xref: /openbsd/lib/libm/src/e_sqrt.c (revision 5af055cd)
1 /* @(#)e_sqrt.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 
13 /* sqrt(x)
14  * Return correctly rounded sqrt.
15  *           ------------------------------------------
16  *	     |  Use the hardware sqrt if you have one |
17  *           ------------------------------------------
18  * Method:
19  *   Bit by bit method using integer arithmetic. (Slow, but portable)
20  *   1. Normalization
21  *	Scale x to y in [1,4) with even powers of 2:
22  *	find an integer k such that  1 <= (y=x*2^(2k)) < 4, then
23  *		sqrt(x) = 2^k * sqrt(y)
24  *   2. Bit by bit computation
25  *	Let q  = sqrt(y) truncated to i bit after binary point (q = 1),
26  *	     i							 0
27  *                                     i+1         2
28  *	    s  = 2*q , and	y  =  2   * ( y - q  ).		(1)
29  *	     i      i            i                 i
30  *
31  *	To compute q    from q , one checks whether
32  *		    i+1       i
33  *
34  *			      -(i+1) 2
35  *			(q + 2      ) <= y.			(2)
36  *     			  i
37  *							      -(i+1)
38  *	If (2) is false, then q   = q ; otherwise q   = q  + 2      .
39  *		 	       i+1   i             i+1   i
40  *
41  *	With some algebric manipulation, it is not difficult to see
42  *	that (2) is equivalent to
43  *                             -(i+1)
44  *			s  +  2       <= y			(3)
45  *			 i                i
46  *
47  *	The advantage of (3) is that s  and y  can be computed by
48  *				      i      i
49  *	the following recurrence formula:
50  *	    if (3) is false
51  *
52  *	    s     =  s  ,	y    = y   ;			(4)
53  *	     i+1      i		 i+1    i
54  *
55  *	    otherwise,
56  *                         -i                     -(i+1)
57  *	    s	  =  s  + 2  ,  y    = y  -  s  - 2  		(5)
58  *           i+1      i          i+1    i     i
59  *
60  *	One may easily use induction to prove (4) and (5).
61  *	Note. Since the left hand side of (3) contain only i+2 bits,
62  *	      it does not necessary to do a full (53-bit) comparison
63  *	      in (3).
64  *   3. Final rounding
65  *	After generating the 53 bits result, we compute one more bit.
66  *	Together with the remainder, we can decide whether the
67  *	result is exact, bigger than 1/2ulp, or less than 1/2ulp
68  *	(it will never equal to 1/2ulp).
69  *	The rounding mode can be detected by checking whether
70  *	huge + tiny is equal to huge, and whether huge - tiny is
71  *	equal to huge for some floating point number "huge" and "tiny".
72  *
73  * Special cases:
74  *	sqrt(+-0) = +-0 	... exact
75  *	sqrt(inf) = inf
76  *	sqrt(-ve) = NaN		... with invalid signal
77  *	sqrt(NaN) = NaN		... with invalid signal for signaling NaN
78  *
79  * Other methods : see the appended file at the end of the program below.
