[d3d11] Optimize UpdateSubresource for small buffer updates

Some games use UpdateSubresource to upload constant buffers in
between draws, so this path should be as fast as possible.

Also fixes a potential issue when using D3D11_COPY_NO_OVERWRITE
on deferred contexts, since the Map requirements don't hold here.

src/d3d11/d3d11_context.cpp

@@ -903,134 +903,6 @@ namespace dxvk {
     });
   }
   
-  
-  void STDMETHODCALLTYPE D3D11DeviceContext::UpdateSubresource(
-          ID3D11Resource*                   pDstResource,
-          UINT                              DstSubresource,
-    const D3D11_BOX*                        pDstBox,
-    const void*                             pSrcData,
-          UINT                              SrcRowPitch,
-          UINT                              SrcDepthPitch) {
-    UpdateSubresource1(pDstResource,
-      DstSubresource, pDstBox, pSrcData,
-      SrcRowPitch, SrcDepthPitch, 0);
-  }
-
-
-  void STDMETHODCALLTYPE D3D11DeviceContext::UpdateSubresource1(
-          ID3D11Resource*                   pDstResource, 
-          UINT                              DstSubresource, 
-    const D3D11_BOX*                        pDstBox, 
-    const void*                             pSrcData, 
-          UINT                              SrcRowPitch, 
-          UINT                              SrcDepthPitch, 
-          UINT                              CopyFlags) {
-    D3D10DeviceLock lock = LockContext();
-
-    if (!pDstResource)
-      return;
-    
-    // Filter out invalid copy flags
-    CopyFlags &= D3D11_COPY_NO_OVERWRITE | D3D11_COPY_DISCARD;
-
-    // We need a different code path for buffers
-    D3D11_RESOURCE_DIMENSION resourceType;
-    pDstResource->GetType(&resourceType);
-    
-    if (resourceType == D3D11_RESOURCE_DIMENSION_BUFFER) {
-      const auto bufferResource = static_cast<D3D11Buffer*>(pDstResource);
-      const auto bufferSlice = bufferResource->GetBufferSlice();
-      
-      VkDeviceSize offset = bufferSlice.offset();
-      VkDeviceSize size   = bufferSlice.length();
-      
-      if (pDstBox != nullptr) {
-        offset = pDstBox->left;
-        size   = pDstBox->right - pDstBox->left;
-      }
-      
-      if (!size || offset + size > bufferSlice.length())
-        return;
-
-      bool useMap = (bufferSlice.buffer()->memFlags() & VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT)
-                 && (size == bufferSlice.length() || CopyFlags);
-      
-      if (useMap) {
-        D3D11_MAP mapType = (CopyFlags & D3D11_COPY_NO_OVERWRITE)
-          ? D3D11_MAP_WRITE_NO_OVERWRITE
-          : D3D11_MAP_WRITE_DISCARD;
-
-        D3D11_MAPPED_SUBRESOURCE mappedSr;
-        if (likely(useMap = SUCCEEDED(Map(pDstResource, 0, mapType, 0, &mappedSr)))) {
-          std::memcpy(reinterpret_cast<char*>(mappedSr.pData) + offset, pSrcData, size);
-          Unmap(pDstResource, 0);
-        }
-      }
-
-      if (!useMap) {
-        DxvkDataSlice dataSlice = AllocUpdateBufferSlice(size);
-        std::memcpy(dataSlice.ptr(), pSrcData, size);
-        
-        EmitCs([
-          cDataBuffer   = std::move(dataSlice),
-          cBufferSlice  = bufferSlice.subSlice(offset, size)
-        ] (DxvkContext* ctx) {
-          ctx->updateBuffer(
-            cBufferSlice.buffer(),
-            cBufferSlice.offset(),
-            cBufferSlice.length(),
-            cDataBuffer.ptr());
-        });
-      }
-    } else {
-      D3D11CommonTexture* dstTexture = GetCommonTexture(pDstResource);
-      
-      if (DstSubresource >= dstTexture->CountSubresources())
-        return;
-      
-      VkFormat packedFormat = dstTexture->GetPackedFormat();
-
-      auto formatInfo = imageFormatInfo(packedFormat);
-      auto subresource = dstTexture->GetSubresourceFromIndex(
-          formatInfo->aspectMask, DstSubresource);
-      
-      VkExtent3D mipExtent = dstTexture->MipLevelExtent(subresource.mipLevel);
-
-      VkOffset3D offset = { 0, 0, 0 };
-      VkExtent3D extent = mipExtent;
-      
-      if (pDstBox != nullptr) {
-        if (pDstBox->left >= pDstBox->right
-         || pDstBox->top >= pDstBox->bottom
-         || pDstBox->front >= pDstBox->back)
-          return;  // no-op, but legal
-        
-        offset.x = pDstBox->left;
-        offset.y = pDstBox->top;
-        offset.z = pDstBox->front;
-        
-        extent.width  = pDstBox->right - pDstBox->left;
-        extent.height = pDstBox->bottom - pDstBox->top;
-        extent.depth  = pDstBox->back - pDstBox->front;
-      }
-      
-      if (!util::isBlockAligned(offset, extent, formatInfo->blockSize, mipExtent)) {
-        Logger::err("D3D11: UpdateSubresource1: Unaligned region");
-        return;
-      }
-
-      auto stagingSlice = AllocStagingBuffer(util::computeImageDataSize(packedFormat, extent));
-
-      util::packImageData(stagingSlice.mapPtr(0),
-        pSrcData, SrcRowPitch, SrcDepthPitch, 0, 0,
-        dstTexture->GetVkImageType(), extent, 1,
-        formatInfo, formatInfo->aspectMask);
-
-      UpdateImage(dstTexture, &subresource,
-        offset, extent, std::move(stagingSlice));
-    }
-  }
-
 
