On 17/12/2020 17:39, Chris Wilson wrote:
Quoting Tvrtko Ursulin (2020-12-17 17:33:05)
On 17/12/2020 17:26, Chris Wilson wrote:
Measure the sample gt-awake time while each engine and every engine is
busy. They should all report the same duration, the elapsed runtime of
the batch.
Signed-off-by: Chris Wilson <chris@xxxxxxxxxxxxxxxxxx>
Cc: Tvrtko Ursulin <tvrtko.ursulin@xxxxxxxxxxxxxxx>
---
include/drm-uapi/i915_drm.h | 3 ++-
tests/i915/perf_pmu.c | 42 +++++++++++++++++++++++++++++++++++++
2 files changed, 44 insertions(+), 1 deletion(-)
diff --git a/include/drm-uapi/i915_drm.h b/include/drm-uapi/i915_drm.h
index ef696d1a5..bf9ea471c 100644
--- a/include/drm-uapi/i915_drm.h
+++ b/include/drm-uapi/i915_drm.h
@@ -177,8 +177,9 @@ enum drm_i915_pmu_engine_sample {
#define I915_PMU_REQUESTED_FREQUENCY __I915_PMU_OTHER(1)
#define I915_PMU_INTERRUPTS __I915_PMU_OTHER(2)
#define I915_PMU_RC6_RESIDENCY __I915_PMU_OTHER(3)
+#define I915_PMU_SOFTWARE_GT_AWAKE_TIME __I915_PMU_OTHER(4)
-#define I915_PMU_LAST I915_PMU_RC6_RESIDENCY
+#define I915_PMU_LAST /* Deprecated - do not use */ I915_PMU_RC6_RESIDENCY
/* Each region is a minimum of 16k, and there are at most 255 of them.
*/
diff --git a/tests/i915/perf_pmu.c b/tests/i915/perf_pmu.c
index c2b186d39..90182074b 100644
--- a/tests/i915/perf_pmu.c
+++ b/tests/i915/perf_pmu.c
@@ -883,6 +883,42 @@ sema_busy(int gem_fd,
close(fd[1]);
}
+static void test_awake(int i915)
+{
+ const struct intel_execution_engine2 *e;
+ unsigned long slept;
+ uint64_t val;
+ int fd;
+
+ fd = perf_i915_open(i915, I915_PMU_SOFTWARE_GT_AWAKE_TIME);
+ igt_skip_on(fd < 0);
+
+ /* Check that each engine is captured by the GT wakeref */
+ __for_each_physical_engine(i915, e) {
+ igt_spin_new(i915, .engine = e->flags);
+
+ val = pmu_read_single(fd);
+ slept = measured_usleep(batch_duration_ns / 1000);
+ val = pmu_read_single(fd) - val;
+
+ gem_quiescent_gpu(i915);
+ assert_within_epsilon(val, slept, tolerance);
+ }
+
+ /* And that the total GT wakeref matches walltime not summation */
+ __for_each_physical_engine(i915, e)
+ igt_spin_new(i915, .engine = e->flags);
+
+ val = pmu_read_single(fd);
+ slept = measured_usleep(batch_duration_ns / 1000);
+ val = pmu_read_single(fd) - val;
+
+ gem_quiescent_gpu(i915);
+ assert_within_epsilon(val, slept, tolerance);
+
Erm who terminates all those spinners? :)
gem_quiescent_gpu() calls igt_terminate_spins(). Hmm, I actually thought
it freed them too, but what's a small leak...
Okay, I'm fine with that.
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@xxxxxxxxx>
Regards,
Tvrtko
_______________________________________________
Intel-gfx mailing list
Intel-gfx@xxxxxxxxxxxxxxxxxxxxx
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
