Rob Suderman
Rob Suderman
We can completely delete the `iree-import-xla` tool once the iree-jax changes land. Hoping to delete this week.
Prioritzed as P3 as TensorFlow is lower priority to other frontends.
Validated compilation in iree-samples repo. Marking as fixed.
@bjacob This should be fixed for given we have a 32-bit version of the lowering now. https://github.com/llvm/llvm-project/commit/9294a1e9a8ab8fed60ee1b0737e944077b5339cd
At this point we generate almost purely `i32` operations that should pass through the vectorization code. The only section that is not i32 attempts to "encourage" the compile towards sqdmulh....
So I think this is not a legal `mhlo.scatter`. I cannot generate this operation via `jax` and the nearest replication broadcasts to add a length-1 dimension.
Validated in iree-samples that the case is now fixed
It also appears that slice fails on this as well.
Issues are lined up for `tfl.strided_slice`, `tf.FusedFixedBatchMatmul` and `tfl.gather`. Issues still exist for propagating shape computations to TOSA.
I think its related to certain types not lowering to spirv - my guess would be that TOSA import does not handle the sample `i64` to `i32` narrowing.