feat(profiler): integrate cupti backend

GitOrigin-RevId: dec8be1908

CMakeLists.txt

@@ -29,6 +29,7 @@ endif()
 include(GNUInstallDirs)
 include(CheckCXXCompilerFlag)
 include(CheckIPOSupported)
+include(CMakeDependentOption)
 
 check_cxx_compiler_flag(-Wclass-memaccess CXX_SUPPORT_WCLASS_MEMACCESS)
 
@@ -97,6 +98,12 @@ option(MGE_BUILD_WITH_ASAN "Enable build with ASAN, need compiler support" OFF)
 option(MGE_WITH_CUSTOM_OP "Build with Custom op" OFF)
 option(MGE_SYNC_THIRD_PARTY "help sync third_party submodule" OFF)
 
+# TODO: add windows support
+cmake_dependent_option(MGE_WITH_CUPTI "Build with CUPTI" ON
+                       "MGE_WITH_CUDA;MGE_BUILD_IMPERATIVE_RT;NOT MSVC;NOT WIN32" OFF)
+
+set(MGB_CUPTI ${MGE_WITH_CUPTI})
+
 if(MSVC OR WIN32)
   # FIXME: static link Windows vc runtime with some version from Visual Studio have some
   # runtime issue at some call PATH, for example: _imperative_rt.pyd -->
@@ -686,6 +693,10 @@ if(MGB_WITH_FLATBUFFERS)
   include(cmake/flatbuffers.cmake)
 endif()
 
+if(MGE_WITH_CUPTI)
+  include(cmake/cupti.cmake)
+endif()
+
 if(MGE_WITH_CUDA)
   include_directories(${CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES})
   foreach(path ${CMAKE_CUDA_HOST_IMPLICIT_LINK_DIRECTORIES})

cmake/cudnn.cmake

@@ -6,7 +6,7 @@ endif()
 if("${CUDNN_ROOT_DIR}" STREQUAL "" AND NOT "$ENV{CUDNN_ROOT_DIR}" STREQUAL "")
   set(CUDNN_ROOT_DIR $ENV{CUDNN_ROOT_DIR})
 endif()
-message("CUDNN ROOT: " ${CUDNN_ROOT_DIR})
+message(STATUS "CUDNN ROOT: ${CUDNN_ROOT_DIR}")
 if(MGE_CUDA_USE_STATIC AND NOT MGE_WITH_CUDNN_SHARED)
   find_library(
     CUDNN_LIBRARY

cmake/cupti.cmake

@@ -0,0 +1,85 @@
+if("${CUDA_ROOT_DIR}" STREQUAL "" AND NOT "$ENV{CUDA_ROOT_DIR}" STREQUAL "")
+  set(CUDA_ROOT_DIR $ENV{CUDA_ROOT_DIR})
+endif()
+if("${CUDA_ROOT_DIR}" STREQUAL "" AND NOT "$ENV{CUDA_PATH}" STREQUAL "")
+  set(CUDA_ROOT_DIR $ENV{CUDA_PATH})
+endif()
+if("${CUDA_ROOT_DIR}" STREQUAL "" AND NOT "$ENV{CUDA_BIN_PATH}" STREQUAL "")
+  set(CUDA_ROOT_DIR $ENV{CUDA_BIN_PATH})
+endif()
+if("${CUDA_ROOT_DIR}" STREQUAL "")
+  message(
+    FATAL_ERROR
+      "Can not find CUDA, please export cuda sdk path to CUDA_ROOT_DIR or CUDA_PATH or CUDA_BIN_PATH"
+  )
+endif()
+# TODO: find_library(CUDA_ROOT_DIR) in cmake/cuda.cmake
+
+set(MGE_CUPTI_USE_STATIC ${MGE_CUDA_USE_STATIC})
+
+# relates https://stackoverflow.com/questions/67485114
+if(${MGE_CUDA_USE_STATIC} AND ${CXX_SUPPORT_GOLD})
+  message(WARNING "static linking CuPTI with gold may break exception handling,\
+                  use shared one instead")
+  set(MGE_CUPTI_USE_STATIC OFF)
+endif()
+
+if(MGE_CUPTI_USE_STATIC)
+  find_library(
+    CUPTI_LIBRARY
+    NAMES libcupti_static.a
+    HINTS ${CUDA_ROOT_DIR} ${CUDA_ROOT_DIR}/extras/CUPTI
+    PATH_SUFFIXES lib lib64
+    DOC "CuPTI library.")
+
+  if("${CUPTI_LIBRARY}" STREQUAL "CUPTI_LIBRARY-NOTFOUND")
+    message(WARNING "Can not find static CuPTI Library, use shared one instead")
+    set(MGE_CUPTI_USE_STATIC OFF)
+  endif()
+endif()
+
+if(NOT ${MGE_CUPTI_USE_STATIC})
+  find_library(
+    CUPTI_LIBRARY
+    NAMES libcupti.so
+    HINTS ${CUDA_ROOT_DIR} ${CUDA_ROOT_DIR}/extras/CUPTI
+    PATH_SUFFIXES lib lib64
+    DOC "CuPTI library.")
+  set(CUPTI_LIBRARY_TYPE SHARED)
+else()
+  set(CUPTI_LIBRARY_TYPE STATIC)
+endif()
+
+if("${CUPTI_LIBRARY}" STREQUAL "CUPTI_LIBRARY-NOTFOUND")
+  message(FATAL_ERROR "Can not find CuPTI Library")
+endif()
+
+find_path(
+  CUPTI_INCLUDE_DIR
+  NAMES cupti.h
+  HINTS ${CUDA_ROOT_DIR} ${CUDA_ROOT_DIR}/extras/CUPTI
+  PATH_SUFFIXES include
+  DOC "Path to CuPTI include directory.")
+
+if(CUPTI_INCLUDE_DIR STREQUAL "CUPTI_INCLUDE_DIR-NOTFOUND")
+  message(FATAL_ERROR "Can not find CuPTI INCLUDE")
+endif()
+
+if(EXISTS ${CUPTI_INCLUDE_DIR}/cupti_version.h)
+  file(READ ${CUPTI_INCLUDE_DIR}/cupti_version.h CUPTI_VERSION_FILE_CONTENTS)
+else()
+  file(READ ${CUPTI_INCLUDE_DIR}/cupti.h CUPTI_VERSION_FILE_CONTENTS)
+endif()
+
+string(REGEX MATCH "define CUPTI_API_VERSION * +([0-9]+)" CUPTI_API_VERSION
+             "${CUPTI_VERSION_FILE_CONTENTS}")
+string(REGEX REPLACE "define CUPTI_API_VERSION * +([0-9]+)" "\\1" CUPTI_API_VERSION
+                     "${CUPTI_API_VERSION}")
+
+add_library(libcupti ${CUPTI_LIBRARY_TYPE} IMPORTED)
+
+set_target_properties(
+  libcupti PROPERTIES IMPORTED_LOCATION ${CUPTI_LIBRARY} INTERFACE_INCLUDE_DIRECTORIES
+                                                         ${CUPTI_INCLUDE_DIR})
+
+message(STATUS "Found CuPTI: ${CUPTI_LIBRARY} (found version: ${CUPTI_API_VERSION})")

