What problem does it solve?
Triton kernels often underperform on NVIDIA Blackwell GPUs unless their autotune configurations match TileIR backend expectations. This skill optimizes existing Triton kernels by focusing on TileIR-relevant tuning parameters (notably occupancy and dot-kernel TMA/2CTA patterns) while keeping correctness via compatibility checks.
Core Features & Use Cases
- Kernel classification for targeted optimization: Detects dot-related, norm-like, element-wise, and reduction kernels to choose the right TileIR strategy.
- TileIR-specific autotune config generation: Adds occupancy, num_ctas (2CTA for dot-related kernels), and num_stages guidance appropriate for TileIR.
- TMA descriptor guidance for dot kernels: Provides required wiring patterns (including pre-hook updates) to avoid runtime errors or silent wrong results.
- Validation and benchmarking workflow: Recommends ENABLE_TILE toggling to compare PTX vs TileIR performance and verify correctness.
Quick Start
Run the skill’s classification and apply TileIR optimizations to your kernel by executing a one-step command that classifies and rewrites autotune configurations: python scripts/classify_kernel.py --file kernel.py --apply-optimizations