V4.7.0 Performance improvements, bug fixes, add assembly hpa_hgemm, initial source hpa_igemm

amcamd released this 19 Dec 19:35

· 3478 commits to master since this release

315c23a

Features

add dot2 instructions for fp16/fp32 hpa_hgemm on gfx906
initial i8/i32 hpa_igemm
enable fractional loads
enable precise bounds check

Assets 2