V4.2.0 Performance improvements

amcamd released this 15 May 22:10

· 3930 commits to master since this release

8040b1f

Features

Fractional global capability
Additional ResNet sizes
Round up for half vgprs
Initial code for PersistentKernel (disabled)
Feature inner unroll2
Enable BufferStore and buffer_atomic_cmpswap for GSU>1

Assets 2