I could be interested if the new implementation(s) to be faster for either afcuda or afcpu. Would it be ?

This would be faster on afcuda. I haven't looked into afcpu's implementation. This will require you to modify the CUDA kernel implementation.

I don't think there will be a big improvement on the CPU backend.

ok, a speed up for afcuda would be good enough, please add here more details and I will see if I could do it.

The join function seems to be calling the join kernel for each buffer that needs to be joined. It would be better if we performed this operation in one kernel call instead of multiple calls. Here are the steps you can take to do this:

1. Allocate a buffer to store the pointers to buffers that need to be joined.
2. Allocate a buffer for the offsets. You could do this with the previous buffer but a separate buffer may be easier.
3. Modify the kernel so that you loop over each of the arrays. This may be a simple modification based on what I saw. You probably just need to add another loop to iterate over each of the buffers.
4. Modify the host code to call the new function. Remove the join wrapper as its now unnecessary.

Is this issue still open? If it is, I would like to work on it