e_sqrt.c revision 97409
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#ifndef lint 14static char rcsid[] = "$FreeBSD: head/lib/msun/src/e_sqrt.c 97409 2002-05-28 17:51:46Z alfred $"; 15#endif 16 17/* __ieee754_sqrt(x) 18 * Return correctly rounded sqrt. 19 * ------------------------------------------ 20 * | Use the hardware sqrt if you have one | 21 * ------------------------------------------ 22 * Method: 23 * Bit by bit method using integer arithmetic. (Slow, but portable) 24 * 1. Normalization 25 * Scale x to y in [1,4) with even powers of 2: 26 * find an integer k such that 1 <= (y=x*2^(2k)) < 4, then 27 * sqrt(x) = 2^k * sqrt(y) 28 * 2. Bit by bit computation 29 * Let q = sqrt(y) truncated to i bit after binary point (q = 1), 30 * i 0 31 * i+1 2 32 * s = 2*q , and y = 2 * ( y - q ). (1) 33 * i i i i 34 * 35 * To compute q from q , one checks whether 36 * i+1 i 37 * 38 * -(i+1) 2 39 * (q + 2 ) <= y. (2) 40 * i 41 * -(i+1) 42 * If (2) is false, then q = q ; otherwise q = q + 2 . 43 * i+1 i i+1 i 44 * 45 * With some algebric manipulation, it is not difficult to see 46 * that (2) is equivalent to 47 * -(i+1) 48 * s + 2 <= y (3) 49 * i i 50 * 51 * The advantage of (3) is that s and y can be computed by 52 * i i 53 * the following recurrence formula: 54 * if (3) is false 55 * 56 * s = s , y = y ; (4) 57 * i+1 i i+1 i 58 * 59 * otherwise, 60 * -i -(i+1) 61 * s = s + 2 , y = y - s - 2 (5) 62 * i+1 i i+1 i i 63 * 64 * One may easily use induction to prove (4) and (5). 65 * Note. Since the left hand side of (3) contain only i+2 bits, 66 * it does not necessary to do a full (53-bit) comparison 67 * in (3). 68 * 3. Final rounding 69 * After generating the 53 bits result, we compute one more bit. 70 * Together with the remainder, we can decide whether the 71 * result is exact, bigger than 1/2ulp, or less than 1/2ulp 72 * (it will never equal to 1/2ulp). 73 * The rounding mode can be detected by checking whether 74 * huge + tiny is equal to huge, and whether huge - tiny is 75 * equal to huge for some floating point number "huge" and "tiny". 76 * 77 * Special cases: 78 * sqrt(+-0) = +-0 ... exact 79 * sqrt(inf) = inf 80 * sqrt(-ve) = NaN ... with invalid signal 81 * sqrt(NaN) = NaN ... with invalid signal for signaling NaN 82 * 83 * Other methods : see the appended file at the end of the program below. 84 *--------------- 85 */ 86 87#include "math.h" 88#include "math_private.h" 89 90static const double one = 1.0, tiny=1.0e-300; 91 92 double __generic___ieee754_sqrt(double x) 93{ 94 double z; 95 int32_t sign = (int)0x80000000; 96 int32_t ix0,s0,q,m,t,i; 97 u_int32_t r,t1,s1,ix1,q1; 98 99 EXTRACT_WORDS(ix0,ix1,x); 100 101 /* take care of Inf and NaN */ 102 if((ix0&0x7ff00000)==0x7ff00000) { 103 return x*x+x; /* sqrt(NaN)=NaN, sqrt(+inf)=+inf 104 sqrt(-inf)=sNaN */ 105 } 106 /* take care of zero */ 107 if(ix0<=0) { 108 if(((ix0&(~sign))|ix1)==0) return x;/* sqrt(+-0) = +-0 */ 109 else if(ix0<0) 110 return (x-x)/(x-x); /* sqrt(-ve) = sNaN */ 111 } 112 /* normalize x */ 113 m = (ix0>>20); 114 if(m==0) { /* subnormal x */ 115 while(ix0==0) { 116 m -= 21; 117 ix0 |= (ix1>>11); ix1 <<= 21; 118 } 119 for(i=0;(ix0&0x00100000)==0;i++) ix0<<=1; 120 m -= i-1; 121 ix0 |= (ix1>>(32-i)); 122 ix1 <<= i; 123 } 124 m -= 1023; /* unbias exponent */ 125 ix0 = (ix0&0x000fffff)|0x00100000; 126 if(m&1){ /* odd m, double x to make it even */ 127 ix0 += ix0 + ((ix1&sign)>>31); 128 ix1 += ix1; 129 } 130 m >>= 1; /* m = [m/2] */ 131 132 /* generate sqrt(x) bit by bit */ 133 ix0 += ix0 + ((ix1&sign)>>31); 134 ix1 += ix1; 135 q = q1 = s0 = s1 = 0; /* [q,q1] = sqrt(x) */ 136 r = 0x00200000; /* r = moving bit from right to left */ 137 138 while(r!