[Intel-gfx] [igt-dev] [PATCH i-g-t 1/2] igt/perf_pmu: Aim for a fixed number of iterations for calibrating accuracy

[Tvrtko Ursulin]

On 10/08/2018 14:25, Chris Wilson wrote:
> Quoting Tvrtko Ursulin (2018-08-09 12:54:41)
>>
>> On 08/08/2018 15:59, Chris Wilson wrote:
>>> Our observation is that the systematic error is proportional to the
>>> number of iterations we perform; the suspicion is that it directly
>>> correlates with the number of sleeps. Reduce the number of iterations,
>>> to try and keep the error in check.
>>>
>>> Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
>>> Cc: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
>>> ---
>>>    tests/perf_pmu.c | 34 +++++++++++++++++++++-------------
>>>    1 file changed, 21 insertions(+), 13 deletions(-)
>>>
>>> diff --git a/tests/perf_pmu.c b/tests/perf_pmu.c
>>> index 9a20abb6b..5a26d5272 100644
>>> --- a/tests/perf_pmu.c
>>> +++ b/tests/perf_pmu.c
>>> @@ -1521,14 +1521,13 @@ static void __rearm_spin_batch(igt_spin_t *spin)
>>>    
>>>    static void
>>>    accuracy(int gem_fd, const struct intel_execution_engine2 *e,
>>> -      unsigned long target_busy_pct)
>>> +      unsigned long target_busy_pct,
>>> +      unsigned long target_iters)
>>>    {
>>> -     unsigned long busy_us = 10000 - 100 * (1 + abs(50 - target_busy_pct));
>>> -     unsigned long idle_us = 100 * (busy_us - target_busy_pct *
>>> -                             busy_us / 100) / target_busy_pct;
>>>        const unsigned long min_test_us = 1e6;
>>> -     const unsigned long pwm_calibration_us = min_test_us;
>>> -     const unsigned long test_us = min_test_us;
>>> +     unsigned long pwm_calibration_us;
>>> +     unsigned long test_us;
>>> +     unsigned long cycle_us, busy_us, idle_us;
>>>        double busy_r, expected;
>>>        uint64_t val[2];
>>>        uint64_t ts[2];
>>> @@ -1538,18 +1537,27 @@ accuracy(int gem_fd, const struct intel_execution_engine2 *e,
>>>        /* Sampling platforms cannot reach the high accuracy criteria. */
>>>        igt_require(gem_has_execlists(gem_fd));
>>>    
>>> -     while (idle_us < 2500) {
>>> +     /* Aim for approximately 100 iterations for calibration */
>>> +     cycle_us = min_test_us / target_iters;
>>> +     busy_us = cycle_us * target_busy_pct / 100;
>>> +     idle_us = cycle_us - busy_us;
>>
>> 2% load, 1s / 10 iters
>>          cycles_us = 100ms
>>          busy_us = 2ms
>>          idle_us = 98ms
>> ...
>>
>>> +
>>> +     while (idle_us < 2500 || busy_us < 2500) {
>>>                busy_us *= 2;
>>>                idle_us *= 2;
>>
>> ...
>>
>> busy_us = 4ms
>> idle_us = 196ms
> 
> Currently it is 250ms per 98:2 cycle and about 20ms per 50:50 cycle. So
> we are only doing 4 and 50 iterations respectively.
> 
> 10 cycles is strictly an improvement :-p

Hmm indeed. It seems I misremembered how it works. I'll re-read your 
patches.

Regards,

Tvrtko