Hello,
I’m currently trying to setup MathDx on a Rocky Linux 9.6 machine (using the Eclipse plugin from the CUDA toolkit) in hopes of using cuFFTDx. I followed the instructions Nvidia posted (here: Quick Installation Guide — cuFFTDx and here: Installation Guide — nvCOMPDx ) to the best of my understanding, but I’m still having issues getting it to work.
Unpacking the .tar into my project and trying to build results in an error where one of the compression tests complains about not being able to find nvlz4.h, which I also cannot find.
Removing the examples from the project causes the complex_types.hpp, block_fft.hpp, preprocess_fold.hpp, post_proces_fold.hpp, fft_block_preprocess.hpp, fft_postprocess.hpp, fft_block_redistribute.hpp, and fft_execution.hpp files to error out with complaints about CUDA staples like “threadIdx” and __synchthreads() not being declared in scope.
I have also tried moving the MathDx content to /usr/local and link to it there, but I get identical results to when I remove the examples.
I am able to compile and run other projects in my Eclipse CUDA workspace.
I am running:
OS: Rocky 9.6
G++: 11.5.0 (default C++ 17 standard)
CUDA: 13.0.88
MathDx: nvidia-mathdx-25.06.1 for CUDA 13
Card: A100
Driver: 580.95.05
I have linked:
(path)/nvidia/mathdx/25.06/lib/libnvcompdx.a
(path)/nvidia/mathdx/25.06/lib/libcusolverdx.a
I have included:
(path)/nvidia/mathdx/25.06/include
(path)/nvidia/mathdx/20.06/external/cutlass/include
Generating SM 8.0 SASS and SM 8.0 PTX
Does anyone know what I’m missing?
Thanks for your time/help.