Use main buffers for utility draws

Source/Core/VideoBackends/D3D/GeometryShaderCache.cpp

@@ -4,7 +4,6 @@
 
 #include <string>
 
-#include "Common/Align.h"
 #include "Common/FileUtil.h"
 #include "Common/StringUtil.h"
 
@@ -18,8 +17,6 @@
 
 #include "VideoCommon/Debugger.h"
 #include "VideoCommon/GeometryShaderGen.h"
-#include "VideoCommon/GeometryShaderManager.h"
-#include "VideoCommon/Statistics.h"
 #include "VideoCommon/VideoConfig.h"
 
 namespace DX11
@@ -36,25 +33,6 @@ ID3D11GeometryShader* GeometryShaderCache::GetCopyGeometryShader()
   return (g_ActiveConfig.stereo_mode != StereoMode::Off) ? CopyGeometryShader : nullptr;
 }
 
-ID3D11Buffer* gscbuf = nullptr;
-
-ID3D11Buffer*& GeometryShaderCache::GetConstantBuffer()
-{
-  // TODO: divide the global variables of the generated shaders into about 5 constant buffers to
-  // speed this up
-  if (GeometryShaderManager::dirty)
-  {
-    D3D11_MAPPED_SUBRESOURCE map;
-    D3D::context->Map(gscbuf, 0, D3D11_MAP_WRITE_DISCARD, 0, &map);
-    memcpy(map.pData, &GeometryShaderManager::constants, sizeof(GeometryShaderConstants));
-    D3D::context->Unmap(gscbuf, 0);
-    GeometryShaderManager::dirty = false;
-
-    ADDSTAT(stats.thisFrame.bytesUniformStreamed, sizeof(GeometryShaderConstants));
-  }
-  return gscbuf;
-}
-
 const char clear_shader_code[] = {
     "struct VSOUTPUT\n"
     "{\n"
@@ -116,15 +94,6 @@ const char copy_shader_code[] = {
 
 void GeometryShaderCache::Init()
 {
-  unsigned int gbsize = Common::AlignUp(static_cast<unsigned int>(sizeof(GeometryShaderConstants)),
-                                        16);  // must be a multiple of 16
-  D3D11_BUFFER_DESC gbdesc = CD3D11_BUFFER_DESC(gbsize, D3D11_BIND_CONSTANT_BUFFER,
-                                                D3D11_USAGE_DYNAMIC, D3D11_CPU_ACCESS_WRITE);
-  HRESULT hr = D3D::device->CreateBuffer(&gbdesc, nullptr, &gscbuf);
-  CHECK(hr == S_OK, "Create geometry shader constant buffer (size=%u)", gbsize);
-  D3D::SetDebugObjectName(gscbuf,
-                          "geometry shader constant buffer used to emulate the GX pipeline");
-
   // used when drawing clear quads
   ClearGeometryShader = D3D::CompileAndCreateGeometryShader(clear_shader_code);
   CHECK(ClearGeometryShader != nullptr, "Create clear geometry shader");
@@ -138,8 +107,6 @@ void GeometryShaderCache::Init()
 
 void GeometryShaderCache::Shutdown()
 {
-  SAFE_RELEASE(gscbuf);
-
   SAFE_RELEASE(ClearGeometryShader);
   SAFE_RELEASE(CopyGeometryShader);
 }

Source/Core/VideoBackends/D3D/GeometryShaderCache.h

@@ -20,7 +20,8 @@ public:
   static ID3D11GeometryShader* GetClearGeometryShader();
   static ID3D11GeometryShader* GetCopyGeometryShader();
 
-  static ID3D11Buffer*& GetConstantBuffer();
+  static ID3D11Buffer* GetConstantBuffer();
+  static void UpdateConstantBuffer(const void* data, u32 data_size);
 };
 
 }  // namespace DX11

Source/Core/VideoBackends/D3D/PixelShaderCache.cpp

@@ -4,7 +4,6 @@
 
 #include <string>
 
-#include "Common/Align.h"
 #include "Common/CommonTypes.h"
 #include "Common/FileUtil.h"
 #include "Common/MsgHandler.h"
@@ -20,8 +19,6 @@
 
 #include "VideoCommon/Debugger.h"
 #include "VideoCommon/PixelShaderGen.h"
-#include "VideoCommon/PixelShaderManager.h"
-#include "VideoCommon/Statistics.h"
 #include "VideoCommon/VideoConfig.h"
 
 namespace DX11
@@ -32,7 +29,6 @@ ID3D11PixelShader* s_AnaglyphProgram = nullptr;
 ID3D11PixelShader* s_DepthResolveProgram = nullptr;
 ID3D11PixelShader* s_rgba6_to_rgb8[2] = {nullptr};
 ID3D11PixelShader* s_rgb8_to_rgba6[2] = {nullptr};
-ID3D11Buffer* pscbuf = nullptr;
 
 const char clear_program_code[] = {"void main(\n"
                                    "out float4 ocol0 : SV_Target,\n"
@@ -277,36 +273,8 @@ ID3D11PixelShader* PixelShaderCache::GetDepthResolveProgram()
   return s_DepthResolveProgram;
 }
 
-static void UpdateConstantBuffers()
-{
-  if (PixelShaderManager::dirty)
-  {
-    D3D11_MAPPED_SUBRESOURCE map;
-    D3D::context->Map(pscbuf, 0, D3D11_MAP_WRITE_DISCARD, 0, &map);
-    memcpy(map.pData, &PixelShaderManager::constants, sizeof(PixelShaderConstants));
-    D3D::context->Unmap(pscbuf, 0);
-    PixelShaderManager::dirty = false;
-
-    ADDSTAT(stats.thisFrame.bytesUniformStreamed, sizeof(PixelShaderConstants));
-  }
-}
-
-ID3D11Buffer* PixelShaderCache::GetConstantBuffer()
-{
-  UpdateConstantBuffers();
-  return pscbuf;
-}
-
 void PixelShaderCache::Init()
 {
-  unsigned int cbsize = Common::AlignUp(static_cast<unsigned int>(sizeof(PixelShaderConstants)),
-                                        16);  // must be a multiple of 16
-  D3D11_BUFFER_DESC cbdesc = CD3D11_BUFFER_DESC(cbsize, D3D11_BIND_CONSTANT_BUFFER,
-                                                D3D11_USAGE_DYNAMIC, D3D11_CPU_ACCESS_WRITE);
-  D3D::device->CreateBuffer(&cbdesc, nullptr, &pscbuf);
-  CHECK(pscbuf != nullptr, "Create pixel shader constant buffer");
-  D3D::SetDebugObjectName(pscbuf, "pixel shader constant buffer used to emulate the GX pipeline");
-
   // used when drawing clear quads
   s_ClearProgram = D3D::CompileAndCreatePixelShader(clear_program_code);
   CHECK(s_ClearProgram != nullptr, "Create clear pixel shader");
@@ -334,8 +302,6 @@ void PixelShaderCache::InvalidateMSAAShaders()
 
 void PixelShaderCache::Shutdown()
 {
-  SAFE_RELEASE(pscbuf);
-
   SAFE_RELEASE(s_ClearProgram);
   SAFE_RELEASE(s_AnaglyphProgram);
   SAFE_RELEASE(s_DepthResolveProgram);

Source/Core/VideoBackends/D3D/PixelShaderCache.h

@@ -21,8 +21,6 @@ public:
   static void Init();
   static void Shutdown();
 
-  static ID3D11Buffer* GetConstantBuffer();
-
   static ID3D11PixelShader* GetColorCopyProgram(bool multisampled);
   static ID3D11PixelShader* GetClearProgram();
   static ID3D11PixelShader* GetAnaglyphProgram();

Source/Core/VideoBackends/D3D/Render.cpp

@@ -167,16 +167,6 @@ void Renderer::SetupDeviceObjects()
   D3D::SetDebugObjectName(m_reset_rast_state, "rasterizer state for Renderer::ResetAPIState");
 
   m_screenshot_texture = nullptr;
-
-  CD3D11_BUFFER_DESC vbo_desc(UTILITY_VBO_SIZE, D3D11_BIND_VERTEX_BUFFER, D3D11_USAGE_DYNAMIC,
-                              D3D11_CPU_ACCESS_WRITE);
-  hr = D3D::device->CreateBuffer(&vbo_desc, nullptr, &m_utility_vertex_buffer);
-  CHECK(SUCCEEDED(hr), "Create utility VBO");
-
-  CD3D11_BUFFER_DESC ubo_desc(UTILITY_UBO_SIZE, D3D11_BIND_CONSTANT_BUFFER, D3D11_USAGE_DYNAMIC,
-                              D3D11_CPU_ACCESS_WRITE);
-  hr = D3D::device->CreateBuffer(&ubo_desc, nullptr, &m_utility_uniform_buffer);
-  CHECK(SUCCEEDED(hr), "Create utility UBO");
 }
 
 // Kill off all device objects
@@ -196,8 +186,6 @@ void Renderer::TeardownDeviceObjects()
   SAFE_RELEASE(m_reset_rast_state);
   SAFE_RELEASE(m_screenshot_texture);
   SAFE_RELEASE(m_3d_vision_texture);
-  SAFE_RELEASE(m_utility_vertex_buffer);
-  SAFE_RELEASE(m_utility_uniform_buffer);
 }
 
 void Renderer::Create3DVisionTexture(int width, int height)
@@ -273,25 +261,6 @@ std::unique_ptr<AbstractPipeline> Renderer::CreatePipeline(const AbstractPipelin
   return DXPipeline::Create(config);
 }
 
-void Renderer::UpdateUtilityUniformBuffer(const void* uniforms, u32 uniforms_size)
-{
-  DEBUG_ASSERT(uniforms_size > 0 && uniforms_size < UTILITY_UBO_SIZE);
-  D3D11_MAPPED_SUBRESOURCE mapped;
-  HRESULT hr = D3D::context->Map(m_utility_uniform_buffer, 0, D3D11_MAP_WRITE_DISCARD, 0, &mapped);
-  CHECK(SUCCEEDED(hr), "Map utility UBO");
-  std::memcpy(mapped.pData, uniforms, uniforms_size);
-  D3D::context->Unmap(m_utility_uniform_buffer, 0);
-}
-
-void Renderer::UpdateUtilityVertexBuffer(const void* vertices, u32 vertex_stride, u32 num_vertices)
-{
-  D3D11_MAPPED_SUBRESOURCE mapped;
-  HRESULT hr = D3D::context->Map(m_utility_vertex_buffer, 0, D3D11_MAP_WRITE_DISCARD, 0, &mapped);
-  CHECK(SUCCEEDED(hr), "Map utility VBO");
-  std::memcpy(mapped.pData, vertices, num_vertices * vertex_stride);
-  D3D::context->Unmap(m_utility_vertex_buffer, 0);
-}
-
 void Renderer::SetPipeline(const AbstractPipeline* pipeline)
 {
   const DXPipeline* dx_pipeline = static_cast<const DXPipeline*>(pipeline);
@@ -308,54 +277,6 @@ void Renderer::SetPipeline(const AbstractPipeline* pipeline)
   D3D::stateman->SetPixelShader(dx_pipeline->GetPixelShader());
 }
 
