pool2d_op.cc

/* Copyright (c) 2018 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/plugin/pool_op_plugin.h"

namespace paddle {
namespace framework {
class Scope;

namespace proto {
class OpDesc;
}  // namespace proto
}  // namespace framework
}  // namespace paddle

namespace paddle {
namespace inference {
namespace tensorrt {

inline void DealCeilMode(const nvinfer1::Dims &input_shape,
                         std::vector<int> ksize, std::vector<int> strides,
                         std::vector<int> paddings, nvinfer1::DimsHW *pre_pad,
                         nvinfer1::DimsHW *post_pad, int input_dims) {
  int input_height = input_shape.d[input_dims - 2];
  int input_width = input_shape.d[input_dims - 1];
  int floor_h_output_size =
      (input_height - ksize[0] + 2 * paddings[0]) / strides[0] + 1;
  int ceil_h_output_size =
      (input_height - ksize[0] + 2 * paddings[0] + strides[0] - 1) /
          strides[0] +
      1;

  int floor_w_output_size =
      (input_width - ksize[1] + 2 * paddings[1]) / strides[1] + 1;
  int ceil_w_output_size =
      (input_width - ksize[1] + 2 * paddings[1] + strides[1] - 1) / strides[1] +
      1;
  if (floor_h_output_size != ceil_h_output_size) {
    post_pad->h() = strides[0] - 1;
  }

  if (floor_w_output_size != ceil_w_output_size) {
    post_pad->w() = strides[1] - 1;
  }
}

/*
 * Pool2dOp, IPoolingLayer in TRT. This Layer doesn't has weights.
 */
class Pool2dOpConverter : public OpConverter {
 public:
  void operator()(const framework::proto::OpDesc &op,
                  const framework::Scope &scope, bool test_mode) override {
    VLOG(4)
        << "convert a fluid pool2d op to tensorrt pool2d layer without bias";
    framework::OpDesc op_desc(op, nullptr);
    auto *input1 = engine_->GetITensor(op_desc.Input("X")[0]);
    nvinfer1::Dims input_shape = input1->getDimensions();
    int input_dims = input_shape.nbDims;

    bool global_pooling =
        BOOST_GET_CONST(bool, op_desc.GetAttr("global_pooling"));
    std::string pool_type =
        BOOST_GET_CONST(std::string, op_desc.GetAttr("pooling_type"));
    std::vector<int> ksize =
        BOOST_GET_CONST(std::vector<int>, op_desc.GetAttr("ksize"));
    std::vector<int> strides =
        BOOST_GET_CONST(std::vector<int>, op_desc.GetAttr("strides"));
    std::vector<int> paddings =
        BOOST_GET_CONST(std::vector<int>, op_desc.GetAttr("paddings"));
    bool exclusive = op_desc.HasAttr("exclusive")
                         ? BOOST_GET_CONST(bool, op_desc.GetAttr("exclusive"))
                         : true;
    bool ceil_mode = BOOST_GET_CONST(bool, op_desc.GetAttr("ceil_mode"));
    bool adaptive = false;
    if (op_desc.HasAttr("adaptive"))
      adaptive = BOOST_GET_CONST(bool, op_desc.GetAttr("adaptive"));
    std::string padding_algorithm = "EXPLICIT";
    if (op_desc.HasAttr("padding_algorithm"))
      padding_algorithm =
          BOOST_GET_CONST(std::string, op_desc.GetAttr("padding_algorithm"));

    nvinfer1::PoolingType nv_pool_type = nvinfer1::PoolingType::kMAX;
    nvinfer1::ReduceOperation reduce_operation =
        nvinfer1::ReduceOperation::kMAX;
    plugin::PoolPlugin::PoolType plugin_pool_type =
        plugin::PoolPlugin::PoolType::max;
    if (pool_type == "max") {
      nv_pool_type = nvinfer1::PoolingType::kMAX;
      reduce_operation = nvinfer1::ReduceOperation::kMAX;
      plugin_pool_type = plugin::PoolPlugin::PoolType::max;
    } else if (pool_type == "avg") {
      nv_pool_type = nvinfer1::PoolingType::kAVERAGE;
      reduce_operation = nvinfer1::ReduceOperation::kAVG;
      plugin_pool_type = plugin::PoolPlugin::PoolType::avg;
    }
    if (global_pooling || adaptive) {
      std::fill(paddings.begin(), paddings.end(), 0);
    }

