melodyshih/cufinufft-bench

Benchmarking for GPU NUFFT libraries

cuFINUFFT-bench

Benchmarking for cuFINUFFT library

CPU code

• FINUFFT: https://github.com/flatironinstitute/finufft

GPU code

• CUNFFT: https://github.com/sukunis/CUNFFT
• cuFINUFFT: https://github.com/flatironinstitute/cufinufft
• gpuNUFFT: https://github.com/andyschwarzl/gpuNUFFT

Installing cuFINUFFT

module load cuda/10.0.130_410.48
module load gcc/7.4.0
make lib

Installing FINUFFT

module load cuda/10.0.130_410.48
module load gcc/7.4.0
make lib

Installing CUNFFT

module load cuda/10.0.130_410.48
module load gcc/7.4.0
cd CUNFFT
cd build/  
make distclean  
cmake .. -DCUT_OFF=3 -DCUNFFT_DOUBLE_PRECISION=OFF -DMILLI_SEC=ON \
         -DMEASURED_TIMES=ON -DCUDA_CUDA_LIBRARY=/usr/lib64/libcuda.so -DPRINT_CONFIG=OFF  
make 
make install

cmake FLAG setting

• -DCUT_OFF (the spreading width is FILTER_SIZE = (2*CUT_OFF+2), defined in cunfft_typedefs.h)
• -DCUNFFT_DOUBLE_PRECISION
• -DMEASURED_TIMES
• -DCUDA_CUDA_LIBRARY
• -DMILLI_SEC

Installing gpuNUFFT

Note: gpuNUFFT uses single precision by default.

module load cuda/10.0.130_410.48
module load gcc/7.4.0
module load matlab
cd gpuNUFFT
cd CUDA
mkdir -p build
cd build
cmake .. -DMATLAB_ROOT_DIR=/cm/shared/sw/pkg/vendor/matlab/R2020a
make

Usage

• [Step 1] Install all the libraries
• [Step 2] Add shared libraries path to LD_LIBRARY_PATH

export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:~/finufft/lib:~/CUNFFT/lib:~/cuFINUFFT/lib

• [Step 3] Execute python scripts

python nuffttype1.py

Details for different libraries

Type 1 transform: nonuniform pts to uniform pts

1. FINUFFT/cuFINUFFT:

Input:

• x: a size M float array of x coordinates of M source points
• y: a size M float array of y coordinates of M source points
• z: a size M float array of z coordinates of M source points, or NULL
• c: a size M complex float array of M strengths
• eps: relative l2 tolerance, i.e. . The spreading width is -log_10 (eps/10)

Output:

• f_k: a size N1xN2x...xNd complex float array of N1xN2x...xNd output modes. The modes are ordered from -N/2 to N/2-1 in each dimension and are ordered first in dimension x, then in y and last in z, i.e. in 3D, f_k[k1+k2*N1+k3*N1*N2] approximates .

2. CUNFFT

Input:

• x: a size dM float array of coordinates of M source points, the t-th coordinate of the j-th source point is the d*j+t entry of x.
• c: a complex float array of M strengths

Output:

• f_k = a size N1xN2x...xNdcomplex float array of N1xN2x...xNd modes. The modes are ordered from -N/2 to N/2-1 in each dimension. Modes are ordered first in dimension z, then in y and last in x, i.e. in 3D, f_k[k_3+k_2*N3+k_1*N3*N2] is the Fourier coefficient approximates .

3. gpuNUFFT

Input:

• k: a size d by M float matrix of coordinates of M source points, the t-th coordinate of the j-th source point is the (t,j) entry of k.
• c: a complex float array of M strengths.

Output:

• f_k = a size N1 by N2 (by N3) complex float d-dimensional array of N1xN2x...xNd modes. The modes are ordered from -N/2 to N/2-1 in each dimension. f_k(k1, k2, k3) is the Fourier coefficient approximates .