VK: Add a timestamp query to track frame time and report via com_showFPS.

Author: DustinHLand

neo/renderer/RenderBackend.h

@@ -56,7 +56,8 @@ struct backEndCounters_t {
 		totalMicroSec( 0 ),
 		shadowMicroSec( 0 ),
 		interactionMicroSec( 0 ),
-		shaderPassMicroSec( 0 ) {
+		shaderPassMicroSec( 0 ),
+		gpuMicroSec( 0 ) {
 	}
 
 	int		c_surfaces;
@@ -77,6 +78,7 @@ struct backEndCounters_t {
 	uint64	depthMicroSec;
 	uint64	interactionMicroSec;
 	uint64	shaderPassMicroSec;
+	uint64	gpuMicroSec;
 };
 
 struct gfxImpParms_t {
@@ -266,6 +268,8 @@ class idRenderBackend {
 
 	void				CreateSemaphores();
 
+	void				CreateQueryPool();
+
 	void				CreateSurface();
 
 	void				CreateSwapChain();
@@ -369,6 +373,10 @@ class idRenderBackend {
 	idArray< VkFramebuffer, NUM_FRAME_DATA >	m_frameBuffers;
 	idArray< VkSemaphore, NUM_FRAME_DATA >		m_acquireSemaphores;
 	idArray< VkSemaphore, NUM_FRAME_DATA >		m_renderCompleteSemaphores;
+
+	idArray< uint32, NUM_FRAME_DATA >			m_queryIndex;
+	idArray< idArray< uint64, NUM_TIMESTAMP_QUERIES >, NUM_FRAME_DATA >	m_queryResults;
+	idArray< VkQueryPool, NUM_FRAME_DATA >		m_queryPools;
 };
 
 #endif

neo/renderer/RenderCommon.h

@@ -78,6 +78,7 @@ static const int MAX_DESC_SET_WRITES		= 32;
 static const int MAX_DESC_SET_UNIFORMS		= 48;
 static const int MAX_IMAGE_PARMS			= 16;
 static const int MAX_UBO_PARMS				= 2;
+static const int NUM_TIMESTAMP_QUERIES		= 16;
 #endif
 
 // vertCacheHandle_t packs size, offset, and frame number into 64 bits

neo/renderer/RenderSystem.cpp

@@ -1338,10 +1338,6 @@ idRenderSystemLocal::SwapCommandBuffers_FinishRendering
 void idRenderSystemLocal::SwapCommandBuffers_FinishRendering( frameTiming_t * frameTiming )  {
 	SCOPED_PROFILE_EVENT( "SwapCommandBuffers_FinishRendering" );
 
-	if ( frameTiming != NULL ) {
-		frameTiming->gpuTime = 0;		// until shown otherwise
-	}
-
 	if ( !m_bInitialized ) {
 		return;
 	}
@@ -1360,6 +1356,7 @@ void idRenderSystemLocal::SwapCommandBuffers_FinishRendering( frameTiming_t * fr
 		frameTiming->depthTime = m_backend.m_pc.depthMicroSec;
 		frameTiming->interactionTime = m_backend.m_pc.interactionMicroSec;
 		frameTiming->shaderTime = m_backend.m_pc.shaderPassMicroSec;
+		frameTiming->gpuTime = m_backend.m_pc.gpuMicroSec;
 	}
 
 	// print any other statistics and clear all of them

neo/renderer/Vulkan/RenderBackend_VK.cpp

@@ -862,6 +862,22 @@ void idRenderBackend::CreateSemaphores() {
 	}
 }
 
+/*
+===============
+idRenderBackend::CreateQueryPool
+===============
+*/
+void idRenderBackend::CreateQueryPool() {
+	VkQueryPoolCreateInfo createInfo = {};
+	createInfo.sType = VK_STRUCTURE_TYPE_QUERY_POOL_CREATE_INFO;
+	createInfo.queryType = VK_QUERY_TYPE_TIMESTAMP;
+	createInfo.queryCount = NUM_TIMESTAMP_QUERIES;
+
+	for ( int i = 0; i < NUM_FRAME_DATA; ++i ) {
+		ID_VK_CHECK( vkCreateQueryPool( vkcontext.device, &createInfo, NULL, &m_queryPools[ i ] ) );
+	}
+}
+
 /*
 =============
 idRenderBackend::CreateRenderTargets
@@ -1165,7 +1181,7 @@ idRenderBackend::Clear
 void idRenderBackend::Clear() {
 	m_instance = VK_NULL_HANDLE;
 	m_physicalDevice = VK_NULL_HANDLE;
-	
+
 	debugReportCallback = VK_NULL_HANDLE;
 	m_instanceExtensions.Clear();
 	m_deviceExtensions.Clear();
@@ -1184,6 +1200,12 @@ void idRenderBackend::Clear() {
 	m_frameBuffers.Zero();
 	m_acquireSemaphores.Zero();
 	m_renderCompleteSemaphores.Zero();
+
+	m_queryIndex.Zero();
+	for ( int i = 0; i < NUM_FRAME_DATA; ++i ) {
+		m_queryResults[ i ].Zero();
+	}
+	m_queryPools.Zero();
 }
 
 /*
@@ -1228,6 +1250,9 @@ void idRenderBackend::Init() {
 	// Create semaphores for image acquisition and rendering completion
 	CreateSemaphores();
 
+	// Create Query Pool
+	CreateQueryPool();
+
 	// Create Command Pool
 	CreateCommandPool();
 
