[Hackathon NO.74] 为 Paddle-TRT 添加 grid_sampler 算子

paddle/fluid/inference/api/analysis_predictor.cc

@@ -2560,6 +2560,9 @@ USE_TRT_CONVERTER(preln_groupnorm_act)
 USE_TRT_CONVERTER(flash_multihead_matmul)
 USE_TRT_CONVERTER(cross_multihead_matmul)
 #endif
+#if IS_TRT_VERSION_GE(8510)
+USE_TRT_CONVERTER(grid_sampler)
+#endif
 #if IS_TRT_VERSION_GE(8200)
 USE_TRT_CONVERTER(set_value)
 #endif

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

@@ -27,6 +27,7 @@ list(
   multihead_matmul_roformer_op.cc
   flash_multihead_matmul_op.cc
   cross_multihead_matmul_op.cc
+  grid_sampler_op.cc
   shuffle_channel_op.cc
   fill_any_like_op.cc
   where_op.cc

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

@@ -0,0 +1,80 @@
+/* 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 "paddle/fluid/inference/tensorrt/convert/op_converter.h"
+
+namespace paddle {
+namespace inference {
+namespace tensorrt {
+
+/*
+ * GridSampler Op
+ */
+class GridSamplerOpConverter : public OpConverter {
+ public:
+  void operator()(const framework::proto::OpDesc& op,
+                  const framework::Scope& scope,
+                  bool test_mode) override {
+#if IS_TRT_VERSION_GE(8510)
+    VLOG(3) << "convert a fluid grid_sampler op to tensorrt GridSample layer";
+    framework::OpDesc op_desc(op, nullptr);
+    std::string input_x_name = op_desc.Input("X").front();
+    std::string input_grid_name = op_desc.Input("Grid").front();
+    std::string output_name = op_desc.Output("Output").front();
+    auto* input_x_tensor = engine_->GetITensor(input_x_name);
+    auto* input_grid_tensor = engine_->GetITensor(input_grid_name);
+
+    auto* layer = TRT_ENGINE_ADD_LAYER(
+        engine_, GridSample, *input_x_tensor, *input_grid_tensor);
+
+    const std::string mode =
+        PADDLE_GET_CONST(std::string, op_desc.GetAttr("mode"));
+    const std::string padding_mode =
+        PADDLE_GET_CONST(std::string, op_desc.GetAttr("padding_mode"));
+    const bool align_corners =
+        PADDLE_GET_CONST(bool, op_desc.GetAttr("align_corners"));
+
+    nvinfer1::InterpolationMode interpolationMode{
+        nvinfer1::InterpolationMode::kNEAREST};
+    if (mode == "nearest") {
+      interpolationMode = nvinfer1::ResizeMode::kNEAREST;
+    } else if (mode == "bilinear") {
+      interpolationMode = nvinfer1::ResizeMode::kLINEAR;
+    }
+
+    nvinfer1::SampleMode sampleMode{nvinfer1::SampleMode::kFILL};
+    if (padding_mode == "zeros") {
+      sampleMode = nvinfer1::SampleMode::kFILL;
+    } else if (padding_mode == "border") {
+      sampleMode = nvinfer1::SampleMode::kCLAMP;
+    } else if (padding_mode == "reflection") {
+      sampleMode = nvinfer1::SampleMode::kREFLECT;
+    }
+
+    layer->setInterpolationMode(interpolationMode);
+    layer->setSampleMode(sampleMode);
+    layer->setAlignCorners(align_corners);
+
+    RreplenishLayerAndOutput(layer, "grid_sampler", {output_name}, test_mode);
+#else
+    VLOG(3) << "grid_sampler is not supported when TensorRT < 8.5.1";
+#endif
+  }
+};
+
+}  // namespace tensorrt
+}  // namespace inference
+}  // namespace paddle
+
+REGISTER_TRT_OP_CONVERTER(grid_sampler, GridSamplerOpConverter);

paddle/fluid/inference/tensorrt/dynamic_shape_infermeta.cc

@@ -377,7 +377,7 @@ nvinfer1::DimsExprs GridSamplerInferMeta(
     output.d[2] = grid_dims.d[1];
     output.d[3] = grid_dims.d[2];
   } else {
-    output.nbDims = 4;
+    output.nbDims = 5;
     output.d[0] = x_dims.d[0];
     output.d[1] = x_dims.d[1];
     output.d[2] = grid_dims.d[1];

paddle/fluid/inference/tensorrt/op_teller.cc

@@ -2542,6 +2542,48 @@ struct SimpleOpTypeSetTeller : public Teller {
       }
     }
 
