未验证 提交 b7db8457 编写于 作者: Z zhoutianzi666 提交者: GitHub

[Paddle-TRT] reshape fill_constant (#44314)

* reshape fill_constant

* commit

* commit
上级 fd6dcdfe
......@@ -2089,6 +2089,7 @@ USE_TRT_CONVERTER(top_k)
USE_TRT_CONVERTER(top_k_v2)
USE_TRT_CONVERTER(squeeze2)
USE_TRT_CONVERTER(unsqueeze2)
USE_TRT_CONVERTER(fill_constant)
USE_TRT_CONVERTER(fused_token_prune)
#if PADDLE_WITH_CUSPARSELT && IS_TRT_VERSION_GE(8000)
USE_TRT_CONVERTER(sparse_fc)
......
......@@ -69,6 +69,7 @@ list(
top_k_op.cc
squeeze2_op.cc
unsqueeze2_op.cc
fill_constant_op.cc
fused_token_prune_op.cc)
if(CUSPARSELT_FOUND AND ${TENSORRT_MAJOR_VERSION} GREATER_EQUAL 8)
......
/* 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 {
class FillConstantOpConverter : public OpConverter {
public:
void operator()(const framework::proto::OpDesc& op,
const framework::Scope& scope,
bool test_mode) override {
VLOG(4)
<< "convert a fluid fill_constant op to tensorrt fill_constant layer";
framework::OpDesc op_desc(op, nullptr);
int dtype = BOOST_GET_CONST(int, op_desc.GetAttr("dtype"));
std::string str_value =
BOOST_GET_CONST(std::string, op_desc.GetAttr("str_value"));
std::vector<int64_t> shape =
BOOST_GET_CONST(std::vector<int64_t>, op_desc.GetAttr("shape"));
std::unique_ptr<framework::Tensor> out_tensor(new framework::Tensor());
out_tensor->Resize(phi::make_ddim(shape));
nvinfer1::DataType trt_dtype = nvinfer1::DataType::kFLOAT;
void* trt_data = nullptr;
size_t trt_num;
if (dtype == 2 || dtype == 3) { // int,int64
auto* tmp_ptr = out_tensor->mutable_data<int>(platform::CPUPlace());
for (int64_t i = 0; i < out_tensor->numel(); i++)
tmp_ptr[i] = std::stoi(str_value);
trt_dtype = nvinfer1::DataType::kINT32;
trt_data = static_cast<void*>(tmp_ptr);
} else if (dtype == 5) { // float
auto* tmp_ptr = out_tensor->mutable_data<float>(platform::CPUPlace());
for (int64_t i = 0; i < out_tensor->numel(); i++)
tmp_ptr[i] = std::stof(str_value);
trt_data = static_cast<void*>(tmp_ptr);
}
trt_num = static_cast<size_t>(out_tensor->numel());
engine_->SetWeights("fill_constant_value", std::move(out_tensor));
TensorRTEngine::Weight weight{trt_dtype, trt_data, trt_num};
nvinfer1::Dims trt_in_shape;
trt_in_shape.nbDims = shape.size();
for (size_t i = 0; i < shape.size(); i++) trt_in_shape.d[i] = shape[i];
nvinfer1::ILayer* layer =
TRT_ENGINE_ADD_LAYER(engine_, Constant, trt_in_shape, weight.get());
auto output_name = op_desc.Output("Out")[0];
RreplenishLayerAndOutput(layer, "fill_constant", {output_name}, test_mode);
}
};
} // namespace tensorrt
} // namespace inference
} // namespace paddle
REGISTER_TRT_OP_CONVERTER(fill_constant, FillConstantOpConverter);
......@@ -35,14 +35,29 @@ class ReshapeOpConverter : public OpConverter {
framework::OpDesc op_desc(op, nullptr);
// Declare inputs
auto* input = engine_->GetITensor(op_desc.Input("X")[0]);
std::vector<int> shape =
BOOST_GET_CONST(std::vector<int>, op_desc.GetAttr("shape"));
int nbDims_num = shape.size();
