[Hackathon NO.77] 为 Paddle-TRT 添加 bitwise 算子 (#51971)

* add bitwise_not trt converter

* run pre-commit

* modify neg_one_tensor_dims init way

* fix BOOL type support requires TensorRT 8.4

* fix int8 & uint8 type

* improve data type readability

* modify filter logic

* fix coverage CI

paddle/fluid/inference/api/analysis_predictor.cc

@@ -2460,6 +2460,7 @@ USE_TRT_CONVERTER(conv2d_transpose);
 USE_TRT_CONVERTER(leaky_relu);
 USE_TRT_CONVERTER(shuffle_channel);
 USE_TRT_CONVERTER(where);
+USE_TRT_CONVERTER(bitwise_not);
 USE_TRT_CONVERTER(one_hot);
 USE_TRT_CONVERTER(one_hot_v2);
 USE_TRT_CONVERTER(swish);

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

@@ -32,6 +32,7 @@ list(
   shuffle_channel_op.cc
   fill_any_like_op.cc
   where_op.cc
+  bitwise_not_op.cc
   one_hot_op.cc
   swish_op.cc
   silu_op.cc

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

+/* Copyright (c) 2023 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 <NvInferRuntimeCommon.h>
+#include <cstddef>
+#include <iostream>
+#include "paddle/fluid/inference/tensorrt/convert/op_converter.h"
+
+namespace paddle {
+namespace inference {
+namespace tensorrt {
+
+class BitwiseNotConverter : public OpConverter {
+ public:
+  void operator()(const framework::proto::OpDesc& op,
+                  const framework::Scope& scope,
+                  bool test_mode) override {
+    VLOG(4) << "convert bitwise_not op to tensorrt layer";
+
+    framework::OpDesc op_desc(op, nullptr);
+    nvinfer1::ILayer* layer = nullptr;
+
+    auto* input_tensor = engine_->GetITensor(op_desc.Input("X")[0]);
+    nvinfer1::DataType data_type = input_tensor->getType();
+
+    // for bool type: use UnaryOperation::kNOT, for int type: !x = -x - 1
+    if (data_type == nvinfer1::DataType::kBOOL) {
+      layer = TRT_ENGINE_ADD_LAYER(
+          engine_, Unary, *input_tensor, nvinfer1::UnaryOperation::kNOT);
+    } else {
+      nvinfer1::Dims input_dims = input_tensor->getDimensions();
+
+      // set up a elementwise -1 tensor, can not get the dims info for
+      // dynamic_shape so just let it broadcaste
+      nvinfer1::Dims neg_one_tensor_dims;
+      neg_one_tensor_dims.nbDims = input_dims.nbDims;
+      for (int i = 0; i < input_dims.nbDims; ++i) {
+        neg_one_tensor_dims.d[i] = 1;
+      }
+
+      nvinfer1::Weights weights{nvinfer1::DataType::kINT32, new int(-1), 1};
+      auto neg_one_tensor =
+          TRT_ENGINE_ADD_LAYER(engine_, Constant, neg_one_tensor_dims, weights)
+              ->getOutput(0);
+
+      auto mul_neg_one =
+          TRT_ENGINE_ADD_LAYER(engine_,
+                               ElementWise,
+                               *input_tensor,
+                               *neg_one_tensor,
+                               nvinfer1::ElementWiseOperation::kPROD);
+
+      layer = TRT_ENGINE_ADD_LAYER(engine_,
+                                   ElementWise,
+                                   *(mul_neg_one->getOutput(0)),
+                                   *neg_one_tensor,
+                                   nvinfer1::ElementWiseOperation::kSUM);
+    }
+
+    auto output_name = op_desc.Output("Out")[0];
+    RreplenishLayerAndOutput(layer, "bitwise_not", {output_name}, test_mode);
+  }
+};
+
+}  // namespace tensorrt
+}  // namespace inference
+}  // namespace paddle
+
+REGISTER_TRT_OP_CONVERTER(bitwise_not, BitwiseNotConverter);

paddle/fluid/inference/tensorrt/op_teller.cc

@@ -1954,6 +1954,21 @@ struct SimpleOpTypeSetTeller : public Teller {
       }
     }
 