   HRESULT STDMETHODCALLTYPE D3D11DeviceContext::UpdateTileMappings(
           ID3D11Resource*                   pTiledResource,
@@ -3668,6 +3540,102 @@ namespace dxvk {
   }
 
 
+  void D3D11DeviceContext::UpdateBuffer(
+          D3D11Buffer*                      pDstBuffer,
+          UINT                              Offset,
+          UINT                              Length,
+    const void*                             pSrcData) {
+    DxvkBufferSlice bufferSlice = pDstBuffer->GetBufferSlice(Offset, Length);
+
+    if (Length <= 65536) {
+      // The backend has special code paths for small buffer updates
+      DxvkDataSlice dataSlice = AllocUpdateBufferSlice(Length);
+      std::memcpy(dataSlice.ptr(), pSrcData, Length);
+
+      EmitCs([
+        cDataBuffer   = std::move(dataSlice),
+        cBufferSlice  = std::move(bufferSlice)
+      ] (DxvkContext* ctx) {
+        ctx->updateBuffer(
+          cBufferSlice.buffer(),
+          cBufferSlice.offset(),
+          cBufferSlice.length(),
+          cDataBuffer.ptr());
+      });
+    } else {
+      // Otherwise, to avoid large data copies on the CS thread,
+      // write directly to a staging buffer and dispatch a copy
+      DxvkBufferSlice stagingSlice = AllocStagingBuffer(Length);
+      std::memcpy(stagingSlice.mapPtr(0), pSrcData, Length);
+
+      EmitCs([
+        cStagingSlice = std::move(stagingSlice),
+        cBufferSlice  = std::move(bufferSlice)
+      ] (DxvkContext* ctx) {
+        ctx->copyBuffer(
+          cBufferSlice.buffer(),
+          cBufferSlice.offset(),
+          cStagingSlice.buffer(),
+          cStagingSlice.offset(),
+          cBufferSlice.length());
+      });
+    }
+  }
+
+
+  void D3D11DeviceContext::UpdateTexture(
+          D3D11CommonTexture*               pDstTexture,
+          UINT                              DstSubresource,
+    const D3D11_BOX*                        pDstBox,
+    const void*                             pSrcData,
+          UINT                              SrcRowPitch,
+          UINT                              SrcDepthPitch) {
+    if (DstSubresource >= pDstTexture->CountSubresources())
+      return;
+
+    VkFormat packedFormat = pDstTexture->GetPackedFormat();
+
+    auto formatInfo = imageFormatInfo(packedFormat);
+    auto subresource = pDstTexture->GetSubresourceFromIndex(
+        formatInfo->aspectMask, DstSubresource);
+
+    VkExtent3D mipExtent = pDstTexture->MipLevelExtent(subresource.mipLevel);
+
+    VkOffset3D offset = { 0, 0, 0 };
+    VkExtent3D extent = mipExtent;
+
+    if (pDstBox != nullptr) {
+      if (pDstBox->left >= pDstBox->right
+        || pDstBox->top >= pDstBox->bottom
+        || pDstBox->front >= pDstBox->back)
+        return;  // no-op, but legal
+
+      offset.x = pDstBox->left;
+      offset.y = pDstBox->top;
+      offset.z = pDstBox->front;
+
+      extent.width  = pDstBox->right - pDstBox->left;
+      extent.height = pDstBox->bottom - pDstBox->top;
+      extent.depth  = pDstBox->back - pDstBox->front;
+    }
+
+    if (!util::isBlockAligned(offset, extent, formatInfo->blockSize, mipExtent)) {
+      Logger::err("D3D11: UpdateSubresource1: Unaligned region");
+      return;
+    }
+
+    auto stagingSlice = AllocStagingBuffer(util::computeImageDataSize(packedFormat, extent));
+
+    util::packImageData(stagingSlice.mapPtr(0),
+      pSrcData, SrcRowPitch, SrcDepthPitch, 0, 0,
+      pDstTexture->GetVkImageType(), extent, 1,
+      formatInfo, formatInfo->aspectMask);
+
+    UpdateImage(pDstTexture, &subresource,
+      offset, extent, std::move(stagingSlice));
+  }
+
+
   void D3D11DeviceContext::UpdateImage(
           D3D11CommonTexture*               pDstTexture,
     const VkImageSubresource*               pDstSubresource,

