V4.1.1 Performance improvements

amcamd released this 26 Apr 20:55

· 3997 commits to master since this release

227344b

Features

Support LSHL_ADD
Vectorize the store-C path
Enable DirectToLds for half
Fix sync with DirectToLds when PrefetchLocalRead=0
Optimize solution merging using lookup
Align MAC blocks when using half datatype
Add mi25 Device 6860 to vega10
Train for DataInitTypeBeta: 0
Add ResNet1x1 to Exact sizes

Assets 2