Add arm64 binary for Jetson Nano

[jalfonsi]

Please, provide optimized binaries for pip installing CuPy on Nvidia Jetson Nano arm64 architecture under latest CUDA 10.x . Compiling from source with pip takes a lot of time.

[emcastillo]

@leofang @jakirkham should be possible to do this in conda?

[leofang]

AFAIK My understanding is that currently Anaconda only supports nvcc for Linux + x86-64, so we can't yet build an arm64 package. (It's the same issue for ppc64 btw.)

@jakirkham How difficult is it to support arm64 in nvcc-feedstock?

[emcastillo]

i think i’ve seen aarch64 pacakges for numpy in conda. But can’t look at it right now 😥

[leofang]

ping @jakirkham

[jakirkham]

Ah sorry for missing this. 😞

There are aarch64 packages for things like numpy. However we don't currently have cudatoolkit packages for aarch64, which we would need for CUDA enabled packages on that platform. Have raised issue ( AnacondaRecipes/cudatoolkit-feedstock#3 ) about aarch64 packages for cudatoolkit.

[leofang]

Thanks, @jakirkham! Do we also need a aarch64 package for nvcc-feedstock?

[jakirkham]

Sorry for the slow reply. We would, but it would be straightforward once we have the other pieces in place.