diff options
Diffstat (limited to 'sysdeps')
-rw-r--r-- | sysdeps/i386/fpu/bits/mathinline.h | 6 | ||||
-rw-r--r-- | sysdeps/i386/i686/multiarch/s_fma.c | 2 | ||||
-rw-r--r-- | sysdeps/ieee754/dbl-64/s_fma.c | 146 | ||||
-rw-r--r-- | sysdeps/ieee754/dbl-64/s_fmaf.c | 2 | ||||
-rw-r--r-- | sysdeps/ieee754/ldbl-128/s_fma.c | 50 | ||||
-rw-r--r-- | sysdeps/ieee754/ldbl-96/s_fma.c | 70 | ||||
-rw-r--r-- | sysdeps/ieee754/ldbl-opt/s_fma.c | 2 | ||||
-rw-r--r-- | sysdeps/powerpc/fpu/s_fma.S | 33 | ||||
-rw-r--r-- | sysdeps/powerpc/fpu/s_fmaf.S | 28 | ||||
-rw-r--r-- | sysdeps/powerpc/powerpc32/fpu/s_fma.S | 5 | ||||
-rw-r--r-- | sysdeps/powerpc/powerpc64/fpu/s_fma.S | 5 | ||||
-rw-r--r-- | sysdeps/s390/fpu/s_fma.c | 37 | ||||
-rw-r--r-- | sysdeps/s390/fpu/s_fmaf.c | 32 | ||||
-rw-r--r-- | sysdeps/unix/sysv/linux/s390/fpu/s_fma.c | 5 | ||||
-rw-r--r-- | sysdeps/x86_64/multiarch/s_fma.c | 4 |
15 files changed, 418 insertions, 9 deletions
diff --git a/sysdeps/i386/fpu/bits/mathinline.h b/sysdeps/i386/fpu/bits/mathinline.h index 8005997520..666ec2bc25 100644 --- a/sysdeps/i386/fpu/bits/mathinline.h +++ b/sysdeps/i386/fpu/bits/mathinline.h @@ -1,6 +1,6 @@ /* Inline math functions for i387. - Copyright (C) 1995,1996,1997,1998,1999,2000,2001,2003,2004,2006,2007,2009 - Free Software Foundation, Inc. + Copyright (C) 1995,1996,1997,1998,1999,2000,2001,2003,2004,2006,2007,2009, + 2010 Free Software Foundation, Inc. This file is part of the GNU C Library. Contributed by John C. Bowman <bowman@math.ualberta.ca>, 1995. @@ -657,8 +657,6 @@ __NTH (ldexpl (long double __x, int __y)) __ldexp_code; } -__inline_mathcodeNP3 (fma, __x, __y, __z, return (__x * __y) + __z) - __inline_mathopNP (rint, "frndint") # endif /* __FAST_MATH__ */ diff --git a/sysdeps/i386/i686/multiarch/s_fma.c b/sysdeps/i386/i686/multiarch/s_fma.c index d9291b0be8..2373caabc4 100644 --- a/sysdeps/i386/i686/multiarch/s_fma.c +++ b/sysdeps/i386/i686/multiarch/s_fma.c @@ -33,4 +33,4 @@ weak_alias (__fma, fma) # define __fma __fma_ia32 #endif -#include <math/s_fma.c> +#include <sysdeps/ieee754/ldbl-96/s_fma.c> diff --git a/sysdeps/ieee754/dbl-64/s_fma.c b/sysdeps/ieee754/dbl-64/s_fma.c new file mode 100644 index 0000000000..ca7300c11e --- /dev/null +++ b/sysdeps/ieee754/dbl-64/s_fma.c @@ -0,0 +1,146 @@ +/* Compute x * y + z as ternary operation. + Copyright (C) 2010 Free Software Foundation, Inc. + This file is part of the GNU C Library. + Contributed by Jakub Jelinek <jakub@redhat.com>, 2010. + + The GNU C Library is free software; you can redistribute it and/or + modify it under the terms of the GNU Lesser General Public + License as published by the Free Software Foundation; either + version 2.1 of the License, or (at your option) any later version. + + The GNU C Library is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU + Lesser General Public License for more details. + + You should have received a copy of the GNU Lesser General Public + License along with the GNU C Library; if not, write to the Free + Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA + 02111-1307 USA. */ + +#include <float.h> +#include <math.h> +#include <fenv.h> +#include <ieee754.h> + +/* This implementation uses rounding to odd to avoid problems with + double rounding. See a paper by Boldo and Melquiond: + http://www.lri.fr/~melquion/doc/08-tc.pdf */ + +double +__fma (double x, double y, double z) +{ + union ieee754_double u, v, w; + int adjust = 0; + u.d = x; + v.d = y; + w.d = z; + if (__builtin_expect (u.ieee.exponent + v.ieee.exponent + >= 0x7ff + IEEE754_DOUBLE_BIAS - DBL_MANT_DIG, 0) + || __builtin_expect (u.ieee.exponent >= 0x7ff - DBL_MANT_DIG, 