Alexander Edwards
Alexander Edwards
This branch is very much a WIP currently, but for anyone interested here is roughly how I plan to structure things, aiming to roughly mesh the shape of the original...
> @ArthurZucker @ae99 let me know if you need help with anything - think this is a super cool addition! Hey @patrickvonplaten and @ArthurZucker! I think the general structure of...
> Hey @ae99 , are you still working on the integration? If not then let me know, I would be happy to continue from where you left. Hey @akashe, that'd...
@nimlgen with this change, the primary generated kernel uses global_dim=128 and localdim=32, for example see one of the generated kernels below: ```c++ kernel void r_128_32_1024_4(device float* data0, const device float*...
@geohot I've updated the PR to be a bit cleaner. I've checked and it appears to me that existing `test/test_ops.py` test cases cover off correctness of this optimization quite well....
> this kernel is still not very efficient @nimlgen agreed! I've been working on a new approach since then, it's able to get all the way to `0.95x` faster than...
@geohot or @wozeparrot, I believe this is in another ready to review state when you get the time!
Ah apologies for delay, time zones. I'll look into why this broke things today.
When using FFMPEG backend (and passing a URL or FileLike) and interacting with the underlying StreamingMediaDecoder, FFMPEG only ends up reading first `n` bytes to get `num_frames`. As far I...