GL timestamps queried via ARB_timer_query should have nanosecond units but I had a suspicion we wouldn't have correct units for Gen graphics on Skylake and so extended arb_timer_query/timestamp-get.c to query the gpu timestamp either side of a sleep on the cpu with a known duration. This was able to confirm my suspicion for Skylake.
I was a little surprised initially to also find that the clock drift for my Haswell machine was greater than I would have guessed, but still it's probably totally reasonable for typical GL_TIMESTAMP uses. I was seeing a drift of ~2 milliseconds in one second. This test allows a drift of up to 3 milliseconds / sec, which is in line with the kind of tolerance the test was already looking for, just for back-to-back GL_TIMESTAMP queries. Robert Bragg (2): util: Add piglit_delay_ns() api arb_timer_query: Add wall clock sanity checks tests/spec/arb_timer_query/timestamp-get.c | 51 +++++++++++++++++++++++++++++- tests/util/piglit-util.c | 29 ++++++++++++++--- tests/util/piglit-util.h | 15 +++++++++ 3 files changed, 90 insertions(+), 5 deletions(-) -- 2.9.2 _______________________________________________ Piglit mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/piglit
