xref: /csrg-svn/lib/libm/common/trig.c (revision 24720)
1 /*
2  * Copyright (c) 1985 Regents of the University of California.
3  *
4  * Use and reproduction of this software are granted  in  accordance  with
5  * the terms and conditions specified in  the  Berkeley  Software  License
6  * Agreement (in particular, this entails acknowledgement of the programs'
7  * source, and inclusion of this notice) with the additional understanding
8  * that  all  recipients  should regard themselves as participants  in  an
9  * ongoing  research  project and hence should  feel  obligated  to report
10  * their  experiences (good or bad) with these elementary function  codes,
11  * using "sendbug 4bsd-bugs@BERKELEY", to the authors.
12  */
13 
14 #ifndef lint
15 static char sccsid[] =
16 "@(#)trig.c	1.2 (Berkeley) 8/22/85; 1.2 (ucb.elefunt) 09/12/85";
17 #endif not lint
18 
19 /* SIN(X), COS(X), TAN(X)
20  * RETURN THE SINE, COSINE, AND TANGENT OF X RESPECTIVELY
21  * DOUBLE PRECISION (VAX D format 56 bits, IEEE DOUBLE 53 BITS)
22  * CODED IN C BY K.C. NG, 1/8/85;
23  * REVISED BY W. Kahan and K.C. NG, 8/17/85.
24  *
25  * Required system supported functions:
26  *      copysign(x,y)
27  *      finite(x)
28  *      drem(x,p)
29  *
30  * Static kernel functions:
31  *      sin__S(z)       ....sin__S(x*x) return (sin(x)-x)/x
32  *      cos__C(z)       ....cos__C(x*x) return cos(x)-1-x*x/2
33  *
34  * Method.
35  *      Let S and C denote the polynomial approximations to sin and cos
36  *      respectively on [-PI/4, +PI/4].
37  *
38  *      SIN and COS:
39  *      1. Reduce the argument into [-PI , +PI] by the remainder function.
40  *      2. For x in (-PI,+PI), there are three cases:
41  *			case 1:	|x| < PI/4
42  *			case 2:	PI/4 <= |x| < 3PI/4
43  *			case 3:	3PI/4 <= |x|.
44  *	   SIN and COS of x are computed by:
45  *
46  *                   sin(x)      cos(x)       remark
47  *     ----------------------------------------------------------
48  *        case 1     S(x)         C(x)
49  *        case 2 sign(x)*C(y)     S(y)      y=PI/2-|x|
50  *        case 3     S(y)        -C(y)      y=sign(x)*(PI-|x|)
51  *     ----------------------------------------------------------
52  *
53  *      TAN:
54  *      1. Reduce the argument into [-PI/2 , +PI/2] by the remainder function.
55  *      2. For x in (-PI/2,+PI/2), there are two cases:
56  *			case 1:	|x| < PI/4
57  *			case 2:	PI/4 <= |x| < PI/2
58  *         TAN of x is computed by:
59  *
60  *                   tan (x)            remark
61  *     ----------------------------------------------------------
62  *        case 1     S(x)/C(x)
63  *        case 2     C(y)/S(y)     y=sign(x)*(PI/2-|x|)
64  *     ----------------------------------------------------------
65  *
66  *   Notes:
67  *      1. S(y) and C(y) were computed by:
68  *              S(y) = y+y*sin__S(y*y)
69  *              C(y) = 1-(y*y/2-cos__C(x*x))          ... if y*y/2 <  thresh,
70  *                   = 0.5-((y*y/2-0.5)-cos__C(x*x))  ... if y*y/2 >= thresh.
71  *         where
72  *              thresh = 0.5*(acos(3/4)**2)
73  *
74  *      2. For better accuracy, we use the following formula for S/C for tan
75  *         (k=0): let ss=sin__S(y*y), and cc=cos__C(y*y), then
76  *
77  *                            y+y*ss             (y*y/2-cc)+ss
78  *             S(y)/C(y)   = -------- = y + y * ---------------.
79  *                               C                     C
80  *
81  *
82  * Special cases:
83  *      Let trig be any of sin, cos, or tan.
84  *      trig(+-INF)  is NaN, with signals;
85  *      trig(NaN)    is that NaN;
86  *      trig(n*PI/2) is exact for any integer n, provided n*PI is
87  *      representable; otherwise, trig(x) is inexact.
88  *
89  * Accuracy:
90  *      trig(x) returns the exact trig(x*pi/PI) nearly rounded, where
91  *
92  *      Decimal:
93  *              pi = 3.141592653589793 23846264338327 .....
94  *    53 bits   PI = 3.141592653589793 115997963 ..... ,
95  *    56 bits   PI = 3.141592653589793 227020265 ..... ,
96  *
97  *      Hexadecimal:
98  *              pi = 3.243F6A8885A308D313198A2E....
