1 /* @(#)e_sqrt.c 1.3 95/01/18 */ 2 /* 3 * ==================================================== 4 * Copyright (C) 1993 by Sun Microsystems, Inc. All rights reserved. 5 * 6 * Developed at SunSoft, 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 /* __ieee754_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 * ieee_sqrt(y) 24 * 2. Bit by bit computation 25 * Let q = ieee_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 "fdlibm.h" 84 85 #ifdef __STDC__ 86 static const double one = 1.0, tiny=1.0e-300; 87 #else 88 static double one = 1.0, tiny=1.0e-300; 89 #endif 90 91 #ifdef __STDC__ __ieee754_sqrt(double x)92 double __ieee754_sqrt(double x) 93 #else 94 double __ieee754_sqrt(x) 95 double x; 96 #endif 97 { 98 double z; 99 int sign = (int)0x80000000; 100 unsigned r,t1,s1,ix1,q1; 101 int ix0,s0,q,m,t,i; 102 103 ix0 = __HI(x); /* high word of x */ 104 ix1 = __LO(x); /* low word of x */ 105 106 /* take care of Inf and NaN */ 107 if((ix0&0x7ff00000)==0x7ff00000) { 108 return x*x+x; /* ieee_sqrt(NaN)=NaN, ieee_sqrt(+inf)=+inf 109 ieee_sqrt(-inf)=sNaN */ 110 } 111 /* take care of zero */ 112 if(ix0<=0) { 113 if(((ix0&(~sign))|ix1)==0) return x;/* ieee_sqrt(+-0) = +-0 */ 114 else if(ix0<0) 115 return (x-x)/(x-x); /* ieee_sqrt(-ve) = sNaN */ 116 } 117 /* normalize x */ 118 m = (ix0>>20); 119 if(m==0) { /* subnormal x */ 120 while(ix0==0) { 121 m -= 21; 122 ix0 |= (ix1>>11); ix1 <<= 21; 123 } 124 for(i=0;(ix0&0x00100000)==0;i++) ix0<<=1; 125 m -= i-1; 126 ix0 |= (ix1>>(32-i)); 127 ix1 <<= i; 128 } 129 m -= 1023; /* unbias exponent */ 130 ix0 = (ix0&0x000fffff)|0x00100000; 131 if(m&1){ /* odd m, double x to make it even */ 132 ix0 += ix0 + ((ix1&sign)>>31); 133 ix1 += ix1; 134 } 135 m >>= 1; /* m = [m/2] */ 136 137 /* generate ieee_sqrt(x) bit by bit */ 138 ix0 += ix0 + ((ix1&sign)>>31); 139 ix1 += ix1; 140 q = q1 = s0 = s1 = 0; /* [q,q1] = ieee_sqrt(x) */ 141 r = 0x00200000; /* r = moving bit from right to left */ 142 143 while(r!=0) { 144 t = s0+r; 145 if(t<=ix0) { 146 s0 = t+r; 147 ix0 -= t; 148 q += r; 149 } 150 ix0 += ix0 + ((ix1&sign)>>31); 151 ix1 += ix1; 152 r>>=1; 153 } 154 155 r = sign; 156 while(r!=0) { 157 t1 = s1+r; 158 t = s0; 159 if((t<ix0)||((t==ix0)&&(t1<=ix1))) { 160 s1 = t1+r; 161 if(((t1&sign)==sign)&&(s1&sign)==0) s0 += 1; 162 ix0 -= t; 163 if (ix1 < t1) ix0 -= 1; 164 ix1 -= t1; 165 q1 += r; 166 } 167 ix0 += ix0 + ((ix1&sign)>>31); 168 ix1 += ix1; 169 r>>=1; 170 } 171 172 /* use floating add to find out rounding direction */ 173 if((ix0|ix1)!=0) { 174 z = one-tiny; /* trigger inexact flag */ 175 if (z>=one) { 176 z = one+tiny; 177 if (q1==(unsigned)0xffffffff) { q1=0; q += 1;} 178 else if (z>one) { 179 if (q1==(unsigned)0xfffffffe) q+=1; 180 q1+=2; 181 } else 182 q1 += (q1&1); 183 } 184 } 185 ix0 = (q>>1)+0x3fe00000; 186 ix1 = q1>>1; 187 if ((q&1)==1) ix1 |= sign; 188 ix0 += (m <<20); 189 __HI(z) = ix0; 190 __LO(z) = ix1; 191 return z; 192 } 193 194 /* 195 Other methods (use floating-point arithmetic) 196 ------------- 197 (This is a copy of a drafted paper by Prof W. Kahan 198 and K.C. Ng, written in May, 1986) 199 200 Two algorithms are given here to implement ieee_sqrt(x) 201 (IEEE double precision arithmetic) in software. 