basic tracy support via cmake and some fixes in code (allows sample based-profiling, no instrumentation included)

CMakeLists.txt

@@ -219,6 +219,8 @@ cmake_dependent_option(YUZU_CMD "Compile the eden-cli executable" ON "NOT ANDROI
 
 cmake_dependent_option(YUZU_CRASH_DUMPS "Compile crash dump (Minidump) support" OFF "WIN32 OR PLATFORM_LINUX" OFF)
 
+option(ENABLE_TRACY "Enable Tracy profiler support" OFF)
+
 option(YUZU_DOWNLOAD_TIME_ZONE_DATA "Always download time zone binaries" ON)
 set(YUZU_TZDB_PATH "" CACHE STRING "Path to a pre-downloaded timezone database")
 
@@ -662,6 +664,29 @@ if (WIN32 AND YUZU_CRASH_DUMPS)
     # target_include_directories(libbreakpad_client INTERFACE "${BREAKPAD_CLIENT_INCLUDE_DIR}")
 endif()
 
+if (ENABLE_TRACY)
+    AddJsonPackage(tracy)
+
+    if (tracy_ADDED)
+        include_directories(${tracy_SOURCE_DIR}/public)
+        # build tracy client as its own static library even if overkill
+        add_library(tracy_client STATIC ${tracy_SOURCE_DIR}/public/TracyClient.cpp)
+        target_include_directories(tracy_client PRIVATE src)
+
+        # Avoid tracy errors (This sets it to W0 instead of W3 though? TODO: fix this)
+        if (MSVC AND CXX_CLANG)
+            target_compile_options(tracy_client PRIVATE /W3)
+        endif()
+
+        add_compile_definitions(TRACY_ENABLE)
+        add_compile_definitions(TRACY_SAMPLING_PROFILER_MANUAL_START) # See yuzu\main.cpp
+        add_compile_definitions(TRACY_ON_DEMAND) # Use on demand mode to avoid profiling emulator start up and game load screens
+        add_compile_definitions(TRACY_FIBERS)
+
+        add_library(Tracy::client ALIAS tracy_client)
+    endif()
+endif()
+
 # Include source code
 # ===================
 

cpmfile.json

@@ -114,5 +114,12 @@
             "QUAZIP_INSTALL OFF",
             "QUAZIP_ENABLE_QTEXTCODEC OFF"
         ]
+    },
+    "tracy": {
+        "repo": "wolfpld/tracy",
+        "sha": "05cceee",
+        "hash": "fdf8f3eb0f44c17760e9e559ece6907606580da20568b7900e97e40f3ea773b9e6dbac7f0b04ef32e363d779ec013af63c85adbe2a3807db9205ec48887a546c",
+        "version": "0.13.1",
+        "download_only": true
     }
 }

src/common/fiber.cpp

@@ -13,6 +13,9 @@
 
 #include <boost/context/detail/fcontext.hpp>
 
+//#include "common/tracy_instrumentation.h"
+//#include <sstream>
+
 namespace Common {
 
 #ifdef __OPENORBIS__
@@ -42,6 +45,13 @@ struct Fiber::FiberImpl {
     u8* rewind_stack_limit = nullptr;
     bool is_thread_fiber = false;
     bool released = false;
+
+// Support tracy fibers, show them using incrementing counter in profiler
+// TODO: Commented out because I could not properly test because CPU threads have messed up ghost zones due to jit code failing to support stack traces
+//#if TRACY_ENABLE
+//    std::string tracy_fiber_name;
+//    static inline int tracy_fiber_next_id = 1;
+//#endif
 };
 
 void Fiber::SetRewindPoint(std::function<void()>&& rewind_func) {
@@ -49,12 +59,20 @@ void Fiber::SetRewindPoint(std::function<void()>&& rewind_func) {
 }
 
 Fiber::Fiber(std::function<void()>&& entry_point_func) : impl{std::make_unique<FiberImpl>()} {
+//#if TRACY_ENABLE
+//    std::ostringstream ss;
+//    ss << "fiber #" << impl->tracy_fiber_next_id++;
+//    impl->tracy_fiber_name = std::move(ss).str();
+//#endif
     impl->entry_point = std::move(entry_point_func);
     impl->stack_limit = impl->stack.data();
     impl->rewind_stack_limit = impl->rewind_stack.data();
     u8* stack_base = impl->stack_limit + DEFAULT_STACK_SIZE;
     impl->context = boost::context::detail::make_fcontext(stack_base, impl->stack.size(), [](boost::context::detail::transfer_t transfer) -> void {
         auto* fiber = static_cast<Fiber*>(transfer.data);
+    //#if TRACY_ENABLE
+    //    TracyFiberEnter(fiber->impl->tracy_fiber_name.c_str());
+    //#endif
         ASSERT(fiber && fiber->impl && fiber->impl->previous_fiber && fiber->impl->previous_fiber->impl);
         ASSERT(fiber->impl->canary_1 == CANARY_VALUE);
         ASSERT(fiber->impl->canary_2 == CANARY_VALUE);
@@ -91,7 +109,16 @@ void Fiber::YieldTo(std::weak_ptr<Fiber> weak_from, Fiber& to) {
     to.impl->guard.lock();
     to.impl->previous_fiber = weak_from.lock();
 
