Add OpenMP Union Find

[WingCode]

closes #73

[quantumgizmos]

Thank you for the PR. Could you provide benchmarks to show the performance under different numbers of cores for surface code decoding?

Thanks
JR

[WingCode]

@quantumgizmos
I have included the benchmarked test case using 'python_test/pcms/hx_surface_20.npz'

Here are the benchmark results:

Threads	Time (µs)
1	48.493 µs
4	41.062 µs
8	39.989 µs

[quantumgizmos]

It's reassuring to see that the time is decreasing. However, I don't think timing over a single syndrome provides a realstic benchmark.

Could you time over a number of syndromes dervied from randomly generated error patterns? It would also be useful to show how the advantage of parallisation changes with different pysical error rates $p$.

Benchmaking procedure

1. For p=0.01 sample >1000 bit strings
2. Use the parity check matrix to calculate syndromes for each of the randomly generated bit strings.
3. For thread_count={1,2,4,8} time how long it takes to decode all the syndromes using the UnionFind Decoder. Calculate the average time per decode call.
4. Repeat for values of p between 0.01 and 0.09

[WingCode]

@quantumgizmos I have enhanced the benchmark to incorporate your feedback. Running this benchmark for p=0.01 and 0.03. The p=0.05 on single thread is taking a quite bit of time to run... I will update once I have more info.

p	1 thread	2 threads	4 threads	8 threads
0.01	158 µs	173 µs	182 µs	183 µs
0.03	296 µs	269 µs	261 µs	256 µs

It's reassuring to see that the time is decreasing. However, I don't think timing over a single syndrome provides a realstic benchmark.

Could you time over a number of syndromes dervied from randomly generated error patterns? It would also be useful to show how the advantage of parallisation changes with different pysical error rates p .

Benchmaking procedure

1. For p=0.01 sample >1000 bit strings
2. Use the parity check matrix to calculate syndromes for each of the randomly generated bit strings.
3. For thread_count={1,2,4,8} time how long it takes to decode all the syndromes using the UnionFind Decoder. Calculate the average time per decode call.
4. Repeat for values of p between 0.01 and 0.09

[quantumgizmos]

Thanks for the new bencmarks. It seems like the parallisation overhead is slowing the decoder. Could you review sections where you've used the OMP critical directive to see whether further parallelisation is possible?

[WingCode]

@quantumgizmos
The latest changes have noticeably boosted the benchmark—thanks for that! I’m seeing excellent single-thread gains, but the improvement becomes less clear as we scale to more threads. Whenever you have a moment, could you share where we could focus next to squeeze out additional multithreaded performance? Your seasoned perspective would help me zero in on the right spots :)

p	1 thread	2 threads	4 threads	8 threads
0.01	61 µs	58 µs	62 µs	55 µs
0.03	105 µs	103 µs	105 µs	105 µs

[WingCode]

@quantumgizmos Also, just a gentle reminder that Monday was mentioned as the last day for reviewing open PRs. If there are any specific changes or clarifications you'd like me to make on this one, I’d be happy to take care of them promptly late Sunday. And if it’s helpful to keep iterating, I’d be more than happy to continue working on it if the issue gets assigned to me.

[quantumgizmos]

Hi @WingCode. Thank you for you work on this. So that you can be issued the UnitaryReward, could you make a comment on issue #73 . I can then assign the issue to you and the UnitaryFund will be in touch.

[quantumgizmos]

If you are interested on working on this further, we can set up a call to discuss. I think we will have to rethink the way the algorithm has been implemented from the bottom up. It doesn't look like OpenMP makes much of a difference with the current data structures.

[WingCode]

If you are interested on working on this further, we can set up a call to discuss. I think we will have to rethink the way the algorithm has been implemented from the bottom up. It doesn't look like OpenMP makes much of a difference with the current data structures.

I'd be happy to continue working on this! Is there a preferred channel where I can reach out to you?

[quantumgizmos]

@WingCode Great! For future discussions, my email is joschka@roffe.eu