[AutoTVM] Move `with tophub_context` to a more proper place?

zhiics · October 3, 2018, 12:03am

dmlc/tvm/blob/master/nnvm/python/nnvm/compiler/build_module.py#L244-L250


if isinstance(autotvm.DispatchContext.current, autotvm.FallbackContext):
    tophub_context = autotvm.tophub.context(target)
else:
    tophub_context = autotvm.util.EmptyContext()


with tophub_context:
    shape = shape if shape else {}

I think we probably need to move with tophub_context here to some other places as a couple of reasons.

This seems not working for multiple targets if we want have heterogeneous compilation in the build function.
Do we need to have users download the pre-tuned parameters in the build? Sometimes they may don’t even have network connection.

Are these the major places we need them?

github.com

dmlc/tvm/blob/master/topi/python/topi/arm_cpu/conv2d.py#L530-L531


cfg = autotvm.DispatchContext.current.query(tvm.target.current_target(), workload)

github.com

dmlc/tvm/blob/master/topi/python/topi/cuda/conv2d_winograd.py#L369-L370


cfg = autotvm.DispatchContext.current.query(tvm.target.current_target(), workload)

Thoughts?

@merrymercy, @tqchen, @kevinthesun, @yzhliu, @Laurawly

Thanks.

merrymercy · October 3, 2018, 7:50am

If there is no network connection, currently a warning will be raised and the compilation can continue
I added support for multiple targets by https://github.com/dmlc/tvm/pull/1803