Only update PushBuffer if any uniform has actually changed

src/engine/renderer/BufferBind.h

@@ -42,9 +42,10 @@ namespace BufferBind {
 		MATERIALS = 0,
 		TEX_DATA = 1,
 		LIGHTMAP_DATA = 2,
-		LIGHTS = 3,
+		GLOBAL_DATA = 3,
+		LIGHTS = 4,
 
-		SURFACE_BATCHES = 4,
+		SURFACE_BATCHES = 5,
 
 		// SSBO
 		SURFACE_DESCRIPTORS = 0,

src/engine/renderer/GLMemory.cpp

@@ -37,10 +37,13 @@ SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #include "GLMemory.h"
 
+#include "gl_shader.h"
+
 // 128 MB, should be enough to fit anything in BAR without going overboard
 const GLsizeiptr GLStagingBuffer::SIZE = 128 * 1024 * 1024 / sizeof( uint32_t );
 
 GLStagingBuffer stagingBuffer;
+PushBuffer pushBuffer;
 
 void GLBufferCopy( GLBuffer* src, GLBuffer* dst, GLintptr srcOffset, GLintptr dstOffset, GLsizeiptr size ) {
 	glCopyNamedBufferSubData( src->id, dst->id,
@@ -114,6 +117,10 @@ void GLStagingBuffer::FlushAll() {
 	FlushStagingCopyQueue();
 }
 
+bool GLStagingBuffer::Active() const {
+	return buffer.id;
+}
+
 void GLStagingBuffer::InitGLBuffer() {
 	buffer.GenBuffer();
 
@@ -130,3 +137,36 @@ void GLStagingBuffer::FreeGLBuffer() {
 	current = 0;
 	last = 0;
 }
+
+void PushBuffer::InitGLBuffers() {
+	globalUBO.GenBuffer();
+
+	globalUBO.BufferStorage( pushBuffer.constUniformsSize + pushBuffer.frameUniformsSize, 1, nullptr );
+
+	globalUBO.BindBufferBase();
+}
+
+void PushBuffer::FreeGLBuffers() {
+	globalUBO.DelBuffer();
+}
+
+uint32_t* PushBuffer::MapGlobalUniformData( const int updateType ) {
+	switch ( updateType ) {
+		case GLUniform::CONST:
+			globalUBOData = stagingBuffer.MapBuffer( constUniformsSize );
+			stagingBuffer.QueueStagingCopy( &globalUBO, 0 );
+			break;
+		case GLUniform::FRAME:
+			globalUBOData = stagingBuffer.MapBuffer( frameUniformsSize );
+			stagingBuffer.QueueStagingCopy( &globalUBO, constUniformsSize );
+			break;
+		default:
+			ASSERT_UNREACHABLE();
+	}
+
+	return globalUBOData;
+}
+
+void PushBuffer::PushGlobalUniforms() {
+	stagingBuffer.FlushAll();
+}

src/engine/renderer/GLMemory.h

@@ -175,6 +175,7 @@ class GLBuffer {
 
 	void DelBuffer() {
 		glDeleteBuffers( 1, &id );
+		id = 0;
 		mapped = false;
 	}
 
@@ -303,6 +304,8 @@ class GLStagingBuffer {
 	void FlushStagingCopyQueue();
 	void FlushAll();
 
+	bool Active() const;
+
 	void InitGLBuffer();
 	void FreeGLBuffer();
 
@@ -321,6 +324,28 @@ class GLStagingBuffer {
 		GL_MAP_FLUSH_EXPLICIT_BIT | GL_MAP_INVALIDATE_BUFFER_BIT );
 };
 
+struct PushBuffer {
+	uint32_t constUniformsSize;
+	uint32_t frameUniformsSize;
+	uint32_t globalUBOSize;
+	uint32_t* globalUBOData;
+
+	uint32_t pushStart = UINT32_MAX;
+	uint32_t pushEnd = 0;
+
+	uint32_t sector = 0;
+	const uint32_t MAX_SECTORS = 1024;
+
+	GLUBO globalUBO = GLUBO( "globalUniforms", BufferBind::GLOBAL_DATA, 0, 0 );
+
+	void InitGLBuffers();
+	void FreeGLBuffers();
+
+	uint32_t* MapGlobalUniformData( const int updateType );
+	void PushGlobalUniforms();
+};
+
 extern GLStagingBuffer stagingBuffer;
+extern PushBuffer pushBuffer;
 
-#endif // GLMEMORY_H
+#endif // GLMEMORY_H

src/engine/renderer/Material.cpp

@@ -161,7 +161,7 @@ void UpdateSurfaceDataGeneric3D( uint32_t* materials, shaderStage_t* pStage, boo
 
 	gl_genericShaderMaterial->SetUniform_DepthScale( pStage->depthFadeValue );
 
