xref: /illumos-gate/usr/src/lib/libmvec/common/__vrsqrt.c (revision b1e2e3fb17324e9ddf43db264a0c64da7756d9e6)
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 /*
23  * Copyright 2011 Nexenta Systems, Inc.  All rights reserved.
24  */
25 /*
26  * Copyright 2006 Sun Microsystems, Inc.  All rights reserved.
27  * Use is subject to license terms.
28  */
29 
30 #include <sys/isa_defs.h>
31 #include "libm_inlines.h"
32 
33 #ifdef _LITTLE_ENDIAN
34 #define HI(x)	*(1+(int*)x)
35 #define LO(x)	*(unsigned*)x
36 #else
37 #define HI(x)	*(int*)x
38 #define LO(x)	*(1+(unsigned*)x)
39 #endif
40 
41 #ifdef __RESTRICT
42 #define restrict _Restrict
43 #else
44 #define restrict
45 #endif
46 
47 /* double rsqrt(double x)
48  *
49  * Method :
50  *	1. Special cases:
51  *		for x = NaN				=> QNaN;
52  *		for x = +Inf				=> 0;
53  *		for x is negative, -Inf			=> QNaN + invalid;
54  *		for x = +0				=> +Inf + divide-by-zero;
55  *		for x = -0				=> -Inf + divide-by-zero.
56  *	2. Computes reciprocal square root from:
57  *		x = m * 2**n
58  *	Where:
59  *		m = [0.5, 2),
60  *		n = ((exponent + 1) & ~1).
61  *	Then:
62  *		rsqrt(x) = 1/sqrt( m * 2**n ) = (2 ** (-n/2)) * (1/sqrt(m))
63  *	2. Computes 1/sqrt(m) from:
64  *		1/sqrt(m) = (1/sqrt(m0)) * (1/sqrt(1 + (1/m0)*dm))
65  *	Where:
66  *		m = m0 + dm,
67  *		m0 = 0.5 * (1 + k/64) for m = [0.5,         0.5+127/256), k = [0, 63];
68  *		m0 = 1.0 * (0 + k/64) for m = [0.5+127/256, 1.0+127/128), k = [64, 127];
69  *		m0 = 2.0              for m = [1.0+127/128, 2.0),         k = 128.
70  *	Then:
71  *		1/sqrt(m0) is looked up in a table,
72  *		1/m0 is computed as (1/sqrt(m0)) * (1/sqrt(m0)).
73  *		1/sqrt(1 + (1/m0)*dm) is computed using approximation:
74  *			1/sqrt(1 + z) = (((((a6 * z + a5) * z + a4) * z + a3)
75  *						* z + a2) * z + a1) * z + a0
76  *			where z = [-1/128, 1/128].
77  *
78  * Accuracy:
79  *	The maximum relative error for the approximating
80  *	polynomial is 2**(-56.26).
81  *	Maximum error observed: less than 0.563 ulp after 1.500.000.000
82  *	results.
83  */
84 
85 extern double sqrt (double);
86 extern const double __vlibm_TBL_rsqrt[];
87 
88 static void
89 __vrsqrt_n(int n, double * restrict px, int stridex, double * restrict py, int stridey);
90 
91 #pragma no_inline(__vrsqrt_n)
92 
93 #define RETURN(ret)						\
94 {								\
95 	*py = (ret);						\
96 	py += stridey;						\
97 	if (n_n == 0)						\
98 	{							\
99 		spx = px; spy = py;				\
100 		hx = HI(px);					\
101 		continue;					\
102 	}							\
103 	n--;							\
104 	break;							\
105 }
106 
107 static const double
108 	DONE = 1.0,
109 	K1 = -5.00000000000005209867e-01,
110 	K2 =  3.75000000000004884257e-01,
111 	K3 = -3.12499999317136886551e-01,
112 	K4 =  2.73437499359815081532e-01,
113 	K5 = -2.46116125605037803130e-01,
114 	K6 =  2.25606914648617522896e-01;
