3f16 3\n"
16141 " %m3x4f16 = OpTypeMatrix %v4f16 3\n"
16142 " %m4x2f16 = OpTypeMatrix %v2f16 4\n"
16143 " %m4x3f16 = OpTypeMatrix %v3f16 4\n"
16144 " %m4x4f16 = OpTypeMatrix %v4f16 4\n"
16146 " %up_f16 = OpTypePointer Uniform %f16 \n"
16147 " %up_v2f16 = OpTypePointer Uniform %v2f16 \n"
16148 " %up_v3f16 = OpTypePointer Uniform %v3f16 \n"
16149 " %up_v4f16 = OpTypePointer Uniform %v4f16 \n"
16150 " %up_m2x2f16 = OpTypePointer Uniform %m2x2f16\n"
16151 " %up_m2x3f16 = OpTypePointer Uniform %m2x3f16\n"
16152 " %up_m2x4f16 = OpTypePointer Uniform %m2x4f16\n"
16153 " %up_m3x2f16 = OpTypePointer Uniform %m3x2f16\n"
16154 " %up_m3x3f16 = OpTypePointer Uniform %m3x3f16\n"
16155 " %up_m3x4f16 = OpTypePointer Uniform %m3x4f16\n"
16156 " %up_m4x2f16 = OpTypePointer Uniform %m4x2f16\n"
16157 " %up_m4x3f16 = OpTypePointer Uniform %m4x3f16\n"
16158 " %up_m4x4f16 = OpTypePointer Uniform %m4x4f16\n"
16160 " %ra_f16 = OpTypeArray %f16 %c_i32_ndp\n"
16161 " %ra_v2f16 = OpTypeArray %v2f16 %c_i32_ndp\n"
16162 " %ra_v3f16 = OpTypeArray %v3f16 %c_i32_ndp\n"
16163 " %ra_v4f16 = OpTypeArray %v4f16 %c_i32_ndp\n"
16164 " %ra_m2x2f16 = OpTypeArray %m2x2f16 %c_i32_ndp\n"
16165 " %ra_m2x3f16 = OpTypeArray %m2x3f16 %c_i32_ndp\n"
16166 " %ra_m2x4f16 = OpTypeArray %m2x4f16 %c_i32_ndp\n"
16167 " %ra_m3x2f16 = OpTypeArray %m3x2f16 %c_i32_ndp\n"
16168 " %ra_m3x3f16 = OpTypeArray %m3x3f16 %c_i32_ndp\n"
16169 " %ra_m3x4f16 = OpTypeArray %m3x4f16 %c_i32_ndp\n"
16170 " %ra_m4x2f16 = OpTypeArray %m4x2f16 %c_i32_ndp\n"
16171 " %ra_m4x3f16 = OpTypeArray %m4x3f16 %c_i32_ndp\n"
16172 " %ra_m4x4f16 = OpTypeArray %m4x4f16 %c_i32_ndp\n"
16174 " %SSBO_f16 = OpTypeStruct %ra_f16 \n"
16175 " %SSBO_v2f16 = OpTypeStruct %ra_v2f16 \n"
16176 " %SSBO_v3f16 = OpTypeStruct %ra_v3f16 \n"
16177 " %SSBO_v4f16 = OpTypeStruct %ra_v4f16 \n"
16178 " %SSBO_m2x2f16 = OpTypeStruct %ra_m2x2f16\n"
16179 " %SSBO_m2x3f16 = OpTypeStruct %ra_m2x3f16\n"
16180 " %SSBO_m2x4f16 = OpTypeStruct %ra_m2x4f16\n"
16181 " %SSBO_m3x2f16 = OpTypeStruct %ra_m3x2f16\n"
16182 " %SSBO_m3x3f16 = OpTypeStruct %ra_m3x3f16\n"
16183 " %SSBO_m3x4f16 = OpTypeStruct %ra_m3x4f16\n"
16184 " %SSBO_m4x2f16 = OpTypeStruct %ra_m4x2f16\n"
16185 " %SSBO_m4x3f16 = OpTypeStruct %ra_m4x3f16\n"
16186 " %SSBO_m4x4f16 = OpTypeStruct %ra_m4x4f16\n"
16188 "%up_SSBO_f16 = OpTypePointer Uniform %SSBO_f16 \n"
16189 "%up_SSBO_v2f16 = OpTypePointer Uniform %SSBO_v2f16 \n"
16190 "%up_SSBO_v3f16 = OpTypePointer Uniform %SSBO_v3f16 \n"
16191 "%up_SSBO_v4f16 = OpTypePointer Uniform %SSBO_v4f16 \n"
16192 "%up_SSBO_m2x2f16 = OpTypePointer Uniform %SSBO_m2x2f16\n"
16193 "%up_SSBO_m2x3f16 = OpTypePointer Uniform %SSBO_m2x3f16\n"
16194 "%up_SSBO_m2x4f16 = OpTypePointer Uniform %SSBO_m2x4f16\n"
16195 "%up_SSBO_m3x2f16 = OpTypePointer Uniform %SSBO_m3x2f16\n"
16196 "%up_SSBO_m3x3f16 = OpTypePointer Uniform %SSBO_m3x3f16\n"
16197 "%up_SSBO_m3x4f16 = OpTypePointer Uniform %SSBO_m3x4f16\n"
16198 "%up_SSBO_m4x2f16 = OpTypePointer Uniform %SSBO_m4x2f16\n"
16199 "%up_SSBO_m4x3f16 = OpTypePointer Uniform %SSBO_m4x3f16\n"
16200 "%up_SSBO_m4x4f16 = OpTypePointer Uniform %SSBO_m4x4f16\n"
16202 " %fp_v2i32 = OpTypePointer Function %v2i32\n"
16203 " %fp_v3i32 = OpTypePointer Function %v3i32\n"
16204 " %fp_v4i32 = OpTypePointer Function %v4i32\n"
16208 const StringTemplate decoration
16210 "OpDecorate %ra_f16 ArrayStride 2 \n"
16211 "OpDecorate %ra_v2f16 ArrayStride 4 \n"
16212 "OpDecorate %ra_v3f16 ArrayStride 8 \n"
16213 "OpDecorate %ra_v4f16 ArrayStride 8 \n"
16214 "OpDecorate %ra_m2x2f16 ArrayStride 8 \n"
16215 "OpDecorate %ra_m2x3f16 ArrayStride 16\n"
16216 "OpDecorate %ra_m2x4f16 ArrayStride 16\n"
16217 "OpDecorate %ra_m3x2f16 ArrayStride 16\n"
16218 "OpDecorate %ra_m3x3f16 ArrayStride 32\n"
16219 "OpDecorate %ra_m3x4f16 ArrayStride 32\n"
16220 "OpDecorate %ra_m4x2f16 ArrayStride 16\n"
16221 "OpDecorate %ra_m4x3f16 ArrayStride 32\n"
16222 "OpDecorate %ra_m4x4f16 ArrayStride 32\n"
16224 "OpMemberDecorate %SSBO_f16 0 Offset 0\n"
16225 "OpMemberDecorate %SSBO_v2f16 0 Offset 0\n"
16226 "OpMemberDecorate %SSBO_v3f16 0 Offset 0\n"
16227 "OpMemberDecorate %SSBO_v4f16 0 Offset 0\n"
16228 "OpMemberDecorate %SSBO_m2x2f16 0 Offset 0\n"
16229 "OpMemberDecorate %SSBO_m2x3f16 0 Offset 0\n"
16230 "OpMemberDecorate %SSBO_m2x4f16 0 Offset 0\n"
16231 "OpMemberDecorate %SSBO_m3x2f16 0 Offset 0\n"
16232 "OpMemberDecorate %SSBO_m3x3f16 0 Offset 0\n"
16233 "OpMemberDecorate %SSBO_m3x4f16 0 Offset 0\n"
16234 "OpMemberDecorate %SSBO_m4x2f16 0 Offset 0\n"
16235 "OpMemberDecorate %SSBO_m4x3f16 0 Offset 0\n"
16236 "OpMemberDecorate %SSBO_m4x4f16 0 Offset 0\n"
16238 "OpDecorate %SSBO_f16 BufferBlock\n"
16239 "OpDecorate %SSBO_v2f16 BufferBlock\n"
16240 "OpDecorate %SSBO_v3f16 BufferBlock\n"
16241 "OpDecorate %SSBO_v4f16 BufferBlock\n"
16242 "OpDecorate %SSBO_m2x2f16 BufferBlock\n"
16243 "OpDecorate %SSBO_m2x3f16 BufferBlock\n"
16244 "OpDecorate %SSBO_m2x4f16 BufferBlock\n"
16245 "OpDecorate %SSBO_m3x2f16 BufferBlock\n"
16246 "OpDecorate %SSBO_m3x3f16 BufferBlock\n"
16247 "OpDecorate %SSBO_m3x4f16 BufferBlock\n"
16248 "OpDecorate %SSBO_m4x2f16 BufferBlock\n"
16249 "OpDecorate %SSBO_m4x3f16 BufferBlock\n"
16250 "OpDecorate %SSBO_m4x4f16 BufferBlock\n"
16252 "OpMemberDecorate %SSBO_m2x2f16 0 ColMajor\n"
16253 "OpMemberDecorate %SSBO_m2x3f16 0 ColMajor\n"
16254 "OpMemberDecorate %SSBO_m2x4f16 0 ColMajor\n"
16255 "OpMemberDecorate %SSBO_m3x2f16 0 ColMajor\n"
16256 "OpMemberDecorate %SSBO_m3x3f16 0 ColMajor\n"
16257 "OpMemberDecorate %SSBO_m3x4f16 0 ColMajor\n"
16258 "OpMemberDecorate %SSBO_m4x2f16 0 ColMajor\n"
16259 "OpMemberDecorate %SSBO_m4x3f16 0 ColMajor\n"
16260 "OpMemberDecorate %SSBO_m4x4f16 0 ColMajor\n"
16262 "OpMemberDecorate %SSBO_m2x2f16 0 MatrixStride 4\n"
16263 "OpMemberDecorate %SSBO_m2x3f16 0 MatrixStride 8\n"
16264 "OpMemberDecorate %SSBO_m2x4f16 0 MatrixStride 8\n"
16265 "OpMemberDecorate %SSBO_m3x2f16 0 MatrixStride 4\n"
16266 "OpMemberDecorate %SSBO_m3x3f16 0 MatrixStride 8\n"
16267 "OpMemberDecorate %SSBO_m3x4f16 0 MatrixStride 8\n"
16268 "OpMemberDecorate %SSBO_m4x2f16 0 MatrixStride 4\n"
16269 "OpMemberDecorate %SSBO_m4x3f16 0 MatrixStride 8\n"
16270 "OpMemberDecorate %SSBO_m4x4f16 0 MatrixStride 8\n"
16272 "${arg_decorations}"
16275 const StringTemplate testFun
16277 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
16278 " %param = OpFunctionParameter %v4f32\n"
16279 " %entry = OpLabel\n"
16281 " %i = OpVariable %fp_i32 Function\n"
16282 "${arg_infunc_vars}"
16283 " OpStore %i %c_i32_0\n"
16284 " OpBranch %loop\n"
16286 " %loop = OpLabel\n"
16287 " %i_cmp = OpLoad %i32 %i\n"
16288 " %lt = OpSLessThan %bool %i_cmp %c_i32_ndp\n"
16289 " OpLoopMerge %merge %next None\n"
16290 " OpBranchConditional %lt %write %merge\n"
16292 " %write = OpLabel\n"
16293 " %ndx = OpLoad %i32 %i\n"
16297 " OpBranch %next\n"
16299 " %next = OpLabel\n"
16300 " %i_cur = OpLoad %i32 %i\n"
16301 " %i_new = OpIAdd %i32 %i_cur %c_i32_1\n"
16302 " OpStore %i %i_new\n"
16303 " OpBranch %loop\n"
16305 " %merge = OpLabel\n"
16306 " OpReturnValue %param\n"
16310 const Math16ArgFragments argFragment1 =
16312 " %src0 = OpAccessChain %up_${t0} %ssbo_src0 %c_i32_0 %ndx\n"
16313 " %val_src0 = OpLoad %${t0} %src0\n"
16314 " %val_dst = ${op} %${tr} ${ext_inst} %val_src0\n"
16315 " %dst = OpAccessChain %up_${tr} %ssbo_dst %c_i32_0 %ndx\n"
16316 " OpStore %dst %val_dst\n",
16322 const Math16ArgFragments argFragment2 =
16324 " %src0 = OpAccessChain %up_${t0} %ssbo_src0 %c_i32_0 %ndx\n"
16325 " %val_src0 = OpLoad %${t0} %src0\n"
16326 " %src1 = OpAccessChain %up_${t1} %ssbo_src1 %c_i32_0 %ndx\n"
16327 " %val_src1 = OpLoad %${t1} %src1\n"
16328 " %val_dst = ${op} %${tr} ${ext_inst} %val_src0 %val_src1\n"
16329 " %dst = OpAccessChain %up_${tr} %ssbo_dst %c_i32_0 %ndx\n"
16330 " OpStore %dst %val_dst\n",
16336 const Math16ArgFragments argFragment3 =
16338 " %src0 = OpAccessChain %up_${t0} %ssbo_src0 %c_i32_0 %ndx\n"
16339 " %val_src0 = OpLoad %${t0} %src0\n"
16340 " %src1 = OpAccessChain %up_${t1} %ssbo_src1 %c_i32_0 %ndx\n"
16341 " %val_src1 = OpLoad %${t1} %src1\n"
16342 " %src2 = OpAccessChain %up_${t2} %ssbo_src2 %c_i32_0 %ndx\n"
16343 " %val_src2 = OpLoad %${t2} %src2\n"
16344 " %val_dst = ${op} %${tr} ${ext_inst} %val_src0 %val_src1 %val_src2\n"
16345 " %dst = OpAccessChain %up_${tr} %ssbo_dst %c_i32_0 %ndx\n"
16346 " OpStore %dst %val_dst\n",
16352 const Math16ArgFragments argFragmentLdExp =
16354 " %src0 = OpAccessChain %up_${t0} %ssbo_src0 %c_i32_0 %ndx\n"
16355 " %val_src0 = OpLoad %${t0} %src0\n"
16356 " %src1 = OpAccessChain %up_${t1} %ssbo_src1 %c_i32_0 %ndx\n"
16357 " %val_src1 = OpLoad %${t1} %src1\n"
16358 "%val_src1i = OpConvertFToS %${dr}i32 %val_src1\n"
16359 " %val_dst = ${op} %${tr} ${ext_inst} %val_src0 %val_src1i\n"
16360 " %dst = OpAccessChain %up_${tr} %ssbo_dst %c_i32_0 %ndx\n"
16361 " OpStore %dst %val_dst\n",
16370 const Math16ArgFragments argFragmentModfFrac =
16372 " %src0 = OpAccessChain %up_${t0} %ssbo_src0 %c_i32_0 %ndx\n"
16373 " %val_src0 = OpLoad %${t0} %src0\n"
16374 " %val_dst = ${op} %${tr} ${ext_inst} %val_src0 %tmp\n"
16375 " %dst = OpAccessChain %up_${tr} %ssbo_dst %c_i32_0 %ndx\n"
16376 " OpStore %dst %val_dst\n",
16378 " %fp_tmp = OpTypePointer Function %${tr}\n",
16382 " %tmp = OpVariable %fp_tmp Function\n",
16385 const Math16ArgFragments argFragmentModfInt =
16387 " %src0 = OpAccessChain %up_${t0} %ssbo_src0 %c_i32_0 %ndx\n"
16388 " %val_src0 = OpLoad %${t0} %src0\n"
16389 "%val_dummy = ${op} %${tr} ${ext_inst} %val_src0 %tmp\n"
16390 " %tmp0 = OpAccessChain %fp_tmp %tmp\n"
16391 " %val_dst = OpLoad %${tr} %tmp0\n"
16392 " %dst = OpAccessChain %up_${tr} %ssbo_dst %c_i32_0 %ndx\n"
16393 " OpStore %dst %val_dst\n",
16395 " %fp_tmp = OpTypePointer Function %${tr}\n",
16399 " %tmp = OpVariable %fp_tmp Function\n",
16402 const Math16ArgFragments argFragmentModfStruct =
16404 " %src0 = OpAccessChain %up_${t0} %ssbo_src0 %c_i32_0 %ndx\n"
16405 " %val_src0 = OpLoad %${t0} %src0\n"
16406 " %val_tmp = ${op} %st_tmp ${ext_inst} %val_src0\n"
16407 "%tmp_ptr_s = OpAccessChain %fp_tmp %tmp\n"
16408 " OpStore %tmp_ptr_s %val_tmp\n"
16409 "%tmp_ptr_l = OpAccessChain %fp_${tr} %tmp %c_${struct_member}\n"
16410 " %val_dst = OpLoad %${tr} %tmp_ptr_l\n"
16411 " %dst = OpAccessChain %up_${tr} %ssbo_dst %c_i32_0 %ndx\n"
16412 " OpStore %dst %val_dst\n",
16414 " %fp_${tr} = OpTypePointer Function %${tr}\n"
16415 " %st_tmp = OpTypeStruct %${tr} %${tr}\n"
16416 " %fp_tmp = OpTypePointer Function %st_tmp\n"
16417 " %c_frac = OpConstant %i32 0\n"
16418 " %c_int = OpConstant %i32 1\n",
16420 "OpMemberDecorate %st_tmp 0 Offset 0\n"
16421 "OpMemberDecorate %st_tmp 1 Offset ${struct_stride}\n",
16423 " %tmp = OpVariable %fp_tmp Function\n",
16426 const Math16ArgFragments argFragmentFrexpStructS =
16428 " %src0 = OpAccessChain %up_${t0} %ssbo_src0 %c_i32_0 %ndx\n"
16429 " %val_src0 = OpLoad %${t0} %src0\n"
16430 " %val_tmp = ${op} %st_tmp ${ext_inst} %val_src0\n"
16431 "%tmp_ptr_s = OpAccessChain %fp_tmp %tmp\n"
16432 " OpStore %tmp_ptr_s %val_tmp\n"
16433 "%tmp_ptr_l = OpAccessChain %fp_${tr} %tmp %c_i32_0\n"
16434 " %val_dst = OpLoad %${tr} %tmp_ptr_l\n"
16435 " %dst = OpAccessChain %up_${tr} %ssbo_dst %c_i32_0 %ndx\n"
16436 " OpStore %dst %val_dst\n",
16438 " %fp_${tr} = OpTypePointer Function %${tr}\n"
16439 " %st_tmp = OpTypeStruct %${tr} %${dr}i32\n"
16440 " %fp_tmp = OpTypePointer Function %st_tmp\n",
16442 "OpMemberDecorate %st_tmp 0 Offset 0\n"
16443 "OpMemberDecorate %st_tmp 1 Offset ${struct_stride}\n",
16445 " %tmp = OpVariable %fp_tmp Function\n",
16448 const Math16ArgFragments argFragmentFrexpStructE =
16450 " %src0 = OpAccessChain %up_${t0} %ssbo_src0 %c_i32_0 %ndx\n"
16451 " %val_src0 = OpLoad %${t0} %src0\n"
16452 " %val_tmp = ${op} %st_tmp ${ext_inst} %val_src0\n"
16453 "%tmp_ptr_s = OpAccessChain %fp_tmp %tmp\n"
16454 " OpStore %tmp_ptr_s %val_tmp\n"
16455 "%tmp_ptr_l = OpAccessChain %fp_${dr}i32 %tmp %c_i32_1\n"
16456 "%val_dst_i = OpLoad %${dr}i32 %tmp_ptr_l\n"
16457 " %val_dst = OpConvertSToF %${tr} %val_dst_i\n"
16458 " %dst = OpAccessChain %up_${tr} %ssbo_dst %c_i32_0 %ndx\n"
16459 " OpStore %dst %val_dst\n",
16461 " %st_tmp = OpTypeStruct %${tr} %${dr}i32\n"
16462 " %fp_tmp = OpTypePointer Function %st_tmp\n",
16464 "OpMemberDecorate %st_tmp 0 Offset 0\n"
16465 "OpMemberDecorate %st_tmp 1 Offset ${struct_stride}\n",
16467 " %tmp = OpVariable %fp_tmp Function\n",