-	gl_genericShaderMaterial->WriteUniformsToBuffer( materials );
+	gl_genericShaderMaterial->WriteUniformsToBuffer( materials, GLShader::MATERIAL );
 }
 
 void UpdateSurfaceDataLightMapping( uint32_t* materials, shaderStage_t* pStage, bool, bool vertexLit, bool fullbright ) {
@@ -207,7 +207,7 @@ void UpdateSurfaceDataLightMapping( uint32_t* materials, shaderStage_t* pStage,
 
 	gl_lightMappingShaderMaterial->SetUniform_SpecularExponent( specExpMin, specExpMax );
 
-	gl_lightMappingShaderMaterial->WriteUniformsToBuffer( materials );
+	gl_lightMappingShaderMaterial->WriteUniformsToBuffer( materials, GLShader::MATERIAL );
 }
 
 void UpdateSurfaceDataReflection( uint32_t* materials, shaderStage_t* pStage, bool, bool, bool ) {
@@ -244,7 +244,7 @@ void UpdateSurfaceDataReflection( uint32_t* materials, shaderStage_t* pStage, bo
 	gl_reflectionShaderMaterial->SetUniform_ReliefDepthScale( depthScale );
 	gl_reflectionShaderMaterial->SetUniform_ReliefOffsetBias( shader->reliefOffsetBias );
 
-	gl_reflectionShaderMaterial->WriteUniformsToBuffer( materials );
+	gl_reflectionShaderMaterial->WriteUniformsToBuffer( materials, GLShader::MATERIAL );
 }
 
 void UpdateSurfaceDataSkybox( uint32_t* materials, shaderStage_t* pStage, bool, bool, bool ) {
@@ -255,7 +255,7 @@ void UpdateSurfaceDataSkybox( uint32_t* materials, shaderStage_t* pStage, bool,
 	// u_AlphaThreshold
 	gl_skyboxShaderMaterial->SetUniform_AlphaTest( GLS_ATEST_NONE );
 
-	gl_skyboxShaderMaterial->WriteUniformsToBuffer( materials );
+	gl_skyboxShaderMaterial->WriteUniformsToBuffer( materials, GLShader::MATERIAL );
 }
 
 void UpdateSurfaceDataScreen( uint32_t* materials, shaderStage_t* pStage, bool, bool, bool ) {
@@ -268,7 +268,7 @@ void UpdateSurfaceDataScreen( uint32_t* materials, shaderStage_t* pStage, bool,
 	this seems to be the only material system shader that might need it to not be global */
 	gl_screenShaderMaterial->SetUniform_CurrentMapBindless( BindAnimatedImage( 0, &pStage->bundle[TB_COLORMAP] ) );
 
-	gl_screenShaderMaterial->WriteUniformsToBuffer( materials );
+	gl_screenShaderMaterial->WriteUniformsToBuffer( materials, GLShader::MATERIAL );
 }
 
 void UpdateSurfaceDataHeatHaze( uint32_t* materials, shaderStage_t* pStage, bool, bool, bool ) {
@@ -285,7 +285,7 @@ void UpdateSurfaceDataHeatHaze( uint32_t* materials, shaderStage_t* pStage, bool
 	// bind u_NormalScale
 	gl_heatHazeShaderMaterial->SetUniform_NormalScale( normalScale );
 
-	gl_heatHazeShaderMaterial->WriteUniformsToBuffer( materials );
+	gl_heatHazeShaderMaterial->WriteUniformsToBuffer( materials, GLShader::MATERIAL );
 }
 
 void UpdateSurfaceDataLiquid( uint32_t* materials, shaderStage_t* pStage, bool, bool, bool ) {
@@ -331,15 +331,15 @@ void UpdateSurfaceDataLiquid( uint32_t* materials, shaderStage_t* pStage, bool,
 
 	gl_liquidShaderMaterial->SetUniform_NormalScale( normalScale );
 
-	gl_liquidShaderMaterial->WriteUniformsToBuffer( materials );
+	gl_liquidShaderMaterial->WriteUniformsToBuffer( materials, GLShader::MATERIAL );
 }
 
 void UpdateSurfaceDataFog( uint32_t* materials, shaderStage_t* pStage, bool, bool, bool ) {
 	// shader_t* shader = pStage->shader;
 
 	materials += pStage->bufferOffset;
 
-	gl_fogQuake3ShaderMaterial->WriteUniformsToBuffer( materials );
+	gl_fogQuake3ShaderMaterial->WriteUniformsToBuffer( materials, GLShader::MATERIAL );
 }
 
