https://hgpu.org/?p=12795
Using hybrid GPU/CPU kernel splitting to accelerate spherical convolutions