add flatten_contiguous_range OpConvert for Paddle-TRT (#38922)

* add trt_convert_flatten_contiguous_rang op * trt version >7,support trt_convert_flatten_contiguous_rang * trt version >7,support trt_convert_flatten_contiguous_rang * trt version >7,support trt_convert_flatten_contiguous_rang * test cast add trt version >=7 skip
PaddlePaddle · Jan 14, 2022 · 050aa6f · 050aa6f
1 parent 87ee3e4
commit 050aa6f
Show file tree

Hide file tree

Showing 6 changed files with 290 additions and 3 deletions.
diff --git a/paddle/fluid/inference/analysis/ir_passes/tensorrt_subgraph_pass.cc b/paddle/fluid/inference/analysis/ir_passes/tensorrt_subgraph_pass.cc
@@ -46,8 +46,11 @@ void analysis::TensorRtSubgraphPass::ApplyImpl(
               << " is diabled by config in TensorRT";
       return false;
     }
-    return tensorrt::OpTeller::Global().Tell(node, no_calib_int8,
-                                             with_dynamic_shape);
+    bool is_ok = tensorrt::OpTeller::Global().Tell(node, no_calib_int8,
+                                                   with_dynamic_shape);
+    if (!is_ok)
+      VLOG(3) << node->Op()->Type().c_str() << " op is not in TensorRT";
+    return is_ok;
   };
 
   framework::ir::SubGraphFuser fuser(

diff --git a/paddle/fluid/inference/api/analysis_predictor.cc b/paddle/fluid/inference/api/analysis_predictor.cc
@@ -1416,6 +1416,7 @@ USE_TRT_CONVERTER(elementwise_min_tensor);
 USE_TRT_CONVERTER(elementwise_pow_tensor);
 USE_TRT_CONVERTER(transpose);
 USE_TRT_CONVERTER(flatten);
+USE_TRT_CONVERTER(flatten_contiguous_range);
 USE_TRT_CONVERTER(matmul);
 USE_TRT_CONVERTER(conv2d);
 USE_TRT_CONVERTER(relu);

diff --git a/paddle/fluid/inference/tensorrt/convert/CMakeLists.txt b/paddle/fluid/inference/tensorrt/convert/CMakeLists.txt
@@ -3,7 +3,7 @@ nv_library(tensorrt_converter
            SRCS matmul_op.cc conv2d_op.cc fc_op.cc pool2d_op.cc elementwise_op.cc
                 batch_norm_op.cc activation_op.cc softmax_op.cc concat_op.cc dropout_op.cc group_norm_op.cc
                 pad_op.cc split_op.cc prelu_op.cc leaky_relu_op.cc gelu_op.cc layer_norm_op.cc multihead_matmul_op.cc
-                shuffle_channel_op.cc swish_op.cc instance_norm_op.cc stack_op.cc transpose_op.cc flatten_op.cc
+                shuffle_channel_op.cc swish_op.cc instance_norm_op.cc stack_op.cc transpose_op.cc flatten_op.cc flatten_contiguous_range_op.cc
                 emb_eltwise_layernorm.cc skip_layernorm.cc scale_op.cc slice_op.cc hard_sigmoid_op.cc hard_swish_op.cc clip_op.cc
                 gather_op.cc
                 anchor_generator_op.cc

