Remove allocations in `FeasibilityFormNLS`

[tmigot]

I run the following script:

using Pkg
Pkg.activate(".")
using LinearAlgebra, NLPModels, NLPModelsModifiers, using NLPModelsTest

# Test FeasibilityFormNLS Models:
problems = setdiff(NLPModelsTest.nls_problems)
map(
  nlp -> print_nlp_allocations(nlp, test_allocs_nlsmodels(nlp, exclude = [hess])),
  map(x -> FeasibilityFormNLS(eval(Symbol(x))()), problems),
) # this is all 0's!!
map(
  nlp -> print_nlp_allocations(nlp, test_allocs_nlpmodels(nlp, linear_api = true, exclude = [hess])),
  map(x -> FeasibilityFormNLS(eval(Symbol(x))()), problems),
)

and it returns

Problem name: LLS_manual-ffnls
                     jprod_nln!: ⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅ 0.0
         jac_op_transpose_prod!: ███████⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅ 224.0
                      cons_lin!: ⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅ 0.0
     jac_nln_op_transpose_prod!: ███⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅ 80.0
                 jac_structure!: ⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅ 0.0
             jac_lin_structure!: ⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅ 0.0
                      cons_nln!: ███⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅ 80.0
                    jtprod_nln!: ███⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅ 80.0
             jac_nln_structure!: ⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅ 0.0
                    jtprod_lin!: ⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅ 0.0
                   jac_op_prod!: █████████████⋅⋅⋅⋅⋅⋅⋅ 400.0
     jac_lin_op_transpose_prod!: ⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅ 0.0
              hess_lag_op_prod!: ██████████████████⋅⋅ 560.0
               jac_lin_op_prod!: ⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅ 0.0
                         hprod!: ████████████████████ 656.0
                         jprod!: █████████████⋅⋅⋅⋅⋅⋅⋅ 400.0
               jac_nln_op_prod!: ⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅ 0.0
                          cons!: ████████⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅ 240.0
                          grad!: ⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅ 0.0
                  hess_op_prod!: ████████████████████ 656.0
                     hprod_lag!: ██████████████████⋅⋅ 560.0
                        jtprod!: ███████⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅ 224.0
                            obj: ⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅ 0.0
                     jprod_lin!: ⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅ 0.0
                     jac_coord!: ⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅ 0.0
                 jac_nln_coord!: ⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅ 0.0
                 jac_lin_coord!: ⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅ 0.0

  Problem name: MGH01_manual-ffnls
                     jprod_nln!: ⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅ 0.0
         jac_op_transpose_prod!: ████⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅ 80.0
     jac_nln_op_transpose_prod!: ████⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅ 80.0
                 jac_structure!: ⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅ 0.0
                      cons_nln!: ████⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅ 80.0
                    jtprod_nln!: ████⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅ 80.0
             jac_nln_structure!: ⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅ 0.0
                   jac_op_prod!: █████████████████⋅⋅⋅ 400.0
              hess_lag_op_prod!: █████████████████⋅⋅⋅ 400.0
                         hprod!: ████████████████████ 480.0
                         jprod!: █████████████████⋅⋅⋅ 400.0
               jac_nln_op_prod!: ⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅ 0.0
                          cons!: ██████████⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅ 240.0
                          grad!: ⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅ 0.0
                  hess_op_prod!: ████████████████████ 480.0
                     hprod_lag!: █████████████████⋅⋅⋅ 400.0
                        jtprod!: ████⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅ 80.0
                            obj: ⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅ 0.0
                     jac_coord!: ⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅ 0.0
                 jac_nln_coord!: ⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅ 0.0

  Problem name: NLSHS20_manual-ffnls
                     jprod_nln!: ⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅ 0.0
         jac_op_transpose_prod!: ██████████⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅ 240.0
     jac_nln_op_transpose_prod!: ███████⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅ 160.0
                 jac_structure!: █████⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅ 112.0
                      cons_nln!: ████⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅ 80.0
                    jtprod_nln!: ███████⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅ 160.0
             jac_nln_structure!: █████⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅ 112.0
                   jac_op_prod!: ████████████████████ 480.0
              hess_lag_op_prod!: █████████████████⋅⋅⋅ 400.0
                         hprod!: ████████████████████ 496.0
                         jprod!: ████████████████████ 480.0
               jac_nln_op_prod!: ⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅ 0.0
                          cons!: █████████████⋅⋅⋅⋅⋅⋅⋅ 320.0
                          grad!: ⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅ 0.0
                  hess_op_prod!: ████████████████████ 496.0
                     hprod_lag!: █████████████████⋅⋅⋅ 400.0
                        jtprod!: ██████████⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅ 240.0
                            obj: ⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅ 0.0
                     jac_coord!: ⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅ 0.0
                 jac_nln_coord!: ⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅ 0.0

