GPU/HW: Only copy the dirty area to the VRAM read texture

src/common/rectangle.h

@@ -87,6 +87,33 @@ struct Rectangle
 
 #undef RELATIONAL_OPERATOR
 
+  // Arithmetic operators.
+#define ARITHMETIC_OPERATOR(op)                                                                                        \
+  constexpr Rectangle& operator op##=(const T amount)                                                                  \
+  {                                                                                                                    \
+    left op## = amount;                                                                                                \
+    top op## = amount;                                                                                                 \
+    right op## = amount;                                                                                               \
+    bottom op## = amount;                                                                                              \
+  }                                                                                                                    \
+  constexpr Rectangle operator op(const T amount)                                                                      \
+  {                                                                                                                    \
+    return Rectangle(left op amount, top op amount, right op amount, bottom op amount);                                \
+  }
+
+  ARITHMETIC_OPERATOR(+);
+  ARITHMETIC_OPERATOR(-);
+  ARITHMETIC_OPERATOR(*);
+  ARITHMETIC_OPERATOR(/);
+  ARITHMETIC_OPERATOR(%);
+  ARITHMETIC_OPERATOR(>>);
+  ARITHMETIC_OPERATOR(<<);
+  ARITHMETIC_OPERATOR(|);
+  ARITHMETIC_OPERATOR(&);
+  ARITHMETIC_OPERATOR(^);
+
+#undef ARITHMETIC_OPERATOR
+
 #ifdef _WINDEF_
   /// Casts this rectangle to a Win32 RECT structure if compatible.
   template<bool _ = true, typename = typename std::enable_if_t<std::is_same_v<T, s32> && _>>

src/core/gpu.h

@@ -385,9 +385,11 @@ protected:
     }
 
     bool IsTexturePageChanged() const { return texture_page_changed; }
+    void SetTexturePageChanged() { texture_page_changed = true; }
     void ClearTexturePageChangedFlag() { texture_page_changed = false; }
 
     bool IsTextureWindowChanged() const { return texture_window_changed; }
+    void SetTextureWindowChanged() { texture_window_changed = true; }
     void ClearTextureWindowChangedFlag() { texture_window_changed = false; }
 
     void SetFromPolygonTexcoord(u32 texcoord0, u32 texcoord1);

src/core/gpu_hw.cpp

@@ -1,6 +1,7 @@
 #include "gpu_hw.h"
 #include "YBaseLib/Assert.h"
 #include "YBaseLib/Log.h"
+#include "common/state_wrapper.h"
 #include "settings.h"
 #include "system.h"
 #include <imgui.h>
@@ -11,15 +12,6 @@ GPU_HW::GPU_HW() = default;
 
 GPU_HW::~GPU_HW() = default;
 
-void GPU_HW::Reset()
-{
-  GPU::Reset();
-
-  m_batch = {};
-  m_batch_ubo_data = {};
-  m_batch_ubo_dirty = true;
-}
-
 bool GPU_HW::Initialize(HostDisplay* host_display, System* system, DMA* dma, InterruptController* interrupt_controller,
                         Timers* timers)
 {
@@ -33,6 +25,29 @@ bool GPU_HW::Initialize(HostDisplay* host_display, System* system, DMA* dma, Int
   return true;
 }
 
+void GPU_HW::Reset()
+{
+  GPU::Reset();
+
+  m_batch = {};
+  m_batch_ubo_data = {};
+  m_batch_ubo_dirty = true;
+
+  SetFullVRAMDirtyRectangle();
+}
+
+bool GPU_HW::DoState(StateWrapper& sw)
+{
+  if (!GPU::DoState(sw))
+    return false;
+
+  // invalidate the whole VRAM read texture when loading state
+  if (sw.IsReading())
+    SetFullVRAMDirtyRectangle();
+
+  return true;
+}
+
 void GPU_HW::UpdateSettings()
 {
   GPU::UpdateSettings();
@@ -228,11 +243,11 @@ void GPU_HW::DispatchRenderCommand(RenderCommand rc, u32 num_vertices, const u32
       if (m_vram_dirty_rect.Valid() && (m_render_state.GetTexturePageRectangle().Intersects(m_vram_dirty_rect) ||
                                         m_render_state.GetTexturePaletteRectangle().Intersects(m_vram_dirty_rect)))
       {
+        Log_WarningPrintf("Invalidating VRAM read cache due to drawing area overlap");
         if (!IsFlushed())
           FlushRender();
 
