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
The maximum linear index with UInt32 is 4,294,967,295 so an array of about 4GB. With GPUs having upwards of 40GB or more memory in the data canter, it's not unlikely that a user want to process something larger than that.
@giordano the challenge is on the lower-level. We could have parameterized on the Context argument in KA 0.9, but in KA 0.10 we will have an interface that won't require a hidden context argument (see #635)
Activity
vchuravy commentedon Jun 28, 2023
@luraess also mentioned that it would make sense to configure the hardware dimension index into the Kernel struct.
simonbyrne commentedon Jul 16, 2023
Could you provide a function that would evaluate differently depending on the device? e.g.
brabreda commentedon Sep 1, 2023
In which case or device would int32 not be sufficient?
vchuravy commentedon Sep 1, 2023
The maximum linear index with
UInt32is 4,294,967,295 so an array of about 4GB. With GPUs having upwards of 40GB or more memory in the data canter, it's not unlikely that a user want to process something larger than that.In particular ML
giordano commentedon Nov 14, 2025
How about a scoped variable to choose the index size locally? 👀
vchuravy commentedon Nov 16, 2025
@giordano the challenge is on the lower-level. We could have parameterized on the Context argument in KA 0.9, but in KA 0.10 we will have an interface that won't require a hidden context argument (see #635)