Re: [PATCH v2 i-g-t] tests/xe_gt_freq: Avoid RPe usage in subtests

["Belgaumkar, Vinay" <vinay.belgaumkar@intel.com>]

On 7/29/2024 9:47 PM, Riana Tauro wrote:
> Hi Vinay
>
> On 7/30/2024 5:47 AM, Vinay Belgaumkar wrote:
>> We are seeing several instances where the RPe, which can be altered by
>> pcode dynamically, is causing subtests to fail randomly. Instead of 
>> relying
>> on it, we can use a mid frequency value for these subtests and avoid 
>> these
>> failures.
>>
>> v2: Fix bug in the tolerance function. Remove rpe usage from freq_range*
>> as well (Badal). Fix test documentation to reflect change (Riana).
>>
>> Link: https://gitlab.freedesktop.org/drm/xe/kernel/-/issues/2200
>> Link: https://gitlab.freedesktop.org/drm/xe/kernel/-/issues/2196
>> Link: https://gitlab.freedesktop.org/drm/xe/kernel/-/issues/2262
>> Link: https://gitlab.freedesktop.org/drm/xe/kernel/-/issues/2256
>> Link: https://gitlab.freedesktop.org/drm/xe/kernel/-/issues/2203
>> Link: https://gitlab.freedesktop.org/drm/xe/kernel/-/issues/2412
>>
>> Cc: Riana Tauro <riana.tauro@intel.com>
>> Cc: Badal Nilawar <badal.nilawar@intel.com>
>> Signed-off-by: Vinay Belgaumkar <vinay.belgaumkar@intel.com>
>> ---
>>   tests/intel/xe_gt_freq.c | 53 ++++++++++++++++++++++++++++------------
>>   1 file changed, 37 insertions(+), 16 deletions(-)
>>
>> diff --git a/tests/intel/xe_gt_freq.c b/tests/intel/xe_gt_freq.c
>> index 93ebb5ed0..1ada37834 100644
>> --- a/tests/intel/xe_gt_freq.c
>> +++ b/tests/intel/xe_gt_freq.c
>> @@ -26,6 +26,9 @@
>>   #include <sys/time.h>
>>     #define MAX_N_EXEC_QUEUES 16
>> +#define GT_FREQUENCY_MULTIPLIER    50
>> +#define GT_FREQUENCY_SCALER    3
>> +#define FREQ_UNIT_MHZ DIV_ROUND_CLOSEST(GT_FREQUENCY_MULTIPLIER, 
>> GT_FREQUENCY_SCALER)
>>     /*
>>    * Too many intermediate components and steps before freq is adjusted
>> @@ -70,9 +73,14 @@ static uint32_t get_freq(int fd, int gt_id, const 
>> char *freq_name)
>>       return freq;
>>   }
>>   -static uint32_t rpe(int fd, int gt_id)
>> +static bool within_expected_range(uint32_t freq, uint32_t val)
>>   {
>> -    return get_freq(fd, gt_id, "rpe");
>> +    /*
>> +     * GT Frequencies are requested at units of 16.66 Mhz, so allow
>> +     * that tolerance.
>> +     */
>> +    return (freq <= val + FREQ_UNIT_MHZ) &&
>> +        (freq >= val - FREQ_UNIT_MHZ);
>>   }
>>     static uint32_t get_throttle(int fd, int gt_id, const char 
>> *throttle_file)
>> @@ -128,6 +136,8 @@ static void test_freq_basic_api(int fd, int gt_id)
>>   {
>>       uint32_t rpn = get_freq(fd, gt_id, "rpn");
>>       uint32_t rp0 = get_freq(fd, gt_id, "rp0");
>> +    uint32_t rpmid = (rp0 + rpn) / 2;
>> +    uint32_t min_freq, max_freq;
>>         /*
>>        * Negative bound tests
>> @@ -142,16 +152,18 @@ static void test_freq_basic_api(int fd, int gt_id)
>>       /* Assert min requests are respected from rp0 to rpn */
>>       igt_assert(set_freq(fd, gt_id, "min", rp0) > 0);
>>       igt_assert(get_freq(fd, gt_id, "min") == rp0);
>> -    igt_assert(set_freq(fd, gt_id, "min", rpe(fd, gt_id)) > 0);
>> -    igt_assert(get_freq(fd, gt_id, "min") == rpe(fd, gt_id));
>> +    igt_assert(set_freq(fd, gt_id, "min", rpmid) > 0);
>> +    min_freq = get_freq(fd, gt_id, "min");
> Seeing the first issue after this change. Do we need a delay?
>
> https://intel-gfx-ci.01.org/tree/intel-xe/IGTPW_11490/bat-adlp-7/igt@xe_gt_freq@freq_basic_api.html 
>

