[Paddle-TRT] reshape fill_constant

paddle/fluid/inference/api/analysis_predictor.cc

@@ -2089,6 +2089,7 @@ USE_TRT_CONVERTER(top_k)
 USE_TRT_CONVERTER(top_k_v2)
 USE_TRT_CONVERTER(squeeze2)
 USE_TRT_CONVERTER(unsqueeze2)
+USE_TRT_CONVERTER(fill_constant)
 #if PADDLE_WITH_CUSPARSELT && IS_TRT_VERSION_GE(8000)
 USE_TRT_CONVERTER(sparse_fc)
 USE_TRT_CONVERTER(sparse_multihead_matmul)

paddle/fluid/inference/tensorrt/convert/CMakeLists.txt

@@ -68,7 +68,8 @@ list(
   c_allreduce_op.cc
   top_k_op.cc
   squeeze2_op.cc
-  unsqueeze2_op.cc)
+  unsqueeze2_op.cc
+  fill_constant_op.cc)
 
 if(CUSPARSELT_FOUND AND ${TENSORRT_MAJOR_VERSION} GREATER_EQUAL 8)
   list(APPEND CONVERT_FILES sparse_fc_op.cc sparse_multihead_matmul_op.cc)

paddle/fluid/inference/tensorrt/convert/fill_constant_op.cc

@@ -0,0 +1,71 @@
+/* Copyright (c) 2018 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 inference {
+namespace tensorrt {
+
+class FillConstantOpConverter : public OpConverter {
+ public:
+  void operator()(const framework::proto::OpDesc& op,
+                  const framework::Scope& scope,
+                  bool test_mode) override {
+    VLOG(4)
+        << "convert a fluid fill_constant op to tensorrt fill_constant layer";
+
+    framework::OpDesc op_desc(op, nullptr);
+    int dtype = BOOST_GET_CONST(int, op_desc.GetAttr("dtype"));
+    std::string str_value =
+        BOOST_GET_CONST(std::string, op_desc.GetAttr("str_value"));
+    std::vector<int64_t> shape =
+        BOOST_GET_CONST(std::vector<int64_t>, op_desc.GetAttr("shape"));
+    std::unique_ptr<framework::Tensor> out_tensor(new framework::Tensor());
+    out_tensor->Resize(phi::make_ddim(shape));
+    nvinfer1::DataType trt_dtype = nvinfer1::DataType::kFLOAT;
+    void* trt_data = nullptr;
+    size_t trt_num;
+    if (dtype == 2 || dtype == 3) {  // int, int64
+      auto* tmp_ptr = out_tensor->mutable_data<int>(platform::CPUPlace());
+      for (int64_t i = 0; i < out_tensor->numel(); i++)
+        tmp_ptr[i] = std::stoi(str_value);
+      trt_dtype = nvinfer1::DataType::kINT32;
+      trt_data = static_cast<void*>(tmp_ptr);
+    } else if (dtype == 5) {  // float
+      auto* tmp_ptr = out_tensor->mutable_data<float>(platform::CPUPlace());
+      for (int64_t i = 0; i < out_tensor->numel(); i++)
+        tmp_ptr[i] = std::stof(str_value);
+      trt_data = static_cast<void*>(tmp_ptr);
+    }
+
+    trt_num = static_cast<size_t>(out_tensor->numel());
+    engine_->SetWeights("fill_constant_value", std::move(out_tensor));
+    TensorRTEngine::Weight weight{trt_dtype, trt_data, trt_num};
+
+    nvinfer1::Dims trt_in_shape;
+    trt_in_shape.nbDims = shape.size();
+    for (size_t i = 0; i < shape.size(); i++) trt_in_shape.d[i] = shape[i];
+    nvinfer1::ILayer* layer =
+        TRT_ENGINE_ADD_LAYER(engine_, Constant, trt_in_shape, weight.get());
+    auto output_name = op_desc.Output("Out")[0];
+    RreplenishLayerAndOutput(layer, "fill_constant", {output_name}, test_mode);
+  }
+};
+
+}  // namespace tensorrt
+}  // namespace inference
+}  // namespace paddle
+
+REGISTER_TRT_OP_CONVERTER(fill_constant, FillConstantOpConverter);

