Remove unnecessary dependence for profiler (#15899)

* refile profiler
test=develop

* follow comment
test=develop

paddle/fluid/platform/CMakeLists.txt

@@ -87,11 +87,11 @@ nv_test(transform_test SRCS transform_test.cu DEPS memory place device_context)
 cc_library(timer SRCS timer.cc)
 cc_test(timer_test SRCS timer_test.cc DEPS timer)
 
-cc_library(device_tracer SRCS device_tracer.cc DEPS boost profiler_proto framework_proto device_context ${GPU_CTX_DEPS})
+cc_library(device_tracer SRCS device_tracer.cc DEPS boost profiler_proto framework_proto ${GPU_CTX_DEPS})
 if(WITH_GPU)
-    nv_library(profiler SRCS profiler.cc profiler.cu DEPS device_context device_tracer)
+    nv_library(profiler SRCS profiler.cc profiler.cu DEPS device_tracer gpu_info enforce)
 else()
-    cc_library(profiler SRCS profiler.cc DEPS device_context device_tracer)
+    cc_library(profiler SRCS profiler.cc DEPS device_tracer enforce)
 endif()
 cc_test(profiler_test SRCS profiler_test.cc DEPS profiler)
 

paddle/fluid/platform/device_tracer.h

@@ -17,6 +17,7 @@ limitations under the License. */
 #include <string>
 
 #include "paddle/fluid/platform/dynload/cupti.h"
+#include "paddle/fluid/platform/event.h"
 #include "paddle/fluid/platform/port.h"
 #include "paddle/fluid/platform/profiler.pb.h"
 
@@ -32,8 +33,6 @@ inline uint64_t PosixInNsec() {
   return 1000 * (static_cast<uint64_t>(tv.tv_sec) * 1000000 + tv.tv_usec);
 }
 
-class Event;
-
 // DeviceTracer performs the following tasks:
 // 1. Register cuda callbacks for various events: kernel, memcpy, etc.
 // 2. Collect cuda statistics: start/end ts, memory, etc.

paddle/fluid/platform/event.h

+/* Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
+
+licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+#include <string>
+
+namespace paddle {
+namespace platform {
+
+enum EventType { kMark, kPushRange, kPopRange };
+
+class Event {
+ public:
+  // The DeviceContext is used to get the cuda stream.
+  // If CPU profiling mode, can pass nullptr.
+  Event(EventType type, std::string name, uint32_t thread_id);
+
+  const EventType& type() const;
+  std::string name() const { return name_; }
+  uint32_t thread_id() const { return thread_id_; }
+
+#ifdef PADDLE_WITH_CUDA
+#ifndef PADDLE_WITH_CUPTI
+  cudaEvent_t event() const { return event_; }
+  int device() const { return device_; }
+#endif
+#endif
+
+  double CpuElapsedMs(const Event& e) const;
+  double CudaElapsedMs(const Event& e) const;
+
+ private:
+  EventType type_;
+  std::string name_;
+  uint32_t thread_id_;
+  int64_t cpu_ns_;
+#ifdef PADDLE_WITH_CUDA
+#ifdef PADDLE_WITH_CUPTI
+  int64_t gpu_ns_ = 0;
+
+ public:
+  void AddCudaElapsedTime(int64_t start_ns, int64_t end_ns) {
+    gpu_ns_ += end_ns - start_ns;
+  }
+
+ private:
+#else
+  cudaEvent_t event_ = nullptr;
+  int device_ = -1;
+#endif
+#endif
+};
+}  // namespace platform
+}  // namespace paddle

paddle/fluid/platform/profiler.cu

@@ -12,9 +12,8 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
 
-#include "paddle/fluid/platform/profiler.h"
-
 #include <cuda.h>
+#include "paddle/fluid/platform/profiler.h"
 
 namespace paddle {
 namespace platform {
@@ -22,26 +21,27 @@ namespace platform {
 __global__ void DummyKernel(int *a) { a[0] = 0; }
 
 static void ForEachDevice(std::function<void(int)> func) {
-  auto original_device = GetCurrentDeviceId();
-  int count = GetCUDADeviceCount();
+  auto original_device = platform::GetCurrentDeviceId();
+  int count = platform::GetCUDADeviceCount();
   for (int i = 0; i < count; i++) {
-    SetDeviceId(i);
+    platform::SetDeviceId(i);
     func(i);
   }
-  SetDeviceId(original_device);
+  platform::SetDeviceId(original_device);
 }
 
 void DummyKernelAndEvent() {
   for (int i = 0; i < 5; i++) {
     ForEachDevice([](int d) {
-      CUDADeviceContext *dev_ctx = new CUDADeviceContext(CUDAPlace(d));
+      platform::SetDeviceId(d);
+      cudaStream_t stream;
+      PADDLE_ENFORCE(cudaStreamCreate(&stream));
       Mark("_cuda_startup_");
       int *ptr;
       PADDLE_ENFORCE(cudaMalloc(&ptr, sizeof(int)));
-      DummyKernel<<<1, 1, 0, dev_ctx->stream()>>>(ptr);
-      dev_ctx->Wait();
+      DummyKernel<<<1, 1, 0, stream>>>(ptr);
+      PADDLE_ENFORCE(cudaStreamSynchronize(stream));
       PADDLE_ENFORCE(cudaFree(ptr));
-      delete dev_ctx;
     });
   }
 }

paddle/fluid/platform/profiler.h

@@ -17,54 +17,13 @@ limitations under the License. */
 #include <list>
 #include <string>
 #include <vector>
-#include "paddle/fluid/platform/device_context.h"
-
-namespace paddle {
-namespace platform {
-
-enum EventType { kMark, kPushRange, kPopRange };
-
-class Event {
- public:
-  // The DeviceContext is used to get the cuda stream.
-  // If CPU profiling mode, can pass nullptr.
-  Event(EventType type, std::string name, uint32_t thread_id);
-
-  const EventType& type() const;
-  std::string name() const { return name_; }
-  uint32_t thread_id() const { return thread_id_; }
-
+#include "paddle/fluid/platform/enforce.h"
+#include "paddle/fluid/platform/event.h"
 #ifdef PADDLE_WITH_CUDA
-#ifndef PADDLE_WITH_CUPTI
-  cudaEvent_t event() const { return event_; }
-  int device() const { return device_; }
-#endif
+#include "paddle/fluid/platform/gpu_info.h"
 #endif
-
-  double CpuElapsedMs(const Event& e) const;
-  double CudaElapsedMs(const Event& e) const;
-
- private:
-  EventType type_;
-  std::string name_;
-  uint32_t thread_id_;
-  int64_t cpu_ns_;
-#ifdef PADDLE_WITH_CUDA
-#ifdef PADDLE_WITH_CUPTI
-  int64_t gpu_ns_ = 0;
-
- public:
-  void AddCudaElapsedTime(int64_t start_ns, int64_t end_ns) {
-    gpu_ns_ += end_ns - start_ns;
-  }
-
- private:
-#else
-  cudaEvent_t event_ = nullptr;
-  int device_ = -1;
-#endif
-#endif
-};
+namespace paddle {
+namespace platform {
 
 enum ProfilerState {
   kDisabled,  // disabled state

paddle/fluid/platform/profiler_test.cc

@@ -33,7 +33,6 @@ TEST(Event, CpuElapsedTime) {
 }
 
 TEST(RecordEvent, RecordEvent) {
-  using paddle::platform::DeviceContext;
   using paddle::platform::Event;
   using paddle::platform::EventType;
   using paddle::platform::RecordEvent;