bitrep - Bit-reproductible math functions

Original source by Andrea Arteaga, ported to GPU through OpenACC and with a Fortran interface.

Bit-reproductible results have been tested on:

• NVHPC 25.3 compiler: CPU ARM vs GPU NVIDIA A100 @ HPCC Turpan (CALMIP)

Build instructions

The following will build a static library libbitrep.a and a test binary that will test if identical result are obtained both on the CPU and the GPU (see ./tests/test_bitrep.f90).

cmake -B build -S . -DCMAKE_VERBOSE_MAKEFILE=ON # -DBUILD_TESTINGS=OFF -DCUDA_CC=cc70
cd build
./test-bitrep
# Output:
# GPU::sin         vs CPU::sin         are NOT identical
# GPU::br_sin      vs CPU::br_sin      are identical
# CPU::sin         vs CPU::br_sin      are NOT identical
# GPU::cos         vs CPU::cos         are NOT identical
# GPU::br_cos      vs CPU::br_cos      are identical
# CPU::cos         vs CPU::br_cos      are NOT identical
# GPU::exp         vs CPU::exp         are NOT identical
# GPU::br_exp      vs CPU::br_exp      are identical
# CPU::exp         vs CPU::br_exp      are NOT identical

On NVHPC, specifying the CUDA compute capability (e.g. cc70 for Volta GPUs) may be necessary when bulding fails to generate code for all capabilities.

Other implementations

• Original source by Andrea Arteaga
• GPU OpenACC port of transcendental functions by Philippe Wautelet

Reference

@INPROCEEDINGS{6877351,
  author={Arteaga, Andrea and Fuhrer, Oliver and Hoefler, Torsten},
  booktitle={2014 IEEE 28th International Parallel and Distributed Processing Symposium},
  title={Designing Bit-Reproducible Portable High-Performance Applications},
  year={2014},
  volume={},
  number={},
  pages={1235-1244},
  keywords={Standards;Libraries;Meteorology;Computer architecture;Accuracy;Computational modeling;Graphics processing units;determinism;reproducibility;parallelism;IEEE-754 standard},
  doi={10.1109/IPDPS.2014.127}}