[D3D12] DXBC color output remapping

src/xenia/gpu/dxbc_shader_translator.cc

@@ -40,7 +40,92 @@ void DxbcShaderTranslator::Reset() {
   std::memset(&stat_, 0, sizeof(stat_));
 }
 
+void DxbcShaderTranslator::CompleteVertexShaderCode() {}
+
+void DxbcShaderTranslator::CompletePixelShaderCode() {
+  // Remap color outputs from Xbox 360 to Direct3D 12 indices.
+  // temp = xe_color_output_map;
+  // [unroll] for (uint i = 0; i < 4; ++i) {
+  //   o[i] = x1[temp[i]];
+  // }
+  shader_code_.push_back(ENCODE_D3D10_SB_OPCODE_TYPE(D3D10_SB_OPCODE_MOV) |
+                         ENCODE_D3D10_SB_TOKENIZED_INSTRUCTION_LENGTH(7));
+  shader_code_.push_back(
+      ENCODE_D3D10_SB_OPERAND_NUM_COMPONENTS(D3D10_SB_OPERAND_4_COMPONENT) |
+      ENCODE_D3D10_SB_OPERAND_4_COMPONENT_SELECTION_MODE(
+          D3D10_SB_OPERAND_4_COMPONENT_MASK_MODE) |
+      D3D10_SB_OPERAND_4_COMPONENT_MASK_ALL |
+      ENCODE_D3D10_SB_OPERAND_TYPE(D3D10_SB_OPERAND_TYPE_TEMP) |
+      ENCODE_D3D10_SB_OPERAND_INDEX_DIMENSION(D3D10_SB_OPERAND_INDEX_1D) |
+      ENCODE_D3D10_SB_OPERAND_INDEX_REPRESENTATION(
+          0, D3D10_SB_OPERAND_INDEX_IMMEDIATE32));
+  shader_code_.push_back(uint32_t(TempRegister::kColorOutputMap));
+  shader_code_.push_back(
+      ENCODE_D3D10_SB_OPERAND_NUM_COMPONENTS(D3D10_SB_OPERAND_4_COMPONENT) |
+      ENCODE_D3D10_SB_OPERAND_4_COMPONENT_SELECTION_MODE(
+          D3D10_SB_OPERAND_4_COMPONENT_SWIZZLE_MODE) |
+      D3D10_SB_OPERAND_4_COMPONENT_NOSWIZZLE |
+      ENCODE_D3D10_SB_OPERAND_TYPE(D3D10_SB_OPERAND_TYPE_CONSTANT_BUFFER) |
+      ENCODE_D3D10_SB_OPERAND_INDEX_DIMENSION(D3D10_SB_OPERAND_INDEX_3D) |
+      ENCODE_D3D10_SB_OPERAND_INDEX_REPRESENTATION(
+          0, D3D10_SB_OPERAND_INDEX_IMMEDIATE32) |
+      ENCODE_D3D10_SB_OPERAND_INDEX_REPRESENTATION(
+          1, D3D10_SB_OPERAND_INDEX_IMMEDIATE32) |
+      ENCODE_D3D10_SB_OPERAND_INDEX_REPRESENTATION(
+          2, D3D10_SB_OPERAND_INDEX_IMMEDIATE32));
+  shader_code_.push_back(uint32_t(RdefConstantBufferIndex::kSystemConstants));
+  shader_code_.push_back(uint32_t(CbufferRegister::kSystemConstants));
+  shader_code_.push_back(kSysConst_ColorOutputMap_Vec);
+  ++stat_.instruction_count;
+  for (uint32_t i = 0; i < 4; ++i) {
+    shader_code_.push_back(ENCODE_D3D10_SB_OPCODE_TYPE(D3D10_SB_OPCODE_MOV) |
+                           ENCODE_D3D10_SB_TOKENIZED_INSTRUCTION_LENGTH(7));
+    shader_code_.push_back(
+        ENCODE_D3D10_SB_OPERAND_NUM_COMPONENTS(D3D10_SB_OPERAND_4_COMPONENT) |
+        ENCODE_D3D10_SB_OPERAND_4_COMPONENT_SELECTION_MODE(
+            D3D10_SB_OPERAND_4_COMPONENT_MASK_MODE) |
+        D3D10_SB_OPERAND_4_COMPONENT_MASK_ALL |
+        ENCODE_D3D10_SB_OPERAND_TYPE(D3D10_SB_OPERAND_TYPE_OUTPUT) |
+        ENCODE_D3D10_SB_OPERAND_INDEX_DIMENSION(D3D10_SB_OPERAND_INDEX_1D) |
+        ENCODE_D3D10_SB_OPERAND_INDEX_REPRESENTATION(
+            0, D3D10_SB_OPERAND_INDEX_IMMEDIATE32));
+    shader_code_.push_back(i);
+    shader_code_.push_back(
+        ENCODE_D3D10_SB_OPERAND_NUM_COMPONENTS(D3D10_SB_OPERAND_4_COMPONENT) |
+        ENCODE_D3D10_SB_OPERAND_4_COMPONENT_SELECTION_MODE(
+            D3D10_SB_OPERAND_4_COMPONENT_SWIZZLE_MODE) |
+        D3D10_SB_OPERAND_4_COMPONENT_NOSWIZZLE |
+        ENCODE_D3D10_SB_OPERAND_TYPE(D3D10_SB_OPERAND_TYPE_INDEXABLE_TEMP) |
+        ENCODE_D3D10_SB_OPERAND_INDEX_DIMENSION(D3D10_SB_OPERAND_INDEX_2D) |
+        ENCODE_D3D10_SB_OPERAND_INDEX_REPRESENTATION(
+            0, D3D10_SB_OPERAND_INDEX_IMMEDIATE32) |
+        ENCODE_D3D10_SB_OPERAND_INDEX_REPRESENTATION(
+            1, D3D10_SB_OPERAND_INDEX_RELATIVE));
+    shader_code_.push_back(1);
+    shader_code_.push_back(
+        ENCODE_D3D10_SB_OPERAND_NUM_COMPONENTS(D3D10_SB_OPERAND_4_COMPONENT) |
+        ENCODE_D3D10_SB_OPERAND_4_COMPONENT_SELECTION_MODE(
+            D3D10_SB_OPERAND_4_COMPONENT_SELECT_1_MODE) |
+        ENCODE_D3D10_SB_OPERAND_4_COMPONENT_SELECT_1(i) |
+        ENCODE_D3D10_SB_OPERAND_TYPE(D3D10_SB_OPERAND_TYPE_TEMP) |
+        ENCODE_D3D10_SB_OPERAND_INDEX_DIMENSION(D3D10_SB_OPERAND_INDEX_1D) |
+        ENCODE_D3D10_SB_OPERAND_INDEX_REPRESENTATION(
+            0, D3D10_SB_OPERAND_INDEX_IMMEDIATE32));
+    shader_code_.push_back(uint32_t(TempRegister::kColorOutputMap));
+    ++stat_.instruction_count;
+    ++stat_.array_instruction_count;
+    ++stat_.mov_instruction_count;
+  }
+}
+
 void DxbcShaderTranslator::CompleteShaderCode() {
+  // Write stage-specific epilogue.
+  if (is_vertex_shader()) {
+    CompleteVertexShaderCode();
+  } else if (is_pixel_shader()) {
+    CompletePixelShaderCode();
+  }
+
   // Return from `main`.
   shader_code_.push_back(ENCODE_D3D10_SB_OPCODE_TYPE(D3D10_SB_OPCODE_RET) |
                          ENCODE_D3D10_SB_TOKENIZED_INSTRUCTION_LENGTH(1));
@@ -880,6 +965,24 @@ void DxbcShaderTranslator::WriteShaderCode() {
     stat_.temp_array_count += 4;
   }
 