  Problem name: NLSLINCON-ffnls
                     jprod_nln!: ⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅ 0.0
         jac_op_transpose_prod!: ██⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅ 496.0
                      cons_lin!: ⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅ 0.0
     jac_nln_op_transpose_prod!: █⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅ 176.0
                 jac_structure!: ⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅ 0.0
             jac_lin_structure!: ⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅ 0.0
                      cons_nln!: █⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅ 176.0
                    jtprod_nln!: █⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅ 176.0
             jac_nln_structure!: ⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅ 0.0
                    jtprod_lin!: ⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅ 0.0
                   jac_op_prod!: ████⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅ 880.0
     jac_lin_op_transpose_prod!: ⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅ 0.0
              hess_lag_op_prod!: ████████████████████ 5456.0
               jac_lin_op_prod!: ⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅ 0.0
                         hprod!: ████████████████████ 5728.0
                         jprod!: ████⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅ 880.0
               jac_nln_op_prod!: ⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅ 0.0
                          cons!: ██⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅ 528.0
                          grad!: ⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅ 0.0
                  hess_op_prod!: ████████████████████ 5728.0
                     hprod_lag!: ████████████████████ 5456.0
                        jtprod!: ██⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅ 496.0
                            obj: ⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅ 0.0
                     jprod_lin!: ⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅ 0.0
                     jac_coord!: ⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅ 0.0
                 jac_nln_coord!: ⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅ 0.0
                 jac_lin_coord!: ⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅ 0.0

so there is potentially the following list to handle:

• cons_nln Save allocations cons_nln in FeasibilityFormNLS #76
• cons! (most likely connected to cons_nln)
• jtprod_nln!
• jtprod!
• hprod!
• jprod!
• jac_nln_structure!

[tmigot]

After #90 , only jtprod! allocates and I have no real idea why... There is something happening with the views that I don't really understand. Here is an illustration:

using NLPModels, NLPModelsModifiers, NLPModelsTest
nlp = NLSLC()
nls = FeasibilityFormNLS(nlp)
x = get_x0(nls)
y = get_y0(nls)
Jtv = similar(x)
jtprod!(nls, x, y, Jtv)
@allocated jtprod!(nls, x, y, Jtv) # 144 - Why would this allocate?
Jl = rand(nls.meta.nlin)
jtprod_lin!(nls, x, Jl, Jtv)
@allocated jtprod_lin!(nls, x, Jl, Jtv) # 0
Jl = rand(nls.meta.nnln)
jtprod_nln!(nls, x, Jl, Jtv)
@allocated jtprod_nln!(nls, x, Jl, Jtv) # 0

# Test on the subproblem with views
x = get_x0(nlp)
y = get_y0(nlp)
Jl = rand(nlp.meta.nlin)
Jtv = similar(x)
jtprod_lin!(nlp, x, Jl, Jtv)
@allocated jtprod_lin!(nlp, x, Jl, Jtv) # 0

jtprod_lin!(nlp, view(x, 1:15), Jl, Jtv)
@allocated jtprod_lin!(nlp, view(x, 1:15), Jl, Jtv) # 48 - Why would this allocate?

[dpo]

Are you saying that others don’t allocate when you create a view? Isn’t it the view itself that allocates?

Julia> @allocated view(x, 1:15)
48

[tmigot]

From my understanding, views of connected indices shouldn't allocate. You are right in your example, but this:

julia> x = rand(20);

julia> @allocated view(x, 1:15)
1499872

julia> @allocated view(x, 1:15)
48

julia> function fun(x) view(x, 1:15); return nothing end
fun (generic function with 1 method)

julia> @allocated fun(x)
758064

julia> @allocated fun(x)
0

[dpo]

Sure, but that’s not really what you’re doing above with jtprod!(). What you are doing is basically:

julia> function fun(x) x .= 2; return nothing end
fun (generic function with 1 method)

julia> x = ones(50);

julia> @allocated fun(x)
1072856

julia> @allocated fun(x)
0

julia> @allocated fun(view(x, 1:15))
3420376

julia> @allocated fun(view(x, 1:15))
48

[tmigot]

Ok, I see, so the problem would be solve if we had re-implemented jtprod! for the initial problem.

After,

function NLPModels.jtprod!(nlp::NLSLC, x::AbstractVector, v::AbstractVector, Jtv::AbstractVector) 
  return jtprod_lin!(nlp, x, v, Jtv)
end

the allocation is 0.