[QST] JIT Compilation

[ahoffus]

I'm trying to compile matx in my program to make use of JIT methods. I'm getting lots of compilation errors which mainly arise from the fact that

#include <libcufftdx.h>

doesn't seem to exist. As far as I can tell there's only cufftdx.hpp available in the mathdx include directory ? Then if I adjust that include file there's a bunch of errors about things like "cufftdxCreateDescriptor" is undefined.

Am I missing something obvious ? I'm using nvidia-mathdx-25.12.0-cuda13

[cliffburdick]

Hi @ahoffus, which CMake flags are you using? If you pass -DMATX_EN_MATHDX=ON CMake should download the MathDx library and libmathdx so the headers exist. If you don't need to use the MathDX functionality (JIT FFTs) then you can just use MATX_EN_JIT=ON and your regular MatX kernels can be jitted as well. There is not as much benefit of this math since we fuse most things anyways.

[ahoffus]

I'm using bazel and was injecting matx manually - I must have missed whichever part downloaded that during my translation of the CMake so I'll go back to the drawing board ! Thanks for the response 👍

[cliffburdick]

Thanks @ahoffus . Please keep us updated on your JIT uses and let us know any pain points.