-        Log_WarningPrintf("Invalidating VRAM read cache due to drawing area overlap");
+        UpdateVRAMReadTexture();
-        m_vram_read_texture_dirty = true;
       }
     }
 

src/core/gpu_hw.h

@@ -30,6 +30,7 @@ public:
   virtual bool Initialize(HostDisplay* host_display, System* system, DMA* dma,
                           InterruptController* interrupt_controller, Timers* timers) override;
   virtual void Reset() override;
+  virtual bool DoState(StateWrapper& sw) override;
   virtual void UpdateSettings() override;
 
 protected:
@@ -111,6 +112,14 @@ protected:
   }
 
   virtual void MapBatchVertexPointer(u32 required_vertices) = 0;
+  virtual void UpdateVRAMReadTexture() = 0;
+
+  void SetFullVRAMDirtyRectangle()
+  {
+    m_vram_dirty_rect.Set(0, 0, VRAM_WIDTH, VRAM_HEIGHT);
+    m_render_state.SetTexturePageChanged();
+  }
+  void ClearVRAMDirtyRectangle() { m_vram_dirty_rect.SetInvalid(); }
 
   u32 GetBatchVertexSpace() const { return static_cast<u32>(m_batch_end_vertex_ptr - m_batch_current_vertex_ptr); }
   u32 GetBatchVertexCount() const { return static_cast<u32>(m_batch_current_vertex_ptr - m_batch_start_vertex_ptr); }
@@ -151,7 +160,6 @@ protected:
 
   // Changed state
   bool m_batch_ubo_dirty = true;
-  bool m_vram_read_texture_dirty = false;
 
 private:
   enum : u32

src/core/gpu_hw_d3d11.cpp

@@ -184,7 +184,7 @@ bool GPU_HW_D3D11::CreateFramebuffer()
   }
 
   m_context->OMSetRenderTargets(1, m_vram_texture.GetD3DRTVArray(), nullptr);
-  m_vram_read_texture_dirty = true;
+  SetFullVRAMDirtyRectangle();
   return true;
 }
 
@@ -192,7 +192,7 @@ void GPU_HW_D3D11::ClearFramebuffer()
 {
   static constexpr std::array<float, 4> color = {};
   m_context->ClearRenderTargetView(m_vram_texture.GetD3DRTV(), color.data());
-  m_vram_read_texture_dirty = true;
+  SetFullVRAMDirtyRectangle();
 }
 
 void GPU_HW_D3D11::DestroyFramebuffer()
@@ -491,9 +491,6 @@ void GPU_HW_D3D11::SetDrawState(BatchRenderMode render_mode)
     UploadUniformBlock(&m_batch_ubo_data, sizeof(m_batch_ubo_data));
     m_batch_ubo_dirty = false;
   }
-
-  if (m_vram_read_texture_dirty)
-    UpdateVRAMReadTexture();
 }
 
 void GPU_HW_D3D11::SetScissorFromDrawingArea()
@@ -648,12 +645,13 @@ void GPU_HW_D3D11::CopyVRAM(u32 src_x, u32 src_y, u32 dst_x, u32 dst_y, u32 widt
 
 void GPU_HW_D3D11::UpdateVRAMReadTexture()
 {
-  m_renderer_stats.num_vram_read_texture_updates++;
+  const auto scaled_rect = m_vram_dirty_rect * m_resolution_scale;
-  m_vram_read_texture_dirty = false;
+  const CD3D11_BOX src_box(scaled_rect.left, scaled_rect.top, 0, scaled_rect.right, scaled_rect.bottom, 1);
-  m_vram_dirty_rect.SetInvalid();
+  m_context->CopySubresourceRegion(m_vram_read_texture, 0, scaled_rect.left, scaled_rect.top, 0, m_vram_texture, 0,
+                                   &src_box);
 