diff --git a/paddle/fluid/inference/tensorrt/convert/flatten_contiguous_range_op.cc b/paddle/fluid/inference/tensorrt/convert/flatten_contiguous_range_op.cc
@@ -0,0 +1,136 @@
+/* Copyright (c) 2022 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 framework {
+class Scope;
+namespace proto {
+class OpDesc;
+}  // namespace proto
+}  // namespace framework
+}  // namespace paddle
+
+namespace paddle {
+namespace inference {
+namespace tensorrt {
+/*
+ * flatten_contiguous_range trt converter
+ */
+class FlattenContiguousRangeOpConverter : public OpConverter {
+ public:
+  void operator()(const framework::proto::OpDesc& op,
+                  const framework::Scope& scope, bool test_mode) override {
+    framework::OpDesc op_desc(op, nullptr);
+    // Declare inputs
+    auto* input = engine_->GetITensor(op_desc.Input("X")[0]);
+    int dims = input->getDimensions().nbDims;
+    int start_axis = BOOST_GET_CONST(int, op_desc.GetAttr("start_axis"));
+    int stop_axis = BOOST_GET_CONST(int, op_desc.GetAttr("stop_axis"));
+
+    nvinfer1::IShuffleLayer* layer = nullptr;
+    if (!engine_->with_dynamic_shape()) {
+      if (start_axis < 0) start_axis += dims + 1;
+      if (stop_axis < 0) stop_axis += dims + 1;
+      int dim_prod = 1;
+      nvinfer1::Dims flatten_dim;
+      flatten_dim.nbDims = dims - (stop_axis - start_axis);
+      for (int i = 0, j = 0; i < dims; ++i) {
+        if (start_axis <= i + 1 && i + 1 <= stop_axis) {
+          int dim_i = input->getDimensions().d[i];
+          PADDLE_ENFORCE_GT(dim_i, 0, platform::errors::InvalidArgument(
+                                          "flatten_contiguous_range input dim "
+                                          "should be > 0, but got %d.",
+                                          dim_i));
+          dim_prod *= dim_i;
+          if (i + 1 == stop_axis) {
+            flatten_dim.d[j++] = dim_prod;
+          }
+        } else {
+          flatten_dim.d[j++] = input->getDimensions().d[i];
+        }
+      }
+      layer = TRT_ENGINE_ADD_LAYER(engine_, Shuffle, *input);
+      layer->setReshapeDimensions(flatten_dim);
+    } else {
+      if (start_axis < 0) start_axis += dims;
+      if (stop_axis < 0) stop_axis += dims;
+      auto* shape_layer = TRT_ENGINE_ADD_LAYER(engine_, Shape, *input);
+      auto* shape_layer_itensor = shape_layer->getOutput(0);
+
+      nvinfer1::Dims start_dim, size_dim, stride_dim;
+      start_dim.nbDims = 1;
+      size_dim.nbDims = 1;
+      stride_dim.nbDims = 1;
+      start_dim.d[0] = start_axis;
+      size_dim.d[0] = stop_axis - start_axis + 1;
+      stride_dim.d[0] = 1;
+      auto* slice_layer =
+          TRT_ENGINE_ADD_LAYER(engine_, Slice, *shape_layer_itensor, start_dim,
+                               size_dim, stride_dim);
+      uint32_t reduce_dim = 1;
+      auto* reduce_prod_layer = TRT_ENGINE_ADD_LAYER(
+          engine_, Reduce, *(slice_layer->getOutput(0)),
+          nvinfer1::ReduceOperation::kPROD, reduce_dim, true);
+
+      nvinfer1::ITensor* input_shape = nullptr;
+      if (start_axis == 0 && stop_axis == dims - 1) {
+        input_shape = reduce_prod_layer->getOutput(0);
+      } else {
+        std::vector<nvinfer1::ITensor*> itensors;
+        if (start_axis > 0) {
+          nvinfer1::Dims left_start_dim, left_size_dim, left_stride_dim;
+          left_start_dim.nbDims = 1;
+          left_size_dim.nbDims = 1;
+          left_stride_dim.nbDims = 1;
+          left_start_dim.d[0] = 0;
+          left_size_dim.d[0] = start_axis;
+          left_stride_dim.d[0] = 1;
+          auto* slice_layer_left = TRT_ENGINE_ADD_LAYER(
+              engine_, Slice, *shape_layer_itensor, left_start_dim,
+              left_size_dim, left_stride_dim);
+          itensors.push_back(slice_layer_left->getOutput(0));
+        }
+        itensors.push_back(reduce_prod_layer->getOutput(0));
+        if (stop_axis < dims - 1) {
+          nvinfer1::Dims right_start_dim, right_size_dim, right_stride_dim;
+          right_start_dim.nbDims = 1;
+          right_size_dim.nbDims = 1;
+          right_stride_dim.nbDims = 1;
+          right_start_dim.d[0] = stop_axis + 1;
+          right_size_dim.d[0] = dims - stop_axis - 1;
+          right_stride_dim.d[0] = 1;
+          auto* slice_layer_right = TRT_ENGINE_ADD_LAYER(
+              engine_, Slice, *shape_layer_itensor, right_start_dim,
+              right_size_dim, right_stride_dim);
+          itensors.push_back(slice_layer_right->getOutput(0));
+        }
+        auto* concat_layer = TRT_ENGINE_ADD_LAYER(
+            engine_, Concatenation, itensors.data(), itensors.size());
+        concat_layer->setAxis(0);
+        input_shape = concat_layer->getOutput(0);
+      }
+      layer = TRT_ENGINE_ADD_LAYER(engine_, Shuffle, *input);
+      layer->setInput(1, *input_shape);
+    }
+    auto output_name = op_desc.Output("Out")[0];
+    RreplenishLayerAndOutput(layer, "flatten_contiguous_range", {output_name},
+                             test_mode);
+  }
+};
+
+}  // namespace tensorrt
+}  // namespace inference
+}  // namespace paddle
+
+REGISTER_TRT_OP_CONVERTER(flatten_contiguous_range,
+                          FlattenContiguousRangeOpConverter);
diff --git a/paddle/fluid/inference/tensorrt/op_teller.cc b/paddle/fluid/inference/tensorrt/op_teller.cc
@@ -55,6 +55,7 @@ struct SimpleOpTypeSetTeller : public Teller {
 // #endif
 #if IS_TRT_VERSION_GE(7000)
     teller_set.insert("tile");
+    teller_set.insert("flatten_contiguous_range");
 #endif
 #if CUDA_VERSION >= 10020
     teller_set.insert("reshape");
@@ -531,6 +532,37 @@ bool OpTeller::Tell(const framework::ir::Node* node, bool use_no_calib_int8,
         if (axis != 1) return false;
       }
     }
+    if (op_type == "flatten_contiguous_range") {
+      if (!with_dynamic_shape) {
+        int start_axis = BOOST_GET_CONST(int, desc.GetAttr("start_axis"));
+        int stop_axis = BOOST_GET_CONST(int, desc.GetAttr("stop_axis"));
+        auto x_var_name = desc.Input("X")[0];
+        auto* block = desc.Block();
+        if (block == nullptr) {
+          VLOG(3) << "The block desc is nullptr, we can't continue to analyze. "
+                     "Developers need to check whether block_desc is passed in "
+                     "the pass.";
+          return false;
+        }
+        auto* x_var_desc = block->FindVar(x_var_name);
+        const auto x_shape = x_var_desc->GetShape();
+        int dims = x_shape.size();
+        if (start_axis < 0) start_axis += dims;
+        if (start_axis == 0) {
+          VLOG(3) << "TRT flatten_contiguous_range not support the "
+                     "batch-dimension being changed";
+          return false;
+        }
+        if (stop_axis < 0) stop_axis += dims;
+        for (int i = start_axis; i <= stop_axis; ++i) {
+          if (x_shape[i] < 0) {
+            VLOG(3) << "On TRT static shape,flatten_contiguous_range input dim "
+                       "should be > 0";
+            return false;
+          }
+        }
+      }
+    }
 
