tree 2d01db8e924fed140ccb1672c759f351d2cc1013
parent aeb38b0a65f5748a39ef82bdf5dfb641f7372ff6
author bjacob <benoitjacob@google.com> 1698688525 -0400
committer GitHub <noreply@github.com> 1698688525 -0400
gpgsig -----BEGIN PGP SIGNATURE-----
 
 wsBcBAABCAAQBQJlP+4NCRBK7hj4Ov3rIwAAKk8IAJuXDTXPQE0iPa3Ox8SURK8D
 lfG2yuUJQyUMHC1SAY2QcnSGnxDT+xa1UTus/1AnGbPQwb7CqQEWCplXsN6Lo4fX
 ii1pn+x49KJfKDzy9KdF2GpPALDotttig6nrt/QZT27DNH53hnSHQHO2JnxDoy8H
 F93ac1ujoT5e2IoKyqf5BeeZUhKHyqJreCzmcoFRdTQDvoewsgh60CJQddFXyIUn
 RIoOs2aytPty1jRmkyjy6qd517J9lv1Gzcp7B82prVvBFJY9RFy6PYj4s2tIgeGW
 TMjbKZ8XmNsFD6+4CDaJ8IOOPKvNb0kDnm/mfn6ryvrNhB5ZQkbt96l5ZARfM68=
 =iuUP
 -----END PGP SIGNATURE-----
 

ukernels: stop abusing signless as signed (#15338)

In MLIR we are chronically abusing signless types as signed. Ongoing
discussion at #15241, etc. At least in ukernels we are insulated enough
from the compiler code that we don't have to replicate the same
confusion. So, just because `linalg` named matmul ops want `int8 x int8
-> int32` matmuls to be represented as `i8 x i8 -> i32` even though
these are signless types and this is silently interpreting these as
signed, we didn't have to let that lower to a `_i8i8i32` ukernel, we
could instead more properly represent that as a `_s8s8s32` ukernel.

Fixing this now paves the way for unsigned LHS/RHS support in ukernels,
a step in #15158, without adding more tech debt.

The bulk of is PR is just a regex substitution:

```
find runtime/src/ compiler/src/ -type f | xargs sed -i 's/IREE_UK_FLAG_MMT4D_TYPE_I8I8I32/IREE_UK_FLAG_MMT4D_TYPE_S8S8S32/g;s/iree_uk_mmt4d_type_i8i8i32/iree_uk_mmt4d_type_s8s8s32/g;s/iree_uk_mmt4d_tile_i8i8i32/iree_uk_mmt4d_tile_s8s8s32/g;s/iree_mmt4d_reference_innerloop_i8i8i32/iree_mmt4d_reference_innerloop_s8s8s32/g'
```