115 
116 void
117 __vrsqrt(int n, double * restrict px, int stridex, double * restrict py, int stridey)
118 {
119 	double		*spx, *spy;
120 	int		ax, lx, hx, n_n;
121 	double		res;
122 
123 	while (n > 1)
124 	{
125 		n_n = 0;
126 		spx = px;
127 		spy = py;
128 		hx = HI(px);
129 		for (; n > 1 ; n--)
130 		{
131 			px += stridex;
132 			if (hx >= 0x7ff00000)		/* X = NaN or Inf	*/
133 			{
134 				res = *(px - stridex);
135 				RETURN (DONE / res)
136 			}
137 
138 			py += stridey;
139 
140 			if (hx < 0x00100000)		/* X = denormal, zero or negative	*/
141 			{
142 				py -= stridey;
143 				ax = hx & 0x7fffffff;
144 				lx = LO((px - stridex));
145 				res = *(px - stridex);
146 
147 				if ((ax | lx) == 0)	/* |X| = zero	*/
148 				{
149 					RETURN (DONE / res)
150 				}
151 				else if (hx >= 0)	/* X = denormal	*/
152 				{
153 					double		res_c0, dsqrt_exp0;
154 					int		ind0, sqrt_exp0;
155 					double		xx0, dexp_hi0, dexp_lo0;
156 					int		hx0, resh0, res_ch0;
157 
158 					res = *(long long*)&res;
159 
160 					hx0 = HI(&res);
161 					sqrt_exp0 = (0x817 - (hx0 >> 21)) << 20;
162 					ind0 = (((hx0 >> 10) & 0x7f8) + 8) & -16;
163 
164 					resh0 = (hx0 & 0x001fffff) | 0x3fe00000;
165 					res_ch0 = (resh0 + 0x00002000) & 0x7fffc000;
166 					HI(&res) = resh0;
167 					HI(&res_c0) = res_ch0;
168 					LO(&res_c0) = 0;
169 
170 					dexp_hi0 = ((double*)((char*)__vlibm_TBL_rsqrt + ind0))[0];
171 					dexp_lo0 = ((double*)((char*)__vlibm_TBL_rsqrt + ind0))[1];
172 					xx0 = dexp_hi0 * dexp_hi0;
173 					xx0 = (res - res_c0) * xx0;
174 					res = (((((K6 * xx0 + K5) * xx0 + K4) * xx0 + K3) * xx0 + K2) * xx0 + K1) * xx0;
175 
176 					res = dexp_hi0 * res + dexp_lo0 + dexp_hi0;
177 
178 					HI(&dsqrt_exp0) = sqrt_exp0;
179 					LO(&dsqrt_exp0) = 0;
180 					res *= dsqrt_exp0;
181 
182 					RETURN (res)
183 				}
184 				else	/* X = negative	*/
185 				{
186 					RETURN (sqrt(res))
187 				}
188 			}
189 			n_n++;
190 			hx = HI(px);
191 		}
192 		if (n_n > 0)
193 			__vrsqrt_n(n_n, spx, stridex, spy, stridey);
194 	}
195 	if (n > 0)
196 	{
197 		hx = HI(px);
198 
199 		if (hx >= 0x7ff00000)		/* X = NaN or Inf	*/
200 		{
201 			res = *px;
202 			*py = DONE / res;
203 		}
204 		else if (hx < 0x00100000)	/* X = denormal, zero or negative	*/
205 		{
206 			ax = hx & 0x7fffffff;
207 			lx = LO(px);
208 			res = *px;
209 
210 			if ((ax | lx) == 0)	/* |X| = zero	*/
211 			{
212 				*py = DONE / res;
213 			}
214 			else if (hx >= 0)	/* X = denormal	*/
215 			{
216 				double		res_c0, dsqrt_exp0;
217 				int		ind0, sqrt_exp0;
218 				double		xx0, dexp_hi0, dexp_lo0;
219 				int		hx0, resh0, res_ch0;
220 
221 				res = *(long long*)&res;
222 
223 				hx0 = HI(&res);
224 				sqrt_exp0 = (0x817 - (hx0 >> 21)) << 20;
225 				ind0 = (((hx0 >> 10) & 0x7f8) + 8) & -16;