cmake/tensorrt.cmake

@@ -36,7 +36,7 @@ else()
     PATH_SUFFIXES lib lib64
     DOC "TRT plugin library.")
 endif()
-message("TRT_LIBRARY" ${TRT_LIBRARY})
+message(STATUS "TRT_LIBRARY: ${TRT_LIBRARY}")
 if(TRT_LIBRARY STREQUAL "TRT_LIBRARY-NOTFOUND")
   message(
     FATAL_ERROR

imperative/CMakeLists.txt

@@ -51,6 +51,10 @@ if(ANDROID)
   target_link_libraries(${MODULE_NAME} PRIVATE ${PYTHON_LIBRARIES})
 endif()
 
+if(MGE_WITH_CUPTI)
+  target_link_libraries(${MODULE_NAME} PRIVATE libcupti)
+endif()
+
 add_subdirectory(${PROJECT_SOURCE_DIR}/third_party/range-v3
                  ${PROJECT_BINARY_DIR}/third_party/range-v3)
 target_link_libraries(${MODULE_NAME} PRIVATE range-v3)

imperative/python/megengine/utils/profiler.py

@@ -16,6 +16,10 @@ from weakref import WeakSet
 
 from .. import _atexit
 from ..core._imperative_rt.core2 import (
+    cupti_available,
+    disable_cupti,
+    enable_cupti,
+    full_sync,
     pop_scope,
     push_scope,
     start_profile,
@@ -50,13 +54,18 @@ class Profiler(ContextDecorator):
 
               with profiler:
                  # your code here
-    
+
            # Then open the profile file in chrome timeline window
     """
 
     CHROME_TIMELINE = "chrome_timeline.json"
 
-    valid_options = {"sample_rate": 0, "profile_device": 1, "num_tensor_watch": 10}
+    valid_options = {
+        "sample_rate": 0,
+        "profile_device": 1,
+        "num_tensor_watch": 10,
+        "enable_cupti": 0,
+    }
     valid_formats = {"chrome_timeline.json", "memory_flow.svg"}
 
     def __init__(
@@ -83,6 +92,11 @@ class Profiler(ContextDecorator):
             self._options[opt] = int(kwargs.pop(opt, optval))
         self._pid = "<PID>"
         self._dump_callback = None
+        if self._options.get("enable_cupti", 0):
+            if cupti_available():
+                enable_cupti()
+            else:
+                get_logger().warning("CuPTI unavailable")
 
     @property
     def path(self):
@@ -116,7 +130,7 @@ class Profiler(ContextDecorator):
 
         assert _running_profiler is self
         _running_profiler = None
-        sync()
+        full_sync()
         self._dump_callback = stop_profile()
         self._pid = os.getpid()
         _living_profilers.add(self)
@@ -160,6 +174,9 @@ class Profiler(ContextDecorator):
         return func
 
     def __del__(self):
+        if self._options.get("enable_cupti", 0):
+            if cupti_available():
+                disable_cupti()
         self.dump()
 
 

imperative/python/src/tensor.cpp

@@ -11,6 +11,7 @@
 
 #include "megbrain/common.h"
 #include "megbrain/dtype.h"
+#include "megbrain/imperative/cpp_cupti.h"
 #include "megbrain/imperative/ops/autogen.h"
 #include "megbrain/imperative/ops/backward_graph.h"
 #include "megbrain/imperative/ops/utility.h"
@@ -982,6 +983,7 @@ void init_tensor(py::module m) {
     m.def("stop_profile", [channel]() -> std::function<void(std::string, std::string)> {
         channel->stop_profile();
         channel->sync();
+        CompNode::sync_all();
         imperative::Profiler::stop_profile();
         auto results = std::make_shared<imperative::Profiler::bundle_t>(
                 imperative::Profiler::collect());
@@ -990,6 +992,9 @@ void init_tensor(py::module m) {
             results = nullptr;
         };
     });
+    m.def("enable_cupti", &cupti::enable);
+    m.def("disable_cupti", &cupti::disable);
+    m.def("cupti_available", &cupti::available);
     m.def("sync", [channel]() {
         if (channel->check_available()) {
             channel->sync();

