Merge branch 'buffer_storage'

CMakeLists.txt

@@ -661,7 +661,7 @@ else()
 		if(NOT ${CMAKE_SYSTEM_NAME} MATCHES "Darwin")
 			include(FindGLEW)
 		endif()
-		if(NOT GLEW_FOUND OR NOT GLEW_HAS_1_9_METHODS)
+		if(NOT GLEW_FOUND OR NOT GLEW_HAS_1_10_METHODS)
 			message("Using static GLEW from Externals")
 			add_subdirectory(Externals/GLew)
 			include_directories(Externals/GLew/include)

CMakeTests/FindGLEW.cmake

@@ -10,13 +10,14 @@ macro(test_glew)
 	#include <GL/glew.h>
 	int main()
 	{
-	#ifdef GLEW_ARB_shader_image_load_store
+	#ifdef GLEW_ARB_buffer_storage
 	return 0;
 	#else
 	return 1;
 	#endif
 	}"
-	GLEW_HAS_1_9_METHODS)
+	GLEW_HAS_1_10_METHODS)
+
 endmacro()
 
 if(PKG_CONFIG_FOUND AND NOT ${var}_FOUND)
@@ -25,7 +26,7 @@ endif()
 
 if(GLEW_FOUND)
 	test_glew()
-	if (GLEW_HAS_1_9_METHODS)
+	if (GLEW_HAS_1_10_METHODS)
 		include_directories(${GLEW_INCLUDE_DIRS})
 		message("GLEW found")
 	endif()

Source/Core/VideoBackends/OGL/GLUtil.h

@@ -30,10 +30,13 @@
 #define GL_READ_WRITE                     0x88BA
 #define GL_SRC1_ALPHA 0
 #define GL_BGRA GL_RGBA
+#define GL_MAP_COHERENT_BIT 0
+#define GL_MAP_PERSISTENT_BIT 0
 #define glDrawElementsBaseVertex(...)
 #define glDrawRangeElementsBaseVertex(...)
 #define glRenderbufferStorageMultisampleCoverageNV(...)
 #define glViewportIndexedf(...)
+#define glBufferStorage(...)
 #endif
 #else
 #define TEX2D	GL_TEXTURE_RECTANGLE_ARB

Source/Core/VideoBackends/OGL/Render.cpp

@@ -486,7 +486,9 @@ Renderer::Renderer()
 	g_ogl_config.bSupportsGLSLCache = TO_BOOL(GLEW_ARB_get_program_binary);
 	g_ogl_config.bSupportsGLPinnedMemory = TO_BOOL(GLEW_AMD_pinned_memory);
 	g_ogl_config.bSupportsGLSync = TO_BOOL(GLEW_ARB_sync);
-	g_ogl_config.bSupportsGLBaseVertex = TO_BOOL(GLEW_ARB_draw_elements_base_vertex);
+	g_ogl_config.bSupportsGLBaseVertex = TO_BOOL(GLEW_ARB_draw_elements_base_vertex) &&
+						!DriverDetails::HasBug(DriverDetails::BUG_BROKENPINNEDMEMORY);
+	g_ogl_config.bSupportsGLBufferStorage = TO_BOOL(GLEW_ARB_buffer_storage);
 	g_ogl_config.bSupportCoverageMSAA = TO_BOOL(GLEW_NV_framebuffer_multisample_coverage);
 	g_ogl_config.bSupportSampleShading = TO_BOOL(GLEW_ARB_sample_shading);
 	g_ogl_config.bSupportOGL31 = TO_BOOL(GLEW_VERSION_3_1);
@@ -555,7 +557,7 @@ Renderer::Renderer()
 				g_ogl_config.gl_renderer,
 				g_ogl_config.gl_version), 5000);
 
-	WARN_LOG(VIDEO,"Missing OGL Extensions: %s%s%s%s%s%s%s%s%s%s",
+	WARN_LOG(VIDEO,"Missing OGL Extensions: %s%s%s%s%s%s%s%s%s%s%s",
 			g_ActiveConfig.backend_info.bSupportsDualSourceBlend ? "" : "DualSourceBlend ",
 			g_ActiveConfig.backend_info.bSupportsGLSLUBO ? "" : "UniformBuffer ",
 			g_ActiveConfig.backend_info.bSupportsPrimitiveRestart ? "" : "PrimitiveRestart ",
@@ -563,6 +565,7 @@ Renderer::Renderer()
 			g_ogl_config.bSupportsGLPinnedMemory ? "" : "PinnedMemory ",
 			g_ogl_config.bSupportsGLSLCache ? "" : "ShaderCache ",
 			g_ogl_config.bSupportsGLBaseVertex ? "" : "BaseVertex ",
+			g_ogl_config.bSupportsGLBufferStorage ? "" : "BufferStorage ",
 			g_ogl_config.bSupportsGLSync ? "" : "Sync ",
 			g_ogl_config.bSupportCoverageMSAA ? "" : "CSAA ",
 			g_ogl_config.bSupportSampleShading ? "" : "SSAA "

