author | alanb |
Thu, 01 Dec 2016 08:57:53 +0000 | |
changeset 42338 | a60f280f803c |
parent 32992 | 19ed8781c2ba |
child 42751 | 38d28e784f44 |
permissions | -rw-r--r-- |
32655 | 1 |
/* |
2 |
* Copyright (c) 1998, 2015, Oracle and/or its affiliates. All rights reserved. |
|
3 |
* DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. |
|
4 |
* |
|
5 |
* This code is free software; you can redistribute it and/or modify it |
|
6 |
* under the terms of the GNU General Public License version 2 only, as |
|
7 |
* published by the Free Software Foundation. Oracle designates this |
|
8 |
* particular file as subject to the "Classpath" exception as provided |
|
9 |
* by Oracle in the LICENSE file that accompanied this code. |
|
10 |
* |
|
11 |
* This code is distributed in the hope that it will be useful, but WITHOUT |
|
12 |
* ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or |
|
13 |
* FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License |
|
14 |
* version 2 for more details (a copy is included in the LICENSE file that |
|
15 |
* accompanied this code). |
|
16 |
* |
|
17 |
* You should have received a copy of the GNU General Public License version |
|
18 |
* 2 along with this work; if not, write to the Free Software Foundation, |
|
19 |
* Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. |
|
20 |
* |
|
21 |
* Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA |
|
22 |
* or visit www.oracle.com if you need additional information or have any |
|
23 |
* questions. |
|
24 |
*/ |
|
25 |
||
26 |
package java.lang; |
|
27 |
||
28 |
/** |
|
32765 | 29 |
* Port of the "Freely Distributable Math Library", version 5.3, from |
30 |
* C to Java. |
|
32655 | 31 |
* |
32 |
* <p>The C version of fdlibm relied on the idiom of pointer aliasing |
|
33 |
* a 64-bit double floating-point value as a two-element array of |
|
34 |
* 32-bit integers and reading and writing the two halves of the |
|
35 |
* double independently. This coding pattern was problematic to C |
|
36 |
* optimizers and not directly expressible in Java. Therefore, rather |
|
37 |
* than a memory level overlay, if portions of a double need to be |
|
38 |
* operated on as integer values, the standard library methods for |
|
39 |
* bitwise floating-point to integer conversion, |
|
40 |
* Double.longBitsToDouble and Double.doubleToRawLongBits, are directly |
|
32765 | 41 |
* or indirectly used. |
32655 | 42 |
* |
43 |
* <p>The C version of fdlibm also took some pains to signal the |
|
44 |
* correct IEEE 754 exceptional conditions divide by zero, invalid, |
|
45 |
* overflow and underflow. For example, overflow would be signaled by |
|
46 |
* {@code huge * huge} where {@code huge} was a large constant that |
|
47 |
* would overflow when squared. Since IEEE floating-point exceptional |
|
48 |
* handling is not supported natively in the JVM, such coding patterns |
|
49 |
* have been omitted from this port. For example, rather than {@code |
|
