hipfire-kernel-atlas

hipfire-kernel-atlas

Use this skill when the task is to explain or visualize how a hipfire quant format and kernel use an AMD GPU ISA target. The primary tool is scripts/kernel_atlas.py; this skill is a thin agent wrapper around that CLI.

Core Workflow

1. Collect or locate Atlas rows

   • Prefer existing JSONL under .codeinsight+research/kernel-atlas/runs/.
   • For AR prefill/decode, collect with collect-ar.
   • Use --profile-prefill / --profile-decode for AR rows when the user wants the ISA view scoped to runtime-hot kernels and tagged by op role.
   • For speculative decode, collect with collect-dflash.
   • Keep raw run data in .codeinsight+research/; it is ignored and may be private.

2. Attach ISA metadata

   • Use --isa-file for one known HSACO/code object.
   • Use --isa-dir .hipfire_kernels/<arch> plus --isa-filter for a bounded set.
   • Prefer --isa-output <path>.json so multiple rows reference one manifest.

3. Attach dispatch/source provenance

   • Use --dispatch-provenance when rows have profiled kernel names.
   • Prefer --dispatch-output <path>.json so multiple rows reference one manifest.
   • Treat dispatch references as evidence to inspect, not proof of a unique runtime branch.
   • Prefer rows with a known arch; source ranking is target-arch-aware when arch-specific kernel files exist.

4. Render the ISA Fit View

   • Use .agents/skills/hipfire-kernel-atlas/render-fit.sh.
   • If a row has artifacts.profile_kernels, the view joins profiled kernel names to ISA object kernel names/symbols and summarizes only matched objects.
   • If a row has dispatch provenance, the view prints hot-kernel op/source/dispatch attribution.
   • Report the visual plus a short readout of likely limit and left on table.

5. Ask Atlas for candidate experiments

   • Use python3 scripts/kernel_atlas.py suggest --row ... --isa ... --dispatch ....
   • Prefer --format markdown for humans and JSON for automation.
   • Let suggest auto-load default history from .codeinsight+research/kernel-atlas/tasks/; use --history only for extra history paths.
   • Treat suggestions as an experiment queue, not as predicted wins.
   • Each suggestion should name the lever type, hot kernel, files, risk, rationale, and eval contract.

6. Create an optimization task

   • Use python3 scripts/kernel_atlas.py task to turn a row into task.json and TASK.md.
   • Include --allowed-file for every path an agent may edit.
   • Include correctness commands for DFlash or risky runtime changes.
   • Generated tasks strip known profiling/instrumentation env from eval and preserve the original row env as baseline.row_env.

7. Evaluate a candidate

   • Use python3 scripts/kernel_atlas.py eval --task ... --runs 5 --warmup-runs 1 --output-dir ....
   • Use --refresh-baseline first to write baseline.json; use --baseline <baseline.json> for candidate comparisons.
   • Report result.json status, selected metric median, speedup, stability, and any failed command output tail.
   • Treat the local ledger.jsonl as experiment lineage, not a public benchmark.
   • If status is needs_baseline, do not claim a speedup; refresh or provide a clean baseline first.

Commands

Render an existing row:

.agents/skills/hipfire-kernel-atlas/render-fit.sh \
  --row .codeinsight+research/kernel-atlas/runs/atlas.jsonl \
  --row-index 0 \
  --isa .codeinsight+research/kernel-atlas/runs/isa.json

Collect a small AR smoke with ISA:

python3 scripts/kernel_atlas.py collect-ar \
  --model ~/.hipfire/models/qwen3.5-0.8b.mq4 \
  --workload qwen3.5-0.8b \
  --model-size 0.8b \
  --quant mq4 \
  --prefill 32 \
  --gen 5 \
  --kv-mode asym3 \
  --profile-prefill \
  --profile-decode \
  --isa-dir .hipfire_kernels/gfx1030 \
  --isa-filter 'gemm_hfq4g256|gemv_hfq4g256' \
  --isa-output .codeinsight+research/kernel-atlas/runs/isa-gfx1030.json \
  --dispatch-provenance \
  --dispatch-output .codeinsight+research/kernel-atlas/runs/dispatch-gfx1030.json \
  --output .codeinsight+research/kernel-atlas/runs/atlas-gfx1030.jsonl

