A pure, low-level tensor program representation enabling tensor program optimization via program rewriting. See the web demo at https://gussmith23.github.io/glenside-web-demo/
68
stars
10
forks
source link
Explore data layout transformation optimizations with Glenside #151
This has come up multiple times. You want to avoid layout transformations when going back and forth to accelerators. I just haven't found a good model to do an experiment over.
One lead might be Alpha Fold. Zihao was showing me that there is a lot of expensive data layout transformation happening.
This has come up multiple times. You want to avoid layout transformations when going back and forth to accelerators. I just haven't found a good model to do an experiment over.
One lead might be Alpha Fold. Zihao was showing me that there is a lot of expensive data layout transformation happening.