Sun May 12 08:56:45 EDT 2019
Feldspar and fused representations
There was a trick there.
In some earlier rai notes mentioned I need a loop transformation
algebra. Maybe start there?
1. It should be simple to create a generic unfused version, where all
arrays are explicit, and time feedback is not in the picture.
2. Perform fusion on that
3. Introduce state on top of it?
So the new focus is to separate feedback from vectors.
The core of the target language doesn't have anything to do with
feedback at the inner loops.