nvinfer1::Dims reshape_dim;
if (engine_->with_dynamic_shape()) { // running the TRT Dynamic Shape mode
reshape_dim.nbDims = nbDims_num;
for (int i = 0; i < nbDims_num; ++i) {
reshape_dim.d[i] = shape[i];
nvinfer1::ITensor* real_shape_tensor = nullptr;
std::vector<nvinfer1::ITensor*> concat_inputs;
bool one_input = false;
if (engine_->with_dynamic_shape()) {
if (op_desc.Inputs().find("ShapeTensor") != op_desc.Inputs().end() &&
op_desc.Input("ShapeTensor").size() > 0) {
for (auto name : op_desc.Input("ShapeTensor"))
concat_inputs.push_back(engine_->GetITensor(name));
real_shape_tensor = Concat(concat_inputs);
} else if (op_desc.Inputs().find("Shape") != op_desc.Inputs().end() &&
op_desc.Input("Shape").size() > 0) {
real_shape_tensor = engine_->GetITensor(op_desc.Input("Shape")[0]);
} else {
reshape_dim.nbDims = nbDims_num;
for (int i = 0; i < nbDims_num; ++i) {
reshape_dim.d[i] = shape[i];
}
one_input = true;
}
} else { // running the TRT Static Shape mode
reshape_dim.nbDims = nbDims_num - 1;
......@@ -51,7 +66,10 @@ class ReshapeOpConverter : public OpConverter {
}
}
auto* layer = TRT_ENGINE_ADD_LAYER(engine_, Shuffle, *input);
layer->setReshapeDimensions(reshape_dim);
if (!engine_->with_dynamic_shape() || one_input)
layer->setReshapeDimensions(reshape_dim);
else
layer->setInput(1, *real_shape_tensor);
auto output_name = op_desc.Output("Out")[0];
RreplenishLayerAndOutput(layer, "reshape", {output_name}, test_mode);
}
......
......@@ -169,6 +169,7 @@ struct SimpleOpTypeSetTeller : public Teller {
"transformer_input_convert",
"recover_padding",
"remove_padding",
"fill_constant",
"squeeze2",
"unsqueeze2"};
std::unordered_set<std::string> teller_set{
......@@ -274,6 +275,7 @@ struct SimpleOpTypeSetTeller : public Teller {
"transformer_input_convert",
"recover_padding",
"remove_padding",
"fill_constant",
"squeeze2",
"unsqueeze2",
"fused_token_prune"};
......@@ -1448,6 +1450,27 @@ bool OpTeller::Tell(const framework::ir::Node* node,
}
}
if (op_type == "fill_constant") {
auto fill_constant_inputs = desc.Inputs();
if (fill_constant_inputs.find("ValueTensor") !=
fill_constant_inputs.end()) {
if (desc.Input("ValueTensor").size()) return false;
}
if (fill_constant_inputs.find("ShapeTensor") !=
fill_constant_inputs.end()) {
if (desc.Input("ShapeTensor").size()) return false;
}
if (fill_constant_inputs.find("ShapeTensorList") !=
fill_constant_inputs.end()) {
if (desc.Input("ShapeTensorList").size()) return false;
}
int dtype = BOOST_GET_CONST(int, desc.GetAttr("dtype"));
// only support int32, int64, float32
if (!(dtype == 2 || dtype == 3 || dtype == 5)) {
return false;
}
}
if (op_type == "instance_norm") {
if (with_dynamic_shape) {
VLOG(3) << "trt instance_norm op does not support dynamic shape ";
......@@ -1801,6 +1824,9 @@ bool OpTeller::Tell(const framework::ir::Node* node,
}
if (op_type == "reshape" || op_type == "reshape2") {
if (with_dynamic_shape) {
return true;
}
if (!desc.HasAttr("shape")) {
return false;
}
......
