Accelerating single-step evaluations through GPU offloading


  • Alexander Freudenberg
  • Martin Schlather University of Mannheim
  • Jeremie Vandenplas Animal Breeding and Genomics, Wageningen UR
  • Torsten Pook Animal Breeding and Genomics, Wageningen UR


Single Nucleotide Polymorphism (SNP) genotype datasets used in empirical research are steadily growing in size which has introduced challenges in the calculation of population statistics that are based on large parts of the genome. In other fields, similar computational challenges have been tackled with the help of GPUs. We have developed a range of algorithms for the calculation of SNP genotype matrix operations widely used in empirical studies, which take advantage of modern NVIDIA GPUs. We provide an implementation in the C library miraculix and exemplary interfaces in Julia and Fortran. To ease adaptation, we also supply functions to calculate a number of derivatives, such as the genomic relationship matrix (GRM), linkage disequilibrium (LD) statistics, the genomic BLUP, and principal components analysis. Source code is released under the Apache 2.0 licence and is freely available at GitHub. The library is developed in C, C++ and CUDA.