[Inference] add squeeze2/unsqueeze2 trt layer (#42782)

* add squeeze2

* add squeeze

* add squeeze2,unsqueeze2

* merge develop

* fix format

* add conditions for squeeze2 and unsqueeze in op_teller

* merge develop

* add squeeze unsqueeze

* add squeeze unsqueeze

* add squeeze unsqueeze

* remove unsqueeze2_eltwise_fuse_pass

* add squeeze/unsqueeze

paddle/fluid/inference/api/analysis_predictor.cc

@@ -1962,6 +1962,8 @@ USE_TRT_CONVERTER(recover_padding)
 USE_TRT_CONVERTER(remove_padding)
 USE_TRT_CONVERTER(top_k)
 USE_TRT_CONVERTER(top_k_v2)
+USE_TRT_CONVERTER(squeeze2)
+USE_TRT_CONVERTER(unsqueeze2)
 #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

@@ -61,7 +61,9 @@ list(
   transformer_input_convert_op.cc
   remove_padding_op.cc
   recover_padding_op.cc
-  top_k_op.cc)
+  top_k_op.cc
+  squeeze2_op.cc
+  unsqueeze2_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/squeeze2_op.cc

+/* 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 Squeeze2OpConverter : public OpConverter {
+ public:
+  void operator()(const framework::proto::OpDesc& op,
+                  const framework::Scope& scope, bool test_mode) override {
+    VLOG(4) << "convert a fluid squeeze2 op to tensorrt shuffle layer";
+
+    framework::OpDesc op_desc(op, nullptr);
+    // Declare inputs
+    auto* input = engine_->GetITensor(op_desc.Input("X")[0]);
+    auto input_dims = input->getDimensions();
+    auto output_name = op_desc.Output("Out")[0];
+
+    // Get Attrs
+    std::vector<int> axes =
+        BOOST_GET_CONST(std::vector<int>, op_desc.GetAttr("axes"));
+    PADDLE_ENFORCE_GT(
+        axes.size(), 0,
+        platform::errors::InvalidArgument(
+            "Attr(axes).size should be > 0 in squeeze2 op in TensorRT,"
+            "but received axes.size() = %d.",
+            axes.size()));
+
+    std::vector<bool> should_squeeze(input_dims.nbDims, false);
+    for (size_t i = 0; i < axes.size(); i++) {
+      if (engine_->with_dynamic_shape()) {
+        axes[i] += (axes[i] < 0) ? input_dims.nbDims : 0;
+      } else {
+        axes[i] += (axes[i] < 0) ? input_dims.nbDims : -1;
+      }
+      should_squeeze[axes[i]] = true;
+    }
+
+    nvinfer1::Dims trt_out_dims;
+    trt_out_dims.nbDims = 0;
+    std::vector<int32_t> gather_indices;
+    for (size_t i = 0; i < should_squeeze.size(); i++) {
+      if (should_squeeze[i]) continue;
+      gather_indices.push_back(i);
+      // for static shape
+      trt_out_dims.d[trt_out_dims.nbDims] = input_dims.d[i];
+      trt_out_dims.nbDims++;
+    }
+
+    auto* layer = TRT_ENGINE_ADD_LAYER(engine_, Shuffle, *input);
+    if (engine_->with_dynamic_shape()) {
+      auto* shape_tensor = Shape(input);
+      auto* real_shape_tensor = Gather(shape_tensor, gather_indices);
+      layer->setInput(1, *real_shape_tensor);
+    } else {
+      layer->setReshapeDimensions(trt_out_dims);
+    }
+    RreplenishLayerAndOutput(layer, "squeeze2", {output_name}, test_mode);
+  }
+};
+
+}  // namespace tensorrt
+}  // namespace inference
+}  // namespace paddle
+
+REGISTER_TRT_OP_CONVERTER(squeeze2, Squeeze2OpConverter);

paddle/fluid/inference/tensorrt/convert/unsqueeze2_op.cc

+/* 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 Unsqueeze2OpConverter : public OpConverter {
+ public:
+  void operator()(const framework::proto::OpDesc& op,
+                  const framework::Scope& scope, bool test_mode) override {
+    VLOG(4) << "convert a fluid unsqueeze2 op to tensorrt shuffle layer";
+
+    framework::OpDesc op_desc(op, nullptr);
+    // Declare inputs
+    auto* input = engine_->GetITensor(op_desc.Input("X")[0]);
+    auto input_dims = input->getDimensions();
+    auto output_name = op_desc.Output("Out")[0];
+
+    // Get Attrs
+    std::vector<int> axes =
+        BOOST_GET_CONST(std::vector<int>, op_desc.GetAttr("axes"));
+    PADDLE_ENFORCE_GT(
+        axes.size(), 0,
+        platform::errors::InvalidArgument(
+            "Attr(axes).size should be > 0 in unsqueeze2 op in TensorRT,"
+            "but received axes.size() = %d.",
+            axes.size()));
+
+    std::vector<bool> should_unsqueeze(input_dims.nbDims + axes.size(), false);
+    int cur_out_rank = input_dims.nbDims;
+    for (size_t i = 0; i < axes.size(); i++) {
+      cur_out_rank++;
+      if (engine_->with_dynamic_shape()) {
+        axes[i] += (axes[i] < 0) ? cur_out_rank : 0;
+      } else {
+        axes[i] += (axes[i] < 0) ? cur_out_rank : -1;
+      }
+      // axes[i] is relative to cur_out_rank
+      // we make [axes[i], cur_out_rank - 2] shift right
+      // and make (axes[i]) to true!
+      for (int j = cur_out_rank - 1; j > axes[i]; j--) {
+        should_unsqueeze[j] = should_unsqueeze[j - 1];
+      }
+      if (axes[i] >= cur_out_rank)
+        should_unsqueeze[cur_out_rank - 1] = true;
+      else
+        should_unsqueeze[axes[i]] = true;
+    }
+
+    nvinfer1::Dims trt_out_dims;
+    trt_out_dims.nbDims = should_unsqueeze.size();
+    std::vector<int32_t> gather_indices;
+    int in_rank_i = 0;
+    for (size_t i = 0; i < should_unsqueeze.size(); i++) {
+      if (should_unsqueeze[i]) {
+        trt_out_dims.d[i] = 1;
+        gather_indices.push_back(input_dims.nbDims);
+        continue;
+      }
+      trt_out_dims.d[i] = input_dims.d[in_rank_i];
+      gather_indices.push_back(in_rank_i);
+      in_rank_i++;
+    }
+
+    auto* layer = TRT_ENGINE_ADD_LAYER(engine_, Shuffle, *input);
+    if (engine_->with_dynamic_shape()) {
+      auto* shape_tensor = Shape(input);
+      std::vector<int32_t> all_one(axes.size(), 1);
+      auto* all_one_tensor = Add1DConstantLayer(all_one);
+      std::vector<nvinfer1::ITensor*> concat_inputs = {shape_tensor,
+                                                       all_one_tensor};
+      auto* real_shape_tensor = Gather(Concat(concat_inputs), gather_indices);
+      layer->setInput(1, *real_shape_tensor);
+    } else {
+      layer->setReshapeDimensions(trt_out_dims);
+    }
+    RreplenishLayerAndOutput(layer, "unsqueeze2", {output_name}, test_mode);
+  }
+};
+
+}  // namespace tensorrt
+}  // namespace inference
+}  // namespace paddle
+
+REGISTER_TRT_OP_CONVERTER(unsqueeze2, Unsqueeze2OpConverter);

paddle/fluid/inference/tensorrt/op_teller.cc

@@ -153,7 +153,9 @@ struct SimpleOpTypeSetTeller : public Teller {
       "preln_skip_layernorm",
       "transformer_input_convert",
       "recover_padding",
-      "remove_padding"};
+      "remove_padding",
+      "squeeze2",
+      "unsqueeze2"};
   std::unordered_set<std::string> teller_set{
       "mul",
       "matmul",
@@ -242,7 +244,9 @@ struct SimpleOpTypeSetTeller : public Teller {
       "multiclass_nms3",
       "transformer_input_convert",
       "recover_padding",
-      "remove_padding"};
+      "remove_padding",
+      "squeeze2",
+      "unsqueeze2"};
 };
 
 bool OpTeller::Tell(const framework::ir::Node* node, bool use_no_calib_int8,
@@ -887,6 +891,44 @@ bool OpTeller::Tell(const framework::ir::Node* node, bool use_no_calib_int8,
       }
     }
 
+    if (op_type == "squeeze2") {
+      std::vector<int> axes;
+      if (desc.HasAttr("axes")) {
+        axes = BOOST_GET_CONST(std::vector<int>, desc.GetAttr("axes"));
+      }
+      if (axes.size() == 0) {
+        VLOG(3) << "The necessary attributes of the squeeze2 operator axes is "
+                   "missing.";
+        return false;
+      }
+      if (!with_dynamic_shape) {
+        if (std::find(axes.begin(), axes.end(), 0) != axes.end()) {
+          VLOG(3) << "Invalid squeeze axes. Axes having batch axis is not "
+                     "supported in static shape";
+          return false;
+        }
+      }
+    }
+
+    if (op_type == "unsqueeze2") {
+      std::vector<int> axes;
+      if (desc.HasAttr("axes")) {
+        axes = BOOST_GET_CONST(std::vector<int>, desc.GetAttr("axes"));
+      }
+      if (axes.size() == 0) {
+        VLOG(3) << "The necessary attributes of the squeeze2 operator axes is "
+                   "missing.";
+        return false;
+      }
+      if (!with_dynamic_shape) {
+        if (std::find(axes.begin(), axes.end(), 0) != axes.end()) {
+          VLOG(3) << "Invalid squeeze axes. Axes having batch axis is not "
+                     "supported in static shape";
+          return false;
+        }
+      }
+    }
+
     if (op_type == "batch_norm") {
       const std::vector<std::string> bn_inputs = {"X", "Bias", "Mean", "Scale",
                                                   "Variance"};

