Merge pull request #1266 from amdrexu/feature

Implement the extension GL_AMD_gpu_shader_half_float_fetch

SPIRV/GLSL.ext.AMD.h

@@ -27,13 +27,8 @@
 #ifndef GLSLextAMD_H
 #define GLSLextAMD_H
 
-enum BuiltIn;
-enum Capability;
-enum Decoration;
-enum Op;
-
 static const int GLSLextAMDVersion = 100;
-static const int GLSLextAMDRevision = 6;
+static const int GLSLextAMDRevision = 7;
 
 // SPV_AMD_shader_ballot
 static const char* const E_SPV_AMD_shader_ballot = "SPV_AMD_shader_ballot";
@@ -107,4 +102,10 @@ static const char* const E_SPV_AMD_shader_image_load_store_lod = "SPV_AMD_shader
 // SPV_AMD_shader_fragment_mask
 static const char* const E_SPV_AMD_shader_fragment_mask = "SPV_AMD_shader_fragment_mask";
 
+// SPV_AMD_gpu_shader_half_float_fetch
+static const char* const E_SPV_AMD_gpu_shader_half_float_fetch = "SPV_AMD_gpu_shader_half_float_fetch";
+
+enum Capability;
+static const Capability CapabilityFloat16ImageAMD = static_cast<Capability>(5008);
+
 #endif  // #ifndef GLSLextAMD_H

SPIRV/GlslangToSpv.cpp

@@ -2340,6 +2340,12 @@ spv::Id TGlslangToSpvTraverser::getSampledType(const glslang::TSampler& sampler)
 {
     switch (sampler.type) {
         case glslang::EbtFloat:    return builder.makeFloatType(32);
+#ifdef AMD_EXTENSIONS
+        case glslang::EbtFloat16:
+            builder.addExtension(spv::E_SPV_AMD_gpu_shader_half_float_fetch);
+            builder.addCapability(spv::CapabilityFloat16ImageAMD);
+            return builder.makeFloatType(16);
+#endif
         case glslang::EbtInt:      return builder.makeIntType(32);
         case glslang::EbtUint:     return builder.makeUintType(32);
         default:
@@ -3159,9 +3165,15 @@ void TGlslangToSpvTraverser::translateArguments(const glslang::TIntermAggregate&
 
     glslang::TSampler sampler = {};
     bool cubeCompare = false;
+#ifdef AMD_EXTENSIONS
+    bool f16ShadowCompare = false;
+#endif
     if (node.isTexture() || node.isImage()) {
         sampler = glslangArguments[0]->getAsTyped()->getType().getSampler();
         cubeCompare = sampler.dim == glslang::EsdCube && sampler.arrayed && sampler.shadow;
+#ifdef AMD_EXTENSIONS
+        f16ShadowCompare = sampler.shadow && glslangArguments[1]->getAsTyped()->getType().getBasicType() == glslang::EbtFloat16;
+#endif
     }
 