50 |
* return huge * huge}, this port will use {@code return INFINITY}. |
|
32765 | 51 |
* |
52 |
* <p>Various comparison and arithmetic operations in fdlibm could be |
|
53 |
* done either based on the integer view of a value or directly on the |
|
54 |
* floating-point representation. Which idiom is faster may depend on |
|
55 |
* platform specific factors. However, for code clarity if no other |
|
56 |
* reason, this port will favor expressing the semantics of those |
|
57 |
* operations in terms of floating-point operations when convenient to |
|
58 |
* do so. |
|
32655 | 59 |
*/ |
60 |
class FdLibm { |
|
61 |
// Constants used by multiple algorithms |
|
62 |
private static final double INFINITY = Double.POSITIVE_INFINITY; |
|
63 |
||
64 |
private FdLibm() { |
|
32765 | 65 |
throw new UnsupportedOperationException("No FdLibm instances for you."); |
32655 | 66 |
} |
67 |
||
68 |
/** |
|
69 |
* Return the low-order 32 bits of the double argument as an int. |
|
70 |
*/ |
|
71 |
private static int __LO(double x) { |
|
72 |
long transducer = Double.doubleToRawLongBits(x); |
|
73 |
return (int)transducer; |
|
74 |
} |
|
75 |
||
76 |
/** |
|
77 |
* Return a double with its low-order bits of the second argument |
|
78 |
* and the high-order bits of the first argument.. |
|
79 |
*/ |
|
80 |
private static double __LO(double x, int low) { |
|
81 |
long transX = Double.doubleToRawLongBits(x); |
|
82 |
return Double.longBitsToDouble((transX & 0xFFFF_FFFF_0000_0000L)|low ); |
|
83 |
} |
|
84 |
||
85 |
/** |
|
86 |
* Return the high-order 32 bits of the double argument as an int. |
|
87 |
*/ |
|
88 |
private static int __HI(double x) { |
|
89 |
long transducer = Double.doubleToRawLongBits(x); |
|
90 |
return (int)(transducer >> 32); |
|
91 |
} |
|
92 |
||
93 |
/** |
|
94 |
* Return a double with its high-order bits of the second argument |
|
95 |
* and the low-order bits of the first argument.. |
|
96 |
*/ |
|
97 |
private static double __HI(double x, int high) { |
|
98 |
long transX = Double.doubleToRawLongBits(x); |
|
99 |
return Double.longBitsToDouble((transX & 0x0000_0000_FFFF_FFFFL)|( ((long)high)) << 32 ); |
|
100 |
} |
|
101 |
||
102 |
/** |
|
32992 | 103 |
* cbrt(x) |
104 |
* Return cube root of x |
|
105 |
*/ |
|
106 |
public static class Cbrt { |
|
107 |
// unsigned |
|
108 |
private static final int B1 = 715094163; /* B1 = (682-0.03306235651)*2**20 */ |
|
109 |
private static final int B2 = 696219795; /* B2 = (664-0.03306235651)*2**20 */ |
|
110 |
||
111 |
private static final double C = 0x1.15f15f15f15f1p-1; // 19/35 ~= 5.42857142857142815906e-01 |
|
112 |
private static final double D = -0x1.691de2532c834p-1; // -864/1225 ~= 7.05306122448979611050e-01 |
|
113 |
private static final double E = 0x1.6a0ea0ea0ea0fp0; // 99/70 ~= 1.41428571428571436819e+00 |
|
114 |
private static final double F = 0x1.9b6db6db6db6ep0; // 45/28 ~= 1.60714285714285720630e+00 |
|
115 |