16470 const Math16ArgFragments argFragmentFrexpS =
16472 " %src0 = OpAccessChain %up_${t0} %ssbo_src0 %c_i32_0 %ndx\n"
16473 " %val_src0 = OpLoad %${t0} %src0\n"
16474 " %out_exp = OpAccessChain %fp_${dr}i32 %tmp\n"
16475 " %val_dst = ${op} %${tr} ${ext_inst} %val_src0 %out_exp\n"
16476 " %dst = OpAccessChain %up_${tr} %ssbo_dst %c_i32_0 %ndx\n"
16477 " OpStore %dst %val_dst\n",
16483 " %tmp = OpVariable %fp_${dr}i32 Function\n",
16486 const Math16ArgFragments argFragmentFrexpE =
16488 " %src0 = OpAccessChain %up_${t0} %ssbo_src0 %c_i32_0 %ndx\n"
16489 " %val_src0 = OpLoad %${t0} %src0\n"
16490 " %out_exp = OpAccessChain %fp_${dr}i32 %tmp\n"
16491 "%val_dummy = ${op} %${tr} ${ext_inst} %val_src0 %out_exp\n"
16492 "%val_dst_i = OpLoad %${dr}i32 %out_exp\n"
16493 " %val_dst = OpConvertSToF %${tr} %val_dst_i\n"
16494 " %dst = OpAccessChain %up_${tr} %ssbo_dst %c_i32_0 %ndx\n"
16495 " OpStore %dst %val_dst\n",
16501 " %tmp = OpVariable %fp_${dr}i32 Function\n",
16504 const Math16TestType& testType = testTypes[testTypeIdx];
16505 const string funcNameString = string(testFunc.funcName) + string(testFunc.funcSuffix);
16506 const string testName = de::toLower(funcNameString);
16507 const Math16ArgFragments* argFragments = DE_NULL;
16508 const size_t typeStructStride = testType.typeStructStride;
16509 const bool extInst = !(testFunc.funcName[0] == 'O' && testFunc.funcName[1] == 'p');
16510 const size_t numFloatsPerArg0Type = testTypes[testFunc.typeArg0].typeArrayStride / sizeof(deFloat16);
16511 const size_t iterations = numDataPoints / numFloatsPerArg0Type;
16512 const size_t numFloatsPerResultType = testTypes[testFunc.typeResult].typeArrayStride / sizeof(deFloat16);
16513 const vector<deFloat16> float16DummyOutput (iterations * numFloatsPerResultType, 0);
16514 VulkanFeatures features;
16515 SpecResource specResource;
16516 map<string, string> specs;
16517 map<string, string> fragments;
16518 vector<string> extensions;
16520 string funcVariables;
16522 string declarations;
16523 string decorations;
16525 switch (testFunc.funcArgsCount)
16529 argFragments = &argFragment1;
16531 if (funcNameString == "ModfFrac") argFragments = &argFragmentModfFrac;
16532 if (funcNameString == "ModfInt") argFragments = &argFragmentModfInt;
16533 if (funcNameString == "ModfStructFrac") argFragments = &argFragmentModfStruct;
16534 if (funcNameString == "ModfStructInt") argFragments = &argFragmentModfStruct;
16535 if (funcNameString == "FrexpS") argFragments = &argFragmentFrexpS;
16536 if (funcNameString == "FrexpE") argFragments = &argFragmentFrexpE;
16537 if (funcNameString == "FrexpStructS") argFragments = &argFragmentFrexpStructS;
16538 if (funcNameString == "FrexpStructE") argFragments = &argFragmentFrexpStructE;
16544 argFragments = &argFragment2;
16546 if (funcNameString == "Ldexp") argFragments = &argFragmentLdExp;
16552 argFragments = &argFragment3;
16558 TCU_THROW(InternalError, "Invalid number of arguments");
16562 if (testFunc.funcArgsCount == 1)
16565 " %ssbo_src0 = OpVariable %up_SSBO_${t0} Uniform\n"
16566 " %ssbo_dst = OpVariable %up_SSBO_${tr} Uniform\n";
16569 "OpDecorate %ssbo_src0 DescriptorSet 0\n"
16570 "OpDecorate %ssbo_src0 Binding 0\n"
16571 "OpDecorate %ssbo_dst DescriptorSet 0\n"
16572 "OpDecorate %ssbo_dst Binding 1\n";
16574 else if (testFunc.funcArgsCount == 2)
16577 " %ssbo_src0 = OpVariable %up_SSBO_${t0} Uniform\n"
16578 " %ssbo_src1 = OpVariable %up_SSBO_${t1} Uniform\n"
16579 " %ssbo_dst = OpVariable %up_SSBO_${tr} Uniform\n";
16582 "OpDecorate %ssbo_src0 DescriptorSet 0\n"
16583 "OpDecorate %ssbo_src0 Binding 0\n"
16584 "OpDecorate %ssbo_src1 DescriptorSet 0\n"
16585 "OpDecorate %ssbo_src1 Binding 1\n"
16586 "OpDecorate %ssbo_dst DescriptorSet 0\n"
16587 "OpDecorate %ssbo_dst Binding 2\n";
16589 else if (testFunc.funcArgsCount == 3)
16592 " %ssbo_src0 = OpVariable %up_SSBO_${t0} Uniform\n"
16593 " %ssbo_src1 = OpVariable %up_SSBO_${t1} Uniform\n"
16594 " %ssbo_src2 = OpVariable %up_SSBO_${t2} Uniform\n"
16595 " %ssbo_dst = OpVariable %up_SSBO_${tr} Uniform\n";
16598 "OpDecorate %ssbo_src0 DescriptorSet 0\n"
16599 "OpDecorate %ssbo_src0 Binding 0\n"
16600 "OpDecorate %ssbo_src1 DescriptorSet 0\n"
16601 "OpDecorate %ssbo_src1 Binding 1\n"
16602 "OpDecorate %ssbo_src2 DescriptorSet 0\n"
16603 "OpDecorate %ssbo_src2 Binding 2\n"
16604 "OpDecorate %ssbo_dst DescriptorSet 0\n"
16605 "OpDecorate %ssbo_dst Binding 3\n";
16609 TCU_THROW(InternalError, "Invalid number of function arguments");
16612 variables += argFragments->variables;
16613 decorations += argFragments->decorations;
16615 specs["dr"] = testTypes[testFunc.typeResult].typePrefix;
16616 specs["d0"] = testTypes[testFunc.typeArg0].typePrefix;
16617 specs["d1"] = testTypes[testFunc.typeArg1].typePrefix;
16618 specs["d2"] = testTypes[testFunc.typeArg2].typePrefix;
16619 specs["tr"] = string(testTypes[testFunc.typeResult].typePrefix) + componentType;
16620 specs["t0"] = string(testTypes[testFunc.typeArg0].typePrefix) + componentType;
16621 specs["t1"] = string(testTypes[testFunc.typeArg1].typePrefix) + componentType;
16622 specs["t2"] = string(testTypes[testFunc.typeArg2].typePrefix) + componentType;
16623 specs["struct_stride"] = de::toString(typeStructStride);
16624 specs["op"] = extInst ? "OpExtInst" : testFunc.funcName;
16625 specs["ext_inst"] = extInst ? string("%ext_import ") + testFunc.funcName : "";
16626 specs["struct_member"] = de::toLower(testFunc.funcSuffix);
16628 variables = StringTemplate(variables).specialize(specs);
16629 decorations = StringTemplate(decorations).specialize(specs);
16630 funcVariables = StringTemplate(argFragments->funcVariables).specialize(specs);
16631 funcCall = StringTemplate(argFragments->bodies).specialize(specs);
16633 specs["num_data_points"] = de::toString(iterations);
16634 specs["arg_vars"] = variables;
16635 specs["arg_decorations"] = decorations;
16636 specs["arg_infunc_vars"] = funcVariables;
16637 specs["arg_func_call"] = funcCall;
16639 fragments["extension"] = "OpExtension \"SPV_KHR_16bit_storage\"\n%ext_import = OpExtInstImport \"GLSL.std.450\"";
16640 fragments["capability"] = "OpCapability Matrix\nOpCapability StorageUniformBufferBlock16";
16641 fragments["decoration"] = decoration.specialize(specs);
16642 fragments["pre_main"] = preMain.specialize(specs);
16643 fragments["testfun"] = testFun.specialize(specs);
16645 for (size_t inputArgNdx = 0; inputArgNdx < testFunc.funcArgsCount; ++inputArgNdx)
16647 const size_t numFloatsPerItem = (inputArgNdx == 0) ? testTypes[testFunc.typeArg0].typeArrayStride / sizeof(deFloat16)
16648 : (inputArgNdx == 1) ? testTypes[testFunc.typeArg1].typeArrayStride / sizeof(deFloat16)
16649 : (inputArgNdx == 2) ? testTypes[testFunc.typeArg2].typeArrayStride / sizeof(deFloat16)
16651 const vector<deFloat16> inputData = testFunc.getInputDataFunc(seed, numFloatsPerItem * iterations, testTypeIdx, numFloatsPerItem, testFunc.funcArgsCount, inputArgNdx);
16653 specResource.inputs.push_back(Resource(BufferSp(new Float16Buffer(inputData)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
16656 specResource.outputs.push_back(Resource(BufferSp(new Float16Buffer(float16DummyOutput)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
16657 specResource.verifyIO = testFunc.verifyFunc;
16659 extensions.push_back("VK_KHR_16bit_storage");
16660 extensions.push_back("VK_KHR_shader_float16_int8");
16662 features.ext16BitStorage = EXT16BITSTORAGEFEATURES_UNIFORM_BUFFER_BLOCK;
16663 features.extFloat16Int8 = EXTFLOAT16INT8FEATURES_FLOAT16;
16665 finalizeTestsCreation(specResource, fragments, testCtx, testGroup, testName, features, extensions, IVec3(1, 1, 1));
16668 template<size_t C, class SpecResource>
16669 tcu::TestCaseGroup* createFloat16ArithmeticSet (tcu::TestContext& testCtx)
16671 DE_STATIC_ASSERT(C >= 1 && C <= 4);
16673 const std::string testGroupName (string("arithmetic_") + de::toString(C));
16674 de::MovePtr<tcu::TestCaseGroup> testGroup (new tcu::TestCaseGroup(testCtx, testGroupName.c_str(), "Float 16 arithmetic and related tests"));
16675 const Math16TestFunc testFuncs[] =
16677 { "OpFNegate", "", 1, C, C, 0, 0, &getInputData, compareFP16ArithmeticFunc< C, C, 0, 0, fp16OpFNegate> },
16678 { "Round", "", 1, C, C, 0, 0, &getInputData, compareFP16ArithmeticFunc< C, C, 0, 0, fp16Round> },
16679 { "RoundEven", "", 1, C, C, 0, 0, &getInputData, compareFP16ArithmeticFunc< C, C, 0, 0, fp16RoundEven> },
16680 { "Trunc", "", 1, C, C, 0, 0, &getInputData, compareFP16ArithmeticFunc< C, C, 0, 0, fp16Trunc> },
16681 { "FAbs", "", 1, C, C, 0, 0, &getInputData, compareFP16ArithmeticFunc< C, C, 0, 0, fp16FAbs> },
16682 { "FSign", "", 1, C, C, 0, 0, &getInputData, compareFP16ArithmeticFunc< C, C, 0, 0, fp16FSign> },
16683 { "Floor", "", 1, C, C, 0, 0, &getInputData, compareFP16ArithmeticFunc< C, C, 0, 0, fp16Floor> },
16684 { "Ceil", "", 1, C, C, 0, 0, &getInputData, compareFP16ArithmeticFunc< C, C, 0, 0, fp16Ceil> },
16685 { "Fract", "", 1, C, C, 0, 0, &getInputData, compareFP16ArithmeticFunc< C, C, 0, 0, fp16Fract> },
16686 { "Radians", "", 1, C, C, 0, 0, &getInputData, compareFP16ArithmeticFunc< C, C, 0, 0, fp16Radians> },
16687 { "Degrees", "", 1, C, C, 0, 0, &getInputData, compareFP16ArithmeticFunc< C, C, 0, 0, fp16Degrees> },
16688 { "Sin", "", 1, C, C, 0, 0, &getInputDataPI, compareFP16ArithmeticFunc< C, C, 0, 0, fp16Sin> },
16689 { "Cos", "", 1, C, C, 0, 0, &getInputDataPI, compareFP16ArithmeticFunc< C, C, 0, 0, fp16Cos> },
16690 { "Tan", "", 1, C, C, 0, 0, &getInputDataPI, compareFP16ArithmeticFunc< C, C, 0, 0, fp16Tan> },
16691 { "Asin", "", 1, C, C, 0, 0, &getInputDataA, compareFP16ArithmeticFunc< C, C, 0, 0, fp16Asin> },
16692 { "Acos", "", 1, C, C, 0, 0, &getInputDataA, compareFP16ArithmeticFunc< C, C, 0, 0, fp16Acos> },
16693 { "Atan", "", 1, C, C, 0, 0, &getInputData, compareFP16ArithmeticFunc< C, C, 0, 0, fp16Atan> },
16694 { "Sinh", "", 1, C, C, 0, 0, &getInputData, compareFP16ArithmeticFunc< C, C, 0, 0, fp16Sinh> },
16695 { "Cosh", "", 1, C, C, 0, 0, &getInputData, compareFP16ArithmeticFunc< C, C, 0, 0, fp16Cosh> },
16696 { "Tanh", "", 1, C, C, 0, 0, &getInputData, compareFP16ArithmeticFunc< C, C, 0, 0, fp16Tanh> },
16697 { "Asinh", "", 1, C, C, 0, 0, &getInputData, compareFP16ArithmeticFunc< C, C, 0, 0, fp16Asinh> },
16698 { "Acosh", "", 1, C, C, 0, 0, &getInputDataAC, compareFP16ArithmeticFunc< C, C, 0, 0, fp16Acosh> },
16699 { "Atanh", "", 1, C, C, 0, 0, &getInputDataA, compareFP16ArithmeticFunc< C, C, 0, 0, fp16Atanh> },
16700 { "Exp", "", 1, C, C, 0, 0, &getInputData, compareFP16ArithmeticFunc< C, C, 0, 0, fp16Exp> },
16701 { "Log", "", 1, C, C, 0, 0, &getInputDataP, compareFP16ArithmeticFunc< C, C, 0, 0, fp16Log> },
16702 { "Exp2", "", 1, C, C, 0, 0, &getInputData, compareFP16ArithmeticFunc< C, C, 0, 0, fp16Exp2> },
16703 { "Log2", "", 1, C, C, 0, 0, &getInputDataP, compareFP16ArithmeticFunc< C, C, 0, 0, fp16Log2> },
16704 { "Sqrt", "", 1, C, C, 0, 0, &getInputDataP, compareFP16ArithmeticFunc< C, C, 0, 0, fp16Sqrt> },
16705 { "InverseSqrt", "", 1, C, C, 0, 0, &getInputDataP, compareFP16ArithmeticFunc< C, C, 0, 0, fp16InverseSqrt> },
16706 { "Modf", "Frac", 1, C, C, 0, 0, &getInputData, compareFP16ArithmeticFunc< C, C, 0, 0, fp16ModfFrac> },
16707 { "Modf", "Int", 1, C, C, 0, 0, &getInputData, compareFP16ArithmeticFunc< C, C, 0, 0, fp16ModfInt> },
16708 { "ModfStruct", "Frac", 1, C, C, 0, 0, &getInputData, compareFP16ArithmeticFunc< C, C, 0, 0, fp16ModfFrac> },
16709 { "ModfStruct", "Int", 1, C, C, 0, 0, &getInputData, compareFP16ArithmeticFunc< C, C, 0, 0, fp16ModfInt> },
16710 { "Frexp", "S", 1, C, C, 0, 0, &getInputData, compareFP16ArithmeticFunc< C, C, 0, 0, fp16FrexpS> },
16711 { "Frexp", "E", 1, C, C, 0, 0, &getInputData, compareFP16ArithmeticFunc< C, C, 0, 0, fp16FrexpE> },
16712 { "FrexpStruct", "S", 1, C, C, 0, 0, &getInputData, compareFP16ArithmeticFunc< C, C, 0, 0, fp16FrexpS> },
16713 { "FrexpStruct", "E", 1, C, C, 0, 0, &getInputData, compareFP16ArithmeticFunc< C, C, 0, 0, fp16FrexpE> },
16714 { "OpFAdd", "", 2, C, C, C, 0, &getInputData, compareFP16ArithmeticFunc< C, C, C, 0, fp16OpFAdd> },
16715 { "OpFSub", "", 2, C, C, C, 0, &getInputData, compareFP16ArithmeticFunc< C, C, C, 0, fp16OpFSub> },
16716 { "OpFMul", "", 2, C, C, C, 0, &getInputData, compareFP16ArithmeticFunc< C, C, C, 0, fp16OpFMul> },
16717 { "OpFDiv", "", 2, C, C, C, 0, &getInputData, compareFP16ArithmeticFunc< C, C, C, 0, fp16OpFDiv> },
