[D3D12] Fetch constants in PS for exp bias and 3D/stacked

src/xenia/gpu/d3d12/d3d12_command_processor.cc

@@ -86,10 +86,10 @@ ID3D12RootSignature* D3D12CommandProcessor::GetRootSignature(
     parameter.ParameterType = D3D12_ROOT_PARAMETER_TYPE_DESCRIPTOR_TABLE;
     parameter.DescriptorTable.NumDescriptorRanges = 1;
     parameter.DescriptorTable.pDescriptorRanges = ⦥
-    parameter.ShaderVisibility = D3D12_SHADER_VISIBILITY_VERTEX;
+    parameter.ShaderVisibility = D3D12_SHADER_VISIBILITY_ALL;
     range.RangeType = D3D12_DESCRIPTOR_RANGE_TYPE_CBV;
     range.NumDescriptors = 1;
-    range.BaseShaderRegister = 10;
+    range.BaseShaderRegister = 2;
     range.RegisterSpace = 0;
     range.OffsetInDescriptorsFromTableStart = 0;
   }
@@ -104,7 +104,7 @@ ID3D12RootSignature* D3D12CommandProcessor::GetRootSignature(
     parameter.ShaderVisibility = D3D12_SHADER_VISIBILITY_VERTEX;
     range.RangeType = D3D12_DESCRIPTOR_RANGE_TYPE_CBV;
     range.NumDescriptors = 8;
-    range.BaseShaderRegister = 2;
+    range.BaseShaderRegister = 3;
     range.RegisterSpace = 0;
     range.OffsetInDescriptorsFromTableStart = 0;
   }
@@ -119,7 +119,7 @@ ID3D12RootSignature* D3D12CommandProcessor::GetRootSignature(
     parameter.ShaderVisibility = D3D12_SHADER_VISIBILITY_PIXEL;
     range.RangeType = D3D12_DESCRIPTOR_RANGE_TYPE_CBV;
     range.NumDescriptors = 8;
-    range.BaseShaderRegister = 2;
+    range.BaseShaderRegister = 3;
     range.RegisterSpace = 0;
     range.OffsetInDescriptorsFromTableStart = 0;
   }
@@ -1083,8 +1083,6 @@ void D3D12CommandProcessor::UpdateSystemConstantValues(Endian index_endian) {
   system_constants_.ssaa_inv_scale[0] = ssaa_inv_scale_x;
   system_constants_.ssaa_inv_scale[1] = ssaa_inv_scale_y;
 
-  // TODO(Triang3l): Whether textures are 3D or stacked.
-
   cbuffer_bindings_system_.up_to_date &= dirty;
 }
 
@@ -1129,9 +1127,9 @@ bool D3D12CommandProcessor::UpdateBindings(
 
   // Begin updating descriptors.
   bool write_common_constant_views = false;
+  bool write_fetch_constant_view = false;
   bool write_vertex_float_constant_views = false;
   bool write_pixel_float_constant_views = false;
-  bool write_fetch_constant_view = false;
   // TODO(Triang3l): Update textures and samplers only if shaders or binding
   // hash change.
   bool write_textures = texture_count != 0;
@@ -1171,6 +1169,18 @@ bool D3D12CommandProcessor::UpdateBindings(
     cbuffer_bindings_bool_loop_.up_to_date = true;
     write_common_constant_views = true;
   }
+  if (!cbuffer_bindings_fetch_.up_to_date) {
+    uint8_t* fetch_constants = constant_buffer_pool_->RequestFull(
+        768, nullptr, nullptr, &cbuffer_bindings_fetch_.buffer_address);
+    if (fetch_constants == nullptr) {
+      return false;
+    }
+    std::memcpy(fetch_constants,
+                &regs[XE_GPU_REG_SHADER_CONSTANT_FETCH_00_0].u32,
+                32 * 6 * sizeof(uint32_t));
+    cbuffer_bindings_fetch_.up_to_date = true;
+    write_fetch_constant_view = true;
+  }
   for (uint32_t i = 0; i < 16; ++i) {
     ConstantBufferBinding& float_binding = cbuffer_bindings_float_[i];
     if (float_binding.up_to_date) {
@@ -1191,18 +1201,6 @@ bool D3D12CommandProcessor::UpdateBindings(
       write_pixel_float_constant_views = true;
     }
   }
-  if (!cbuffer_bindings_fetch_.up_to_date) {
-    uint8_t* fetch_constants = constant_buffer_pool_->RequestFull(
-        768, nullptr, nullptr, &cbuffer_bindings_fetch_.buffer_address);
-    if (fetch_constants == nullptr) {
-      return false;
-    }
-    std::memcpy(fetch_constants,
-                &regs[XE_GPU_REG_SHADER_CONSTANT_FETCH_00_0].u32,
-                32 * 6 * sizeof(uint32_t));
-    cbuffer_bindings_fetch_.up_to_date = true;
-    write_fetch_constant_view = true;
-  }
 
