Python simple cuda

[MaximEremenko]

This pull request introduces a simplified Python interface for Type 3 NUFFTs (nonuniform-to-nonuniform) in 1D, 2D, and 3D within the cufinufft library. The new functions—nufft1d3, nufft2d3, and nufft3d3—allow users to perform Type 3 transforms with minimal setup, similar to the existing Type 1 and Type 2 interfaces.
Changes

• Added nufft1d3, nufft2d3, and nufft3d3 to cufinufft/_simple.py, providing user-friendly access to Type 3 NUFFTs.
• Updated cufinufft/init.py to include the new functions in the module’s public API.
• Modified _invoke_plan to correctly handle Type 3 parameters (s, t, u) and plan creation for Type 3 transforms.

With this update, all NUFFT types (1, 2, and 3) now offer a consistent and simplified Python interface.

[DiamonDinoia]

This looks great.

I would ask @janden to review.

My suggestion it to extend https://github.com/flatironinstitute/finufft/blob/master/python/cufinufft/tests/test_simple.py with type3

[DiamonDinoia]

Hi @MaximEremenko,

I noticed the tests failing do you need help with it?

[MaximEremenko]

Hi @DiamonDinoia ,
I have version that can pass tests. Problem is here:
in function: verify_type3

ind = (int(0.1789 * n_target_pts),) and assert type3_rel_err < 100 * tol

Original CPU code looks a little different.

I made verify_type3 to ind = tuple(n-1 for n in n_tr + (int(0.1789 *n_pts),)) and it works but it looks like GPU version gives bigger error compare to CPU and general test case tuple(np.random.randint(0, n) for n in n_tr + (n_pts,)) does not work correctly.

[ahbarnett]

Really the python tests should not rely on the error of a single output (ind), since this can fluctuate too much. In the C++ and matlab/octave tests I updated this (eg for the latter see https://github.com/flatironinstitute/finufft/blob/master/matlab/test/fullmathtest.m ). They use small tests where computing the answer directly is cheap, then using relative l2-norm of difference.

…

On Wed, May 28, 2025 at 11:01 AM Maxim Eremenko ***@***.***> wrote: *MaximEremenko* left a comment (flatironinstitute/finufft#685) <#685 (comment)> Hi, I have version that can pass tests. Problem is here: in function: verify_type3 ind = (int(0.1789 * n_target_pts),) and assert type3_rel_err < 100 * tol Original CPU code looks a little different. I made verify_type3 to ind = tuple(n-1 for n in n_tr + (int(0.1789 *n_pts),)) and it works but it looks like GPU version gives bigger error compere to CPU and general test case tuple(np.random.randint(0, n) for n in n_tr + (n_pts,)) does not work correctly. — Reply to this email directly, view it on GitHub <#685 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ACNZRSSOZFHEX2Y4PVGOAS33AXFTXAVCNFSM6AAAAAB5ZTXJB6VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDSMJWGY3TENZXGM> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

-- *-------------------------------------------------------------------~^`^~._.~' |\ Alex Barnett Center for Computational Mathematics, Flatiron Institute | \ http://users.flatironinstitute.org/~ahb 646-876-5942

[DiamonDinoia]

Hi @MaximEremenko,

Yes, let's leave something sensible that passes here.
@janden might want to change the tests to align to C++/matlab in a different PR.