/*------------------------------------------------------------------------- * drawElements Quality Program OpenGL ES 2.0 Module * ------------------------------------------------- * * Copyright 2014 The Android Open Source Project * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. * *//*! * \file * \brief Optimized vs unoptimized shader performance tests. *//*--------------------------------------------------------------------*/ #include "es2pShaderOptimizationTests.hpp" #include "glsShaderPerformanceMeasurer.hpp" #include "gluRenderContext.hpp" #include "gluShaderProgram.hpp" #include "tcuTestLog.hpp" #include "tcuVector.hpp" #include "tcuStringTemplate.hpp" #include "deSharedPtr.hpp" #include "deStringUtil.hpp" #include "deMath.h" #include "glwFunctions.hpp" #include #include #include using glu::ShaderProgram; using tcu::TestLog; using tcu::Vec4; using de::SharedPtr; using de::toString; using std::vector; using std::string; namespace deqp { using gls::ShaderPerformanceMeasurer; namespace gles2 { namespace Performance { static inline std::map singleMap (const string& key, const string& value) { std::map res; res[key] = value; return res; } static inline string repeat (const string& str, int numRepeats, const string& delim = "") { string result = str; for (int i = 1; i < numRepeats; i++) result += delim + str; return result; } static inline string repeatIndexedTemplate (const string& strTempl, int numRepeats, const string& delim = "", int ndxStart = 0) { const tcu::StringTemplate templ(strTempl); string result; std::map params; for (int i = 0; i < numRepeats; i++) { params["PREV_NDX"] = toString(i + ndxStart - 1); params["NDX"] = toString(i + ndxStart); result += (i > 0 ? delim : "") + templ.specialize(params); } return result; } namespace { enum CaseShaderType { CASESHADERTYPE_VERTEX = 0, CASESHADERTYPE_FRAGMENT, CASESHADERTYPE_LAST }; static inline string getShaderPrecision (CaseShaderType shaderType) { switch (shaderType) { case CASESHADERTYPE_VERTEX: return "highp"; case CASESHADERTYPE_FRAGMENT: return "mediump"; default: DE_ASSERT(false); return DE_NULL; } } struct ProgramData { glu::ProgramSources sources; vector attributes; //!< \note Shouldn't contain a_position; that one is set by gls::ShaderPerformanceMeasurer. ProgramData (void) {} ProgramData (const glu::ProgramSources& sources_, const vector& attributes_ = vector()) : sources(sources_), attributes(attributes_) {} ProgramData (const glu::ProgramSources& sources_, const gls::AttribSpec& attribute) : sources(sources_), attributes(1, attribute) {} }; //! Shader boilerplate helper; most cases have similar basic shader structure. static inline ProgramData defaultProgramData (CaseShaderType shaderType, const string& funcDefs, const string& mainStatements) { const bool isVertexCase = shaderType == CASESHADERTYPE_VERTEX; const bool isFragmentCase = shaderType == CASESHADERTYPE_FRAGMENT; const string vtxPrec = getShaderPrecision(CASESHADERTYPE_VERTEX); const string fragPrec = getShaderPrecision(CASESHADERTYPE_FRAGMENT); return ProgramData(glu::ProgramSources() << glu::VertexSource( "attribute " + vtxPrec + " vec4 a_position;\n" "attribute " + vtxPrec + " vec4 a_value;\n" "varying " + fragPrec + " vec4 v_value;\n" + (isVertexCase ? funcDefs : "") + "void main (void)\n" "{\n" " gl_Position = a_position;\n" " " + vtxPrec + " vec4 value = a_value;\n" + (isVertexCase ? mainStatements : "") + " v_value = value;\n" "}\n") << glu::FragmentSource( "varying " + fragPrec + " vec4 v_value;\n" + (isFragmentCase ? funcDefs : "") + "void main (void)\n" "{\n" " " + fragPrec + " vec4 value = v_value;\n" + (isFragmentCase ? mainStatements : "") + " gl_FragColor = value;\n" "}\n"), gls::AttribSpec("a_value", Vec4(1.0f, 0.0f, 0.0f, 0.0f), Vec4(0.0f, 1.0f, 0.0f, 0.0f), Vec4(0.0f, 0.0f, 1.0f, 0.0f), Vec4(0.0f, 0.0f, 0.0f, 1.0f))); } static inline ProgramData defaultProgramData (CaseShaderType shaderType, const string& mainStatements) { return defaultProgramData(shaderType, "", mainStatements); } class ShaderOptimizationCase : public TestCase { public: ShaderOptimizationCase (Context& context, const char* name, const char* description, CaseShaderType caseShaderType) : TestCase (context, tcu::NODETYPE_PERFORMANCE, name, description) , m_caseShaderType (caseShaderType) , m_state (STATE_LAST) , m_measurer (context.getRenderContext(), caseShaderType == CASESHADERTYPE_VERTEX ? gls::CASETYPE_VERTEX : caseShaderType == CASESHADERTYPE_FRAGMENT ? gls::CASETYPE_FRAGMENT : gls::CASETYPE_LAST) , m_unoptimizedResult (-1.0f, -1.0f) , m_optimizedResult (-1.0f, -1.0f) { } virtual ~ShaderOptimizationCase (void) {} void init (void); IterateResult iterate (void); protected: virtual