private static final double G = 0x1.6db6db6db6db7p-2; // 5/14 ~= 3.57142857142857150787e-01 |
|
116 |
||
117 |
public static strictfp double compute(double x) { |
|
118 |
double t = 0.0; |
|
119 |
double sign; |
|
120 |
||
121 |
if (x == 0.0 || !Double.isFinite(x)) |
|
122 |
return x; // Handles signed zeros properly |
|
123 |
||
124 |
sign = (x < 0.0) ? -1.0: 1.0; |
|
125 |
||
126 |
x = Math.abs(x); // x <- |x| |
|
127 |
||
128 |
// Rough cbrt to 5 bits |
|
129 |
if (x < 0x1.0p-1022) { // subnormal number |
|
130 |
t = 0x1.0p54; // set t= 2**54 |
|
131 |
t *= x; |
|
132 |
t = __HI(t, __HI(t)/3 + B2); |
|
133 |
} else { |
|
134 |
int hx = __HI(x); // high word of x |
|
135 |
t = __HI(t, hx/3 + B1); |
|
136 |
} |
|
137 |
||
138 |
// New cbrt to 23 bits, may be implemented in single precision |
|
139 |
double r, s, w; |
|
140 |
r = t * t/x; |
|
141 |
s = C + r*t; |
|
142 |
t *= G + F/(s + E + D/s); |
|
143 |
||
144 |
// Chopped to 20 bits and make it larger than cbrt(x) |
|
145 |
t = __LO(t, 0); |
|
146 |
t = __HI(t, __HI(t) + 0x00000001); |
|
147 |
||
148 |
// One step newton iteration to 53 bits with error less than 0.667 ulps |
|
149 |
s = t * t; // t*t is exact |
|
150 |
r = x / s; |
|
151 |
w = t + t; |
|
152 |
r = (r - t)/(w + r); // r-s is exact |
|
153 |
t = t + t*r; |
|
154 |
||
155 |
// Restore the original sign bit |
|
156 |
return sign * t; |
|
157 |
} |
|
158 |
} |
|
159 |
||
160 |
/** |
|
32765 | 161 |
* hypot(x,y) |
162 |
* |
|
163 |
* Method : |
|
164 |
* If (assume round-to-nearest) z = x*x + y*y |
|
165 |
* has error less than sqrt(2)/2 ulp, than |
|
166 |
* sqrt(z) has error less than 1 ulp (exercise). |
|
167 |
* |
|
168 |
* So, compute sqrt(x*x + y*y) with some care as |
|
169 |
* follows to get the error below 1 ulp: |
|
170 |
* |
|
171 |
* Assume x > y > 0; |
|
172 |
* (if possible, set rounding to round-to-nearest) |
|
173 |
* 1. if x > 2y use |
|
174 |
* x1*x1 + (y*y + (x2*(x + x1))) for x*x + y*y |
|
175 |
* where x1 = x with lower 32 bits cleared, x2 = x - x1; else |
|
176 |
* 2. if x <= 2y use |
|
177 |
* t1*y1 + ((x-y) * (x-y) + (t1*y2 + t2*y)) |
|
178 |
* where t1 = 2x with lower 32 bits cleared, t2 = 2x - t1, |
|
179 |
* y1= y with lower 32 bits chopped, y2 = y - y1. |
|
180 |
* |
|
181 |
* NOTE: scaling may be necessary if some argument is too |
|
182 |
* large or too tiny |
|
183 |
* |
|
184 |
* Special cases: |
|
185 |
* hypot(x,y) is INF if x or y is +INF or -INF; else |
|
186 |
* hypot(x,y) is NAN if x or y is NAN. |
|
187 |
* |
|
188 |
* Accuracy: |
|
189 |
* hypot(x,y) returns sqrt(x^2 + y^2) with error less |
|
190 |
* than 1 ulp (unit in the last place) |
|
191 |
*/ |
|
192 |
public static class Hypot { |
|
193 |
public static final double TWO_MINUS_600 = 0x1.0p-600; |
|
194 |
public static final double TWO_PLUS_600 = 0x1.0p+600; |
|
195 |
||
196 |
public static strictfp double compute(double x, double y) { |
|
197 |
double a = Math.abs(x); |
|
198 |
double b = Math.abs(y); |
|
199 |
||
200 |
