MatX FFT
There was a chance to test FFT of NVIDIA's C++ numerical calculation library MatX.
Along with this, I made a shared libray which can be used by not only from C++. The test environment
is as below.
- Xeon W-2133 3.6GHz 6-core 12-thread 16GB
- NVIDIA RTX A2000
- Ubuntu 22.04.4 LTS
- CUDA 12.4.131
MatX supports only Linux, then I tested on Ubuntu, the test data is as below.
- 2D data: 4096x4096, complex number(single precision)
2D FFT execution time was as below.
- 2D FFT execution time: 12[ms]
- 2D FFT execution time + CPU->GPU/ GPU->CPU transfer time: 70[ms]
The data transfer time becomes a problem rather than FFT execution time. In addtion, 70[ms] was about
the same speed to Python's CuPy.
- Python 3.11.8
- CuPy 13.0.0
For those who are interested, the made library(libMatxFft.so) and sample programs(C/C++, C#, Python)
can be downloaded.
MatxFftLibSample[Ubuntu, zip-1.0MB]