lava-nc
Speed of QP solver is very slow
The functioning of the QP solver is extremely slow compared to a pure numpy implementation. This is in reference to Issue https://github.com/lava-nc/lava/issues/36
What is this story about? Could you provide more context?
- Is this after Harry's fix?
- How does this relate to the original issue and the first solution from Harry?
- What do you expect to happen next?
- ...
Before Harry's fix in https://github.com/lava-nc/lava/pull/87, the CSP based implementation was 5000x slower than pure numpy implementation as pointed out in lava-nc/lava#36. After the fix there is an ~50x improvement in speed as compared to before but it's still 100x slower than a pure numpy implementation. (Intel Core i5-5257U with 8GB RAM, running Ubuntu 20.04.2 LTS with lava v0.2.0 vs v0.1.1) This issue was created to be a reference in lava-optimization for the speed improvement. An issue for further speed up (if possible and necessary) will be created separately.
@ashishrao7 what is the status of this issue? was there any progress or is the SW lava backend still slower than numpy?
No the software backed is now much faster than it was when the issue was made. I think the speed issue after Harry's fix was because of a hacky primitive profiling tool that I had developed and was using. One also has to keep in mind that using get()/set() can also make it slow because of the serial nature of get()/set(). Closing this issue now.