Getting Started

These examples utilize the following toolsets:

• cuFFT
• cuFFTDx (Requires joining CUDA Math Library Early Access Program) https://developer.nvidia.com/CUDAMathLibraryEA
• C++11
• CUB 1.13+

Hardware

Volta+

cuFFT_vs_cuFFTDx

This code runs three scenarios

1. cuFFT using cudaMalloc
2. cuFFT using cudaMallocManaged
3. cuFFTDx using cudaMalloc

Objectives

1. Compare coding styles between cuFFT, using cudaMalloc and cudaMallocManaged
2. Compare performance between cuFFT, using cudaMalloc and cudaMallocManaged
3. Compare performance and results between cuFFT and cuFFTDx

Execution

For float

mkdir build;
cd build
cmake -DCMAKE_CUDA_ARCHITECTURES=75 -DCUB_DIR=${HOME}/workStuff/git_examples/cub -DCUFFTDX_DIR=${HOME}/workStuff/cufft/libcufftdx/include ..
make -j

If you don't pass -DCMAKE_CUDA_ARCHITECTURES=XX versions CC60, CC70, CC75, and CC80 will be built.

Output

$ D2Z_Z2D/D2Z_Z2D 
cufftExecD2Z/Z2D - FFT/IFFT - Managed   29.65 ms
cufftExecD2Z/Z2D - FFT/IFFT - Managed   20.99 ms
cufftExecC2C - FFT/IFFT - Dx            23.31 ms

Compare results [Malloc/Managed]
All values match!

Compare results [Malloc/Dx]
All values match!

Notes

1. This code utilizes cuFFT Callbacks

• https://devblogs.nvidia.com/cuda-pro-tip-use-cufft-callbacks-custom-data-processing/

2. This code utilizes separate compilation and linking

• https://devblogs.nvidia.com/separate-compilation-linking-cuda-device-code/