imperative/src/impl/cpp_cupti.cpp

@@ -0,0 +1,273 @@
+#include "megbrain/imperative/cpp_cupti.h"
+
+#include <cinttypes>
+#include <cstddef>
+#include <cstdlib>
+
+#include "megbrain/exception.h"
+#include "megbrain/imperative/profiler.h"
+#include "megbrain/imperative/utils/platform.h"
+
+#include "./profiler/events.h"
+
+#if MGB_CUPTI
+#include "cupti.h"
+
+#define CUPTI_CALL(call)                                                     \
+    do {                                                                     \
+        CUptiResult _status = call;                                          \
+        if (_status != CUPTI_SUCCESS) {                                      \
+            const char* errstr;                                              \
+            cuptiGetResultString(_status, &errstr);                          \
+            mgb_assert(_status == CUPTI_SUCCESS, "cupti error: %s", errstr); \
+        }                                                                    \
+    } while (0)
+#endif
+
+namespace mgb::imperative::cupti {
+
+#if MGB_CUPTI
+namespace {
+CUpti_SubscriberHandle cuptiSubscriber;
+
+void cuptiSubscriberCallback(
+        void* userdata, CUpti_CallbackDomain domain, CUpti_CallbackId cb_id,
+        const void* cb_info) {
+    using namespace profiler;
+    switch (domain) {
+        case CUPTI_CB_DOMAIN_DRIVER_API: {
+            auto cb_data = (const CUpti_CallbackData*)cb_info;
+            switch (cb_id) {
+                case CUPTI_DRIVER_TRACE_CBID_cuLaunchKernel: {
+                    if (cb_data->callbackSite == CUPTI_API_ENTER) {
+                        MGB_RECORD_EVENT(
+                                CUPTIKernelLaunchEvent, cb_data->correlationId,
+                                cb_data->symbolName);
+                    } else if (cb_data->callbackSite == CUPTI_API_EXIT) {
+                        MGB_RECORD_EVENT(
+                                CUPTIKernelLaunchFinishEvent, cb_data->correlationId,
+                                cb_data->symbolName);
+                    }
+                    break;
+                }
+                case CUPTI_DRIVER_TRACE_CBID_cuMemcpyHtoA: {
+                }
+                case CUPTI_DRIVER_TRACE_CBID_cuMemcpyAsync: {
+                    if (cb_data->callbackSite == CUPTI_API_ENTER) {
+                        MGB_RECORD_EVENT(
+                                CUPTIMemcpyLaunchEvent, cb_data->correlationId);
+                    } else if (cb_data->callbackSite == CUPTI_API_EXIT) {
+                        MGB_RECORD_EVENT(
+                                CUPTIMemcpyLaunchFinishEvent, cb_data->correlationId);
+                    }
+                    break;
+                }
+                default: {
+                    if (cb_data->callbackSite == CUPTI_API_ENTER) {
+                        MGB_RECORD_EVENT(
+                                CUPTIDriverEvent, cb_data->correlationId,
+                                cb_data->functionName);
+                    } else if (cb_data->callbackSite == CUPTI_API_EXIT) {
+                        MGB_RECORD_EVENT(
+                                CUPTIDriverFinishEvent, cb_data->correlationId,
+                                cb_data->functionName);
+                    }
+                }
+            }
+            break;
+        }
+        case CUPTI_CB_DOMAIN_RUNTIME_API: {
+            auto cb_data = (const CUpti_CallbackData*)cb_info;
+            if (cb_data->callbackSite == CUPTI_API_ENTER) {
+                MGB_RECORD_EVENT(
+                        CUPTIRuntimeEvent, cb_data->correlationId,
+                        cb_data->functionName);
+            } else if (cb_data->callbackSite == CUPTI_API_EXIT) {
+                MGB_RECORD_EVENT(
+                        CUPTIRuntimeFinishEvent, cb_data->correlationId,
+                        cb_data->functionName);
+            }
+            break;
+        }
+    }
+}
+
+void handleActivity(CUpti_Activity* record) {
+    using namespace std::chrono_literals;
+    auto delta = 16ns;
+    switch (record->kind) {
+        case CUPTI_ACTIVITY_KIND_KERNEL:
+        case CUPTI_ACTIVITY_KIND_CONCURRENT_KERNEL: {
+            auto kernel = cupti::activity<CUpti_ActivityKernel4>(record);
+            MGB_RECORD_EVENT(
+                    profiler::CUPTIKernelExecuteEvent, kernel->correlationId,
+                    kernel->name, kernel.stream(), kernel.start(),
+                    kernel.end() - delta);
+            break;
+        }
+        case CUPTI_ACTIVITY_KIND_MEMCPY: {
+            auto memcpy = cupti::activity<CUpti_ActivityMemcpy>(record);
+            MGB_RECORD_EVENT(
+                    profiler::CUPTIMemcpyEvent, memcpy->correlationId, memcpy->srcKind,
+                    memcpy->dstKind, memcpy->bytes, memcpy.stream(), memcpy.start(),
+                    memcpy.end());
+            break;
+        }
+        case CUPTI_ACTIVITY_KIND_MEMSET: {
+            auto memset = cupti::activity<CUpti_ActivityMemset>(record);
+            MGB_RECORD_EVENT(
+                    profiler::CUPTIMemsetEvent, memset->correlationId, memset->value,
+                    memset->bytes, memset.stream(), memset.start(),
+                    memset.end() - delta);
+            break;
+        }
+        default:
+            break;
+    }
+}
+
+using activity_buffer_t =
+        std::aligned_storage_t<8 * 1024 * 1024, ACTIVITY_RECORD_ALIGNMENT>;
+
+void bufferRequested(uint8_t** buffer, size_t* size, size_t* maxNumRecords) {
+    *buffer = reinterpret_cast<uint8_t*>(new activity_buffer_t());
+    *size = sizeof(activity_buffer_t);
+    *maxNumRecords = 0;
+}
+
+void bufferCompleted(
+        CUcontext ctx, uint32_t streamId, uint8_t* buffer, size_t size,
+        size_t validSize) {
+    CUptiResult status;
+    CUpti_Activity* record = NULL;
+
+    if (validSize > 0) {
+        do {
+            status = cuptiActivityGetNextRecord(buffer, validSize, &record);
+            if (status == CUPTI_SUCCESS) {
+                handleActivity(record);
+            } else if (status == CUPTI_ERROR_MAX_LIMIT_REACHED)
+                break;
+            else {
+                CUPTI_CALL(status);
+            }
+        } while (1);
+
+        size_t dropped;
+        CUPTI_CALL(cuptiActivityGetNumDroppedRecords(ctx, streamId, &dropped));
+        mgb_assert(dropped == 0, "%zu records dropped", dropped);
+    }
+
+    delete reinterpret_cast<activity_buffer_t*>(buffer);
+}
+
+static bool initialized = false;
+}  // namespace
+
+bool available() {
+    uint32_t compiletime_version = (CUPTI_API_VERSION);
+    uint32_t runtime_version;
+    CUPTI_CALL(cuptiGetVersion(&runtime_version));
+    if (compiletime_version != runtime_version) {
+        static std::once_flag once;
+        std::call_once(once, [&] {
+            mgb_log_warn(
+                    "CuPTI version %d mismatch against compiletime version %d. "
+                    "This may caused by user config LD_LIBRARY_PATH"
+                    "at unix-like env or config PATH at Windows env",
+                    (int)compiletime_version, (int)runtime_version);
+        });
+        return false;
+    }
+    return true;
+}
+
+void enable() {
+    // not thread safe
+    mgb_assert(!initialized, "cupti already initialized");
+    // callback
+    CUPTI_CALL(cuptiSubscribe(
+            &cuptiSubscriber, (CUpti_CallbackFunc)cuptiSubscriberCallback,
+            (void*)nullptr));
+    CUPTI_CALL(cuptiEnableDomain(1, cuptiSubscriber, CUPTI_CB_DOMAIN_DRIVER_API));
+    CUPTI_CALL(cuptiEnableDomain(1, cuptiSubscriber, CUPTI_CB_DOMAIN_RUNTIME_API));
+
+    // activity
+    CUPTI_CALL(cuptiActivityEnable(CUPTI_ACTIVITY_KIND_DEVICE));
+    CUPTI_CALL(cuptiActivityEnable(CUPTI_ACTIVITY_KIND_CONTEXT));
+    CUPTI_CALL(cuptiActivityEnable(CUPTI_ACTIVITY_KIND_DRIVER));
+    CUPTI_CALL(cuptiActivityEnable(CUPTI_ACTIVITY_KIND_RUNTIME));
+    CUPTI_CALL(cuptiActivityEnable(CUPTI_ACTIVITY_KIND_MEMCPY));
+    CUPTI_CALL(cuptiActivityEnable(CUPTI_ACTIVITY_KIND_MEMSET));
+    CUPTI_CALL(cuptiActivityEnable(CUPTI_ACTIVITY_KIND_NAME));
+    CUPTI_CALL(cuptiActivityEnable(CUPTI_ACTIVITY_KIND_MARKER));
+    CUPTI_CALL(cuptiActivityEnable(CUPTI_ACTIVITY_KIND_CONCURRENT_KERNEL));
+    CUPTI_CALL(cuptiActivityEnable(CUPTI_ACTIVITY_KIND_OVERHEAD));
+    CUPTI_CALL(cuptiActivityRegisterCallbacks(bufferRequested, bufferCompleted));
+    initialized = true;
+}
+
+void disable() {
+    mgb_assert(initialized, "cupti not initialized yet");
+    flush();
+    CUPTI_CALL(cuptiFinalize());
+    initialized = false;
+}
+
+void flush() {
+    if (initialized) {
+        CUPTI_CALL(cuptiActivityFlushAll(1));
+    }
+}
+
+bool enabled() {
+    return initialized;
+}
+
+time_point clock::now() {
+    uint64_t timestamp;
+    CUPTI_CALL(cuptiGetTimestamp(&timestamp));
+    using namespace std::chrono;
+    // overflow?
+    return time_point(duration((int64_t)timestamp));
+}
+
+#else
+
+class CuPTIUnavailableError : public MegBrainError {
+public:
+    CuPTIUnavailableError()
+            : MegBrainError(
+#if MGB_CUDA
+                      "CuPTI disabled at compile time"
+#else
+                      "CuPTI unsupported on non cuda platform"
+#endif
+              ) {
+    }
+};
+
+bool available() {
+    return false;
+}
+
+void enable() {
+    throw CuPTIUnavailableError();
+}
+
+void disable() {
+    throw CuPTIUnavailableError();
+}
+
+void flush() {}
+
+bool enabled() {
+    return false;
+}
+
+time_point clock::now() {
+    throw CuPTIUnavailableError();
+}
+
+#endif
+}  // namespace mgb::imperative::cupti