Suggest candidate experiments from a profiled row:

python3 scripts/kernel_atlas.py suggest \
  --row .codeinsight+research/kernel-atlas/runs/atlas-gfx1201.jsonl \
  --row-index 1 \
  --isa .codeinsight+research/kernel-atlas/runs/isa-gfx1201.json \
  --dispatch .codeinsight+research/kernel-atlas/runs/dispatch-gfx1201.json \
  --format markdown

Create a bounded task from a profiled row:

python3 scripts/kernel_atlas.py task \
  --row .codeinsight+research/kernel-atlas/runs/atlas-gfx1201.jsonl \
  --row-index 1 \
  --isa .codeinsight+research/kernel-atlas/runs/isa-gfx1201.json \
  --dispatch .codeinsight+research/kernel-atlas/runs/dispatch-gfx1201.json \
  --allowed-file kernels/src/gemv_hfq4g256_multirow.hip \
  --output-dir .codeinsight+research/kernel-atlas/tasks/gfx1201-gemv-r4

Create a PyTorch-shape task for non-Qwen work:

python3 scripts/kernel_atlas.py task-pytorch \
  --name llama-rmsnorm-shape \
  --op rmsnorm \
  --input-shape 1,2048,4096 \
  --dtype float16 \
  --eval-command 'python3 bench_rmsnorm.py' \
  --allowed-file kernels/src/rmsnorm_candidate.hip \
  --output-dir .codeinsight+research/kernel-atlas/tasks/llama-rmsnorm-shape

Refresh a stable baseline and then evaluate a candidate:

python3 scripts/kernel_atlas.py eval \
  --task .codeinsight+research/kernel-atlas/tasks/gfx1201-gemv-r4/task.json \
  --runs 5 \
  --warmup-runs 1 \
  --refresh-baseline \
  --output-dir .codeinsight+research/kernel-atlas/tasks/gfx1201-gemv-r4/eval-baseline

python3 scripts/kernel_atlas.py eval \
  --task .codeinsight+research/kernel-atlas/tasks/gfx1201-gemv-r4/task.json \
  --baseline .codeinsight+research/kernel-atlas/tasks/gfx1201-gemv-r4/eval-baseline/baseline.json \
  --runs 5 \
  --warmup-runs 1 \
  --output-dir .codeinsight+research/kernel-atlas/tasks/gfx1201-gemv-r4/eval-001

Interpretation Rules

• Treat the view as ISA fit, not full hardware occupancy. True occupancy also needs counters, wave residency, clocks, cache behavior, and launch overlap.
• If matrix units are available but observed matrix ops are zero, ask whether the workload phase should route through WMMA/MFMA or whether it is a decode GEMV path where memory/launch dominates.
• If VGPR/SGPR/spills are high, prioritize register pressure and spill removal before claiming a bandwidth win.
• If the row is DFlash, do not treat tok/s alone as correctness evidence. Run the DFlash coherence gate before claiming a spec-decode improvement.
• If eval reports unstable, do not claim a win or regression; tighten the run shape or rerun after DPM/thermal state settles.
• For PyTorch-shape tasks, treat the eval command as the source of truth until Atlas has a real PyTorch profiler/extractor producer.
• If the worktree is dirty, cite the row's provenance.diff_md5 and avoid comparing it as a shipped baseline.

Good Agent Output

Include:

• the rendered ASCII fit view, or the most relevant section of it
• the row path and ISA manifest path
• arch, quant, phase, and shape bucket
• runtime metric used for the readout
• one concise interpretation of likely limit and left on table

Avoid:

• calling the heuristic a roofline model
• claiming a perf win from smoke runs
• mixing rows from different prompts or dirty binaries without saying so