Optimized perf by fixing several issues with the GPU<->CPU syncing

neo/renderer/NVRHI/Framebuffer_NVRHI.cpp

@@ -38,11 +38,7 @@ extern DeviceManager* deviceManager;
 
 static void R_ListFramebuffers_f( const idCmdArgs& args )
 {
-	if( !glConfig.framebufferObjectAvailable )
-	{
-		common->Printf( "GL_EXT_framebuffer_object is not available.\n" );
-		return;
-	}
+	// TODO
 }
 
 Framebuffer::Framebuffer( const char* name, int w, int h )

neo/renderer/NVRHI/Image_NVRHI.cpp

@@ -357,6 +357,7 @@ void idImage::AllocImage()
 	uint scaledWidth = originalWidth;
 	uint scaledHeight = originalHeight;
 
+#if 0
 	uint maxTextureSize = 0;
 
 	if( maxTextureSize > 0 &&
@@ -375,6 +376,7 @@ void idImage::AllocImage()
 			scaledHeight = maxTextureSize;
 		}
 	}
+#endif
 
 	auto textureDesc = nvrhi::TextureDesc()
 					   .setDebugName( GetName() )

neo/renderer/NVRHI/RenderBackend_NVRHI.cpp

@@ -61,6 +61,7 @@ public:
 
 	int										currentImageParm = 0;
 	idArray< idImage*, MAX_IMAGE_PARMS >	imageParms;
+	idScreenRect							scissor;		// set by GL_Scissor
 	//nvrhi::GraphicsPipelineHandle			pipeline;
 	//bool									fullscreen = false;
 };
@@ -172,9 +173,6 @@ idRenderBackend::DrawElementsWithCounters
 */
 void idRenderBackend::DrawElementsWithCounters( const drawSurf_t* surf )
 {
-	// Only update the constant buffer if it was updated at all.
-	renderProgManager.CommitConstantBuffer( commandList );
-
 	// Get vertex buffer
 	const vertCacheHandle_t vbHandle = surf->ambientCache;
 	idVertexBuffer* vertexBuffer;
@@ -265,6 +263,9 @@ void idRenderBackend::DrawElementsWithCounters( const drawSurf_t* surf )
 		changeState = true;
 	}
 
+	// TODO: Only update the constant buffer if it was updated at all.
+	renderProgManager.CommitConstantBuffer( commandList );
+
 	if( changeState )
 	{
 		nvrhi::GraphicsState state;
@@ -285,11 +286,17 @@ void idRenderBackend::DrawElementsWithCounters( const drawSurf_t* surf )
 								  ( float )currentViewport.y2,
 								  currentViewport.zmin,
 								  currentViewport.zmax };
-		state.viewport.addViewportAndScissorRect( viewport );
+		state.viewport.addViewport( viewport );
 
-		if( !currentScissor.IsEmpty() )
+#if 0
+		if( !context.scissor.IsEmpty() )
 		{
-			state.viewport.addScissorRect( nvrhi::Rect( currentScissor.x1, currentScissor.x2, currentScissor.y1, currentScissor.y2 ) );
+			state.viewport.addScissorRect( nvrhi::Rect( context.scissor.x1, context.scissor.x2, context.scissor.y1, context.scissor.y2 ) );
+		}
+		else
+#endif
+		{
+			state.viewport.addScissorRect( nvrhi::Rect( viewport ) );
 		}
 
 		commandList->setGraphicsState( state );
@@ -677,28 +684,6 @@ void idRenderBackend::GL_EndFrame()
 	commandList->close();
 
 	deviceManager->GetDevice()->executeCommandList( commandList );
-
-	// Make sure that all frames have finished rendering
-	deviceManager->GetDevice()->waitForIdle();
-
-	// Release all in-flight references to the render targets
-	deviceManager->GetDevice()->runGarbageCollection();
-
-	// Present to the swap chain.
-	deviceManager->Present();
-}
-
-void idRenderBackend::GL_EndRenderPass()
-{
-#if defined( USE_NVRHI )
-	commandList->close();
-
-	deviceManager->GetDevice()->executeCommandList( commandList );
-
-	deviceManager->GetDevice()->runGarbageCollection();
-
-	commandList->open();
-#endif
 }
 
 /*
@@ -712,6 +697,27 @@ void idRenderBackend::GL_BlockingSwapBuffers()
 {
 	// Make sure that all frames have finished rendering
 	deviceManager->GetDevice()->waitForIdle();
+
+	// Release all in-flight references to the render targets
+	deviceManager->GetDevice()->runGarbageCollection();
+
+	// Present to the swap chain.
+	deviceManager->Present();
+
+	renderLog.EndFrame();
+}
+
+void idRenderBackend::GL_EndRenderPass()
+{
+#if 0
+	commandList->close();
+
+	deviceManager->GetDevice()->executeCommandList( commandList );
+
+	deviceManager->GetDevice()->runGarbageCollection();
+
+	commandList->open();
+#endif
 }
 
