[mlir] -test-vector-unrolling-patterns pass failed with assertion error "cast<Ty>() argument of incompatible type!"
#64075
Comments
EugeneZelenko
added
mlir
crash
|
@llvm/issue-subscribers-mlir |
|
I have found the root cause, but I give up on fixing this for now because I cannot figure it what to change. Maybe later. The root cause is that A minimum example that also crashes: func.func @multi_reduction(%0: vector<4x2xf32>, %acc1: f32) -> f32 {
%2 = vector.multi_reduction <add>, %0, %acc1 [0, 1] : vector<4x2xf32> to f32
return %2 : f32
}To see that these return types make sense, take the following Julia example: julia> A = reshape(1:(3*3), 3, 3)
3×3 reshape(::UnitRange{Int64}, 3, 3) with eltype Int64:
1 4 7
2 5 8
3 6 9
julia> A[2:2, 2:2] # extract strided slice with offset = [0, 0], sizes = [1, 1], stride = [1, 1]
1×1 Matrix{Int64}:
5So, indeed, it makes sense that a strided slice returns a n-dimensional vector even for a single scalar result. Next, reduce doesn't necessarily have to return a vector: julia> reduce(min, A[2:2, 2:2])
5This problem with mismatching types most likely lies somewhere in the llvm-project/mlir/lib/Dialect/Vector/Transforms/VectorUnroll.cpp Lines 348 to 415 in c6f0f88 |
|
I was looking forward to working on this issue. Can this issue be assigned to me? |
|
@Prakhar-Dixit yea sure - good luck and god speed 😄 |
|
After investigating this issue, I discovered that the Minimal Example of the Crashing IR: The IR appears valid because it lowers correctly and shows no signs of being invalid. Therefore, I believe the proper fix is to bail out instead of handling scalars, as I assume it was never intended to do so when the operation is asked to reduce a vector to a scalar. For example: Would love to hear your thoughts on this, |
…29694) Fixes issue #64075 Referencing this comment for more detailed view -> #64075 (comment) **Minimal example crashing :** ``` func.func @multi_reduction(%0: vector<4x2xf32>, %acc1: f32) -> f32 { %2 = vector.multi_reduction <add>, %0, %acc1 [0, 1] : vector<4x2xf32> to f32 return %2 : f32 } ```
… scalar (#129694) Fixes issue #64075 Referencing this comment for more detailed view -> llvm/llvm-project#64075 (comment) **Minimal example crashing :** ``` func.func @multi_reduction(%0: vector<4x2xf32>, %acc1: f32) -> f32 { %2 = vector.multi_reduction <add>, %0, %acc1 [0, 1] : vector<4x2xf32> to f32 return %2 : f32 } ```
MLIR built at commit 585cbe3
Reproduced with:
mlir-opt -test-vector-unrolling-patterns=unroll-based-on-type temp.mlirtemp.mlir:
trace:
The text was updated successfully, but these errors were encountered: