perf(mge/profiler): reduce profiler overhead

GitOrigin-RevId: 4e1fff9dc43341e1fd2cc414823cf31efd8e5831

imperative/src/impl/interpreter/interpreter_impl.cpp

@@ -525,19 +525,23 @@ void ChannelImpl::process_one_task(IdentifiedCommand& icmd) {
                 // Before wait
                 //TODO: split operator wait and execute so that OpWait could be corrected recorded.
                 // Before execute
-                m_worker_state.profiler->record_host<HostOpExecuteEvent>(event_data);
-                for (auto&& device: devices) {
-                    sync_device_scope(device);
-                    m_worker_state.profiler->record_device<DeviceOpExecuteEvent>(device, event_data);
+                if (m_worker_state.profiler->is_profiling()) {
+                    m_worker_state.profiler->record_host<HostOpExecuteEvent>(event_data);
+                    for (auto&& device: devices) {
+                        sync_device_scope(device);
+                        m_worker_state.profiler->record_device<DeviceOpExecuteEvent>(device, event_data);
+                    }
                 }
                 // Apply op
                 // Here std::move is REQUIRED for removing duplicated references.
                 auto tensor_outputs = OpDef::apply_on_physical_tensor(
                     *cmd.op, std::move(tensor_inputs));
                 // After execute
-                m_worker_state.profiler->record_host<HostOpFinishEvent>(event_data);
-                for (auto&& device: devices) {
-                    m_worker_state.profiler->record_device<DeviceOpFinishEvent>(device, event_data);
+                if (m_worker_state.profiler->is_profiling()) {
+                    m_worker_state.profiler->record_host<HostOpFinishEvent>(event_data);
+                    for (auto&& device: devices) {
+                        m_worker_state.profiler->record_device<DeviceOpFinishEvent>(device, event_data);
+                    }
                 }
                 // End profiling operator
                 mgb_assert(tensor_outputs.size() == cmd.outputs.size());

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

@@ -140,27 +140,32 @@ public:
 public:
     template <typename TEvent, typename... TArgs>
     void record_host(TArgs&&... args) {
-        auto instant = HostInstant{std::this_thread::get_id(), m_host_timer.get_msecs()};
         MGB_LOCK_GUARD(m_lock);
         if (!m_event_mask.test(index_of<TEvent>())) {
             return;
         }
         mgb_assert(m_status != Stopped, "record after stop");
+        auto instant = HostInstant{std::this_thread::get_id(), m_host_timer.get_msecs()};
         m_record_list.emplace_back(EventRecord<TEvent>{std::move(instant), {std::forward<TArgs>(args)...}});
     }
     template <typename TEvent, typename... TArgs>
     void record_device(Device device, TArgs&&... args) {
-        auto before = m_host_timer.get_msecs();
-        auto event = m_device_timer.get_device_time(device);
-        auto after = m_host_timer.get_msecs();
-        auto instant = DeviceInstant{before, event, after};
         MGB_LOCK_GUARD(m_lock);
         if (!m_event_mask.test(index_of<TEvent>())) {
             return;
         }
         mgb_assert(m_status != Stopped, "record after stop");
+        auto before = m_host_timer.get_msecs();
+        auto event = m_device_timer.get_device_time(device);
+        auto after = m_host_timer.get_msecs();
+        auto instant = DeviceInstant{before, event, after};
         m_record_list.emplace_back(EventRecord<TEvent>{std::move(instant), {std::forward<TArgs>(args)...}});
     }
+    // unsafe
+    bool is_profiling() {
+        MGB_LOCK_GUARD(m_lock);
+        return m_status == Profiling;
+    }
     void start(Mask mask) {
         MGB_LOCK_GUARD(m_lock);
         mgb_assert(m_status == NotStarted, "profiler already started");