Kernel Forge: generate, check, and profile
Create CUDA templates, select a visible GPU, and separate measured timings from modeled Roofline limits.
The problem
A kernel experiment needs a reproducible starting implementation, correct GPU selection, a numerical check, and an interpretable timing result. A theoretical ceiling alone cannot tell you how fast a kernel ran.
What I built
I built a Python CLI for device discovery, CUDA template generation, checked benchmarking, and Roofline analysis. Device selection follows CUDA-visible ordinals so local masks and multi-GPU hosts can be handled explicitly.
Usable tool
Who it is for: CUDA developers learning or automating kernel generation and performance analysis.
First task: Install from the checkout and run forge doctor --json before generating a kernel.
What you can produce: CUDA-visible device information; generated templates can then be checked and timed on your GPU.
Current scope: A Python CLI with CUDA templates and modeled Roofline analysis. Actual execution requires a CUDA GPU/toolkit; modeled limits are separate from measured timings.
Start with the example Source and documentation Step-by-step tutorials
A first useful result
Generate vector addition and inspect its measured latency and modeled arithmetic intensity. Then compare naive and tiled matrix templates at the same shape. The tutorial saves doctor output and JSON results before drawing a conclusion.
Evidence and limits
Portable tests cover device selection and CLI behavior; GPU checks require actual CUDA execution. Roofline ceilings use reference/fallback specifications and modeled traffic, not measured DRAM counters. Current templates are CUDA; this page does not imply an implemented Triton path.