diff --git a/paddle/fluid/inference/tensorrt/convert/activation_op.cc b/paddle/fluid/inference/tensorrt/convert/activation_op.cc index 9244b9af0bbd6cfc392b1b940d81c04b0dd0cde9..e6a0ecf4aececcba012923f631b2dcfd8f69743d 100644 --- a/paddle/fluid/inference/tensorrt/convert/activation_op.cc +++ b/paddle/fluid/inference/tensorrt/convert/activation_op.cc @@ -52,11 +52,6 @@ class ActivationOpConverter : public OpConverter { engine_->GetITensor(op_desc.Input("X")[0]); auto op_pair = ops.find(op_type_); - if (op_pair == ops.end()) { - PADDLE_THROW(platform::errors::Fatal( - "Wrong activation op type, the trt do not support the %s act type.", - op_type_)); - } nvinfer1::IActivationLayer* layer = TRT_ENGINE_ADD_LAYER( engine_, Activation, *const_cast(input_tensor), diff --git a/paddle/fluid/inference/tensorrt/convert/affine_channel_op.cc b/paddle/fluid/inference/tensorrt/convert/affine_channel_op.cc index 813342c08483b7e9124929d3f00d8155d337e67e..eba67c3c098ca60b7608ecf6db50b46e233955a5 100644 --- a/paddle/fluid/inference/tensorrt/convert/affine_channel_op.cc +++ b/paddle/fluid/inference/tensorrt/convert/affine_channel_op.cc @@ -55,16 +55,6 @@ class AffineChannelOpConverter : public OpConverter { auto* bias_t = bias_v->GetMutable(); float* bias_ptr = engine_->GetWeightCPUData(bias_name, bias_t, false); - auto data_layout = framework::StringToDataLayout( - BOOST_GET_CONST(std::string, op_desc.GetAttr("data_layout"))); - - PADDLE_ENFORCE_EQ( - data_layout, framework::DataLayout::kNCHW, - platform::errors::InvalidArgument( - "TensorRT affine channel converter can only convert NCHW format. " - "Other format should be run in fluid mode. Report a bug on github " - "issue if you see this line.")); - // tensorrt scalend layer only support spatial dims >= 2, // so nhwc is not availabe (spatial dims == 0) const int channel_axis = engine_->with_dynamic_shape(); diff --git a/paddle/fluid/inference/tensorrt/convert/elementwise_op.cc b/paddle/fluid/inference/tensorrt/convert/elementwise_op.cc index 19d79510547ecc5e46803a0830b74cde1b94158e..5419933e4073673f56c72d06c49f488167421dbe 100644 --- a/paddle/fluid/inference/tensorrt/convert/elementwise_op.cc +++ b/paddle/fluid/inference/tensorrt/convert/elementwise_op.cc @@ -25,10 +25,6 @@ static bool CheckDims(const nvinfer1::Dims& dims_x, return false; } for (int i = 0; i < dims_x.nbDims; i++) { - // conservative judgment - if (dims_x.d[i] == -1 || dims_y.d[i] == -1) { - return false; - } if (dims_x.d[i] != dims_y.d[i]) { return false; } diff --git a/paddle/fluid/inference/tensorrt/op_teller.cc b/paddle/fluid/inference/tensorrt/op_teller.cc index 48c7b7fdd0d79dbacd705896aef1c12ac15ccf42..6db81cefb46a1bb85548073e07267c37d76b0a44 100644 --- a/paddle/fluid/inference/tensorrt/op_teller.cc +++ b/paddle/fluid/inference/tensorrt/op_teller.cc @@ -225,6 +225,27 @@ bool OpTeller::Tell(const framework::ir::Node* node, bool use_no_calib_int8, << desc.Output("Output").size() << " output."; return false; } + +// strides > 1 and 'SAME' is only supported by trt7.0 above +#if !IS_TRT_VERSION_GE(7000) + if (op_type == "conv2d" || op_type == "conv2d_fusion" || + op_type == "depthwise_conv2d") { + if (desc.HasAttr("padding_algorithm") && with_dynamic_shape) { + auto padding_algorithm = + BOOST_GET_CONST(std::string, desc.GetAttr("padding_algorithm")); + if (padding_algorithm == "SAME" && desc.HasAttr("strides")) { + const std::vector strides = + BOOST_GET_CONST(std::vector, desc.GetAttr("strides")); + // there is no issue if strides.size() less than 2 + if (strides.size() > 1) { + for (size_t i = 0; i < strides.size(); i++) { + if (strides[i] > 1) return false; + } + } + } + } + } +#endif } if (op_type == "matmul") { diff --git a/python/paddle/fluid/tests/unittests/ir/inference/test_trt_conv_pass.py b/python/paddle/fluid/tests/unittests/ir/inference/test_trt_conv_pass.py index ec3955a9ae1441cdaa4efa5b0e87ff8b74a0b689..7f613c4765963da8e02d30aa7cf35553fc4d31fa 100644 --- a/python/paddle/fluid/tests/unittests/ir/inference/test_trt_conv_pass.py +++ b/python/paddle/fluid/tests/unittests/ir/inference/test_trt_conv_pass.py @@ -161,5 +161,70 @@ class TensorRTSubgraphPassDepthwiseConvTransposeTest( self.use_cudnn = False +class DynamicShapeTensorRTSubgraphPassConvTest(InferencePassTest): + def setUp(self): + self.set_params() + with fluid.program_guard(self.main_program, self.startup_program): + data = fluid.data( + name="data", shape=[-1, 6, -1, -1], dtype="float32") + conv_out = fluid.layers.conv2d( + input=data, + num_filters=self.conv_num_filters, + filter_size=self.conv_filter_size, + groups=self.conv_groups, + padding=self.conv_padding, + bias_attr=False, + use_cudnn=self.use_cudnn, + stride=self.stride, + act=None) + self.feeds = { + "data": np.random.random([32, 6, 64, 64]).astype("float32"), + } + self.enable_trt = True + self.trt_parameters = DynamicShapeTensorRTSubgraphPassConvTest.TensorRTParam( + 1 << 30, 32, 0, AnalysisConfig.Precision.Float32, False, False) + self.dynamic_shape_params = DynamicShapeTensorRTSubgraphPassConvTest.DynamicShapeParam( + { + "conv2d_0.tmp_0": [1, 6, 8, 8], + "data": [1, 6, 8, 8], + "depthwise_conv2d_0.tmp_0": [1, 6, 8, 8] + }, { + "conv2d_0.tmp_0": [32, 6, 64, 64], + "data": [32, 6, 64, 64], + "depthwise_conv2d_0.tmp_0": [32, 6, 64, 64] + }, { + "conv2d_0.tmp_0": [16, 6, 16, 16], + "data": [16, 6, 16, 16], + "depthwise_conv2d_0.tmp_0": [32, 6, 64, 64] + }, False) + self.fetch_list = [conv_out] + + def set_params(self): + self.conv_num_filters = 6 + self.conv_filter_size = 6 + self.conv_groups = 6 + self.conv_padding = 'SAME' + self.use_cudnn = True + self.stride = [2, 2] + + def test_check_output(self): + if core.is_compiled_with_cuda(): + use_gpu = True + self.check_output_with_option(use_gpu) + self.assertTrue( + PassVersionChecker.IsCompatible('tensorrt_subgraph_pass')) + + +class DynamicShapeTensorRTSubgraphPassDepthwiseConvTransposeTest( + DynamicShapeTensorRTSubgraphPassConvTest): + def set_params(self): + self.conv_num_filters = 6 + self.conv_filter_size = 6 + self.conv_groups = 6 + self.conv_padding = 'SAME' + self.use_cudnn = False + self.stride = [2, 2] + + if __name__ == "__main__": unittest.main()