From 4f5f0be769228a61bb91baf63faa6ca3a71382d0 Mon Sep 17 00:00:00 2001
From: Luo Tao <luotao02@baidu.com>
Date: Mon, 14 May 2018 21:19:34 +0800
Subject: [PATCH] use the latest buffer to update the convert

---
 paddle/fluid/inference/engine.h                           | 4 ++--
 paddle/fluid/inference/tensorrt/CMakeLists.txt            | 1 -
 paddle/fluid/inference/tensorrt/convert/CMakeLists.txt    | 2 +-
 paddle/fluid/inference/tensorrt/convert/io_converter.cc   | 3 ++-
 .../inference/tensorrt/convert/test_activation_op.cc      | 8 ++++----
 5 files changed, 9 insertions(+), 9 deletions(-)

diff --git a/paddle/fluid/inference/engine.h b/paddle/fluid/inference/engine.h
index c0e5f3b201b..ce2b8161715 100644
--- a/paddle/fluid/inference/engine.h
+++ b/paddle/fluid/inference/engine.h
@@ -59,8 +59,8 @@ class EngineBase {
 
 struct Buffer {
   void* buffer{nullptr};               // buffer should be allocated only once.
-  int max_size;                        // buffer allocated space.
-  int size;                            // data size.
+  size_t max_size;                     // buffer allocated space.
+  size_t size;                         // data size.
   DeviceType device{DeviceType::UNK};  // tells which device this buffer is on.
 };
 
diff --git a/paddle/fluid/inference/tensorrt/CMakeLists.txt b/paddle/fluid/inference/tensorrt/CMakeLists.txt
index 677b3e04af8..b52d083f280 100644
--- a/paddle/fluid/inference/tensorrt/CMakeLists.txt
+++ b/paddle/fluid/inference/tensorrt/CMakeLists.txt
@@ -1,5 +1,4 @@
 nv_library(tensorrt_engine SRCS engine.cc DEPS framework_proto)
 nv_test(test_tensorrt SRCS test_tensorrt.cc DEPS dynload_cuda device_context dynamic_loader)
 nv_test(test_tensorrt_engine SRCS test_engine.cc DEPS dynload_cuda tensorrt_engine)
-
 add_subdirectory(convert)
diff --git a/paddle/fluid/inference/tensorrt/convert/CMakeLists.txt b/paddle/fluid/inference/tensorrt/convert/CMakeLists.txt
index 5178c54c084..4fb4511d991 100644
--- a/paddle/fluid/inference/tensorrt/convert/CMakeLists.txt
+++ b/paddle/fluid/inference/tensorrt/convert/CMakeLists.txt
@@ -1,4 +1,4 @@
 nv_test(test_op_converter SRCS test_op_converter.cc mul_op.cc conv2d_op.cc DEPS ${FLUID_CORE_MODULES})
-nv_test(test_trt_activation_op SRCS test_activation_op.cc activation_op.cc
+nv_test(test_trt_activation_op SRCS test_activation_op.cc activation_op.cc io_converter.cc
   DEPS ${FLUID_CORE_MODULES} activation_op tensorrt_engine)
 nv_test(test_io_converter SRCS test_io_converter.cc io_converter.cc DEPS dynload_cuda dynamic_loader lod_tensor)
diff --git a/paddle/fluid/inference/tensorrt/convert/io_converter.cc b/paddle/fluid/inference/tensorrt/convert/io_converter.cc
index 13bc2b37595..2d583c00123 100644
--- a/paddle/fluid/inference/tensorrt/convert/io_converter.cc
+++ b/paddle/fluid/inference/tensorrt/convert/io_converter.cc
@@ -58,7 +58,7 @@ class DefaultIOConverter : public EngineIOConverter {
                                            cudaMemcpyDeviceToHost, *stream_));
     } else if (is_gpu_place(place)) {
       PADDLE_ENFORCE_EQ(0, cudaMemcpyAsync(out->data<float>(), in, size,
-                                           cudaMemcpyHostToHost, *stream_));
+                                           cudaMemcpyDeviceToDevice, *stream_));
     } else {
       PADDLE_THROW("Unknown device for converter");
     }
@@ -66,6 +66,7 @@ class DefaultIOConverter : public EngineIOConverter {
   }
 };
 
+// fluid LodTensor <-> tensorrt ITensor
 REGISTER_TENSORRT_IO_CONVERTER(default, DefaultIOConverter);
 
 }  // namespace tensorrt
diff --git a/paddle/fluid/inference/tensorrt/convert/test_activation_op.cc b/paddle/fluid/inference/tensorrt/convert/test_activation_op.cc
index c43f7202127..3275b65bc6d 100644
--- a/paddle/fluid/inference/tensorrt/convert/test_activation_op.cc
+++ b/paddle/fluid/inference/tensorrt/convert/test_activation_op.cc
@@ -74,13 +74,13 @@ void Compare(const std::string op_type, float input, float expect) {
 
   // convert LoDTensor to ITensor
   size_t size = x_tensor->memory_size();
-  EngineIOConverter::ConvertInput(op_type, *x_tensor, engine->buffer("X"), size,
-                                  &stream);
+  EngineIOConverter::ConvertInput(op_type, *x_tensor,
+                                  engine->buffer("X").buffer, size, &stream);
   // run tensorrt Outp
   engine->Execute(1);
   // convert ITensor to LoDTensor
-  EngineIOConverter::ConvertOutput(op_type, engine->buffer("Out"), out_tensor,
-                                   size, &stream);
+  EngineIOConverter::ConvertOutput(op_type, engine->buffer("Out").buffer,
+                                   out_tensor, size, &stream);
   // get tensorrt output
   std::vector<float> out2;
   framework::TensorToVector(*out_tensor, ctx, &out2);
-- 
GitLab