tracy: added Vulkan GPU profiling

src/common/debug.h

@@ -33,7 +33,7 @@ enum MarkersPallete : int {
     EmulatorMarkerColor = 0x264653,
     RendererMarkerColor = 0x2a9d8f,
     HleMarkerColor = 0xe9c46a,
-    Reserved0 = 0xf4a261,
+    GpuMarkerColor = 0xf4a261,
     Reserved1 = 0xe76f51,
 };
 
@@ -48,4 +48,7 @@ enum MarkersPallete : int {
 #define TRACE_CRIT(msg)                                                                            \
     [](const auto& msg) { TracyMessageC(msg.c_str(), msg.size(), tracy::Color::HotPink); }(msg)
 
+#define GPU_SCOPE_LOCATION(name, color)                                                            \
+    tracy::SourceLocationData{name, TracyFunction, TracyFile, (uint32_t)TracyLine, color};
+
 #define FRAME_END FrameMark

src/video_core/renderer_vulkan/renderer_vulkan.cpp

@@ -2,6 +2,7 @@
 // SPDX-License-Identifier: GPL-2.0-or-later
 
 #include "common/config.h"
+#include "common/debug.h"
 #include "common/singleton.h"
 #include "core/file_format/splash.h"
 #include "core/libraries/system/systemservice.h"
@@ -270,6 +271,9 @@ void RendererVulkan::Present(Frame* frame) {
     };
     const vk::CommandBuffer cmdbuf = frame->cmdbuf;
     cmdbuf.begin(begin_info);
+    {
+        TracyVkZoneC(instance.GetProfilerContext(), cmdbuf, "Host frame",
+                     MarkersPallete::GpuMarkerColor);
 
         const vk::Extent2D extent = swapchain.GetExtent();
         const std::array pre_barriers{
@@ -324,8 +328,8 @@ void RendererVulkan::Present(Frame* frame) {
         };
 
         cmdbuf.pipelineBarrier(vk::PipelineStageFlagBits::eColorAttachmentOutput,
-                           vk::PipelineStageFlagBits::eTransfer, vk::DependencyFlagBits::eByRegion,
+                               vk::PipelineStageFlagBits::eTransfer,
-                           {}, {}, pre_barriers);
+                               vk::DependencyFlagBits::eByRegion, {}, {}, pre_barriers);
 
         cmdbuf.blitImage(frame->image, vk::ImageLayout::eTransferSrcOptimal, swapchain_image,
                          vk::ImageLayout::eTransferDstOptimal,
@@ -335,7 +339,8 @@ void RendererVulkan::Present(Frame* frame) {
         cmdbuf.pipelineBarrier(vk::PipelineStageFlagBits::eAllCommands,
                                vk::PipelineStageFlagBits::eAllCommands,
                                vk::DependencyFlagBits::eByRegion, {}, {}, post_barrier);
-
+    }
+    TracyVkCollect(instance.GetProfilerContext(), cmdbuf);
     cmdbuf.end();
 
     static constexpr std::array<vk::PipelineStageFlags, 2> wait_stage_masks = {

src/video_core/renderer_vulkan/vk_instance.cpp

@@ -160,6 +160,7 @@ bool Instance::CreateDevice() {
     // The next two extensions are required to be available together in order to support write masks
     color_write_en = add_extension(VK_EXT_COLOR_WRITE_ENABLE_EXTENSION_NAME);
     color_write_en &= add_extension(VK_EXT_EXTENDED_DYNAMIC_STATE_3_EXTENSION_NAME);
+    const auto calibrated_timestamps = add_extension(VK_EXT_CALIBRATED_TIMESTAMPS_EXTENSION_NAME);
 
     const auto family_properties = physical_device.getQueueFamilyProperties();
     if (family_properties.empty()) {
@@ -212,6 +213,7 @@ bool Instance::CreateDevice() {
         },
         vk::PhysicalDeviceVulkan12Features{
             .scalarBlockLayout = true,
+            .hostQueryReset = true,
             .timelineSemaphore = true,
         },
         vk::PhysicalDeviceVulkan13Features{
@@ -251,6 +253,27 @@ bool Instance::CreateDevice() {
     graphics_queue = device->getQueue(queue_family_index, 0);
     present_queue = device->getQueue(queue_family_index, 0);
 