if (!Double.isFinite(a) || !Double.isFinite(b)) { |
|
201 |
if (a == INFINITY || b == INFINITY) |
|
202 |
return INFINITY; |
|
203 |
else |
|
204 |
return a + b; // Propagate NaN significand bits |
|
205 |
} |
|
206 |
||
207 |
if (b > a) { |
|
208 |
double tmp = a; |
|
209 |
a = b; |
|
210 |
b = tmp; |
|
211 |
} |
|
212 |
assert a >= b; |
|
213 |
||
214 |
// Doing bitwise conversion after screening for NaN allows |
|
215 |
// the code to not worry about the possibility of |
|
216 |
// "negative" NaN values. |
|
217 |
||
218 |
// Note: the ha and hb variables are the high-order |
|
219 |
// 32-bits of a and b stored as integer values. The ha and |
|
220 |
// hb values are used first for a rough magnitude |
|
221 |
// comparison of a and b and second for simulating higher |
|
222 |
// precision by allowing a and b, respectively, to be |
|
223 |
// decomposed into non-overlapping portions. Both of these |
|
224 |
// uses could be eliminated. The magnitude comparison |
|
225 |
// could be eliminated by extracting and comparing the |
|
226 |
// exponents of a and b or just be performing a |
|
227 |
// floating-point divide. Splitting a floating-point |
|
228 |
// number into non-overlapping portions can be |
|
229 |
// accomplished by judicious use of multiplies and |
|
230 |
// additions. For details see T. J. Dekker, A Floating |
|
231 |
// Point Technique for Extending the Available Precision , |
|
232 |
// Numerische Mathematik, vol. 18, 1971, pp.224-242 and |
|
233 |
// subsequent work. |
|
234 |
||
235 |
int ha = __HI(a); // high word of a |
|
236 |
int hb = __HI(b); // high word of b |
|
237 |
||
238 |
if ((ha - hb) > 0x3c00000) { |
|
239 |
return a + b; // x / y > 2**60 |
|
240 |
} |
|
241 |
||
242 |
int k = 0; |
|
32928
a3f03999ed62
8138823: Correct bug in port of fdlibm hypot to Java
darcy
parents:
32844
diff
changeset
|
243 |
if (a > 0x1.00000_ffff_ffffp500) { // a > ~2**500 |
32765 | 244 |
// scale a and b by 2**-600 |
245 |
ha -= 0x25800000; |
|
246 |
hb -= 0x25800000; |
|
247 |
a = a * TWO_MINUS_600; |
|
248 |
b = b * TWO_MINUS_600; |
|
249 |
k += 600; |
|
250 |
} |
|
251 |
double t1, t2; |
|
252 |
if (b < 0x1.0p-500) { // b < 2**-500 |
|
253 |
if (b < Double.MIN_NORMAL) { // subnormal b or 0 */ |
|
254 |
if (b == 0.0) |
|
255 |
return a; |
|
256 |
t1 = 0x1.0p1022; // t1 = 2^1022 |
|
257 |
b *= t1; |
|
258 |
a *= t1; |
|
259 |
k -= 1022; |
|
260 |
} else { // scale a and b by 2^600 |
|
261 |
ha += 0x25800000; // a *= 2^600 |
|
262 |
hb += 0x25800000; // b *= 2^600 |
|
263 |
a = a * TWO_PLUS_600; |
|
264 |
b = b * TWO_PLUS_600; |
|
265 |
k -= 600; |
|
266 |
} |
|
267 |
} |
|
268 |
// medium size a and b |
|
269 |
double w = a - b; |
|
270 |
if (w > b) { |
|
271 |
t1 = 0; |
|
272 |
t1 = __HI(t1, ha); |
|
273 |
t2 = a - t1; |
|
274 |
w = Math.sqrt(t1*t1 - (b*(-b) - t2 * (a + t1))); |
|
275 |
} else { |
|
276 |
double y1, y2; |
|
277 |
a = a + a; |
|
278 |
y1 = 0; |
|
279 |
y1 = __HI(y1, hb); |
|
280 |
