k_rem_pio2.c 10.0 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358
  1. /* @(#)k_rem_pio2.c 5.1 93/09/24 */
  2. /*
  3. * ====================================================
  4. * Copyright (C) 1993 by Sun Microsystems, Inc. All rights reserved.
  5. *
  6. * Developed at SunPro, a Sun Microsystems, Inc. business.
  7. * Permission to use, copy, modify, and distribute this
  8. * software is freely granted, provided that this notice
  9. * is preserved.
  10. * ====================================================
  11. */
  12. #if defined(LIBM_SCCS) && !defined(lint)
  13. static const char rcsid[] =
  14. "$NetBSD: k_rem_pio2.c,v 1.7 1995/05/10 20:46:25 jtc Exp $";
  15. #endif
  16. /*
  17. * __kernel_rem_pio2(x,y,e0,nx,prec,ipio2)
  18. * double x[],y[]; int e0,nx,prec; int ipio2[];
  19. *
  20. * __kernel_rem_pio2 return the last three digits of N with
  21. * y = x - N*pi/2
  22. * so that |y| < pi/2.
  23. *
  24. * The method is to compute the integer (mod 8) and fraction parts of
  25. * (2/pi)*x without doing the full multiplication. In general we
  26. * skip the part of the product that are known to be a huge integer (
  27. * more accurately, = 0 mod 8 ). Thus the number of operations are
  28. * independent of the exponent of the input.
  29. *
  30. * (2/pi) is represented by an array of 24-bit integers in ipio2[].
  31. *
  32. * Input parameters:
  33. * x[] The input value (must be positive) is broken into nx
  34. * pieces of 24-bit integers in double precision format.
  35. * x[i] will be the i-th 24 bit of x. The scaled exponent
  36. * of x[0] is given in input parameter e0 (i.e., x[0]*2^e0
  37. * match x's up to 24 bits.
  38. *
  39. * Example of breaking a double positive z into x[0]+x[1]+x[2]:
  40. * e0 = ilogb(z)-23
  41. * z = scalbn(z,-e0)
  42. * for i = 0,1,2
  43. * x[i] = floor(z)
  44. * z = (z-x[i])*2**24
  45. *
  46. *
  47. * y[] ouput result in an array of double precision numbers.
  48. * The dimension of y[] is:
  49. * 24-bit precision 1
  50. * 53-bit precision 2
  51. * 64-bit precision 2
  52. * 113-bit precision 3
  53. * The actual value is the sum of them. Thus for 113-bit
  54. * precison, one may have to do something like:
  55. *
  56. * long double t,w,r_head, r_tail;
  57. * t = (long double)y[2] + (long double)y[1];
  58. * w = (long double)y[0];
  59. * r_head = t+w;
  60. * r_tail = w - (r_head - t);
  61. *
  62. * e0 The exponent of x[0]
  63. *
  64. * nx dimension of x[]
  65. *
  66. * prec an integer indicating the precision:
  67. * 0 24 bits (single)
  68. * 1 53 bits (double)
  69. * 2 64 bits (extended)
  70. * 3 113 bits (quad)
  71. *
  72. * ipio2[]
  73. * integer array, contains the (24*i)-th to (24*i+23)-th
  74. * bit of 2/pi after binary point. The corresponding
  75. * floating value is
  76. *
  77. * ipio2[i] * 2^(-24(i+1)).
  78. *
  79. * External function:
  80. * double scalbn(), floor();
  81. *
  82. *
  83. * Here is the description of some local variables:
  84. *
  85. * jk jk+1 is the initial number of terms of ipio2[] needed
  86. * in the computation. The recommended value is 2,3,4,
  87. * 6 for single, double, extended,and quad.
  88. *
  89. * jz local integer variable indicating the number of
  90. * terms of ipio2[] used.
  91. *
  92. * jx nx - 1
  93. *
  94. * jv index for pointing to the suitable ipio2[] for the
  95. * computation. In general, we want
  96. * ( 2^e0*x[0] * ipio2[jv-1]*2^(-24jv) )/8
  97. * is an integer. Thus
  98. * e0-3-24*jv >= 0 or (e0-3)/24 >= jv
  99. * Hence jv = max(0,(e0-3)/24).
  100. *
  101. * jp jp+1 is the number of terms in PIo2[] needed, jp = jk.
  102. *
  103. * q[] double array with integral value, representing the
  104. * 24-bits chunk of the product of x and 2/pi.
  105. *
  106. * q0 the corresponding exponent of q[0]. Note that the
  107. * exponent for q[i] would be q0-24*i.
  108. *
  109. * PIo2[] double precision array, obtained by cutting pi/2
  110. * into 24 bits chunks.
