android-12.0.0_r34/s

/* Copyright 2017 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.
==============================================================================*/
#include "tensorflow/compiler/tf2xla/kernels/image_resize_ops.h"

#include "absl/strings/str_format.h"
#include "absl/types/span.h"
#include "tensorflow/compiler/jit/xla_activity.pb.h"
#include "tensorflow/compiler/jit/xla_activity_listener.h"
#include "tensorflow/compiler/tf2xla/shape_util.h"
#include "tensorflow/compiler/tf2xla/type_util.h"
#include "tensorflow/compiler/tf2xla/xla_helpers.h"
#include "tensorflow/compiler/tf2xla/xla_op_kernel.h"
#include "tensorflow/compiler/tf2xla/xla_op_registry.h"
#include "tensorflow/compiler/xla/array4d.h"
#include "tensorflow/compiler/xla/client/lib/constants.h"
#include "tensorflow/compiler/xla/client/xla_builder.h"
#include "tensorflow/core/framework/kernel_def_builder.h"
#include "tensorflow/core/framework/register_types.h"
#include "tensorflow/core/lib/math/math_util.h"

namespace tensorflow {
namespace {

// We implement bilinear interpolation by upsampling followed by convolution.
// The basic idea is as follows. To scale from NxN to RxR:
//
//    1. S := (N - 1) /  gcd(N-1, R-1)
//    2. k := (R - 1) /  gcd(N-1, R-1)
//    3. Convolution((2k-1)x(2k-1), stride=S, lhs_dilation=k, padding=k-1)
//
// For example, to Scale from 7x7 -> 15x15:
//
//    1. S := (7-1) / gcd(7-1, 15-1) = 6 / gcd(6, 14) = 6 / 2 = 3
//    2. k := (15 - 1) / gcd(7-1, 15-1) = 14 / gcd(6, 14) = 14 / 2 = 7
//    3. Convolution(15x15, stride=3, lhs_dilation=7, padding=2)
//
//
// The 7x7 -> 15x15 case is much too large to write out in full as an
// example. The smallest interesting example is 3x3 -> 4x4.
//
// S := 2
// k := 3
//
// 00 03 06    00 00 00 00 00 00 00 00 00 00 00      00 02 04 06
// 09 12 15 -> 00 00 00 00 00 00 00 00 00 00 00   -> 06 08 10 12
// 18 21 24    00 00 00 00 00 03 00 00 06 00 00      12 14 16 18
//             00 00 00 00 00 00 00 00 00 00 00      18 20 22 24
//             00 00 00 00 00 00 00 00 00 00 00
//             00 00 09 00 00 12 00 00 15 00 00
//             00 00 00 00 00 00 00 00 00 00 00
//             00 00 00 00 00 00 00 00 00 00 00
//             00 00 18 00 00 21 00 00 24 00 00
//             00 00 00 00 00 00 00 00 00 00 00
//             00 00 00 00 00 00 00 00 00 00 00
//
// with the following convolutional kernel, with stride [2, 2]:
//       1 2 3 2 1
//       2 4 6 4 2
// 1/9 * 3 6 9 6 3
//       2 4 6 4 2
//       1 2 3 2 1
// Note that the convolution kernel matrix is separable and thus we can instead
// use 2 consecutive 1D kernel of the dimension 2k-1, along each axis.

// Computes the size of the convolutional kernel and stride to use when resizing
// from in_size to out_size.
struct ResizeConvolutionDims {
  // Size of the kernel to use.
  std::vector<int64> kernel_size;  // k

  // Stride of the convolution to use.
  std::vector<int64> stride;  // S
};
ResizeConvolutionDims ComputeResizeConvolutionParameters(
    absl::Span<const int64> in_size, absl::Span<const int64> out_size,
    bool align_corners) {
  CHECK_EQ(in_size.size(), out_size.size());
  int num_spatial_dims = in_size.size();
  ResizeConvolutionDims dims;
  dims.kernel_size.resize(num_spatial_dims);
  dims.stride.resize(num_spatial_dims);
  for (int i = 0; i < num_spatial_dims; ++i) {
    if (in_size[i] == 1) {
      // We must handle input size 1 specially because XLA convolution does
      // not allow stride 0.
      dims.stride[i] = dims.kernel_size[i] = 1;
    } else if (out_size[i] == 1) {
      // If in_size[i] > 1 but out_size[i] == 1, then we slice out the first
      // entry before resizing.
      dims.stride[i] = dims.kernel_size[i] = 1;
    } else {
      // The scaling factor changes depending on the alignment of corners.
      const int64 in_size_factor = align_corners ? in_size[i] - 1 : in_size[i];
      const int64 out_size_factor =
          align_corners ? out_size[i] - 1 : out_size[i];

