laplacekorea / blis_apple Goto Github PK

Current MD file explains this port of BLIS to Apple's matrix coprocessor.

• Here is the original BLIS README.

As of Jul., 2021, the coprocessor is undocumented but not protected either. (Any user / program is allowed to invoke this coprocessor and it's supposed to be safe.) This work is based on Dougall Johnson's effort on analyzing the related instructions.

Known issues:

• Generic-strided is not supported by our microkernels for the destination matrix. Program would assert(false) upon encountering such a situation.
• TRSM might fail. Try commenting out function call to bli_cntx_set_packm_kers in config/aaplmx/bli_cntx_init_aaplmx.c if your need TRSM to work.
• Performance is around 75% the hardware peak. This is worse than the Accelerate framework on small size matrices while better on large size one.