[vulkan] Add gpu profiling zones

This relies on my fork of tracy: https://github.com/taniwha/tracy on the wip-c-vulkan branch. Everything is still rather flaky though. This necessitated the jump to vulkan 1.2 as a requirement.
2025-04-04 16:31:30 +00:00 · 2023-12-04 21:38:43 +09:00 · 2023-12-04 21:38:43 +09:00 · fccd06c5bf
commit fccd06c5bf
parent 57282702ba
30 changed files with 195 additions and 48 deletions
--- a/Makefile.am
+++ b/Makefile.am
@ -63,8 +63,10 @@ PTHREAD_CFLAGS=@PTHREAD_CFLAGS@

 if HAVE_TRACY
 tracy_src=tracy/public/TracyClient.cpp
+tracyvk_src=tracy/public/client/TracyVkProfiler.cpp
 else
 tracy_src=
+tracyvk_src=
 endif

 lib_ldflags=-version-info $(QUAKE_LIBRARY_VERSION_INFO) \
--- a/include/QF/Vulkan/render.h
+++ b/include/QF/Vulkan/render.h
@ -419,6 +419,9 @@ typedef struct qfv_renderframe_s {
 	VkSemaphore imageAvailableSemaphore;
 	VkSemaphore renderDoneSemaphore;
 	qfv_cmdpoolmgr_t cmdpool;
+#ifdef TRACY_ENABLE
+	qftVkCtx_t *qftVkCtx;
+#endif
 } qfv_renderframe_t;

 typedef struct qfv_renderframeset_s
@ -440,6 +443,7 @@ typedef struct qfv_renderctx_s {

 typedef struct qfv_taskctx_s {
 	struct vulkan_ctx_s *ctx;
+	qfv_renderframe_t *frame;
 	qfv_pipeline_t *pipeline;
 	qfv_renderpass_t *renderpass;
 	VkCommandBuffer cmd;
--- a/include/qftracy.h
+++ b/include/qftracy.h
@ -7,6 +7,7 @@

 // enable tracy
 #define TRACY_ENABLE
+#define TRACY_VK_USE_SYMBOL_TABLE
 #include "tracy/TracyC.h"

 static inline void __qfZoneEnd (TracyCZoneCtx **ctxptr)
@ -14,6 +15,8 @@ static inline void __qfZoneEnd (TracyCZoneCtx **ctxptr)
 	TracyCZoneEnd (**ctxptr);
 }

+#define qftVkCtx_t struct ___tracy_vkctx
+
 #define qfConcatInternal(a,b) a##b
 #define qfConcat(a,b) qfConcatInternal(a, b)

@ -50,4 +53,6 @@ static inline void __qfZoneEnd (TracyCZoneCtx **ctxptr)
 #define qfZoneNamedN(varname, name, active)
 #define qfMessageL(msg)

+#define qftVkCtx_t void
+
 #endif
--- a/include/vid_vulkan.h
+++ b/include/vid_vulkan.h
@ -10,6 +10,43 @@
 #include "QF/qtypes.h"
 #include "QF/simd/types.h"

+#ifdef HAVE_TRACY
+#include "tracy/TracyCVulkan.h"
+
+#define qftCVkContextHostCalibrated(instance, physdev, device, instanceProcAddr, deviceProcAddr) TracyCVkContextHostCalibrated(instance, physdev, device, instanceProcAddr, deviceProcAddr)
+#define qftCVkContextDestroy(ctx) TracyCVkDestroy( ctx )
+#define qftCVkContextName(ctx, name, size) TracyCVkContextName(ctx, name, size)
+#define qftCVkCollect(ctx, cmdbuf) TracyCVkCollect (ctx, cmdbuf)
+
+static inline void __qftVkZoneEnd (___tracy_vkctx_scope ***zone)
+{
+	TracyCVkZoneEnd (**zone);
+}
+
+#define qftVkZone(ctx, cmdbuf, name) TracyCVkZone (ctx, cmdbuf, name)
+#define qftVkScopedZone(ctx, cmdbuf, name) \
+	qftVkZone (ctx, cmdbuf, name) \
+	__attribute__((cleanup(__qftVkZoneEnd))) \
+	___tracy_vkctx_scope **qfConcat(__qfScoped##varname, __COUNTER__) = &___tracy_gpu_zone
+#define qftVkZoneC(ctx, cmdbuf, name, color) TracyCVkZone (ctx, cmdbuf, name, color)
+#define qftVkZoneEnd(varname) TracyCVkZoneEnd (varname)
+
+#else
+
+#define qftVkContext(instance, physdev, device, queue, cmdbuf, instanceProcAddr, deviceProcAddr)
+#define qftVkContextCalibrated(instance, physdev, device, queue, cmdbuf, instanceProcAddr, deviceProcAddr)
+#define qftCVkContextHostCalibrated(instance, physdev, device, instanceProcAddr, deviceProcAddr)
+#define qftCVkContextDestroy(ctx)
+#define qftCVkContextName(ctx, name, size)
+#define qftCVkContextCollect(ctx, cmdbuf)
+
+#define qftVkZone(ctx, cmdbuf, name)
+#define qftVkScopedZone(ctx, cmdbuf, name)
+#define qftVkZoneC(ctx, cmdbuf, name, color)
+#define qftVkZoneEnd(varname)
+
+#endif
+
 #define VA_CTX_COUNT 64

