Iterative Kernel Optimization Loop

Iterative Kernel Optimization Loop

Overview

This skill is the orchestration layer that chains the other skills into an automated, iterative optimization cycle. It turns awesome_kernel_skills from a static knowledge collection into an agent-driven optimization engine, similar to AutoKernel's program.md.

The agent (Cursor / Claude Code / Codex) acts as the execution engine -- this skill tells it what to do, in what order, and when to stop.

The Loop

                    ┌──────────────────────────────┐
                    │  INPUT                        │
                    │  - kernel source (.py)        │
                    │  - target metric & threshold  │
                    │  - target GPU (nvidia/amd)    │
                    └──────────────┬───────────────┘
                                   ▼
              ┌─── Phase 0: BASELINE ───────────────────┐
              │  Run test_*.py → record baseline perf   │
              │  If no test exists, create one first     │
              └──────────────┬──────────────────────────┘
                             ▼
        ┌────────────────────────────────────────────┐
        │  Phase 1: PROFILE                          │
   ┌───▶│  Skill: profiling/SKILL.md                 │
   │    │  Action: run NCU or rocprof                │
   │    │  Output: metrics.json                      │
   │    └──────────────┬─────────────────────────────┘
   │                   ▼
   │    ┌────────────────────────────────────────────┐
   │    │  Phase 2: DIAGNOSE                         │
   │    │  Skill: bottleneck-diagnosis/SKILL.md      │
   │    │  Action: classify memory/compute/latency   │
   │    │  Output: bottleneck type + tier priority    │
   │    └──────────────┬─────────────────────────────┘
   │                   ▼
   │    ┌────────────────────────────────────────────┐
   │    │  Phase 3: OPTIMIZE                         │
   │    │  Skills: tier1~5 SKILL.md (by priority)    │
   │    │  Action: apply ONE optimization at a time  │
   │    │  Rule: modify kernel, keep changes small   │
   │    └──────────────┬─────────────────────────────┘
   │                   ▼
   │    ┌────────────────────────────────────────────┐
   │    │  Phase 4: VERIFY                           │
   │    │  Skill: verification/SKILL.md              │
   │    │  Action: run test → correctness check      │
   │    │  If FAIL → revert change, try next tier    │
   │    └──────────────┬─────────────────────────────┘
   │                   ▼
   │    ┌────────────────────────────────────────────┐
   │    │  Phase 5: BENCHMARK                        │
   │    │  Skill: benchmark/SKILL.md                 │
   │    │  Action: measure TFLOPS or GBps            │
   │    │  Compare vs previous iteration             │
   │    └──────────────┬─────────────────────────────┘
   │                   ▼
   │    ┌────────────────────────────────────────────┐
   │    │  Phase 6: DECIDE                           │
   │    │  - Target met?          → DONE, log result │
   │    │  - Perf improved?       → log, continue    │
   │    │  - Perf regressed?      → revert, try next │
   │    │  - All tiers exhausted? → DONE, log best   │
   │    │  - Max iterations hit?  → DONE, log best   │
   │    └──────────────┬─────────────────────────────┘
   │                   │ not done
   └───────────────────┘

Configuration

Before starting the loop, establish these parameters:

kernel_path: skills/kernels/softmax/triton_template.py
test_path: skills/kernels/softmax/test_softmax.py
target_metric: gbps            # "tflops" or "gbps"
target_value: 800              # target GBps (or TFLOPS)
target_pct_of_peak: 0.70       # alternative: 70% of HBM peak
max_iterations: 5              # stop after N optimization rounds
gpu: auto                      # "nvidia", "amd", or "auto" (detect)

Phase Details

Phase 0: Baseline

1. Detect GPU via scripts/detect_gpu.py
2. Run test_*.py --json to get baseline correctness + performance
3. Record in optimization log:

## Iteration 0 (Baseline)
- GPU: NVIDIA RTX 4090
- Correctness: PASS
- GBps: 1747.0
- Target: 2000 GBps (70% of peak)
- Gap: 253 GBps (12.6%)

Phase 1: Profile

Read profiling/SKILL.md and execute:

• NVIDIA: ncu --set full -o iter{N}_profile python test_*.py
• AMD: rocprof --stats python test_*.py

Extract key metrics into structured form.

