From db505b794c697631f65e6b91ff106496debb86ac Mon Sep 17 00:00:00 2001 From: Szabolcs Nagy Date: Sun, 22 Oct 2017 14:19:20 +0000 Subject: math: new logf from https://github.com/ARM-software/optimized-routines, commit 04884bd04eac4b251da4026900010ea7d8850edc, with minor changes to better fit into musl. code size change: +289 bytes. benchmark on x86_64 before, after, speedup: -Os: logf rthruput: 8.40 ns/call 6.14 ns/call 1.37x logf latency: 31.79 ns/call 24.33 ns/call 1.31x -O3: logf rthruput: 8.43 ns/call 5.58 ns/call 1.51x logf latency: 32.04 ns/call 20.88 ns/call 1.53x --- src/math/logf.c | 110 ++++++++++++++++++++++++++++---------------------------- 1 file changed, 56 insertions(+), 54 deletions(-) (limited to 'src/math/logf.c') diff --git a/src/math/logf.c b/src/math/logf.c index 52230a1b..7ee5d7fe 100644 --- a/src/math/logf.c +++ b/src/math/logf.c @@ -1,69 +1,71 @@ -/* origin: FreeBSD /usr/src/lib/msun/src/e_logf.c */ /* - * Conversion to float by Ian Lance Taylor, Cygnus Support, ian@cygnus.com. - */ -/* - * ==================================================== - * Copyright (C) 1993 by Sun Microsystems, Inc. All rights reserved. + * Single-precision log function. * - * 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) 2017-2018, Arm Limited. + * SPDX-License-Identifier: MIT */ #include #include +#include "libm.h" +#include "logf_data.h" + +/* +LOGF_TABLE_BITS = 4 +LOGF_POLY_ORDER = 4 + +ULP error: 0.818 (nearest rounding.) +Relative error: 1.957 * 2^-26 (before rounding.) +*/ -static const float -ln2_hi = 6.9313812256e-01, /* 0x3f317180 */ -ln2_lo = 9.0580006145e-06, /* 0x3717f7d1 */ -/* |(log(1+s)-log(1-s))/s - Lg(s)| < 2**-34.24 (~[-4.95e-11, 4.97e-11]). */ -Lg1 = 0xaaaaaa.0p-24, /* 0.66666662693 */ -Lg2 = 0xccce13.0p-25, /* 0.40000972152 */ -Lg3 = 0x91e9ee.0p-25, /* 0.28498786688 */ -Lg4 = 0xf89e26.0p-26; /* 0.24279078841 */ +#define T __logf_data.tab +#define A __logf_data.poly +#define Ln2 __logf_data.ln2 +#define N (1 << LOGF_TABLE_BITS) +#define OFF 0x3f330000 float logf(float x) { - union {float f; uint32_t i;} u = {x}; - float_t hfsq,f,s,z,R,w,t1,t2,dk; - uint32_t ix; - int k; + double_t z, r, r2, y, y0, invc, logc; + uint32_t ix, iz, tmp; + int k, i; - ix = u.i; - k = 0; - if (ix < 0x00800000 || ix>>31) { /* x < 2**-126 */ - if (ix<<1 == 0) - return -1/(x*x); /* log(+-0)=-inf */ - if (ix>>31) - return (x-x)/0.0f; /* log(-#) = NaN */ - /* subnormal number, scale up x */ - k -= 25; - x *= 0x1p25f; - u.f = x; - ix = u.i; - } else if (ix >= 0x7f800000) { - return x; - } else if (ix == 0x3f800000) + ix = asuint(x); + /* Fix sign of zero with downward rounding when x==1. */ + if (WANT_ROUNDING && predict_false(ix == 0x3f800000)) return 0; + if (predict_false(ix - 0x00800000 >= 0x7f800000 - 0x00800000)) { + /* x < 0x1p-126 or inf or nan. */ + if (ix * 2 == 0) + return __math_divzerof(1); + if (ix == 0x7f800000) /* log(inf) == inf. */ + return x; + if ((ix & 0x80000000) || ix * 2 >= 0xff000000) + return __math_invalidf(x); + /* x is subnormal, normalize it. */ + ix = asuint(x * 0x1p23f); + ix -= 23 << 23; + } + + /* x = 2^k z; where z is in range [OFF,2*OFF] and exact. + The range is split into N subintervals. + The ith subinterval contains z and c is near its center. */ + tmp = ix - OFF; + i = (tmp >> (23 - LOGF_TABLE_BITS)) % N; + k = (int32_t)tmp >> 23; /* arithmetic shift */ + iz = ix - (tmp & 0x1ff << 23); + invc = T[i].invc; + logc = T[i].logc; + z = (double_t)asfloat(iz); - /* reduce x into [sqrt(2)/2, sqrt(2)] */ - ix += 0x3f800000 - 0x3f3504f3; - k += (int)(ix>>23) - 0x7f; - ix = (ix&0x007fffff) + 0x3f3504f3; - u.i = ix; - x = u.f; + /* log(x) = log1p(z/c-1) + log(c) + k*Ln2 */ + r = z * invc - 1; + y0 = logc + (double_t)k * Ln2; - f = x - 1.0f; - s = f/(2.0f + f); - z = s*s; - w = z*z; - t1= w*(Lg2+w*Lg4); - t2= z*(Lg1+w*Lg3); - R = t2 + t1; - hfsq = 0.5f*f*f; - dk = k; - return s*(hfsq+R) + dk*ln2_lo - hfsq + f + dk*ln2_hi; + /* Pipelined polynomial evaluation to approximate log1p(r). */ + r2 = r * r; + y = A[1] * r + A[2]; + y = A[0] * r2 + y; + y = y * r2 + (y0 + r); + return eval_as_float(y); } -- cgit v1.2.1