+  // Initialize the depth output if used, which must be initialized on every
+  // execution path.
+  if (is_pixel_shader() && writes_depth_) {
+    shader_object_.push_back(ENCODE_D3D10_SB_OPCODE_TYPE(D3D10_SB_OPCODE_MOV) |
+                             ENCODE_D3D10_SB_TOKENIZED_INSTRUCTION_LENGTH(4));
+    shader_object_.push_back(
+        ENCODE_D3D10_SB_OPERAND_NUM_COMPONENTS(D3D10_SB_OPERAND_1_COMPONENT) |
+        ENCODE_D3D10_SB_OPERAND_TYPE(D3D10_SB_OPERAND_TYPE_OUTPUT_DEPTH) |
+        ENCODE_D3D10_SB_OPERAND_INDEX_DIMENSION(D3D10_SB_OPERAND_INDEX_0D));
+    shader_object_.push_back(
+        ENCODE_D3D10_SB_OPERAND_NUM_COMPONENTS(D3D10_SB_OPERAND_1_COMPONENT) |
+        ENCODE_D3D10_SB_OPERAND_TYPE(D3D10_SB_OPERAND_TYPE_IMMEDIATE32) |
+        ENCODE_D3D10_SB_OPERAND_INDEX_DIMENSION(D3D10_SB_OPERAND_INDEX_0D));
+    shader_object_.push_back(0);
+    ++stat_.instruction_count;
+    ++stat_.mov_instruction_count;
+  }
+
   // Write the translated shader code.
   size_t code_size_dwords = shader_code_.size();
   // So [] won't crash in case the size is zero somehow.