 typedef struct qfv_renderpassset_s
--- a/libs/client/cl_temp_entities.c
+++ b/libs/client/cl_temp_entities.c
@ -113,6 +113,7 @@ static vec4f_t beam_rolls[360];
 void
 CL_TEnts_Precache (void)
 {
+	qfZoneScoped (true);
 	cl_sfx_wizhit = S_PrecacheSound ("wizard/hit.wav");
 	cl_sfx_knighthit = S_PrecacheSound ("hknight/hit.wav");
 	cl_sfx_tink1 = S_PrecacheSound ("weapons/tink1.wav");
@ -558,6 +559,7 @@ static const TE_Effect nqEffects[256] = {
 void
 CL_ParseTEnt_nq (qmsg_t *net_message, double time, TEntContext_t *ctx)
 {
+	qfZoneScoped (true);
 	byte        type = MSG_ReadByte (net_message);
 	parse_tent (net_message, time, ctx, nqEffects[type]);
 }
@ -680,6 +682,7 @@ CL_UpdateTEnts (double time, TEntContext_t *ctx)
 void
 CL_ParseParticleEffect (qmsg_t *net_message)
 {
+	qfZoneScoped (true);
 	int         i, count, color;
 	vec4f_t     org = {0, 0, 0, 1}, dir = {};

--- a/libs/client/cl_view.c
+++ b/libs/client/cl_view.c
@ -470,6 +470,7 @@ V_DriftPitch (viewstate_t *vs)
 void
 V_ParseDamage (qmsg_t *net_message, viewstate_t *vs)
 {
+	qfZoneScoped (true);
 	float       count, side;
 	int         armor, blood;
 	vec4f_t     origin = vs->player_origin;
--- a/libs/client/cl_world.c
+++ b/libs/client/cl_world.c
@ -73,6 +73,7 @@ CL_World_Init (void)
 void
 CL_ParseBaseline (qmsg_t *msg, entity_state_t *baseline, int version)
 {
+	qfZoneScoped (true);
 	int         bits = 0;

 	if (version == 2)
@ -107,6 +108,7 @@ CL_ParseBaseline (qmsg_t *msg, entity_state_t *baseline, int version)
 void
 CL_ParseStatic (qmsg_t *msg, int version)
 {
+	qfZoneScoped (true);
 	entity_t	    ent;
 	entity_state_t	es;

@ -240,6 +242,7 @@ CL_World_NewMap (const char *mapname, const char *skyname)
 void
 CL_World_Clear (void)
 {
+	qfZoneScoped (true);
 	Scene_FreeAllEntities (cl_world.scene);
 	CL_ClearTEnts ();
 }
--- a/libs/models/model.c
+++ b/libs/models/model.c
@ -174,6 +174,7 @@ mod_unload_model (size_t ind)
 VISIBLE void
 Mod_ClearAll (void)
 {
+	qfZoneScoped (true);
 	size_t      i;

 	for (i = 0; i < mod_numknown; i++) {
@ -338,6 +339,7 @@ Mod_ForName (const char *name, bool crash)
 VISIBLE void
 Mod_TouchModel (const char *name)
 {
+	qfZoneScoped (true);
 	model_t    *mod;

 	mod = Mod_FindName (name);
--- a/libs/scene/scene.c
+++ b/libs/scene/scene.c
@ -305,6 +305,7 @@ Scene_CreateEntity (scene_t *scene)
 void
 Scene_DestroyEntity (scene_t *scene, entity_t ent)
 {
+	qfZoneScoped (true);
 	ECS_DelEntity (scene->reg, ent.id);
 }

--- a/libs/video/renderer/r_screen.c
+++ b/libs/video/renderer/r_screen.c
@ -497,6 +497,7 @@ SCR_Shutdown (void)
 void
 SCR_NewScene (scene_t *scene)
 {
+	qfZoneScoped (true);
 	scr_scene = scene;
 	if (scene) {
 		mod_brush_t *brush = &scr_scene->worldmodel->brush;
--- a/libs/video/renderer/vid_render_vulkan.c
+++ b/libs/video/renderer/vid_render_vulkan.c
@ -143,6 +143,7 @@ vulkan_R_Init (void)
 static void
 vulkan_R_ClearState (void)
 {
+	qfZoneScoped (true);
 	QFV_DeviceWaitIdle (vulkan_ctx->device);
 	//FIXME clear scene correctly
 	r_refdef.worldmodel = 0;
@ -160,6 +161,7 @@ vulkan_R_LoadSkys (const char *skyname)
 static void
 vulkan_R_NewScene (scene_t *scene)
 {
+	qfZoneScoped (true);
 	Vulkan_NewScene (scene, vulkan_ctx);
 }

