Skip to content

Scientific library for high-precision computations and research

License

Notifications You must be signed in to change notification settings

Maratyszcza/FPplus

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

9 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

FPplus

Scientific library for high-precision computations and research

FPplus was originally developed for a research project on instructions to accelerate high-precision computations, but it is also useful as a general-purpose library. FPplus features:

  • Header-only library for error-free transforms and double-double computations
    • Implements error-free addition, multiplication, and fused multiply-add
    • Implements double-double addition and multiplication in multiple variants
  • Compatible with C99, C++, OpenCL, and CUDA
  • Special versions of error-free transforms in SIMD intrinsics:
    • x86 SIMD (128-bit and 256-bit AVX + FMA, 512-bit wide MIC and AVX-512)
    • IBM VSX (POWER 7 and POWER 8) and QPX (Blue Gene/Q)
    • ARMv8 SIMD
  • Extensive documentation with references to scientific literature
  • Testsuite based on MPFR and Google Test
  • Examples and code-generators for high-precision algorithms:
    • Polynomial evaluation with compensated Horner scheme
    • Compensated dot product algorithm
    • Inner kernel of matrix multiplication (GEMM) operation in double-double precision

Requirements

CPU targets:
  • gcc-compatible compiler (tested on gcc, clang and icc)
  • Hardware FMA support
  • Precise floating-point semantics
    • No -ffast-math option when compiling with gcc or clang
    • -fp-model precise when compiling with icc
OpenCL targets:
  • cl_khr_fp64, cl_amd_fp64, or cl_APPLE_fp64_basic_ops extension
  • Hardware FMA support (FP_FAST_FMA must be defined by OpenCL compiler)
  • Precise floating-point semantics
    • No -cl-fast-relaxed-math option
CUDA targets:
  • Compute capability 2.0 or higher

Using FPplus

#include <fpplus.h>

Publications

Marat Dukhan, Richard Vuduc and Jason Riedy "Wanted: Floating-Point Add Round-off Error instruction". arXiv preprint 1603.00491 (2016)

Acknowledgements

HPC Garage logo Georgia Tech College of Computing logo

The library was developed by Marat Dukhan as a research project at Richard Vuduc's HPC Garage lab in the Georgia Institute of Technology, College of Computing, School of Computational Science and Engineering. FPplus is based on algorithms in Handbook of Floating-Point Arithmetics, David Bailey's QD library, the works of Jonathan Shewchuk, Theodorus Dekker, Donald Knuth, and Sylvie Boldo and Jean-Michel Muller. We thank Jason Riedy for his feedback and support.

This material is based upon work supported by the U.S. National Science Foundation (NSF) Award Number 1339745 and the U.S. Dept. of Energy (DOE), Office of Science, Advanced Scientific Computing Research under award DE-FC02-10ER26006/DE-SC0004915. Any opinions, findings and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect those of NSF or DOE.

About

Scientific library for high-precision computations and research

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published