ProgramData generateProgramData (bool optimized) const = 0; const CaseShaderType m_caseShaderType; private: enum State { STATE_INIT_UNOPTIMIZED = 0, STATE_MEASURE_UNOPTIMIZED, STATE_INIT_OPTIMIZED, STATE_MEASURE_OPTIMIZED, STATE_FINISHED, STATE_LAST }; ProgramData& programData (bool optimized) { return optimized ? m_optimizedData : m_unoptimizedData; } SharedPtr& program (bool optimized) { return optimized ? m_optimizedProgram : m_unoptimizedProgram; } ShaderPerformanceMeasurer::Result& result (bool optimized) { return optimized ? m_optimizedResult : m_unoptimizedResult; } State m_state; ShaderPerformanceMeasurer m_measurer; ProgramData m_unoptimizedData; ProgramData m_optimizedData; SharedPtr m_unoptimizedProgram; SharedPtr m_optimizedProgram; ShaderPerformanceMeasurer::Result m_unoptimizedResult; ShaderPerformanceMeasurer::Result m_optimizedResult; }; void ShaderOptimizationCase::init (void) { const glu::RenderContext& renderCtx = m_context.getRenderContext(); TestLog& log = m_testCtx.getLog(); m_measurer.logParameters(log); for (int ndx = 0; ndx < 2; ndx++) { const bool optimized = ndx == 1; programData(optimized) = generateProgramData(optimized); for (int i = 0; i < (int)programData(optimized).attributes.size(); i++) DE_ASSERT(programData(optimized).attributes[i].name != "a_position"); // \note Position attribute is set by m_measurer. program(optimized) = SharedPtr(new ShaderProgram(renderCtx, programData(optimized).sources)); { const tcu::ScopedLogSection section(log, optimized ? "OptimizedProgram" : "UnoptimizedProgram", optimized ? "Hand-optimized program" : "Unoptimized program"); log << *program(optimized); } if (!program(optimized)->isOk()) TCU_FAIL("Shader compilation failed"); } m_state = STATE_INIT_UNOPTIMIZED; } ShaderOptimizationCase::IterateResult ShaderOptimizationCase::iterate (void) { TestLog& log = m_testCtx.getLog(); if (m_state == STATE_INIT_UNOPTIMIZED || m_state == STATE_INIT_OPTIMIZED) { const bool optimized = m_state == STATE_INIT_OPTIMIZED; m_measurer.init(program(optimized)->getProgram(), programData(optimized).attributes, 1); m_state = optimized ? STATE_MEASURE_OPTIMIZED : STATE_MEASURE_UNOPTIMIZED; return CONTINUE; } else if (m_state == STATE_MEASURE_UNOPTIMIZED || m_state == STATE_MEASURE_OPTIMIZED) { m_measurer.iterate(); if (m_measurer.isFinished()) { const bool optimized = m_state == STATE_MEASURE_OPTIMIZED; const tcu::ScopedLogSection section (log, optimized ? "OptimizedResult" : "UnoptimizedResult", optimized ? "Measurement results for hand-optimized program" : "Measurement result for unoptimized program"); m_measurer.logMeasurementInfo(log); result(optimized) = m_measurer.getResult(); m_measurer.deinit(); m_state = optimized ? STATE_FINISHED : STATE_INIT_OPTIMIZED; } return CONTINUE; } else { DE_ASSERT(m_state == STATE_FINISHED); const float unoptimizedRelevantResult = m_caseShaderType == CASESHADERTYPE_VERTEX ? m_unoptimizedResult.megaVertPerSec : m_unoptimizedResult.megaFragPerSec; const float optimizedRelevantResult = m_caseShaderType == CASESHADERTYPE_VERTEX ? m_optimizedResult.megaVertPerSec : m_optimizedResult.megaFragPerSec; const char* const relevantResultName = m_caseShaderType == CASESHADERTYPE_VERTEX ? "vertex" : "fragment"; const float ratio = unoptimizedRelevantResult / optimizedRelevantResult; const int handOptimizationGain = (int)deFloatRound(100.0f/ratio) - 100; log << TestLog::Message << "Unoptimized / optimized " << relevantResultName << " performance ratio: " << ratio << TestLog::EndMessage; if (handOptimizationGain >= 0) log << TestLog::Message << "Note: " << handOptimizationGain << "% performance gain was achieved with hand-optimized version" << TestLog::EndMessage; else log << TestLog::Message << "Note: hand-optimization degraded performance by " << -handOptimizationGain << "%" << TestLog::EndMessage; m_testCtx.setTestResult(QP_TEST_RESULT_PASS, de::floatToString(ratio, 2).c_str()); return STOP; } } class LoopUnrollCase : public ShaderOptimizationCase { public: enum CaseType { CASETYPE_INDEPENDENT = 0, CASETYPE_DEPENDENT, CASETYPE_LAST }; LoopUnrollCase (Context& context, const char* name, const char* description, CaseShaderType caseShaderType, CaseType caseType, int numRepetitions) : ShaderOptimizationCase (context, name, description, caseShaderType) , m_numRepetitions (numRepetitions) , m_caseType (caseType) { } protected: ProgramData generateProgramData (bool optimized) const { const string repetition = optimized ? repeatIndexedTemplate("\t" + expressionTemplate(m_caseType) + ";\n", m_numRepetitions) : loop(m_numRepetitions, expressionTemplate(m_caseType)); return defaultProgramData(m_caseShaderType, "\t" + getShaderPrecision(m_caseShaderType) + " vec4 