Update to support SCS 3.0 about scs HOT 8 CLOSED

Hi Brendan,

Thank you for the info and sorry for the late reply.
This week I am still quite bussy but I will have a look at it next week.

from scs.

Hi @FlorianSchwendinger and @bodono , I've already made all the changes to the R scs package for 3.0. I also added all the tests that Brendan had and they pass. I'll commit and make a PR very soon.

from scs.

@bnaras how do you want proceed? Should I wait for the changes you suggested? We could also upload the new (current) version now and later make the name changes. What do you think?

@FlorianSchwendinger I will commit these changes later this week and make a PR.

from scs.

Continuing pull/8#issuecomment-939350343 I tried the openMP support but saw in htop that it is currently using already all of my cores which solving bigger problems and setting the environment variables did not change that. So I will look later into this.

@bnaras how do you want proceed? Should I wait for the changes you suggested? We could also upload the new (current) version now and later make the name changes. What do you think?

from scs.

I think the only way it could be using more cores without openMP is from linking in a multi-threaded blas / lapack. It uses blas / lapack to do the Anderson acceleration and to project onto the positive semidefinite cone.

from scs.

Continuing pull/8#issuecomment-939350343 I tried the openMP support but saw in htop that it is currently using already all of my cores which solving bigger problems and setting the environment variables did not change that. So I will look later into this.

I tried looking into this myself too and I think that this requires linking to indirect version of the solver. We currently only link to the direct version for the R package. Could that make a difference @bodono ?

from scs.

The direct version of the solver is the main version, the indirect version is really only required for huge problems. Other than the indirect solver, openMP is only used to parallelize some of the cone projections (eg, exponential cone). I think it's fine to not have it on by default if it's causing issues.

from scs.

The direct version of the solver is the main version, the indirect version is really only required for huge problems. Other than the indirect solver, openMP is only used to parallelize some of the cone projections (eg, exponential cone). I think it's fine to not have it on by default if it's causing issues.

Ok, very interesting. I can have a look at adding the indirect solver, for the next release.

from scs.