Profiler refine and add CUDA runtime api tracer

paddle/fluid/framework/details/all_reduce_op_handle.cc

@@ -53,7 +53,7 @@ AllReduceOpHandle::AllReduceOpHandle(ir::Node *node,
 #endif
 
 void AllReduceOpHandle::RunImpl() {
-  platform::RecordEvent record_event(Name(), dev_ctxes_.cbegin()->second);
+  platform::RecordEvent record_event(Name());
 
   WaitInputVarGenerated();
   auto in_var_handles = DynamicCast<VarHandle>(this->Inputs());

paddle/fluid/framework/details/broadcast_op_handle.cc

@@ -22,7 +22,7 @@ namespace framework {
 namespace details {
 
 void BroadcastOpHandle::RunImpl() {
-  platform::RecordEvent record_event(Name(), dev_ctxes_.begin()->second);
+  platform::RecordEvent record_event(Name());
 
   if (places_.size() == 1) return;
 

paddle/fluid/framework/details/fused_broadcast_op_handle.cc

@@ -22,7 +22,7 @@ namespace framework {
 namespace details {
 
 void FusedBroadcastOpHandle::RunImpl() {
-  platform::RecordEvent record_event(Name(), dev_ctxes_.begin()->second);
+  platform::RecordEvent record_event(Name());
 
   if (places_.size() == 1UL) return;
 

paddle/fluid/framework/details/reduce_op_handle.cc

@@ -139,7 +139,7 @@ void ReduceOpHandle::GatherSelectedRows(
 #endif
 
 void ReduceOpHandle::RunImpl() {
-  platform::RecordEvent record_event(Name(), dev_ctxes_.cbegin()->second);
+  platform::RecordEvent record_event(Name());
 
   if (places_.size() == 1) return;
   // the input and output may have dummy var.

paddle/fluid/framework/details/scope_buffered_ssa_graph_executor.cc

@@ -63,7 +63,7 @@ FeedFetchList ScopeBufferedSSAGraphExecutor::Run(
     eptr = std::current_exception();
   }
 
-  platform::RecordEvent e("ScopeBufferedSSAGraphExecutorAfterRun", nullptr);
+  platform::RecordEvent e("ScopeBufferedSSAGraphExecutorAfterRun");
   ++drop_scope_counter_;
 
   bool stream_end = false;

paddle/fluid/framework/details/threaded_ssa_graph_executor.cc

@@ -37,7 +37,7 @@ ThreadedSSAGraphExecutor::ThreadedSSAGraphExecutor(
 FeedFetchList ThreadedSSAGraphExecutor::Run(
     const std::vector<std::string> &fetch_tensors) {
   std::unique_ptr<platform::RecordEvent> event(
-      new platform::RecordEvent("ThreadedSSAGraphExecutorPrepare", nullptr));
+      new platform::RecordEvent("ThreadedSSAGraphExecutorPrepare"));
   std::unordered_map<OpHandleBase *, size_t> pending_ops;
   std::unordered_set<VarHandleBase *> pending_vars;
   auto ready_vars = std::make_shared<BlockingQueue<VarHandleBase *>>();

paddle/fluid/framework/operator.cc

@@ -177,9 +177,7 @@ void OperatorBase::Run(const Scope& scope, const platform::Place& place) {
     // in concurrency scenerio. Here use an `if` to fix this issue.
     // Please not remove the `if`, ask @Superjomn if there are any concern.
     if (platform::IsProfileEnabled()) {
-      platform::DeviceContextPool& pool =
-          platform::DeviceContextPool::Instance();
-      platform::RecordEvent record_event(Type(), pool.Get(place));
+      platform::RecordEvent record_event(Type());
       RunImpl(scope, place);
     } else {
       RunImpl(scope, place);

paddle/fluid/inference/tests/test_helper.h

@@ -171,9 +171,7 @@ void TestInference(const std::string& dirname,
   // Enable the profiler
   paddle::platform::EnableProfiler(state);
   {
-    paddle::platform::RecordEvent record_event(
-        "init_program",
-        paddle::platform::DeviceContextPool::Instance().Get(place));
+    paddle::platform::RecordEvent record_event("init_program");
     inference_program = InitProgram(&executor, scope, dirname, is_combined);
   }
 