      int64 gcd = MathUtil::GCD(static_cast<uint64>(in_size_factor),
                                static_cast<uint64>(out_size_factor));
      dims.stride[i] = in_size_factor / gcd;
      dims.kernel_size[i] = out_size_factor / gcd;
    }
  }
  return dims;
}

// The upper padding of the input needed by ConvGeneralDilated calls is
// determined by solving two related relationships (assuming rhs_dilation == 0):
// 1. dilated_input_dim = lower_padding + upper_padding
//                        + lhs_dilation * (in_size - 1) + 1
// 2. dilated_input_dim = (2 * dims.kernel-size - 1)
//                        + dims.stride * (out_size - 1)
int64 CalculateUpperPadding(int64 in_size, int64 out_size, int64 kernel_size,
                            int64 stride) {
  int64 padding = (2 * kernel_size - 1) + (out_size - 1) * stride -
                  (kernel_size - 1) - 1 - (kernel_size * (in_size - 1));

  return padding;
}

// Form a 2D convolution kernel like:
//       1 2 3 2 1
//       2 4 6 4 2
// 1/9 * 3 6 9 6 3
//       2 4 6 4 2
//       1 2 3 2 1
// by multiplying two 1D kernels of the form:
// 1/3 * [1 2 3 2 1]
// If the 2D kernel would be very large, the 1D kernel can be applied once in
// each dimension due to the symmetry of the kernel along all axis to reduce the
// computational intensity.
xla::XlaOp MakeBilinear1DKernel(xla::XlaBuilder* builder,
                                xla::PrimitiveType type, int64 n) {
  std::vector<float> kernel(n * 2 - 1);
  for (int64 i = 0; i < n; ++i) {
    float v = (i + 1.0f) / n;
    kernel[i] = v;
    kernel[n * 2 - 2 - i] = v;
  }
  return xla::ConvertElementType(xla::ConstantR1<float>(builder, kernel), type);
}

// Unlike the bilinear kernel, which is triangular, the nearest neighbor
// kernel is a square. For example, a 1D kernel with n=3 would look like
// [0 1 1 1 0]
// and n=4 would look like
// [0 0 1 1 1 1 0].
// Note that in the second case, the kernel is not symmetric and we default
// to the right (because an existing non TPU kernel
// for nearest neighbor resize already chose to default to the right,
// so we want to be consistent).
xla::XlaOp MakeNearestNeighbor1DKernel(xla::XlaBuilder* builder,
                                       xla::PrimitiveType type, int64 n) {
  std::vector<float> kernel(n * 2 - 1, 0.0f);
  std::fill(&kernel[n / 2], &kernel[(3 * n) / 2], 1.0f);

  return xla::ConvertElementType(xla::ConstantR1<float>(builder, kernel), type);
}

// Kernels with more than 16 spatial elements are considered intense and the
// kernel should be applied to each dimension independently.
const int64 kMax2DKernelSize = 16;

xla::XlaOp MakeGeneralResizeKernel(xla::XlaBuilder* builder,
                                   xla::PrimitiveType type,
                                   absl::Span<const int64> kernel_size,
                                   int64 channels, bool is_kernel_bilinear) {
  auto make_kernel_func =
      is_kernel_bilinear ? MakeBilinear1DKernel : MakeNearestNeighbor1DKernel;

  std::vector<int64> depthwise_kernel_sizes = {
      (2 * kernel_size[0] - 1), (2 * kernel_size[1] - 1), channels, 1};
  auto depthwise_kernel =
      xla::BroadcastInDim(make_kernel_func(builder, type, kernel_size[1]),
                          depthwise_kernel_sizes, /*broadcast_dimensions=*/{1});

  return xla::Mul(depthwise_kernel,
                  make_kernel_func(builder, type, kernel_size[0]),
                  /*broadcast_dimensions=*/{0});
}