0) + || __builtin_expect (v.ieee.exponent >= 0x7ff - DBL_MANT_DIG, 0) + || __builtin_expect (w.ieee.exponent >= 0x7ff - DBL_MANT_DIG, 0)) + { + /* If x or y or z is Inf/NaN or if fma will certainly overflow, + compute as x * y + z. */ + if (u.ieee.exponent == 0x7ff + || v.ieee.exponent == 0x7ff + || w.ieee.exponent == 0x7ff + || u.ieee.exponent + v.ieee.exponent + > 0x7ff + IEEE754_DOUBLE_BIAS) + return x * y + z; + if (u.ieee.exponent + v.ieee.exponent + >= 0x7ff + IEEE754_DOUBLE_BIAS - DBL_MANT_DIG) + { + /* Compute 1p-53 times smaller result and multiply + at the end. */ + if (u.ieee.exponent > v.ieee.exponent) + u.ieee.exponent -= DBL_MANT_DIG; + else + v.ieee.exponent -= DBL_MANT_DIG; + /* If x + y exponent is very large and z exponent is very small, + it doesn't matter if we don't adjust it. */ + if (w.ieee.exponent > DBL_MANT_DIG) + w.ieee.exponent -= DBL_MANT_DIG; + adjust = 1; + } + else if (w.ieee.exponent >= 0x7ff - DBL_MANT_DIG) + { + /* Similarly. + If z exponent is very large and x and y exponents are + very small, it doesn't matter if we don't adjust it. */ + if (u.ieee.exponent > v.ieee.exponent) + { + if (u.ieee.exponent > DBL_MANT_DIG) + u.ieee.exponent -= DBL_MANT_DIG; + } + else if (v.ieee.exponent > DBL_MANT_DIG) + v.ieee.exponent -= DBL_MANT_DIG; + w.ieee.exponent -= DBL_MANT_DIG; + adjust = 1; + } + else if (u.ieee.exponent >= 0x7ff - DBL_MANT_DIG) + { + u.ieee.exponent -= DBL_MANT_DIG; + if (v.ieee.exponent) + v.ieee.exponent += DBL_MANT_DIG; + else + v.d *= 0x1p53; + } + else + { + v.ieee.exponent -= DBL_MANT_DIG; + if (u.ieee.exponent) + u.ieee.exponent += DBL_MANT_DIG; + else + u.d *= 0x1p53; + } + x = u.d; + y = v.d; + z = w.d; + } + /* Multiplication m1 + m2 = x * y using Dekker's algorithm. */ +#define C ((1 << (DBL_MANT_DIG + 1) / 2) + 1) + double x1 = x * C; + double y1 = y * C; + double m1 = x * y; + x1 = (x - x1) + x1; + y1 = (y - y1) + y1; + double x2 = x - x1; + double y2 = y - y1; + double m2 = (((x1 * y1 - m1) + x1 * y2) + x2 * y1) + x2 * y2; + + /* Addition a1 + a2 = z + m1 using Knuth's algorithm. */ + double a1 = z + m1; + double t1 = a1 - z; + double t2 = a1 - t1; + t1 = m1 - t1; + t2 = z - t2; + double a2 = t1 + t2; + + fenv_t env; + feholdexcept (&env); + fesetround (FE_TOWARDZERO); + /* Perform m2 + a2 addition with round to odd. */ + u.d = a2 + m2; + if ((u.ieee.mantissa1 & 1) == 0 && u.ieee.exponent != 0x7ff) + u.ieee.mantissa1 |= fetestexcept (FE_INEXACT) != 0; + feupdateenv (&env); + + /* Add that to a1. */ + a1 = a1 + u.d; + + /* And adjust exponent if needed. */ + if (__builtin_expect (adjust, 0)) + a1 *= 0x1p53; + + return a1; +} +#ifndef __fma +weak_alias (__fma, fma) +#endif + +#ifdef NO_LONG_DOUBLE +strong_alias (__fma, __fmal) +weak_alias (__fmal, fmal) +#endif diff --git a/sysdeps/ieee754/dbl-64/s_fmaf.c b/sysdeps/ieee754/dbl-64/s_fmaf.c index 9721876601..cd16cd1dce 100644 --- a/sysdeps/ieee754/dbl-64/s_fmaf.c +++ b/sysdeps/ieee754/dbl-64/s_fmaf.c @@ -39,7 +39,7 @@ __fmaf (float x, float y, float z) fesetround (FE_TOWARDZERO); /* Perform addition with round to odd. */ u.d = temp + (double) z; - if ((u.ieee.mantissa1 & 1) == 0 && u.ieee.exponent != 0xff) + if ((u.ieee.mantissa1 & 1) == 0 && u.ieee.exponent != 0x7ff) u.ieee.mantissa1 |= fetestexcept (FE_INEXACT) != 0; feupdateenv (&env); /* And finally truncation with round to nearest. */ diff --git a/sysdeps/ieee754/ldbl-128/s_fma.c b/sysdeps/ieee754/ldbl-128/s_fma.c new file mode 100644 index 0000000000..78055e5c6c --- /dev/null +++ b/sysdeps/ieee754/ldbl-128/s_fma.c @@ -0,0 +1,50 @@ +/* Compute x * y + z as ternary operation. + Copyright (C) 2010 Free Software