valueOrig = value;\n" + repetition); } private: const int m_numRepetitions; const CaseType m_caseType; static inline string expressionTemplate (CaseType caseType) { switch (caseType) { case CASETYPE_INDEPENDENT: return "value += sin(float(${NDX}+1)*valueOrig)"; case CASETYPE_DEPENDENT: return "value = sin(value)"; default: DE_ASSERT(false); return DE_NULL; } } static inline string loop (int iterations, const string& innerExpr) { return "\tfor (int i = 0; i < " + toString(iterations) + "; i++)\n\t\t" + tcu::StringTemplate(innerExpr).specialize(singleMap("NDX", "i")) + ";\n"; } }; class LoopInvariantCodeMotionCase : public ShaderOptimizationCase { public: LoopInvariantCodeMotionCase (Context& context, const char* name, const char* description, CaseShaderType caseShaderType, int numLoopIterations) : ShaderOptimizationCase (context, name, description, caseShaderType) , m_numLoopIterations (numLoopIterations) { } protected: ProgramData generateProgramData (bool optimized) const { float scale = 0.0f; for (int i = 0; i < m_numLoopIterations; i++) scale += 3.2f*(float)i + 4.6f; scale = 1.0f / scale; const string precision = getShaderPrecision(m_caseShaderType); const string statements = optimized ? " " + precision + " vec4 valueOrig = value;\n" " " + precision + " vec4 y = sin(cos(sin(valueOrig)));\n" " for (int i = 0; i < " + toString(m_numLoopIterations) + "; i++)\n" " {\n" " " + precision + " float x = 3.2*float(i) + 4.6;\n" " value += x*y;\n" " }\n" " value *= " + toString(scale) + ";\n" : " " + precision + " vec4 valueOrig = value;\n" " for (int i = 0; i < " + toString(m_numLoopIterations) + "; i++)\n" " {\n" " " + precision + " float x = 3.2*float(i) + 4.6;\n" " " + precision + " vec4 y = sin(cos(sin(valueOrig)));\n" " value += x*y;\n" " }\n" " value *= " + toString(scale) + ";\n"; return defaultProgramData(m_caseShaderType, statements); } private: const int m_numLoopIterations; }; class FunctionInliningCase : public ShaderOptimizationCase { public: FunctionInliningCase (Context& context, const char* name, const char* description, CaseShaderType caseShaderType, int callNestingDepth) : ShaderOptimizationCase (context, name, description, caseShaderType) , m_callNestingDepth (callNestingDepth) { } protected: ProgramData generateProgramData (bool optimized) const { const string precision = getShaderPrecision(m_caseShaderType); const string expression = "value*vec4(0.8, 0.7, 0.6, 0.9)"; const string maybeFuncDefs = optimized ? "" : funcDefinitions(m_callNestingDepth, precision, expression); const string mainValueStatement = (optimized ? "\tvalue = " + expression : "\tvalue = func" + toString(m_callNestingDepth-1) + "(value)") + ";\n"; return defaultProgramData(m_caseShaderType, maybeFuncDefs, mainValueStatement); } private: const int m_callNestingDepth; static inline string funcDefinitions (int callNestingDepth, const string& precision, const string& expression) { string result = precision + " vec4 func0 (" + precision + " vec4 value) { return " + expression + "; }\n"; for (int i = 1; i < callNestingDepth; i++) result += precision + " vec4 func" + toString(i) + " (" + precision + " vec4 v) { return func" + toString(i-1) + "(v); }\n"; return result; } }; class ConstantPropagationCase : public ShaderOptimizationCase { public: enum CaseType { CASETYPE_BUILT_IN_FUNCTIONS = 0, CASETYPE_ARRAY, CASETYPE_STRUCT, CASETYPE_LAST }; ConstantPropagationCase (Context& context, const char* name, const char* description, CaseShaderType caseShaderType, CaseType caseType, bool useConstantExpressionsOnly) : ShaderOptimizationCase (context, name, description, caseShaderType) , m_caseType (caseType) , m_useConstantExpressionsOnly (useConstantExpressionsOnly) { DE_ASSERT(!(m_caseType == CASETYPE_ARRAY && m_useConstantExpressionsOnly)); // \note Would need array constructors, which GLSL ES 1 doesn't have. } protected: ProgramData generateProgramData (bool optimized) const { const bool isVertexCase = m_caseShaderType == CASESHADERTYPE_VERTEX; const string precision = getShaderPrecision(m_caseShaderType); const string statements = m_caseType == CASETYPE_BUILT_IN_FUNCTIONS ? builtinFunctionsCaseStatements (optimized, m_useConstantExpressionsOnly, precision, isVertexCase) : m_caseType == CASETYPE_ARRAY ? arrayCaseStatements (optimized, precision, isVertexCase) : m_caseType == CASETYPE_STRUCT ? structCaseStatements (optimized, m_useConstantExpressionsOnly, precision, isVertexCase) : DE_NULL; return defaultProgramData(m_caseShaderType, statements); } private: const CaseType m_caseType; const bool m_useConstantExpressionsOnly; static inline string builtinFunctionsCaseStatements (bool optimized, bool constantExpressionsOnly, const string& precision, bool useHeavierWorkload) { const string constMaybe = constantExpressionsOnly ? "const " : ""; const int numSinRows = useHeavierWorkload ? 