kernel-exp-history

Kernel Experiment History

Overview

Use the local kernel experiment database to look up prior optimization attempts and record new results after an optimization iteration completes.

Workflow

1) Find prior experiments for inspiration

• Read references/kernel_exp_dataclass.py to understand the database helpers and schema.
• Start with top_experiments(max_results=20) to get a score-sorted list of high-impact experiments.
• If more context is needed, load full entries using get_experiment(exp_id) or list_experiments() and filter by operator_sig, dtype_sig, env, or base_commit.
• Summarize the most relevant patterns (block sizes, memory changes, profiling signals, etc.) before proposing new optimizations.

Query Examples

Example 1: Find similar kernel optimizations

# Search for cache kernel optimizations
from kernel_exp_dataclass import list_experiments

experiments = list_experiments()
cache_exps = [e for e in experiments if 'cache' in e.operator_sig.lower()]

# Sort by score
cache_exps_sorted = sorted(cache_exps, key=lambda x: x.score, reverse=True)

print("Top cache kernel optimizations:")
for exp in cache_exps_sorted[:5]:
    print(f"  {exp.score:.4f}x - {exp.change_summary}")

Example 2: Find best unroll factor

# Compare different unroll factors
unroll_exps = [e for e in experiments if 'unroll' in e.change_summary.lower()]

for exp in unroll_exps:
    factor = 'unknown'
    if 'unroll 4' in exp.detailed_description.lower():
        factor = '4'
    elif 'unroll 8' in exp.detailed_description.lower():
        factor = '8'
    print(f"Unroll {factor}: {exp.score:.4f}x - {exp.operator_sig[:50]}")

Example 3: Learn from failures

# Find what NOT to do
failures = [e for e in experiments if e.score < 0.98 or e.is_buggy]

print("Failed optimizations (learn from these!):")
for exp in failures:
    print(f"  ❌ {exp.change_summary}")
    print(f"     Why: {exp.detailed_description[:100]}...")