=0) { 139 t = s0+r; 140 if(t<=ix0) { 141 s0 = t+r; 142 ix0 -= t; 143 q += r; 144 } 145 ix0 += ix0 + ((ix1&sign)>>31); 146 ix1 += ix1; 147 r>>=1; 148 } 149 150 r = sign; 151 while(r!=0) { 152 t1 = s1+r; 153 t = s0; 154 if((t<ix0)||((t==ix0)&&(t1<=ix1))) { 155 s1 = t1+r; 156 if(((t1&sign)==sign)&&(s1&sign)==0) s0 += 1; 157 ix0 -= t; 158 if (ix1 < t1) ix0 -= 1; 159 ix1 -= t1; 160 q1 += r; 161 } 162 ix0 += ix0 + ((ix1&sign)>>31); 163 ix1 += ix1; 164 r>>=1; 165 } 166 167 /* use floating add to find out rounding direction */ 168 if((ix0|ix1)!=0) { 169 z = one-tiny; /* trigger inexact flag */ 170 if (z>=one) { 171 z = one+tiny; 172 if (q1==(u_int32_t)0xffffffff) { q1=0; q += 1;} 173 else if (z>one) { 174 if (q1==(u_int32_t)0xfffffffe) q+=1; 175 q1+=2; 176 } else 177 q1 += (q1&1); 178 } 179 } 180 ix0 = (q>>1)+0x3fe00000; 181 ix1 = q1>>1; 182 if ((q&1)==1) ix1 |= sign; 183 ix0 += (m <<20); 184 INSERT_WORDS(z,ix0,ix1); 185 return z; 186} 187 188/* 189Other methods (use floating-point arithmetic) 190------------- 191(This is a copy of a drafted paper by Prof W. Kahan 192and K.C. Ng, written in May, 1986) 193 194 Two algorithms are given here to implement sqrt(x) 195 (IEEE double precision arithmetic) in software. 196 Both supply sqrt(x) correctly rounded. The first algorithm (in 197 Section A) uses newton iterations and involves four divisions. 198 The second one uses reciproot iterations to avoid division, but 199 requires more multiplications. Both algorithms need the ability 200 to chop results of arithmetic operations instead of round them, 201 and the INEXACT flag to indicate when an arithmetic operation 202 is executed exactly with no roundoff error, all part of the 203 standard (IEEE 754-1985). The ability to perform shift, add, 204 subtract and logical AND operations upon 32-bit words is needed 205 too, though not part of the standard. 206 207A. sqrt(x) by Newton Iteration 208 209 (1) Initial approximation 210 211 Let x0 and x1 be the leading and the trailing 32-bit words of 212 a floating point number x (in IEEE double format) respectively 213 214 1 11 52 ...widths 215 ------------------------------------------------------ 216 x: |s| e | f | 217 ------------------------------------------------------ 218 msb lsb msb lsb ...order 219 220 221 ------------------------ ------------------------ 222 x0: |s| e | f1 | x1: | f2 | 223 ------------------------ ------------------------ 224 225 By performing shifts and subtracts on x0 and x1 (both regarded 226 as integers), we obtain an 8-bit approximation of sqrt(x) as 227 follows. 228 229 k := (x0>>1) + 0x1ff80000; 230 y0 := k - T1[31&(k>>15)]. ... y ~ sqrt(x) to 8 bits 231 Here k is a 32-bit integer and T1[] is an integer array containing 232 correction terms. Now magically the floating value of y (y's 233 leading 32-bit word is y0, the value of its trailing word is 0) 234 approximates sqrt(x) to almost 8-bit. 235 236 Value of T1: 237 static int T1[32]= { 238 0, 1024, 3062, 5746, 9193, 13348, 18162, 23592, 239 29598, 36145, 43202, 50740, 58733, 67158, 75992, 85215, 240 83599, 71378, 60428, 50647, 41945, 34246, 27478, 21581, 241 16499, 12183, 8588, 5674, 3403, 1742, 661, 130,}; 242 243 (2) Iterative refinement 244 245 Apply Heron's rule three times to y, we have y approximates 246 sqrt(x) to within 1 ulp (Unit in the Last Place): 247 248 y := (y+x/y)/2 ... almost 17 sig. bits 249 y := (y+x/y)/2 ... almost 35 sig. bits 250 y := y-(y-x/y)/2 ... within 1 ulp 251 252 253 Remark 1. 