80  *---------------
81  */
82 
83 #include <float.h>
84 #include <math.h>
85 
86 #include "math_private.h"
87 
88 static	const double	one	= 1.0, tiny=1.0e-300;
89 
90 double
91 sqrt(double x)
92 {
93 	double z;
94 	int32_t sign = (int)0x80000000;
95 	int32_t ix0,s0,q,m,t,i;
96 	u_int32_t r,t1,s1,ix1,q1;
97 
98 	EXTRACT_WORDS(ix0,ix1,x);
99 
100     /* take care of Inf and NaN */
101 	if((ix0&0x7ff00000)==0x7ff00000) {
102 	    return x*x+x;		/* sqrt(NaN)=NaN, sqrt(+inf)=+inf
103 					   sqrt(-inf)=sNaN */
104 	}
105     /* take care of zero */
106 	if(ix0<=0) {
107 	    if(((ix0&(~sign))|ix1)==0) return x;/* sqrt(+-0) = +-0 */
108 	    else if(ix0<0)
109 		return (x-x)/(x-x);		/* sqrt(-ve) = sNaN */
110 	}
111     /* normalize x */
112 	m = (ix0>>20);
113 	if(m==0) {				/* subnormal x */
114 	    while(ix0==0) {
115 		m -= 21;
116 		ix0 |= (ix1>>11); ix1 <<= 21;
117 	    }
118 	    for(i=0;(ix0&0x00100000)==0;i++) ix0<<=1;
119 	    m -= i-1;
120 	    ix0 |= (ix1>>(32-i));
121 	    ix1 <<= i;
122 	}
123 	m -= 1023;	/* unbias exponent */
124 	ix0 = (ix0&0x000fffff)|0x00100000;
125 	if(m&1){	/* odd m, double x to make it even */
126 	    ix0 += ix0 + ((ix1&sign)>>31);
127 	    ix1 += ix1;
128 	}
129 	m >>= 1;	/* m = [m/2] */
130 
131     /* generate sqrt(x) bit by bit */
132 	ix0 += ix0 + ((ix1&sign)>>31);
133 	ix1 += ix1;
134 	q = q1 = s0 = s1 = 0;	/* [q,q1] = sqrt(x) */
135 	r = 0x00200000;		/* r = moving bit from right to left */
136 
137 	while(r!=0) {
138 	    t = s0+r;
139 	    if(t<=ix0) {
140 		s0   = t+r;
141 		ix0 -= t;
142 		q   += r;
143 	    }
144 	    ix0 += ix0 + ((ix1&sign)>>31);
145 	    ix1 += ix1;
146 	    r>>=1;
147 	}
148 
149 	r = sign;
150 	while(r!=0) {
151 	    t1 = s1+r;
152 	    t  = s0;
153 	    if((t<ix0)||((t==ix0)&&(t1<=ix1))) {
154 		s1  = t1+r;
155 		if(((t1&sign)==sign)&&(s1&sign)==0) s0 += 1;
156 		ix0 -= t;
157 		if (ix1 < t1) ix0 -= 1;
158 		ix1 -= t1;
159 		q1  += r;
160 	    }
161 	    ix0 += ix0 + ((ix1&sign)>>31);
162 	    ix1 += ix1;
163 	    r>>=1;
164 	}
165 
166     /* use floating add to find out rounding direction */
167 	if((ix0|ix1)!=0) {
168 	    z = one-tiny; /* trigger inexact flag */
169 	    if (z>=one) {
170 	        z = one+tiny;
171 	        if (q1==(u_int32_t)0xffffffff) { q1=0; q += 1;}
172 		else if (z>one) {
173 		    if (q1==(u_int32_t)0xfffffffe) q+=1;
174 		    q1+=2;
175 		} else
176 	            q1 += (q1&1);
177 	    }
178 	}
179 	ix0 = (q>>1)+0x3fe00000;
180 	ix1 =  q1>>1;
181 	if ((q&1)==1) ix1 |= sign;
182 	ix0 += (m <<20);
183 	INSERT_WORDS(z,ix0,ix1);
184 	return z;
185 }
186 
187 /*
188 Other methods  (use floating-point arithmetic)
189 -------------
190 (This is a copy of a drafted paper by Prof W. Kahan
191 and K.C. Ng, written in May, 1986)
192 
193 	Two algorithms are given here to implement sqrt(x)
194 	(IEEE double precision arithmetic) in software.
195 	Both supply sqrt(x) correctly rounded. The first algorithm (in
196 	Section A) uses newton iterations and involves four divisions.
197 	The second one uses reciproot iterations to avoid division, but
198 	requires more multiplications. Both algorithms need the ability
199 	to chop results of arithmetic operations instead of round them,
200 	and the INEXACT flag to indicate when an arithmetic operation
201 	is executed exactly with no roundoff error, all part of the
202 	standard (IEEE 754-1985). The ability to perform shift, add,
203 	subtract and logical AND operations upon 32-bit words is needed
204 	too, though not part of the standard.
