From 4d38213f2351a88c1720840ef36b99d61e6d176d Mon Sep 17 00:00:00 2001
From: Connor McLaughlin <mclaughc@outlook.com>
Date: Wed, 18 Sep 2019 00:58:30 +1000
Subject: [PATCH] GPU: Implement VRAM-to-VRAM copies

---
 src/pse/gpu.cpp           | 42 +++++++++++++++++++++++++++++++++++++--
 src/pse/gpu.h             |  2 ++
 src/pse/gpu_hw_opengl.cpp | 13 ++++++++++++
 src/pse/gpu_hw_opengl.h   |  1 +
 4 files changed, 56 insertions(+), 2 deletions(-)

diff --git a/src/pse/gpu.cpp b/src/pse/gpu.cpp
index 0958cec23..c66dd62da 100644
--- a/src/pse/gpu.cpp
+++ b/src/pse/gpu.cpp
@@ -343,7 +343,10 @@ void GPU::WriteGP0(u32 value)
       case 0x00: // NOP
         break;
 
-      case 0x02: // Fill Rectnagle
+      case 0x01: // Clear cache
+        break;
+
+      case 0x02: // Fill Rectangle
       {
         if (!HandleFillRectangleCommand())
           return;
@@ -357,13 +360,20 @@ void GPU::WriteGP0(u32 value)
       }
       break;
 
-      case 0xC0: // Copy Rectnagle VRAM->CPU
+      case 0xC0: // Copy Rectangle VRAM->CPU
       {
         if (!HandleCopyRectangleVRAMToCPUCommand())
           return;
       }
       break;
 
+      case 0x80: // Copy Rectangle VRAM->VRAM
+      {
+        if (!HandleCopyRectangleVRAMToVRAMCommand())
+          return;
+      }
+      break;
+
       case 0xE1: // Set draw mode
       {
         // 0..10 bits match GPUSTAT
@@ -669,6 +679,32 @@ bool GPU::HandleCopyRectangleVRAMToCPUCommand()
   return true;
 }
 
+bool GPU::HandleCopyRectangleVRAMToVRAMCommand()
+{
+  if (m_GP0_command.size() < 4)
+    return false;
+
+  const u32 src_x = m_GP0_command[1] & UINT32_C(0xFFFF);
+  const u32 src_y = m_GP0_command[1] >> 16;
+  const u32 dst_x = m_GP0_command[2] & UINT32_C(0xFFFF);
+  const u32 dst_y = m_GP0_command[2] >> 16;
+  const u32 width = m_GP0_command[3] & UINT32_C(0xFFFF);
+  const u32 height = m_GP0_command[3] >> 16;
+
+  Log_DebugPrintf("Copy rectangle from VRAM to VRAM src=(%u,%u), dst=(%u,%u), size=(%u,%u)", src_x, src_y, dst_x, dst_y,
+                  width, height);
+
+  if ((src_x + width) > VRAM_WIDTH || (src_y + height) > VRAM_HEIGHT || (dst_x + width) > VRAM_WIDTH ||
+      (dst_y + height) > VRAM_HEIGHT)
+  {
+    Panic("Out of bounds VRAM copy");
+    return true;
+  }
+
+  CopyVRAM(src_x, src_y, dst_x, dst_y, width, height);
+  return true;
+}
+
 void GPU::UpdateDisplay()
 {
   m_texture_config.page_changed = true;
@@ -681,6 +717,8 @@ void GPU::FillVRAM(u32 x, u32 y, u32 width, u32 height, u32 color) {}
 
 void GPU::UpdateVRAM(u32 x, u32 y, u32 width, u32 height, const void* data) {}
 
+void GPU::CopyVRAM(u32 src_x, u32 src_y, u32 dst_x, u32 dst_y, u32 width, u32 height) {}
+
 void GPU::DispatchRenderCommand(RenderCommand rc, u32 num_vertices) {}
 
 void GPU::FlushRender() {}
diff --git a/src/pse/gpu.h b/src/pse/gpu.h
index b8022f5dc..1d63cbbd0 100644
--- a/src/pse/gpu.h
+++ b/src/pse/gpu.h
@@ -131,12 +131,14 @@ protected:
   bool HandleFillRectangleCommand();
   bool HandleCopyRectangleCPUToVRAMCommand();
   bool HandleCopyRectangleVRAMToCPUCommand();
+  bool HandleCopyRectangleVRAMToVRAMCommand();
 
   // Rendering in the backend
   virtual void UpdateDisplay();
   virtual void ReadVRAM(u32 x, u32 y, u32 width, u32 height, void* buffer);
   virtual void FillVRAM(u32 x, u32 y, u32 width, u32 height, u32 color);
   virtual void UpdateVRAM(u32 x, u32 y, u32 width, u32 height, const void* data);
+  virtual void CopyVRAM(u32 src_x, u32 src_y, u32 dst_x, u32 dst_y, u32 width, u32 height);
   virtual void DispatchRenderCommand(RenderCommand rc, u32 num_vertices);
   virtual void FlushRender();
 
diff --git a/src/pse/gpu_hw_opengl.cpp b/src/pse/gpu_hw_opengl.cpp
index bc752d80a..7c5df30bb 100644
--- a/src/pse/gpu_hw_opengl.cpp
+++ b/src/pse/gpu_hw_opengl.cpp
@@ -300,6 +300,19 @@ void GPU_HW_OpenGL::UpdateVRAM(u32 x, u32 y, u32 width, u32 height, const void*
                   rgba_data.data());
 }
 
+void GPU_HW_OpenGL::CopyVRAM(u32 src_x, u32 src_y, u32 dst_x, u32 dst_y, u32 width, u32 height)
+{
+  glDisable(GL_SCISSOR_TEST);
+
+  // lower-left origin flip
+  src_y = VRAM_HEIGHT - src_y - height;
+  dst_y = VRAM_HEIGHT - dst_y - height;
+
+  glBindFramebuffer(GL_FRAMEBUFFER, m_framebuffer_fbo_id);
+  glBlitFramebuffer(src_x, src_y, src_x + width, src_y + height, dst_x, dst_y, dst_x + width, dst_y + height,
+                    GL_COLOR_BUFFER_BIT, GL_NEAREST);
+}
+
 void GPU_HW_OpenGL::UpdateTexturePageTexture()
 {
   glBindFramebuffer(GL_FRAMEBUFFER, m_texture_page_fbo_id);
diff --git a/src/pse/gpu_hw_opengl.h b/src/pse/gpu_hw_opengl.h
index 98f5ea7a3..3c5cb6df3 100644
--- a/src/pse/gpu_hw_opengl.h
+++ b/src/pse/gpu_hw_opengl.h
@@ -21,6 +21,7 @@ protected:
   void ReadVRAM(u32 x, u32 y, u32 width, u32 height, void* buffer) override;
   void FillVRAM(u32 x, u32 y, u32 width, u32 height, u32 color) override;
   void UpdateVRAM(u32 x, u32 y, u32 width, u32 height, const void* data) override;
+  void CopyVRAM(u32 src_x, u32 src_y, u32 dst_x, u32 dst_y, u32 width, u32 height) override;
   void UpdateTexturePageTexture() override;
   void FlushRender() override;