Skip to content

[Paddle Inference] Add trt tile converter for dynamic shape. #50841

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 2 commits into from
Mar 2, 2023
Merged

[Paddle Inference] Add trt tile converter for dynamic shape. #50841

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
fff13f8
add_trt_tile
xiaoxiaohehe001 Feb 23, 2023
50c36f9
tile_trt
xiaoxiaohehe001 Feb 28, 2023
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
141 changes: 98 additions & 43 deletions paddle/fluid/inference/tensorrt/convert/tile_op.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,8 +1,11 @@
/* 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.
Expand All @@ -11,65 +14,117 @@ limitations under the License. */

#include "paddle/fluid/inference/tensorrt/convert/op_converter.h"

namespace paddle {
namespace framework {
class Scope;
namespace proto {
class OpDesc;
} // namespace proto
} // namespace framework
} // namespace paddle

namespace paddle {
namespace inference {
namespace tensorrt {

/*
* ReshapeOp
*/
class TileOpConverter : public OpConverter {
public:
void operator()(const framework::proto::OpDesc& op,
const framework::Scope& scope,
bool test_mode) override {
#if IS_TRT_VERSION_GE(7000)
VLOG(4) << "convert a fluid tile op to tensorrt tile layer";
VLOG(4) << "convert a tile op to tensorrt tile layer";

framework::OpDesc op_desc(op, nullptr);
// Declare inputs
auto* input = engine_->GetITensor(op_desc.Input("X")[0]);
nvinfer1::Dims input_shape = input->getDimensions();
std::vector<int> repeat_times =
PADDLE_GET_CONST(std::vector<int>, op_desc.GetAttr("repeat_times"));

nvinfer1::Dims output_dim = input_shape;
nvinfer1::Dims output_stride;
// If input_dims.nbDims + 1 < repeat_times.size() means we
// should expand 1 on batchsize. trt doesn't support this behavior.
PADDLE_ENFORCE_GE(input_shape.nbDims + 1,
repeat_times.size(),
platform::errors::InvalidArgument(
"Can't change batchsize, please check repeat_times"));
int diff = input_shape.nbDims + 1 - repeat_times.size();
if (diff > 0) repeat_times.insert(repeat_times.begin(), diff, 1);

// Can't expand on batchsize
PADDLE_ENFORCE_EQ(
repeat_times[0],
1,
platform::errors::InvalidArgument(
"Can't expand on batchsize, please check repeat_times"));
output_stride.nbDims = input_shape.nbDims;
for (int i = 0; i < input_shape.nbDims; i++) {
output_dim.d[i] = output_dim.d[i] * repeat_times[i + 1];
output_stride.d[i] = 1;
auto inputs = op_desc.Inputs();
auto input_shape = input->getDimensions();
auto rank = input_shape.nbDims;
auto output_name = op_desc.Output("Out")[0];

if (engine_->with_dynamic_shape()) {
std::vector<int32_t> start(rank, 0);
std::vector<int32_t> stride(rank, 1);
auto start_tensor =
Add1DConstantLayer(start, output_name + "start_tensor");
auto stride_tensor =
Add1DConstantLayer(stride, output_name + "stride_tensor");
auto input_shape_tensor = Shape(input);

nvinfer1::ITensor* repeat_tensor = nullptr;
int32_t repeat_rank = 0;
if (inputs.find("RepeatTimes") != inputs.end() &&
op_desc.Input("RepeatTimes").size() >= 1) {
repeat_tensor = engine_->GetITensor(op_desc.Input("RepeatTimes")[0]);
repeat_rank = repeat_tensor->getDimensions().d[0];
} else if (inputs.find("repeat_times_tensor") != inputs.end() &&
op_desc.Input("repeat_times_tensor").size() >= 1) {
int32_t repeat_size = op_desc.Input("repeat_times_tensor").size();
std::vector<nvinfer1::ITensor*> repeat_tensors;
for (int32_t i = 0; i < repeat_size; ++i) {
repeat_tensors.push_back(
engine_->GetITensor(op_desc.Input("repeat_times_tensor")[i]));
}
repeat_tensor = Concat(repeat_tensors);
repeat_rank = repeat_size;
} else {
std::vector<int32_t> repeat_times = PADDLE_GET_CONST(
std::vector<int32_t>, op_desc.GetAttr("repeat_times"));
repeat_tensor =
Add1DConstantLayer(repeat_times, output_name + "_shape_tensor_");
repeat_rank = repeat_times.size();
}

