D3D11Device: Fix blend state not applying to MRTs

src/util/d3d11_device.h

@@ -118,9 +118,14 @@ protected:
   void DestroyDevice() override;
 
 private:
+  using BlendStateMapKey = std::pair<u64, u32>;
+  struct BlendStateMapHash
+  {
+    size_t operator()(const BlendStateMapKey& key) const;
+  };
   using RasterizationStateMap = std::unordered_map<u8, ComPtr<ID3D11RasterizerState>>;
   using DepthStateMap = std::unordered_map<u8, ComPtr<ID3D11DepthStencilState>>;
-  using BlendStateMap = std::unordered_map<u64, ComPtr<ID3D11BlendState>>;
+  using BlendStateMap = std::unordered_map<BlendStateMapKey, ComPtr<ID3D11BlendState>, BlendStateMapHash>;
   using InputLayoutMap =
     std::unordered_map<GPUPipeline::InputLayout, ComPtr<ID3D11InputLayout>, GPUPipeline::InputLayoutHash>;
 
@@ -146,7 +151,7 @@ private:
 
   ComPtr<ID3D11RasterizerState> GetRasterizationState(const GPUPipeline::RasterizationState& rs, Error* error);
   ComPtr<ID3D11DepthStencilState> GetDepthState(const GPUPipeline::DepthState& ds, Error* error);
-  ComPtr<ID3D11BlendState> GetBlendState(const GPUPipeline::BlendState& bs, Error* error);
+  ComPtr<ID3D11BlendState> GetBlendState(const GPUPipeline::BlendState& bs, u32 num_rts, Error* error);
   ComPtr<ID3D11InputLayout> GetInputLayout(const GPUPipeline::InputLayout& il, const D3D11Shader* vs, Error* error);
 
   bool CreateTimestampQueries();

src/util/d3d11_pipeline.cpp

@@ -6,6 +6,7 @@
 #include "d3d_common.h"
 
 #include "common/error.h"
+#include "common/hash_combine.h"
 
 #include "fmt/format.h"
 
@@ -206,11 +207,19 @@ D3D11Device::ComPtr<ID3D11DepthStencilState> D3D11Device::GetDepthState(const GP
   return dds;
 }
 
-D3D11Device::ComPtr<ID3D11BlendState> D3D11Device::GetBlendState(const GPUPipeline::BlendState& bs, Error* error)
+size_t D3D11Device::BlendStateMapHash::operator()(const BlendStateMapKey& key) const
+{
+  size_t h = std::hash<u64>()(key.first);
+  hash_combine(h, key.second);
+  return h;
+}
+
+D3D11Device::ComPtr<ID3D11BlendState> D3D11Device::GetBlendState(const GPUPipeline::BlendState& bs, u32 num_rts, Error* error)
 {
   ComPtr<ID3D11BlendState> dbs;
 
-  const auto it = m_blend_states.find(bs.key);
+  const std::pair<u64, u32> key(bs.key, num_rts);
+  const auto it = m_blend_states.find(key);
   if (it != m_blend_states.end())
   {
     dbs = it->second;
@@ -243,7 +252,9 @@ D3D11Device::ComPtr<ID3D11BlendState> D3D11Device::GetBlendState(const GPUPipeli
   }};
 
   D3D11_BLEND_DESC blend_desc = {};
-  D3D11_RENDER_TARGET_BLEND_DESC& tgt_desc = blend_desc.RenderTarget[0];
+  for (u32 i = 0; i < num_rts; i++)
+  {
+    D3D11_RENDER_TARGET_BLEND_DESC& tgt_desc = blend_desc.RenderTarget[i];
     tgt_desc.BlendEnable = bs.enable;
     tgt_desc.RenderTargetWriteMask = bs.write_mask;
     if (bs.enable)
@@ -255,12 +266,13 @@ D3D11Device::ComPtr<ID3D11BlendState> D3D11Device::GetBlendState(const GPUPipeli
       tgt_desc.DestBlendAlpha = blend_mapping[static_cast<u8>(bs.dst_alpha_blend.GetValue())];
       tgt_desc.BlendOpAlpha = op_mapping[static_cast<u8>(bs.alpha_blend_op.GetValue())];
     }
+  }
 
   HRESULT hr = m_device->CreateBlendState(&blend_desc, dbs.GetAddressOf());
   if (FAILED(hr)) [[unlikely]]
     Error::SetHResult(error, "CreateBlendState() failed: ", hr);
   else
-    m_blend_states.emplace(bs.key, dbs);
+    m_blend_states.emplace(key, dbs);
 
   return dbs;
 }
@@ -322,7 +334,7 @@ std::unique_ptr<GPUPipeline> D3D11Device::CreatePipeline(const GPUPipeline::Grap
 {
   ComPtr<ID3D11RasterizerState> rs = GetRasterizationState(config.rasterization, error);
   ComPtr<ID3D11DepthStencilState> ds = GetDepthState(config.depth, error);
-  ComPtr<ID3D11BlendState> bs = GetBlendState(config.blend, error);
+  ComPtr<ID3D11BlendState> bs = GetBlendState(config.blend, config.GetRenderTargetCount(), error);
   if (!rs || !ds || !bs)
     return {};
 

src/util/gpu_device.cpp

@@ -167,6 +167,17 @@ void GPUPipeline::GraphicsConfig::SetTargetFormats(GPUTexture::Format color_form
   depth_format = depth_format_;
 }
 
+u32 GPUPipeline::GraphicsConfig::GetRenderTargetCount() const
+{
+  u32 num_rts = 0;
+  for (; num_rts < static_cast<u32>(std::size(color_formats)); num_rts++)
+  {
+    if (color_formats[num_rts] == GPUTexture::Format::Unknown)
+      break;
+  }
+  return num_rts;
+}
+
 GPUTextureBuffer::GPUTextureBuffer(Format format, u32 size) : m_format(format), m_size_in_elements(size)
 {
 }

src/util/gpu_device.h

@@ -413,6 +413,7 @@ public:
 
     void SetTargetFormats(GPUTexture::Format color_format,
                           GPUTexture::Format depth_format_ = GPUTexture::Format::Unknown);
+    u32 GetRenderTargetCount() const;
   };
 
   GPUPipeline();