--- a/libs/video/renderer/vulkan/device.c
+++ b/libs/video/renderer/vulkan/device.c
@ -154,20 +154,29 @@ QFV_CreateDevice (vulkan_ctx_t *ctx, const char **extensions)
 			.multiview = 1,
 			.multiviewGeometryShader = 1,
 		};
-		VkPhysicalDeviceFeatures features = {
-			.imageCubeArray = 1,
-			.independentBlend = 1,
-			.geometryShader = 1,
-			.multiViewport = 1,
-			.fragmentStoresAndAtomics = 1,
-			.fillModeNonSolid = 1,
+		VkPhysicalDeviceVulkan12Features features12 = {
+			.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_FEATURES,
+			.pNext = &multiview_features,
+			.hostQueryReset = 1,
+		};
+		VkPhysicalDeviceFeatures2 features = {
+			.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FEATURES_2,
+			.pNext = &features12,
+			.features = {
+				.imageCubeArray = 1,
+				.independentBlend = 1,
+				.geometryShader = 1,
+				.multiViewport = 1,
+				.fragmentStoresAndAtomics = 1,
+				.fillModeNonSolid = 1,
+			},
 		};
 		VkDeviceCreateInfo dCreateInfo = {
-			VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO, &multiview_features, 0,
+			VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO, &features, 0,
 			1, &qCreateInfo,
 			nlay, lay,
 			next, ext,
-			&features
+			0
 		};
 		qfv_device_t *device = calloc (1, sizeof (qfv_device_t)
 										  + sizeof (qfv_devfuncs_t));
@ -203,6 +212,7 @@ QFV_DestroyDevice (qfv_device_t *device)
 int
 QFV_DeviceWaitIdle (qfv_device_t *device)
 {
+	qfZoneScoped (true);
 	qfv_devfuncs_t *dfunc = device->funcs;
 	return dfunc->vkDeviceWaitIdle (device->dev) == VK_SUCCESS;
 }
--- a/libs/video/renderer/vulkan/instance.c
+++ b/libs/video/renderer/vulkan/instance.c
@ -208,7 +208,7 @@ QFV_CreateInstance (vulkan_ctx_t *ctx,
 		VK_STRUCTURE_TYPE_APPLICATION_INFO, 0,
 		appName, appVersion,
 		PACKAGE_STRING, 0x000702ff, //FIXME version
-		VK_API_VERSION_1_1,
+		VK_API_VERSION_1_2,
 	};
 	VkInstanceCreateInfo createInfo = {
 		VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO, 0, 0,
--- a/libs/video/renderer/vulkan/render.c
+++ b/libs/video/renderer/vulkan/render.c
@ -47,6 +47,7 @@
 #include "QF/Vulkan/device.h"
 #include "QF/Vulkan/dsmanager.h"
 #include "QF/Vulkan/image.h"
+#include "QF/Vulkan/instance.h"
 #include "QF/Vulkan/pipeline.h"
 #include "QF/Vulkan/render.h"
 #include "QF/Vulkan/resource.h"
@ -141,9 +142,13 @@ run_subpass (qfv_subpass_t *sp, qfv_taskctx_t *taskctx)
 	qfv_devfuncs_t *dfunc = device->funcs;
 	dfunc->vkBeginCommandBuffer (taskctx->cmd, &sp->beginInfo);

-	for (uint32_t i = 0; i < sp->pipeline_count; i++) {
-		__auto_type pipeline = &sp->pipelines[i];
-		run_pipeline (pipeline, taskctx);
+	{
+		qftVkScopedZone (taskctx->frame->qftVkCtx, taskctx->cmd, "subpass");
+
+		for (uint32_t i = 0; i < sp->pipeline_count; i++) {
+			__auto_type pipeline = &sp->pipelines[i];
+			run_pipeline (pipeline, taskctx);
+		}
 	}

 	dfunc->vkEndCommandBuffer (taskctx->cmd);
@ -158,6 +163,7 @@ run_renderpass (qfv_renderpass_t *rp, vulkan_ctx_t *ctx, void *data)
 	qfv_device_t *device = ctx->device;
 	qfv_devfuncs_t *dfunc = device->funcs;
 	__auto_type rctx = ctx->render_context;
+	auto frame = &rctx->frames.a[ctx->curFrame];
 	__auto_type job = rctx->job;

