Inconsistent behavior in sparse matrices when interacting with NaNs

[gabgoh]

julia> spzeros(2,2)*[NaN;NaN]    
2-element Array{Float64,1}:      
 0.0                             
 0.0                             

julia> zeros(2,2)*[NaN;NaN]      
2-element Array{Float64,1}:      
 NaN                             
 NaN            

I believe the latter behavior is the correct one

[tkelman]

Yes, this is wrong, our current sparse matrix implementation does not give proper IEEE754 results for operations on the implicit non-stored zeros. Should be fixable with a bit of care.

related: #14973

[KristofferC]

Is there any programming language or library that does this properly for all operations?

[JaredCrean2]

Oh, this again. There was also some discussion in #15579 revolving around handling Infs and NaNs in sparse algorithms.

[ViralBShah]

Cc @Sacha0

[tkelman]

I'll repeat myself here too:

Ignoring the corner cases doesn't make Julia's behavior materially more correct here. Fixing it, for the cases where that's not terribly complicated to do [...] means we're ahead of the curve in fixing bugs that have either not been reported yet, not been fixed yet, or been left unaddressed in existing libraries. We don't need to repeat the mistakes of our predecessors here.

Calling this a feature instead of a bug runs counter to the purpose of sparse matrices as a representation optimization that shouldn't change visible behavior in a generic programming sense.

[JaredCrean2]

@tkelman I think you and I should get matching shirts and wear them to the next JuliaCon: https://www.amazon.com/Arguing-Engineer-Like-Wrestling-shirt/dp/B01HJ85TTC

Ok, if you are really intent on doing things this way, how about structuring the API in this way:

function func_unchecked(args...)
   # implement the algorithm, not performing any checks
end

function func_checked(args...)
  # do checks here
  func_unchecked(args...)
end

and export both versions, and also link this up to the @fastmath machinery so when the user decides they don't care about Infs and NaNs, calls to func_checked turn into func_unchecked?

[tkelman]

that seems fine, except that I don't think we should export the unchecked version

[JaredCrean2]

Why not? It just unchecked, not unsafe. The functions that do arithmetic don't check for overflow, underflow etc.

[tkelman]

namespace pollution - we should probably be putting these things in an Unchecked module.

[JaredCrean2]

Sounds reasonable.

[KristofferC]

Closing in favor of https://github.com/JuliaLang/julia/issues/22733