You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
VORTEXM4: STRMM/DTRMM at about half of Accelerate at n = 1024 #6147
STRMM and DTRMM on VORTEXM4 (SME2 kernels from #6074) are slower than the other level-3 routines also at large sizes. On an M4 Pro, one thread, n = 1024: all 16 STRMM variants run at 780-888 GFLOPS, Accelerate at 1536-1562 (0.52x); DTRMM is at 0.61x. At n = 4096 STRMM is at 0.78-0.86x. SYMM, which uses the same GEMM kernel through the level-3 driver, reaches 0.93x at n = 1024.
The difference is therefore specific to the TRMM path. Candidates, not yet measured separately:
the TRMM kernel (gemm_kernel_sme2.c with TRMMKERNEL) zeroes ZA and stores C = alpha * ZA with a per-row scale-only store (s2_c_store_alpha), not the four-row fast path of the GEMM case;
Activity
arm64: VORTEXM4 TRMM kernel stores four rows of ZA per move
arm64: VORTEXM4 left TRMM kernel skips zero tiles at the diagonal
arm64: VORTEXM4 TRMM panel copies in one streaming-mode call
tesch1 commentedon Nov 15, 2026
Fixed by #6159 (merged), closing.