-void Renderer::DrawUtilityPipeline(const void* uniforms, u32 uniforms_size, const void* vertices,
-                                   u32 vertex_stride, u32 num_vertices)
-{
-  // Copy in uniforms.
-  if (uniforms_size > 0)
-  {
-    UpdateUtilityUniformBuffer(uniforms, uniforms_size);
-    D3D::stateman->SetVertexConstants(m_utility_uniform_buffer);
-    D3D::stateman->SetPixelConstants(m_utility_uniform_buffer);
-    D3D::stateman->SetGeometryConstants(m_utility_uniform_buffer);
-  }
-
-  // If the vertices are larger than our buffer, we need to break it up into multiple draws.
-  const char* vertices_ptr = static_cast<const char*>(vertices);
-  while (num_vertices > 0)
-  {
-    u32 vertices_this_draw = num_vertices;
-    if (vertices_ptr)
-    {
-      vertices_this_draw = std::min(vertices_this_draw, UTILITY_VBO_SIZE / vertex_stride);
-      DEBUG_ASSERT(vertices_this_draw > 0);
-      UpdateUtilityVertexBuffer(vertices_ptr, vertex_stride, vertices_this_draw);
-      D3D::stateman->SetVertexBuffer(m_utility_vertex_buffer, vertex_stride, 0);
-    }
-
-    // Apply pending state and draw.
-    D3D::stateman->Apply();
-    D3D::context->Draw(vertices_this_draw, 0);
-    vertices_ptr += vertex_stride * vertices_this_draw;
-    num_vertices -= vertices_this_draw;
-  }
-}
-
-void Renderer::DispatchComputeShader(const AbstractShader* shader, const void* uniforms,
-                                     u32 uniforms_size, u32 groups_x, u32 groups_y, u32 groups_z)
-{
-  D3D::stateman->SetComputeShader(static_cast<const DXShader*>(shader)->GetD3DComputeShader());
-
-  if (uniforms_size > 0)
-  {
-    UpdateUtilityUniformBuffer(uniforms, uniforms_size);
-    D3D::stateman->SetComputeConstants(m_utility_uniform_buffer);
-  }
-
-  D3D::stateman->Apply();
-  D3D::context->Dispatch(groups_x, groups_y, groups_z);
-}
-
 TargetRectangle Renderer::ConvertEFBRectangle(const EFBRectangle& rc)
 {
   TargetRectangle result;
@@ -560,6 +481,18 @@ void Renderer::SetViewport(float x, float y, float width, float height, float ne
   D3D::context->RSSetViewports(1, &vp);
 }
 
+void Renderer::Draw(u32 base_vertex, u32 num_vertices)
+{
+  D3D::stateman->Apply();
+  D3D::context->Draw(num_vertices, base_vertex);
+}
+
+void Renderer::DrawIndexed(u32 base_index, u32 num_indices, u32 base_vertex)
+{
+  D3D::stateman->Apply();
+  D3D::context->DrawIndexed(num_indices, base_index, base_vertex);
+}
+
 void Renderer::ClearScreen(const EFBRectangle& rc, bool colorEnable, bool alphaEnable, bool zEnable,
                            u32 color, u32 z)
 {

Source/Core/VideoBackends/D3D/Render.h

@@ -50,6 +50,8 @@ public:
   void SetInterlacingMode() override;
   void SetViewport(float x, float y, float width, float height, float near_depth,
                    float far_depth) override;
+  void Draw(u32 base_vertex, u32 num_vertices) override;
+  void DrawIndexed(u32 base_index, u32 num_indices, u32 base_vertex) override;
   void SetFullscreen(bool enable_fullscreen) override;
   bool IsFullscreen() const override;
 
@@ -73,11 +75,6 @@ public:
 
   void ReinterpretPixelData(unsigned int convtype) override;
 
-  void DrawUtilityPipeline(const void* uniforms, u32 uniforms_size, const void* vertices,
-                           u32 vertex_stride, u32 num_vertices) override;
-  void DispatchComputeShader(const AbstractShader* shader, const void* uniforms, u32 uniforms_size,
-                             u32 groups_x, u32 groups_y, u32 groups_z) override;
-
 private:
   void SetupDeviceObjects();
   void TeardownDeviceObjects();
@@ -89,9 +86,6 @@ private:
   void BlitScreen(TargetRectangle src, TargetRectangle dst, D3DTexture2D* src_texture,
                   u32 src_width, u32 src_height);
 
-  void UpdateUtilityUniformBuffer(const void* uniforms, u32 uniforms_size);
-  void UpdateUtilityVertexBuffer(const void* vertices, u32 vertex_stride, u32 num_vertices);
-
   StateCache m_state_cache;
 
   std::array<ID3D11BlendState*, 4> m_clear_blend_states{};
@@ -103,9 +97,6 @@ private:
   ID3D11Texture2D* m_screenshot_texture = nullptr;
   D3DTexture2D* m_3d_vision_texture = nullptr;
 
-  ID3D11Buffer* m_utility_vertex_buffer = nullptr;
-  ID3D11Buffer* m_utility_uniform_buffer = nullptr;
-
   u32 m_last_multisamples = 1;
   bool m_last_stereo_mode = false;
   bool m_last_fullscreen_state = false;

Source/Core/VideoBackends/D3D/VertexManager.cpp

@@ -6,6 +6,7 @@
 
 #include <d3d11.h>
 
+#include "Common/Align.h"
 #include "Common/CommonTypes.h"
 
 #include "VideoBackends/D3D/BoundingBox.h"
@@ -19,11 +20,14 @@
 
 #include "VideoCommon/BoundingBox.h"
 #include "VideoCommon/Debugger.h"
+#include "VideoCommon/GeometryShaderManager.h"
 #include "VideoCommon/IndexGenerator.h"
 #include "VideoCommon/NativeVertexFormat.h"
+#include "VideoCommon/PixelShaderManager.h"
 #include "VideoCommon/RenderBase.h"
 #include "VideoCommon/Statistics.h"
 #include "VideoCommon/VertexLoaderManager.h"
+#include "VideoCommon/VertexShaderManager.h"
 #include "VideoCommon/VideoConfig.h"
 
 namespace DX11
@@ -33,15 +37,34 @@ const u32 MAX_IBUFFER_SIZE = VertexManager::MAXIBUFFERSIZE * sizeof(u16) * 8;
 const u32 MAX_VBUFFER_SIZE = VertexManager::MAXVBUFFERSIZE;
 const u32 MAX_BUFFER_SIZE = MAX_IBUFFER_SIZE + MAX_VBUFFER_SIZE;
 
+static ID3D11Buffer* AllocateConstantBuffer(u32 size)
+{
+  const u32 cbsize = Common::AlignUp(size, 16u);  // must be a multiple of 16
+  const CD3D11_BUFFER_DESC cbdesc(cbsize, D3D11_BIND_CONSTANT_BUFFER, D3D11_USAGE_DYNAMIC,
+                                  D3D11_CPU_ACCESS_WRITE);
+  ID3D11Buffer* cbuf;
+  const HRESULT hr = D3D::device->CreateBuffer(&cbdesc, nullptr, &cbuf);
+  CHECK(hr == S_OK, "shader constant buffer (size=%u)", cbsize);
+  D3D::SetDebugObjectName(cbuf, "constant buffer used to emulate the GX pipeline");
+  return cbuf;
+}
+
+static void UpdateConstantBuffer(ID3D11Buffer* const buffer, const void* data, u32 data_size)
+{
+  D3D11_MAPPED_SUBRESOURCE map;
+  D3D::context->Map(buffer, 0, D3D11_MAP_WRITE_DISCARD, 0, &map);
+  memcpy(map.pData, data, data_size);
+  D3D::context->Unmap(buffer, 0);
+
+  ADDSTAT(stats.thisFrame.bytesUniformStreamed, data_size);
+}
+
 void VertexManager::CreateDeviceObjects()
 {
   D3D11_BUFFER_DESC bufdesc =
       CD3D11_BUFFER_DESC(MAX_BUFFER_SIZE, D3D11_BIND_INDEX_BUFFER | D3D11_BIND_VERTEX_BUFFER,
                          D3D11_USAGE_DYNAMIC, D3D11_CPU_ACCESS_WRITE);
 
-  m_vertexDrawOffset = 0;
-  m_indexDrawOffset = 0;
-
   for (int i = 0; i < MAX_BUFFER_COUNT; i++)
   {
     m_buffers[i] = nullptr;
@@ -50,12 +73,18 @@ void VertexManager::CreateDeviceObjects()
     D3D::SetDebugObjectName(m_buffers[i], "Buffer of VertexManager");
   }
 
-  m_currentBuffer = 0;
+  m_buffer_cursor = MAX_BUFFER_SIZE;
-  m_bufferCursor = MAX_BUFFER_SIZE;
+
+  m_vertex_constant_buffer = AllocateConstantBuffer(sizeof(VertexShaderConstants));
+  m_geometry_constant_buffer = AllocateConstantBuffer(sizeof(GeometryShaderConstants));
+  m_pixel_constant_buffer = AllocateConstantBuffer(sizeof(PixelShaderConstants));
 }
 
 void VertexManager::DestroyDeviceObjects()
 {
+  SAFE_RELEASE(m_pixel_constant_buffer);
+  SAFE_RELEASE(m_geometry_constant_buffer);
+  SAFE_RELEASE(m_vertex_constant_buffer);
   for (int i = 0; i < MAX_BUFFER_COUNT; i++)
   {
     SAFE_RELEASE(m_buffers[i]);
@@ -64,12 +93,12 @@ void VertexManager::DestroyDeviceObjects()
 
 VertexManager::VertexManager()
 {
-  LocalVBuffer.resize(MAXVBUFFERSIZE);
+  m_staging_vertex_buffer.resize(MAXVBUFFERSIZE);
 
-  m_cur_buffer_pointer = m_base_buffer_pointer = &LocalVBuffer[0];
+  m_cur_buffer_pointer = m_base_buffer_pointer = &m_staging_vertex_buffer[0];
-  m_end_buffer_pointer = m_base_buffer_pointer + LocalVBuffer.size();
+  m_end_buffer_pointer = m_base_buffer_pointer + m_staging_vertex_buffer.size();
 
-  LocalIBuffer.resize(MAXIBUFFERSIZE);
+  m_staging_index_buffer.resize(MAXIBUFFERSIZE);
 
   CreateDeviceObjects();
 }
@@ -79,71 +108,103 @@ VertexManager::~VertexManager()
   DestroyDeviceObjects();
 }
 
-void VertexManager::PrepareDrawBuffers(u32 stride)
+void VertexManager::UploadUtilityUniforms(const void* uniforms, u32 uniforms_size)
+{
+  // Just use the one buffer for all three.
+  UpdateConstantBuffer(m_vertex_constant_buffer, uniforms, uniforms_size);
+  D3D::stateman->SetVertexConstants(m_vertex_constant_buffer);
+  D3D::stateman->SetGeometryConstants(m_vertex_constant_buffer);
+  D3D::stateman->SetPixelConstants(m_vertex_constant_buffer);
+  VertexShaderManager::dirty = true;
+  GeometryShaderManager::dirty = true;
+  PixelShaderManager::dirty = true;
+}
+
+void VertexManager::ResetBuffer(u32 vertex_stride, bool cull_all)
+{
+  m_cur_buffer_pointer = m_base_buffer_pointer;
+  IndexGenerator::Start(m_staging_index_buffer.data());
+}
+
+void VertexManager::CommitBuffer(u32 num_vertices, u32 vertex_stride, u32 num_indices,
+                                 u32* out_base_vertex, u32* out_base_index)
 {
   D3D11_MAPPED_SUBRESOURCE map;
 
-  u32 vertexBufferSize = u32(m_cur_buffer_pointer - m_base_buffer_pointer);
+  u32 vertexBufferSize = Common::AlignUp(num_vertices * vertex_stride, sizeof(u16));
-  u32 indexBufferSize = IndexGenerator::GetIndexLen() * sizeof(u16);
+  u32 indexBufferSize = num_indices * sizeof(u16);
   u32 totalBufferSize = vertexBufferSize + indexBufferSize;
 
-  u32 cursor = m_bufferCursor;
+  u32 cursor = m_buffer_cursor;
-  u32 padding = m_bufferCursor % stride;
+  u32 padding = vertex_stride > 0 ? (m_buffer_cursor % vertex_stride) : 0;
   if (padding)
   {
-    cursor += stride - padding;
+    cursor += vertex_stride - padding;
   }
 
   D3D11_MAP MapType = D3D11_MAP_WRITE_NO_OVERWRITE;
   if (cursor + totalBufferSize >= MAX_BUFFER_SIZE)
   {
     // Wrap around
-    m_currentBuffer = (m_currentBuffer + 1) % MAX_BUFFER_COUNT;
+    m_current_buffer = (m_current_buffer + 1) % MAX_BUFFER_COUNT;
     cursor = 0;
     MapType = D3D11_MAP_WRITE_DISCARD;
   }
 
-  m_vertexDrawOffset = cursor;
+  *out_base_vertex = vertex_stride > 0 ? (cursor / vertex_stride) : 0;
-  m_indexDrawOffset = cursor + vertexBufferSize;
+  *out_base_index = (cursor + vertexBufferSize) / sizeof(u16);
 