   // Allocate the descriptors.
   uint32_t view_count_partial_update = 0;
@@ -1210,6 +1208,10 @@ bool D3D12CommandProcessor::UpdateBindings(
     // System and bool/loop constants.
     view_count_partial_update += 2;
   }
+  if (write_fetch_constant_view) {
+    // Fetch constants.
+    ++view_count_partial_update;
+  }
   if (write_vertex_float_constant_views) {
     // Vertex float constants.
     view_count_partial_update += 8;
@@ -1218,10 +1220,6 @@ bool D3D12CommandProcessor::UpdateBindings(
     // Pixel float constants.
     view_count_partial_update += 8;
   }
-  if (write_fetch_constant_view) {
-    // Fetch constants.
-    ++view_count_partial_update;
-  }
   if (write_textures) {
     view_count_partial_update += texture_count;
   }
@@ -1254,9 +1252,9 @@ bool D3D12CommandProcessor::UpdateBindings(
     // Need to update all view descriptors.
     draw_view_full_update_ = view_full_update_index;
     write_common_constant_views = true;
+    write_fetch_constant_view = true;
     write_vertex_float_constant_views = true;
     write_pixel_float_constant_views = true;
-    write_fetch_constant_view = true;
     write_textures = texture_count != 0;
     // If updating fully, write the shared memory descriptor (t0, space1).
     shared_memory_->CreateSRV(view_cpu_handle);
@@ -1293,9 +1291,20 @@ bool D3D12CommandProcessor::UpdateBindings(
     current_graphics_root_up_to_date_ &=
         ~(1u << kRootParameter_CommonConstants);
   }
+  if (write_fetch_constant_view) {
+    gpu_handle_fetch_constants_ = view_gpu_handle;
+    // Fetch constants (b2).
+    constant_buffer_desc.BufferLocation =
+        cbuffer_bindings_fetch_.buffer_address;
+    constant_buffer_desc.SizeInBytes = 768;
+    device->CreateConstantBufferView(&constant_buffer_desc, view_cpu_handle);
+    view_cpu_handle.ptr += view_handle_size;
+    view_gpu_handle.ptr += view_handle_size;
+    current_graphics_root_up_to_date_ &= ~(1u << kRootParameter_FetchConstants);
+  }
   if (write_vertex_float_constant_views) {
     gpu_handle_vertex_float_constants_ = view_gpu_handle;
-    // Vertex float constants (b2-b9).
+    // Vertex float constants (b3-b10).
     for (uint32_t i = 0; i < 8; ++i) {
       constant_buffer_desc.BufferLocation =
           cbuffer_bindings_float_[i].buffer_address;
@@ -1309,7 +1318,7 @@ bool D3D12CommandProcessor::UpdateBindings(
   }
   if (write_pixel_float_constant_views) {
     gpu_handle_pixel_float_constants_ = view_gpu_handle;
-    // Pixel float constants (b2-b9).
+    // Pixel float constants (b3-b10).
     for (uint32_t i = 0; i < 8; ++i) {
       constant_buffer_desc.BufferLocation =
           cbuffer_bindings_float_[8 + i].buffer_address;
@@ -1321,17 +1330,6 @@ bool D3D12CommandProcessor::UpdateBindings(
     current_graphics_root_up_to_date_ &=
         ~(1u << kRootParameter_PixelFloatConstants);
   }
-  if (write_fetch_constant_view) {
-    gpu_handle_fetch_constants_ = view_gpu_handle;
-    // Fetch constants (b10).
-    constant_buffer_desc.BufferLocation =
-        cbuffer_bindings_fetch_.buffer_address;
-    constant_buffer_desc.SizeInBytes = 768;
-    device->CreateConstantBufferView(&constant_buffer_desc, view_cpu_handle);
-    view_cpu_handle.ptr += view_handle_size;
-    view_gpu_handle.ptr += view_handle_size;
-    current_graphics_root_up_to_date_ &= ~(1u << kRootParameter_FetchConstants);
-  }
   if (write_textures) {
     if (pixel_texture_count != 0) {
       assert_true(current_graphics_root_extras_.pixel_textures !=

src/xenia/gpu/d3d12/d3d12_command_processor.h

@@ -95,14 +95,14 @@ class D3D12CommandProcessor : public CommandProcessor {
     // These are always present.
 
