The CORE-MATH project

CORE-MATH Mission: provide on-the-shelf high performance open-source mathematical functions with correct rounding that can be integrated into current mathematical libraries (GNU libc, Intel Math Library, AMD Libm, Newlib, OpenLibm, Musl, Apple Libm, llvm-libc, Microsoft libm, CUDA libm, ROCm)

Performance of CORE-MATH functions (revision 81d5ea0) on an Intel Xeon Silver 4214, with GCC 13.2.0, compared to GNU libc 2.37, the Intel Math Library from icx 2023.2.0 (with -fp-model=strict -fno-fast-math and LLVM libc (revision 099dbb):

binary32 functions: reciprocal throughput, latency (with revision 81d5ea0 of CORE-MATH)
binary64 functions: reciprocal throughput, latency (with revision 81d5ea0 of CORE-MATH)
new C23 functions: reciprocal throughput, latency (with revision 65a4a9d of CORE-MATH)

Available CORE-MATH functions (under MIT license for the stand-alone functions):

function	binary32	binary64	binary80	binary128
acos	code	code	glibc	glibc (1)
acosh	code	code
acospi	code
asin	code	code
asinh	code	code
asinpi	code
atan	code	code
atan2	code
atan2pi	code
atanh	code	code
atanpi	code
cbrt	code	code (proof)	glibc	glibc
cos	code	code
cosh	code	code
cospi	code	code
erf	code	code
erfc	code	code
exp	code	code
exp10	code	code
exp10m1	code	code
exp2	code	code
exp2m1	code	code
expm1	code	code
hypot	code	code
log	code	code (with Gappa proof)
log10	code	code
log10p1	code
log1p	code	code
log2	code	code
log2p1	code	code
pow	code	code
rsqrt	code	code
sin	code	code
sinh	code	code
sinpi	code	code
tan	code	code
tanh	code	code
tanpi	code	code

Caption:

glibc: patch for GNU libc (details here on how to use it)
llvm: pointer to a correctly-rounded implementation in llvm-libc
proof: a paper proof of the correctness of the algorithm
reserved: this implementation is reserved for a CORE-MATH contributor
(1) tested on hard-to-round cases |x| < 2^-34
(2) only for rounding to nearest

Notes:

correct rounding is claimed for all IEEE-754 rounding modes. For univariate binary32 functions it is checked by exhaustive search. For bivariate functions or larger formats it is checked only with respect to known hard-to-round cases. If you find an input which is not correctly rounded, please send it to paul dot zimmermann @ inria dot fr.
these implementations only care about correct rounding for regular numbers, they do not necessarily correctly deal with NaN, infinities or zero, set the underflow and overflow exceptions, nor the inexact flag.
these implementations were tested on x86_64-linux, with and without the use of fma (fused multiply add).

Other correctly-rounded implementations:

CRlibm this is a mirror of the original CRlibm project (which is no more available). CRlibm is no longer maintained, but contains very useful resources: provides binary64 exp, log, cos, sin, tan, cospi, sinpi, tanpi, atan, atanpi, cosh, sinh, log2, log10, asin, acos, asinpi, acospi, expm1, log1p, exp2, pow (only for rounding to nearest for pow).
RLIBM provides binary32 log, log2, log10, exp, exp2, exp10, sinh, cosh, sinpi, cospi, sin, cos, tan, atan, asin, acos for all rounding modes.
LLVM-libc provides some correctly rounded functions (see links below): claimed accuracy of LLVM functions
JuliaIntervals provides elementary functions with directed rounding wrapping CRlibm.

function	binary32	binary64	binary80	binary128
acos	rlibm llvm	crlibm
acosh	llvm
acospi		crlibm
asin	rlibm llvm	crlibm
asinh	llvm
asinpi		crlibm
atan	rlibm llvm	crlibm
atan2
atanh	llvm	crlibm
atanpi		crlibm
cbrt
cos	rlibm llvm,llvm	crlibm
cosh	rlibm llvm llvm	crlibm
cospi	rlibm	crlibm
erf	llvm
erfc
exp	rlibm llvm,llvm	crlibm, llvm
exp10	rlibm llvm	llvm
exp2	rlibm llvm llvm llvm	crlibm, llvm
expm1	llvm,llvm	crlibm,llvm
hypot	llvm,llvm	llvm
log	rlibm llvm	crlibm, llvm
log10	rlibm llvm	crlibm, llvm
log1p	llvm	crlibm, llvm
log2	rlibm llvm	crlibm, llvm
pow		crlibm (2)
sin	rlibm llvm,llvm	crlibm
sincos	llvm
sinh	rlibm llvm llvm	crlibm
sinpi	rlibm	crlibm
tan	rlibm llvm,llvm	crlibm
tanh	llvm llvm
tanpi		crlibm