+    if (op_type == "grid_sampler") {
+#if !IS_TRT_VERSION_GE(8510)
+      VLOG(3) << "grid_sampler is not supported when TensorRT < 8.5.1";
+      return false;
+#else
+      if (!with_dynamic_shape) {
+        VLOG(3) << "the grid_sampler does not support "
+                   "static shape yet";
+        return false;
+      }
+
+      if (!desc.HasAttr("mode") || !desc.HasAttr("padding_mode") ||
+          !desc.HasAttr("align_corners")) {
+        VLOG(3) << "grid_sampler need attributes : mode, padding_mode, "
+                   "align_corners";
+        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;
+      }
+      auto input_name = desc.Input("X")[0];
+      auto* input_desc = block->FindVar(input_name);
+      const auto input_shape = input_desc->GetShape();
+
+      auto grid_name = desc.Input("Grid")[0];
+      auto* grid_desc = block->FindVar(grid_name);
+      const auto grid_shape = grid_desc->GetShape();
+
+      if (input_shape.size() != 4 || grid_shape.size() != 4) {
+        VLOG(3) << "The input and grid tensors must be shape tensors of rank 4 "
+                   "using TRT GridSample layer.";
+        return false;
+      }
+
+#endif
+    }
+
     if (use_no_calib_int8) {
       return int8_teller_set.count(op_type);
     } else {
@@ -2701,7 +2743,8 @@ struct SimpleOpTypeSetTeller : public Teller {
       "expand_v2",
       "fuse_eleadd_transpose",
       "skip_groupnorm_act",
-      "preln_groupnorm_act"};
+      "preln_groupnorm_act",
+      "grid_sampler"};
 
   std::unordered_set<std::string> teller_set{
       "mul",
@@ -2853,7 +2896,8 @@ struct SimpleOpTypeSetTeller : public Teller {
       "expand_v2",
       "fuse_eleadd_transpose",
       "skip_groupnorm_act",
-      "preln_groupnorm_act"};
+      "preln_groupnorm_act",
+      "grid_sampler"};
 };
 
 struct GenericPluginTeller : public Teller {

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

@@ -25,61 +25,103 @@
 
 class TrtConvertGridSampler(TrtLayerAutoScanTest):
     def is_program_valid(self, program_config: ProgramConfig) -> bool:
+        self.trt_param.workspace_size = 1073741824
         return True
 
     def sample_program_configs(self):
         def generate_input1():
-            return np.random.random([1, 3, 32, 32]).astype(np.float32)
+            if self.dims == 4:
+                self.input_shape = [1, 3, 32, 32]
+                return np.random.random([1, 3, 32, 32]).astype(np.float32)
+            elif self.dims == 5:
+                self.input_shape = [1, 3, 32, 32, 64]
+                return np.random.random([1, 3, 32, 32, 64]).astype(np.float32)
 
