xref: /illumos-gate/usr/src/lib/libm/common/C/__rem_pio2m.c (revision d48be21240dfd051b689384ce2b23479d757f2d8)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright 2011 Nexenta Systems, Inc.  All rights reserved.
23  */
24 /*
25  * Copyright 2005 Sun Microsystems, Inc.  All rights reserved.
26  * Use is subject to license terms.
27  */
28 
29 /*
30  * int __rem_pio2m(x,y,e0,nx,prec,ipio2)
31  * double x[],y[]; int e0,nx,prec; const int ipio2[];
32  *
33  * __rem_pio2m return the last three digits of N with
34  *		y = x - N*pi/2
35  * so that |y| < pi/4.
36  *
37  * The method is to compute the integer (mod 8) and fraction parts of
38  * (2/pi)*x without doing the full multiplication. In general we
39  * skip the part of the product that are known to be a huge integer (
40  * more accurately, = 0 mod 8 ). Thus the number of operations are
41  * independent of the exponent of the input.
42  *
43  * (2/PI) is represented by an array of 24-bit integers in ipio2[].
44  * Here PI could as well be a machine value pi.
45  *
46  * Input parameters:
47  *	x[]	The input value (must be positive) is broken into nx
48  *		pieces of 24-bit integers in double precision format.
49  *		x[i] will be the i-th 24 bit of x. The scaled exponent
50  *		of x[0] is given in input parameter e0 (i.e., x[0]*2^e0
51  *		match x's up to 24 bits.
52  *
53  *		Example of breaking a double z into x[0]+x[1]+x[2]:
54  *			e0 = ilogb(z)-23
55  *			z  = scalbn(z,-e0)
56  *		for i = 0,1,2
57  *			x[i] =  floor(z)
58  *			z    = (z-x[i])*2**24
59  *
60  *
61  *	y[]	ouput result in an array of double precision numbers.
62  *		The dimension of y[] is:
63  *			24-bit  precision	1
64  *			53-bit  precision	2
65  *			64-bit  precision	2
66  *			113-bit precision	3
67  *		The actual value is the sum of them. Thus for 113-bit
68  *		precsion, one may have to do something like:
69  *
70  *		long double t,w,r_head, r_tail;
71  *		t = (long double)y[2] + (long double)y[1];
72  *		w = (long double)y[0];
73  *		r_head = t+w;
74  *		r_tail = w - (r_head - t);
75  *
76  *	e0	The exponent of x[0]
77  *
78  *	nx	dimension of x[]
79  *
80  *	prec	an interger indicating the precision:
81  *			0	24  bits (single)
82  *			1	53  bits (double)
83  *			2	64  bits (extended)
84  *			3	113 bits (quad)
85  *
86  *	ipio2[]
87  *		integer array, contains the (24*i)-th to (24*i+23)-th
88  *		bit of 2/pi or 2/PI after binary point. The corresponding
89  *		floating value is
90  *
91  *			ipio2[i] * 2^(-24(i+1)).
92  *
93  * External function:
94  *	double scalbn( ), floor( );
95  *
96  *
97  * Here is the description of some local variables:
98  *
99  *	jk	jk+1 is the initial number of terms of ipio2[] needed
100  *		in the computation. The recommended value is 3,4,4,
101  *		6 for single, double, extended,and quad.
102  *
103  *	jz	local integer variable indicating the number of
104  *		terms of ipio2[] used.
105  *
106  *	jx	nx - 1
107  *
108  *	jv	index for pointing to the suitable ipio2[] for the
109  *		computation. In general, we want
110  *			( 2^e0*x[0] * ipio2[jv-1]*2^(-24jv) )/8
111  *		is an integer. Thus
112  *			e0-3-24*jv >= 0 or (e0-3)/24 >= jv
113  *		Hence jv = max(0,(e0-3)/24).
114  *
115  *	jp	jp+1 is the number of terms in pio2[] needed, jp = jk.
116  *
117  *	q[]	double array with integral value, representing the
118  *		24-bits chunk of the product of x and 2/pi.
119  *
120  *	q0	the corresponding exponent of q[0]. Note that the
121  *		exponent for q[i] would be q0-24*i.
122  *
123  *	pio2[]	double precision array, obtained by cutting pi/2
124  *		into 24 bits chunks.
