Skip to content

list_kernels.md

Latest commit

 

History

History
158 lines (116 loc) · 3.73 KB

list_kernels.md

File metadata and controls

158 lines (116 loc) · 3.73 KB
Raw

Kernel Selection Tool

Helper script to list and filter TritonBench kernels for batch evaluation in KernelBand experiments.

Location

scripts/list_kernels.py

Purpose

When running KernelBand optimization experiments, you often need to select specific kernels based on:

  • Difficulty level (1-5)
  • Specific index range
  • Total count for testing

This tool helps you:

  1. View available kernels and their difficulty ratings
  2. Filter kernels by difficulty or index
  3. Generate YAML configuration for direct use
  4. Update config files with selected kernels

Usage

View Statistics

# Show kernel count by difficulty
python scripts/list_kernels.py --stats

Output:

======================================================================
Kernel Statistics
======================================================================

Total kernels: 163

Kernels by difficulty:
  Difficulty 1: 45 kernels (easiest)
  Difficulty 2: 58 kernels
  Difficulty 3: 35 kernels
  Difficulty 4: 15 kernels
  Difficulty 5: 10 kernels (hardest)
======================================================================

List Kernels by Difficulty

# List all easy kernels (difficulty 1-2)
python scripts/list_kernels.py --difficulty 1 2

# List medium difficulty kernels
python scripts/list_kernels.py --difficulty 3

# List hard kernels
python scripts/list_kernels.py --difficulty 4 5

List Kernels by Index Range

# List first 10 kernels
python scripts/list_kernels.py --range 0 10

# List kernels 20-30
python scripts/list_kernels.py --range 20 30

Combine Filters

# First 5 easy kernels
python scripts/list_kernels.py --difficulty 1 2 --range 0 5

Generate YAML Output

# Output in YAML format for config
python scripts/list_kernels.py --yaml --difficulty 1 2

Output:

target_kernels:
  - "add_example.py"
  - "vector_addition.py"
  - "elementwise_mul.py"
  # ...

Update Config File Directly

# Update target_kernels in existing config
python scripts/list_kernels.py --config configs/examples/my_config.yaml --difficulty 1 2

# Update with specific range
python scripts/list_kernels.py --config configs/examples/test_config.yaml --difficulty 1 --range 0 5

This will automatically replace the target_kernels field in your config file.

Common Workflows

1. Quick Testing (1-2 Easy Kernels)

# Update test config with 2 easy kernels
python scripts/list_kernels.py --config configs/examples/test_config.yaml --difficulty 1 --range 0 2

2. Progressive Difficulty Experiments

# Create configs for different difficulty levels
python scripts/list_kernels.py --config configs/easy.yaml --difficulty 1 2
python scripts/list_kernels.py --config configs/moderate.yaml --difficulty 3
python scripts/list_kernels.py --config configs/hard.yaml --difficulty 4 5

3. Full Benchmark

# Update config with all kernels
python scripts/list_kernels.py --config configs/full_benchmark.yaml --all

Options Reference

Option Description
--stats Show kernel statistics by difficulty
--all List all kernels
--difficulty N [N ...] Filter by difficulty levels (1-5)
--range START END Filter by index range
--yaml Output in YAML format
--config PATH Update target_kernels in config file

Data Source

The tool reads kernel information from:

data/TritonBench/statistics.json

Each kernel entry contains:

  • file: Kernel filename (e.g., "add_example.py")
  • difficulty: Difficulty rating (1-5)
  • instruction: Natural language description
  • label: Reference implementation

Related Documentation