Re: [igt-dev] [PATCH i-g-t v2 1/2] tests/intel/xe_guc_pc: Add freq-power test

[Riana Tauro <riana.tauro@intel.com>]

On 9/25/2023 7:03 PM, Rodrigo Vivi wrote:
> On Mon, Sep 25, 2023 at 03:24:14PM +0530, Riana Tauro wrote:
>> Hi Rodrigo
>>
>> On 9/21/2023 9:28 PM, Rodrigo Vivi wrote:
>>> On Thu, Sep 21, 2023 at 02:25:07PM +0530, Riana Tauro wrote:
>>>> An assumption is that at lower frequencies,
>>>> not only do we run slower, but we save power compared to
>>>> higher frequencies.
>>>>
>>>> This test runs a spinner and sets the min and max frequencies
>>>> to rp0 and rpn respectively. It then checks if power consumed
>>>> at lower frequencies is lesser than higher frequencies.
>>>>
>>>> v2: Remove Run Type
>>>>       change test name (Kamil)
>>>>       change test documentation and comments (Vinay)
>>>>
>>>> Signed-off-by: Riana Tauro <riana.tauro@intel.com>
>>>> Reviewed-by: Vinay Belgaumkar <vinay.belgaumkar@intel.com>
>>>> ---
>>>>    tests/intel/xe_guc_pc.c | 88 +++++++++++++++++++++++++++++++++++++++++
>>>>    1 file changed, 88 insertions(+)
>>>>
>>>> diff --git a/tests/intel/xe_guc_pc.c b/tests/intel/xe_guc_pc.c
>>>> index 0327d8e0e..2b3d08fcb 100644
>>>> --- a/tests/intel/xe_guc_pc.c
>>>> +++ b/tests/intel/xe_guc_pc.c
>>>> @@ -13,6 +13,7 @@
>>>>    #include "igt.h"
>>>>    #include "lib/igt_syncobj.h"
>>>> +#include "igt_power.h"
>>>>    #include "igt_sysfs.h"
>>>>    #include "xe_drm.h"
>>>> @@ -382,6 +383,82 @@ static void test_reset(int fd, int gt_id, int cycles)
>>>>    	}
>>>>    }
>>>> +static int cmp_u64(const void *a, const void *b)
>>>> +{
>>>> +	return (*(u64 *)a - *(u64 *)b);
>>>> +}
>>>> +
>>>> +static uint64_t measure_power(int fd, struct igt_power *gpu)
>>>> +{
>>>> +	struct power_sample sample[2];
>>>> +	uint64_t power[5];
>>>> +
>>>> +	for (int i = 0; i < 5; i++) {
>>>> +		igt_power_get_energy(gpu, &sample[0]);
>>>> +		usleep(10000); /* 10 ms */
>>>> +		igt_power_get_energy(gpu, &sample[1]);
>>>> +
>>>> +		power[i] = igt_power_get_mW(gpu, &sample[0], &sample[1]);
>>>> +	}
>>>> +	/* Sort in ascending order and use a triangular filter */
>>>> +	qsort(power, 5, sizeof(*power), cmp_u64);
>>>> +	return DIV_ROUND_UP(power[1] + 2 * power[2] + power[3], 4);
>>>> +}
>>>> +
>>>> +/**
>>>> + * SUBTEST: freq-power
>>>> + * Description: Validates power consumed at higher frequencies is more than
>>>> + *              power consumed at lower frequencies.
>>>> + */
>>>> +static void test_freq_power(int fd, int gt_id, struct drm_xe_engine_class_instance *hwe)
>>>> +{
>>>> +	uint32_t rp0, rpn, vm;
>>>> +	uint64_t ahnd;
>>>> +	struct igt_power gpu;
>>>> +	struct {
>>>> +		uint64_t power;
>>>> +		uint32_t freq;
>>>> +	} min, max;
>>>> +	igt_spin_t *spin;
>>>> +
>>>> +	/* Run for engines belonging to the gt */
>>>> +	if (gt_id != hwe->gt_id)
>>>> +		return;
>>>> +
>>>> +	igt_power_open(fd, &gpu, "gpu");
>>>> +
>>>> +	rpn = get_freq(fd, gt_id, "rpn");
>>>> +	rp0 = get_freq(fd, gt_id, "rp0");
>>>> +
>>>> +	vm = xe_vm_create(fd, 0, 0);
>>>> +	ahnd = intel_allocator_open(fd, vm, INTEL_ALLOCATOR_RELOC);
>>>> +	spin = igt_spin_new(fd, .ahnd = ahnd, .vm = vm, .hwe = hwe);