src/d3d11/d3d11_context.h

@@ -133,23 +133,6 @@ namespace dxvk {
     void STDMETHODCALLTYPE GenerateMips(
             ID3D11ShaderResourceView*         pShaderResourceView);
     
-    void STDMETHODCALLTYPE UpdateSubresource(
-            ID3D11Resource*                   pDstResource,
-            UINT                              DstSubresource,
-      const D3D11_BOX*                        pDstBox,
-      const void*                             pSrcData,
-            UINT                              SrcRowPitch,
-            UINT                              SrcDepthPitch);
-    
-    void STDMETHODCALLTYPE UpdateSubresource1(
-            ID3D11Resource*                   pDstResource,
-            UINT                              DstSubresource,
-      const D3D11_BOX*                        pDstBox,
-      const void*                             pSrcData,
-            UINT                              SrcRowPitch,
-            UINT                              SrcDepthPitch,
-            UINT                              CopyFlags);
-
     HRESULT STDMETHODCALLTYPE UpdateTileMappings(
             ID3D11Resource*                   pTiledResource,
             UINT                              NumTiledResourceRegions,
@@ -807,6 +790,82 @@ namespace dxvk {
             ID3D11Resource*                   pResource,
             UINT                              Subresource);
 
+    template<typename ContextType>
+    static void UpdateResource(
+            ContextType*                      pContext,
+            ID3D11Resource*                   pDstResource,
+            UINT                              DstSubresource,
+      const D3D11_BOX*                        pDstBox,
+      const void*                             pSrcData,
+            UINT                              SrcRowPitch,
+            UINT                              SrcDepthPitch,
+            UINT                              CopyFlags) {
+      D3D10DeviceLock lock = pContext->LockContext();
+
+      if (!pDstResource)
+        return;
+
+      // We need a different code path for buffers
+      D3D11_RESOURCE_DIMENSION resourceType;
+      pDstResource->GetType(&resourceType);
+
+      if (likely(resourceType == D3D11_RESOURCE_DIMENSION_BUFFER)) {
+        const auto bufferResource = static_cast<D3D11Buffer*>(pDstResource);
+        uint64_t bufferSize = bufferResource->Desc()->ByteWidth;
+
+        // Provide a fast path for mapped buffer updates since some
+        // games use UpdateSubresource to update constant buffers.
+        if (likely(bufferResource->GetMapMode() == D3D11_COMMON_BUFFER_MAP_MODE_DIRECT) && likely(!pDstBox)) {
+          pContext->UpdateMappedBuffer(bufferResource, 0, bufferSize, pSrcData, 0);
+          return;
+        }
+
+        // Validate buffer range to update
+        uint64_t offset = 0;
+        uint64_t length = bufferSize;
+
+        if (pDstBox) {
+          offset = pDstBox->left;
+          length = pDstBox->right - offset;
+        }
+
+        if (unlikely(offset + length > bufferSize))
+          return;
+
+        // Still try to be fast if a box is provided but we update the full buffer
+        if (likely(bufferResource->GetMapMode() == D3D11_COMMON_BUFFER_MAP_MODE_DIRECT)) {
+          CopyFlags &= D3D11_COPY_DISCARD | D3D11_COPY_NO_OVERWRITE;
+
+          if (likely(length == bufferSize) || unlikely(CopyFlags != 0)) {
+            pContext->UpdateMappedBuffer(bufferResource, offset, length, pSrcData, CopyFlags);
+            return;
+          }
+        }
+
+        // Otherwise we can't really do anything fancy, so just do a GPU copy
+        pContext->UpdateBuffer(bufferResource, offset, length, pSrcData);
+      } else {
+        D3D11CommonTexture* textureResource = GetCommonTexture(pDstResource);
+
+        pContext->UpdateTexture(textureResource,
+          DstSubresource, pDstBox, pSrcData, SrcRowPitch, SrcDepthPitch);
+      }
+    }
+
+    void UpdateBuffer(
+            D3D11Buffer*                      pDstBuffer,
+            UINT                              Offset,
+            UINT                              Length,
+      const void*                             pSrcData);
+
+    void UpdateTexture(
+            D3D11CommonTexture*               pDstTexture,
+            UINT                              DstSubresource,
+      const D3D11_BOX*                        pDstBox,
+      const void*                             pSrcData,
+            UINT                              SrcRowPitch,
+            UINT                              SrcDepthPitch);
+
     void UpdateImage(
             D3D11CommonTexture*               pDstTexture,
       const VkImageSubresource*               pDstSubresource,
@@ -921,7 +980,7 @@ namespace dxvk {
             UINT                              NumViews,
             ID3D11RenderTargetView* const*    ppRenderTargetViews,
             ID3D11DepthStencilView*           pDepthStencilView);
-    
+
     VkClearValue ConvertColorValue(
       const FLOAT                             Color[4],
       const DxvkFormatInfo*                   pFormatInfo);