    if (padding_algorithm == "VALID") {
      std::fill(paddings.begin(), paddings.end(), 0);
    }
    nvinfer1::DimsHW nv_ksize(ksize[0], ksize[1]);
    nvinfer1::DimsHW nv_strides(strides[0], strides[1]);
    nvinfer1::DimsHW nv_paddings(paddings[0], paddings[1]);

    nvinfer1::ILayer *layer = nullptr;
    nvinfer1::DimsHW g_pre_pad(0, 0);
    nvinfer1::DimsHW g_post_pad(0, 0);
    // paddle Non ceil_mode : Output size = (input size - filter size + 2 *
    // padding) / stride (stride size) + 1
    // tensorrt EXPLICIT_ROUND_DOWN: O = floor((M - DK) / S) + 1
    // so if M - DK < 0 we need extra padding
    if (input_shape.d[input_dims - 2] - ksize[0] + 2 * paddings[0] < 0) {
      g_post_pad.h() = strides[0] - 1;
    }
    if (input_shape.d[input_dims - 1] - ksize[1] + 2 * paddings[1] < 0) {
      g_post_pad.w() = strides[1] - 1;
    }

    if (op_desc.HasAttr("enable_int8")) {
      CHECK(op_desc.HasAttr("Input_scale"));
      float input_scale =
          BOOST_GET_CONST(float, op_desc.GetAttr("Input_scale"));
      engine_->SetTensorDynamicRange(input1, input_scale);
    }

    std::vector<int> real_paddings = paddings;
    for (int i = 0; i < 2; ++i) {
      int copy_pad = *(paddings.begin() + i);
      real_paddings.insert(real_paddings.begin() + 2 * i + 1, copy_pad);
    }
    // SAME
    if (padding_algorithm == "SAME") {
      // expand
      for (int i = 0; i < 2; ++i) {
        int copy_pad = *(paddings.begin() + 2 * i);
        paddings.insert(paddings.begin() + 2 * i + 1, copy_pad);
      }
      // compute
      for (int i = 0; i < 2; ++i) {
        int out_size = (input_shape.d[2 + i] + strides[i] - 1) / strides[i];
        int pad_sum = std::max(
            (out_size - 1) * strides[i] + ksize[i] - input_shape.d[2 + i], 0);
        int pad_0 = pad_sum / 2;
        int pad_1 = pad_sum - pad_0;
        paddings[i * 2] = pad_0;
        paddings[i * 2 + 1] = pad_1;
      }
      real_paddings = paddings;
      // slice
      for (int i = 0; i < 2; ++i) {
        paddings.erase(paddings.begin() + i + 1);
      }
    }
    // VALID
    if (padding_algorithm == "VALID") {
      std::fill(real_paddings.begin(), real_paddings.end(), 0);
    }

    if (global_pooling == true && !engine_->with_dynamic_shape()) {
      nv_ksize.d[0] = input_shape.d[input_dims - 2];
      nv_ksize.d[1] = input_shape.d[input_dims - 1];
      ksize[0] = input_shape.d[input_dims - 2];
      ksize[1] = input_shape.d[input_dims - 1];
    }

    if (engine_->with_dynamic_shape()) {
      if (!adaptive && !global_pooling && !ceil_mode) {
        // input_shape.d < 0 means we can't get shape info here.
        // we may suffer from issue if shape is not met finally.
        if ((padding_algorithm != "SAME") &&
            ((g_post_pad.w() > 0 && input_shape.d[input_dims - 2] > 0) ||
             (g_post_pad.h() > 0 && input_shape.d[input_dims - 1] > 0))) {
          auto *pad_layer = TRT_ENGINE_ADD_LAYER(engine_, Padding, *input1,
                                                 g_pre_pad, g_post_pad);
          PADDLE_ENFORCE_NOT_NULL(
              pad_layer, platform::errors::Fatal(
                             "Pad layer in poolOp converter could not be "
                             "created. The pointer to pad layer is `NULL`."));
          input1 = pad_layer->getOutput(0);
        }

        auto *pool_layer = TRT_ENGINE_ADD_LAYER(engine_, Pooling, *input1,
                                                nv_pool_type, nv_ksize);
        pool_layer->setStride(nv_strides);
        pool_layer->setPadding(nv_paddings);
        pool_layer->setAverageCountExcludesPadding(exclusive);
        if (padding_algorithm == "SAME") {
          pool_layer->setPaddingMode(nvinfer1::PaddingMode::kSAME_UPPER);
        }
        layer = pool_layer;
      } else if (!adaptive && !global_pooling && ceil_mode) {
        auto *pool_layer = TRT_ENGINE_ADD_LAYER(engine_, Pooling, *input1,
                                                nv_pool_type, nv_ksize);
        pool_layer->setStride(nv_strides);
        pool_layer->setPadding(nv_paddings);
        pool_layer->setAverageCountExcludesPadding(exclusive);
        if (padding_algorithm == "SAME") {
          pool_layer->setPaddingMode(nvinfer1::PaddingMode::kSAME_UPPER);
        } else {
          pool_layer->setPaddingMode(nvinfer1::PaddingMode::kEXPLICIT_ROUND_UP);
        }
        layer = pool_layer;
      } else if (global_pooling && !adaptive) {
        auto *reduce_layer = TRT_ENGINE_ADD_LAYER(engine_, Reduce, *input1,
                                                  reduce_operation, 12, true);
        layer = reduce_layer;
      } else {
#if IS_TRT_VERSION_GE(6000)
        plugin::PoolPluginDynamic *plugin = new plugin::PoolPluginDynamic(
            ceil_mode, pool_type, adaptive, exclusive, ksize, strides, paddings,
            global_pooling);
        layer = engine_->AddDynamicPlugin(&input1, 1, plugin);
#endif
      }
      auto output_name = op_desc.Output("Out")[0];
      layer->setName(("pool2d (Output: " + output_name + ")").c_str());
      layer->getOutput(0)->setName(output_name.c_str());
      engine_->SetITensor(output_name, layer->getOutput(0));
      if (test_mode) {
        engine_->DeclareOutput(output_name);
      }
      return;
    }

    if (global_pooling == true && adaptive == false) {
      auto *pool_layer = TRT_ENGINE_ADD_LAYER(engine_, Pooling, *input1,
                                              nv_pool_type, nv_ksize);
      PADDLE_ENFORCE_NOT_NULL(
          pool_layer, platform::errors::Fatal(
                          "trt pool layer in converter could not be created."));
      auto output_name = op_desc.Output("Out")[0];
      pool_layer->setName(("pool2d (Output: " + output_name + ")").c_str());
      pool_layer->getOutput(0)->setName(output_name.c_str());
      engine_->SetITensor(output_name, pool_layer->getOutput(0));
      layer = pool_layer;
      if (test_mode) {
        engine_->DeclareOutput(output_name);
      }
      return;
    }

    if (!adaptive) {
      if (ceil_mode) {
        if (nv_ksize.d[0] % nv_strides.d[0] == 0 &&
            nv_ksize.d[1] % nv_strides.d[1] == 0) {
          nvinfer1::DimsHW pre_pad(0, 0);
          nvinfer1::DimsHW post_pad(0, 0);
          // If ceil mode is true, we will pad the appropriate size to the
          // input.
          DealCeilMode(input_shape, ksize, strides, paddings, &pre_pad,
                       &post_pad, input_dims);
          auto *pad_layer = TRT_ENGINE_ADD_LAYER(engine_, Padding, *input1,
                                                 pre_pad, post_pad);

