core/kernels/depthwise_conv_op.h

/* Copyright 2015 The TensorFlow 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.
==============================================================================*/

#ifndef TENSORFLOW_CORE_KERNELS_DEPTHWISE_CONV_OP_H_
#define TENSORFLOW_CORE_KERNELS_DEPTHWISE_CONV_OP_H_

#include "third_party/eigen3/unsupported/Eigen/CXX11/Tensor"
#include "tensorflow/core/framework/types.h"
#include "tensorflow/core/util/tensor_format.h"

namespace tensorflow {

struct DepthwiseArgs {
  // Input layer dimensions
  int batch;
  int in_rows;
  int in_cols;
  int in_depth;
  int filter_rows;
  int filter_cols;
  int depth_multiplier;
  int stride;
  int pad_rows;
  int pad_cols;

  // Output layer dimensions
  int out_rows;
  int out_cols;
  int out_depth;

  DepthwiseArgs()
      : batch(0),
        in_rows(0),
        in_cols(0),
        in_depth(0),
        filter_rows(0),
        filter_cols(0),
        depth_multiplier(0),
        stride(0),
        pad_rows(0),
        pad_cols(0),
        out_rows(0),
        out_cols(0),
        out_depth(0) {}
};

// Forward declaration.
class OpKernelContext;

template <typename Device, typename T>
struct LaunchDepthwiseConvOp {
  void operator()(OpKernelContext* ctx, const DepthwiseArgs& args,
                  const T* input, const T* filter, T* output,
                  TensorFormat data_format);
};

template <typename Device, typename T>
struct LaunchDepthwiseConvBackpropInputOp {
  void operator()(OpKernelContext* ctx, const DepthwiseArgs& args,
                  const T* out_backprop, const T* filter, T* in_backprop,
                  TensorFormat data_format);
};

template <typename Device, typename T>
struct LaunchDepthwiseConvBackpropFilterOp {
  void operator()(OpKernelContext* ctx, const DepthwiseArgs& args,
                  const T* out_backprop, const T* input, T* filter_backprop,
                  TensorFormat data_format);
};

#if GOOGLE_CUDA
template <typename T>
struct LaunchDepthwiseConvOp<Eigen::GpuDevice, T> {
  void operator()(OpKernelContext* ctx, const DepthwiseArgs& args,
                  const T* input, const T* filter, T* output,
                  TensorFormat data_format);
};

template <typename T>
struct LaunchDepthwiseConvBackpropInputOp<Eigen::GpuDevice, T> {
  void operator()(class OpKernelContext* ctx, const DepthwiseArgs& args,
                  const T* out_backprop, const T* filter, T* in_backprop,
                  TensorFormat data_format);
};

template <typename T>
struct LaunchDepthwiseConvBackpropFilterOp<Eigen::GpuDevice, T> {
  void operator()(class OpKernelContext* ctx, const DepthwiseArgs& args,
                  const T* out_backprop, const T* input, T* filter_backprop,
                  TensorFormat data_format);
};
#endif

}  // namespace tensorflow

namespace tensorflow {
namespace functor {

// Pads 'filter' to vector-register boundary along its inner dimension:
//   filter_inner_dim_size = in_depth * depth_multiplier
// Requires 'filter' to have the following storage order:
//   [filter_rows, filter_cols, in_depth, depth_multiplier]
// Returns zero-padded filter in 'padded_filter'.
//
// EX:
//   in_depth = 3, depth_multiplier = 2, filter [2, 2], register_width = 4
//   So we have a total of 3 * 2 = 6 filters, each of spatial size 2 x 2.
//
//   filter [rows, cols, in_depth, depth_multiplier]
//     [u0, v0, w0, x0] [y0, z0, u1, v1] [w1, x1, y1, z1]
//     [u2, v2, w2, x2] [y2, z2, u3, v3] [w3, x3, y3, z3]
//
//   padded_filter [rows, cols, in_depth, depth_multiplier]
//     [u0, v0, w0, x0] [y0, z0, 0, 0] [u1, v1, w1, x1] [y1, z1, 0, 0]
//     [u2, v2, w2, x2] [y2, z2, 0, 0] [u3, v3, w3, x3] [y3, z3, 0, 0]

template <typename T>
struct DepthwiseFilterPadOp {
  void operator()(const DepthwiseArgs& args, const T* filter,
                  T* padded_filter) {
    typedef typename Eigen::internal::packet_traits<T>::type Packet;
    static const int64 kPacketSize = (sizeof(Packet) / sizeof(T));

    // Calculate vectorized and scalar lengths of filter's inner dimension.
    const int64 filter_inner_dim_size = args.out_depth;
    const int64 vectorized_size =
        (filter_inner_dim_size / kPacketSize) * kPacketSize;
    const int64 scalar_size = filter_inner_dim_size - vectorized_size;
    // Calculate required padding and padded output buffer stride.
    const int64 pad_size = scalar_size > 0 ? kPacketSize - scalar_size : 0;
    const int64 padded_filter_stride = vectorized_size + kPacketSize;

    const int64 filter_spatial_size = args.filter_rows * args.filter_cols;
    for (int64 i = 0; i < filter_spatial_size; ++i) {
      const int64 input_base = i * filter_inner_dim_size;
      const int64 output_base = i * padded_filter_stride;
      // Write vectorized length of filter's inner dimension to output.
      for (int64 j = 0; j < vectorized_size; j += kPacketSize) {
        const auto v = Eigen::internal::ploadu<Packet>(filter + input_base + j);
        Eigen::internal::pstoreu<T>(padded_filter + output_base + j, v);
      }
      // Write scalar length of filter's inner dimension to output.
      for (int64 j = 0; j < scalar_size; ++j) {
        padded_filter[output_base + vectorized_size + j] =
            filter[input_base + vectorized_size + j];
      }
      // Pad the remainder of output to vector-register boundary.
      for (int64 j = 0; j < pad_size; ++j) {
        padded_filter[output_base + vectorized_size + scalar_size + j] =
            static_cast<T>(0);
      }
    }
  }
};

