[Code Health] Re-tier log levels: replace V0..V9 with debug/info/timing/warn/error; fix CANN dlog INFO seeding (a2a3 + a5)

[ChaoZheng109]

Category

Technical Debt (cleanup, refactor)

Component

Other — logging subsystem (host HostLogger + AICPU device_log + --log-level surface), spanning both a2a3 and a5.

Description

Two coupled problems with the current logging design:

1. simpler self-seeds CANN's driver INFO stream (device performance tax).
When ASCEND_GLOBAL_LOG_LEVEL is unset, simpler_init calls dlog_setlevel(-1, HostLogger::level(), 0). HostLogger::level() defaults to INFO and -1 targets every CANN module, so we open the CANN driver's own INFO logs as a side effect of wanting our own host verbosity. Those device-side writes land inside the measured device wall (~4–6% on some workloads; see #1409). It is not a CANN default — it is self-inflicted.

2. The V0..V9 INFO sub-tier scheme is a two-axis (severity, info_v) design that is more complex than what the code uses. In practice device logging is binary (V0/V9 only — see #1428), and the v-tiers force a --log-level surface (v0..v9) wired through conftest.py, _log.py::_split_threshold, log_config.py, and the docs.

Target: a single-axis ladder debug / info / timing / warn / error, default timing, replacing V0..V9. STRACE moves from LOG_INFO_V9 to the timing level; simpler_init maps the device (CANN) level so timing → WARN. Result:

• default timing → host emits only STRACE (+ warn/error); device level = WARN → no driver-log tax on any run (not just rounds > 1);
• info / debug → device INFO back on for debugging, without the DEBUG flood;
• WARN/ERROR always visible; one --log-level still drives both host and device.

Benchmark measurement is unaffected either way — [STRACE] rides host stderr + the rtMemcpy device-phase readback, not the CANN log.

This supersedes the per-run ASCEND_GLOBAL_LOG_LEVEL env-var mitigation in #1409, and depends on the footprint reduction in #1428 landing first.

Location

• Device seeding: src/common/platform/onboard/host/c_api_shared.cpp (simpler_init → dlog_setlevel(-1, HostLogger::level(), 0))
• STRACE emission: src/common/log/include/common/strace.h (LOG_INFO_V9 → timing)
• Two-axis machinery: python/simpler/_log.py (_split_threshold), simpler_setup/log_config.py (LOG_LEVEL_CHOICES), conftest.py (V0..V9 registration), src/common/log/host_log.cpp (info_v), src/common/platform/onboard/aicpu/device_log.cpp (g_log_info_v), src/common/log/include/common/unified_log.h (LOG_INFO_V0..V9)
• Call-site migration, both arches: src/a2a3/** and src/a5/** LOG_INFO_V* sites
• Docs: docs/logging.md, docs/testing.md, docs/troubleshooting/device-error-codes.md, .claude/rules/running-onboard.md

Proposed Fix

Phased (after #1428):

1. Introduce the ladder debug / info / timing / warn / error (single severity axis, timing between info and warn); make timing the default. Emit STRACE at timing.
2. Wire the device mapping in simpler_init: debug→DEBUG, info→INFO, timing→WARN, warn→WARN, error→ERROR — so default timing no longer seeds CANN INFO.
3. Remove V0..V9 (or keep as deprecated aliases: v9→timing, v5–v8→info, v0–v4→debug) and migrate remaining call sites + the --log-level surface + docs.

Open decisions to settle in the issue: hard-remove vs deprecated aliases for v0..v9; whether the 109 must-see device diagnostics map to info or debug.

Priority

Medium (real device-side overhead, but #1409 provides a stopgap; depends on #1428)

---

Context / root-cause discussion: #1409. Depends on: #1428. Related: #409 (hot-path logging overhead).