[d3d11] Optimize UAV binding

- UpdateBuffer is faster than ClearBuffer for small updates.
- We shouldn't dispatch *two* CS commands for each UAV, one is enough.

src/d3d11/d3d11_context.cpp

@@ -2181,14 +2181,9 @@ namespace dxvk {
       DxbcProgramType::ComputeShader,
       m_state.cs.unorderedAccessViews,
       StartSlot, NumUAVs,
-      ppUnorderedAccessViews);
+      ppUnorderedAccessViews,
-    
-    if (pUAVInitialCounts != nullptr) {
-      InitUnorderedAccessViewCounters(
-        NumUAVs, ppUnorderedAccessViews,
       pUAVInitialCounts);
   }
-  }
   
   
   void STDMETHODCALLTYPE D3D11DeviceContext::CSGetShader(
@@ -2289,17 +2284,12 @@ namespace dxvk {
       // UAVs are made available to all shader stages in
       // the graphics pipeline even though this code may
       // suggest that they are limited to the pixel shader.
-      // This behaviour is only required for FL_11_1.
       SetUnorderedAccessViews(
         DxbcProgramType::PixelShader,
         m_state.ps.unorderedAccessViews,
         UAVStartSlot, NumUAVs,
-        ppUnorderedAccessViews);
+        ppUnorderedAccessViews,
-      
+        pUAVInitialCounts);
-      if (pUAVInitialCounts != nullptr) {
-        InitUnorderedAccessViewCounters(NumUAVs,
-          ppUnorderedAccessViews, pUAVInitialCounts);
-      }
     }
 
     BindFramebuffer(spillOnBind);
@@ -2882,14 +2872,24 @@ namespace dxvk {
   void D3D11DeviceContext::BindUnorderedAccessView(
           UINT                              UavSlot,
           UINT                              CtrSlot,
-          D3D11UnorderedAccessView*         pUav) {
+          D3D11UnorderedAccessView*         pUav,
+          UINT                              Counter) {
     EmitCs([
       cUavSlotId    = UavSlot,
       cCtrSlotId    = CtrSlot,
       cImageView    = pUav != nullptr ? pUav->GetImageView()    : nullptr,
       cBufferView   = pUav != nullptr ? pUav->GetBufferView()   : nullptr,
-      cCounterSlice = pUav != nullptr ? pUav->GetCounterSlice() : DxvkBufferSlice()
+      cCounterSlice = pUav != nullptr ? pUav->GetCounterSlice() : DxvkBufferSlice(),
+      cCounterValue = Counter
     ] (DxvkContext* ctx) {
+      if (cCounterSlice.defined() && cCounterValue != ~0u) {
+        ctx->updateBuffer(
+          cCounterSlice.buffer(),
+          cCounterSlice.offset(),
+          sizeof(uint32_t),
+          &cCounterValue);
+      }
+
       ctx->bindResourceView   (cUavSlotId, cImageView, cBufferView);
       ctx->bindResourceBuffer (cCtrSlotId, cCounterSlice);
     });
@@ -2989,7 +2989,8 @@ namespace dxvk {
           D3D11UnorderedAccessBindings&     Bindings,
           UINT                              StartSlot,
           UINT                              NumUAVs,
-          ID3D11UnorderedAccessView* const* ppUnorderedAccessViews) {
+          ID3D11UnorderedAccessView* const* ppUnorderedAccessViews,
+    const UINT*                             pUAVInitialCounts) {
     const uint32_t uavSlotId = computeResourceSlotId(
       ShaderStage, DxbcBindingType::UnorderedAccessView,
       StartSlot);
@@ -3003,7 +3004,8 @@ namespace dxvk {
       
       if (Bindings[StartSlot + i] != uav) {
         Bindings[StartSlot + i] = uav;
-        BindUnorderedAccessView(uavSlotId + i, ctrSlotId + i, uav);
+        BindUnorderedAccessView(uavSlotId + i, ctrSlotId + i, uav,
+          pUAVInitialCounts ? pUAVInitialCounts[i] : ~0u);
       }
     }
   }
@@ -3028,31 +3030,6 @@ namespace dxvk {
   }
   