Foundation, Inc. + This file is part of the GNU C Library. + Contributed by Jakub Jelinek <jakub@redhat.com>, 2010. + + The GNU C Library is free software; you can redistribute it and/or + modify it under the terms of the GNU Lesser General Public + License as published by the Free Software Foundation; either + version 2.1 of the License, or (at your option) any later version. + + The GNU C Library is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU + Lesser General Public License for more details. + + You should have received a copy of the GNU Lesser General Public + License along with the GNU C Library; if not, write to the Free + Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA + 02111-1307 USA. */ + +#include <math.h> +#include <fenv.h> +#include <ieee754.h> + +/* This implementation relies on long double being more than twice as + precise as double and uses rounding to odd in order to avoid problems + with double rounding. + See a paper by Boldo and Melquiond: + http://www.lri.fr/~melquion/doc/08-tc.pdf */ + +double +__fma (double x, double y, double z) +{ + fenv_t env; + /* Multiplication is always exact. */ + long double temp = (long double) x * (long double) y; + union ieee854_long_double u; + feholdexcept (&env); + fesetround (FE_TOWARDZERO); + /* Perform addition with round to odd. */ + u.d = temp + (long double) z; + if ((u.ieee.mantissa3 & 1) == 0 && u.ieee.exponent != 0x7fff) + u.ieee.mantissa3 |= fetestexcept (FE_INEXACT) != 0; + feupdateenv (&env); + /* And finally truncation with round to nearest. */ + return (double) u.d; +} +#ifndef __fma +weak_alias (__fma, fma) +#endif diff --git a/sysdeps/ieee754/ldbl-96/s_fma.c b/sysdeps/ieee754/ldbl-96/s_fma.c new file mode 100644 index 0000000000..90c6d1f29d --- /dev/null +++ b/sysdeps/ieee754/ldbl-96/s_fma.c @@ -0,0 +1,70 @@ +/* Compute x * y + z as ternary operation. + Copyright (C) 2010 Free Software Foundation, Inc. + This file is part of the GNU C Library. + Contributed by Jakub Jelinek <jakub@redhat.com>, 2010. + + The GNU C Library is free software; you can redistribute it and/or + modify it under the terms of the GNU Lesser General Public + License as published by the Free Software Foundation; either + version 2.1 of the License, or (at your option) any later version. + + The GNU C Library is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU + Lesser General Public License for more details. + + You should have received a copy of the GNU Lesser General Public + License along with the GNU C Library; if not, write to the Free + Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA + 02111-1307 USA. */ + +#include <float.h> +#include <math.h> +#include <fenv.h> +#include <ieee754.h> + +/* This implementation uses rounding to odd to avoid problems with + double rounding. See a paper by Boldo and Melquiond: + http://www.lri.fr/~melquion/doc/08-tc.pdf */ + +double +__fma (double x, double y, double z) +{ + /* Multiplication m1 + m2 = x * y using Dekker's algorithm. */ +#define C ((1ULL << (LDBL_MANT_DIG + 1) / 2) + 1) + long double x1 = x * C; + long double y1 = y * C; + long double m1 = x * y; + x1 = (x - x1) + x1; + y1 = (y - y1) + y1; + long double x2 = x - x1; + long double y2 = y - y1; + long double m2 = (((x1 * y1 - m1) + x1 * y2) + x2 * y1) + x2 * y2; + + /* Addition a1 + a2 = z + m1 using Knuth's algorithm. */ + long double a1 = z + m1; + long double t1 = a1 - z; + long double t2 = a1 - t1; + t1 = m1 - t1; + t2 = z - t2; + long double a2 = t1 + t2; + + fenv_t env; + feholdexcept (&env); + fesetround (FE_TOWARDZERO); + /* Perform m2 + a2 addition with round to odd. */ + a2 = a2 + m2; + + /* Add that