202 Both supply ieee_sqrt(x) correctly rounded. The first algorithm (in 203 Section A) uses newton iterations and involves four divisions. 204 The second one uses reciproot iterations to avoid division, but 205 requires more multiplications. Both algorithms need the ability 206 to chop results of arithmetic operations instead of round them, 207 and the INEXACT flag to indicate when an arithmetic operation 208 is executed exactly with no roundoff error, all part of the 209 standard (IEEE 754-1985). The ability to perform shift, add, 210 subtract and logical AND operations upon 32-bit words is needed 211 too, though not part of the standard. 212 213 A. ieee_sqrt(x) by Newton Iteration 214 215 (1) Initial approximation 216 217 Let x0 and x1 be the leading and the trailing 32-bit words of 218 a floating point number x (in IEEE double format) respectively 219 220 1 11 52 ...widths 221 ------------------------------------------------------ 222 x: |s| e | f | 223 ------------------------------------------------------ 224 msb lsb msb lsb ...order 225 226 227 ------------------------ ------------------------ 228 x0: |s| e | f1 | x1: | f2 | 229 ------------------------ ------------------------ 230 231 By performing shifts and subtracts on x0 and x1 (both regarded 232 as integers), we obtain an 8-bit approximation of ieee_sqrt(x) as 233 follows. 234 235 k := (x0>>1) + 0x1ff80000; 236 y0 := k - T1[31&(k>>15)]. ... y ~ ieee_sqrt(x) to 8 bits 237 Here k is a 32-bit integer and T1[] is an integer array containing 238 correction terms. Now magically the floating value of y (y's 239 leading 32-bit word is y0, the value of its trailing word is 0) 240 approximates ieee_sqrt(x) to almost 8-bit. 241 242 Value of T1: 243 static int T1[32]= { 244 0, 1024, 3062, 5746, 9193, 13348, 18162, 23592, 245 29598, 36145, 43202, 50740, 58733, 67158, 75992, 85215, 246 83599, 71378, 60428, 50647, 41945, 34246, 27478, 21581, 247 16499, 12183, 8588, 5674, 3403, 1742, 661, 130,}; 248 249 (2) Iterative refinement 250 251 Apply Heron's rule three times to y, we have y approximates 252 sqrt(x) to within 1 ulp (Unit in the Last Place): 253 254 y := (y+x/y)/2 ... almost 17 sig. bits 255 y := (y+x/y)/2 ... almost 35 sig. bits 256 y := y-(y-x/y)/2 ... within 1 ulp 257 258 259 Remark 1. 260 Another way to improve y to within 1 ulp is: 261 262 y := (y+x/y) ... almost 17 sig. bits to 2*ieee_sqrt(x) 263 y := y - 0x00100006 ... almost 18 sig. bits to ieee_sqrt(x) 264 265 2 266 (x-y )*y 267 y := y + 2* ---------- ...within 1 ulp 268 2 269 3y + x 270 271 272 This formula has one division fewer than the one above; however, 273 it requires more multiplications and additions. Also x must be 274 scaled in advance to avoid spurious overflow in evaluating the 275 expression 3y*y+x. Hence it is not recommended uless division 276 is slow. If division is very slow, then one should use the 277 reciproot algorithm given in section B. 278 279 (3) Final adjustment 280 281 By twiddling y's last bit it is possible to force y to be 282 correctly rounded according