226 
227 				resh0 = (hx0 & 0x001fffff) | 0x3fe00000;
228 				res_ch0 = (resh0 + 0x00002000) & 0x7fffc000;
229 				HI(&res) = resh0;
230 				HI(&res_c0) = res_ch0;
231 				LO(&res_c0) = 0;
232 
233 				dexp_hi0 = ((double*)((char*)__vlibm_TBL_rsqrt + ind0))[0];
234 				dexp_lo0 = ((double*)((char*)__vlibm_TBL_rsqrt + ind0))[1];
235 				xx0 = dexp_hi0 * dexp_hi0;
236 				xx0 = (res - res_c0) * xx0;
237 				res = (((((K6 * xx0 + K5) * xx0 + K4) * xx0 + K3) * xx0 + K2) * xx0 + K1) * xx0;
238 
239 				res = dexp_hi0 * res + dexp_lo0 + dexp_hi0;
240 
241 				HI(&dsqrt_exp0) = sqrt_exp0;
242 				LO(&dsqrt_exp0) = 0;
243 				res *= dsqrt_exp0;
244 
245 				*py = res;
246 			}
247 			else	/* X = negative	*/
248 			{
249 				*py = sqrt(res);
250 			}
251 		}
252 		else
253 		{
254 			double		res_c0, dsqrt_exp0;
255 			int		ind0, sqrt_exp0;
256 			double		xx0, dexp_hi0, dexp_lo0;
257 			int		resh0, res_ch0;
258 
259 			sqrt_exp0 = (0x5fe - (hx >> 21)) << 20;
260 			ind0 = (((hx >> 10) & 0x7f8) + 8) & -16;
261 
262 			resh0 = (hx & 0x001fffff) | 0x3fe00000;
263 			res_ch0 = (resh0 + 0x00002000) & 0x7fffc000;
264 			HI(&res) = resh0;
265 			LO(&res) = LO(px);
266 			HI(&res_c0) = res_ch0;
267 			LO(&res_c0) = 0;
268 
269 			dexp_hi0 = ((double*)((char*)__vlibm_TBL_rsqrt + ind0))[0];
270 			dexp_lo0 = ((double*)((char*)__vlibm_TBL_rsqrt + ind0))[1];
271 			xx0 = dexp_hi0 * dexp_hi0;
272 			xx0 = (res - res_c0) * xx0;
273 			res = (((((K6 * xx0 + K5) * xx0 + K4) * xx0 + K3) * xx0 + K2) * xx0 + K1) * xx0;
274 
275 			res = dexp_hi0 * res + dexp_lo0 + dexp_hi0;
276 
277 			HI(&dsqrt_exp0) = sqrt_exp0;
278 			LO(&dsqrt_exp0) = 0;
279 			res *= dsqrt_exp0;
280 
281 			*py = res;
282 		}
283 	}
284 }
285 
286 static void
287 __vrsqrt_n(int n, double * restrict px, int stridex, double * restrict py, int stridey)
288 {
289 	double		res0, res_c0, dsqrt_exp0;
290 	double		res1, res_c1, dsqrt_exp1;
291 	double		res2, res_c2, dsqrt_exp2;
292 	int		ind0, sqrt_exp0;
293 	int		ind1, sqrt_exp1;
294 	int		ind2, sqrt_exp2;
295 	double		xx0, dexp_hi0, dexp_lo0;
296 	double		xx1, dexp_hi1, dexp_lo1;
297 	double		xx2, dexp_hi2, dexp_lo2;
298 	int		hx0, resh0, res_ch0;
299 	int		hx1, resh1, res_ch1;
300 	int		hx2, resh2, res_ch2;
301 
302 	LO(&dsqrt_exp0) = 0;
303 	LO(&dsqrt_exp1) = 0;
304 	LO(&dsqrt_exp2) = 0;
305 	LO(&res_c0) = 0;
306 	LO(&res_c1) = 0;
307 	LO(&res_c2) = 0;
308 
309 	for(; n > 2 ; n -= 3)
310 	{
311 		hx0 = HI(px);
312 		LO(&res0) = LO(px);
313 		px += stridex;
314 
315 		hx1 = HI(px);
316 		LO(&res1) = LO(px);
317 		px += stridex;
318 
319 		hx2 = HI(px);
320 		LO(&res2) = LO(px);
321 		px += stridex;
322 
323 		sqrt_exp0 = (0x5fe - (hx0 >> 21)) << 20;
324 		sqrt_exp1 = (0x5fe - (hx1 >> 21)) << 20;
325 		sqrt_exp2 = (0x5fe - (hx2 >> 21)) << 20;
326 		ind0 = (((hx0 >> 10) & 0x7f8) + 8) & -16;