to a1 again using rounding to odd. */ + union ieee854_long_double u; + u.d = a1 + a2; + if ((u.ieee.mantissa1 & 1) == 0 && u.ieee.exponent != 0x7fff) + u.ieee.mantissa1 |= fetestexcept (FE_INEXACT) != 0; + feupdateenv (&env); + + /* Add finally round to double precision. */ + return u.d; +} +#ifndef __fma +weak_alias (__fma, fma) +#endif diff --git a/sysdeps/ieee754/ldbl-opt/s_fma.c b/sysdeps/ieee754/ldbl-opt/s_fma.c index 2ee7214131..1723c5c306 100644 --- a/sysdeps/ieee754/ldbl-opt/s_fma.c +++ b/sysdeps/ieee754/ldbl-opt/s_fma.c @@ -1,5 +1,5 @@ #include <math_ldbl_opt.h> -#include <math/s_fma.c> +#include <sysdeps/ieee754/dbl-64/s_fma.c> #if LONG_DOUBLE_COMPAT(libm, GLIBC_2_1) compat_symbol (libm, __fma, fmal, GLIBC_2_1); #endif diff --git a/sysdeps/powerpc/fpu/s_fma.S b/sysdeps/powerpc/fpu/s_fma.S new file mode 100644 index 0000000000..9c650eab07 --- /dev/null +++ b/sysdeps/powerpc/fpu/s_fma.S @@ -0,0 +1,33 @@ +/* Compute x * y + z as ternary operation. PowerPC version. + Copyright (C) 2010 Free Software Foundation, Inc. + This file is part of the GNU C Library. + + The GNU C Library is free software; you can redistribute it and/or + modify it under the terms of the GNU Lesser General Public + License as published by the Free Software Foundation; either + version 2.1 of the License, or (at your option) any later version. + + The GNU C Library is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU + Lesser General Public License for more details. + + You should have received a copy of the GNU Lesser General Public + License along with the GNU C Library; if not, write to the Free + Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA + 02111-1307 USA. */ + +#include <sysdep.h> + +ENTRY(__fma) +/* double [f1] fma (double [f1] x, double [f2] y, double [f3] z); */ + fmadd fp1,fp1,fp2,fp3 + blr +END(__fma) + +weak_alias (__fma,fma) + +#ifdef NO_LONG_DOUBLE +weak_alias (__fma,__fmal) +weak_alias (__fma,fmal) +#endif diff --git a/sysdeps/powerpc/fpu/s_fmaf.S b/sysdeps/powerpc/fpu/s_fmaf.S new file mode 100644 index 0000000000..1a755fe2db --- /dev/null +++ b/sysdeps/powerpc/fpu/s_fmaf.S @@ -0,0 +1,28 @@ +/* Compute x * y + z as ternary operation. PowerPC version. + Copyright (C) 2010 Free Software Foundation, Inc. + This file is part of the GNU C Library. + + The GNU C Library is free software; you can redistribute it and/or + modify it under the terms of the GNU Lesser General Public + License as published by the Free Software Foundation; either + version 2.1 of the License, or (at your option) any later version. + + The GNU C Library is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU + Lesser General Public License for more details. + + You should have received a copy of the GNU Lesser General Public + License along with the GNU C Library; if not, write to the Free + Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA + 02111-1307 USA. */ + +#include <sysdep.h> + +ENTRY(__fmaf) +/* float [f1] fmaf (float [f1] x, float [f2] y, float [f3] z); */ + fmadds fp1,fp1,fp2,fp3 + blr +END(__fmaf) + +weak_alias (__fmaf,fmaf) diff --git a/sysdeps/powerpc/powerpc32/fpu/s_fma.S b/sysdeps/powerpc/powerpc32/fpu/s_fma.S new file mode 100644 index 0000000000..d40695c633 --- /dev/null +++ b/sysdeps/powerpc/powerpc32/fpu/s_fma.S @@ -0,0 +1,5 @@ +#include <math_ldbl_opt.h> +#include <sysdeps/powerpc/fpu/s_fma.S> +#if LONG_DOUBLE_COMPAT(libm, GLIBC_2_1) +compat_symbol (libm, __fma, fmal, GLIBC_2_1) +#endif diff --git a/sysdeps/powerpc/powerpc64/fpu/s_fma.S b/sysdeps/powerpc/powerpc64/fpu/s_fma.S new file mode 100644 index 0000000000..d40695c633 --- /dev/null +++ b/sysdeps/powerpc/powerpc64/fpu/s_fma.S @@ -0,0 +1,5 @@ +#include <math_ldbl_opt.h> +#include <sysdeps/powerpc/fpu/s_fma.S> +#if LONG_DOUBLE_COMPAT(libm, GLIBC_2_1) +compat_symbol (libm, __fma, fmal, GLIBC_2_1) +#endif diff --git a/sysdeps/s390/fpu/s_fma.c b/sysdeps/s390/fpu/s_fma.c new file mode 100644 index 0000000000..f682362313 --- /dev/null +++ b/sysdeps/s390/fpu/s_fma.c @@ -0,0 +1,37 @@ +/* Compute x * y + z as ternary operation. S/390 version. + Copyright (C) 2010 Free Software Foundation, Inc. + This file is part of the GNU C Library. + Contributed by Jakub Jelinek <jakub@redhat.com>, 2010. + + The GNU C Library is free software; you can redistribute it and/or + modify it under the terms of the GNU Lesser General Public + License as published by the Free Software Foundation; either + version 2.1 of the License, or (at your option) any later version. + + The GNU C Library is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU + Lesser General Public License for more details. + + You should have received a copy of the GNU Lesser General Public + License along with the GNU C Library; if not, write to the Free + Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA + 02111-1307 USA. */ + +#include <math.h> + +double +__fma (double x, double y, double z) +{ + double r; + asm ("madbr %0,%1,%2" : "=f" (r) : "%f" (x), "fR" (y), "0" (z)); + return r; +} +#ifndef __fma +weak_alias (__fma, fma) +#endif + +#ifdef NO_LONG_DOUBLE +strong_alias (__fma, __fmal) +weak_alias (__fmal, fmal) +#endif diff --git a/sysdeps/s390/fpu/s_fmaf.c b/sysdeps/s390/fpu/s_fmaf.c new file mode 100644 index 0000000000..030d7b229b --- /dev/null +++ b/sysdeps/s390/fpu/s_fmaf.c @@ -0,0 +1,32 @@ +/* Compute x * y + z as ternary operation. S/390 version. + Copyright (C) 2010 Free Software Foundation, Inc. + This file is part of the GNU C Library. + Contributed by Jakub Jelinek <jakub@redhat.com>, 2010. + + The GNU C Library is free software; you can redistribute it and/or + modify it under the terms of the GNU Lesser General Public + License as published by the Free Software Foundation; either + version 2.1 of the License, or (at your option) any later version. + + The GNU C Library is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU + Lesser General Public License for more details. + + You should have received a copy of the GNU Lesser General Public + License along with the GNU C Library; if not, write to the Free + Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA + 02111-1307 USA. */ + +#include <math.h> + +float +__fmaf (float x, float y, float z) +{ + float r; + asm ("maebr %0,%1,%2" : "=f" (r) : "%f" (x), "fR" (y), "0" (z)); + return r; +} +#ifndef __fmaf +weak_alias (__fmaf, fmaf) +#endif diff --git a/sysdeps/unix/sysv/linux/s390/fpu/s_fma.c b/sysdeps/unix/sysv/linux/s390/fpu/s_fma.c new file mode 100644 index 0000000000..2b2d505030 --- /dev/null +++ b/sysdeps/unix/sysv/linux/s390/fpu/s_fma.c @@ -0,0 +1,5 @@ +#include <math_ldbl_opt.h> +#include <sysdeps/s390/fpu/s_fma.c> +#if LONG_DOUBLE_COMPAT(libm, GLIBC_2_1) +compat_symbol (libm, __fma, fmal, GLIBC_2_1); +#endif diff --git a/sysdeps/x86_64/multiarch/s_fma.c b/sysdeps/x86_64/multiarch/s_fma.c index cfecf9b41d..9a680c68fc 100644 --- a/sysdeps/x86_64/multiarch/s_fma.c +++ b/sysdeps/x86_64/multiarch/s_fma.c @@ -1,5 +1,5 @@ /* FMA version of fma. - Copyright (C) 2009 Free Software Foundation, Inc. + Copyright (C) 2009, 2010 Free Software Foundation, Inc. Contributed by Intel Corporation. This file is part of the GNU C Library. @@ -40,4 +40,4 @@ weak_alias (__fma, fma) # define __fma __fma_sse2 #endif -#include <math/s_fma.c> +#include <sysdeps/ieee754/dbl-64/s_fma.c> |