cuFINUFFT

A GPU implementation of 2,3 dimension type 1,2 non-uniform FFT based on FINUFFT (https://github.com/flatironinstitute/finufft).

This is a work as a summer intern at Flatiron Institute advised by CCM project leader Alex Barnett.

Code dependency

• CUB library (https://github.com/NVlabs/cub)

Installation

• Get the CUB library - git clone https://github.com/NVlabs/cub.git
• Modify make.inc - set the INC with -I$(CUDA_DIR)/samples/common/inc/ -I$(CUDA_DIR)/include/ -I$(CUB_DIR)
• Compile - make all
• Run a test code - ./cufinufft2d1_test 2 128 128 10 1e-6

Interface

cuFINUFFT API contains 5 stages:

• Set cufinufft default options - int ier=cufinufft_default_opts(type1, dim, dplan.opts);
• Make cufinufft plan - ier=cufinufft_makeplan(type1, dim, nmodes, iflag, ntransf, tol, ntransfcufftplan, &dplan);
• Set the locations of non-uniform points x,y,z - ier=cufinufft_setNUpts(M, x, y, z, 0, NULL, NULL, NULL, &dplan);
• Apply the transformation with data c,fk - ier=cufinufft_exec(c, fk, &dplan);
• Destroy cufinufft plan - ier=cufinufft_destroy(&dplan);

Preprocessors

• SINGLE - single-precision
• TIME - timing for each stage
• SPREADTIME - more detailed timing from spreading and interpolation
• DEBUG - debug mode outputs all the middle stages' result

Other

• If you're running the code on GPU with Compute Capability less than 5.0 (ex. Kepler, Fermi), change the -arch=sm_50 flag to lower number. (See http://arnon.dk/matching-sm-architectures-arch-and-gencode-for-various-nvidia-cards/)