[Paddle-TRT]TensorRT support isnan_v2

paddle/fluid/inference/api/analysis_predictor.cc

@@ -3618,6 +3618,7 @@ USE_TRT_CONVERTER(assign)
 USE_TRT_CONVERTER(p_norm)
 USE_TRT_CONVERTER(unbind)
 USE_TRT_CONVERTER(flip)
+USE_TRT_CONVERTER(isnan_v2)
 USE_TRT_CONVERTER(share_data)
 #if IS_TRT_VERSION_GE(8522)
 USE_TRT_CONVERTER(flash_multihead_matmul)

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

@@ -111,6 +111,7 @@ list(
   temporal_shift_op.cc
   einsum_op.cc
   unbind_op.cc
+  isnan_v2_op.cc
   assign_op.cc
   p_norm_op.cc
   flip_op.cc

paddle/fluid/inference/tensorrt/convert/isnan_v2_op.cc

@@ -0,0 +1,54 @@
+/* Copyright (c) 2024 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 {
+
+/*
+ * IsnanV2 Op
+ */
+class IsnanV2OpConverter : public OpConverter {
+ public:
+  void operator()(const framework::proto::OpDesc& op,
+                  const framework::Scope& scope,
+                  bool test_mode) override {
+    VLOG(3) << "convert a isnan_v2 op to tensorrt layer";
+    framework::OpDesc op_desc(op, nullptr);
+    std::string input_x_name = op_desc.Input("X").front();
+    std::string output_name = op_desc.Output("Out").front();
+    auto* input_x_tensor = engine_->GetITensor(input_x_name);
+#if IS_TRT_VERSION_GE(10100)
+    auto* layer = TRT_ENGINE_ADD_LAYER(
+        engine_, Unary, *input_x_tensor, nvinfer1::UnaryOperation::kISNAN);
+    ReplenishLayerAndOutput(layer, "isnan_v2", {output_name}, test_mode);
+#else
+    auto* equal_layer =
+        TRT_ENGINE_ADD_LAYER(engine_,
+                             ElementWise,
+                             *input_x_tensor,
+                             *input_x_tensor,
+                             nvinfer1::ElementWiseOperation::kEQUAL);
+    auto* layer = TRT_ENGINE_ADD_LAYER(engine_,
+                                       Unary,
+                                       *equal_layer->getOutput(0),
+                                       nvinfer1::UnaryOperation::kNOT);
+    ReplenishLayerAndOutput(layer, "isnan_v2", {output_name}, test_mode);
+#endif
+  }
+};
+
+}  // namespace tensorrt
+}  // namespace inference
+}  // namespace paddle
+REGISTER_TRT_OP_CONVERTER(isnan_v2, IsnanV2OpConverter);

paddle/fluid/inference/tensorrt/op_teller.cc

@@ -2793,6 +2793,21 @@ struct SimpleOpTypeSetTeller : public Teller {
       }
     }
 
+    if (op_type == "isnan_v2") {
+      if (!with_dynamic_shape) {
+        VLOG(3) << "the isnan_v2 does not support "
+                   "static shape yet";
+        return false;
+      }
+      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;
+      }
+    }
+
     if (op_type == "p_norm") {
       auto* block = desc.Block();
       if (block == nullptr) {
@@ -3077,6 +3092,7 @@ struct SimpleOpTypeSetTeller : public Teller {
       "grid_sampler",
       "cumsum",
       "unbind",
+      "isnan_v2",
       "p_norm",
       "assign",
       "flip",
@@ -3250,6 +3266,7 @@ struct SimpleOpTypeSetTeller : public Teller {
       "grid_sampler",
       "cumsum",
       "unbind",
+      "isnan_v2",
       "p_norm",
       "assign",
       "flip",