254 Another way to improve y to within 1 ulp is: 255 256 y := (y+x/y) ... almost 17 sig. bits to 2*sqrt(x) 257 y := y - 0x00100006 ... almost 18 sig. bits to sqrt(x) 258 259 2 260 (x-y )*y 261 y := y + 2* ---------- ...within 1 ulp 262 2 263 3y + x 264 265 266 This formula has one division fewer than the one above; however, 267 it requires more multiplications and additions. Also x must be 268 scaled in advance to avoid spurious overflow in evaluating the 269 expression 3y*y+x. Hence it is not recommended uless division 270 is slow. If division is very slow, then one should use the 271 reciproot algorithm given in section B. 272 273 (3) Final adjustment 274 275 By twiddling y's last bit it is possible to force y to be 276 correctly rounded according to the prevailing rounding mode 277 as follows. Let r and i be copies of the rounding mode and 278 inexact flag before entering the square root program. Also we 279 use the expression y+-ulp for the next representable floating 280 numbers (up and down) of y. Note that y+-ulp = either fixed 281 point y+-1, or multiply y by nextafter(1,+-inf) in chopped 282 mode. 283 284 I := FALSE; ... reset INEXACT flag I 285 R := RZ; ... set rounding mode to round-toward-zero 286 z := x/y; ... chopped quotient, possibly inexact 287 If(not I) then { ... if the quotient is exact 288 if(z=y) { 289 I := i; ... restore inexact flag 290 R := r; ... restore rounded mode 291 return sqrt(x):=y. 292 } else { 293 z := z - ulp; ... special rounding 294 } 295 } 296 i := TRUE; ... sqrt(x) is inexact 297 If (r=RN) then z=z+ulp ... rounded-to-nearest 298 If (r=RP) then { ... round-toward-+inf 299 y = y+ulp; z=z+ulp; 300 } 301 y := y+z; ... chopped sum 302 y0:=y0-0x00100000; ... y := y/2 is correctly rounded. 303 I := i; ... restore inexact flag 304 R := r; ... restore rounded mode 305 return sqrt(x):=y. 306 307 (4) Special cases 308 309 Square root of +inf, +-0, or NaN is itself; 310 Square root of a negative number is NaN with invalid signal. 311 312 313B. sqrt(x) by Reciproot Iteration 314 315 (1) Initial approximation 316 317 Let x0 and x1 be the leading and the trailing 32-bit words of 318 a floating point number x (in IEEE double format) respectively 319 (see section A). By performing shifs and subtracts on x0 and y0, 320 we obtain a 7.8-bit approximation of 1/sqrt(x) as follows. 321 322 k := 0x5fe80000 - (x0>>1); 323 y0:= k - T2[63&(k>>14)]. ... y ~ 1/sqrt(x) to 7.8 bits 324 325 Here k is a 32-bit integer and T2[] is an integer array 326 containing correction terms. Now magically the floating 327 value of y (y's leading 32-bit word is y0, the value of 328 its trailing word y1 is set to zero) approximates 1/sqrt(x) 329 to almost 7.8-bit. 330 331 Value of T2: 332 static int T2[64]= { 333 0x1500, 0x2ef8, 0x4d67, 0x6b02, 0x87be, 0xa395, 0xbe7a, 0xd866, 334 0xf14a, 0x1091b,0x11fcd,0x13552,0x14999,0x15c98,0x16e34,0x17e5f, 335 0x18d03,0x19a01,0x1a545,0x1ae8a,0x1b5c4,0x1bb01,0x1bfde,0x1c28d, 336 0x1c2de,0x1c0db,0x1ba73,0x1b11c,0x1a4b5,0x1953d,0x18266,0x16be0, 337 0x1683e,0x179d8,0x18a4d,0x19992,0x1a789,0x1b445,0x1bf61,0x1c989, 338 0x1d16d,0x1d77b,0x1dddf,0x1e2ad,0x1e5bf,0x1e6e8,0x1e654,0x1e3cd, 339 0x1df2a,0x1d635,0x1cb16,0x1be2c,0x1ae4e,0x19bde,0x1868e,0x16e2e, 340 0x1527f,0x1334a,0x11051,0xe951, 0xbe01, 0x8e0d, 0x5924, 0x1edd,}; 341 342 (2) Iterative refinement 343 344 Apply Reciproot iteration three times to y and multiply the 345 result by x to get an approximation z that matches sqrt(x) 346 to about 1 ulp. To be exact, we will have 347 -1ulp < sqrt(x)-z<1.0625ulp. 348 349 ... set rounding mode to Round-to-nearest 350 y := y*(1.5-0.5*x*y*y) ... almost 15 sig. bits to 1/sqrt(x) 351 y := y*((1.5-2^-30)+0.5*x*y*y)... about 29 sig. bits to 1/sqrt(x) 352 ... special arrangement for better accuracy 353 z := x*y ... 