16718 { "Atan2", "", 2, C, C, C, 0, &getInputData, compareFP16ArithmeticFunc< C, C, C, 0, fp16Atan2> },
16719 { "Pow", "", 2, C, C, C, 0, &getInputDataP, compareFP16ArithmeticFunc< C, C, C, 0, fp16Pow> },
16720 { "FMin", "", 2, C, C, C, 0, &getInputData, compareFP16ArithmeticFunc< C, C, C, 0, fp16FMin> },
16721 { "FMax", "", 2, C, C, C, 0, &getInputData, compareFP16ArithmeticFunc< C, C, C, 0, fp16FMax> },
16722 { "Step", "", 2, C, C, C, 0, &getInputData, compareFP16ArithmeticFunc< C, C, C, 0, fp16Step> },
16723 { "Ldexp", "", 2, C, C, C, 0, &getInputData, compareFP16ArithmeticFunc< C, C, C, 0, fp16Ldexp> },
16724 { "FClamp", "", 3, C, C, C, C, &getInputData, compareFP16ArithmeticFunc< C, C, C, C, fp16FClamp> },
16725 { "FMix", "", 3, C, C, C, C, &getInputDataD, compareFP16ArithmeticFunc< C, C, C, C, fp16FMix> },
16726 { "SmoothStep", "", 3, C, C, C, C, &getInputDataSS, compareFP16ArithmeticFunc< C, C, C, C, fp16SmoothStep> },
16727 { "Fma", "", 3, C, C, C, C, &getInputData, compareFP16ArithmeticFunc< C, C, C, C, fp16Fma> },
16728 { "Length", "", 1, 1, C, 0, 0, &getInputData, compareFP16ArithmeticFunc< 1, C, 0, 0, fp16Length> },
16729 { "Distance", "", 2, 1, C, C, 0, &getInputData, compareFP16ArithmeticFunc< 1, C, C, 0, fp16Distance> },
16730 { "Cross", "", 2, C, C, C, 0, &getInputDataD, compareFP16ArithmeticFunc< C, C, C, 0, fp16Cross> },
16731 { "Normalize", "", 1, C, C, 0, 0, &getInputData, compareFP16ArithmeticFunc< C, C, 0, 0, fp16Normalize> },
16732 { "FaceForward", "", 3, C, C, C, C, &getInputDataD, compareFP16ArithmeticFunc< C, C, C, C, fp16FaceForward> },
16733 { "Reflect", "", 2, C, C, C, 0, &getInputDataD, compareFP16ArithmeticFunc< C, C, C, 0, fp16Reflect> },
16734 { "Refract", "", 3, C, C, C, 1, &getInputDataN, compareFP16ArithmeticFunc< C, C, C, 1, fp16Refract> },
16735 { "OpDot", "", 2, 1, C, C, 0, &getInputDataD, compareFP16ArithmeticFunc< 1, C, C, 0, fp16Dot> },
16736 { "OpVectorTimesScalar", "", 2, C, C, 1, 0, &getInputDataV, compareFP16ArithmeticFunc< C, C, 1, 0, fp16VectorTimesScalar> },
16739 for (deUint32 testFuncIdx = 0; testFuncIdx < DE_LENGTH_OF_ARRAY(testFuncs); ++testFuncIdx)
16741 const Math16TestFunc& testFunc = testFuncs[testFuncIdx];
16742 const string funcNameString = testFunc.funcName;
16744 if ((C != 3) && funcNameString == "Cross")
16747 if ((C < 2) && funcNameString == "OpDot")
16750 if ((C < 2) && funcNameString == "OpVectorTimesScalar")
16753 createFloat16ArithmeticFuncTest<SpecResource>(testCtx, *testGroup.get(), C, testFunc);
16756 return testGroup.release();
16759 template<class SpecResource>
16760 tcu::TestCaseGroup* createFloat16ArithmeticSet (tcu::TestContext& testCtx)
16762 const std::string testGroupName ("arithmetic");
16763 de::MovePtr<tcu::TestCaseGroup> testGroup (new tcu::TestCaseGroup(testCtx, testGroupName.c_str(), "Float 16 arithmetic and related tests"));
16764 const Math16TestFunc testFuncs[] =
16766 { "OpTranspose", "2x2", 1, MAT2X2, MAT2X2, 0, 0, &getInputDataM, compareFP16ArithmeticFunc< 4, 4, 0, 0, fp16Transpose<2,2> > },
16767 { "OpTranspose", "3x2", 1, MAT2X3, MAT3X2, 0, 0, &getInputDataM, compareFP16ArithmeticFunc< 8, 8, 0, 0, fp16Transpose<3,2> > },
16768 { "OpTranspose", "4x2", 1, MAT2X4, MAT4X2, 0, 0, &getInputDataM, compareFP16ArithmeticFunc< 8, 8, 0, 0, fp16Transpose<4,2> > },
16769 { "OpTranspose", "2x3", 1, MAT3X2, MAT2X3, 0, 0, &getInputDataM, compareFP16ArithmeticFunc< 8, 8, 0, 0, fp16Transpose<2,3> > },
16770 { "OpTranspose", "3x3", 1, MAT3X3, MAT3X3, 0, 0, &getInputDataM, compareFP16ArithmeticFunc< 16, 16, 0, 0, fp16Transpose<3,3> > },
16771 { "OpTranspose", "4x3", 1, MAT3X4, MAT4X3, 0, 0, &getInputDataM, compareFP16ArithmeticFunc< 16, 16, 0, 0, fp16Transpose<4,3> > },
16772 { "OpTranspose", "2x4", 1, MAT4X2, MAT2X4, 0, 0, &getInputDataM, compareFP16ArithmeticFunc< 8, 8, 0, 0, fp16Transpose<2,4> > },
16773 { "OpTranspose", "3x4", 1, MAT4X3, MAT3X4, 0, 0, &getInputDataM, compareFP16ArithmeticFunc< 16, 16, 0, 0, fp16Transpose<3,4> > },
16774 { "OpTranspose", "4x4", 1, MAT4X4, MAT4X4, 0, 0, &getInputDataM, compareFP16ArithmeticFunc< 16, 16, 0, 0, fp16Transpose<4,4> > },
16775 { "OpMatrixTimesScalar", "2x2", 2, MAT2X2, MAT2X2, 1, 0, &getInputDataD, compareFP16ArithmeticFunc< 4, 4, 1, 0, fp16MatrixTimesScalar<2,2> > },
16776 { "OpMatrixTimesScalar", "2x3", 2, MAT2X3, MAT2X3, 1, 0, &getInputDataD, compareFP16ArithmeticFunc< 8, 8, 1, 0, fp16MatrixTimesScalar<2,3> > },
16777 { "OpMatrixTimesScalar", "2x4", 2, MAT2X4, MAT2X4, 1, 0, &getInputDataD, compareFP16ArithmeticFunc< 8, 8, 1, 0, fp16MatrixTimesScalar<2,4> > },
16778 { "OpMatrixTimesScalar", "3x2", 2, MAT3X2, MAT3X2, 1, 0, &getInputDataD, compareFP16ArithmeticFunc< 8, 8, 1, 0, fp16MatrixTimesScalar<3,2> > },
16779 { "OpMatrixTimesScalar", "3x3", 2, MAT3X3, MAT3X3, 1, 0, &getInputDataD, compareFP16ArithmeticFunc< 16, 16, 1, 0, fp16MatrixTimesScalar<3,3> > },
16780 { "OpMatrixTimesScalar", "3x4", 2, MAT3X4, MAT3X4, 1, 0, &getInputDataD, compareFP16ArithmeticFunc< 16, 16, 1, 0, fp16MatrixTimesScalar<3,4> > },
16781 { "OpMatrixTimesScalar", "4x2", 2, MAT4X2, MAT4X2, 1, 0, &getInputDataD, compareFP16ArithmeticFunc< 8, 8, 1, 0, fp16MatrixTimesScalar<4,2> > },
16782 { "OpMatrixTimesScalar", "4x3", 2, MAT4X3, MAT4X3, 1, 0, &getInputDataD, compareFP16ArithmeticFunc< 16, 16, 1, 0, fp16MatrixTimesScalar<4,3> > },
16783 { "OpMatrixTimesScalar", "4x4", 2, MAT4X4, MAT4X4, 1, 0, &getInputDataD, compareFP16ArithmeticFunc< 16, 16, 1, 0, fp16MatrixTimesScalar<4,4> > },
16784 { "OpVectorTimesMatrix", "2x2", 2, VEC2, VEC2, MAT2X2, 0, &getInputDataD, compareFP16ArithmeticFunc< 2, 2, 4, 0, fp16VectorTimesMatrix<2,2> > },
16785 { "OpVectorTimesMatrix", "2x3", 2, VEC2, VEC3, MAT2X3, 0, &getInputDataD, compareFP16ArithmeticFunc< 2, 3, 8, 0, fp16VectorTimesMatrix<2,3> > },
16786 { "OpVectorTimesMatrix", "2x4", 2, VEC2, VEC4, MAT2X4, 0, &getInputDataD, compareFP16ArithmeticFunc< 2, 4, 8, 0, fp16VectorTimesMatrix<2,4> > },
16787 { "OpVectorTimesMatrix", "3x2", 2, VEC3, VEC2, MAT3X2, 0, &getInputDataD, compareFP16ArithmeticFunc< 3, 2, 8, 0, fp16VectorTimesMatrix<3,2> > },
16788 { "OpVectorTimesMatrix", "3x3", 2, VEC3, VEC3, MAT3X3, 0, &getInputDataD, compareFP16ArithmeticFunc< 3, 3, 16, 0, fp16VectorTimesMatrix<3,3> > },
16789 { "OpVectorTimesMatrix", "3x4", 2, VEC3, VEC4, MAT3X4, 0, &getInputDataD, compareFP16ArithmeticFunc< 3, 4, 16, 0, fp16VectorTimesMatrix<3,4> > },
16790 { "OpVectorTimesMatrix", "4x2", 2, VEC4, VEC2, MAT4X2, 0, &getInputDataD, compareFP16ArithmeticFunc< 4, 2, 8, 0, fp16VectorTimesMatrix<4,2> > },
16791 { "OpVectorTimesMatrix", "4x3", 2, VEC4, VEC3, MAT4X3, 0, &getInputDataD, compareFP16ArithmeticFunc< 4, 3, 16, 0, fp16VectorTimesMatrix<4,3> > },
16792 { "OpVectorTimesMatrix", "4x4", 2, VEC4, VEC4, MAT4X4, 0, &getInputDataD, compareFP16ArithmeticFunc< 4, 4, 16, 0, fp16VectorTimesMatrix<4,4> > },
16793 { "OpMatrixTimesVector", "2x2", 2, VEC2, MAT2X2, VEC2, 0, &getInputDataD, compareFP16ArithmeticFunc< 2, 4, 2, 0, fp16MatrixTimesVector<2,2> > },
16794 { "OpMatrixTimesVector", "2x3", 2, VEC3, MAT2X3, VEC2, 0, &getInputDataD, compareFP16ArithmeticFunc< 3, 8, 2, 0, fp16MatrixTimesVector<2,3> > },
16795 { "OpMatrixTimesVector", "2x4", 2, VEC4, MAT2X4, VEC2, 0, &getInputDataD, compareFP16ArithmeticFunc< 4, 8, 2, 0, fp16MatrixTimesVector<2,4> > },
16796 { "OpMatrixTimesVector", "3x2", 2, VEC2, MAT3X2, VEC3, 0, &getInputDataD, compareFP16ArithmeticFunc< 2, 8, 3, 0, fp16MatrixTimesVector<3,2> > },
16797 { "OpMatrixTimesVector", "3x3", 2, VEC3, MAT3X3, VEC3, 0, &getInputDataD, compareFP16ArithmeticFunc< 3, 16, 3, 0, fp16MatrixTimesVector<3,3> > },
16798 { "OpMatrixTimesVector", "3x4", 2, VEC4, MAT3X4, VEC3, 0, &getInputDataD, compareFP16ArithmeticFunc< 4, 16, 3, 0, fp16MatrixTimesVector<3,4> > },
16799 { "OpMatrixTimesVector", "4x2", 2, VEC2, MAT4X2, VEC4, 0, &getInputDataD, compareFP16ArithmeticFunc< 2, 8, 4, 0, fp16MatrixTimesVector<4,2> > },
16800 { "OpMatrixTimesVector", "4x3", 2, VEC3, MAT4X3, VEC4, 0, &getInputDataD, compareFP16ArithmeticFunc< 3, 16, 4, 0, fp16MatrixTimesVector<4,3> > },
16801 { "OpMatrixTimesVector", "4x4", 2, VEC4, MAT4X4, VEC4, 0, &getInputDataD, compareFP16ArithmeticFunc< 4, 16, 4, 0, fp16MatrixTimesVector<4,4> > },
16802 { "OpMatrixTimesMatrix", "2x2_2x2", 2, MAT2X2, MAT2X2, MAT2X2, 0, &getInputDataD, compareFP16ArithmeticFunc< 4, 4, 4, 0, fp16MatrixTimesMatrix<2,2,2,2> > },
16803 { "OpMatrixTimesMatrix", "2x2_3x2", 2, MAT3X2, MAT2X2, MAT3X2, 0, &getInputDataD, compareFP16ArithmeticFunc< 8, 4, 8, 0, fp16MatrixTimesMatrix<2,2,3,2> > },
16804 { "OpMatrixTimesMatrix", "2x2_4x2", 2, MAT4X2, MAT2X2, MAT4X2, 0, &getInputDataD, compareFP16ArithmeticFunc< 8, 4, 8, 0, fp16MatrixTimesMatrix<2,2,4,2> > },
16805 { "OpMatrixTimesMatrix", "2x3_2x2", 2, MAT2X3, MAT2X3, MAT2X2, 0, &getInputDataD, compareFP16ArithmeticFunc< 8, 8, 4, 0, fp16MatrixTimesMatrix<2,3,2,2> > },
16806 { "OpMatrixTimesMatrix", "2x3_3x2", 2, MAT3X3, MAT2X3, MAT3X2, 0, &getInputDataD, compareFP16ArithmeticFunc< 16, 8, 8, 0, fp16MatrixTimesMatrix<2,3,3,2> > },
16807 { "OpMatrixTimesMatrix", "2x3_4x2", 2, MAT4X3, MAT2X3, MAT4X2, 0, &getInputDataD, compareFP16ArithmeticFunc< 16, 8, 8, 0, fp16MatrixTimesMatrix<2,3,4,2> > },
16808 { "OpMatrixTimesMatrix", "2x4_2x2", 2, MAT2X4, MAT2X4, MAT2X2, 0, &getInputDataD, compareFP16ArithmeticFunc< 8, 8, 4, 0, fp16MatrixTimesMatrix<2,4,2,2> > },
16809 { "OpMatrixTimesMatrix", "2x4_3x2", 2, MAT3X4, MAT2X4, MAT3X2, 0, &getInputDataD, compareFP16ArithmeticFunc< 16, 8, 8, 0, fp16MatrixTimesMatrix<2,4,3,2> > },
16810 { "OpMatrixTimesMatrix", "2x4_4x2", 2, MAT4X4, MAT2X4, MAT4X2, 0, &getInputDataD, compareFP16ArithmeticFunc< 16, 8, 8, 0, fp16MatrixTimesMatrix<2,4,4,2> > },
16811 { "OpMatrixTimesMatrix", "3x2_2x3", 2, MAT2X2, MAT3X2, MAT2X3, 0, &getInputDataD, compareFP16ArithmeticFunc< 4, 8, 8, 0, fp16MatrixTimesMatrix<3,2,2,3> > },
16812 { "OpMatrixTimesMatrix", "3x2_3x3", 2, MAT3X2, MAT3X2, MAT3X3, 0, &getInputDataD, compareFP16ArithmeticFunc< 8, 8, 16, 0, fp16MatrixTimesMatrix<3,2,3,3> > },
16813 { "OpMatrixTimesMatrix", "3x2_4x3", 2, MAT4X2, MAT3X2, MAT4X3, 0, &getInputDataD, compareFP16ArithmeticFunc< 8, 8, 16, 0, fp16MatrixTimesMatrix<3,2,4,3> > },
16814 { "OpMatrixTimesMatrix", "3x3_2x3", 2, MAT2X3, MAT3X3, MAT2X3, 0, &getInputDataD, compareFP16ArithmeticFunc< 8, 16, 8, 0, fp16MatrixTimesMatrix<3,3,2,3> > },
16815 { "OpMatrixTimesMatrix", "3x3_3x3", 2, MAT3X3, MAT3X3, MAT3X3, 0, &getInputDataD, compareFP16ArithmeticFunc< 16, 16, 16, 0, fp16MatrixTimesMatrix<3,3,3,3> > },
16816 { "OpMatrixTimesMatrix", "3x3_4x3", 2, MAT4X3, MAT3X3, MAT4X3, 0, &getInputDataD, compareFP16ArithmeticFunc< 16, 16, 16, 0, fp16MatrixTimesMatrix<3,3,4,3> > },
16817 { "OpMatrixTimesMatrix", "3x4_2x3", 2, MAT2X4, MAT3X4, MAT2X3, 0, &getInputDataD, compareFP16ArithmeticFunc< 8, 16, 8, 0, fp16MatrixTimesMatrix<3,4,2,3> > },
16818 { "OpMatrixTimesMatrix", "3x4_3x3", 2, MAT3X4, MAT3X4, MAT3X3, 0, &getInputDataD, compareFP16ArithmeticFunc< 16, 16, 16, 0, fp16MatrixTimesMatrix<3,4,3,3> > },
16819 { "OpMatrixTimesMatrix", "3x4_4x3", 2, MAT4X4, MAT3X4, MAT4X3, 0, &getInputDataD, compareFP16ArithmeticFunc< 16, 16, 16, 0, fp16MatrixTimesMatrix<3,4,4,3> > },
16820 { "OpMatrixTimesMatrix", "4x2_2x4", 2, MAT2X2, MAT4X2, MAT2X4, 0, &getInputDataD, compareFP16ArithmeticFunc< 4, 8, 8, 0, fp16MatrixTimesMatrix<4,2,2,4> > },
16821 { "OpMatrixTimesMatrix", "4x2_3x4", 2, MAT3X2, MAT4X2, MAT3X4, 0, &getInputDataD, compareFP16ArithmeticFunc< 8, 8, 16, 0, fp16MatrixTimesMatrix<4,2,3,4> > },
16822 { "OpMatrixTimesMatrix", "4x2_4x4", 2, MAT4X2, MAT4X2, MAT4X4, 0, &getInputDataD, compareFP16ArithmeticFunc< 8, 8, 16, 0, fp16MatrixTimesMatrix<4,2,4,4> > },
16823 { "OpMatrixTimesMatrix", "4x3_2x4", 2, MAT2X3, MAT4X3, MAT2X4, 0, &getInputDataD, compareFP16ArithmeticFunc< 8, 16, 8, 0, fp16MatrixTimesMatrix<4,3,2,4> > },
16824 { "OpMatrixTimesMatrix", "4x3_3x4", 2, MAT3X3, MAT4X3, MAT3X4, 0, &getInputDataD, compareFP16ArithmeticFunc< 16, 16, 16, 0, fp16MatrixTimesMatrix<4,3,3,4> > },
16825 { "OpMatrixTimesMatrix", "4x3_4x4", 2, MAT4X3, MAT4X3, MAT4X4, 0, &getInputDataD, compareFP16ArithmeticFunc< 16, 16, 16, 0, fp16MatrixTimesMatrix<4,3,4,4> > },
16826 { "OpMatrixTimesMatrix", "4x4_2x4", 2, MAT2X4, MAT4X4, MAT2X4, 0, &getInputDataD, compareFP16ArithmeticFunc< 8, 16, 8, 0, fp16MatrixTimesMatrix<4,4,2,4> > },
