Struct anira_plan_info#

struct anira_plan_info#

One plan of the report, addressed by its dense index: the variant it runs, the backend it runs on and its budget. Tier 2, struct_size first; enumerated by anira_plan_report_plans at the caller’s stride. budget_ms is the budget of this one plan: under a MEASURED budget what the warm-up pass measured for it, under an EXPLICIT one the contract’s figure; the Hard promise is the worst case across every plan, the session sized for the largest. The warm-up fields (warmup_iterations, warmup_outcome, measured_ms, median_ms) say what the pass measured, and the real-time fields (rt_factor, rt_parallel, rt_verdict) whether the plan keeps up on this machine, an estimate at prepare; the rows are read at any time after the prepare.

Public Members

uint32_t struct_size#

sizeof(anira_plan_info) of the caller’s header.

uint32_t variant#

The index of the variant in the inference stage; 0 in this pre-release.

uint32_t engine#

anira_engine; ANIRA_ENGINE_CUSTOM for a custom engine, which engine_id names.

uint32_t provider#

anira_provider of the plan; ANIRA_PROVIDER_CUSTOM beside a provider_id.

const char *engine_id#

NULL for a built-in engine; the registered name of a custom engine, valid while the report is.

double budget_ms#

The per-inference budget of this plan in milliseconds: its warm-up figure under a MEASURED budget (the maximum of its last four inferences), the contract’s figure under an EXPLICIT one.

uint32_t engine_flags#

The flags of a custom engine’s descriptor (anira_engine_desc.flags: the ANIRA_ENGINE_FLAG_* promises); 0 for a built-in engine. A tail field: a caller whose header ends before it gets its rows without it.

uint32_t reserved#

0.

const char *provider_id#

NULL for a provider the enum names; the name of the plan’s custom provider (anira_provider), valid while the report is. A tail field: a caller whose header ends before it gets its rows without it.

uint32_t memory_kind#

anira_memory_kind: the physical memory the plan’s provider computes in, as the handler’s context probed it at the prepare that loaded the plan’s model (anira_edge_info.memory_kind): the fact a device provider’s slot classes follow. ANIRA_MEMORY_UNIFIED on a provider that runs over host memory. A tail field: a caller whose header ends before it gets its rows without it.

uint32_t warmup_iterations#

The inferences the warm-up pass ran for this plan at this prepare (those of the stored measurement under ANIRA_WARMUP_OUTCOME_REUSED); 0 without a pass. A tail field: a caller whose header ends before it gets its rows without it.

double measured_ms#

What the pass measured for this plan, in milliseconds: the maximum of its last four inference times (of all of them below four), each timed on the inference thread that ran it, never below one clock step; under an explicit budget the observation beside the declared figure; 0 without a pass. A tail field: a caller whose header ends before it gets its rows without it.

double median_ms#

The median of the same times, the jitter at a glance beside measured_ms; 0 without a pass. A tail field: a caller whose header ends before it gets its rows without it.

uint32_t warmup_outcome#

anira_warmup_outcome: how this plan’s pass ended. A tail field: a caller whose header ends before it gets its rows without it.

uint32_t rt_parallel#

The inferences of the session that run at once, n: 1 for an exclusive session (a stateful model, a declared State pair), else the model’s instances as the pool clamps them; the same in every row of a session. A tail field: a caller whose header ends before it gets its rows without it.

double rt_factor#

The real-time factor: the plan’s inference time over the anchor’s hop period (T * host_rate / (1000 * hop), the latency model’s kappa), T being measured_ms, or budget_ms without a measurement (ANIRA_RT_UNMEASURED). 1.0 is an inference as long as the audio it produces; the plan keeps up while rt_factor / rt_parallel stays below 1. A tail field: a caller whose header ends before it gets its rows without it.

uint32_t rt_verdict#

anira_rt_verdict: the thresholds applied to rt_factor / rt_parallel (below 0.75 comfortable, below 1 marginal, else not real-time), or ANIRA_RT_UNMEASURED without a measurement. A host that plays live holds a NOT_REALTIME plan until its user agrees, and never gates an offline render. A tail field: a caller whose header ends before it gets its rows without it.

uint32_t reserved2#

0.