-  const CD3D11_BOX src_box(0, 0, 0, m_vram_texture.GetWidth(), m_vram_texture.GetHeight(), 1);
+  m_renderer_stats.num_vram_read_texture_updates++;
-  m_context->CopySubresourceRegion(m_vram_read_texture, 0, 0, 0, 0, m_vram_texture, 0, &src_box);
+  ClearVRAMDirtyRectangle();
 }
 
 void GPU_HW_D3D11::FlushRender()

src/core/gpu_hw_d3d11.h

@@ -34,13 +34,13 @@ protected:
   void CopyVRAM(u32 src_x, u32 src_y, u32 dst_x, u32 dst_y, u32 width, u32 height) override;
   void FlushRender() override;
   void MapBatchVertexPointer(u32 required_vertices) override;
+  void UpdateVRAMReadTexture() override;
 
 private:
   void SetCapabilities();
   bool CreateFramebuffer();
   void ClearFramebuffer();
   void DestroyFramebuffer();
-  void UpdateVRAMReadTexture();
 
   bool CreateVertexBuffer();
   bool CreateUniformBuffer();

src/core/gpu_hw_opengl.cpp

@@ -169,7 +169,7 @@ void GPU_HW_OpenGL::CreateFramebuffer()
     std::make_unique<GL::Texture>(texture_width, texture_height, GL_RGBA, GL_UNSIGNED_BYTE, nullptr, false, true);
 
   m_vram_texture->BindFramebuffer(GL_DRAW_FRAMEBUFFER);
-  m_vram_read_texture_dirty = true;
+  SetFullVRAMDirtyRectangle();
 }
 
 void GPU_HW_OpenGL::ClearFramebuffer()
@@ -178,7 +178,7 @@ void GPU_HW_OpenGL::ClearFramebuffer()
   glClearColor(0.0f, 0.0f, 0.0f, 0.0f);
   glClear(GL_COLOR_BUFFER_BIT);
   glEnable(GL_SCISSOR_TEST);
-  m_vram_read_texture_dirty = true;
+  SetFullVRAMDirtyRectangle();
 }
 
 void GPU_HW_OpenGL::DestroyFramebuffer()
@@ -352,9 +352,6 @@ void GPU_HW_OpenGL::SetDrawState(BatchRenderMode render_mode)
     UploadUniformBlock(&m_batch_ubo_data, sizeof(m_batch_ubo_data));
     m_batch_ubo_dirty = false;
   }
-
-  if (m_vram_read_texture_dirty)
-    UpdateVRAMReadTexture();
 }
 
 void GPU_HW_OpenGL::SetScissorFromDrawingArea()
@@ -678,13 +675,16 @@ void GPU_HW_OpenGL::CopyVRAM(u32 src_x, u32 src_y, u32 dst_x, u32 dst_y, u32 wid
 
 void GPU_HW_OpenGL::UpdateVRAMReadTexture()
 {
-  m_renderer_stats.num_vram_read_texture_updates++;
+  // TODO: Fallback blit path.
-  m_vram_read_texture_dirty = false;
+  const auto scaled_rect = m_vram_dirty_rect * m_resolution_scale;
-  m_vram_dirty_rect.SetInvalid();
+  const u32 flipped_y = m_vram_texture->GetHeight() - scaled_rect.top - scaled_rect.GetHeight();
 
-  // TODO: Fallback blit path, and partial updates.
+  glCopyImageSubData(m_vram_texture->GetGLId(), GL_TEXTURE_2D, 0, scaled_rect.left, flipped_y, 0,
-  glCopyImageSubData(m_vram_texture->GetGLId(), GL_TEXTURE_2D, 0, 0, 0, 0, m_vram_read_texture->GetGLId(),
+                     m_vram_read_texture->GetGLId(), GL_TEXTURE_2D, 0, scaled_rect.left, flipped_y, 0,
-                     GL_TEXTURE_2D, 0, 0, 0, 0, m_vram_texture->GetWidth(), m_vram_texture->GetHeight(), 1);
+                     scaled_rect.GetWidth(), scaled_rect.GetHeight(), 1);
+
+  m_renderer_stats.num_vram_read_texture_updates++;
+  ClearVRAMDirtyRectangle();
 }
 
 void GPU_HW_OpenGL::FlushRender()

src/core/gpu_hw_opengl.h

@@ -30,6 +30,7 @@ protected:
   void CopyVRAM(u32 src_x, u32 src_y, u32 dst_x, u32 dst_y, u32 width, u32 height) override;
   void FlushRender() override;
   void MapBatchVertexPointer(u32 required_vertices) override;
+  void UpdateVRAMReadTexture() override;
 
 private:
   struct GLStats
@@ -48,7 +49,6 @@ private:
   void CreateFramebuffer();
   void ClearFramebuffer();
   void DestroyFramebuffer();
-  void UpdateVRAMReadTexture();
 
   void CreateVertexBuffer();
   void CreateUniformBuffer();