Closed mattfel1 closed 6 years ago
This specific issue was because I was using the same implicit vT regardless of iteration difference segmentation in the unroller. I found some other problems with really complicated iteration difference cases but those are now also fixed and test cases added to IterationDiffs app
Found this in
RegFileMultiWrite
. We delayval value_1 = rf(p, k)
by 1, andk
is parallelized by 2. Because of iteration analysis, it splits the vectorized read into 2 separate reads. The retime transformer is delaying the first one, which is a vec with 1 element but the DelayLine has type vec of size 2.