Hi,
On Tue, Nov 07, 2023 at 05:51:15PM +0800, heqiong wrote:
> Reducing the impact of the cntvct_el0 register and isb() operation
> on microbenchmark test results to improve testing accuracy and reduce
> latency in test results.
> ---
> arm/micro-bench.c | 16 ++++++++++------
> 1 file changed, 10 insertions(+), 6 deletions(-)
>
> diff --git a/arm/micro-bench.c b/arm/micro-bench.c
> index fbe59d03..65f4c4dd 100644
> --- a/arm/micro-bench.c
> +++ b/arm/micro-bench.c
> @@ -346,17 +346,21 @@ static void loop_test(struct exit_test *test)
> }
> }
>
> + dsb(ish);
> + isb();
> + start = read_sysreg(cntvct_el0);
> + isb();
> while (ntimes < test->times && total_ns.ns < NS_5_SECONDS) {
^^^^^^^^^^^^^^^^^^^^^^^^^^
This will always evaluate to true because total_ns is now computed at the
end of the loop instead of every iteration.
Do we want to drop the upper bound on how long a test takes to execute? I
don't have an opinion about it.
Thanks,
Alex
> - isb();
> - start = read_sysreg(cntvct_el0);
> test->exec();
> - isb();
> - end = read_sysreg(cntvct_el0);
>
> ntimes++;
> - total_ticks += (end - start);
> - ticks_to_ns_time(total_ticks, &total_ns);
> }
> + dsb(ish);
> + isb();
> + end = read_sysreg(cntvct_el0);
> +
> + total_ticks = end - start;
> + ticks_to_ns_time(total_ticks, &total_ns);
>
> if (test->post) {
> test->post(ntimes, &total_ticks);
> --
> 2.39.3
>