16827 { "OpMatrixTimesMatrix", "4x4_3x4", 2, MAT3X4, MAT4X4, MAT3X4, 0, &getInputDataD, compareFP16ArithmeticFunc< 16, 16, 16, 0, fp16MatrixTimesMatrix<4,4,3,4> > },
16828 { "OpMatrixTimesMatrix", "4x4_4x4", 2, MAT4X4, MAT4X4, MAT4X4, 0, &getInputDataD, compareFP16ArithmeticFunc< 16, 16, 16, 0, fp16MatrixTimesMatrix<4,4,4,4> > },
16829 { "OpOuterProduct", "2x2", 2, MAT2X2, VEC2, VEC2, 0, &getInputDataD, compareFP16ArithmeticFunc< 4, 2, 2, 0, fp16OuterProduct<2,2> > },
16830 { "OpOuterProduct", "2x3", 2, MAT2X3, VEC3, VEC2, 0, &getInputDataD, compareFP16ArithmeticFunc< 8, 3, 2, 0, fp16OuterProduct<2,3> > },
16831 { "OpOuterProduct", "2x4", 2, MAT2X4, VEC4, VEC2, 0, &getInputDataD, compareFP16ArithmeticFunc< 8, 4, 2, 0, fp16OuterProduct<2,4> > },
16832 { "OpOuterProduct", "3x2", 2, MAT3X2, VEC2, VEC3, 0, &getInputDataD, compareFP16ArithmeticFunc< 8, 2, 3, 0, fp16OuterProduct<3,2> > },
16833 { "OpOuterProduct", "3x3", 2, MAT3X3, VEC3, VEC3, 0, &getInputDataD, compareFP16ArithmeticFunc< 16, 3, 3, 0, fp16OuterProduct<3,3> > },
16834 { "OpOuterProduct", "3x4", 2, MAT3X4, VEC4, VEC3, 0, &getInputDataD, compareFP16ArithmeticFunc< 16, 4, 3, 0, fp16OuterProduct<3,4> > },
16835 { "OpOuterProduct", "4x2", 2, MAT4X2, VEC2, VEC4, 0, &getInputDataD, compareFP16ArithmeticFunc< 8, 2, 4, 0, fp16OuterProduct<4,2> > },
16836 { "OpOuterProduct", "4x3", 2, MAT4X3, VEC3, VEC4, 0, &getInputDataD, compareFP16ArithmeticFunc< 16, 3, 4, 0, fp16OuterProduct<4,3> > },
16837 { "OpOuterProduct", "4x4", 2, MAT4X4, VEC4, VEC4, 0, &getInputDataD, compareFP16ArithmeticFunc< 16, 4, 4, 0, fp16OuterProduct<4,4> > },
16838 { "Determinant", "2x2", 1, SCALAR, MAT2X2, NONE, 0, &getInputDataC, compareFP16ArithmeticFunc< 1, 4, 0, 0, fp16Determinant<2> > },
16839 { "Determinant", "3x3", 1, SCALAR, MAT3X3, NONE, 0, &getInputDataC, compareFP16ArithmeticFunc< 1, 16, 0, 0, fp16Determinant<3> > },
16840 { "Determinant", "4x4", 1, SCALAR, MAT4X4, NONE, 0, &getInputDataC, compareFP16ArithmeticFunc< 1, 16, 0, 0, fp16Determinant<4> > },
16841 { "MatrixInverse", "2x2", 1, MAT2X2, MAT2X2, NONE, 0, &getInputDataC, compareFP16ArithmeticFunc< 4, 4, 0, 0, fp16Inverse<2> > },
16844 for (deUint32 testFuncIdx = 0; testFuncIdx < DE_LENGTH_OF_ARRAY(testFuncs); ++testFuncIdx)
16846 const Math16TestFunc& testFunc = testFuncs[testFuncIdx];
16848 createFloat16ArithmeticFuncTest<SpecResource>(testCtx, *testGroup.get(), 0, testFunc);
16851 return testGroup.release();
16854 const string getNumberTypeName (const NumberType type)
16856 if (type == NUMBERTYPE_INT32)
16860 else if (type == NUMBERTYPE_UINT32)
16864 else if (type == NUMBERTYPE_FLOAT32)
16875 deInt32 getInt(de::Random& rnd)
16877 return rnd.getInt(std::numeric_limits<int>::min(), std::numeric_limits<int>::max());
16880 const string repeatString (const string& str, int times)
16883 for (int i = 0; i < times; ++i)
16890 const string getRandomConstantString (const NumberType type, de::Random& rnd)
16892 if (type == NUMBERTYPE_INT32)
16894 return numberToString<deInt32>(getInt(rnd));
16896 else if (type == NUMBERTYPE_UINT32)
16898 return numberToString<deUint32>(rnd.getUint32());
16900 else if (type == NUMBERTYPE_FLOAT32)
16902 return numberToString<float>(rnd.getFloat());
16911 void createVectorCompositeCases (vector<map<string, string> >& testCases, de::Random& rnd, const NumberType type)
16913 map<string, string> params;
16916 for (int width = 2; width <= 4; ++width)
16918 const string randomConst = numberToString(getInt(rnd));
16919 const string widthStr = numberToString(width);
16920 const string composite_type = "${customType}vec" + widthStr;
16921 const int index = rnd.getInt(0, width-1);
16923 params["type"] = "vec";
16924 params["name"] = params["type"] + "_" + widthStr;
16925 params["compositeDecl"] = composite_type + " = OpTypeVector ${customType} " + widthStr +"\n";
16926 params["compositeType"] = composite_type;
16927 params["filler"] = string("%filler = OpConstant ${customType} ") + getRandomConstantString(type, rnd) + "\n";
16928 params["compositeConstruct"] = "%instance = OpCompositeConstruct " + composite_type + repeatString(" %filler", width) + "\n";
16929 params["indexes"] = numberToString(index);
16930 testCases.push_back(params);
16934 void createArrayCompositeCases (vector<map<string, string> >& testCases, de::Random& rnd, const NumberType type)
16936 const int limit = 10;
16937 map<string, string> params;
16939 for (int width = 2; width <= limit; ++width)
16941 string randomConst = numberToString(getInt(rnd));
16942 string widthStr = numberToString(width);
16943 int index = rnd.getInt(0, width-1);
16945 params["type"] = "array";
16946 params["name"] = params["type"] + "_" + widthStr;
16947 params["compositeDecl"] = string("%arraywidth = OpConstant %u32 " + widthStr + "\n")
16948 + "%composite = OpTypeArray ${customType} %arraywidth\n";
16949 params["compositeType"] = "%composite";
16950 params["filler"] = string("%filler = OpConstant ${customType} ") + getRandomConstantString(type, rnd) + "\n";
16951 params["compositeConstruct"] = "%instance = OpCompositeConstruct %composite" + repeatString(" %filler", width) + "\n";
16952 params["indexes"] = numberToString(index);
16953 testCases.push_back(params);
16957 void createStructCompositeCases (vector<map<string, string> >& testCases, de::Random& rnd, const NumberType type)
16959 const int limit = 10;
16960 map<string, string> params;
16962 for (int width = 2; width <= limit; ++width)
16964 string randomConst = numberToString(getInt(rnd));
16965 int index = rnd.getInt(0, width-1);
16967 params["type"] = "struct";
16968 params["name"] = params["type"] + "_" + numberToString(width);
16969 params["compositeDecl"] = "%composite = OpTypeStruct" + repeatString(" ${customType}", width) + "\n";
16970 params["compositeType"] = "%composite";
16971 params["filler"] = string("%filler = OpConstant ${customType} ") + getRandomConstantString(type, rnd) + "\n";
16972 params["compositeConstruct"] = "%instance = OpCompositeConstruct %composite" + repeatString(" %filler", width) + "\n";
16973 params["indexes"] = numberToString(index);
16974 testCases.push_back(params);
16978 void createMatrixCompositeCases (vector<map<string, string> >& testCases, de::Random& rnd, const NumberType type)
16980 map<string, string> params;
16983 for (int width = 2; width <= 4; ++width)
16985 string widthStr = numberToString(width);
16987 for (int column = 2 ; column <= 4; ++column)
16989 int index_0 = rnd.getInt(0, column-1);
16990 int index_1 = rnd.getInt(0, width-1);
16991 string columnStr = numberToString(column);
16993 params["type"] = "matrix";
16994 params["name"] = params["type"] + "_" + widthStr + "x" + columnStr;
16995 params["compositeDecl"] = string("%vectype = OpTypeVector ${customType} " + widthStr + "\n")
16996 + "%composite = OpTypeMatrix %vectype " + columnStr + "\n";
16997 params["compositeType"] = "%composite";
16999 params["filler"] = string("%filler = OpConstant ${customType} ") + getRandomConstantString(type, rnd) + "\n"
17000 + "%fillerVec = OpConstantComposite %vectype" + repeatString(" %filler", width) + "\n";
17002 params["compositeConstruct"] = "%instance = OpCompositeConstruct %composite" + repeatString(" %fillerVec", column) + "\n";
17003 params["indexes"] = numberToString(index_0) + " " + numberToString(index_1);
17004 testCases.push_back(params);
17009 void createCompositeCases (vector<map<string, string> >& testCases, de::Random& rnd, const NumberType type)
17011 createVectorCompositeCases(testCases, rnd, type);
17012 createArrayCompositeCases(testCases, rnd, type);
17013 createStructCompositeCases(testCases, rnd, type);
17014 // Matrix only supports float types
17015 if (type == NUMBERTYPE_FLOAT32)
17017 createMatrixCompositeCases(testCases, rnd, type);
17021 const string getAssemblyTypeDeclaration (const NumberType type)
17025 case NUMBERTYPE_INT32: return "OpTypeInt 32 1";
17026 case NUMBERTYPE_UINT32: return "OpTypeInt 32 0";
17027 case NUMBERTYPE_FLOAT32: return "OpTypeFloat 32";
17028 default: DE_ASSERT(false); return "";
17032 const string getAssemblyTypeName (const NumberType type)
17036 case NUMBERTYPE_INT32: return "%i32";
17037 case NUMBERTYPE_UINT32: return "%u32";
17038 case NUMBERTYPE_FLOAT32: return "%f32";
17039 default: DE_ASSERT(false); return "";
17043 const string specializeCompositeInsertShaderTemplate (const NumberType type, const map<string, string>& params)
17045 map<string, string> parameters(params);
17047 const string customType = getAssemblyTypeName(type);
17048 map<string, string> substCustomType;
17049 substCustomType["customType"] = customType;
17050 parameters["compositeDecl"] = StringTemplate(parameters.at("compositeDecl")).specialize(substCustomType);
17051 parameters["compositeType"] = StringTemplate(parameters.at("compositeType")).specialize(substCustomType);
17052 parameters["compositeConstruct"] = StringTemplate(parameters.at("compositeConstruct")).specialize(substCustomType);
17053 parameters["filler"] = StringTemplate(parameters.at("filler")).specialize(substCustomType);
17054 parameters["customType"] = customType;
17055 parameters["compositeDecorator"] = (parameters["type"] == "array") ? "OpDecorate %composite ArrayStride 4\n" : "";
17057 if (parameters.at("compositeType") != "%u32vec3")
17059 parameters["u32vec3Decl"] = "%u32vec3 = OpTypeVector %u32 3\n";
17062 return StringTemplate(
17063 "OpCapability Shader\n"
17064 "OpCapability Matrix\n"
17065 "OpMemoryModel Logical GLSL450\n"
17066 "OpEntryPoint GLCompute %main \"main\" %id\n"
17067 "OpExecutionMode %main LocalSize 1 1 1\n"
17069 "OpSource GLSL 430\n"
17070 "OpName %main \"main\"\n"
17071 "OpName %id \"gl_GlobalInvocationID\"\n"
17074 "OpDecorate %id BuiltIn GlobalInvocationId\n"
17075 "OpDecorate %buf BufferBlock\n"
17076 "OpDecorate %indata DescriptorSet 0\n"
17077 "OpDecorate %indata Binding 0\n"
17078 "OpDecorate %outdata DescriptorSet 0\n"
17079 "OpDecorate %outdata Binding 1\n"
17080 "OpDecorate %customarr ArrayStride 4\n"
17081 "${compositeDecorator}"
17082 "OpMemberDecorate %buf 0 Offset 0\n"
17085 "%void = OpTypeVoid\n"
17086 "%voidf = OpTypeFunction %void\n"
17087 "%u32 = OpTypeInt 32 0\n"
17088 "%i32 = OpTypeInt 32 1\n"
17089 "%f32 = OpTypeFloat 32\n"
17091 // Composite declaration
17097 "${u32vec3Decl:opt}"
17098 "%uvec3ptr = OpTypePointer Input %u32vec3\n"
17100 // Inherited from custom
17101 "%customptr = OpTypePointer Uniform ${customType}\n"
17102 "%customarr = OpTypeRuntimeArray ${customType}\n"
17103 "%buf = OpTypeStruct %customarr\n"
17104 "%bufptr = OpTypePointer Uniform %buf\n"
17106 "%indata = OpVariable %bufptr Uniform\n"
17107 "%outdata = OpVariable %bufptr Uniform\n"
17109 "%id = OpVariable %uvec3ptr Input\n"
17110 "%zero = OpConstant %i32 0\n"
17112 "%main = OpFunction %void None %voidf\n"
17113 "%label = OpLabel\n"
17114 "%idval = OpLoad %u32vec3 %id\n"
17115 "%x = OpCompositeExtract %u32 %idval 0\n"
17117 "%inloc = OpAccessChain %customptr %indata %zero %x\n"
17118 "%outloc = OpAccessChain %customptr %outdata %zero %x\n"
17119 // Read the input value
17120 "%inval = OpLoad ${customType} %inloc\n"
17121 // Create the composite and fill it
17122 "${compositeConstruct}"
17123 // Insert the input value to a place
17124 "%instance2 = OpCompositeInsert ${compositeType} %inval %instance ${indexes}\n"
17125 // Read back the value from the position
17126 "%out_val = OpCompositeExtract ${customType} %instance2 ${indexes}\n"
17127 // Store it in the output position
17128 " OpStore %outloc %out_val\n"
17131 ).specialize(parameters);
17134 template<typename T>
17135 BufferSp createCompositeBuffer(T number)
17137 return BufferSp(new Buffer<T>(vector<T>(1, number)));
17140 tcu::TestCaseGroup* createOpCompositeInsertGroup (tcu::TestContext& testCtx)
17142 de::MovePtr<tcu::TestCaseGroup> group (new tcu::TestCaseGroup(testCtx, "opcompositeinsert", "Test the OpCompositeInsert instruction"));
17143 de::Random rnd (deStringHash(group->getName()));
17145 for (int type = NUMBERTYPE_INT32; type != NUMBERTYPE_END32; ++type)
17147 NumberType numberType = NumberType(type);
17148 const string typeName = getNumberTypeName(numberType);
17149 const string description = "Test the OpCompositeInsert instruction with " + typeName + "s";
17150 de::MovePtr<tcu::TestCaseGroup> subGroup (new tcu::TestCaseGroup(testCtx, typeName.c_str(), description.c_str()));
17151 vector<map<string, string> > testCases;
17153 createCompositeCases(testCases, rnd, numberType);
