The action item #1 should be completed with ⚙ D152644 [mlir][Vector] Add support for 0-D vectors to vector.insert/extract.
I’m now looking at #2 and, in particular, at adding support for Value indices to vector.extract/insert ops. They only support constant indices right now. The implementation is almost done but there are around 88 lit test files that would need to be updated to the new vector.extract/insert format. Unfortunately, I can’t think of a ninja regex way to update them so I would really appreciate some help here. Basically, we have to replace check rules like:
// CHECK-NEXT: %[[R2:.+]] = vector.extract %{{.*}}[0, 0] : vector<1x1x16xf32>
with something like:
// CHECK-DAG: %[[C0:.+]] = arith.constant 0 : index
// CHECK-DAG: %[[C0_1:.+]] = arith.constant 0 : index
...
// CHECK-NEXT: %[[R2:.+]] = vector.extract %{{.*}}[%[[C0]], %[[C0_1]]] : vector<1x1x16xf32>
where C0, C0_1, etc. can be random indices and be located anywhere in the output.
If anybody has cycles to work on this, please call out the tests you want to work on in the review. You can share your diffs over GH also there. I’ll make sure those contributing are added as co-authors. To repro, you can just apply the ongoing diff and run the check-mlir target.
Thanks in advance!
Diego
cc: @banach-space, @javiersetoain, @nicolasvasilache, @ThomasRaoux, @hanchung