Bug 223767

Summary: [Metal ANGLE] fast/canvas/webgl/out-of-bounds-simulated-vertexAttrib0-drawArrays.html causes GPURestarts on some machines
Product: WebKit Reporter: Kyle Piddington <kpiddington>
Component: ANGLEAssignee: Nobody <webkit-unassigned>
Status: RESOLVED FIXED    
Severity: Normal CC: darin, dino, kkinnunen, webkit-bug-importer
Priority: P2 Keywords: InRadar
Version: WebKit Nightly Build   
Hardware: Unspecified   
OS: Unspecified   
Bug Depends on: 223926    
Bug Blocks: 220076    
Attachments:
Description Flags
Patch darin: review+

Kyle Piddington
Reported 2021-03-25 15:32:26 PDT
When running fast/canvas/webgl/out-of-bounds-simulated-vertexAttrib0-drawArrays.html, some Intel machines experience GPURestarts. These restarts can cause waterfall effects, making other tests fail in parallel, and sometimes blacklisting the test running process from executing GPU work. this results in a cascade of unrelated failures. The actual issue is unrelated to ANGLE, but rather the driver on the host systems. Submitting a large render workload (1 trillion+ points) causes the Intel GPUDriver to restart. To mitigate this issue, we can submit a smaller workload for this test, until the underlying driver bug is fixed.
Attachments
Patch (1.70 KB, patch)
2021-03-25 15:36 PDT, Kyle Piddington 		darin: review+
DetailsFormatted DiffDiff
View All Add attachment proposed patch, testcase, etc.
Kyle Piddington
Comment 1 2021-03-25 15:33:59 PDT
Note, these restart issues were only seen in automation on MacMini8'1, and iMac16'2
Kyle Piddington
Comment 2 2021-03-25 15:36:03 PDT
Created attachment 424287 [details] Patch
Dean Jackson
Comment 3 2021-03-25 20:20:52 PDT
Committed r275075 (235787@main): <https://commits.webkit.org/235787@main>
Radar WebKit Bug Importer
Comment 4 2021-03-25 20:21:15 PDT
<rdar://problem/75869407>
Kimmo Kinnunen
Comment 5 2021-03-25 23:51:04 PDT
I think the test originally tested the thing it tried to test? If a test tests 0x40000000 to test for a bug, it's not really ok to change it to 0x1, 0x40, or 0x400000, is it? It appears that the bug that the test tests for really manifests. Would it have been better to revert this change and skip the test instead? If a test fails in other place due to a bug in implementation, we typically don't change the test...
Note You need to log in before you can comment on or make changes to this bug.