tests/periph_gpio: reduce benchmark iterations

[kaspar030]

Contribution description

Our MCU's are predictable enough that 100k iterations should suffice.
No need to wait tens of seconds.

Testing procedure

Issues/PRs references

Found in #12457, was about to increase test time for z1.

[benpicco]

If Murdock is happy, this sould be fine.
(But can Murdock even be truly happy with the current nrf52dk situation?)

[gschorcht]

I'm wondering whether it will solve the benchmark problem on ATmega. When I was benchmarking the new GPIO API in PR #14607, I had either to reduce the number of tests to 1000 or to remove the irq_disable and irq_enable calls. With disabled interrupts, the results of the benchmarks were completely nonsense probabyl because the timer was not working correctly. I will try later today.

[kaspar030]

I'm wondering whether it will solve the benchmark problem on ATmega. When I was benchmarking the new GPIO API in PR #14607, I had either to reduce the number of tests to 1000 or to remove the irq_disable and irq_enable calls. With disabled interrupts, the results of the benchmarks were completely nonsense probabyl because the timer was not working correctly. I will try later today.

that (results completely nonsense) probably was caused by #14639!

[kaspar030]

remove the irq_disable and irq_enable calls.

ah, didn't read good, you had that figured out already...

[gschorcht]

Unfortunatly, 100k iterations are is still too much for ATmega. With bench 0 4, I get

                 gpio_set:    231624us  ---   2.316us per call  ---     431734 calls per sec
               gpio_clear:      6984us  ---   0.069us per call  ---   14318442 calls per sec
              gpio_toggle:      1444us  ---   0.014us per call  ---   69252077 calls per sec
                gpio_read:      6984us  ---   0.069us per call  ---   14318442 calls per sec
               gpio_write:     13232us  ---   0.132us per call  ---    7557436 calls per sec

It makes no sense that gpio_toggle would need only 0.014 us while gpio_set takes 2.316 us although toggle is realized as a combination of first reading the pin and then setting or clearing it 🤔

Just for information, real values are:

                 gpio_set:   4940248us  ---   4.940us per call  ---     202418 calls per sec
               gpio_clear:   5315464us  ---   5.315us per call  ---     188130 calls per sec
              gpio_toggle:  10505756us  ---  10.505us per call  ---      95185 calls per sec
                gpio_read:   5315388us  ---   5.315us per call  ---     188133 calls per sec
               gpio_write:   5377964us  ---   5.377us per call  ---     185943 calls per sec

which makes much more sense.

[gschorcht]

remove the irq_disable and irq_enable calls.

ah, didn't read good, you had that figured out already...

Yeah, a year ago when I was already working on the new GPIO API. But this should be solved with PR #14639.

[gschorcht]

remove the irq_disable and irq_enable calls.

ah, didn't read good, you had that figured out already...

That's why, I also used an own version in tests/periph_gpio_exp in PR #14602 😉

RIOT/tests/periph_gpio_exp/main.c

Lines 33 to 48 in 70db5f7

	/**
	* BENCHMARK_FUNC disables the interrupts and does not work correctly on
	* some platforms, e.g. ATmega. On other platforms, such as STM32, the
	* benchmark test gives the same results with interrupts not disabled.
	* Therefore a modified version is used here that does not disable the
	* interrupts.
	*/
	#define MY_BENCHMARK_FUNC(name, runs, func) \
	{ \
	uint32_t _benchmark_time = xtimer_now_usec(); \
	for (unsigned long i = 0; i < runs; i++) { \
	func; \
	} \
	_benchmark_time = (xtimer_now_usec() - _benchmark_time);\
	benchmark_print_time(_benchmark_time, runs, name); \
	}

So, I can remove it once this PR is merged.

[gschorcht]

It seems that Murdock failed due to a reason that is not related to this PR.

[kaspar030]

Unfortunatly, 100k iterations are is still too much for ATmega. With bench 0 4, I get

too much shouldn't happen, just take longer. I think this pr didn't include the irq disable fix PR. I've rebased, so it should now. Can you double check?

[gschorcht]

Unfortunatly, 100k iterations are is still too much for ATmega. With bench 0 4, I get

too much shouldn't happen, just take longer. I think this pr didn't include the irq disable fix PR. I've rebased, so it should now. Can you double check?

Cool, now it is correct 👍

                 gpio_set:    494064us  ---   4.940us per call  ---     202402 calls per sec
               gpio_clear:    531564us  ---   5.315us per call  ---     188124 calls per sec
              gpio_toggle:   1050596us  ---  10.505us per call  ---      95184 calls per sec
                gpio_read:    531636us  ---   5.316us per call  ---     188098 calls per sec
               gpio_write:    537812us  ---   5.378us per call  ---     185938 calls per sec

[kaspar030]

Unrelated tests failed:

--- run_test job results (3 failed, 202 passed, 205 total):
    failed:
    tests/pkg_relic/nrf52dk:gnu
    tests/suit_manifest/nrf52dk:gnu
    tests/xtimer_usleep/nrf52dk:llvm

will re-trigger.

[kaspar030]

Cool, now it is correct +1

thanks for confirming!

[gschorcht]

Murdock is happy, failed tests before were not related to this PR.