   
-  void D3D11DeviceContext::InitUnorderedAccessViewCounters(
-          UINT                              NumUAVs,
-          ID3D11UnorderedAccessView* const* ppUnorderedAccessViews,
-    const UINT*                             pUAVInitialCounts) {
-    for (uint32_t i = 0; i < NumUAVs; i++) {
-      auto uav = static_cast<D3D11UnorderedAccessView*>(ppUnorderedAccessViews[i]);
-      
-      if (uav != nullptr) {
-        const DxvkBufferSlice counterSlice = uav->GetCounterSlice();
-        const D3D11UavCounter counterValue = { pUAVInitialCounts[i] };
-        
-        if (counterSlice.defined() && counterValue.atomicCtr != 0xFFFFFFFFu) {
-          EmitCs([counterSlice, counterValue] (DxvkContext* ctx) {
-            ctx->clearBuffer(
-              counterSlice.buffer(),
-              counterSlice.offset(),
-              counterSlice.length(),
-              counterValue.atomicCtr);
-          });
-        }
-      }
-    }
-  }
-  
-  
   void D3D11DeviceContext::GetConstantBuffers(
     const D3D11ConstantBufferBindings&      Bindings,
           UINT                              StartSlot,
@@ -3175,7 +3152,7 @@ namespace dxvk {
     for (uint32_t i = 0; i < Bindings.size(); i++) {
       BindUnorderedAccessView(
         uavSlotId + i, ctrSlotId + i,
-        Bindings[i].ptr());
+        Bindings[i].ptr(), ~0u);
     }
   }
   

src/d3d11/d3d11_context.h

@@ -701,7 +701,8 @@ namespace dxvk {
     void BindUnorderedAccessView(
             UINT                              UavSlot,
             UINT                              CtrSlot,
-            D3D11UnorderedAccessView*         pUav);
+            D3D11UnorderedAccessView*         pUav,
+            UINT                              Counter);
     
     void DiscardBuffer(
             D3D11Buffer*                      pBuffer);
@@ -734,18 +735,14 @@ namespace dxvk {
             D3D11UnorderedAccessBindings&     Bindings,
             UINT                              StartSlot,
             UINT                              NumUAVs,
-            ID3D11UnorderedAccessView* const* ppUnorderedAccessViews);
+            ID3D11UnorderedAccessView* const* ppUnorderedAccessViews,
+      const UINT*                             pUAVInitialCounts);
     
     void SetRenderTargets(
             UINT                              NumViews,
             ID3D11RenderTargetView* const*    ppRenderTargetViews,
             ID3D11DepthStencilView*           pDepthStencilView);
     
-    void InitUnorderedAccessViewCounters(
-            UINT                              NumUAVs,
-            ID3D11UnorderedAccessView* const* ppUnorderedAccessViews,
-      const UINT*                             pUAVInitialCounts);
-    
     void GetConstantBuffers(
       const D3D11ConstantBufferBindings&      Bindings,
             UINT                              StartSlot,

src/d3d11/d3d11_device.cpp

@@ -1404,7 +1404,7 @@ namespace dxvk {
     const auto& devInfo = m_dxvkAdapter->deviceProperties();
 
     VkDeviceSize uavCounterSliceLength = align<VkDeviceSize>(
-      sizeof(D3D11UavCounter), devInfo.limits.minStorageBufferOffsetAlignment);
+      sizeof(uint32_t), devInfo.limits.minStorageBufferOffsetAlignment);
 
     DxvkBufferCreateInfo uavCounterInfo;
     uavCounterInfo.size   = 4096 * uavCounterSliceLength;

src/d3d11/d3d11_view_uav.h

@@ -8,17 +8,6 @@ namespace dxvk {
   
   class D3D11Device;
   
-  /**
-   * \brief UAV counter structure
-   * 
-   * Data structure passed to shaders that use
-   * append/consume buffer functionality.
-   */
-  struct D3D11UavCounter {
-    uint32_t atomicCtr;
-  };
-
-
   /**
    * \brief Unordered access view
    *