Source/Core/VideoBackends/OGL/Render.h

@@ -22,6 +22,7 @@ extern struct VideoConfig {
 	bool bSupportsGLPinnedMemory;
 	bool bSupportsGLSync;
 	bool bSupportsGLBaseVertex;
+	bool bSupportsGLBufferStorage;
 	bool bSupportCoverageMSAA;
 	bool bSupportSampleShading;
 	GLSL_VERSION eSupportedGLSLVersion;

Source/Core/VideoBackends/OGL/StreamBuffer.cpp

@@ -33,7 +33,11 @@ StreamBuffer::StreamBuffer(u32 type, size_t size, StreamType uploadType)
 			g_Config.bHackedBufferUpload = false;
 		}
 
-		if(!g_ogl_config.bSupportsGLBaseVertex && (m_uploadtype & BUFFERSUBDATA)
+		if (g_ogl_config.bSupportsGLBufferStorage && 
+			!(DriverDetails::HasBug(DriverDetails::BUG_BROKENBUFFERSTORAGE) && type == GL_ARRAY_BUFFER) && 
+			(m_uploadtype & BUFFERSTORAGE))
+			m_uploadtype = BUFFERSTORAGE;
+		else if(!g_ogl_config.bSupportsGLBaseVertex && (m_uploadtype & BUFFERSUBDATA)
 			&& !DriverDetails::HasBug(DriverDetails::BUG_BROKENBUFFERSTREAM))
 			m_uploadtype = BUFFERSUBDATA;
 		else if(!g_ogl_config.bSupportsGLBaseVertex && (m_uploadtype & BUFFERDATA))
@@ -81,9 +85,9 @@ void StreamBuffer::Alloc ( size_t size, u32 stride )
 		break;
 	case MAP_AND_SYNC:
 	case PINNED_MEMORY:
-
+	case BUFFERSTORAGE:
 		// insert waiting slots for used memory
-		for(size_t i=SLOT(m_used_iterator); i<SLOT(m_iterator); i++)
+		for (size_t i = SLOT(m_used_iterator); i < SLOT(m_iterator); i++)
 		{
 			fences[i] = glFenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, 0);
 		}
@@ -98,7 +102,7 @@ void StreamBuffer::Alloc ( size_t size, u32 stride )
 		m_free_iterator = iter_end;
 
 		// if buffer is full
-		if(iter_end >= m_size) {
+		if (iter_end >= m_size) {
 
 			// insert waiting slots in unused space at the end of the buffer
 			for (size_t i = SLOT(m_used_iterator); i < SYNC_POINTS; i++)
@@ -111,7 +115,7 @@ void StreamBuffer::Alloc ( size_t size, u32 stride )
 			iter_end = size;
 