src/d3d11/d3d11_context_def.cpp

@@ -246,6 +246,31 @@ namespace dxvk {
   }
   
   
+  void STDMETHODCALLTYPE D3D11DeferredContext::UpdateSubresource(
+          ID3D11Resource*                   pDstResource,
+          UINT                              DstSubresource,
+    const D3D11_BOX*                        pDstBox,
+    const void*                             pSrcData,
+          UINT                              SrcRowPitch,
+          UINT                              SrcDepthPitch) {
+    UpdateResource<D3D11DeferredContext>(this, pDstResource,
+      DstSubresource, pDstBox, pSrcData, SrcRowPitch, SrcDepthPitch, 0);
+  }
+
+
+  void STDMETHODCALLTYPE D3D11DeferredContext::UpdateSubresource1(
+          ID3D11Resource*                   pDstResource,
+          UINT                              DstSubresource,
+    const D3D11_BOX*                        pDstBox,
+    const void*                             pSrcData,
+          UINT                              SrcRowPitch,
+          UINT                              SrcDepthPitch,
+          UINT                              CopyFlags) {
+    UpdateResource<D3D11DeferredContext>(this, pDstResource,
+      DstSubresource, pDstBox, pSrcData, SrcRowPitch, SrcDepthPitch, CopyFlags);
+  }
+
+
   void STDMETHODCALLTYPE D3D11DeferredContext::SwapDeviceContextState(
           ID3DDeviceContextState*           pState,
           ID3DDeviceContextState**          ppPreviousState) {
@@ -342,6 +367,35 @@ namespace dxvk {
   }
   
   
+  void D3D11DeferredContext::UpdateMappedBuffer(
+          D3D11Buffer*                  pDstBuffer,
+          UINT                          Offset,
+          UINT                          Length,
+    const void*                         pSrcData,
+          UINT                          CopyFlags) {
+    void* mapPtr = nullptr;
+
+    if (unlikely(CopyFlags == D3D11_COPY_NO_OVERWRITE)) {
+      auto entry = FindMapEntry(pDstBuffer, 0);
+
+      if (entry != m_mappedResources.rend())
+        mapPtr = entry->MapPointer;
+    }
+
+    if (likely(!mapPtr)) {
+      // The caller validates the map mode, so we can
+      // safely ignore the MapBuffer return value here
+      D3D11DeferredContextMapEntry entry;
+      MapBuffer(pDstBuffer, &entry);
+
+      mapPtr = entry.MapPointer;
+      m_mappedResources.push_back(std::move(entry));
+    }
+
+    std::memcpy(reinterpret_cast<char*>(mapPtr) + Offset, pSrcData, Length);
+  }
+
+
   void D3D11DeferredContext::FinalizeQueries() {
     for (auto& query : m_queriesBegun) {
       m_commandList->AddQuery(query.ptr());

src/d3d11/d3d11_context_def.h

@@ -19,7 +19,7 @@ namespace dxvk {
   };
   
   class D3D11DeferredContext : public D3D11DeviceContext {
-    
+    friend class D3D11DeviceContext;
   public:
     
     D3D11DeferredContext(
@@ -76,6 +76,23 @@ namespace dxvk {
             ID3D11Resource*             pResource,
             UINT                        Subresource);
     
