[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(8500)
+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 +1,93 @@
+/* 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"
+
+namespace paddle {
+namespace framework {
+class Scope;
+
+namespace proto {
+class OpDesc;
+}  // namespace proto
+}  // namespace framework
+}  // namespace paddle
+
+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 {
+    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);
+    int32_t const inputRank = input_x_tensor->getDimensions().nbDims;
+    auto* input_grid_tensor = engine_->GetITensor(input_grid_name);
+    int32_t const gridRank = input_grid_tensor->getDimensions().nbDims;
+
+    if (inputRank != gridRank) {
+      PADDLE_THROW(platform::errors::InvalidArgument(
+          "The input tensor and the grid tensor must have the same rank"));
+    }
+
+    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);
+  }
+};
+
+}  // namespace tensorrt
+}  // namespace inference
+}  // namespace paddle
+
+REGISTER_TRT_OP_CONVERTER(grid_sampler, GridSamplerOpConverter);

paddle/fluid/inference/tensorrt/dynamic_shape_infermeta.cc

@@ -360,6 +360,7 @@ nvinfer1::DimsExprs PNormInferMeta(
   return output;
 }
 
+#if !IS_TRT_VERSION_GE(8500)
 nvinfer1::DimsExprs GridSamplerInferMeta(
     int output_index,
     const nvinfer1::DimsExprs* inputs,
@@ -386,6 +387,7 @@ nvinfer1::DimsExprs GridSamplerInferMeta(
   }
   return output;
 }
+#endif
 
 PD_REGISTER_DYNAMIC_INFER_META_FN(gather_nd, GatherNdInferMeta);
 PD_REGISTER_DYNAMIC_INFER_META_FN(yolo_box, YoloBoxInferMeta);
@@ -395,7 +397,9 @@ PD_REGISTER_DYNAMIC_INFER_META_FN(scatter_nd_add, ScatterNdAddInferMeta);
 PD_REGISTER_DYNAMIC_INFER_META_FN(inverse, UnchangedInferMeta);
 PD_REGISTER_DYNAMIC_INFER_META_FN(moe, MoeInferMeta);
 PD_REGISTER_DYNAMIC_INFER_META_FN(pad3d, Pad3dInferMeta);
+#if !IS_TRT_VERSION_GE(8500)
 PD_REGISTER_DYNAMIC_INFER_META_FN(grid_sampler, GridSamplerInferMeta);
+#endif
 }  // namespace tensorrt
 }  // namespace inference
 }  // namespace paddle

paddle/fluid/inference/tensorrt/dynamic_shape_infermeta_registry.h

@@ -15,6 +15,7 @@
 #pragma once
 
 #include "paddle/fluid/inference/tensorrt/dynamic_shape_infermeta_factory.h"
+#include "paddle/fluid/inference/tensorrt/helper.h"
 
 namespace paddle {
 namespace inference {
@@ -27,7 +28,9 @@ USE_TRT_DYNAMIC_INFER_META_FN(unfold);
 USE_TRT_DYNAMIC_INFER_META_FN(scatter_nd_add);
 USE_TRT_DYNAMIC_INFER_META_FN(pad3d);
 USE_TRT_DYNAMIC_INFER_META_FN(inverse);
+#if !IS_TRT_VERSION_GE(8500)
 USE_TRT_DYNAMIC_INFER_META_FN(grid_sampler);
+#endif
 }  // namespace tensorrt
 }  // namespace inference
 }  // namespace paddle

paddle/fluid/inference/tensorrt/op_teller.cc

@@ -2542,6 +2542,34 @@ struct SimpleOpTypeSetTeller : public Teller {
       }
     }
 
+    if (op_type == "grid_sampler") {
+#if !IS_TRT_VERSION_GE(8500)
+      VLOG(3) << "grid_sampler is not supported when TensorRT < 8.5";
+      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;
+      }
+#endif
+    }
+
     if (use_no_calib_int8) {
       return int8_teller_set.count(op_type);
     } else {
@@ -2701,7 +2729,8 @@ 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 +2882,8 @@ 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

@@ -25,6 +25,10 @@
 
 class TrtConvertGridSampler(TrtLayerAutoScanTest):
     def is_program_valid(self, program_config: ProgramConfig) -> bool:
+        self.trt_param.workspace_size = 1073741824
+        ver = paddle_infer.get_trt_compile_version()
+        if ver[0] * 1000 + ver[1] * 100 + ver[2] * 10 < 8500:
+            return False
         return True
 
     def sample_program_configs(self):
@@ -34,6 +38,12 @@ def generate_input1():
         def generate_input2():
             return np.random.random([1, 3, 3, 2]).astype(np.float32)
 
+        desc = {
+            "mode": "bilinear",
+            "padding_mode": "border",
+            "align_corners": True,
+        }
+
         ops_config = [
             {
                 "op_type": "grid_sampler",
@@ -42,7 +52,7 @@ def generate_input2():
                     "Grid": ["grid_data"],
                 },
                 "op_outputs": {"Output": ["output_data"]},
-                "op_attrs": {},
+                "op_attrs": desc,
             }
         ]