Enable reftests on Android for SW-WR
Categories
(Core :: Graphics: WebRender, task)
Tracking
()
Tracking | Status | |
---|---|---|
firefox90 | --- | fixed |
People
(Reporter: sotaro, Assigned: sotaro)
References
(Blocks 2 open bugs)
Details
Attachments
(3 files, 3 obsolete files)
We should be running SW-WR reftests on Android.
Assignee | ||
Updated•4 years ago
|
Assignee | ||
Updated•4 years ago
|
Assignee | ||
Comment 1•4 years ago
|
||
Assignee | ||
Comment 2•4 years ago
|
||
Just enabling sw-wr reftest caused crash in CompositorOGL::BeginFrame()
https://treeherder.mozilla.org/jobs?repo=try&revision=35f4becd1411f29ce6df944c17a05cc76d4d79c5
Assignee | ||
Comment 3•4 years ago
|
||
Updated•4 years ago
|
Comment 4•4 years ago
|
||
Comment on attachment 9215038 [details]
Bug 1704000 - RenderCompositorLayersSWGL::BeforeWrRender()
Revision D111583 was moved to bug 1704431. Setting attachment 9215038 [details] to obsolete.
Assignee | ||
Comment 5•4 years ago
|
||
With current patch, test failures were basically addressed. But some apz tests results seem to have problems. Black rectangle exists :(
https://treeherder.mozilla.org/jobs?repo=try&revision=671e4d5541b065bd7248b73d19d433760c3a312a
Assignee | ||
Comment 6•4 years ago
|
||
Reftests still have some failures :(
With PBO
https://treeherder.mozilla.org/jobs?repo=try&revision=b8d0a1b6feb80513e91fd0fa56f39dc2f816f129
Without PBO
https://treeherder.mozilla.org/jobs?repo=try&revision=37c7b3d0a7e657b57f13e9882d25c117e10cc325
Assignee | ||
Comment 7•4 years ago
|
||
On release build, frequency of failures became less frequent.
https://treeherder.mozilla.org/jobs?repo=try&revision=bb6247a1f2f2b02cb1b0c4be7ca31ea5d69623da
Assignee | ||
Comment 8•4 years ago
|
||
zoom performance is not good(bug 1703709). It might be related.
Assignee | ||
Updated•4 years ago
|
Assignee | ||
Comment 9•4 years ago
|
||
(In reply to Sotaro Ikeda [:sotaro] from comment #6)
Reftests still have some failures :(
With PBO
https://treeherder.mozilla.org/jobs?repo=try&revision=b8d0a1b6feb80513e91fd0fa56f39dc2f816f129Without PBO
https://treeherder.mozilla.org/jobs?repo=try&revision=37c7b3d0a7e657b57f13e9882d25c117e10cc325
It seems that, azp did not work as expected with sw-wr. But sw-wr does not directly interact with apz.
:hiro, do you have any ideas about what could cause the problem?
Comment 10•4 years ago
|
||
Given that failures happen on tests specifying reftest-async-scroll and those failures are not perma, it's high frequent intermittent, it looks like with SW-WR we take a snapshot before applying the async scroll offset even though it's a sync IPC call, the call ends up deferring the task to the updater thread, that's the problem?
Comment 11•4 years ago
|
||
(In reply to Hiroyuki Ikezoe (:hiro) from comment #10)
Given that failures happen on tests specifying reftest-async-scroll and those failures are not perma, it's high frequent intermittent, it looks like with SW-WR we take a snapshot before applying the async scroll offset even though it's a sync IPC call, the call ends up deferring the task to the updater thread, that's the problem?
And moreover, even if applying the offset value has finished on the updater thread at the moment we take a snapshot, the offset needs to be sampled on the sampler thread, then it needs to be rendered on the renderer thread? I am pretty sure about the sampler thread, but not 100% about the renderer thread.
Assignee | ||
Comment 12•4 years ago
|
||
When RenderCompositorOGLSWGL always request full render, the failures did not happen. The problem seems to exist around RenderCompositorOGLSWGL.
https://treeherder.mozilla.org/jobs?repo=try&revision=65962e21ec8f5b2c160f57c72b3b716e82ad4dc9
Assignee | ||
Comment 13•4 years ago
|
||
When the problem happened, RenderCompositorLayersSWGL::StartCompositing() was returned early by aNumDirtyRects==0.
https://treeherder.mozilla.org/jobs?repo=try&revision=28e04ed3b5108980aeeee61482976a2375f1917e
Assignee | ||
Comment 14•4 years ago
|
||
Just apz scroll update might not provide dirty rect.
Assignee | ||
Updated•4 years ago
|
Assignee | ||
Comment 15•4 years ago
|
||
Updated•4 years ago
|
Assignee | ||
Comment 16•4 years ago
|
||
APZ test failure was related to skip rendering during taking snapshot.
RendererOGL::UpdateAndRender() does readback with glReadPixels(). If RenderCompositorOGLSWGL skips rendering at RenderCompositorLayersSWGL::StartCompositing(), frambuffer does not have valid rendering. readback image was old data that existed in android BufferQueue.
When rendering was forced during readback, reftest failures were addressed.
https://treeherder.mozilla.org/jobs?repo=try&revision=2abb60075d8af54b17b021e7db42ed08652feeb9
Assignee | ||
Comment 17•4 years ago
|
||
Updated•4 years ago
|
Assignee | ||
Comment 18•4 years ago
|
||
Assignee | ||
Comment 19•4 years ago
|
||
Updated•4 years ago
|
Comment 20•4 years ago
|
||
Comment 21•4 years ago
|
||
Comment 22•4 years ago
|
||
Comment 23•4 years ago
|
||
Backed out changeset 8e6ae945c88a (bug 1704000) for Reftest failures on Android. CLOSED TREE
Log:
https://treeherder.mozilla.org/logviewer?job_id=340662468&repo=autoland&lineNumber=5695
Push with failures:
https://treeherder.mozilla.org/jobs?repo=autoland&group_state=expanded&revision=8e6ae945c88a59bdf07606465fc7faed71def91c
Backout:
https://hg.mozilla.org/integration/autoland/rev/92eabce732fb
Assignee | ||
Comment 24•4 years ago
|
||
Sorry, I forgot to land dependent patches.
Assignee | ||
Updated•4 years ago
|
Comment 25•4 years ago
|
||
bugherder |
Assignee | ||
Comment 26•4 years ago
|
||
Comment 27•4 years ago
|
||
Assignee | ||
Updated•4 years ago
|
Comment 28•4 years ago
|
||
bugherder |
Description
•