+    if (calibrated_timestamps) {
+        const auto& time_domains = physical_device.getCalibrateableTimeDomainsEXT();
+#if _WIN64
+        const bool has_host_time_domain =
+            std::find(time_domains.cbegin(), time_domains.cend(),
+                      vk::TimeDomainEXT::eQueryPerformanceCounter) != time_domains.cend();
+#else
+        const bool has_host_time_domain =
+            std::find(time_domains.cbegin(), time_domains.cend(),
+                      vk::TimeDomainEXT::eClockMonotonicRaw) != time_domains.cend();
+#endif
+        if (has_host_time_domain) {
+            static constexpr std::string_view context_name{"vk_rasterizer"};
+            profiler_context =
+                TracyVkContextHostCalibrated(*instance, physical_device, *device,
+                                             VULKAN_HPP_DEFAULT_DISPATCHER.vkGetInstanceProcAddr,
+                                             VULKAN_HPP_DEFAULT_DISPATCHER.vkGetDeviceProcAddr);
+            TracyVkContextName(profiler_context, context_name.data(), context_name.size());
+        }
+    }
+
     CreateAllocator();
     return true;
 }

src/video_core/renderer_vulkan/vk_instance.h

@@ -7,6 +7,9 @@
 
 #include "video_core/renderer_vulkan/vk_platform.h"
 
+#define TRACY_VK_USE_SYMBOL_TABLE
+#include <tracy/TracyVulkan.hpp>
+
 namespace Frontend {
 class WindowSDL;
 }
@@ -67,6 +70,10 @@ public:
         return present_queue;
     }
 
+    TracyVkCtx GetProfilerContext() const {
+        return profiler_context;
+    }
+
     /// Returns true when a known debugging tool is attached.
     bool HasDebuggingToolAttached() const {
         return has_renderdoc || has_nsight_graphics;
@@ -208,6 +215,7 @@ private:
     vk::Queue graphics_queue;
     std::vector<vk::PhysicalDevice> physical_devices;
     std::vector<std::string> available_extensions;
+    TracyVkCtx profiler_context{};
     u32 queue_family_index{0};
     bool image_view_reinterpretation{true};
     bool timeline_semaphores{};

src/video_core/renderer_vulkan/vk_scheduler.cpp

@@ -2,17 +2,21 @@
 // SPDX-License-Identifier: GPL-2.0-or-later
 
 #include <mutex>
+#include "common/debug.h"
 #include "video_core/renderer_vulkan/vk_instance.h"
 #include "video_core/renderer_vulkan/vk_scheduler.h"
 
 namespace Vulkan {
 
 Scheduler::Scheduler(const Instance& instance)
-    : master_semaphore{instance}, command_pool{instance, &master_semaphore} {
+    : instance{instance}, master_semaphore{instance}, command_pool{instance, &master_semaphore} {
+    profiler_scope = reinterpret_cast<tracy::VkCtxScope*>(std::malloc(sizeof(tracy::VkCtxScope)));
     AllocateWorkerCommandBuffers();
 }
 
-Scheduler::~Scheduler() = default;
+Scheduler::~Scheduler() {
+    std::free(profiler_scope);
+}
 
 void Scheduler::Flush(vk::Semaphore signal, vk::Semaphore wait) {
     // When flushing, we only send data to the worker thread; no waiting is necessary.
@@ -41,11 +45,18 @@ void Scheduler::AllocateWorkerCommandBuffers() {
 
     current_cmdbuf = command_pool.Commit();
     current_cmdbuf.begin(begin_info);
+
+    static const auto scope_loc = GPU_SCOPE_LOCATION("Guest Frame", MarkersPallete::GpuMarkerColor);
+    new (profiler_scope)
+        tracy::VkCtxScope{instance.GetProfilerContext(), &scope_loc, current_cmdbuf, true};
 }
 
 void Scheduler::SubmitExecution(vk::Semaphore signal_semaphore, vk::Semaphore wait_semaphore) {
     const u64 signal_value = master_semaphore.NextTick();
 
+    profiler_scope->~VkCtxScope();
+    TracyVkCollect(instance.GetProfilerContext(), current_cmdbuf);
+
     std::scoped_lock lk{submit_mutex};
     master_semaphore.SubmitWork(current_cmdbuf, wait_semaphore, signal_semaphore, signal_value);
     master_semaphore.Refresh();

src/video_core/renderer_vulkan/vk_scheduler.h

@@ -54,10 +54,12 @@ private:
     void SubmitExecution(vk::Semaphore signal_semaphore, vk::Semaphore wait_semaphore);
 
 private:
+    const Instance& instance;
     MasterSemaphore master_semaphore;
     CommandPool command_pool;
     vk::CommandBuffer current_cmdbuf;
     std::condition_variable_any event_cv;
+    tracy::VkCtxScope* profiler_scope{};
 };
 
 } // namespace Vulkan