12 : 1; return optimized ? " value = vec4(0.4, 0.5, 0.6, 0.7) * value; // NOTE: factor doesn't necessarily match the one in unoptimized shader, but shouldn't make a difference performance-wise\n" : " " + constMaybe + precision + " vec4 a = vec4(sin(0.7), cos(0.2), sin(0.9), abs(-0.5));\n" " " + constMaybe + precision + " vec4 b = cos(a) + fract(3.0*a.xzzw);\n" " " + constMaybe + "bvec4 c = bvec4(true, false, true, true);\n" " " + constMaybe + precision + " vec4 d = exp(b + vec4(c));\n" " " + constMaybe + precision + " vec4 e0 = inversesqrt(mix(d+a, d+b, a));\n" + repeatIndexedTemplate(" " + constMaybe + precision + " vec4 e${NDX} = sin(sin(sin(sin(e${PREV_NDX}))));\n", numSinRows, "", 1) + " " + constMaybe + precision + " vec4 f = abs(e" + toString(numSinRows) + ");\n" + " value = f*value;\n"; } static inline string arrayCaseStatements (bool optimized, const string& precision, bool useHeavierWorkload) { const int numSinRows = useHeavierWorkload ? 12 : 1; return optimized ? " value = vec4(0.4, 0.5, 0.6, 0.7) * value; // NOTE: factor doesn't necessarily match the one in unoptimized shader, but shouldn't make a difference performance-wise\n" : " const int arrLen = 4;\n" " " + precision + " vec4 arr[arrLen];\n" " arr[0] = vec4(0.1, 0.5, 0.9, 1.3);\n" " arr[1] = vec4(0.2, 0.6, 1.0, 1.4);\n" " arr[2] = vec4(0.3, 0.7, 1.1, 1.5);\n" " arr[3] = vec4(0.4, 0.8, 1.2, 1.6);\n" " " + precision + " vec4 a = (arr[0] + arr[1] + arr[2] + arr[3]) * 0.25;\n" " " + precision + " vec4 b0 = cos(sin(a));\n" + repeatIndexedTemplate(" " + precision + " vec4 b${NDX} = sin(sin(sin(sin(b${PREV_NDX}))));\n", numSinRows, "", 1) + " " + precision + " vec4 c = abs(b" + toString(numSinRows) + ");\n" + " value = c*value;\n"; } static inline string structCaseStatements (bool optimized, bool constantExpressionsOnly, const string& precision, bool useHeavierWorkload) { const string constMaybe = constantExpressionsOnly ? "const " : ""; const int numSinRows = useHeavierWorkload ? 12 : 1; return optimized ? " value = vec4(0.4, 0.5, 0.6, 0.7) * value; // NOTE: factor doesn't necessarily match the one in unoptimized shader, but shouldn't make a difference performance-wise\n" : " struct S\n" " {\n" " " + precision + " vec4 a;\n" " " + precision + " vec4 b;\n" " " + precision + " vec4 c;\n" " " + precision + " vec4 d;\n" " };\n" "\n" " " + constMaybe + "S s =\n" " S(vec4(0.1, 0.5, 0.9, 1.3),\n" " vec4(0.2, 0.6, 1.0, 1.4),\n" " vec4(0.3, 0.7, 1.1, 1.5),\n" " vec4(0.4, 0.8, 1.2, 1.6));\n" " " + constMaybe + precision + " vec4 a = (s.a + s.b + s.c + s.d) * 0.25;\n" " " + constMaybe + precision + " vec4 b0 = cos(sin(a));\n" + repeatIndexedTemplate(" " + constMaybe + precision + " vec4 b${NDX} = sin(sin(sin(sin(b${PREV_NDX}))));\n", numSinRows, "", 1) + " " + constMaybe + precision + " vec4 c = abs(b" + toString(numSinRows) + ");\n" + " value = c*value;\n"; } }; class CommonSubexpressionCase : public ShaderOptimizationCase { public: enum CaseType { CASETYPE_SINGLE_STATEMENT = 0, CASETYPE_MULTIPLE_STATEMENTS, CASETYPE_STATIC_BRANCH, CASETYPE_LOOP, CASETYPE_LAST }; CommonSubexpressionCase (Context& context, const char* name, const char* description, CaseShaderType caseShaderType, CaseType caseType) : ShaderOptimizationCase (context, name, description, caseShaderType) , m_caseType (caseType) { } protected: ProgramData generateProgramData (bool optimized) const { const bool isVertexCase = m_caseShaderType == CASESHADERTYPE_VERTEX; const string precision = getShaderPrecision(m_caseShaderType); const string statements = m_caseType == CASETYPE_SINGLE_STATEMENT ? singleStatementCaseStatements (optimized, precision, isVertexCase) : m_caseType == CASETYPE_MULTIPLE_STATEMENTS ? multipleStatementsCaseStatements (optimized, precision, isVertexCase) : m_caseType == CASETYPE_STATIC_BRANCH ? staticBranchCaseStatements (optimized, precision, isVertexCase) : m_caseType == CASETYPE_LOOP ? loopCaseStatements (optimized, precision, isVertexCase) : DE_NULL; return defaultProgramData(m_caseShaderType, statements); } private: const CaseType m_caseType; static inline string singleStatementCaseStatements (bool optimized, const string& precision, bool useHeavierWorkload) { const int numTopLevelRepeats = useHeavierWorkload ? 4 : 1; return optimized ? " " + precision + " vec4 s = sin(value);\n" " " + precision + " vec4 cs = cos(s);\n" " " + precision + " vec4 d = fract(s + cs) + sqrt(s + exp(cs));\n" " value = " + repeat("d", numTopLevelRepeats, "+") + ";\n" : " value = " + repeat("fract(sin(value) + cos(sin(value))) + sqrt(sin(value) + exp(cos(sin(value))))", numTopLevelRepeats, "\n\t + ") + ";\n"; } static inline string multipleStatementsCaseStatements (bool optimized, const string& precision, bool useHeavierWorkload) { const int numTopLevelRepeats = useHeavierWorkload ? 