V4.6.0 Performance improvements, Bug fixes, add source hpa_hgemm

amcamd released this 11 Oct 23:34

· 3588 commits to master since this release

Features

Merge gfx906 code into gfx900/gfx803 code
Tune hgemm and sgemm for Resnet50 on gfx906
Add source hpa_hgemm
Use precise bounds check when possible
Tested on ROCm 1.9

Assets 2