[dxbc] Initial shadow sampler support

src/dxbc/dxbc_compiler.cpp

@@ -1209,7 +1209,7 @@ namespace dxvk {
     // TODO support address offset
     // TODO support more sample ops
     
-    // sample has four operands:
+    // All sample instructions have at least these operands:
     //  (dst0) The destination register
     //  (src0) Texture coordinates
     //  (src1) The texture itself
@@ -1227,6 +1227,12 @@ namespace dxvk {
     const DxbcRegisterValue coord = emitRegisterLoad(
       texCoordReg, DxbcRegMask(true, true, true, true));
     
+    // Load reference value for depth-compare operations
+    DxbcRegisterValue referenceValue;
+    
+    if (ins.op == DxbcOpcode::SampleC || ins.op == DxbcOpcode::SampleClz)
+      referenceValue = emitRegisterLoad(ins.src[3], DxbcRegMask(true, false, false, false));
+    
     // Combine the texture and the sampler into a sampled image
     const uint32_t sampledImageType = m_module.defSampledImageType(
       m_textures.at(textureId).textureTypeId);
@@ -1242,18 +1248,41 @@ namespace dxvk {
     
     // Sampling an image in SPIR-V always returns a four-component
     // vector, so we need to declare the corresponding type here
-    // TODO infer sampled type properly
+    // TODO infer sampled types properly
     DxbcRegisterValue result;
-    result.type.ctype  = DxbcScalarType::Float32;
+    
-    result.type.ccount = 4;
+    switch (ins.op) {
-    result.id = m_module.opImageSampleImplicitLod(
+      case DxbcOpcode::Sample: {
-      getVectorTypeId(result.type),
+        result.type.ctype  = DxbcScalarType::Float32;
-      sampledImageId, coord.id);
+        result.type.ccount = 4;
+        result.id = m_module.opImageSampleImplicitLod(
+          getVectorTypeId(result.type),
+          sampledImageId, coord.id);
+      } break;
+      
+      case DxbcOpcode::SampleClz: {
+        result.type.ctype  = DxbcScalarType::Float32;
+        result.type.ccount = 1;
+        result.id = m_module.opImageSampleDrefExplicitLod(
+          getVectorTypeId(result.type),
+          sampledImageId, coord.id,
+          referenceValue.id,
+          m_module.constf32(0.0f));
+      } break;
+      
+      default:
+        Logger::warn(str::format(
+          "DxbcCompiler: Unhandled instruction: ",
+          ins.op));
+        return;
+    }
     
     // Swizzle components using the texture swizzle
     // and the destination operand's write mask
-    result = emitRegisterSwizzle(result,
+    if (result.type.ccount != 1) {
-      textureReg.swizzle, ins.dst[0].mask);
+      result = emitRegisterSwizzle(result,
+        textureReg.swizzle, ins.dst[0].mask);
+    }
     
     emitRegisterStore(ins.dst[0], result);
   }

src/dxbc/dxbc_defs.cpp

@@ -169,10 +169,10 @@ namespace dxvk {
     } },
     /* IMul                                 */
     { 4, DxbcInstClass::VectorImul, {
-      { DxbcOperandKind::DstReg, DxbcScalarType::Float32 },
+      { DxbcOperandKind::DstReg, DxbcScalarType::Sint32 },
-      { DxbcOperandKind::DstReg, DxbcScalarType::Float32 },
+      { DxbcOperandKind::DstReg, DxbcScalarType::Sint32 },
-      { DxbcOperandKind::SrcReg, DxbcScalarType::Float32 },
+      { DxbcOperandKind::SrcReg, DxbcScalarType::Sint32 },
-      { DxbcOperandKind::SrcReg, DxbcScalarType::Float32 },
+      { DxbcOperandKind::SrcReg, DxbcScalarType::Sint32 },
     } },
     /* INe                                  */
     { 3, DxbcInstClass::VectorCmp, {
@@ -300,9 +300,21 @@ namespace dxvk {
       { DxbcOperandKind::SrcReg, DxbcScalarType::Float32 },
     } },
     /* SampleC                              */
-    { },
+    { 5, DxbcInstClass::TextureSample, {
+      { DxbcOperandKind::DstReg, DxbcScalarType::Float32 },
+      { DxbcOperandKind::SrcReg, DxbcScalarType::Float32 },
+      { DxbcOperandKind::SrcReg, DxbcScalarType::Float32 },
+      { DxbcOperandKind::SrcReg, DxbcScalarType::Float32 },
+      { DxbcOperandKind::SrcReg, DxbcScalarType::Float32 },
+    } },
     /* SampleClz                            */
-    { },
+    { 5, DxbcInstClass::TextureSample, {
+      { DxbcOperandKind::DstReg, DxbcScalarType::Float32 },
+      { DxbcOperandKind::SrcReg, DxbcScalarType::Float32 },
+      { DxbcOperandKind::SrcReg, DxbcScalarType::Float32 },
+      { DxbcOperandKind::SrcReg, DxbcScalarType::Float32 },
+      { DxbcOperandKind::SrcReg, DxbcScalarType::Float32 },
+    } },
     /* SampleL                              */
     { },
     /* SampleD                              */
@@ -335,10 +347,10 @@ namespace dxvk {
     { },
     /* UMul                                 */
     { 4, DxbcInstClass::VectorImul, {
-      { DxbcOperandKind::DstReg, DxbcScalarType::Float32 },
+      { DxbcOperandKind::DstReg, DxbcScalarType::Uint32 },
-      { DxbcOperandKind::DstReg, DxbcScalarType::Float32 },
+      { DxbcOperandKind::DstReg, DxbcScalarType::Uint32 },
-      { DxbcOperandKind::SrcReg, DxbcScalarType::Float32 },
+      { DxbcOperandKind::SrcReg, DxbcScalarType::Uint32 },
-      { DxbcOperandKind::SrcReg, DxbcScalarType::Float32 },
+      { DxbcOperandKind::SrcReg, DxbcScalarType::Uint32 },
     } },
     /* UMad                                 */
     { },

src/spirv/spirv_module.cpp

@@ -1533,6 +1533,26 @@ namespace dxvk {
   }
   
   
+  uint32_t SpirvModule::opImageSampleDrefExplicitLod(
+          uint32_t                resultType,
+          uint32_t                sampledImage,
+          uint32_t                coordinates,
+          uint32_t                reference,
+          uint32_t                lod) {
+    uint32_t resultId = this->allocateId();
+    
+    m_code.putIns (spv::OpImageSampleDrefExplicitLod, 8);
+    m_code.putWord(resultType);
+    m_code.putWord(resultId);
+    m_code.putWord(sampledImage);
+    m_code.putWord(coordinates);
+    m_code.putWord(reference);
+    m_code.putWord(spv::ImageOperandsLodMask);
+    m_code.putWord(lod);
+    return resultId;
+  }
+  
+  
   void SpirvModule::opLoopMerge(
           uint32_t                mergeBlock,
           uint32_t                continueTarget,

src/spirv/spirv_module.h

@@ -520,6 +520,13 @@ namespace dxvk {
             uint32_t                sampledImage,
             uint32_t                coordinates);
     
+    uint32_t opImageSampleDrefExplicitLod(
+            uint32_t                resultType,
+            uint32_t                sampledImage,
+            uint32_t                coordinates,
+            uint32_t                reference,
+            uint32_t                lod);
+    
     void opLoopMerge(
             uint32_t                mergeBlock,
             uint32_t                continueTarget,