Use BufferStorage extension for faster glReadPixels.

projects/msvc12/GLideN64.vcxproj

@@ -265,6 +265,7 @@
  <ItemGroup>
  <ClCompile Include="..\..\src\3DMath.cpp" />
  <ClCompile Include="..\..\src\BufferCopy\ColorBufferToRDRAM.cpp" />
+ <ClCompile Include="..\..\src\BufferCopy\ColorBufferToRDRAM_BufferStorageExt.cpp" />
  <ClCompile Include="..\..\src\BufferCopy\ColorBufferToRDRAM_GL.cpp" />
  <ClCompile Include="..\..\src\BufferCopy\DepthBufferToRDRAM.cpp" />
  <ClCompile Include="..\..\src\BufferCopy\RDRAMtoColorBuffer.cpp" />
@@ -380,6 +381,8 @@
  <ItemGroup>
  <ClInclude Include="..\..\src\3DMath.h" />
  <ClInclude Include="..\..\src\BufferCopy\ColorBufferToRDRAM.h" />
+ <ClInclude Include="..\..\src\BufferCopy\ColorBufferToRDRAM_BufferStorageExt.h" />
+ <ClInclude Include="..\..\src\BufferCopy\ColorBufferToRDRAM_GL.h" />
  <ClInclude Include="..\..\src\BufferCopy\DepthBufferToRDRAM.h" />
  <ClInclude Include="..\..\src\BufferCopy\RDRAMtoColorBuffer.h" />
  <ClInclude Include="..\..\src\BufferCopy\WriteToRDRAM.h" />

projects/msvc12/GLideN64.vcxproj.filters

@@ -257,6 +257,9 @@
  <ClCompile Include="..\..\src\Log.cpp">
  <Filter>Source Files</Filter>
  </ClCompile>
+ <ClCompile Include="..\..\src\BufferCopy\ColorBufferToRDRAM_BufferStorageExt.cpp">
+ <Filter>Source Files\BufferCopy</Filter>
+ </ClCompile>
  </ItemGroup>
  <ItemGroup>
  <ClInclude Include="..\..\src\3DMath.h">
@@ -454,5 +457,11 @@
  <ClInclude Include="..\..\src\common\GLFunctions.h">
  <Filter>Header Files\common</Filter>
  </ClInclude>
+ <ClInclude Include="..\..\src\BufferCopy\ColorBufferToRDRAM_BufferStorageExt.h">
+ <Filter>Header Files\BufferCopy</Filter>
+ </ClInclude>
+ <ClInclude Include="..\..\src\BufferCopy\ColorBufferToRDRAM_GL.h">
+ <Filter>Header Files\BufferCopy</Filter>
+ </ClInclude>
  </ItemGroup>
 </Project>

src/BufferCopy/ColorBufferToRDRAM.cpp

@@ -6,11 +6,18 @@
 
 #include <FBOTextureFormats.h>
 #include <FrameBuffer.h>
-#include <Textures.h>
 #include <Config.h>
 #include <N64.h>
 #include <VI.h>
-#include <Log.h>
+#include "Log.h"
+#ifndef GLES2
+#include "ColorBufferToRDRAM_GL.h"
+#include "ColorBufferToRDRAM_BufferStorageExt.h"
+#elif defined(ANDROID) && defined (GLES2)
+#include "ColorBufferToRDRAM_GLES.h"
+#else
+#include "ColorBufferToRDRAMStub.h"
+#endif
 
 ColorBufferToRDRAM::ColorBufferToRDRAM()
  : m_FBO(0)
@@ -264,6 +271,31 @@ void ColorBufferToRDRAM::copyChunkToRDRAM(u32 _address)
  _copy(_address, _address + 0x1000, true);
 }
 
+
+ColorBufferToRDRAM & ColorBufferToRDRAM::get()
+{
+#ifndef GLES2
+
+ static bool supportsBufferStorage = OGLVideo::isExtensionSupported("GL_EXT_buffer_storage") ||
+ OGLVideo::isExtensionSupported("GL_ARB_buffer_storage");
+
+ if (supportsBufferStorage) {
+ static ColorBufferToRDRAM_BufferStorageExt cbCopy;
+ return cbCopy;
+ } else {
+ static ColorBufferToRDRAM_GL cbCopy;
+ return cbCopy;
+ }
+
+#elif defined(ANDROID) && defined (GLES2)
+ static ColorBufferToRDRAM_GLES cbCopy;
+ return cbCopy;
+#else
+ static ColorBufferToRDRAMStub cbCopy;
+ return cbCopy;
+#endif
+}
+
 void copyWhiteToRDRAM(FrameBuffer * _pBuffer)
 {
  if (_pBuffer->m_size == G_IM_SIZ_32b) {

src/BufferCopy/ColorBufferToRDRAMStub.cpp → src/BufferCopy/ColorBufferToRDRAMStub.h