>>>> +
>>>> +	igt_assert(set_freq(fd, gt_id, "min", rpn) > 0);
>>>> +	igt_assert(set_freq(fd, gt_id, "max", rpn) > 0);
>>>> +	min.freq = get_freq(fd, gt_id, "act");
>>>> +	min.power = measure_power(fd, &gpu);
>>>> +
>>>> +	igt_assert(set_freq(fd, gt_id, "min", rp0) > 0);
>>>> +	igt_assert(set_freq(fd, gt_id, "max", rp0) > 0);
>>>> +	max.freq = get_freq(fd, gt_id, "act");
>>>> +	max.power = measure_power(fd, &gpu);
>>>> +
>>>> +	igt_info("Engine %s:%d  min:%lumW @ %uMHz, max:%lumW @ %uMHz\n",
>>>> +		 xe_engine_class_string(hwe->engine_class), hwe->engine_instance,
>>>> +		 min.power, min.freq, max.power, max.freq);
>>>> +
>>>> +	igt_spin_free(fd, spin);
>>>> +	put_ahnd(ahnd);
>>>> +	xe_vm_destroy(fd, vm);
>>>> +	igt_power_close(&gpu);
>>>> +
>>>> +	/* power@max_freq should be at least 10% greater than power@min_freq */
>>>> +	igt_assert_f((11 * min.power < 10 * max.power),
>>>> +		     "%s:%d did not conserve power when setting lower frequency!\n",
>>>> +		     xe_engine_class_string(hwe->engine_class), hwe->engine_instance);
>>>
>>> What exactly are we trying to test here with this case?
>>> This creates an artificial KPI that might not be true for the broader range
>>> of SKUs and generations out there. And then when it fails what should we do?
>>> come here and update the test case?
>> This test was a port from i915 [rps-power & slpc-power]
>>
>> The commit message for the i915 tests is based on an assumption
>>
>> "at lower frequencies, not only do we run
>>   slower, but we save power compared to higher frequencies."
>>
>> I went through the failures there hasn't been a fix anywhere.
>>
>> But shouldn't the power consumed at lower frequencies be lesser than higher
>> (removing the 10% in the above condition)?
> 
> well, our hw is very complex. We have many power gating mechanisms that
> might be saving power even when we select higher frequencies, but are
> not exercising the entire GPU.
> 
> But even if we create a test that are capable of exercising all the EUs,
> CSes, fixed functions and everything, maybe finishing the workload quickly
> on higher frequencies, might trigger more deep lower power states that could
> save power over the slow workload taking longer to execute.
> 
Thanks for the clarification. Will drop this test

Thanks,
Riana
>>
>> I will drop this test if not necessary
> 
> yes, please let's drop it. I'm in favor of killing the i915 as well >
> Thanks,
> Rodrigo.
> 
>>
>> Thanks
>> Riana Tauro
>>>
>>>> +}
>>>> +
>>>>    igt_main
>>>>    {
>>>>    	struct drm_xe_engine_class_instance *hwe;
>>>> @@ -472,6 +549,17 @@ igt_main
>>>>    		}
>>>>    	}
>>>> +	igt_describe("Validate more power is consumed at higher frequencies");
>>>> +	igt_subtest("freq-power") {
>>>> +		/* FIXME: Remove skip once hwmon is added */
>>>> +		igt_skip_on(xe_has_vram(fd));
>>>> +		xe_for_each_gt(fd, gt) {
>>>> +			xe_for_each_hw_engine(fd, hwe) {
>>>> +				test_freq_power(fd, gt, hwe);
>>>> +			}
>>>> +		}
>>>> +	}
>>>> +
>>>>    	igt_fixture {
>>>>    		xe_for_each_gt(fd, gt) {
>>>>    			set_freq(fd, gt, "min", stash_min);
>>>> -- 
>>>> 2.40.0
>>>>