4 : 2; DE_ASSERT(numTopLevelRepeats >= 2); return optimized ? " " + precision + " vec4 a = sin(value) + cos(exp(value));\n" " " + precision + " vec4 b = cos(cos(a));\n" " a = fract(exp(sqrt(b)));\n" "\n" + repeat("\tvalue += a*b;\n", numTopLevelRepeats) : repeatIndexedTemplate( " " + precision + " vec4 a${NDX} = sin(value) + cos(exp(value));\n" " " + precision + " vec4 b${NDX} = cos(cos(a${NDX}));\n" " a${NDX} = fract(exp(sqrt(b${NDX})));\n" "\n", numTopLevelRepeats) + repeatIndexedTemplate( " value += a${NDX}*b${NDX};\n", numTopLevelRepeats); } static inline string staticBranchCaseStatements (bool optimized, const string& precision, bool useHeavierWorkload) { const int numTopLevelRepeats = useHeavierWorkload ? 4 : 2; DE_ASSERT(numTopLevelRepeats >= 2); if (optimized) { return " " + precision + " vec4 a = sin(value) + cos(exp(value));\n" " " + precision + " vec4 b = cos(a);\n" " b = cos(b);\n" " a = fract(exp(sqrt(b)));\n" "\n" + repeat(" value += a*b;\n", numTopLevelRepeats); } else { string result; for (int i = 0; i < numTopLevelRepeats; i++) { result += " " + precision + " vec4 a" + toString(i) + " = sin(value) + cos(exp(value));\n" " " + precision + " vec4 b" + toString(i) + " = cos(a" + toString(i) + ");\n"; if (i % 3 == 0) result += " if (1 < 2)\n" " b" + toString(i) + " = cos(b" + toString(i) + ");\n"; else if (i % 3 == 1) result += " b" + toString(i) + " = cos(b" + toString(i) + ");\n"; else if (i % 3 == 2) result += " if (2 < 1);\n" " else\n" " b" + toString(i) + " = cos(b" + toString(i) + ");\n"; else DE_ASSERT(false); result += " a" + toString(i) + " = fract(exp(sqrt(b" + toString(i) + ")));\n\n"; } result += repeatIndexedTemplate(" value += a${NDX}*b${NDX};\n", numTopLevelRepeats); return result; } } static inline string loopCaseStatements (bool optimized, const string& precision, bool useHeavierWorkload) { const int numLoopIterations = useHeavierWorkload ? 32 : 4; return optimized ? " " + precision + " vec4 acc = value;\n" " for (int i = 0; i < " + toString(numLoopIterations) + "; i++)\n" " acc = sin(acc);\n" "\n" " value += acc;\n" " value += acc;\n" : " " + precision + " vec4 acc0 = value;\n" " for (int i = 0; i < " + toString(numLoopIterations) + "; i++)\n" " acc0 = sin(acc0);\n" "\n" " " + precision + " vec4 acc1 = value;\n" " for (int i = 0; i < " + toString(numLoopIterations) + "; i++)\n" " acc1 = sin(acc1);\n" "\n" " value += acc0;\n" " value += acc1;\n"; } }; class DeadCodeEliminationCase : public ShaderOptimizationCase { public: enum CaseType { CASETYPE_DEAD_BRANCH_SIMPLE = 0, CASETYPE_DEAD_BRANCH_COMPLEX, CASETYPE_DEAD_BRANCH_COMPLEX_NO_CONST, CASETYPE_DEAD_BRANCH_FUNC_CALL, CASETYPE_UNUSED_VALUE_BASIC, CASETYPE_UNUSED_VALUE_LOOP, CASETYPE_UNUSED_VALUE_DEAD_BRANCH, CASETYPE_UNUSED_VALUE_AFTER_RETURN, CASETYPE_UNUSED_VALUE_MUL_ZERO, CASETYPE_LAST }; DeadCodeEliminationCase (Context& context, const char* name, const char* description, CaseShaderType caseShaderType, CaseType caseType) : ShaderOptimizationCase (context, name, description, caseShaderType) , m_caseType (caseType) { } protected: ProgramData generateProgramData (bool optimized) const { const bool isVertexCase = m_caseShaderType == CASESHADERTYPE_VERTEX; const string precision = getShaderPrecision(m_caseShaderType); const string funcDefs = m_caseType == CASETYPE_DEAD_BRANCH_FUNC_CALL ? deadBranchFuncCallCaseFuncDefs (optimized, precision) : m_caseType == CASETYPE_UNUSED_VALUE_AFTER_RETURN ? unusedValueAfterReturnCaseFuncDefs (optimized, precision, isVertexCase) : ""; const string statements = m_caseType == CASETYPE_DEAD_BRANCH_SIMPLE ? deadBranchSimpleCaseStatements (optimized, isVertexCase) : m_caseType == CASETYPE_DEAD_BRANCH_COMPLEX ? deadBranchComplexCaseStatements (optimized, precision, true, isVertexCase) : m_caseType == CASETYPE_DEAD_BRANCH_COMPLEX_NO_CONST ? deadBranchComplexCaseStatements (optimized, precision, false, isVertexCase) : m_caseType == CASETYPE_DEAD_BRANCH_FUNC_CALL ? deadBranchFuncCallCaseStatements (optimized, isVertexCase) : m_caseType == CASETYPE_UNUSED_VALUE_BASIC ? unusedValueBasicCaseStatements (optimized, precision, isVertexCase) : m_caseType == CASETYPE_UNUSED_VALUE_LOOP ? unusedValueLoopCaseStatements (optimized, precision, isVertexCase) : m_caseType == CASETYPE_UNUSED_VALUE_DEAD_BRANCH ? unusedValueDeadBranchCaseStatements (optimized, precision, isVertexCase) : m_caseType == CASETYPE_UNUSED_VALUE_AFTER_RETURN ? unusedValueAfterReturnCaseStatements () : m_caseType == CASETYPE_UNUSED_VALUE_MUL_ZERO ? unusedValueMulZeroCaseStatements (optimized, precision, isVertexCase) : DE_NULL; return defaultProgramData(m_caseShaderType, funcDefs, statements); } private: const CaseType m_caseType; static inline string deadBranchSimpleCaseStatements (bool optimized, bool useHeavierWorkload) { const int numLoopIterations = useHeavierWorkload ? 