src/xenia/gpu/dxbc_shader_translator.h

@@ -24,6 +24,7 @@ class DxbcShaderTranslator : public ShaderTranslator {
   ~DxbcShaderTranslator() override;
 
   // IF SYSTEM CONSTANTS ARE CHANGED OR ADDED, THE FOLLOWING MUST BE UPDATED:
+  // - kSysConst enum (registers and first components).
   // - rdef_constants_.
   // - rdef_constant_buffers_ system constant buffer size.
   // - d3d12/shaders/xenos_draw.hlsli (for geometry shaders).
@@ -76,6 +77,39 @@ class DxbcShaderTranslator : public ShaderTranslator {
     kFloatConstantsLast = kFloatConstantsFirst + kFloatConstantPageCount - 1,
   };
 
+  enum : uint32_t {
+    kSysConst_MulRcpW_Vec = 0,
+    kSysConst_MulRcpW_Comp = 0,
+    kSysConst_VertexBaseIndex_Vec = 0,
+    kSysConst_VertexBaseIndex_Comp = 3,
+
+    kSysConst_NDCScale_Vec = 1,
+    kSysConst_NDCScale_Comp = 0,
+    kSysConst_VertexIndexEndian_Vec = 1,
+    kSysConst_VertexIndexEndian_Comp = 3,
+
+    kSysConst_NDCOffset_Vec = 2,
+    kSysConst_NDCOffset_Comp = 0,
+    kSysConst_PixelHalfPixelOffset_Vec = 2,
+    kSysConst_PixelHalfPixelOffset_Comp = 3,
+
+    kSysConst_PointSize_Vec = 3,
+    kSysConst_PointSize_Comp = 0,
+    kSysConst_SSAAInvScale_Vec = 0,
+    kSysConst_SSAAInvScale_Comp = 2,
+
+    kSysConst_PixelPosReg_Vec = 4,
+    kSysConst_PixelPosReg_Comp = 0,
+    kSysConst_AlphaTest_Vec = 4,
+    kSysConst_AlphaTest_Comp = 1,
+    kSysConst_AlphaTestRange_Vec = 4,
+    kSysConst_AlphaTestRange_Comp = 2,
+
+    kSysConst_ColorExpBias_Vec = 5,
+
+    kSysConst_ColorOutputMap_Vec = 6,
+  };
+
   static constexpr uint32_t kInterpolatorCount = 16;
   static constexpr uint32_t kPointParametersTexCoord = kInterpolatorCount;
 
@@ -95,7 +129,15 @@ class DxbcShaderTranslator : public ShaderTranslator {
   static constexpr uint32_t kPSInPositionRegister =
       kPSInPointParametersRegister + 1;
 
-  // Writes the epilogue.
+  enum class TempRegister {
+    kColorOutputMap,
+
+    kCount,
+  };
+
+  // Writing the epilogue.
+  void CompleteVertexShaderCode();
+  void CompletePixelShaderCode();
   void CompleteShaderCode();
 
   // Appends a string to a DWORD stream, returns the DWORD-aligned length.
@@ -233,6 +275,7 @@ class DxbcShaderTranslator : public ShaderTranslator {
     uint32_t temp_register_count;
     // Unknown in Wine.
     uint32_t def_count;
+    // Only inputs and outputs.
     uint32_t dcl_count;
     uint32_t float_instruction_count;
     uint32_t int_instruction_count;
@@ -250,6 +293,7 @@ class DxbcShaderTranslator : public ShaderTranslator {
     uint32_t texture_comp_instructions;
     uint32_t texture_bias_instructions;
     uint32_t texture_gradient_instructions;
+    // Moving to outputs only.
     uint32_t mov_instruction_count;
     // Unknown in Wine.
     uint32_t movc_instruction_count;