// Copies data from local region in 'input' specified by 'out_r' and 'out_'c'
// to 'input_buffer'. The copied data is replicated by factor
// 'args.depth_mulitplier', and padded to vector register-width boundaries so
// that it is aligned for efficient traversal and vector multiply-add by the
// depthwise kernel.
//
// EX:
//   in_depth = 3, depth_multiplier = 2, filter [2, 2], register_width = 4
//
//   input: [batch, in_rows, in_cols, in_depth]
//
//     [a0, a1, a2, b0, b1, b2, ..., e0, e1, e2, f0, f1, f2, ...]
//
//   input_buffer (register boundaries shown):
//     [a0, a0, a1, a1] [a2, a2, 0, 0]   in_row = 0, in_col = 0
//     [b0, b0, b1, b1] [b2, b2, 0, 0]   in_row = 0, in_col = 1
//     [e0, e0, e1, e1] [e2, e2, 0, 0]   in_row = 1, in_col = 0
//     [f0, f0, f1, f1] [f2, f2, 0, 0]   in_row = 1, in_col = 1
//
// Returns replicated and padded data from specified input region in
// 'input_buffer'.

template <typename T>
struct DepthwiseInputCopyOp {
  void operator()(const DepthwiseArgs& args,
                  const int64 padded_filter_inner_dim_size, const int64 out_r,
                  const int64 out_c, const T* input, T* input_buffer) {
    typedef typename Eigen::internal::packet_traits<T>::type Packet;
    static const int64 kPacketSize = (sizeof(Packet) / sizeof(T));

    // Calculate vectorized and scalar (residual) lengths for 'in_depth'.
    const int64 input_vectorized_size =
        (args.in_depth / kPacketSize) * kPacketSize;
    const int64 input_scalar_size = args.in_depth % kPacketSize;

    // Calculate vectorized and scalar (residual) lengths for
    // 'depth_multiplier'. This is used to efficiently replicate data for
    // when 'depth_multiplier' > kPacketSize.
    const int64 dm_vectorized_size =
        (args.depth_multiplier / kPacketSize) * kPacketSize;
    const int64 dm_scalar_size = args.depth_multiplier % kPacketSize;

    // Calculate output padding length.
    const int64 output_scalar_size = args.out_depth % kPacketSize;
    const int64 output_pad_size =
        output_scalar_size > 0 ? kPacketSize - output_scalar_size : 0;

    const int64 replicated_packet_size = kPacketSize * args.depth_multiplier;

    // Iterate through all rows x cols reading 'in_depth' from 'input' and
    // replicating by 'depth_multiplier' into 'input_buffer' (otherwise
    // zero-padding input buffer as needed).
    auto* in_buf = input_buffer;
    const int64 in_r_start = out_r * args.stride - args.pad_rows;
    const int64 in_c_start = out_c * args.stride - args.pad_cols;

    for (int64 f_r = 0; f_r < args.filter_rows; ++f_r) {
      const int64 in_r = in_r_start + f_r;

      for (int64 f_c = 0; f_c < args.filter_cols; ++f_c) {
        const int64 in_c = in_c_start + f_c;

        if (in_r >= 0 && in_r < args.in_rows && in_c >= 0 &&
            in_c < args.in_cols) {
          auto* in = input + (in_r * args.in_cols + in_c) * args.in_depth;
          // Copy vectorized portion of inner dimension.
          for (int64 d = 0; d < input_vectorized_size; d += kPacketSize) {
            auto v = Eigen::internal::ploadu<Packet>(in + d);
            for (int dm = 0; dm < args.depth_multiplier; ++dm) {
              Eigen::internal::pscatter<T, Packet>(in_buf + dm, v,
                                                   args.depth_multiplier);
            }
            in_buf += replicated_packet_size;
          }

          // Copy scalar portion of inner dimension.
          for (int64 d = 0; d < input_scalar_size; ++d) {
            T v = in[input_vectorized_size + d];
            const int64 base = d * args.depth_multiplier;
            if (dm_vectorized_size > 0) {
              // Copy vectorized portion of replicated output.
              // This branch is only taken if 'args.depth_multiplier' is
              // vectorizable (i.e. args.depth_multiplier >= register width).
              auto p = Eigen::internal::pset1<Packet>(v);
              for (int64 dm = 0; dm < dm_vectorized_size; dm += kPacketSize) {
                Eigen::internal::pstoreu<T>(in_buf + base + dm, p);
              }
              // Copy scalar portion of replicated output.
              for (int64 dm = 0; dm < dm_scalar_size; ++dm) {
                in_buf[base + dm_vectorized_size + dm] = v;
              }
            } else {
              // Depth multiplier is less than one packet: scalar copy.
              for (int dm = 0; dm < args.depth_multiplier; ++dm) {
                in_buf[base + dm] = v;
              }
            }
          }
          in_buf += input_scalar_size * args.depth_multiplier;

          // Pad the remainder of the output to vector register boundary.
          for (int64 d = 0; d < output_pad_size; ++d) {
            in_buf[d] = static_cast<T>(0);
          }
          in_buf += output_pad_size;

        } else {
          // Zero pad.
          memset(in_buf, 0, sizeof(T) * padded_filter_inner_dim_size);
          in_buf += padded_filter_inner_dim_size;
        }
      }
    }
  }
};

}  // namespace functor
}  // namespace tensorflow

#endif  // TENSORFLOW_CORE_KERNELS_DEPTHWISE_CONV_OP_H_