Make sparse operations less dependent on inference #19611
Conversation
| (&)(A::SparseMatrixCSC, B::SparseMatrixCSC) = (_checksameshape(A, B); broadcast(&, A, B)) | ||
| (|)(A::SparseMatrixCSC, B::SparseMatrixCSC) = (_checksameshape(A, B); broadcast(|, A, B)) | ||
| xor(A::SparseMatrixCSC, B::SparseMatrixCSC) = (_checksameshape(A, B); broadcast(xor, A, B)) | ||
| (+)(A::SparseMatrixCSC, B::SparseMatrixCSC) = broadcast(+, A, B) |
There was a problem hiding this comment.
why remove the size check?
There was a problem hiding this comment.
It is also performed within the broadcast method.
There was a problem hiding this comment.
But shouldn't broadcast allow for 1-by-n .+ n-by-1 ?
There was a problem hiding this comment.
Shut! I missed these where binary arithmetic operations. I'll change it back.
There was a problem hiding this comment.
we should probably have a test that that throws an error with + but works with .+, if we don't already
pabloferz
force-pushed
the
pz/mapbcsparse
branch
from
8dab405 to
2a6a83d
Compare
|
I might have something potentially better in the works. Give me a few hours to see whether it works out... |
|
See #19623 for a potential alternative. The downside of converting to dense and back that I see is that something like |
|
Closing in favor of #19623 |
| @@ -1828,3 +1828,8 @@ let | |||
| @test_throws DimensionMismatch broadcast(+, A, B, speye(N)) | |||
| @test_throws DimensionMismatch broadcast!(+, X, A, B, speye(N)) | |||
| end | |||
|
|
|||
| let A = sparse(Real[1 1]) | |||
| @test (A + A)::SparseMatrixCSC{Int,Int} == [2 2] #19595 | |||
There was a problem hiding this comment.
the test will be good to port over
There was a problem hiding this comment.
Would you say we should verify the return type to have eltype Int? Presently
julia> Real[1 1] + Real[1 1]
1×2 Array{Real,2}:
2 2so maybe Real would be an acceptable return type, too?
There was a problem hiding this comment.
Personally, I'd prefer the dense case to return an Array{Int,2}, but we'd need a revision of the behavior of arithmetic operation on dense arrays. So, checking for eltype(A + A) === Int here seems better to me.
This should address #19595 and also fix the examples in #19561, although it might still be worth to have something like #19589.
This falls back, when the return type cannot be inferred, to treating the
SparseMatrixCSCs asArrays and then callssparseover the result. Granted, this is not the most efficient fallback, but since it applies to type-unstable functions, it shouldn't be much of a problem. We could eventually improve the performance if it seems necessary.