 	VkCommandBuffer cmd = QFV_GetCmdBuffer (ctx, false);
@ -167,30 +173,34 @@ run_renderpass (qfv_renderpass_t *rp, vulkan_ctx_t *ctx, void *data)
 		VK_STRUCTURE_TYPE_COMMAND_BUFFER_BEGIN_INFO,
 	};
 	dfunc->vkBeginCommandBuffer (cmd, &beginInfo);
-	QFV_duCmdBeginLabel (device, cmd, rp->label.name,
-						 {VEC4_EXP (rp->label.color)});
-	dfunc->vkCmdBeginRenderPass (cmd, &rp->beginInfo, rp->subpassContents);
-	for (uint32_t i = 0; i < rp->subpass_count; i++) {
-		__auto_type sp = &rp->subpasses[i];
-		QFV_duCmdBeginLabel (device, cmd, sp->label.name,
-							 {VEC4_EXP (sp->label.color)});
-		qfv_taskctx_t taskctx = {
-			.ctx = ctx,
-			.renderpass = rp,
-			.cmd = QFV_GetCmdBuffer (ctx, true),
-			.data = data,
-		};
-		run_subpass (sp, &taskctx);
-		dfunc->vkCmdExecuteCommands (cmd, 1, &taskctx.cmd);
-		QFV_duCmdEndLabel (device, cmd);
-		//FIXME comment is a bit off as exactly one buffer is always submitted
-		//
-		//Regardless of whether any commands were submitted for this
-		//subpass, must step through each and every subpass, otherwise
-		//the attachments won't be transitioned correctly.
-		//However, only if not the last (or only) subpass.
-		if (i < rp->subpass_count - 1) {
-			dfunc->vkCmdNextSubpass (cmd, rp->subpassContents);
+	{
+		qftVkScopedZone (frame->qftVkCtx, cmd, "renderpass");
+		QFV_duCmdBeginLabel (device, cmd, rp->label.name,
+							 {VEC4_EXP (rp->label.color)});
+		dfunc->vkCmdBeginRenderPass (cmd, &rp->beginInfo, rp->subpassContents);
+		for (uint32_t i = 0; i < rp->subpass_count; i++) {
+			__auto_type sp = &rp->subpasses[i];
+			QFV_duCmdBeginLabel (device, cmd, sp->label.name,
+								 {VEC4_EXP (sp->label.color)});
+			qfv_taskctx_t taskctx = {
+				.ctx = ctx,
+				.frame = frame,
+				.renderpass = rp,
+				.cmd = QFV_GetCmdBuffer (ctx, true),
+				.data = data,
+			};
+			run_subpass (sp, &taskctx);
+			dfunc->vkCmdExecuteCommands (cmd, 1, &taskctx.cmd);
+			QFV_duCmdEndLabel (device, cmd);
+			//FIXME comment is a bit off as exactly one buffer is always submitted
+			//
+			//Regardless of whether any commands were submitted for this
+			//subpass, must step through each and every subpass, otherwise
+			//the attachments won't be transitioned correctly.
+			//However, only if not the last (or only) subpass.
+			if (i < rp->subpass_count - 1) {
+				dfunc->vkCmdNextSubpass (cmd, rp->subpassContents);
+			}
 		}
 	}
 	dfunc->vkCmdEndRenderPass (cmd);
@ -208,10 +218,14 @@ run_compute_pipeline (qfv_pipeline_t *pipeline, VkCommandBuffer cmd,
 	}
 	qfv_device_t *device = ctx->device;
 	qfv_devfuncs_t *dfunc = device->funcs;
+	auto rctx = ctx->render_context;
+	auto frame = &rctx->frames.a[ctx->curFrame];
+	qftVkScopedZone (frame->qftVkCtx, cmd, "compute");
 	dfunc->vkCmdBindPipeline (cmd, pipeline->bindPoint, pipeline->pipeline);

 	qfv_taskctx_t taskctx = {
 		.ctx = ctx,
+		.frame = frame,
 		.pipeline = pipeline,
 		.cmd = cmd,
 	};
@ -257,15 +271,40 @@ run_compute (qfv_compute_t *comp, vulkan_ctx_t *ctx, qfv_step_t *step)
 static void
 run_process (qfv_process_t *proc, vulkan_ctx_t *ctx)
 {
+	auto rctx = ctx->render_context;
+	auto frame = &rctx->frames.a[ctx->curFrame];
 	qfZoneNamed (zone, true);
 	qfZoneName (zone, proc->label.name, proc->label.name_len);
 	qfZoneColor (zone, proc->label.color32);
 	qfv_taskctx_t taskctx = {
 		.ctx = ctx,
+		.frame = frame,
 	};
 	run_tasks (proc->task_count, proc->tasks, &taskctx);
 }

