video_VDAPerf: "delivery_time" regression between 9878.0.0 and 9879.0.0 on eve |
|||||||||||
Issue description"delivery_time" metrics got worse between 62.9831.0.0 and 62.9901.21.0 on eve. The same regression happened on other x86 boards (e.g. poppy). https://chromeperf.appspot.com/report?sid=77971c40a22811935dedea6d36eaef568f33cc8a1ad9601d020752eec7c5f5e8 As far as I bisected, this regression has happened since ChromeOS 62.9879.0.0. It is the ChromeOS version libva version update was landed. crosreview.com/557361 I run ChromeOS autotest 62.9878.0.0 on ChromeOS 62.9879.0.0 with old version libva. The regression didn't happen. It indicates this regression is caused by libva version update. According to the graph, the regression was affected especially for VP8 and VP9. H264 was also affected, but the effect was smaller than VP8 and VP9. This is probably because Libva developers tuned and measured the performance with H264. The attached files are chart-flies.json I collected. * ChromeOS9878_libva39_VDAPerf_vp9_result-chart.json The result of video_VDAPerf.vp9 (ChromeOS 9878) on ChromeOS 9878. No regression happened. * ChromeOS9879_libva39_VDAPerf_vp9_result-chart.json The result of video_VDAPerf.vp9 (ChromeOS 9878) on ChromeOS 9879, where libva 0.39 was used by reverting libva version update CL. No regression happened. * ChromeOS9879_libva40_VDAPerf_vp9_result-chart.json The result of video_VDAPerf.vp9 (ChromeOS 9879) on ChromeOS 9879, where libva 0.40 was used. The regression happened.
,
Oct 5 2017
,
Oct 7 2017
,
Oct 10 2017
This regression has occurred from libva 1.7.3. I continue to bisect commits between libva 1.7.2 and libva 1.7.3.
,
Oct 10 2017
The bad commit causing the regression is the following commit in libva-intel-driver. https://github.com/01org/intel-vaapi-driver/commit/78b6841fa3e247f15005ea3450d8404c59938ac3#diff-34f570ee8d123f1a978d80988ebf447c
,
Oct 10 2017
I confirm the regression is resolved by just reverting the commit in libva 1.7.3.
,
Oct 11 2017
The following revision refers to this bug: https://chromium.googlesource.com/chromiumos/overlays/chromiumos-overlay/+/990c41eeb413d15b2a46b780a4abb6d4f20d3cd6 commit 990c41eeb413d15b2a46b780a4abb6d4f20d3cd6 Author: Daniel Charles <daniel.charles@intel.com> Date: Wed Oct 11 23:11:11 2017 intel-vaapi-driver: do not use media read with vpp Revert of upstream patch that uses media read instead of AVS for data reading at vpp. The upstream patch seek to fix a problem shown with jpeg encoding and high resolutions typically > 4K. This fixes a problem seen where decoding+vpp as measured by ChromeOS tools was higher than before. Using AVS again will not affect decoding+vpp time. BUG=b:67626373, chromium:771860 TEST="test_that video_VDAPerf, check the output json files for less delivery_time" Change-Id: I704e49a5dd8d16f52a722dd3e513dfcbc3459430 Signed-off-by: Daniel Charles <daniel.charles@intel.com> Reviewed-on: https://chromium-review.googlesource.com/710734 Commit-Ready: Hirokazu Honda <hiroh@chromium.org> Reviewed-by: Hirokazu Honda <hiroh@chromium.org> Reviewed-by: Alexandre Courbot <acourbot@chromium.org> [modify] https://crrev.com/990c41eeb413d15b2a46b780a4abb6d4f20d3cd6/x11-libs/libva-intel-driver/libva-intel-driver-1.8.3.ebuild [rename] https://crrev.com/990c41eeb413d15b2a46b780a4abb6d4f20d3cd6/x11-libs/libva-intel-driver/libva-intel-driver-1.8.3-r3.ebuild [add] https://crrev.com/990c41eeb413d15b2a46b780a4abb6d4f20d3cd6/x11-libs/libva-intel-driver/files/Revert-Use-Media-Read-message-if-possible-on-Gen8.patch
,
Oct 16 2017
Looks like the regressions has been fixed since the above CL was merged. Enough data are not shown yet though. Let us close this issue, and reopen if it is not really fixed. https://chromeperf.appspot.com/report?sid=08378ca2ee7f799071ffd123aefe13f3c5910d1bc2133632acf470a19440a2f3 |
|||||||||||
►
Sign in to add a comment |
|||||||||||
Comment 1 Deleted