Merge pull request KhronosGroup#2977 from AMD-dwang/AMD_shader_early_…

…and_late_fragment_tests

Add SPV_AMD_shader_early_and_late_fragment_tests

SPIRV/GLSL.ext.KHR.h

@@ -53,5 +53,6 @@ static const char* const E_SPV_KHR_terminate_invocation         = "SPV_KHR_termi
 static const char* const E_SPV_KHR_workgroup_memory_explicit_layout = "SPV_KHR_workgroup_memory_explicit_layout";
 static const char* const E_SPV_KHR_subgroup_uniform_control_flow = "SPV_KHR_subgroup_uniform_control_flow";
 static const char* const E_SPV_KHR_fragment_shader_barycentric = "SPV_KHR_fragment_shader_barycentric";
+static const char* const E_SPV_AMD_shader_early_and_late_fragment_tests = "SPV_AMD_shader_early_and_late_fragment_tests";
 
 #endif  // #ifndef GLSLextKHR_H

SPIRV/GlslangToSpv.cpp

@@ -1623,6 +1623,12 @@ TGlslangToSpvTraverser::TGlslangToSpvTraverser(unsigned int spvVersion,
         if (glslangIntermediate->getEarlyFragmentTests())
             builder.addExecutionMode(shaderEntry, spv::ExecutionModeEarlyFragmentTests);
 
+        if (glslangIntermediate->getEarlyAndLateFragmentTestsAMD())
+        {
+            builder.addExecutionMode(shaderEntry, spv::ExecutionModeEarlyAndLateFragmentTestsAMD);
+            builder.addExtension(spv::E_SPV_AMD_shader_early_and_late_fragment_tests);
+        }
+
         if (glslangIntermediate->getPostDepthCoverage()) {
             builder.addCapability(spv::CapabilitySampleMaskPostDepthCoverage);
             builder.addExecutionMode(shaderEntry, spv::ExecutionModePostDepthCoverage);
@@ -1632,6 +1638,9 @@ TGlslangToSpvTraverser::TGlslangToSpvTraverser(unsigned int spvVersion,
         if (glslangIntermediate->isDepthReplacing())
             builder.addExecutionMode(shaderEntry, spv::ExecutionModeDepthReplacing);
 
+        if (glslangIntermediate->isStencilReplacing())
+            builder.addExecutionMode(shaderEntry, spv::ExecutionModeStencilRefReplacingEXT);
+
 #ifndef GLSLANG_WEB
 
         switch(glslangIntermediate->getDepth()) {
@@ -1640,6 +1649,20 @@ TGlslangToSpvTraverser::TGlslangToSpvTraverser(unsigned int spvVersion,
         case glslang::EldUnchanged: mode = spv::ExecutionModeDepthUnchanged; break;
         default:                    mode = spv::ExecutionModeMax;            break;
         }
+
+        if (mode != spv::ExecutionModeMax)
+            builder.addExecutionMode(shaderEntry, (spv::ExecutionMode)mode);
+
+        switch (glslangIntermediate->getStencil()) {
+        case glslang::ElsRefUnchangedFrontAMD:  mode = spv::ExecutionModeStencilRefUnchangedFrontAMD; break;
+        case glslang::ElsRefGreaterFrontAMD:    mode = spv::ExecutionModeStencilRefGreaterFrontAMD;   break;
+        case glslang::ElsRefLessFrontAMD:       mode = spv::ExecutionModeStencilRefLessFrontAMD;      break;
+        case glslang::ElsRefUnchangedBackAMD:   mode = spv::ExecutionModeStencilRefUnchangedBackAMD;  break;
+        case glslang::ElsRefGreaterBackAMD:     mode = spv::ExecutionModeStencilRefGreaterBackAMD;    break;
+        case glslang::ElsRefLessBackAMD:        mode = spv::ExecutionModeStencilRefLessBackAMD;       break;
+        default:                       mode = spv::ExecutionModeMax;                         break;
+        }
+
         if (mode != spv::ExecutionModeMax)
             builder.addExecutionMode(shaderEntry, (spv::ExecutionMode)mode);
         switch (glslangIntermediate->getInterlockOrdering()) {

SPIRV/doc.cpp

@@ -173,28 +173,32 @@ const char* ExecutionModeString(int mode)
     case 31: return "ContractionOff";
     case 32: return "Bad";
 