imperative/src/impl/profiler.cpp

@@ -12,7 +12,9 @@
 #include "megbrain/imperative/profiler.h"
 
 #include <chrono>
+#include <unordered_map>
 
+#include "megbrain/imperative/cpp_cupti.h"
 #include "megbrain/imperative/ops/opr_attr.h"
 #include "megbrain/imperative/physical_tensor.h"
 
@@ -48,6 +50,21 @@ bool Profiler::sm_profiling = false;
 thread_local Profiler* Profiler::tm_profiler = nullptr;
 std::atomic_size_t Profiler::sm_preferred_capacity;
 
+void Profiler::start_profile() {
+    mgb_assert(!sm_profiling);
+    sm_start_at = Timer::record_host();
+    sm_profiling = true;
+    if (cupti::enabled()) {
+        MGB_RECORD_EVENT(profiler::CUPTITimestampEvent, cupti::clock::now());
+    }
+}
+
+void Profiler::stop_profile() {
+    mgb_assert(sm_profiling);
+    cupti::flush();
+    sm_profiling = false;
+}
+
 auto Profiler::get_thread_dict() -> thread_dict_t {
     thread_dict_t thread_dict;
     for (auto&& [tid, profiler] : sm_profilers) {

imperative/src/impl/profiler/chrome_timeline.cpp

@@ -19,6 +19,7 @@
 
 #include "nlohmann/json.hpp"
 
+#include "megbrain/imperative/utils/platform.h"
 #include "megbrain/utils/debug.h"
 
