Allow index inside dimension tensors.

Ever since https://reviews.llvm.org/rG3dceb6d2460550ef8c2e53df2181c8a11a7e6730 MLIR now allows `index` type inside tensors, addressing this longstanding TODO.

PiperOrigin-RevId: 306391413
Change-Id: I797623de02ff031b9ffabfaee81afba91ead6692
This commit is contained in:
Sean Silva 2020-04-14 01:00:05 -07:00 committed by TensorFlower Gardener
parent 242129341e
commit b2418b7fee

View File

@ -55,11 +55,10 @@ def HLO_Tuple : NestedTupleOf<[HLO_Tensor, HLO_Token]>;
def HLO_TensorOrTuple : AnyTypeOf<[HLO_Tensor, HLO_Tuple]>;
// Dynamic representation of a shape vector as a tensor. Ideally this would be
// an index type (as it stores indices) but that is currently disallowed in
// MLIR.
// Dynamic representation of a shape vector as a tensor.
def HLO_DimensionTensor : ShapedContainerType<
[AnySignlessInteger], And<[IsTensorTypePred, HasAnyRankOfPred<[1]>]>,
[Index, AnySignlessInteger],
And<[IsTensorTypePred, HasAnyRankOfPred<[1]>]>,
"a 1D tensor of dimensions">;
// In general, static shaped tensor constraints should be avoided unless