init roll convert (#41689)

* init roll convert

* add ut for roll convert

* roll convert don't support trt6.0

* fix: change ut for trt 7.0.0.1

paddle/fluid/inference/api/analysis_predictor.cc

@@ -1755,6 +1755,7 @@ USE_TRT_CONVERTER(deformable_conv);
 USE_TRT_CONVERTER(pool3d)
 USE_TRT_CONVERTER(fused_preln_embedding_eltwise_layernorm)
 USE_TRT_CONVERTER(preln_skip_layernorm)
+USE_TRT_CONVERTER(roll)
 USE_TRT_CONVERTER(strided_slice)
 #endif
 

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

@@ -25,6 +25,7 @@ nv_library(tensorrt_converter
                 preln_emb_eltwise_layernorm.cc
 		strided_slice_op.cc
                 preln_skip_layernorm.cc
+		roll_op.cc
            DEPS tensorrt_engine tensorrt_plugin operator scope framework_proto op_registry)
 
 nv_test(test_op_converter SRCS test_op_converter.cc DEPS

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

+/* 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"
+#include "paddle/fluid/inference/tensorrt/helper.h"
+
+namespace paddle {
+namespace framework {
+class Scope;
+namespace proto {
+class OpDesc;
+}  // namespace proto
+}  // namespace framework
+}  // namespace paddle
+
+namespace paddle {
+namespace inference {
+namespace tensorrt {
+/*
+ * Stack converter from fluid to tensorRT.
+ */
+class RollOpConverter : public OpConverter {
+ public:
+  void operator()(const framework::proto::OpDesc& op,
+                  const framework::Scope& scope, bool test_mode) override {
+    VLOG(4) << "convert fluid Roll op to tensorrt Slice layer";
+
+    framework::OpDesc op_desc(op, nullptr);
+    auto* input = engine_->GetITensor(op_desc.Input("X")[0]);
+    nvinfer1::Dims input_dims = input->getDimensions();
+
+    std::vector<int64_t> axis =
+        BOOST_GET_CONST(std::vector<int64_t>, op_desc.GetAttr("axis"));
+    std::vector<int64_t> shifts =
+        BOOST_GET_CONST(std::vector<int64_t>, op_desc.GetAttr("shifts"));
+
+    nvinfer1::Dims start;
+    start.nbDims = input_dims.nbDims;
+    for (int i = 0; i < start.nbDims; i++) {
+      start.d[i] = 0;
+    }
+    int axis_size = axis.size();
+    for (int i = 0; i < axis_size; i++) {
+      start.d[axis[i]] = (-shifts[i]) % input_dims.d[axis[i]];
+    }
+
+    nvinfer1::Dims stride;
+    stride.nbDims = input_dims.nbDims;
+    for (int i = 0; i < stride.nbDims; i++) {
+      stride.d[i] = 1;
+    }
+
+    nvinfer1::Dims size;
+    size.nbDims = input_dims.nbDims;
+    for (int i = 0; i < size.nbDims; i++) {
+      size.d[i] = 1;
+    }
+
+    auto output_name = op_desc.Output("Out")[0];
+
+    auto shape_layer = TRT_ENGINE_ADD_LAYER(engine_, Shape, *input);
+
+    auto* layer =
+        TRT_ENGINE_ADD_LAYER(engine_, Slice, *input, start, size, stride);
+    layer->setInput(2, *shape_layer->getOutput(0));
+#if IS_TRT_VERSION_GE(7000)
+    layer->setMode(nvinfer1::SliceMode::kWRAP);
+#endif
+
+    RreplenishLayerAndOutput(layer, "roll", {output_name}, test_mode);
+  }
+};
+
+}  // namespace tensorrt
+}  // namespace inference
+}  // namespace paddle
+
+REGISTER_TRT_OP_CONVERTER(roll, RollOpConverter);

paddle/fluid/inference/tensorrt/op_teller.cc

@@ -119,6 +119,7 @@ struct SimpleOpTypeSetTeller : public Teller {
       "slice",
       "strided_slice",
       "fused_preln_embedding_eltwise_layernorm",
+      "roll",
       "preln_skip_layernorm"};
   std::unordered_set<std::string> teller_set{
       "mul",
@@ -182,6 +183,7 @@ struct SimpleOpTypeSetTeller : public Teller {
       "strided_slice",
       "fused_preln_embedding_eltwise_layernorm",
       "preln_skip_layernorm",
+      "roll",
       "multiclass_nms3"};
 };
 
@@ -928,6 +930,28 @@ bool OpTeller::Tell(const framework::ir::Node* node, bool use_no_calib_int8,
       }
     }
 
+    if (op_type == "roll") {
+#if !IS_TRT_VERSION_GE(7000)
+      VLOG(3) << "roll converter does not support trt versions below 7.0";
+      return false;
+#endif
+      if (!with_dynamic_shape) {
+        return false;
+      }
+    }
+
+    if (op_type == "strided_slice") {
+      if (!with_dynamic_shape) {
+        return false;
+      }
+      if (!desc.HasAttr("axes") || !desc.HasAttr("starts") ||
+          !desc.HasAttr("ends") || !desc.HasAttr("strides")) {
+        VLOG(3)
+            << "The necessary attributes of the strided_slice operator miss ";
+        return false;
+      }
+    }
+
     if (op_type == "slice") {
       if (desc.HasAttr("decrease_axis")) {
         std::vector<int> decrease_axis =

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

+# 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.
+
+from trt_layer_auto_scan_test import TrtLayerAutoScanTest, SkipReasons
+from program_config import TensorConfig, ProgramConfig
+import numpy as np
+import paddle.inference as paddle_infer
+from functools import partial
+from typing import Optional, List, Callable, Dict, Any, Set
+import unittest
+
+
+class TrtConvertRollTest(TrtLayerAutoScanTest):
+    def is_program_valid(self, program_config: ProgramConfig) -> bool:
+        inputs = program_config.inputs
+        weights = program_config.weights
+        attrs = [
+            program_config.ops[i].attrs
+            for i in range(len(program_config.ops))
+        ]
+        return True
+
+    def sample_program_configs(self):
+        def generate_input1(attrs: List[Dict[str, Any]]):
+            return np.ones([1, 56, 56, 192]).astype(np.float32)
+
+        for axis in [[1, 2]]:
+            for shifts in [[-1, -1], [-3, -3]]:
+                dics = [{
+                    "axis": axis,
+                    "shifts": shifts,
+                }]
+
+                ops_config = [{
+                    "op_type": "roll",
+                    "op_inputs": {
+                        "X": ["input_data"]
+                    },
+                    "op_outputs": {
+                        "Out": ["roll_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, dics))
+                    },
+                    outputs=["roll_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": [1, 56, 56, 192]
+            }
+            self.dynamic_shape.max_input_shape = {
+                "input_data": [8, 56, 56, 192]
+            }
+            self.dynamic_shape.opt_input_shape = {
+                "input_data": [4, 56, 56, 192]
+            }
+
+        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):
+            inputs = program_config.inputs
+
+            if not dynamic_shape:
+                return 0, 3
+            ver = paddle_infer.get_trt_compile_version()
+            if ver[0] * 1000 + ver[1] * 100 + ver[2] * 10 < 7000:
+                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-4
+
+        # 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-4
+
+    def test(self):
+        self.run_test()
+
+
+if __name__ == "__main__":
+    unittest.main()