GL3: Optimize volatile vertex buffers

--HG--
branch : opengl3

code/nel/src/3d/driver/opengl3/driver_opengl_extension.cpp

@@ -125,6 +125,9 @@ PFNGLUNMAPBUFFERPROC 							nglUnmapBuffer;
 PFNGLGETBUFFERPARAMETERIVPROC 					nglGetBufferParameteriv;
 PFNGLGETBUFFERPOINTERVPROC 						nglGetBufferPointerv;
 
+PFNGLMAPBUFFERRANGEPROC							nglMapBufferRange;
+PFNGLFLUSHMAPPEDBUFFERRANGEPROC					nglFlushMappedBufferRange;
+
 PFNGLGENQUERIESPROC								nglGenQueries;
 PFNGLDELETEQUERIESPROC							nglDeleteQueries;
 PFNGLISQUERYPROC								nglIsQuery;
@@ -420,6 +423,9 @@ static bool setupGLCore(std::vector<const char *> &glext)
 	CHECK_ADDRESS(PFNGLGETBUFFERPARAMETERIVPROC, glGetBufferParameteriv);
 	CHECK_ADDRESS(PFNGLGETBUFFERPOINTERVPROC, glGetBufferPointerv);
 
+	CHECK_ADDRESS(PFNGLMAPBUFFERRANGEPROC, glMapBufferRange);
+	CHECK_ADDRESS(PFNGLFLUSHMAPPEDBUFFERRANGEPROC, glFlushMappedBufferRange);
+
 	CHECK_ADDRESS(PFNGLGENQUERIESPROC, glGenQueries);
 	CHECK_ADDRESS(PFNGLDELETEQUERIESPROC, glDeleteQueries);
 	CHECK_ADDRESS(PFNGLISQUERYPROC, glIsQuery);

code/nel/src/3d/driver/opengl3/driver_opengl_extension.h

@@ -212,6 +212,9 @@ extern PFNGLUNMAPBUFFERPROC								nglUnmapBuffer;
 extern PFNGLGETBUFFERPARAMETERIVPROC					nglGetBufferParameteriv;
 extern PFNGLGETBUFFERPOINTERVPROC						nglGetBufferPointerv;
 
+extern PFNGLMAPBUFFERRANGEPROC							nglMapBufferRange;
+extern PFNGLFLUSHMAPPEDBUFFERRANGEPROC					nglFlushMappedBufferRange;
+
 extern PFNGLGENQUERIESPROC								nglGenQueries;
 extern PFNGLDELETEQUERIESPROC							nglDeleteQueries;
 extern PFNGLISQUERYPROC									nglIsQuery;

code/nel/src/3d/driver/opengl3/driver_opengl_vertex_buffer.cpp

@@ -87,7 +87,10 @@ void *CVertexBufferGL::lock()
 {
 	H_AUTO_OGL(CVertexBufferGLARB_lock);
 
-	if (m_VertexPtr) return m_VertexPtr; // already locked
+	if (m_VertexPtr) return m_VertexPtr; // already locked :|
+
+	const uint size = VB->getNumVertices() * VB->getVertexSize();
+
 	if (m_Invalid)
 	{
 		if (VB->getLocation() != CVertexBuffer::NotResident)
@@ -106,7 +109,6 @@ void *CVertexBufferGL::lock()
 			m_Driver->incrementResetCounter();
 			return &m_DummyVB[0];
 		}
-		const uint size = VB->getNumVertices() * VB->getVertexSize();
 		m_Driver->_DriverGLStates.forceBindARBVertexBuffer(vertexBufferID);
 		nglBufferData(GL_ARRAY_BUFFER, size, NULL, m_Driver->vertexBufferUsageGL3(m_MemType));
 		if (glGetError() != GL_NO_ERROR)
@@ -130,7 +132,27 @@ void *CVertexBufferGL::lock()
 	m_Driver->_DriverGLStates.bindARBVertexBuffer(VertexObjectId);
 
 
+	// m_VertexPtr = nglMapBuffer(GL_ARRAY_BUFFER, GL_WRITE_ONLY);
+	
+	// PERFORMANCE: AMD: This brings framerate from 24fps to 38fps, glitches with volatile buffers such as animated models and gui, likely glitches with others
+	// m_VertexPtr = nglMapBufferRange(GL_ARRAY_BUFFER, 0, size, GL_MAP_WRITE_BIT | GL_MAP_UNSYNCHRONIZED_BIT);
+
+	// PERFORMANCE: AMD: This brings framerate from 24fps to 38fps, glitches with landscape rendering
+	// m_VertexPtr = nglMapBufferRange(GL_ARRAY_BUFFER, 0, size, GL_MAP_WRITE_BIT | GL_MAP_INVALIDATE_BUFFER_BIT);
+
+	// Invalidate when updating volatile buffers, framerate from 24fps to 38fps in reference test on AMD platform
+	// TODO: Find where we can optimize with GL_MAP_UNSYNCHRONIZED_BIT
+	switch (m_MemType)
+	{
+	case CVertexBuffer::AGPVolatile:
+	case CVertexBuffer::RAMVolatile:
+		m_VertexPtr = nglMapBufferRange(GL_ARRAY_BUFFER, 0, size, GL_MAP_WRITE_BIT | GL_MAP_INVALIDATE_BUFFER_BIT);
+		break;
+	default:
 		m_VertexPtr = nglMapBuffer(GL_ARRAY_BUFFER, GL_WRITE_ONLY);
+		break;
+	}
+
 	if (!m_VertexPtr)
 	{
 		nglUnmapBuffer(GL_ARRAY_BUFFER);