 /*
@@ -773,9 +779,9 @@ idRenderBackend::GL_Scissor
 void idRenderBackend::GL_Scissor( int x /* left*/, int y /* bottom */, int w, int h )
 {
 	// TODO Check if this is right.
-	//currentScissor.Clear();
-	//currentScissor.AddPoint( x, y );
-	//currentScissor.AddPoint( x + w, y + h );
+	context.scissor.Clear();
+	context.scissor.AddPoint( x, y );
+	context.scissor.AddPoint( x + w, y + h );
 }
 
 /*
@@ -921,23 +927,6 @@ void idRenderBackend::CheckCVars()
 		}
 	}*/
 
-	if( r_useSeamlessCubeMap.IsModified() )
-	{
-		r_useSeamlessCubeMap.ClearModified();
-		if( glConfig.seamlessCubeMapAvailable )
-		{
-			if( r_useSeamlessCubeMap.GetBool() )
-			{
-				//glEnable( GL_TEXTURE_CUBE_MAP_SEAMLESS );
-			}
-			else
-			{
-				//glDisable( GL_TEXTURE_CUBE_MAP_SEAMLESS );
-			}
-		}
-	}
-
-
 	// SRS - Enable SDL-driven vync changes without restart for UNIX-like OSs
 #if defined(__linux__) || defined(__FreeBSD__) || defined(__APPLE__)
 	extern idCVar r_swapInterval;
@@ -958,6 +947,7 @@ void idRenderBackend::CheckCVars()
 	}
 #endif
 	// SRS end
+
 	if( r_antiAliasing.IsModified() )
 	{
 		switch( r_antiAliasing.GetInteger() )
@@ -977,6 +967,7 @@ void idRenderBackend::CheckCVars()
 		}
 	}
 
+	/*
 	if( r_usePBR.IsModified() ||
 			r_useHDR.IsModified() ||
 			r_useHalfLambertLighting.IsModified() ||
@@ -1002,20 +993,7 @@ void idRenderBackend::CheckCVars()
 		renderProgManager.KillAllShaders();
 		renderProgManager.LoadAllShaders();
 	}
-
-	// RB: turn off shadow mapping for OpenGL drivers that are too slow
-	switch( glConfig.driverType )
-	{
-		case GLDRV_OPENGL_ES2:
-		case GLDRV_OPENGL_ES3:
-			//case GLDRV_OPENGL_MESA:
-			r_useShadowMapping.SetInteger( 0 );
-			break;
-
-		default:
-			break;
-	}
-	// RB end
+	*/
 }
 
 /*
@@ -1324,6 +1302,7 @@ idRenderBackend::idRenderBackend()
 	memset( &glConfig, 0, sizeof( glConfig ) );
 
 	//glConfig.gpuSkinningAvailable = true;
+	glConfig.timerQueryAvailable = true;
 }
 
 /*

neo/renderer/NVRHI/RenderProgs_NVRHI.cpp

@@ -314,4 +314,6 @@ idRenderProgManager::CommitConstantBuffer
 void idRenderProgManager::CommitConstantBuffer( nvrhi::ICommandList* commandList )
 {
 	commandList->writeBuffer( constantBuffer, uniforms.Ptr(), uniforms.Allocated() );
+
+	//commandList->setPushConstants( uniforms.Ptr(), uniforms.Allocated() );
 }

neo/renderer/RenderBackend.cpp

@@ -5910,12 +5910,12 @@ void idRenderBackend::ExecuteBackEndCommands( const emptyCommand_t* cmds )
 	if( renderSystem->GetStereo3DMode() != STEREO3D_OFF )
 	{
 		StereoRenderExecuteBackEndCommands( cmds );
-		renderLog.EndFrame();
+		//renderLog.EndFrame();
 		return;
 	}
 
-	renderLog.StartFrame();
 	GL_StartFrame();
+	renderLog.StartFrame( commandList );
 
 	void* textureId = globalImages->hierarchicalZbufferImage->GetTextureID();
 	globalImages->LoadDeferredImages( commandList );
@@ -6030,7 +6030,7 @@ void idRenderBackend::ExecuteBackEndCommands( const emptyCommand_t* cmds )
 		pc.c_copyFrameBuffer = 0;
 	}
 
