Epi4Tensor is an implementation on modern NVIDIA GPU microarchitectures of an algorithm for efficient fourth-order exhaustive epistasis detection, a bioinformatics application that searches for associations between single nucleotide polymorphisms (SNPs) and a given observable trait. NVIDIA GPUs with third generation tensor cores (Ampere microarchitecture) are efficiently exploited through the extensive use of AND+POPC 1-bit precision operations, while compatibility with second generation tensor cores (Turing microarchitecture) is achieved relying on XOR+POPC 1-bit precision operations. This software has been subject to optimization on systems with A100 (Ampere/SM80) and Titan RTX (Turing/SM75) GPUs. However, it should also work well with many other Turing and Ampere GPUs.
- CUDA Toolkit 11.4 and CUTLASS 2.5 (more recent)
- NVIDIA Turing (SM75) or Ampere (SM80/SM86) GPU
Download and uncompress the CUTLASS library to your preferred location, reflecting it in the Makefile. Compile the tool using the make command.
To test the produced binary you can pass one of the provided sample datasets (download from here) as input, e.g:
$ ./bin/epi4tensor datasets/db_256snps_131072samples.txtThe tool has multi-GPU support. If your system has multiple compatible GPUs, you can customize the tool by setting NUM_GPUS in the Makefile to a number that represents your particular configuration.
Ricardo Nobre, Aleksandar Ilic, Sergio Santander-Jiménez, and Leonel Sousa. 2022. Tensor-Accelerated Fourth-Order Epistasis Detection on GPUs. In 51st International Conference on Parallel Processing (ICPP ’22), August 29-September 1, 2022, Bordeaux, France. ACM, New York, NY, USA, 11 pages. https://doi.org/10.1145/3545008.3545066