@@ -1317,6 +1342,11 @@ void idRenderBackend::Shutdown() {
 	// Destroy Command Pool
 	vkDestroyCommandPool( vkcontext.device, vkcontext.commandPool, NULL );
 
+	// Destroy Query Pools
+	for ( int i = 0; i < NUM_FRAME_DATA; ++i ) {
+		vkDestroyQueryPool( vkcontext.device, m_queryPools[ i ], NULL );
+	}
+
 	// Destroy Semaphores
 	for ( int i = 0; i < NUM_FRAME_DATA; ++i ) {
 		vkDestroySemaphore( vkcontext.device, m_acquireSemaphores[ i ], NULL );
@@ -1555,17 +1585,37 @@ void idRenderBackend::GL_StartFrame() {
 	stagingManager.Flush();
 	renderProgManager.StartFrame();
 
+	VkCommandBuffer cmdBuffer = vkcontext.commandBuffer[ vkcontext.currentFrameData ];
+	VkQueryPool queryPool = m_queryPools[ vkcontext.currentFrameData ];
+	idArray< uint64, NUM_TIMESTAMP_QUERIES > & results = m_queryResults[ vkcontext.currentFrameData ];
+
+	if ( m_queryIndex[ vkcontext.currentFrameData ] > 0 ) {
+		vkGetQueryPoolResults( vkcontext.device, queryPool, 0, 2, 
+			results.ByteSize(), results.Ptr(), sizeof( uint64 ), VK_QUERY_RESULT_64_BIT | VK_QUERY_RESULT_WAIT_BIT );
+
+		const uint64 gpuStart = results[ 0 ];
+		const uint64 gpuEnd = results[ 1 ];
+		const uint64 tick = ( 1000 * 1000 * 1000 ) / vkcontext.gpu->props.limits.timestampPeriod;
+		m_pc.gpuMicroSec = ( ( gpuEnd - gpuStart ) * 1000 * 1000 ) / tick;
+
+		m_queryIndex[ vkcontext.currentFrameData ] = 0;
+	}
+
 	VkCommandBufferBeginInfo commandBufferBeginInfo = {};
 	commandBufferBeginInfo.sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_BEGIN_INFO;
-	ID_VK_CHECK( vkBeginCommandBuffer( vkcontext.commandBuffer[ vkcontext.currentFrameData ], &commandBufferBeginInfo ) );
+	ID_VK_CHECK( vkBeginCommandBuffer( cmdBuffer, &commandBufferBeginInfo ) );
+
+	vkCmdResetQueryPool( cmdBuffer, queryPool, 0, NUM_TIMESTAMP_QUERIES );
 
 	VkRenderPassBeginInfo renderPassBeginInfo = {};
 	renderPassBeginInfo.sType = VK_STRUCTURE_TYPE_RENDER_PASS_BEGIN_INFO;
 	renderPassBeginInfo.renderPass = vkcontext.renderPass;
 	renderPassBeginInfo.framebuffer = m_frameBuffers[ m_currentSwapIndex ];
 	renderPassBeginInfo.renderArea.extent = m_swapchainExtent;
 
-	vkCmdBeginRenderPass( vkcontext.commandBuffer[ vkcontext.currentFrameData ], &renderPassBeginInfo, VK_SUBPASS_CONTENTS_INLINE );
+	vkCmdBeginRenderPass( cmdBuffer, &renderPassBeginInfo, VK_SUBPASS_CONTENTS_INLINE );
+
+	vkCmdWriteTimestamp( cmdBuffer, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT, queryPool, m_queryIndex[ vkcontext.currentFrameData ]++ );
 }
 
 /*
@@ -1574,7 +1624,11 @@ idRenderBackend::GL_EndFrame
 ==================
 */
 void idRenderBackend::GL_EndFrame() {
-	vkCmdEndRenderPass( vkcontext.commandBuffer[ vkcontext.currentFrameData ] );
+	VkCommandBuffer cmdBuffer = vkcontext.commandBuffer[ vkcontext.currentFrameData ];
+
+	vkCmdWriteTimestamp( cmdBuffer, VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT, m_queryPools[ vkcontext.currentFrameData ], m_queryIndex[ vkcontext.currentFrameData ]++ );
+
+	vkCmdEndRenderPass( cmdBuffer );
 
 	// Transition our swap image to present.
 	// Do this instead of having the renderpass do the transition
@@ -1601,7 +1655,7 @@ void idRenderBackend::GL_EndFrame() {
 		VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT,
 		0, 0, NULL, 0, NULL, 1, &barrier );
 
-	ID_VK_CHECK( vkEndCommandBuffer( vkcontext.commandBuffer[ vkcontext.currentFrameData ] ) )
+	ID_VK_CHECK( vkEndCommandBuffer( cmdBuffer ) )
 	vkcontext.commandBufferRecorded[ vkcontext.currentFrameData ] = true;
 
 	VkSemaphore * acquire = &m_acquireSemaphores[ vkcontext.currentFrameData ];
@@ -1612,7 +1666,7 @@ void idRenderBackend::GL_EndFrame() {
 	VkSubmitInfo submitInfo = {};
 	submitInfo.sType = VK_STRUCTURE_TYPE_SUBMIT_INFO;
 	submitInfo.commandBufferCount = 1;
-	submitInfo.pCommandBuffers = &vkcontext.commandBuffer[ vkcontext.currentFrameData ];
+	submitInfo.pCommandBuffers = &cmdBuffer;
 	submitInfo.waitSemaphoreCount = 1;
 	submitInfo.pWaitSemaphores = acquire;
 	submitInfo.signalSemaphoreCount = 1;