paddle/fluid/inference/tensorrt/convert/reshape_op.cc

@@ -35,14 +35,29 @@ class ReshapeOpConverter : public OpConverter {
     framework::OpDesc op_desc(op, nullptr);
     // Declare inputs
     auto* input = engine_->GetITensor(op_desc.Input("X")[0]);
+
     std::vector<int> shape =
         BOOST_GET_CONST(std::vector<int>, op_desc.GetAttr("shape"));
     int nbDims_num = shape.size();
     nvinfer1::Dims reshape_dim;
-    if (engine_->with_dynamic_shape()) {  // running the TRT Dynamic Shape mode
-      reshape_dim.nbDims = nbDims_num;
-      for (int i = 0; i < nbDims_num; ++i) {
-        reshape_dim.d[i] = shape[i];
+    nvinfer1::ITensor* real_shape_tensor = nullptr;
+    std::vector<nvinfer1::ITensor*> concat_inputs;
+    bool one_input = false;
+    if (engine_->with_dynamic_shape()) {
+      if (op_desc.Inputs().find("ShapeTensor") != op_desc.Inputs().end() &&
+          op_desc.Input("ShapeTensor").size() > 0) {
+        for (auto name : op_desc.Input("ShapeTensor"))
+          concat_inputs.push_back(engine_->GetITensor(name));
+        real_shape_tensor = Concat(concat_inputs);
+      } else if (op_desc.Inputs().find("Shape") != op_desc.Inputs().end() &&
+                 op_desc.Input("Shape").size() > 0) {
+        real_shape_tensor = engine_->GetITensor(op_desc.Input("Shape")[0]);
+      } else {
+        reshape_dim.nbDims = nbDims_num;
+        for (int i = 0; i < nbDims_num; ++i) {
+          reshape_dim.d[i] = shape[i];
+        }
+        one_input = true;
       }
     } else {  // running the TRT Static Shape mode
       reshape_dim.nbDims = nbDims_num - 1;
@@ -51,7 +66,10 @@ class ReshapeOpConverter : public OpConverter {
       }
     }
     auto* layer = TRT_ENGINE_ADD_LAYER(engine_, Shuffle, *input);
-    layer->setReshapeDimensions(reshape_dim);
+    if (!engine_->with_dynamic_shape() || one_input)
+      layer->setReshapeDimensions(reshape_dim);
+    else
+      layer->setInput(1, *real_shape_tensor);
     auto output_name = op_desc.Output("Out")[0];
     RreplenishLayerAndOutput(layer, "reshape", {output_name}, test_mode);
   }

paddle/fluid/inference/tensorrt/op_teller.cc

@@ -169,6 +169,7 @@ struct SimpleOpTypeSetTeller : public Teller {
       "transformer_input_convert",
       "recover_padding",
       "remove_padding",
+      "fill_constant",
       "squeeze2",
       "unsqueeze2"};
   std::unordered_set<std::string> teller_set{
@@ -274,6 +275,7 @@ struct SimpleOpTypeSetTeller : public Teller {
       "transformer_input_convert",
       "recover_padding",
       "remove_padding",
+      "fill_constant",
       "squeeze2",
       "unsqueeze2"};
 };
@@ -1447,6 +1449,27 @@ bool OpTeller::Tell(const framework::ir::Node* node,
       }
     }
 
+    if (op_type == "fill_constant") {
+      auto fill_constant_inputs = desc.Inputs();
+      if (fill_constant_inputs.find("ValueTensor") !=
+          fill_constant_inputs.end()) {
+        if (desc.Input("ValueTensor").size()) return false;
+      }
+      if (fill_constant_inputs.find("ShapeTensor") !=
+          fill_constant_inputs.end()) {
+        if (desc.Input("ShapeTensor").size()) return false;
+      }
+      if (fill_constant_inputs.find("ShapeTensorList") !=
+          fill_constant_inputs.end()) {
+        if (desc.Input("ShapeTensorList").size()) return false;
+      }
+      int dtype = BOOST_GET_CONST(int, desc.GetAttr("dtype"));
+      // only support int32, int64, float32
+      if (!(dtype == 2 || dtype == 3 || dtype == 5)) {
+        return false;
+      }
+    }
+
     if (op_type == "instance_norm") {
       if (with_dynamic_shape) {
         VLOG(3) << "trt instance_norm op does not support dynamic shape ";
@@ -1800,6 +1823,9 @@ bool OpTeller::Tell(const framework::ir::Node* node,
     }
 
