Merge pull request #1806 from KhronosGroup/nan-clamp

SPV: Add a switch for favoring non-NaN operands in min, max, and clamp.

SPIRV/GlslangToSpv.cpp

@@ -224,6 +224,7 @@ protected:
     bool linkageOnly;                  // true when visiting the set of objects in the AST present only for establishing interface, whether or not they were statically used
     std::set<spv::Id> iOSet;           // all input/output variables from either static use or declaration of interface
     const glslang::TIntermediate* glslangIntermediate;
+    bool nanMinMaxClamp;               // true if use NMin/NMax/NClamp instead of FMin/FMax/FClamp
     spv::Id stdBuiltins;
     std::unordered_map<const char*, spv::Id> extBuiltinMap;
 
@@ -1313,7 +1314,8 @@ TGlslangToSpvTraverser::TGlslangToSpvTraverser(unsigned int spvVersion, const gl
       sequenceDepth(0), logger(buildLogger),
       builder(spvVersion, (glslang::GetKhronosToolId() << 16) | glslang::GetSpirvGeneratorVersion(), logger),
       inEntryPoint(false), entryPointTerminated(false), linkageOnly(false),
-      glslangIntermediate(glslangIntermediate)
+      glslangIntermediate(glslangIntermediate),
+      nanMinMaxClamp(glslangIntermediate->getNanMinMaxClamp())
 {
     spv::ExecutionModel executionModel = TranslateExecutionModel(glslangIntermediate->getStage());
 
@@ -7014,7 +7016,7 @@ spv::Id TGlslangToSpvTraverser::createMiscOperation(glslang::TOperator op, spv::
     switch (op) {
     case glslang::EOpMin:
         if (isFloat)
-            libCall = spv::GLSLstd450FMin;
+            libCall = nanMinMaxClamp ? spv::GLSLstd450NMin : spv::GLSLstd450FMin;
         else if (isUnsigned)
             libCall = spv::GLSLstd450UMin;
         else
@@ -7026,7 +7028,7 @@ spv::Id TGlslangToSpvTraverser::createMiscOperation(glslang::TOperator op, spv::
         break;
     case glslang::EOpMax:
         if (isFloat)
-            libCall = spv::GLSLstd450FMax;
+            libCall = nanMinMaxClamp ? spv::GLSLstd450NMax : spv::GLSLstd450FMax;
         else if (isUnsigned)
             libCall = spv::GLSLstd450UMax;
         else
@@ -7045,7 +7047,7 @@ spv::Id TGlslangToSpvTraverser::createMiscOperation(glslang::TOperator op, spv::
 
     case glslang::EOpClamp:
         if (isFloat)
-            libCall = spv::GLSLstd450FClamp;
+            libCall = nanMinMaxClamp ? spv::GLSLstd450NClamp : spv::GLSLstd450FClamp;
         else if (isUnsigned)
             libCall = spv::GLSLstd450UClamp;
         else

SPIRV/disassemble.cpp

@@ -648,6 +648,9 @@ static void GLSLstd450GetDebugNames(const char** names)
     names[GLSLstd450InterpolateAtCentroid]   = "InterpolateAtCentroid";
     names[GLSLstd450InterpolateAtSample]     = "InterpolateAtSample";
     names[GLSLstd450InterpolateAtOffset]     = "InterpolateAtOffset";
+    names[GLSLstd450NMin]                    = "NMin";
+    names[GLSLstd450NMax]                    = "NMax";
+    names[GLSLstd450NClamp]                  = "NClamp";
 }
 
 #ifdef AMD_EXTENSIONS

StandAlone/StandAlone.cpp

@@ -104,6 +104,7 @@ enum TOptions {
 bool targetHlslFunctionality1 = false;
 bool SpvToolsDisassembler = false;
 bool SpvToolsValidate = false;
+bool NaNClamp = false;
 
 //
 // Return codes from main/exit().
@@ -522,6 +523,8 @@ void ProcessArguments(std::vector<std::unique_ptr<glslang::TWorkItem>>& workItem
                     } else if (lowerword == "keep-uncalled" || // synonyms
                                lowerword == "ku") {
                         Options |= EOptionKeepUncalled;
+                    } else if (lowerword == "nan-clamp") {
+                        NaNClamp = true;
                     } else if (lowerword == "no-storage-format" || // synonyms
                                lowerword == "nsf") {
                         Options |= EOptionNoStorageFormat;
@@ -981,6 +984,7 @@ void CompileAndLinkShaderUnits(std::vector<ShaderCompUnit> compUnits)
 
         shader->setFlattenUniformArrays((Options & EOptionFlattenUniformArrays) != 0);
         shader->setNoStorageFormat((Options & EOptionNoStorageFormat) != 0);
+        shader->setNanMinMaxClamp(NaNClamp);
         shader->setResourceSetBinding(baseResourceSetBinding[compUnit.stage]);
 
