kernels

A collection of high-performance CUDA kernels. Each kernel lives in its own directory with a self-contained build, correctness tests, and benchmark results.

kernels

kernel	category	target	peak BW
rmsnorm	normalization	H100 SXM (sm_90a)	80.6% @ prefill-70B
qr	linear algebra	H100 SXM (sm_90a)	baseline: torch.geqrf

structure

kernels/
└── <category>/
    └── <kernel>/
        ├── <kernel>.cuh   # header-only implementation
        ├── test_<kernel>.cu
        ├── Makefile
        └── README.md

building

Each kernel is self-contained:

cd normalization/rmsnorm
make           # H100
make GPU=A100  # A100
make test      # build + run correctness + benchmarks