Merge pull request #7592 from stenzek/imgui-prereq

VideoBackends: Share GX vertex/index/uniform buffers with utility draws

6 files changed, 104 insertions, 31 deletions

diff --git a/Source/Core/VideoCommon/IndexGenerator.cpp b/Source/Core/VideoCommon/IndexGenerator.cpp
index 6cb49aceba..5e45cc6a5f 100644
--- a/Source/Core/VideoCommon/IndexGenerator.cpp
+++ b/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,
diff --git a/Source/Core/VideoCommon/IndexGenerator.h b/Source/Core/VideoCommon/IndexGenerator.h
index 8d2a5e8700..67f16b66c1 100644
--- a/Source/Core/VideoCommon/IndexGenerator.h
+++ b/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); }
diff --git a/Source/Core/VideoCommon/RenderBase.cpp b/Source/Core/VideoCommon/RenderBase.cpp
index 2e42ed984a..ba60e1636c 100644
--- a/Source/Core/VideoCommon/RenderBase.cpp
+++ b/Source/Core/VideoCommon/RenderBase.cpp
@@ -78,8 +78,10 @@ static float AspectToWidescreen(float aspect)
   return aspect * ((16.0f / 9.0f) / (4.0f / 3.0f));
 }
 
-Renderer::Renderer(int backbuffer_width, int backbuffer_height)
-    : m_backbuffer_width(backbuffer_width), m_backbuffer_height(backbuffer_height)
+Renderer::Renderer(int backbuffer_width, int backbuffer_height,
+                   AbstractTextureFormat backbuffer_format)
+    : m_backbuffer_width(backbuffer_width), m_backbuffer_height(backbuffer_height),
+      m_backbuffer_format(backbuffer_format)
 {
   UpdateActiveConfig();
   UpdateDrawRectangle();
@@ -93,6 +95,11 @@ Renderer::Renderer(int backbuffer_width, int backbuffer_height)
 
 Renderer::~Renderer() = default;
 
+bool Renderer::Initialize()
+{
+  return true;
+}
+
 void Renderer::Shutdown()
 {
   // First stop any framedumping, which might need to dump the last xfb frame. This process
@@ -697,6 +704,11 @@ void Renderer::Swap(u32 xfbAddr, u32 fbWidth, u32 fbStride, u32 fbHeight, const
 
       m_last_xfb_region = xfb_rect;
 
+      // Since we use the common pipelines here and draw vertices if a batch is currently being
+      // built by the vertex loader, we end up trampling over its pointer, as we share the buffer
+      // with the loader, and it has not been unmapped yet. Force a pipeline flush to avoid this.
+      g_vertex_manager->Flush();
+
       // TODO: merge more generic parts into VideoCommon
       {
         std::lock_guard<std::mutex> guard(m_swap_mutex);
diff --git a/Source/Core/VideoCommon/RenderBase.h b/Source/Core/VideoCommon/RenderBase.h
index d5dd761c40..bea1f5ba0a 100644
--- a/Source/Core/VideoCommon/RenderBase.h
+++ b/Source/Core/VideoCommon/RenderBase.h
@@ -32,6 +32,7 @@
 #include "VideoCommon/BPMemory.h"
 #include "VideoCommon/FPSCounter.h"
 #include "VideoCommon/RenderState.h"
+#include "VideoCommon/TextureConfig.h"
 #include "VideoCommon/VideoCommon.h"
 
 class AbstractFramebuffer;
@@ -72,13 +73,16 @@ enum class OSDMessage : s32
 class Renderer
 {
 public:
-  Renderer(int backbuffer_width, int backbuffer_height);
+  Renderer(int backbuffer_width, int backbuffer_height, AbstractTextureFormat backbuffer_format);
   virtual ~Renderer();
 
   using ClearColor = std::array<float, 4>;
 
   virtual bool IsHeadless() const = 0;
 
+  virtual bool Initialize();
+  virtual void Shutdown();
+
   virtual void SetPipeline(const AbstractPipeline* pipeline) {}
   virtual void SetScissorRect(const MathUtil::Rectangle<int>& rc) {}
   virtual void SetTexture(u32 index, const AbstractTexture* texture) {}
@@ -110,6 +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;
@@ -188,18 +196,6 @@ public:
 
   virtual std::unique_ptr<VideoCommon::AsyncShaderCompiler> CreateAsyncShaderCompiler();
 
-  virtual void Shutdown();
-
-  // 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:
@@ -229,6 +225,7 @@ protected:
   // Backbuffer (window) size and render area
   int m_backbuffer_width = 0;
   int m_backbuffer_height = 0;
+  AbstractTextureFormat m_backbuffer_format = AbstractTextureFormat::Undefined;
   TargetRectangle m_target_rectangle = {};
 
   FPSCounter m_fps_counter;
diff --git a/Source/Core/VideoCommon/VertexManagerBase.cpp b/Source/Core/VideoCommon/VertexManagerBase.cpp
index 8364b039c5..0e178e4d9b 100644
--- a/Source/Core/VideoCommon/VertexManagerBase.cpp
+++ b/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
-    GeometryShaderManager::SetConstants();
-    PixelShaderManager::SetConstants();
+      DrawCurrentBatch(base_index, num_indices, base_vertex);
+      INCSTAT(stats.thisFrame.numDrawCalls);
 
-    if (PerfQueryBase::ShouldEmulate())
-      g_perf_query->EnableQuery(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);
+      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)
diff --git a/Source/Core/VideoCommon/VertexManagerBase.h b/Source/Core/VideoCommon/VertexManagerBase.h
index 88d9d9fbe0..cd3e4ed552 100644
--- a/Source/Core/VideoCommon/VertexManagerBase.h
+++ b/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) {}
-  virtual void ResetBuffer(u32 stride) = 0;
+  // Vertex buffers/index buffer creation.
+  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();
 };