99  *    53 bits   PI = 3.243F6A8885A30  =  2 * 1.921FB54442D18    error=.276ulps
100  *    56 bits   PI = 3.243F6A8885A308 =  4 * .C90FDAA22168C2    error=.206ulps
101  *
102  *      In a test run with 1,024,000 random arguments on a VAX, the maximum
103  *      observed errors (compared with the exact trig(x*pi/PI)) were
104  *                      tan(x) : 2.09 ulps (around 4.716340404662354)
105  *                      sin(x) : .861 ulps
106  *                      cos(x) : .857 ulps
107  *
108  * Constants:
109  * The hexadecimal values are the intended ones for the following constants.
110  * The decimal values may be used, provided that the compiler will convert
111  * from decimal to binary accurately enough to produce the hexadecimal values
112  * shown.
113  */
114 
115 #ifdef VAX
116 /*thresh =  2.6117239648121182150E-1    , Hex  2^ -1   *  .85B8636B026EA0 */
117 /*PIo4   =  7.8539816339744830676E-1    , Hex  2^  0   *  .C90FDAA22168C2 */
118 /*PIo2   =  1.5707963267948966135E0     , Hex  2^  1   *  .C90FDAA22168C2 */
119 /*PI3o4  =  2.3561944901923449203E0     , Hex  2^  2   *  .96CBE3F9990E92 */
120 /*PI     =  3.1415926535897932270E0     , Hex  2^  2   *  .C90FDAA22168C2 */
121 /*PI2    =  6.2831853071795864540E0     ; Hex  2^  3   *  .C90FDAA22168C2 */
122 static long    threshx[] = { 0xb8633f85, 0x6ea06b02};
123 #define   thresh    (*(double*)threshx)
124 static long      PIo4x[] = { 0x0fda4049, 0x68c2a221};
125 #define     PIo4    (*(double*)PIo4x)
126 static long      PIo2x[] = { 0x0fda40c9, 0x68c2a221};
127 #define     PIo2    (*(double*)PIo2x)
128 static long      PI3o4x[] = { 0xcbe34116, 0x0e92f999};
129 #define     PI3o4    (*(double*)PI3o4x)
130 static long        PIx[] = { 0x0fda4149, 0x68c2a221};
131 #define       PI    (*(double*)PIx)
132 static long       PI2x[] = { 0x0fda41c9, 0x68c2a221};
133 #define      PI2    (*(double*)PI2x)
134 #else   /* IEEE double  */
135 static double
136 thresh =  2.6117239648121182150E-1    , /*Hex  2^ -2   *  1.0B70C6D604DD4 */
137 PIo4   =  7.8539816339744827900E-1    , /*Hex  2^ -1   *  1.921FB54442D18 */
138 PIo2   =  1.5707963267948965580E0     , /*Hex  2^  0   *  1.921FB54442D18 */
139 PI3o4  =  2.3561944901923448370E0     , /*Hex  2^  1   *  1.2D97C7F3321D2 */
140 PI     =  3.1415926535897931160E0     , /*Hex  2^  1   *  1.921FB54442D18 */
141 PI2    =  6.2831853071795862320E0     ; /*Hex  2^  2   *  1.921FB54442D18 */
142 #endif
143 static double zero=0, one=1, negone= -1, half=1.0/2.0,
144 	      small=1E-10, /* 1+small**2==1; better values for small:
145 					small = 1.5E-9 for VAX D
146 					      = 1.2E-8 for IEEE Double
147 					      = 2.8E-10 for IEEE Extended */
148 	      big=1E20;    /* big = 1/(small**2) */
149 
150 double tan(x)
151 double x;
152 {
153         double copysign(),drem(),cos__C(),sin__S(),a,z,ss,cc,c;
154         int finite(),k;
155 
156         /* tan(NaN) and tan(INF) must be NaN */
157             if(!finite(x))  return(x-x);
158         x=drem(x,PI);        /* reduce x into [-PI/2, PI/2] */
159         a=copysign(x,one);   /* ... = abs(x) */
160 	if ( a >= PIo4 ) {k=1; x = copysign( PIo2 - a , x ); }
161 	   else { k=0; if(a < small ) { big + a; return(x); }}
162 
163         z  = x*x;
164         cc = cos__C(z);
165         ss = sin__S(z);
166 	z  = z*half ;		/* Next get c = cos(x) accurately */
167 	c  = (z >= thresh )? half-((z-half)-cc) : one-(z-cc);
