AMD RDNA3 GPU driver via raw KFD ioctls. No ROCm, no OpenCL.
- Device discovery and topology queries
- VRAM/GTT/Userptr memory with shared address space
- PM4 compute queue and kernel dispatch
- Hand-written RDNA3 ASM kernels: matmul (3100 GFLOP/s), matvec, superlinear
- Tile-safe buffer padding for OOB protection
- Event-based interrupt wait (no CPU polling)
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Tommi Niemi
66a9edb898