         if (Options & EOptionHlslIoMapping)
@@ -1533,9 +1537,11 @@ void usage()
            "                                    works independently of source language\n"
            "  --hlsl-iomap                      perform IO mapping in HLSL register space\n"
            "  --hlsl-enable-16bit-types         allow 16-bit types in SPIR-V for HLSL\n"
-           "  --hlsl-dx9-compatible             interprets sampler declarations as a texture/sampler combo like DirectX9 would."
+           "  --hlsl-dx9-compatible             interprets sampler declarations as a\n"
+           "                                    texture/sampler combo like DirectX9 would.\n"
            "  --invert-y | --iy                 invert position.Y output in vertex shader\n"
            "  --keep-uncalled | --ku            don't eliminate uncalled functions\n"
+           "  --nan-clamp                       favor non-NaN operand in min, max, and clamp\n"
            "  --no-storage-format | --nsf       use Unknown image format\n"
            "  --reflect-strict-array-suffix     use strict array suffix rules when\n"
            "                                    reflecting\n"

Test/runtests

@@ -239,6 +239,13 @@ $EXE -D -E hlsl.pp.expand.frag > $TARGETDIR/hlsl.pp.expand.frag.out 2> $TARGETDI
 diff -b $BASEDIR/hlsl.pp.expand.frag.out $TARGETDIR/hlsl.pp.expand.frag.out || HASERROR=1
 diff -b $BASEDIR/hlsl.pp.expand.frag.err $TARGETDIR/hlsl.pp.expand.frag.err || HASERROR=1
 
+#
+# Test --nan-clamp
+#
+echo "Testing nan-clamp"
+$EXE --nan-clamp -H --aml --amb spv.400.frag > $TARGETDIR/spv.400.frag.nanclamp.out
+diff -b $BASEDIR/spv.400.frag.nanclamp.out $TARGETDIR/spv.400.frag.nanclamp.out || HASERROR=1
+
 #
 # Final checking
 #

glslang/MachineIndependent/ShaderLang.cpp

@@ -1791,6 +1791,7 @@ void TShader::setUniformLocationBase(int base)
 void TShader::setHlslIoMapping(bool hlslIoMap)          { intermediate->setHlslIoMapping(hlslIoMap); }
 void TShader::setFlattenUniformArrays(bool flatten)     { intermediate->setFlattenUniformArrays(flatten); }
 void TShader::setNoStorageFormat(bool useUnknownFormat) { intermediate->setNoStorageFormat(useUnknownFormat); }
+void TShader::setNanMinMaxClamp(bool useNonNan)         { intermediate->setNanMinMaxClamp(useNonNan); }
 void TShader::setResourceSetBinding(const std::vector<std::string>& base)   { intermediate->setResourceSetBinding(base); }
 void TShader::setTextureSamplerTransformMode(EShTextureSamplerTransformMode mode) { intermediate->setTextureSamplerTransformMode(mode); }
 

glslang/MachineIndependent/localintermediate.h

@@ -266,7 +266,8 @@ public:
         needToLegalize(false),
         binaryDoubleOutput(false),
         usePhysicalStorageBuffer(false),
-        uniformLocationBase(0)
+        uniformLocationBase(0),
+        nanMinMaxClamp(false)
     {
         localSize[0] = 1;
         localSize[1] = 1;
@@ -767,6 +768,9 @@ public:
     void setUniformLocationBase(int base) { uniformLocationBase = base; }
     int getUniformLocationBase() const { return uniformLocationBase; }
 
+    void setNanMinMaxClamp(bool setting) { nanMinMaxClamp = setting; }
+    bool getNanMinMaxClamp() const { return nanMinMaxClamp; }
+
     void setNeedsLegalization() { needToLegalize = true; }
     bool needsLegalization() const { return needToLegalize; }
 
@@ -900,6 +904,7 @@ protected:
 
     std::unordered_map<std::string, int> uniformLocationOverrides;
     int uniformLocationBase;
+    bool nanMinMaxClamp;            // true if desiring min/max/clamp to favor non-NaN over NaN
 
 private:
     void operator=(TIntermediate&); // prevent assignments

glslang/Public/ShaderLang.h

@@ -435,6 +435,7 @@ public:
     void setHlslIoMapping(bool hlslIoMap);
     void setFlattenUniformArrays(bool flatten);
     void setNoStorageFormat(bool useUnknownFormat);
+    void setNanMinMaxClamp(bool nanMinMaxClamp);
     void setTextureSamplerTransformMode(EShTextureSamplerTransformMode mode);
 
     // For setting up the environment (cleared to nothingness in the constructor).