 #include "./formats.h"
@@ -198,6 +199,8 @@ struct ChromeTimelineEventVisitor : EventVisitor<ChromeTimelineEventVisitor> {
     decltype(getpid()) pid = getpid();
     std::string pid_str = std::to_string(pid);
 
+    ChromeTimelineEventVisitor() {}
+
     ChromeTraceEvent& new_event(
             std::string name, char ph, size_t tid, profiler::HostTime time) {
         return trace_events.new_event().name(name).ph(ph).pid(pid).tid(tid).ts(
@@ -213,8 +216,13 @@ struct ChromeTimelineEventVisitor : EventVisitor<ChromeTimelineEventVisitor> {
                 .ts(since_start(current->time));
     }
 
+    ChromeTraceEvent& new_cupti_event(
+            std::string name, char ph, cupti::stream_t stream,
+            cupti::time_point timestamp) {
+        return new_event(name, ph, to_tid(stream), time_from_cupti(timestamp));
+    }
+
     ChromeTraceEvent& new_device_event(std::string name, char ph, CompNode device) {
-        using namespace std::literals::chrono_literals;
         auto time = since_start(to_device_time(current->time, device));
         return trace_events.new_event()
                 .name(name)
@@ -391,6 +399,80 @@ struct ChromeTimelineEventVisitor : EventVisitor<ChromeTimelineEventVisitor> {
             auto device_ahead = std::chrono::duration_cast<std::chrono::milliseconds>(
                     current_device_time - current_host_time);
             new_host_event("device_ahead_ms", 'C').arg("value", device_ahead.count());
+        } else if constexpr (std::is_same_v<TEvent, CUPTIKernelLaunchEvent>) {
+            new_host_event(demangle(event.name), 'B');
+            new_host_event(pid_str, 's')
+                    .id(event.correlation_id)
+                    .cat("KernelLink")
+                    .scope(pid_str);
+        } else if constexpr (std::is_same_v<TEvent, CUPTIKernelLaunchFinishEvent>) {
+            new_host_event(demangle(event.name), 'E');
+        } else if constexpr (std::is_same_v<TEvent, CUPTIKernelExecuteEvent>) {
+            new_cupti_event(demangle(event.name), 'B', event.stream, event.start)
+                    .arg("execution_time", (event.end - event.start).count());
+            new_cupti_event(pid_str, 'f', event.stream, event.end)
+                    .id(event.correlation_id)
+                    .bp('e')
+                    .cat("KernelLink")
+                    .scope(pid_str);
+            new_cupti_event(demangle(event.name), 'E', event.stream, event.end)
+                    .arg("execution_time", (event.end - event.start).count());
+        } else if constexpr (std::is_same_v<TEvent, CUPTIMemcpyLaunchEvent>) {
+            new_host_event("Memcpy", 'B');
+            new_host_event(pid_str, 's')
+                    .id(event.correlation_id)
+                    .cat("CUPTILink")
+                    .scope(pid_str);
+        } else if constexpr (std::is_same_v<TEvent, CUPTIMemcpyLaunchFinishEvent>) {
+            new_host_event("Memcpy", 'E');
+        } else if constexpr (std::is_same_v<TEvent, CUPTIMemcpyEvent>) {
+            auto memkind2str = [](uint8_t kind) {
+                const char* const valid_kinds[] = {
+                        "CUPTI_ACTIVITY_MEMORY_KIND_UNKNOWN",
+                        "CUPTI_ACTIVITY_MEMORY_KIND_PAGEABLE",
+                        "CUPTI_ACTIVITY_MEMORY_KIND_PINNED",
+                        "CUPTI_ACTIVITY_MEMORY_KIND_DEVICE",
+                        "CUPTI_ACTIVITY_MEMORY_KIND_ARRAY",
+                        "CUPTI_ACTIVITY_MEMORY_KIND_MANAGED",
+                        "CUPTI_ACTIVITY_MEMORY_KIND_DEVICE_STATIC",
+                        "CUPTI_ACTIVITY_MEMORY_KIND_MANAGED_STATIC"};
+                if (kind > (sizeof(valid_kinds) / sizeof(const char*))) {
+                    return "invalid";
+                }
+                return valid_kinds[kind];
+            };
+            new_cupti_event("Memcpy", 'B', event.stream, event.start)
+                    .arg("bytes", imperative::to_string(event.bytes))
+                    .arg("src_kind", memkind2str(event.src_kind))
+                    .arg("dst_kind", memkind2str(event.dst_kind));
+            new_cupti_event(pid_str, 'f', event.stream, event.start)
+                    .id(event.correlation_id)
+                    .bp('e')
+                    .cat("CUPTILink")
+                    .scope(pid_str);
+            new_cupti_event("Memcpy", 'E', event.stream, event.end)
+                    .arg("bytes", imperative::to_string(event.bytes))
+                    .arg("src_kind", memkind2str(event.src_kind))
+                    .arg("dst_kind", memkind2str(event.dst_kind));
+        } else if constexpr (std::is_same_v<TEvent, CUPTIMemsetEvent>) {
+            new_cupti_event("Memset", 'B', event.stream, event.start)
+                    .arg("value", imperative::to_string(event.value))
+                    .arg("bytes", imperative::to_string(event.bytes));
+            new_cupti_event("Memset", 'E', event.stream, event.start)
+                    .arg("value", imperative::to_string(event.value))
+                    .arg("bytes", imperative::to_string(event.bytes));
+        } else if constexpr (std::is_same_v<TEvent, CUPTIRuntimeEvent>) {
+            new_host_event(event.name, 'B');
+        } else if constexpr (std::is_same_v<TEvent, CUPTIRuntimeFinishEvent>) {
+            new_host_event(event.name, 'E');
+        } else if constexpr (std::is_same_v<TEvent, CUPTIDriverEvent>) {
+            new_host_event(event.name, 'B');
+            new_host_event(pid_str, 's')
+                    .id(event.correlation_id)
+                    .cat("CUPTILink")
+                    .scope(pid_str);
+        } else if constexpr (std::is_same_v<TEvent, CUPTIDriverFinishEvent>) {
+            new_host_event(event.name, 'E');
         }
     }
 