@@ -230,9 +228,7 @@ void TestInference(const std::string& dirname,
 
     // Run repeat times to profile the performance
     for (int i = 0; i < repeat; ++i) {
-      paddle::platform::RecordEvent record_event(
-          "run_inference",
-          paddle::platform::DeviceContextPool::Instance().Get(place));
+      paddle::platform::RecordEvent record_event("run_inference");
 
       if (PrepareContext) {
         // Note: if you change the inference_program, you need to call

paddle/fluid/operators/distributed/brpc/brpc_client.cc

@@ -80,7 +80,7 @@ VarHandlePtr BRPCClient::AsyncSendVar(const std::string& ep,
     google::protobuf::Closure* done = brpc::NewCallback(
         &HandleSendResponse, cntl, response, var_h, ch_ptr, ch_ctx, this);
 
-    platform::RecordRPCEvent record_event(method, p_ctx);
+    platform::RecordRPCEvent record_event(method);
 
     ch_ctx->stub->SendVariable(cntl, &request, response, done);
 
@@ -184,7 +184,7 @@ VarHandlePtr BRPCClient::_AsyncGetVar(const std::string& ep,
     google::protobuf::Closure* done = brpc::NewCallback(
         &HandleGetResponse, cntl, response, var_h, ch_ptr, ch_ctx, this);
 
-    platform::RecordRPCEvent record_event(method, p_ctx);
+    platform::RecordRPCEvent record_event(method);
 
     if (method_name == kGetMonomerRPC) {
       ch_ctx->stub->GetMonomerVariable(cntl, &req, response, done);
@@ -272,7 +272,7 @@ VarHandlePtr BRPCClient::AsyncPrefetchVar(const std::string& ep,
                                   &cntl->request_attachment(), out_var_name_val,
                                   false, 0, table_name_val);
 
-    platform::RecordRPCEvent record_event(method, p_ctx);
+    platform::RecordRPCEvent record_event(method);
 
     google::protobuf::Closure* done = brpc::NewCallback(
         &HandleGetResponse, cntl, response, var_h, ch_ptr, ch_ctx, this);
@@ -311,7 +311,7 @@ VarHandlePtr BRPCClient::AsyncSendFetchBarrier(const std::string& ep,
   VarHandlePtr var_h(
       new VarHandle(ep, method, FETCH_BARRIER_MESSAGE, nullptr, nullptr));
 
-  platform::RecordRPCEvent record_event(method, nullptr);
+  platform::RecordRPCEvent record_event(method);
 
   google::protobuf::Closure* done = brpc::NewCallback(
       &HandleFetchBarrierResponse, cntl, response, var_h, ch_ptr, ch_ctx, this);
@@ -406,7 +406,7 @@ VarHandlePtr BRPCClient::AsyncSendVarMessage(
   sendrecv::VoidMessage* response = new sendrecv::VoidMessage();
   cntl->set_timeout_ms(time_out);
 
-  platform::RecordRPCEvent record_event(method_name, nullptr);
+  platform::RecordRPCEvent record_event(method_name);
 
   VarHandlePtr var_h(
       new VarHandle(ep, method_name, req.varname(), nullptr, nullptr));

paddle/fluid/operators/distributed/grpc/grpc_client.cc

@@ -89,7 +89,7 @@ VarHandlePtr GRPCClient::AsyncSendVar(const std::string& ep,
     // stub context
     s->response_call_back_ = nullptr;
 
-    platform::RecordRPCEvent record_event(method, p_ctx);
+    platform::RecordRPCEvent record_event(method);
 
     auto call = s->stub_g_.PrepareUnaryCall(
         s->context_.get(), "/sendrecv.SendRecvService/SendVariable", req, &cq_);
@@ -184,7 +184,7 @@ VarHandlePtr GRPCClient::_AsyncGetVar(
         // stub context
         s->response_call_back_ = ProcGetResponse;
 
-        platform::RecordRPCEvent record_event(method, p_ctx);
+        platform::RecordRPCEvent record_event(method);
 
         auto call =
             s->stub_g_.PrepareUnaryCall(s->context_.get(), rpc_path, buf, &cq_);
@@ -235,7 +235,7 @@ VarHandlePtr GRPCClient::AsyncPrefetchVar(const std::string& ep,
     // stub context
     s->response_call_back_ = ProcGetResponse;
 
-    platform::RecordRPCEvent record_event(method, p_ctx);
+    platform::RecordRPCEvent record_event(method);
 
     auto call = s->stub_g_.PrepareUnaryCall(
         s->context_.get(), "/sendrecv.SendRecvService/PrefetchVariable", req,
@@ -265,7 +265,7 @@ VarHandlePtr GRPCClient::AsyncSendBatchBarrier(const std::string& ep,
   sendrecv::VariableMessage req;
   req.set_varname(BATCH_BARRIER_MESSAGE);
 
