提交 0cb88c34 编写于 作者: N nhzlx

add op converter

上级 afdc7309
...@@ -70,7 +70,7 @@ class DfgPassManagerImpl final : public DfgPassManager { ...@@ -70,7 +70,7 @@ class DfgPassManagerImpl final : public DfgPassManager {
auto trt_teller = [&](const Node* node) { auto trt_teller = [&](const Node* node) {
std::unordered_set<std::string> teller_set( std::unordered_set<std::string> teller_set(
{"mul", "conv2d", "pool2d", "relu", "softmax", "sigmoid", {"mul", "conv2d", "pool2d", "relu", "softmax", "sigmoid",
"depthwise_conv2d", "batch_norm", "concat", "tanh", "depthwise_conv2d", "batch_norm", "concat", "tanh", "pad",
"elementwise_add", "dropout"}); "elementwise_add", "dropout"});
if (!node->IsFunction()) return false; if (!node->IsFunction()) return false;
......
...@@ -185,3 +185,4 @@ USE_TRT_CONVERTER(softmax); ...@@ -185,3 +185,4 @@ USE_TRT_CONVERTER(softmax);
USE_TRT_CONVERTER(batch_norm); USE_TRT_CONVERTER(batch_norm);
USE_TRT_CONVERTER(concat); USE_TRT_CONVERTER(concat);
USE_TRT_CONVERTER(dropout); USE_TRT_CONVERTER(dropout);
USE_TRT_CONVERTER(pad);
# Add TRT tests # Add TRT tests
nv_library(tensorrt_converter nv_library(tensorrt_converter
SRCS mul_op.cc conv2d_op.cc fc_op.cc pool2d_op.cc elementwise_op.cc SRCS mul_op.cc conv2d_op.cc fc_op.cc pool2d_op.cc elementwise_op.cc
batch_norm_op.cc activation_op.cc softmax_op.cc concat_op.cc dropout_op.cc batch_norm_op.cc activation_op.cc softmax_op.cc concat_op.cc dropout_op.cc pad_op.cc
DEPS tensorrt_engine operator scope framework_proto op_registry) DEPS tensorrt_engine operator scope framework_proto op_registry)
nv_test(test_op_converter SRCS test_op_converter.cc DEPS nv_test(test_op_converter SRCS test_op_converter.cc DEPS
...@@ -26,6 +26,8 @@ nv_test(test_trt_batch_norm_op SRCS test_batch_norm_op.cc batch_norm_op.cc ...@@ -26,6 +26,8 @@ nv_test(test_trt_batch_norm_op SRCS test_batch_norm_op.cc batch_norm_op.cc
DEPS ${FLUID_CORE_MODULES} tensorrt_engine batch_norm_op SERIAL) DEPS ${FLUID_CORE_MODULES} tensorrt_engine batch_norm_op SERIAL)
nv_test(test_trt_concat_op SRCS test_concat_op.cc concat_op.cc nv_test(test_trt_concat_op SRCS test_concat_op.cc concat_op.cc
DEPS ${FLUID_CORE_MODULES} tensorrt_engine concat_op SERIAL) DEPS ${FLUID_CORE_MODULES} tensorrt_engine concat_op SERIAL)
nv_test(test_trt_dropout_op SRCS test_dropout_op.cc dropout_op.cc nv_test(test_trt_dropout_op SRCS test_dropout_op.cc dropout_op.cc
DEPS ${FLUID_CORE_MODULES} tensorrt_engine dropout_op SERIAL) DEPS ${FLUID_CORE_MODULES} tensorrt_engine dropout_op SERIAL)
nv_test(test_trt_pad_op SRCS test_pad_op.cc pad_op.cc
DEPS ${FLUID_CORE_MODULES} tensorrt_engine pad_op SERIAL)
/* 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/tensorrt/convert/op_converter.h"
namespace paddle {
namespace inference {
namespace tensorrt {
/*
* PadOp.
*/
class PadOpConverter : public OpConverter {
public:
void operator()(const framework::proto::OpDesc& op,
const framework::Scope& scope, bool test_mode) override {
VLOG(4) << "convert a fluid transpose op to tensorrt tranpose layer";
framework::OpDesc op_desc(op, nullptr);
// Declare inputs
auto* input = engine_->GetITensor(op_desc.Input("X")[0]);
const std::vector<int> paddings =
boost::get<std::vector<int>>(op_desc.GetAttr("paddings"));
const float pad_value = boost::get<float>(op_desc.GetAttr("pad_value"));
nvinfer1::Dims input_shape = input->getDimensions();
int nbDims = input_shape.nbDims;
int pad_size = static_cast<int>(paddings.size());
PADDLE_ENFORCE_GE(nbDims, 2);
PADDLE_ENFORCE_EQ((nbDims + 1) * 2, pad_size);
PADDLE_ENFORCE(pad_value == 0.0, "The pad layer of TRT only support zero.");
nvinfer1::DimsHW pre_pad(paddings[pad_size - 4], paddings[pad_size - 2]);
nvinfer1::DimsHW post_pad(paddings[pad_size - 3], paddings[pad_size - 1]);
auto* layer = TRT_ENGINE_ADD_LAYER(engine_, Padding,
*const_cast<nvinfer1::ITensor*>(input),
pre_pad, post_pad);
PADDLE_ENFORCE(layer != nullptr);
auto output_name = op_desc.Output("Out")[0];
engine_->SetITensor(output_name, layer->getOutput(0));
layer->setName(("scale (Output: " + output_name + ")").c_str());
layer->getOutput(0)->setName(output_name.c_str());
if (test_mode) { // the test framework can not determine which is the
// output, so place the declaration inside.
engine_->DeclareOutput(output_name);
}
}
};
} // namespace tensorrt
} // namespace inference
} // namespace paddle
REGISTER_TRT_OP_CONVERTER(pad, PadOpConverter);
/* 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 <gtest/gtest.h>
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/inference/tensorrt/convert/ut_helper.h"
namespace paddle {
namespace inference {
namespace tensorrt {
TEST(PadConverter, main) {
framework::Scope scope;
std::unordered_set<std::string> parameters;
TRTConvertValidation validator(10, parameters, scope, 1000);
validator.DeclInputVar("pad-X", nvinfer1::Dims3(3, 2, 2));
validator.DeclOutputVar("pad-Out", nvinfer1::Dims3(3, 3, 5));
// Prepare Op description
framework::OpDesc desc;
desc.SetType("pad");
desc.SetInput("X", {"pad-X"});
desc.SetOutput("Out", {"pad-Out"});
std::vector<int> paddings = {0, 0, 0, 0, 0, 1, 1, 2};
float pad_value = 0.0;
desc.SetAttr("paddings", paddings);
desc.SetAttr("pad_value", pad_value);
LOG(INFO) << "set OP";
validator.SetOp(*desc.Proto());
LOG(INFO) << "execute";
validator.Execute(2);
}
} // namespace tensorrt
} // namespace inference
} // namespace paddle
USE_OP(pad);
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册