Investigate use of `std::atomic_flag` instead of `std::binary_semaphore`

[DeveloperPaul123]

std::atomic_flag is the only atomic primitive guaranteed to be lock free. It would be interesting to see if this has any positive impact on performance over std::binary_semaphore.

[DeveloperPaul123]

Did a quick benchmark on quick-bench and it seems that std::binary_semaphore has the best performance when it comes to ping/pong which I think matches well with how we're using it in the thread pool (as a signal mechanism).

https://quick-bench.com/q/JkZjpTgsjQkSiyI20IRcZtEFNso

[DeveloperPaul123]

I wrote a quick benchmark and ran it locally on windows and am getting inconsistent results. I think I will need to do a more proper test with pyperf on a Linux machine to get better numbers.

relative	ms/op	op/s	err%	total	Thread signaling
100.0%	451.75	2.21	1.8%	53.95	std::atomic_flag
99.2%	455.40	2.20	2.0%	54.42	std::binary_semaphore

relative	ms/op	op/s	err%	total	Thread signaling
100.0%	444.76	2.25	2.1%	53.20	std::atomic_flag
98.4%	452.16	2.21	3.0%	53.78	std::binary_semaphore

relative	ms/op	op/s	err%	total	Thread signaling
100.0%	485.05	2.06	0.3%	57.51	std::atomic_flag
103.0%	470.99	2.12	0.8%	57.44	std::binary_semaphore

relative	ms/op	op/s	err%	total	Thread signaling
100.0%	457.13	2.19	1.9%	55.43	std::atomic_flag
96.9%	471.90	2.12	3.8%	56.81	std::binary_semaphore

relative	ms/op	op/s	err%	total	Thread signaling
100.0%	481.77	2.08	2.8%	56.86	std::atomic_flag
105.2%	457.91	2.18	4.1%	54.98	std::binary_semaphore

relative	ms/op	op/s	err%	total	Thread signaling
100.0%	465.74	2.15	0.4%	55.85	std::atomic_flag
101.3%	459.84	2.17	0.6%	55.00	std::binary_semaphore

relative	ms/op	op/s	err%	total	Thread signaling
100.0%	453.31	2.21	0.9%	54.72	std::atomic_flag
95.0%	477.07	2.10	2.4%	56.92	std::binary_semaphore

relative	ms/op	op/s	err%	total	Thread signaling
100.0%	477.71	2.09	0.8%	57.59	std::atomic_flag
101.6%	470.13	2.13	0.6%	56.57	std::binary_semaphore

relative	ms/op	op/s	err%	total	Thread signaling
100.0%	477.80	2.09	1.1%	57.12	std::atomic_flag
100.4%	475.79	2.10	1.8%	56.92	std::binary_semaphore

relative	ms/op	op/s	err%	total	Thread signaling
100.0%	479.51	2.09	1.2%	58.10	std::atomic_flag
102.9%	465.78	2.15	4.3%	55.52	std::binary_semaphore

[jtd-formlabs]

here are 10 runs on a linux system with pyperf system tune set up:

relative	ms/op	op/s	err%	total	Thread signaling
100.0%	406.13	2.46	1.1%	48.78	std::atomic_flag
93.8%	432.78	2.31	1.9%	52.17	std::binary_semaphore

relative	ms/op	op/s	err%	total	Thread signaling
100.0%	409.78	2.44	0.7%	49.23	std::atomic_flag
106.2%	385.80	2.59	0.7%	46.45	std::binary_semaphore

relative	ms/op	op/s	err%	total	Thread signaling
100.0%	402.90	2.48	0.7%	48.30	std::atomic_flag
101.1%	398.52	2.51	0.8%	47.77	std::binary_semaphore

relative	ms/op	op/s	err%	total	Thread signaling
100.0%	397.33	2.52	0.4%	47.64	std::atomic_flag
104.8%	379.24	2.64	0.7%	45.69	std::binary_semaphore

relative	ms/op	op/s	err%	total	Thread signaling
100.0%	372.06	2.69	0.9%	44.75	std::atomic_flag
88.7%	419.39	2.38	2.2%	50.05	std::binary_semaphore

relative	ms/op	op/s	err%	total	Thread signaling
100.0%	420.61	2.38	0.9%	50.24	std::atomic_flag
112.4%	374.31	2.67	0.8%	45.01	std::binary_semaphore

relative	ms/op	op/s	err%	total	Thread signaling
100.0%	394.11	2.54	0.9%	47.54	std::atomic_flag
97.8%	403.07	2.48	0.6%	48.64	std::binary_semaphore

relative	ms/op	op/s	err%	total	Thread signaling
100.0%	406.72	2.46	0.7%	48.58	std::atomic_flag
105.2%	386.67	2.59	1.1%	46.27	std::binary_semaphore

relative	ms/op	op/s	err%	total	Thread signaling
100.0%	409.09	2.44	0.3%	49.27	std::atomic_flag
107.2%	381.71	2.62	1.0%	45.74	std::binary_semaphore

relative	ms/op	op/s	err%	total	Thread signaling
100.0%	394.07	2.54	0.7%	47.26	std::atomic_flag
90.8%	434.10	2.30	0.6%	52.22	std::binary_semaphore

[DeveloperPaul123]

@jtd-formlabs Thanks for doing that. It seems like they're essentially the same. At most we'd be saving tens of milliseconds, so it seems like it's not worth it to me.