-	renderLog.EndFrame();
+	//renderLog.EndFrame();
 }
 
 

neo/renderer/RenderBackend.h

@@ -497,6 +497,7 @@ private:
 	idRenderMatrix		shadowV[6];				// shadow depth view matrix
 	idRenderMatrix		shadowP[6];				// shadow depth projection matrix
 
+	// TODO remove
 	float				hdrAverageLuminance;
 	float				hdrMaxLuminance;
 	float				hdrTime;

neo/renderer/RenderLog.cpp

@@ -106,14 +106,20 @@ PC_BeginNamedEvent
 FIXME: this is not thread safe on the PC
 ========================
 */
-void PC_BeginNamedEvent( const char* szName, const idVec4& color )
+void PC_BeginNamedEvent( const char* szName, const idVec4& color, nvrhi::ICommandList* commandList )
 {
 	if( r_logLevel.GetInteger() <= 0 )
 	{
 		return;
 	}
 
-#if defined( USE_VULKAN )
+#if defined( USE_NVRHI )
+	if( commandList )
+	{
+		commandList->beginMarker( szName );
+	}
+
+#elif defined( USE_VULKAN )
 
 	// start an annotated group of calls under the this name
 	// SRS - Prefer VK_EXT_debug_utils over VK_EXT_debug_marker/VK_EXT_debug_report (deprecated by VK_EXT_debug_utils)
@@ -141,8 +147,6 @@ void PC_BeginNamedEvent( const char* szName, const idVec4& color )
 
 		qvkCmdDebugMarkerBeginEXT( vkcontext.commandBuffer[ vkcontext.frameParity ], &label );
 	}
-#elif defined(USE_NVRHI)
-	// SP: TODO
 #else
 	// RB: colors are not supported in OpenGL
 
@@ -196,14 +200,20 @@ void PC_BeginNamedEvent( const char* szName, const idVec4& color )
 PC_EndNamedEvent
 ========================
 */
-void PC_EndNamedEvent()
+void PC_EndNamedEvent( nvrhi::ICommandList* commandList )
 {
 	if( r_logLevel.GetInteger() <= 0 )
 	{
 		return;
 	}
 
-#if defined( USE_VULKAN )
+#if defined( USE_NVRHI )
+	if( commandList )
+	{
+		commandList->endMarker();
+	}
+
+#elif defined( USE_VULKAN )
 	// SRS - Prefer VK_EXT_debug_utils over VK_EXT_debug_marker/VK_EXT_debug_report (deprecated by VK_EXT_debug_utils)
 	if( vkcontext.debugUtilsSupportAvailable )
 	{
@@ -300,8 +310,6 @@ idRenderLog
 idRenderLog	renderLog;
 
 
-
-// RB begin
 /*
 ========================
 idRenderLog::idRenderLog
@@ -309,9 +317,22 @@ idRenderLog::idRenderLog
 */
 idRenderLog::idRenderLog()
 {
+	frameCounter = 0;
+	frameParity = 0;
 }
 
-#if 1
+void idRenderLog::StartFrame( nvrhi::ICommandList* _commandList )
+{
+	commandList = _commandList;
+}
+
+void idRenderLog::EndFrame()
+{
+	frameCounter++;
+	frameParity = frameCounter % NUM_FRAME_DATA;
+}
+
+
 
 /*
 ========================
@@ -329,7 +350,11 @@ void idRenderLog::OpenMainBlock( renderLogMainBlock_t block, nvrhi::ICommandList
 	{
 		mainBlock = block;
 
-#if defined( USE_VULKAN )
+#if defined( USE_NVRHI )
+
+		// SP: use nvrhi timer queries
+
+#elif defined( USE_VULKAN )
 		if( vkcontext.queryIndex[ vkcontext.frameParity ] >= ( NUM_TIMESTAMP_QUERIES - 1 ) )
 		{
 			return;
@@ -341,10 +366,6 @@ void idRenderLog::OpenMainBlock( renderLogMainBlock_t block, nvrhi::ICommandList
 		uint32 queryIndex = vkcontext.queryAssignedIndex[ vkcontext.frameParity ][ mainBlock * 2 + 0 ] = vkcontext.queryIndex[ vkcontext.frameParity ]++;
 		vkCmdWriteTimestamp( commandBuffer, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT, queryPool, queryIndex );
 
-#elif defined(USE_NVRHI)
-
-		// SP: use nvrhi timer queries
-
 #elif defined(__APPLE__)
 		// SRS - For OSX use elapsed time query for Apple OpenGL 4.1 using GL_TIME_ELAPSED vs GL_TIMESTAMP (which is not implemented on OSX)
 		// SRS - OSX AMD drivers have a rendering bug (flashing colours) with an elasped time query when Shadow Mapping is on - turn off query for that case unless r_skipAMDWorkarounds is set
@@ -413,7 +434,6 @@ void idRenderLog::CloseMainBlock()
 	}
 }
 