+static void
+run_collect (vulkan_ctx_t *ctx)
+{
+#ifdef TRACY_ENABLE
+	auto device = ctx->device;
+	auto dfunc = device->funcs;
+	auto rctx = ctx->render_context;
+	auto frame = &rctx->frames.a[ctx->curFrame];
+
+	VkCommandBuffer cmd = QFV_GetCmdBuffer (ctx, false);
+	QFV_duSetObjectName (device, VK_OBJECT_TYPE_COMMAND_BUFFER, cmd,
+						 va (ctx->va_ctx, "cmd:render:%s", "tracy"));
+	VkCommandBufferBeginInfo beginInfo = {
+		VK_STRUCTURE_TYPE_COMMAND_BUFFER_BEGIN_INFO,
+	};
+	dfunc->vkBeginCommandBuffer (cmd, &beginInfo);
+	qftCVkCollect (frame->qftVkCtx, cmd);
+	dfunc->vkEndCommandBuffer (cmd);
+	QFV_AppendCmdBuffer (ctx, cmd);
+#endif
+}
+
 void
 QFV_RunRenderPass (vulkan_ctx_t *ctx, qfv_renderpass_t *renderpass,
 				   uint32_t width, uint32_t height, void *data)
@ -309,6 +348,7 @@ QFV_RunRenderJob (vulkan_ctx_t *ctx)
 		update_time (&step->time, step_start, Sys_LongTime ());
 	}

+	qfMessageL ("submit");
 	auto device = ctx->device;
 	auto dfunc = device->funcs;
 	auto queue = &device->queue;
@ -324,6 +364,7 @@ QFV_RunRenderJob (vulkan_ctx_t *ctx)
 	dfunc->vkResetFences (device->dev, 1, &frame->fence);
 	dfunc->vkQueueSubmit (queue->queue, 1, &submitInfo, frame->fence);

+	qfMessageL ("present");
 	VkPresentInfoKHR presentInfo = {
 		VK_STRUCTURE_TYPE_PRESENT_INFO_KHR, 0,
 		1, &frame->renderDoneSemaphore,
@ -332,6 +373,7 @@ QFV_RunRenderJob (vulkan_ctx_t *ctx)
 	};
 	dfunc->vkQueuePresentKHR (queue->queue, &presentInfo);

+	qfMessageL ("update_time");
 	if (++ctx->curFrame >= rctx->frames.size) {
 		ctx->curFrame = 0;
 	}
@ -451,6 +493,7 @@ wait_on_fence (const exprval_t **params, exprval_t *result, exprctx_t *ectx)
 	QFV_CmdPoolManager_Reset (&frame->cmdpool);
 	auto job = ctx->render_context->job;
 	DARRAY_RESIZE (&job->commands, 0);
+	run_collect (ctx);
 }

 static void
@ -530,6 +573,14 @@ QFV_Render_Init (vulkan_ctx_t *ctx)
 		frame->renderDoneSemaphore = QFV_CreateSemaphore (device);
 		QFV_CmdPoolManager_Init (&frame->cmdpool, device,
 								 va (ctx->va_ctx, "render pool:%zd", i));
+#ifdef TRACY_ENABLE
+		auto instance = ctx->instance->instance;
+		auto physdev = ctx->device->physDev->dev;
+		auto gipa = ctx->vkGetInstanceProcAddr;
+		auto gdpa = ctx->instance->funcs->vkGetDeviceProcAddr;
+		frame->qftVkCtx = qftCVkContextHostCalibrated (instance, physdev,
+													   device->dev, gipa, gdpa);
+#endif
 	}
 }

--- a/libs/video/renderer/vulkan/resource.c
+++ b/libs/video/renderer/vulkan/resource.c
@ -281,6 +281,7 @@ QFV_CreateResource (qfv_device_t *device, qfv_resource_t *resource)
 void
 QFV_DestroyResource (qfv_device_t *device, qfv_resource_t *resource)
 {
+	qfZoneScoped (true);
 	qfv_devfuncs_t *dfunc = device->funcs;

 	for (unsigned i = 0; i < resource->num_objects; i++) {
@ -291,6 +292,7 @@ QFV_DestroyResource (qfv_device_t *device, qfv_resource_t *resource)
 				break;
 			case qfv_res_buffer_view:
 				{
+					qfZoneScoped (true);
 					__auto_type buffview = &obj->buffer_view;
 					dfunc->vkDestroyBufferView (device->dev, buffview->view, 0);
 					buffview->view = 0;
@ -298,6 +300,7 @@ QFV_DestroyResource (qfv_device_t *device, qfv_resource_t *resource)
 				break;
 			case qfv_res_image_view:
 				{
+					qfZoneScoped (true);
 					__auto_type imgview = &obj->image_view;
 					dfunc->vkDestroyImageView (device->dev, imgview->view, 0);
 					imgview->view = 0;
@ -310,6 +313,7 @@ QFV_DestroyResource (qfv_device_t *device, qfv_resource_t *resource)
 		switch (obj->type) {
 			case qfv_res_buffer:
 				{
+					qfZoneScoped (true);
 					__auto_type buffer = &obj->buffer;
 					dfunc->vkDestroyBuffer (device->dev, buffer->buffer, 0);
 					buffer->buffer = 0;
@ -317,6 +321,7 @@ QFV_DestroyResource (qfv_device_t *device, qfv_resource_t *resource)
 				break;
 			case qfv_res_image:
 				{
+					qfZoneScoped (true);
 					__auto_type image = &obj->image;
 					dfunc->vkDestroyImage (device->dev, image->image, 0);
 					image->image = 0;
@ -327,7 +332,9 @@ QFV_DestroyResource (qfv_device_t *device, qfv_resource_t *resource)
 				break;
 		}
 	}
+	qfMessageL("free memory");
 	dfunc->vkFreeMemory (device->dev, resource->memory, 0);
+	qfMessageL("memory freed");
 	resource->memory = 0;
 }

