apache / tvm

Open deep learning compiler stack for cpu, gpu and specialized accelerators
https://tvm.apache.org/
Apache License 2.0
11.83k stars 3.48k forks source link

[AOT] Explicitly support heterogeneous execution with AOT Executor #10034

Closed areusch closed 2 months ago

areusch commented 2 years ago

Currently, AOTExecutor only supports heterogeneous execution by leveraging tir.call_extern via BYOC. In order to properly model a heterogeneous environment (in particular, to model multiple disjoint memories and leverage #8404), AOTExecutor needs to formally support invoking the Device API. It's currently not possible to do this from TIR via the AOT run-function.

cc @Mousius @alanmacd @lhutton1

driazati commented 2 years ago

@areusch is this still planned?

areusch commented 2 years ago

yes