205 
206 A.  sqrt(x) by Newton Iteration
207 
208    (1)	Initial approximation
209 
210 	Let x0 and x1 be the leading and the trailing 32-bit words of
211 	a floating point number x (in IEEE double format) respectively
212 
213 	    1    11		     52				  ...widths
214 	   ------------------------------------------------------
215 	x: |s|	  e     |	      f				|
216 	   ------------------------------------------------------
217 	      msb    lsb  msb				      lsb ...order
218 
219 
220 	     ------------------------  	     ------------------------
221 	x0:  |s|   e    |    f1     |	 x1: |          f2           |
222 	     ------------------------  	     ------------------------
223 
224 	By performing shifts and subtracts on x0 and x1 (both regarded
225 	as integers), we obtain an 8-bit approximation of sqrt(x) as
226 	follows.
227 
228 		k  := (x0>>1) + 0x1ff80000;
229 		y0 := k - T1[31&(k>>15)].	... y ~ sqrt(x) to 8 bits
230 	Here k is a 32-bit integer and T1[] is an integer array containing
231 	correction terms. Now magically the floating value of y (y's
232 	leading 32-bit word is y0, the value of its trailing word is 0)
233 	approximates sqrt(x) to almost 8-bit.
234 
235 	Value of T1:
236 	static int T1[32]= {
237 	0,	1024,	3062,	5746,	9193,	13348,	18162,	23592,
238 	29598,	36145,	43202,	50740,	58733,	67158,	75992,	85215,
239 	83599,	71378,	60428,	50647,	41945,	34246,	27478,	21581,
240 	16499,	12183,	8588,	5674,	3403,	1742,	661,	130,};
241 
242     (2)	Iterative refinement
243 
244 	Apply Heron's rule three times to y, we have y approximates
245 	sqrt(x) to within 1 ulp (Unit in the Last Place):
246 
247 		y := (y+x/y)/2		... almost 17 sig. bits
248 		y := (y+x/y)/2		... almost 35 sig. bits
249 		y := y-(y-x/y)/2	... within 1 ulp
250 
251 
252 	Remark 1.
253 	    Another way to improve y to within 1 ulp is:
254 
255 		y := (y+x/y)		... almost 17 sig. bits to 2*sqrt(x)
256 		y := y - 0x00100006	... almost 18 sig. bits to sqrt(x)
257 
258 				2
259 			    (x-y )*y
260 		y := y + 2* ----------	...within 1 ulp
261 			       2
262 			     3y  + x
263 
264 
265 	This formula has one division fewer than the one above; however,
266 	it requires more multiplications and additions. Also x must be
267 	scaled in advance to avoid spurious overflow in evaluating the
268 	expression 3y*y+x. Hence it is not recommended uless division
269 	is slow. If division is very slow, then one should use the
270 	reciproot algorithm given in section B.
271 
272     (3) Final adjustment
273 
274 	By twiddling y's last bit it is possible to force y to be
275 	correctly rounded according to the prevailing rounding mode
276 	as follows. Let r and i be copies of the rounding mode and
277 	inexact flag before entering the square root program. Also we
278 	use the expression y+-ulp for the next representable floating
279 	numbers (up and down) of y. Note that y+-ulp = either fixed
280 	point y+-1, or multiply y by nextafter(1,+-inf) in chopped
281 	mode.
282 
283 		I := FALSE;	... reset INEXACT flag I
284 		R := RZ;	... set rounding mode to round-toward-zero
285 		z := x/y;	... chopped quotient, possibly inexact
286 		If(not I) then {	... if the quotient is exact
287 		    if(z=y) {
288 		        I := i;	 ... restore inexact flag
289 		        R := r;  ... restore rounded mode
290 		        return sqrt(x):=y.
