[d3d11] Sub-allocate from larger update buffers for UpdateSubresources

Reduces the allocation overhead when applications frequently
call UpdateSubresources to update small buffers and textures.

src/d3d11/d3d11_context.cpp

@@ -569,15 +569,18 @@ namespace dxvk {
         std::memcpy(mappedSr.pData, pSrcData, size);
         Unmap(pDstResource, 0);
       } else {
+        DxvkDataSlice dataSlice = AllocUpdateBufferSlice(size);
+        std::memcpy(dataSlice.ptr(), pSrcData, size);
+        
         EmitCs([
-          cDataBuffer   = Rc<DxvkDataBuffer>(new DxvkDataBuffer(pSrcData, size)),
+          cDataBuffer   = std::move(dataSlice),
           cBufferSlice  = bufferSlice.subSlice(offset, size)
         ] (DxvkContext* ctx) {
           ctx->updateBuffer(
             cBufferSlice.buffer(),
             cBufferSlice.offset(),
             cBufferSlice.length(),
-            cDataBuffer->data());
+            cDataBuffer.ptr());
         });
       }
     } else {
@@ -620,10 +623,10 @@ namespace dxvk {
       const VkDeviceSize bytesPerLayer = regionExtent.height * bytesPerRow;
       const VkDeviceSize bytesTotal    = regionExtent.depth  * bytesPerLayer;
       
-      Rc<DxvkDataBuffer> imageDataBuffer = new DxvkDataBuffer(bytesTotal);
+      DxvkDataSlice imageDataBuffer = AllocUpdateBufferSlice(bytesTotal);
       
       util::packImageData(
-        reinterpret_cast<char*>(imageDataBuffer->data()),
+        reinterpret_cast<char*>(imageDataBuffer.ptr()),
         reinterpret_cast<const char*>(pSrcData),
         regionExtent, formatInfo->elementSize,
         SrcRowPitch, SrcDepthPitch);
@@ -638,7 +641,7 @@ namespace dxvk {
         cSrcBytesPerLayer = bytesPerLayer
       ] (DxvkContext* ctx) {
         ctx->updateImage(cDstImage, cDstLayers,
-          cDstOffset, cDstExtent, cSrcData->data(),
+          cDstOffset, cDstExtent, cSrcData.ptr(),
           cSrcBytesPerRow, cSrcBytesPerLayer);
       });
     }
@@ -2113,4 +2116,27 @@ namespace dxvk {
     return m_device->createSampler(info);
   }
   
+  
+  DxvkDataSlice D3D11DeviceContext::AllocUpdateBufferSlice(size_t Size) {
+    constexpr size_t UpdateBufferSize = 4 * 1024 * 1024;
+    
+    if (Size >= UpdateBufferSize) {
+      Rc<DxvkDataBuffer> buffer = new DxvkDataBuffer(Size);
+      return buffer->alloc(Size);
+    } else {
+      if (m_updateBuffer == nullptr)
+        m_updateBuffer = new DxvkDataBuffer(Size);
+      
+      DxvkDataSlice slice = m_updateBuffer->alloc(Size);
+      
+      if (slice.ptr() == nullptr) {
+        m_updateBuffer = new DxvkDataBuffer(Size);
+        slice = m_updateBuffer->alloc(Size);
+      }
+      
+      return slice;
+    }
+  }
+    
+  
 }

src/d3d11/d3d11_context.h

@@ -520,6 +520,7 @@ namespace dxvk {
     Rc<DxvkDevice>              m_device;
     Rc<DxvkCsChunk>             m_csChunk;
     Rc<DxvkSampler>             m_defaultSampler;
+    Rc<DxvkDataBuffer>          m_updateBuffer;
     
     Com<D3D11BlendState>        m_defaultBlendState;
     Com<D3D11DepthStencilState> m_defaultDepthStencilState;
@@ -565,6 +566,8 @@ namespace dxvk {
     
     Rc<DxvkSampler> CreateDefaultSampler();
     
