Support Type 3?

[dfm]

There aren't currently any plans to support the Type 3 transform for a few reasons:

• I'm not totally sure of the use cases,
• The logic will probably be somewhat more complicated than the existing implementations, and
• cuFINUFFT doesn't seem to support Type 3.

Do we want to work around these issues?

[tyler-a-cox]

Hi folks,

I wanted to bump this to see if there was any interest in having this feature added? I've found the Type 3 transform in finufft extremely useful and would love to have autodiff capabilities of that function.

Thanks!

[dfm]

Thanks @tyler-a-cox! It's great to know that you have a use case. One question, how important is GPU support? AFAIK, there still isn't any GPU support for Type 3 transforms (@lgarrison, @ahbarnett - is that right?).

I'm happy to chat through what a CPU implementation would take, but I'm probably currently too over-committed to dedicate much coding time myself right now!

[ahbarnett]

It's true that GPU type 3 is some way off (maybe this summer?). I also would like to know use-cases for that (eg MRI uses t1 and t2) to understand how to prioritize that.

However, wrapping t3 CPU would be not too hard a generalization of t1, since the source derivatives are similar, but you need to add target derivs too (as in a t2). Much easier project than understanding how the t3 works :)

PS Leslie & I recently realized
flatironinstitute/finufft#415
you could (at least for t1, t2) push the derivatives through to the kernel, so that a combo of value plus derivs could be done with one FFT. Maybe not a game-changer for 1d1 (at most 2x speedup) but, for 3d3 it would turn 7 FFTs into 1.
For now, just a plain vectorized stack of 7 3d3 calls is the way to evaluate all derivs for 3d3.

[tyler-a-cox]

Thanks for the responses, @dfm and @ahbarnett! I'm personally interested in using the type 3 transform for simulating radio interferometric measurements. The sources and measurements are often not gridded, requiring us to use the type 3 transform. On CPU alone, we've found that our implementation with the type 3 is an order of magnitude or two faster than our existing simulator, under certain circumstances. We would definitely be interested in a GPU implementation in the future, but for now the CPU implementation would be very useful!

I will take a look at what was done for the type 1 and 2 transforms and see if I can get a CPU implementation running. Thanks again for the help!