y2 = b - y1; |
|
281 |
t1 = 0; |
|
282 |
t1 = __HI(t1, ha + 0x00100000); |
|
283 |
t2 = a - t1; |
|
284 |
w = Math.sqrt(t1*y1 - (w*(-w) - (t1*y2 + t2*b))); |
|
285 |
} |
|
286 |
if (k != 0) { |
|
287 |
return Math.powerOfTwoD(k) * w; |
|
288 |
} else |
|
289 |
return w; |
|
290 |
} |
|
291 |
} |
|
292 |
||
293 |
/** |
|
32655 | 294 |
* Compute x**y |
295 |
* n |
|
296 |
* Method: Let x = 2 * (1+f) |
|
297 |
* 1. Compute and return log2(x) in two pieces: |
|
298 |
* log2(x) = w1 + w2, |
|
299 |
* where w1 has 53 - 24 = 29 bit trailing zeros. |
|
32765 | 300 |
* 2. Perform y*log2(x) = n+y' by simulating multi-precision |
32655 | 301 |
* arithmetic, where |y'| <= 0.5. |
302 |
* 3. Return x**y = 2**n*exp(y'*log2) |
|
303 |
* |
|
304 |
* Special cases: |
|
305 |
* 1. (anything) ** 0 is 1 |
|
306 |
* 2. (anything) ** 1 is itself |
|
307 |
* 3. (anything) ** NAN is NAN |
|
308 |
* 4. NAN ** (anything except 0) is NAN |
|
309 |
* 5. +-(|x| > 1) ** +INF is +INF |
|
310 |
* 6. +-(|x| > 1) ** -INF is +0 |
|
311 |
* 7. +-(|x| < 1) ** +INF is +0 |
|
312 |
* 8. +-(|x| < 1) ** -INF is +INF |
|
313 |
* 9. +-1 ** +-INF is NAN |
|
314 |
* 10. +0 ** (+anything except 0, NAN) is +0 |
|
315 |
* 11. -0 ** (+anything except 0, NAN, odd integer) is +0 |
|
316 |
* 12. +0 ** (-anything except 0, NAN) is +INF |
|
317 |
* 13. -0 ** (-anything except 0, NAN, odd integer) is +INF |
|
318 |
* 14. -0 ** (odd integer) = -( +0 ** (odd integer) ) |
|
319 |
* 15. +INF ** (+anything except 0,NAN) is +INF |
|
320 |
* 16. +INF ** (-anything except 0,NAN) is +0 |
|
321 |
* 17. -INF ** (anything) = -0 ** (-anything) |
|
322 |
* 18. (-anything) ** (integer) is (-1)**(integer)*(+anything**integer) |
|
323 |
* 19. (-anything except 0 and inf) ** (non-integer) is NAN |
|
324 |
* |
|
325 |
* Accuracy: |
|
326 |
* pow(x,y) returns x**y nearly rounded. In particular |
|
327 |
* pow(integer,integer) |
|
328 |
* always returns the correct integer provided it is |
|
329 |
* representable. |
|
330 |
*/ |
|
331 |
public static class Pow { |
|
332 |
public static strictfp double compute(final double x, final double y) { |
|
333 |
double z; |
|
334 |
double r, s, t, u, v, w; |
|
335 |
int i, j, k, n; |
|
336 |
||
337 |
// y == zero: x**0 = 1 |
|
338 |
if (y == 0.0) |
|
339 |
return 1.0; |
|
340 |
||
341 |
// +/-NaN return x + y to propagate NaN significands |
|
342 |
if (Double.isNaN(x) || Double.isNaN(y)) |
|
343 |
return x + y; |
|
344 |
||
345 |
final double y_abs = Math.abs(y); |
|
346 |
double x_abs = Math.abs(x); |
|
347 |
// Special values of y |
|
348 |
if (y == 2.0) { |
|
349 |
return x * x; |
|
350 |
} else if (y == 0.5) { |
|
351 |
if (x >= -Double.MAX_VALUE) // Handle x == -infinity later |
|
352 |
return Math.sqrt(x + 0.0); // Add 0.0 to properly handle x == -0.0 |
|
353 |
} else if (y_abs == 1.0) { // y is +/-1 |
|
354 |
return (y == 1.0) ? x : 1.0 / x; |
|
355 |
} else if (y_abs == INFINITY) { // y is +/-infinity |
|
356 |
if (x_abs == 1.0) |