125  *
126  *	f[]	ipio2[] in floating point
127  *
128  *	iq[]	integer array by breaking up q[] in 24-bits chunk.
129  *
130  *	fq[]	final product of x*(2/pi) in fq[0],..,fq[jk]
131  *
132  *	ih	integer. If >0 it indicats q[] is >= 0.5, hence
133  *		it also indicates the *sign* of the result.
134  *
135  */
136 
137 #include <assert.h>
138 #include "libm.h"
139 
140 #if defined(__i386) && !defined(__amd64)
141 extern int __swapRP(int);
142 #endif
143 
144 static const int init_jk[] = { 3, 4, 4, 6 }; /* initial value for jk */
145 
146 static const double pio2[] = {
147 	1.57079625129699707031e+00,
148 	7.54978941586159635335e-08,
149 	5.39030252995776476554e-15,
150 	3.28200341580791294123e-22,
151 	1.27065575308067607349e-29,
152 	1.22933308981111328932e-36,
153 	2.73370053816464559624e-44,
154 	2.16741683877804819444e-51,
155 };
156 
157 static const double
158 	zero	= 0.0,
159 	one	= 1.0,
160 	half	= 0.5,
161 	eight	= 8.0,
162 	eighth	= 0.125,
163 	two24	= 16777216.0,
164 	twon24	= 5.960464477539062500E-8;
165 
166 int
167 __rem_pio2m(double *x, double *y, int e0, int nx, int prec, const int *ipio2)
168 {
169 	int	jz, jx, jv, jp, jk, carry, n, iq[20];
170 	int	i, j, k, m, q0, ih;
171 	double	z, fw, f[20], fq[20], q[20];
172 #if defined(__i386) && !defined(__amd64)
173 	int	rp;
174 
175 	rp = __swapRP(fp_extended);
176 #endif
177 
178 	fq[0] = NAN;	/* Make gcc happy */
179 	/* initialize jk */
180 	jp = jk = init_jk[prec];
181 
182 	/* determine jx,jv,q0, note that 3>q0 */
183 	jx = nx - 1;
184 	jv = (e0 - 3) / 24;
185 	if (jv < 0)
186 		jv = 0;
187 	q0 = e0 - 24 * (jv + 1);
188 
189 	/* set up f[0] to f[jx+jk] where f[jx+jk] = ipio2[jv+jk] */
190 	j = jv - jx;
191 	m = jx + jk;
192 	for (i = 0; i <= m; i++, j++)
193 		f[i] = (j < 0)? zero : (double)ipio2[j];
194 
195 	/* compute q[0],q[1],...q[jk] */
196 	for (i = 0; i <= jk; i++) {
197 		for (j = 0, fw = zero; j <= jx; j++)
198 			fw += x[j] * f[jx+i-j];
199 		q[i] = fw;
200 	}
201 
202 	jz = jk;
203 recompute:
204 	/* distill q[] into iq[] reversingly */
205 	for (i = 0, j = jz, z = q[jz]; j > 0; i++, j--) {
206 		fw = (double)((int)(twon24 * z));
207 		iq[i] = (int)(z - two24 * fw);
208 		z = q[j-1] + fw;
209 	}
210 
211 	/* compute n */
212 	z = scalbn(z, q0);		/* actual value of z */
213 	z -= eight * floor(z * eighth);	/* trim off integer >= 8 */
214 	n = (int)z;
215 	z -= (double)n;
216 	ih = 0;
217 	if (q0 > 0) {			/* need iq[jz-1] to determine n */
218 		i = (iq[jz-1] >> (24 - q0));
219 		n += i;
220 		iq[jz-1] -= i << (24 - q0);
221 		ih = iq[jz-1] >> (23 - q0);
222 	} else if (q0 == 0) {
223 		ih = iq[jz-1] >> 23;
224 	} else if (z >= half) {
225 		ih = 2;
226 	}
227 
228 	if (ih > 0) {	/* q > 0.5 */
229 		n += 1;
230 		carry = 0;
231 		for (i = 0; i < jz; i++) {	/* compute 1-q */
232 			j = iq[i];
233 			if (carry == 0) {
234 				if (j != 0) {
235 					carry = 1;
236 					iq[i] = 0x1000000 - j;
