CUDA arrays not working well with broadcast!(), and other in-place operations inside a loop

[maleadt]

@ayushinav

As an end user, I am more concerned about using GPUs to reduce the computation time without going into the depths of it. Flux.jl had made my life easy because not much was required to do to shift from CPUs to GPUs, but it has come to haunt me now.

I’m not sure what’s the problem that’s haunting you here?

@ayushinav

While we are at it, could you please also explain why the first iteration runs fast (and as expected) while the later ones don’t.

Same explanation, you’re measuring time wrong. You can only use @time when you synchronize the GPU, which you aren’t.

@ayushinav

If there was some issue with measuring time, I guess the first iteration would also have given approx. 7 seconds.

Why to you assume that? The GPU can only queue a number of asynchronous operations, so it’s likely that the first sequence of operations without a synchronize() gets queued up without anything more, while subsequent iterations will need to wait until there’s space in the queue.