Merge pull request #14500 from NHZlX/refine_trt

Fix gpu load model and demo_ci on trt

Author: NHZlX

paddle/fluid/inference/analysis/CMakeLists.txt

@@ -7,16 +7,17 @@ set(analysis_deps # analysis_deps can be extended accross the project
 add_subdirectory(ir_passes)
 add_subdirectory(passes)
 
-cc_library(ir_pass_manager SRCS ir_pass_manager.cc DEPS graph pass ${INFER_IR_PASSES})
+cc_library(analysis_helper SRCS helper.cc DEPS framework_proto proto_desc graph paddle_fluid_api)
+
+cc_library(ir_pass_manager SRCS ir_pass_manager.cc DEPS graph pass ${INFER_IR_PASSES} analysis_helper)
 
 cc_library(argument SRCS argument.cc DEPS scope proto_desc)
 cc_library(analysis_pass SRCS analysis_pass.cc DEPS proto_desc)
 
 cc_library(analysis SRCS
   analyzer.cc
-  helper.cc
   analysis_pass
-  DEPS ${analysis_deps}
+  DEPS ${analysis_deps} analysis_helper
   )
 
 cc_test(test_dot SRCS dot_tester.cc DEPS analysis)

paddle/fluid/inference/analysis/analyzer_tester.cc

@@ -30,6 +30,7 @@ TEST(Analyzer, analysis_without_tensorrt) {
   Argument argument;
   argument.SetModelDir(FLAGS_inference_model_dir);
   argument.SetIrAnalysisPasses({"infer_clean_graph_pass"});
+  argument.SetUseGPU(false);
 
   Analyzer analyser;
   analyser.Run(&argument);
@@ -41,6 +42,7 @@ TEST(Analyzer, analysis_with_tensorrt) {
   argument.SetTensorRtWorkspaceSize(1 << 20);
   argument.SetModelDir(FLAGS_inference_model_dir);
   argument.SetIrAnalysisPasses({"infer_clean_graph_pass"});
+  argument.SetUseGPU(false);
 
   Analyzer analyser;
   analyser.Run(&argument);

paddle/fluid/inference/analysis/argument.h

@@ -116,6 +116,7 @@ struct Argument {
                       std::vector<std::string>);
 
   DECL_ARGUMENT_FIELD(use_gpu, UseGPU, bool);
+  DECL_ARGUMENT_FIELD(gpu_device_id, GPUDeviceId, int);
   DECL_ARGUMENT_FIELD(use_tensorrt, UseTensorRT, bool);
   DECL_ARGUMENT_FIELD(tensorrt_node_teller, TensorRtNodeTeller,
                       std::function<bool(const framework::ir::Node*)>);

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

@@ -4,4 +4,6 @@ set(analysis_deps ${analysis_deps}
         subgraph_detector tensorrt_subgraph_pass
         CACHE INTERNAL "")
 
+set(pass_file ${PADDLE_BINARY_DIR}/paddle/fluid/inference/api/paddle_inference_pass.h)
+file(APPEND ${pass_file} "USE_PASS(tensorrt_subgraph_pass);\n")
 set(INFER_IR_PASSES ${INFER_IR_PASSES} tensorrt_subgraph_pass CACHE INTERNAL "")

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

@@ -30,15 +30,28 @@ void IrGraphBuildPass::RunImpl(Argument *argument) {
   if (!argument->scope_valid()) {
     argument->SetScope(new framework::Scope);
   }
+  PADDLE_ENFORCE(argument->use_gpu_valid());
+
+  // The load program should run on the same device with the inference program,
+  // 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();
+  }
 
   if (argument->model_dir_valid()) {
-    auto program = LoadModel(argument->model_dir(), argument->scope_ptr());
+    auto program =
+        LoadModel(argument->model_dir(), argument->scope_ptr(), place);
     argument->SetMainProgram(program.release());
   } else if (argument->model_program_path_valid() &&
              argument->model_params_path_valid()) {
     auto program =
         LoadModel(argument->model_program_path(), argument->model_params_path(),
-                  argument->scope_ptr());
+                  argument->scope_ptr(), place);
     argument->SetMainProgram(program.release());
   } else {
     PADDLE_THROW(
@@ -52,16 +65,15 @@ void IrGraphBuildPass::RunImpl(Argument *argument) {
 }
 
 std::unique_ptr<framework::ProgramDesc> IrGraphBuildPass::LoadModel(
-    const std::string &path, framework::Scope *scope) {
-  platform::CPUPlace place;
+    const std::string &path, framework::Scope *scope,
+    const platform::Place &place) {
   framework::Executor exe(place);
   return Load(&exe, scope, path);
 }
 
 std::unique_ptr<framework::ProgramDesc> IrGraphBuildPass::LoadModel(
     const std::string &program_path, const std::string &params_path,
-    framework::Scope *scope) {
-  platform::CPUPlace place;
+    framework::Scope *scope, const platform::Place &place) {
   framework::Executor exe(place);
   return Load(&exe, scope, program_path, params_path);
 }