--- a/libs/video/renderer/vulkan/vulkan_draw.c
+++ b/libs/video/renderer/vulkan/vulkan_draw.c
@ -939,6 +939,7 @@ slice_draw (const exprval_t **params, exprval_t *result, exprctx_t *ectx)
 		return;
 	}

+	qftVkScopedZone (taskctx->frame->qftVkCtx, taskctx->cmd, "slice_draw");
 	VkDeviceMemory memory = dctx->draw_resource[1].memory;
 	size_t      atom = device->physDev->properties->limits.nonCoherentAtomSize;
 	size_t      atom_mask = atom - 1;
--- a/libs/video/renderer/vulkan/vulkan_lighting.c
+++ b/libs/video/renderer/vulkan/vulkan_lighting.c
@ -1313,6 +1313,7 @@ Vulkan_Lighting_Setup (vulkan_ctx_t *ctx)
 static void
 clear_shadows (vulkan_ctx_t *ctx)
 {
+	qfZoneScoped (true);
 	qfv_device_t *device = ctx->device;
 	lightingctx_t *lctx = ctx->lighting_context;

@ -2040,6 +2041,7 @@ scene_lightleaf_ui (void *comp, imui_ctx_t *imui_ctx,
 void
 Vulkan_LoadLights (scene_t *scene, vulkan_ctx_t *ctx)
 {
+	qfZoneScoped (true);
 	lightingctx_t *lctx = ctx->lighting_context;

 	lctx->scene = scene;
--- a/libs/video/renderer/vulkan/vulkan_output.c
+++ b/libs/video/renderer/vulkan/vulkan_output.c
@ -221,6 +221,7 @@ output_draw (qfv_taskctx_t *taskctx,
 	auto layout = taskctx->pipeline->layout;
 	auto cmd = taskctx->cmd;

+	qftVkScopedZone (taskctx->frame->qftVkCtx, taskctx->cmd, "flat");
 	VkDescriptorSet set[] = {
 		Vulkan_Matrix_Descriptors (ctx, ctx->curFrame),
 		oframe->set,
--- a/libs/video/renderer/vulkan/vulkan_vid_common.c
+++ b/libs/video/renderer/vulkan/vulkan_vid_common.c
@ -98,6 +98,9 @@ static const char *instance_extensions[] = {

 static const char *device_extensions[] = {
 	VK_KHR_SWAPCHAIN_EXTENSION_NAME,
+#ifdef TRACY_ENABLE
+	VK_EXT_CALIBRATED_TIMESTAMPS_EXTENSION_NAME,
+#endif
 	0,
 };

--- a/libs/video/targets/Makemodule.am
+++ b/libs/video/targets/Makemodule.am
@ -31,7 +31,7 @@ js_libs = libs/util/libQFutil.la

 libs_video_targets_libvid_common_la_SOURCES = \
 	libs/video/targets/vid.c
-libs_video_targets_libvid_common_la_CFLAGS=	@PREFER_NON_PIC@
+libs_video_targets_libvid_common_la_CFLAGS=	$(AM_CFLAGS) @PREFER_NON_PIC@
 libs_video_targets_libvid_common_la_LDFLAGS=	@STATIC@

 libs_video_targets_libvid_x11_la_SOURCES = \
@ -39,7 +39,7 @@ libs_video_targets_libvid_x11_la_SOURCES = \
 	libs/video/targets/context_x11.c \
 	libs/video/targets/dga_check.c \
 	$(x11_src)
-libs_video_targets_libvid_x11_la_CFLAGS=		@PREFER_NON_PIC@ $(X_CFLAGS)
+libs_video_targets_libvid_x11_la_CFLAGS=		$(AM_CFLAGS) @PREFER_NON_PIC@ $(X_CFLAGS)
 libs_video_targets_libvid_x11_la_LDFLAGS=		@STATIC@
 EXTRA_libs_video_targets_libvid_x11_la_SOURCES= $(x11_vulkan_src)

@ -47,7 +47,7 @@ libs_video_targets_libvid_win_la_SOURCES = \
 	libs/video/targets/in_win.c \
 	libs/video/targets/context_win.c \
 	$(win_src)
-libs_video_targets_libvid_win_la_CFLAGS=		@PREFER_NON_PIC@ $(X_CFLAGS)
+libs_video_targets_libvid_win_la_CFLAGS=		$(AM_CFLAGS) @PREFER_NON_PIC@ $(X_CFLAGS)
 libs_video_targets_libvid_win_la_LDFLAGS=		@STATIC@
 EXTRA_libs_video_targets_libvid_win_la_SOURCES= $(win_vulkan_src)