-  D3D::context->Map(m_buffers[m_currentBuffer], 0, MapType, 0, &map);
+  D3D::context->Map(m_buffers[m_current_buffer], 0, MapType, 0, &map);
   u8* mappedData = reinterpret_cast<u8*>(map.pData);
-  memcpy(mappedData + m_vertexDrawOffset, m_base_buffer_pointer, vertexBufferSize);
+  if (vertexBufferSize > 0)
-  memcpy(mappedData + m_indexDrawOffset, GetIndexBuffer(), indexBufferSize);
+    std::memcpy(mappedData + cursor, m_base_buffer_pointer, vertexBufferSize);
-  D3D::context->Unmap(m_buffers[m_currentBuffer], 0);
+  if (indexBufferSize > 0)
+    std::memcpy(mappedData + cursor + vertexBufferSize, m_staging_index_buffer.data(),
+                indexBufferSize);
+  D3D::context->Unmap(m_buffers[m_current_buffer], 0);
 
-  m_bufferCursor = cursor + totalBufferSize;
+  m_buffer_cursor = cursor + totalBufferSize;
 
   ADDSTAT(stats.thisFrame.bytesVertexStreamed, vertexBufferSize);
   ADDSTAT(stats.thisFrame.bytesIndexStreamed, indexBufferSize);
+
+  D3D::stateman->SetVertexBuffer(m_buffers[m_current_buffer], vertex_stride, 0);
+  D3D::stateman->SetIndexBuffer(m_buffers[m_current_buffer]);
 }
 
-void VertexManager::Draw(u32 stride)
+void VertexManager::UploadConstants()
 {
-  u32 indices = IndexGenerator::GetIndexLen();
+  if (VertexShaderManager::dirty)
+  {
+    UpdateConstantBuffer(m_vertex_constant_buffer, &VertexShaderManager::constants,
+                         sizeof(VertexShaderConstants));
+    VertexShaderManager::dirty = false;
+  }
+  if (GeometryShaderManager::dirty)
+  {
+    UpdateConstantBuffer(m_geometry_constant_buffer, &GeometryShaderManager::constants,
+                         sizeof(GeometryShaderConstants));
+    GeometryShaderManager::dirty = false;
+  }
+  if (PixelShaderManager::dirty)
+  {
+    UpdateConstantBuffer(m_pixel_constant_buffer, &PixelShaderManager::constants,
+                         sizeof(PixelShaderConstants));
+    PixelShaderManager::dirty = false;
+  }
 
-  D3D::stateman->SetVertexBuffer(m_buffers[m_currentBuffer], stride, 0);
+  D3D::stateman->SetPixelConstants(m_pixel_constant_buffer, g_ActiveConfig.bEnablePixelLighting ?
-  D3D::stateman->SetIndexBuffer(m_buffers[m_currentBuffer]);
+                                                                m_vertex_constant_buffer :
-
+                                                                nullptr);
-  u32 baseVertex = m_vertexDrawOffset / stride;
+  D3D::stateman->SetVertexConstants(m_vertex_constant_buffer);
-  u32 startIndex = m_indexDrawOffset / sizeof(u16);
+  D3D::stateman->SetGeometryConstants(m_geometry_constant_buffer);
-
-  D3D::stateman->Apply();
-  D3D::context->DrawIndexed(indices, startIndex, baseVertex);
-
-  INCSTAT(stats.thisFrame.numDrawCalls);
 }
 
-void VertexManager::vFlush()
+void VertexManager::DrawCurrentBatch(u32 base_index, u32 num_indices, u32 base_vertex)
 {
-  u32 stride = VertexLoaderManager::GetCurrentVertexFormat()->GetVertexStride();
-  PrepareDrawBuffers(stride);
-
-  if (!m_current_pipeline_object)
-    return;
-
   FramebufferManager::SetIntegerEFBRenderTarget(
       m_current_pipeline_config.blending_state.logicopenable);
+
   if (g_ActiveConfig.backend_info.bSupportsBBox && BoundingBox::active)
   {
     D3D::context->OMSetRenderTargetsAndUnorderedAccessViews(
@@ -151,21 +212,7 @@ void VertexManager::vFlush()
         nullptr);
   }
 
-  g_renderer->SetPipeline(m_current_pipeline_object);
+  D3D::stateman->Apply();
-
+  D3D::context->DrawIndexed(num_indices, base_index, base_vertex);
-  ID3D11Buffer* vertexConstants = VertexShaderCache::GetConstantBuffer();
-  D3D::stateman->SetPixelConstants(PixelShaderCache::GetConstantBuffer(),
-                                   g_ActiveConfig.bEnablePixelLighting ? vertexConstants : nullptr);
-  D3D::stateman->SetVertexConstants(vertexConstants);
-  D3D::stateman->SetGeometryConstants(GeometryShaderCache::GetConstantBuffer());
-
-  Draw(stride);
 }
-
+}  // namespace DX11
-void VertexManager::ResetBuffer(u32 stride)
-{
-  m_cur_buffer_pointer = m_base_buffer_pointer;
-  IndexGenerator::Start(GetIndexBuffer());
-}
-
-}  // namespace

Source/Core/VideoBackends/D3D/VertexManager.h

@@ -42,32 +42,32 @@ public:
   std::unique_ptr<NativeVertexFormat>
   CreateNativeVertexFormat(const PortableVertexDeclaration& vtx_decl) override;
 
-  void CreateDeviceObjects() override;
+  void UploadUtilityUniforms(const void* uniforms, u32 uniforms_size) override;
-  void DestroyDeviceObjects() override;
 
 protected:
-  void ResetBuffer(u32 stride) override;
+  void CreateDeviceObjects() override;
-  u16* GetIndexBuffer() { return &LocalIBuffer[0]; }
+  void DestroyDeviceObjects() override;
+  void ResetBuffer(u32 vertex_stride, bool cull_all) override;
+  void CommitBuffer(u32 num_vertices, u32 vertex_stride, u32 num_indices, u32* out_base_vertex,
+                    u32* out_base_index) override;
+  void UploadConstants() override;
+  void DrawCurrentBatch(u32 base_index, u32 num_indices, u32 base_vertex) override;
 
 private:
-  void PrepareDrawBuffers(u32 stride);
-  void Draw(u32 stride);
-  // temp
-  void vFlush() override;
-
-  u32 m_vertexDrawOffset;
-  u32 m_indexDrawOffset;
-  u32 m_currentBuffer;
-  u32 m_bufferCursor;
-
   enum
   {
     MAX_BUFFER_COUNT = 2
   };
-  ID3D11Buffer* m_buffers[MAX_BUFFER_COUNT];
+  ID3D11Buffer* m_buffers[MAX_BUFFER_COUNT] = {};
+  u32 m_current_buffer = 0;
+  u32 m_buffer_cursor = 0;
 
-  std::vector<u8> LocalVBuffer;
+  std::vector<u8> m_staging_vertex_buffer;
-  std::vector<u16> LocalIBuffer;
+  std::vector<u16> m_staging_index_buffer;
+
+  ID3D11Buffer* m_vertex_constant_buffer = nullptr;
+  ID3D11Buffer* m_geometry_constant_buffer = nullptr;
+  ID3D11Buffer* m_pixel_constant_buffer = nullptr;
 };
 
-}  // namespace
+}  // namespace DX11

Source/Core/VideoBackends/D3D/VertexShaderCache.cpp

@@ -4,7 +4,6 @@
 
 #include <string>
 
-#include "Common/Align.h"
 #include "Common/CommonTypes.h"
 #include "Common/FileUtil.h"
 #include "Common/MsgHandler.h"
@@ -23,7 +22,6 @@
 #include "VideoCommon/UberShaderVertex.h"
 #include "VideoCommon/VertexLoaderManager.h"
 #include "VideoCommon/VertexShaderGen.h"
-#include "VideoCommon/VertexShaderManager.h"
 
 namespace DX11
 {
@@ -49,25 +47,6 @@ ID3D11InputLayout* VertexShaderCache::GetClearInputLayout()
   return ClearLayout;
 }
 
-ID3D11Buffer* vscbuf = nullptr;
-
-ID3D11Buffer*& VertexShaderCache::GetConstantBuffer()
-{
-  // TODO: divide the global variables of the generated shaders into about 5 constant buffers to
-  // speed this up
-  if (VertexShaderManager::dirty)
-  {
-    D3D11_MAPPED_SUBRESOURCE map;
-    D3D::context->Map(vscbuf, 0, D3D11_MAP_WRITE_DISCARD, 0, &map);
-    memcpy(map.pData, &VertexShaderManager::constants, sizeof(VertexShaderConstants));
-    D3D::context->Unmap(vscbuf, 0);
-    VertexShaderManager::dirty = false;
-
-    ADDSTAT(stats.thisFrame.bytesUniformStreamed, sizeof(VertexShaderConstants));
-  }
-  return vscbuf;
-}
-
 // this class will load the precompiled shaders into our cache
 template <typename UidType>
 class VertexShaderCacheInserter : public LinearDiskCacheReader<UidType, u8>
@@ -121,14 +100,6 @@ void VertexShaderCache::Init()
       {"COLOR", 0, DXGI_FORMAT_R8G8B8A8_UNORM, 0, 12, D3D11_INPUT_PER_VERTEX_DATA, 0},
   };
 
-  unsigned int cbsize = Common::AlignUp(static_cast<unsigned int>(sizeof(VertexShaderConstants)),
-                                        16);  // must be a multiple of 16
-  D3D11_BUFFER_DESC cbdesc = CD3D11_BUFFER_DESC(cbsize, D3D11_BIND_CONSTANT_BUFFER,
-                                                D3D11_USAGE_DYNAMIC, D3D11_CPU_ACCESS_WRITE);
-  HRESULT hr = D3D::device->CreateBuffer(&cbdesc, nullptr, &vscbuf);
-  CHECK(hr == S_OK, "Create vertex shader constant buffer (size=%u)", cbsize);
-  D3D::SetDebugObjectName(vscbuf, "vertex shader constant buffer used to emulate the GX pipeline");
-
   D3DBlob* blob;
   D3D::CompileVertexShader(simple_shader_code, &blob);
   D3D::device->CreateInputLayout(simpleelems, 2, blob->Data(), blob->Size(), &SimpleLayout);
@@ -156,8 +127,6 @@ void VertexShaderCache::Init()
 
 void VertexShaderCache::Shutdown()
 {
-  SAFE_RELEASE(vscbuf);
-
   SAFE_RELEASE(SimpleVertexShader);
   SAFE_RELEASE(ClearVertexShader);
 

Source/Core/VideoBackends/D3D/VertexShaderCache.h

@@ -23,8 +23,6 @@ public:
   static void Init();
   static void Shutdown();
 
-  static ID3D11Buffer*& GetConstantBuffer();
-
   static ID3D11VertexShader* GetSimpleVertexShader();
   static ID3D11VertexShader* GetClearVertexShader();
   static ID3D11InputLayout* GetSimpleInputLayout();

Source/Core/VideoBackends/Null/VertexManager.cpp

@@ -22,6 +22,10 @@ VertexManager::CreateNativeVertexFormat(const PortableVertexDeclaration& vtx_dec
   return std::make_unique<NullNativeVertexFormat>(vtx_decl);
 }
 
+void VertexManager::UploadUtilityUniforms(const void* uniforms, u32 uniforms_size)
+{
+}
+
 VertexManager::VertexManager() : m_local_v_buffer(MAXVBUFFERSIZE), m_local_i_buffer(MAXIBUFFERSIZE)
 {
 }
@@ -30,15 +34,24 @@ VertexManager::~VertexManager()
 {
 }
 
-void VertexManager::ResetBuffer(u32 stride)
+void VertexManager::ResetBuffer(u32 vertex_stride, bool cull_all)
 {
   m_cur_buffer_pointer = m_base_buffer_pointer = m_local_v_buffer.data();
   m_end_buffer_pointer = m_cur_buffer_pointer + m_local_v_buffer.size();
   IndexGenerator::Start(&m_local_i_buffer[0]);
 }
 
-void VertexManager::vFlush()
+void VertexManager::CommitBuffer(u32 num_vertices, u32 vertex_stride, u32 num_indices,
+                                 u32* out_base_vertex, u32* out_base_index)
 {
 }
 