@@ -13,9 +13,3 @@ class ColorBufferToRDRAMStub : public ColorBufferToRDRAM
  bool _readPixels(GLint _x0, GLint _y0, GLsizei _width, GLsizei _height, u32 _size, bool _sync) override {}
  void _cleanUp() override {}
 };
-
-ColorBufferToRDRAM & ColorBufferToRDRAM::get()
-{
- static ColorBufferToRDRAMStub cbCopy;
- return cbCopy;
-}

src/BufferCopy/ColorBufferToRDRAM_BufferStorageExt.cpp

@@ -0,0 +1,104 @@
+#include <OpenGL.h>
+#include <Textures.h>
+#include <FBOTextureFormats.h>
+
+#if defined(EGL) || defined(GLESX)
+#include <inc/ARB_buffer_storage.h>
+#include <EGL/egl.h>
+#include <EGL/eglext.h>
+#endif
+
+#include "ColorBufferToRDRAM_BufferStorageExt.h"
+
+ColorBufferToRDRAM_BufferStorageExt::ColorBufferToRDRAM_BufferStorageExt()
+ : ColorBufferToRDRAM(), m_curIndex(0)
+{
+#ifdef GLESX
+ glBufferStorage = (PFNGLBUFFERSTORAGEPROC)eglGetProcAddress("glBufferStorageEXT");
+#endif
+}
+
+void ColorBufferToRDRAM_BufferStorageExt::_init()
+{
+ // Generate Pixel Buffer Objects
+ glGenBuffers(_numPBO, m_PBO);
+}
+
+void ColorBufferToRDRAM_BufferStorageExt::_destroy()
+{
+ for (int index = 0; index < _numPBO; ++index) {
+ if (m_PBOData[index] != nullptr) {
+ glBindBuffer(GL_PIXEL_PACK_BUFFER, m_PBO[index]);
+ glUnmapBuffer(GL_PIXEL_PACK_BUFFER);
+ }
+ }
+
+ glDeleteBuffers(_numPBO, m_PBO);
+
+ for (int index = 0; index < _numPBO; ++index) {
+ m_PBO[index] = 0;
+ }
+}
+
+void ColorBufferToRDRAM_BufferStorageExt::_initBuffers(void)
+{
+ // Initialize Pixel Buffer Objects
+ for (int index = 0; index < _numPBO; ++index) {
+ glBindBuffer(GL_PIXEL_PACK_BUFFER, m_PBO[index]);
+ m_fence[index] = 0;
+ glBufferStorage(GL_PIXEL_PACK_BUFFER, m_pTexture->textureBytes, nullptr, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
+ m_PBOData[index] = glMapBufferRange(GL_PIXEL_PACK_BUFFER, 0, m_pTexture->textureBytes, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT );
+ }
+
+ glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
+}
+
+bool ColorBufferToRDRAM_BufferStorageExt::_readPixels(GLint _x0, GLint _y0, GLsizei _width, GLsizei _height, u32 _size, bool _sync)
+{
+ GLenum colorFormat, colorType, colorFormatBytes;
+ if (_size > G_IM_SIZ_8b) {
+ colorFormat = fboFormats.colorFormat;
+ colorType = fboFormats.colorType;
+ colorFormatBytes = fboFormats.colorFormatBytes;
+ } else {
+ colorFormat = fboFormats.monochromeFormat;
+ colorType = fboFormats.monochromeType;
+ colorFormatBytes = fboFormats.monochromeFormatBytes;
+ }
+
+ glBindBuffer(GL_PIXEL_PACK_BUFFER, m_PBO[m_curIndex]);
+ glReadPixels(_x0, _y0, m_pTexture->realWidth, _height, colorFormat, colorType, 0);
+
+ //Setup a fence sync object so that we know when glReadPixels completes
+ glMemoryBarrier(GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
+ m_fence[m_curIndex] = glFenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, 0);
+
+ if (!_sync) {
+ m_curIndex = (m_curIndex+1)%_numPBO;
+ }
+
+ //Wait for glReadPixels to complete for the currently selected PBO
+ if (m_fence[m_curIndex] != 0) {
+ glClientWaitSync(m_fence[m_curIndex], 0, 1e8);
+ glDeleteSync(m_fence[m_curIndex]);
+ }
+
+ GLubyte* pixelData = reinterpret_cast<GLubyte*>(m_PBOData[m_curIndex]);
+ if (pixelData == nullptr)
+ return false;
+
+ int widthBytes = _width * colorFormatBytes;
+ int strideBytes = m_pTexture->realWidth * colorFormatBytes;
+
+ GLubyte* pixelDataAlloc = m_pixelData.data();
+ for (unsigned int lnIndex = 0; lnIndex < _height; ++lnIndex) {
+ memcpy(pixelDataAlloc + lnIndex*widthBytes, pixelData + (lnIndex*strideBytes), widthBytes);
+ }
+
+ return true;
+}
+
+void ColorBufferToRDRAM_BufferStorageExt::_cleanUp()
+{
+ glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
+}