xla::XlaOp MakeGeneralResizeKernelInDim(xla::XlaBuilder* builder,
                                        xla::PrimitiveType type,
                                        absl::Span<const int64> kernel_size,
                                        int64 channels, int64 dim,
                                        bool is_kernel_bilinear) {
  auto make_kernel_func =
      is_kernel_bilinear ? MakeBilinear1DKernel : MakeNearestNeighbor1DKernel;

  std::vector<int64> depthwise_kernel_sizes = {
      dim == 0 ? (2 * kernel_size[0] - 1) : 1,
      dim == 1 ? (2 * kernel_size[1] - 1) : 1, channels, 1};
  return xla::BroadcastInDim(make_kernel_func(builder, type, kernel_size[dim]),
                             depthwise_kernel_sizes,
                             /*broadcast_dimensions=*/{dim});
}

xla::XlaOp BroadcastSpatialDimensions(xla::XlaBuilder* builder,
                                      const xla::XlaOp& input,
                                      int32 spatial_dimensions_offset,
                                      absl::Span<const int64> in_size,
                                      absl::Span<const int64> out_size) {
  // Add broadcasts to handle expanding from a size == 1 dimension to a
  // size > 1 dimension.
  auto broadcast_shape_or_status = builder->GetShape(input);
  if (!broadcast_shape_or_status.ok()) {
    return builder->ReportError(broadcast_shape_or_status.status());
  }
  xla::Shape broadcast_shape = broadcast_shape_or_status.ValueOrDie();
  for (int32 i = 0; i < in_size.size(); ++i) {
    if (in_size[i] == 1 && out_size[i] > 1) {
      broadcast_shape.set_dimensions(spatial_dimensions_offset + i,
                                     out_size[i]);
    }
  }
  return xla::BroadcastInDim(input, broadcast_shape.dimensions(),
                             /*broadcast_dimensions=*/{0, 1, 2, 3});
}

xla::XlaOp ResizeUsingDilationAndConvolution(
    xla::XlaBuilder* builder, const xla::XlaOp& input, xla::PrimitiveType type,
    const int num_spatial_dims, absl::Span<const int64> in_size,
    absl::Span<const int64> out_size, const int64 channels,
    const bool align_corners, bool is_kernel_bilinear) {
  // Picture for a 1x3 to 1x4 bilinear resize:
  // stride = 2, kernel size = 3
  // Input:
  // 3 6 9
  // Input with dilation and padding:
  // 0 0 3 0 0 6 0 0 9 0 0
  // Convolution kernel:
  // 1/3 * [1 2 3 2 1]
  // Output:
  // 3 5 7 9
  xla::ConvolutionDimensionNumbers dimension_numbers;
  dimension_numbers.set_input_batch_dimension(0);
  dimension_numbers.set_output_batch_dimension(0);
  dimension_numbers.set_input_feature_dimension(num_spatial_dims + 1);
  dimension_numbers.set_output_feature_dimension(num_spatial_dims + 1);
  for (int i = 0; i < num_spatial_dims; ++i) {
    dimension_numbers.add_input_spatial_dimensions(1 + i);
    dimension_numbers.add_output_spatial_dimensions(1 + i);
    dimension_numbers.add_kernel_spatial_dimensions(i);
  }
  dimension_numbers.set_kernel_input_feature_dimension(num_spatial_dims + 1);
  dimension_numbers.set_kernel_output_feature_dimension(num_spatial_dims);

  ResizeConvolutionDims dims =
      ComputeResizeConvolutionParameters(in_size, out_size, align_corners);

  if (dims.kernel_size[0] * dims.kernel_size[1] >
      kMax2DKernelSize * kMax2DKernelSize) {
    BroadcastOptimizationRemark(
        XlaOptimizationRemark::SLOW_IMAGE_RESIZE_DIMENSIONS,
        absl::StrFormat("%dx%d", dims.kernel_size[0], dims.kernel_size[1]))
        .IgnoreError();
  }

  xla::XlaOp output;

