Add a limit-querying API for GPUAdapter #1100
Conversation
|
Do we think there are going to be limits that are vectors of values? |
kainino0x
changed the title
Maybe, but it should be fine to change the signatures later if needed. Perhaps supportsLimit wouldn't be allowed for them. |
kainino0x
force-pushed
the
adapter-limits
branch
from
6a5eac7 to
efb6aa9
Compare
|
PTAL |
Kangz
left a comment
Kangz
left a comment
There was a problem hiding this comment.
It's interesting to design an API with the privacy budget in mind but I don't think there's state of the art. Maybe we should ask privacy budget folks to review this or an explainer of this change?
Extracted from gpuweb#1100 (and without the introduction of GPULimitName).
kainino0x
force-pushed
the
adapter-limits
branch
from
b2d0dd9 to
573f0c3
Compare
Extracted from #1100 (and without the introduction of GPULimitName).
This allows apps to request information about individual available adapter limits, which lets the UA track privacy budgeting with better granularity. See also gpuweb#489, gpuweb#495; gpuweb#1098.
kainino0x
force-pushed
the
adapter-limits
branch
from
e026466 to
09b8d89
Compare
|
The objection we had to #489 applies to On a more general note, I thought the group agreed to bucketing devices into similar-sized buckets for privacy. This kind of API should probably just tell the author which bucket they're in. |
There's the Metal feature set tables that could be used by implementations to map from Metal device families and feature set to concrete limits. This will be necessary to do even if the limits are exposed only on the
The UAs can bucket for sure, but I don't think we agreed that this group should define buckets (yet?) because the definition of the buckets will likely be difficult discussions based on the devices each participant cares about. |
Thanks for the info. It's been a long time and I wasn't sure whether the objection was to the inclusion of e.g. getLimit, or the lack of e.g. supportsLimit. I don't have any further stuff to say right now (other than what Corentin said above + the meeting discussion), but I do still think that getLimit is implementable and should be included. |
I think it's been determined that this API shouldn't consider privacy budgeting in it's design. - UAs can still lazily-decide on GPULimits values. - UAs can still do bucketing ahead of time, or even based on queries to GPULimits members. - Apps can attempt requestDevice without checking limits, if they just want to know if their requested limits are supported. Exact IDL interface is TBD, but editorial. Needs more text to be filled out later. Also adds some other TODOs. See also #489, #495; #1098, #1100.
5 participants
This allows apps to request information about individual available
adapter limits, which lets the UA track privacy budgeting with
better granularity.
See also #489, #495; #1098.
@litherum PTAL; does this address concerns with #489?
Preview | Diff