-#endif
 
 /*
 ========================
@@ -422,14 +442,7 @@ idRenderLog::OpenBlock
 */
 void idRenderLog::OpenBlock( const char* label, const idVec4& color )
 {
-#if defined( USE_NVRHI )
-	if( commandList && r_logLevel.GetInteger() > 0 )
-	{
-		commandList->beginMarker( label );
-	}
-#else
-	PC_BeginNamedEvent( label, color );
-#endif
+	PC_BeginNamedEvent( label, color, commandList );
 }
 
 /*
@@ -439,13 +452,6 @@ idRenderLog::CloseBlock
 */
 void idRenderLog::CloseBlock()
 {
-#if defined( USE_NVRHI )
-	if( commandList && r_logLevel.GetInteger() > 0 )
-	{
-		commandList->endMarker();
-	}
-#else
-	PC_EndNamedEvent();
-#endif
+	PC_EndNamedEvent( commandList );
 }
 // RB end

neo/renderer/RenderLog.h

@@ -89,13 +89,16 @@ private:
 
 #if defined( USE_NVRHI )
 	nvrhi::CommandListHandle		commandList;
+
+	uint64							frameCounter;
+	uint32							frameParity;
 #endif
 
 public:
 	idRenderLog();
 
-	void		StartFrame() {}
-	void		EndFrame() {}
+	void		StartFrame( nvrhi::ICommandList* _commandList );
+	void		EndFrame();
 	void		Close() {}
 	int			Active()
 	{

neo/renderer/RenderProgs_GLSL.cpp

@@ -31,13 +31,14 @@ If you have questions concerning this license or the applicable additional terms
 #include "precompiled.h"
 #pragma hdrstop
 
+#if !defined( USE_NVRHI )
+
 #include "RenderCommon.h"
 #include "RenderProgs_embedded.h"
 
+
 idCVar r_skipStripDeadCode( "r_skipStripDeadCode", "0", CVAR_BOOL, "Skip stripping dead code" );
 
-
-
 struct idCGBlock
 {
 	idStr prefix;	// tokens that comes before the name
@@ -1680,7 +1681,7 @@ const char* idRenderProgManager::GetGLSLMacroName( shaderFeature_t sf ) const
 	return GLSLMacroNames[ sf ];
 }
 
-#if !defined( USE_NVRHI )
+
 /*
 ================================================================================================
 idRenderProgManager::FindGLSLProgram

neo/renderer/RenderSystem.h

@@ -173,16 +173,11 @@ struct backEndCounters_t
 struct glconfig_t
 {
 	graphicsVendor_t	vendor;
-	graphicsDriverType_t driverType;
 
 	const char* 		renderer_string;
 	const char* 		vendor_string;
 	const char* 		version_string;
 	const char* 		extensions_string;
-	const char* 		wgl_extensions_string;
-	const char* 		shading_language_string;
-
-	float				glVersion;				// atof( version_string )
 
 	int					maxTextureSize;			// queried from GL
 	int					maxTextureCoords;
@@ -190,6 +185,36 @@ struct glconfig_t
 	int					uniformBufferOffsetAlignment;
 	float				maxTextureAnisotropy;
 
+	bool				timerQueryAvailable;
+	bool				gpuSkinningAvailable;
+
+	stereo3DMode_t		stereo3Dmode;
+	int					nativeScreenWidth; // this is the native screen width resolution of the renderer
+	int					nativeScreenHeight; // this is the native screen height resolution of the renderer
+
+	int					displayFrequency;
+
+	int					isFullscreen;					// monitor number
+	bool				isStereoPixelFormat;
+	bool				stereoPixelFormatAvailable;
+	int					multisamples;
+
+	// Screen separation for stereoscopic rendering is set based on this.
+	// PC vid code sets this, converting from diagonals / inches / whatever as needed.
+	// If the value can't be determined, set something reasonable, like 50cm.
+	float				physicalScreenWidthInCentimeters;
+
+	float				pixelAspect;
+
+#if !defined(USE_NVRHI)
+
+	graphicsDriverType_t driverType;
+
+	const char* 		wgl_extensions_string;
+	const char* 		shading_language_string;
+
+	float				glVersion;				// atof( version_string )
+
 	int					colorBits;
 	int					depthBits;
 	int					stencilBits;
@@ -210,12 +235,11 @@ struct glconfig_t
 	bool				twoSidedStencilAvailable;
 	bool				depthBoundsTestAvailable;
 	bool				syncAvailable;
-	bool				timerQueryAvailable;
+
 	bool				occlusionQueryAvailable;
 	bool				debugOutputAvailable;
 	bool				swapControlTearAvailable;
 
