use the latest buffer to update the convert

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.
 };
 

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)

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)

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

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);