GPU: Basic/hacky CPU->VRAM transfers

src/pse/gpu.cpp

@@ -103,7 +103,7 @@ u32 GPU::ReadGPUREAD()
 void GPU::WriteGP0(u32 value)
 {
   m_GP0_command.push_back(value);
-  Assert(m_GP0_command.size() <= 128);
+  Assert(m_GP0_command.size() <= 1048576);
 
   const u8 command = Truncate8(m_GP0_command[0] >> 24);
   const u32 param = m_GP0_command[0] & UINT32_C(0x00FFFFFF);
@@ -121,6 +121,13 @@ void GPU::WriteGP0(u32 value)
       case 0x00: // NOP
         break;
 
+      case 0xA0: // Copy Rectangle CPU->VRAM
+      {
+        if (!HandleCopyRectangleCPUToVRAMCommand())
+          return;
+      }
+      break;
+
       case 0xE1: // Set draw mode
       {
         // 0..10 bits match GPUSTAT
@@ -288,6 +295,36 @@ bool GPU::HandleRenderCommand()
   return true;
 }
 
+bool GPU::HandleCopyRectangleCPUToVRAMCommand()
+{
+  if (m_GP0_command.size() < 3)
+    return false;
+
+  const u32 copy_width = m_GP0_command[2] & UINT32_C(0xFFFF);
+  const u32 copy_height = m_GP0_command[2] >> 16;
+  const u32 num_pixels = copy_width * copy_height;
+  const u32 num_words = 3 + ((num_pixels + 1) / 2);
+  if (m_GP0_command.size() < num_words)
+    return false;
+
+  const u32 dst_x = m_GP0_command[1] & UINT32_C(0xFFFF);
+  const u32 dst_y = m_GP0_command[1] >> 16;
+
+  Log_DebugPrintf("Copy rectangle from CPU to VRAM offset=(%u,%u), size=(%u,%u)", dst_x, dst_y, copy_width,
+                  copy_height);
+
+  if ((dst_x + copy_width) > VRAM_WIDTH || (dst_y + copy_height) > VRAM_HEIGHT)
+  {
+    Panic("Out of bounds VRAM copy");
+    return true;
+  }
+
+  UpdateVRAM(dst_x, dst_y, copy_width, copy_height, &m_GP0_command[3]);
+  return true;
+}
+
+void GPU::UpdateVRAM(u32 x, u32 y, u32 width, u32 height, const void* data) {}
+
 void GPU::DispatchRenderCommand(RenderCommand rc, u32 num_vertices) {}
 
 void GPU::FlushRender() {}

src/pse/gpu.h

@@ -97,8 +97,10 @@ protected:
 
   // Rendering commands, returns false if not enough data is provided
   bool HandleRenderCommand();
+  bool HandleCopyRectangleCPUToVRAMCommand();
 
   // Rendering in the backend
+  virtual void UpdateVRAM(u32 x, u32 y, u32 width, u32 height, const void* data);
   virtual void DispatchRenderCommand(RenderCommand rc, u32 num_vertices);
   virtual void FlushRender();
 

src/pse/gpu_hw.cpp

@@ -85,7 +85,7 @@ void main()
 {
   // 0..+1023 -> -1..1
   float pos_x = (float(a_position.x) / 511.5) - 1.0;
-  float pos_y = (float(a_position.y) / 255.5) + 1.0;
+  float pos_y = (float(a_position.y) / -255.5) + 1.0;
   gl_Position = vec4(pos_x, pos_y, 0.0, 1.0);
 
   v_color = a_color;

src/pse/gpu_hw_opengl.cpp

@@ -51,7 +51,7 @@ void GPU_HW_OpenGL::ClearFramebuffer()
   glClear(GL_COLOR_BUFFER_BIT);
   glBindFramebuffer(GL_FRAMEBUFFER, 0);
 
-  m_system->GetHostInterface()->SetDisplayTexture(m_framebuffer_texture.get(), 0, 0, VRAM_WIDTH, VRAM_HEIGHT);
+  //m_system->GetHostInterface()->SetDisplayTexture(m_framebuffer_texture.get(), 0, 0, VRAM_WIDTH, VRAM_HEIGHT);
 }
 
 void GPU_HW_OpenGL::DestroyFramebuffer()
@@ -128,6 +128,45 @@ void GPU_HW_OpenGL::SetViewport()
 
 void GPU_HW_OpenGL::SetScissor() {}
 
+inline u32 ConvertRGBA5551ToRGBA8888(u16 color)
+{
+  u8 r = Truncate8(color & 31);
+  u8 g = Truncate8((color >> 5) & 31);
+  u8 b = Truncate8((color >> 10) & 31);
+  u8 a = Truncate8((color >> 15) & 1);
+
+  // 00012345 -> 1234545
+  b = (b << 3) | (b >> 3);
+  g = (g << 3) | (g >> 3);
+  r = (r << 3) | (r >> 3);
+  a = a ? 255 : 0;
+
+  return ZeroExtend32(r) | (ZeroExtend32(g) << 8) | (ZeroExtend32(b) << 16) | (ZeroExtend32(a) << 24);
+}
+
+void GPU_HW_OpenGL::UpdateVRAM(u32 x, u32 y, u32 width, u32 height, const void* data)
+{
+  const u32 pixel_count = width * height;
+  std::vector<u32> rgba_data;
+  rgba_data.reserve(pixel_count);
+
+  const u8* source_ptr = static_cast<const u8*>(data);
+  for (u32 i = 0; i < pixel_count; i++)
+  {
+    u16 src_col;
+    std::memcpy(&src_col, source_ptr, sizeof(src_col));
+    source_ptr += sizeof(src_col);
+
+    const u32 dst_col = ConvertRGBA5551ToRGBA8888(src_col);
+    rgba_data.push_back(dst_col);
+  }
+
+  m_framebuffer_texture->Bind();
+  glTexSubImage2D(GL_TEXTURE_2D, 0, x, y, width, height, GL_RGBA, GL_UNSIGNED_BYTE,
+                  rgba_data.data());
+  m_system->GetHostInterface()->SetDisplayTexture(m_framebuffer_texture.get(), 0, 0, VRAM_WIDTH, VRAM_HEIGHT);
+}
+
 void GPU_HW_OpenGL::DispatchRenderCommand(RenderCommand rc, u32 num_vertices)
 {
   LoadVertices(rc, num_vertices);

src/pse/gpu_hw_opengl.h

@@ -16,8 +16,9 @@ public:
   void Reset() override;
 
 protected:
-  virtual void DispatchRenderCommand(RenderCommand rc, u32 num_vertices) override;
+  void UpdateVRAM(u32 x, u32 y, u32 width, u32 height, const void* data) override;
-  virtual void FlushRender() override;
+  void DispatchRenderCommand(RenderCommand rc, u32 num_vertices) override;
+  void FlushRender() override;
 
 private:
   void CreateFramebuffer();