Merge pull request #9648 from stenzek/dx11-gl-bbox-caching

Cache bounding box values between register reads in DX11/GL

Source/Core/VideoBackends/D3D/D3DBoundingBox.cpp

@@ -2,18 +2,25 @@
 // Licensed under GPLv2+
 // Refer to the license.txt file included.
 
-#include "VideoBackends/D3D/D3DBoundingBox.h"
+#include <algorithm>
+#include <array>
+
 #include "Common/CommonTypes.h"
 #include "Common/MsgHandler.h"
+#include "VideoBackends/D3D/D3DBoundingBox.h"
 #include "VideoBackends/D3D/D3DState.h"
 #include "VideoBackends/D3DCommon/D3DCommon.h"
 #include "VideoCommon/VideoConfig.h"
 
 namespace DX11
 {
+static constexpr u32 NUM_BBOX_VALUES = 4;
 static ComPtr<ID3D11Buffer> s_bbox_buffer;
 static ComPtr<ID3D11Buffer> s_bbox_staging_buffer;
 static ComPtr<ID3D11UnorderedAccessView> s_bbox_uav;
+static std::array<s32, NUM_BBOX_VALUES> s_bbox_values;
+static std::array<bool, NUM_BBOX_VALUES> s_bbox_dirty;
+static bool s_bbox_valid = false;
 
 ID3D11UnorderedAccessView* BBox::GetUAV()
 {
@@ -22,16 +29,17 @@ ID3D11UnorderedAccessView* BBox::GetUAV()
 
 void BBox::Init()
 {
-  if (g_ActiveConfig.backend_info.bSupportsBBox)
+  if (!g_ActiveConfig.backend_info.bSupportsBBox)
-  {
+    return;
+
   // Create 2 buffers here.
   // First for unordered access on default pool.
-    auto desc = CD3D11_BUFFER_DESC(4 * sizeof(s32), D3D11_BIND_UNORDERED_ACCESS,
+  auto desc = CD3D11_BUFFER_DESC(NUM_BBOX_VALUES * sizeof(s32), D3D11_BIND_UNORDERED_ACCESS,
-                                   D3D11_USAGE_DEFAULT, 0, 0, 4);
+                                 D3D11_USAGE_DEFAULT, 0, 0, sizeof(s32));
-    int initial_values[4] = {0, 0, 0, 0};
+  const s32 initial_values[NUM_BBOX_VALUES] = {0, 0, 0, 0};
   D3D11_SUBRESOURCE_DATA data;
   data.pSysMem = initial_values;
-    data.SysMemPitch = 4 * sizeof(s32);
+  data.SysMemPitch = NUM_BBOX_VALUES * sizeof(s32);
   data.SysMemSlicePitch = 0;
   HRESULT hr;
   hr = D3D::device->CreateBuffer(&desc, &data, &s_bbox_buffer);
@@ -52,12 +60,14 @@ void BBox::Init()
   UAVdesc.ViewDimension = D3D11_UAV_DIMENSION_BUFFER;
   UAVdesc.Buffer.FirstElement = 0;
   UAVdesc.Buffer.Flags = 0;
-    UAVdesc.Buffer.NumElements = 4;
+  UAVdesc.Buffer.NumElements = NUM_BBOX_VALUES;
   hr = D3D::device->CreateUnorderedAccessView(s_bbox_buffer.Get(), &UAVdesc, &s_bbox_uav);
   CHECK(SUCCEEDED(hr), "Create BoundingBox UAV.");
   D3DCommon::SetDebugObjectName(s_bbox_uav.Get(), "BoundingBox UAV");
   D3D::stateman->SetOMUAV(s_bbox_uav.Get());
-  }
+
+  s_bbox_dirty = {};
+  s_bbox_valid = true;
 }
 
 void BBox::Shutdown()
@@ -67,23 +77,73 @@ void BBox::Shutdown()
   s_bbox_buffer.Reset();
 }
 
-void BBox::Set(int index, int value)
+void BBox::Flush()
 {
-  D3D11_BOX box{index * sizeof(s32), 0, 0, (index + 1) * sizeof(s32), 1, 1};
+  s_bbox_valid = false;
-  D3D::context->UpdateSubresource(s_bbox_buffer.Get(), 0, &box, &value, 0, 0);
+
+  if (std::none_of(s_bbox_dirty.begin(), s_bbox_dirty.end(), [](bool dirty) { return dirty; }))
+    return;
+
+  for (u32 start = 0; start < NUM_BBOX_VALUES;)
+  {
+    if (!s_bbox_dirty[start])
+    {
+      start++;
+      continue;
     }
 