test/ir/inference/test_trt_convert_isnan_v2.py

@@ -0,0 +1,164 @@
+# Copyright (c) 2024 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 os
+import unittest
+from functools import partial
+from typing import 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 TrtConvertIsnanV2Test(TrtLayerAutoScanTest):
+    def sample_program_configs(self):
+        def generate_input1(dims):
+            if dims == 1:
+                data = np.random.random([3]).astype(np.float32)
+                mask = np.random.random([3]).astype(np.float32) < 0.3
+                data[mask] = np.nan
+                return data
+            elif dims == 2:
+                data = np.random.random([3, 64]).astype(np.float32)
+                mask = np.random.random([3, 64]).astype(np.float32) < 0.3
+                data[mask] = np.nan
+                return data
+            elif dims == 3:
+                data = np.random.random([3, 64, 64]).astype(np.float32)
+                mask = np.random.random([3, 64, 64]).astype(np.float32) < 0.3
+                data[mask] = np.nan
+                return data
+            else:
+                data = np.random.random([1, 3, 64, 64]).astype(np.float32)
+                mask = np.random.random([1, 3, 64, 64]).astype(np.float32) < 0.3
+                data[mask] = np.nan
+                return data
+
+        for dims in [1, 2, 3, 4]:
+            self.dims = dims
+            ops_config = [
+                {
+                    "op_type": "isnan_v2",
+                    "op_inputs": {
+                        "X": ["input_data"],
+                    },
+                    "op_outputs": {
+                        "Out": ["isnan_v2_output_data"],
+                    },
+                    "op_attrs": {},
+                },
+                {
+                    "op_type": "cast",
+                    "op_inputs": {"X": ["isnan_v2_output_data"]},
+                    "op_outputs": {"Out": ["output_data"]},
+                    "op_attrs": {
+                        "in_dtype": 0,
+                        "out_dtype": 5,
+                    },
+                },
+            ]
+            ops = self.generate_op_config(ops_config)
+
+            program_config = ProgramConfig(
+                ops=ops,
+                weights={},
+                inputs={
+                    "input_data": TensorConfig(
+                        data_gen=partial(generate_input1, dims)
+                    )
+                },
+                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": [128]}
+                self.dynamic_shape.opt_input_shape = {"input_data": [64]}
+            elif self.dims == 2:
+                self.dynamic_shape.min_input_shape = {"input_data": [1, 32]}
+                self.dynamic_shape.max_input_shape = {"input_data": [4, 64]}
+                self.dynamic_shape.opt_input_shape = {"input_data": [3, 64]}
+            elif self.dims == 3:
+                self.dynamic_shape.min_input_shape = {"input_data": [1, 32, 32]}
+                self.dynamic_shape.max_input_shape = {
+                    "input_data": [10, 64, 64]
+                }
+                self.dynamic_shape.opt_input_shape = {"input_data": [3, 64, 64]}
+            else:
+                self.dynamic_shape.min_input_shape = {
+                    "input_data": [1, 3, 32, 32]
+                }
+                self.dynamic_shape.max_input_shape = {
+                    "input_data": [4, 3, 64, 64]
+                }
+                self.dynamic_shape.opt_input_shape = {
+                    "input_data": [1, 3, 64, 64]
+                }
+
+        def generate_trt_nodes_num(attrs, dynamic_shape):
+            if not dynamic_shape:
+                return 0, 4
+            return 1, 2
+
+        def clear_dynamic_shape():
+            self.dynamic_shape.max_input_shape = {}
+            self.dynamic_shape.min_input_shape = {}
+            self.dynamic_shape.opt_input_shape = {}
+
+        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
+        program_config.set_input_type(np.float32)
+        yield self.create_inference_config(), generate_trt_nodes_num(
+            attrs, False
+        ), 1e-5
+        self.trt_param.precision = paddle_infer.PrecisionType.Half
+        program_config.set_input_type(np.float16)
+        yield self.create_inference_config(), generate_trt_nodes_num(
+            attrs, False
+        ), (1e-3, 1e-3)
+
+        # for dynamic_shape mode
+        generate_dynamic_shape(attrs)
+        self.trt_param.precision = paddle_infer.PrecisionType.Float32
+        program_config.set_input_type(np.float32)
+        yield self.create_inference_config(), generate_trt_nodes_num(
+            attrs, True
+        ), 1e-5
+        self.trt_param.precision = paddle_infer.PrecisionType.Half
+        program_config.set_input_type(np.float16)
+        yield self.create_inference_config(), generate_trt_nodes_num(
+            attrs, True
+        ), (1e-3, 1e-3)
+
+    def test(self):
+        if os.name != 'nt':
+            self.run_test()
+
+
+if __name__ == "__main__":
+    unittest.main()