         def generate_input2():
-            return np.random.random([1, 3, 3, 2]).astype(np.float32)
-
-        ops_config = [
-            {
-                "op_type": "grid_sampler",
-                "op_inputs": {
-                    "X": ["input_data"],
-                    "Grid": ["grid_data"],
-                },
-                "op_outputs": {"Output": ["output_data"]},
-                "op_attrs": {},
-            }
-        ]
-
-        ops = self.generate_op_config(ops_config)
-        for i in range(10):
-            program_config = ProgramConfig(
-                ops=ops,
-                weights={},
-                inputs={
-                    "input_data": TensorConfig(
-                        data_gen=partial(generate_input1)
-                    ),
-                    "grid_data": TensorConfig(
-                        data_gen=partial(generate_input2)
-                    ),
-                },
-                outputs=["output_data"],
-            )
-
-        yield program_config
+            if self.dims == 4:
+                self.input_shape = [1, 3, 3, 2]
+                return np.random.random([1, 3, 3, 2]).astype(np.float32)
+            elif self.dims == 5:
+                self.input_shape = [1, 3, 3, 2, 3]
+                return np.random.random([1, 3, 3, 2, 3]).astype(np.float32)
+
+        mode = ["bilinear", "nearest"]
+        padding_mode = ["zeros", "reflection", "border"]
+        align_corners = [True, False]
+        descs = []
+        for m in mode:
+            for p in padding_mode:
+                for a in align_corners:
+                    descs.append(
+                        {
+                            "mode": m,
+                            "padding_mode": p,
+                            "align_corners": a,
+                        }
+                    )
+
+        for dims in [4, 5]:
+            for desc in descs:
+                self.dims = dims
+                ops_config = [
+                    {
+                        "op_type": "grid_sampler",
+                        "op_inputs": {
+                            "X": ["input_data"],
+                            "Grid": ["grid_data"],
+                        },
+                        "op_outputs": {"Output": ["output_data"]},
+                        "op_attrs": desc,
+                    }
+                ]
+                ops = self.generate_op_config(ops_config)
+
+                program_config = ProgramConfig(
+                    ops=ops,
+                    weights={},
+                    inputs={
+                        "input_data": TensorConfig(
+                            data_gen=partial(generate_input1)
+                        ),
+                        "grid_data": TensorConfig(
+                            data_gen=partial(generate_input2)
+                        ),
+                    },
+                    outputs=["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, 3, 32, 32],
-                "grid_data": [1, 3, 3, 2],
-            }
-            self.dynamic_shape.max_input_shape = {
-                "input_data": [1, 3, 64, 64],
-                "grid_data": [1, 3, 4, 4],
-            }
-            self.dynamic_shape.opt_input_shape = {
-                "input_data": [1, 3, 32, 32],
-                "grid_data": [1, 3, 3, 2],
-            }
+        def generate_dynamic_shape():
+            if self.dims == 4:
+                self.dynamic_shape.min_input_shape = {
+                    "input_data": [1, 3, 32, 32],
+                    "grid_data": [1, 3, 3, 2],
+                }
+                self.dynamic_shape.max_input_shape = {
+                    "input_data": [1, 3, 64, 64],
+                    "grid_data": [1, 3, 6, 2],
+                }
+                self.dynamic_shape.opt_input_shape = {
+                    "input_data": [1, 3, 32, 32],
+                    "grid_data": [1, 3, 3, 2],
+                }
+            elif self.dims == 5:
+                self.dynamic_shape.min_input_shape = {
+                    "input_data": [1, 3, 32, 32, 64],
+                    "grid_data": [1, 3, 3, 2, 3],
+                }
+                self.dynamic_shape.max_input_shape = {
+                    "input_data": [1, 3, 64, 64, 128],
+                    "grid_data": [1, 3, 3, 6, 3],
+                }
+                self.dynamic_shape.opt_input_shape = {
+                    "input_data": [1, 3, 32, 32, 64],
+                    "grid_data": [1, 3, 3, 2, 3],
+                }
 
         def clear_dynamic_shape():
             self.dynamic_shape.max_input_shape = {}
@@ -92,13 +134,9 @@ def clear_dynamic_shape():
 
         # for static_shape
         clear_dynamic_shape()
-        self.trt_param.precision = paddle_infer.PrecisionType.Float32
-        yield self.create_inference_config(), (0, 4), 1e-5
-        self.trt_param.precision = paddle_infer.PrecisionType.Half
-        yield self.create_inference_config(), (0, 4), 1e-3
 
         # for dynamic_shape
-        generate_dynamic_shape(attrs)
+        generate_dynamic_shape()
         self.trt_param.precision = paddle_infer.PrecisionType.Float32
         yield self.create_inference_config(), (1, 3), 1e-5
         self.trt_param.precision = paddle_infer.PrecisionType.Half