-    case ExecutionModeInitializer:              return "Initializer";
-    case ExecutionModeFinalizer:                return "Finalizer";
-    case ExecutionModeSubgroupSize:             return "SubgroupSize";
-    case ExecutionModeSubgroupsPerWorkgroup:    return "SubgroupsPerWorkgroup";
-    case ExecutionModeSubgroupsPerWorkgroupId:  return "SubgroupsPerWorkgroupId";
-    case ExecutionModeLocalSizeId:              return "LocalSizeId";
-    case ExecutionModeLocalSizeHintId:          return "LocalSizeHintId";
-
-    case ExecutionModePostDepthCoverage:        return "PostDepthCoverage";
-    case ExecutionModeDenormPreserve:           return "DenormPreserve";
-    case ExecutionModeDenormFlushToZero:        return "DenormFlushToZero";
-    case ExecutionModeSignedZeroInfNanPreserve: return "SignedZeroInfNanPreserve";
-    case ExecutionModeRoundingModeRTE:          return "RoundingModeRTE";
-    case ExecutionModeRoundingModeRTZ:          return "RoundingModeRTZ";
-    case ExecutionModeStencilRefReplacingEXT:   return "StencilRefReplacingEXT";
+    case ExecutionModeInitializer:                   return "Initializer";
+    case ExecutionModeFinalizer:                     return "Finalizer";
+    case ExecutionModeSubgroupSize:                  return "SubgroupSize";
+    case ExecutionModeSubgroupsPerWorkgroup:         return "SubgroupsPerWorkgroup";
+    case ExecutionModeSubgroupsPerWorkgroupId:       return "SubgroupsPerWorkgroupId";
+    case ExecutionModeLocalSizeId:                   return "LocalSizeId";
+    case ExecutionModeLocalSizeHintId:               return "LocalSizeHintId";
+
+    case ExecutionModePostDepthCoverage:             return "PostDepthCoverage";
+    case ExecutionModeDenormPreserve:                return "DenormPreserve";
+    case ExecutionModeDenormFlushToZero:             return "DenormFlushToZero";
+    case ExecutionModeSignedZeroInfNanPreserve:      return "SignedZeroInfNanPreserve";
+    case ExecutionModeRoundingModeRTE:               return "RoundingModeRTE";
+    case ExecutionModeRoundingModeRTZ:               return "RoundingModeRTZ";
+    case ExecutionModeEarlyAndLateFragmentTestsAMD:  return "EarlyAndLateFragmentTestsAMD";
+    case ExecutionModeStencilRefUnchangedFrontAMD:   return "StencilRefUnchangedFrontAMD";
+    case ExecutionModeStencilRefLessFrontAMD:        return "StencilRefLessFrontAMD";
+    case ExecutionModeStencilRefGreaterBackAMD:      return "StencilRefGreaterBackAMD";
+    case ExecutionModeStencilRefReplacingEXT:        return "StencilRefReplacingEXT";
     case ExecutionModeSubgroupUniformControlFlowKHR: return "SubgroupUniformControlFlow";
 
-    case ExecutionModeOutputLinesNV:            return "OutputLinesNV";
-    case ExecutionModeOutputPrimitivesNV:       return "OutputPrimitivesNV";
-    case ExecutionModeOutputTrianglesNV:        return "OutputTrianglesNV";
-    case ExecutionModeDerivativeGroupQuadsNV:   return "DerivativeGroupQuadsNV";
-    case ExecutionModeDerivativeGroupLinearNV:  return "DerivativeGroupLinearNV";
+    case ExecutionModeOutputLinesNV:                 return "OutputLinesNV";
+    case ExecutionModeOutputPrimitivesNV:            return "OutputPrimitivesNV";
+    case ExecutionModeOutputTrianglesNV:             return "OutputTrianglesNV";
+    case ExecutionModeDerivativeGroupQuadsNV:        return "DerivativeGroupQuadsNV";
+    case ExecutionModeDerivativeGroupLinearNV:       return "DerivativeGroupLinearNV";
 
     case ExecutionModePixelInterlockOrderedEXT:         return "PixelInterlockOrderedEXT";
     case ExecutionModePixelInterlockUnorderedEXT:       return "PixelInterlockUnorderedEXT";

SPIRV/spirv.hpp

@@ -158,7 +158,14 @@ enum ExecutionMode {
     ExecutionModeSignedZeroInfNanPreserve = 4461,
     ExecutionModeRoundingModeRTE = 4462,
     ExecutionModeRoundingModeRTZ = 4463,
+    ExecutionModeEarlyAndLateFragmentTestsAMD = 5017,
     ExecutionModeStencilRefReplacingEXT = 5027,
+    ExecutionModeStencilRefUnchangedFrontAMD = 5079,
+    ExecutionModeStencilRefGreaterFrontAMD = 5080,
+    ExecutionModeStencilRefLessFrontAMD = 5081,
+    ExecutionModeStencilRefUnchangedBackAMD = 5082,
+    ExecutionModeStencilRefGreaterBackAMD = 5083,
+    ExecutionModeStencilRefLessBackAMD = 5084,
     ExecutionModeOutputLinesNV = 5269,
     ExecutionModeOutputPrimitivesNV = 5270,
     ExecutionModeDerivativeGroupQuadsNV = 5289,