17155 for (vector<map<string, string> >::const_iterator test = testCases.begin(); test != testCases.end(); ++test)
17157 ComputeShaderSpec spec;
17159 spec.assembly = specializeCompositeInsertShaderTemplate(numberType, *test);
17161 switch (numberType)
17163 case NUMBERTYPE_INT32:
17165 deInt32 number = getInt(rnd);
17166 spec.inputs.push_back(createCompositeBuffer<deInt32>(number));
17167 spec.outputs.push_back(createCompositeBuffer<deInt32>(number));
17170 case NUMBERTYPE_UINT32:
17172 deUint32 number = rnd.getUint32();
17173 spec.inputs.push_back(createCompositeBuffer<deUint32>(number));
17174 spec.outputs.push_back(createCompositeBuffer<deUint32>(number));
17177 case NUMBERTYPE_FLOAT32:
17179 float number = rnd.getFloat();
17180 spec.inputs.push_back(createCompositeBuffer<float>(number));
17181 spec.outputs.push_back(createCompositeBuffer<float>(number));
17188 spec.numWorkGroups = IVec3(1, 1, 1);
17189 subGroup->addChild(new SpvAsmComputeShaderCase(testCtx, test->at("name").c_str(), "OpCompositeInsert test", spec));
17191 group->addChild(subGroup.release());
17193 return group.release();
17196 struct AssemblyStructInfo
17198 AssemblyStructInfo (const deUint32 comp, const deUint32 idx)
17199 : components (comp)
17203 deUint32 components;
17207 const string specializeInBoundsShaderTemplate (const NumberType type, const AssemblyStructInfo& structInfo, const map<string, string>& params)
17209 // Create the full index string
17210 string fullIndex = numberToString(structInfo.index) + " " + params.at("indexes");
17211 // Convert it to list of indexes
17212 vector<string> indexes = de::splitString(fullIndex, ' ');
17214 map<string, string> parameters (params);
17215 parameters["structType"] = repeatString(" ${compositeType}", structInfo.components);
17216 parameters["structConstruct"] = repeatString(" %instance", structInfo.components);
17217 parameters["insertIndexes"] = fullIndex;
17219 // In matrix cases the last two index is the CompositeExtract indexes
17220 const deUint32 extractIndexes = (parameters["type"] == "matrix") ? 2 : 1;
17222 // Construct the extractIndex
17223 for (vector<string>::const_iterator index = indexes.end() - extractIndexes; index != indexes.end(); ++index)
17225 parameters["extractIndexes"] += " " + *index;
17228 // Remove the last 1 or 2 element depends on matrix case or not
17229 indexes.erase(indexes.end() - extractIndexes, indexes.end());
17232 // Generate AccessChain index expressions (except for the last one, because we use ptr to the composite)
17233 for (vector<string>::const_iterator index = indexes.begin(); index != indexes.end(); ++index)
17235 string indexId = "%index_" + numberToString(id++);
17236 parameters["accessChainConstDeclaration"] += indexId + " = OpConstant %u32 " + *index + "\n";
17237 parameters["accessChainIndexes"] += " " + indexId;
17240 parameters["compositeDecorator"] = (parameters["type"] == "array") ? "OpDecorate %composite ArrayStride 4\n" : "";
17242 const string customType = getAssemblyTypeName(type);
17243 map<string, string> substCustomType;
17244 substCustomType["customType"] = customType;
17245 parameters["compositeDecl"] = StringTemplate(parameters.at("compositeDecl")).specialize(substCustomType);
17246 parameters["compositeType"] = StringTemplate(parameters.at("compositeType")).specialize(substCustomType);
17247 parameters["compositeConstruct"] = StringTemplate(parameters.at("compositeConstruct")).specialize(substCustomType);
17248 parameters["filler"] = StringTemplate(parameters.at("filler")).specialize(substCustomType);
17249 parameters["customType"] = customType;
17251 const string compositeType = parameters.at("compositeType");
17252 map<string, string> substCompositeType;
17253 substCompositeType["compositeType"] = compositeType;
17254 parameters["structType"] = StringTemplate(parameters.at("structType")).specialize(substCompositeType);
17255 if (compositeType != "%u32vec3")
17257 parameters["u32vec3Decl"] = "%u32vec3 = OpTypeVector %u32 3\n";
17260 return StringTemplate(
17261 "OpCapability Shader\n"
17262 "OpCapability Matrix\n"
17263 "OpMemoryModel Logical GLSL450\n"
17264 "OpEntryPoint GLCompute %main \"main\" %id\n"
17265 "OpExecutionMode %main LocalSize 1 1 1\n"
17267 "OpSource GLSL 430\n"
17268 "OpName %main \"main\"\n"
17269 "OpName %id \"gl_GlobalInvocationID\"\n"
17271 "OpDecorate %id BuiltIn GlobalInvocationId\n"
17272 "OpDecorate %buf BufferBlock\n"
17273 "OpDecorate %indata DescriptorSet 0\n"
17274 "OpDecorate %indata Binding 0\n"
17275 "OpDecorate %outdata DescriptorSet 0\n"
17276 "OpDecorate %outdata Binding 1\n"
17277 "OpDecorate %customarr ArrayStride 4\n"
17278 "${compositeDecorator}"
17279 "OpMemberDecorate %buf 0 Offset 0\n"
17281 "%void = OpTypeVoid\n"
17282 "%voidf = OpTypeFunction %void\n"
17283 "%i32 = OpTypeInt 32 1\n"
17284 "%u32 = OpTypeInt 32 0\n"
17285 "%f32 = OpTypeFloat 32\n"
17288 // %u32vec3 if not already declared in ${compositeDecl}
17289 "${u32vec3Decl:opt}"
17290 "%uvec3ptr = OpTypePointer Input %u32vec3\n"
17291 // Inherited from composite
17292 "%composite_p = OpTypePointer Function ${compositeType}\n"
17293 "%struct_t = OpTypeStruct${structType}\n"
17294 "%struct_p = OpTypePointer Function %struct_t\n"
17297 "${accessChainConstDeclaration}"
17298 // Inherited from custom
17299 "%customptr = OpTypePointer Uniform ${customType}\n"
17300 "%customarr = OpTypeRuntimeArray ${customType}\n"
17301 "%buf = OpTypeStruct %customarr\n"
17302 "%bufptr = OpTypePointer Uniform %buf\n"
17303 "%indata = OpVariable %bufptr Uniform\n"
17304 "%outdata = OpVariable %bufptr Uniform\n"
17306 "%id = OpVariable %uvec3ptr Input\n"
17307 "%zero = OpConstant %u32 0\n"
17308 "%main = OpFunction %void None %voidf\n"
17309 "%label = OpLabel\n"
17310 "%struct_v = OpVariable %struct_p Function\n"
17311 "%idval = OpLoad %u32vec3 %id\n"
17312 "%x = OpCompositeExtract %u32 %idval 0\n"
17313 // Create the input/output type
17314 "%inloc = OpInBoundsAccessChain %customptr %indata %zero %x\n"
17315 "%outloc = OpInBoundsAccessChain %customptr %outdata %zero %x\n"
17316 // Read the input value
17317 "%inval = OpLoad ${customType} %inloc\n"
17318 // Create the composite and fill it
17319 "${compositeConstruct}"
17320 // Create the struct and fill it with the composite
17321 "%struct = OpCompositeConstruct %struct_t${structConstruct}\n"
17322 // Insert the value
17323 "%comp_obj = OpCompositeInsert %struct_t %inval %struct ${insertIndexes}\n"
17324 // Store the object
17325 " OpStore %struct_v %comp_obj\n"
17326 // Get deepest possible composite pointer
17327 "%inner_ptr = OpInBoundsAccessChain %composite_p %struct_v${accessChainIndexes}\n"
17328 "%read_obj = OpLoad ${compositeType} %inner_ptr\n"
17329 // Read back the stored value
17330 "%read_val = OpCompositeExtract ${customType} %read_obj${extractIndexes}\n"
17331 " OpStore %outloc %read_val\n"
17334 ).specialize(parameters);
17337 tcu::TestCaseGroup* createOpInBoundsAccessChainGroup (tcu::TestContext& testCtx)
17339 de::MovePtr<tcu::TestCaseGroup> group (new tcu::TestCaseGroup(testCtx, "opinboundsaccesschain", "Test the OpInBoundsAccessChain instruction"));
17340 de::Random rnd (deStringHash(group->getName()));
17342 for (int type = NUMBERTYPE_INT32; type != NUMBERTYPE_END32; ++type)
17344 NumberType numberType = NumberType(type);
17345 const string typeName = getNumberTypeName(numberType);
17346 const string description = "Test the OpInBoundsAccessChain instruction with " + typeName + "s";
17347 de::MovePtr<tcu::TestCaseGroup> subGroup (new tcu::TestCaseGroup(testCtx, typeName.c_str(), description.c_str()));
17349 vector<map<string, string> > testCases;
17350 createCompositeCases(testCases, rnd, numberType);
17352 for (vector<map<string, string> >::const_iterator test = testCases.begin(); test != testCases.end(); ++test)
17354 ComputeShaderSpec spec;
17356 // Number of components inside of a struct
17357 deUint32 structComponents = rnd.getInt(2, 8);
17358 // Component index value
17359 deUint32 structIndex = rnd.getInt(0, structComponents - 1);
17360 AssemblyStructInfo structInfo(structComponents, structIndex);
17362 spec.assembly = specializeInBoundsShaderTemplate(numberType, structInfo, *test);
17364 switch (numberType)
17366 case NUMBERTYPE_INT32:
17368 deInt32 number = getInt(rnd);
17369 spec.inputs.push_back(createCompositeBuffer<deInt32>(number));
17370 spec.outputs.push_back(createCompositeBuffer<deInt32>(number));
17373 case NUMBERTYPE_UINT32:
17375 deUint32 number = rnd.getUint32();
17376 spec.inputs.push_back(createCompositeBuffer<deUint32>(number));
17377 spec.outputs.push_back(createCompositeBuffer<deUint32>(number));
17380 case NUMBERTYPE_FLOAT32:
17382 float number = rnd.getFloat();
17383 spec.inputs.push_back(createCompositeBuffer<float>(number));
17384 spec.outputs.push_back(createCompositeBuffer<float>(number));
17390 spec.numWorkGroups = IVec3(1, 1, 1);
17391 subGroup->addChild(new SpvAsmComputeShaderCase(testCtx, test->at("name").c_str(), "OpInBoundsAccessChain test", spec));
17393 group->addChild(subGroup.release());
17395 return group.release();
17398 // If the params missing, uninitialized case
17399 const string specializeDefaultOutputShaderTemplate (const NumberType type, const map<string, string>& params = map<string, string>())
17401 map<string, string> parameters(params);
17403 parameters["customType"] = getAssemblyTypeName(type);
17405 // Declare the const value, and use it in the initializer
17406 if (params.find("constValue") != params.end())
17408 parameters["variableInitializer"] = " %const";
17410 // Uninitialized case
17413 parameters["commentDecl"] = ";";
17416 return StringTemplate(
17417 "OpCapability Shader\n"
17418 "OpMemoryModel Logical GLSL450\n"
17419 "OpEntryPoint GLCompute %main \"main\" %id\n"
17420 "OpExecutionMode %main LocalSize 1 1 1\n"
17421 "OpSource GLSL 430\n"
17422 "OpName %main \"main\"\n"
17423 "OpName %id \"gl_GlobalInvocationID\"\n"
17425 "OpDecorate %id BuiltIn GlobalInvocationId\n"
17426 "OpDecorate %indata DescriptorSet 0\n"
17427 "OpDecorate %indata Binding 0\n"
17428 "OpDecorate %outdata DescriptorSet 0\n"
17429 "OpDecorate %outdata Binding 1\n"
17430 "OpDecorate %in_arr ArrayStride 4\n"
17431 "OpDecorate %in_buf BufferBlock\n"
17432 "OpMemberDecorate %in_buf 0 Offset 0\n"
17434 "%void = OpTypeVoid\n"
17435 "%voidf = OpTypeFunction %void\n"
17436 "%u32 = OpTypeInt 32 0\n"
17437 "%i32 = OpTypeInt 32 1\n"
17438 "%f32 = OpTypeFloat 32\n"
17439 "%uvec3 = OpTypeVector %u32 3\n"
17440 "%uvec3ptr = OpTypePointer Input %uvec3\n"
17441 "${commentDecl:opt}%const = OpConstant ${customType} ${constValue:opt}\n"
17443 "%in_ptr = OpTypePointer Uniform ${customType}\n"
17444 "%in_arr = OpTypeRuntimeArray ${customType}\n"
17445 "%in_buf = OpTypeStruct %in_arr\n"
17446 "%in_bufptr = OpTypePointer Uniform %in_buf\n"
17447 "%indata = OpVariable %in_bufptr Uniform\n"
17448 "%outdata = OpVariable %in_bufptr Uniform\n"
17449 "%id = OpVariable %uvec3ptr Input\n"
17450 "%var_ptr = OpTypePointer Function ${customType}\n"
17452 "%zero = OpConstant %i32 0\n"
17454 "%main = OpFunction %void None %voidf\n"
17455 "%label = OpLabel\n"
17456 "%out_var = OpVariable %var_ptr Function${variableInitializer:opt}\n"
17457 "%idval = OpLoad %uvec3 %id\n"
17458 "%x = OpCompositeExtract %u32 %idval 0\n"
17459 "%inloc = OpAccessChain %in_ptr %indata %zero %x\n"
17460 "%outloc = OpAccessChain %in_ptr %outdata %zero %x\n"
17462 "%outval = OpLoad ${customType} %out_var\n"
17463 " OpStore %outloc %outval\n"
17466 ).specialize(parameters);
17469 bool compareFloats (const std::vector<Resource>&, const vector<AllocationSp>& outputAllocs, const std::vector<Resource>& expectedOutputs, TestLog& log)
17471 DE_ASSERT(outputAllocs.size() != 0);
17472 DE_ASSERT(outputAllocs.size() == expectedOutputs.size());
17474 // Use custom epsilon because of the float->string conversion
17475 const float epsilon = 0.00001f;
17477 for (size_t outputNdx = 0; outputNdx < outputAllocs.size(); ++outputNdx)
17479 vector<deUint8> expectedBytes;
17483 expectedOutputs[outputNdx].getBytes(expectedBytes);
17484 memcpy(&expected, &expectedBytes.front(), expectedBytes.size());
17485 memcpy(&actual, outputAllocs[outputNdx]->getHostPtr(), expectedBytes.size());
17487 // Test with epsilon
17488 if (fabs(expected - actual) > epsilon)
17490 log << TestLog::Message << "Error: The actual and expected values not matching."
