Closed ericphanson closed 1 month ago
All modified and coverable lines are covered by tests :white_check_mark:
Project coverage is 98.09%. Comparing base (
d19955e
) to head (7b10c27
).
:umbrella: View full report in Codecov by Sentry.
:loudspeaker: Have feedback on the report? Share it here.
I guess to hook everything up, we would need to move the sets to MOI and update Hypatia to support them directly
Yes, I think @blegat has a plan for this.
this is the second example from https://github.com/jump-dev/Convex.jl/pull/418#issuecomment-832190560
I thought it could be nice to have some examples using some of the CVXQUAD functionality. I put in a lot of quantum information theory preliminaries, but maybe we could trim some back if it's too much.
This example originally had 23s of formulation time and 4s solve time. I profiled and found
permutedims_matrix
was very slow and allocating (and it gets used a lot, since it is used fortranspose
and thus forvcat
). I was able to optimize it by constructing the sparse matrix directly (in COO format). Now we have 3s formulation time and the same 4s solve time, which seems much more acceptable (maybe a little slow still, but I didn't see any other low hanging fruit).BTW, here the reformulations and bridges really blow up the problem size:
using
I think Hypatia can be used to avoid some of the reformulations; I guess to hook everything up, we would need to move the sets to MOI and update Hypatia to support them directly (?).