 			// wait for space at the start
-			for(u32 i=0; i<=SLOT(iter_end); i++)
+			for (u32 i = 0; i <= SLOT(iter_end); i++)
 			{
 				glClientWaitSync(fences[i], GL_SYNC_FLUSH_COMMANDS_BIT, GL_TIMEOUT_IGNORED);
 				glDeleteSync(fences[i]);
@@ -130,7 +134,7 @@ void StreamBuffer::Alloc ( size_t size, u32 stride )
 		m_iterator_aligned = 0;
 		break;
 	case STREAM_DETECT:
-	case DETECT_MASK: // Just to shutup warnings
+	case DETECT_MASK: // To shutup compiler warnings
 		break;
 	}
 	m_iterator = m_iterator_aligned;
@@ -151,8 +155,9 @@ size_t StreamBuffer::Upload ( u8* data, size_t size )
 		break;
 	case PINNED_MEMORY:
 	case MAP_AND_RISK:
-		if(pointer)
+	case BUFFERSTORAGE:
-			memcpy(pointer+m_iterator, data, size);
+		if (pointer)
+			memcpy(pointer + m_iterator, data, size);
 		break;
 	case BUFFERSUBDATA:
 		glBufferSubData(m_buffertype, m_iterator, size, data);
@@ -161,7 +166,7 @@ size_t StreamBuffer::Upload ( u8* data, size_t size )
 		glBufferData(m_buffertype, size, data, GL_STREAM_DRAW);
 		break;
 	case STREAM_DETECT:
-	case DETECT_MASK: // Just to shutup warnings
+	case DETECT_MASK: // To shutup compiler warnings
 		break;
 	}
 	size_t ret = m_iterator;
@@ -206,6 +211,26 @@ void StreamBuffer::Init()
 			Init();
 		}
 		break;
+
+	case BUFFERSTORAGE:
+		glGetError(); // errors before this allocation should be ignored
+		fences = new GLsync[SYNC_POINTS];
+		for (u32 i = 0; i<SYNC_POINTS; i++)
+			fences[i] = glFenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, 0);
+
+		glBindBuffer(m_buffertype, m_buffer);
+
+		// PERSISTANT_BIT to make sure that the buffer can be used while mapped
+		// COHERENT_BIT is set so we don't have to use a MemoryBarrier on write
+		// CLIENT_STORAGE_BIT is set since we access the buffer more frequently on the client side then server side
+		glBufferStorage(m_buffertype, m_size, NULL, 
+			GL_MAP_WRITE_BIT | GL_MAP_PERSISTENT_BIT | GL_MAP_COHERENT_BIT | GL_CLIENT_STORAGE_BIT); 
+		pointer = (u8*)glMapBufferRange(m_buffertype, 0, m_size, 
+			GL_MAP_WRITE_BIT | GL_MAP_PERSISTENT_BIT | GL_MAP_COHERENT_BIT); 
+		if(!pointer)
+			ERROR_LOG(VIDEO, "Buffer allocation failed");
+		break;
+
 	case MAP_AND_RISK:
 		glBindBuffer(m_buffertype, m_buffer);
 		glBufferData(m_buffertype, m_size, NULL, GL_STREAM_DRAW);
@@ -218,7 +243,7 @@ void StreamBuffer::Init()
 		glBindBuffer(m_buffertype, m_buffer);
 		break;
 	case STREAM_DETECT:
-	case DETECT_MASK: // Just to shutup warnings
+	case DETECT_MASK: // To shutup compiler warnings
 		break;
 	}
 }
@@ -240,8 +265,14 @@ void StreamBuffer::Shutdown()
 		glFinish(); // ogl pipeline must be flushed, else this buffer can be in use
 		FreeAlignedMemory(pointer);
 		break;
+	case BUFFERSTORAGE:
+		DeleteFences();
+		glUnmapBuffer(m_buffertype);
+		glBindBuffer(m_buffertype, 0);
+		glFinish(); // ogl pipeline must be flushed, else this buffer can be in use
+		break;
 	case STREAM_DETECT:
-	case DETECT_MASK: // Just to shutup warnings
+	case DETECT_MASK: // To shutup compiler warnings
 		break;
 	}
 }

Source/Core/VideoBackends/OGL/StreamBuffer.h

@@ -18,14 +18,15 @@
 namespace OGL
 {
 enum StreamType {
-	DETECT_MASK = 0x7F,
+	DETECT_MASK = 0xFF,
 	STREAM_DETECT = (1 << 0),
 	MAP_AND_ORPHAN = (1 << 1),
 	MAP_AND_SYNC = (1 << 2),
 	MAP_AND_RISK = (1 << 3),
 	PINNED_MEMORY = (1 << 4),
 	BUFFERSUBDATA = (1 << 5),
-	BUFFERDATA = (1 << 6)
+	BUFFERDATA = (1 << 6),
+	BUFFERSTORAGE = (1 << 7),
 };
 
 class StreamBuffer {

Source/Core/VideoCommon/DriverDetails.cpp

@@ -54,6 +54,7 @@ namespace DriverDetails
 		{OS_ALL,    VENDOR_ATI,      DRIVER_ATI,          -1, BUG_BROKENHACKEDBUFFER,  -1.0, -1.0, true},
 		{OS_LINUX,  VENDOR_ATI,      DRIVER_ATI,          -1, BUG_BROKENPINNEDMEMORY,  -1.0, -1.0, true},
 		{OS_ALL,    VENDOR_MESA,     DRIVER_NOUVEAU,      -1, BUG_BROKENHACKEDBUFFER,  -1.0, -1.0, true},
+		{OS_ALL,    VENDOR_NVIDIA,   DRIVER_NVIDIA,       -1, BUG_BROKENBUFFERSTORAGE, -1.0, -1.0, true},
 		{OS_OSX,    VENDOR_INTEL,    DRIVER_INTEL,      3000, BUG_PRIMITIVERESTART,    -1.0, -1.0, true},
 	};
 

Source/Core/VideoCommon/DriverDetails.h

@@ -141,6 +141,14 @@ namespace DriverDetails
 		// Ended Version: -1
 		// If a shader includes a textureSize function call then the shader compiler will call abort()
 		BUG_BROKENTEXTURESIZE,
+		// Bug: ARB_buffer_storage doesn't work with ARRAY_BUFFER type streams 
+		// Affected devices: Geforce 4xx+
+		// Started Version: -1
+		// Ended Version: -1
+		// The buffer_storage streaming method is required for greater speed gains in our buffer streaming
+		// It reduces what is needed for streaming to basically a memcpy call
+		// It seems to work for all buffer types except GL_ARRAY_BUFFER
+		BUG_BROKENBUFFERSTORAGE,
 		// Bug: Intel HD 3000 on OS X has broken primitive restart
 		// Affected devices: Intel HD 3000
 		// Affected OS: OS X