On 2026-08-26 at 11:41 +1000, John Hubbard <[email protected]> wrote... > On 8/25/26 6:27 PM, Alistair Popple wrote: > > On 2026-08-26 at 07:13 +1000, John Hubbard <[email protected]> wrote... > >> On 8/25/26 1:53 PM, Danilo Krummrich wrote: > >>> On Tue Aug 25, 2026 at 9:13 AM CEST, Alistair Popple wrote: > >> ... > > Sure. Two questions: > > > > 1) Do we also need to provide the decoded implementation value? > > User space needs to: > > a) Know enough about the GPU to be able to choose an SM version (and > probably a few other things) which in turn is eventually required in > order to run GPU programs, as shown in your merge request[1].
Right - currently that comes from either chip-id lookup table or direct decoding of the chip-id (although [1] doesn't currently look at it). If we're not going to allow either of those for architecture it doesn't make sense to need it for implementation either. IOW I think kernel should provide the decoded value for implementation too. > b) Display GPU information to the user. > > c) I'm sure to be overlooking something. > > > 2) If we provide the decoded values why do we need to expose the chip-id at > > all? > OK, so maybe we don't? Yeah, I agree. Doesn't make sense to if we're going to do the decoding in the kernel. - Alistair > [1] > https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/43352/diffs?commit_id=945a8e7c51939577526974b342c8dc6bafc3372f#3bb2fa8303917a2ad83e2239c2cae842dfaa594f > > thanks, > -- > John Hubbard >
