aditew01 commented 2 weeks ago

Description

This enables jit:uni reorder for dtype-> bf16 -> f32

Checklist

[ ] Do all unit and benchdnn tests (make test and make test_benchdnn_*) pass locally for each commit? - Yes
[ ] Have you formatted the code using clang-format? - Yes

Adding a small testcase logs to demonstrate perf numbers before and after.

TC: ONEDNN_VERBOSE=all ./benchdnn --reorder --sdt=bf16 --ddt=f32 --mode=p 1x4096x4096

(new) jit:any :

onednn_verbose,v1,primitive,exec,cpu,reorder,jit:uni,undef,src:bf16::blocked:abc::f0 dst:f32::blocked:abc::f0,,,1x4096x4096,0.493896

(old) simple:any :

onednn_verbose,v1,primitive,exec,cpu,reorder,simple:any,undef,src:bf16::blocked:abc::f0 dst:f32::blocked:abc::f0,,,1x4096x4096,11.9778

mgouicem commented 2 weeks ago

Thanks for the PR. Please rename the commit message to adhere to the oneDNN coding guideline.

aditew01 commented 2 weeks ago

cc: @theComputeKid

aditew01 commented 1 week ago

cc: @Sqvid

aditew01 commented 1 week ago

cc: @mgouicem