-}  // namespace
+void VertexManager::UploadConstants()
+{
+}
+
+void VertexManager::DrawCurrentBatch(u32 base_index, u32 num_indices, u32 base_vertex)
+{
+}
+
+}  // namespace Null

Source/Core/VideoBackends/Null/VertexManager.h

@@ -20,11 +20,16 @@ public:
   std::unique_ptr<NativeVertexFormat>
   CreateNativeVertexFormat(const PortableVertexDeclaration& vtx_decl) override;
 
+  void UploadUtilityUniforms(const void* uniforms, u32 uniforms_size) override;
+
 protected:
-  void ResetBuffer(u32 stride) override;
+  void ResetBuffer(u32 vertex_stride, bool cull_all) override;
+  void CommitBuffer(u32 num_vertices, u32 vertex_stride, u32 num_indices, u32* out_base_vertex,
+                    u32* out_base_index) override;
+  void UploadConstants() override;
+  void DrawCurrentBatch(u32 base_index, u32 num_indices, u32 base_vertex) override;
 
 private:
-  void vFlush() override;
   std::vector<u8> m_local_v_buffer;
   std::vector<u16> m_local_i_buffer;
 };

Source/Core/VideoBackends/OGL/ProgramShaderCache.cpp

@@ -259,6 +259,21 @@ void ProgramShaderCache::UploadConstants()
   }
 }
 
+void ProgramShaderCache::UploadConstants(const void* data, u32 data_size)
+{
+  // allocate and copy
+  const u32 alloc_size = Common::AlignUp(data_size, s_ubo_align);
+  auto buffer = s_buffer->Map(alloc_size, s_ubo_align);
+  std::memcpy(buffer.first, data, data_size);
+  s_buffer->Unmap(alloc_size);
+
+  // bind the same sub-buffer to all stages
+  for (u32 index = 1; index <= 3; index++)
+    glBindBufferRange(GL_UNIFORM_BUFFER, index, s_buffer->m_buffer, buffer.second, data_size);
+
+  ADDSTAT(stats.thisFrame.bytesUniformStreamed, data_size);
+}
+
 bool ProgramShaderCache::CompileShader(SHADER& shader, const std::string& vcode,
                                        const std::string& pcode, const std::string& gcode)
 {
@@ -539,6 +554,11 @@ void ProgramShaderCache::BindVertexFormat(const GLVertexFormat* vertex_format)
   s_last_VAO = new_VAO;
 }
 
+bool ProgramShaderCache::IsValidVertexFormatBound()
+{
+  return s_last_VAO != 0 && s_last_VAO != s_attributeless_VAO;
+}
+
 void ProgramShaderCache::InvalidateVertexFormat()
 {
   s_last_VAO = 0;

Source/Core/VideoBackends/OGL/ProgramShaderCache.h

@@ -69,6 +69,7 @@ class ProgramShaderCache
 {
 public:
   static void BindVertexFormat(const GLVertexFormat* vertex_format);
+  static bool IsValidVertexFormatBound();
   static void InvalidateVertexFormat();
   static void InvalidateLastProgram();
 
@@ -83,6 +84,7 @@ public:
   static u32 GetUniformBufferAlignment();
   static void InvalidateConstants();
   static void UploadConstants();
+  static void UploadConstants(const void* data, u32 data_size);
 
   static void Init();
   static void Shutdown();

Source/Core/VideoBackends/OGL/Render.cpp

@@ -1184,6 +1184,27 @@ void Renderer::SetViewport(float x, float y, float width, float height, float ne
   glDepthRangef(near_depth, far_depth);
 }
 
+void Renderer::Draw(u32 base_vertex, u32 num_vertices)
+{
+  glDrawArrays(static_cast<const OGLPipeline*>(m_graphics_pipeline)->GetGLPrimitive(), base_vertex,
+               num_vertices);
+}
+
+void Renderer::DrawIndexed(u32 base_index, u32 num_indices, u32 base_vertex)
+{
+  if (g_ogl_config.bSupportsGLBaseVertex)
+  {
+    glDrawElementsBaseVertex(static_cast<const OGLPipeline*>(m_graphics_pipeline)->GetGLPrimitive(),
+                             num_indices, GL_UNSIGNED_SHORT,
+                             static_cast<u16*>(nullptr) + base_index, base_vertex);
+  }
+  else
+  {
+    glDrawElements(static_cast<const OGLPipeline*>(m_graphics_pipeline)->GetGLPrimitive(),
+                   num_indices, GL_UNSIGNED_SHORT, static_cast<u16*>(nullptr) + base_index);
+  }
+}
+
 void Renderer::ClearScreen(const EFBRectangle& rc, bool colorEnable, bool alphaEnable, bool zEnable,
                            u32 color, u32 z)
 {
@@ -1675,54 +1696,6 @@ void Renderer::SetInterlacingMode()
   // TODO
 }
 
-void Renderer::DrawUtilityPipeline(const void* uniforms, u32 uniforms_size, const void* vertices,
-                                   u32 vertex_stride, u32 num_vertices)
-{
-  // Copy in uniforms.
-  if (uniforms_size > 0)
-    UploadUtilityUniforms(uniforms, uniforms_size);
-
-  // Draw from base index if there is vertex data.
-  if (vertices)
-  {
-    StreamBuffer* vbuf = static_cast<VertexManager*>(g_vertex_manager.get())->GetVertexBuffer();
-    auto buf = vbuf->Map(vertex_stride * num_vertices, vertex_stride);
-    std::memcpy(buf.first, vertices, vertex_stride * num_vertices);
-    vbuf->Unmap(vertex_stride * num_vertices);
-    glDrawArrays(m_graphics_pipeline->GetGLPrimitive(), buf.second / vertex_stride, num_vertices);
-  }
-  else
-  {
-    glDrawArrays(m_graphics_pipeline->GetGLPrimitive(), 0, num_vertices);
-  }
-}
-
-void Renderer::UploadUtilityUniforms(const void* uniforms, u32 uniforms_size)
-{
-  DEBUG_ASSERT(uniforms_size > 0);
-
-  auto buf = ProgramShaderCache::GetUniformBuffer()->Map(
-      uniforms_size, ProgramShaderCache::GetUniformBufferAlignment());
-  std::memcpy(buf.first, uniforms, uniforms_size);
-  ProgramShaderCache::GetUniformBuffer()->Unmap(uniforms_size);
-  glBindBufferRange(GL_UNIFORM_BUFFER, 1, ProgramShaderCache::GetUniformBuffer()->m_buffer,
-                    buf.second, uniforms_size);
-
-  // This is rather horrible, but because of how the UBOs are bound, this forces it to rebind.
-  ProgramShaderCache::InvalidateConstants();
-}
-
-void Renderer::DispatchComputeShader(const AbstractShader* shader, const void* uniforms,
-                                     u32 uniforms_size, u32 groups_x, u32 groups_y, u32 groups_z)
-{
-  glUseProgram(static_cast<const OGLShader*>(shader)->GetGLComputeProgramID());
-  if (uniforms_size > 0)
-    UploadUtilityUniforms(uniforms, uniforms_size);
-
-  glDispatchCompute(groups_x, groups_y, groups_z);
-  ProgramShaderCache::InvalidateLastProgram();
-}
-
 std::unique_ptr<VideoCommon::AsyncShaderCompiler> Renderer::CreateAsyncShaderCompiler()
 {
   return std::make_unique<SharedContextAsyncShaderCompiler>();

Source/Core/VideoBackends/OGL/Render.h

@@ -116,6 +116,8 @@ public:
   void SetInterlacingMode() override;
   void SetViewport(float x, float y, float width, float height, float near_depth,
                    float far_depth) override;
+  void Draw(u32 base_vertex, u32 num_vertices) override;
+  void DrawIndexed(u32 base_index, u32 num_indices, u32 base_vertex) override;
 
   void RenderText(const std::string& text, int left, int top, u32 color) override;
 
@@ -137,18 +139,14 @@ public:
 
   void ReinterpretPixelData(unsigned int convtype) override;
 
-  void DrawUtilityPipeline(const void* uniforms, u32 uniforms_size, const void* vertices,
-                           u32 vertex_stride, u32 num_vertices) override;
-
-  void DispatchComputeShader(const AbstractShader* shader, const void* uniforms, u32 uniforms_size,
-                             u32 groups_x, u32 groups_y, u32 groups_z) override;
-
   std::unique_ptr<VideoCommon::AsyncShaderCompiler> CreateAsyncShaderCompiler() override;
 
   // Only call methods from this on the GPU thread.
   GLContext* GetMainGLContext() const { return m_main_gl_context.get(); }
   bool IsGLES() const { return m_main_gl_context->IsGLES(); }
 
+  const OGLPipeline* GetCurrentGraphicsPipeline() const { return m_graphics_pipeline; }
+
 private:
   void UpdateEFBCache(EFBAccessType type, u32 cacheRectIdx, const EFBRectangle& efbPixelRc,
                       const TargetRectangle& targetPixelRc, const void* data);
@@ -165,7 +163,6 @@ private:
   void ApplyBlendingState(const BlendingState state, bool force = false);
   void ApplyRasterizationState(const RasterizationState state, bool force = false);
   void ApplyDepthState(const DepthState state, bool force = false);
-  void UploadUtilityUniforms(const void* uniforms, u32 uniforms_size);
 
   std::unique_ptr<GLContext> m_main_gl_context;
   std::array<const AbstractTexture*, 8> m_bound_textures{};

Source/Core/VideoBackends/OGL/StreamBuffer.h

@@ -19,6 +19,8 @@ public:
   static std::unique_ptr<StreamBuffer> Create(u32 type, u32 size);
   virtual ~StreamBuffer();
 
+  u32 GetCurrentOffset() const { return m_iterator; }
+
   /* This mapping function will return a pair of:
    * - the pointer to the mapped buffer
    * - the offset into the real GPU buffer (always multiple of stride)

Source/Core/VideoBackends/OGL/VertexManager.cpp

@@ -15,6 +15,7 @@
 #include "Common/StringUtil.h"
 
 #include "VideoBackends/OGL/BoundingBox.h"
+#include "VideoBackends/OGL/OGLPipeline.h"
 #include "VideoBackends/OGL/ProgramShaderCache.h"
 #include "VideoBackends/OGL/Render.h"
 #include "VideoBackends/OGL/StreamBuffer.h"
@@ -31,11 +32,6 @@ namespace OGL
 const u32 MAX_IBUFFER_SIZE = 2 * 1024 * 1024;
 const u32 MAX_VBUFFER_SIZE = 32 * 1024 * 1024;
 
-static std::unique_ptr<StreamBuffer> s_vertexBuffer;
-static std::unique_ptr<StreamBuffer> s_indexBuffer;
-static size_t s_baseVertex;
-static size_t s_index_offset;
-
 VertexManager::VertexManager() : m_cpu_v_buffer(MAX_VBUFFER_SIZE), m_cpu_i_buffer(MAX_IBUFFER_SIZE)
 {
   CreateDeviceObjects();
@@ -48,58 +44,45 @@ VertexManager::~VertexManager()
 
 void VertexManager::CreateDeviceObjects()
 {
-  s_vertexBuffer = StreamBuffer::Create(GL_ARRAY_BUFFER, MAX_VBUFFER_SIZE);
+  m_vertex_buffer = StreamBuffer::Create(GL_ARRAY_BUFFER, MAX_VBUFFER_SIZE);
-  m_vertex_buffers = s_vertexBuffer->m_buffer;
+  m_index_buffer = StreamBuffer::Create(GL_ELEMENT_ARRAY_BUFFER, MAX_IBUFFER_SIZE);
-
-  s_indexBuffer = StreamBuffer::Create(GL_ELEMENT_ARRAY_BUFFER, MAX_IBUFFER_SIZE);
-  m_index_buffers = s_indexBuffer->m_buffer;
 }
 
 void VertexManager::DestroyDeviceObjects()
 {
-  s_vertexBuffer.reset();
+  m_vertex_buffer.reset();
-  s_indexBuffer.reset();
+  m_index_buffer.reset();
 }
 
