I saw from a post from 2019 that jitclass should be available on GPU one day and I see it’s still isn’t, wondering if there are any foreseen time estimate for that?
Thank you!
I’m afraid there isn’t. Although we are working on improving the scope of what can be done with the CUDA target, jitclasses are not high enough up the priority list to have them in any near-term plans.
Thank you for the quick reply, I will try to find a workaround for that.
No problem - feel free to post back with specific patterns in code you’re trying to adapt, and I can try and see what they might best map onto in the CUDA target.
I have an operation pattern that takes as input an arbitrary number of matrices, also a set of parameters that can be numeric or strings. The operation processes in inputs to produce an output a matrix of similar shape. We can assume all input and output are from the same shape for simplicity.
So currently I have a jitclass from which the input matrices can be accessed as well as the parameters. Each operation is implemented by overloading only the functionality part so the API is the same for all the different operations which makes it easy to extend and maintain.
It’s working great on CPU, but obviously can’t make it run as-is on GPU.