--- a/libs/video/targets/vid.c
+++ b/libs/video/targets/vid.c
@ -296,6 +296,7 @@ VID_InitGamma (const byte *pal)
 void
 VID_ClearMemory (void)
 {
+	qfZoneScoped (true);
 	if (vi->flush_caches) {
 		vi->flush_caches (vi->ctx);
 	}
--- a/nq/source/Makemodule.am
+++ b/nq/source/Makemodule.am
@ -91,7 +91,7 @@ nq_x11_libs=  \
 			libs/models/libQFmodels.la \
 			libs/video/targets/libQFx11.la \
 			$(nq_client_LIBS)
-nq_x11_SOURCES= 	nq/source/sys_unix.c $(tracy_src)
+nq_x11_SOURCES= 	nq/source/sys_unix.c $(tracy_src) $(tracyvk_src)
 nq_x11_LDADD=		$(nq_x11_libs) \
 			$(VIDMODE_LIBS) $(DGA_LIBS) ${XFIXES_LIBS} $(XI2_LIBS) $(X_LIBS) \
 			-lX11 $(X_EXTRA_LIBS) $(X_SHM_LIB) $(NET_LIBS) $(DL_LIBS)
@ -106,7 +106,7 @@ nq_win_libs=  \
 			libs/models/libQFmodels.la \
 			libs/video/targets/libQFwin.la \
 			$(nq_client_LIBS)
-nq_win_SOURCES= 	nq/source/sys_win.c $(tracy_src)
+nq_win_SOURCES= 	nq/source/sys_win.c $(tracy_src) $(tracyvk_src)
 nq_win_LDADD=		$(nq_win_libs) -lgdi32 -lcomctl32 -lwinmm $(NET_LIBS)
 nq_win_LDFLAGS= 	$(common_ldflags)
 nq_win_DEPENDENCIES=	$(nq_win_libs)
--- a/nq/source/cl_ents.c
+++ b/nq/source/cl_ents.c
@ -69,6 +69,7 @@ set_t       cl_forcelink = SET_STATIC_INIT (MAX_EDICTS, alloc_forcelink);
 void
 CL_ClearEnts (void)
 {
+	qfZoneScoped (true);
 	size_t      i;

 	for (i = 0; i < MAX_EDICTS; i++) {
--- a/nq/source/cl_main.c
+++ b/nq/source/cl_main.c
@ -224,6 +224,7 @@ CL_Shutdown (void *data)
 void
 CL_ClearMemory (void)
 {
+	qfZoneScoped (true);
 	VID_ClearMemory ();
 	SCR_SetFullscreen (0);

@ -280,6 +281,7 @@ CL_InitCvars (void)
 void
 CL_ClearState (void)
 {
+	qfZoneScoped (true);
 	CL_ClearMemory ();
 	if (!sv.active)
 		Host_ClearMemory ();
--- a/nq/source/cl_parse.c
+++ b/nq/source/cl_parse.c
@ -146,6 +146,7 @@ CL_EntityNum (int num)
 static void
 CL_ParseStartSoundPacket (void)
 {
+	qfZoneScoped (true);
 	float       attenuation;
 	int         channel, ent, field_mask, sound_num, volume;

@ -250,6 +251,7 @@ CL_KeepaliveMessage (void)
 static void
 CL_NewMap (const char *mapname)
 {
+	qfZoneScoped (true);
 	CL_World_NewMap (mapname, 0);
 	V_NewScene (&cl.viewstate, cl_world.scene);

@ -264,6 +266,7 @@ CL_NewMap (const char *mapname)
 static void
 CL_ParseServerInfo (void)
 {
+	qfZoneScoped (true);
 	char        model_precache[MAX_MODELS][MAX_QPATH];
 	char        sound_precache[MAX_SOUNDS][MAX_QPATH];
 	const char *str;
@ -540,6 +543,7 @@ CL_ParseUpdate (int bits)
 static void
 CL_ParseClientdata (void)
 {
+	qfZoneScoped (true);
 	int         i, j;
 	int         bits;

@ -687,6 +691,7 @@ CL_ParseClientdata (void)
 static void
 CL_ParseStaticSound (int version)
 {
+	qfZoneScoped (true);
 	int         sound_num;
 	float       vol, atten;
 	vec4f_t     org = { 0, 0, 0, 1 };
@ -718,7 +723,7 @@ CL_SetStat (int stat, int value)
 void
 CL_ParseServerMessage (void)
 {
-	qfZoneNamedN (psm_zone, "CL_ParseServerMessage", true);
+	qfZoneScoped (true);
 	int         cmd = 0, i, j;
 	const char *str;
 	static dstring_t *stuffbuf;
--- a/nq/source/host.c
+++ b/nq/source/host.c
@ -609,6 +609,7 @@ Host_ShutdownServer (bool crash)
 void
 Host_ClearMemory (void)
 {
+	qfZoneScoped (true);
 	Sys_MaskPrintf (SYS_dev, "Clearing memory\n");
 	Mod_ClearAll ();
 	if (host_hunklevel)
--- a/nq/source/sys_win.c
+++ b/nq/source/sys_win.c
@ -217,6 +217,7 @@ WinMain (HINSTANCE hInstance, HINSTANCE hPrevInstance, LPSTR lpCmdLine,