-StreamBuffer* VertexManager::GetVertexBuffer() const
+void VertexManager::UploadUtilityUniforms(const void* uniforms, u32 uniforms_size)
 {
-  return s_vertexBuffer.get();
+  ProgramShaderCache::InvalidateConstants();
-}
+  ProgramShaderCache::UploadConstants(uniforms, uniforms_size);
-
-OGL::StreamBuffer* VertexManager::GetIndexBuffer() const
-{
-  return s_indexBuffer.get();
 }
 
 GLuint VertexManager::GetVertexBufferHandle() const
 {
-  return m_vertex_buffers;
+  return m_vertex_buffer->m_buffer;
 }
 
 GLuint VertexManager::GetIndexBufferHandle() const
 {
-  return m_index_buffers;
+  return m_index_buffer->m_buffer;
 }
 
-void VertexManager::PrepareDrawBuffers(u32 stride)
+static void CheckBufferBinding()
 {
-  u32 vertex_data_size = IndexGenerator::GetNumVerts() * stride;
-  u32 index_data_size = IndexGenerator::GetIndexLen() * sizeof(u16);
-
   // The index buffer is part of the VAO state, therefore we need to bind it first.
-  const GLVertexFormat* vertex_format =
+  if (!ProgramShaderCache::IsValidVertexFormatBound())
-      static_cast<GLVertexFormat*>(VertexLoaderManager::GetCurrentVertexFormat());
+  {
-  ProgramShaderCache::BindVertexFormat(vertex_format);
+    ProgramShaderCache::BindVertexFormat(
-  s_vertexBuffer->Unmap(vertex_data_size);
+        static_cast<GLVertexFormat*>(VertexLoaderManager::GetCurrentVertexFormat()));
-  s_indexBuffer->Unmap(index_data_size);
+  }
-
-  ADDSTAT(stats.thisFrame.bytesVertexStreamed, vertex_data_size);
-  ADDSTAT(stats.thisFrame.bytesIndexStreamed, index_data_size);
 }
 
-void VertexManager::ResetBuffer(u32 stride)
+void VertexManager::ResetBuffer(u32 vertex_stride, bool cull_all)
 {
-  if (m_cull_all)
+  if (cull_all)
   {
     // This buffer isn't getting sent to the GPU. Just allocate it on the cpu.
     m_cur_buffer_pointer = m_base_buffer_pointer = m_cpu_v_buffer.data();
@@ -109,68 +92,41 @@ void VertexManager::ResetBuffer(u32 stride)
   }
   else
   {
-    // The index buffer is part of the VAO state, therefore we need to bind it first.
+    CheckBufferBinding();
-    const GLVertexFormat* vertex_format =
-        static_cast<GLVertexFormat*>(VertexLoaderManager::GetCurrentVertexFormat());
-    ProgramShaderCache::BindVertexFormat(vertex_format);
 
-    auto buffer = s_vertexBuffer->Map(MAXVBUFFERSIZE, stride);
+    auto buffer = m_vertex_buffer->Map(MAXVBUFFERSIZE, vertex_stride);
     m_cur_buffer_pointer = m_base_buffer_pointer = buffer.first;
     m_end_buffer_pointer = buffer.first + MAXVBUFFERSIZE;
-    s_baseVertex = buffer.second / stride;
 
-    buffer = s_indexBuffer->Map(MAXIBUFFERSIZE * sizeof(u16));
+    buffer = m_index_buffer->Map(MAXIBUFFERSIZE * sizeof(u16));
     IndexGenerator::Start((u16*)buffer.first);
-    s_index_offset = buffer.second;
   }
 }
 
-void VertexManager::Draw(u32 stride)
+void VertexManager::CommitBuffer(u32 num_vertices, u32 vertex_stride, u32 num_indices,
+                                 u32* out_base_vertex, u32* out_base_index)
 {
-  u32 index_size = IndexGenerator::GetIndexLen();
+  u32 vertex_data_size = num_vertices * vertex_stride;
-  u32 max_index = IndexGenerator::GetNumVerts();
+  u32 index_data_size = num_indices * sizeof(u16);
-  GLenum primitive_mode = 0;
 
-  switch (m_current_primitive_type)
+  *out_base_vertex = vertex_stride > 0 ? (m_vertex_buffer->GetCurrentOffset() / vertex_stride) : 0;
-  {
+  *out_base_index = m_index_buffer->GetCurrentOffset() / sizeof(u16);
-  case PrimitiveType::Points:
-    primitive_mode = GL_POINTS;
-    break;
-  case PrimitiveType::Lines:
-    primitive_mode = GL_LINES;
-    break;
-  case PrimitiveType::Triangles:
-    primitive_mode = GL_TRIANGLES;
-    break;
-  case PrimitiveType::TriangleStrip:
-    primitive_mode = GL_TRIANGLE_STRIP;
-    break;
-  }
 
-  if (g_ogl_config.bSupportsGLBaseVertex)
+  CheckBufferBinding();
-  {
+  m_vertex_buffer->Unmap(vertex_data_size);
-    glDrawRangeElementsBaseVertex(primitive_mode, 0, max_index, index_size, GL_UNSIGNED_SHORT,
+  m_index_buffer->Unmap(index_data_size);
-                                  (u8*)nullptr + s_index_offset, (GLint)s_baseVertex);
-  }
-  else
-  {
-    glDrawRangeElements(primitive_mode, 0, max_index, index_size, GL_UNSIGNED_SHORT,
-                        (u8*)nullptr + s_index_offset);
-  }
 
-  INCSTAT(stats.thisFrame.numDrawCalls);
+  ADDSTAT(stats.thisFrame.bytesVertexStreamed, vertex_data_size);
+  ADDSTAT(stats.thisFrame.bytesIndexStreamed, index_data_size);
 }
 
-void VertexManager::vFlush()
+void VertexManager::UploadConstants()
 {
-  GLVertexFormat* nativeVertexFmt = (GLVertexFormat*)VertexLoaderManager::GetCurrentVertexFormat();
-  u32 stride = nativeVertexFmt->GetVertexStride();
-
-  PrepareDrawBuffers(stride);
-
-  // upload global constants
   ProgramShaderCache::UploadConstants();
+}
 
+void VertexManager::DrawCurrentBatch(u32 base_index, u32 num_indices, u32 base_vertex)
+{
   if (::BoundingBox::active && !g_Config.BBoxUseFragmentShaderImplementation())
   {
     glEnable(GL_STENCIL_TEST);
@@ -178,8 +134,8 @@ void VertexManager::vFlush()
 
   if (m_current_pipeline_object)
   {
-    g_renderer->SetPipeline(m_current_pipeline_object);
+    static_cast<Renderer*>(g_renderer.get())->SetPipeline(m_current_pipeline_object);
-    Draw(stride);
+    static_cast<Renderer*>(g_renderer.get())->DrawIndexed(base_index, num_indices, base_vertex);
   }
 
   if (::BoundingBox::active && !g_Config.BBoxUseFragmentShaderImplementation())
@@ -191,5 +147,4 @@ void VertexManager::vFlush()
   g_Config.iSaveTargetId++;
   ClearEFBCache();
 }
-
+}  // namespace OGL
-}  // namespace

Source/Core/VideoBackends/OGL/VertexManager.h

@@ -35,27 +35,26 @@ public:
   std::unique_ptr<NativeVertexFormat>
   CreateNativeVertexFormat(const PortableVertexDeclaration& vtx_decl) override;
 
-  void CreateDeviceObjects() override;
+  void UploadUtilityUniforms(const void* uniforms, u32 uniforms_size) override;
-  void DestroyDeviceObjects() override;
 
-  StreamBuffer* GetVertexBuffer() const;
-  StreamBuffer* GetIndexBuffer() const;
   GLuint GetVertexBufferHandle() const;
   GLuint GetIndexBufferHandle() const;
 
 protected:
-  void ResetBuffer(u32 stride) override;
+  void CreateDeviceObjects() override;
+  void DestroyDeviceObjects() override;
+  void ResetBuffer(u32 vertex_stride, bool cull_all) override;
+  void CommitBuffer(u32 num_vertices, u32 vertex_stride, u32 num_indices, u32* out_base_vertex,
+                    u32* out_base_index) override;
+  void UploadConstants() override;
+  void DrawCurrentBatch(u32 base_index, u32 num_indices, u32 base_vertex) override;
 
 private:
-  void Draw(u32 stride);
+  std::unique_ptr<StreamBuffer> m_vertex_buffer;
-  void vFlush() override;
+  std::unique_ptr<StreamBuffer> m_index_buffer;
-  void PrepareDrawBuffers(u32 stride);
-
-  GLuint m_vertex_buffers;
-  GLuint m_index_buffers;
 
   // Alternative buffers in CPU memory for primatives we are going to discard.
   std::vector<u8> m_cpu_v_buffer;
   std::vector<u16> m_cpu_i_buffer;
 };
-}
+}  // namespace OGL

Source/Core/VideoBackends/Software/SWVertexLoader.cpp

@@ -48,14 +48,29 @@ SWVertexLoader::~SWVertexLoader()
 {
 }
 
-void SWVertexLoader::ResetBuffer(u32 stride)
+void SWVertexLoader::UploadUtilityUniforms(const void* uniforms, u32 uniforms_size)
+{
+}
+
+void SWVertexLoader::ResetBuffer(u32 vertex_stride, bool cull_all)
 {
   m_cur_buffer_pointer = m_base_buffer_pointer = m_local_vertex_buffer.data();
   m_end_buffer_pointer = m_cur_buffer_pointer + m_local_vertex_buffer.size();
   IndexGenerator::Start(m_local_index_buffer.data());
 }
 
-void SWVertexLoader::vFlush()
+void SWVertexLoader::CommitBuffer(u32 num_vertices, u32 vertex_stride, u32 num_indices,
+                                  u32* out_base_vertex, u32* out_base_index)
+{
+  *out_base_vertex = 0;
+  *out_base_index = 0;
+}
+
+void SWVertexLoader::UploadConstants()
+{
+}
+
+void SWVertexLoader::DrawCurrentBatch(u32 base_index, u32 num_indices, u32 base_vertex)
 {
   DebugUtil::OnObjectBegin();
 

Source/Core/VideoBackends/Software/SWVertexLoader.h

@@ -23,9 +23,14 @@ public:
   std::unique_ptr<NativeVertexFormat>
   CreateNativeVertexFormat(const PortableVertexDeclaration& vdec) override;
 
-private:
+  void UploadUtilityUniforms(const void* uniforms, u32 uniforms_size) override;
-  void ResetBuffer(u32 stride) override;
+
-  void vFlush() override;
+protected:
+  void ResetBuffer(u32 vertex_stride, bool cull_all) override;
+  void CommitBuffer(u32 num_vertices, u32 vertex_stride, u32 num_indices, u32* out_base_vertex,
+                    u32* out_base_index) override;
+  void UploadConstants() override;
+  void DrawCurrentBatch(u32 base_index, u32 num_indices, u32 base_vertex) override;
 
   void SetFormat(u8 attributeIndex, u8 primitiveType);
   void ParseVertex(const PortableVertexDeclaration& vdec, int index);

Source/Core/VideoBackends/Vulkan/Renderer.cpp

@@ -259,158 +259,6 @@ void Renderer::SetPipeline(const AbstractPipeline* pipeline)
   StateTracker::GetInstance()->SetPipeline(static_cast<const VKPipeline*>(pipeline));
 }
 
