Re: [PATCH i-g-t 2/2] tests/xe_pmu: Add frequency test

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

Hi Vinay

On 4/10/2025 7:03 AM, Vinay Belgaumkar wrote:
> Add a basic test that uses PMU to read GT actual and requested
> frequencies while running a workload.
> 
> v2: Rebase and comments (Riana)
> v3: Address review comments (Kamil and Riana)
> 
> Cc: Lucas De Marchi <lucas.demarchi@intel.com>
> Cc: Rodrigo Vivi <rodrigo.vivi@intel.com>
> Cc: Riana Tauro <riana.tauro@intel.com>
> Cc: Kamil Konieczny <kamil.konieczny@linux.intel.com>
> Signed-off-by: Vinay Belgaumkar <vinay.belgaumkar@intel.com>
> ---
>   tests/intel/xe_pmu.c | 147 +++++++++++++++++++++++++++++++++++++++++++
>   1 file changed, 147 insertions(+)
> 
> diff --git a/tests/intel/xe_pmu.c b/tests/intel/xe_pmu.c
> index 175bbf374..589c7cbde 100644
> --- a/tests/intel/xe_pmu.c
> +++ b/tests/intel/xe_pmu.c
> @@ -376,6 +376,94 @@ static void test_gt_c6_idle(int xe, unsigned int gt)
>   	close(pmu_fd);
>   }
>   
> +/**
> + * SUBTEST: gt-frequency
> + * Description: Validate we can collect accurate frequency PMU stats
> + *		while running a workload.
> + */
> +static void test_gt_frequency(int fd, struct drm_xe_engine_class_instance *eci)
> +{
> +	struct xe_cork *cork = NULL;
> +	uint64_t end[2], start[2];
> +	unsigned long config_rq_freq, config_act_freq;
> +	double min[2], max[2];
> +	uint32_t gt = eci->gt_id;
> +	uint32_t orig_min = xe_gt_get_freq(fd, eci->gt_id, "min");
> +	uint32_t orig_max = xe_gt_get_freq(fd, eci->gt_id, "max");
> +	uint32_t vm;
> +	int pmu_fd[2];
> +
> +	config_rq_freq = get_event_config(gt, NULL, "gt-requested-frequency");
> +	pmu_fd[0] = open_group(fd, config_rq_freq, -1);
> +
> +	config_act_freq = get_event_config(gt, NULL, "gt-actual-frequency");
> +	pmu_fd[1] = open_group(fd, config_act_freq, pmu_fd[0]);
> +
> +	vm = xe_vm_create(fd, 0, 0);
> +
> +	cork = xe_cork_create_opts(fd, eci, vm, 1, 1);
> +	xe_cork_sync_start(fd, cork);
> +
> +	/*
> +	 * Set GPU to min frequency and read PMU counters.
> +	 */
> +	igt_assert(xe_gt_set_freq(fd, gt, "max", orig_min) > 0);
> +	igt_assert(xe_gt_get_freq(fd, gt, "max") == orig_min);
> +
> +	pmu_read_multi(pmu_fd[0], 2, start);
> +	usleep(SLEEP_DURATION * USEC_PER_SEC);
> +	pmu_read_multi(pmu_fd[0], 2, end);
> +
> +	min[0] = (end[0] - start[0]);
> +	min[1] = (end[1] - start[1]);
> +
> +	/*
> +	 * Set GPU to max frequency and read PMU counters.
> +	 */
> +	igt_assert(xe_gt_set_freq(fd, gt, "max", orig_max) > 0);
> +	igt_assert(xe_gt_get_freq(fd, gt, "max") == orig_max);
> +	igt_assert(xe_gt_set_freq(fd, gt, "min", orig_max) > 0);
> +	igt_assert(xe_gt_get_freq(fd, gt, "min") == orig_max);
> +
> +	pmu_read_multi(pmu_fd[0], 2, start);
> +	usleep(SLEEP_DURATION * USEC_PER_SEC);
> +	pmu_read_multi(pmu_fd[0], 2, end);
> +
> +	max[0] = (end[0] - start[0]);
> +	max[1] = (end[1] - start[1]);
> +
> +	xe_cork_sync_end(fd, cork);
> +
> +	/*
> +	 * Restore min/max.
> +	 */
> +	igt_assert(xe_gt_set_freq(fd, gt, "min", orig_min) > 0);