+    if (op_type == "bitwise_not") {
+#if !IS_TRT_VERSION_GE(8400)
+      auto* block = desc.Block();
+      auto x_var_name = desc.Input("X")[0];
+      auto* x_var_desc = block->FindVar(x_var_name);
+      auto dtype = x_var_desc->GetDataType();
+      if (dtype == framework::proto::VarType::BOOL ||
+          dtype == framework::proto::VarType::INT8 ||
+          dtype == framework::proto::VarType::UINT8) {
+        VLOG(3) << "BOOL / INT8 / UINT8 type support requires TensorRT 8.4";
+        return false;
+      }
+#endif
+    }
+
     if (op_type == "one_hot" || op_type == "one_hot_v2") {
 #if IS_TRT_VERSION_LT(8510)
       VLOG(3) << "one_hot/one_hot_v2 is not supported when TensorRT < 8.5.1";
@@ -2778,6 +2793,7 @@ struct SimpleOpTypeSetTeller : public Teller {
       "fc",
       "shuffle_channel",
       "where",
+      "bitwise_not",
       "one_hot",
       "one_hot_v2",
       "swish",
@@ -2935,6 +2951,7 @@ struct SimpleOpTypeSetTeller : public Teller {
       "fc",
       "shuffle_channel",
       "where",
+      "bitwise_not",
       "one_hot",
       "one_hot_v2",
       "swish",

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

+# Copyright (c) 2023 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.
+
+import unittest
+from functools import partial
+from typing import Any, Dict, List
+
+import numpy as np
+from program_config import ProgramConfig, TensorConfig
+from trt_layer_auto_scan_test import TrtLayerAutoScanTest
+
+import paddle.inference as paddle_infer
+
+
+class TrtConvertActivationTest(TrtLayerAutoScanTest):
+    def is_program_valid(self, program_config: ProgramConfig) -> bool:
+        return True
+
+    def sample_program_configs(self):
+        self.trt_param.workspace_size = 1073741824
+
+        def generate_input1(dims, batch, attrs: List[Dict[str, Any]]):
+            if dims == 1:
+                return np.random.random([32]).astype(np.bool8)
+            elif dims == 2:
+                return np.random.random([3, 32]).astype(np.int8)
+            elif dims == 3:
+                return np.random.random([3, 32, 32]).astype(np.int32)
+            else:
+                return np.random.random([batch, 3, 32, 32]).astype(np.int64)
+
+        for dims in [1, 2, 3, 4]:
+            for batch in [1, 4]:
+                self.dims = dims
+                dics = [{}]
+
+                ops_config = [
+                    {
+                        "op_type": 'bitwise_not',
+                        "op_inputs": {"X": ["input_data"]},
+                        "op_outputs": {"Out": ["output_data"]},
+                        "op_attrs": dics[0],
+                    }
+                ]
+                ops = self.generate_op_config(ops_config)
+
+                program_config = ProgramConfig(
+                    ops=ops,
+                    weights={},
+                    inputs={
+                        "input_data": TensorConfig(
+                            data_gen=partial(generate_input1, dims, batch, dics)
+                        )
+                    },
+                    outputs=["output_data"],
+                )
+                yield program_config
+
+    def sample_predictor_configs(
+        self, program_config
+    ) -> (paddle_infer.Config, List[int], float):
+        def generate_dynamic_shape(attrs):
+            if self.dims == 1:
+                self.dynamic_shape.min_input_shape = {"input_data": [1]}
+                self.dynamic_shape.max_input_shape = {"input_data": [64]}
+                self.dynamic_shape.opt_input_shape = {"input_data": [32]}
+            elif self.dims == 2:
+                self.dynamic_shape.min_input_shape = {"input_data": [1, 16]}
+                self.dynamic_shape.max_input_shape = {"input_data": [4, 32]}
+                self.dynamic_shape.opt_input_shape = {"input_data": [3, 32]}
+            elif self.dims == 3:
+                self.dynamic_shape.min_input_shape = {"input_data": [1, 16, 16]}
+                self.dynamic_shape.max_input_shape = {"input_data": [4, 32, 32]}
+                self.dynamic_shape.opt_input_shape = {"input_data": [3, 32, 32]}
+            else:
+                self.dynamic_shape.min_input_shape = {
+                    "input_data": [1, 3, 16, 16]
+                }
+                self.dynamic_shape.max_input_shape = {
+                    "input_data": [4, 3, 32, 32]
+                }
+                self.dynamic_shape.opt_input_shape = {
+                    "input_data": [1, 3, 32, 32]
+                }
+
+        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):
+            ver = paddle_infer.get_trt_compile_version()
+            trt_version = ver[0] * 1000 + ver[1] * 100 + ver[2] * 10
+            if trt_version >= 8400:
+                if self.dims == 1 and not dynamic_shape:
+                    return 0, 3
+                return 1, 2
+            else:
+                if (self.dims == 1 and not dynamic_shape) or (
+                    program_config.inputs['input_data'].dtype
+                    in ['bool', 'int8', 'uint8']
+                ):
+                    return 0, 3
+                return 1, 2
+
+        attrs = [
+            program_config.ops[i].attrs for i in range(len(program_config.ops))
+        ]
+
+        # 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 test(self):
+        self.run_test()
+
+
+if __name__ == "__main__":
+    unittest.main()