Test/baseResults/spv.earlyAndlateFragmentTests.frag.out

@@ -0,0 +1,41 @@
+spv.earlyAndlateFragmentTests.frag
+// Module Version 10000
+// Generated by (magic number): 8000a
+// Id's are bound by 16
+
+                              Capability Shader
+                              Extension  "SPV_AMD_shader_early_and_late_fragment_tests"
+               1:             ExtInstImport  "GLSL.std.450"
+                              MemoryModel Logical GLSL450
+                              EntryPoint Fragment 4  "main" 8 11
+                              ExecutionMode 4 OriginUpperLeft
+                              ExecutionMode 4 EarlyAndLateFragmentTestsAMD
+                              ExecutionMode 4 DepthReplacing
+                              ExecutionMode 4 DepthLess
+                              Source GLSL 450
+                              SourceExtension  "GL_ARB_fragment_shader_interlock"
+                              SourceExtension  "GL_ARB_shader_stencil_export"
+                              SourceExtension  "GL_EXT_fragment_shading_rate"
+                              Name 4  "main"
+                              Name 8  "gl_FragDepth"
+                              Name 11  "instanceIndex"
+                              Decorate 8(gl_FragDepth) BuiltIn FragDepth
+                              Decorate 11(instanceIndex) Flat
+                              Decorate 11(instanceIndex) Location 0
+               2:             TypeVoid
+               3:             TypeFunction 2
+               6:             TypeFloat 32
+               7:             TypePointer Output 6(float)
+ 8(gl_FragDepth):      7(ptr) Variable Output
+               9:             TypeInt 32 1
+              10:             TypePointer Input 9(int)
+11(instanceIndex):     10(ptr) Variable Input
+              14:    6(float) Constant 1117913088
+         4(main):           2 Function None 3
+               5:             Label
+              12:      9(int) Load 11(instanceIndex)
+              13:    6(float) ConvertSToF 12
+              15:    6(float) FDiv 13 14
+                              Store 8(gl_FragDepth) 15
+                              Return
+                              FunctionEnd

Test/baseResults/spv.shaderStencilExport.frag.out

@@ -10,6 +10,7 @@ spv.shaderStencilExport.frag
                               MemoryModel Logical GLSL450
                               EntryPoint Fragment 4  "main" 8
                               ExecutionMode 4 OriginUpperLeft
+                              ExecutionMode 4 StencilRefReplacingEXT
                               Source GLSL 450
                               SourceExtension  "GL_ARB_shader_stencil_export"
                               Name 4  "main"

Test/spv.earlyAndlateFragmentTests.frag

@@ -0,0 +1,12 @@
+#version 450 core
+#extension GL_EXT_fragment_shading_rate : enable
+#extension GL_ARB_shader_stencil_export : enable
+#extension GL_ARB_fragment_shader_interlock : enable
+#extension GL_AMD_shader_early_and_late_fragment_tests : enable
+layout(location = 0) flat in int instanceIndex;
+layout(early_and_late_fragment_tests_amd) in;
+layout(depth_less) out float gl_FragDepth;
+void main()
+{
+  gl_FragDepth = float(instanceIndex) / float(81);
+}

glslang/Include/BaseTypes.h

@@ -128,6 +128,7 @@ enum TStorageQualifier {
     // built-ins written by fragment shader
     EvqFragColor,
     EvqFragDepth,
+    EvqFragStencil,
 
     // end of list
     EvqLast
@@ -353,6 +354,7 @@ __inline const char* GetStorageQualifierString(TStorageQualifier q)
     case EvqPointCoord:     return "gl_PointCoord";  break;
     case EvqFragColor:      return "fragColor";      break;
     case EvqFragDepth:      return "gl_FragDepth";   break;
+    case EvqFragStencil:    return "gl_FragStencilRefARB"; break;
     case EvqPayload:        return "rayPayloadNV";     break;
     case EvqPayloadIn:      return "rayPayloadInNV";   break;
     case EvqHitAttr:        return "hitAttributeNV";   break;

glslang/Include/Types.h

@@ -443,6 +443,18 @@ enum TLayoutDepth {
     EldCount
 };
 