17491 << " Expected: " << expected << " Actual: " << actual << " Epsilon: " << epsilon << TestLog::EndMessage;
17498 // Checks if the driver crash with uninitialized cases
17499 bool passthruVerify (const std::vector<Resource>&, const vector<AllocationSp>& outputAllocs, const std::vector<Resource>& expectedOutputs, TestLog&)
17501 DE_ASSERT(outputAllocs.size() != 0);
17502 DE_ASSERT(outputAllocs.size() == expectedOutputs.size());
17504 // Copy and discard the result.
17505 for (size_t outputNdx = 0; outputNdx < outputAllocs.size(); ++outputNdx)
17507 vector<deUint8> expectedBytes;
17508 expectedOutputs[outputNdx].getBytes(expectedBytes);
17510 const size_t width = expectedBytes.size();
17511 vector<char> data (width);
17513 memcpy(&data[0], outputAllocs[outputNdx]->getHostPtr(), width);
17518 tcu::TestCaseGroup* createShaderDefaultOutputGroup (tcu::TestContext& testCtx)
17520 de::MovePtr<tcu::TestCaseGroup> group (new tcu::TestCaseGroup(testCtx, "shader_default_output", "Test shader default output."));
17521 de::Random rnd (deStringHash(group->getName()));
17523 for (int type = NUMBERTYPE_INT32; type != NUMBERTYPE_END32; ++type)
17525 NumberType numberType = NumberType(type);
17526 const string typeName = getNumberTypeName(numberType);
17527 const string description = "Test the OpVariable initializer with " + typeName + ".";
17528 de::MovePtr<tcu::TestCaseGroup> subGroup (new tcu::TestCaseGroup(testCtx, typeName.c_str(), description.c_str()));
17530 // 2 similar subcases (initialized and uninitialized)
17531 for (int subCase = 0; subCase < 2; ++subCase)
17533 ComputeShaderSpec spec;
17534 spec.numWorkGroups = IVec3(1, 1, 1);
17536 map<string, string> params;
17538 switch (numberType)
17540 case NUMBERTYPE_INT32:
17542 deInt32 number = getInt(rnd);
17543 spec.inputs.push_back(createCompositeBuffer<deInt32>(number));
17544 spec.outputs.push_back(createCompositeBuffer<deInt32>(number));
17545 params["constValue"] = numberToString(number);
17548 case NUMBERTYPE_UINT32:
17550 deUint32 number = rnd.getUint32();
17551 spec.inputs.push_back(createCompositeBuffer<deUint32>(number));
17552 spec.outputs.push_back(createCompositeBuffer<deUint32>(number));
17553 params["constValue"] = numberToString(number);
17556 case NUMBERTYPE_FLOAT32:
17558 float number = rnd.getFloat();
17559 spec.inputs.push_back(createCompositeBuffer<float>(number));
17560 spec.outputs.push_back(createCompositeBuffer<float>(number));
17561 spec.verifyIO = &compareFloats;
17562 params["constValue"] = numberToString(number);
17569 // Initialized subcase
17572 spec.assembly = specializeDefaultOutputShaderTemplate(numberType, params);
17573 subGroup->addChild(new SpvAsmComputeShaderCase(testCtx, "initialized", "OpVariable initializer tests.", spec));
17575 // Uninitialized subcase
17578 spec.assembly = specializeDefaultOutputShaderTemplate(numberType);
17579 spec.verifyIO = &passthruVerify;
17580 subGroup->addChild(new SpvAsmComputeShaderCase(testCtx, "uninitialized", "OpVariable initializer tests.", spec));
17583 group->addChild(subGroup.release());
17585 return group.release();
17588 tcu::TestCaseGroup* createOpNopTests (tcu::TestContext& testCtx)
17590 de::MovePtr<tcu::TestCaseGroup> testGroup (new tcu::TestCaseGroup(testCtx, "opnop", "Test OpNop"));
17591 RGBA defaultColors[4];
17592 map<string, string> opNopFragments;
17594 getDefaultColors(defaultColors);
17596 opNopFragments["testfun"] =
17597 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
17598 "%param1 = OpFunctionParameter %v4f32\n"
17599 "%label_testfun = OpLabel\n"
17608 "%a = OpVectorExtractDynamic %f32 %param1 %c_i32_0\n"
17609 "%b = OpFAdd %f32 %a %a\n"
17611 "%c = OpFSub %f32 %b %a\n"
17612 "%ret = OpVectorInsertDynamic %v4f32 %param1 %c %c_i32_0\n"
17615 "OpReturnValue %ret\n"
17618 createTestsForAllStages("opnop", defaultColors, defaultColors, opNopFragments, testGroup.get());
17620 return testGroup.release();
17623 tcu::TestCaseGroup* createOpNameTests (tcu::TestContext& testCtx)
17625 de::MovePtr<tcu::TestCaseGroup> testGroup (new tcu::TestCaseGroup(testCtx, "opname","Test OpName"));
17626 RGBA defaultColors[4];
17627 map<string, string> opNameFragments;
17629 getDefaultColors(defaultColors);
17631 opNameFragments["testfun"] =
17632 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
17633 "%param1 = OpFunctionParameter %v4f32\n"
17634 "%label_func = OpLabel\n"
17635 "%a = OpVectorExtractDynamic %f32 %param1 %c_i32_0\n"
17636 "%b = OpFAdd %f32 %a %a\n"
17637 "%c = OpFSub %f32 %b %a\n"
17638 "%ret = OpVectorInsertDynamic %v4f32 %param1 %c %c_i32_0\n"
17639 "OpReturnValue %ret\n"
17642 opNameFragments["debug"] =
17643 "OpName %BP_main \"not_main\"";
17645 createTestsForAllStages("opname", defaultColors, defaultColors, opNameFragments, testGroup.get());
17647 return testGroup.release();
17650 tcu::TestCaseGroup* createFloat16Tests (tcu::TestContext& testCtx)
17652 de::MovePtr<tcu::TestCaseGroup> testGroup (new tcu::TestCaseGroup(testCtx, "float16", "Float 16 tests"));
17654 testGroup->addChild(createOpConstantFloat16Tests(testCtx));
17655 testGroup->addChild(createFloat16LogicalSet<GraphicsResources>(testCtx, TEST_WITH_NAN));
17656 testGroup->addChild(createFloat16LogicalSet<GraphicsResources>(testCtx, TEST_WITHOUT_NAN));
17657 testGroup->addChild(createFloat16FuncSet<GraphicsResources>(testCtx));
17658 testGroup->addChild(createDerivativeTests<256, 1>(testCtx));
17659 testGroup->addChild(createDerivativeTests<256, 2>(testCtx));
17660 testGroup->addChild(createDerivativeTests<256, 4>(testCtx));
17661 testGroup->addChild(createFloat16VectorExtractSet<GraphicsResources>(testCtx));
17662 testGroup->addChild(createFloat16VectorInsertSet<GraphicsResources>(testCtx));
17663 testGroup->addChild(createFloat16VectorShuffleSet<GraphicsResources>(testCtx));
17664 testGroup->addChild(createFloat16CompositeConstructSet<GraphicsResources>(testCtx));
17665 testGroup->addChild(createFloat16CompositeInsertExtractSet<GraphicsResources>(testCtx, "OpCompositeExtract"));
17666 testGroup->addChild(createFloat16CompositeInsertExtractSet<GraphicsResources>(testCtx, "OpCompositeInsert"));
17667 testGroup->addChild(createFloat16ArithmeticSet<GraphicsResources>(testCtx));
17668 testGroup->addChild(createFloat16ArithmeticSet<1, GraphicsResources>(testCtx));
17669 testGroup->addChild(createFloat16ArithmeticSet<2, GraphicsResources>(testCtx));
17670 testGroup->addChild(createFloat16ArithmeticSet<3, GraphicsResources>(testCtx));
17671 testGroup->addChild(createFloat16ArithmeticSet<4, GraphicsResources>(testCtx));
17673 return testGroup.release();
17676 tcu::TestCaseGroup* createFloat16Group (tcu::TestContext& testCtx)
17678 de::MovePtr<tcu::TestCaseGroup> testGroup (new tcu::TestCaseGroup(testCtx, "float16", "Float 16 tests"));
17680 testGroup->addChild(createFloat16OpConstantCompositeGroup(testCtx));
17681 testGroup->addChild(createFloat16LogicalSet<ComputeShaderSpec>(testCtx, TEST_WITH_NAN));
17682 testGroup->addChild(createFloat16LogicalSet<ComputeShaderSpec>(testCtx, TEST_WITHOUT_NAN));
17683 testGroup->addChild(createFloat16FuncSet<ComputeShaderSpec>(testCtx));
17684 testGroup->addChild(createFloat16VectorExtractSet<ComputeShaderSpec>(testCtx));
17685 testGroup->addChild(createFloat16VectorInsertSet<ComputeShaderSpec>(testCtx));
17686 testGroup->addChild(createFloat16VectorShuffleSet<ComputeShaderSpec>(testCtx));
17687 testGroup->addChild(createFloat16CompositeConstructSet<ComputeShaderSpec>(testCtx));
17688 testGroup->addChild(createFloat16CompositeInsertExtractSet<ComputeShaderSpec>(testCtx, "OpCompositeExtract"));
17689 testGroup->addChild(createFloat16CompositeInsertExtractSet<ComputeShaderSpec>(testCtx, "OpCompositeInsert"));
17690 testGroup->addChild(createFloat16ArithmeticSet<ComputeShaderSpec>(testCtx));
17691 testGroup->addChild(createFloat16ArithmeticSet<1, ComputeShaderSpec>(testCtx));
17692 testGroup->addChild(createFloat16ArithmeticSet<2, ComputeShaderSpec>(testCtx));
17693 testGroup->addChild(createFloat16ArithmeticSet<3, ComputeShaderSpec>(testCtx));
17694 testGroup->addChild(createFloat16ArithmeticSet<4, ComputeShaderSpec>(testCtx));
17696 return testGroup.release();
17699 tcu::TestCaseGroup* createBoolMixedBitSizeGroup (tcu::TestContext& testCtx)
17701 de::MovePtr<tcu::TestCaseGroup> group (new tcu::TestCaseGroup(testCtx, "mixed_bitsize", "Tests boolean operands produced from instructions of different bit-sizes"));
17703 de::Random rnd (deStringHash(group->getName()));
17704 const int numElements = 100;
17705 vector<float> inputData (numElements, 0);
17706 vector<float> outputData (numElements, 0);
17707 fillRandomScalars(rnd, 0.0f, 100.0f, &inputData[0], 100);
17709 const StringTemplate shaderTemplate (
17711 "OpMemoryModel Logical GLSL450\n"
17712 "OpEntryPoint GLCompute %main \"main\" %id\n"
17713 "OpExecutionMode %main LocalSize 1 1 1\n"
17714 "OpSource GLSL 430\n"
17715 "OpName %main \"main\"\n"
17716 "OpName %id \"gl_GlobalInvocationID\"\n"
17718 "OpDecorate %id BuiltIn GlobalInvocationId\n"
17720 + string(getComputeAsmInputOutputBufferTraits()) + string(getComputeAsmCommonTypes()) + string(getComputeAsmInputOutputBuffer()) +
17722 "%id = OpVariable %uvec3ptr Input\n"
17724 "%main = OpFunction %void None %voidf\n"
17725 "%label = OpLabel\n"
17726 "%idval = OpLoad %uvec3 %id\n"
17727 "%x = OpCompositeExtract %u32 %idval 0\n"
17728 "%inloc = OpAccessChain %f32ptr %indata %c0i32 %x\n"
17732 "%outloc = OpAccessChain %f32ptr %outdata %c0i32 %x\n"
17733 " OpStore %outloc %res\n"
17738 // Each test case produces 4 boolean values, and we want each of these values
17739 // to come froma different combination of the available bit-sizes, so compute
17740 // all possible combinations here.
17741 vector<deUint32> widths;
17742 widths.push_back(32);
17743 widths.push_back(16);
17744 widths.push_back(8);
17746 vector<IVec4> cases;
17747 for (size_t width0 = 0; width0 < widths.size(); width0++)
17749 for (size_t width1 = 0; width1 < widths.size(); width1++)
17751 for (size_t width2 = 0; width2 < widths.size(); width2++)
17753 for (size_t width3 = 0; width3 < widths.size(); width3++)
17755 cases.push_back(IVec4(widths[width0], widths[width1], widths[width2], widths[width3]));
17761 for (size_t caseNdx = 0; caseNdx < cases.size(); caseNdx++)
17763 /// Skip cases where all bitsizes are the same, we are only interested in testing booleans produced from instructions with different native bit-sizes
17764 if (cases[caseNdx][0] == cases[caseNdx][1] && cases[caseNdx][0] == cases[caseNdx][2] && cases[caseNdx][0] == cases[caseNdx][3])
17767 map<string, string> specializations;
17768 ComputeShaderSpec spec;
17770 // Inject appropriate capabilities and reference constants depending
17771 // on the bit-sizes required by this test case
17772 bool hasFloat32 = cases[caseNdx][0] == 32 || cases[caseNdx][1] == 32 || cases[caseNdx][2] == 32 || cases[caseNdx][3] == 32;
17773 bool hasFloat16 = cases[caseNdx][0] == 16 || cases[caseNdx][1] == 16 || cases[caseNdx][2] == 16 || cases[caseNdx][3] == 16;
17774 bool hasInt8 = cases[caseNdx][0] == 8 || cases[caseNdx][1] == 8 || cases[caseNdx][2] == 8 || cases[caseNdx][3] == 8;
17776 string capsStr = "OpCapability Shader\n";
17778 "%c0i32 = OpConstant %i32 0\n"
17779 "%c1f32 = OpConstant %f32 1.0\n"
17780 "%c0f32 = OpConstant %f32 0.0\n";
17785 "%c10f32 = OpConstant %f32 10.0\n"
17786 "%c25f32 = OpConstant %f32 25.0\n"
17787 "%c50f32 = OpConstant %f32 50.0\n"
17788 "%c90f32 = OpConstant %f32 90.0\n";
17793 capsStr += "OpCapability Float16\n";
17795 "%f16 = OpTypeFloat 16\n"
17796 "%c10f16 = OpConstant %f16 10.0\n"
17797 "%c25f16 = OpConstant %f16 25.0\n"
17798 "%c50f16 = OpConstant %f16 50.0\n"
17799 "%c90f16 = OpConstant %f16 90.0\n";
17804 capsStr += "OpCapability Int8\n";
17806 "%i8 = OpTypeInt 8 1\n"
17807 "%c10i8 = OpConstant %i8 10\n"
17808 "%c25i8 = OpConstant %i8 25\n"
17809 "%c50i8 = OpConstant %i8 50\n"
17810 "%c90i8 = OpConstant %i8 90\n";
17813 // Each invocation reads a different float32 value as input. Depending on
17814 // the bit-sizes required by the particular test case, we also produce
17815 // float16 and/or and int8 values by converting from the 32-bit float.
17816 string testStr = "";
17817 testStr += "%inval32 = OpLoad %f32 %inloc\n";
17819 testStr += "%inval16 = OpFConvert %f16 %inval32\n";
17821 testStr += "%inval8 = OpConvertFToS %i8 %inval32\n";
17823 // Because conversions from Float to Int round towards 0 we want our "greater" comparisons to be >=,
17824 // that way a float32/float16 comparison such as 50.6f >= 50.0f will preserve its result
17825 // when converted to int8, since FtoS(50.6f) results in 50. For "less" comparisons, it is the
17826 // other way around, so in this case we want < instead of <=.