+    void STDMETHODCALLTYPE UpdateSubresource(
+            ID3D11Resource*                   pDstResource,
+            UINT                              DstSubresource,
+      const D3D11_BOX*                        pDstBox,
+      const void*                             pSrcData,
+            UINT                              SrcRowPitch,
+            UINT                              SrcDepthPitch);
+
+    void STDMETHODCALLTYPE UpdateSubresource1(
+            ID3D11Resource*                   pDstResource,
+            UINT                              DstSubresource,
+      const D3D11_BOX*                        pDstBox,
+      const void*                             pSrcData,
+            UINT                              SrcRowPitch,
+            UINT                              SrcDepthPitch,
+            UINT                              CopyFlags);
+
     void STDMETHODCALLTYPE SwapDeviceContextState(
            ID3DDeviceContextState*           pState,
            ID3DDeviceContextState**          ppPreviousState);
@@ -103,7 +120,14 @@ namespace dxvk {
             ID3D11Resource*               pResource,
             UINT                          Subresource,
             D3D11DeferredContextMapEntry* pMapEntry);
-    
+
+    void UpdateMappedBuffer(
+            D3D11Buffer*                  pDstBuffer,
+            UINT                          Offset,
+            UINT                          Length,
+      const void*                         pSrcData,
+            UINT                          CopyFlags);
+
     void FinalizeQueries();
 
     Com<D3D11CommandList> CreateCommandList();

src/d3d11/d3d11_context_imm.cpp

@@ -288,11 +288,8 @@ namespace dxvk {
     const void*                             pSrcData,
           UINT                              SrcRowPitch,
           UINT                              SrcDepthPitch) {
-    FlushImplicit(FALSE);
+    UpdateResource<D3D11ImmediateContext>(this, pDstResource,
-
+      DstSubresource, pDstBox, pSrcData, SrcRowPitch, SrcDepthPitch, 0);
-    D3D11DeviceContext::UpdateSubresource(
-      pDstResource, DstSubresource, pDstBox,
-      pSrcData, SrcRowPitch, SrcDepthPitch);
   }
 
   
@@ -304,12 +301,8 @@ namespace dxvk {
           UINT                              SrcRowPitch,
           UINT                              SrcDepthPitch,
           UINT                              CopyFlags) {
-    FlushImplicit(FALSE);
+    UpdateResource<D3D11ImmediateContext>(this, pDstResource,
-
+      DstSubresource, pDstBox, pSrcData, SrcRowPitch, SrcDepthPitch, CopyFlags);
-    D3D11DeviceContext::UpdateSubresource1(
-      pDstResource, DstSubresource, pDstBox,
-      pSrcData, SrcRowPitch, SrcDepthPitch,
-      CopyFlags);
   }
   
   
@@ -497,6 +490,31 @@ namespace dxvk {
   }
   
   
+  void D3D11ImmediateContext::UpdateMappedBuffer(
+          D3D11Buffer*                  pDstBuffer,
+          UINT                          Offset,
+          UINT                          Length,
+    const void*                         pSrcData,
+          UINT                          CopyFlags) {
+    DxvkBufferSliceHandle slice;
+
+    if (likely(CopyFlags != D3D11_COPY_NO_OVERWRITE)) {
+      slice = pDstBuffer->DiscardSlice();
+
+      EmitCs([
+        cBuffer      = pDstBuffer->GetBuffer(),
+        cBufferSlice = slice
+      ] (DxvkContext* ctx) {
+        ctx->invalidateBuffer(cBuffer, cBufferSlice);
+      });
+    } else {
+      slice = pDstBuffer->GetMappedSlice();
+    }
+
+    std::memcpy(reinterpret_cast<char*>(slice.mapPtr) + Offset, pSrcData, Length);
+  }
+
+
   void STDMETHODCALLTYPE D3D11ImmediateContext::SwapDeviceContextState(
           ID3DDeviceContextState*           pState,
           ID3DDeviceContextState**          ppPreviousState) {

src/d3d11/d3d11_context_imm.h

@@ -16,6 +16,7 @@ namespace dxvk {
   class D3D11ImmediateContext : public D3D11DeviceContext {
     friend class D3D11SwapChain;
     friend class D3D11VideoContext;
+    friend class D3D11DeviceContext;
   public:
     
     D3D11ImmediateContext(
@@ -144,6 +145,13 @@ namespace dxvk {
             D3D11CommonTexture*         pResource,
             UINT                        Subresource);
     
+    void UpdateMappedBuffer(
+            D3D11Buffer*                  pDstBuffer,
+            UINT                          Offset,
+            UINT                          Length,
+      const void*                         pSrcData,
+            UINT                          CopyFlags);
+
     void SynchronizeDevice();
     
     bool WaitForResource(