16 : 4; return optimized ? " value = vec4(0.6, 0.7, 0.8, 0.9) * value;\n" : " value = vec4(0.6, 0.7, 0.8, 0.9) * value;\n" " if (2 < 1)\n" " {\n" " value = cos(exp(sin(value))*log(sqrt(value)));\n" " for (int i = 0; i < " + toString(numLoopIterations) + "; i++)\n" " value = sin(value);\n" " }\n"; } static inline string deadBranchComplexCaseStatements (bool optimized, const string& precision, bool useConst, bool useHeavierWorkload) { const string constMaybe = useConst ? "const " : ""; const int numLoopIterations = useHeavierWorkload ? 16 : 4; return optimized ? " value = vec4(0.6, 0.7, 0.8, 0.9) * value;\n" : " value = vec4(0.6, 0.7, 0.8, 0.9) * value;\n" " " + constMaybe + precision + " vec4 a = vec4(sin(0.7), cos(0.2), sin(0.9), abs(-0.5));\n" " " + constMaybe + precision + " vec4 b = cos(a) + fract(3.0*a.xzzw);\n" " " + constMaybe + "bvec4 c = bvec4(true, false, true, true);\n" " " + constMaybe + precision + " vec4 d = exp(b + vec4(c));\n" " " + constMaybe + precision + " vec4 e = 1.8*abs(sin(sin(inversesqrt(mix(d+a, d+b, a)))));\n" " if (e.x > 1.0)\n" " {\n" " value = cos(exp(sin(value))*log(sqrt(value)));\n" " for (int i = 0; i < " + toString(numLoopIterations) + "; i++)\n" " value = sin(value);\n" " }\n"; } static inline string deadBranchFuncCallCaseFuncDefs (bool optimized, const string& precision) { return optimized ? "" : precision + " float func (" + precision + " float x) { return 2.0*x; }\n"; } static inline string deadBranchFuncCallCaseStatements (bool optimized, bool useHeavierWorkload) { const int numLoopIterations = useHeavierWorkload ? 16 : 4; return optimized ? " value = vec4(0.6, 0.7, 0.8, 0.9) * value;\n" : " value = vec4(0.6, 0.7, 0.8, 0.9) * value;\n" " if (func(0.3) > 1.0)\n" " {\n" " value = cos(exp(sin(value))*log(sqrt(value)));\n" " for (int i = 0; i < " + toString(numLoopIterations) + "; i++)\n" " value = sin(value);\n" " }\n"; } static inline string unusedValueBasicCaseStatements (bool optimized, const string& precision, bool useHeavierWorkload) { const int numSinRows = useHeavierWorkload ? 12 : 1; return optimized ? " " + precision + " vec4 used = vec4(0.6, 0.7, 0.8, 0.9) * value;\n" " value = used;\n" : " " + precision + " vec4 used = vec4(0.6, 0.7, 0.8, 0.9) * value;\n" " " + precision + " vec4 unused = cos(exp(sin(value))*log(sqrt(value))) + used;\n" + repeat(" unused = sin(sin(sin(sin(unused))));\n", numSinRows) + " value = used;\n"; } static inline string unusedValueLoopCaseStatements (bool optimized, const string& precision, bool useHeavierWorkload) { const int numLoopIterations = useHeavierWorkload ? 16 : 4; return optimized ? " " + precision + " vec4 used = vec4(0.6, 0.7, 0.8, 0.9) * value;\n" " value = used;\n" : " " + precision + " vec4 used = vec4(0.6, 0.7, 0.8, 0.9) * value;\n" " " + precision + " vec4 unused = cos(exp(sin(value))*log(sqrt(value)));\n" " for (int i = 0; i < " + toString(numLoopIterations) + "; i++)\n" " unused = sin(unused + used);\n" " value = used;\n"; } static inline string unusedValueAfterReturnCaseFuncDefs (bool optimized, const string& precision, bool useHeavierWorkload) { const int numSinRows = useHeavierWorkload ? 12 : 1; return optimized ? precision + " vec4 func (" + precision + " vec4 v)\n" "{\n" " " + precision + " vec4 used = vec4(0.6, 0.7, 0.8, 0.9) * v;\n" " return used;\n" "}\n" : precision + " vec4 func (" + precision + " vec4 v)\n" "{\n" " " + precision + " vec4 used = vec4(0.6, 0.7, 0.8, 0.9) * v;\n" " " + precision + " vec4 unused = cos(exp(sin(v))*log(sqrt(v)));\n" + repeat(" unused = sin(sin(sin(sin(unused))));\n", numSinRows) + " return used;\n" " used = used*unused;" " return used;\n" "}\n"; } static inline string unusedValueAfterReturnCaseStatements (void) { return " value = func(value);\n"; } static inline string unusedValueDeadBranchCaseStatements (bool optimized, const string& precision, bool useHeavierWorkload) { const int numSinRows = useHeavierWorkload ? 12 : 1; return optimized ? " " + precision + " vec4 used = vec4(0.6, 0.7, 0.8, 0.9) * value;\n" " value = used;\n" : " " + precision + " vec4 used = vec4(0.6, 0.7, 0.8, 0.9) * value;\n" " " + precision + " vec4 unused = cos(exp(sin(value))*log(sqrt(value)));\n" + repeat(" unused = sin(sin(sin(sin(unused))));\n", numSinRows) + " if (2 < 1)\n" " used = used*unused;\n" " value = used;\n"; } static inline string unusedValueMulZeroCaseStatements (bool optimized, const string& precision, bool useHeavierWorkload) { const int numSinRows = useHeavierWorkload ? 