  // Concatenation and padding below currently assumes num_spatial_dims is 2 to
  // prevent needless code complexity.
  CHECK_EQ(num_spatial_dims, 2)
      << "ResizeUsingDilationAndConvolution pads only 2 dimensions currently.";
  std::vector<int64> upper_padding(num_spatial_dims);
  for (int i = 0; i < num_spatial_dims; ++i) {
    upper_padding[i] = dims.kernel_size[i] - 1;
  }
  xla::XlaOp input_data = input;

  if (!align_corners) {
    // When Tensorflow does not align_corners, the resize indexing can access
    // beyond the upper bound and is instead clamped to prevent out of bounds
    // reads. This is conceptually the same as extending the edges of the input.
    // We emulate this by copying the last row/column of the input.
    // Calculate what padding would be needed then determine how far to extend
    // the border before lhs dilation.
    std::vector<int64> num_extended(num_spatial_dims);
    upper_padding[0] = CalculateUpperPadding(
        in_size[0], out_size[0], dims.kernel_size[0], dims.stride[0]);
    upper_padding[1] = CalculateUpperPadding(
        in_size[1], out_size[1], dims.kernel_size[1], dims.stride[1]);
    num_extended[0] = upper_padding[0] / (dims.kernel_size[0]);
    num_extended[1] = upper_padding[1] / (dims.kernel_size[1]);

    const int64 batch_dim_size =
        builder->GetShape(input).ValueOrDie().dimensions(0);
    if (num_extended[0] > 0) {
      auto slice = xla::Slice(
          input_data, {0, in_size[0] - 1, 0, 0},
          {batch_dim_size, in_size[0], in_size[1], channels}, {1, 1, 1, 1});
      for (int i = 0; i < num_extended[0]; i++) {
        input_data = xla::ConcatInDim(builder, {input_data, slice}, 1);
      }
    }

    if (num_extended[1] > 0) {
      auto slice = xla::Slice(
          input_data, {0, 0, in_size[1] - 1, 0},
          {batch_dim_size, in_size[0] + num_extended[0], in_size[1], channels},
          {1, 1, 1, 1});
      for (int i = 0; i < num_extended[1]; i++) {
        input_data = xla::ConcatInDim(builder, {input_data, slice}, 2);
      }
    }

    // Setting in_size to (in_size + num_extended) due to the above Slice and
    // ConcatInDim. Recalculate needed padding after the above Slice/Concat.
    upper_padding[0] =
        CalculateUpperPadding(in_size[0] + num_extended[0], out_size[0],
                              dims.kernel_size[0], dims.stride[0]);
    upper_padding[1] =
        CalculateUpperPadding(in_size[1] + num_extended[1], out_size[1],
                              dims.kernel_size[1], dims.stride[1]);
  }

  // Split convolutions into independent dimensions if they would be a very
  // large kernel or if one or more of the dimensions are already equal.
  bool decompose_resize =
      in_size[0] == out_size[0] || in_size[1] == out_size[1] ||
      dims.kernel_size[0] * dims.kernel_size[1] >= kMax2DKernelSize;
  if (!decompose_resize) {
    xla::XlaOp kernel = MakeGeneralResizeKernel(builder, type, dims.kernel_size,
                                                channels, is_kernel_bilinear);
    output =
        xla::ConvGeneralDilated(input_data, kernel, dims.stride,
                                /*padding=*/
                                {{dims.kernel_size[0] - 1, upper_padding[0]},
                                 {dims.kernel_size[1] - 1, upper_padding[1]}},
                                /*lhs_dilation=*/dims.kernel_size,
                                /*rhs_dilation=*/{1, 1}, dimension_numbers,
                                /*feature_group_count=*/channels);
  } else {
    output = input_data;
    if (in_size[0] != out_size[0]) {
      xla::XlaOp kernel0 = MakeGeneralResizeKernelInDim(
          builder, type, dims.kernel_size, channels, 0, is_kernel_bilinear);
      output = xla::ConvGeneralDilated(
          output, kernel0, {dims.stride[0], 1},
          /*padding=*/
          {{dims.kernel_size[0] - 1, upper_padding[0]}, {0, 0}},
          /*lhs_dilation=*/{dims.kernel_size[0], 1},
          /*rhs_dilation=*/{1, 1}, dimension_numbers,
          /*feature_group_count=*/channels);
    }

    if (in_size[1] != out_size[1]) {
      xla::XlaOp kernel1 = MakeGeneralResizeKernelInDim(
          builder, type, dims.kernel_size, channels, 1, is_kernel_bilinear);
      output = xla::ConvGeneralDilated(
          output, kernel1, {1, dims.stride[1]},
          /*padding=*/
          {{0, 0}, {dims.kernel_size[1] - 1, upper_padding[1]}},
          /*lhs_dilation=*/{1, dims.kernel_size[1]},
          /*rhs_dilation=*/{1, 1}, dimension_numbers,
          /*feature_group_count=*/channels);
    }
  }

