[d3d9, dxso] Implement FETCH4

src/d3d9/d3d9_device.cpp

@@ -3507,6 +3507,16 @@ namespace dxvk {
         m_dirtySamplerStates |= 1u << StateSampler;
       else if (Type == D3DSAMP_SRGBTEXTURE)
         BindTexture(StateSampler);
+
+      constexpr DWORD Fetch4Enabled  = MAKEFOURCC('G', 'E', 'T', '4');
+      constexpr DWORD Fetch4Disabled = MAKEFOURCC('G', 'E', 'T', '1');
+
+      if (Type == D3DSAMP_MIPMAPLODBIAS) {
+        if (Value == Fetch4Enabled)
+          m_fetch4Enabled |= 1u << StateSampler;
+        else if (Value == Fetch4Disabled)
+          m_fetch4Enabled &= ~(1u << StateSampler);
+      }
     }
 
     return D3D_OK;
@@ -5603,21 +5613,26 @@ namespace dxvk {
     if (m_flags.test(D3D9DeviceFlag::DirtyInputLayout))
       BindInputLayout();
 
-    auto UpdateSamplerTypes = [&](uint32_t types, uint32_t projections) {
+    auto UpdateSamplerTypes = [&](uint32_t types, uint32_t projections, uint32_t fetch4) {
       if (m_lastSamplerTypeBitfield != types)
         UpdateSamplerSpecConsant(types);
 
       if (m_lastProjectionBitfield != projections)
         UpdateProjectionSpecConstant(projections);
+
+      if (m_lastFetch4 != fetch4)
+        UpdateFetch4SpecConstant(fetch4);
     };
 
     if (likely(UseProgrammablePS())) {
       UploadConstants<DxsoProgramTypes::PixelShader>();
 
+      uint32_t fetch4 = m_fetch4Enabled & (m_activeTextures & m_psShaderMasks.samplerMask);
+
       if (GetCommonShader(m_state.pixelShader)->GetInfo().majorVersion() >= 2)
-        UpdateSamplerTypes(m_d3d9Options.forceSamplerTypeSpecConstants ? m_samplerTypeBitfield : 0u, 0u);
+        UpdateSamplerTypes(m_d3d9Options.forceSamplerTypeSpecConstants ? m_samplerTypeBitfield : 0u, 0u, fetch4);
       else
-        UpdateSamplerTypes(m_samplerTypeBitfield, m_projectionBitfield); // For implicit samplers...
+        UpdateSamplerTypes(m_samplerTypeBitfield, m_projectionBitfield, fetch4); // For implicit samplers...
 
       UpdateBoolSpecConstantPixel(
         m_state.psConsts.bConsts[0] &
@@ -5625,7 +5640,7 @@ namespace dxvk {
     }
     else {
       UpdateBoolSpecConstantPixel(0);
-      UpdateSamplerTypes(0u, 0u);
+      UpdateSamplerTypes(0u, 0u, 0u);
 
       UpdateFixedFunctionPS();
     }
@@ -6316,6 +6331,15 @@ namespace dxvk {
   }
 
 
+  void D3D9DeviceEx::UpdateFetch4SpecConstant(uint32_t value) {
+    EmitCs([cBitfield = value](DxvkContext* ctx) {
+      ctx->setSpecConstant(VK_PIPELINE_BIND_POINT_GRAPHICS, D3D9SpecConstantId::Fetch4, cBitfield);
+      });
+
+    m_lastFetch4 = value;
+  }
+
+
   void D3D9DeviceEx::ApplyPrimitiveType(
     DxvkContext*      pContext,
     D3DPRIMITIVETYPE  PrimType) {

src/d3d9/d3d9_device.h

@@ -1036,6 +1036,9 @@ namespace dxvk {
     uint32_t                        m_activeTextures   = 0;
     uint32_t                        m_activeTexturesToUpload = 0;
 
+    uint32_t                        m_fetch4Enabled = 0;
+    uint32_t                        m_lastFetch4    = 0;
+
     uint32_t                        m_activeHazardsDS = 0;
     uint32_t                        m_lastHazardsDS   = 0;
 
@@ -1189,6 +1192,8 @@ namespace dxvk {
 
     void UpdateProjectionSpecConstant(uint32_t value);
 
+    void UpdateFetch4SpecConstant(uint32_t value);
+
   };
 
 }

src/d3d9/d3d9_spec_constants.h

@@ -17,6 +17,7 @@ namespace dxvk {
 
     VertexShaderBools = 8,
     PixelShaderBools  = 9,
+    Fetch4            = 10
   };
 
 }

src/dxso/dxso_compiler.cpp

@@ -468,6 +468,10 @@ namespace dxvk {
       }
     }
 
+    m_ps.fetch4Spec = m_module.specConst32(m_module.defIntType(32, 0), 0);
+    m_module.decorateSpecId(m_ps.fetch4Spec, getSpecId(D3D9SpecConstantId::Fetch4));
+    m_module.setDebugName(m_ps.fetch4Spec, "s_fetch4");
+
     this->setupRenderStateInfo();
     this->emitPsSharedConstants();
 
