Merge pull request #14662 from NHZlX/cherry_pick_sync_pass

cherry-pick from #14649

Author: NHZlX

paddle/fluid/inference/analysis/passes/CMakeLists.txt

@@ -1,6 +1,7 @@
 cc_library(ir_graph_build_pass SRCS ir_graph_build_pass.cc DEPS analysis_pass argument ir_pass_manager)
 cc_library(ir_analysis_pass SRCS ir_analysis_pass.cc DEPS analysis_pass argument ir_pass_manager)
-cc_library(analysis_passes SRCS passes.cc DEPS ir_graph_build_pass ir_analysis_pass)
+cc_library(ir_params_sync_among_devices_pass SRCS ir_params_sync_among_devices_pass.cc DEPS analysis_pass argument ir_pass_manager)
+cc_library(analysis_passes SRCS passes.cc DEPS ir_graph_build_pass ir_analysis_pass ir_params_sync_among_devices_pass)
 
 set(analysis_deps ${analysis_deps}
         ir_graph_build_pass

paddle/fluid/inference/analysis/passes/ir_analysis_compose_pass.cc

@@ -61,6 +61,7 @@ void IrAnalysisComposePass::InitTensorRTAttrs(Argument *argument) {
 void IrAnalysisComposePass::ApplyIrPasses(Argument *argument) {
   std::vector<std::string> passes({
       "ir_graph_build_pass", "ir_analysis_pass",
+      "ir_params_sync_among_devices_pass",
   });
   for (const auto &pass : passes) {
     VLOG(2) << "Run pass " << pass;

paddle/fluid/inference/analysis/passes/ir_graph_build_pass.cc

@@ -36,12 +36,7 @@ void IrGraphBuildPass::RunImpl(Argument *argument) {
   // so that the parameters will on the same device, or they will keep copying
   // between difference devices.
   platform::Place place;
-  if (argument->use_gpu()) {
-    PADDLE_ENFORCE(argument->gpu_device_id_valid());
-    place = platform::CUDAPlace(argument->gpu_device_id());
-  } else {
-    place = platform::CPUPlace();
-  }
+  place = platform::CPUPlace();
 
   if (argument->model_dir_valid()) {
     auto program =

paddle/fluid/inference/analysis/passes/ir_params_sync_among_devices_pass.cc

@@ -0,0 +1,74 @@
+// Copyright (c) 2018 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.
+
+#include "paddle/fluid/inference/analysis/passes/ir_params_sync_among_devices_pass.h"
+#include "paddle/fluid/framework/data_layout.h"
+#include "paddle/fluid/framework/lod_tensor.h"
+#include "paddle/fluid/framework/tensor_util.h"
+#include "paddle/fluid/platform/enforce.h"
+
+namespace paddle {
+namespace inference {
+namespace analysis {
+
+void IrParamsSyncAmongDevicesPass::RunImpl(Argument *argument) {
+  PADDLE_ENFORCE(argument->scope_valid());
+  PADDLE_ENFORCE(argument->use_gpu_valid());
+
+  platform::Place place;
+
+  // The parameters are on the cpu, therefore, synchronization is not necessary.
+  if (!argument->use_gpu()) return;
+
+  LOG(INFO) << "Sync params from CPU to GPU";
+
+  PADDLE_ENFORCE(argument->gpu_device_id_valid());
+  place = platform::CUDAPlace(argument->gpu_device_id());
+
+  auto *scope = argument->scope_ptr();
+  std::vector<std::string> all_vars = scope->LocalVarNames();
+
+  // We get all the vars from local_scope instead of the ProgramDesc.
+  // Because there exists the case that new parameter variables are not added to
+  // the program in the analysis pass.
+  for (auto &var_name : all_vars) {
+    auto *var = scope->FindLocalVar(var_name);
+    PADDLE_ENFORCE(var != nullptr);
+    if (var->IsType<framework::LoDTensor>() ||
+        var->IsType<framework::Tensor>()) {
+      auto *t = var->GetMutable<framework::LoDTensor>();
+
+      platform::CPUPlace cpu_place;
+      framework::LoDTensor temp_tensor;
+      temp_tensor.Resize(t->dims());
+      temp_tensor.mutable_data<float>(cpu_place);
+
+      // Copy the parameter data to a tmp tensor.
+      TensorCopySync(*t, cpu_place, &temp_tensor);
+      // Reallocation the space on GPU
+      t->mutable_data<float>(place);
+
+      // Copy parameter data to newly allocated GPU space.
+      TensorCopySync(temp_tensor, place, t);
+    }
+  }
+}
+
+std::string IrParamsSyncAmongDevicesPass::repr() const {
+  return "ir-params-sync-among-devices-pass";
+}
+
+}  // namespace analysis
+}  // namespace inference
+}  // namespace paddle

