[PATCH i-g-t 3/6] tests/intel/xe_oa: Sanity check PEC report data for TestOa metric set

[Umesh Nerlige Ramappa]

On Tue, Apr 08, 2025 at 11:12:07AM -0700, Ashutosh Dixit wrote:
>Implement sanity checking for Xe2 PEC OA reports. Previously there was
>sanity checking only for Xe1 OA reports, but no sanity checking for Xe2 PEC
>OA reports.
>
>Signed-off-by: Ashutosh Dixit <ashutosh.dixit at intel.com>
>---
> tests/intel/xe_oa.c | 121 +++++++++++++++++++++++++++++++++++++++++++-
> 1 file changed, 119 insertions(+), 2 deletions(-)
>
>diff --git a/tests/intel/xe_oa.c b/tests/intel/xe_oa.c
>index 2440101900..eaf97ae0df 100644
>--- a/tests/intel/xe_oa.c
>+++ b/tests/intel/xe_oa.c
>@@ -966,6 +966,115 @@ accumulator_print(struct accumulator *accumulator, const char *title)
> 		igt_debug("\tC%u = %"PRIu64"\n", i, deltas[idx++]);
> }
>
>+
>+/*
>+ * pec_sanity_check_reports() uses the following properties of the TestOa
>+ * metric set with the "576B_PEC64LL" or XE_OA_FORMAT_PEC64u64 format. See
>+ * e.g. lib/xe/oa-configs/oa-lnl.xml.
>+ *
>+ * If pec[] is the array of pec qwords following the report header (Bspec
>+ * 60942) then we have:
>+ *
>+ *	pec[2]  : test_event1_cycles
>+ *	pec[3]  : test_event1_cycles_xecore0
>+ *	pec[4]  : test_event1_cycles_xecore1
>+ *	pec[5]  : test_event1_cycles_xecore2
>+ *	pec[6]  : test_event1_cycles_xecore3
>+ *	pec[21] : test_event1_cycles_xecore4
>+ *	pec[22] : test_event1_cycles_xecore5
>+ *	pec[23] : test_event1_cycles_xecore6
>+ *	pec[24] : test_event1_cycles_xecore7
>+ *
>+ * test_event1_cycles_xecore* increment with every clock, so they increment
>+ * the same as gpu_ticks in report headers in successive reports. And
>+ * test_event1_cycles increment by 'gpu_ticks * num_xecores'.
>+ *
>+ * These equations are not exact due to fluctuations, but are precise when
>+ * averaged over long periods.
>+ */
>+static void pec_sanity_check_one(const u32 *report)
>+{
>+	int xecore_idx[] = {3, 4, 5, 6, 21, 22, 23, 24};
>+	u64 first, *pec = (u64 *)(report + 8);
>+
>+	igt_debug("\ttest_event1_cycles: %#lx\n", pec[2]);
>+	for (int i = 0; i < ARRAY_SIZE(xecore_idx); i++)
>+		igt_debug("\ttest_event1_cycles_xecore %d: %#lx\n", i, pec[xecore_idx[i]]);
>+
>+	/* Compare against the first non-zero test_event1_cycles_xecore* */
>+	for (int i = 0; i < ARRAY_SIZE(xecore_idx); i++) {
>+		first = pec[xecore_idx[i]];
>+		if (first)
>+			break;
>+	}
>+
>+	/* test_event1_cycles_xecore* should be within an epsilon of each other */
>+	for (int i = 0; i < ARRAY_SIZE(xecore_idx); i++) {
>+		igt_debug("n %d: pec[n] %#lx, first %#lx\n",
>+			  xecore_idx[i], pec[xecore_idx[i]], first);
>+		/* 0 value for pec[xecore_idx[i]] indicates missing xecore */
>+		if (pec[xecore_idx[i]])
>+			assert_within_epsilon(pec[xecore_idx[i]], first, 
>0.1);
>+	}
>+
>+	igt_debug("first * num_xecores: %#lx, pec[2] %#lx\n",
>+		  first * intel_xe_perf->devinfo.n_eu_sub_slices, pec[2]);
>+	/* test_event1_cycles should be close to (test_event1_cycles_xecore* * num_xecores) */
>+	assert_within_epsilon(first * intel_xe_perf->devinfo.n_eu_sub_slices, pec[2], 0.1);
>+}
>+
>+static void pec_sanity_check_two(const u32 *report0, const u32 *report1,
>+				 struct intel_xe_perf_metric_set *set)

I would just s/pec_sanity_check_two/pec_sanity_check/ to validate 2 
reports and drop the "pec_sanity_check_one" altogether. We only care 
about delta between 2 counters.

>+{
>+	u64 tick_delta = oa_tick_delta(report1, report0, set->perf_oa_format);
>+	int xecore_idx[] = {3, 4, 5, 6, 21, 22, 23, 24};
>+	u64 *pec0 = (u64 *)(report0 + 8);
>+	u64 *pec1 = (u64 *)(report1 + 8);
>+
>+	igt_debug("tick delta = %#lx\n", tick_delta);
>+
>+	/* Difference in test_event1_cycles_xecore* values should be close to tick_delta */
>+	for (int i = 0; i < ARRAY_SIZE(xecore_idx); i++) {

Maybe, within the loop you can have,

n = xecore_idx[i];

and that can be used in the below code, for ex:

		igt_debug("pec1[%d] - pec0[%d] %#lx, tick delta %#lx\n", n, pec1[n] - pec0[n], tick_delta);

>+		igt_debug("n %d: pec1[n] - pec0[n] %#lx, tick delta %#lx\n",
>+			  xecore_idx[i], pec1[xecore_idx[i]] - pec0[xecore_idx[i]], tick_delta);


>+		/* 0 value for pec[xecore_idx[i]] indicates missing xecore */
>+		if (pec1[xecore_idx[i]] && pec0[xecore_idx[i]])
>+			assert_within_epsilon(pec1[xecore_idx[i]] - pec0[xecore_idx[i]],
>+					      tick_delta, 0.1);
>+		/* Same test_event1_cycles_xecore* should be present in all reports */
>+		if (pec1[xecore_idx[i]])
>+			igt_assert(pec0[xecore_idx[i]]);
>+	}
>+
>+	igt_debug("pec1[2] - pec0[2] %#lx, tick_delta * num_xecores: %#lx\n",
>+		  pec1[2] - pec0[2], tick_delta * intel_xe_perf->devinfo.n_eu_sub_slices);
>+	/* Difference in test_event1_cycles should be close to (tick_delta * num_xecores) */
>+	assert_within_epsilon(pec1[2] - pec0[2],
>+			      tick_delta * intel_xe_perf->devinfo.n_eu_sub_slices, 0.1);
>+}

Thanks,
Umesh