# Copyright (c) 2021 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.
from trt_layer_auto_scan_test import TrtLayerAutoScanTest, SkipReasons
from program_config import TensorConfig, ProgramConfig
import unittest
import numpy as np
import paddle.inference as paddle_infer
from functools import partial
from typing import Optional, List, Callable, Dict, Any, Set
class TrtConvertSplitTest(TrtLayerAutoScanTest):
def is_program_valid(self, program_config: ProgramConfig) -> bool:
return True
def sample_program_configs(self):
def generate_value_data(attrs: List[Dict[str, Any]]):
return np.array([1]).astype(np.int32)
def generate_shape_data(attrs: List[Dict[str, Any]]):
return np.array([4, 23]).astype(np.int32)
def generate_shapelist_data(attrs: List[Dict[str, Any]]):
return np.array([4]).astype(np.int32)
for shape in [[2, 3, 4]]:
for num_input in [0, 1, 2, 3]:
for dtype in [5, 2, 3]:
for str_value in ["2", "23", "-1"]:
self.num_input = num_input
dics = [{
"str_value": str_value,
"shape": shape,
"dtype": dtype
}, {
"axis": -1
}]
dics_intput = [{
"ValueTensor": ["value_data"]
}, {
"ShapeTensor": ["shape_data"],
}, {
"ShapeTensorList": ["shapeT1_data", "shapeT2_data"],
}, {}]
ops_config = [
{
"op_type": "fill_constant",
"op_inputs": dics_intput[num_input],
"op_outputs": {
"Out": ["out_data"],
},
"op_attrs": dics[0]
},
]
def generate_input():
return np.random.random([1, 1]).astype(np.float32)
ops = self.generate_op_config(ops_config)
program_config = ProgramConfig(
ops=ops,
weights={},
inputs={
"value_data":
TensorConfig(data_gen=partial(
generate_value_data, dics)),
"shape_data":
TensorConfig(data_gen=partial(
generate_shape_data, dics)),
"shapeT1_data":
TensorConfig(data_gen=partial(
generate_shapelist_data, dics)),
"shapeT2_data":
TensorConfig(data_gen=partial(
generate_shapelist_data, dics)),
},
outputs=["out_data"])
yield program_config
def sample_predictor_configs(
self, program_config) -> (paddle_infer.Config, List[int], float):
def generate_dynamic_shape(attrs):
self.input_shape = [1, 1]
max_shape = list(self.input_shape)
min_shape = list(self.input_shape)
opt_shape = list(self.input_shape)
for i in range(len(self.input_shape)):
max_shape[i] = max_shape[i] + 1
self.dynamic_shape.min_input_shape = {"Y_data": min_shape}
self.dynamic_shape.max_input_shape = {"Y_data": max_shape}
self.dynamic_shape.opt_input_shape = {"Y_data": opt_shape}
def clear_dynamic_shape():
self.dynamic_shape.min_input_shape = {}
self.dynamic_shape.max_input_shape = {}
self.dynamic_shape.opt_input_shape = {}
def generate_trt_nodes_num(attrs, dynamic_shape):
if (self.num_input < 3):
return 0, 6
return 1, 5
attrs = [
program_config.ops[i].attrs for i in range(len(program_config.ops))
]
# Don't test static shape
# for dynamic_shape
generate_dynamic_shape(attrs)
self.trt_param.precision = paddle_infer.PrecisionType.Float32
yield self.create_inference_config(), generate_trt_nodes_num(
attrs, True), 1e-5
self.trt_param.precision = paddle_infer.PrecisionType.Half
yield self.create_inference_config(), generate_trt_nodes_num(
attrs, True), 1e-5
def add_skip_trt_case(self):
pass
def test(self):
self.add_skip_trt_case()
self.run_test()
if __name__ == "__main__":
unittest.main()
......@@ -48,12 +48,16 @@ class TrtConvertReshapeTest(TrtLayerAutoScanTest):
def generate_input1(attrs: List[Dict[str, Any]]):
if self.dims == 4:
self.input_shape = [1, 2, 4, 6]
return np.ones([1, 2, 4, 6]).astype(np.float32)
elif self.dims == 3:
self.input_shape = [1, 8, 6]
return np.ones([1, 8, 6]).astype(np.float32)
elif self.dims == 2:
self.input_shape = [1, 48]
return np.ones([1, 48]).astype(np.float32)
elif self.dims == 1:
self.input_shape = [48]
return np.ones([48]).astype(np.float32)