     if (op_type == "reshape" || op_type == "reshape2") {
+      if (with_dynamic_shape) {
+        return true;
+      }
       if (!desc.HasAttr("shape")) {
         return false;
       }

python/paddle/fluid/tests/unittests/ir/inference/test_trt_convert_fill_constant.py

@@ -0,0 +1,142 @@
+# 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 TrtConvertSplitTest(TrtLayerAutoScanTest):
+
+    def is_program_valid(self, program_config: ProgramConfig) -> bool:
+        return True
+
+    def sample_program_configs(self):
+
+        def generate_value_data(attrs: List[Dict[str, Any]]):
+            return np.array([1]).astype(np.int32)
+
+        def generate_shape_data(attrs: List[Dict[str, Any]]):
+            return np.array([4, 23]).astype(np.int32)
+
+        def generate_shapelist_data(attrs: List[Dict[str, Any]]):
+            return np.array([4]).astype(np.int32)
+
+        for shape in [[2, 3, 4]]:
+            for num_input in [0, 1, 2, 3]:
+                for dtype in [5, 2, 3]:
+                    for str_value in ["2", "23", "-1"]:
+                        self.num_input = num_input
+                        dics = [{
+                            "str_value": str_value,
+                            "shape": shape,
+                            "dtype": dtype
+                        }, {
+                            "axis": -1
+                        }]
+                        dics_intput = [{
+                            "ValueTensor": ["value_data"]
+                        }, {
+                            "ShapeTensor": ["shape_data"],
+                        }, {
+                            "ShapeTensorList": ["shapeT1_data", "shapeT2_data"],
+                        }, {}]
+                        ops_config = [
+                            {
+                                "op_type": "fill_constant",
+                                "op_inputs": dics_intput[num_input],
+                                "op_outputs": {
+                                    "Out": ["out_data"],
+                                },
+                                "op_attrs": dics[0]
+                            },
+                        ]
+
+                        def generate_input():
+                            return np.random.random([1, 1]).astype(np.float32)
+
+                        ops = self.generate_op_config(ops_config)
+                        program_config = ProgramConfig(
+                            ops=ops,
+                            weights={},
+                            inputs={
+                                "value_data":
+                                TensorConfig(data_gen=partial(
+                                    generate_value_data, dics)),
+                                "shape_data":
+                                TensorConfig(data_gen=partial(
+                                    generate_shape_data, dics)),
+                                "shapeT1_data":
+                                TensorConfig(data_gen=partial(
+                                    generate_shapelist_data, dics)),
+                                "shapeT2_data":
+                                TensorConfig(data_gen=partial(
+                                    generate_shapelist_data, dics)),
+                            },
+                            outputs=["out_data"])
+
+                        yield program_config
+
+    def sample_predictor_configs(
+            self, program_config) -> (paddle_infer.Config, List[int], float):
+
+        def generate_dynamic_shape(attrs):
+            self.input_shape = [1, 1]
+            max_shape = list(self.input_shape)
+            min_shape = list(self.input_shape)
+            opt_shape = list(self.input_shape)
+            for i in range(len(self.input_shape)):
+                max_shape[i] = max_shape[i] + 1
+            self.dynamic_shape.min_input_shape = {"Y_data": min_shape}
+            self.dynamic_shape.max_input_shape = {"Y_data": max_shape}
+            self.dynamic_shape.opt_input_shape = {"Y_data": opt_shape}
+
+        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 (self.num_input < 3):
+                return 0, 6
+            return 1, 5
+
+        attrs = [
+            program_config.ops[i].attrs for i in range(len(program_config.ops))
+        ]
+        # Don't test static shape
+
+        # 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):
+        pass
+
+    def test(self):
+        self.add_skip_trt_case()
+        self.run_test()
+
+
+if __name__ == "__main__":
+    unittest.main()