Matlab Mexcuda support for cufinufft

[lu1and10]

To add support gpuArray in Matlab for Matlab interface with cufinufft as the backend.

• Matlab cuda guru interface and simple interface
• Test cpu examples/tests don't break with gpu checks included in simple interface
• Check Matlab cufinufft_plan.m guru interface
• Add gpuArray version of check_finufft.m, check_finufft_single.m, math tests and examples.
• Check simple interface works for gpuArray
• Add cufinufft_plan.docsrc
• Adjust simple interface docsrc to incorporate cufinufft support, add gpu opts docs(cpu opts are different)
• Add instruction to build mexcuda in docs/install_gpu.rst
• Update docs/matlab.rst
• fix clang format to exclude finufft.cpp mwrap-generated code and cufinufft.cu

[DiamonDinoia]

we don't reallly need to specify the architectures as it should use native by default.

[lu1and10]

Yes, we don't have to. But I find it's a bit annoying compiled on my FI cluster workstation with P4000 card, later testing on v100, a100 and v100 recompiling every time, at least for cluster users compiling with different arches save some time.

[ahbarnett]

Nice progress. Are you making a PR for mwrap github too? (so we can see the new formats for gpuarrays). Will be exciting to have that as a general tool too (I could add examples to mwrapdemo repo...). Let me know if need any help.

[lu1and10]

Are you making a PR for mwrap github too? (so we can see the new formats for gpuarrays). Will be exciting to have that as a general tool too (I could add examples to mwrapdemo repo...).

Yes, I should make a PR for mwrap, I tested mwrap with simple demo https://github.com/lu1and10/mwrap/blob/gpu/testing/test_gpu.mw and our cufinufft code wrap. I should do more cleaning and have more tests C/Fortran, different combinations of gpu array of integers, complex, floating point types with in/out/inout tokens. Maybe I should make a draft PR @zgimbutas and add more tests for mwrp later.

Let me know if need any help.

Yes, I think it will be useful to go through the mexcuda install/compile process on different machines, and Matlab versions to see what kind of problem we could meet. For now, I tested on FI cluster and @haiszhu tested on his ubuntu machine.

[DiamonDinoia]

Then I suggest using all-major as the supported architectures depends on the cuda version installed. This fails with cuda 11.3 for example.

…

On Tuesday, February 18, 2025, Libin Lu ***@***.***> wrote: ***@***.**** commented on this pull request. ------------------------------ In docs/install_gpu.rst <#634 (comment)> : > @@ -92,3 +92,35 @@ Assuming ``pytest`` is installed (otherwise, just run ``pip install pytest``), y In contrast to the C interface tests, these check for correctness, so a successful test run signifies that the library is working correctly. Note that you can specify other framework (``pycuda``, ``torch``, or ``numba``) for testing using the ``--framework`` argument. + + +Matlab interface +---------------- + +.. _install-matlab-gpu: + +In addition to the Python interface, cuFINUFFT also comes with a Matlab interface. To install this, you first build the shared library. +For example, assuming in the root directory of finufft, then run + +.. code-block:: bash + + cmake -S . -B build -D FINUFFT_USE_CUDA=ON -D FINUFFT_STATIC_LINKING=OFF -D CMAKE_VERBOSE_MAKEFILE:BOOL=ON -D FINUFFT_CUDA_ARCHITECTURES="60;70;80;90" Yes, we don't have to. But I find it's a bit annoying compiled on my FI cluster workstation with P4000 card, later testing on v100, a100 and v100 recompiling every time, at least for cluster users compiling with different arches save some time. — Reply to this email directly, view it on GitHub <#634 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ACGKNQMGLLSQM2UMUZ6QB4D2QO4IFAVCNFSM6AAAAABXMHEAAKVHI2DSMVQWIX3LMV43YUDVNRWFEZLROVSXG5CSMV3GSZLXHMZDMMRVGI2DENJUG4> . You are receiving this because you commented.Message ID: ***@***.***>

[lu1and10]

Then I suggest using all-major as the supported architectures depends on the cuda version installed. This fails with cuda 11.3 for example.

That is why there is line saying that "You may adjust FINUFFT_CUDA_ARCHITECTURES to generate the code for different compute capabilities.".
I think it's a use case of FINUFFT_CUDA_ARCHITECTURES which you also use here https://github.com/DiamonDinoia/finufft/blob/d10220f23ae472e660006d62fe5d4638e1a7d4e6/.github/workflows/build_cufinufft_wheels.yml#L61

[zgimbutas]

I would be happy to help with the pull request to mwrap once the patch is ready. Thanks for taking care of gpu/cuda bindings.On Feb 18, 2025, at 5:27 PM, Libin Lu ***@***.***> wrote: Then I suggest using all-major as the supported architectures depends on the cuda version installed. This fails with cuda 11.3 for example. That is why there is line saying that "You may adjust FINUFFT_CUDA_ARCHITECTURES to generate the code for different compute capabilities.". I think it's a use case of FINUFFT_CUDA_ARCHITECTURES which you also use here https://github.com/DiamonDinoia/finufft/blob/d10220f23ae472e660006d62fe5d4638e1a7d4e6/.github/workflows/build_cufinufft_wheels.yml#L61—Reply to this email directly, view it on GitHub, or unsubscribe.You are receiving this because you were mentioned.Message ID: ***@***.***> lu1and10 left a comment (flatironinstitute/finufft#634) Then I suggest using all-major as the supported architectures depends on the cuda version installed. This fails with cuda 11.3 for example. That is why there is line saying that "You may adjust FINUFFT_CUDA_ARCHITECTURES to generate the code for different compute capabilities.". I think it's a use case of FINUFFT_CUDA_ARCHITECTURES which you also use here https://github.com/DiamonDinoia/finufft/blob/d10220f23ae472e660006d62fe5d4638e1a7d4e6/.github/workflows/build_cufinufft_wheels.yml#L61 —Reply to this email directly, view it on GitHub, or unsubscribe.You are receiving this because you were mentioned.Message ID: ***@***.***>

[lu1and10]

Then I suggest using all-major as the supported architectures depends on the cuda version installed. This fails with cuda 11.3 for example.

@DiamonDinoia btw, a side finding is that the current https://finufft.readthedocs.io/en/latest/install_gpu.html#cmake-installation does not provide proper install instruction, i.e., cmake -D FINUFFT_USE_CUDA=ON -D CMAKE_CUDA_ARCHITECTURES=80 .. will not generate code for compute_cap 8.0. CMAKE_CUDA_ARCHITECTURES is not assigned to FINUFFT_CUDA_ARCHITECTURES in the current CMakeLists.txt. (

finufft/CMakeLists.txt

Line 305 in b8cea9d

if(NOT DEFINED FINUFFT_CUDA_ARCHITECTURES)

is always false, since FINUFFT_CUDA_ARCHITECTURES defaults to native, always defined.)