Bug 289147 – [WebGPU] Memory barriers in vertex shader validation routines are problematic on M1

• Status: RESOLVED
• Resolution: FIXED
• Component: WebGPU

[Mike Wyrzykowski]

[WebGPU] Visual difference on some sites between M1 and M2 Macs

[Radar WebKit Bug Importer]

<rdar://problem/146219177>

[Mike Wyrzykowski]

M1s are triggering a context lost due to MTLCommandBuffer failure:

error	23:00:40.699875-0800	com.apple.WebKit.GPU	-[_MTLCommandBuffer didScheduleWithStartTime:endTime:error:], line 1020: error 'Execution of the command buffer was aborted due to an error during execution. Internal Error (00000100:Internal Error)'
error	23:00:40.699908-0800	com.apple.WebKit.GPU	Execution of the command buffer was aborted due to an error during execution. Internal Error (00000100:Internal Error)
default	23:00:40.719243-0800	com.apple.WebKit.GPU	Encountered fatal command buffer error Error Domain=MTLCommandBufferErrorDomain Code=1 "Internal Error (00000100:Internal Error)" UserInfo={NSLocalizedDescription=Internal Error (00000100:Internal Error), NSUnderlyingError=0x7067adcb0 {Error Domain=IOGPUCommandQueueErrorDomain Code=256 "(null)"}}, underlying error Error Domain=IOGPUCommandQueueErrorDomain Code=256 "(null)"

[Mike Wyrzykowski]

If we ignore this error on M1s the page works as expected, need to figure out what error code 256 is
```
Encountered non-fatal command buffer error Error Domain=MTLCommandBufferErrorDomain Code=1 "Internal Error (00000100:Internal Error)" UserInfo={NSLocalizedDescription=Internal Error (00000100:Internal Error), NSUnderlyingError=0x42dd31fb0 {Error Domain=IOGPUCommandQueueErrorDomain Code=256 "(null)"}}, underlying error Error Domain=IOGPUCommandQueueErrorDomain Code=256 "(null)"
Encountered non-fatal command buffer error Error Domain=MTLCommandBufferErrorDomain Code=1 "Internal Error (00000100:Internal Error)" UserInfo={NSLocalizedDescription=Internal Error (00000100:Internal Error), NSUnderlyingError=0x43510cb00 {Error Domain=IOGPUCommandQueueErrorDomain Code=256 "(null)"}}, underlying error Error Domain=IOGPUCommandQueueErrorDomain Code=256 "(null)"
Encountered non-fatal command buffer error Error Domain=MTLCommandBufferErrorDomain Code=1 "Internal Error (00000100:Internal Error)" UserInfo={NSLocalizedDescription=Internal Error (00000100:Internal Error), NSUnderlyingError=0x42ddf50e0 {Error Domain=IOGPUCommandQueueErrorDomain Code=256 "(null)"}}, underlying error Error Domain=IOGPUCommandQueueErrorDomain Code=256 "(null)"
Encountered non-fatal command buffer error Error Domain=MTLCommandBufferErrorDomain Code=1 "Internal Error (00000100:Internal Error)" UserInfo={NSLocalizedDescription=Internal Error (00000100:Internal Error), NSUnderlyingError=0x142c9c810 {Error Domain=IOGPUCommandQueueErrorDomain Code=256 "(null)"}}, underlying error Error Domain=IOGPUCommandQueueErrorDomain Code=256 "(null)"
Encountered non-fatal command buffer error Error Domain=MTLCommandBufferErrorDomain Code=1 "Internal Error (00000100:Internal Error)" UserInfo={NSLocalizedDescription=Internal Error (00000100:Internal Error), NSUnderlyingError=0x436b94e40 {Error Domain=IOGPUCommandQueueErrorDomain Code=256 "(null)"}}, underlying error Error Domain=IOGPUCommandQueueErrorDomain Code=256 "(null)"
Encountered non-fatal command buffer error Error Domain=MTLCommandBufferErrorDomain Code=1 "Internal Error (00000100:Internal Error)" UserInfo={NSLocalizedDescription=Internal Error (00000100:Internal Error), NSUnderlyingError=0x142ca24f0 {Error Domain=IOGPUCommandQueueErrorDomain Code=256 "(null)"}}, underlying error Error Domain=IOGPUCommandQueueErrorDomain Code=256 "(null)"
com.apple.WebKit.GPU.Development(40094) MallocStackLogging: msl_handle_memory_pressure_event: approaching memory limit. Starting stack-logging.
```

[Mike Wyrzykowski]

Ok great I root caused it. Too many completion handlers are being added to the MTLCommandBuffer, which appears to be exhausting some driver / kernel resource.

WebKit can manage this ourselves quite easily, I will change that

[Mike Wyrzykowski]

Further analysis indicates this is not due to the total number of completion handlers, but rather the way we are using memory barriers does not seem to work on M1 Macs.

We could split passes on M1s as this is not a codepath which should not be regularly triggered

[Mike Wyrzykowski]

Splitting render passes to validate like we do for Intel appears to resolve the issue.

[Mike Wyrzykowski]

Splitting the render passes is not needed, there is a simpler solution, I will implement that instead

[Mike Wyrzykowski]

Pull request: https://github.com/WebKit/WebKit/pull/42069

[EWS]

Committed 291897@main (8f1919db1fce): <https://commits.webkit.org/291897@main>

Reviewed commits have been landed. Closing PR #42069 and removing active labels.