From 2b17cfd3653f817bcc820b865010d7920dbc4fd3 Mon Sep 17 00:00:00 2001
From: Connor McLaughlin <mclaughc@outlook.com>
Date: Sat, 2 Nov 2019 22:31:15 +1000
Subject: [PATCH] GPU: Use streaming buffers for CPU->VRAM transfers

---
 src/core/gpu_hw.h          |  1 +
 src/core/gpu_hw_opengl.cpp | 21 ++++++++++++++++-----
 src/core/gpu_hw_opengl.h   |  3 +++
 3 files changed, 20 insertions(+), 5 deletions(-)

diff --git a/src/core/gpu_hw.h b/src/core/gpu_hw.h
index 7828ad714..235074464 100644
--- a/src/core/gpu_hw.h
+++ b/src/core/gpu_hw.h
@@ -83,6 +83,7 @@ protected:
     }
   };
 
+  static constexpr u32 VRAM_UPDATE_TEXTURE_BUFFER_SIZE = VRAM_WIDTH * VRAM_HEIGHT * sizeof(u32);
   static constexpr u32 VERTEX_BUFFER_SIZE = 1 * 1024 * 1024;
   static constexpr u32 MIN_BATCH_VERTEX_COUNT = 6;
   static constexpr u32 MAX_BATCH_VERTEX_COUNT = VERTEX_BUFFER_SIZE / sizeof(HWVertex);
diff --git a/src/core/gpu_hw_opengl.cpp b/src/core/gpu_hw_opengl.cpp
index 3003d7521..c0d06f5b7 100644
--- a/src/core/gpu_hw_opengl.cpp
+++ b/src/core/gpu_hw_opengl.cpp
@@ -23,6 +23,7 @@ bool GPU_HW_OpenGL::Initialize(System* system, DMA* dma, InterruptController* in
 
   CreateFramebuffer();
   CreateVertexBuffer();
+  CreateTextureBuffer();
   if (!CompilePrograms())
     return false;
 
@@ -251,6 +252,13 @@ void GPU_HW_OpenGL::CreateVertexBuffer()
   glGenVertexArrays(1, &m_attributeless_vao_id);
 }
 
+void GPU_HW_OpenGL::CreateTextureBuffer()
+{
+  m_texture_stream_buffer = GL::StreamBuffer::Create(GL_PIXEL_UNPACK_BUFFER, VRAM_UPDATE_TEXTURE_BUFFER_SIZE);
+  if (!m_texture_stream_buffer)
+    Panic("Failed to create texture stream buffer");
+}
+
 bool GPU_HW_OpenGL::CompilePrograms()
 {
   for (u32 render_mode = 0; render_mode < 4; render_mode++)
@@ -564,12 +572,13 @@ void GPU_HW_OpenGL::FillVRAM(u32 x, u32 y, u32 width, u32 height, u16 color)
 
 void GPU_HW_OpenGL::UpdateVRAM(u32 x, u32 y, u32 width, u32 height, const void* data)
 {
-  std::vector<u32> rgba_data;
-  rgba_data.reserve(width * height);
+  const u32 num_pixels = width * height;
+  const auto map_result = m_texture_stream_buffer->Map(sizeof(u32), num_pixels * sizeof(u32));
 
   // reverse copy the rows so it matches opengl's lower-left origin
   const u32 source_stride = width * sizeof(u16);
   const u8* source_ptr = static_cast<const u8*>(data) + (source_stride * (height - 1));
+  u32* dest_ptr = static_cast<u32*>(map_result.pointer);
   for (u32 row = 0; row < height; row++)
   {
     const u8* source_row_ptr = source_ptr;
@@ -580,13 +589,14 @@ void GPU_HW_OpenGL::UpdateVRAM(u32 x, u32 y, u32 width, u32 height, const void*
       std::memcpy(&src_col, source_row_ptr, sizeof(src_col));
       source_row_ptr += sizeof(src_col);
 
-      const u32 dst_col = RGBA5551ToRGBA8888(src_col);
-      rgba_data.push_back(dst_col);
+      *(dest_ptr++) = RGBA5551ToRGBA8888(src_col);
     }
 
     source_ptr -= source_stride;
   }
 
+  m_texture_stream_buffer->Unmap(num_pixels * sizeof(u32));
+
   // have to write to the 1x texture first
   if (m_resolution_scale > 1)
     m_vram_downsample_texture->Bind();
@@ -597,7 +607,8 @@ void GPU_HW_OpenGL::UpdateVRAM(u32 x, u32 y, u32 width, u32 height, const void*
   const u32 flipped_y = VRAM_HEIGHT - y - height;
 
   // update texture data
-  glTexSubImage2D(GL_TEXTURE_2D, 0, x, flipped_y, width, height, GL_RGBA, GL_UNSIGNED_BYTE, rgba_data.data());
+  glTexSubImage2D(GL_TEXTURE_2D, 0, x, flipped_y, width, height, GL_RGBA, GL_UNSIGNED_BYTE,
+                  reinterpret_cast<void*>(map_result.index_aligned * sizeof(u32)));
   InvalidateVRAMReadCache();
 
   if (m_resolution_scale > 1)
diff --git a/src/core/gpu_hw_opengl.h b/src/core/gpu_hw_opengl.h
index 81236a943..8ba2eac34 100644
--- a/src/core/gpu_hw_opengl.h
+++ b/src/core/gpu_hw_opengl.h
@@ -53,6 +53,7 @@ private:
   void UpdateVRAMReadTexture();
 
   void CreateVertexBuffer();
+  void CreateTextureBuffer();
 
   bool CompilePrograms();
   bool CompileProgram(GL::Program& prog, HWBatchRenderMode render_mode, TextureMode texture_mode, bool dithering);
@@ -68,6 +69,8 @@ private:
   GLuint m_vao_id = 0;
   GLuint m_attributeless_vao_id = 0;
 
+  std::unique_ptr<GL::StreamBuffer> m_texture_stream_buffer;
+
   bool m_vram_read_texture_dirty = true;
   bool m_drawing_area_changed = true;
   bool m_show_renderer_statistics = false;