 	oldtime = Sys_DoubleTime () - 0.1;
 	while (1) {							// Main message loop
+		qfFrameMark;
 		if (!isDedicated) {
 			// yield the CPU for a little while when paused, minimized, or
 			// not the focus
--- a/qtv/source/Makemodule.am
+++ b/qtv/source/Makemodule.am
@ -40,7 +40,7 @@ qtv_LIBS= \
 		libs/ui/libQFui.la \
 		libs/util/libQFutil.la

-qtv_server_SOURCES=	qtv/source/client.c qtv/source/connection.c qtv/source/qtv.c qtv/source/sbar.c qtv/source/server.c qtv/source/sv_parse.c
+qtv_server_SOURCES=	qtv/source/client.c qtv/source/connection.c qtv/source/qtv.c qtv/source/sbar.c qtv/source/server.c qtv/source/sv_parse.c $(tracy_src)
 qtv_server_LDADD=	$(qtv_LIBS) $(NET_LIBS) $(DL_LIBS) $(CURSES_LIBS)
 qtv_server_LDFLAGS=	$(common_ldflags)
 qtv_server_DEPENDENCIES=	$(qtv_LIBS)
--- a/qw/source/Makemodule.am
+++ b/qw/source/Makemodule.am
@ -71,7 +71,7 @@ qw_server_LIBS= \
 		libs/util/libQFutil.la

 qw_server_deps=qw/source/libqw_server.a qw/source/libqw_common.a $(qw_server_LIBS)
-qw_server_SOURCES=		$(syssv_SRC)
+qw_server_SOURCES=		$(syssv_SRC) $(tracy_src)
 qw_server_LDADD=		$(qw_server_deps) $(NET_LIBS) $(DL_LIBS) $(CURSES_LIBS)
 qw_server_LDFLAGS=		$(common_ldflags)
 qw_server_DEPENDENCIES=	$(qw_server_deps)
@ -117,7 +117,7 @@ qw_client_x11_libs= \
 			libs/models/libQFmodels.la \
 			libs/video/targets/libQFx11.la \
 			$(qw_client_LIBS)
-qw_client_x11_SOURCES=	qw/source/cl_sys_unix.c $(tracy_src)
+qw_client_x11_SOURCES=	qw/source/cl_sys_unix.c $(tracy_src) $(tracyvk_src)
 qw_client_x11_LDADD= $(qw_client_x11_libs) \
 			$(VIDMODE_LIBS) $(DGA_LIBS) ${XFIXES_LIBS} $(XI2_LIBS) $(X_LIBS) \
 			-lX11 $(X_EXTRA_LIBS) $(X_SHM_LIB) $(NET_LIBS) $(LIBCURL_LIBS) \
@ -133,7 +133,7 @@ qw_client_win_libs= \
 			libs/models/libQFmodels.la \
 			libs/video/targets/libQFwin.la \
 			$(qw_client_LIBS)
-qw_client_win_SOURCES=	qw/source/cl_sys_win.c $(tracy_src)
+qw_client_win_SOURCES=	qw/source/cl_sys_win.c $(tracy_src) $(tracyvk_src)
 qw_client_win_LDADD= $(qw_client_win_libs) -lgdi32 -lwinmm $(NET_LIBS) $(LIBCURL_LIBS)
 qw_client_win_LDFLAGS=  $(common_ldflags)
 qw_client_win_DEPENDENCIES= $(qw_client_win_libs)
--- a/ruamoko/qwaq/Makemodule.am
+++ b/ruamoko/qwaq/Makemodule.am
@ -120,7 +120,7 @@ qwaq_x11_libs= \
 ruamoko_qwaq_qwaq_x11_SOURCES= \
 	ruamoko/qwaq/builtins/main.c \
 	ruamoko/qwaq/builtins/qwaq-graphics.c \
-	ruamoko/qwaq/builtins/graphics.c $(tracy_src)
+	ruamoko/qwaq/builtins/graphics.c $(tracy_src) $(tracyvk_src)
 ruamoko_qwaq_qwaq_x11_LDADD= $(qwaq_x11_libs) $(QWAQ_LIBS) \
 			$(VIDMODE_LIBS) $(DGA_LIBS) ${XFIXES_LIBS} $(XI2_LIBS) $(X_LIBS) \
 			-lX11 $(X_EXTRA_LIBS) $(X_SHM_LIB) $(PTHREAD_LDFLAGS) $(DL_LIBS)