Remove the memory copy of feeding data in C++ inference API (#14577)

* Remove the memory copy for feeding data in C++ inference API
* Fix compling dependence
* Fix compling in ONLY_CPU mode

paddle/fluid/inference/api/CMakeLists.txt

@@ -30,7 +30,9 @@ cc_library(paddle_pass_builder SRCS paddle_pass_builder.cc)
 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_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 zero_copy_tensor reset_tensor_array)
 
 cc_test(test_paddle_inference_api
         SRCS api_tester.cc

paddle/fluid/inference/api/analysis_predictor.cc

@@ -31,6 +31,7 @@
 #include "paddle/fluid/inference/tensorrt/convert/op_converter.h"
 #endif
 #include "paddle/fluid/inference/utils/singleton.h"
+#include "paddle/fluid/memory/memcpy.h"
 #include "paddle/fluid/platform/cpu_helper.h"
 #include "paddle/fluid/platform/profiler.h"
 
@@ -214,17 +215,29 @@ bool AnalysisPredictor::SetFeed(const std::vector<PaddleTensor> &inputs,
     framework::DDim ddim = framework::make_ddim(inputs[i].shape);
     void *input_ptr;
     if (inputs[i].dtype == PaddleDType::INT64) {
-      input_ptr = input.mutable_data<int64_t>(ddim, platform::CPUPlace());
+      input_ptr = input.mutable_data<int64_t>(ddim, place_);
     } else if (inputs[i].dtype == PaddleDType::FLOAT32) {
-      input_ptr = input.mutable_data<float>(ddim, platform::CPUPlace());
+      input_ptr = input.mutable_data<float>(ddim, place_);
     } else {
       LOG(ERROR) << "unsupported feed type " << inputs[i].dtype;
       return false;
     }
 
+    if (platform::is_cpu_place(place_)) {
       // TODO(panyx0718): Init LoDTensor from existing memcpy to save a copy.
       std::memcpy(static_cast<void *>(input_ptr), inputs[i].data.data(),
                   inputs[i].data.length());
+    } else {
+#ifdef PADDLE_WITH_CUDA
+      auto dst_gpu_place = boost::get<platform::CUDAPlace>(place_);
+      memory::Copy(dst_gpu_place, static_cast<void *>(input_ptr),
+                   platform::CPUPlace(), inputs[i].data.data(),
+                   inputs[i].data.length(),
+                   0);  // stream 0 for sync copy
+#else
+      PADDLE_THROW("Not compile with CUDA, should not reach here.");
+#endif
+    }
     // TODO(Superjomn) Low performance, need optimization for heavy LoD copy.
     framework::LoD lod;
     for (auto &level : inputs[i].lod) {

paddle/fluid/inference/api/api_impl.cc

@@ -24,6 +24,7 @@ limitations under the License. */
 #include "paddle/fluid/inference/api/api_impl.h"
 #include "paddle/fluid/inference/api/details/reset_tensor_array.h"
 #include "paddle/fluid/inference/api/helper.h"
+#include "paddle/fluid/memory/memcpy.h"
 #include "paddle/fluid/platform/cpu_helper.h"
 #include "paddle/fluid/platform/profiler.h"
 
@@ -193,17 +194,30 @@ bool NativePaddlePredictor::SetFeed(const std::vector<PaddleTensor> &inputs,
     framework::DDim ddim = framework::make_ddim(inputs[i].shape);
     void *input_ptr;
     if (inputs[i].dtype == PaddleDType::INT64) {
-      input_ptr = input.mutable_data<int64_t>(ddim, platform::CPUPlace());
+      input_ptr = input.mutable_data<int64_t>(ddim, place_);
     } else if (inputs[i].dtype == PaddleDType::FLOAT32) {
-      input_ptr = input.mutable_data<float>(ddim, platform::CPUPlace());
+      input_ptr = input.mutable_data<float>(ddim, place_);
     } else {
       LOG(ERROR) << "unsupported feed type " << inputs[i].dtype;
       return false;
     }
 
+    if (platform::is_cpu_place(place_)) {
       // TODO(panyx0718): Init LoDTensor from existing memcpy to save a copy.
       std::memcpy(static_cast<void *>(input_ptr), inputs[i].data.data(),
                   inputs[i].data.length());
+    } else {
+#ifdef PADDLE_WITH_CUDA
+      auto dst_gpu_place = boost::get<platform::CUDAPlace>(place_);
+      memory::Copy(dst_gpu_place, static_cast<void *>(input_ptr),
+                   platform::CPUPlace(), inputs[i].data.data(),
+                   inputs[i].data.length(),
+                   0);  // stream 0 for sync copy
+#else
+      PADDLE_THROW("Not compile with CUDA, should not reach here.");
+#endif
+    }
+
     // TODO(Superjomn) Low performance, need optimization for heavy LoD copy.
     framework::LoD lod;
     for (auto &level : inputs[i].lod) {