Faster GPU kernels, checked one optimization at a time.
What it does
Reduces a cuTile kernel's latency through isolated optimizations, correctness checks, benchmarks, and keep-or-revert decisions.
When to use it
Use when asked to optimize or iteratively benchmark and refine a cuTile GPU kernel in TileGym.
How to use it
Give it a cuTile kernel in the project; it tunes the kernel, verifies correctness, benchmarks performance, and records the results.
What you provide
- An existing project
Good to know
- acts without asking
What you need · 4
The setup requires an available Blackwell or Ampere GPU node, where subsequent benchmarks must run.
Requires a Python runtime to run the prescribed correctness tests and performance benchmarks.
Requires pytest to run the project's correctness tests and performance benchmarks.
The cuTile kernel code uses cuda.tile; the debug reference specifies its import.
Access · 0
No additional actions listed.