Show HN: Libcusort – A radix sort for modern Nvidia GPUs
https://github.com/IlyaGrebnov/libcusort
I wrote a single-header CUDA radix sort library that's 1-6% faster than CUB for 32-bit keys and up to 38% faster for small types.Why bother? CUB's radix sort is already memory-bound, which is theoretically optimal. But CUB's codebase prioritizes backward compatibility and generality across GPU generations, leaving room for targeted optimizations on modern hardware.Benchmarks (RTX 5090, 67M...