def generate_weight1(attrs: List[Dict[str, Any]]):
......@@ -66,69 +70,36 @@ class TrtConvertReshapeTest(TrtLayerAutoScanTest):
return np.array([24]).astype(np.int32)
for dims in [4, 3, 2, 1]:
for num_input in [0, 1, 2, 3]:
for shape in [[1, 6, 8], [1, 2, 4, 6], [1, 1, 0, 12], [1, 0, 6],
[1, -1, 12], [2, -1], [3, 16], [3, 4, 4], [48]]:
dics = [{
for shape in [[1, 6, 8], [1, 2, 4, 6], [1, 1, 0, 12], [1, 0, 6],
[1, -1, 12], [2, -1], [3, 16], [3, 4, 4], [48],
[-1, 48]]:
dics = [
{
"shape": shape,
}, {}]
self.num_input = num_input
self.dims = dims
dics_intput = [{
"X": ["reshape_input"],
"Shape": ["shape_data"],
"ShapeTensor": ["shapeT1_data", "shapeT2_data"],
}, {
"X": ["reshape_input"],
"Shape": ["shape_data"],
}, {
"X": ["reshape_input"],
"ShapeTensor": ["shapeT1_data", "shapeT2_data"],
}, {
"X": ["reshape_input"]
}]
dics_weight = [{
"shape_data":
TensorConfig(data_gen=partial(generate_weight1, dics)),
"shapeT1_data":
TensorConfig(
data_gen=partial(generate_shapeT1_data, dics)),
"shapeT2_data":
TensorConfig(
data_gen=partial(generate_shapeT2_data, dics))
}, {
"shape_data":
TensorConfig(data_gen=partial(generate_weight1, dics))
}, {
"shapeT1_data":
TensorConfig(
data_gen=partial(generate_shapeT1_data, dics)),
"shapeT2_data":
TensorConfig(
data_gen=partial(generate_shapeT2_data, dics))
}, {}]
ops_config = [{
"op_type": "reshape",
"op_inputs": dics_intput[num_input],
"op_outputs": {
"Out": ["reshape_out"]
},
"op_attrs": dics[0]
}]
ops = self.generate_op_config(ops_config)
program_config = ProgramConfig(
ops=ops,
weights=dics_weight[num_input],
inputs={
"reshape_input":
TensorConfig(
data_gen=partial(generate_input1, dics))
},
outputs=["reshape_out"])
},
]
self.dims = dims
dics_intput = [{"X": ["reshape_input"]}]
ops_config = [{
"op_type": "reshape",
"op_inputs": dics_intput[0],
"op_outputs": {
"Out": ["reshape_out"]
},
"op_attrs": dics[0]
}]
ops = self.generate_op_config(ops_config)
program_config = ProgramConfig(
ops=ops,
weights={},
inputs={
"reshape_input":
TensorConfig(data_gen=partial(generate_input1, dics))
},
outputs=["reshape_out"])
yield program_config
yield program_config
def sample_predictor_configs(
self, program_config) -> (paddle_infer.Config, List[int], float):
......@@ -169,22 +140,31 @@ class TrtConvertReshapeTest(TrtLayerAutoScanTest):
self.dynamic_shape.opt_input_shape = {}
def generate_trt_nodes_num(attrs, dynamic_shape):
# in static shape mode, here is consistent with op_teller.cc
if (not dynamic_shape):
if (attrs[0]['shape'][0] == 0):
return 1, 2
elif (len(attrs[0]['shape']) == 1):
return 0, 3
elif (np.prod(attrs[0]['shape'][1:]) == np.prod(
self.input_shape[1:])):
return 1, 2
else:
return 0, 3
return 1, 2
attrs = [
program_config.ops[i].attrs for i in range(len(program_config.ops))
]
if attrs[0]['shape'][0] > 1 and len(attrs[0]['shape']) > 1:
pass
else:
# for static_shape
clear_dynamic_shape()
self.trt_param.precision = paddle_infer.PrecisionType.Float32
yield self.create_inference_config(), generate_trt_nodes_num(
attrs, False), 1e-5
self.trt_param.precision = paddle_infer.PrecisionType.Half
yield self.create_inference_config(), generate_trt_nodes_num(
attrs, False), 1e-5
# for static_shape
clear_dynamic_shape()
self.trt_param.precision = paddle_infer.PrecisionType.Float32
yield self.create_inference_config(), generate_trt_nodes_num(
attrs, False), 1e-5
self.trt_param.precision = paddle_infer.PrecisionType.Half
yield self.create_inference_config(), generate_trt_nodes_num(
attrs, False), 1e-5
# for dynamic_shape
generate_dynamic_shape(attrs)
......@@ -196,14 +176,243 @@ class TrtConvertReshapeTest(TrtLayerAutoScanTest):
attrs, True), 1e-5
def add_skip_trt_case(self):
pass
def test(self):
self.add_skip_trt_case()
self.run_test()