src/BufferCopy/ColorBufferToRDRAM_BufferStorageExt.h

@@ -0,0 +1,26 @@
+#include "ColorBufferToRDRAM.h"
+
+#include "inc/ARB_buffer_storage.h"
+
+
+class ColorBufferToRDRAM_BufferStorageExt : public ColorBufferToRDRAM
+{
+public:
+ ColorBufferToRDRAM_BufferStorageExt();
+ ~ColorBufferToRDRAM_BufferStorageExt() = default;
+
+private:
+ void _init() override;
+ void _destroy() override;
+ bool _readPixels(GLint _x0, GLint _y0, GLsizei _width, GLsizei _height, u32 _size, bool _sync) override;
+ void _cleanUp() override;
+ void _initBuffers(void) override;
+ static const int _numPBO = 2;
+ GLuint m_PBO[_numPBO];
+ void* m_PBOData[_numPBO];
+ u32 m_curIndex;
+ GLsync m_fence[_numPBO];
+#ifdef GLESX
+ PFNGLBUFFERSTORAGEPROC glBufferStorage;
+#endif
+};

src/BufferCopy/ColorBufferToRDRAM_GL.cpp

@@ -1,30 +1,6 @@
 #include "ColorBufferToRDRAM.h"
-
-#include <Textures.h>
 #include <FBOTextureFormats.h>
-
-class ColorBufferToRDRAM_GL : public ColorBufferToRDRAM
-{
-public:
- ColorBufferToRDRAM_GL();
- ~ColorBufferToRDRAM_GL() {};
-
-private:
- void _init() override;
- void _destroy() override;
- bool _readPixels(GLint _x0, GLint _y0, GLsizei _width, GLsizei _height, u32 _size, bool _sync) override;
- void _cleanUp() override;
- void _initBuffers(void) override;
- static const int _numPBO = 3;
- GLuint m_PBO[_numPBO];
- u32 m_curIndex;
-};
-
-ColorBufferToRDRAM & ColorBufferToRDRAM::get()
-{
- static ColorBufferToRDRAM_GL cbCopy;
- return cbCopy;
-}
+#include "ColorBufferToRDRAM_GL.h"
 
 ColorBufferToRDRAM_GL::ColorBufferToRDRAM_GL()
  : ColorBufferToRDRAM()
@@ -81,7 +57,7 @@ bool ColorBufferToRDRAM_GL::_readPixels(GLint _x0, GLint _y0, GLsizei _width, GL
  glReadPixels(_x0, _y0, m_pTexture->realWidth, _height, colorFormat, colorType, 0);
  glBindBuffer(GL_PIXEL_PACK_BUFFER, m_PBO[nextIndex]);
  } else {
- glBindBuffer(GL_PIXEL_PACK_BUFFER, m_PBO[2]);
+ glBindBuffer(GL_PIXEL_PACK_BUFFER, m_PBO[_numPBO - 1]);
  glReadPixels(_x0, _y0, m_pTexture->realWidth, _height, colorFormat, colorType, 0);
  }
 

src/BufferCopy/ColorBufferToRDRAM_GL.h

@@ -0,0 +1,21 @@
+#include "ColorBufferToRDRAM.h"
+
+#include <Textures.h>
+#include <FBOTextureFormats.h>
+
+class ColorBufferToRDRAM_GL : public ColorBufferToRDRAM
+{
+public:
+ ColorBufferToRDRAM_GL();
+ ~ColorBufferToRDRAM_GL() {};
+
+private:
+ void _init() override;
+ void _destroy() override;
+ bool _readPixels(GLint _x0, GLint _y0, GLsizei _width, GLsizei _height, u32 _size, bool _sync) override;
+ void _cleanUp() override;
+ void _initBuffers(void) override;
+ static const int _numPBO = 3;
+ GLuint m_PBO[_numPBO];
+ u32 m_curIndex;
+};

src/BufferCopy/ColorBufferToRDRAM_GLES.cpp

@@ -1,5 +1,6 @@
 #include <algorithm>
 #include "ColorBufferToRDRAM.h"
+#include "ColorBufferToRDRAM_GLES.h"
 