|
357 |
return y - y; // inf**+/-1 is NaN |
|
358 |
else if (x_abs > 1.0) // (|x| > 1)**+/-inf = inf, 0 |
|
359 |
return (y >= 0) ? y : 0.0; |
|
360 |
else // (|x| < 1)**-/+inf = inf, 0 |
|
361 |
return (y < 0) ? -y : 0.0; |
|
362 |
} |
|
363 |
||
364 |
final int hx = __HI(x); |
|
365 |
int ix = hx & 0x7fffffff; |
|
366 |
||
367 |
/* |
|
368 |
* When x < 0, determine if y is an odd integer: |
|
369 |
* y_is_int = 0 ... y is not an integer |
|
370 |
* y_is_int = 1 ... y is an odd int |
|
371 |
* y_is_int = 2 ... y is an even int |
|
372 |
*/ |
|
373 |
int y_is_int = 0; |
|
374 |
if (hx < 0) { |
|
375 |
if (y_abs >= 0x1.0p53) // |y| >= 2^53 = 9.007199254740992E15 |
|
376 |
y_is_int = 2; // y is an even integer since ulp(2^53) = 2.0 |
|
377 |
else if (y_abs >= 1.0) { // |y| >= 1.0 |
|
378 |
long y_abs_as_long = (long) y_abs; |
|
379 |
if ( ((double) y_abs_as_long) == y_abs) { |
|
380 |
y_is_int = 2 - (int)(y_abs_as_long & 0x1L); |
|
381 |
} |
|
382 |
} |
|
383 |
} |
|
384 |
||
385 |
// Special value of x |
|
386 |
if (x_abs == 0.0 || |
|
387 |
x_abs == INFINITY || |
|
388 |
x_abs == 1.0) { |
|
389 |
z = x_abs; // x is +/-0, +/-inf, +/-1 |
|
390 |
if (y < 0.0) |
|
391 |
z = 1.0/z; // z = (1/|x|) |
|
392 |
if (hx < 0) { |
|
393 |
if (((ix - 0x3ff00000) | y_is_int) == 0) { |
|
394 |
z = (z-z)/(z-z); // (-1)**non-int is NaN |
|
395 |
} else if (y_is_int == 1) |
|
396 |
z = -1.0 * z; // (x < 0)**odd = -(|x|**odd) |
|
397 |
} |
|
398 |
return z; |
|
399 |
} |
|
400 |
||
401 |
n = (hx >> 31) + 1; |
|
402 |
||
403 |
// (x < 0)**(non-int) is NaN |
|
404 |
if ((n | y_is_int) == 0) |
|
405 |
return (x-x)/(x-x); |
|
406 |
||
407 |
s = 1.0; // s (sign of result -ve**odd) = -1 else = 1 |
|
408 |
if ( (n | (y_is_int - 1)) == 0) |
|
409 |
s = -1.0; // (-ve)**(odd int) |
|
410 |
||
411 |
double p_h, p_l, t1, t2; |
|
412 |
// |y| is huge |
|
32844 | 413 |
if (y_abs > 0x1.00000_ffff_ffffp31) { // if |y| > ~2**31 |
32655 | 414 |
final double INV_LN2 = 0x1.7154_7652_b82fep0; // 1.44269504088896338700e+00 = 1/ln2 |
415 |
final double INV_LN2_H = 0x1.715476p0; // 1.44269502162933349609e+00 = 24 bits of 1/ln2 |
|
416 |
final double INV_LN2_L = 0x1.4ae0_bf85_ddf44p-26; // 1.92596299112661746887e-08 = 1/ln2 tail |
|
417 |
||
418 |
// Over/underflow if x is not close to one |
|
32844 | 419 |
if (x_abs < 0x1.fffff_0000_0000p-1) // |x| < ~0.9999995231628418 |
32655 | 420 |
return (y < 0.0) ? s * INFINITY : s * 0.0; |
32844 | 421 |
if (x_abs > 0x1.00000_ffff_ffffp0) // |x| > ~1.0 |
32655 | 422 |
return (y > 0.0) ? s * INFINITY : s * 0.0; |
423 |
/* |
|
424 |
* now |1-x| is tiny <= 2**-20, sufficient to compute |
|
425 |
* log(x) by x - x^2/2 + x^3/3 - x^4/4 |
|
426 |
*/ |
|
427 |
t = x_abs - 1.0; // t has 20 trailing zeros |
|
428 |
w = (t * t) * (0.5 - t * (0.3333333333333333333333 - t * 0.25)); |
|
429 |
u = INV_LN2_H * t; // INV_LN2_H has 21 sig. bits |
|
430 |
v = t * INV_LN2_L - w * INV_LN2; |
|
431 |
t1 = u + v; |
|
432 |