  111. *
  112. * f[] ipio2[] in floating point
  113. *
  114. * iq[] integer array by breaking up q[] in 24-bits chunk.
  115. *
  116. * fq[] final product of x*(2/pi) in fq[0],..,fq[jk]
  117. *
  118. * ih integer. If >0 it indicates q[] is >= 0.5, hence
  119. * it also indicates the *sign* of the result.
  120. *
  121. */
  122. /*
  123. * Constants:
  124. * The hexadecimal values are the intended ones for the following
  125. * constants. The decimal values may be used, provided that the
  126. * compiler will convert from decimal to binary accurately enough
  127. * to produce the hexadecimal values shown.
  128. */
  129. #include "math_libm.h"
  130. #include "math_private.h"
  131. libm_hidden_proto(scalbn)
  132. libm_hidden_proto(floor)
  133. #ifdef __STDC__
  134. static const int init_jk[] = { 2, 3, 4, 6 }; /* initial value for jk */
  135. #else
  136. static int init_jk[] = { 2, 3, 4, 6 };
  137. #endif
  138. #ifdef __STDC__
  139. static const double PIo2[] = {
  140. #else
  141. static double PIo2[] = {
  142. #endif
  143. 1.57079625129699707031e+00, /* 0x3FF921FB, 0x40000000 */
  144. 7.54978941586159635335e-08, /* 0x3E74442D, 0x00000000 */
  145. 5.39030252995776476554e-15, /* 0x3CF84698, 0x80000000 */
  146. 3.28200341580791294123e-22, /* 0x3B78CC51, 0x60000000 */
  147. 1.27065575308067607349e-29, /* 0x39F01B83, 0x80000000 */
  148. 1.22933308981111328932e-36, /* 0x387A2520, 0x40000000 */
  149. 2.73370053816464559624e-44, /* 0x36E38222, 0x80000000 */
  150. 2.16741683877804819444e-51, /* 0x3569F31D, 0x00000000 */
  151. };
  152. #ifdef __STDC__
  153. static const double
  154. #else
  155. static double
  156. #endif
  157. zero = 0.0, one = 1.0, two24 = 1.67772160000000000000e+07, /* 0x41700000, 0x00000000 */
  158. twon24 = 5.96046447753906250000e-08; /* 0x3E700000, 0x00000000 */
  159. #ifdef __STDC__
  160. int attribute_hidden
  161. __kernel_rem_pio2(double *x, double *y, int e0, int nx, int prec,
  162. const int32_t * ipio2)
  163. #else
  164. int attribute_hidden
  165. __kernel_rem_pio2(x, y, e0, nx, prec, ipio2)
  166. double x[], y[];
  167. int e0, nx, prec;
  168. int32_t ipio2[];
  169. #endif
  170. {
  171. int32_t jz, jx, jv, jp, jk, carry, n, iq[20], i, j, k, m, q0, ih;
  172. double z, fw, f[20], fq[20], q[20];
  173. /* initialize jk */
  174. jk = init_jk[prec];
  175. jp = jk;
  176. /* determine jx,jv,q0, note that 3>q0 */
  177. jx = nx - 1;
  178. jv = (e0 - 3) / 24;
  179. if (jv < 0)
  180. jv = 0;
  181. q0 = e0 - 24 * (jv + 1);
  182. /* set up f[0] to f[jx+jk] where f[jx+jk] = ipio2[jv+jk] */
  183. j = jv - jx;
  184. m = jx + jk;
  185. for (i = 0; i <= m; i++, j++)
  186. f[i] = (j < 0) ? zero : (double) ipio2[j];
  187. /* compute q[0],q[1],...q[jk] */
  188. for (i = 0; i <= jk; i++) {
  189. for (j = 0, fw = 0.0; j <= jx; j++)
  190. fw += x[j] * f[jx + i - j];
  191. q[i] = fw;
  192. }
  193. jz = jk;
  194. recompute:
  195. /* distill q[] into iq[] reversingly */
  196. for (i = 0, j = jz, z = q[jz]; j > 0; i++, j--) {
  197. fw = (double) ((int32_t) (twon24 * z));
  198. iq[i] = (int32_t) (z - two24 * fw);
  199. z = q[j - 1] + fw;
  200. }
  201. /* compute n */
  202. z = scalbn(z, q0); /* actual value of z */
  203. z -= 8.0 * floor(z * 0.125); /* trim off integer >= 8 */
  204. n = (int32_t) z;
  205. z -= (double) n;