17827 if (cases[caseNdx][0] == 32)
17828 testStr += "%cmp1 = OpFOrdGreaterThanEqual %bool %inval32 %c25f32\n";
17829 else if (cases[caseNdx][0] == 16)
17830 testStr += "%cmp1 = OpFOrdGreaterThanEqual %bool %inval16 %c25f16\n";
17832 testStr += "%cmp1 = OpSGreaterThanEqual %bool %inval8 %c25i8\n";
17834 if (cases[caseNdx][1] == 32)
17835 testStr += "%cmp2 = OpFOrdLessThan %bool %inval32 %c50f32\n";
17836 else if (cases[caseNdx][1] == 16)
17837 testStr += "%cmp2 = OpFOrdLessThan %bool %inval16 %c50f16\n";
17839 testStr += "%cmp2 = OpSLessThan %bool %inval8 %c50i8\n";
17841 if (cases[caseNdx][2] == 32)
17842 testStr += "%cmp3 = OpFOrdLessThan %bool %inval32 %c10f32\n";
17843 else if (cases[caseNdx][2] == 16)
17844 testStr += "%cmp3 = OpFOrdLessThan %bool %inval16 %c10f16\n";
17846 testStr += "%cmp3 = OpSLessThan %bool %inval8 %c10i8\n";
17848 if (cases[caseNdx][3] == 32)
17849 testStr += "%cmp4 = OpFOrdGreaterThanEqual %bool %inval32 %c90f32\n";
17850 else if (cases[caseNdx][3] == 16)
17851 testStr += "%cmp4 = OpFOrdGreaterThanEqual %bool %inval16 %c90f16\n";
17853 testStr += "%cmp4 = OpSGreaterThanEqual %bool %inval8 %c90i8\n";
17855 testStr += "%and1 = OpLogicalAnd %bool %cmp1 %cmp2\n";
17856 testStr += "%or1 = OpLogicalOr %bool %cmp3 %cmp4\n";
17857 testStr += "%or2 = OpLogicalOr %bool %and1 %or1\n";
17858 testStr += "%not1 = OpLogicalNot %bool %or2\n";
17859 testStr += "%res = OpSelect %f32 %not1 %c1f32 %c0f32\n";
17861 specializations["CAPS"] = capsStr;
17862 specializations["CONST"] = constStr;
17863 specializations["TEST"] = testStr;
17865 // Compute expected result by evaluating the boolean expression computed in the shader for each input value
17866 for (size_t ndx = 0; ndx < numElements; ++ndx)
17867 outputData[ndx] = !((inputData[ndx] >= 25.0f && inputData[ndx] < 50.0f) || (inputData[ndx] < 10.0f || inputData[ndx] >= 90.0f));
17869 spec.assembly = shaderTemplate.specialize(specializations);
17870 spec.inputs.push_back(BufferSp(new Float32Buffer(inputData)));
17871 spec.outputs.push_back(BufferSp(new Float32Buffer(outputData)));
17872 spec.numWorkGroups = IVec3(numElements, 1, 1);
17874 spec.requestedVulkanFeatures.extFloat16Int8 |= EXTFLOAT16INT8FEATURES_FLOAT16;
17876 spec.requestedVulkanFeatures.extFloat16Int8 |= EXTFLOAT16INT8FEATURES_INT8;
17877 spec.extensions.push_back("VK_KHR_shader_float16_int8");
17879 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]);
17880 group->addChild(new SpvAsmComputeShaderCase(testCtx, testName.c_str(), "", spec));
17883 return group.release();
17886 tcu::TestCaseGroup* createBoolGroup (tcu::TestContext& testCtx)
17888 de::MovePtr<tcu::TestCaseGroup> testGroup (new tcu::TestCaseGroup(testCtx, "bool", "Boolean tests"));
17890 testGroup->addChild(createBoolMixedBitSizeGroup(testCtx));
17892 return testGroup.release();
17895 tcu::TestCaseGroup* createOpNameAbuseTests (tcu::TestContext& testCtx)
17897 de::MovePtr<tcu::TestCaseGroup> abuseGroup(new tcu::TestCaseGroup(testCtx, "opname_abuse", "OpName abuse tests"));
17898 vector<CaseParameter> abuseCases;
17899 RGBA defaultColors[4];
17900 map<string, string> opNameFragments;
17902 getOpNameAbuseCases(abuseCases);
17903 getDefaultColors(defaultColors);
17905 opNameFragments["testfun"] =
17906 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
17907 "%param1 = OpFunctionParameter %v4f32\n"
17908 "%label_func = OpLabel\n"
17909 "%a = OpVectorExtractDynamic %f32 %param1 %c_i32_0\n"
17910 "%b = OpFAdd %f32 %a %a\n"
17911 "%c = OpFSub %f32 %b %a\n"
17912 "%ret = OpVectorInsertDynamic %v4f32 %param1 %c %c_i32_0\n"
17913 "OpReturnValue %ret\n"
17916 for (unsigned int i = 0; i < abuseCases.size(); i++)
17919 casename = string("main") + abuseCases[i].name;
17921 opNameFragments["debug"] =
17922 "OpName %BP_main \"" + abuseCases[i].param + "\"";
17924 createTestsForAllStages(casename, defaultColors, defaultColors, opNameFragments, abuseGroup.get());
17927 for (unsigned int i = 0; i < abuseCases.size(); i++)
17930 casename = string("b") + abuseCases[i].name;
17932 opNameFragments["debug"] =
17933 "OpName %b \"" + abuseCases[i].param + "\"";
17935 createTestsForAllStages(casename, defaultColors, defaultColors, opNameFragments, abuseGroup.get());
17939 opNameFragments["debug"] =
17940 "OpName %test_code \"name1\"\n"
17941 "OpName %param1 \"name2\"\n"
17942 "OpName %a \"name3\"\n"
17943 "OpName %b \"name4\"\n"
17944 "OpName %c \"name5\"\n"
17945 "OpName %ret \"name6\"\n";
17947 createTestsForAllStages("everything_named", defaultColors, defaultColors, opNameFragments, abuseGroup.get());
17951 opNameFragments["debug"] =
17952 "OpName %test_code \"the_same\"\n"
17953 "OpName %param1 \"the_same\"\n"
17954 "OpName %a \"the_same\"\n"
17955 "OpName %b \"the_same\"\n"
17956 "OpName %c \"the_same\"\n"
17957 "OpName %ret \"the_same\"\n";
17959 createTestsForAllStages("everything_named_the_same", defaultColors, defaultColors, opNameFragments, abuseGroup.get());
17963 opNameFragments["debug"] =
17964 "OpName %BP_main \"to_be\"\n"
17965 "OpName %BP_main \"or_not\"\n"
17966 "OpName %BP_main \"to_be\"\n";
17968 createTestsForAllStages("main_has_multiple_names", defaultColors, defaultColors, opNameFragments, abuseGroup.get());
17972 opNameFragments["debug"] =
17973 "OpName %b \"to_be\"\n"
17974 "OpName %b \"or_not\"\n"
17975 "OpName %b \"to_be\"\n";
17977 createTestsForAllStages("b_has_multiple_names", defaultColors, defaultColors, opNameFragments, abuseGroup.get());
17980 return abuseGroup.release();
17984 tcu::TestCaseGroup* createOpMemberNameAbuseTests (tcu::TestContext& testCtx)
17986 de::MovePtr<tcu::TestCaseGroup> abuseGroup(new tcu::TestCaseGroup(testCtx, "opmembername_abuse", "OpName abuse tests"));
17987 vector<CaseParameter> abuseCases;
17988 RGBA defaultColors[4];
17989 map<string, string> opMemberNameFragments;
17991 getOpNameAbuseCases(abuseCases);
17992 getDefaultColors(defaultColors);
17994 opMemberNameFragments["pre_main"] =
17995 "%f3str = OpTypeStruct %f32 %f32 %f32\n";
17997 opMemberNameFragments["testfun"] =
17998 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
17999 "%param1 = OpFunctionParameter %v4f32\n"
18000 "%label_func = OpLabel\n"
18001 "%a = OpVectorExtractDynamic %f32 %param1 %c_i32_0\n"
18002 "%b = OpFAdd %f32 %a %a\n"
18003 "%c = OpFSub %f32 %b %a\n"
18004 "%cstr = OpCompositeConstruct %f3str %c %c %c\n"
18005 "%d = OpCompositeExtract %f32 %cstr 0\n"
18006 "%ret = OpVectorInsertDynamic %v4f32 %param1 %d %c_i32_0\n"
18007 "OpReturnValue %ret\n"
18010 for (unsigned int i = 0; i < abuseCases.size(); i++)
18013 casename = string("f3str_x") + abuseCases[i].name;
18015 opMemberNameFragments["debug"] =
18016 "OpMemberName %f3str 0 \"" + abuseCases[i].param + "\"";
18018 createTestsForAllStages(casename, defaultColors, defaultColors, opMemberNameFragments, abuseGroup.get());
18022 opMemberNameFragments["debug"] =
18023 "OpMemberName %f3str 0 \"name1\"\n"
18024 "OpMemberName %f3str 1 \"name2\"\n"
18025 "OpMemberName %f3str 2 \"name3\"\n";
18027 createTestsForAllStages("everything_named", defaultColors, defaultColors, opMemberNameFragments, abuseGroup.get());
18031 opMemberNameFragments["debug"] =
18032 "OpMemberName %f3str 0 \"the_same\"\n"
18033 "OpMemberName %f3str 1 \"the_same\"\n"
18034 "OpMemberName %f3str 2 \"the_same\"\n";
18036 createTestsForAllStages("everything_named_the_same", defaultColors, defaultColors, opMemberNameFragments, abuseGroup.get());
18040 opMemberNameFragments["debug"] =
18041 "OpMemberName %f3str 0 \"to_be\"\n"
18042 "OpMemberName %f3str 1 \"or_not\"\n"
18043 "OpMemberName %f3str 0 \"to_be\"\n"
18044 "OpMemberName %f3str 2 \"makes_no\"\n"
18045 "OpMemberName %f3str 0 \"difference\"\n"
18046 "OpMemberName %f3str 0 \"to_me\"\n";
18049 createTestsForAllStages("f3str_x_has_multiple_names", defaultColors, defaultColors, opMemberNameFragments, abuseGroup.get());
18052 return abuseGroup.release();
18055 vector<deUint32> getSparseIdsAbuseData (const deUint32 numDataPoints, const deUint32 seed)
18057 vector<deUint32> result;
18058 de::Random rnd (seed);
18060 result.reserve(numDataPoints);
18062 for (deUint32 dataPointNdx = 0; dataPointNdx < numDataPoints; ++dataPointNdx)
18063 result.push_back(rnd.getUint32());
18068 vector<deUint32> getSparseIdsAbuseResults (const vector<deUint32>& inData1, const vector<deUint32>& inData2)
18070 vector<deUint32> result;
18072 result.reserve(inData1.size());
18074 for (size_t dataPointNdx = 0; dataPointNdx < inData1.size(); ++dataPointNdx)
18075 result.push_back(inData1[dataPointNdx] + inData2[dataPointNdx]);
18080 template<class SpecResource>
18081 void createSparseIdsAbuseTest (tcu::TestContext& testCtx, de::MovePtr<tcu::TestCaseGroup>& testGroup)
18083 const deUint32 numDataPoints = 16;
18084 const std::string testName ("sparse_ids");
18085 const deUint32 seed (deStringHash(testName.c_str()));
18086 const vector<deUint32> inData1 (getSparseIdsAbuseData(numDataPoints, seed + 1));
18087 const vector<deUint32> inData2 (getSparseIdsAbuseData(numDataPoints, seed + 2));
18088 const vector<deUint32> outData (getSparseIdsAbuseResults(inData1, inData2));
18089 const StringTemplate preMain
18091 "%c_i32_ndp = OpConstant %i32 ${num_data_points}\n"
18092 " %up_u32 = OpTypePointer Uniform %u32\n"
18093 " %ra_u32 = OpTypeArray %u32 %c_i32_ndp\n"
18094 " %SSBO32 = OpTypeStruct %ra_u32\n"
18095 "%up_SSBO32 = OpTypePointer Uniform %SSBO32\n"
18096 "%ssbo_src0 = OpVariable %up_SSBO32 Uniform\n"
18097 "%ssbo_src1 = OpVariable %up_SSBO32 Uniform\n"
18098 " %ssbo_dst = OpVariable %up_SSBO32 Uniform\n"
18100 const StringTemplate decoration
18102 "OpDecorate %ra_u32 ArrayStride 4\n"
18103 "OpMemberDecorate %SSBO32 0 Offset 0\n"
18104 "OpDecorate %SSBO32 BufferBlock\n"
18105 "OpDecorate %ssbo_src0 DescriptorSet 0\n"
18106 "OpDecorate %ssbo_src0 Binding 0\n"
18107 "OpDecorate %ssbo_src1 DescriptorSet 0\n"
18108 "OpDecorate %ssbo_src1 Binding 1\n"
18109 "OpDecorate %ssbo_dst DescriptorSet 0\n"
18110 "OpDecorate %ssbo_dst Binding 2\n"
18112 const StringTemplate testFun
18114 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
18115 " %param = OpFunctionParameter %v4f32\n"
18117 " %entry = OpLabel\n"
18118 " %i = OpVariable %fp_i32 Function\n"
18119 " OpStore %i %c_i32_0\n"
18120 " OpBranch %loop\n"
18122 " %loop = OpLabel\n"
18123 " %i_cmp = OpLoad %i32 %i\n"
18124 " %lt = OpSLessThan %bool %i_cmp %c_i32_ndp\n"
18125 " OpLoopMerge %merge %next None\n"
18126 " OpBranchConditional %lt %write %merge\n"
18128 " %write = OpLabel\n"
18129 " %ndx = OpLoad %i32 %i\n"
18131 " %127 = OpAccessChain %up_u32 %ssbo_src0 %c_i32_0 %ndx\n"
18132 " %128 = OpLoad %u32 %127\n"
18134 // The test relies on SPIR-V compiler option SPV_TEXT_TO_BINARY_OPTION_PRESERVE_NUMERIC_IDS set in assembleSpirV()
18135 " %4194000 = OpAccessChain %up_u32 %ssbo_src1 %c_i32_0 %ndx\n"
18136 " %4194001 = OpLoad %u32 %4194000\n"
18138 " %2097151 = OpIAdd %u32 %128 %4194001\n"
18139 " %2097152 = OpAccessChain %up_u32 %ssbo_dst %c_i32_0 %ndx\n"
18140 " OpStore %2097152 %2097151\n"
18141 " OpBranch %next\n"
18143 " %next = OpLabel\n"
18144 " %i_cur = OpLoad %i32 %i\n"
18145 " %i_new = OpIAdd %i32 %i_cur %c_i32_1\n"
18146 " OpStore %i %i_new\n"
18147 " OpBranch %loop\n"
18149 " %merge = OpLabel\n"
18150 " OpReturnValue %param\n"
18154 SpecResource specResource;
18155 map<string, string> specs;
18156 VulkanFeatures features;
18157 map<string, string> fragments;
18158 vector<string> extensions;
18160 specs["num_data_points"] = de::toString(numDataPoints);
18162 fragments["decoration"] = decoration.specialize(specs);
18163 fragments["pre_main"] = preMain.specialize(specs);
18164 fragments["testfun"] = testFun.specialize(specs);
18166 specResource.inputs.push_back(Resource(BufferSp(new Uint32Buffer(inData1)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
18167 specResource.inputs.push_back(Resource(BufferSp(new Uint32Buffer(inData2)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
18168 specResource.outputs.push_back(Resource(BufferSp(new Uint32Buffer(outData)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
18170 features.coreFeatures.vertexPipelineStoresAndAtomics = true;
18171 features.coreFeatures.fragmentStoresAndAtomics = true;
18173 finalizeTestsCreation(specResource, fragments, testCtx, *testGroup.get(), testName, features, extensions, IVec3(1, 1, 1));
18176 vector<deUint32> getLotsIdsAbuseData (const deUint32 numDataPoints, const deUint32 seed)
18178 vector<deUint32> result;
18179 de::Random rnd (seed);
18181 result.reserve(numDataPoints);
18184 result.push_back(1u);
18187 for (deUint32 dataPointNdx = 1; dataPointNdx < numDataPoints; ++dataPointNdx)
18188 result.push_back(rnd.getUint8());
18193 vector<deUint32> getLotsIdsAbuseResults (const vector<deUint32>& inData1, const vector<deUint32>& inData2, const deUint32 count)
18195 vector<deUint32> result;
18197 result.reserve(inData1.size());
18199 for (size_t dataPointNdx = 0; dataPointNdx < inData1.size(); ++dataPointNdx)
18200 result.push_back(inData1[dataPointNdx] + count * inData2[dataPointNdx]);
18205 template<class SpecResource>
18206 void createLotsIdsAbuseTest (tcu::TestContext& testCtx, de::MovePtr<tcu::TestCaseGroup>& testGroup)
18208 const deUint32 numDataPoints = 16;
18209 const deUint32 firstNdx = 100u;
18210 const deUint32 sequenceCount = 10000u;
18211 const std::string testName ("lots_ids");
18212 const deUint32 seed (deStringHash(testName.c_str()));
18213 const vector<deUint32> inData1 (getLotsIdsAbuseData(numDataPoints, seed + 1));
18214 const vector<deUint32> inData2 (getLotsIdsAbuseData(numDataPoints, seed + 2));
18215 const vector<deUint32> outData (getLotsIdsAbuseResults(inData1, inData2, sequenceCount));
18216 const StringTemplate preMain
18218 "%c_i32_ndp = OpConstant %i32 ${num_data_points}\n"
18219 " %up_u32 = OpTypePointer Uniform %u32\n"
18220 " %ra_u32 = OpTypeArray %u32 %c_i32_ndp\n"
18221 " %SSBO32 = OpTypeStruct %ra_u32\n"
18222 "%up_SSBO32 = OpTypePointer Uniform %SSBO32\n"
18223 "%ssbo_src0 = OpVariable %up_SSBO32 Uniform\n"
18224 "%ssbo_src1 = OpVariable %up_SSBO32 Uniform\n"
18225 " %ssbo_dst = OpVariable %up_SSBO32 Uniform\n"
18227 const StringTemplate decoration
