HIP : RDNA3.5 MMVQ/MMQ kernels and per-type dispatch thresholds for quantised matmul - #56

Open

SimonTeixidor wants to merge 15 commits into

halo-box:masterfrom

SimonTeixidor:pr/rdna35-quant-matmul

Commits on Sep 14, 2026

CUDA/HIP: let IQ2_XS/IQ2_S reach hipBLAS above ne11=128 on RDNA3.5
SimonTeixidor
committed
b358475
HIP: branch-free SWAR for __vsub4 / __vcmpne4 / __vcmpeq4
SimonTeixidor
committed
2ae2ee9
CUDA/HIP: avoid emulated __vcmpne4/__vsub4 in the IQ2/IQ3 sign path
SimonTeixidor
committed
66efda4
CUDA/HIP: route the fork-only IQ3_S sign sites through apply_signs4
SimonTeixidor
committed
e1430b1
HIP: per-type MMVQ batch thresholds on RDNA3.5
SimonTeixidor
committed
df3e4d7
HIP: two MMVQ rows per block on RDNA3.5 for ncols_dst 2..8
SimonTeixidor
committed
856d62f
HIP: prefer the vector kernels below one MMQ tile on RDNA3.5
SimonTeixidor
committed
b9bbaa7
HIP: full-rate 24-bit multiply for the IQ2_XS/IQ2_S/IQ3_S scale epilogue
SimonTeixidor
committed
c6e682c
HIP: two output rows per block in the RDNA3.5 four-column MMVQ kernels
SimonTeixidor
committed
9a08368
CUDA/HIP: make calc_rows_per_block a per-type table on RDNA3.5
SimonTeixidor
committed
dada19e
CUDA/HIP: cheaper MMVQ epilogue and sign application for the codebook quants
SimonTeixidor
committed
7ed89d9
HIP: env override for the RDNA3.5 per-type MMVQ/MMQ crossover
SimonTeixidor
committed
c4d485b
hip: extend the RDNA3.5 MMQ prefetch to the J = 16 tile for Q4_K/Q5_K/Q6_K
SimonTeixidor
committed
151cdf1
HIP: re-tune the whole RDNA3.5 MMVQ/MMQ crossover table
SimonTeixidor
committed
16730c2
HIP: extend the RDNA3.5 crossover re-tune to the eleven unmeasured types
SimonTeixidor
committed
79194d3