291 		    } else {
292 			z := z - ulp;	... special rounding
293 		    }
294 		}
295 		i := TRUE;		... sqrt(x) is inexact
296 		If (r=RN) then z=z+ulp	... rounded-to-nearest
297 		If (r=RP) then {	... round-toward-+inf
298 		    y = y+ulp; z=z+ulp;
299 		}
300 		y := y+z;		... chopped sum
301 		y0:=y0-0x00100000;	... y := y/2 is correctly rounded.
302 	        I := i;	 		... restore inexact flag
303 	        R := r;  		... restore rounded mode
304 	        return sqrt(x):=y.
305 
306     (4)	Special cases
307 
308 	Square root of +inf, +-0, or NaN is itself;
309 	Square root of a negative number is NaN with invalid signal.
310 
311 
312 B.  sqrt(x) by Reciproot Iteration
313 
314    (1)	Initial approximation
315 
316 	Let x0 and x1 be the leading and the trailing 32-bit words of
317 	a floating point number x (in IEEE double format) respectively
318 	(see section A). By performing shifs and subtracts on x0 and y0,
319 	we obtain a 7.8-bit approximation of 1/sqrt(x) as follows.
320 
321 	    k := 0x5fe80000 - (x0>>1);
322 	    y0:= k - T2[63&(k>>14)].	... y ~ 1/sqrt(x) to 7.8 bits
323 
324 	Here k is a 32-bit integer and T2[] is an integer array
325 	containing correction terms. Now magically the floating
326 	value of y (y's leading 32-bit word is y0, the value of
327 	its trailing word y1 is set to zero) approximates 1/sqrt(x)
328 	to almost 7.8-bit.
329 
330 	Value of T2:
331 	static int T2[64]= {
332 	0x1500,	0x2ef8,	0x4d67,	0x6b02,	0x87be,	0xa395,	0xbe7a,	0xd866,
333 	0xf14a,	0x1091b,0x11fcd,0x13552,0x14999,0x15c98,0x16e34,0x17e5f,
334 	0x18d03,0x19a01,0x1a545,0x1ae8a,0x1b5c4,0x1bb01,0x1bfde,0x1c28d,
335 	0x1c2de,0x1c0db,0x1ba73,0x1b11c,0x1a4b5,0x1953d,0x18266,0x16be0,
336 	0x1683e,0x179d8,0x18a4d,0x19992,0x1a789,0x1b445,0x1bf61,0x1c989,
337 	0x1d16d,0x1d77b,0x1dddf,0x1e2ad,0x1e5bf,0x1e6e8,0x1e654,0x1e3cd,
338 	0x1df2a,0x1d635,0x1cb16,0x1be2c,0x1ae4e,0x19bde,0x1868e,0x16e2e,
339 	0x1527f,0x1334a,0x11051,0xe951,	0xbe01,	0x8e0d,	0x5924,	0x1edd,};
340 
341     (2)	Iterative refinement
342 
343 	Apply Reciproot iteration three times to y and multiply the
344 	result by x to get an approximation z that matches sqrt(x)
345 	to about 1 ulp. To be exact, we will have
346 		-1ulp < sqrt(x)-z<1.0625ulp.
347 
348 	... set rounding mode to Round-to-nearest
349 	   y := y*(1.5-0.5*x*y*y)	... almost 15 sig. bits to 1/sqrt(x)
350 	   y := y*((1.5-2^-30)+0.5*x*y*y)... about 29 sig. bits to 1/sqrt(x)
351 	... special arrangement for better accuracy
352 	   z := x*y			... 29 bits to sqrt(x), with z*y<1
353 	   z := z + 0.5*z*(1-z*y)	... about 1 ulp to sqrt(x)
354 
355 	Remark 2. The constant 1.5-2^-30 is chosen to bias the error so that
356 	(a) the term z*y in the final iteration is always less than 1;
357 	(b) the error in the final result is biased upward so that
358 		-1 ulp < sqrt(x) - z < 1.0625 ulp
359 	    instead of |sqrt(x)-z|<1.03125ulp.