Hmm, rpmid here would have been 650. SLPC ended up setting min to RPe, 
which is probably 700 here. Will change this conditional check.

Thanks,

Vinay.

>
> Thanks,
> Riana
>
>
>> + igt_assert(within_expected_range(min_freq, rpmid));
>>       igt_assert(set_freq(fd, gt_id, "min", rpn) > 0);
>>       igt_assert(get_freq(fd, gt_id, "min") == rpn);
>>         /* Assert max requests are respected from rpn to rp0 */
>>       igt_assert(set_freq(fd, gt_id, "max", rpn) > 0);
>>       igt_assert(get_freq(fd, gt_id, "max") == rpn);
>> -    igt_assert(set_freq(fd, gt_id, "max", rpe(fd, gt_id)) > 0);
>> -    igt_assert(get_freq(fd, gt_id, "max") == rpe(fd, gt_id));
>> +    igt_assert(set_freq(fd, gt_id, "max", rpmid) > 0);
>> +    max_freq = get_freq(fd, gt_id, "max");
>> +    igt_assert(within_expected_range(max_freq, rpmid));
>>       igt_assert(set_freq(fd, gt_id, "max", rp0) > 0);
>>       igt_assert(get_freq(fd, gt_id, "max") == rp0);
>>   }
>> @@ -168,13 +180,16 @@ static void test_freq_fixed(int fd, int gt_id, 
>> bool gt_idle)
>>   {
>>       uint32_t rpn = get_freq(fd, gt_id, "rpn");
>>       uint32_t rp0 = get_freq(fd, gt_id, "rp0");
>> +    uint32_t rpmid = (rp0 + rpn) / 2;
>> +    uint32_t cur_freq, act_freq;
>>         igt_debug("Starting testing fixed request\n");
>>         /*
>>        * For Fixed freq we need to set both min and max to the 
>> desired value
>>        * Then we check if hardware is actually operating at the 
>> desired freq
>> -     * And let's do this for all the 3 known Render Performance (RP) 
>> values.
>> +     * And let's do this for all the 2 known Render Performance (RP) 
>> values
>> +     * RP0 and RPn and something in between.
>>        */
>>       igt_assert(set_freq(fd, gt_id, "min", rpn) > 0);
>>       igt_assert(set_freq(fd, gt_id, "max", rpn) > 0);
>> @@ -190,17 +205,19 @@ static void test_freq_fixed(int fd, int gt_id, 
>> bool gt_idle)
>>           igt_assert(get_freq(fd, gt_id, "act") == rpn);
>>       }
>>   -    igt_assert(set_freq(fd, gt_id, "min", rpe(fd, gt_id)) > 0);
>> -    igt_assert(set_freq(fd, gt_id, "max", rpe(fd, gt_id)) > 0);
>> +    igt_assert(set_freq(fd, gt_id, "min", rpmid) > 0);
>> +    igt_assert(set_freq(fd, gt_id, "max", rpmid) > 0);
>>       usleep(ACT_FREQ_LATENCY_US);
>> -    igt_assert(get_freq(fd, gt_id, "cur") == rpe(fd, gt_id));
>> +    cur_freq = get_freq(fd, gt_id, "cur");
>> +    igt_assert(within_expected_range(cur_freq, rpmid));
>>         if (gt_idle) {
>>           igt_assert_f(igt_wait(xe_is_gt_in_c6(fd, gt_id), 1000, 10),
