feat(tracy): integrate Tracy performance profiling support

CMakeLists.txt

@@ -300,6 +300,15 @@ if(MLLM_BUILD_SDK_C_BINDING)
     RUNTIME DESTINATION bin)
 endif()
 
+if(MLLM_TRACY_ENABLE)
+  install(
+    TARGETS MllmTracy
+    EXPORT MllmTargets
+    LIBRARY DESTINATION lib
+    ARCHIVE DESTINATION lib
+    RUNTIME DESTINATION bin)
+endif()
+
 install(
   DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}/mllm/
   DESTINATION include/mllm

mllm/CMakeLists.txt

@@ -80,6 +80,7 @@ endif()
 
 if(MLLM_TRACY_ENABLE)
   add_subdirectory(tracy_perf)
+  target_link_libraries(MllmRT PUBLIC MllmTracy)
 endif()
 
 # Host backend will be build by default

mllm/backends/cpu/CPUAllocator.cpp

@@ -3,6 +3,7 @@
 
 #include "mllm/backends/cpu/CPUAllocator.hpp"
 #include "mllm/backends/cpu/kernels/Kernels.hpp"
+#include "mllm/tracy_perf/Tracy.hpp"
 
 namespace mllm::cpu {
 
@@ -18,12 +19,20 @@ void align_alloc(void** ptr, size_t required_bytes, size_t align) {
     *ptr = nullptr;
     return;
   }
+#if defined(MLLM_TRACY_ENABLE) && MLLM_TRACY_ENABLE == 1
+  TracyAlloc(p1, required_bytes + offset);
+#endif
   p2 = (void**)(((size_t)(p1) + offset) & ~(align - 1));  // NOLINT
   p2[-1] = p1;
   *ptr = p2;
 }
 
-void align_free(void* ptr) { free(((void**)ptr)[-1]); }
+void align_free(void* ptr) {
+#if defined(MLLM_TRACY_ENABLE) && MLLM_TRACY_ENABLE == 1
+  TracyFree(((void**)ptr)[-1]);
+#endif
+  free(((void**)ptr)[-1]);
+}
 
 bool CPUAllocator::alloc(Storage* storage) {
   void* ptr;

mllm/backends/cpu/CPUDispatcher.cpp

@@ -5,6 +5,7 @@
 #include "mllm/engine/Dispatcher.hpp"
 #include "mllm/utils/Common.hpp"
 #include "mllm/nn/Module.hpp"
+#include "mllm/tracy_perf/Tracy.hpp"
 
 #ifdef MLLM_PERFETTO_ENABLE
 #include "mllm/engine/Perf.hpp"
@@ -39,6 +40,7 @@ TaskResult::sender_t CPUDispatcher::asyncReceive(const Task::ptr_t& task) {
 }
 
 void CPUDispatcher::process(const Task::ptr_t& task) {
+  MLLM_TRACY_ZONE_SCOPED;
   switch (task->type) {
     case TaskTypes::kExecuteOp: {
 #ifdef MLLM_PERFETTO_ENABLE

mllm/engine/Context.cpp

@@ -7,6 +7,7 @@
 #include "mllm/engine/Context.hpp"
 #include "mllm/engine/SessionTCB.hpp"
 #include "mllm/engine/DispatcherManager.hpp"
+#include "mllm/tracy_perf/Tracy.hpp"
 
 namespace mllm {
 
@@ -42,6 +43,7 @@ Backend::ptr_t Context::getBackend(const DeviceTypes& device) {
 
 std::vector<Tensor> Context::buildOpAndSubmitTask(OpTypes op_type, const BaseOpOptionsBase& base_options,
                                                   const std::vector<Tensor>& inputs, DeviceTypes special_device) {
+  MLLM_TRACY_ZONE_SCOPED;
   auto device = special_device != kDeviceTypes_End ? special_device : inputs[0].device();
 
   // If input device and special device are different, prefer non-CPU device

mllm/engine/DispatcherManager.cpp

@@ -5,6 +5,7 @@
 #include "exec/static_thread_pool.hpp"
 #include "mllm/utils/Common.hpp"
 #include "mllm/engine/Context.hpp"
+#include "mllm/tracy_perf/Tracy.hpp"
 
 namespace mllm {
 
@@ -14,7 +15,10 @@ DispatcherManager::DispatcherManager(const DispatcherManagerOptions& options)
   exec::numa_policy numa{exec::no_numa_policy{}};
 }
 