to the prevailing rounding mode 283 as follows. Let r and i be copies of the rounding mode and 284 inexact flag before entering the square root program. Also we 285 use the expression y+-ulp for the next representable floating 286 numbers (up and down) of y. Note that y+-ulp = either fixed 287 point y+-1, or multiply y by ieee_nextafter(1,+-inf) in chopped 288 mode. 289 290 I := FALSE; ... reset INEXACT flag I 291 R := RZ; ... set rounding mode to round-toward-zero 292 z := x/y; ... chopped quotient, possibly inexact 293 If(not I) then { ... if the quotient is exact 294 if(z=y) { 295 I := i; ... restore inexact flag 296 R := r; ... restore rounded mode 297 return ieee_sqrt(x):=y. 298 } else { 299 z := z - ulp; ... special rounding 300 } 301 } 302 i := TRUE; ... ieee_sqrt(x) is inexact 303 If (r=RN) then z=z+ulp ... rounded-to-nearest 304 If (r=RP) then { ... round-toward-+inf 305 y = y+ulp; z=z+ulp; 306 } 307 y := y+z; ... chopped sum 308 y0:=y0-0x00100000; ... y := y/2 is correctly rounded. 309 I := i; ... restore inexact flag 310 R := r; ... restore rounded mode 311 return ieee_sqrt(x):=y. 312 313 (4) Special cases 314 315 Square root of +inf, +-0, or NaN is itself; 316 Square root of a negative number is NaN with invalid signal. 317 318 319 B. ieee_sqrt(x) by Reciproot Iteration 320 321 (1) Initial approximation 322 323 Let x0 and x1 be the leading and the trailing 32-bit words of 324 a floating point number x (in IEEE double format) respectively 325 (see section A). By performing shifs and subtracts on x0 and y0, 326 we obtain a 7.8-bit approximation of 1/ieee_sqrt(x) as follows. 327 328 k := 0x5fe80000 - (x0>>1); 329 y0:= k - T2[63&(k>>14)]. ... y ~ 1/ieee_sqrt(x) to 7.8 bits 330 331 Here k is a 32-bit integer and T2[] is an integer array 332 containing correction terms. Now magically the floating 333 value of y (y's leading 32-bit word is y0, the value of 334 its trailing word y1 is set to zero) approximates 1/ieee_sqrt(x) 335 to almost 7.8-bit. 336 337 Value of T2: 338 static int T2[64]= { 339 0x1500, 0x2ef8, 0x4d67, 0x6b02, 0x87be, 0xa395, 0xbe7a, 0xd866, 340 0xf14a, 0x1091b,0x11fcd,0x13552,0x14999,0x15c98,0x16e34,0x17e5f, 341 0x18d03,0x19a01,0x1a545,0x1ae8a,0x1b5c4,0x1bb01,0x1bfde,0x1c28d, 342 0x1c2de,0x1c0db,0x1ba73,0x1b11c,0x1a4b5,0x1953d,0x18266,0x16be0, 343 0x1683e,0x179d8,0x18a4d,0x19992,0x1a789,0x1b445,0x1bf61,0x1c989, 344 0x1d16d,0x1d77b,0x1dddf,0x1e2ad,0x1e5bf,0x1e6e8,0x1e654,0x1e3cd, 345 0x1df2a,0x1d635,0x1cb16,0x1be2c,0x1ae4e,0x19bde,0x1868e,0x16e2e, 346 0x1527f,0x1334a,0x11051,0xe951, 0xbe01, 0x8e0d, 0x5924, 0x1edd,}; 347 348 (2) Iterative refinement 349 350 Apply Reciproot iteration three times to y and multiply the 351 result by x to get an approximation z that matches ieee_sqrt(x) 352 to about 1 ulp. To be exact, we will have 353 -1ulp < ieee_sqrt(x)-z<1.0625ulp. 354 355 ... set rounding mode to Round-to-nearest 356 y := y*(1.5-0.5*x*y*y) ... almost 15 sig. bits to 1/ieee_sqrt(x) 357 y := y*((1.5-2^-30)+0.5*x*y*y)... about 29 sig. bits to 1/ieee_sqrt(x) 358 ... special arrangement for better accuracy 359 z := x*y ... 