}
};
std::unique_ptr<TargetHardwareOperation> CreateArithmeticOp() {
  return std::make_unique<GpuArithmeticOp>();
}

// Currently used for these ops:
// tfl.concatenation / tfl.reshape
class GpuConcatOp : public TargetHardwareOperation {
  double GetOpCost(mlir::Operation* op) const override {
    int64_t count;
    if (ArithmeticCountUtilHelper::GetInputTensorTotalSize(op, &count))
      return kGPUCopyUnitCost * count;