OLD | NEW |
| (Empty) |
1 /* | |
2 * Copyright 2012 Google Inc. | |
3 * | |
4 * Use of this source code is governed by a BSD-style license that can be | |
5 * found in the LICENSE file. | |
6 */ | |
7 // http://metamerist.com/cbrt/CubeRoot.cpp | |
8 // | |
9 | |
10 #include <math.h> | |
11 #include "CubicUtilities.h" | |
12 | |
13 #define TEST_ALTERNATIVES 0 | |
14 #if TEST_ALTERNATIVES | |
15 typedef float (*cuberootfnf) (float); | |
16 typedef double (*cuberootfnd) (double); | |
17 | |
18 // estimate bits of precision (32-bit float case) | |
19 inline int bits_of_precision(float a, float b) | |
20 { | |
21 const double kd = 1.0 / log(2.0); | |
22 | |
23 if (a==b) | |
24 return 23; | |
25 | |
26 const double kdmin = pow(2.0, -23.0); | |
27 | |
28 double d = fabs(a-b); | |
29 if (d < kdmin) | |
30 return 23; | |
31 | |
32 return int(-log(d)*kd); | |
33 } | |
34 | |
35 // estiamte bits of precision (64-bit double case) | |
36 inline int bits_of_precision(double a, double b) | |
37 { | |
38 const double kd = 1.0 / log(2.0); | |
39 | |
40 if (a==b) | |
41 return 52; | |
42 | |
43 const double kdmin = pow(2.0, -52.0); | |
44 | |
45 double d = fabs(a-b); | |
46 if (d < kdmin) | |
47 return 52; | |
48 | |
49 return int(-log(d)*kd); | |
50 } | |
51 | |
52 // cube root via x^(1/3) | |
53 static float pow_cbrtf(float x) | |
54 { | |
55 return (float) pow(x, 1.0f/3.0f); | |
56 } | |
57 | |
58 // cube root via x^(1/3) | |
59 static double pow_cbrtd(double x) | |
60 { | |
61 return pow(x, 1.0/3.0); | |
62 } | |
63 | |
64 // cube root approximation using bit hack for 32-bit float | |
65 static float cbrt_5f(float f) | |
66 { | |
67 unsigned int* p = (unsigned int *) &f; | |
68 *p = *p/3 + 709921077; | |
69 return f; | |
70 } | |
71 #endif | |
72 | |
73 // cube root approximation using bit hack for 64-bit float | |
74 // adapted from Kahan's cbrt | |
75 static double cbrt_5d(double d) | |
76 { | |
77 const unsigned int B1 = 715094163; | |
78 double t = 0.0; | |
79 unsigned int* pt = (unsigned int*) &t; | |
80 unsigned int* px = (unsigned int*) &d; | |
81 pt[1]=px[1]/3+B1; | |
82 return t; | |
83 } | |
84 | |
85 #if TEST_ALTERNATIVES | |
86 // cube root approximation using bit hack for 64-bit float | |
87 // adapted from Kahan's cbrt | |
88 #if 0 | |
89 static double quint_5d(double d) | |
90 { | |
91 return sqrt(sqrt(d)); | |
92 | |
93 const unsigned int B1 = 71509416*5/3; | |
94 double t = 0.0; | |
95 unsigned int* pt = (unsigned int*) &t; | |
96 unsigned int* px = (unsigned int*) &d; | |
97 pt[1]=px[1]/5+B1; | |
98 return t; | |
99 } | |
100 #endif | |
101 | |
102 // iterative cube root approximation using Halley's method (float) | |
103 static float cbrta_halleyf(const float a, const float R) | |
104 { | |
105 const float a3 = a*a*a; | |
106 const float b= a * (a3 + R + R) / (a3 + a3 + R); | |
107 return b; | |
108 } | |
109 #endif | |
110 | |
111 // iterative cube root approximation using Halley's method (double) | |
112 static double cbrta_halleyd(const double a, const double R) | |
113 { | |
114 const double a3 = a*a*a; | |
115 const double b= a * (a3 + R + R) / (a3 + a3 + R); | |
116 return b; | |
117 } | |
118 | |
119 #if TEST_ALTERNATIVES | |
120 // iterative cube root approximation using Newton's method (float) | |
121 static float cbrta_newtonf(const float a, const float x) | |
122 { | |
123 // return (1.0 / 3.0) * ((a + a) + x / (a * a)); | |
124 return a - (1.0f / 3.0f) * (a - x / (a*a)); | |
125 } | |
126 | |
127 // iterative cube root approximation using Newton's method (double) | |
128 static double cbrta_newtond(const double a, const double x) | |
129 { | |
130 return (1.0/3.0) * (x / (a*a) + 2*a); | |
131 } | |
132 | |
133 // cube root approximation using 1 iteration of Halley's method (double) | |
134 static double halley_cbrt1d(double d) | |
135 { | |
