add multi threading

[Roger-luo]

Add multi threading support. Pending for benchmark results.

[Roger-luo]

I think controldo can use multi threading as well, but I'll check this later.

[codecov]

Codecov Report

Merging #19 into master will decrease coverage by 0.01%.
The diff coverage is 89.47%.

@@            Coverage Diff             @@
##           master      #19      +/-   ##
==========================================
- Coverage   77.18%   77.16%   -0.02%     
==========================================
  Files           8        8              
  Lines         434      438       +4     
==========================================
+ Hits          335      338       +3     
- Misses         99      100       +1

Impacted Files	Coverage Δ
src/instruct.jl	92.5% <89.47%> (-0.45%)	⬇️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 2f785b6...69e218a. Read the comment docs.

[Roger-luo]

https://gist.github.com/Roger-luo/a6e660228955696f3b6054a9f7129d1d

[Roger-luo]

need to add a threshold to disable multi threading when the number of qubits is not large.

Or should we just let user to decide this? since user can always close it by tweaking JULIA_NUM_THREADS @GiggleLiu

[Roger-luo]

OK so it looks like the overhead of multi threads, should define a threshold to avoid them.

https://gist.github.com/Roger-luo/26682142d70f2d24c8a8bea36518900f

[GiggleLiu]

need to add a threshold to disable multi threading when the number of qubits is not large.

Or should we just let user to decide this? since user can always close it by tweaking JULIA_NUM_THREADS @GiggleLiu

Cool, it is better to let user decide this.

[GiggleLiu]

Please show the number of threads. It looks like always a lot of overhead. Lol

[Roger-luo]

It's in the benchmark report @GiggleLiu

[GiggleLiu]

please check the GPU implementation for some inspiration of multi-threading. controldo can be parallelized.

https://github.com/QuantumBFS/CuYao.jl

[codecov]

Codecov Report

Merging #19 into master will decrease coverage by 1.35%.
The diff coverage is 77.35%.

@@            Coverage Diff             @@
##           master      #19      +/-   ##
==========================================
- Coverage   80.82%   79.47%   -1.36%     
==========================================
  Files           8        8              
  Lines         558      570      +12     
==========================================
+ Hits          451      453       +2     
- Misses        107      117      +10     

Impacted Files	Coverage Δ
src/instruct.jl	90.99% <77.35%> (-4.49%)	⬇️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 932c243...2898e0a. Read the comment docs.

[Roger-luo]

seems to have memory leak on Linux (works fine on Mac for me). Might related to JuliaLang/julia#31923

We need to wait for this issue to be fixed I think.

[Roger-luo]

I think it is fixed on master now: JuliaLang/julia#32217

[Roger-luo]

we need more benchmark reports before we ship this feature.

[GiggleLiu]

Any new benchmark reports?

[Roger-luo]

Not yet. I'll have time on this after JuliaCon

[Roger-luo]

I think if we merge this (after 1.3 is out), but we might need to ask our users to use julia 1.3 for best performance. (I'm not sure if thee new multi-threading support will be backported)

[Roger-luo]

Some updates on this. The @threads macro is compatible since 1.0, thus we can have this for all 1.x, but the performance might be different. I'll encourage people to use latest version if you need multi-threading. I'm still not sure why there's some regression for single thread benchmark on this yet.

I suspect we have some cache line problem when enabling multi-threading, the scaling is not good.

[Roger-luo]

We need to refactor this branch after #30 @GiggleLiu and I need to explore some more benchmark on cache line problem. Hopefully we can deliver this feature this year.

[Roger-luo]

@GiggleLiu This looks good now. I think we could merge this first. I'll make a few following PR to fix some performance issue later.