 /*
@@ -1593,9 +1593,24 @@ void MaterialSystem::UpdateDynamicSurfaces() {
 	GL_CheckErrors();
 }
 
+void MaterialSystem::SetConstUniforms() {
+	globalUBOProxy->SetUniform_SurfaceDescriptorsCount( surfaceDescriptorsCount );
+	uint32_t globalWorkGroupX = surfaceDescriptorsCount % MAX_COMMAND_COUNTERS == 0 ?
+		surfaceDescriptorsCount / MAX_COMMAND_COUNTERS : surfaceDescriptorsCount / MAX_COMMAND_COUNTERS + 1;
+
+	globalUBOProxy->SetUniform_FirstPortalGroup( globalWorkGroupX );
+	globalUBOProxy->SetUniform_TotalPortals( totalPortals );
+}
+
+void MaterialSystem::SetFrameUniforms() {
+	globalUBOProxy->SetUniform_Frame( nextFrame );
+
+	globalUBOProxy->SetUniform_UseFrustumCulling( r_gpuFrustumCulling.Get() );
+	globalUBOProxy->SetUniform_UseOcclusionCulling( r_gpuOcclusionCulling.Get() );
+}
+
 void MaterialSystem::UpdateFrameData() {
 	gl_clearSurfacesShader->BindProgram( 0 );
-	gl_clearSurfacesShader->SetUniform_Frame( nextFrame );
 	gl_clearSurfacesShader->DispatchCompute( MAX_VIEWS, 1, 1 );
 
 	GL_CheckErrors();
@@ -1624,8 +1639,10 @@ void MaterialSystem::DepthReduction() {
 	uint32_t globalWorkgroupX = ( width + 7 ) / 8;
 	uint32_t globalWorkgroupY = ( height + 7 ) / 8;
 
-	// FIXME: u_DepthMap object on the shader is not actually used
-	GL_Bind( tr.currentDepthImage );
+	gl_depthReductionShader->SetUniform_DepthMapBindless(
+		GL_BindToTMU( 0, tr.currentDepthImage )
+	);
+
 	glBindImageTexture( 2, depthImage->texnum, 0, GL_FALSE, 0, GL_WRITE_ONLY, GL_R32F );
 
 	gl_depthReductionShader->SetUniform_InitialDepthLevel( true );
@@ -1681,15 +1698,9 @@ void MaterialSystem::CullSurfaces() {
 		uint32_t globalWorkGroupX = surfaceDescriptorsCount % MAX_COMMAND_COUNTERS == 0 ?
 			surfaceDescriptorsCount / MAX_COMMAND_COUNTERS : surfaceDescriptorsCount / MAX_COMMAND_COUNTERS + 1;
 		GL_Bind( depthImage );
-		gl_cullShader->SetUniform_Frame( nextFrame );
 		gl_cullShader->SetUniform_ViewID( view );
-		gl_cullShader->SetUniform_SurfaceDescriptorsCount( surfaceDescriptorsCount );
-		gl_cullShader->SetUniform_UseFrustumCulling( r_gpuFrustumCulling.Get() );
-		gl_cullShader->SetUniform_UseOcclusionCulling( r_gpuOcclusionCulling.Get() );
 		gl_cullShader->SetUniform_CameraPosition( origin );
 		gl_cullShader->SetUniform_ModelViewMatrix( viewMatrix );
-		gl_cullShader->SetUniform_FirstPortalGroup( globalWorkGroupX );
-		gl_cullShader->SetUniform_TotalPortals( totalPortals );
 		gl_cullShader->SetUniform_ViewWidth( depthImage->width );
 		gl_cullShader->SetUniform_ViewHeight( depthImage->height );
 		gl_cullShader->SetUniform_SurfaceCommandsOffset( surfaceCommandsCount * ( MAX_VIEWS * nextFrame + view ) );
@@ -1721,7 +1732,6 @@ void MaterialSystem::CullSurfaces() {
 		gl_cullShader->DispatchCompute( globalWorkGroupX, 1, 1 );
 
 		gl_processSurfacesShader->BindProgram( 0 );
-		gl_processSurfacesShader->SetUniform_Frame( nextFrame );
 		gl_processSurfacesShader->SetUniform_ViewID( view );
 		gl_processSurfacesShader->SetUniform_SurfaceCommandsOffset( surfaceCommandsCount * ( MAX_VIEWS * nextFrame + view ) );
 

src/engine/renderer/Material.h

@@ -362,6 +362,8 @@ class MaterialSystem {
 
 	void StartFrame();
 	void EndFrame();
+	void SetConstUniforms();
+	void SetFrameUniforms();
 
 	void GenerateDepthImages( const int width, const int height, imageParams_t imageParms );