29 bits to sqrt(x), with z*y<1 354 z := z + 0.5*z*(1-z*y) ... about 1 ulp to sqrt(x) 355 356 Remark 2. The constant 1.5-2^-30 is chosen to bias the error so that 357 (a) the term z*y in the final iteration is always less than 1; 358 (b) the error in the final result is biased upward so that 359 -1 ulp < sqrt(x) - z < 1.0625 ulp 360 instead of |sqrt(x)-z|<1.03125ulp. 361 362 (3) Final adjustment 363 364 By twiddling y's last bit it is possible to force y to be 365 correctly rounded according to the prevailing rounding mode 366 as follows. Let r and i be copies of the rounding mode and 367 inexact flag before entering the square root program. Also we 368 use the expression y+-ulp for the next representable floating 369 numbers (up and down) of y. Note that y+-ulp = either fixed 370 point y+-1, or multiply y by nextafter(1,+-inf) in chopped 371 mode. 372 373 R := RZ; ... set rounding mode to round-toward-zero 374 switch(r) { 375 case RN: ... round-to-nearest 376 if(x<= z*(z-ulp)...chopped) z = z - ulp; else 377 if(x<= z*(z+ulp)...chopped) z = z; else z = z+ulp; 378 break; 379 case RZ:case RM: ... round-to-zero or round-to--inf 380 R:=RP; ... reset rounding mod to round-to-+inf 381 if(x<z*z ... rounded up) z = z - ulp; else 382 if(x>=(z+ulp)*(z+ulp) ...rounded up) z = z+ulp; 383 break; 384 case RP: ... round-to-+inf 385 if(x>(z+ulp)*(z+ulp)...chopped) z = z+2*ulp; else 386 if(x>z*z ...chopped) z = z+ulp; 387 break; 388 } 389 390 Remark 3. The above comparisons can be done in fixed point. For 391 example, to compare x and w=z*z chopped, it suffices to compare 392 x1 and w1 (the trailing parts of x and w), regarding them as 393 two's complement integers. 394 395 ...Is z an exact square root? 396 To determine whether z is an exact square root of x, let z1 be the 397 trailing part of z, and also let x0 and x1 be the leading and 398 trailing parts of x. 399 400 If ((z1&0x03ffffff)!=0) ... not exact if trailing 26 bits of z!=0 401 I := 1; ... Raise Inexact flag: z is not exact 402 else { 403 j := 1 - [(x0>>20)&1] ... j = logb(x) mod 2 404 k := z1 >> 26; ... get z's 25-th and 26-th 405 fraction bits 406 I := i or (k&j) or ((k&(j+j+1))!=(x1&3)); 407 } 408 R:= r ... restore rounded mode 409 return sqrt(x):=z. 410 411 If multiplication is cheaper then the foregoing red tape, the 412 Inexact flag can be evaluated by 413 414 I := i; 415 I := (z*z!=x) or I. 416 417 Note that z*z can overwrite I; this value must be sensed if it is 418 True. 419 420 Remark 4. If z*z = x exactly, then bit 25 to bit 0 of z1 must be 421 zero. 422 423 -------------------- 424 z1: | f2 | 425 -------------------- 426 bit 31 bit 0 427 428 Further more, bit 27 and 26 of z1, bit 0 and 1 of x1, and the odd 429 or even of logb(x) have the following relations: 430 431 ------------------------------------------------- 432 bit 27,26 of z1 bit 1,0 of x1 logb(x) 433 ------------------------------------------------- 434 00 00 odd and even 435 01 01 even 436 10 10 odd 437 10 00 even 438 11 01 even 439 ------------------------------------------------- 440 441 (4) Special cases (see (4) of Section A). 442 443 */ 444 445