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

@@ -17,6 +17,7 @@
 #include <string>
 #include "paddle/fluid/framework/scope.h"
 #include "paddle/fluid/inference/analysis/analysis_pass.h"
+#include "paddle/fluid/platform/place.h"
 
 namespace paddle {
 namespace inference {
@@ -32,11 +33,12 @@ class IrGraphBuildPass : public AnalysisPass {
   std::string repr() const override;
 
  private:
-  std::unique_ptr<framework::ProgramDesc> LoadModel(const std::string &path,
-                                                    framework::Scope *scope);
+  std::unique_ptr<framework::ProgramDesc> LoadModel(
+      const std::string &path, framework::Scope *scope,
+      const platform::Place &place);
   std::unique_ptr<framework::ProgramDesc> LoadModel(
       const std::string &program_path, const std::string &params_path,
-      framework::Scope *scope);
+      framework::Scope *scope, const platform::Place &place);
 
   std::string model_binary_str_;
 };

paddle/fluid/inference/api/CMakeLists.txt

@@ -27,11 +27,10 @@ endif()
 cc_library(reset_tensor_array SRCS details/reset_tensor_array.cc DEPS lod_tensor scope)
 cc_library(analysis_config SRCS analysis_config.cc DEPS lod_tensor paddle_pass_builder)
 cc_library(paddle_pass_builder SRCS paddle_pass_builder.cc)
-cc_library(paddle_inference_api SRCS api.cc api_impl.cc helper.cc DEPS lod_tensor scope paddle_pass_builder reset_tensor_array analysis_config analysis_config paddle_pass_builder)
-cc_library(analysis_predictor SRCS analysis_predictor.cc DEPS paddle_inference_api analysis naive_executor zero_copy_tensor reset_tensor_array analysis_config paddle_pass_builder)
-cc_library(zero_copy_tensor SRCS details/zero_copy_tensor.cc DEPS paddle_inference_api)
-cc_library(zero_copy_tensor_dummy SRCS details/zero_copy_tensor_dummy.cc DEPS paddle_inference_api)
-
+cc_library(analysis_predictor SRCS analysis_predictor.cc DEPS paddle_inference_api analysis naive_executor zero_copy_tensor reset_tensor_array analysis_config paddle_pass_builder ir_pass_manager)
+cc_library(zero_copy_tensor SRCS details/zero_copy_tensor.cc DEPS scope lod_tensor enforce)
+cc_library(zero_copy_tensor_dummy SRCS details/zero_copy_tensor_dummy.cc)
+cc_library(paddle_inference_api SRCS api.cc api_impl.cc helper.cc DEPS lod_tensor scope paddle_pass_builder reset_tensor_array analysis_config analysis_config paddle_pass_builder DEPS zero_copy_tensor)
 
 cc_test(test_paddle_inference_api
         SRCS api_tester.cc

paddle/fluid/inference/api/analysis_predictor.cc

@@ -285,6 +285,7 @@ void AnalysisPredictor::OptimizeInferenceProgram() {
   status_program_optimized_ = true;
 
   argument_.SetUseGPU(config_.use_gpu);
+  argument_.SetGPUDeviceId(config_.device);
   // Analyze inference_program
   if (!config_.model_dir.empty()) {
     argument_.SetModelDir(config_.model_dir);
@@ -491,8 +492,7 @@ bool AnalysisPredictor::LoadParameters() {
   }
 
   // Use NaiveExecutor to Load parameters.
-  platform::CPUPlace place;
-  framework::NaiveExecutor e(place);
+  framework::NaiveExecutor e(place_);
   e.Prepare(scope_.get(), *load_program, 0, false);
   e.Run();
   VLOG(3) << "get " << scope_->LocalVarNames().size() << " vars after load";

paddle/fluid/inference/api/paddle_pass_builder.h

@@ -116,8 +116,12 @@ 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", "conv_bn_fuse_pass",
+        "infer_clean_graph_pass",
+        // "infer_clean_graph_pass", "conv_bn_fuse_pass",
     });
   }
 

paddle/fluid/inference/tests/api/CMakeLists.txt

@@ -1,5 +1,9 @@
 set(INFERENCE_EXTRA_DEPS paddle_inference_api paddle_fluid_api ir_pass_manager analysis_predictor)
 