> +	igt_assert(xe_gt_get_freq(fd, gt, "min") == orig_min);
> +
> +	igt_info("Minimum frequency: requested %.1f, actual %.1f\n",
> +		 min[0], min[1]);
> +	igt_info("Maximum frequency: requested %.1f, actual %.1f\n",
> +		 max[0], max[1]);
> +
> +	close(pmu_fd[0]);
> +	close(pmu_fd[1]);
> +
> +	if (cork)
> +		xe_cork_destroy(fd, cork);
> +
> +	xe_vm_destroy(fd, vm);
> +
> +	close(pmu_fd[0]);
> +	close(pmu_fd[1]);
> +
> +	assert_within_epsilon(min[0], orig_min, tolerance);
> +	/*
> +	 * On thermally throttled devices we cannot be sure maximum frequency
> +	 * can be reached so use larger tolerance downwards.
> +	 */
> +	assert_within_epsilon_up_down(max[0], orig_max, tolerance, 0.15f);
> +}
> +
>   static unsigned int enable_and_provision_vfs(int fd)
>   {
>   	unsigned int gt, num_vfs;
> @@ -427,6 +515,35 @@ static void disable_vfs(int fd)
>   		       "Failed to restore sriov_drivers_autoprobe value\n");
>   }
>   
> +static void stash_gt_freq(int fd, uint32_t **stash_min, uint32_t **stash_max)
> +{
> +	int num_gts, gt;
> +
> +	num_gts = xe_number_gt(fd);
> +
> +	*stash_min = (uint32_t *) malloc(sizeof(uint32_t) * num_gts);
> +	*stash_max = (uint32_t *) malloc(sizeof(uint32_t) * num_gts);
> +
> +	igt_skip_on(*stash_min == NULL || *stash_max == NULL);
> +
> +	xe_for_each_gt(fd, gt) {
> +		*stash_min[gt] = xe_gt_get_freq(fd, gt, "min");
> +		*stash_max[gt] = xe_gt_get_freq(fd, gt, "max");
> +	}
> +}
> +
> +static void restore_gt_freq(int fd, uint32_t *stash_min, uint32_t *stash_max)
> +{
> +	int gt;
> +
> +	xe_for_each_gt(fd, gt) {
> +		xe_gt_set_freq(fd, gt, "max", stash_max[gt]);
> +		xe_gt_set_freq(fd, gt, "min", stash_min[gt]);
> +	}
> +	free(stash_min);
> +	free(stash_max);
> +}
> +
>   igt_main
>   {
>   	int fd, gt;
> @@ -482,6 +599,36 @@ igt_main
>   			disable_vfs(fd);
>   	}
>   
> +	igt_subtest_group {
> +		bool has_freq0_node, needs_freq_restore = false;
> +		uint32_t *stash_min, *stash_max;
> +
> +		igt_fixture {
> +			has_freq0_node = xe_sysfs_gt_has_node(fd, 0, "freq0");
> +		}
> +
> +		igt_describe("Validate PMU GT freq measured is within the tolerance");
> +		igt_subtest_with_dynamic("gt-frequency") {
> +			igt_skip_on(!has_freq0_node);
> +			stash_gt_freq(fd, &stash_min, &stash_max);
> +			needs_freq_restore = true;
> +			xe_for_each_gt(fd, gt) {
Since it's one test, i thought it could be something like this, to avoid 
subtest group

	int orig_min = xe_gt_get_freq(fd, gt, "min");
	int orig_max = xe_gt_get_freq(fd, gt, "max");> +			 
igt_dynamic_f("gt%u", gt)
> +				xe_for_each_engine(fd, eci) {
> +					if (gt == eci->gt_id) {
> +						test_gt_frequency(fd, eci);
> +						break;
> +					}
> +				}
		xe_gt_set_freq(fd, gt, "max", orig_max);
		xe_gt_set_freq(fd, gt, "min", orig_min);

But if there are additional tests planned
Reviewed-by: Riana Tauro <riana.tauro@intel.com>>

> +			}
> +		}
> +
> +		igt_fixture {
> +			if (needs_freq_restore)
> +				restore_gt_freq(fd, stash_min, stash_max);
> +		}
> +	}
> +
>   	igt_fixture {
>   		close(fd);
>   	}