+enum TLayoutStencil {
+    ElsNone,
+    ElsRefUnchangedFrontAMD,
+    ElsRefGreaterFrontAMD,
+    ElsRefLessFrontAMD,
+    ElsRefUnchangedBackAMD,
+    ElsRefGreaterBackAMD,
+    ElsRefLessBackAMD,
+
+    ElsCount
+};
+
 enum TBlendEquationShift {
     // No 'EBlendNone':
     // These are used as bit-shift amounts.  A mask of such shifts will have type 'int',
@@ -705,6 +717,7 @@ class TQualifier {
         case EvqVaryingOut:
         case EvqFragColor:
         case EvqFragDepth:
+        case EvqFragStencil:
             return true;
         default:
             return false;
@@ -772,6 +785,7 @@ class TQualifier {
         case EvqVaryingOut:
         case EvqFragColor:
         case EvqFragDepth:
+        case EvqFragStencil:
             return true;
         default:
             return false;
@@ -1239,6 +1253,18 @@ class TQualifier {
         default:           return "none";
         }
     }
+    static const char* getLayoutStencilString(TLayoutStencil s)
+    {
+        switch (s) {
+        case ElsRefUnchangedFrontAMD: return "stencil_ref_unchanged_front_amd";
+        case ElsRefGreaterFrontAMD:   return "stencil_ref_greater_front_amd";
+        case ElsRefLessFrontAMD:      return "stencil_ref_less_front_amd";
+        case ElsRefUnchangedBackAMD:  return "stencil_ref_unchanged_back_amd";
+        case ElsRefGreaterBackAMD:    return "stencil_ref_greater_back_amd";
+        case ElsRefLessBackAMD:       return "stencil_ref_less_back_amd";
+        default:                      return "none";
+        }
+    }
     static const char* getBlendEquationString(TBlendEquationShift e)
     {
         switch (e) {
@@ -1336,7 +1362,9 @@ struct TShaderQualifiers {
 #ifndef GLSLANG_WEB
     bool earlyFragmentTests;  // fragment input
     bool postDepthCoverage;   // fragment input
+    bool earlyAndLateFragmentTestsAMD; //fragment input
     TLayoutDepth layoutDepth;
+    TLayoutStencil layoutStencil;
     bool blendEquation;       // true if any blend equation was specified
     int numViews;             // multiview extenstions
     TInterlockOrdering interlockOrdering;
@@ -1346,6 +1374,7 @@ struct TShaderQualifiers {
     int primitives;                     // mesh shader "max_primitives"DerivativeGroupLinear;   // true if layout derivative_group_linearNV set
     bool layoutPrimitiveCulling;        // true if layout primitive_culling set
     TLayoutDepth getDepth() const { return layoutDepth; }
+    TLayoutStencil getStencil() const { return layoutStencil; }
 #else
     TLayoutDepth getDepth() const { return EldNone; }
 #endif
@@ -1371,8 +1400,10 @@ struct TShaderQualifiers {
         localSizeSpecId[2] = TQualifier::layoutNotSet;
 #ifndef GLSLANG_WEB
         earlyFragmentTests = false;
+        earlyAndLateFragmentTestsAMD = false;
         postDepthCoverage = false;
         layoutDepth = EldNone;
+        layoutStencil = ElsNone;
         blendEquation = false;
         numViews = TQualifier::layoutNotSet;
         layoutOverrideCoverage      = false;
@@ -1424,10 +1455,14 @@ struct TShaderQualifiers {
 #ifndef GLSLANG_WEB
         if (src.earlyFragmentTests)
             earlyFragmentTests = true;
+        if (src.earlyAndLateFragmentTestsAMD)
+            earlyAndLateFragmentTestsAMD = true;
         if (src.postDepthCoverage)
             postDepthCoverage = true;
         if (src.layoutDepth)
             layoutDepth = src.layoutDepth;
+        if (src.layoutStencil)
+            layoutStencil = src.layoutStencil;
         if (src.blendEquation)
             blendEquation = src.blendEquation;
         if (src.numViews != TQualifier::layoutNotSet)

glslang/MachineIndependent/Initialize.cpp

@@ -8065,6 +8065,7 @@ void TBuiltIns::identifyBuiltIns(int version, EProfile profile, const SpvVersion
         SpecialQualifier("gl_FragDepth",        EvqFragDepth,  EbvFragDepth,        symbolTable);
 #ifndef GLSLANG_WEB
         SpecialQualifier("gl_FragDepthEXT",     EvqFragDepth,  EbvFragDepth,        symbolTable);
+        SpecialQualifier("gl_FragStencilRefARB", EvqFragStencil, EbvFragStencilRef, symbolTable);
         SpecialQualifier("gl_HelperInvocation", EvqVaryingIn,  EbvHelperInvocation, symbolTable);
 
         BuiltInVariable("gl_ClipDistance",    EbvClipDistance,   symbolTable);