136 double a = cbrt_5d(d); | |
137 return cbrta_halleyd(a, d); | |
138 } | |
139 | |
140 // cube root approximation using 1 iteration of Halley's method (float) | |
141 static float halley_cbrt1f(float d) | |
142 { | |
143 float a = cbrt_5f(d); | |
144 return cbrta_halleyf(a, d); | |
145 } | |
146 | |
147 // cube root approximation using 2 iterations of Halley's method (double) | |
148 static double halley_cbrt2d(double d) | |
149 { | |
150 double a = cbrt_5d(d); | |
151 a = cbrta_halleyd(a, d); | |
152 return cbrta_halleyd(a, d); | |
153 } | |
154 #endif | |
155 | |
156 // cube root approximation using 3 iterations of Halley's method (double) | |
157 static double halley_cbrt3d(double d) | |
158 { | |
159 double a = cbrt_5d(d); | |
160 a = cbrta_halleyd(a, d); | |
161 a = cbrta_halleyd(a, d); | |
162 return cbrta_halleyd(a, d); | |
163 } | |
164 | |
165 #if TEST_ALTERNATIVES | |
166 // cube root approximation using 2 iterations of Halley's method (float) | |
167 static float halley_cbrt2f(float d) | |
168 { | |
169 float a = cbrt_5f(d); | |
170 a = cbrta_halleyf(a, d); | |
171 return cbrta_halleyf(a, d); | |
172 } | |
173 | |
174 // cube root approximation using 1 iteration of Newton's method (double) | |
175 static double newton_cbrt1d(double d) | |
176 { | |
177 double a = cbrt_5d(d); | |
178 return cbrta_newtond(a, d); | |
179 } | |
180 | |
181 // cube root approximation using 2 iterations of Newton's method (double) | |
182 static double newton_cbrt2d(double d) | |
183 { | |
184 double a = cbrt_5d(d); | |
185 a = cbrta_newtond(a, d); | |
186 return cbrta_newtond(a, d); | |
187 } | |
188 | |
189 // cube root approximation using 3 iterations of Newton's method (double) | |
190 static double newton_cbrt3d(double d) | |
191 { | |
192 double a = cbrt_5d(d); | |
193 a = cbrta_newtond(a, d); | |
194 a = cbrta_newtond(a, d); | |
195 return cbrta_newtond(a, d); | |
196 } | |
197 | |
198 // cube root approximation using 4 iterations of Newton's method (double) | |
199 static double newton_cbrt4d(double d) | |
200 { | |
201 double a = cbrt_5d(d); | |
202 a = cbrta_newtond(a, d); | |
203 a = cbrta_newtond(a, d); | |
204 a = cbrta_newtond(a, d); | |
205 return cbrta_newtond(a, d); | |
206 } | |
207 | |
208 // cube root approximation using 2 iterations of Newton's method (float) | |
209 static float newton_cbrt1f(float d) | |
210 { | |
211 float a = cbrt_5f(d); | |
212 return cbrta_newtonf(a, d); | |
213 } | |
214 | |
215 // cube root approximation using 2 iterations of Newton's method (float) | |
216 static float newton_cbrt2f(float d) | |
217 { | |
218 float a = cbrt_5f(d); | |
219 a = cbrta_newtonf(a, d); | |
220 return cbrta_newtonf(a, d); | |
221 } | |
222 | |
223 // cube root approximation using 3 iterations of Newton's method (float) | |
224 static float newton_cbrt3f(float d) | |
225 { | |
226 float a = cbrt_5f(d); | |
227 a = cbrta_newtonf(a, d); | |
228 a = cbrta_newtonf(a, d); | |
229 return cbrta_newtonf(a, d); | |
230 } | |
231 | |
232 // cube root approximation using 4 iterations of Newton's method (float) | |
233 static float newton_cbrt4f(float d) | |
234 { | |
235 float a = cbrt_5f(d); | |
236 a = cbrta_newtonf(a, d); | |
237 a = cbrta_newtonf(a, d); | |
238 a = cbrta_newtonf(a, d); | |
239 return cbrta_newtonf(a, d); | |
240 } | |
241 | |
242 static double TestCubeRootf(const char* szName, cuberootfnf cbrt, double rA, dou
ble rB, int rN) | |
243 { | |
244 const int N = rN; | |
245 | |
246 float dd = float((rB-rA) / N); | |
247 | |
248 // calculate 1M numbers | |
249 int i=0; | |
250 float d = (float) rA; | |
251 | |
252 double s = 0.0; | |
253 | |
254 for(d=(float) rA, i=0; i<N; i++, d += dd) | |
255 { | |
256 s += cbrt(d); | |
257 } | |
258 | |
259 double bits = 0.0; | |
260 double worstx=0.0; | |
261 double worsty=0.0; | |
262 int minbits=64; | |
263 | |