-void DispatcherManager::submit(dispatcher_id_t id, const Task::ptr_t& task) { dispatchers_[id]->receive(task); }
+void DispatcherManager::submit(dispatcher_id_t id, const Task::ptr_t& task) {
+  MLLM_TRACY_ZONE_SCOPED;
+  dispatchers_[id]->receive(task);
+}
 
 TaskResult::sender_t DispatcherManager::asyncSubmit(dispatcher_id_t id, const Task::ptr_t& task) {
   return dispatchers_[id]->asyncReceive(task);

mllm/engine/MemoryManager.cpp

@@ -3,6 +3,7 @@
 
 #include "mllm/utils/Common.hpp"
 #include "mllm/engine/MemoryManager.hpp"
+#include "mllm/tracy_perf/Tracy.hpp"
 
 #ifdef MLLM_PERFETTO_ENABLE
 #include "mllm/engine/Perf.hpp"
@@ -25,6 +26,7 @@ void MemoryManager::registerAllocator(const DeviceTypes& device, const Allocator
 }
 
 void MemoryManager::alloc(Storage* s) {
+  MLLM_TRACY_ZONE_SCOPED;
   auto& allocator = allocators_[s->device_];
   auto try_to_alloc_size = allocator->allocSize(s);
 
@@ -58,6 +60,7 @@ void MemoryManager::alloc(Storage* s) {
 void MemoryManager::alloc(const std::shared_ptr<Storage>& s) { alloc(s.get()); }
 
 void MemoryManager::free(Storage* s) {
+  MLLM_TRACY_ZONE_SCOPED;
   auto& allocator = allocators_[s->device_];
   auto try_to_alloc_size = allocator->allocSize(s);
 

mllm/tracy_perf/CMakeLists.txt

@@ -1,5 +1,6 @@
 if(MLLM_TRACY_ENABLE)
   add_library(MllmTracy SHARED Tracy.cpp)
-  target_link_libraries(MllmTracy PUBLIC tracy)
+  target_link_libraries(MllmTracy PUBLIC Tracy::TracyClient)
+  target_include_directories(MllmTracy PUBLIC ${MLLM_INCLUDE_DIR})
   target_compile_definitions(MllmTracy PUBLIC MLLM_TRACY_ENABLE)
 endif()

mllm/tracy_perf/Tracy.cpp

@@ -1,3 +1,6 @@
+// Copyright (c) MLLM Team.
+// Licensed under the MIT License.
+
 #ifdef MLLM_TRACY_ENABLE
-#include "tracy/Tracy.cpp"
-#endif
+#include "mllm/tracy_perf/Tracy.hpp"
+#endif

mllm/tracy_perf/Tracy.hpp

@@ -1,8 +1,10 @@
-#ifndef MLLM_TRACY_HPP
-#define MLLM_TRACY_HPP
+// Copyright (c) MLLM Team.
+// Licensed under the MIT License.
+
+#pragma once
 
 #ifdef MLLM_TRACY_ENABLE
-#include "tracy/Tracy.hpp"
+#include <tracy/Tracy.hpp>
 #define MLLM_TRACY_ZONE_SCOPED ZoneScoped
 #define MLLM_TRACY_ZONE_SCOPED_NAMED(name) ZoneScopedN(name)
 #define MLLM_TRACY_FRAME_MARK FrameMark
@@ -11,5 +13,3 @@
 #define MLLM_TRACY_ZONE_SCOPED_NAMED(name)
 #define MLLM_TRACY_FRAME_MARK
 #endif
-
-#endif  // MLLM_TRACY_HPP

tasks/build_osx_apple_silicon_accelerate.yaml

@@ -10,6 +10,7 @@ Tasks:
         - "-DMLLM_KERNEL_USE_THREADS=ON"
         - "-DMLLM_KERNEL_THREADS_VENDOR_OPENMP=OFF"
         - "-DMLLM_KERNEL_THREADS_VENDOR_APPLE_GCD=ON"
+        - "-DMLLM_TRACY_ENABLE=OFF"
 
   - CMakeBuildTask:
       cmake_cfg_path: "build-osx-accelerate"