-int BBox::Get(int index)
+    u32 end = start + 1;
+    s_bbox_dirty[start] = false;
+    for (; end < NUM_BBOX_VALUES; end++)
+    {
+      if (!s_bbox_dirty[end])
+        break;
+
+      s_bbox_dirty[end] = false;
+    }
+
+    D3D11_BOX box{start * sizeof(s32), 0, 0, end * sizeof(s32), 1, 1};
+    D3D::context->UpdateSubresource(s_bbox_buffer.Get(), 0, &box, &s_bbox_values[start], 0, 0);
+  }
+}
+
+void BBox::Readback()
 {
-  int data = 0;
   D3D::context->CopyResource(s_bbox_staging_buffer.Get(), s_bbox_buffer.Get());
+
   D3D11_MAPPED_SUBRESOURCE map;
   HRESULT hr = D3D::context->Map(s_bbox_staging_buffer.Get(), 0, D3D11_MAP_READ, 0, &map);
   if (SUCCEEDED(hr))
   {
-    data = ((s32*)map.pData)[index];
+    for (u32 i = 0; i < NUM_BBOX_VALUES; i++)
+    {
+      if (!s_bbox_dirty[i])
+      {
+        std::memcpy(&s_bbox_values[i], reinterpret_cast<const u8*>(map.pData) + sizeof(s32) * i,
+                    sizeof(s32));
       }
+    }
+
     D3D::context->Unmap(s_bbox_staging_buffer.Get(), 0);
-  return data;
+  }
+
+  s_bbox_valid = true;
+}
+
+void BBox::Set(int index, int value)
+{
+  if (s_bbox_valid && s_bbox_values[index] == value)
+    return;
+
+  s_bbox_values[index] = value;
+  s_bbox_dirty[index] = true;
+}
+
+int BBox::Get(int index)
+{
+  if (!s_bbox_valid)
+    Readback();
+
+  return s_bbox_values[index];
 }
 };  // namespace DX11

Source/Core/VideoBackends/D3D/D3DBoundingBox.h

@@ -14,6 +14,9 @@ public:
   static void Init();
   static void Shutdown();
 
+  static void Flush();
+  static void Readback();
+
   static void Set(int index, int value);
   static int Get(int index);
 };

Source/Core/VideoBackends/D3D/D3DRender.cpp

@@ -274,6 +274,11 @@ void Renderer::BBoxWriteImpl(int index, u16 value)
   BBox::Set(index, value);
 }
 
+void Renderer::BBoxFlushImpl()
+{
+  BBox::Flush();
+}
+
 void Renderer::Flush()
 {
   D3D::context->Flush();

Source/Core/VideoBackends/D3D/D3DRender.h

@@ -63,6 +63,7 @@ public:
 
   u16 BBoxReadImpl(int index) override;
   void BBoxWriteImpl(int index, u16 value) override;
+  void BBoxFlushImpl() override;
 
   void Flush() override;
   void WaitForGPUIdle() override;

Source/Core/VideoBackends/OGL/OGLBoundingBox.cpp

@@ -2,6 +2,8 @@
 // Licensed under GPLv2+
 // Refer to the license.txt file included.
 
+#include <algorithm>
+#include <array>
 #include <cstring>
 
 #include "Common/GL/GLUtil.h"
@@ -12,7 +14,15 @@
 #include "VideoCommon/DriverDetails.h"
 #include "VideoCommon/VideoConfig.h"
 
+enum : u32
+{
+  NUM_BBOX_VALUES = 4,
+};
+
 static GLuint s_bbox_buffer_id;
+static std::array<s32, NUM_BBOX_VALUES> s_bbox_values;
+static std::array<bool, NUM_BBOX_VALUES> s_bbox_dirty;
+static bool s_bbox_valid = false;
 
 namespace OGL
 {
@@ -21,10 +31,14 @@ void BoundingBox::Init()
   if (!g_ActiveConfig.backend_info.bSupportsBBox)
     return;
 
-  int initial_values[4] = {0, 0, 0, 0};
+  const s32 initial_values[NUM_BBOX_VALUES] = {0, 0, 0, 0};
+  std::memcpy(s_bbox_values.data(), initial_values, sizeof(s_bbox_values));
+  s_bbox_dirty = {};
+  s_bbox_valid = true;
+
   glGenBuffers(1, &s_bbox_buffer_id);
   glBindBuffer(GL_SHADER_STORAGE_BUFFER, s_bbox_buffer_id);
-  glBufferData(GL_SHADER_STORAGE_BUFFER, 4 * sizeof(s32), initial_values, GL_DYNAMIC_DRAW);
+  glBufferData(GL_SHADER_STORAGE_BUFFER, sizeof(initial_values), initial_values, GL_DYNAMIC_DRAW);
   glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 0, s_bbox_buffer_id);
 }
 
@@ -36,22 +50,42 @@ void BoundingBox::Shutdown()
   glDeleteBuffers(1, &s_bbox_buffer_id);
 }
 
