Value rhs = op.getOperand(1);
  TransposeOp rhs_transpose_op = CreateTransposeOp(
      /*input=*/rhs, kNchwToNhwcPermutation, rewriter);

  auto new_binary_op =
      rewriter.create<OpT>(op.getLoc(), lhs_pre_transpose, rhs_transpose_op);

  // NHWC -> NCHW for the output, to match the shapes of `op`'s users.
  TransposeOp output_transpose_op = CreateTransposeOp(
      /*input=*/new_binary_op, kNhwcToNchwPermutation, rewriter);

  for (int64_t i = 0, e = out_transpose.size(); i < e; ++i) {
    out_inverse_transpose->at(out_transpose[i]) = i;
  }

  *lhs = createTransposeOp(*lhs, loc, lhs_transpose, rewriter);
  *rhs = createTransposeOp(*rhs, loc, rhs_transpose, rewriter);
  return success();
}

template <int I>
inline int64_t ProdShapeWithIndexInTuple(
    ArrayRef<int64_t> shape,