paddle/fluid/inference/analysis/passes/ir_params_sync_among_devices_pass.h

@@ -0,0 +1,39 @@
+// Copyright (c) 2018 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>
+#include <vector>
+
+#include "paddle/fluid/framework/scope.h"
+#include "paddle/fluid/inference/analysis/analysis_pass.h"
+#include "paddle/fluid/platform/place.h"
+
+namespace paddle {
+namespace inference {
+namespace analysis {
+
+/*
+ * Sync parameter from CPU to GPU.
+ */
+class IrParamsSyncAmongDevicesPass : public AnalysisPass {
+ public:
+  void RunImpl(Argument *argument) override;
+  std::string repr() const override;
+};
+
+}  // namespace analysis
+}  // namespace inference
+}  // namespace paddle

paddle/fluid/inference/analysis/passes/passes.cc

@@ -16,6 +16,7 @@
 #include "paddle/fluid/inference/analysis/passes/ir_analysis_compose_pass.cc"
 #include "paddle/fluid/inference/analysis/passes/ir_analysis_pass.h"
 #include "paddle/fluid/inference/analysis/passes/ir_graph_build_pass.h"
+#include "paddle/fluid/inference/analysis/passes/ir_params_sync_among_devices_pass.h"
 
 namespace paddle {
 namespace inference {
@@ -27,6 +28,9 @@ PassRegistry::PassRegistry() {
                   std::unique_ptr<AnalysisPass>(new IrGraphBuildPass));
   passes_.emplace("ir_analysis_compose_pass",
                   std::unique_ptr<AnalysisPass>(new IrAnalysisComposePass));
+  passes_.emplace(
+      "ir_params_sync_among_devices_pass",
+      std::unique_ptr<AnalysisPass>(new IrParamsSyncAmongDevicesPass));
 }
 
 }  // namespace analysis

paddle/fluid/inference/api/paddle_pass_builder.h

@@ -116,12 +116,8 @@ class CpuPassStrategy : public PassStrategy {
 class GpuPassStrategy : public PassStrategy {
  public:
   GpuPassStrategy() : PassStrategy({}) {
-    // TODO(NHZlX) Problem with Data synchronization between GPU and CPU
-    // When running in GPU mode, the parameters are all on GPU. But the
-    // opearations of "conv_bn_fuse_pass" are on CPU.
     passes_.assign({
-        "infer_clean_graph_pass",
-        // "infer_clean_graph_pass", "conv_bn_fuse_pass",
+        "infer_clean_graph_pass", "conv_bn_fuse_pass",
     });
   }
 

paddle/fluid/inference/utils/benchmark.cc

@@ -33,7 +33,7 @@ std::string Benchmark::SerializeToString() const {
   ss << batch_size_ << "\t";
   ss << num_threads_ << "\t";
   ss << latency_ << "\t";
-  ss << 1000 / latency_;
+  ss << 1000.0 / latency_;
   ss << '\n';
   return ss.str();
 }

paddle/fluid/inference/utils/benchmark.h

@@ -11,9 +11,11 @@
 // 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 <fstream>
 #include <iostream>
+#include <string>
 
 namespace paddle {
 namespace inference {
@@ -31,8 +33,8 @@ struct Benchmark {
   bool use_gpu() const { return use_gpu_; }
   void SetUseGpu() { use_gpu_ = true; }
 
-  int latency() const { return latency_; }
-  void SetLatency(int x) { latency_ = x; }
+  float latency() const { return latency_; }
+  void SetLatency(float x) { latency_ = x; }
 
   const std::string& name() const { return name_; }
   void SetName(const std::string& name) { name_ = name; }
@@ -43,7 +45,7 @@ struct Benchmark {
  private:
   bool use_gpu_{false};
   int batch_size_{0};
-  int latency_;
+  float latency_;
   int num_threads_{1};
   std::string name_;
 };