From 09f507c284f1e71345152fc1846a13db6a5ede85 Mon Sep 17 00:00:00 2001
From: Philip Rebohle <philip.rebohle@tu-dortmund.de>
Date: Sat, 26 Oct 2019 15:46:40 +0200
Subject: [PATCH] [d3d11] Allocate xfb counter buffer per buffer

Same idea as with UAV counters, allows for more efficient updates.
---
 src/d3d11/d3d11_buffer.cpp | 26 +++++++++++++++++++++++---
 src/d3d11/d3d11_buffer.h   |  8 ++++++--
 src/d3d11/d3d11_device.cpp | 22 ----------------------
 src/d3d11/d3d11_device.h   |  4 ----
 4 files changed, 29 insertions(+), 31 deletions(-)

diff --git a/src/d3d11/d3d11_buffer.cpp b/src/d3d11/d3d11_buffer.cpp
index 5fe8621f6..2738345b9 100644
--- a/src/d3d11/d3d11_buffer.cpp
+++ b/src/d3d11/d3d11_buffer.cpp
@@ -76,13 +76,12 @@ namespace dxvk {
 
     // For Stream Output buffers we need a counter
     if (pDesc->BindFlags & D3D11_BIND_STREAM_OUTPUT)
-      m_soCounter = m_device->AllocXfbCounterSlice();
+      m_soCounter = CreateSoCounterBuffer();
   }
   
   
   D3D11Buffer::~D3D11Buffer() {
-    if (m_soCounter.defined())
-      m_device->FreeXfbCounterSlice(m_soCounter);
+
   }
   
   
@@ -255,6 +254,27 @@ namespace dxvk {
     
     return memoryFlags;
   }
+
+
+  Rc<DxvkBuffer> D3D11Buffer::CreateSoCounterBuffer() {
+    Rc<DxvkDevice> device = m_device->GetDXVKDevice();
+
+    DxvkBufferCreateInfo info;
+    info.size   = sizeof(D3D11SOCounter);
+    info.usage  = VK_BUFFER_USAGE_TRANSFER_DST_BIT
+                | VK_BUFFER_USAGE_TRANSFER_SRC_BIT
+                | VK_BUFFER_USAGE_INDIRECT_BUFFER_BIT
+                | VK_BUFFER_USAGE_TRANSFORM_FEEDBACK_COUNTER_BUFFER_BIT_EXT;
+    info.stages = VK_PIPELINE_STAGE_TRANSFER_BIT
+                | VK_PIPELINE_STAGE_DRAW_INDIRECT_BIT
+                | VK_PIPELINE_STAGE_TRANSFORM_FEEDBACK_BIT_EXT;
+    info.access = VK_ACCESS_TRANSFER_READ_BIT
+                | VK_ACCESS_TRANSFER_WRITE_BIT
+                | VK_ACCESS_INDIRECT_COMMAND_READ_BIT
+                | VK_ACCESS_TRANSFORM_FEEDBACK_COUNTER_READ_BIT_EXT
+                | VK_ACCESS_TRANSFORM_FEEDBACK_COUNTER_WRITE_BIT_EXT;
+    return device->createBuffer(info, VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT);
+  }
   
 
   D3D11Buffer* GetCommonBuffer(ID3D11Resource* pResource) {
diff --git a/src/d3d11/d3d11_buffer.h b/src/d3d11/d3d11_buffer.h
index 99d96cba6..f19e410bb 100644
--- a/src/d3d11/d3d11_buffer.h
+++ b/src/d3d11/d3d11_buffer.h
@@ -92,7 +92,9 @@ namespace dxvk {
     }
 
     DxvkBufferSlice GetSOCounter() {
-      return m_soCounter;
+      return m_soCounter != nullptr
+        ? DxvkBufferSlice(m_soCounter)
+        : DxvkBufferSlice();
     }
     
     DxvkBufferSliceHandle AllocSlice() {
@@ -127,7 +129,7 @@ namespace dxvk {
     const D3D11_BUFFER_DESC     m_desc;
     
     Rc<DxvkBuffer>              m_buffer;
-    DxvkBufferSlice             m_soCounter;
+    Rc<DxvkBuffer>              m_soCounter;
     DxvkBufferSliceHandle       m_mapped;
 
     D3D11DXGIResource           m_resource;
@@ -139,6 +141,8 @@ namespace dxvk {
     
     VkMemoryPropertyFlags GetMemoryFlags() const;
 
