[dxvk] Initial work on instance data fetch rates

Adds a pipeline flag to indicate that instanced
draw calls need to be emulated.

src/d3d11/d3d11_device.cpp

@@ -731,17 +731,10 @@ namespace dxvk {
         // set by D3D11DeviceContext::IASetVertexBuffers.
         DxvkVertexBinding binding;
         binding.binding   = pInputElementDescs[i].InputSlot;
-        binding.inputRate = VK_VERTEX_INPUT_RATE_VERTEX;
+        binding.fetchRate = pInputElementDescs[i].InstanceDataStepRate;
-        
+        binding.inputRate = pInputElementDescs[i].InputSlotClass == D3D11_INPUT_PER_INSTANCE_DATA
-        if (pInputElementDescs[i].InputSlotClass == D3D11_INPUT_PER_INSTANCE_DATA) {
+          ? VK_VERTEX_INPUT_RATE_INSTANCE
-          binding.inputRate = VK_VERTEX_INPUT_RATE_INSTANCE;
+          : VK_VERTEX_INPUT_RATE_VERTEX;
-          
-          if (pInputElementDescs[i].InstanceDataStepRate != 1) {
-            Logger::warn(str::format(
-              "D3D11Device: Unsupported instance data step rate: ",
-              pInputElementDescs[i].InstanceDataStepRate));
-          }
-        }
         
         // Check if the binding was already defined. If so, the
         // parameters must be identical (namely, the input rate).

src/dxbc/dxbc_compiler.cpp

@@ -4930,6 +4930,8 @@ namespace dxvk {
     m_module.setDebugName(typeId, "s_push_constant");
     m_module.setDebugMemberName(typeId, PerVertex_Position, "instance_id");
     
+    // There's only ever going to be one single push constant
+    // block per shader, so we'll declare the variable here
     uint32_t ptrTypeId = m_module.defPointerType(
       typeId, spv::StorageClassPushConstant);
     

src/dxvk/dxvk_cmdlist.h

@@ -333,6 +333,17 @@ namespace dxvk {
     }
     
     
+    void cmdPushConstants(
+            VkPipelineLayout        layout,
+            VkShaderStageFlags      stageFlags,
+            uint32_t                offset,
+            uint32_t                size,
+      const void*                   pValues) {
+      m_vkd->vkCmdPushConstants(m_buffer,
+        layout, stageFlags, offset, size, pValues);
+    }
+    
+    
     void cmdResolveImage(
             VkImage                 srcImage,
             VkImageLayout           srcImageLayout,

src/dxvk/dxvk_constant_state.h

@@ -136,6 +136,7 @@ namespace dxvk {
    */
   struct DxvkVertexBinding {
     uint32_t          binding;
+    uint32_t          fetchRate;
     VkVertexInputRate inputRate;
   };
   

src/dxvk/dxvk_context.cpp

@@ -572,9 +572,16 @@ namespace dxvk {
     this->commitGraphicsState();
     
     if (m_gpActivePipeline != VK_NULL_HANDLE) {
-      m_cmd->cmdDraw(
+      if (!m_flags.test(DxvkContextFlag::GpEmulateInstanceFetchRate)) {
-        vertexCount, instanceCount,
+        m_cmd->cmdDraw(
-        firstVertex, firstInstance);
+          vertexCount, instanceCount,
+          firstVertex, firstInstance);
+      } else {
+        static bool errorShown = false;
+        
+        if (!std::exchange(errorShown, true))
+          Logger::warn("Dxvk: GpEmulateInstanceFetchRate not supported for direct draws");
+      }
     }
   }
   
@@ -588,10 +595,17 @@ namespace dxvk {
     if (m_gpActivePipeline != VK_NULL_HANDLE) {
       auto physicalSlice = buffer.physicalSlice();
       
-      m_cmd->cmdDrawIndirect(
+      if (!m_flags.test(DxvkContextFlag::GpEmulateInstanceFetchRate)) {
-        physicalSlice.handle(),
+        m_cmd->cmdDrawIndirect(
-        physicalSlice.offset(),
+          physicalSlice.handle(),
-        count, stride);
+          physicalSlice.offset(),
+          count, stride);
+      } else {
+        static bool errorShown = false;
+        
+        if (!std::exchange(errorShown, true))
+          Logger::warn("Dxvk: GpEmulateInstanceFetchRate not supported for indirect draws");
+      }
     }
   }
   
