[D3D12] ROV: Empty pixel shader for depth-only drawing

src/xenia/gpu/d3d12/pipeline_cache.cc

@@ -15,6 +15,7 @@
 #include <cinttypes>
 #include <cmath>
 #include <cstring>
+#include <utility>
 
 #include "xenia/base/assert.h"
 #include "xenia/base/logging.h"
@@ -42,6 +43,11 @@ PipelineCache::PipelineCache(D3D12CommandProcessor* command_processor,
       edram_rov_used_(edram_rov_used) {
   shader_translator_ = std::make_unique<DxbcShaderTranslator>(edram_rov_used_);
 
+  if (edram_rov_used_) {
+    depth_only_pixel_shader_ =
+        std::move(shader_translator_->CreateDepthOnlyPixelShader());
+  }
+
   // Set pipeline state description values we never change.
   // Zero out tessellation, stream output, blend state and formats for render
   // targets 4+, node mask, cached PSO, flags and other things.
@@ -325,8 +331,13 @@ PipelineCache::UpdateStatus PipelineCache::UpdateShaderStages(
     update_desc_.PS.pShaderBytecode = pixel_shader->translated_binary().data();
     update_desc_.PS.BytecodeLength = pixel_shader->translated_binary().size();
   } else {
-    update_desc_.PS.pShaderBytecode = nullptr;
+    if (edram_rov_used_) {
-    update_desc_.PS.BytecodeLength = 0;
+      update_desc_.PS.pShaderBytecode = depth_only_pixel_shader_.data();
+      update_desc_.PS.BytecodeLength = depth_only_pixel_shader_.size();
+    } else {
+      update_desc_.PS.pShaderBytecode = nullptr;
+      update_desc_.PS.BytecodeLength = 0;
+    }
   }
   switch (primitive_type) {
     case PrimitiveType::kPointList:

src/xenia/gpu/d3d12/pipeline_cache.h

@@ -11,6 +11,7 @@
 #define XENIA_GPU_D3D12_PIPELINE_CACHE_H_
 
 #include <unordered_map>
+#include <vector>
 
 #include "third_party/xxhash/xxhash.h"
 
@@ -93,6 +94,10 @@ class PipelineCache {
   // All loaded shaders mapped by their guest hash key.
   std::unordered_map<uint64_t, D3D12Shader*> shader_map_;
 
+  // Empty depth-only pixel shader for writing to depth buffer via ROV when no
+  // Xenos pixel shader provided.
+  std::vector<uint8_t> depth_only_pixel_shader_;
+
   // Hash state used to incrementally produce pipeline hashes during update.
   // By the time the full update pass has run the hash will represent the
   // current state in a way that can uniquely identify the produced

src/xenia/gpu/dxbc_shader_translator.h

@@ -376,6 +376,10 @@ class DxbcShaderTranslator : public ShaderTranslator {
   static bool GetBlendConstants(uint32_t blend_control, uint32_t& blend_x_out,
                                 uint32_t& blend_y_out);
 
+  // Creates a special pixel shader without color outputs - this resets the
+  // state of the translator.
+  std::vector<uint8_t> CreateDepthOnlyPixelShader();
+
  protected:
   void Reset() override;
 
@@ -603,6 +607,15 @@ class DxbcShaderTranslator : public ShaderTranslator {
            (index_representation_1 << 25) | (index_representation_2 << 28);
   }
 
+  // Use these instead of is_vertex_shader/is_pixel_shader because they don't
+  // take is_depth_only_pixel_shader_ into account.
+  inline bool IsDXBCVertexShader() const {
+    return !is_depth_only_pixel_shader_ && is_vertex_shader();
+  }
+  inline bool IsDXBCPixelShader() const {
+    return is_depth_only_pixel_shader_ || is_pixel_shader();
+  }
+
   // Allocates a new r# register for internal use and returns its index.
   uint32_t PushSystemTemp(bool zero = false);
   // Frees the last allocated internal r# registers for later reuse.
@@ -783,6 +796,10 @@ class DxbcShaderTranslator : public ShaderTranslator {
   // Whether the output merger should be emulated in pixel shaders.
   bool edram_rov_used_;
 
+  // Is currently writing the empty depth-only pixel shader, for
+  // CompleteTranslation.
+  bool is_depth_only_pixel_shader_;
+
   // Data types used in constants buffers. Listed in dependency order.
   enum class RdefTypeIndex {
     kFloat,