+    Rc<DxvkBuffer> CreateSoCounterBuffer();
+
   };
 
 
diff --git a/src/d3d11/d3d11_device.cpp b/src/d3d11/d3d11_device.cpp
index 2083a8a72..c02883af1 100644
--- a/src/d3d11/d3d11_device.cpp
+++ b/src/d3d11/d3d11_device.cpp
@@ -45,7 +45,6 @@ namespace dxvk {
     m_context     = new D3D11ImmediateContext(this, m_dxvkDevice);
     m_d3d10Device = new D3D10Device(this, m_context);
 
-    m_xfbCounters = CreateXFBCounterBuffer();
     m_predicates  = CreatePredicateBuffer();
   }
   
@@ -1969,27 +1968,6 @@ namespace dxvk {
   }
   
   
-  Rc<D3D11CounterBuffer> D3D11Device::CreateXFBCounterBuffer() {
-    DxvkBufferCreateInfo xfbCounterInfo;
-    xfbCounterInfo.size   = 4096 * sizeof(D3D11SOCounter);
-    xfbCounterInfo.usage  = VK_BUFFER_USAGE_TRANSFER_DST_BIT
-                          | VK_BUFFER_USAGE_TRANSFER_SRC_BIT
-                          | VK_BUFFER_USAGE_INDIRECT_BUFFER_BIT
-                          | VK_BUFFER_USAGE_TRANSFORM_FEEDBACK_COUNTER_BUFFER_BIT_EXT;
-    xfbCounterInfo.stages = VK_PIPELINE_STAGE_TRANSFER_BIT
-                          | VK_PIPELINE_STAGE_DRAW_INDIRECT_BIT
-                          | VK_PIPELINE_STAGE_TRANSFORM_FEEDBACK_BIT_EXT;
-    xfbCounterInfo.access = VK_ACCESS_TRANSFER_READ_BIT
-                          | VK_ACCESS_TRANSFER_WRITE_BIT
-                          | VK_ACCESS_INDIRECT_COMMAND_READ_BIT
-                          | VK_ACCESS_TRANSFORM_FEEDBACK_COUNTER_READ_BIT_EXT
-                          | VK_ACCESS_TRANSFORM_FEEDBACK_COUNTER_WRITE_BIT_EXT;
-    
-    return new D3D11CounterBuffer(m_dxvkDevice,
-      xfbCounterInfo, sizeof(D3D11SOCounter));
-  }
-  
-  
   Rc<D3D11CounterBuffer> D3D11Device::CreatePredicateBuffer() {
     DxvkBufferCreateInfo predCounterInfo;
     predCounterInfo.size   = 4096 * sizeof(uint32_t);
diff --git a/src/d3d11/d3d11_device.h b/src/d3d11/d3d11_device.h
index 8d24e2a0a..80808f2ea 100644
--- a/src/d3d11/d3d11_device.h
+++ b/src/d3d11/d3d11_device.h
@@ -425,10 +425,8 @@ namespace dxvk {
       return m_d3d10Device;
     }
     
-    DxvkBufferSlice AllocXfbCounterSlice() { return m_xfbCounters->AllocSlice(); }
     DxvkBufferSlice AllocPredicateSlice () { return m_predicates ->AllocSlice(); }
     
-    void FreeXfbCounterSlice(const DxvkBufferSlice& Slice) { m_xfbCounters->FreeSlice(Slice); }
     void FreePredicateSlice (const DxvkBufferSlice& Slice) { m_predicates ->FreeSlice(Slice); }
     
     static bool CheckFeatureLevelSupport(
@@ -459,7 +457,6 @@ namespace dxvk {
     D3D11ImmediateContext*          m_context     = nullptr;
     D3D10Device*                    m_d3d10Device = nullptr;
 
-    Rc<D3D11CounterBuffer>          m_xfbCounters;
     Rc<D3D11CounterBuffer>          m_predicates;
     
     D3D11StateObjectSet<D3D11BlendState>        m_bsStateObjects;
@@ -469,7 +466,6 @@ namespace dxvk {
     D3D11ShaderModuleSet                        m_shaderModules;
     D3D11CommandListAllocator                   m_commandListAllocator;
     
-    Rc<D3D11CounterBuffer> CreateXFBCounterBuffer();
     Rc<D3D11CounterBuffer> CreatePredicateBuffer();
 
     HRESULT CreateShaderModule(