29 bits to ieee_sqrt(x), with z*y<1 360 z := z + 0.5*z*(1-z*y) ... about 1 ulp to ieee_sqrt(x) 361 362 Remark 2. The constant 1.5-2^-30 is chosen to bias the error so that 363 (a) the term z*y in the final iteration is always less than 1; 364 (b) the error in the final result is biased upward so that 365 -1 ulp < ieee_sqrt(x) - z < 1.0625 ulp 366 instead of |ieee_sqrt(x)-z|<1.03125ulp. 367 368 (3) Final adjustment 369 370 By twiddling y's last bit it is possible to force y to be 371 correctly rounded according to the prevailing rounding mode 372 as follows. Let r and i be copies of the rounding mode and 373 inexact flag before entering the square root program. Also we 374 use the expression y+-ulp for the next representable floating 375 numbers (up and down) of y. Note that y+-ulp = either fixed 376 point y+-1, or multiply y by ieee_nextafter(1,+-inf) in chopped 377 mode. 378 379 R := RZ; ... set rounding mode to round-toward-zero 380 switch(r) { 381 case RN: ... round-to-nearest 382 if(x<= z*(z-ulp)...chopped) z = z - ulp; else 383 if(x<= z*(z+ulp)...chopped) z = z; else z = z+ulp; 384 break; 385 case RZ:case RM: ... round-to-zero or round-to--inf 386 R:=RP; ... reset rounding mod to round-to-+inf 387 if(x<z*z ... rounded up) z = z - ulp; else 388 if(x>=(z+ulp)*(z+ulp) ...rounded up) z = z+ulp; 389 break; 390 case RP: ... round-to-+inf 391 if(x>(z+ulp)*(z+ulp)...chopped) z = z+2*ulp; else 392 if(x>z*z ...chopped) z = z+ulp; 393 break; 394 } 395 396 Remark 3. The above comparisons can be done in fixed point. For 397 example, to compare x and w=z*z chopped, it suffices to compare 398 x1 and w1 (the trailing parts of x and w), regarding them as 399 two's complement integers. 400 401 ...Is z an exact square root? 402 To determine whether z is an exact square root of x, let z1 be the 403 trailing part of z, and also let x0 and x1 be the leading and 404 trailing parts of x. 405 406 If ((z1&0x03ffffff)!=0) ... not exact if trailing 26 bits of z!=0 407 I := 1; ... Raise Inexact flag: z is not exact 408 else { 409 j := 1 - [(x0>>20)&1] ... j = ieee_logb(x) mod 2 410 k := z1 >> 26; ... get z's 25-th and 26-th 411 fraction bits 412 I := i or (k&j) or ((k&(j+j+1))!=(x1&3)); 413 } 414 R:= r ... restore rounded mode 415 return ieee_sqrt(x):=z. 416 417 If multiplication is cheaper then the foregoing red tape, the 418 Inexact flag can be evaluated by 419 420 I := i; 421 I := (z*z!=x) or I. 422 423 Note that z*z can overwrite I; this value must be sensed if it is 424 True. 425 426 Remark 4. If z*z = x exactly, then bit 25 to bit 0 of z1 must be 427 zero. 428 429 -------------------- 430 z1: | f2 | 431 -------------------- 432 bit 31 bit 0 433 434 Further more, bit 27 and 26 of z1, bit 0 and 1 of x1, and the odd 435 or even of ieee_logb(x) have the following relations: 436 437 ------------------------------------------------- 438 bit 27,26 of z1 bit 1,0 of x1 logb(x) 439 ------------------------------------------------- 440 00 00 odd and even 441 01 01 even 442 10 10 odd 443 10 00 even 444 11 01 even 445 ------------------------------------------------- 446 447 (4) Special cases (see (4) of Section A). 448 449 */ 450 451