[Performance] Runtime performance optimization tracking

[zhusy54]

Overview

This issue tracks ongoing runtime performance optimization work for the tensormap_and_ringbuffer runtime on the a2a3 (Ascend 910B/C) platform. Each subtask below represents an independent optimization point.

Platform

All / Unknown

Runtime Variant

tensormap_and_ringbuffer

Git Commit ID

6644bc7

CANN Version

8.5.0.alpha001

Host Platform

Linux (aarch64)

---

Optimization Tasks

Subtask 1: Parallel for dependence optimization

Optimize parallel_for loops by analyzing data dependences to enable more aggressive parallelism.

Currently, parallel_for constructs may be overly conservative in their dependence assumptions, preventing loop iterations from running in parallel when they could safely do so. By introducing dependence analysis, we can identify loops with no loop-carried dependences and schedule them with full parallelism.

Status: Open

---

Subtask 2: Dual slot scheduling for mix subgraph tasks

Support dual slot scheduling for mix subgraphs (subgraphs that contain both AIC and AIV tasks).

Currently, mix subgraph tasks are scheduled conservatively with a single slot, serializing AIC and AIV work even when they could be dispatched concurrently into two hardware slots. Enabling dual slot scheduling for mix subgraphs would allow AIC and AIV kernels to overlap in execution, reducing end-to-end latency.

Status: ✅ Done

Related PRs:

• Add: dual-slot AICPU dispatch payload and two-phase pipelining scheduler #477 Add: dual-slot AICPU dispatch payload and two-phase pipelining scheduler
• Refactor: unify two-phase dispatch and fix sync_start spin loop #553 Refactor: unify two-phase dispatch and fix sync_start spin loop

---

Reproduction

python examples/scripts/run_example.py \
    -k tests/st/a2a3/tensormap_and_ringbuffer/paged_attention/kernels \
    -g tests/st/a2a3/tensormap_and_ringbuffer/paged_attention/golden.py \
    -p a2a3 -d 5 -n 10

Expected Performance

Each subtask is expected to reduce end-to-end latency. Specific numbers TBD after profiling each optimization.

Actual Performance

Current baseline (before optimizations). No regression — these are proactive optimization opportunities.