t1 =__LO(t1, 0); |
|
433 |
t2 = v - (t1 - u); |
|
434 |
} else { |
|
435 |
final double CP = 0x1.ec70_9dc3_a03fdp-1; // 9.61796693925975554329e-01 = 2/(3ln2) |
|
436 |
final double CP_H = 0x1.ec709ep-1; // 9.61796700954437255859e-01 = (float)cp |
|
437 |
final double CP_L = -0x1.e2fe_0145_b01f5p-28; // -7.02846165095275826516e-09 = tail of CP_H |
|
438 |
||
439 |
double z_h, z_l, ss, s2, s_h, s_l, t_h, t_l; |
|
440 |
n = 0; |
|
441 |
// Take care of subnormal numbers |
|
442 |
if (ix < 0x00100000) { |
|
443 |
x_abs *= 0x1.0p53; // 2^53 = 9007199254740992.0 |
|
444 |
n -= 53; |
|
445 |
ix = __HI(x_abs); |
|
446 |
} |
|
447 |
n += ((ix) >> 20) - 0x3ff; |
|
448 |
j = ix & 0x000fffff; |
|
449 |
// Determine interval |
|
450 |
ix = j | 0x3ff00000; // Normalize ix |
|
451 |
if (j <= 0x3988E) |
|
452 |
k = 0; // |x| <sqrt(3/2) |
|
453 |
else if (j < 0xBB67A) |
|
454 |
k = 1; // |x| <sqrt(3) |
|
455 |
else { |
|
456 |
k = 0; |
|
457 |
n += 1; |
|
458 |
ix -= 0x00100000; |
|
459 |
} |
|
460 |
x_abs = __HI(x_abs, ix); |
|
461 |
||
462 |
// Compute ss = s_h + s_l = (x-1)/(x+1) or (x-1.5)/(x+1.5) |
|
463 |
||
464 |
final double BP[] = {1.0, |
|
465 |
1.5}; |
|
466 |
final double DP_H[] = {0.0, |
|
467 |
0x1.2b80_34p-1}; // 5.84962487220764160156e-01 |
|
468 |
final double DP_L[] = {0.0, |
|
469 |
0x1.cfde_b43c_fd006p-27};// 1.35003920212974897128e-08 |
|
470 |
||
471 |
// Poly coefs for (3/2)*(log(x)-2s-2/3*s**3 |
|
472 |
final double L1 = 0x1.3333_3333_33303p-1; // 5.99999999999994648725e-01 |
|
473 |
final double L2 = 0x1.b6db_6db6_fabffp-2; // 4.28571428578550184252e-01 |
|
474 |
final double L3 = 0x1.5555_5518_f264dp-2; // 3.33333329818377432918e-01 |
|
475 |
final double L4 = 0x1.1746_0a91_d4101p-2; // 2.72728123808534006489e-01 |
|
476 |
final double L5 = 0x1.d864_a93c_9db65p-3; // 2.30660745775561754067e-01 |
|
477 |
final double L6 = 0x1.a7e2_84a4_54eefp-3; // 2.06975017800338417784e-01 |
|
478 |
u = x_abs - BP[k]; // BP[0]=1.0, BP[1]=1.5 |
|
479 |
v = 1.0 / (x_abs + BP[k]); |
|
480 |
ss = u * v; |
|
481 |
s_h = ss; |
|
482 |
s_h = __LO(s_h, 0); |
|
483 |
// t_h=x_abs + BP[k] High |
|
484 |
t_h = 0.0; |
|
485 |
t_h = __HI(t_h, ((ix >> 1) | 0x20000000) + 0x00080000 + (k << 18) ); |
|
486 |
t_l = x_abs - (t_h - BP[k]); |
|
487 |
s_l = v * ((u - s_h * t_h) - s_h * t_l); |
|
488 |
// Compute log(x_abs) |
|
489 |
s2 = ss * ss; |
|
490 |
r = s2 * s2* (L1 + s2 * (L2 + s2 * (L3 + s2 * (L4 + s2 * (L5 + s2 * L6))))); |
|
491 |
r += s_l * (s_h + ss); |
|
492 |
s2 = s_h * s_h; |
|
493 |
t_h = 3.0 + s2 + r; |
|
494 |
t_h = __LO(t_h, 0); |
|
495 |
t_l = r - ((t_h - 3.0) - s2); |
|
496 |
// u+v = ss*(1+...) |
|
497 |
u = s_h * t_h; |
|
498 |
v = s_l * t_h + t_l * ss; |
|
499 |
// 2/(3log2)*(ss + ...) |
|
500 |
p_h = u + v; |
|
501 |
p_h = __LO(p_h, 0); |
|
502 |
p_l = v - (p_h - u); |
|
503 |
z_h = CP_H * p_h; // CP_H + CP_L = 2/(3*log2) |
|
504 |
z_l = CP_L * p_h + p_l * CP + DP_L[k]; |
|
505 |
// log2(x_abs) = (ss + ..)*2/(3*log2) = n + DP_H + z_h + z_l |
