drm/i915/selftests: Repeat the rps clock frequency measurement

Repeat the measurement of the clock frequency a few times and use the
median to try and reduce the systematic measurement error.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20200504044903.7626-6-chris@chris-wilson.co.uk

drivers/gpu/drm/i915/gt/selftest_rps.c

@@ -56,6 +56,18 @@ static int cmp_u64(const void *A, const void *B)
 		return 0;
 }
 
+static int cmp_u32(const void *A, const void *B)
+{
+	const u32 *a = A, *b = B;
+
+	if (a < b)
+		return -1;
+	else if (a > b)
+		return 1;
+	else
+		return 0;
+}
+
 static struct i915_vma *
 create_spin_counter(struct intel_engine_cs *engine,
 		    struct i915_address_space *vm,
@@ -236,8 +248,8 @@ int live_rps_clock_interval(void *arg)
 	for_each_engine(engine, gt, id) {
 		unsigned long saved_heartbeat;
 		struct i915_request *rq;
-		ktime_t dt;
 		u32 cycles;
+		u64 dt;
 
 		if (!intel_engine_can_store_dword(engine))
 			continue;
@@ -286,15 +298,29 @@ int live_rps_clock_interval(void *arg)
 				  engine->name);
 			err = -ENODEV;
 		} else {
-			preempt_disable();
-			dt = ktime_get();
-			cycles = -intel_uncore_read_fw(gt->uncore,
-						       GEN6_RP_CUR_UP_EI);
-			udelay(1000);
-			dt = ktime_sub(ktime_get(), dt);
-			cycles += intel_uncore_read_fw(gt->uncore,
-						       GEN6_RP_CUR_UP_EI);
-			preempt_enable();
+			ktime_t dt_[5];
+			u32 cycles_[5];
+			int i;
+
+			for (i = 0; i < 5; i++) {
+				preempt_disable();
+
+				dt_[i] = ktime_get();
+				cycles_[i] = -intel_uncore_read_fw(gt->uncore, GEN6_RP_CUR_UP_EI);
+
+				udelay(1000);
+
+				dt_[i] = ktime_sub(ktime_get(), dt_[i]);
+				cycles_[i] += intel_uncore_read_fw(gt->uncore, GEN6_RP_CUR_UP_EI);
+
+				preempt_enable();
+			}
+
+			/* Use the median of both cycle/dt; close enough */
+			sort(cycles_, 5, sizeof(*cycles_), cmp_u32, NULL);
+			cycles = (cycles_[1] + 2 * cycles_[2] + cycles_[3]) / 4;
+			sort(dt_, 5, sizeof(*dt_), cmp_u64, NULL);
+			dt = div_u64(dt_[1] + 2 * dt_[2] + dt_[3], 4);
 		}
 
 		intel_uncore_write_fw(gt->uncore, GEN6_RP_CONTROL, 0);
@@ -306,14 +332,14 @@ int live_rps_clock_interval(void *arg)
 		if (err == 0) {
 			u64 time = intel_gt_pm_interval_to_ns(gt, cycles);
 			u32 expected =
-				intel_gt_ns_to_pm_interval(gt, ktime_to_ns(dt));
+				intel_gt_ns_to_pm_interval(gt, dt);
 
 			pr_info("%s: rps counted %d C0 cycles [%lldns] in %lldns [%d cycles], using GT clock frequency of %uKHz\n",
-				engine->name, cycles, time, ktime_to_ns(dt), expected,
+				engine->name, cycles, time, dt, expected,
 				gt->clock_frequency / 1000);
 
-			if (10 * time < 8 * ktime_to_ns(dt) ||
-			    8 * time > 10 * ktime_to_ns(dt)) {
+			if (10 * time < 8 * dt ||
+			    8 * time > 10 * dt) {
 				pr_err("%s: rps clock time does not match walltime!\n",
 				       engine->name);
 				err = -EINVAL;