matrix-less dQ #59
matrix-less dQ #59
Conversation
Codecov Report
@@ Coverage Diff @@
## master #59 +/- ##
==========================================
- Coverage 91.77% 91.39% -0.39%
==========================================
Files 4 4
Lines 304 302 -2
==========================================
- Hits 279 276 -3
- Misses 25 26 +1
Continue to review full report at Codecov.
|
|
Good to merge |
|
if someone wants to have a final look before merging this |
src/MOI_wrapper.jl
Outdated
| end | ||
|
|
||
| du, dv, dw = dz[1:n], dz[n+1:n+m], dz[n+m+1] | ||
| @inbounds (du, dv, dw) = (dz[1:n], dz[n+1:n+m], dz[n+m+1]) |
There was a problem hiding this comment.
This @inbounds may trigger more segfault (when we change this function and forget something) than have any visible effect on perf no ? This isn't a critical part of the code.
There was a problem hiding this comment.
yes right, normally we control the size of dz deduced from the RHS we form but the bound check might be more future-proof
In backward_conic,
dQis materialized only for a matrix-vector product, it is eliminated from the pass.There should also be a way to avoid materializing
Q, and ultimately reasoning on the sparse construction of the problem and not its explicit matrix form