|
506 |
t = (double)n; |
|
507 |
t1 = (((z_h + z_l) + DP_H[k]) + t); |
|
508 |
t1 = __LO(t1, 0); |
|
509 |
t2 = z_l - (((t1 - t) - DP_H[k]) - z_h); |
|
510 |
} |
|
511 |
||
512 |
// Split up y into (y1 + y2) and compute (y1 + y2) * (t1 + t2) |
|
513 |
double y1 = y; |
|
514 |
y1 = __LO(y1, 0); |
|
515 |
p_l = (y - y1) * t1 + y * t2; |
|
516 |
p_h = y1 * t1; |
|
517 |
z = p_l + p_h; |
|
518 |
j = __HI(z); |
|
519 |
i = __LO(z); |
|
520 |
if (j >= 0x40900000) { // z >= 1024 |
|
521 |
if (((j - 0x40900000) | i)!=0) // if z > 1024 |
|
522 |
return s * INFINITY; // Overflow |
|
523 |
else { |
|
524 |
final double OVT = 8.0085662595372944372e-0017; // -(1024-log2(ovfl+.5ulp)) |
|
525 |
if (p_l + OVT > z - p_h) |
|
526 |
return s * INFINITY; // Overflow |
|
527 |
} |
|
528 |
} else if ((j & 0x7fffffff) >= 0x4090cc00 ) { // z <= -1075 |
|
529 |
if (((j - 0xc090cc00) | i)!=0) // z < -1075 |
|
530 |
return s * 0.0; // Underflow |
|
531 |
else { |
|
532 |
if (p_l <= z - p_h) |
|
533 |
return s * 0.0; // Underflow |
|
534 |
} |
|
535 |
} |
|
536 |
/* |
|
537 |
* Compute 2**(p_h+p_l) |
|
538 |
*/ |
|
539 |
// Poly coefs for (3/2)*(log(x)-2s-2/3*s**3 |
|
540 |
final double P1 = 0x1.5555_5555_5553ep-3; // 1.66666666666666019037e-01 |
|
541 |
final double P2 = -0x1.6c16_c16b_ebd93p-9; // -2.77777777770155933842e-03 |
|
542 |
final double P3 = 0x1.1566_aaf2_5de2cp-14; // 6.61375632143793436117e-05 |
|
543 |
final double P4 = -0x1.bbd4_1c5d_26bf1p-20; // -1.65339022054652515390e-06 |
|
544 |
final double P5 = 0x1.6376_972b_ea4d0p-25; // 4.13813679705723846039e-08 |
|
545 |
final double LG2 = 0x1.62e4_2fef_a39efp-1; // 6.93147180559945286227e-01 |
|
546 |
final double LG2_H = 0x1.62e43p-1; // 6.93147182464599609375e-01 |
|
547 |
final double LG2_L = -0x1.05c6_10ca_86c39p-29; // -1.90465429995776804525e-09 |
|
548 |
i = j & 0x7fffffff; |
|
549 |
k = (i >> 20) - 0x3ff; |
|
550 |
n = 0; |
|
551 |
if (i > 0x3fe00000) { // if |z| > 0.5, set n = [z + 0.5] |
|
552 |
n = j + (0x00100000 >> (k + 1)); |
|
553 |
k = ((n & 0x7fffffff) >> 20) - 0x3ff; // new k for n |
|
554 |
t = 0.0; |
|
555 |
t = __HI(t, (n & ~(0x000fffff >> k)) ); |
|
556 |
n = ((n & 0x000fffff) | 0x00100000) >> (20 - k); |
|
557 |
if (j < 0) |
|
558 |
n = -n; |
|
559 |
p_h -= t; |
|
560 |
} |
|
561 |
t = p_l + p_h; |
|
562 |
t = __LO(t, 0); |
|
563 |
u = t * LG2_H; |
|
564 |
v = (p_l - (t - p_h)) * LG2 + t * LG2_L; |
|
565 |
z = u + v; |
|
566 |
w = v - (z - u); |
|
567 |
t = z * z; |
|
568 |
t1 = z - t * (P1 + t * (P2 + t * (P3 + t * (P4 + t * P5)))); |
|
569 |
r = (z * t1)/(t1 - 2.0) - (w + z * w); |
|
570 |
z = 1.0 - (r - z); |
|
571 |
j = __HI(z); |
|
572 |
j += (n << 20); |
|
573 |
if ((j >> 20) <= 0) |
|
574 |
z = Math.scalb(z, n); // subnormal output |
|
575 |
else { |
|
576 |
int z_hi = __HI(z); |
|
577 |
z_hi += (n << 20); |
|
578 |
z = __HI(z, z_hi); |
|
579 |
} |
|
580 |
return s * z; |
|
581 |
} |
|
582 |
} |
|
583 |
} |