-  platform::RecordRPCEvent record_event(method, nullptr);
+  platform::RecordRPCEvent record_event(method);
 
   auto rpc = s->stub_->AsyncSendVariable(s->context_.get(), req, &cq_);
   rpc->Finish(&s->reply_, &s->status_, reinterpret_cast<void*>(s));
@@ -290,7 +290,7 @@ VarHandlePtr GRPCClient::AsyncSendFetchBarrier(const std::string& ep,
   sendrecv::VariableMessage req;
   req.set_varname(FETCH_BARRIER_MESSAGE);
 
-  platform::RecordRPCEvent record_event(method, nullptr);
+  platform::RecordRPCEvent record_event(method);
 
   auto rpc = s->stub_->AsyncGetVariable(s->context_.get(), req, &cq_);
   rpc->Finish(&s->reply_, &s->status_, reinterpret_cast<void*>(s));
@@ -317,7 +317,7 @@ VarHandlePtr GRPCClient::AsyncGetMonomerBarrier(const std::string& ep,
   sendrecv::VariableMessage req;
   req.set_varname(var_name);
 
-  platform::RecordRPCEvent record_event(method, nullptr);
+  platform::RecordRPCEvent record_event(method);
 
   auto rpc = s->stub_->AsyncGetMonomerBarrier(s->context_.get(), req, &cq_);
   rpc->Finish(&s->reply_, &s->status_, reinterpret_cast<void*>(s));
@@ -342,7 +342,7 @@ VarHandlePtr GRPCClient::AsyncSendComplete(const std::string& ep,
   sendrecv::VariableMessage req;
   req.set_varname(COMPLETE_MESSAGE);
 
-  platform::RecordRPCEvent record_event(method, nullptr);
+  platform::RecordRPCEvent record_event(method);
 
   auto rpc = s->stub_->AsyncSendVariable(s->context_.get(), req, &cq_);
   rpc->Finish(&s->reply_, &s->status_, reinterpret_cast<void*>(s));
@@ -372,7 +372,7 @@ VarHandlePtr GRPCClient::AsyncCheckpointNotify(const std::string& ep,
   req.set_varname(CHECKPOINT_SAVE_MESSAGE);
   req.set_out_varname(dir);
 
-  platform::RecordRPCEvent record_event(method, nullptr);
+  platform::RecordRPCEvent record_event(method);
 
   auto rpc = s->stub_->AsyncCheckpointNotify(s->context_.get(), req, &cq_);
   rpc->Finish(&s->reply_, &s->status_, reinterpret_cast<void*>(s));

paddle/fluid/operators/distributed/grpc/grpc_serde.cc

@@ -38,7 +38,7 @@ void SerializeToByteBuffer(const std::string& name, framework::Variable* var,
                            ::grpc::ByteBuffer* msg, const std::string& out_name,
                            const int trainer_id,
                            const std::string& table_name) {
-  platform::RecordRPCEvent record_event("serial", &ctx);
+  platform::RecordRPCEvent record_event("serial");
   VarMsg request;
   TensorPayload* payload = nullptr;
 
@@ -147,7 +147,7 @@ void DeserializeFromByteBuffer(const ::grpc::ByteBuffer& msg,
                                const platform::DeviceContext& ctx,
                                const framework::Scope* scope,
                                framework::Variable** var, int* trainer_id) {
-  platform::RecordRPCEvent record_event("deserial", &ctx);
+  platform::RecordRPCEvent record_event("deserial");
   operators::distributed::GRPCVariableResponse resp(scope, &ctx);
   PADDLE_ENFORCE(resp.Parse(msg) == 0, "parse bytebuffer to tensor error!");
   *var = resp.GetVar();

paddle/fluid/operators/reader/read_op.cc

@@ -85,9 +85,7 @@ class ReadOp : public framework::OperatorBase {
     std::vector<framework::LoDTensor> ins;
 
     // For profiling
-    platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
-    auto& ctx = *pool.Get(dev_place);
-    platform::RecordEvent record_event(Type(), &ctx);
+    platform::RecordEvent record_event(Type());
 
     reader->ReadNext(&ins);
     if (ins.empty()) {

paddle/fluid/platform/CMakeLists.txt

@@ -88,7 +88,11 @@ 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 ${GPU_CTX_DEPS})
-cc_library(profiler SRCS profiler.cc DEPS device_context device_tracer)
+if(WITH_GPU)
+    nv_library(profiler SRCS profiler.cc profiler.cu DEPS device_context device_tracer)
+else()
+    cc_library(profiler SRCS profiler.cc DEPS device_context device_tracer)
+endif()
 cc_test(profiler_test SRCS profiler_test.cc DEPS profiler)
 
 nv_test(float16_gpu_test SRCS float16_test.cu DEPS lod_tensor)