@@ -2766,12 +2770,22 @@ void DxsoCompiler::emitControlFlowGenericLoop(
         }
       }
 
+      uint32_t fetch4 = 0;
+      if (m_programInfo.type() == DxsoProgramType::PixelShader) {
+        fetch4 = m_module.opBitFieldUExtract(
+          m_module.defIntType(32, 0), m_ps.fetch4Spec,
+          m_module.consti32(samplerIdx), m_module.consti32(1));
+
+        fetch4 = m_module.opIEqual(m_module.defBoolType(), fetch4, m_module.constu32(1));
+      }
+
       result.id = this->emitSample(
         projDivider != 0,
         typeId,
         imageVarId,
         texcoordVar.id,
         reference,
+        fetch4,
         imageOperands);
 
       if (switchProjResult) {
@@ -2783,6 +2797,7 @@ void DxsoCompiler::emitControlFlowGenericLoop(
           imageVarId,
           texcoordVar.id,
           reference,
+          fetch4,
           imageOperands);
 
         uint32_t shouldProj = m_module.opBitFieldUExtract(
@@ -3028,40 +3043,56 @@ void DxsoCompiler::emitControlFlowGenericLoop(
           uint32_t                sampledImage,
           uint32_t                coordinates,
           uint32_t                reference,
+          uint32_t                fetch4,
     const SpirvImageOperands&     operands) {
     const bool depthCompare = reference != 0;
     const bool explicitLod  =
        (operands.flags & spv::ImageOperandsLodMask)
     || (operands.flags & spv::ImageOperandsGradMask);
 
+    uint32_t val;
+
+    // No Fetch 4
     if (projected) {
       if (depthCompare) {
         if (explicitLod)
-          return m_module.opImageSampleProjDrefExplicitLod(resultType, sampledImage, coordinates, reference, operands);
+          val = m_module.opImageSampleProjDrefExplicitLod(resultType, sampledImage, coordinates, reference, operands);
         else
-          return m_module.opImageSampleProjDrefImplicitLod(resultType, sampledImage, coordinates, reference, operands);
+          val = m_module.opImageSampleProjDrefImplicitLod(resultType, sampledImage, coordinates, reference, operands);
       }
       else {
         if (explicitLod)
-          return m_module.opImageSampleProjExplicitLod(resultType, sampledImage, coordinates, operands);
+          val = m_module.opImageSampleProjExplicitLod(resultType, sampledImage, coordinates, operands);
         else
-          return m_module.opImageSampleProjImplicitLod(resultType, sampledImage, coordinates, operands);
+          val = m_module.opImageSampleProjImplicitLod(resultType, sampledImage, coordinates, operands);
       }
     }
     else {
       if (depthCompare) {
         if (explicitLod)
-          return m_module.opImageSampleDrefExplicitLod(resultType, sampledImage, coordinates, reference, operands);
+          val = m_module.opImageSampleDrefExplicitLod(resultType, sampledImage, coordinates, reference, operands);
         else
-          return m_module.opImageSampleDrefImplicitLod(resultType, sampledImage, coordinates, reference, operands);
+          val = m_module.opImageSampleDrefImplicitLod(resultType, sampledImage, coordinates, reference, operands);
       }
       else {
         if (explicitLod)
-          return m_module.opImageSampleExplicitLod(resultType, sampledImage, coordinates, operands);
+          val = m_module.opImageSampleExplicitLod(resultType, sampledImage, coordinates, operands);
         else
-          return m_module.opImageSampleImplicitLod(resultType, sampledImage, coordinates, operands);
+          val = m_module.opImageSampleImplicitLod(resultType, sampledImage, coordinates, operands);
       }
     }
+
+
+    if (fetch4 && !depthCompare) {
+      uint32_t fetch4Val = m_module.opImageGather(resultType, sampledImage, coordinates, m_module.consti32(0), operands);
+      // B R G A swizzle... Funny D3D9 order.
+      const std::array<uint32_t, 4> indices = { 2, 0, 1, 3 };
+      fetch4Val = m_module.opVectorShuffle(resultType, fetch4Val, fetch4Val, indices.size(), indices.data());
+
+      val = m_module.opSelect(resultType, fetch4, fetch4Val, val);
+    }
+
+    return val;
   }
 
 

src/dxso/dxso_compiler.h

@@ -157,6 +157,7 @@ namespace dxvk {
     uint32_t functionId         = 0;
     uint32_t samplerTypeSpec    = 0;
     uint32_t projectionSpec     = 0;
+    uint32_t fetch4Spec         = 0;
 
     //////////////
     // Misc Types
@@ -642,6 +643,7 @@ namespace dxvk {
             uint32_t                sampledImage,
             uint32_t                coordinates,
             uint32_t                reference,
+            uint32_t                fetch4,
       const SpirvImageOperands&     operands);
 
     ///////////////////////////////