diff options
Diffstat (limited to 'sysdeps/ieee754/dbl-64/e_log.c')
-rw-r--r-- | sysdeps/ieee754/dbl-64/e_log.c | 127 |
1 files changed, 16 insertions, 111 deletions
diff --git a/sysdeps/ieee754/dbl-64/e_log.c b/sysdeps/ieee754/dbl-64/e_log.c index 6a18ebb904..2483dd8551 100644 --- a/sysdeps/ieee754/dbl-64/e_log.c +++ b/sysdeps/ieee754/dbl-64/e_log.c @@ -23,11 +23,10 @@ /* FUNCTION:ulog */ /* */ /* FILES NEEDED: dla.h endian.h mpa.h mydefs.h ulog.h */ -/* mpexp.c mplog.c mpa.c */ /* ulog.tbl */ /* */ /* An ultimate log routine. Given an IEEE double machine number x */ -/* it computes the correctly rounded (to nearest) value of log(x). */ +/* it computes the rounded (to nearest) value of log(x). */ /* Assumption: Machine arithmetic operations are performed in */ /* round to nearest mode of IEEE 754 standard. */ /* */ @@ -40,34 +39,26 @@ #include "MathLib.h" #include <math.h> #include <math_private.h> -#include <stap-probe.h> #ifndef SECTION # define SECTION #endif -void __mplog (mp_no *, mp_no *, int); - /*********************************************************************/ -/* An ultimate log routine. Given an IEEE double machine number x */ -/* it computes the correctly rounded (to nearest) value of log(x). */ +/* An ultimate log routine. Given an IEEE double machine number x */ +/* it computes the rounded (to nearest) value of log(x). */ /*********************************************************************/ double SECTION __ieee754_log (double x) { -#define M 4 - static const int pr[M] = { 8, 10, 18, 32 }; - int i, j, n, ux, dx, p; + int i, j, n, ux, dx; double dbl_n, u, p0, q, r0, w, nln2a, luai, lubi, lvaj, lvbj, - sij, ssij, ttij, A, B, B0, y, y1, y2, polI, polII, sa, sb, - t1, t2, t7, t8, t, ra, rb, ww, - a0, aa0, s1, s2, ss2, s3, ss3, a1, aa1, a, aa, b, bb, c; + sij, ssij, ttij, A, B, B0, polI, polII, t8, a, aa, b, bb, c; #ifndef DLA_FMS - double t3, t4, t5, t6; + double t1, t2, t3, t4, t5; #endif number num; - mp_no mpx, mpy, mpy1, mpy2, mperr; #include "ulog.tbl" #include "ulog.h" @@ -101,7 +92,7 @@ __ieee754_log (double x) if (w == 0.0) return 0.0; - /*--- Stage I, the case abs(x-1) < 0.03 */ + /*--- The case abs(x-1) < 0.03 */ t8 = MHALF * w; EMULV (t8, w, a, aa, t1, t2, t3, t4, t5); @@ -118,50 +109,12 @@ __ieee754_log (double x) polII *= w * w * w; c = (aa + bb) + polII; - /* End stage I, case abs(x-1) < 0.03 */ - if ((y = b + (c + b * E2)) == b + (c - b * E2)) - return y; - - /*--- Stage II, the case abs(x-1) < 0.03 */ - - a = d19.d + w * d20.d; - a = d18.d + w * a; - a = d17.d + w * a; - a = d16.d + w * a; - a = d15.d + w * a; - a = d14.d + w * a; - a = d13.d + w * a; - a = d12.d + w * a; - a = d11.d + w * a; - - EMULV (w, a, s2, ss2, t1, t2, t3, t4, t5); - ADD2 (d10.d, dd10.d, s2, ss2, s3, ss3, t1, t2); - MUL2 (w, 0, s3, ss3, s2, ss2, t1, t2, t3, t4, t5, t6, t7, t8); - ADD2 (d9.d, dd9.d, s2, ss2, s3, ss3, t1, t2); - MUL2 (w, 0, s3, ss3, s2, ss2, t1, t2, t3, t4, t5, t6, t7, t8); - ADD2 (d8.d, dd8.d, s2, ss2, s3, ss3, t1, t2); - MUL2 (w, 0, s3, ss3, s2, ss2, t1, t2, t3, t4, t5, t6, t7, t8); - ADD2 (d7.d, dd7.d, s2, ss2, s3, ss3, t1, t2); - MUL2 (w, 0, s3, ss3, s2, ss2, t1, t2, t3, t4, t5, t6, t7, t8); - ADD2 (d6.d, dd6.d, s2, ss2, s3, ss3, t1, t2); - MUL2 (w, 0, s3, ss3, s2, ss2, t1, t2, t3, t4, t5, t6, t7, t8); - ADD2 (d5.d, dd5.d, s2, ss2, s3, ss3, t1, t2); - MUL2 (w, 0, s3, ss3, s2, ss2, t1, t2, t3, t4, t5, t6, t7, t8); - ADD2 (d4.d, dd4.d, s2, ss2, s3, ss3, t1, t2); - MUL2 (w, 0, s3, ss3, s2, ss2, t1, t2, t3, t4, t5, t6, t7, t8); - ADD2 (d3.d, dd3.d, s2, ss2, s3, ss3, t1, t2); - MUL2 (w, 0, s3, ss3, s2, ss2, t1, t2, t3, t4, t5, t6, t7, t8); - ADD2 (d2.d, dd2.d, s2, ss2, s3, ss3, t1, t2); - MUL2 (w, 0, s3, ss3, s2, ss2, t1, t2, t3, t4, t5, t6, t7, t8); - MUL2 (w, 0, s2, ss2, s3, ss3, t1, t2, t3, t4, t5, t6, t7, t8); - ADD2 (w, 0, s3, ss3, b, bb, t1, t2); + /* Here b contains the high part of the result, and c the low part. + Maximum error is b * 2.334e-19, so accuracy is >61 bits. + Therefore max ULP error of b + c is ~0.502. */ + return b + c; - /* End stage II, case abs(x-1) < 0.03 */ - if ((y = b + (bb + b * E4)) == b + (bb - b * E4)) - return y; - goto stage_n; - - /*--- Stage I, the case abs(x-1) > 0.03 */ + /*--- The case abs(x-1) > 0.03 */ case_03: /* Find n,u such that x = u*2**n, 1/sqrt(2) < u < sqrt(2) */ @@ -203,58 +156,10 @@ case_03: B0 = (((lubi + lvbj) + ssij) + ttij) + dbl_n * LN2B; B = polI + B0; - /* End stage I, case abs(x-1) >= 0.03 */ - if ((y = A + (B + E1)) == A + (B - E1)) - return y; - - - /*--- Stage II, the case abs(x-1) > 0.03 */ - - /* Improve the accuracy of r0 */ - EMULV (p0, r0, sa, sb, t1, t2, t3, t4, t5); - t = r0 * ((1 - sa) - sb); - EADD (r0, t, ra, rb); - - /* Compute w */ - MUL2 (q, 0, ra, rb, w, ww, t1, t2, t3, t4, t5, t6, t7, t8); - - EADD (A, B0, a0, aa0); - - /* Evaluate polynomial III */ - s1 = (c3.d + (c4.d + c5.d * w) * w) * w; - EADD (c2.d, s1, s2, ss2); - MUL2 (s2, ss2, w, ww, s3, ss3, t1, t2, t3, t4, t5, t6, t7, t8); - MUL2 (s3, ss3, w, ww, s2, ss2, t1, t2, t3, t4, t5, t6, t7, t8); - ADD2 (s2, ss2, w, ww, s3, ss3, t1, t2); - ADD2 (s3, ss3, a0, aa0, a1, aa1, t1, t2); - - /* End stage II, case abs(x-1) >= 0.03 */ - if ((y = a1 + (aa1 + E3)) == a1 + (aa1 - E3)) - return y; - - - /* Final stages. Use multi-precision arithmetic. */ -stage_n: - - for (i = 0; i < M; i++) - { - p = pr[i]; - __dbl_mp (x, &mpx, p); - __dbl_mp (y, &mpy, p); - __mplog (&mpx, &mpy, p); - __dbl_mp (e[i].d, &mperr, p); - __add (&mpy, &mperr, &mpy1, p); - __sub (&mpy, &mperr, &mpy2, p); - __mp_dbl (&mpy1, &y1, p); - __mp_dbl (&mpy2, &y2, p); - if (y1 == y2) - { - LIBC_PROBE (slowlog, 3, &p, &x, &y1); - return y1; - } - } - LIBC_PROBE (slowlog_inexact, 3, &p, &x, &y1); - return y1; + /* Here A contains the high part of the result, and B the low part. + Maximum abs error is 6.095e-21 and min log (x) is 0.0295 since x > 1.03. + Therefore max ULP error of A + B is ~0.502. */ + return A + B; } #ifndef __ieee754_log |