 #include <Textures.h>
 #include <FBOTextureFormats.h>
@@ -9,39 +10,10 @@
 #include <EGL/egl.h>
 #include <EGL/eglext.h>
 
-typedef void (GL_APIENTRY* PFNGLEGLIMAGETARGETTEXTURE2DOESPROC) (GLenum target, EGLImageKHR image);
-typedef void (GL_APIENTRY* PFNGLEGLIMAGETARGETRENDERBUFFERSTORAGEOESPROC) (GLenum target, EGLImageKHR image);
-
-using namespace android;
-
-class ColorBufferToRDRAM_GLES : public ColorBufferToRDRAM
-{
-public:
- ColorBufferToRDRAM_GLES();
- ~ColorBufferToRDRAM_GLES() {};
-
-private:
- void _init() override;
- void _destroy() override;
- bool _readPixels(GLint _x0, GLint _y0, GLsizei _width, GLsizei _height, u32 _size, bool _sync) override;
- void _cleanUp() override;
- void _initBuffers(void) override;
-
- GraphicBuffer* m_window;
- EGLImageKHR m_image;
- PFNGLEGLIMAGETARGETTEXTURE2DOESPROC m_glEGLImageTargetTexture2DOES;
-};
-
-ColorBufferToRDRAM & ColorBufferToRDRAM::get()
-{
- static ColorBufferToRDRAM_GLES cbCopy;
- return cbCopy;
-}
-
-ColorBufferToRDRAM_GLES::ColorBufferToRDRAM_GLES() 
- : ColorBufferToRDRAM(),
- m_window(nullptr),
- m_image(0)
+ColorBufferToRDRAM_GLES::ColorBufferToRDRAM_GLES()
+ : ColorBufferToRDRAM(),
+ m_window(nullptr),
+ m_image(0)
 {
 }
 

src/BufferCopy/ColorBufferToRDRAM_GLES.h

@@ -0,0 +1,33 @@
+#include <algorithm>
+#include "ColorBufferToRDRAM.h"
+
+#include <Textures.h>
+#include <FBOTextureFormats.h>
+
+#include <ui/GraphicBuffer.h>
+#include <android/native_window.h>
+#include <EGL/egl.h>
+#include <EGL/eglext.h>
+
+typedef void (GL_APIENTRY* PFNGLEGLIMAGETARGETTEXTURE2DOESPROC) (GLenum target, EGLImageKHR image);
+typedef void (GL_APIENTRY* PFNGLEGLIMAGETARGETRENDERBUFFERSTORAGEOESPROC) (GLenum target, EGLImageKHR image);
+
+using namespace android;
+
+class ColorBufferToRDRAM_GLES : public ColorBufferToRDRAM
+{
+public:
+ ColorBufferToRDRAM_GLES();
+ ~ColorBufferToRDRAM_GLES() = default;
+
+private:
+ void _init() override;
+ void _destroy() override;
+ bool _readPixels(GLint _x0, GLint _y0, GLsizei _width, GLsizei _height, u32 _size, bool _sync) override;
+ void _cleanUp() override;
+ void _initBuffers(void) override;
+
+ GraphicBuffer* m_window;
+ EGLImageKHR m_image;
+ PFNGLEGLIMAGETARGETTEXTURE2DOESPROC m_glEGLImageTargetTexture2DOES;
+};

src/CMakeLists.txt

@@ -241,7 +241,6 @@ if(GLES2)
  )
  list(APPEND GLideN64_SOURCES 
  GLES2/GLSLCombiner_gles2.cpp
- BufferCopy/ColorBufferToRDRAMStub.cpp
  )
  SET(OPENGL_LIBRARIES -lGLESv2 -lEGL)
 else(GLES2)
@@ -255,6 +254,7 @@ else(GLES2)
  list(APPEND GLideN64_SOURCES 
  OGL3X/GLSLCombiner_ogl3x.cpp
  BufferCopy/ColorBufferToRDRAM_GL.cpp
+ BufferCopy/ColorBufferToRDRAM_BufferStorageExt.cpp
  )
 endif(GLES2)