CUSL: A CUDA port of the GNU Scientific Library

Port of some routines of the GNU Scientific Library (GSL) to CUDA.

Dependencies

cuRAND

Available functions

• gsl_ran_dir_3d => sphere.cuh/cu_ran_dir_3d

cu_sf_bessel_K0_scaled_e cu_sf_bessel_K1_scaled_e gsl_sf_bessel_Kn

• gsl_ran_chisq

cheb_eval_e

gsl_sf_lnfact_e gsl_sf_fact_e: factorial gsl_ran_gamma_double: Gamma distribution

cu_sf_psi_int_e: Digamma (Psi) function

cu_poly_eval: polynomial evaluation

Tests

describe tests of consistency between GSL results and CUSL

Work needed: TODO

• cu_ran_dir_3d test
• further tests of distribution of vectors from cu_ran_dir_3d, comparison with GSL. Since we do not have the same RNG available in the GPU and GSL, we cannot test random vectors for the same random numbers
• test cu_sf_bessel_Kn
• cu_ran_chisq: check if distribution is correct and matches GSL
• error checking like GSL: I removed all calls to gsl_sf_result to simplify things, so there is very little automatic error checking in the CUDA routines available here