264 for(d=(float) rA, i=0; i<N; i++, d += dd) | |
265 { | |
266 float a = cbrt((float) d); | |
267 float b = (float) pow((double) d, 1.0/3.0); | |
268 | |
269 int bc = bits_of_precision(a, b); | |
270 bits += bc; | |
271 | |
272 if (b > 1.0e-6) | |
273 { | |
274 if (bc < minbits) | |
275 { | |
276 minbits = bc; | |
277 worstx = d; | |
278 worsty = a; | |
279 } | |
280 } | |
281 } | |
282 | |
283 bits /= N; | |
284 | |
285 printf(" %3d mbp %6.3f abp\n", minbits, bits); | |
286 | |
287 return s; | |
288 } | |
289 | |
290 | |
291 static double TestCubeRootd(const char* szName, cuberootfnd cbrt, double rA, dou
ble rB, int rN) | |
292 { | |
293 const int N = rN; | |
294 | |
295 double dd = (rB-rA) / N; | |
296 | |
297 int i=0; | |
298 | |
299 double s = 0.0; | |
300 double d = 0.0; | |
301 | |
302 for(d=rA, i=0; i<N; i++, d += dd) | |
303 { | |
304 s += cbrt(d); | |
305 } | |
306 | |
307 | |
308 double bits = 0.0; | |
309 double worstx = 0.0; | |
310 double worsty = 0.0; | |
311 int minbits = 64; | |
312 for(d=rA, i=0; i<N; i++, d += dd) | |
313 { | |
314 double a = cbrt(d); | |
315 double b = pow(d, 1.0/3.0); | |
316 | |
317 int bc = bits_of_precision(a, b); // min(53, count_matching_bitsd(a, b)
- 12); | |
318 bits += bc; | |
319 | |
320 if (b > 1.0e-6) | |
321 { | |
322 if (bc < minbits) | |
323 { | |
324 bits_of_precision(a, b); | |
325 minbits = bc; | |
326 worstx = d; | |
327 worsty = a; | |
328 } | |
329 } | |
330 } | |
331 | |
332 bits /= N; | |
333 | |
334 printf(" %3d mbp %6.3f abp\n", minbits, bits); | |
335 | |
336 return s; | |
337 } | |
338 | |
339 static int _tmain() | |
340 { | |
341 // a million uniform steps through the range from 0.0 to 1.0 | |
342 // (doing uniform steps in the log scale would be better) | |
343 double a = 0.0; | |
344 double b = 1.0; | |
345 int n = 1000000; | |
346 | |
347 printf("32-bit float tests\n"); | |
348 printf("----------------------------------------\n"); | |
349 TestCubeRootf("cbrt_5f", cbrt_5f, a, b, n); | |
350 TestCubeRootf("pow", pow_cbrtf, a, b, n); | |
351 TestCubeRootf("halley x 1", halley_cbrt1f, a, b, n); | |
352 TestCubeRootf("halley x 2", halley_cbrt2f, a, b, n); | |
353 TestCubeRootf("newton x 1", newton_cbrt1f, a, b, n); | |
354 TestCubeRootf("newton x 2", newton_cbrt2f, a, b, n); | |
355 TestCubeRootf("newton x 3", newton_cbrt3f, a, b, n); | |
356 TestCubeRootf("newton x 4", newton_cbrt4f, a, b, n); | |
357 printf("\n\n"); | |
358 | |
359 printf("64-bit double tests\n"); | |
360 printf("----------------------------------------\n"); | |
361 TestCubeRootd("cbrt_5d", cbrt_5d, a, b, n); | |
362 TestCubeRootd("pow", pow_cbrtd, a, b, n); | |
363 TestCubeRootd("halley x 1", halley_cbrt1d, a, b, n); | |
364 TestCubeRootd("halley x 2", halley_cbrt2d, a, b, n); | |
365 TestCubeRootd("halley x 3", halley_cbrt3d, a, b, n); | |
366 TestCubeRootd("newton x 1", newton_cbrt1d, a, b, n); | |
367 TestCubeRootd("newton x 2", newton_cbrt2d, a, b, n); | |
368 TestCubeRootd("newton x 3", newton_cbrt3d, a, b, n); | |
369 TestCubeRootd("newton x 4", newton_cbrt4d, a, b, n); | |
370 printf("\n\n"); | |
371 | |
372 return 0; | |
373 } | |
374 #endif | |
375 | |
376 double cube_root(double x) { | |
377 if (approximately_zero_cubed(x)) { | |
378 return 0; | |
379 } | |
380 double result = halley_cbrt3d(fabs(x)); | |
381 if (x < 0) { | |
382 result = -result; | |
383 } | |
384 return result; | |
385 } | |
386 | |
387 #if TEST_ALTERNATIVES | |
388 // http://bytes.com/topic/c/answers/754588-tips-find-cube-root-program-using-c | |
389 /* cube root */ | |
390 int icbrt(int n) { | |
391 int t=0, x=(n+2)/3; /* works for n=0 and n>=1 */ | |
392 for(; t!=x;) { | |
393 int x3=x*x*x; | |
394 t=x; | |
395 x*=(2*n + x3); | |
396 x/=(2*x3 + n); | |
397 } | |
398 return x ; /* always(?) equal to floor(n^(1/3)) */ | |
399 } | |
400 #endif | |
OLD | NEW |