Archives Discussions

[Ceq]

I couldn't find anything about this in documentation:

When accessing gather streams inside a kernel, if you read out of stream bounds the currently behaviour is to read the nearest boundary element, for example, for gstream[8][8]:
gstream[(-1, -1)] is the same as gstream[(0, 0)]
gstream[(1, 10)] is the same as gstream[(1, 7)]

I've tested using CPU backend only, is the same for GPU?
Is this the standard behavior? Should we avoid it?
Can you rely it won't change in future versions?


Also, if you access gstream[(0.75, 0.75)] it is rounded to gstream[(1, 1)]
However, accessing gstream[(0.74, 0.74)] is rounded to gstream[(0, 0)]
This kind of rounding has surprised me a bit... how is it performed?


And last question: I've seen a few articles about Radeon 4870 launching next month and specs look quite promising, would it be supported too?

Thanks

[michael_chu]

Hi Ceq,

I would not recommend relying on the out of stream behaviour remaining the same in future.

Do you have an application use case that needs to rely on out of stream behaviour working properly?

Also, when integer support is available, using floats for indices should no longer be necessary and eliminate a lot of the confusion there.

The next gen Radeon will be supported but not in v1.1-beta. Earliest time to expect support will be in v1.2-beta.

Michael.