  // Add broadcasts to handle expanding from a size == 1 dimension to a
  // size > 1 dimension.
  return BroadcastSpatialDimensions(
      builder, output, /*spatial_dimensions_offset=*/1, in_size, out_size);
}

xla::XlaOp ResizeUsingDilationAndConvolutionGradOp(
    xla::XlaBuilder* builder, const xla::XlaOp& grad, xla::PrimitiveType type,
    const int num_spatial_dims, absl::Span<const int64> in_size,
    absl::Span<const int64> grad_size, const int64 channels,
    const bool align_corners, bool is_kernel_bilinear) {
  ResizeConvolutionDims dims =
      ComputeResizeConvolutionParameters(in_size, grad_size, align_corners);

  // To form the backward convolution, we keep the kernel unchanged (it is
  // already symmetric) and swap the roles of strides and LHS dilation.
  xla::ConvolutionDimensionNumbers dimension_numbers;
  dimension_numbers.set_input_batch_dimension(0);
  dimension_numbers.set_output_batch_dimension(0);
  dimension_numbers.set_input_feature_dimension(num_spatial_dims + 1);
  dimension_numbers.set_output_feature_dimension(num_spatial_dims + 1);
  for (int i = 0; i < num_spatial_dims; ++i) {
    dimension_numbers.add_input_spatial_dimensions(i + 1);
    dimension_numbers.add_output_spatial_dimensions(i + 1);
    dimension_numbers.add_kernel_spatial_dimensions(i);
  }
  dimension_numbers.set_kernel_input_feature_dimension(num_spatial_dims + 1);
  dimension_numbers.set_kernel_output_feature_dimension(num_spatial_dims);
  xla::XlaOp output;
  if (dims.kernel_size[0] * dims.kernel_size[1] < kMax2DKernelSize) {
    xla::XlaOp kernel = MakeGeneralResizeKernel(builder, type, dims.kernel_size,
                                                channels, is_kernel_bilinear);

    // Broadcast the input kernel where the forward op expanded from a size == 1
    // dimension to a size > 1 dimension. This has the effect of summing the
    // gradient contributions in that dimension.
    kernel = BroadcastSpatialDimensions(
        builder, kernel, /*spatial_dimensions_offset=*/0, in_size, grad_size);

    output = xla::ConvGeneralDilated(
        grad, kernel, /*window_strides=*/dims.kernel_size,
        /*padding=*/
        {{dims.kernel_size[0] - 1, dims.kernel_size[0] - 1},
         {dims.kernel_size[1] - 1, dims.kernel_size[1] - 1}},
        /*lhs_dilation=*/dims.stride,
        /*rhs_dilation=*/{1, 1}, dimension_numbers,
        /*feature_group_count=*/channels);
  } else {
    xla::XlaOp kernel0 = MakeGeneralResizeKernelInDim(
        builder, type, dims.kernel_size, channels, 0, is_kernel_bilinear);
    xla::XlaOp kernel1 = MakeGeneralResizeKernelInDim(
        builder, type, dims.kernel_size, channels, 1, is_kernel_bilinear);

    // Broadcast the input kernel where the forward op expanded from a
    // size == 1 dimension to a size > 1 dimension. This has the effect of
    // summing the gradient contributions in that dimension.
    if (in_size[0] == 1 && grad_size[0] > 1) {
      kernel0 = BroadcastSpatialDimensions(builder, kernel0,
                                           /*spatial_dimensions_offset=*/0, {1},
                                           {grad_size[0]});
    }
    if (in_size[1] == 1 && grad_size[1] > 1) {
      kernel1 = BroadcastSpatialDimensions(builder, kernel0,
                                           /*spatial_dimensions_offset=*/0,
                                           in_size, grad_size);
    }

