NeoNUFFT is a rewrite of the fiNUFFT library with extended functionality using highway for portable vectorization. It supports GPU usage through CUDA and ROCm.
Bipp can be configured with the following options:
| Option | Values | Default | Description |
|---|---|---|---|
NEONUFFT_MUTLI_ARCH |
ON, OFF |
ON |
Build for multiple CPU architecture for performance portability |
NEONUFFT_THREADING |
NATIVE, OPENMP, TBB |
OPENMP |
Select GPU backend |
NEONUFFT_GPU |
OFF, CUDA, ROCM |
OFF |
Select GPU backend |
NEONUFFT_BUILD_TESTS |
ON, OFF |
OFF |
Build test executables |
NEONUFFT_INSTALL_LIB |
ON, OFF |
ON |
Add library to install target |
NEONUFFT_BUNDLED_LIBS |
ON, OFF |
ON |
Download and build Google Highway and GoogleTest |
Portable CPU only build:
cmake -B build -DNEONUFFT_THREADING=NATIVE -DNEONUFFT_GPU=OFF -DNEONUFFT_MULTI_ARCH=ON
make -C build
Native build with CUDA for sm_90 architecture:
cmake -B build -DNEONUFFT_THREADING=OPENMP -DNEONUFFT_GPU=CUDA -DNEONUFFT_MULTI_ARCH=OFF -DCMAKE_CXX_FLAGS="-march=native" -DCMAKE_CUDA_ARCHITECTURES=90
make -C build