-void BoundingBox::Set(int index, int value)
+void BoundingBox::Flush()
 {
-  if (!g_ActiveConfig.backend_info.bSupportsBBox)
+  s_bbox_valid = false;
+
+  if (std::none_of(s_bbox_dirty.begin(), s_bbox_dirty.end(), [](bool dirty) { return dirty; }))
     return;
 
   glBindBuffer(GL_SHADER_STORAGE_BUFFER, s_bbox_buffer_id);
-  glBufferSubData(GL_SHADER_STORAGE_BUFFER, index * sizeof(int), sizeof(int), &value);
+
+  for (u32 start = 0; start < NUM_BBOX_VALUES;)
+  {
+    if (!s_bbox_dirty[start])
+    {
+      start++;
+      continue;
+    }
+
+    u32 end = start + 1;
+    s_bbox_dirty[start] = false;
+    for (; end < NUM_BBOX_VALUES; end++)
+    {
+      if (!s_bbox_dirty[end])
+        break;
+
+      s_bbox_dirty[end] = false;
+    }
+
+    glBufferSubData(GL_SHADER_STORAGE_BUFFER, start * sizeof(s32), (end - start) * sizeof(s32),
+                    &s_bbox_values[start]);
+  }
+
   glBindBuffer(GL_SHADER_STORAGE_BUFFER, 0);
 }
 
-int BoundingBox::Get(int index)
+void BoundingBox::Readback()
 {
-  if (!g_ActiveConfig.backend_info.bSupportsBBox)
-    return 0;
-
-  int data = 0;
   glBindBuffer(GL_SHADER_STORAGE_BUFFER, s_bbox_buffer_id);
   if (!DriverDetails::HasBug(DriverDetails::BUG_SLOW_GETBUFFERSUBDATA) &&
       !static_cast<Renderer*>(g_renderer.get())->IsGLES())
@@ -59,20 +93,52 @@ int BoundingBox::Get(int index)
     // Using glMapBufferRange to read back the contents of the SSBO is extremely slow
     // on nVidia drivers. This is more noticeable at higher internal resolutions.
     // Using glGetBufferSubData instead does not seem to exhibit this slowdown.
-    glGetBufferSubData(GL_SHADER_STORAGE_BUFFER, index * sizeof(int), sizeof(int), &data);
+    std::array<s32, NUM_BBOX_VALUES> gpu_values;
+    glGetBufferSubData(GL_SHADER_STORAGE_BUFFER, 0, sizeof(s32) * NUM_BBOX_VALUES,
+                       gpu_values.data());
+    for (u32 i = 0; i < NUM_BBOX_VALUES; i++)
+    {
+      if (!s_bbox_dirty[i])
+        s_bbox_values[i] = gpu_values[i];
+    }
   }
   else
   {
     // Using glMapBufferRange is faster on AMD cards by a measurable margin.
-    void* ptr = glMapBufferRange(GL_SHADER_STORAGE_BUFFER, index * sizeof(int), sizeof(int),
+    void* ptr = glMapBufferRange(GL_SHADER_STORAGE_BUFFER, 0, sizeof(s32) * NUM_BBOX_VALUES,
                                  GL_MAP_READ_BIT);
     if (ptr)
     {
-      memcpy(&data, ptr, sizeof(int));
+      for (u32 i = 0; i < NUM_BBOX_VALUES; i++)
+      {
+        if (!s_bbox_dirty[i])
+        {
+          std::memcpy(&s_bbox_values[i], reinterpret_cast<const u8*>(ptr) + sizeof(s32) * i,
+                      sizeof(s32));
+        }
+      }
+
       glUnmapBuffer(GL_SHADER_STORAGE_BUFFER);
     }
   }
   glBindBuffer(GL_SHADER_STORAGE_BUFFER, 0);
-  return data;
+  s_bbox_valid = true;
+}
+
+void BoundingBox::Set(int index, int value)
+{
+  if (s_bbox_valid && s_bbox_values[index] == value)
+    return;
+
+  s_bbox_values[index] = value;
+  s_bbox_dirty[index] = true;
+}
+
+int BoundingBox::Get(int index)
+{
+  if (!s_bbox_valid)
+    Readback();
+
+  return s_bbox_values[index];
 }
 };  // namespace OGL

Source/Core/VideoBackends/OGL/OGLBoundingBox.h

@@ -12,6 +12,9 @@ public:
   static void Init();
   static void Shutdown();
 
+  static void Flush();
+  static void Readback();
+
   static void Set(int index, int value);
   static int Get(int index);
 };

Source/Core/VideoBackends/OGL/OGLRender.cpp

@@ -882,6 +882,11 @@ void Renderer::BBoxWriteImpl(int index, u16 value)
   BoundingBox::Set(index, swapped_value);
 }
 
+void Renderer::BBoxFlushImpl()
+{
+  BoundingBox::Flush();
+}
+
 void Renderer::SetViewport(float x, float y, float width, float height, float near_depth,
                            float far_depth)
 {

Source/Core/VideoBackends/OGL/OGLRender.h

@@ -128,6 +128,7 @@ public:
 
   u16 BBoxReadImpl(int index) override;
   void BBoxWriteImpl(int index, u16 value) override;
+  void BBoxFlushImpl() override;
 
   void BeginUtilityDrawing() override;
   void EndUtilityDrawing() override;