    output = xla::ConvGeneralDilated(
        grad, kernel0, /*window_strides=*/{dims.kernel_size[0], 1},
        /*padding=*/
        {{dims.kernel_size[0] - 1, dims.kernel_size[0] - 1}, {0, 0}},
        /*lhs_dilation=*/{dims.stride[0], 1},
        /*rhs_dilation=*/{1, 1}, dimension_numbers,
        /*feature_group_count=*/channels);

    output = xla::ConvGeneralDilated(
        output, kernel1, /*window_strides=*/{1, dims.kernel_size[1]},
        /*padding=*/
        {{0, 0}, {dims.kernel_size[1] - 1, dims.kernel_size[1] - 1}},
        /*lhs_dilation=*/{1, dims.stride[1]},
        /*rhs_dilation=*/{1, 1}, dimension_numbers,
        /*feature_group_count=*/channels);
  }

  // If in_size[i] > 1 and grad_size[i] == 1, pad the output in dimension i.
  // Opposite of the slice performed by the forward op.
  xla::PaddingConfig padding = xla::MakeNoPaddingConfig(4);
  bool pad_output = false;
  for (int i = 0; i < num_spatial_dims; ++i) {
    if (in_size[i] > 1 && grad_size[i] == 1) {
      pad_output = true;
      padding.mutable_dimensions(1 + i)->set_edge_padding_high(in_size[i] - 1);
    }
  }
  if (pad_output) {
    output = xla::Pad(output, xla::Zero(builder, type), padding);
  }
  return output;
}

void GeneralCompile(XlaOpKernelContext* ctx, bool align_corners_,
                    bool is_kernel_bilinear) {
  xla::XlaBuilder* b = ctx->builder();

  TensorShape input_shape = ctx->InputShape(0);
  OP_REQUIRES(ctx, input_shape.dims() == 4,
              errors::InvalidArgument("input must be 4-dimensional",
                                      input_shape.DebugString()));
  // First dimension always assumed to be batch
  const int64 batch = input_shape.dim_size(0);
  std::vector<int64> in_size = {input_shape.dim_size(1),
                                input_shape.dim_size(2)};
  // Last/4th dimension always assumed to be num channels
  const int64 channels = input_shape.dim_size(3);
  OP_REQUIRES(ctx, in_size[0] > 0 && in_size[1] > 0,
              errors::InvalidArgument("input size must be positive, got [",
                                      in_size[0], ",", in_size[1], "]"));

  std::vector<int64> out_size;
  OP_REQUIRES_OK(ctx, ctx->ConstantInputAsIntVector(1, &out_size));
  OP_REQUIRES(ctx, out_size.size() == 2,
              errors::InvalidArgument("output size must be length 2, got ",
                                      out_size.size()));
  OP_REQUIRES(ctx, out_size[0] > 0 && out_size[1] > 0,
              errors::InvalidArgument("output size must be positive, got [",
                                      out_size[0], ",", out_size[1], "]"));

  const int num_spatial_dims = 2;

  xla::XlaOp input = ctx->Input(0);
  xla::PrimitiveType input_type = ctx->input_xla_type(0);

  // If in_size[i] > 1 and out_size[i] == 1, slice out the first input in
  // dimension i.
  bool slice_input = false;
  for (int i = 0; i < num_spatial_dims; ++i) {
    if (in_size[i] > 1 && out_size[i] == 1) {
      // If in_size[i] > 1 but out_size[i] == 1, then we slice out the first
      // entry before resizing.
      slice_input = true;
      in_size[i] = 1;
    }
  }
  if (slice_input) {
    input = xla::Slice(input, {0, 0, 0, 0},
                       {batch, in_size[0], in_size[1], channels}, {1, 1, 1, 1});
  }

  // Output is always type float if 'is_kernel_bilinear' is true.
  // GPU with integer input also uses float, because XLA
  // integer convolution on CuDNN is either not supported or not allowed
  // directly.
  xla::PrimitiveType original_input_type = input_type;
  if (is_kernel_bilinear || (xla::primitive_util::IsIntegralType(input_type))) {
    input = xla::ConvertElementType(input, xla::F32);
    input_type = xla::F32;
  }

  for (int dim = 0; dim < in_size.size(); ++dim) {
    // If the pairwise_distance function more accurately estimated performance,
    // this threshold could be reduced.
    constexpr int64 kSmallDimThreshold = 1 << 10;
    if (in_size[dim] > out_size[dim] || out_size[dim] < kSmallDimThreshold) {
      std::vector<int64> next_size = in_size;
      next_size[dim] = out_size[dim];
      input = ResizeUsingDilationAndConvolution(
          b, input, input_type, num_spatial_dims, in_size, next_size, channels,
          align_corners_, is_kernel_bilinear);
      in_size[dim] = next_size[dim];
    }
  }

