[d3d11] Update mapped buffers of staging textures immediately

Improves performance in Lords of the Fallen and The Surge.
Closes #1049.

Co-authored-by: Robin <robin.kertels@outlook.com>

src/d3d11/d3d11_context.cpp

@@ -554,6 +554,9 @@ namespace dxvk {
             cExtent);
         }
       });
+
+      if (dstTextureInfo->CanUpdateMappedBufferEarly())
+        UpdateMappedBuffer(dstTextureInfo, dstSubresource);
     }
   }
   
@@ -604,8 +607,11 @@ namespace dxvk {
           cSrcBuffer.length());
       });
     } else {
-      const Rc<DxvkImage> dstImage = GetCommonTexture(pDstResource)->GetImage();
-      const Rc<DxvkImage> srcImage = GetCommonTexture(pSrcResource)->GetImage();
+      auto dstTexture = GetCommonTexture(pDstResource);
+      auto srcTexture = GetCommonTexture(pSrcResource);
+
+      const Rc<DxvkImage> dstImage = dstTexture->GetImage();
+      const Rc<DxvkImage> srcImage = srcTexture->GetImage();
 
       const DxvkFormatInfo* dstFormatInfo = imageFormatInfo(dstImage->info().format);
       const DxvkFormatInfo* srcFormatInfo = imageFormatInfo(srcImage->info().format);
@@ -652,6 +658,11 @@ namespace dxvk {
             cSrcImage, cSrcLayers, VkOffset3D { 0, 0, 0 },
             cExtent);
         });
+
+        if (dstTexture->CanUpdateMappedBufferEarly()) {
+          for (uint32_t j = 0; j < dstImage->info().numLayers; j++)
+            UpdateMappedBuffer(dstTexture, { dstLayers.aspectMask, i, j });
+        }
       }
     }
   }
@@ -1229,6 +1240,9 @@ namespace dxvk {
             cPackedFormat);
         }
       });
+
+      if (textureInfo->CanUpdateMappedBufferEarly())
+        UpdateMappedBuffer(textureInfo, subresource);
     }
   }
   
@@ -3667,6 +3681,40 @@ namespace dxvk {
         ctrSlotId + i, ~0u);
     }
   }
+
+
+  void D3D11DeviceContext::UpdateMappedBuffer(
+    const D3D11CommonTexture*               pTexture,
+          VkImageSubresource                Subresource) {
+    Rc<DxvkImage>  mappedImage  = pTexture->GetImage();
+    Rc<DxvkBuffer> mappedBuffer = pTexture->GetMappedBuffer();
+
+    VkFormat packedFormat = m_parent->LookupPackedFormat(
+      pTexture->Desc()->Format, pTexture->GetFormatMode()).Format;
+    
+    VkExtent3D levelExtent = mappedImage->mipLevelExtent(Subresource.mipLevel);
+    
+    EmitCs([
+      cImageBuffer  = std::move(mappedBuffer),
+      cImage        = std::move(mappedImage),
+      cSubresources = vk::makeSubresourceLayers(Subresource),
+      cLevelExtent  = levelExtent,
+      cPackedFormat = packedFormat
+    ] (DxvkContext* ctx) {
+      if (cSubresources.aspectMask != (VK_IMAGE_ASPECT_DEPTH_BIT | VK_IMAGE_ASPECT_STENCIL_BIT)) {
+        ctx->copyImageToBuffer(
+          cImageBuffer, 0, VkExtent2D { 0u, 0u },
+          cImage, cSubresources, VkOffset3D { 0, 0, 0 },
+          cLevelExtent);
+      } else {
+        ctx->copyDepthStencilImageToPackedBuffer(
+          cImageBuffer, 0, cImage, cSubresources,
+          VkOffset2D { 0, 0 },
+          VkExtent2D { cLevelExtent.width, cLevelExtent.height },
+          cPackedFormat);
+      }
+    });
+  }
   
   
   bool D3D11DeviceContext::ValidateRenderTargets(

src/d3d11/d3d11_context.h

@@ -803,6 +803,10 @@ namespace dxvk {
             DxbcProgramType                   Stage,
             D3D11UnorderedAccessBindings&     Bindings);
     
+    void UpdateMappedBuffer(
+      const D3D11CommonTexture*               pTexture,
+            VkImageSubresource                Subresource);
+    
     bool ValidateRenderTargets(
             UINT                              NumViews,
             ID3D11RenderTargetView* const*    ppRenderTargetViews,

src/d3d11/d3d11_context_imm.cpp

@@ -383,32 +383,16 @@ namespace dxvk {
         // When using any map mode which requires the image contents
         // to be preserved, and if the GPU has write access to the
         // image, copy the current image contents into the buffer.
-        if (pResource->Desc()->Usage == D3D11_USAGE_STAGING) {
-          auto subresourceLayers = vk::makeSubresourceLayers(subresource);
-          
-          EmitCs([
-            cImageBuffer  = mappedBuffer,
-            cImage        = mappedImage,
-            cSubresources = subresourceLayers,
-            cLevelExtent  = levelExtent,
-            cPackedFormat = packedFormat
-          ] (DxvkContext* ctx) {
-            if (cSubresources.aspectMask != (VK_IMAGE_ASPECT_DEPTH_BIT | VK_IMAGE_ASPECT_STENCIL_BIT)) {
-              ctx->copyImageToBuffer(
-                cImageBuffer, 0, VkExtent2D { 0u, 0u },
-                cImage, cSubresources, VkOffset3D { 0, 0, 0 },
-                cLevelExtent);
-            } else {
-              ctx->copyDepthStencilImageToPackedBuffer(
-                cImageBuffer, 0, cImage, cSubresources,
-                VkOffset2D { 0, 0 },
-                VkExtent2D { cLevelExtent.width, cLevelExtent.height },
-                cPackedFormat);
-            }
-          });
+        if (pResource->Desc()->Usage == D3D11_USAGE_STAGING
+         && !pResource->CanUpdateMappedBufferEarly()) {
+          UpdateMappedBuffer(pResource, subresource);
+          MapFlags &= ~D3D11_MAP_FLAG_DO_NOT_WAIT;
         }
         
-        WaitForResource(mappedBuffer, 0);
+        // Wait for mapped buffer to become available
+        if (!WaitForResource(mappedBuffer, MapFlags))
+          return DXGI_ERROR_WAS_STILL_DRAWING;
+        
         physSlice = mappedBuffer->getSliceHandle();
       }
       

src/d3d11/d3d11_texture.h

@@ -130,6 +130,22 @@ namespace dxvk {
     void ClearMappedSubresource() {
       m_mappedSubresource = VkImageSubresource { };
     }
+
+    /**
+     * \brief Checks whether we can update the mapped buffer early
+     * 
+     * For images which are mapped through a buffer and that are
+     * only used for transfer operations, we can update the mapped
+     * buffer right after performing those transfers to avoid stalls.
+     * \returns \c true if the mapped buffer can be updated early
+     */
+    bool CanUpdateMappedBufferEarly() const {
+      return m_mapMode == D3D11_COMMON_TEXTURE_MAP_MODE_BUFFER
+          && (m_desc.BindFlags & ~D3D11_BIND_SHADER_RESOURCE) == 0
+          && m_desc.Usage     == D3D11_USAGE_STAGING
+          && m_desc.MipLevels == 1
+          && m_desc.ArraySize == 1;
+    }
     
     /**
      * \brief Computes subresource from the subresource index