# reshape having three inputs.
class TrtConvertReshapeTest2(TrtLayerAutoScanTest):
def is_program_valid(self, program_config: ProgramConfig) -> bool:
return True
def sample_program_configs(self):
def teller1(program_config, predictor_config):
if len(program_config.weights) >= 1:
return True
return False
def generate_input1(attrs: List[Dict[str, Any]]):
if self.dims == 4:
return np.random.random([1, 2, 4, 6]).astype(np.float32)
elif self.dims == 3:
return np.random.random([1, 8, 6]).astype(np.float32)
elif self.dims == 2:
return np.random.random([1, 48]).astype(np.float32)
elif self.dims == 1:
return np.random.random([48]).astype(np.float32)
self.add_skip_case(teller1, SkipReasons.TRT_NOT_SUPPORT,
"INPUT ShapeTensor and Shape NOT SUPPORT")
for dims in [4, 3, 2, 1]:
for shape in [[-1, 48]]:
dics = [{
"shape": shape,
}, {}]
self.dims = dims
dics_intput = [
{
"X": ["reshape_input"],
"ShapeTensor": ["shapeT1_data", "shapeT2_data"],
},
]
ops_config = [
{
"op_type": "fill_constant",
"op_inputs": {},
"op_outputs": {
"Out": ["shapeT1_data"]
},
"op_attrs": {
"dtype": 2,
"str_value": "2",
"shape": [1],
},
},
{
"op_type": "fill_constant",
"op_inputs": {},
"op_outputs": {
"Out": ["shapeT2_data"]
},
"op_attrs": {
"dtype": 2,
"str_value": "24",
"shape": [1],
},
},
{
"op_type": "reshape",
"op_inputs": dics_intput[0],
"op_outputs": {
"Out": ["reshape_out"]
},
"op_attrs": dics[0]
},
]
ops = self.generate_op_config(ops_config)
program_config = ProgramConfig(
ops=ops,
weights={},
inputs={
"reshape_input":
TensorConfig(data_gen=partial(generate_input1, dics))
},
outputs=["reshape_out"])
yield program_config
def sample_predictor_configs(
self, program_config) -> (paddle_infer.Config, List[int], float):
def generate_dynamic_shape():
if self.dims == 4:
self.dynamic_shape.min_input_shape = {
"reshape_input": [1, 2, 4, 6]
}
self.dynamic_shape.max_input_shape = {
"reshape_input": [4, 2, 4, 6]
}
self.dynamic_shape.opt_input_shape = {
"reshape_input": [1, 2, 4, 6]
}
elif self.dims == 3:
self.dynamic_shape.min_input_shape = {
"reshape_input": [1, 8, 6]
}
self.dynamic_shape.max_input_shape = {
"reshape_input": [4, 8, 6]
}
self.dynamic_shape.opt_input_shape = {
"reshape_input": [1, 8, 6]
}
elif self.dims == 2:
self.dynamic_shape.min_input_shape = {"reshape_input": [1, 48]}
self.dynamic_shape.max_input_shape = {"reshape_input": [4, 48]}
self.dynamic_shape.opt_input_shape = {"reshape_input": [1, 48]}
elif self.dims == 1:
self.dynamic_shape.min_input_shape = {"reshape_input": [48]}
self.dynamic_shape.max_input_shape = {"reshape_input": [48]}
self.dynamic_shape.opt_input_shape = {"reshape_input": [48]}
# for dynamic_shape
generate_dynamic_shape()
self.trt_param.precision = paddle_infer.PrecisionType.Float32
yield self.create_inference_config(), (1, 2), 1e-5
self.trt_param.precision = paddle_infer.PrecisionType.Half
yield self.create_inference_config(), (1, 2), 1e-5
def add_skip_trt_case(self):
pass
def test(self):
self.add_skip_trt_case()
self.run_test()
# reshape having 2 inputs.
class TrtConvertReshapeTest3(TrtLayerAutoScanTest):
def is_program_valid(self, program_config: ProgramConfig) -> bool:
return True
def sample_program_configs(self):
def generate_input1(attrs: List[Dict[str, Any]]):
if self.dims == 4:
return np.random.random([1, 2, 12, 6]).astype(np.float32)
elif self.dims == 3:
return np.random.random([1, 8, 18]).astype(np.float32)
elif self.dims == 2:
return np.random.random([1, 144]).astype(np.float32)
elif self.dims == 1:
return np.random.random([144]).astype(np.float32)
for dims in [4, 3, 2, 1]:
for shape in [[-1, 144]]:
dics = [{
"shape": shape,
}, {}]
self.dims = dims
dics_intput = [
{
"X": ["reshape_input"],
"shape_data": ["shape_data"],
},
]
ops_config = [
{
"op_type": "fill_constant",
"op_inputs": {},
"op_outputs": {
"Out": ["shape_data"]
},
"op_attrs": {
"dtype": 2,
"str_value": "12",
"shape": [2],
},
},
{
"op_type": "reshape",
"op_inputs": dics_intput[0],
"op_outputs": {
"Out": ["reshape_out"]
},
"op_attrs": dics[0]
},
]
ops = self.generate_op_config(ops_config)
program_config = ProgramConfig(
ops=ops,
weights={},
inputs={
"reshape_input":
TensorConfig(data_gen=partial(generate_input1, dics))
},
outputs=["reshape_out"])
yield program_config
def sample_predictor_configs(
self, program_config) -> (paddle_infer.Config, List[int], float):
def generate_dynamic_shape():
if self.dims == 4:
self.dynamic_shape.min_input_shape = {
"reshape_input": [1, 2, 12, 6]
}
self.dynamic_shape.max_input_shape = {
"reshape_input": [4, 2, 12, 6]
}
self.dynamic_shape.opt_input_shape = {
"reshape_input": [1, 2, 12, 6]
}
elif self.dims == 3:
self.dynamic_shape.min_input_shape = {
"reshape_input": [1, 8, 18]
}
self.dynamic_shape.max_input_shape = {
"reshape_input": [4, 8, 18]
}
self.dynamic_shape.opt_input_shape = {
"reshape_input": [1, 8, 18]
}
elif self.dims == 2:
self.dynamic_shape.min_input_shape = {"reshape_input": [1, 144]}
self.dynamic_shape.max_input_shape = {"reshape_input": [4, 144]}
self.dynamic_shape.opt_input_shape = {"reshape_input": [1, 144]}
elif self.dims == 1:
self.dynamic_shape.min_input_shape = {"reshape_input": [144]}
self.dynamic_shape.max_input_shape = {"reshape_input": [144]}
self.dynamic_shape.opt_input_shape = {"reshape_input": [144]}
# for dynamic_shape
generate_dynamic_shape()
self.trt_param.precision = paddle_infer.PrecisionType.Float32
yield self.create_inference_config(), (1, 2), 1e-5
self.trt_param.precision = paddle_infer.PrecisionType.Half
yield self.create_inference_config(), (1, 2), 1e-5
def add_skip_trt_case(self):
pass
def test(self):
self.add_skip_trt_case()
......
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