Slack models on GPU

[tmigot]

using Test, NLPModels, CUDA, NLPModelsModifiers, NLPModelsTest
nlp = LINSV(CuArray{Float64})
SlackModel(nlp) # fails

fails, because we expected NLPModelMeta{S, T} if nlp is of type AbstractNLPModel{S, T} (see

NLPModelsModifiers.jl/src/slack-model.jl

Line 70 in a9fa774

mutable struct SlackModel{T, S, M <: AbstractNLPModel{T, S}} <: AbstractNLPModel{T, S}

).
However, nlp is {Float64, CuArray{Float64}} while the "slack" meta is {Float64, CuArray{Float64, 1, CUDA.Mem.DeviceBuffer}}

slackMeta = NLPModelsModifiers.slack_meta(meta) # https://github.com/JuliaSmoothOptimizers/NLPModelsModifiers.jl/blob/a9fa77496346fd5da171535295d949b570ac2f5a/src/slack-model.jl#L115

@amontoison an opinion on this? #114 is a suggestion to fix this.

[amontoison]

I suggest to use S = eltype(v) with v a vector allocated with CuVector{Float64}(undef, n) when we create the NLPModel.

[tmigot]

Not sure, I fully understand your suggestion. My question is more do you know the difference between CuArray{Float64, 1, CUDA.Mem.DeviceBuffer} and CuArray{Float64, 1}? and is it a problem to cast CuArray{Float64, 1, CUDA.Mem.DeviceBuffer} into CuArray{Float64, 1}?

[amontoison]

Not sure, I fully understand your suggestion. My question is more do you know the difference between CuArray{Float64, 1, CUDA.Mem.DeviceBuffer} and CuArray{Float64, 1}? and is it a problem to cast CuArray{Float64, 1, CUDA.Mem.DeviceBuffer} into CuArray{Float64, 1}?

Yes, I know the difference.
The last attribute is the type of memory allocated by CUDA.
It can be host, device or unified memory.

I suggest this blog to understand it:
https://juliagpu.org/post/2023-11-07-cuda_5.1/