+if(WITH_GPU AND TENSORRT_FOUND)
+    set(INFERENCE_EXTRA_DEPS ${INFERENCE_EXTRA_DEPS} analysis ${analysis_deps} ir_pass_manager analysis_predictor)
+endif()
+
 function(download_model install_dir model_name)
     if (NOT EXISTS ${install_dir})
         inference_download_and_uncompress(${install_dir} ${INFERENCE_URL} ${model_name})
@@ -75,11 +79,11 @@ endif()
 inference_analysis_api_test(test_analyzer_ocr ${OCR_INSTALL_DIR} analyzer_vis_tester.cc)
 
 # resnet50
-inference_analysis_api_test_with_fake_data(test_analyzer_resnet50 
+inference_analysis_api_test_with_fake_data(test_analyzer_resnet50
   "${INFERENCE_DEMO_INSTALL_DIR}/resnet50" analyzer_resnet50_tester.cc "resnet50_model.tar.gz")
 
 # mobilenet with depthwise_conv op
-inference_analysis_api_test_with_fake_data(test_analyzer_mobilenet 
+inference_analysis_api_test_with_fake_data(test_analyzer_mobilenet
   "${INFERENCE_DEMO_INSTALL_DIR}/mobilenet_depthwise_conv" analyzer_resnet50_tester.cc "mobilenet_model.tar.gz")
 
 # anakin
@@ -89,15 +93,15 @@ if (WITH_ANAKIN AND WITH_MKL) # only needed in CI
    set(ANAKIN_RNN1_INSTALL_DIR "${ANAKIN_INSTALL_DIR}/rnn1")
    inference_download(${ANAKIN_RNN1_INSTALL_DIR} ${INFERENCE_URL} "anakin_test%2Fditu_rnn.anakin2.model.bin")
    inference_download(${ANAKIN_RNN1_INSTALL_DIR} ${INFERENCE_URL} "anakin_test%2Fditu_rnn_data.txt")
-   cc_test(test_anakin_rnn1 SRCS anakin_rnn1_tester.cc 
-           ARGS --model=${ANAKIN_RNN1_INSTALL_DIR}/anakin_test%2Fditu_rnn.anakin2.model.bin 
+   cc_test(test_anakin_rnn1 SRCS anakin_rnn1_tester.cc
+           ARGS --model=${ANAKIN_RNN1_INSTALL_DIR}/anakin_test%2Fditu_rnn.anakin2.model.bin
                 --datapath=${ANAKIN_RNN1_INSTALL_DIR}/anakin_test%2Fditu_rnn_data.txt
            DEPS inference_anakin_api_shared SERIAL)
    # anakin mobilenet
    if(WITH_GPU)
        set(ANAKIN_MOBILENET_INSTALL_DIR "${ANAKIN_INSTALL_DIR}/mobilenet")
        inference_download(${ANAKIN_MOBILENET_INSTALL_DIR} ${INFERENCE_URL} "mobilenet_v2.anakin.bin")
-       cc_test(test_anakin_mobilenet SRCS anakin_mobilenet_tester.cc 
+       cc_test(test_anakin_mobilenet SRCS anakin_mobilenet_tester.cc
                ARGS --model=${ANAKIN_MOBILENET_INSTALL_DIR}/mobilenet_v2.anakin.bin
                DEPS inference_anakin_api_shared dynload_cuda SERIAL)
    endif()
@@ -109,6 +113,6 @@ if(WITH_GPU AND TENSORRT_FOUND)
        inference_download_and_uncompress(${TRT_MODEL_INSTALL_DIR} ${INFERENCE_URL}/tensorrt_test "trt_test_models.tar.gz")
    endif()
    inference_analysis_test(test_trt_models SRCS trt_models_tester.cc
-      EXTRA_DEPS ${INFERENCE_EXTRA_DEPS} analysis ${analysis_deps} ir_pass_manager analysis_predictor
+      EXTRA_DEPS ${INFERENCE_EXTRA_DEPS}
         ARGS --infer_model=${TRT_MODEL_INSTALL_DIR}/trt_test_models SERIAL)
 endif()