360 
361     (3)	Final adjustment
362 
363 	By twiddling y's last bit it is possible to force y to be
364 	correctly rounded according to the prevailing rounding mode
365 	as follows. Let r and i be copies of the rounding mode and
366 	inexact flag before entering the square root program. Also we
367 	use the expression y+-ulp for the next representable floating
368 	numbers (up and down) of y. Note that y+-ulp = either fixed
369 	point y+-1, or multiply y by nextafter(1,+-inf) in chopped
370 	mode.
371 
372 	R := RZ;		... set rounding mode to round-toward-zero
373 	switch(r) {
374 	    case RN:		... round-to-nearest
375 	       if(x<= z*(z-ulp)...chopped) z = z - ulp; else
376 	       if(x<= z*(z+ulp)...chopped) z = z; else z = z+ulp;
377 	       break;
378 	    case RZ:case RM:	... round-to-zero or round-to--inf
379 	       R:=RP;		... reset rounding mod to round-to-+inf
380 	       if(x<z*z ... rounded up) z = z - ulp; else
381 	       if(x>=(z+ulp)*(z+ulp) ...rounded up) z = z+ulp;
382 	       break;
383 	    case RP:		... round-to-+inf
384 	       if(x>(z+ulp)*(z+ulp)...chopped) z = z+2*ulp; else
385 	       if(x>z*z ...chopped) z = z+ulp;
386 	       break;
387 	}
388 
389 	Remark 3. The above comparisons can be done in fixed point. For
390 	example, to compare x and w=z*z chopped, it suffices to compare
391 	x1 and w1 (the trailing parts of x and w), regarding them as
392 	two's complement integers.
393 
394 	...Is z an exact square root?
395 	To determine whether z is an exact square root of x, let z1 be the
396 	trailing part of z, and also let x0 and x1 be the leading and
397 	trailing parts of x.
398 
399 	If ((z1&0x03ffffff)!=0)	... not exact if trailing 26 bits of z!=0
400 	    I := 1;		... Raise Inexact flag: z is not exact
401 	else {
402 	    j := 1 - [(x0>>20)&1]	... j = logb(x) mod 2
403 	    k := z1 >> 26;		... get z's 25-th and 26-th
404 					    fraction bits
405 	    I := i or (k&j) or ((k&(j+j+1))!=(x1&3));
406 	}
407 	R:= r		... restore rounded mode
408 	return sqrt(x):=z.
409 
410 	If multiplication is cheaper then the foregoing red tape, the
411 	Inexact flag can be evaluated by
412 
413 	    I := i;
414 	    I := (z*z!=x) or I.
415 
416 	Note that z*z can overwrite I; this value must be sensed if it is
417 	True.
418 
419 	Remark 4. If z*z = x exactly, then bit 25 to bit 0 of z1 must be
420 	zero.
421 
422 		    --------------------
423 		z1: |        f2        |
424 		    --------------------
425 		bit 31		   bit 0
426 
427 	Further more, bit 27 and 26 of z1, bit 0 and 1 of x1, and the odd
428 	or even of logb(x) have the following relations:
429 
430 	-------------------------------------------------
431 	bit 27,26 of z1		bit 1,0 of x1	logb(x)
432 	-------------------------------------------------
433 	00			00		odd and even
434 	01			01		even
435 	10			10		odd
436 	10			00		even
437 	11			01		even
438 	-------------------------------------------------
439 
440     (4)	Special cases (see (4) of Section A).
441 
442  */
443 
444 #if	LDBL_MANT_DIG == DBL_MANT_DIG
445 __strong_alias(sqrtl, sqrt);
446 #endif	/* LDBL_MANT_DIG == DBL_MANT_DIG */
447