diff --git a/paddle/fluid/inference/tensorrt/convert/pool2d_op.cc b/paddle/fluid/inference/tensorrt/convert/pool2d_op.cc index 1898f28c73ad0b47d1c84fecb249c739c5259383..733a8f64ae5dbaa43ce7393818c40ecb4a72a0ef 100644 --- a/paddle/fluid/inference/tensorrt/convert/pool2d_op.cc +++ b/paddle/fluid/inference/tensorrt/convert/pool2d_op.cc @@ -107,6 +107,9 @@ class Pool2dOpConverter : public OpConverter { plugin_pool_type = plugin::PoolPlugin::PoolType::avg; } + if (padding_algorithm == "VALID") { + std::fill(paddings.begin(), paddings.end(), 0); + } nvinfer1::DimsHW nv_ksize(ksize[0], ksize[1]); nvinfer1::DimsHW nv_strides(strides[0], strides[1]); nvinfer1::DimsHW nv_paddings(paddings[0], paddings[1]); @@ -123,6 +126,30 @@ class Pool2dOpConverter : public OpConverter { if (engine_->with_dynamic_shape()) { if (!adaptive && !global_pooling && !ceil_mode) { + auto *pool_layer = TRT_ENGINE_ADD_LAYER(engine_, Pooling, *input1, + nv_pool_type, nv_ksize); + pool_layer->setStride(nv_strides); + pool_layer->setPadding(nv_paddings); + pool_layer->setAverageCountExcludesPadding(exclusive); + if (padding_algorithm == "SAME") { + pool_layer->setPaddingMode(nvinfer1::PaddingMode::kSAME_UPPER); + } + layer = pool_layer; + } else if (!adaptive && !global_pooling && ceil_mode) { + nvinfer1::DimsHW pre_pad(0, 0); + nvinfer1::DimsHW post_pad(0, 0); + // If ceil mode is true, we will pad the appropriate size to the input. + DealCeilMode(input_shape, ksize, strides, paddings, &pre_pad, &post_pad, + input_dims); + auto *pad_layer = TRT_ENGINE_ADD_LAYER( + engine_, Padding, *const_cast(input1), pre_pad, + post_pad); + PADDLE_ENFORCE_NOT_NULL( + pad_layer, platform::errors::Fatal( + "Pad layer in poolOp converter could not be " + "created. The pointer to pad layer is `NULL`.")); + input1 = pad_layer->getOutput(0); + auto *pool_layer = TRT_ENGINE_ADD_LAYER(engine_, Pooling, *input1, nv_pool_type, nv_ksize); pool_layer->setStride(nv_strides); diff --git a/paddle/fluid/inference/tensorrt/op_teller.cc b/paddle/fluid/inference/tensorrt/op_teller.cc index 7a70ceda60c1fb44ad82f90e512ef0ab15b10be9..ef50aee48e2eb84ed7950793c3d233250cf07ada 100644 --- a/paddle/fluid/inference/tensorrt/op_teller.cc +++ b/paddle/fluid/inference/tensorrt/op_teller.cc @@ -174,22 +174,8 @@ bool OpTeller::Tell(const framework::ir::Node* node, bool use_no_calib_int8, if (op_type == "pool2d") { std::vector paddings = BOOST_GET_CONST(std::vector, desc.GetAttr("paddings")); - if (paddings.size() > 2) return false; - if (desc.HasAttr("exclusive")) { - if (BOOST_GET_CONST(bool, desc.GetAttr("exclusive"))) { - std::vector ksize = - BOOST_GET_CONST(std::vector, desc.GetAttr("ksize")); - for (size_t i = 0; i < ksize.size(); i++) { - if (ksize[i] <= paddings[i]) { - VLOG(3) << "the padding size should be less than the filter size " - "for exclusive-counting pooling."; - return false; - } - } - } - } - if (desc.HasAttr("ceil_mode")) { - if (BOOST_GET_CONST(bool, desc.GetAttr("ceil_mode"))) return false; + if (paddings.size() > 2) { + return false; } if (desc.Input("X").size() != 1) { VLOG(3) << "TRT Pool2d expect 1 input, but got " @@ -211,15 +197,32 @@ bool OpTeller::Tell(const framework::ir::Node* node, bool use_no_calib_int8, << pool_type << " pool type."; return false; } + if (pool_type == "avg") { + if (desc.HasAttr("global_pooling")) { + if (!BOOST_GET_CONST(bool, desc.GetAttr("global_pooling"))) { + if (desc.HasAttr("exclusive")) { + if (BOOST_GET_CONST(bool, desc.GetAttr("exclusive"))) { + std::vector ksize = + BOOST_GET_CONST(std::vector, desc.GetAttr("ksize")); + for (size_t i = 0; i < ksize.size(); i++) { + if (ksize[i] <= paddings[i]) { + VLOG(3) << "the padding size should be less than the " + "filter size " + "for exclusive-counting pooling."; + return false; + } + } + } + } + } + } + } } } if (op_type == "conv2d" || op_type == "conv2d_transpose" || op_type == "conv2d_fusion" || op_type == "depthwise_conv2d" || op_type == "depthwise_conv2d_transpose") { - std::vector paddings = - BOOST_GET_CONST(std::vector, desc.GetAttr("paddings")); - if (desc.Input("Input").size() != 1) { VLOG(3) << "TRT Conv2d expect 1 input, but got " << desc.Input("Input").size() << " input."; diff --git a/python/paddle/fluid/tests/unittests/ir/inference/test_trt_convert_pool2d.py b/python/paddle/fluid/tests/unittests/ir/inference/test_trt_convert_pool2d.py index 3e923b1bd89d60c66d8450a75a827d7681283580..9ec2f83fa5ba0a204afa8bda703bd4eb94054f3d 100644 --- a/python/paddle/fluid/tests/unittests/ir/inference/test_trt_convert_pool2d.py +++ b/python/paddle/fluid/tests/unittests/ir/inference/test_trt_convert_pool2d.py @@ -21,9 +21,22 @@ from typing import Optional, List, Callable, Dict, Any, Set class TrtConvertPool2dTest(TrtLayerAutoScanTest): - def is_program_valid(self, program_config: ProgramConfig) -> bool: + def is_paddings_valid(self, program_config: ProgramConfig) -> bool: + exclusive = program_config.ops[0].attrs['exclusive'] + paddings = program_config.ops[0].attrs['paddings'] + ksize = program_config.ops[0].attrs['ksize'] + pooling_type = program_config.ops[0].attrs['pooling_type'] + global_pooling = program_config.ops[0].attrs['global_pooling'] + if global_pooling == False: + if pooling_type == 'avg': + for index in range(len(ksize)): + if ksize[index] <= paddings[index]: + return False return True + def is_program_valid(self, program_config: ProgramConfig) -> bool: + return self.is_paddings_valid(program_config) + def sample_program_configs(self): self.trt_param.workspace_size = 1073741824 @@ -34,7 +47,7 @@ class TrtConvertPool2dTest(TrtLayerAutoScanTest): return np.random.random([24, 3, 3, 3]).astype(np.float32) for strides in [[1, 1], [2, 2], [1, 2]]: - for paddings in [[0, 2], [0, 3], [1, 2, 3, 4]]: + for paddings in [[0, 2], [0, 3], [0, 1, 2, 3]]: for pooling_type in ['max', 'avg']: for padding_algotithm in ['EXPLICIT', 'SAME', 'VAILD']: for ksize in [[2, 3], [3, 3]]: @@ -43,7 +56,6 @@ class TrtConvertPool2dTest(TrtLayerAutoScanTest): for exclusive in [True, False]: for adaptive in [True, False]: for ceil_mode in [True, False]: - self.paddings = paddings dics = [{ "pooling_type": @@ -102,9 +114,6 @@ class TrtConvertPool2dTest(TrtLayerAutoScanTest): self.dynamic_shape.opt_input_shape = {} def generate_trt_nodes_num(attrs, dynamic_shape): - if self.paddings == [0, 3] or attrs[0][ - 'global_pooling'] == True or attrs[0]['ceil_mode'] == True: - return 0, 3 return 1, 2 attrs = [ @@ -139,6 +148,15 @@ class TrtConvertPool2dTest(TrtLayerAutoScanTest): self.add_skip_case(teller1, SkipReasons.TRT_NOT_IMPLEMENTED, "4-dims paddings are not support for trt now.") + def teller2(program_config, predictor_config): + if program_config.ops[0].attrs['global_pooling'] == True: + return True + return False + + self.add_skip_case( + teller2, SkipReasons.TRT_NOT_IMPLEMENTED, + "It is not support that global_pooling is true for trt now.") + def test(self): self.add_skip_trt_case() self.run_test()