12 : 1; return optimized ? " " + precision + " vec4 used = vec4(0.6, 0.7, 0.8, 0.9) * value;\n" " value = used;\n" : " " + precision + " vec4 used = vec4(0.6, 0.7, 0.8, 0.9) * value;\n" " " + precision + " vec4 unused = cos(exp(sin(value))*log(sqrt(value)));\n" + repeat(" unused = sin(sin(sin(sin(unused))));\n", numSinRows) + " value = used + unused*float(1-1);\n"; } }; } // anonymous ShaderOptimizationTests::ShaderOptimizationTests (Context& context) : TestCaseGroup(context, "optimization", "Shader Optimization Performance Tests") { } ShaderOptimizationTests::~ShaderOptimizationTests (void) { } void ShaderOptimizationTests::init (void) { TestCaseGroup* const unrollGroup = new TestCaseGroup(m_context, "loop_unrolling", "Loop Unrolling Cases"); TestCaseGroup* const loopInvariantCodeMotionGroup = new TestCaseGroup(m_context, "loop_invariant_code_motion", "Loop-Invariant Code Motion Cases"); TestCaseGroup* const inlineGroup = new TestCaseGroup(m_context, "function_inlining", "Function Inlining Cases"); TestCaseGroup* const constantPropagationGroup = new TestCaseGroup(m_context, "constant_propagation", "Constant Propagation Cases"); TestCaseGroup* const commonSubexpressionGroup = new TestCaseGroup(m_context, "common_subexpression_elimination", "Common Subexpression Elimination Cases"); TestCaseGroup* const deadCodeEliminationGroup = new TestCaseGroup(m_context, "dead_code_elimination", "Dead Code Elimination Cases"); addChild(unrollGroup); addChild(loopInvariantCodeMotionGroup); addChild(inlineGroup); addChild(constantPropagationGroup); addChild(commonSubexpressionGroup); addChild(deadCodeEliminationGroup); for (int caseShaderTypeI = 0; caseShaderTypeI < CASESHADERTYPE_LAST; caseShaderTypeI++) { const CaseShaderType caseShaderType = (CaseShaderType)caseShaderTypeI; const char* const caseShaderTypeSuffix = caseShaderType == CASESHADERTYPE_VERTEX ? "_vertex" : caseShaderType == CASESHADERTYPE_FRAGMENT ? "_fragment" : DE_NULL; // Loop unrolling cases. { static const int loopIterationCounts[] = { 4, 8, 32 }; for (int caseTypeI = 0; caseTypeI < LoopUnrollCase::CASETYPE_LAST; caseTypeI++) { const LoopUnrollCase::CaseType caseType = (LoopUnrollCase::CaseType)caseTypeI; const string caseTypeName = caseType == LoopUnrollCase::CASETYPE_INDEPENDENT ? "independent_iterations" : caseType == LoopUnrollCase::CASETYPE_DEPENDENT ? "dependent_iterations" : DE_NULL; const string caseTypeDesc = caseType == LoopUnrollCase::CASETYPE_INDEPENDENT ? "loop iterations don't depend on each other" : caseType == LoopUnrollCase::CASETYPE_DEPENDENT ? "loop iterations depend on each other" : DE_NULL; for (int loopIterNdx = 0; loopIterNdx < DE_LENGTH_OF_ARRAY(loopIterationCounts); loopIterNdx++) { const int loopIterations = loopIterationCounts[loopIterNdx]; const string name = caseTypeName + "_" + toString(loopIterations) + caseShaderTypeSuffix; const string description = toString(loopIterations) + " iterations; " + caseTypeDesc; unrollGroup->addChild(new LoopUnrollCase(m_context, name.c_str(), description.c_str(), caseShaderType, caseType, loopIterations)); } } } // Loop-invariant code motion cases. { static const int loopIterationCounts[] = { 4, 8, 32 }; for (int loopIterNdx = 0; loopIterNdx < DE_LENGTH_OF_ARRAY(loopIterationCounts); loopIterNdx++) { const int loopIterations = loopIterationCounts[loopIterNdx]; const string name = toString(loopIterations) + "_iterations" + caseShaderTypeSuffix; loopInvariantCodeMotionGroup->addChild(new LoopInvariantCodeMotionCase(m_context, name.c_str(), "", caseShaderType, loopIterations)); } } // Function inlining cases. { static const int callNestingDepths[] = { 4, 8, 32 }; for (int nestDepthNdx = 0; nestDepthNdx < DE_LENGTH_OF_ARRAY(callNestingDepths); nestDepthNdx++) { const int nestingDepth = callNestingDepths[nestDepthNdx]; const string name = toString(nestingDepth) + "_nested" + caseShaderTypeSuffix; inlineGroup->addChild(new FunctionInliningCase(m_context, name.c_str(), "", caseShaderType, nestingDepth)); } } // Constant propagation cases. for (int caseTypeI = 0; caseTypeI < ConstantPropagationCase::CASETYPE_LAST; caseTypeI++) { const ConstantPropagationCase::CaseType caseType = (ConstantPropagationCase::CaseType)caseTypeI; const string caseTypeName = caseType == ConstantPropagationCase::CASETYPE_BUILT_IN_FUNCTIONS ? "built_in_functions" : caseType == ConstantPropagationCase::CASETYPE_ARRAY ? "array" : caseType == ConstantPropagationCase::CASETYPE_STRUCT ? "struct" : DE_NULL; for (int constantExpressionsOnlyI = 0; constantExpressionsOnlyI <= 1; constantExpressionsOnlyI++) { const bool constantExpressionsOnly = constantExpressionsOnlyI != 0; const string name = caseTypeName + (constantExpressionsOnly ? "" : "_no_const") + caseShaderTypeSuffix; if (caseType == ConstantPropagationCase::CASETYPE_ARRAY && constantExpressionsOnly) // \note See ConstantPropagationCase's constructor for explanation. continue; constantPropagationGroup->addChild(new ConstantPropagationCase(m_context, name.c_str(), "", caseShaderType, caseType, constantExpressionsOnly)); } } // Common subexpression cases. for (int caseTypeI = 0; caseTypeI < CommonSubexpressionCase::CASETYPE_LAST; caseTypeI++) { const CommonSubexpressionCase::CaseType caseType = (CommonSubexpressionCase::CaseType)caseTypeI; const string caseTypeName = caseType == CommonSubexpressionCase::CASETYPE_SINGLE_STATEMENT ? "single_statement" : caseType == CommonSubexpressionCase::CASETYPE_MULTIPLE_STATEMENTS ? "multiple_statements" : caseType == CommonSubexpressionCase::CASETYPE_STATIC_BRANCH ? "static_branch" : caseType == CommonSubexpressionCase::CASETYPE_LOOP ? "loop" : DE_NULL; const string description = caseType == CommonSubexpressionCase::CASETYPE_SINGLE_STATEMENT ? "A single statement containing multiple uses of same subexpression" : caseType == CommonSubexpressionCase::CASETYPE_MULTIPLE_STATEMENTS ? "Multiple statements performing same computations" : caseType == CommonSubexpressionCase::CASETYPE_STATIC_BRANCH ? "Multiple statements including a static conditional" : caseType == CommonSubexpressionCase::CASETYPE_LOOP ? "Multiple loops performing the same computations" : DE_NULL; commonSubexpressionGroup->addChild(new CommonSubexpressionCase(m_context, (caseTypeName + caseShaderTypeSuffix).c_str(), description.c_str(), caseShaderType, caseType)); } // Dead code elimination cases. for (int caseTypeI = 0; caseTypeI < DeadCodeEliminationCase::CASETYPE_LAST; caseTypeI++) { const DeadCodeEliminationCase::CaseType caseType = (DeadCodeEliminationCase::CaseType)caseTypeI; const char* const caseTypeName = caseType == DeadCodeEliminationCase::CASETYPE_DEAD_BRANCH_SIMPLE ? "dead_branch_simple" : caseType == DeadCodeEliminationCase::CASETYPE_DEAD_BRANCH_COMPLEX ? "dead_branch_complex" : caseType == DeadCodeEliminationCase::CASETYPE_DEAD_BRANCH_COMPLEX_NO_CONST ? "dead_branch_complex_no_const" : caseType == DeadCodeEliminationCase::CASETYPE_DEAD_BRANCH_FUNC_CALL ? "dead_branch_func_call" : caseType == DeadCodeEliminationCase::CASETYPE_UNUSED_VALUE_BASIC ? "unused_value_basic" : caseType == DeadCodeEliminationCase::CASETYPE_UNUSED_VALUE_LOOP ? "unused_value_loop" : caseType == DeadCodeEliminationCase::CASETYPE_UNUSED_VALUE_DEAD_BRANCH ? "unused_value_dead_branch" : caseType == DeadCodeEliminationCase::CASETYPE_UNUSED_VALUE_AFTER_RETURN ? "unused_value_after_return" : caseType == DeadCodeEliminationCase::CASETYPE_UNUSED_VALUE_MUL_ZERO ? "unused_value_mul_zero" : DE_NULL; const char* const caseTypeDescription = caseType == DeadCodeEliminationCase::CASETYPE_DEAD_BRANCH_SIMPLE ? "Do computation inside a branch that is never taken (condition is simple false constant expression)" : caseType == DeadCodeEliminationCase::CASETYPE_DEAD_BRANCH_COMPLEX ? "Do computation inside a branch that is never taken (condition is complex false constant expression)" : caseType == DeadCodeEliminationCase::CASETYPE_DEAD_BRANCH_COMPLEX_NO_CONST ? "Do computation inside a branch that is never taken (condition is complex false expression, not constant expression but still compile-time computable)" : caseType == DeadCodeEliminationCase::CASETYPE_DEAD_BRANCH_FUNC_CALL ? "Do computation inside a branch that is never taken (condition is compile-time computable false expression containing function call to a simple inlineable function)" : caseType == DeadCodeEliminationCase::CASETYPE_UNUSED_VALUE_BASIC ? "Compute a value that is never used even statically" : caseType == DeadCodeEliminationCase::CASETYPE_UNUSED_VALUE_LOOP ? "Compute a value, using a loop, that is never used even statically" : caseType == DeadCodeEliminationCase::CASETYPE_UNUSED_VALUE_DEAD_BRANCH ? "Compute a value that is used only inside a statically dead branch" : caseType == DeadCodeEliminationCase::CASETYPE_UNUSED_VALUE_AFTER_RETURN ? "Compute a value that is used only after a return statement" : caseType == DeadCodeEliminationCase::CASETYPE_UNUSED_VALUE_MUL_ZERO ? "Compute a value that is used but multiplied by a zero constant expression" : DE_NULL; deadCodeEliminationGroup->addChild(new DeadCodeEliminationCase(m_context, (string() + caseTypeName + caseShaderTypeSuffix).c_str(), caseTypeDescription, caseShaderType, caseType)); } } } } // Performance } // gles2 } // deqp