[D3D12] ROV: Check if supported

src/xenia/gpu/d3d12/d3d12_command_processor.cc

@@ -628,8 +628,6 @@ bool D3D12CommandProcessor::SetupContext() {
     return false;
   }
 
-  pipeline_cache_ = std::make_unique<PipelineCache>(this, register_file_);
-
   texture_cache_ = std::make_unique<TextureCache>(this, register_file_,
                                                   shared_memory_.get());
   if (!texture_cache_->Initialize()) {
@@ -644,6 +642,9 @@ bool D3D12CommandProcessor::SetupContext() {
     return false;
   }
 
+  pipeline_cache_ = std::make_unique<PipelineCache>(
+      this, register_file_, render_target_cache_->IsROVUsedForEDRAM());
+
   primitive_converter_ =
       std::make_unique<PrimitiveConverter>(this, register_file_, memory_);
   if (!primitive_converter_->Initialize()) {
@@ -810,12 +811,12 @@ void D3D12CommandProcessor::ShutdownContext() {
 
   primitive_converter_.reset();
 
+  pipeline_cache_.reset();
+
   render_target_cache_.reset();
 
   texture_cache_.reset();
 
-  pipeline_cache_.reset();
-
   // Root signatured are used by pipelines, thus freed after the pipelines.
   for (auto it : root_signatures_) {
     it.second->Release();
@@ -1021,12 +1022,12 @@ void D3D12CommandProcessor::PerformSwap(uint32_t frontbuffer_ptr,
 
     primitive_converter_->ClearCache();
 
+    pipeline_cache_->ClearCache();
+
     render_target_cache_->ClearCache();
 
     texture_cache_->ClearCache();
 
-    pipeline_cache_->ClearCache();
-
     for (auto it : root_signatures_) {
       it.second->Release();
     }

src/xenia/gpu/d3d12/pipeline_cache.cc

@@ -36,9 +36,11 @@ namespace d3d12 {
 #include "xenia/gpu/d3d12/shaders/dxbc/primitive_rectangle_list_gs.h"
 
 PipelineCache::PipelineCache(D3D12CommandProcessor* command_processor,
-                             RegisterFile* register_file)
-    : command_processor_(command_processor), register_file_(register_file) {
-  shader_translator_ = std::make_unique<DxbcShaderTranslator>();
+                             RegisterFile* register_file, bool edram_rov_used)
+    : command_processor_(command_processor),
+      register_file_(register_file),
+      edram_rov_used_(edram_rov_used) {
+  shader_translator_ = std::make_unique<DxbcShaderTranslator>(edram_rov_used_);
 
   // Set pipeline state description values we never change.
   // Zero out tessellation, stream output, blend state and formats for render

src/xenia/gpu/d3d12/pipeline_cache.h

@@ -35,7 +35,7 @@ class PipelineCache {
   };
 
   PipelineCache(D3D12CommandProcessor* command_processor,
-                RegisterFile* register_file);
+                RegisterFile* register_file, bool edram_rov_used);
   ~PipelineCache();
 
   void Shutdown();
@@ -85,6 +85,9 @@ class PipelineCache {
   D3D12CommandProcessor* command_processor_;
   RegisterFile* register_file_;
 
+  // Whether the output merger is emulated in pixel shaders.
+  bool edram_rov_used_;
+
   // Reusable shader translator.
   std::unique_ptr<DxbcShaderTranslator> shader_translator_ = nullptr;
   // All loaded shaders mapped by their guest hash key.

src/xenia/gpu/d3d12/render_target_cache.cc

@@ -9,6 +9,8 @@
 
 #include "xenia/gpu/d3d12/render_target_cache.h"
 
+#include <gflags/gflags.h>
+
 #include <algorithm>
 #include <cmath>
 #include <cstring>
@@ -23,6 +25,10 @@
 #include "xenia/gpu/texture_util.h"
 #include "xenia/ui/d3d12/d3d12_util.h"
 
+DEFINE_bool(d3d12_rov, false,
+            "Use rasterizer-ordered views for render target emulation where "
+            "available (experimental and currently largely unimplemented).");
+
 namespace xe {
 namespace gpu {
 namespace d3d12 {
@@ -330,6 +336,14 @@ void RenderTargetCache::ClearCache() {
   }
 }
 
+bool RenderTargetCache::IsROVUsedForEDRAM() const {
+  if (!FLAGS_d3d12_rov) {
+    return false;
+  }
+  auto provider = command_processor_->GetD3D12Context()->GetD3D12Provider();
+  return provider->AreRasterizerOrderedViewsSupported();
+}
+
 void RenderTargetCache::BeginFrame() {
   ClearBindings();
 

