// LHS transpose permutation must be a NCHW->NHWC permutation.
template <typename OpT>
void DeferRhsTransposeForBinaryOp(OpT op, PatternRewriter& rewriter) {
  auto transpose_op = cast<TransposeOp>(op.getOperand(0).getDefiningOp());
  Value lhs_pre_transpose = transpose_op.getOperand();

  // NCHW -> NHWC for the right-hand side, to match the operand's shape.
  Value rhs = op.getOperand(1);

  // CHECK: %[[RES_TRANSPOSE:[0-9]*]] = "tf.Transpose"(%[[CONV2D]], %[[RES_PERM]])
  // CHECK: return %[[RES_TRANSPOSE]]

  %0 = "tf.Conv2D"(%input, %filter)
       {
         data_format = "NCHW",
         dilations = [1, 2, 3, 4],
         explicit_paddings = [1, 2, 3, 4, 5, 6, 7, 8],
         padding = "EXPLICIT",
         strides = [5, 6, 7, 8]

    reverse[permutation[i]] = i;
  }
  return reverse;
}

SmallVector<int64_t, 4> GetDataFormatPermutation(StringRef from, StringRef to) {
  if (from == "NHWC" && to == "NCHW") {
    return {0, 3, 1, 2};
  } else if (from == "NCHW" && to == "NHWC") {
    return {0, 2, 3, 1};
  } else {
    return {};
  }
}

// Shuffle elements in the `attr` according to the permutation. Optional

  device: "/job:localhost/replica:0/task:0/device:GPU:0"
  attr {
    key: "T"
    value {
      type: DT_HALF
    }
  }
  attr {
    key: "data_format"
    value {
      s: "NCHW"
    }
  }
  attr {
    key: "dilations"
    value {
      list {
        i: 1
        i: 1
        i: 1
        i: 1
      }
    }
  }
  attr {

    const DenseIntElementsAttr& dense_attr, PatternRewriter& builder);

// Returns a NHWC shaped type from an NCHW shaped type op.
// For example- Given a Composite op that wraps a core.aten.avg_pool2d, this
// returns the return type of the tfl.average_pool_2d emitted. Note that the
// aten.avg_pool2d works with the NCHW layout while tfl.average_pool_2d assumes
// NHWC.
ShapedType GetNhwcReturnTypeFromNchw(Operation* old_op);

  device: "/job:localhost/replica:0/task:0/device:GPU:0"
  attr {
    key: "T"
    value {
      type: DT_FLOAT
    }
  }
  attr {
    key: "data_format"
    value {
      s: "NCHW"
    }
  }
  attr {
    key: "dilations"
    value {
      list {
        i: 1
        i: 1
        i: 1
        i: 1
      }
    }
  }
  attr {

  op: "Conv2D"
  input: "input"
  input: "conv_net_2d/conv_2d_0/w/read"
  attr {
    key: "T"
    value {
      type: DT_FLOAT
    }
  }
  attr {
    key: "data_format"
    value {
      s: "NCHW"
    }
  }
  attr {
    key: "dilations"
    value {
      list {
        i: 1
        i: 1
        i: 1
        i: 1
      }
    }
  }
  attr {

inline constexpr StringRef kQuantizationMethodAttr = "_quantization_method";

// Permutation from the NHWC tensor format to NCHW. This is an inverse
// permutation of `kNchwToNhwcPermutation`.
inline constexpr std::array<int64_t, 4> kNhwcToNchwPermutation = {0, 3, 1, 2};

// Permutation from the NCHW tensor format to NHWC. This is an inverse
// permutation of `kNchwToNhwcPermutation`.

  // Also FusedBatchNorm in training mode prefers NCHW data format. Check if all
  // users can efficiently use NHWC data format?
  if (one_by_one && trivial_strides && trivial_dilations) {
    return "NHWC";
  }

  // If filter spatial dimensions are unknown or not 1x1 we prefer NCHW, because
  // it's the fastest option on NVIDIA GPUs with cuDNN library support.
  return "NCHW";
}

//
// Description:
//   It accumulates all the values from out_backprop into the feature dimension.
//   For NHWC data format, the feature dimension is the last. For NCHW data
//   format, the feature dimension is the third-to-last.
Status BiasAddGrad(AbstractContext* ctx,
                   AbstractTensorHandle* const out_backprop,