You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Description
Discovered in #6012 (comment), sometimes (compatible) bindgroups with same content are not deduplicated, therefore not detected as equal (I suspect the problem originates from layouts that come from getBindGroupLayout).
Why exactly BGL needs to work using deduplication, can't we just operate with duplicated bindgrops and check for compat by content, to me it seems like a big footgun.
I'm not sure that the underlying APIs allow that. I tried reducing potential misuse (forgetting to deduplicate) but it didn't work out since the other code-path that creates BGLs is quite different.
Description
Discovered in #6012 (comment), sometimes (compatible) bindgroups with same content are not deduplicated, therefore not detected as equal (I suspect the problem originates from layouts that come from
getBindGroupLayout
).Repro steps
Repro is here: https://github.com/sagudev/wgpu-problem/tree/another-empty-bind-panic-2. This is
webgpu:api,validation,encoding,programmable,pipeline_bind_group_compat:default_bind_group_layouts_never_match,render_pass:pipelineType="auto0";bindingType="auto0";swap=true;empty=false;renderCommand="draw"
but rewritten in rust (and to wgpu).Expected vs observed behavior
Per CTS there should actually be no error (bind groups are the same in content): https://github.com/gpuweb/cts/blob/198d1770062c1a8aba86e7d6e001bb47bea028ee/src/webgpu/api/validation/encoding/programmable/pipeline_bind_group_compat.spec.ts#L960
When this is done it should be safe to revert 60b42f9 (ping me for CTS run on servo)
The text was updated successfully, but these errors were encountered: