Fix convolve3 launch configuration in CUDA backend #2519
Conversation
|
Would it be possible to update the convolve3 documentation to describe the expected batching behavior? |
|
@syurkevi We are already explain the batching behavior for all convolve functions in our documentation. This change is an internal fix for CUDA kernel launch configuration and doesn't change the external batching behavior. |
The documentation you linked is copy-pasted from convolve2 and doesn't apply to convolve3. It would be nice to update the correct batching behavior to the documentation as it is difficult to understand what it should be without reading the source code. |
Batched input of Convolve3 was incorrectly folding batch size into CUDA launch grid's y & z dimensions. Since the batch for 3d inputs can happen along 4th dimension only, folding it onto x dimesion of CUDA grid is sufficient.
|
|
|
@arrayfire/core-devel Alright, here's the new updated documentation for convolution. have a look and share feedback pls. |
Batched input of Convolve3 was incorrectly folding batch size
into CUDA launch grid's y & z dimensions. Since the batch for 3d
inputs can happen along 4th dimension only, folding it onto x dimesion
of CUDA grid is sufficient.
Have to merge arrayfire/arrayfire-data#14 before this is merged.