Refine profiler code.

paddle/framework/executor.cc

@@ -120,8 +120,7 @@ void Executor::Run(const ProgramDesc& pdesc, Scope* scope, int block_id,
     VLOG(3) << op->DebugStringEx(local_scope);
 
     platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
-    auto dev_ctx = const_cast<platform::DeviceContext*>(pool.Get(place_));
-    platform::RecordEvent record_event(op->Type(), dev_ctx);
+    platform::RecordEvent record_event(op->Type(), pool.Get(place_));
 
     op->Run(*local_scope, place_);
     if (FLAGS_do_memory_benchmark) {

paddle/platform/profiler.cc

@@ -47,16 +47,16 @@ inline uint64_t GetTimeInNsec() {
 }
 
 Event::Event(EventKind kind, std::string name, uint32_t thread_id,
-             DeviceContext* dev_ctx)
+             const DeviceContext* dev_ctx)
     : kind_(kind), name_(name), thread_id_(thread_id), has_cuda_(false) {
 #ifdef PADDLE_WITH_CUDA
+  has_cuda_ = dev_ctx ? platform::is_gpu_place(dev_ctx->GetPlace()) : false;
+  if (has_cuda_) {
     auto* cuda_dev_ctx = static_cast<const CUDADeviceContext*>(dev_ctx);
-  if (cuda_dev_ctx) {
     PADDLE_ENFORCE(cudaGetDevice(&device_));
     PADDLE_ENFORCE(cudaEventCreate(&event_));
     auto stream = cuda_dev_ctx->stream();
     PADDLE_ENFORCE(cudaEventRecord(event_, stream));
-    has_cuda_ = true;
   }
 #endif
   cpu_ns_ = GetTimeInNsec();
@@ -114,19 +114,20 @@ inline EventList& GetEventList() {
   return *g_event_list;
 }
 
-void Mark(const std::string& name, DeviceContext* dev_ctx) {
+void Mark(const std::string& name, const DeviceContext* dev_ctx) {
   GetEventList().Record(EventKind::kMark, name, g_thread_id, dev_ctx);
 }
 
-void PushEvent(const std::string& name, DeviceContext* dev_ctx) {
+void PushEvent(const std::string& name, const DeviceContext* dev_ctx) {
   GetEventList().Record(EventKind::kPushRange, name, g_thread_id, dev_ctx);
 }
 
-void PopEvent(const std::string& name, DeviceContext* dev_ctx) {
+void PopEvent(const std::string& name, const DeviceContext* dev_ctx) {
   GetEventList().Record(EventKind::kPopRange, name, g_thread_id, dev_ctx);
 }
 
-RecordEvent::RecordEvent(const std::string& name, DeviceContext* dev_ctx) {
+RecordEvent::RecordEvent(const std::string& name,
+                         const DeviceContext* dev_ctx) {
   if (g_state == ProfilerState::kDisabled) return;
   dev_ctx_ = dev_ctx;
   name_ = name;
@@ -155,6 +156,7 @@ void EnableProfiler(ProfilerState state) {
         DeviceContext* dev_ctx = new CUDADeviceContext(CUDAPlace(d));
         Mark("_cuda_startup_", dev_ctx);
         dev_ctx->Wait();
+        delete dev_ctx;
       });
     }
   }

paddle/platform/profiler.h

@@ -29,7 +29,7 @@ class Event {
   // The DeviceContext is used to get the cuda stream.
   // If CPU profiling mode, can pass nullptr.
   Event(EventKind kind, std::string name, uint32_t thread_id,
-        DeviceContext* dev_ctx);
+        const DeviceContext* dev_ctx);
 
   std::string kind() const;
   std::string name() const { return name_; }
@@ -95,19 +95,19 @@ enum ProfilerState {
   kCUDA,      // GPU profiling state
 };
 
-void Mark(const std::string& name, DeviceContext* dev_ctx);
+void Mark(const std::string& name, const DeviceContext* dev_ctx);
 
-void PushEvent(const std::string& name, DeviceContext* dev_ctx);
+void PushEvent(const std::string& name, const DeviceContext* dev_ctx);
 
-void PopEvent(const std::string& name, DeviceContext* dev_ctx);
+void PopEvent(const std::string& name, const DeviceContext* dev_ctx);
 
 struct RecordEvent {
-  explicit RecordEvent(const std::string& name, DeviceContext* dev_ctx);
+  explicit RecordEvent(const std::string& name, const DeviceContext* dev_ctx);
 
   ~RecordEvent();
 
   // The device context is used by Event to get the current cuda stream.
-  DeviceContext* dev_ctx_;
+  const DeviceContext* dev_ctx_;
   // Event name
   std::string name_;
 };

python/paddle/v2/fluid/profiler.py

@@ -81,10 +81,11 @@ def profiler(state, sorted_key=None):
     to add more records.
 
     Args:
-        state (string) : The profiling state, It should be 'CPU' or 'GPU'.
-            Although users may define CPUPlace or CUDAPlace when using Fluid,
-            the profiler doesn't get the state based on this Place. Since the
-            implementation is an independent part from the Fluid.
+        state (string) : The profiling state, which should be 'CPU' or 'GPU',
+            telling the profiler to use CPU timer or GPU timer for profiling.
+            Although users may have already specified the execution place
+            (CPUPlace/CUDAPlace) in the begining, for flexibility the profiler
+            would not inherit this place.
         sorted_key (string) : If None, the profiling results will be printed
             in the order of first end time of events. Otherwise, the profiling
             results will be sorted by the this flag. This flag should be one

python/paddle/v2/fluid/tests/test_profiler.py

@@ -41,8 +41,8 @@ class TestProfiler(unittest.TestCase):
                 exe.run(fluid.default_main_program(), feed={'data': input})
         os.remove(output_file)
 
-    def profiler(self, state):
-        if state == 'GPU' and core.is_compile_gpu():
+    def net_profiler(self, state):
+        if state == 'GPU' and not core.is_compile_gpu():
             return
         startup_program = fluid.Program()
         main_program = fluid.Program()
@@ -79,11 +79,11 @@ class TestProfiler(unittest.TestCase):
                 acc = np.array(outs[1])
                 pass_acc = accuracy.eval(exe)
 
-    def not_test_cpu_profiler(self):
-        self.profiler('CPU')
+    def test_cpu_profiler(self):
+        self.net_profiler('CPU')
 
-    def not_test_cuda_profiler(self):
-        self.profiler('GPU')
+    def test_cuda_profiler(self):
+        self.net_profiler('GPU')
 
 
 if __name__ == '__main__':