18229 "OpDecorate %ra_u32 ArrayStride 4\n"
18230 "OpMemberDecorate %SSBO32 0 Offset 0\n"
18231 "OpDecorate %SSBO32 BufferBlock\n"
18232 "OpDecorate %ssbo_src0 DescriptorSet 0\n"
18233 "OpDecorate %ssbo_src0 Binding 0\n"
18234 "OpDecorate %ssbo_src1 DescriptorSet 0\n"
18235 "OpDecorate %ssbo_src1 Binding 1\n"
18236 "OpDecorate %ssbo_dst DescriptorSet 0\n"
18237 "OpDecorate %ssbo_dst Binding 2\n"
18239 const StringTemplate testFun
18241 "%test_code = OpFunction %v4f32 None %v4f32_v4f32_function\n"
18242 " %param = OpFunctionParameter %v4f32\n"
18244 " %entry = OpLabel\n"
18245 " %i = OpVariable %fp_i32 Function\n"
18246 " OpStore %i %c_i32_0\n"
18247 " OpBranch %loop\n"
18249 " %loop = OpLabel\n"
18250 " %i_cmp = OpLoad %i32 %i\n"
18251 " %lt = OpSLessThan %bool %i_cmp %c_i32_ndp\n"
18252 " OpLoopMerge %merge %next None\n"
18253 " OpBranchConditional %lt %write %merge\n"
18255 " %write = OpLabel\n"
18256 " %ndx = OpLoad %i32 %i\n"
18258 " %90 = OpAccessChain %up_u32 %ssbo_src1 %c_i32_0 %ndx\n"
18259 " %91 = OpLoad %u32 %90\n"
18261 " %98 = OpAccessChain %up_u32 %ssbo_src0 %c_i32_0 %ndx\n"
18262 " %${zeroth_id} = OpLoad %u32 %98\n"
18266 // The test relies on SPIR-V compiler option SPV_TEXT_TO_BINARY_OPTION_PRESERVE_NUMERIC_IDS set in assembleSpirV()
18267 " %dst = OpAccessChain %up_u32 %ssbo_dst %c_i32_0 %ndx\n"
18268 " OpStore %dst %${last_id}\n"
18269 " OpBranch %next\n"
18271 " %next = OpLabel\n"
18272 " %i_cur = OpLoad %i32 %i\n"
18273 " %i_new = OpIAdd %i32 %i_cur %c_i32_1\n"
18274 " OpStore %i %i_new\n"
18275 " OpBranch %loop\n"
18277 " %merge = OpLabel\n"
18278 " OpReturnValue %param\n"
18282 deUint32 lastId = firstNdx;
18283 SpecResource specResource;
18284 map<string, string> specs;
18285 VulkanFeatures features;
18286 map<string, string> fragments;
18287 vector<string> extensions;
18288 std::string sequence;
18290 for (deUint32 sequenceNdx = 0; sequenceNdx < sequenceCount; ++sequenceNdx)
18292 const deUint32 sequenceId = sequenceNdx + firstNdx;
18293 const std::string sequenceIdStr = de::toString(sequenceId);
18295 sequence += "%" + sequenceIdStr + " = OpIAdd %u32 %91 %" + de::toString(sequenceId - 1) + "\n";
18296 lastId = sequenceId;
18298 if (sequenceNdx == 0)
18299 sequence.reserve((10 + sequence.length()) * sequenceCount);
18302 specs["num_data_points"] = de::toString(numDataPoints);
18303 specs["zeroth_id"] = de::toString(firstNdx - 1);
18304 specs["last_id"] = de::toString(lastId);
18305 specs["seq"] = sequence;
18307 fragments["decoration"] = decoration.specialize(specs);
18308 fragments["pre_main"] = preMain.specialize(specs);
18309 fragments["testfun"] = testFun.specialize(specs);
18311 specResource.inputs.push_back(Resource(BufferSp(new Uint32Buffer(inData1)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
18312 specResource.inputs.push_back(Resource(BufferSp(new Uint32Buffer(inData2)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
18313 specResource.outputs.push_back(Resource(BufferSp(new Uint32Buffer(outData)), VK_DESCRIPTOR_TYPE_STORAGE_BUFFER));
18315 features.coreFeatures.vertexPipelineStoresAndAtomics = true;
18316 features.coreFeatures.fragmentStoresAndAtomics = true;
18318 finalizeTestsCreation(specResource, fragments, testCtx, *testGroup.get(), testName, features, extensions, IVec3(1, 1, 1));
18321 tcu::TestCaseGroup* createSpirvIdsAbuseTests (tcu::TestContext& testCtx)
18323 de::MovePtr<tcu::TestCaseGroup> testGroup (new tcu::TestCaseGroup(testCtx, "spirv_ids_abuse", "SPIR-V abuse tests"));
18325 createSparseIdsAbuseTest<GraphicsResources>(testCtx, testGroup);
18326 createLotsIdsAbuseTest<GraphicsResources>(testCtx, testGroup);
18328 return testGroup.release();
18331 tcu::TestCaseGroup* createSpirvIdsAbuseGroup (tcu::TestContext& testCtx)
18333 de::MovePtr<tcu::TestCaseGroup> testGroup (new tcu::TestCaseGroup(testCtx, "spirv_ids_abuse", "SPIR-V abuse tests"));
18335 createSparseIdsAbuseTest<ComputeShaderSpec>(testCtx, testGroup);
18336 createLotsIdsAbuseTest<ComputeShaderSpec>(testCtx, testGroup);
18338 return testGroup.release();
18341 tcu::TestCaseGroup* createInstructionTests (tcu::TestContext& testCtx)
18343 const bool testComputePipeline = true;
18345 de::MovePtr<tcu::TestCaseGroup> instructionTests (new tcu::TestCaseGroup(testCtx, "instruction", "Instructions with special opcodes/operands"));
18346 de::MovePtr<tcu::TestCaseGroup> computeTests (new tcu::TestCaseGroup(testCtx, "compute", "Compute Instructions with special opcodes/operands"));
18347 de::MovePtr<tcu::TestCaseGroup> graphicsTests (new tcu::TestCaseGroup(testCtx, "graphics", "Graphics Instructions with special opcodes/operands"));
18349 computeTests->addChild(createSpivVersionCheckTests(testCtx, testComputePipeline));
18350 computeTests->addChild(createLocalSizeGroup(testCtx));
18351 computeTests->addChild(createOpNopGroup(testCtx));
18352 computeTests->addChild(createOpFUnordGroup(testCtx, TEST_WITHOUT_NAN));
18353 computeTests->addChild(createOpFUnordGroup(testCtx, TEST_WITH_NAN));
18354 computeTests->addChild(createOpAtomicGroup(testCtx, false));
18355 computeTests->addChild(createOpAtomicGroup(testCtx, true)); // Using new StorageBuffer decoration
18356 computeTests->addChild(createOpAtomicGroup(testCtx, false, 1024, true)); // Return value validation
18357 computeTests->addChild(createOpLineGroup(testCtx));
18358 computeTests->addChild(createOpModuleProcessedGroup(testCtx));
18359 computeTests->addChild(createOpNoLineGroup(testCtx));
18360 computeTests->addChild(createOpConstantNullGroup(testCtx));
18361 computeTests->addChild(createOpConstantCompositeGroup(testCtx));
18362 computeTests->addChild(createOpConstantUsageGroup(testCtx));
18363 computeTests->addChild(createSpecConstantGroup(testCtx));
18364 computeTests->addChild(createOpSourceGroup(testCtx));
18365 computeTests->addChild(createOpSourceExtensionGroup(testCtx));
18366 computeTests->addChild(createDecorationGroupGroup(testCtx));
18367 computeTests->addChild(createOpPhiGroup(testCtx));
18368 computeTests->addChild(createLoopControlGroup(testCtx));
18369 computeTests->addChild(createFunctionControlGroup(testCtx));
18370 computeTests->addChild(createSelectionControlGroup(testCtx));
18371 computeTests->addChild(createBlockOrderGroup(testCtx));
18372 computeTests->addChild(createMultipleShaderGroup(testCtx));
18373 computeTests->addChild(createMemoryAccessGroup(testCtx));
18374 computeTests->addChild(createOpCopyMemoryGroup(testCtx));
18375 computeTests->addChild(createOpCopyObjectGroup(testCtx));
18376 computeTests->addChild(createNoContractionGroup(testCtx));
18377 computeTests->addChild(createOpUndefGroup(testCtx));
18378 computeTests->addChild(createOpUnreachableGroup(testCtx));
18379 computeTests->addChild(createOpQuantizeToF16Group(testCtx));
18380 computeTests->addChild(createOpFRemGroup(testCtx));
18381 computeTests->addChild(createOpSRemComputeGroup(testCtx, QP_TEST_RESULT_PASS));
18382 computeTests->addChild(createOpSRemComputeGroup64(testCtx, QP_TEST_RESULT_PASS));
18383 computeTests->addChild(createOpSModComputeGroup(testCtx, QP_TEST_RESULT_PASS));
18384 computeTests->addChild(createOpSModComputeGroup64(testCtx, QP_TEST_RESULT_PASS));
18385 computeTests->addChild(createConvertComputeTests(testCtx, "OpSConvert", "sconvert"));
18386 computeTests->addChild(createConvertComputeTests(testCtx, "OpUConvert", "uconvert"));
18387 computeTests->addChild(createConvertComputeTests(testCtx, "OpFConvert", "fconvert"));
18388 computeTests->addChild(createConvertComputeTests(testCtx, "OpConvertSToF", "convertstof"));
18389 computeTests->addChild(createConvertComputeTests(testCtx, "OpConvertFToS", "convertftos"));
18390 computeTests->addChild(createConvertComputeTests(testCtx, "OpConvertUToF", "convertutof"));
18391 computeTests->addChild(createConvertComputeTests(testCtx, "OpConvertFToU", "convertftou"));
18392 computeTests->addChild(createOpCompositeInsertGroup(testCtx));
18393 computeTests->addChild(createOpInBoundsAccessChainGroup(testCtx));
18394 computeTests->addChild(createShaderDefaultOutputGroup(testCtx));
18395 computeTests->addChild(createOpNMinGroup(testCtx));
18396 computeTests->addChild(createOpNMaxGroup(testCtx));
18397 computeTests->addChild(createOpNClampGroup(testCtx));
18399 de::MovePtr<tcu::TestCaseGroup> computeAndroidTests (new tcu::TestCaseGroup(testCtx, "android", "Android CTS Tests"));
18401 computeAndroidTests->addChild(createOpSRemComputeGroup(testCtx, QP_TEST_RESULT_QUALITY_WARNING));
18402 computeAndroidTests->addChild(createOpSModComputeGroup(testCtx, QP_TEST_RESULT_QUALITY_WARNING));
18404 computeTests->addChild(computeAndroidTests.release());
18407 computeTests->addChild(create8BitStorageComputeGroup(testCtx));
18408 computeTests->addChild(create16BitStorageComputeGroup(testCtx));
18409 computeTests->addChild(createFloatControlsComputeGroup(testCtx));
18410 computeTests->addChild(createUboMatrixPaddingComputeGroup(testCtx));
18411 computeTests->addChild(createCompositeInsertComputeGroup(testCtx));
18412 computeTests->addChild(createVariableInitComputeGroup(testCtx));
18413 computeTests->addChild(createConditionalBranchComputeGroup(testCtx));
18414 computeTests->addChild(createIndexingComputeGroup(testCtx));
18415 computeTests->addChild(createVariablePointersComputeGroup(testCtx));
18416 computeTests->addChild(createPhysicalPointersComputeGroup(testCtx));
18417 computeTests->addChild(createImageSamplerComputeGroup(testCtx));
18418 computeTests->addChild(createOpNameGroup(testCtx));
18419 computeTests->addChild(createOpMemberNameGroup(testCtx));
18420 computeTests->addChild(createPointerParameterComputeGroup(testCtx));
18421 computeTests->addChild(createFloat16Group(testCtx));
18422 computeTests->addChild(createBoolGroup(testCtx));
18423 computeTests->addChild(createWorkgroupMemoryComputeGroup(testCtx));
18424 computeTests->addChild(createSpirvIdsAbuseGroup(testCtx));
18425 computeTests->addChild(createSignedIntCompareGroup(testCtx));
18427 graphicsTests->addChild(createCrossStageInterfaceTests(testCtx));
18428 graphicsTests->addChild(createSpivVersionCheckTests(testCtx, !testComputePipeline));
18429 graphicsTests->addChild(createOpNopTests(testCtx));
18430 graphicsTests->addChild(createOpSourceTests(testCtx));
18431 graphicsTests->addChild(createOpSourceContinuedTests(testCtx));
18432 graphicsTests->addChild(createOpModuleProcessedTests(testCtx));
18433 graphicsTests->addChild(createOpLineTests(testCtx));
18434 graphicsTests->addChild(createOpNoLineTests(testCtx));
18435 graphicsTests->addChild(createOpConstantNullTests(testCtx));
18436 graphicsTests->addChild(createOpConstantCompositeTests(testCtx));
18437 graphicsTests->addChild(createMemoryAccessTests(testCtx));
18438 graphicsTests->addChild(createOpUndefTests(testCtx));
18439 graphicsTests->addChild(createSelectionBlockOrderTests(testCtx));
18440 graphicsTests->addChild(createModuleTests(testCtx));
18441 graphicsTests->addChild(createSwitchBlockOrderTests(testCtx));
18442 graphicsTests->addChild(createOpPhiTests(testCtx));
18443 graphicsTests->addChild(createNoContractionTests(testCtx));
18444 graphicsTests->addChild(createOpQuantizeTests(testCtx));
18445 graphicsTests->addChild(createLoopTests(testCtx));
18446 graphicsTests->addChild(createSpecConstantTests(testCtx));
18447 graphicsTests->addChild(createSpecConstantOpQuantizeToF16Group(testCtx));
18448 graphicsTests->addChild(createBarrierTests(testCtx));
18449 graphicsTests->addChild(createDecorationGroupTests(testCtx));
18450 graphicsTests->addChild(createFRemTests(testCtx));
18451 graphicsTests->addChild(createOpSRemGraphicsTests(testCtx, QP_TEST_RESULT_PASS));
18452 graphicsTests->addChild(createOpSModGraphicsTests(testCtx, QP_TEST_RESULT_PASS));
18455 de::MovePtr<tcu::TestCaseGroup> graphicsAndroidTests (new tcu::TestCaseGroup(testCtx, "android", "Android CTS Tests"));
18457 graphicsAndroidTests->addChild(createOpSRemGraphicsTests(testCtx, QP_TEST_RESULT_QUALITY_WARNING));
18458 graphicsAndroidTests->addChild(createOpSModGraphicsTests(testCtx, QP_TEST_RESULT_QUALITY_WARNING));
18460 graphicsTests->addChild(graphicsAndroidTests.release());
18462 graphicsTests->addChild(createOpNameTests(testCtx));
18463 graphicsTests->addChild(createOpNameAbuseTests(testCtx));
18464 graphicsTests->addChild(createOpMemberNameAbuseTests(testCtx));
18466 graphicsTests->addChild(create8BitStorageGraphicsGroup(testCtx));
18467 graphicsTests->addChild(create16BitStorageGraphicsGroup(testCtx));
18468 graphicsTests->addChild(createFloatControlsGraphicsGroup(testCtx));
18469 graphicsTests->addChild(createUboMatrixPaddingGraphicsGroup(testCtx));
18470 graphicsTests->addChild(createCompositeInsertGraphicsGroup(testCtx));
18471 graphicsTests->addChild(createVariableInitGraphicsGroup(testCtx));
18472 graphicsTests->addChild(createConditionalBranchGraphicsGroup(testCtx));
18473 graphicsTests->addChild(createIndexingGraphicsGroup(testCtx));
18474 graphicsTests->addChild(createVariablePointersGraphicsGroup(testCtx));
18475 graphicsTests->addChild(createImageSamplerGraphicsGroup(testCtx));
18476 graphicsTests->addChild(createConvertGraphicsTests(testCtx, "OpSConvert", "sconvert"));
18477 graphicsTests->addChild(createConvertGraphicsTests(testCtx, "OpUConvert", "uconvert"));
18478 graphicsTests->addChild(createConvertGraphicsTests(testCtx, "OpFConvert", "fconvert"));
18479 graphicsTests->addChild(createConvertGraphicsTests(testCtx, "OpConvertSToF", "convertstof"));
18480 graphicsTests->addChild(createConvertGraphicsTests(testCtx, "OpConvertFToS", "convertftos"));
18481 graphicsTests->addChild(createConvertGraphicsTests(testCtx, "OpConvertUToF", "convertutof"));
18482 graphicsTests->addChild(createConvertGraphicsTests(testCtx, "OpConvertFToU", "convertftou"));
18483 graphicsTests->addChild(createPointerParameterGraphicsGroup(testCtx));
18484 graphicsTests->addChild(createVaryingNameGraphicsGroup(testCtx));
18485 graphicsTests->addChild(createFloat16Tests(testCtx));
18486 graphicsTests->addChild(createSpirvIdsAbuseTests(testCtx));
18488 instructionTests->addChild(computeTests.release());
18489 instructionTests->addChild(graphicsTests.release());
18491 return instructionTests.release();