+//#if TRACY_ENABLE
+//    // Is this correct?
+//    TracyFiberLeave;
+//#endif
+
     auto transfer = boost::context::detail::jump_fcontext(to.impl->context, &to);
+//#if TRACY_ENABLE
+//    // Switched to executing "to" fiber, inform tracy
+//    TracyFiberEnter(to.impl->tracy_fiber_name.c_str());
+//#endif
     // "from" might no longer be valid if the thread was killed
     if (auto from = weak_from.lock()) {
         if (from->impl->previous_fiber == nullptr) {

src/common/tracy_instrumentation.h

@@ -0,0 +1,22 @@
+#pragma once
+
+// Add our own tracy instrumentation macros which compile to nothing without TRACY_ENABLE
+// tracys macros alread do that but do require you to always include the full tracy code even without TRACY_ENABLE
+// Doing this means any instrumentation in the project keeps compiling even with ENABLE_TRACY=0 in cmake, meaning most devs can compile without even pulling tracy at all
+// Additionally These macros may be more readable than the badly named "ZoneScoped", and we could easily add levels or categories of instrumentation that can be made selectable
+// Unless there is already a tracing framework that is configurable...?
+#if TRACY_ENABLE
+#include <tracy/Tracy.hpp>
+
+#define TRACY_ZONE_SCOPED ZoneScoped;
+#define TRACY_ZONE_SCOPEDN(name) ZoneScopedN(name);
+#define TRACY_ZONE_SCOPEDVN(var, name) ZoneScopedVN(var, name);
+#define TRACY_ZONE_NAMEDN(var, name) ZoneNamedN(var, name, true);
+#define TRACY_MSG_C(text, col) TracyMessageC(text, strlen(text), col);
+#else
+#define TRACY_ZONE_SCOPED
+#define TRACY_ZONE_SCOPEDN(name)
+#define TRACY_ZONE_SCOPEDVN(var, name)
+#define TRACY_ZONE_NAMEDN(var, name)
+#define TRACY_MSG_C(text, col)
+#endif

src/video_core/renderer_vulkan/vk_swapchain.cpp

@@ -19,6 +19,8 @@
 #include "video_core/vulkan_common/vulkan_wrapper.h"
 #include "vulkan/vulkan_core.h"
 
+#include "common/tracy_instrumentation.h"
+
 namespace Vulkan {
 
 namespace {
@@ -226,6 +228,10 @@ void Swapchain::Present(VkSemaphore render_semaphore) {
     if (frame_index >= image_count) {
         frame_index = 0;
     }
+
+    #if TRACY_ENABLE
+        FrameMark;
+    #endif
 }
 
 void Swapchain::CreateSwapchain(const VkSurfaceCapabilitiesKHR& capabilities) {

src/yuzu/CMakeLists.txt

@@ -401,6 +401,10 @@ target_link_libraries(yuzu PRIVATE common core input_common frontend_common netw
 target_link_libraries(yuzu PRIVATE Boost::headers glad Qt6::Widgets Qt6::Charts Qt6::Concurrent)
 target_link_libraries(yuzu PRIVATE ${PLATFORM_LIBRARIES} Threads::Threads)
 
+if (ENABLE_TRACY)
+    target_link_libraries(yuzu PRIVATE Tracy::client)
+endif()
+
 if (UNIX AND NOT APPLE)
     target_link_libraries(yuzu PRIVATE Qt6::DBus)
 endif()

src/yuzu/main.cpp

@@ -20,6 +20,8 @@
 #ifdef _WIN32
 #include <QScreen>
 
+#include "common/tracy_instrumentation.h"
+
 static void OverrideWindowsFont() {
     // Qt5 chooses these fonts on Windows and they have fairly ugly alphanumeric/cyrillic characters
     // Asking to use "MS Shell Dlg 2" gives better other chars while leaving the Chinese Characters.
@@ -91,6 +93,17 @@ int main(int argc, char* argv[]) {
         return 0;
     }
 
+#if TRACY_ENABLE
+    // Avoid cmake static library for tracy being optimized away by "helpful" linker if when not referenced (no manual instrumentation, but still want sample-based mode)
+    //TracyNoop; // This would be needed if not for tracy::BeginSamplingProfiling
+
+    // have to use TRACY_SAMPLING_PROFILER_MANUAL_START because program is started a second time as child process for vulkan check
+    // this messes with tracy sample profiling as windows event tracing in child process stops it for parent process
+    if (!is_child) {
+        tracy::BeginSamplingProfiling();
+    }
+#endif
+
     if (StartupChecks(argv[0], &has_broken_vulkan,
                       Settings::values.perform_vulkan_check.GetValue())) {
         return 0;
@@ -177,5 +190,11 @@ int main(int argc, char* argv[]) {
 
     int result = app.exec();
     detached_tasks.WaitForAllTasks();
+
+#if TRACY_ENABLE
+    if (!is_child) {
+        tracy::EndSamplingProfiling();
+    }
+#endif
     return result;
 }

src/yuzu_cmd/CMakeLists.txt

@@ -55,6 +55,10 @@ if(WIN32)
     endif()
 endif()
 
+if (ENABLE_TRACY)
+    target_link_libraries(yuzu-cmd PRIVATE Tracy::client)
+endif()
+
 create_target_directory_groups(yuzu-cmd)
 
 # needed for vma

src/yuzu_room_standalone/CMakeLists.txt

@@ -11,6 +11,10 @@ if(UNIX AND NOT APPLE)
     install(TARGETS yuzu_room_standalone RUNTIME DESTINATION "${CMAKE_INSTALL_PREFIX}/bin")
 endif()
 
+if (ENABLE_TRACY)
+    target_link_libraries(yuzu_room_standalone PRIVATE Tracy::client)
+endif()
+
 if (YUZU_STATIC_ROOM)
     target_link_options(yuzu_room_standalone PRIVATE "-static")
 endif()