Phase 2: Diagnose

Read bottleneck-diagnosis/SKILL.md and apply the flowchart:

IF dram_bw_pct high AND compute_pct low → memory-bound → Tier 1-2, then 3
ELIF compute_pct high                   → compute-bound → Tier 3, then 1, 5
ELIF both low                           → latency-bound → Tier 4, then 3

Output a tier priority list, e.g. [tier2, tier1, tier3].

Phase 3: Optimize

For the highest-priority untried tier:

1. Read the corresponding tier{N}/SKILL.md
2. Read the kernel's own SKILL.md for kernel-specific guidance
3. Apply ONE focused change (do not change multiple things at once)
4. Common changes by tier:

Rule: Save a copy of the kernel before modifying so you can revert.

Phase 4: Verify

Run test_*.py (correctness portion only):

• PASS → proceed to benchmark
• FAIL → revert the change, mark this tier+change as "tried/failed", pick next

Phase 5: Benchmark

Run test_*.py --json (full benchmark):

• Record TFLOPS or GBps
• Compare vs previous iteration

Phase 6: Decide

if current_metric >= target_value:
    status = "TARGET_MET"
elif current_metric > best_metric:
    best_metric = current_metric
    best_kernel = save_snapshot()
    status = "IMPROVED"
elif current_metric <= prev_metric:
    revert_to_previous()
    mark_tier_change_as_tried()
    status = "REGRESSED"

if iteration >= max_iterations or all_tiers_exhausted():
    restore_best_kernel()
    status = "MAX_ITERATIONS"

Optimization Log Format

Append to experiments/README.md after each iteration:

## Optimization: {kernel_name} on {gpu}

| Iter | Action | Tier | Correctness | Metric | Delta | Status |
|------|--------|------|-------------|--------|-------|--------|
| 0 | baseline | - | PASS | 1747 GBps | - | - |
| 1 | increase BLOCK_SIZE 4096→8192 | T1 | PASS | 1820 GBps | +4.2% | improved |
| 2 | vectorized loads (float4) | T2 | PASS | 1955 GBps | +7.4% | improved |
| 3 | multi-row=8 | T1 | FAIL | - | - | reverted |
| 4 | fuse with layernorm | T3 | PASS | 2010 GBps | +2.8% | TARGET_MET |

Final: 2010 GBps (70.2% of peak), 4 iterations, target met.

Guardrails

1. One change per iteration: Never apply two optimizations simultaneously. You cannot attribute improvement (or regression) otherwise.
2. Always verify before benchmark: Never measure performance of an incorrect kernel.
3. Revert on regression: Do not keep a change that makes things slower unless it enables a larger subsequent optimization (document why).
4. Respect platform: Check tier5-arch-nvidia or tier5-arch-amd-rocm before using arch-specific features. Never use Hopper TMA on AMD.
5. Max 5 iterations default: Diminishing returns kick in fast. If stuck after 5 rounds, the kernel may need an algorithmic redesign rather than tuning.
6. Log everything: Every iteration must be recorded. Future optimization attempts should read the log to avoid repeating failed changes.

When Profiling Is Not Available

If NCU/rocprof cannot be run (e.g. container restrictions, no root), fall back to:

1. Use test_*.py --json benchmark numbers as the only signal
2. Classify based on kernel type from the kernel's SKILL.md (it states bound type)
3. Follow the tier priority for that bound type
4. Iterate on autotune configs (Tier 1) first -- this is always safe

Agent Instructions

When asked to optimize a kernel iteratively:

1. Read this skill first to understand the loop structure
2. Establish target: ask the user or derive from GPU peak specs
3. Run Phase 0 to get baseline
4. Execute the loop: Profile → Diagnose → Optimize → Verify → Benchmark → Decide
5. After each iteration, update the optimization log in experiments/
6. Stop when target is met, max iterations reached, or all viable tiers are exhausted
7. Report final result with summary table

When profiling tools are available, always start with Phase 1 (profile). When they are not, skip to Phase 3 using the kernel's documented bound type.

Relationship to Other Skills

optimize-loop (this skill) ─── orchestrates ───▶ all other skills
  │
  ├─ Phase 1 → profiling/SKILL.md
  ├─ Phase 2 → bottleneck-diagnosis/SKILL.md
  ├─ Phase 3 → tier1~5/SKILL.md + kernels/*/SKILL.md
  ├─ Phase 4 → verification/SKILL.md
  └─ Phase 5 → benchmark/SKILL.md

References

• AutoKernel program.md -- The original iterative agent-driven kernel optimization loop
• KernelSkill dual-level memory -- Reusable optimization skills + short-term backtracking memory