himanshugoel2797 / SRW_FFT_Example

Simplified test for gpu optimization of SRW's FFT

SRW FFT Example

This is a simple program demonstrating porting of components of SRW to CUDA. The FFT implementation (via FFTW3) is taken from SRW, and modified to use cufft, the additional processing done on the FFT output has also been ported to CUDA.

Compiling and Running

Compilation requires nvcc and cufft to be installed and available on the PATH variable. To compile without plan reuse:

make

To compile with plan reuse:

make CFLAGS="-DPLAN_REUSE"

On Windows this will require GNU/Make to be installed, which for Chocolatey users is available via:

choco install make

To run (on Linux):

./program 8

To run (on Windows):

.\program.exe 8

where the argument specifies the length of the FFT to be performed in powers of 10.

Example Output

$ ./program.exe 7
Starting benchmark
Benchmark took 5.63443 ms as measured from CPU.
Benchmark took 5.64782 ms as measured from GPU.

Code Structure

• The program entry point and benchmark setup is in kernel.cu
• The FFT wrapper is in gmfft.cpp
• CUDA ports of various portions of FFT pre/post-processing are in ```fft_kernels.cu