aboutsummaryrefslogtreecommitdiff
path: root/sysdeps/ieee754
diff options
context:
space:
mode:
authorAdhemerval Zanella <adhemerval.zanella@linaro.org>2024-10-25 15:21:49 -0300
committerAdhemerval Zanella <adhemerval.zanella@linaro.org>2024-11-01 11:27:35 -0300
commitbbd578b38df4ef417acc9af332fa2ddbb4952df4 (patch)
treecc71c3652b5d47db815a62283c1633b21c946163 /sysdeps/ieee754
parent5c22fd25c15ff56583ca8588f137d8e59d775752 (diff)
downloadglibc-bbd578b38df4ef417acc9af332fa2ddbb4952df4.zip
glibc-bbd578b38df4ef417acc9af332fa2ddbb4952df4.tar.gz
glibc-bbd578b38df4ef417acc9af332fa2ddbb4952df4.tar.bz2
math: Use expm1f from CORE-MATH
The CORE-MATH implementation is correctly rounded (for any rounding mode) and shows better performance compared to the generic expm1f. The code was adapted to glibc style and to use the definition of math_config.h (to handle errno, overflow, and underflow). Benchtest on x64_64 (Ryzen 9 5900X, gcc 14.2.1), aarch64 (Neoverse-N1, gcc 13.3.1), and powerpc (POWER10, gcc 13.2.1): Latency master patched improvement x86_64 96.7402 36.4026 62.37% x86_64v2 97.5391 33.4625 65.69% x86_64v3 82.1778 30.8668 62.44% i686 120.58 94.8302 21.35% aarch64 32.3558 12.8881 60.17% power10 23.5087 9.8574 58.07% powerpc 23.4776 9.06325 61.40% reciprocal-throughput master patched improvement x86_64 27.8224 15.9255 42.76% x86_64v2 27.8364 9.6438 65.36% x86_64v3 20.3227 9.6146 52.69% i686 63.5629 59.4718 6.44% aarch64 17.4838 7.1082 59.34% power10 12.4644 8.7829 29.54% powerpc 14.2152 5.94765 58.16% Signed-off-by: Alexei Sibidanov <sibid@uvic.ca> Signed-off-by: Paul Zimmermann <Paul.Zimmermann@inria.fr> Signed-off-by: Adhemerval Zanella <adhemerval.zanella@linaro.org> Reviewed-by: DJ Delorie <dj@redhat.com>
Diffstat (limited to 'sysdeps/ieee754')
-rw-r--r--sysdeps/ieee754/flt-32/s_exp2m1f.c2
-rw-r--r--sysdeps/ieee754/flt-32/s_expm1f.c232
2 files changed, 113 insertions, 121 deletions
diff --git a/sysdeps/ieee754/flt-32/s_exp2m1f.c b/sysdeps/ieee754/flt-32/s_exp2m1f.c
index f899152..1ed7068 100644
--- a/sysdeps/ieee754/flt-32/s_exp2m1f.c
+++ b/sysdeps/ieee754/flt-32/s_exp2m1f.c
@@ -3,7 +3,7 @@
Copyright (c) 2022-2024 Alexei Sibidanov.
The original version of this file was copied from the CORE-MATH
-project (file src/binary32/exp2m1/exp2m1f.c, revision bc385c2).
+project (file src/binary32/exp2m1/exp2m1f.c, revision baf5f6b).
Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
diff --git a/sysdeps/ieee754/flt-32/s_expm1f.c b/sysdeps/ieee754/flt-32/s_expm1f.c
index 35f7b52..edd7c9a 100644
--- a/sysdeps/ieee754/flt-32/s_expm1f.c
+++ b/sysdeps/ieee754/flt-32/s_expm1f.c
@@ -1,132 +1,124 @@
-/* s_expm1f.c -- float version of s_expm1.c.
- */
+/* Correctly-rounded natural exponent function biased by 1 for binary32 value.
-/*
- * ====================================================
- * Copyright (C) 1993 by Sun Microsystems, Inc. All rights reserved.
- *
- * Developed at SunPro, a Sun Microsystems, Inc. business.
- * Permission to use, copy, modify, and distribute this
- * software is freely granted, provided that this notice
- * is preserved.
- * ====================================================
- */
+Copyright (c) 2022-2024 Alexei Sibidanov.
+
+This file is part of the CORE-MATH project
+project (file src/binary32/expm1/expm1f.c, revision bc385c2).