@@ -403,7 +485,8 @@ struct ChromeTimelineEventVisitor : EventVisitor<ChromeTimelineEventVisitor> {
             if (thread_dict.count(host)) {
                 trace_events.new_event()
                         .name("thread_name")
-                        .pid('M')
+                        .ph('M')
+                        .pid(pid)
                         .tid(to_tid(host))
                         .arg("name", thread_dict.at(host));
             }
@@ -411,7 +494,8 @@ struct ChromeTimelineEventVisitor : EventVisitor<ChromeTimelineEventVisitor> {
         for (auto&& device : devices()) {
             trace_events.new_event()
                     .name("thread_name")
-                    .pid('M')
+                    .ph('M')
+                    .pid(pid)
                     .tid(to_tid(device))
                     .arg("name", device.to_string_logical());
         }
@@ -419,7 +503,7 @@ struct ChromeTimelineEventVisitor : EventVisitor<ChromeTimelineEventVisitor> {
 };
 
 void dump_chrome_timeline(std::string filename, Profiler::bundle_t result) {
-    ChromeTimelineEventVisitor visitor;
+    ChromeTimelineEventVisitor visitor{};
     visitor.process_events(result);
     visitor.name_threads(result.thread_dict);
     auto trace_events = std::move(visitor.trace_events);

imperative/src/impl/profiler/events.h

@@ -16,6 +16,7 @@
 
 #include "../interpreter/stack_manager.h"
 #include "../op_trait.h"
+#include "megbrain/imperative/cpp_cupti.h"
 
 namespace mgb::imperative::profiler {
 
@@ -181,6 +182,60 @@ DEF_DUR_EVENT(HostToDevice, {
     void* device_ptr;
 });
 
+// cupti events
+DEF_EVENT(CUPTITimestamp, { cupti::clock::time_point timestamp; });
+
+DEF_DUR_EVENT(CUPTIKernelLaunch, {
+    uint32_t correlation_id;
+    const char* name;
+});
+
+DEF_EVENT(CUPTIKernelExecute, {
+    uint32_t correlation_id;
+    const char* name;
+    cupti::stream_t stream;
+    cupti::time_point start;
+    cupti::time_point end;
+});
+
+DEF_DUR_EVENT(CUPTIMemcpyLaunch, { uint32_t correlation_id; });
+
+DEF_EVENT(CUPTIMemcpy, {
+    uint32_t correlation_id;
+    uint8_t src_kind;
+    uint8_t dst_kind;
+    uint64_t bytes;
+    cupti::stream_t stream;
+    cupti::time_point start;
+    cupti::time_point end;
+});
+
+DEF_EVENT(CUPTIMemset, {
+    uint32_t correlation_id;
+    uint32_t value;
+    uint64_t bytes;
+    cupti::stream_t stream;
+    cupti::time_point start;
+    cupti::time_point end;
+});
+
+DEF_EVENT(CUPTIUnknownDevice, {});
+
+DEF_DUR_EVENT(CUPTIRuntime, {
+    uint32_t correlation_id;
+    const char* name;
+});
+
+DEF_DUR_EVENT(CUPTIDriver, {
+    uint32_t correlation_id;
+    const char* name;
+});
+
+DEF_EVENT(CUPTIIdentifyStream, {
+    cupti::stream_t stream;
+    CompNode device;
+});
+
 #undef DEF_EVENT
 #undef DEF_DUR_EVENT
 

imperative/src/impl/profiler/states.h

@@ -180,10 +180,13 @@ private:
     HostTime m_start_time;
     CompNode::UnorderedMap<size_t> m_device_tid_table;
     std::unordered_map<std::thread::id, size_t> m_host_tid_table;
+    std::unordered_map<cupti::stream_t, size_t> m_cupti_tid_table;
     CompNode::UnorderedMap<std::map<profiler::HostTime, profiler::RealDuration>>
             m_device_timeline;
     std::unordered_map<std::thread::id, std::vector<Trace>> m_trace_stack;
     std::unordered_map<std::string, int64_t> m_counter_table;
+    std::optional<std::pair<profiler::HostTime, cupti::time_point>> m_cupti_timestamp =
+            {};
 
 protected:
     Profiler::Record* current;
@@ -191,6 +194,11 @@ protected:
     ProfileTensorState* current_tensor;
 
 protected:
+    size_t next_tid() {
+        return m_host_tid_table.size() + m_device_tid_table.size() +
+               m_cupti_tid_table.size();
+    }
+
     profiler::Duration since_start(profiler::HostTime time) {
         return time - m_start_time;
     }
@@ -229,6 +237,10 @@ protected:
 
     size_t to_tid(CompNode device) { return m_device_tid_table.at(device); }
 
+    size_t to_tid(cupti::stream_t cupti_stream) {
+        return m_cupti_tid_table.at(cupti_stream);
+    }
+
     SmallVector<std::thread::id> host_threads() {
         SmallVector<std::thread::id> host_threads;
         for (auto&& [host, _] : m_host_tid_table) {
@@ -254,6 +266,13 @@ protected:
         value += delta;
     }
 
