optimize-op

[//]: # "SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved." [//]: # "SPDX-License-Identifier: Apache-2.0" # Optimize Op Read `.agents/guidance/OPTIMIZATION_GUIDELINES.md` and follow it as the source of truth. Use `tools/optimize_op.py` for the deterministic gates and the versioned MR summary. Do not duplicate or weaken the guidance here. ## Workflow 1. Run `python3 tools/optimize_op.py <Op> --phase preflight`. Resolve every readiness `GAP` in a separate non-`perf:` coverage commit before optimizing. 2. Survey and baseline the full union expanded from all declared `basic` and `advanced` configurations. Iterate profile → lead → one change → correctness/performance evaluation → review/refactor. Run `python3 tools/refactor_op.py <Op> --phase assess` before every kept `perf:` commit. Re-profile after each win and stop only at the guidance's measured ridge or three-post-win-strike condition. 3. Initialize the concise v1 MR summary from exact optimized case keys: ```bash python3 tools/optimize_op.py <Op> --phase summary \ --benchmark-base <ref> --optimized-cases-file <file> \ --state provisional --bottleneck Memory-bound \ --profile-evidence '<measured evidence>' --out <description> ``` Use `Compute-bound` when supported by the profile. Keep the human-authored assessment, six hard-evidence checklist lines, and one to five learnings current. Regenerate scope, categories, per-SKU Impact, and Layout comparison; do not hand-edit those derived fields. Maintain the canonical companion `## Memory footprint` section outside the bounded v1 block; check its summary item only after ODO-9 passes. 4. Refresh the description with reference-SKU evidence and finalize it: ```bash python3 tools/optimize_op.py <Op> --phase summary \ --results <description> --state final --out <description> ``` Require every configured reference SKU. Keep the visible warning for any non-reference local SKU retained for context. 5. Run the final gate: ```bash python3 tools/optimize_op.py <Op> --phase evidence \ --base <ref> --results <description> ``` Require pixelwise-equality evidence for every changed configuration, updated and non-regressing comparable baselines, full-surface profile/lead coverage (agent-verified per the guidance; the gate does not check it), hard evidence for all six checklist items, exhausted leads, stable API/ABI, and `perf:` commit hygiene. Treat ODO-9 as the memory-footprint authority; its `MANUAL` result requires explicit human review before submission. Resolve every `GAP` and unresolved `MANUAL`, then rerun. Never weaken or silently retune tests, fabricate baselines, check an evidence item without its hard proof, weaken the ODO-9 memory policy, or change public API/ABI.