On 07/23/2015 09:13 AM, Vince Weaver wrote:
> On Thu, 23 Jul 2015, William Cohen wrote:
>
>> Right now the output of the test is looking wrong on the cortex ac53 board
>> and I need to do some additional investigation. The ac53 events are very
>> similar to other armv8 hardware. I should get an xgene or seattle machine
>> and do some comparisons between them. The problem could be in places other
>> than libpfm4 such as the kernel.
>
> The problem is in PAPI, the workaround for the lack of kernel/user
> separation pre Cortex-A15 was never updated.
>
> see the code in the perf_event component:
>
> /* ARM */
> if ( _papi_hwi_system_info.hw_info.vendor == PAPI_VENDOR_ARM) {
> /* FIXME: this will change with Cortex A15 */
> vector->cmp_info.available_domains |=
> PAPI_DOM_USER | PAPI_DOM_KERNEL | PAPI_DOM_SUPERVISOR;
> vector->cmp_info.default_domain =
> PAPI_DOM_USER | PAPI_DOM_KERNEL | PAPI_DOM_SUPERVISOR;
> }
>
> Vince
>
Hi Vince,
Thanks for pointing that out. Is there some way to get information describing
the version of the pmu hardware rather than directly using the processor
identification? It would be nice to avoid enumerating every possible processor
that support the domain selection.
Below are some quick tests with perf stat to what is returned for setting the
masks.
on cortex a9 machine
Performance counter stats for 'du':
12,522,383 cycles
0 cycles:u
0 cycles:k
0.152377046 seconds time elapsed
on cortex a53 machine:
Performance counter stats for 'du':
6262550 cycles
1753114 cycles:u [30.50%]
<not counted> cycles:k
0.037436820 seconds time elapsed
-Will
------------------------------------------------------------------------------
_______________________________________________
perfmon2-devel mailing list
perfmon2-devel@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/perfmon2-devel
