POWER10/11 MMA acceleration for all quantized formats (silicon-validated)#100

Open

mavin2009 wants to merge 21 commits into

PrismML-Eng:prismfrom

mavin2009:power-mma

Commits on Jul 23, 2026

power-mma: Q1_0/Q2_0 MMA GEMM + no-packing GEMV for POWER10/11
mavin2009
committed
power-mma: Q4_K/Q5_K/Q2_K/Q6_K MMA sgemm
mavin2009
committed
power-mma: Q3_K and IQ4_NL/IQ4_XS MMA sgemm
mavin2009
committed
power-mma: Q4_1/Q5_0/Q5_1 legacy-format MMA sgemm
mavin2009
committed
power-mma: grid/ternary formats via two shared signed kernels
mavin2009
committed
power-mma: abort loudly on pack-buffer allocation failure
mavin2009
committed
power-mma: tensor-keyed pack cache + MXFP4/NVFP4 coverage
mavin2009
committed
power-mma: rollout hardening -- cache fingerprint, admission policy, shared B-pack
mavin2009
committed
power-mma: prefetch the next packed chunk in the GEMM kernels
mavin2009
committed
power-mma: IQGRID_PINGPONG accumulator-alternation variant (flag)
mavin2009
committed
power-mma: route Q8_0/Q4_0 through the pack-cached MMA kernels
mavin2009
committed
power-mma: GCC 11 compatibility
mavin2009
committed
power-mma: exact-integer 128*codesum corrections
mavin2009
committed
power-mma: row-partition fallback for small n on cache-hit paths
mavin2009
committed
power-mma: n<8 hands generation back to vec_dot; cache slots raised
mavin2009
committed
power-mma: pack-cache slot table grows; refusals are counted and loud
mavin2009
committed
power-mma: keep the cached packed path at n=1 for IQ1_M and TQ1_0
mavin2009
committed
power-mma: ISA 3.1 experiments -- dcbt streams default for iq4, lxvp flag
mavin2009
committed
power-mma: slice-parallel first-touch packing
mavin2009
committed
power-mma: two-row VSX GEMV experiment, kept unwired
mavin2009
committed
power-mma: GER GEMV over cached tiles at n=1, routed per format by measurement
mavin2009
committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

POWER10/11 MMA acceleration for all quantized formats (silicon-validated)#100

POWER10/11 MMA acceleration for all quantized formats (silicon-validated)#100
mavin2009 wants to merge 21 commits into
PrismML-Eng:prismfrom
mavin2009:power-mma

Commits on Jul 23, 2026

power-mma: Q1_0/Q2_0 MMA GEMM + no-packing GEMV for POWER10/11

power-mma: Q4_K/Q5_K/Q2_K/Q6_K MMA sgemm

power-mma: Q3_K and IQ4_NL/IQ4_XS MMA sgemm

power-mma: Q4_1/Q5_0/Q5_1 legacy-format MMA sgemm

power-mma: grid/ternary formats via two shared signed kernels

power-mma: abort loudly on pack-buffer allocation failure

power-mma: tensor-keyed pack cache + MXFP4/NVFP4 coverage

power-mma: rollout hardening -- cache fingerprint, admission policy, shared B-pack

power-mma: prefetch the next packed chunk in the GEMM kernels

power-mma: IQGRID_PINGPONG accumulator-alternation variant (flag)

power-mma: route Q8_0/Q4_0 through the pack-cached MMA kernels

power-mma: GCC 11 compatibility

power-mma: exact-integer 128*codesum corrections

power-mma: row-partition fallback for small n on cache-hit paths

power-mma: n<8 hands generation back to vec_dot; cache slots raised

power-mma: pack-cache slot table grows; refusals are counted and loud

power-mma: keep the cached packed path at n=1 for IQ1_M and TQ1_0

power-mma: ISA 3.1 experiments -- dcbt streams default for iq4, lxvp flag

power-mma: slice-parallel first-touch packing

power-mma: two-row VSX GEMV experiment, kept unwired

power-mma: GER GEMV over cached tiles at n=1, routed per format by measurement