CPU kernel parity: decode and tile kernels at llama.cpp speed or better on zen2, zen4, Intel Granite Rapids and M4 Pro - #3924
Merged
Merged
Commits
Commits on Sep 2, 2026
kernel parity pass, step 0: the kq kernel microbench, the reference rows, and the two research memos
andcommitted- andcommitted
- andcommitted
- committed
- andcommitted
- andcommitted
- andcommitted
- andcommitted
- andcommitted
- committed
- committed
- committed
- committed
- committed
- committed
- committed
- committed
- committed
- committed
- committed
- committed
- committed
- committed
- committed
- committed
- committed
- andcommitted
- andcommitted
- committed
- andcommitted
- andcommitted
- andcommitted
- andcommitted
- andcommitted
- committed
- committed
- andcommitted
- andcommitted
- andcommitted
- andcommitted
- andcommitted
- committed
- committed
- andcommitted
- andcommitted
- andcommitted
- committed
- committed
- andcommitted
- committed
- committed
- committed
- committed
- committed
- committed
- committed
- andcommitted
- committed
- committed
- andcommitted
- committed
- andcommitted
- committed
- committed
- andcommitted
- committed
- committed
- committed
- andcommitted
- andcommitted
- andcommitted
- committed
- committed
- committed
- andcommitted
grid gemv: the VBMI symbol lattice as a perm row for iq2xxs and iq3xxs (grid = "vbmi"; lab, unraced)
andcommitted- committed
- andcommitted
- committed
- committed
- committed
- committed
- committed
- committed
- committed
- committed
- committed
- committed
- andcommitted
- committed
- committed
- andcommitted
- andcommitted
- committed
- committed
- andcommitted
- andcommitted
- andcommitted
- committed
- committed
- committed
- andcommitted
- committed
- committed
- andcommitted
- andcommitted
- andcommitted
- andcommitted
- andcommitted
- committed
- committed
- committed
- andcommitted
- andcommitted
- committed
- andcommitted
- committed
- andcommitted
- andcommitted
- andcommitted
- andcommitted
- andcommitted
- andcommitted