dialects: (arith, core) Constant op to accept tensor and memref

[n-io]

Upstream mlir accepts the following arith.constant:

%t_const = arith.constant dense<1.234500e-01> : tensor<16xf32>
%m_const = arith.constant dense<1.678900e-01> : memref<64xf32>

As for tensors, xdsl currently accepts it in generic form and print it in both generic and non-generic form, but will fail to parse the non-generic form.

As for memrefs, xdsl will not accept it and requires an extensive change to the attribute parser to add support for memrefs in RankedVectorOrTensorOf. Some code had to be moved after the definition of MemRefType to support this change.

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 89.85%. Comparing base (ddbe80c) to head (13c5808).
Report is 5 commits behind head on main.

Additional details and impacted files

@@           Coverage Diff           @@
##             main    #2969   +/-   ##
=======================================
  Coverage   89.84%   89.85%           
=======================================
  Files         408      408           
  Lines       51002    51028   +26     
  Branches     7909     7912    +3     
=======================================
+ Hits        45825    45850   +25     
- Misses       3927     3928    +1     
  Partials     1250     1250           

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[superlopuh]

Why move this down? What's the change here? Git isn't being very helpful

[n-io]

As briefly mentioned in the PR description, though likely not explained in sufficient detail, I had to move RankedVectorOrTensorOf after MemRefType, as I changed it to now also accept MemRefType. As a follow-up, DenseIntOrFPElementsAttr also had to be moved down since it uses RankedVectorOrTensorOf.

As far as I remember, the only change required to DenseIntOrFPElementsAttr was to change the signature of the type argument in from_list from RankedVectorOrTensorOf[AnyFloat | IntegerType | IndexType] to:

RankedVectorOrTensorOf[AnyFloat | IntegerType | IndexType]
            | RankedVectorOrTensorOf[AnyFloat]
            | RankedVectorOrTensorOf[IntegerType]
            | RankedVectorOrTensorOf[IndexType]

Not fully clear why, but it does make pyright happy.

BTW, accepting new name suggestions for RankedVectorOrTensorOf.

[superlopuh]

Makes a lot of sense, thank you

[superlopuh]

Code looks good to me, I would probably wait for someone more familiar with MLIR than I am to comment on how similar this is to how things happen over there. My understanding is that some of the names follow the MLIR API.

Mabe @math-fehr or @zero9178 can advise?

[n-io]

Code looks good to me, I would probably wait for someone more familiar with MLIR than I am to comment on how similar this is to how things happen over there. My understanding is that some of the names follow the MLIR API.

Mabe @math-fehr or @zero9178 can advise?

I'd appreciate that, not super familiar with that part of the API myself, but it appears RankedVectorOrTensorOf was an xdsl concept, where MLIR if I'm not mistaken has no similar concept and would more generically use Attribute

[PapyChacal]

Nice!

[AntonLydike]

I don't think memref makes all that much sense to be accepted here? Seeing as arith does not work on memref anyway? (big debate on this happening rn)...

[AntonLydike]

It appears that arith.constant can instantiate memrefs (although I don't know how it would even lower that?), but other arith ops can't operate on memrefs...

[PapyChacal]

I don't think memref makes all that much sense to be accepted here? Seeing as arith does not work on memref anyway? (big debate on this happening rn)...

I don't think it makes much sense either; but this is about matching MLIR's implementation, not changing it, IMO?

[n-io]

It appears that arith.constant can instantiate memrefs (although I don't know how it would even lower that?), but other arith ops can't operate on memrefs...

It's lowered during bufferization, e.g.:

%0 = arith.constant dense<1.234500e-01> : tensor<8xf32>

mlir-opt %s -allow-unregistered-dialect --eliminate-empty-tensors --one-shot-bufferize="allow-unknown-ops"

module {
  memref.global "private" constant @__constant_8xf32 : memref<8xf32> = dense<1.234500e-01> {alignment = 64 : i64}
  %0 = memref.get_global @__constant_8xf32 : memref<8xf32>
}