     for (int i = 0; i < (int)glslangArguments.size(); ++i) {
@@ -3186,6 +3198,21 @@ void TGlslangToSpvTraverser::translateArguments(const glslang::TIntermAggregate&
             if ((sampler.ms && i == 3) || (! sampler.ms && i == 2))
                 lvalue = true;
             break;
+#ifdef AMD_EXTENSIONS
+        case glslang::EOpSparseTexture:
+            if (((cubeCompare || f16ShadowCompare) && i == 3) || (! (cubeCompare || f16ShadowCompare) && i == 2))
+                lvalue = true;
+            break;
+        case glslang::EOpSparseTextureClamp:
+            if (((cubeCompare || f16ShadowCompare) && i == 4) || (! (cubeCompare || f16ShadowCompare) && i == 3))
+                lvalue = true;
+            break;
+        case glslang::EOpSparseTextureLod:
+        case glslang::EOpSparseTextureOffset:
+            if  ((f16ShadowCompare && i == 4) || (! f16ShadowCompare && i == 3))
+                lvalue = true;
+            break;
+#else
         case glslang::EOpSparseTexture:
             if ((cubeCompare && i == 3) || (! cubeCompare && i == 2))
                 lvalue = true;
@@ -3199,6 +3226,7 @@ void TGlslangToSpvTraverser::translateArguments(const glslang::TIntermAggregate&
             if (i == 3)
                 lvalue = true;
             break;
+#endif
         case glslang::EOpSparseTextureFetch:
             if ((sampler.dim != glslang::EsdRect && i == 3) || (sampler.dim == glslang::EsdRect && i == 2))
                 lvalue = true;
@@ -3207,6 +3235,23 @@ void TGlslangToSpvTraverser::translateArguments(const glslang::TIntermAggregate&
             if ((sampler.dim != glslang::EsdRect && i == 4) || (sampler.dim == glslang::EsdRect && i == 3))
                 lvalue = true;
             break;
+#ifdef AMD_EXTENSIONS
+        case glslang::EOpSparseTextureLodOffset:
+        case glslang::EOpSparseTextureGrad:
+        case glslang::EOpSparseTextureOffsetClamp:
+            if ((f16ShadowCompare && i == 5) || (! f16ShadowCompare && i == 4))
+                lvalue = true;
+            break;
+        case glslang::EOpSparseTextureGradOffset:
+        case glslang::EOpSparseTextureGradClamp:
+            if ((f16ShadowCompare && i == 6) || (! f16ShadowCompare && i == 5))
+                lvalue = true;
+            break;
+        case glslang::EOpSparseTextureGradOffsetClamp:
+            if ((f16ShadowCompare && i == 7) || (! f16ShadowCompare && i == 6))
+                lvalue = true;
+            break;
+#else
         case glslang::EOpSparseTextureLodOffset:
         case glslang::EOpSparseTextureGrad:
         case glslang::EOpSparseTextureOffsetClamp:
@@ -3222,6 +3267,7 @@ void TGlslangToSpvTraverser::translateArguments(const glslang::TIntermAggregate&
             if (i == 6)
                 lvalue = true;
             break;
+#endif
         case glslang::EOpSparseTextureGather:
             if ((sampler.shadow && i == 3) || (! sampler.shadow && i == 2))
                 lvalue = true;
@@ -3274,6 +3320,12 @@ spv::Id TGlslangToSpvTraverser::createImageTextureFunctionCall(glslang::TIntermO
     // Process a GLSL texturing op (will be SPV image)
     const glslang::TSampler sampler = node->getAsAggregate() ? node->getAsAggregate()->getSequence()[0]->getAsTyped()->getType().getSampler()
                                                              : node->getAsUnaryNode()->getOperand()->getAsTyped()->getType().getSampler();
+#ifdef AMD_EXTENSIONS
+    bool f16ShadowCompare = (sampler.shadow && node->getAsAggregate())
+                                ? node->getAsAggregate()->getSequence()[1]->getAsTyped()->getType().getBasicType() == glslang::EbtFloat16
+                                : false;
+#endif
+
     std::vector<spv::Id> arguments;
     if (node->getAsAggregate())
         translateArguments(*node->getAsAggregate(), arguments);
@@ -3517,6 +3569,9 @@ spv::Id TGlslangToSpvTraverser::createImageTextureFunctionCall(glslang::TIntermO
 #ifdef AMD_EXTENSIONS
         if (cracked.gather)
             ++nonBiasArgCount; // comp argument should be present when bias argument is present
+
+        if (f16ShadowCompare)
+            ++nonBiasArgCount;
 #endif
         if (cracked.offset)
             ++nonBiasArgCount;
@@ -3560,7 +3615,11 @@ spv::Id TGlslangToSpvTraverser::createImageTextureFunctionCall(glslang::TIntermO
     bool noImplicitLod = false;
 
     // sort out where Dref is coming from
+#ifdef AMD_EXTENSIONS
+    if (cubeCompare || f16ShadowCompare) {
+#else
     if (cubeCompare) {
+#endif
         params.Dref = arguments[2];
         ++extraArgs;
     } else if (sampler.shadow && cracked.gather) {

SPIRV/SpvBuilder.cpp

@@ -1751,7 +1751,11 @@ Id Builder::createTextureQueryCall(Op opCode, const TextureParameters& parameter
         break;
     }
     case OpImageQueryLod:
+#ifdef AMD_EXTENSIONS
+        resultType = makeVectorType(getScalarTypeId(getTypeId(parameters.coords)), 2);
+#else
         resultType = makeVectorType(makeFloatType(32), 2);
+#endif
         break;
     case OpImageQueryLevels:
     case OpImageQuerySamples:

SPIRV/doc.cpp

@@ -849,6 +849,7 @@ const char* CapabilityString(int info)
     case 5013: return "StencilExportEXT";
 
 #ifdef AMD_EXTENSIONS
+    case 5008: return "Float16ImageAMD";
     case 5009: return "ImageGatherBiasLodAMD";
     case 5010: return "FragmentMaskAMD";
     case 5015: return "ImageReadWriteLodAMD";