python/paddle/fluid/tests/unittests/ir/inference/test_trt_convert_squeeze2.py

+# 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:
+        inputs = program_config.inputs
+        attrs = [
+            program_config.ops[i].attrs for i in range(len(program_config.ops))
+        ]
+        if len(inputs['in_data'].shape) <= max(attrs[0]['axes']):
+            return False
+        return True
+
+    def sample_program_configs(self):
+        for dims in [2, 3, 4]:
+            for batch in [3, 4]:
+                for axes in [[2], [2, 3], [-1]]:
+                    self.batch = batch
+                    self.dims = dims
+                    self.axes = axes
+                    dics = [{"axes": axes}]
+                    ops_config = [{
+                        "op_type": "squeeze2",
+                        "op_inputs": {
+                            "X": ["in_data"]
+                        },
+                        "op_outputs": {
+                            "Out": ["out_data"],
+                            "XShape": ["XShape_data"]
+                        },
+                        "op_attrs": dics[0]
+                    }]
+                    # new_axes is the update of axes
+                    new_axes = list(axes)
+                    for i in range(len(new_axes)):
+                        if (new_axes[i] < 0):
+                            new_axes[i] += dims
+                    if (max(new_axes) >= dims):
+                        continue
+                # generate input data
+                    self.input_shape = [1] * dims
+                    for i in range(dims):
+                        self.input_shape[i] = np.random.randint(1, 20)
+
+                    def generate_input1(attrs: List[Dict[str, Any]], batch):
+                        self.input_shape[0] = batch
+                        for i in new_axes:
+                            self.input_shape[i] = 1
+                        return np.random.random(self.input_shape).astype(
+                            np.float32)
+
+                    ops = self.generate_op_config(ops_config)
+                    program_config = ProgramConfig(
+                        ops=ops,
+                        weights={},
+                        inputs={
+                            "in_data":
+                            TensorConfig(
+                                data_gen=partial(generate_input1, dics, batch))
+                        },
+                        outputs=["out_data"])
+
+                    yield program_config
+
+    def sample_predictor_configs(
+            self, program_config) -> (paddle_infer.Config, List[int], float):
+
+        def generate_dynamic_shape(attrs):
+            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 = {"in_data": min_shape}
+            self.dynamic_shape.max_input_shape = {"in_data": max_shape}
+            self.dynamic_shape.opt_input_shape = {"in_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):
+            return 1, 2
+
+        attrs = [
+            program_config.ops[i].attrs for i in range(len(program_config.ops))
+        ]
+        self.trt_param.max_batch_size = 9
+        # for static_shape
+        clear_dynamic_shape()
+        self.trt_param.precision = paddle_infer.PrecisionType.Float32
+        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 add_skip_trt_case(self):
+        pass
+
+    def test(self):
+        self.add_skip_trt_case()
+        self.run_test()
+
+
+if __name__ == "__main__":
+    unittest.main()

python/paddle/fluid/tests/unittests/ir/inference/test_trt_convert_unsqueeze2.py

+# 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):
+        for dims in [2, 3, 4]:
+            for batch in [3, 4]:
+                for axes in [[-2, 3], [1], [2], [2, 3]]:
+                    self.batch = batch
+                    self.dims = dims
+                    self.axes = axes
+                    dics = [{"axes": axes}]
+                    ops_config = [{
+                        "op_type": "unsqueeze2",
+                        "op_inputs": {
+                            "X": ["in_data"]
+                        },
+                        "op_outputs": {
+                            "Out": ["out_data"],
+                            "XShape": ["XShape_data"]
+                        },
+                        "op_attrs": dics[0]
+                    }]
+
+                    # generate input data
+                    self.input_shape = [1] * dims
+                    for i in range(dims):
+                        self.input_shape[i] = np.random.randint(1, 20)
+
+                    def generate_input1(attrs: List[Dict[str, Any]], batch):
+                        self.input_shape[0] = batch
+                        return np.random.random(self.input_shape).astype(
+                            np.float32)
+
+                    ops = self.generate_op_config(ops_config)
+                    program_config = ProgramConfig(
+                        ops=ops,
+                        weights={},
+                        inputs={
+                            "in_data":
+                            TensorConfig(
+                                data_gen=partial(generate_input1, dics, batch))
+                        },
+                        outputs=["out_data"])
+
+                    yield program_config
+
+    def sample_predictor_configs(
+            self, program_config) -> (paddle_infer.Config, List[int], float):
+
+        def generate_dynamic_shape(attrs):
+            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 = {"in_data": min_shape}
+            self.dynamic_shape.max_input_shape = {"in_data": max_shape}
+            self.dynamic_shape.opt_input_shape = {"in_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):
+            return 1, 2
+
+        attrs = [
+            program_config.ops[i].attrs for i in range(len(program_config.ops))
+        ]
+        self.trt_param.max_batch_size = 9
+        # for static_shape
+        clear_dynamic_shape()
+        self.trt_param.precision = paddle_infer.PrecisionType.Float32
+        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 add_skip_trt_case(self):
+        pass
+
+    def test(self):
+        self.add_skip_trt_case()
+        self.run_test()
+
+
+if __name__ == "__main__":
+    unittest.main()