• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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)
bits_of_precision(float a,float b)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)
bits_of_precision(double a,double b)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)
pow_cbrtf(float x)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)
pow_cbrtd(double x)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
cbrt_5f(float f)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
cbrt_5d(double d)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)
cbrta_halleyf(const float a,const float R)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)
cbrta_halleyd(const double a,const double R)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)
cbrta_newtonf(const float a,const float x)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)
cbrta_newtond(const double a,const double x)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)
halley_cbrt1d(double d)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)
halley_cbrt1f(float d)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)
halley_cbrt2d(double d)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)
halley_cbrt3d(double d)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)
halley_cbrt2f(float d)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)
newton_cbrt1d(double d)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)
newton_cbrt2d(double d)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)
newton_cbrt3d(double d)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)
newton_cbrt4d(double d)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)
newton_cbrt1f(float d)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)
newton_cbrt2f(float d)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)
newton_cbrt3f(float d)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)
newton_cbrt4f(float d)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 
TestCubeRootf(const char * szName,cuberootfnf cbrt,double rA,double rB,int rN)242 static double TestCubeRootf(const char* szName, cuberootfnf cbrt, double rA, double 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 
TestCubeRootd(const char * szName,cuberootfnd cbrt,double rA,double rB,int rN)291 static double TestCubeRootd(const char* szName, cuberootfnd cbrt, double rA, double 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 
_tmain()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 
cube_root(double x)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 */
icbrt(int n)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
401