[d3d11] Use multiDrawIndirect for subsequent indirect draw calls

Significantly improves performance in AC:Odyssey when CPU bound.
Only has an effect when no state changes between draw calls, and
when the draw parameter buffer is tightly packed.

src/d3d11/d3d11_cmd.h

@@ -0,0 +1,21 @@
+#pragma once
+
+#include "d3d11_include.h"
+
+namespace dxvk {
+
+  enum class D3D11CmdType {
+    DrawIndirect,
+    DrawIndirectIndexed,
+  };
+
+  struct D3D11CmdData {
+    D3D11CmdType        type;
+  };
+
+  struct D3D11CmdDrawIndirectData : public D3D11CmdData {
+    uint32_t            offset;
+    uint32_t            count;
+  };
+
+}

src/d3d11/d3d11_context.cpp

@@ -18,7 +18,8 @@ namespace dxvk {
     m_multithread(this, false),
     m_device    (Device),
     m_csFlags   (CsFlags),
-    m_csChunk   (AllocCsChunk()) {
+    m_csChunk   (AllocCsChunk()),
+    m_cmdData   (nullptr) {
     // Create default state objects. We won't ever return them
     // to the application, but we'll use them to apply state.
     Com<ID3D11BlendState>         defaultBlendState;
@@ -1401,10 +1402,27 @@ namespace dxvk {
     
     SetDrawBuffer(pBufferForArgs);
     
-    EmitCs([cOffset = AlignedByteOffsetForArgs]
-    (DxvkContext* ctx) {
-      ctx->drawIndexedIndirect(cOffset, 1, 0);
-    });
+    // If possible, batch up multiple indirect draw calls of
+    // the same type into one single multiDrawIndirect call
+    constexpr VkDeviceSize stride = sizeof(VkDrawIndexedIndirectCommand);
+    auto cmdData = static_cast<D3D11CmdDrawIndirectData*>(m_cmdData);
+
+    bool useMultiDraw = cmdData && cmdData->type == D3D11CmdType::DrawIndirectIndexed
+      && cmdData->offset + cmdData->count * stride == AlignedByteOffsetForArgs
+      && m_device->features().core.features.multiDrawIndirect;
+    
+    if (useMultiDraw) {
+      cmdData->count += 1;
+    } else {
+      cmdData = EmitCsCmd<D3D11CmdDrawIndirectData>(
+        [] (DxvkContext* ctx, const D3D11CmdDrawIndirectData* data) {
+          ctx->drawIndexedIndirect(data->offset, data->count, stride);
+        });
+      
+      cmdData->type   = D3D11CmdType::DrawIndirectIndexed;
+      cmdData->offset = AlignedByteOffsetForArgs;
+      cmdData->count  = 1;
+    }
   }
   
   
@@ -1414,11 +1432,28 @@ namespace dxvk {
     D3D10DeviceLock lock = LockContext();
     
     SetDrawBuffer(pBufferForArgs);
+
+    // If possible, batch up multiple indirect draw calls of
+    // the same type into one single multiDrawIndirect call
+    constexpr VkDeviceSize stride = sizeof(VkDrawIndirectCommand);
+    auto cmdData = static_cast<D3D11CmdDrawIndirectData*>(m_cmdData);
+
+    bool useMultiDraw = cmdData && cmdData->type == D3D11CmdType::DrawIndirectIndexed
+      && cmdData->offset + cmdData->count * stride == AlignedByteOffsetForArgs
+      && m_device->features().core.features.multiDrawIndirect;
     
-    EmitCs([cOffset = AlignedByteOffsetForArgs]
-    (DxvkContext* ctx) {
-      ctx->drawIndirect(cOffset, 1, 0);
-    });
+    if (useMultiDraw) {
+      cmdData->count += 1;
+    } else {
+      cmdData = EmitCsCmd<D3D11CmdDrawIndirectData>(
+        [] (DxvkContext* ctx, const D3D11CmdDrawIndirectData* data) {
+          ctx->drawIndirect(data->offset, data->count, stride);
+        });
+      
+      cmdData->type   = D3D11CmdType::DrawIndirect;
+      cmdData->offset = AlignedByteOffsetForArgs;
+      cmdData->count  = 1;
+    }
   }
   
   

src/d3d11/d3d11_context.h

@@ -7,6 +7,7 @@
 #include "../d3d10/d3d10_multithread.h"
 
 #include "d3d11_annotation.h"
+#include "d3d11_cmd.h"
 #include "d3d11_context_state.h"
 #include "d3d11_device_child.h"
 #include "d3d11_texture.h"
@@ -662,6 +663,7 @@ namespace dxvk {
     Com<D3D11RasterizerState>   m_defaultRasterizerState;
     
     D3D11ContextState           m_state;
+    D3D11CmdData*               m_cmdData;
     
     void ApplyInputLayout();
     
@@ -814,6 +816,8 @@ namespace dxvk {
 
     template<typename Cmd>
     void EmitCs(Cmd&& command) {
+      m_cmdData = nullptr;
+
       if (!m_csChunk->push(command)) {
         EmitCsChunk(std::move(m_csChunk));
         
@@ -821,11 +825,29 @@ namespace dxvk {
         m_csChunk->push(command);
       }
     }
+
+    template<typename M, typename Cmd, typename... Args>
+    M* EmitCsCmd(Cmd&& command, Args&&... args) {
+      M* data = m_csChunk->pushCmd<M, Cmd, Args...>(
+        command, std::forward<Args>(args)...);
+
+      if (!data) {
+        EmitCsChunk(std::move(m_csChunk));
+        
+        m_csChunk = AllocCsChunk();
+        data = m_csChunk->pushCmd<M, Cmd, Args...>(
+          command, std::forward<Args>(args)...);
+      }
+
+      m_cmdData = data;
+      return data;
+    }
     
     void FlushCsChunk() {
       if (m_csChunk->commandCount() != 0) {
         EmitCsChunk(std::move(m_csChunk));
         m_csChunk = AllocCsChunk();
+        m_cmdData = nullptr;
       }
     }