Use block_gmres for batched linear solves #129

Activity

1. thorek1

   changed the title [-]Consider block-gmres[/-] [+]block-gmres[/+] on Dec 8, 2023
2. gdalle commented on Dec 12, 2023

   

   gdalle

   on Dec 12, 2023

   Member

   More actions

   Thanks for the idea. I don't have much time to work on this at the moment though
3. mohdibntarek commented on Dec 12, 2023

   

   mohdibntarek

   on Dec 12, 2023

   Collaborator

   More actions

   how can block-gmres help ID.jl?
4. thorek1 commented on Dec 12, 2023

   

   thorek1

   on Dec 12, 2023

   ContributorAuthor

   More actions

   for example here:

       dy = ntuple(Val(N)) do k
           dₖx = partials.(x_and_dx, k)
           dₖx_vec = vec(dₖx)
           dₖc_vec = similar(y_vec)
           mul!(dₖc_vec, B_vec, dₖx_vec)
           dₖy_vec = solve(implicit.linear_solver, A_vec_presolved, -dₖc_vec)
           reshape(dₖy_vec, size(y))
       end
   
       y_and_dy = map(eachindex(IndexCartesian(), y)) do i
           Dual{T}(y[i], Partials(ntuple(k -> dy[k][i], Val(N))))
       end

   instead of solving with gmres for each slice you could use block-gmres once. I didn't check the block-gmres implementation but I think it is worth testing whether this speeds up computation and this might also allow to solve #118

   👍React with 👍2Reacted by Guillaume Dalle and Mohamed Tarek
5. gdalle

   added

   performanceImprove performance or stabilityImprove performance or stability

   on Feb 21, 2024
6. gdalle commented on Apr 10, 2024

   

   gdalle

   on Apr 10, 2024

   Member

   More actions

   This is implemented in #135 but it still requires JuliaSmoothOptimizers/Krylov.jl#854 to be fixed
7. gdalle

   changed the title [-]block-gmres[/-] [+]Use `block_gmres` for batched linear solves[/+] on Apr 10, 2024