diff --git a/paddle/fluid/inference/tensorrt/convert/split_op.cc b/paddle/fluid/inference/tensorrt/convert/split_op.cc index 47a6dd783a70cf7b4a8c3d7beb988fbc0f6a8786..591eb06a362024d675814975dbc168652a4dc5eb 100644 --- a/paddle/fluid/inference/tensorrt/convert/split_op.cc +++ b/paddle/fluid/inference/tensorrt/convert/split_op.cc @@ -49,43 +49,21 @@ class SplitOpConverter : public OpConverter { } else { axis += (axis < 0) ? input_dims.nbDims : -1; } - - PADDLE_ENFORCE_NE(input_dims.d[axis], -1, - platform::errors::InvalidArgument( - "The (%d) dim of input should not be -1", axis)); if (num > 0) { int64_t in_axis_dim = input_dims.d[axis]; - PADDLE_ENFORCE_EQ( - in_axis_dim % num, 0, - platform::errors::InvalidArgument( - "Invalid number to split. Tensor split does not result" - " in an equal division of dimensions. Axis dim = %d %% num = %d " - "!= 0", - in_axis_dim, num)); size_t out_axis_dim = in_axis_dim / num; for (int i = 0; i < num; ++i) { output_lengths.push_back(out_axis_dim); } } - PADDLE_ENFORCE_EQ( - output_lengths.size(), output_num, - platform::errors::InvalidArgument( - "The output_length should be equal to the output size.")); - nvinfer1::ILayer* layer = nullptr; if (engine_->with_dynamic_shape()) { -#if IS_TRT_VERSION_GE(6000) bool with_fp16 = engine_->WithFp16() && !engine_->disable_trt_plugin_fp16(); plugin::SplitPluginDynamic* plugin = new plugin::SplitPluginDynamic(axis, output_lengths, with_fp16); layer = engine_->AddDynamicPlugin(&input, input_num, plugin); -#else - PADDLE_THROW(platform::errors::Fatal( - "You are running the TRT Dynamic Shape mode, need to confirm that " - "your TRT version is no less than 6.0")); -#endif } else { bool with_fp16 = engine_->WithFp16() && !engine_->disable_trt_plugin_fp16(); diff --git a/paddle/fluid/inference/tensorrt/op_teller.cc b/paddle/fluid/inference/tensorrt/op_teller.cc index 3067c2893825d3ce2e092959ad24d2ba8b1f2a01..ad63f3ecfa7a0ac5028b47209c4450029b06b0a9 100644 --- a/paddle/fluid/inference/tensorrt/op_teller.cc +++ b/paddle/fluid/inference/tensorrt/op_teller.cc @@ -577,16 +577,78 @@ bool OpTeller::Tell(const framework::ir::Node* node, bool use_no_calib_int8, << desc.Input("X").size() << "."; return false; } + auto split_inputs = desc.Inputs(); + if (split_inputs.find("AxisTensor") != split_inputs.end()) { + if (desc.Input("AxisTensor").size() >= 1) { + return false; + } + } + if (split_inputs.find("SectionsTensorList") != split_inputs.end()) { + if (desc.Input("SectionsTensorList").size() >= 1) { + return false; + } + } if (!desc.HasAttr("axis")) { return false; - } else { - int axis = BOOST_GET_CONST(int, desc.GetAttr("axis")); - if (axis == 0) { - VLOG(3) << "Invalid split axis. Split on batch is not supported in " - "TensorRT"; - return false; + } + int axis = BOOST_GET_CONST(int, desc.GetAttr("axis")); + + if (axis == 0) { + VLOG(3) << "Invalid split axis. Split on batch is not supported in " + "TensorRT"; + return false; + } + auto* block = desc.Block(); + auto x_var_name = desc.Input("X")[0]; + auto* x_var_desc = block->FindVar(x_var_name); + const auto x_shape = x_var_desc->GetShape(); + size_t output_num = desc.Output("Out").size(); + std::vector output_lengths; + int num = 0; + if (desc.HasAttr("num")) { + num = BOOST_GET_CONST(int, desc.GetAttr("num")); + } + if (desc.HasAttr("sections")) { + output_lengths = + BOOST_GET_CONST(std::vector, desc.GetAttr("sections")); + } + if (output_lengths.size() == 0 && num == 0) { + VLOG(3) << "sections and num cannot be equal to 0 at the same time"; + return false; + } + if (with_dynamic_shape) { +#if IS_TRT_VERSION_GE(6000) +#else + VLOG(3) << "You are running the TRT Dynamic Shape mode, need to " + "confirm that " + "your TRT version is no less than 6.0"; + return false; +#endif + } + axis += (axis < 0) ? x_shape.size() : 0; + if (x_shape[axis] == -1) { + VLOG(3) << "The (" << axis << ") dim of input should not be -1"; + return false; + } + if (output_lengths.size() == 0) { + if (num > 0) { + int64_t in_axis_dim = x_shape[axis]; + if (in_axis_dim % num != 0) { + VLOG(3) << "Invalid number to split. Tensor split does not result" + " in an equal division of dimensions. Axis dim = " + << in_axis_dim << " num = " << num << "!= 0"; + return false; + } + size_t out_axis_dim = in_axis_dim / num; + for (int i = 0; i < num; ++i) { + output_lengths.push_back(out_axis_dim); + } } } + if (output_lengths.size() != output_num) { + VLOG(3) << "The output_length should be equal to the output size."; + return false; + } } if (op_type == "scale") { auto scale_inputs = desc.Inputs(); diff --git a/python/paddle/fluid/tests/unittests/ir/inference/test_trt_convert_split.py b/python/paddle/fluid/tests/unittests/ir/inference/test_trt_convert_split.py new file mode 100644 index 0000000000000000000000000000000000000000..2db60ccc61b950c6abcdcc71d271e5a9fb275d83 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/ir/inference/test_trt_convert_split.py @@ -0,0 +1,235 @@ +# 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 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: + inputs = program_config.inputs + weights = program_config.weights + outputs = program_config.outputs + + attrs = [ + program_config.ops[i].attrs + for i in range(len(program_config.ops)) + ] + # the dimensions of input and axis match + if len(inputs['split_input'].shape) <= attrs[0]['axis']: + return False + + #Sections and num cannot both be equal to 0. + if len(attrs[0]['sections']) == 0: + if attrs[0]['num'] == 0: + return False + + #When sections and num are not both equal to 0, sections has higher priority. + #The sum of sections should be equal to the input size. + if len(attrs[0]['sections']) != 0: + if attrs[0]['num'] != 0: + return False + if len(outputs) != len(attrs[0]['sections']): + return False + sum = 0 + for num in attrs[0]['sections']: + sum += num + if sum != inputs['split_input'].shape[attrs[0]['axis']]: + return False + + #The size of num should be equal to the input dimension. + if attrs[0]['num'] != 0: + if len(outputs) != attrs[0]['num']: + return False + + #Test AxisTensor and SectionsTensorList + if self.num_input == 0: + if self.dims == 2 and attrs[0]['sections'] == [10, 14] and len( + outputs) == 2: + return True + else: + return False + + return True + + def sample_program_configs(self): + def generate_input1(attrs: List[Dict[str, Any]], batch): + if self.dims == 4: + return np.ones([batch, 3, 3, 24]).astype(np.float32) + elif self.dims == 3: + return np.ones([batch, 3, 24]).astype(np.float32) + elif self.dims == 2: + return np.ones([batch, 24]).astype(np.float32) + elif self.dims == 1: + return np.ones([24]).astype(np.float32) + + def generate_AxisTensor(attrs: List[Dict[str, Any]]): + return np.ones([1]).astype(np.int32) + + def generate_SectionsTensorList1(attrs: List[Dict[str, Any]]): + return np.array([10]).astype(np.int32) + + def generate_SectionsTensorList2(attrs: List[Dict[str, Any]]): + return np.array([14]).astype(np.int32) + + for num_input in [0, 1]: + for dims in [1, 2, 3, 4]: + for batch in [3, 6, 9]: + for Out in [["output_var0", "output_var1"], + ["output_var0", "output_var1", "output_var2"]]: + for sections in [[], [1, 2], [2, 1], [10, 14], + [1, 1, 1], [2, 2, 2], [3, 3, 3], + [3, 7, 14]]: + for num in [0, 3]: + for axis in [0, 1, 2, 3]: + self.batch = batch + self.num_input = num_input + self.dims = dims + dics = [{ + "sections": sections, + "num": num, + "axis": axis + }, {}] + + dics_intput = [{ + "X": ["split_input"], + "AxisTensor": ["AxisTensor"], + "SectionsTensorList": [ + "SectionsTensorList1", + "SectionsTensorList2" + ] + }, { + "X": ["split_input"] + }] + dics_intputs = [{ + "AxisTensor": + TensorConfig(data_gen=partial( + generate_AxisTensor, dics)), + "SectionsTensorList1": TensorConfig( + data_gen=partial( + generate_SectionsTensorList1, + dics)), + "SectionsTensorList2": + TensorConfig(data_gen=partial( + generate_SectionsTensorList2, dics)) + }, {}] + + ops_config = [{ + "op_type": "split", + "op_inputs": dics_intput[num_input], + "op_outputs": { + "Out": Out + }, + "op_attrs": dics[0] + }] + ops = self.generate_op_config(ops_config) + program_config = ProgramConfig( + ops=ops, + weights=dics_intputs[num_input], + inputs={ + "split_input": + TensorConfig(data_gen=partial( + generate_input1, dics, batch)) + }, + outputs=Out) + + yield program_config + + def sample_predictor_configs( + self, program_config) -> (paddle_infer.Config, List[int], float): + def generate_dynamic_shape(attrs): + if self.dims == 4: + self.dynamic_shape.min_input_shape = { + "split_input": [1, 3, 3, 24] + } + self.dynamic_shape.max_input_shape = { + "split_input": [9, 3, 3, 24] + } + self.dynamic_shape.opt_input_shape = { + "split_input": [1, 3, 3, 24] + } + elif self.dims == 3: + self.dynamic_shape.min_input_shape = {"split_input": [1, 3, 24]} + self.dynamic_shape.max_input_shape = {"split_input": [9, 3, 24]} + self.dynamic_shape.opt_input_shape = {"split_input": [1, 3, 24]} + elif self.dims == 2: + self.dynamic_shape.min_input_shape = {"split_input": [1, 24]} + self.dynamic_shape.max_input_shape = {"split_input": [9, 24]} + self.dynamic_shape.opt_input_shape = {"split_input": [1, 24]} + elif self.dims == 1: + self.dynamic_shape.min_input_shape = {"split_input": [24]} + self.dynamic_shape.max_input_shape = {"split_input": [24]} + self.dynamic_shape.opt_input_shape = {"split_input": [24]} + + 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 len(program_config.outputs) == 2: + if attrs[0]['axis'] != 0: + return 1, 3 + else: + return 0, 4 + else: + if attrs[0]['axis'] != 0: + return 1, 4 + else: + return 0, 5 + + attrs = [ + program_config.ops[i].attrs + for i in range(len(program_config.ops)) + ] + self.trt_param.max_batch_size = 9 + # 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) + 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): + def teller1(program_config, predictor_config): + if len(program_config.weights) == 3: + return True + return False + + self.add_skip_case( + teller1, SkipReasons.TRT_NOT_SUPPORT, + "INPUT AxisTensor AND SectionsTensorList NOT SUPPORT.") + + def test(self): + self.add_skip_trt_case() + self.run_test() + + +if __name__ == "__main__": + unittest.main()