327 		ind1 = (((hx1 >> 10) & 0x7f8) + 8) & -16;
328 		ind2 = (((hx2 >> 10) & 0x7f8) + 8) & -16;
329 
330 		resh0 = (hx0 & 0x001fffff) | 0x3fe00000;
331 		resh1 = (hx1 & 0x001fffff) | 0x3fe00000;
332 		resh2 = (hx2 & 0x001fffff) | 0x3fe00000;
333 		res_ch0 = (resh0 + 0x00002000) & 0x7fffc000;
334 		res_ch1 = (resh1 + 0x00002000) & 0x7fffc000;
335 		res_ch2 = (resh2 + 0x00002000) & 0x7fffc000;
336 		HI(&res0) = resh0;
337 		HI(&res1) = resh1;
338 		HI(&res2) = resh2;
339 		HI(&res_c0) = res_ch0;
340 		HI(&res_c1) = res_ch1;
341 		HI(&res_c2) = res_ch2;
342 
343 		dexp_hi0 = ((double*)((char*)__vlibm_TBL_rsqrt + ind0))[0];
344 		dexp_hi1 = ((double*)((char*)__vlibm_TBL_rsqrt + ind1))[0];
345 		dexp_hi2 = ((double*)((char*)__vlibm_TBL_rsqrt + ind2))[0];
346 		dexp_lo0 = ((double*)((char*)__vlibm_TBL_rsqrt + ind0))[1];
347 		dexp_lo1 = ((double*)((char*)__vlibm_TBL_rsqrt + ind1))[1];
348 		dexp_lo2 = ((double*)((char*)__vlibm_TBL_rsqrt + ind2))[1];
349 		xx0 = dexp_hi0 * dexp_hi0;
350 		xx1 = dexp_hi1 * dexp_hi1;
351 		xx2 = dexp_hi2 * dexp_hi2;
352 		xx0 = (res0 - res_c0) * xx0;
353 		xx1 = (res1 - res_c1) * xx1;
354 		xx2 = (res2 - res_c2) * xx2;
355 		res0 = (((((K6 * xx0 + K5) * xx0 + K4) * xx0 + K3) * xx0 + K2) * xx0 + K1) * xx0;
356 		res1 = (((((K6 * xx1 + K5) * xx1 + K4) * xx1 + K3) * xx1 + K2) * xx1 + K1) * xx1;
357 		res2 = (((((K6 * xx2 + K5) * xx2 + K4) * xx2 + K3) * xx2 + K2) * xx2 + K1) * xx2;
358 
359 		res0 = dexp_hi0 * res0 + dexp_lo0 + dexp_hi0;
360 		res1 = dexp_hi1 * res1 + dexp_lo1 + dexp_hi1;
361 		res2 = dexp_hi2 * res2 + dexp_lo2 + dexp_hi2;
362 
363 		HI(&dsqrt_exp0) = sqrt_exp0;
364 		HI(&dsqrt_exp1) = sqrt_exp1;
365 		HI(&dsqrt_exp2) = sqrt_exp2;
366 		res0 *= dsqrt_exp0;
367 		res1 *= dsqrt_exp1;
368 		res2 *= dsqrt_exp2;
369 
370 		*py = res0;
371 		py += stridey;
372 
373 		*py = res1;
374 		py += stridey;
375 
376 		*py = res2;
377 		py += stridey;
378 	}
379 
380 	for(; n > 0 ; n--)
381 	{
382 		hx0 = HI(px);
383 
384 		sqrt_exp0 = (0x5fe - (hx0 >> 21)) << 20;
385 		ind0 = (((hx0 >> 10) & 0x7f8) + 8) & -16;
386 
387 		resh0 = (hx0 & 0x001fffff) | 0x3fe00000;
388 		res_ch0 = (resh0 + 0x00002000) & 0x7fffc000;
389 		HI(&res0) = resh0;
390 		LO(&res0) = LO(px);
391 		HI(&res_c0) = res_ch0;
392 		LO(&res_c0) = 0;
393 
394 		px += stridex;
395 
396 		dexp_hi0 = ((double*)((char*)__vlibm_TBL_rsqrt + ind0))[0];
397 		dexp_lo0 = ((double*)((char*)__vlibm_TBL_rsqrt + ind0))[1];
398 		xx0 = dexp_hi0 * dexp_hi0;
399 		xx0 = (res0 - res_c0) * xx0;
400 		res0 = (((((K6 * xx0 + K5) * xx0 + K4) * xx0 + K3) * xx0 + K2) * xx0 + K1) * xx0;
401 
402 		res0 = dexp_hi0 * res0 + dexp_lo0 + dexp_hi0;
403 
404 		HI(&dsqrt_exp0) = sqrt_exp0;
405 		LO(&dsqrt_exp0) = 0;
406 		res0 *= dsqrt_exp0;
407 
408 		*py = res0;
409 		py += stridey;
410 	}
411 }
412