>>                    "GT %d should be in C6\n", gt_id);
>>           igt_assert(get_freq(fd, gt_id, "act") == 0);
>>       } else {
>> -        igt_assert(get_freq(fd, gt_id, "act") == rpe(fd, gt_id));
>> +        act_freq = get_freq(fd, gt_id, "act");
>> +        igt_assert(within_expected_range(act_freq, cur_freq));
>>       }
>>         igt_assert(set_freq(fd, gt_id, "min", rp0) > 0);
>> @@ -232,15 +249,17 @@ static void test_freq_fixed(int fd, int gt_id, 
>> bool gt_idle)
>>   static void test_freq_range(int fd, int gt_id, bool gt_idle)
>>   {
>>       uint32_t rpn = get_freq(fd, gt_id, "rpn");
>> +    uint32_t rp0 = get_freq(fd, gt_id, "rp0");
>> +    uint32_t rpmid = (rp0 + rpn) / 2;
>>       uint32_t cur, act;
>>         igt_debug("Starting testing range request\n");
>>         igt_assert(set_freq(fd, gt_id, "min", rpn) > 0);
>> -    igt_assert(set_freq(fd, gt_id, "max", rpe(fd, gt_id)) > 0);
>> +    igt_assert(set_freq(fd, gt_id, "max", rpmid) > 0);
>>       usleep(ACT_FREQ_LATENCY_US);
>>       cur = get_freq(fd, gt_id, "cur");
>> -    igt_assert(rpn <= cur && cur <= rpe(fd, gt_id));
>> +    igt_assert(rpn <= cur && cur <= rpmid + FREQ_UNIT_MHZ);
>>         if (gt_idle) {
>>           igt_assert_f(igt_wait(xe_is_gt_in_c6(fd, gt_id), 1000, 10),
>> @@ -248,7 +267,7 @@ static void test_freq_range(int fd, int gt_id, 
>> bool gt_idle)
>>           igt_assert(get_freq(fd, gt_id, "act") == 0);
>>       } else {
>>           act = get_freq(fd, gt_id, "act");
>> -        igt_assert(rpn <= act && act <= rpe(fd, gt_id));
>> +        igt_assert(rpn <= act && within_expected_range(act, cur));
>>       }
>>         igt_debug("Finished testing range request\n");
>> @@ -262,17 +281,19 @@ static void test_freq_range(int fd, int gt_id, 
>> bool gt_idle)
>>   static void test_freq_low_max(int fd, int gt_id)
>>   {
>>       uint32_t rpn = get_freq(fd, gt_id, "rpn");
>> +    uint32_t rp0 = get_freq(fd, gt_id, "rp0");
>> +    uint32_t rpmid = (rp0 + rpn) / 2;
>>         /*
>>        *  When max request < min request, max is ignored and min 
>> works like
>>        * a fixed one. Let's assert this assumption
>>        */
>> -    igt_assert(set_freq(fd, gt_id, "min", rpe(fd, gt_id)) > 0);
>> +    igt_assert(set_freq(fd, gt_id, "min", rpmid) > 0);
>>       igt_assert(set_freq(fd, gt_id, "max", rpn) > 0);
>>       usleep(ACT_FREQ_LATENCY_US);
>>         /* Refresh value of rpe, pcode could have adjusted it */
>> -    igt_assert(get_freq(fd, gt_id, "cur") == rpe(fd, gt_id));
>> +    igt_assert(within_expected_range(get_freq(fd, gt_id, "cur"), 
>> rpmid));
>>   }
>>     /**