There are four configuration objects for each Band’s component.

Each configuration field is optional or required. If a field is optional, then it is guaranteed that the default value exists. If a field is required, a configuration cannot be generated by RuntimeConfigBuilder without specifying the field.

Enumeration Types

  • BandSchedulerType

    • kFixedDeviceFixedWorker
    • kRoundRobin
    • kShortestExpectedLatency
    • kFixedDeviceGlobalQueue
    • kHeterogeneousEarliestFinishTime
    • kLeastSlackTimeFirst
    • kHeterogeneousEarliestFinishTimeReserved
  • CPUMaskFlags

    • kAll
    • kLittle
    • kBig
    • kPrimary
  • DeviceFlags

    • kCPU
    • kGPU
    • kDSP
    • kNPU
  • SubgraphPreparationType

    • kNoFallbackSubgraph
    • kFallbackPerWorker
    • kUnitSubgraph
    • kMergeUnitSubgraph


  • online [type: bool, default: true]: Profile online if true, offline if false.
  • num_warmups [type: int, default: 1]: The number of warmup runs before profile.
  • num_runs [type: int, default: 1]: The number of runs for profile
  • copy_computation_ratio [type: std::vector<int>, default: [30000, ...]]: The ratio of computation to input-output copy. Used for latency estimation. The size of the list should be the same as the number of devices.
  • smoothing_factor [type: float, default: 0.1]: The momentum to reflect current profiled data. <updateed_profile> = <smoothing_factor> * <curr_profile> + (1. - <smoothing_factor>) * <prev_profile>.
  • profile_data_path [type: std::string, default: ""]: The input path to the file for offline profile results. If not specified, this will be ignored and will not generate the result file.


  • schedule_window_size [type: int, default: INT_MAX]: The size of window that scheduler will use.
  • schedulers [type: std::vector<SchedulerType>, required]: The types of schedulers. If N schedulers are specified, N queues will be generated.
  • cpu_mask [type: CPUMaskFlags, default: kAll]: CPU masks to set CPU affinity.
  • log_path [type: std::string, default: ""]: The output path to the file for planner’s log. If not specified, this will be ignored and will not generate the result file.


  • workers [type: std::vector<DeviceFlags>, default: [kCPU, kGPU, ...]]: The list of target devices. By default, one worker per device is generated.
  • cpu_masks [type: std::vector<CPUMaskFlags>, default: [kAll, kAll, ...]]: CPU masks to set CPU affinity. The size of the list must be the same as the size of workers.
  • num_threads [type: std::vector<int>, default: [1, 1, ...]]: The number of threads. The size of the list must be the same as the size of workers.
  • allow_worksteal [type: bool, default: false]: Work-stealing is enabled if true, disabled if false.
  • availability_check_interval_ms [type: int, default: 30_000]: The interval for checking availability of devices. Used for detecting thermal throttling.


  • RuntimeConfig contains ProfileConfig, PlannerConfig and WorkerConfig.
  • minimum_subgraph_size [type: int, default: 7]: The minimum subgraph size. If candidate subgraph size is smaller than this, the subgraph will not be created.
  • subgraph_preparation_type [type: SubgraphPreparationType, default: kMergeUnitSubgraph]: For fallback schedulers, determine how to generate candidate subgraphs.
  • cpu_mask [type: CPUMaskFlags, default: kAll]: The CPU mask for Band Engine.

RuntimeConfigBuilder API

RuntimeConfigBuilder delegates all builder that inherits ConfigBuilder. It is a friend class of all the other ConfigBuilder classes, so make sure to not change their members in RuntimeConfigBuilder.

Exmaple Usage

RuntimeConfigBuilder b;
auto config = b.AddOnline(false)  // Default was `true`
                  .AddSmoothingFactor(0.3)  // Default was `0.1`
                  .AddSchedulers({SchedulerType::kRoundRobin, SchedulerType::kLeastSlackTimeFirst})  // Required field.


All Add* methods are idempotent, i.e. multiple calls behaves the same as a single call.

  • AddOnline(bool online)
  • AddNumWarmups(int num_warmups)
  • AddNumRuns(int num_runs)
  • AddCopyComputationRatio(std::vector<int> copy_computation_ratio)
  • AddSmoothingFactor(float smoothing_factor)
  • AddProfileLogPath(std::string profile_data_path)
  • AddPlannerLogPath(std::string planner_log_path)
  • AddScheduleWindowSize(int schedule_window_size)
  • AddSchedulers(std::vector<SchedulerType> schedulers)
  • AddPlannerCPUMask(CPUMaskFlags cpu_masks)
  • AddWorkers(std::vector<DeviceFlags> workers)
  • AddWorkerCPUMasks(std::vector<CPUMaskFlags> cpu_masks)
  • AddWorkerNumThreads(std::vector<int> num_threads)
  • AddAllowWorkSteal(bool allow_worksteal)
  • AddAvailabilityCheckIntervalMs(int32_t availability_check_interval_ms)
  • AddMinimumSubgraphSize(int minimum_subgraph_size)
  • AddSubgraphPreparationType(SubgraphPreparationType subgraph_preparation_type)
  • AddCPUMask(CPUMaskFlags cpu_mask)