168 	if (k==0) return ( x + (x*(z-(cc-ss)))/c );  /* sin/cos */
169 	return( c/(x+x*ss) );	/*                  ... cos/sin */
170 
171 
172 }
173 double sin(x)
174 double x;
175 {
176         double copysign(),drem(),sin__S(),cos__C(),a,c,z;
177         int finite();
178 
179         /* sin(NaN) and sin(INF) must be NaN */
180             if(!finite(x))  return(x-x);
181 	x=drem(x,PI2);         /*    reduce x into [-PI, PI] */
182         a=copysign(x,one);
183 	if( a >= PIo4 ) {
184 	     if( a >= PI3o4 )   /* 	.. in [3PI/4,  PI ]  */
185 		x=copysign((a=PI-a),x);
186 
187 	     else {	       /* 	.. in [PI/4, 3PI/4]  */
188 		a=PIo2-a;      /* return sign(x)*C(PI/2-|x|) */
189 		z=a*a;
190 		c=cos__C(z);
191 		z=z*half;
192 		a=(z>=thresh)?half-((z-half)-c):one-(z-c);
193 		return(copysign(a,x));
194 		}
195              }
196 
197         /* return S(x) */
198             if( a < small) { big + a; return(x);}
199             return(x+x*sin__S(x*x));
200 }
201 
202 double cos(x)
203 double x;
204 {
205         double copysign(),drem(),sin__S(),cos__C(),a,c,z,s=1.0;
206         int finite();
207 
208         /* cos(NaN) and cos(INF) must be NaN */
209             if(!finite(x))  return(x-x);
210 	x=drem(x,PI2);         /*    reduce x into [-PI, PI] */
211         a=copysign(x,one);
212 	if ( a >= PIo4 ) {
213 	     if ( a >= PI3o4 )  /* 	.. in [3PI/4,  PI ]  */
214 		{ a=PI-a; s= negone; }
215 
216 	     else 	       /* 	.. in [PI/4, 3PI/4]  */
217                                /*        return  S(PI/2-|x|) */
218 		{ a=PIo2-a; return(a+a*sin__S(a*a));}
219 	     }
220 
221 
222         /* return s*C(a) */
223             if( a < small) { big + a; return(s);}
224 	    z=a*a;
225 	    c=cos__C(z);
226 	    z=z*half;
227 	    a=(z>=thresh)?half-((z-half)-c):one-(z-c);
228 	    return(copysign(a,s));
229 }
230 
231 
232 /* sin__S(x*x)
233  * DOUBLE PRECISION (VAX D format 56 bits, IEEE DOUBLE 53 BITS)
234  * STATIC KERNEL FUNCTION OF SIN(X), COS(X), AND TAN(X)
235  * CODED IN C BY K.C. NG, 1/21/85;
236  * REVISED BY K.C. NG on 8/13/85.
237  *
238  *	    sin(x*k) - x
239  * RETURN  --------------- on [-PI/4,PI/4] , where k=pi/PI, PI is the rounded
240  *	            x
241  * value of pi in machine precision:
242  *
243  *	Decimal:
244  *		pi = 3.141592653589793 23846264338327 .....
245  *    53 bits   PI = 3.141592653589793 115997963 ..... ,
246  *    56 bits   PI = 3.141592653589793 227020265 ..... ,
247  *
248  *	Hexadecimal:
249  *		pi = 3.243F6A8885A308D313198A2E....
250  *    53 bits   PI = 3.243F6A8885A30  =  2 * 1.921FB54442D18
251  *    56 bits   PI = 3.243F6A8885A308 =  4 * .C90FDAA22168C2
252  *
253  * Method:
254  *	1. Let z=x*x. Create a polynomial approximation to
255  *	    (sin(k*x)-x)/x  =  z*(S0 + S1*z^1 + ... + S5*z^5).
256  *	Then
257  *      sin__S(x*x) = z*(S0 + S1*z^1 + ... + S5*z^5)
258  *
259  *	The coefficient S's are obtained by a special Remez algorithm.
260  *
261  * Accuracy:
262  *	In the absence of rounding error, the approximation has absolute error
263  *	less than 2**(-61.11) for VAX D FORMAT, 2**(-57.45) for IEEE DOUBLE.
264  *
265  * Constants:
266  * The hexadecimal values are the intended ones for the following constants.
267  * The decimal values may be used, provided that the compiler will convert
268  * from decimal to binary accurately enough to produce the hexadecimal values
269  * shown.