nvinfer1::ITensor* repeat_expand_tensor;
if (rank > repeat_rank) {
auto* one_rank_tensor =
Add1DConstantLayer(std::vector<int32_t>(rank - repeat_rank, 1),
output_name + "_one_rank_tensor_");
std::vector<nvinfer1::ITensor*> itensors;
itensors.push_back(one_rank_tensor);
itensors.push_back(repeat_tensor);
repeat_expand_tensor = Concat(itensors);
} else {
repeat_expand_tensor = repeat_tensor;
}
auto output_shape_tensor = Prod(input_shape_tensor, repeat_expand_tensor);
auto layer = TRT_ENGINE_ADD_LAYER(engine_,
Slice,
*input,
nvinfer1::Dims{},
nvinfer1::Dims{},
nvinfer1::Dims{});

layer->setInput(1, *start_tensor);
layer->setInput(2, *output_shape_tensor);
layer->setInput(3, *stride_tensor);
layer->setMode(nvinfer1::SliceMode::kWRAP);
RreplenishLayerAndOutput(layer, "tile", {output_name}, test_mode);

} else {
std::vector<int> repeat_times =
PADDLE_GET_CONST(std::vector<int>, op_desc.GetAttr("repeat_times"));
auto output_dim = input_shape;
auto output_stride = input_shape;
// If input_dims.nbDims + 1 < repeat_times.size() means we
// should expand 1 on batchsize. trt doesn't support this behavior.
PADDLE_ENFORCE_GE(
rank + 1,
repeat_times.size(),
platform::errors::InvalidArgument(
"Can't change batchsize, please check repeat_times"));
int32_t diff = rank + 1 - repeat_times.size();
if (diff > 0) repeat_times.insert(repeat_times.begin(), diff, 1);

// Can't expand on batchsize
PADDLE_ENFORCE_EQ(
repeat_times[0],
1,
platform::errors::InvalidArgument(
"Can't expand on batchsize, please check repeat_times"));
output_stride.nbDims = rank;
for (int32_t i = 0; i < rank; i++) {
output_dim.d[i] = output_dim.d[i] * repeat_times[i + 1];
output_stride.d[i] = 1;
}
auto layer = TRT_ENGINE_ADD_LAYER(
engine_, Slice, *input, input_shape, output_dim, output_stride);
layer->setMode(nvinfer1::SliceMode::kWRAP);
RreplenishLayerAndOutput(layer, "tile", {output_name}, test_mode);
}

auto* layer = TRT_ENGINE_ADD_LAYER(
engine_, Slice, *input, input_shape, output_dim, output_stride);
layer->setMode(nvinfer1::SliceMode::kWRAP);
auto output_name = op_desc.Output("Out")[0];
RreplenishLayerAndOutput(layer, "tile", {output_name}, test_mode);
#endif
}
};
Expand Down
19 changes: 10 additions & 9 deletions paddle/fluid/inference/tensorrt/op_teller.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2232,18 +2232,19 @@ struct SimpleOpTypeSetTeller : public Teller {
if (op_type == "tile") {
// Paddle-TRT does not support the input tensors.
auto tile_inputs = desc.Inputs();
if (tile_inputs.find("repeat_times_tensor") != tile_inputs.end()) {
if (desc.Input("repeat_times_tensor").size() >= 1) {
return false;
if (!with_dynamic_shape) {
if (tile_inputs.find("repeat_times_tensor") != tile_inputs.end()) {
if (desc.Input("repeat_times_tensor").size() >= 1) {
return false;
}
}
}
if (tile_inputs.find("RepeatTimes") != tile_inputs.end()) {
if (desc.Input("RepeatTimes").size() >= 1) {
return false;
if (tile_inputs.find("RepeatTimes") != tile_inputs.end()) {
if (desc.Input("RepeatTimes").size() >= 1) {
return false;
}
}
if (!desc.HasAttr("repeat_times")) return false;
}
if (with_dynamic_shape) return false;
if (!with_dynamic_shape && !desc.HasAttr("repeat_times")) return false;
}
#endif

Expand Down
Loading