+    DxvkDataSlice AllocUpdateBufferSlice(size_t Size);
+    
     template<typename Cmd>
     void EmitCs(Cmd&& command) {
       if (!m_csChunk->push(command)) {
@@ -584,6 +587,7 @@ namespace dxvk {
     
     virtual void EmitCsChunk(Rc<DxvkCsChunk>&& chunk) = 0;
     
+    
   };
   
 }

src/dxvk/dxvk_data.cpp

@@ -7,16 +7,18 @@ namespace dxvk {
   DxvkDataBuffer:: DxvkDataBuffer() { }
   DxvkDataBuffer::~DxvkDataBuffer() { }
   
-  DxvkDataBuffer::DxvkDataBuffer(
+  DxvkDataBuffer::DxvkDataBuffer(size_t size) {
-          size_t  size) {
     m_data.resize(size);
   }
   
-  DxvkDataBuffer::DxvkDataBuffer(
+  
-    const void*   data,
+  DxvkDataSlice DxvkDataBuffer::alloc(size_t n) {
-          size_t  size) {
+    const size_t offset = m_offset;
-    m_data.resize(size);
+    
-    std::memcpy(m_data.data(), data, size);
+    if (offset + n <= m_data.size()) {
+      m_offset += align(n, CACHE_LINE_SIZE);
+      return DxvkDataSlice(this, offset, n);
+    } return DxvkDataSlice();
   }
   
 }

src/dxvk/dxvk_data.h

@@ -4,41 +4,78 @@
 
 namespace dxvk {
   
+  class DxvkDataSlice;
+  
   /**
    * \brief Data buffer
    * 
-   * Stores immutable data. Used for temporary
+   * Provides a fixed-size buffer with a linear memory
-   * copies of data that can be transferred to
+   * allocator for arbitrary data. Can be used to copy
-   * or from DXVK resources.
+   * data to or from resources. Note that allocations
+   * will be aligned to a cache line boundary.
    */
   class DxvkDataBuffer : public RcObject {
-    
+    friend class DxvkDataSlice;
   public:
     
     DxvkDataBuffer();
-    DxvkDataBuffer(
+    DxvkDataBuffer(size_t size);
-            size_t  size);
-    DxvkDataBuffer(
-      const void*   data,
-            size_t  size);
     ~DxvkDataBuffer();
     
-    size_t size() const {
+    /**
-      return m_data.size();
+     * \brief Allocates a slice
-    }
+     * 
-    
+     * If the desired slice length is larger than the
-    void* data() {
+     * number of bytes left in the buffer, this will
-      return m_data.data();
+     * fail and the returned slice points to \c nullptr.
-    }
+     * \param [in] n Number of bytes to allocate
-    
+     * \returns The slice, or an empty slice on failure
-    const void* data() const {
+     */
-      return m_data.data();
+    DxvkDataSlice alloc(size_t n);
-    }
     
   private:
     
     std::vector<char> m_data;
+    size_t            m_offset = 0;
     
   };
   
+  
+  /**
+   * \brief Data buffer slice
+   * 
+   * A slice of a \ref DxvkDataBuffer which stores
+   * a strong reference to the backing buffer object.
+   */
+  class DxvkDataSlice {
+    
+  public:
+    
+    DxvkDataSlice() { }
+    DxvkDataSlice(
+      const Rc<DxvkDataBuffer>& buffer,
+            size_t              offset,
+            size_t              length)
+    : m_buffer(buffer),
+      m_offset(offset),
+      m_length(length) { }
+    
+    void* ptr() const {
+      return m_buffer != nullptr
+        ? m_buffer->m_data.data() + m_offset
+        : nullptr;
+    }
+    
+    size_t offset() const { return m_offset; }
+    size_t length() const { return m_length; }
+    
+  private:
+    
+    Rc<DxvkDataBuffer> m_buffer;
+    size_t             m_offset = 0;
+    size_t             m_length = 0;
+    
+  };
+  
+  
 }

src/util/util_math.h

@@ -2,6 +2,8 @@
 
 namespace dxvk {
   
+  constexpr size_t CACHE_LINE_SIZE = 64;
+  
   template<typename T>
   T clamp(T n, T lo, T hi) {
     if (n < lo) return lo;