DFX and Profiling¶
Every profiling / diagnostics reference lives in this directory. Start with the framework doc if you are adding a collector, or jump straight to the per-feature doc for the data you want.
Analysis CLIs that consume these outputs are documented in simpler_setup/tools/README.md.
Framework and conventions¶
| Document | What it covers |
|---|---|
| Profiling Framework | Collector architecture, host/device split, how perf levels gate collection |
| Profiling / DFX Configuration Naming Rules | Naming conventions every new DFX knob must follow |
| Profiling Name Map | func_id → human-readable name mapping consumed by the trace tools |
| DFX Backpressure | Per-lane block-on-contention across collectors |
| DFX Buffer Capacity Audit (a2a3) | Per-buffer capacity accounting and overflow behavior |
Timing and scheduling¶
| Document | What it covers |
|---|---|
| L2 Timing | host_wall / device_wall / Effective / Orch / Sched breakdown from [STRACE] |
| Host runtime trace markers | The [STRACE] marker set emitted by the host runtime, and simpler.trace for emitting your own |
| The host_build_graph bind phases | What the bind segments are, the switches that expose them, and how to compare two branches on the qwen and dsv4 decode cases |
| Device-side phase timing | Fixed AICPU phases and the variable-phase plan |
| Scheduler-Overhead Model | Is the scheduler the bottleneck, or starved — the model behind --overhead |
Per-task traces¶
| Document | What it covers |
|---|---|
| Chip Swimlane Profiling | Per-task timing and scheduler phases across cores |
| Core Swimlane Profiling | Intra-core AICore pipeline trace for a single task |
| PMU Profiling | Per-task AICore hardware counters |
Data and dependencies¶
| Document | What it covers |
|---|---|
| Args Dump | Per-task argument capture, and replaying a task from it |
| dep_gen | Complete per-submit dependency graph, tensor-annotated |
| Scope Stats | Per-scope resource-usage peaks (ring fill, dep pool) |
Related, outside this directory¶
| Document | What it covers |
|---|---|
| Log System | Log levels and sinks — the host/device logging path, distinct from profiling collectors |
| troubleshooting/ | Device error codes and timeout defaults, for when a run fails rather than runs slowly |