@@ -1042,6 +1056,8 @@ namespace dxvk {
     m_flags.set(
       DxvkContextFlag::GpDirtyPipelineState,
       DxvkContextFlag::GpDirtyVertexBuffers);
+    m_flags.clr(
+      DxvkContextFlag::GpEmulateInstanceFetchRate);
     
     for (uint32_t i = 0; i < attributeCount; i++) {
       m_state.gp.state.ilAttributes[i].location = attributes[i].location;
@@ -1056,6 +1072,10 @@ namespace dxvk {
     for (uint32_t i = 0; i < bindingCount; i++) {
       m_state.gp.state.ilBindings[i].binding    = bindings[i].binding;
       m_state.gp.state.ilBindings[i].inputRate  = bindings[i].inputRate;
+      m_state.vi.vertexFetchRates[bindings[i].binding] = bindings[i].fetchRate;
+      
+      if (bindings[i].inputRate == VK_VERTEX_INPUT_RATE_INSTANCE && bindings[i].fetchRate != 1)
+        m_flags.set(DxvkContextFlag::GpEmulateInstanceFetchRate);
     }
     
     for (uint32_t i = bindingCount; i < m_state.gp.state.ilBindingCount; i++)

src/dxvk/dxvk_context.h

@@ -558,12 +558,12 @@ namespace dxvk {
     
     void updateShaderResources(
             VkPipelineBindPoint     bindPoint,
-      const Rc<DxvkPipelineLayout>&  layout);
+      const Rc<DxvkPipelineLayout>& layout);
     
     void updateShaderDescriptors(
             VkPipelineBindPoint     bindPoint,
       const DxvkBindingState&       bindingState,
-      const Rc<DxvkPipelineLayout>&  layout);
+      const Rc<DxvkPipelineLayout>& layout);
     
     void updateViewports();
     void updateBlendConstants();

src/dxvk/dxvk_context_state.h

@@ -21,16 +21,17 @@ namespace dxvk {
    * be updated.
    */
   enum class DxvkContextFlag : uint64_t  {
-    GpRenderPassBound,      ///< Render pass is currently bound
+    GpRenderPassBound,          ///< Render pass is currently bound
-    GpDirtyPipeline,        ///< Graphics pipeline binding is out of date
+    GpDirtyPipeline,            ///< Graphics pipeline binding is out of date
-    GpDirtyPipelineState,   ///< Graphics pipeline needs to be recompiled
+    GpDirtyPipelineState,       ///< Graphics pipeline needs to be recompiled
-    GpDirtyResources,       ///< Graphics pipeline resource bindings are out of date
+    GpDirtyResources,           ///< Graphics pipeline resource bindings are out of date
-    GpDirtyVertexBuffers,   ///< Vertex buffer bindings are out of date
+    GpDirtyVertexBuffers,       ///< Vertex buffer bindings are out of date
-    GpDirtyIndexBuffer,     ///< Index buffer binding are out of date
+    GpDirtyIndexBuffer,         ///< Index buffer binding are out of date
+    GpEmulateInstanceFetchRate, ///< The current input layout uses fetch rates != 1
     
-    CpDirtyPipeline,        ///< Compute pipeline binding are out of date
+    CpDirtyPipeline,            ///< Compute pipeline binding are out of date
-    CpDirtyPipelineState,   ///< Compute pipeline needs to be recompiled
+    CpDirtyPipelineState,       ///< Compute pipeline needs to be recompiled
-    CpDirtyResources,       ///< Compute pipeline resource bindings are out of date
+    CpDirtyResources,           ///< Compute pipeline resource bindings are out of date
   };
   
   using DxvkContextFlags = Flags<DxvkContextFlag>;
@@ -42,23 +43,25 @@ namespace dxvk {
     uint32_t        bindingMask = 0;
     
     std::array<DxvkBufferSlice,
-      DxvkLimits::MaxNumVertexBindings> vertexBuffers;
+      DxvkLimits::MaxNumVertexBindings> vertexBuffers = { };
     std::array<uint32_t,
-      DxvkLimits::MaxNumVertexBindings> vertexStrides;
+      DxvkLimits::MaxNumVertexBindings> vertexStrides = { };
+    std::array<uint32_t,
+      DxvkLimits::MaxNumVertexBindings> vertexFetchRates = { };
   };
   
   
   struct DxvkViewportState {
-    std::array<VkViewport, DxvkLimits::MaxNumViewports> viewports;
+    std::array<VkViewport, DxvkLimits::MaxNumViewports> viewports    = { };
-    std::array<VkRect2D,   DxvkLimits::MaxNumViewports> scissorRects;
+    std::array<VkRect2D,   DxvkLimits::MaxNumViewports> scissorRects = { };
   };
   
   
   struct DxvkOutputMergerState {
     Rc<DxvkFramebuffer> framebuffer;
     
-    DxvkBlendConstants  blendConstants;
+    DxvkBlendConstants  blendConstants    = { 0.0f, 0.0f, 0.0f, 0.0f };
-    uint32_t            stencilReference;
+    uint32_t            stencilReference  = 0;
   };
   
   
@@ -73,9 +76,9 @@ namespace dxvk {
     DxvkShaderStage  tes;
     DxvkShaderStage  gs;
     DxvkShaderStage  fs;
-    
+
     DxvkGraphicsPipelineStateInfo state;
-    Rc<DxvkGraphicsPipeline> pipeline;
+    Rc<DxvkGraphicsPipeline>      pipeline;
   };
   
   

src/dxvk/dxvk_limits.h

@@ -14,6 +14,7 @@ namespace dxvk {
     MaxNumActiveBindings        =  128,
     MaxNumQueuedCommandBuffers  =    8,
     MaxVertexBindingStride      = 2048,
+    MaxPushConstantSize         =  128,
   };
   
 }