Crate veloq_nsys_query 

veloq-nsys-query — per-subcommand query implementations.

Each subcommand owns one module here. Phase 0 ships summary; stats, search, inspect, timeline, gaps, correlate follow.

Re-exports

pub use error::NsysQueryError;

pub use error::NsysQueryResult;

pub use error::SqlPhase;

pub use event_ref::EventRef;

pub use event_ref::NvtxContext;

pub use kind_filter::KindFilter;

pub use row_id::EventKind;

pub use row_id::RowId;

Modules

column_map

Schema-probe helpers shared by inspect and search.

concurrency

veloq concurrency <trace> — GPU kernel/transfer overlap extraction.

correlate

veloq correlate <row_id>... — single-event causal-chain reverse lookup.

docgen

Auto-generated reference-doc bodies.

error

event_ref

EventRef — the shared “row that references one trace event” shape every list-of-events response returns.

gaps

veloq gaps --min Nms — GPU idle-bubble detection.

graph_replays

veloq graph-replays <trace> — CUDA graph replay decomposition.

hardware

veloq hardware <trace> — CPU / GPU / NIC inventory.

inspect

veloq inspect <trace> <row_id> [<row_id> …] — full event details.

kind_filter

Type-safe “which event kinds is this request about” selector.

kind_policy

Shared request-validation policies for the kind-aware verbs (stats, search). Two silent-drop traps these policies prevent:

kind_sql

Per-event-kind SQL + label fragments shared across query commands.

metrics

veloq metrics — hardware-performance counter / CPU sample / scheduler-event queries.

ncu_command

Generate an Nsight Compute command for one CUDA kernel event.

nvtx_attribution

Shared NVTX→GPU attribution CTE.

nvtx_parent

Rank-and-pick-innermost NVTX parent attribution — SQL plumbing.

nvtx_projection

Shared NVTX→GPU-event projection CTE templates.

nvtx_reverse

Reverse NVTX attribution — “which NVTX range was this event launched inside?”.

row_id

Wire-format event identifiers.

search

veloq search <trace> ... — filter events into a list of row_ids plus a few headline columns. Designed as the inspect entry-point.

slices

veloq slices --pattern <glob> — NVTX-range attribution views.

stats

veloq stats <trace> — aggregated GPU work statistics.

stats_by_size

veloq stats --by size — bytes-as-aggregate-unit stats.

summary

veloq summary <trace> — one-shot overview of a trace.

timeline

veloq timeline <trace> --interval Nms — time-bucketed GPU activity.

viz_timeline

NSys static timeline SVG figure export.

Functions

check_limit

Reject limit == 0 at the public-API boundary. The CLI also guards via CommonFilters::limit_or, but library callers can hand-build a request with limit: 0, which silently zeroes total_matched (the count comes off SQL rows that LIMIT 0 suppressed). Call this at the top of every run().

decode_global_tid

Decode an nsys globalTid into (pid, tid). NSys packs four fields into the 64-bit slot:

module_basename

NSys records modules as absolute paths (/usr/lib/x86_64-linux-gnu/libc.so.6) or Windows-style (C:\Windows\system32\foo.dll). For hotspot tables / callchains agents (and humans) want the basename — libc.so.6 / foo.dll. Centralised here so the metrics --type cpu-sampling path and inspect cpu_sample:N agree on what “module name” means without two copies of the slice-on-/ logic drifting.

open_scoped

Shared verb preamble: validate the limit, open the trace, and resolve the --from/--to window to an absolute (start_ns, end_ns). Used by the verbs whose run() opens with exactly this sequence (stats / search / stats_by_size). Verbs that interleave other validation between these steps — gaps’ --min check, timeline‘s --interval check, slices’ deferred window resolution — keep their own preamble so error precedence is unchanged.

parse_positive_duration

Parse a CLI duration flag (100us / 1.2s / 42ns / …) into ns, rejecting non-positive results. Wraps veloq_core::time::parse_duration_ns with a flag-name aware typed error and a “must be positive” guard. Used by every command that accepts a bucket/interval-like duration flag.