-	// RB begin
 	bool				gremedyStringMarkerAvailable;
 	bool				khronosDebugAvailable;
 	bool				vertexHalfFloatAvailable;
@@ -226,31 +250,11 @@ struct glconfig_t
 //	bool				framebufferPackedDepthStencilAvailable;
 	bool				framebufferBlitAvailable;
 
-	// only true with uniform buffer support and an OpenGL driver that supports GLSL >= 1.50
-	bool				gpuSkinningAvailable;
-	// RB end
-
-	stereo3DMode_t		stereo3Dmode;
-	int					nativeScreenWidth; // this is the native screen width resolution of the renderer
-	int					nativeScreenHeight; // this is the native screen height resolution of the renderer
-
-	int					displayFrequency;
-
-	int					isFullscreen;					// monitor number
-	bool				isStereoPixelFormat;
-	bool				stereoPixelFormatAvailable;
-	int					multisamples;
-
-	// Screen separation for stereoscopic rendering is set based on this.
-	// PC vid code sets this, converting from diagonals / inches / whatever as needed.
-	// If the value can't be determined, set something reasonable, like 50cm.
-	float				physicalScreenWidthInCentimeters;
-
-	float				pixelAspect;
-
-#if !defined(USE_NVRHI) && !defined(USE_VULKAN)
+#if !defined(USE_VULKAN)
 	GLuint				global_vao;
 #endif
+
+#endif
 };
 
 

neo/renderer/RenderSystem_init.cpp

@@ -1510,6 +1510,7 @@ void GfxInfo_f( const idCmdArgs& args )
 	common->Printf( "GL_RENDERER: %s\n", glConfig.renderer_string );
 	common->Printf( "GL_VERSION: %s\n", glConfig.version_string );
 	common->Printf( "GL_EXTENSIONS: %s\n", glConfig.extensions_string );
+#if !defined( USE_NVRHI )
 	if( glConfig.wgl_extensions_string )
 	{
 		common->Printf( "WGL_EXTENSIONS: %s\n", glConfig.wgl_extensions_string );
@@ -1517,12 +1518,13 @@ void GfxInfo_f( const idCmdArgs& args )
 	common->Printf( "GL_MAX_TEXTURE_SIZE: %d\n", glConfig.maxTextureSize );
 	common->Printf( "GL_MAX_TEXTURE_COORDS_ARB: %d\n", glConfig.maxTextureCoords );
 	common->Printf( "GL_MAX_TEXTURE_IMAGE_UNITS_ARB: %d\n", glConfig.maxTextureImageUnits );
+#endif
 
 	// print all the display adapters, monitors, and video modes
 	//void DumpAllDisplayDevices();
 	//DumpAllDisplayDevices();
 
-	common->Printf( "\nPIXELFORMAT: color(%d-bits) Z(%d-bit) stencil(%d-bits)\n", glConfig.colorBits, glConfig.depthBits, glConfig.stencilBits );
+	//common->Printf( "\nPIXELFORMAT: color(%d-bits) Z(%d-bit) stencil(%d-bits)\n", glConfig.colorBits, glConfig.depthBits, glConfig.stencilBits );
 	common->Printf( "MODE: %d, %d x %d %s hz:", r_vidMode.GetInteger(), renderSystem->GetWidth(), renderSystem->GetHeight(), fsstrings[r_fullscreen.GetBool()] );
 	if( glConfig.displayFrequency )
 	{

neo/sys/win32/DeviceManager_DX12.cpp

@@ -600,7 +600,17 @@ void DeviceManager_DX12::Present()
 	auto bufferIndex = m_SwapChain->GetCurrentBackBufferIndex();
 
 	UINT presentFlags = 0;
-	if( !deviceParms.vsyncEnabled && !glConfig.isFullscreen && glConfig.swapControlTearAvailable )
+
+	if( r_swapInterval.GetInteger() == 1 )
+	{
+		SetVsyncEnabled( false );
+	}
+	else if( r_swapInterval.GetInteger() == 2 )
+	{
+		SetVsyncEnabled( true );
+	}
+
+	if( !deviceParms.vsyncEnabled && !glConfig.isFullscreen && m_TearingSupported && r_swapInterval.GetInteger() == 0 )
 	{
 		presentFlags |= DXGI_PRESENT_ALLOW_TEARING;
 	}