Bug in dense_dims for reshaped views #157

Activity

1. Tokazama commented on Jun 10, 2022

   

   Tokazama

   on Jun 10, 2022

   Member

   More actions

   resolved with #161
2. Tokazama

   closed this as completedon Jun 10, 2022
3. ranocha commented on Jun 10, 2022

   

   ranocha

   on Jun 10, 2022

   Last edited by ranocha

   ContributorAuthor

   More actions

   I fixed it only for the case of reshaped versions of abstract vectors. It still fails for reshaped views of higher-dimensional arrays, e.g.,

   (@v1.7) pkg> activate --temp
     Activating new project at `/tmp/jl_4vH1Gz`
   
   (jl_4vH1Gz) pkg> add ArrayInterface
       Updating registry at `~/.julia/registries/General.toml`
      Resolving package versions...
       Updating `/tmp/jl_4vH1Gz/Project.toml`
     [4fba245c] + ArrayInterface v6.0.14
   
   julia> using ArrayInterface
   
   julia> u_base = randn(10, 10, 10); u_view = view(u_base, 3, :, :); u_reshaped_view = reshape(u_view, 1, size(u_base, 2) * size(u_base, 3));
   
   julia> ArrayInterface.dense_dims(u_reshaped_view)
   
   julia> ArrayInterface.dense_dims(u_reshaped_view) |> isnothing
   true

   Thus, this issue is not resolved in general.
4. Tokazama commented on Jun 10, 2022

   

   Tokazama

   on Jun 10, 2022

   Member

   More actions

   I guess in that example we could know that no dimensions are dense because u_view doesn't have any dense dimensions, but I'm not sure we can do any better at identifying instances where they are dense.
5. ChrisRackauckas

   reopened this on Jun 11, 2022
6. Tokazama commented on Jun 11, 2022

   

   Tokazama

   on Jun 11, 2022

   Member

   More actions

   #307 tests all the examples in this issue. I'm assuming there are still times when we don't detect dense dimensions just because we can't know in some situations without knowing the size.