+    profiler::HostTime time_from_cupti(cupti::time_point timestamp) {
+        mgb_assert(m_cupti_timestamp.has_value());
+        return m_cupti_timestamp->first +
+               std::chrono::duration_cast<profiler::HostTime::duration>(
+                       timestamp - m_cupti_timestamp->second);
+    }
+
 public:
     void process_events(Profiler::bundle_t& bundle) {
         m_start_time = bundle.start_at;
@@ -272,7 +291,11 @@ public:
                 TensorCommandEvent, TensorCommandFinishEvent, AutoEvictEvent,
                 AutoEvictFinishEvent, CustomEvent, CustomFinishEvent, RecordDeviceEvent,
                 ScopeEvent, ScopeFinishEvent, HostToDeviceEvent,
-                HostToDeviceFinishEvent>
+                HostToDeviceFinishEvent, CUPTITimestampEvent, CUPTIKernelLaunchEvent,
+                CUPTIKernelLaunchFinishEvent, CUPTIKernelExecuteEvent,
+                CUPTIMemcpyLaunchEvent, CUPTIMemcpyLaunchFinishEvent, CUPTIMemcpyEvent,
+                CUPTIRuntimeEvent, CUPTIRuntimeFinishEvent, CUPTIDriverEvent,
+                CUPTIDriverFinishEvent, CUPTIMemsetEvent>
                 converter;
 
         auto for_each_entry = [&](auto&& handler) {
@@ -289,7 +312,9 @@ public:
             std::shared_ptr<CompNode::Event> device;
         };
         CompNode::UnorderedMap<DeviceStartPair> device_start_table;
+        std::unordered_map<cupti::stream_t, CompNode> cupti_stream_table;
 
+        // record device time
         for_each_entry([&](auto&& event) {
             using T = std::decay_t<decltype(event)>;
             if constexpr (std::is_same_v<T, RecordDeviceEvent>) {
@@ -313,8 +338,7 @@ public:
         // register host threads
         for_each_entry([&](auto&& event) {
             if (!m_host_tid_table.count(current->tid)) {
-                m_host_tid_table[current->tid] = {
-                        m_device_tid_table.size() + m_host_tid_table.size()};
+                m_host_tid_table[current->tid] = next_tid();
             }
         });
 
@@ -340,14 +364,39 @@ public:
             } else if constexpr (std::is_same_v<T, TensorProduceEvent>) {
                 auto& tensor = m_tensors[event.tensor_id];
                 if (!m_device_tid_table.count(event.device)) {
-                    m_device_tid_table[event.device] = {
-                            m_device_tid_table.size() + m_host_tid_table.size()};
+                    m_device_tid_table[event.device] = next_tid();
                 }
                 tensor.device = event.device;
                 tensor.layout = event.layout;
             }
         });
 
+        for_each_entry([&](auto&& event) {
+            using T = std::decay_t<decltype(event)>;
+            if constexpr (std::is_same_v<T, CUPTIIdentifyStreamEvent>) {
+                if (!m_cupti_tid_table.count(event.stream)) {
+                    m_cupti_tid_table[event.stream] =
+                            m_device_tid_table.at(event.device);
+                }
+            }
+        });
+
+        // record cupti streams
+        for_each_entry([&](auto&& event) {
+            using T = std::decay_t<decltype(event)>;
+            if constexpr (
+                    std::is_same_v<T, CUPTIKernelExecuteEvent> ||
+                    std::is_same_v<T, CUPTIMemcpyEvent> ||
+                    std::is_same_v<T, CUPTIMemsetEvent>) {
+                if (!m_cupti_tid_table.count(event.stream)) {
+                    m_cupti_tid_table[event.stream] = next_tid();
+                }
+            } else if constexpr (std::is_same_v<T, CUPTITimestampEvent>) {
+                mgb_assert(!m_cupti_timestamp.has_value());
+                m_cupti_timestamp.emplace(current->time, event.timestamp);
+            }
+        });
+
         // replay execution
         using namespace std::placeholders;
         for_each_entry([&](auto&& event) {

imperative/src/impl/utils/platform.cpp

@@ -0,0 +1,25 @@
+#include "megbrain/imperative/utils/platform.h"
+
+#ifdef __GNUG__
+#include <cxxabi.h>
+#include <cstdlib>
+#include <memory>
+#endif
+
+using namespace mgb;
+using namespace imperative;
+
+/*
+ * demangle typeid, see
+ * http://stackoverflow.com/questions/281818/unmangling-the-result-of-stdtype-infoname
+ */
+std::string mgb::imperative::demangle(std::string mangled) {
+#ifdef __GNUG__
+    int status = -1;
+    std::unique_ptr<char, void (*)(void*)> res{
+            abi::__cxa_demangle(mangled.c_str(), nullptr, nullptr, &status), std::free};
+    return (status == 0) ? res.get() : mangled;
+#else
+    return mangled;
+#endif
+}

imperative/src/include/megbrain/imperative/cpp_cupti.h

@@ -0,0 +1,86 @@
+#pragma once
+
+#include <chrono>
+#include <ctime>
+
+#include "megbrain/common.h"
+#include "megbrain/imperative/utils/to_string.h"
+
+namespace mgb::imperative::cupti {
+
+struct clock {
+    typedef std::chrono::nanoseconds duration;
+    typedef duration::rep rep;
+    typedef duration::period period;
+    typedef std::chrono::time_point<clock> time_point;
+    static const bool is_steady = false;
+
+    static time_point now() /* noexcept */;
+};
+
+using time_point = clock::time_point;
+
+using duration = clock::duration;
+
+struct device_t {
+    uint32_t device_id;
+
+    bool operator==(const device_t& rhs) const { return device_id == rhs.device_id; }
+};
+
+struct context_t : device_t {
+    uint32_t context_id;
+
+    bool operator==(const context_t& rhs) const {
+        return device_t::operator==(rhs) && context_id == rhs.context_id;
+    }
+};
+
+struct stream_t : context_t {
+    uint32_t stream_id;
+
+    bool operator==(const stream_t& rhs) const {
+        return context_t::operator==(rhs) && stream_id == rhs.stream_id;
+    }
+};
+
+bool available();
+
+void enable();
+
+void disable();
+
+void flush();
+
+bool enabled();
+
+template <typename TActivity>
+struct activity {
+private:
+    TActivity* m_ptr;
+
+public:
+    activity(void* ptr) : m_ptr((TActivity*)ptr) {}
+
+    time_point start() const { return time_point(duration(m_ptr->start)); }
+
+    time_point end() const { return time_point(duration(m_ptr->end)); }
+
+    device_t device() const { return {m_ptr->deviceId}; }
+
+    context_t context() const { return {device(), m_ptr->contextId}; }
+
+    stream_t stream() const { return {context(), m_ptr->streamId}; }
+
+    TActivity* operator->() const { return m_ptr; }
+};
+
+}  // namespace mgb::imperative::cupti
+
+template <>
+class std::hash<mgb::imperative::cupti::stream_t> {
+public:
+    size_t operator()(const mgb::imperative::cupti::stream_t& value) const {
+        return value.stream_id;
+    }
+};

imperative/src/include/megbrain/imperative/profiler.h

@@ -194,16 +194,9 @@ public:
 
     static bool is_profiling() { return sm_profiling; }
 
-    static void start_profile() {
-        mgb_assert(!sm_profiling);
-        sm_start_at = Timer::record_host();
-        sm_profiling = true;
-    }
+    static void start_profile();
 