  // This function approximates the cost of a bilinear resize from a src_size to
  // a dst_size. The accuracy is okay, but empirically, the algorithm makes some
  // suboptimal choices. A better cost model would improve performance.
  auto pairwise_distance = [align_corners_](int64 src_size, int64 dst_size) {
    auto params = ComputeResizeConvolutionParameters({src_size}, {dst_size},
                                                     align_corners_);
    return params.stride[0];
  };

  for (int dim = 0; dim < in_size.size(); ++dim) {
    std::vector<int64> distances(out_size[dim] + 1);
    std::vector<int64> next_step(out_size[dim] + 1);
    for (int64 i = distances.size() - 2; i >= in_size[dim]; --i) {
      distances[i] = INT64_MAX;
      for (int64 j = i + 1; j < distances.size(); ++j) {
        int64 distance = pairwise_distance(i, j) + distances[j];
        if (distance < distances[i]) {
          distances[i] = distance;
          next_step[i] = j;
        }
      }
    }

    while (in_size[dim] != out_size[dim]) {
      auto next_size = in_size;
      next_size[dim] = next_step[in_size[dim]];
      input = ResizeUsingDilationAndConvolution(
          b, input, input_type, num_spatial_dims, in_size, next_size, channels,
          align_corners_, is_kernel_bilinear);
      in_size[dim] = next_size[dim];
    }
  }

  // Bilinear always outputs float, but nearest neighbor keeps the original type
  if (!is_kernel_bilinear && original_input_type != input_type) {
    input = xla::ConvertElementType(input, original_input_type);
  }
  ctx->SetOutput(0, input);
}
}  // namespace

ResizeNearestNeighborOp::ResizeNearestNeighborOp(OpKernelConstruction* ctx)
    : XlaOpKernel(ctx) {
  OP_REQUIRES_OK(ctx, ctx->GetAttr("align_corners", &align_corners_));
  OP_REQUIRES(
      ctx, align_corners_ == true,
      errors::Unimplemented("ResizeNearestNeighbor with align_corners=False "
                            "is not yet implemented"));
  OP_REQUIRES_OK(ctx, ctx->GetAttr("half_pixel_centers", &half_pixel_centers_));
  OP_REQUIRES(ctx, half_pixel_centers_ == false,
              errors::Unimplemented(
                  "ResizeNearestNeighbor with half_pixel_centers=True is "
                  "not yet implemented"));
}

void ResizeNearestNeighborOp::Compile(XlaOpKernelContext* ctx) {
  GeneralCompile(ctx, align_corners_, is_kernel_bilinear_);
}

REGISTER_XLA_OP(Name("ResizeNearestNeighbor").CompileTimeConstantInput("size"),
                ResizeNearestNeighborOp);

ResizeBilinearOp::ResizeBilinearOp(OpKernelConstruction* ctx)
    : XlaOpKernel(ctx) {
  OP_REQUIRES_OK(ctx, ctx->GetAttr("align_corners", &align_corners_));
  OP_REQUIRES_OK(ctx, ctx->GetAttr("half_pixel_centers", &half_pixel_centers_));
  OP_REQUIRES(
      ctx, half_pixel_centers_ == false,
      errors::Unimplemented("ResizeBilinear with half_pixel_centers=True is "
                            "not yet implemented"));
}

void ResizeBilinearOp::Compile(XlaOpKernelContext* ctx) {
  GeneralCompile(ctx, align_corners_, is_kernel_bilinear_);
}

REGISTER_XLA_OP(Name("ResizeBilinear").CompileTimeConstantInput("size"),
                ResizeBilinearOp);