-void Renderer::DrawUtilityPipeline(const void* uniforms, u32 uniforms_size, const void* vertices,
-                                   u32 vertex_stride, u32 num_vertices)
-{
-  // Binding the utility pipeline layout breaks the standard layout.
-  StateTracker::GetInstance()->SetPendingRebind();
-
-  // Upload uniforms.
-  VkBuffer uniform_buffer = g_object_cache->GetUtilityShaderUniformBuffer()->GetBuffer();
-  u32 uniform_buffer_offset = 0;
-  if (uniforms_size > 0)
-    std::tie(uniform_buffer, uniform_buffer_offset) =
-        UpdateUtilityUniformBuffer(uniforms, uniforms_size);
-
-  // Upload vertices.
-  VkBuffer vertex_buffer = VK_NULL_HANDLE;
-  VkDeviceSize vertex_buffer_offset = 0;
-  if (vertices)
-  {
-    u32 vertices_size = vertex_stride * num_vertices;
-    StreamBuffer* vbo_buf = g_object_cache->GetUtilityShaderVertexBuffer();
-    if (!vbo_buf->ReserveMemory(vertices_size, vertex_stride))
-    {
-      Util::ExecuteCurrentCommandsAndRestoreState(true);
-      if (!vbo_buf->ReserveMemory(vertices_size, vertex_stride))
-      {
-        PanicAlert("Failed to reserve vertex buffer space for utility draw.");
-        return;
-      }
-    }
-
-    vertex_buffer = vbo_buf->GetBuffer();
-    vertex_buffer_offset = vbo_buf->GetCurrentOffset();
-    std::memcpy(vbo_buf->GetCurrentHostPointer(), vertices, vertices_size);
-    vbo_buf->CommitMemory(vertices_size);
-  }
-
-  // Allocate descriptor sets.
-  std::array<VkDescriptorSet, 2> dsets;
-  dsets[0] = g_command_buffer_mgr->AllocateDescriptorSet(
-      g_object_cache->GetDescriptorSetLayout(DESCRIPTOR_SET_LAYOUT_SINGLE_UNIFORM_BUFFER));
-  dsets[1] = g_command_buffer_mgr->AllocateDescriptorSet(
-      g_object_cache->GetDescriptorSetLayout(DESCRIPTOR_SET_LAYOUT_PIXEL_SHADER_SAMPLERS));
-
-  // Flush first if failed.
-  if (dsets[0] == VK_NULL_HANDLE || dsets[1] == VK_NULL_HANDLE)
-  {
-    Util::ExecuteCurrentCommandsAndRestoreState(true);
-    dsets[0] = g_command_buffer_mgr->AllocateDescriptorSet(
-        g_object_cache->GetDescriptorSetLayout(DESCRIPTOR_SET_LAYOUT_SINGLE_UNIFORM_BUFFER));
-    dsets[1] = g_command_buffer_mgr->AllocateDescriptorSet(
-        g_object_cache->GetDescriptorSetLayout(DESCRIPTOR_SET_LAYOUT_PIXEL_SHADER_SAMPLERS));
-
-    if (dsets[0] == VK_NULL_HANDLE || dsets[1] == VK_NULL_HANDLE)
-    {
-      PanicAlert("Failed to allocate descriptor sets in utility draw.");
-      return;
-    }
-  }
-
-  // Build UBO descriptor set.
-  std::array<VkWriteDescriptorSet, 2> dswrites;
-  VkDescriptorBufferInfo dsbuffer = {uniform_buffer, 0, std::max(uniforms_size, 4u)};
-  dswrites[0] = {VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET,    nullptr, dsets[0],  0,      0, 1,
-                 VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC, nullptr, &dsbuffer, nullptr};
-  dswrites[1] = {VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET,
-                 nullptr,
-                 dsets[1],
-                 0,
-                 0,
-                 NUM_PIXEL_SHADER_SAMPLERS,
-                 VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER,
-                 StateTracker::GetInstance()->GetPSSamplerBindings().data(),
-                 nullptr,
-                 nullptr};
-
-  // Build commands.
-  VkCommandBuffer command_buffer = g_command_buffer_mgr->GetCurrentCommandBuffer();
-  vkCmdBindPipeline(command_buffer, VK_PIPELINE_BIND_POINT_GRAPHICS,
-                    StateTracker::GetInstance()->GetPipeline()->GetVkPipeline());
-  if (vertex_buffer != VK_NULL_HANDLE)
-    vkCmdBindVertexBuffers(command_buffer, 0, 1, &vertex_buffer, &vertex_buffer_offset);
-
-  // Update and bind descriptors.
-  VkPipelineLayout pipeline_layout = g_object_cache->GetPipelineLayout(PIPELINE_LAYOUT_UTILITY);
-  vkUpdateDescriptorSets(g_vulkan_context->GetDevice(), static_cast<u32>(dswrites.size()),
-                         dswrites.data(), 0, nullptr);
-  vkCmdBindDescriptorSets(command_buffer, VK_PIPELINE_BIND_POINT_GRAPHICS, pipeline_layout, 0,
-                          static_cast<u32>(dsets.size()), dsets.data(), 1, &uniform_buffer_offset);
-
-  // Ensure we're in a render pass before drawing, just in case we had to flush.
-  StateTracker::GetInstance()->BeginRenderPass();
-  vkCmdDraw(command_buffer, num_vertices, 1, 0, 0);
-}
-
-void Renderer::DispatchComputeShader(const AbstractShader* shader, const void* uniforms,
-                                     u32 uniforms_size, u32 groups_x, u32 groups_y, u32 groups_z)
-{
-  // Binding the utility pipeline layout breaks the standard layout.
-  StateTracker::GetInstance()->SetPendingRebind();
-  StateTracker::GetInstance()->EndRenderPass();
-
-  // Upload uniforms.
-  VkBuffer uniform_buffer = g_object_cache->GetUtilityShaderUniformBuffer()->GetBuffer();
-  u32 uniform_buffer_offset = 0;
-  if (uniforms_size > 0)
-    std::tie(uniform_buffer, uniform_buffer_offset) =
-        UpdateUtilityUniformBuffer(uniforms, uniforms_size);
-
-  // Flush first if failed.
-  VkDescriptorSet dset = g_command_buffer_mgr->AllocateDescriptorSet(
-      g_object_cache->GetDescriptorSetLayout(DESCRIPTOR_SET_LAYOUT_COMPUTE));
-  if (dset == VK_NULL_HANDLE)
-  {
-    Util::ExecuteCurrentCommandsAndRestoreState(true);
-    dset = g_command_buffer_mgr->AllocateDescriptorSet(
-        g_object_cache->GetDescriptorSetLayout(DESCRIPTOR_SET_LAYOUT_COMPUTE));
-    if (dset == VK_NULL_HANDLE)
-    {
-      PanicAlert("Failed to allocate descriptor sets in utility dispatch.");
-      return;
-    }
-  }
-
-  std::array<VkWriteDescriptorSet, 2> dswrites;
-  VkDescriptorBufferInfo dsbuffer = {uniform_buffer, 0, std::max(uniforms_size, 4u)};
-  dswrites[0] = {VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET,    nullptr, dset,      0,      0, 1,
-                 VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC, nullptr, &dsbuffer, nullptr};
-  dswrites[1] = {VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET,
-                 nullptr,
-                 dset,
-                 1,
-                 0,
-                 NUM_PIXEL_SHADER_SAMPLERS,
-                 VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER,
-                 StateTracker::GetInstance()->GetPSSamplerBindings().data(),
-                 nullptr,
-                 nullptr};
-
-  // TODO: Texel buffers, storage images.
-
-  // Build commands.
-  VkCommandBuffer command_buffer = g_command_buffer_mgr->GetCurrentCommandBuffer();
-  VkPipelineLayout pipeline_layout = g_object_cache->GetPipelineLayout(PIPELINE_LAYOUT_UTILITY);
-  vkCmdBindPipeline(command_buffer, VK_PIPELINE_BIND_POINT_COMPUTE,
-                    static_cast<const VKShader*>(shader)->GetComputePipeline());
-  vkUpdateDescriptorSets(g_vulkan_context->GetDevice(), static_cast<u32>(dswrites.size()),
-                         dswrites.data(), 0, nullptr);
-  vkCmdBindDescriptorSets(command_buffer, VK_PIPELINE_BIND_POINT_COMPUTE, pipeline_layout, 0, 1,
-                          &dset, 1, &uniform_buffer_offset);
-  vkCmdDispatch(command_buffer, groups_x, groups_y, groups_z);
-}
-
 void Renderer::RenderText(const std::string& text, int left, int top, u32 color)
 {
   u32 backbuffer_width = m_swap_chain->GetWidth();
@@ -1185,6 +1033,23 @@ void Renderer::SetViewport(float x, float y, float width, float height, float ne
   StateTracker::GetInstance()->SetViewport(viewport);
 }
 
+void Renderer::Draw(u32 base_vertex, u32 num_vertices)
+{
+  if (StateTracker::GetInstance()->Bind())
+    return;
+
+  vkCmdDraw(g_command_buffer_mgr->GetCurrentCommandBuffer(), num_vertices, 1, base_vertex, 0);
+}
+
+void Renderer::DrawIndexed(u32 base_index, u32 num_indices, u32 base_vertex)
+{
+  if (!StateTracker::GetInstance()->Bind())
+    return;
+
+  vkCmdDrawIndexed(g_command_buffer_mgr->GetCurrentCommandBuffer(), num_indices, 1, base_index,
+                   base_vertex, 0);
+}
+
 void Renderer::RecompileShaders()
 {
   DestroyShaders();

Source/Core/VideoBackends/Vulkan/Renderer.h

@@ -87,11 +87,8 @@ public:
   void SetInterlacingMode() override;
   void SetViewport(float x, float y, float width, float height, float near_depth,
                    float far_depth) override;
-
+  void Draw(u32 base_vertex, u32 num_vertices) override;
-  void DrawUtilityPipeline(const void* uniforms, u32 uniforms_size, const void* vertices,
+  void DrawIndexed(u32 base_index, u32 num_indices, u32 base_vertex) override;
-                           u32 vertex_stride, u32 num_vertices) override;
-  void DispatchComputeShader(const AbstractShader* shader, const void* uniforms, u32 uniforms_size,
-                             u32 groups_x, u32 groups_y, u32 groups_z) override;
 
 private:
   bool CreateSemaphores();

Source/Core/VideoBackends/Vulkan/StateTracker.cpp

@@ -218,6 +218,37 @@ void StateTracker::UpdatePixelShaderConstants()
   PixelShaderManager::dirty = false;
 }
 
+void StateTracker::UpdateConstants(const void* data, u32 data_size)
+{
+  if (!m_uniform_stream_buffer->ReserveMemory(
+          data_size, g_vulkan_context->GetUniformBufferAlignment(), true, true, false))
+  {
+    WARN_LOG(VIDEO, "Executing command buffer while waiting for ext space in uniform buffer");
+    Util::ExecuteCurrentCommandsAndRestoreState(false);
+  }
+
+  for (u32 binding = 0; binding < NUM_UBO_DESCRIPTOR_SET_BINDINGS; binding++)
+  {
+    if (m_bindings.uniform_buffer_bindings[binding].buffer != m_uniform_stream_buffer->GetBuffer())
+    {
+      m_bindings.uniform_buffer_bindings[binding].buffer = m_uniform_stream_buffer->GetBuffer();
+      m_dirty_flags |= DIRTY_FLAG_VS_UBO << binding;
+    }
+    m_bindings.uniform_buffer_offsets[binding] =
+        static_cast<uint32_t>(m_uniform_stream_buffer->GetCurrentOffset());
+  }
+  m_dirty_flags |= DIRTY_FLAG_DYNAMIC_OFFSETS;
+
+  std::memcpy(m_uniform_stream_buffer->GetCurrentHostPointer(), data, data_size);
+  ADDSTAT(stats.thisFrame.bytesUniformStreamed, data_size);
+  m_uniform_stream_buffer->CommitMemory(data_size);
+
+  // Cached data is now out-of-sync.
+  VertexShaderManager::dirty = true;
+  GeometryShaderManager::dirty = true;
+  PixelShaderManager::dirty = true;
+}
+
 bool StateTracker::ReserveConstantStorage()
 {
   // Since we invalidate all constants on command buffer execution, it doesn't matter if this
@@ -473,16 +504,16 @@ bool StateTracker::Bind(bool rebind_all /*= false*/)
   {
     vkCmdBindDescriptorSets(command_buffer, VK_PIPELINE_BIND_POINT_GRAPHICS,
                             m_pipeline->GetVkPipelineLayout(), 0, m_num_active_descriptor_sets,
-                            m_descriptor_sets.data(), NUM_UBO_DESCRIPTOR_SET_BINDINGS,
+                            m_descriptor_sets.data(), m_num_dynamic_offsets,
                             m_bindings.uniform_buffer_offsets.data());
   }
   else if (m_dirty_flags & DIRTY_FLAG_DYNAMIC_OFFSETS)
   {
-    vkCmdBindDescriptorSets(
+    vkCmdBindDescriptorSets(command_buffer, VK_PIPELINE_BIND_POINT_GRAPHICS,
-        command_buffer, VK_PIPELINE_BIND_POINT_GRAPHICS, m_pipeline->GetVkPipelineLayout(),
+                            m_pipeline->GetVkPipelineLayout(),
-        DESCRIPTOR_SET_BIND_POINT_UNIFORM_BUFFERS, 1,
+                            DESCRIPTOR_SET_BIND_POINT_UNIFORM_BUFFERS, 1,
-        &m_descriptor_sets[DESCRIPTOR_SET_BIND_POINT_UNIFORM_BUFFERS],
+                            &m_descriptor_sets[DESCRIPTOR_SET_BIND_POINT_UNIFORM_BUFFERS],
-        NUM_UBO_DESCRIPTOR_SET_BINDINGS, m_bindings.uniform_buffer_offsets.data());
+                            m_num_dynamic_offsets, m_bindings.uniform_buffer_offsets.data());
   }
 
