diff --git a/paddle/fluid/inference/tensorrt/convert/prelu_op.cc b/paddle/fluid/inference/tensorrt/convert/prelu_op.cc index a8a36e1238168ad368a02bf2ebed915939c3d5c1..94f5708e03c1245945e9f57c94931d7847a1c902 100644 --- a/paddle/fluid/inference/tensorrt/convert/prelu_op.cc +++ b/paddle/fluid/inference/tensorrt/convert/prelu_op.cc @@ -34,11 +34,7 @@ class PReluOpConverter : public OpConverter { auto* input = engine_->GetITensor(op_desc.Input("X")[0]); // Get attrs std::string mode = BOOST_GET_CONST(std::string, op_desc.GetAttr("mode")); - // auto* alpha_var = scope.FindVar(op_desc.Input("Alpha")[0]); - PADDLE_ENFORCE_NOT_NULL( - alpha_var, platform::errors::NotFound( - "Variable Alpha of prelu TRT converter is not found.")); auto* alpha_tensor = alpha_var->GetMutable(); platform::CPUPlace cpu_place; @@ -50,15 +46,9 @@ class PReluOpConverter : public OpConverter { nvinfer1::ILayer* layer = nullptr; if (engine_->with_dynamic_shape()) { -#if IS_TRT_VERSION_GE(6000) plugin::PReluPluginDynamic* plugin = new plugin::PReluPluginDynamic( alpha_data, alpha_tensor_temp->numel(), mode); 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 { #if IS_TRT_VERSION_GE(7000) float* alpha_weight_data = engine_->GetWeightCPUData( diff --git a/paddle/fluid/inference/tensorrt/op_teller.cc b/paddle/fluid/inference/tensorrt/op_teller.cc index 2104d38ebc59f851f8eb85ceadc88662924b54f5..95fdb09cb6807e26a9e1781d0b8d36ac142c4640 100644 --- a/paddle/fluid/inference/tensorrt/op_teller.cc +++ b/paddle/fluid/inference/tensorrt/op_teller.cc @@ -661,6 +661,28 @@ bool OpTeller::Tell(const framework::ir::Node* node, bool use_no_calib_int8, << desc.Output("Out").size() << "."; return false; } + + auto* block = desc.Block(); + auto* var_desc = block->FindVar(desc.Input("Alpha")[0]); + if (!var_desc) { + VLOG(3) << "Variable Alpha of prelu TRT converter not found."; + return false; + } + + 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(); + if (x_shape.size() == 1) { + VLOG(3) << "prelu op does not support input's dim is 1 in tensorrt."; + return false; + } + + if (!with_dynamic_shape) { + if (x_shape.size() == 2) { + VLOG(3) << "prelu op does not support input's dim is 2 in tensorrt."; + return false; + } + } } if (op_type == "roi_align") { diff --git a/python/paddle/fluid/tests/unittests/ir/inference/CMakeLists.txt b/python/paddle/fluid/tests/unittests/ir/inference/CMakeLists.txt index 1757ad9d943cb17f3782fcaee5a4490479f0fa4e..54229533935a42ece7bfa0113d52f56846e8a5ba 100644 --- a/python/paddle/fluid/tests/unittests/ir/inference/CMakeLists.txt +++ b/python/paddle/fluid/tests/unittests/ir/inference/CMakeLists.txt @@ -31,7 +31,7 @@ if(WITH_GPU AND TENSORRT_FOUND) foreach(target ${TEST_TRT_CONVERTER}) py_test_modules(${target} MODULES ${target}) - set_tests_properties(${target} PROPERTIES TIMEOUT 100) + set_tests_properties(${target} PROPERTIES TIMEOUT 300) endforeach() endif() diff --git a/python/paddle/fluid/tests/unittests/ir/inference/test_trt_convert_prelu.py b/python/paddle/fluid/tests/unittests/ir/inference/test_trt_convert_prelu.py new file mode 100644 index 0000000000000000000000000000000000000000..4122e2623cb5a7c90bac8fac0914c38e0d5d475e --- /dev/null +++ b/python/paddle/fluid/tests/unittests/ir/inference/test_trt_convert_prelu.py @@ -0,0 +1,195 @@ +# 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 TrtConvertPreluTest(TrtLayerAutoScanTest): + def is_program_valid(self, program_config: ProgramConfig) -> bool: + return True + + def sample_program_configs(self): + def generate_input(batch, dim1, dim2, dim3): + shape = [batch] + if dim1 != 0: + shape.append(dim1) + if dim2 != 0: + shape.append(dim2) + if dim3 != 0: + shape.append(dim3) + return np.random.random(shape).astype(np.float32) + + def generate_alpha(attrs: List[Dict[str, Any]], dim1, dim2, dim3): + if attrs[0]["mode"] == "all": + return np.random.random(size=(1)).astype(np.float32) + elif attrs[0]["mode"] == "channel": + shape = [1] + if dim1 != 0: + shape.append(dim1) + if dim2 != 0: + shape.append(1) + if dim3 != 0: + shape.append(1) + return np.random.random(size=shape).astype(np.float32) + elif attrs[0]["mode"] == "element": + shape = [1] + if dim1 != 0: + shape.append(dim1) + if dim2 != 0: + shape.append(dim2) + if dim3 != 0: + shape.append(dim3) + return np.random.random(size=shape).astype(np.float32) + + for batch in [1, 4]: + for dim1 in [0, 3]: + for dim2 in [0, 16]: + for dim3 in [0, 32]: + self.dim1 = dim1 + self.dim2 = dim2 + self.dim3 = dim3 + + if dim1 == 0 and dim2 != 0: + continue + if dim1 == 0 and dim2 == 0 and dim3 != 0: + continue + + for mode in ["all", "channel", "element"]: + if mode == "channel" and dim1 == 0: + continue + dics = [{"mode": mode}] + ops_config = [{ + "op_type": "prelu", + "op_inputs": { + "X": ["input_data"], + "Alpha": ["alpha_weight"] + }, + "op_outputs": { + "Out": ["output_data"] + }, + "op_attrs": dics[0] + }] + ops = self.generate_op_config(ops_config) + + program_config = ProgramConfig( + ops=ops, + weights={ + "alpha_weight": TensorConfig( + data_gen=partial(generate_alpha, dics, + dim1, dim2, dim3)) + }, + inputs={ + "input_data": TensorConfig( + data_gen=partial(generate_input, batch, + dim1, dim2, dim3)), + }, + outputs=["output_data"]) + + yield program_config + + def sample_predictor_configs( + self, program_config) -> (paddle_infer.Config, List[int], float): + def generate_dynamic_shape(attrs): + if self.dim1 == 0: + self.dynamic_shape.min_input_shape = {"input_data": [1], } + self.dynamic_shape.max_input_shape = {"input_data": [4], } + self.dynamic_shape.opt_input_shape = {"input_data": [2], } + else: + if self.dim2 == 0 and self.dim3 == 0: + self.dynamic_shape.min_input_shape = { + "input_data": [1, 1], + } + self.dynamic_shape.max_input_shape = { + "input_data": [4, 64], + } + self.dynamic_shape.opt_input_shape = { + "input_data": [2, 3], + } + elif self.dim2 != 0 and self.dim3 != 0: + self.dynamic_shape.min_input_shape = { + "input_data": [1, 1, 1, 1], + } + self.dynamic_shape.max_input_shape = { + "input_data": [4, 64, 128, 128], + } + self.dynamic_shape.opt_input_shape = { + "input_data": [2, 3, 16, 32], + } + elif self.dim3 == 0: + self.dynamic_shape.min_input_shape = { + "input_data": [1, 1, 1], + } + self.dynamic_shape.max_input_shape = { + "input_data": [4, 64, 256], + } + self.dynamic_shape.opt_input_shape = { + "input_data": [2, 3, 128], + } + + def clear_dynamic_shape(): + self.dynamic_shape.max_input_shape = {} + self.dynamic_shape.min_input_shape = {} + self.dynamic_shape.opt_input_shape = {} + + attrs = [ + program_config.ops[i].attrs + for i in range(len(program_config.ops)) + ] + + # for static_shape + clear_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 + + # for dynamic_shape + generate_dynamic_shape(attrs) + 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): + def teller1(program_config, predictor_config): + if self.dim1 == 0 and self.dim2 == 0 and self.dim3 == 0: + return True + return False + + self.add_skip_case(teller1, SkipReasons.TRT_NOT_SUPPORT, + "Trt does not support 1-dimensional input.") + + def teller2(program_config, predictor_config): + if (len(self.dynamic_shape.min_input_shape) == 0): + if self.dim1 != 0 and self.dim2 == 0 and self.dim3 == 0: + return True + return False + + self.add_skip_case( + teller2, SkipReasons.TRT_NOT_SUPPORT, + "Need to repair the case: the output of GPU and tensorrt has diff when the input dimension is 2 in static shape mode." + ) + + def test(self): + self.add_skip_trt_case() + self.run_test() + + +if __name__ == "__main__": + unittest.main()