     if (op_type == "gather") {
       auto gather_inputs = desc.Inputs();

diff --git a/...on/paddle/fluid/tests/unittests/ir/inference/test_trt_convert_flatten_contiguous_range.py b/...on/paddle/fluid/tests/unittests/ir/inference/test_trt_convert_flatten_contiguous_range.py
@@ -0,0 +1,115 @@
+# 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 TrtConvertFlattenContiguousRangeTest(TrtLayerAutoScanTest):
+    def is_program_valid(self, program_config: ProgramConfig) -> bool:
+        return True
+
+    def sample_program_configs(self):
+        def generate_input(batch):
+            return np.random.random([2, batch, 4, 8, 3]).astype(np.float32)
+
+        for batch in [1, 2, 4]:
+            for start_axis in range(5):
+                for stop_axis in range(start_axis, 5):
+                    type = "flatten_contiguous_range"
+                    op_outputs = {
+                        "Out": ["output_data"],
+                        "XShape": ["xshape_data"]
+                    }
+                    ops_config = [{
+                        "op_type": type,
+                        "op_inputs": {
+                            "X": ["input_data"]
+                        },
+                        "op_outputs": op_outputs,
+                        "op_attrs": {
+                            "start_axis": start_axis,
+                            "stop_axis": stop_axis,
+                        }
+                    }]
+                    ops = self.generate_op_config(ops_config)
+
+                    program_config = ProgramConfig(
+                        ops=ops,
+                        weights={},
+                        inputs={
+                            "input_data": TensorConfig(
+                                data_gen=partial(generate_input, batch))
+                        },
+                        outputs=["output_data"])
+                    yield program_config
+
+    def sample_predictor_configs(
+            self, program_config) -> (paddle_infer.Config, List[int], float):
+        def generate_dynamic_shape(attrs):
+            self.dynamic_shape.min_input_shape = {"input_data": [2, 1, 4, 8, 3]}
+            self.dynamic_shape.max_input_shape = {"input_data": [2, 4, 4, 8, 3]}
+            self.dynamic_shape.opt_input_shape = {"input_data": [2, 2, 4, 8, 3]}
+
+        def clear_dynamic_shape():
+            self.dynamic_shape.max_input_shape = {}
+            self.dynamic_shape.min_input_shape = {}
+            self.dynamic_shape.opt_input_shape = {}
+
+        def generate_trt_nodes_num(attrs, dynamic_shape):
+            ver = paddle_infer.get_trt_compile_version()
+            if ver[0] * 1000 + ver[1] * 100 + ver[0] * 10 >= 7000:
+                if dynamic_shape:
+                    return 1, 2
+                else:
+                    if attrs[0]['start_axis'] == 0:
+                        return 0, 3
+                    else:
+                        return 1, 2
+            else:
+                return 0, 3
+
+        attrs = [
+            program_config.ops[i].attrs
+            for i in range(len(program_config.ops))
+        ]
+
+        # for static_shape
+        clear_dynamic_shape()
+        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 test(self):
+        self.run_test()
+
+
+if __name__ == "__main__":
+    unittest.main()