237 				}
238 			} else {
239 				iq[i] = 0xffffff - j;
240 			}
241 		}
242 		if (q0 > 0) {		/* rare case: chance is 1 in 12 */
243 			switch (q0) {
244 			case 1:
245 				iq[jz-1] &= 0x7fffff;
246 				break;
247 			case 2:
248 				iq[jz-1] &= 0x3fffff;
249 				break;
250 			}
251 		}
252 		if (ih == 2) {
253 			z = one - z;
254 			if (carry != 0)
255 				z -= scalbn(one, q0);
256 		}
257 	}
258 
259 	/* check if recomputation is needed */
260 	if (z == zero) {
261 		j = 0;
262 		for (i = jz - 1; i >= jk; i--)
263 			j |= iq[i];
264 		if (j == 0) {	/* need recomputation */
265 			/* set k to no. of terms needed */
266 			for (k = 1; iq[jk-k] == 0; k++)
267 				;
268 
269 			/* add q[jz+1] to q[jz+k] */
270 			for (i = jz + 1; i <= jz + k; i++) {
271 				f[jx+i] = (double)ipio2[jv+i];
272 				for (j = 0, fw = zero; j <= jx; j++)
273 					fw += x[j] * f[jx+i-j];
274 				q[i] = fw;
275 			}
276 			jz += k;
277 			goto recompute;
278 		}
279 	}
280 
281 	/* cut out zero terms */
282 	if (z == zero) {
283 		jz -= 1;
284 		q0 -= 24;
285 		while (iq[jz] == 0) {
286 			jz--;
287 			q0 -= 24;
288 		}
289 	} else {		/* break z into 24-bit if neccessary */
290 		z = scalbn(z, -q0);
291 		if (z >= two24) {
292 			fw = (double)((int)(twon24 * z));
293 			iq[jz] = (int)(z - two24 * fw);
294 			jz += 1;
295 			q0 += 24;
296 			iq[jz] = (int)fw;
297 		} else {
298 			iq[jz] = (int)z;
299 		}
300 	}
301 
302 	/* convert integer "bit" chunk to floating-point value */
303 	fw = scalbn(one, q0);
304 	for (i = jz; i >= 0; i--) {
305 		q[i] = fw * (double)iq[i];
306 		fw *= twon24;
307 	}
308 
309 	/* compute pio2[0,...,jp]*q[jz,...,0] */
310 	for (i = jz; i >= 0; i--) {
311 		for (fw = zero, k = 0; k <= jp && k <= jz - i; k++)
312 			fw += pio2[k] * q[i+k];
313 		fq[jz-i] = fw;
314 	}
315 
316 	/* compress fq[] into y[] */
317 	switch (prec) {
318 	case 0:
319 		fw = zero;
320 		for (i = jz; i >= 0; i--)
321 			fw += fq[i];
322 		y[0] = (ih == 0)? fw : -fw;
323 		break;
324 
325 	case 1:
326 	case 2:
327 		fw = zero;
328 		for (i = jz; i >= 0; i--)
329 			fw += fq[i];
330 		y[0] = (ih == 0)? fw : -fw;
331 
332 		assert(!isnan(fq[0]));
333 		fw = fq[0] - fw;
334 		for (i = 1; i <= jz; i++)
335 			fw += fq[i];
336 		y[1] = (ih == 0)? fw : -fw;
337 		break;
338 
339 	default:
340 		for (i = jz; i > 0; i--) {
341 			fw = fq[i-1] + fq[i];
342 			fq[i] += fq[i-1] - fw;
343 			fq[i-1] = fw;
344 		}
345 		for (i = jz; i > 1; i--) {
346 			fw = fq[i-1] + fq[i];
347 			fq[i] += fq[i-1] - fw;
348 			fq[i-1] = fw;
349 		}
350 		for (fw = zero, i = jz; i >= 2; i--)
351 			fw += fq[i];
352 		if (ih == 0) {
353 			y[0] = fq[0];
354 			y[1] = fq[1];
355 			y[2] = fw;
356 		} else {
357 			y[0] = -fq[0];
358 			y[1] = -fq[1];
359 			y[2] = -fw;
360 		}
361 	}
362 
363 #if defined(__i386) && !defined(__amd64)
364 	(void) __swapRP(rp);
365 #endif
366 	return (n & 7);
367 }
368