   if (m_dirty_flags & DIRTY_FLAG_VIEWPORT || rebind_all)
@@ -639,6 +670,14 @@ void StateTracker::EndClearRenderPass()
 }
 
 bool StateTracker::UpdateDescriptorSet()
+{
+  if (m_pipeline->GetUsage() == AbstractPipelineUsage::GX)
+    return UpdateGXDescriptorSet();
+  else
+    return UpdateUtilityDescriptorSet();
+}
+
+bool StateTracker::UpdateGXDescriptorSet()
 {
   const size_t MAX_DESCRIPTOR_WRITES = NUM_UBO_DESCRIPTOR_SET_BINDINGS +  // UBO
                                        1 +                                // Samplers
@@ -729,6 +768,50 @@ bool StateTracker::UpdateDescriptorSet()
     vkUpdateDescriptorSets(g_vulkan_context->GetDevice(), num_writes, writes.data(), 0, nullptr);
 
   m_num_active_descriptor_sets = NUM_GX_DRAW_DESCRIPTOR_SETS;
+  m_num_dynamic_offsets = NUM_UBO_DESCRIPTOR_SET_BINDINGS;
+  return true;
+}
+
+bool StateTracker::UpdateUtilityDescriptorSet()
+{
+  // Allocate descriptor sets.
+  m_descriptor_sets[0] = g_command_buffer_mgr->AllocateDescriptorSet(
+      g_object_cache->GetDescriptorSetLayout(DESCRIPTOR_SET_LAYOUT_SINGLE_UNIFORM_BUFFER));
+  m_descriptor_sets[1] = g_command_buffer_mgr->AllocateDescriptorSet(
+      g_object_cache->GetDescriptorSetLayout(DESCRIPTOR_SET_LAYOUT_PIXEL_SHADER_SAMPLERS));
+  if (m_descriptor_sets[0] == VK_NULL_HANDLE || m_descriptor_sets[1] == VK_NULL_HANDLE)
+  {
+    return false;
+  }
+
+  // Build UBO descriptor set.
+  std::array<VkWriteDescriptorSet, 2> dswrites;
+  dswrites[0] = {VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET,
+                 nullptr,
+                 m_descriptor_sets[0],
+                 0,
+                 0,
+                 1,
+                 VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC,
+                 nullptr,
+                 &m_bindings.uniform_buffer_bindings[UBO_DESCRIPTOR_SET_BINDING_VS],
+                 nullptr};
+  dswrites[1] = {VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET,
+                 nullptr,
+                 m_descriptor_sets[1],
+                 0,
+                 0,
+                 NUM_PIXEL_SHADER_SAMPLERS,
+                 VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER,
+                 m_bindings.ps_samplers.data(),
+                 nullptr,
+                 nullptr};
+
+  vkUpdateDescriptorSets(g_vulkan_context->GetDevice(), static_cast<uint32_t>(dswrites.size()),
+                         dswrites.data(), 0, nullptr);
+  m_num_active_descriptor_sets = NUM_UTILITY_DRAW_DESCRIPTOR_SETS;
+  m_num_dynamic_offsets = 1;
+  m_dirty_flags |= DIRTY_FLAG_DESCRIPTOR_SET_BINDING;
   return true;
 }
 

Source/Core/VideoBackends/Vulkan/StateTracker.h

@@ -30,10 +30,6 @@ public:
   static bool CreateInstance();
   static void DestroyInstance();
 
-  const std::array<VkDescriptorImageInfo, NUM_PIXEL_SHADER_SAMPLERS>& GetPSSamplerBindings() const
-  {
-    return m_bindings.ps_samplers;
-  }
   VkFramebuffer GetFramebuffer() const { return m_framebuffer; }
   const VKPipeline* GetPipeline() const { return m_pipeline; }
   void SetVertexBuffer(VkBuffer buffer, VkDeviceSize offset);
@@ -47,6 +43,9 @@ public:
   void UpdateGeometryShaderConstants();
   void UpdatePixelShaderConstants();
 
+  // Updates constants from external data, e.g. utility draws.
+  void UpdateConstants(const void* data, u32 data_size);
+
   void SetTexture(size_t index, VkImageView view);
   void SetSampler(size_t index, VkSampler sampler);
 
@@ -104,7 +103,8 @@ private:
   // Number of descriptor sets for game draws.
   enum
   {
-    NUM_GX_DRAW_DESCRIPTOR_SETS = DESCRIPTOR_SET_BIND_POINT_STORAGE_OR_TEXEL_BUFFER + 1
+    NUM_GX_DRAW_DESCRIPTOR_SETS = DESCRIPTOR_SET_BIND_POINT_STORAGE_OR_TEXEL_BUFFER + 1,
+    NUM_UTILITY_DRAW_DESCRIPTOR_SETS = 2
   };
 
   enum DITRY_FLAG : u32
@@ -133,6 +133,8 @@ private:
   bool IsViewportWithinRenderArea() const;
 
   bool UpdateDescriptorSet();
+  bool UpdateGXDescriptorSet();
+  bool UpdateUtilityDescriptorSet();
 
   // Allocates storage in the uniform buffer of the specified size. If this storage cannot be
   // allocated immediately, the current command buffer will be submitted and all stage's
@@ -167,6 +169,7 @@ private:
   } m_bindings;
   size_t m_uniform_buffer_reserve_size = 0;
   u32 m_num_active_descriptor_sets = 0;
+  u32 m_num_dynamic_offsets = 0;
 
   // rasterization
   VkViewport m_viewport = {0.0f, 0.0f, 1.0f, 1.0f, 0.0f, 1.0f};

Source/Core/VideoBackends/Vulkan/VertexManager.cpp

@@ -69,25 +69,14 @@ VertexManager::CreateNativeVertexFormat(const PortableVertexDeclaration& vtx_dec
   return std::make_unique<VertexFormat>(vtx_decl);
 }
 
-void VertexManager::PrepareDrawBuffers(u32 stride)
+void VertexManager::UploadUtilityUniforms(const void* uniforms, u32 uniforms_size)
 {
-  size_t vertex_data_size = IndexGenerator::GetNumVerts() * stride;
+  StateTracker::GetInstance()->UpdateConstants(uniforms, uniforms_size);
-  size_t index_data_size = IndexGenerator::GetIndexLen() * sizeof(u16);
-
-  m_vertex_stream_buffer->CommitMemory(vertex_data_size);
-  m_index_stream_buffer->CommitMemory(index_data_size);
-
-  ADDSTAT(stats.thisFrame.bytesVertexStreamed, static_cast<int>(vertex_data_size));
-  ADDSTAT(stats.thisFrame.bytesIndexStreamed, static_cast<int>(index_data_size));
-
-  StateTracker::GetInstance()->SetVertexBuffer(m_vertex_stream_buffer->GetBuffer(), 0);
-  StateTracker::GetInstance()->SetIndexBuffer(m_index_stream_buffer->GetBuffer(), 0,
-                                              VK_INDEX_TYPE_UINT16);
 }
 
-void VertexManager::ResetBuffer(u32 stride)
+void VertexManager::ResetBuffer(u32 vertex_stride, bool cull_all)
 {
-  if (m_cull_all)
+  if (cull_all)
   {
     // Not drawing on the gpu, so store in a heap buffer instead
     m_cur_buffer_pointer = m_base_buffer_pointer = m_cpu_vertex_buffer.data();
@@ -97,7 +86,8 @@ void VertexManager::ResetBuffer(u32 stride)
   }
 
   // Attempt to allocate from buffers
-  bool has_vbuffer_allocation = m_vertex_stream_buffer->ReserveMemory(MAXVBUFFERSIZE, stride);
+  bool has_vbuffer_allocation =
+      m_vertex_stream_buffer->ReserveMemory(MAXVBUFFERSIZE, vertex_stride);
   bool has_ibuffer_allocation =
       m_index_stream_buffer->ReserveMemory(MAXIBUFFERSIZE * sizeof(u16), sizeof(u16));
   if (!has_vbuffer_allocation || !has_ibuffer_allocation)
@@ -108,7 +98,7 @@ void VertexManager::ResetBuffer(u32 stride)
 
     // Attempt to allocate again, this may cause a fence wait
     if (!has_vbuffer_allocation)
-      has_vbuffer_allocation = m_vertex_stream_buffer->ReserveMemory(MAXVBUFFERSIZE, stride);
+      has_vbuffer_allocation = m_vertex_stream_buffer->ReserveMemory(MAXVBUFFERSIZE, vertex_stride);
     if (!has_ibuffer_allocation)
       has_ibuffer_allocation =
           m_index_stream_buffer->ReserveMemory(MAXIBUFFERSIZE * sizeof(u16), sizeof(u16));
@@ -123,34 +113,40 @@ void VertexManager::ResetBuffer(u32 stride)
   m_end_buffer_pointer = m_vertex_stream_buffer->GetCurrentHostPointer() + MAXVBUFFERSIZE;
   m_cur_buffer_pointer = m_vertex_stream_buffer->GetCurrentHostPointer();
   IndexGenerator::Start(reinterpret_cast<u16*>(m_index_stream_buffer->GetCurrentHostPointer()));
-
-  // Update base indices
-  m_current_draw_base_vertex =
-      static_cast<u32>(m_vertex_stream_buffer->GetCurrentOffset() / stride);
-  m_current_draw_base_index =
-      static_cast<u32>(m_index_stream_buffer->GetCurrentOffset() / sizeof(u16));
 }
 
-void VertexManager::vFlush()
+void VertexManager::CommitBuffer(u32 num_vertices, u32 vertex_stride, u32 num_indices,
+                                 u32* out_base_vertex, u32* out_base_index)
 {
-  const VertexFormat* vertex_format =
+  const u32 vertex_data_size = num_vertices * vertex_stride;
-      static_cast<VertexFormat*>(VertexLoaderManager::GetCurrentVertexFormat());
+  const u32 index_data_size = num_indices * sizeof(u16);
-  u32 vertex_stride = vertex_format->GetVertexStride();
 
-  // Figure out the number of indices to draw
+  *out_base_vertex =
-  u32 index_count = IndexGenerator::GetIndexLen();
+      vertex_stride > 0 ?
+          static_cast<u32>(m_vertex_stream_buffer->GetCurrentOffset() / vertex_stride) :
+          0;
+  *out_base_index = static_cast<u32>(m_index_stream_buffer->GetCurrentOffset() / sizeof(u16));
 
-  // Update tracked state
+  m_vertex_stream_buffer->CommitMemory(vertex_data_size);
+  m_index_stream_buffer->CommitMemory(index_data_size);
+
+  ADDSTAT(stats.thisFrame.bytesVertexStreamed, static_cast<int>(vertex_data_size));
+  ADDSTAT(stats.thisFrame.bytesIndexStreamed, static_cast<int>(index_data_size));
+
+  StateTracker::GetInstance()->SetVertexBuffer(m_vertex_stream_buffer->GetBuffer(), 0);
+  StateTracker::GetInstance()->SetIndexBuffer(m_index_stream_buffer->GetBuffer(), 0,
+                                              VK_INDEX_TYPE_UINT16);
+}
+
+void VertexManager::UploadConstants()
+{
   StateTracker::GetInstance()->UpdateVertexShaderConstants();
   StateTracker::GetInstance()->UpdateGeometryShaderConstants();
   StateTracker::GetInstance()->UpdatePixelShaderConstants();
+}
 
