Using Adams2019 autoscheduler for device offload cases

[shubhamp-ca]

I am working on a change in apps/autoscheduler that enables emitting schedule for hexagon offload case. It involves the user specifying which stage to be offloaded using the ".hexagon()" directive and autoscheduler picking the correct vector width while autoscheduling depending on whether a stage is offloaded or not. It is to be noted that it is possible for the autoscheduler to offload a stage without the user explicitly directing the offload of that stage if, it is computed within the loopnest of a stage offloaded by the user.

This change doesn't modify the featurization to accommodate device features separately and also, it uses a single weights file for the cost model as opposed to possibly different weights for host and device. Autotuner performance is good with this change but not so much when using autoscheduler in direct mode as hexagon trained weights aren't available. I am curious to know if there is already any effort going on for other device offload cases (GPU?) and how that is implemented, if at all. Any ideas for implementing device offload in autoscheduler will be very helpful. If interested, I will be happy to open a PR for this change.

[shubhamp-ca]

@abadams Any thoughts?

[abadams]

There is effort going on the for the GPU, but it's for GPU-only schedules, so offloading isn't really considered. In that work the search space, featurization, and model are all somewhat different to the CPU. I don't think there's anything useful to transfer over to a hexagon autoscheduler that tries to do heterogeneous compute.

[pranavb-ca]

So, if I understand this correctly, the GPU effort executes every stage on the GPU like (homogenous compute). Is there a plan to make the GPU autoscheduler effort target heterogenous compute?

[abadams]

No, I don't think that's something anyone is thinking about. The target is typically discrete GPUs where it doesn't make sense to go back and forth between the GPU and CPU

[pranavb-ca]

I see, that makes sense. For us the heterogenous case makes sense because there are things (e.g Floating point) that we'd do on the CPU and some other things on the DSP. @shubhamp-ca is working on a way for the user to say if he'd like a particular stage on Hexagon and teh autoscheduler uses that as input. I'll let him chime in with the details (which may even come in the form of a PR). Would love your thoughts on the same.