GPU type 1: please try much faster output-driven gpu_method=3

[ahbarnett]

Please note that for 3-10x faster type-1 GPU transforms, use the option gpu_method=3, which is a new output-driven algorithm which exploits more threads, by Marco Barbone, building on Julia work of Juan Polanco. It is not the default yet, since we'd like users to try it out first. It doesn't affect type-2, which are already fast.

Enjoy!

[DiamonDinoia]

I would also add that for type 2 we (thanks to Robert Blackwell) tweaked the parallelism to leverage newer GPUs better. On recent hardware this transform can also be 2/4x faster