-    static void stop_profile() {
-        mgb_assert(sm_profiling);
-        sm_profiling = false;
-    }
+    static void stop_profile();
 
     static thread_dict_t get_thread_dict();
 

imperative/src/include/megbrain/imperative/utils/platform.h

@@ -0,0 +1,9 @@
+#pragma once
+
+#include <string>
+
+namespace mgb::imperative {
+
+std::string demangle(std::string mangled);
+
+}

imperative/test/CMakeLists.txt

@@ -37,6 +37,10 @@ if(MGE_WITH_CUDA)
   list(APPEND LINK_LIBS cudart)
 endif()
 
+if(MGE_WITH_CUPTI)
+  list(APPEND LINK_LIBS libcupti)
+endif()
+
 if(MGE_WITH_DISTRIBUTED)
   list(APPEND LINK_LIBS megray)
 endif()

scripts/whl/manylinux2014/build_wheel_common.sh

@@ -61,11 +61,11 @@ echo "Build with ${SDK_NAME}"
 
 if [ $SDK_NAME == "cu101" ];then
     CUDA_COPY_LIB_LIST="${CUDA_LIB_DIR}/libnvrtc.so.10.1"
-    EXTRA_CMAKE_FLAG=" -DMGE_WITH_CUDNN_SHARED=OFF -DMGE_WITH_CUBLAS_SHARED=OFF" 
-    BUILD_GCC8="ON"  
-    REQUIR_CUDA_VERSION="10010" 
-    REQUIR_CUDNN_VERSION="7.6.3" 
-    REQUIR_TENSORRT_VERSION="6.0.1.5" 
+    EXTRA_CMAKE_FLAG=" -DMGE_WITH_CUDNN_SHARED=OFF -DMGE_WITH_CUBLAS_SHARED=OFF"
+    BUILD_GCC8="ON"
+    REQUIR_CUDA_VERSION="10010"
+    REQUIR_CUDNN_VERSION="7.6.3"
+    REQUIR_TENSORRT_VERSION="6.0.1.5"
     REQUIR_CUBLAS_VERSION="10.2.1.243"
 
 elif [ $SDK_NAME == "cu102_JetsonNano" ];then
@@ -87,6 +87,12 @@ elif [ $SDK_NAME == "cu102_JetsonNano" ];then
         ${CUDNN_LIB_DIR}/libcudnn_ops_train.so.8:\
         ${CUDNN_LIB_DIR}/libcudnn.so.8"
 
+    if [ ${machine} == "aarch64" ];then
+        CUDA_COPY_LIB_LIST="\
+        ${CUDA_LIB_DIR}/libcupti.so.10.2:\
+        ${CUDA_COPY_LIB_LIST}"
+    fi
+
     EXTRA_CMAKE_FLAG="-DMGE_WITH_CUDNN_SHARED=ON -DMGE_WITH_CUBLAS_SHARED=ON -DMGE_CUDA_GENCODE=\"-gencode arch=compute_53,code=sm_53\" "
 
 elif [ $SDK_NAME == "cu111" ];then
@@ -118,6 +124,12 @@ elif [ $SDK_NAME == "cu111" ];then
         ${CUDNN_LIB_DIR}/libcudnn_ops_train.so.8:\
         ${CUDNN_LIB_DIR}/libcudnn.so.8"
 
+    if [ ${machine} == "aarch64" ];then
+        CUDA_COPY_LIB_LIST="\
+        ${CUDA_LIB_DIR}/libcupti.so.11.1:\
+        ${CUDA_COPY_LIB_LIST}"
+    fi
+
     if [ ${IN_CI} = "true" ] && [ ${machine} == "aarch64" ]; then
         EXTRA_CMAKE_FLAG=" -DMGE_WITH_CUDNN_SHARED=ON -DMGE_WITH_CUBLAS_SHARED=ON -DMGE_CUDA_GENCODE=\"-gencode arch=compute_75,code=sm_75\" "
     else
@@ -152,9 +164,9 @@ elif [ $SDK_NAME == "cu112" ];then
         -gencode arch=compute_86,code=sm_86 \
         -gencode arch=compute_86,code=compute_86\" "
 
-    REQUIR_CUDA_VERSION="11020" 
-    REQUIR_CUDNN_VERSION="8.0.4" 
-    REQUIR_TENSORRT_VERSION="7.2.2.3" 
+    REQUIR_CUDA_VERSION="11020"
+    REQUIR_CUDNN_VERSION="8.0.4"
+    REQUIR_TENSORRT_VERSION="7.2.2.3"
     REQUIR_CUBLAS_VERSION="11.3.1.68"
 
 elif [ $SDK_NAME == "cpu" ];then

src/megbrain_build_config.h.in

@@ -35,6 +35,7 @@
 #cmakedefine01 MGB_ENABLE_FBS_SERIALIZATION
 #cmakedefine01 MGB_IS_DEV
 #cmakedefine01 MGB_CUSTOM_OP
+#cmakedefine01 MGB_CUPTI
 // DNN related flags
 // Platform macro's
 #cmakedefine01 MEGDNN_WITH_CUDA