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
+*/
-#include <errno.h>
-#include <float.h>
#include <math.h>
-#include <math-barriers.h>
-#include <math_private.h>
#include <math-underflow.h>
#include <libm-alias-float.h>
-
-static const float huge = 1.0e+30;
-static const float tiny = 1.0e-30;
-
-static const float
-one = 1.0,
-o_threshold = 8.8721679688e+01,/* 0x42b17180 */
-ln2_hi = 6.9313812256e-01,/* 0x3f317180 */
-ln2_lo = 9.0580006145e-06,/* 0x3717f7d1 */
-invln2 = 1.4426950216e+00,/* 0x3fb8aa3b */
- /* scaled coefficients related to expm1 */
-Q1 = -3.3333335072e-02, /* 0xbd088889 */
-Q2 = 1.5873016091e-03, /* 0x3ad00d01 */
-Q3 = -7.9365076090e-05, /* 0xb8a670cd */
-Q4 = 4.0082177293e-06, /* 0x36867e54 */
-Q5 = -2.0109921195e-07; /* 0xb457edbb */
+#include "math_config.h"
float
-__expm1f(float x)
+__expm1f (float x)
{
- float y,hi,lo,c,t,e,hxs,hfx,r1;
- int32_t k,xsb;
- uint32_t hx;
-
- GET_FLOAT_WORD(hx,x);
- xsb = hx&0x80000000; /* sign bit of x */
- if(xsb==0) y=x; else y= -x; /* y = |x| */
- hx &= 0x7fffffff; /* high word of |x| */
-
- /* filter out huge and non-finite argument */
- if(hx >= 0x4195b844) { /* if |x|>=27*ln2 */
- if(hx >= 0x42b17218) { /* if |x|>=88.721... */
- if(hx>0x7f800000)
- return x+x; /* NaN */
- if(hx==0x7f800000)
- return (xsb==0)? x:-1.0;/* exp(+-inf)={inf,-1} */
- if(x > o_threshold) {
- __set_errno (ERANGE);
- return huge*huge; /* overflow */
- }
- }
- if(xsb!=0) { /* x < -27*ln2, return -1.0 with inexact */
- math_force_eval(x+tiny);/* raise inexact */
- return tiny-one; /* return -1 */
- }
- }
-
- /* argument reduction */
- if(hx > 0x3eb17218) { /* if |x| > 0.5 ln2 */
- if(hx < 0x3F851592) { /* and |x| < 1.5 ln2 */
- if(xsb==0)
- {hi = x - ln2_hi; lo = ln2_lo; k = 1;}
- else
- {hi = x + ln2_hi; lo = -ln2_lo; k = -1;}
- } else {
- k = invln2*x+((xsb==0)?(float)0.5:(float)-0.5);
- t = k;
- hi = x - t*ln2_hi; /* t*ln2_hi is exact here */
- lo = t*ln2_lo;
- }
- x = hi - lo;
- c = (hi-x)-lo;
- }
- else if(hx < 0x33000000) { /* when |x|<2**-25, return x */
- math_check_force_underflow (x);
- t = huge+x; /* return x with inexact flags when x!=0 */
- return x - (t-(huge+x));
+ static const double c[] =
+ {
+ 1, 0x1.62e42fef4c4e7p-6, 0x1.ebfd1b232f475p-13, 0x1.c6b19384ecd93p-20
+ };
+ static const double ch[] =
+ {
+ 0x1.62e42fefa39efp-6, 0x1.ebfbdff82c58fp-13, 0x1.c6b08d702e0edp-20,
+ 0x1.3b2ab6fb92e5ep-27, 0x1.5d886e6d54203p-35, 0x1.430976b8ce6efp-43
+ };
+ static const double td[] =
+ {
+ 0x1p+0, 0x1.059b0d3158574p+0, 0x1.0b5586cf9890fp+0,
+ 0x1.11301d0125b51p+0, 0x1.172b83c7d517bp+0, 0x1.1d4873168b9aap+0,
+ 0x1.2387a6e756238p+0, 0x1.29e9df51fdee1p+0, 0x1.306fe0a31b715p+0,
+ 0x1.371a7373aa9cbp+0, 0x1.3dea64c123422p+0, 0x1.44e086061892dp+0,
+ 0x1.4bfdad5362a27p+0, 0x1.5342b569d4f82p+0, 0x1.5ab07dd485429p+0,
+ 0x1.6247eb03a5585p+0, 0x1.6a09e667f3bcdp+0, 0x1.71f75e8ec5f74p+0,
+ 0x1.7a11473eb0187p+0, 0x1.82589994cce13p+0, 0x1.8ace5422aa0dbp+0,
+ 0x1.93737b0cdc5e5p+0, 0x1.9c49182a3f09p+0, 0x1.a5503b23e255dp+0,
+ 0x1.ae89f995ad3adp+0, 0x1.b7f76f2fb5e47p+0, 0x1.c199bdd85529cp+0,
+ 0x1.cb720dcef9069p+0, 0x1.d5818dcfba487p+0, 0x1.dfc97337b9b5fp+0,