ResizeBilinearGradOp::ResizeBilinearGradOp(OpKernelConstruction* ctx)
    : XlaOpKernel(ctx) {
  OP_REQUIRES_OK(ctx, ctx->GetAttr("align_corners", &align_corners_));
  OP_REQUIRES_OK(ctx, ctx->GetAttr("half_pixel_centers", &half_pixel_centers_));
  OP_REQUIRES(
      ctx, align_corners_ == true,
      errors::Unimplemented("ResizeBilinearGrad with align_corners=False is "
                            "not yet implemented"));
  OP_REQUIRES(ctx, half_pixel_centers_ == false,
              errors::Unimplemented(
                  "ResizeBilinearGrad with half_pixel_centers=True is "
                  "not yet implemented"));

  DataType output_dtype;
  OP_REQUIRES_OK(ctx, ctx->GetAttr("T", &output_dtype));
  OP_REQUIRES_OK(ctx, DataTypeToPrimitiveType(output_dtype, &output_type_));
}

void ResizeBilinearGradOp::Compile(XlaOpKernelContext* ctx) {
  xla::XlaBuilder* b = ctx->builder();

  TensorShape input_shape = ctx->InputShape(1);
  OP_REQUIRES(ctx, input_shape.dims() == 4,
              errors::InvalidArgument("input must be 4-dimensional",
                                      input_shape.DebugString()));
  const int64 batch = input_shape.dim_size(0);
  std::vector<int64> in_size = {input_shape.dim_size(1),
                                input_shape.dim_size(2)};
  const int64 channels = input_shape.dim_size(3);
  OP_REQUIRES(ctx, in_size[0] > 0 && in_size[1] > 0,
              errors::InvalidArgument("input size must be positive, got [",
                                      in_size[0], ",", in_size[1], "]"));

  TensorShape grad_shape = ctx->InputShape(0);
  OP_REQUIRES(ctx, grad_shape.dims() == 4,
              errors::InvalidArgument("gradient must be 4-dimensional",
                                      grad_shape.DebugString()));
  const int64 grad_batch = grad_shape.dim_size(0);
  const std::vector<int64> grad_size = {grad_shape.dim_size(1),
                                        grad_shape.dim_size(2)};
  const int64 grad_channels = grad_shape.dim_size(3);
  OP_REQUIRES(ctx, batch == grad_batch,
              errors::InvalidArgument(
                  "activations and gradients must have the same batch size (",
                  batch, " vs. ", grad_batch, ")"));
  OP_REQUIRES(ctx, grad_size[0] > 0 && grad_size[1] > 0,
              errors::InvalidArgument("gradient size must be positive, got [",
                                      grad_size[0], ",", grad_size[1], "]"));
  OP_REQUIRES(
      ctx, channels == grad_channels,
      errors::InvalidArgument(
          "activations and gradients must have the same number of channels (",
          channels, " vs. ", grad_channels, ")"));

  const int num_spatial_dims = 2;

  xla::XlaOp grad = ctx->Input(0);

  xla::XlaOp output = grad;
  while (in_size != grad_size) {
    if (in_size[0] != 1 && in_size[1] != 1) {
      std::vector<float> k = {
          (static_cast<float>(grad_size[0]) - 1) / ((in_size[0] - 1) * 2),
          (static_cast<float>(grad_size[1]) - 1) / ((in_size[1] - 1) * 2)};
      if ((k[0] == std::floor(k[0])) && (k[1] == std::floor(k[1])) &&
          k[0] > 1 && k[1] > 1) {
        std::vector<int64> next_grad_size = {(in_size[0] - 1) * 2 + 1,
                                             (in_size[1] - 1) * 2 + 1};
        output = ResizeUsingDilationAndConvolutionGradOp(
            b, grad, xla::F32, num_spatial_dims, in_size, next_grad_size,
            channels, align_corners_, true);
        grad = output;
        in_size = next_grad_size;
      } else {
        output = ResizeUsingDilationAndConvolutionGradOp(
            b, grad, xla::F32, num_spatial_dims, in_size, grad_size, channels,
            align_corners_, true);
        in_size = grad_size;
      }
    } else {
      output = ResizeUsingDilationAndConvolutionGradOp(
          b, grad, xla::F32, num_spatial_dims, in_size, grad_size, channels,
          align_corners_, true);
      in_size = grad_size;
    }
  }

  output = xla::ConvertElementType(output, output_type_);
  ctx->SetOutput(0, output);
}

REGISTER_XLA_OP(Name("ResizeBilinearGrad"), ResizeBilinearGradOp);

}  // namespace tensorflow