Fix several bugs on Hexagon and some cleanup #5570
Conversation
|
@aankit-ca @pranavb-ca please take a look. |
|
Windows failure is unrelated |
|
@dsharletg For vmpa, I think we can simply change a01 to |
|
Rest LGTM |
|
I think we need to do the interleaving in hvx_128.ll, because the interleaving is different. Shuffle::make_interleave interleaves the vectors before splitting to native vector widths, while this implementation interleaves them after splitting to native vector widths. What do you think? |
|
I'm sorry. I meant changing a01 to Shuffle::make_concat({mpys[0].first, mpys[1].first}); for vmpa |
|
But I think the same problem applies. By doing Shuffle::make_concat, we concatenate the vectors before slicing to native vector widths. But we need to concatenate the two operands after slicing to native vector widths, which is what the runtime hvx_128.ll implementation does. |
|
Aah. Right. Yeah we will need to the concatenation in .ll |
|
Cool, thanks for confirming :) This is a tricky issue, I've had this bug before myself. |
This PR fixes bugs:
It also removes VtmpyGenerator. This has been disabled for a long time, and now we have VectorReduce. Additionally, I have a WIP pass in another branch that finds vector reductions in a target independent lowering pass. If we try to bring back automatic finding of stencil vector reductions, we should build it into that pass instead of fixing VtmpyGenerator.