+ 0x1.ea4afa2a490dap+0, 0x1.f50765b6e454p+0
+ };
+ const double iln2 = 0x1.71547652b82fep+5;
+ const double big = 0x1.8p52;
+ double z = x;
+ uint32_t ux = asuint (x);
+ uint32_t ax = ux << 1;
+ if (__glibc_likely (ax < 0x7c400000u))
+ { /* |x| < 0.15625 */
+ if (__glibc_unlikely (ax < 0x676a09e8u))
+ { /* |x| < 0x1.6a09e8p-24 */
+ if (__glibc_unlikely (ax == 0x0u))
+ return x; /* x = +-0 */
+ return fmaf (fabsf (x), 0x1p-25f, x);
}
- else k = 0;
-
- /* x is now in primary range */
- hfx = (float)0.5*x;
- hxs = x*hfx;
- r1 = one+hxs*(Q1+hxs*(Q2+hxs*(Q3+hxs*(Q4+hxs*Q5))));
- t = (float)3.0-r1*hfx;
- e = hxs*((r1-t)/((float)6.0 - x*t));
- if(k==0) return x - (x*e-hxs); /* c is 0 */
- else {
- e = (x*(e-c)-c);
- e -= hxs;
- if(k== -1) return (float)0.5*(x-e)-(float)0.5;
- if(k==1) {
- if(x < (float)-0.25) return -(float)2.0*(e-(x+(float)0.5));
- else return one+(float)2.0*(x-e);
- }
- if (k <= -2 || k>56) { /* suffice to return exp(x)-1 */
- int32_t i;
- y = one-(e-x);
- GET_FLOAT_WORD(i,y);
- SET_FLOAT_WORD(y,i+(k<<23)); /* add k to y's exponent */
- return y-one;
- }
- t = one;
- if(k<23) {
- int32_t i;
- SET_FLOAT_WORD(t,0x3f800000 - (0x1000000>>k)); /* t=1-2^-k */
- y = t-(e-x);
- GET_FLOAT_WORD(i,y);
- SET_FLOAT_WORD(y,i+(k<<23)); /* add k to y's exponent */
- } else {
- int32_t i;
- SET_FLOAT_WORD(t,((0x7f-k)<<23)); /* 2^-k */
- y = x-(e+t);
- y += one;
- GET_FLOAT_WORD(i,y);
- SET_FLOAT_WORD(y,i+(k<<23)); /* add k to y's exponent */
- }
+ static const double b[] =
+ {
+ 0x1.fffffffffffc2p-2, 0x1.55555555555fep-3, 0x1.555555559767fp-5,
+ 0x1.1111111098dc1p-7, 0x1.6c16bca988aa9p-10, 0x1.a01a07658483fp-13,
+ 0x1.a05b04d2c3503p-16, 0x1.71de3a960b5e3p-19
+ };
+ double z2 = z * z, z4 = z2 * z2;
+ double r = z + z2
+ * ((b[0] + z * b[1]) + z2 * (b[2] + z * b[3])
+ + z4 * ((b[4] + z * b[5]) + z2 * (b[6] + z * b[7])));
+ return r;
+ }
+ if (__glibc_unlikely (ax >= 0x8562e430u))
+ { /* |x| > 88.72 */
+ if (ax > (0xffu << 24))
+ return x + x; /* nan */
+ if (__glibc_unlikely (ux >> 31))
+ { /* x < 0 */
+ if (ax == (0xffu << 24))
+ return -1.0f;
+ return -1.0f + 0x1p-26f;
}
- return y;
+ if (ax == (0xffu << 24))
+ return INFINITY;
+ return __math_oflowf (0);
+ }
+ double a = iln2 * z;
+ double ia = roundeven (a);
+ double h = a - ia;
+ double h2 = h * h;
+ uint64_t u = asuint64 (ia + big);
+ double c2 = c[2] + h * c[3], c0 = c[0] + h * c[1];
+ const uint64_t *tdl = (uint64_t *) ((void *) td);
+ double sv = asdouble (tdl[u & 0x1f] + ((u >> 5) << 52));
+ double r = (c0 + h2 * c2) * sv - 1.0;
+ float ub = r, lb = r - sv * 0x1.3b3p-33;
+ if (__glibc_unlikely (ub != lb))
+ {
+ if (__glibc_unlikely (ux > 0xc18aa123u)) /* x < -17.32 */
+ return -1.0f + 0x1p-26f;
+ const double iln2h = 0x1.7154765p+5;
+ const double iln2l = 0x1.5c17f0bbbe88p-26;
+ double s = sv;
+ h = (iln2h * z - ia) + iln2l * z;
+ h2 = h * h;
+ double w = s * h;
+ r = (s - 1) + w
+ * ((ch[0] + h * ch[1])
+ + h2 * ((ch[2] + h * ch[3]) + h2 * (ch[4] + h * ch[5])));
+ ub = r;
+ }
+ return ub;
}
libm_alias_float (__expm1, expm1)
generated by cgit v1.1 at 2025-04-03 08:55:40 +0000