  206. ih = 0;
  207. if (q0 > 0) { /* need iq[jz-1] to determine n */
  208. i = (iq[jz - 1] >> (24 - q0));
  209. n += i;
  210. iq[jz - 1] -= i << (24 - q0);
  211. ih = iq[jz - 1] >> (23 - q0);
  212. } else if (q0 == 0)
  213. ih = iq[jz - 1] >> 23;
  214. else if (z >= 0.5)
  215. ih = 2;
  216. if (ih > 0) { /* q > 0.5 */
  217. n += 1;
  218. carry = 0;
  219. for (i = 0; i < jz; i++) { /* compute 1-q */
  220. j = iq[i];
  221. if (carry == 0) {
  222. if (j != 0) {
  223. carry = 1;
  224. iq[i] = 0x1000000 - j;
  225. }
  226. } else
  227. iq[i] = 0xffffff - j;
  228. }
  229. if (q0 > 0) { /* rare case: chance is 1 in 12 */
  230. switch (q0) {
  231. case 1:
  232. iq[jz - 1] &= 0x7fffff;
  233. break;
  234. case 2:
  235. iq[jz - 1] &= 0x3fffff;
  236. break;
  237. }
  238. }
  239. if (ih == 2) {
  240. z = one - z;
  241. if (carry != 0)
  242. z -= scalbn(one, q0);
  243. }
  244. }
  245. /* check if recomputation is needed */
  246. if (z == zero) {
  247. j = 0;
  248. for (i = jz - 1; i >= jk; i--)
  249. j |= iq[i];
  250. if (j == 0) { /* need recomputation */
  251. for (k = 1; iq[jk - k] == 0; k++); /* k = no. of terms needed */
  252. for (i = jz + 1; i <= jz + k; i++) { /* add q[jz+1] to q[jz+k] */
  253. f[jx + i] = (double) ipio2[jv + i];
  254. for (j = 0, fw = 0.0; j <= jx; j++)
  255. fw += x[j] * f[jx + i - j];
  256. q[i] = fw;
  257. }
  258. jz += k;
  259. goto recompute;
  260. }
  261. }
  262. /* chop off zero terms */
  263. if (z == 0.0) {
  264. jz -= 1;
  265. q0 -= 24;
  266. while (iq[jz] == 0) {
  267. jz--;
  268. q0 -= 24;
  269. }
  270. } else { /* break z into 24-bit if necessary */
  271. z = scalbn(z, -q0);
  272. if (z >= two24) {
  273. fw = (double) ((int32_t) (twon24 * z));
  274. iq[jz] = (int32_t) (z - two24 * fw);
  275. jz += 1;
  276. q0 += 24;
  277. iq[jz] = (int32_t) fw;
  278. } else
  279. iq[jz] = (int32_t) z;
  280. }
  281. /* convert integer "bit" chunk to floating-point value */
  282. fw = scalbn(one, q0);
  283. for (i = jz; i >= 0; i--) {
  284. q[i] = fw * (double) iq[i];
  285. fw *= twon24;
  286. }
  287. /* compute PIo2[0,...,jp]*q[jz,...,0] */
  288. for (i = jz; i >= 0; i--) {
  289. for (fw = 0.0, k = 0; k <= jp && k <= jz - i; k++)
  290. fw += PIo2[k] * q[i + k];
  291. fq[jz - i] = fw;
  292. }
  293. /* compress fq[] into y[] */
  294. switch (prec) {
  295. case 0:
  296. fw = 0.0;
  297. for (i = jz; i >= 0; i--)
  298. fw += fq[i];
  299. y[0] = (ih == 0) ? fw : -fw;
  300. break;
  301. case 1:
  302. case 2:
  303. fw = 0.0;
  304. for (i = jz; i >= 0; i--)
  305. fw += fq[i];
  306. y[0] = (ih == 0) ? fw : -fw;
  307. fw = fq[0] - fw;
  308. for (i = 1; i <= jz; i++)
  309. fw += fq[i];
  310. y[1] = (ih == 0) ? fw : -fw;
  311. break;
  312. case 3: /* painful */
  313. for (i = jz; i > 0; i--) {
  314. fw = fq[i - 1] + fq[i];
  315. fq[i] += fq[i - 1] - fw;
  316. fq[i - 1] = fw;
  317. }
  318. for (i = jz; i > 1; i--) {
  319. fw = fq[i - 1] + fq[i];
  320. fq[i] += fq[i - 1] - fw;
  321. fq[i - 1] = fw;
  322. }
  323. for (fw = 0.0, i = jz; i >= 2; i--)
  324. fw += fq[i];
  325. if (ih == 0) {
  326. y[0] = fq[0];
  327. y[1] = fq[1];
  328. y[2] = fw;
  329. } else {
  330. y[0] = -fq[0];
  331. y[1] = -fq[1];
  332. y[2] = -fw;
  333. }
  334. }
  335. return n & 7;
  336. }