          PADDLE_ENFORCE_NOT_NULL(
              pad_layer, platform::errors::Fatal(
                             "Pad layer in poolOp converter could not be "
                             "created. The pointer to pad layer is `NULL`."));
          input1 = pad_layer->getOutput(0);

          auto *pool_layer = TRT_ENGINE_ADD_LAYER(engine_, Pooling, *input1,
                                                  nv_pool_type, nv_ksize);
          PADDLE_ENFORCE_NOT_NULL(
              pool_layer,
              platform::errors::Fatal(
                  "trt pool layer in converter could not be created."));
          pool_layer->setStride(nv_strides);
          pool_layer->setPadding(nv_paddings);
          if (padding_algorithm == "SAME") {
            pool_layer->setPaddingMode(nvinfer1::PaddingMode::kSAME_UPPER);
          }
          pool_layer->setAverageCountExcludesPadding(exclusive);
          layer = pool_layer;
        } else {
          std::vector<int> input_shape_v;
          for (int i = 0; i < input_dims; i++) {
            input_shape_v.push_back(input_shape.d[i]);
          }
          plugin::PoolPlugin *plugin = new plugin::PoolPlugin(
              ceil_mode, plugin_pool_type, adaptive, exclusive, ksize, strides,
              paddings, input_shape_v, real_paddings);
          auto *pool_layer = engine_->AddPlugin(&input1, 1, plugin);
          PADDLE_ENFORCE_NOT_NULL(
              pool_layer,
              platform::errors::Fatal(
                  "trt pool plugin layer in converter could not be created."));
          layer = pool_layer;
        }
      } else {
#if IS_TRT_VERSION_GE(8000)
        // Exclude padding pixels from the average mean is not supported well by
        // TRT
        // so enable padding for trt8.0 above.
        if ((g_post_pad.w() > 0 || g_post_pad.h() > 0) &&
            (padding_algorithm != "SAME") && !ceil_mode) {
          auto *pad_layer = TRT_ENGINE_ADD_LAYER(engine_, Padding, *input1,
                                                 g_pre_pad, g_post_pad);
          PADDLE_ENFORCE_NOT_NULL(
              pad_layer, platform::errors::Fatal(
                             "Pad layer in poolOp converter could not be "
                             "created. The pointer to pad layer is `NULL`."));
          input1 = pad_layer->getOutput(0);
        }
#endif
        auto *pool_layer = TRT_ENGINE_ADD_LAYER(engine_, Pooling, *input1,
                                                nv_pool_type, nv_ksize);
        PADDLE_ENFORCE_NOT_NULL(
            pool_layer,
            platform::errors::Fatal(
                "trt pool layer in converter could not be created."));
        pool_layer->setStride(nv_strides);
        pool_layer->setPadding(nv_paddings);
        if (padding_algorithm == "SAME") {
          pool_layer->setPaddingMode(nvinfer1::PaddingMode::kSAME_UPPER);
        }
        pool_layer->setAverageCountExcludesPadding(exclusive);
        layer = pool_layer;
      }
    } else {
      // Average pooling needs to exclude the padding pixels from the average
      // mean.
      // It is not supported well by TRT, we use a plugin here
      std::vector<int> input_shape_v;
      for (int i = 0; i < input_dims; i++) {
        input_shape_v.push_back(input_shape.d[i]);
      }
      plugin::PoolPlugin *plugin = new plugin::PoolPlugin(
          ceil_mode, plugin_pool_type, adaptive, exclusive, ksize, strides,
          paddings, input_shape_v, real_paddings);
      auto *pool_layer = engine_->AddPlugin(&input1, 1, plugin);
      PADDLE_ENFORCE_NOT_NULL(
          pool_layer,
          platform::errors::Fatal(
              "trt pool plugin layer in converter could not be created."));
      layer = pool_layer;
    }
    auto output_name = op_desc.Output("Out")[0];
    RreplenishLayerAndOutput(layer, "pool2d", {output_name}, test_mode);
  }
};

}  // namespace tensorrt
}  // namespace inference
}  // namespace paddle

REGISTER_TRT_OP_CONVERTER(pool2d, Pool2dOpConverter);