     // Very frequently changed, especially for UI draws, and for models drawn in
-    // multiple parts - contains fetch constants with vertex addresses (b10).
+    // multiple parts - contains vertex and texture fetch constants (b2).
     kRootParameter_FetchConstants,
     // Quite frequently changed (for one object drawn multiple times, for
     // instance - may contain projection matrices) - 8 pages of float constants
-    // (b2-b9).
+    // (b3-b10).
     kRootParameter_VertexFloatConstants,
     // Less frequently changed (per-material) - 8 pages of float constants
-    // (b2-b9).
+    // (b3-b10).
     kRootParameter_PixelFloatConstants,
     // Rarely changed - system constants like viewport and alpha testing (b0)
     // and loop and bool constants (b1).
@@ -223,9 +223,9 @@ class D3D12CommandProcessor : public CommandProcessor {
 
   // Latest descriptor handles used for handling Xenos draw calls.
   D3D12_GPU_DESCRIPTOR_HANDLE gpu_handle_common_constants_;
+  D3D12_GPU_DESCRIPTOR_HANDLE gpu_handle_fetch_constants_;
   D3D12_GPU_DESCRIPTOR_HANDLE gpu_handle_vertex_float_constants_;
   D3D12_GPU_DESCRIPTOR_HANDLE gpu_handle_pixel_float_constants_;
-  D3D12_GPU_DESCRIPTOR_HANDLE gpu_handle_fetch_constants_;
   D3D12_GPU_DESCRIPTOR_HANDLE gpu_handle_shared_memory_;
   D3D12_GPU_DESCRIPTOR_HANDLE gpu_handle_pixel_textures_;
   D3D12_GPU_DESCRIPTOR_HANDLE gpu_handle_pixel_samplers_;

src/xenia/gpu/hlsl_shader_translator.cc

@@ -176,7 +176,6 @@ std::vector<uint8_t> HlslShaderTranslator::CompleteTranslation() {
       "  float xe_pixel_half_pixel_offset;\n"
       "  float2 xe_ssaa_inv_scale;\n"
       "  uint xe_pixel_pos_reg;\n"
-      "  uint xe_textures_are_3d;\n"
       "};\n"
       "\n"
       "cbuffer xe_loop_bool_constants : register(b1) {\n"
@@ -184,11 +183,15 @@ std::vector<uint8_t> HlslShaderTranslator::CompleteTranslation() {
       "  uint4 xe_loop_constants[32];\n"
       "};\n"
       "\n"
+      "cbuffer xe_fetch_constants : register(b2) {\n"
+      "  uint4 xe_fetch[48];\n"
+      "};\n"
+      "\n"
       "struct XeFloatConstantPage {\n"
       "  float4 c[32];\n"
       "};\n"
       "ConstantBuffer<XeFloatConstantPage> "
-      "xe_float_constants[8] : register(b2);\n"
+      "xe_float_constants[8] : register(b3);\n"
       "\n");
 
   // Textures and samplers.
@@ -232,10 +235,6 @@ std::vector<uint8_t> HlslShaderTranslator::CompleteTranslation() {
     // -1 point size means the geometry shader will use the global setting by
     // default.
     source.AppendFormat(
-        "cbuffer xe_vertex_fetch_constants : register(b10) {\n"
-        "  uint4 xe_vertex_fetch[48];\n"
-        "};\n"
-        "\n"
         "ByteAddressBuffer xe_shared_memory : register(t0, space1);\n"
         "\n"
         "#define XE_BYTE_SWAP_OVERLOAD(XeByteSwapType) \\\n"
@@ -903,8 +902,8 @@ void HlslShaderTranslator::ProcessVertexFetchInstruction(
   }
   EmitSourceDepth("xe_vertex_element%s = XeByteSwap(xe_shared_memory.Load%s(\n",
                   load_swizzle, load_function_suffix);
-  EmitSourceDepth("    (xe_vertex_fetch[%uu].%c & 0x1FFFFFFCu)",
+  EmitSourceDepth("    (xe_fetch[%uu].%c & 0x1FFFFFFCu)", vfetch_index >> 1,
-                  vfetch_index >> 1, (vfetch_index & 1) ? 'z' : 'x');
+                  (vfetch_index & 1) ? 'z' : 'x');
   if (instr.attributes.stride != 0) {
     EmitSource(" + uint(xe_src0.x) * %uu", instr.attributes.stride * 4);
   }
@@ -912,7 +911,7 @@ void HlslShaderTranslator::ProcessVertexFetchInstruction(
     EmitSource(" + %uu", instr.attributes.offset * 4);
   }
   EmitSource("),\n");
-  EmitSourceDepth("    xe_vertex_fetch[%uu].%c);\n", vfetch_index >> 1,
+  EmitSourceDepth("    xe_fetch[%uu].%c);\n", vfetch_index >> 1,
                   (vfetch_index & 1) ? 'w' : 'y');
 
   // Convert to the target format.

src/xenia/gpu/hlsl_shader_translator.h

@@ -37,7 +37,6 @@ class HlslShaderTranslator : public ShaderTranslator {
     // vec4 3
     float ssaa_inv_scale[2];
     uint32_t pixel_pos_reg;
-    uint32_t textures_are_3d;
   };
 
   struct TextureSRV {