PR #8562 Expand Qgemm UDOT kernel to 8x8 block

comments

Chen Fu committed 5 years ago

use v2 v3 for B unrolled in k dimension

Chen Fu committed 5 years ago

move comments before loop prefix

Chen Fu committed 5 years ago

use same accumulators for k-dimension unroll

Chen Fu committed 5 years ago

rearrange registers- prepare to process more column

Chen Fu committed 5 years ago

load A data with 128b load

Chen Fu committed 5 years ago

process 8 rows

Chen Fu committed 5 years ago

dot prepack A 8 rows

Chen Fu committed 5 years ago

debug aid

Chen Fu committed 5 years ago

add linux/android version

Chen Fu committed 5 years ago

optimize m4 loop prefix

Chen Fu committed 5 years ago

revert m4 loop behavior

Chen Fu committed 5 years ago

optimize m1

Chen Fu committed 5 years ago

optimize m2 loop

Chen Fu committed 5 years ago

Revert "debug aid"

Chen Fu committed 5 years ago

copy m1 m2 optimization to aarch64

Chen Fu committed 5 years ago

reduce number of saved registers

Chen Fu committed 5 years ago

pipelining dup instructions into the outter m8 loop

Chen Fu committed 5 years ago

adjust spaces

Chen Fu committed 5 years ago

onnxruntime Expand Qgemm UDOT kernel to 8x8 block #8562 Merged