-  // Commit memory to device.
+void VertexManager::DrawCurrentBatch(u32 base_index, u32 num_indices, u32 base_vertex)
-  // NOTE: This must be done after constant upload, as a constant buffer overrun can cause
+{
-  // the current command buffer to be executed, and we want the buffer space to be associated
-  // with the command buffer that has the corresponding draw.
-  PrepareDrawBuffers(vertex_stride);
-
   // Flush all EFB pokes and invalidate the peek cache.
   FramebufferManager::GetInstance()->InvalidatePeekCache();
   FramebufferManager::GetInstance()->FlushEFBPokes();
@@ -168,19 +164,14 @@ void VertexManager::vFlush()
   }
 
   // Bind all pending state to the command buffer
-  if (m_current_pipeline_object)
+  if (StateTracker::GetInstance()->Bind())
   {
-    g_renderer->SetPipeline(m_current_pipeline_object);
+    vkCmdDrawIndexed(g_command_buffer_mgr->GetCurrentCommandBuffer(), num_indices, 1, base_index,
-    if (!StateTracker::GetInstance()->Bind())
+                     base_vertex, 0);
-    {
+  }
-      WARN_LOG(VIDEO, "Skipped draw of %u indices", index_count);
+  else
-      return;
+  {
-    }
+    WARN_LOG(VIDEO, "Skipped draw of %u indices", num_indices);
-
-    // Execute the draw
-    vkCmdDrawIndexed(g_command_buffer_mgr->GetCurrentCommandBuffer(), index_count, 1,
-                     m_current_draw_base_index, m_current_draw_base_vertex, 0);
-    INCSTAT(stats.thisFrame.numDrawCalls);
   }
 
   StateTracker::GetInstance()->OnDraw();

Source/Core/VideoBackends/Vulkan/VertexManager.h

@@ -27,20 +27,19 @@ public:
   std::unique_ptr<NativeVertexFormat>
   CreateNativeVertexFormat(const PortableVertexDeclaration& vtx_decl) override;
 
-protected:
+  void UploadUtilityUniforms(const void* uniforms, u32 uniforms_size) override;
-  void PrepareDrawBuffers(u32 stride);
-  void ResetBuffer(u32 stride) override;
 
-private:
+protected:
-  void vFlush() override;
+  void ResetBuffer(u32 vertex_stride, bool cull_all) override;
+  void CommitBuffer(u32 num_vertices, u32 vertex_stride, u32 num_indices, u32* out_base_vertex,
+                    u32* out_base_index) override;
+  void UploadConstants() override;
+  void DrawCurrentBatch(u32 base_index, u32 num_indices, u32 base_vertex) override;
 
   std::vector<u8> m_cpu_vertex_buffer;
   std::vector<u16> m_cpu_index_buffer;
 
   std::unique_ptr<StreamBuffer> m_vertex_stream_buffer;
   std::unique_ptr<StreamBuffer> m_index_stream_buffer;
-
-  u32 m_current_draw_base_vertex = 0;
-  u32 m_current_draw_base_index = 0;
 };
 }

Source/Core/VideoCommon/IndexGenerator.cpp

@@ -3,6 +3,7 @@
 // Refer to the license.txt file included.
 
 #include <cstddef>
+#include <cstring>
 
 #include "Common/CommonTypes.h"
 #include "Common/Compiler.h"
@@ -56,6 +57,13 @@ void IndexGenerator::AddIndices(int primitive, u32 numVerts)
   base_index += numVerts;
 }
 
+void IndexGenerator::AddExternalIndices(const u16* indices, u32 num_indices, u32 num_vertices)
+{
+  std::memcpy(index_buffer_current, indices, sizeof(u16) * num_indices);
+  index_buffer_current += num_indices;
+  base_index += num_vertices;
+}
+
 // Triangles
 template <bool pr>
 DOLPHIN_FORCE_INLINE u16* IndexGenerator::WriteTriangle(u16* Iptr, u32 index1, u32 index2,

Source/Core/VideoCommon/IndexGenerator.h

@@ -18,6 +18,8 @@ public:
 
   static void AddIndices(int primitive, u32 numVertices);
 
+  static void AddExternalIndices(const u16* indices, u32 num_indices, u32 num_vertices);
+
   // returns numprimitives
   static u32 GetNumVerts() { return base_index; }
   static u32 GetIndexLen() { return (u32)(index_buffer_current - BASEIptr); }

Source/Core/VideoCommon/RenderBase.h

@@ -114,6 +114,10 @@ public:
   {
   }
 
+  // Drawing with currently-bound pipeline state.
+  virtual void Draw(u32 base_vertex, u32 num_vertices) {}
+  virtual void DrawIndexed(u32 base_index, u32 num_indices, u32 base_vertex) {}
+
   // Shader modules/objects.
   virtual std::unique_ptr<AbstractShader>
   CreateShaderFromSource(ShaderStage stage, const char* source, size_t length) = 0;
@@ -192,16 +196,6 @@ public:
 
   virtual std::unique_ptr<VideoCommon::AsyncShaderCompiler> CreateAsyncShaderCompiler();
 
-  // Drawing utility shaders.
-  virtual void DrawUtilityPipeline(const void* uniforms, u32 uniforms_size, const void* vertices,
-                                   u32 vertex_stride, u32 num_vertices)
-  {
-  }
-  virtual void DispatchComputeShader(const AbstractShader* shader, const void* uniforms,
-                                     u32 uniforms_size, u32 groups_x, u32 groups_y, u32 groups_z)
-  {
-  }
-
   void ShowOSDMessage(OSDMessage message);
 
 protected:

Source/Core/VideoCommon/VertexManagerBase.cpp

@@ -27,6 +27,7 @@
 #include "VideoCommon/PixelShaderManager.h"
 #include "VideoCommon/RenderBase.h"
 #include "VideoCommon/SamplerCommon.h"
+#include "VideoCommon/Statistics.h"
 #include "VideoCommon/TextureCacheBase.h"
 #include "VideoCommon/VertexLoaderManager.h"
 #include "VideoCommon/VertexShaderManager.h"
@@ -131,7 +132,7 @@ DataReader VertexManagerBase::PrepareForAdditionalData(int primitive, u32 count,
   // need to alloc new buffer
   if (m_is_flushed)
   {
-    g_vertex_manager->ResetBuffer(stride);
+    g_vertex_manager->ResetBuffer(stride, cullall);
     m_is_flushed = false;
   }
 
@@ -209,6 +210,28 @@ std::pair<size_t, size_t> VertexManagerBase::ResetFlushAspectRatioCount()
   return val;
 }
 
+void VertexManagerBase::UploadUtilityVertices(const void* vertices, u32 vertex_stride,
+                                              u32 num_vertices, const u16* indices, u32 num_indices,
+                                              u32* out_base_vertex, u32* out_base_index)
+{
+  // The GX vertex list should be flushed before any utility draws occur.
+  ASSERT(m_is_flushed);
+
+  // Copy into the buffers usually used for GX drawing.
+  ResetBuffer(std::max(vertex_stride, 1u), false);
+  if (vertices)
+  {
+    const u32 copy_size = vertex_stride * num_vertices;
+    ASSERT((m_cur_buffer_pointer + copy_size) <= m_end_buffer_pointer);
+    std::memcpy(m_cur_buffer_pointer, vertices, copy_size);
+    m_cur_buffer_pointer += copy_size;
+  }
+  if (indices)
+    IndexGenerator::AddExternalIndices(indices, num_indices, num_vertices);
+
+  CommitBuffer(num_vertices, vertex_stride, num_indices, out_base_vertex, out_base_index);
+}
+
 static void SetSamplerState(u32 index, float custom_tex_scale, bool custom_tex,
                             bool has_arbitrary_mips)
 {
@@ -384,19 +407,35 @@ void VertexManagerBase::Flush()
 
   if (!m_cull_all)
   {
+    // Update and upload constants. Note for the Vulkan backend, this must occur before the
+    // vertex/index buffer is committed, otherwise the data will be associated with the
+    // previous command buffer, instead of the one with the draw if there is an overflow.
+    GeometryShaderManager::SetConstants();
+    PixelShaderManager::SetConstants();
+    UploadConstants();
+
+    // Now the vertices can be flushed to the GPU.
+    const u32 num_indices = IndexGenerator::GetIndexLen();
+    u32 base_vertex, base_index;
+    CommitBuffer(IndexGenerator::GetNumVerts(),
+                 VertexLoaderManager::GetCurrentVertexFormat()->GetVertexStride(), num_indices,
+                 &base_vertex, &base_index);
+
     // Update the pipeline, or compile one if needed.
     UpdatePipelineConfig();
     UpdatePipelineObject();
+    if (m_current_pipeline_object)
+    {
+      g_renderer->SetPipeline(m_current_pipeline_object);
+      if (PerfQueryBase::ShouldEmulate())
+        g_perf_query->EnableQuery(bpmem.zcontrol.early_ztest ? PQG_ZCOMP_ZCOMPLOC : PQG_ZCOMP);
 
-    // set the rest of the global constants
+      DrawCurrentBatch(base_index, num_indices, base_vertex);
-    GeometryShaderManager::SetConstants();
+      INCSTAT(stats.thisFrame.numDrawCalls);
-    PixelShaderManager::SetConstants();
 
-    if (PerfQueryBase::ShouldEmulate())
+      if (PerfQueryBase::ShouldEmulate())
-      g_perf_query->EnableQuery(bpmem.zcontrol.early_ztest ? PQG_ZCOMP_ZCOMPLOC : PQG_ZCOMP);
+        g_perf_query->DisableQuery(bpmem.zcontrol.early_ztest ? PQG_ZCOMP_ZCOMPLOC : PQG_ZCOMP);
-    g_vertex_manager->vFlush();
+    }
-    if (PerfQueryBase::ShouldEmulate())
-      g_perf_query->DisableQuery(bpmem.zcontrol.early_ztest ? PQG_ZCOMP_ZCOMPLOC : PQG_ZCOMP);
   }
 
   GFX_DEBUGGER_PAUSE_AT(NEXT_FLUSH, true);
@@ -413,7 +452,6 @@ void VertexManagerBase::Flush()
 void VertexManagerBase::DoState(PointerWrap& p)
 {
   p.Do(m_zslope);
-  g_vertex_manager->vDoState(p);
 }
 
 void VertexManagerBase::CalculateZSlope(NativeVertexFormat* format)

Source/Core/VideoCommon/VertexManagerBase.h

@@ -69,9 +69,29 @@ public:
     m_pipeline_config_changed = true;
   }
 
+  // Utility pipeline drawing (e.g. EFB copies, post-processing, UI).
+  virtual void UploadUtilityUniforms(const void* uniforms, u32 uniforms_size) = 0;
+  void UploadUtilityVertices(const void* vertices, u32 vertex_stride, u32 num_vertices,
+                             const u16* indices, u32 num_indices, u32* out_base_vertex,
+                             u32* out_base_index);
+
 protected:
-  virtual void vDoState(PointerWrap& p) {}
+  // Vertex buffers/index buffer creation.
-  virtual void ResetBuffer(u32 stride) = 0;
+  virtual void CreateDeviceObjects() {}
+  virtual void DestroyDeviceObjects() {}
+
+  // Prepares the buffer for the next batch of vertices.
+  virtual void ResetBuffer(u32 vertex_stride, bool cull_all) = 0;
+
+  // Commits/uploads the current batch of vertices.
+  virtual void CommitBuffer(u32 num_vertices, u32 vertex_stride, u32 num_indices,
+                            u32* out_base_vertex, u32* out_base_index) = 0;
+
+  // Uploads uniform buffers for GX draws.
+  virtual void UploadConstants() = 0;
+
+  // Issues the draw call for the current batch in the backend.
+  virtual void DrawCurrentBatch(u32 base_index, u32 num_indices, u32 base_vertex) = 0;
 
   u8* m_cur_buffer_pointer = nullptr;
   u8* m_base_buffer_pointer = nullptr;
@@ -98,10 +118,6 @@ private:
   size_t m_flush_count_4_3 = 0;
   size_t m_flush_count_anamorphic = 0;
 
-  virtual void vFlush() = 0;
-
-  virtual void CreateDeviceObjects() {}
-  virtual void DestroyDeviceObjects() {}
   void UpdatePipelineConfig();
   void UpdatePipelineObject();
 };