src/xenia/gpu/d3d12/render_target_cache.h

@@ -185,6 +185,38 @@ class D3D12CommandProcessor;
 // multisampled surface is the same as a single-sampled surface with 2x height
 // and width - however, format size doesn't effect the dimensions. Surface pitch
 // in the surface info register is single-sampled.
+//
+// =============================================================================
+// Rasterizer-ordered view usage:
+// =============================================================================
+//
+// There is a separate output merger emulation path currently in development,
+// using rasterizer-ordered views for writing directly to the 10 MB EDRAM buffer
+// instead of the host output merger for render target output.
+//
+// The convential method of implementing Xenos render targets via host render
+// targets has various flaws that may be impossible to fix:
+// - k_16_16 and k_16_16_16_16 have -32...32 range on Xenos, but there's no
+//   equivalent format on PC APIs. They may be emulated using snorm16 (by
+//   dividing shader color output by 32) or float32, however, blending behaves
+//   incorrectly for both. In the former case, multiplicative blending may not
+//   work correctly - 1 becomes 1/32, and instead of 1 * 1 = 1, you get
+//   1/32 * 1/32 = 1/1024. For 32-bit floats, additive blending result may go up
+//   to infinity.
+// - k_2_10_10_10_FLOAT has similar blending issues, though less prominent, when
+//   emulated via float16 render targets. In addition to a greater range for
+//   RGB (values can go up to 65504 and infinity rather than 31.875), alpha is
+//   represented totally differently - in k_2_10_10_10_FLOAT, it may have only
+//   4 values, and adding, for example, 0.1 to 0.333 will still result in 0.333,
+//   while with float16, it will be increasing, and the limit is infinity.
+// - Due to simultaneously bound host render targets being independent from each
+//   other, and because the height is unknown (and the viewport and scissor are
+//   not always present - D3DPT_RECTLIST is used very commonly, especially for
+//   clearing (Direct3D 9 Clear is implemented this way on the Xbox 360) and
+//   copying, and it's usually drawn without a viewport and with 8192x8192
+//   scissor), there may be cases of simulatenously bound render targets
+//   overlapping each other in the EDRAM in a way that is difficult to resolve,
+//   and stores/loads may destroy data.
 class RenderTargetCache {
  public:
   // Direct3D 12 debug layer does some kaschenit-style trolling by giving errors
@@ -208,6 +240,11 @@ class RenderTargetCache {
   void Shutdown();
   void ClearCache();
 
+  // Should a rasterizer-ordered UAV of the EDRAM buffer with format conversion
+  // and blending performed in pixel shaders be used instead of host render
+  // targets.
+  bool IsROVUsedForEDRAM() const;
+
   void BeginFrame();
   // Called in the beginning of a draw call - may bind pipelines.
   bool UpdateRenderTargets(const D3D12Shader* pixel_shader);

src/xenia/gpu/dxbc_shader_translator.cc

@@ -62,7 +62,8 @@ using namespace ucode;
 //   second buffer in the descriptor array at b2, which is assigned to CB1, the
 //   index would be CB1[3][0].
 
-DxbcShaderTranslator::DxbcShaderTranslator() {
+DxbcShaderTranslator::DxbcShaderTranslator(bool edram_rovs_used)
+    : edram_rovs_used_(edram_rovs_used) {
   // Don't allocate again and again for the first shader.
   shader_code_.reserve(8192);
   shader_object_.reserve(16384);

src/xenia/gpu/dxbc_shader_translator.h

@@ -23,7 +23,7 @@ namespace gpu {
 // Generates shader model 5_1 byte code (for Direct3D 12).
 class DxbcShaderTranslator : public ShaderTranslator {
  public:
-  DxbcShaderTranslator();
+  DxbcShaderTranslator(bool edram_rovs_used);
   ~DxbcShaderTranslator() override;
 
   // Constant buffer bindings in space 0.
@@ -431,6 +431,9 @@ class DxbcShaderTranslator : public ShaderTranslator {
   // generated in the end of translation.
   std::vector<uint32_t> shader_object_;
 
+  // Whether the output merger should be emulated in pixel shaders.
+  bool edram_rovs_used_;
+
   // Data types used in constants buffers. Listed in dependency order.
   enum class RdefTypeIndex {
     kFloat,

src/xenia/ui/d3d12/d3d12_provider.cc

@@ -150,12 +150,13 @@ bool D3D12Provider::Initialize() {
   descriptor_size_dsv_ =
       device->GetDescriptorHandleIncrementSize(D3D12_DESCRIPTOR_HEAP_TYPE_DSV);
 
-  // Check if tiled resources and programmable sample positions (programmable
-  // sample positions added in Creators Update) are supported.
+  // Check if optional features are supported.
+  rasterizer_ordered_views_supported_ = false;
   tiled_resources_tier_ = 0;
   D3D12_FEATURE_DATA_D3D12_OPTIONS options;
   if (SUCCEEDED(device->CheckFeatureSupport(D3D12_FEATURE_D3D12_OPTIONS,
                                             &options, sizeof(options)))) {
+    rasterizer_ordered_views_supported_ = options.ROVsSupported ? true : false;
     tiled_resources_tier_ = uint32_t(options.TiledResourcesTier);
   }
   programmable_sample_positions_tier_ = 0;
@@ -167,8 +168,9 @@ bool D3D12Provider::Initialize() {
   }
   XELOGD3D(
       "Direct3D 12 device supports tiled resources tier %u, programmable "
-      "sample positions tier %u",
-      tiled_resources_tier_, programmable_sample_positions_tier_);
+      "sample positions tier %u; rasterizer-ordered views %ssupported",
+      tiled_resources_tier_, programmable_sample_positions_tier_,
+      rasterizer_ordered_views_supported_ ? "" : "un");
 
   // Get the graphics analysis interface, will silently fail if PIX not
   // attached.

src/xenia/ui/d3d12/d3d12_provider.h

@@ -62,6 +62,9 @@ class D3D12Provider : public GraphicsProvider {
     return start;
   }
 
+  bool AreRasterizerOrderedViewsSupported() const {
+    return rasterizer_ordered_views_supported_;
+  }
   uint32_t GetTiledResourcesTier() const { return tiled_resources_tier_; }
   uint32_t GetProgrammableSamplePositionsTier() const {
     return programmable_sample_positions_tier_;
@@ -82,6 +85,7 @@ class D3D12Provider : public GraphicsProvider {
   uint32_t descriptor_size_rtv_;
   uint32_t descriptor_size_dsv_;
 
+  bool rasterizer_ordered_views_supported_;
   uint32_t tiled_resources_tier_;
   uint32_t programmable_sample_positions_tier_;
 };