Module telemetry 

Launch telemetry: timing, occupancy, and register usage reporting.

This module provides post-launch diagnostics for GPU kernel execution. After a kernel launch, LaunchTelemetry captures grid/block dimensions, GPU-side timing, achieved occupancy, and register usage. A TelemetryCollector accumulates entries and produces a TelemetrySummary with per-kernel aggregation.

Telemetry can be exported to JSON, CSV, or Chrome trace format via TelemetryExporter.

Example

use oxicuda_launch::telemetry::{LaunchTelemetry, TelemetryCollector};

let mut collector = TelemetryCollector::new(1000);
let entry = LaunchTelemetry::new("vector_add", (4, 1, 1), (256, 1, 1))
    .with_elapsed_ms(0.5)
    .with_achieved_occupancy(0.85);
collector.record(entry);
let summary = collector.summary();
assert_eq!(summary.total_launches, 1);

Structs

KernelStats

Aggregated statistics for a single kernel across multiple launches.

LaunchTelemetry

Telemetry data collected after a single kernel launch.

TelemetryCollector

Accumulates LaunchTelemetry entries and produces summaries.

TelemetryExporter

Export telemetry data in various formats.

TelemetrySummary

Summary of all collected telemetry data.

Enums

SmVersion

GPU architecture version for occupancy estimation.

Functions

estimate_occupancy

Estimates theoretical occupancy for a kernel launch configuration.