270  *
271  */
272 
273 #ifdef VAX
274 /*S0     = -1.6666666666666646660E-1    , Hex  2^ -2   * -.AAAAAAAAAAAA71 */
275 /*S1     =  8.3333333333297230413E-3    , Hex  2^ -6   *  .8888888888477F */
276 /*S2     = -1.9841269838362403710E-4    , Hex  2^-12   * -.D00D00CF8A1057 */
277 /*S3     =  2.7557318019967078930E-6    , Hex  2^-18   *  .B8EF1CA326BEDC */
278 /*S4     = -2.5051841873876551398E-8    , Hex  2^-25   * -.D73195374CE1D3 */
279 /*S5     =  1.6028995389845827653E-10   , Hex  2^-32   *  .B03D9C6D26CCCC */
280 /*S6     = -6.2723499671769283121E-13   ; Hex  2^-40   * -.B08D0B7561EA82 */
281 static long        S0x[] = { 0xaaaabf2a, 0xaa71aaaa};
282 #define       S0    (*(double*)S0x)
283 static long        S1x[] = { 0x88883d08, 0x477f8888};
284 #define       S1    (*(double*)S1x)
285 static long        S2x[] = { 0x0d00ba50, 0x1057cf8a};
286 #define       S2    (*(double*)S2x)
287 static long        S3x[] = { 0xef1c3738, 0xbedca326};
288 #define       S3    (*(double*)S3x)
289 static long        S4x[] = { 0x3195b3d7, 0xe1d3374c};
290 #define       S4    (*(double*)S4x)
291 static long        S5x[] = { 0x3d9c3030, 0xcccc6d26};
292 #define       S5    (*(double*)S5x)
293 static long        S6x[] = { 0x8d0bac30, 0xea827561};
294 #define       S6    (*(double*)S6x)
295 #else	/* IEEE double  */
296 static double
297 S0     = -1.6666666666666463126E-1    , /*Hex  2^ -3   * -1.555555555550C */
298 S1     =  8.3333333332992771264E-3    , /*Hex  2^ -7   *  1.111111110C461 */
299 S2     = -1.9841269816180999116E-4    , /*Hex  2^-13   * -1.A01A019746345 */
300 S3     =  2.7557309793219876880E-6    , /*Hex  2^-19   *  1.71DE3209CDCD9 */
301 S4     = -2.5050225177523807003E-8    , /*Hex  2^-26   * -1.AE5C0E319A4EF */
302 S5     =  1.5868926979889205164E-10   ; /*Hex  2^-33   *  1.5CF61DF672B13 */
303 #endif
304 
305 static double sin__S(z)
306 double z;
307 {
308 #ifdef VAX
309 	return(z*(S0+z*(S1+z*(S2+z*(S3+z*(S4+z*(S5+z*S6)))))));
310 #else 	/* IEEE double */
311 	return(z*(S0+z*(S1+z*(S2+z*(S3+z*(S4+z*S5))))));
312 #endif
313 }
314 
315 
316 /* cos__C(x*x)
317  * DOUBLE PRECISION (VAX D FORMAT 56 BITS, IEEE DOUBLE 53 BITS)
318  * STATIC KERNEL FUNCTION OF SIN(X), COS(X), AND TAN(X)
319  * CODED IN C BY K.C. NG, 1/21/85;
320  * REVISED BY K.C. NG on 8/13/85.
321  *
322  *	   		    x*x
323  * RETURN   cos(k*x) - 1 + ----- on [-PI/4,PI/4],  where k = pi/PI,
324  *	  		     2
325  * PI is the rounded value of pi in machine precision :
326  *
327  *	Decimal:
328  *		pi = 3.141592653589793 23846264338327 .....
329  *    53 bits   PI = 3.141592653589793 115997963 ..... ,
330  *    56 bits   PI = 3.141592653589793 227020265 ..... ,
331  *
332  *	Hexadecimal:
333  *		pi = 3.243F6A8885A308D313198A2E....
334  *    53 bits   PI = 3.243F6A8885A30  =  2 * 1.921FB54442D18
335  *    56 bits   PI = 3.243F6A8885A308 =  4 * .C90FDAA22168C2
336  *
337  *
338  * Method:
339  *	1. Let z=x*x. Create a polynomial approximation to
340  *	    cos(k*x)-1+z/2  =  z*z*(C0 + C1*z^1 + ... + C5*z^5)
341  *	then
342  *      cos__C(z) =  z*z*(C0 + C1*z^1 + ... + C5*z^5)
343  *
344  *	The coefficient C's are obtained by a special Remez algorithm.
345  *
346  * Accuracy:
347  *	In the absence of rounding error, the approximation has absolute error
348  *	less than 2**(-64) for VAX D FORMAT, 2**(-58.3) for IEEE DOUBLE.
349  *
350  *
351  * Constants:
352  * The hexadecimal values are the intended ones for the following constants.
353  * The decimal values may be used, provided that the compiler will convert
354  * from decimal to binary accurately enough to produce the hexadecimal values
355  * shown.
356  *
357  */
358 
359 #ifdef VAX
360 /*C0     =  4.1666666666666504759E-2    , Hex  2^ -4   *  .AAAAAAAAAAA9F0 */
361 /*C1     = -1.3888888888865302059E-3    , Hex  2^ -9   * -.B60B60B60A0CCA */
362 /*C2     =  2.4801587285601038265E-5    , Hex  2^-15   *  .D00D00CDCD098F */
363 /*C3     = -2.7557313470902390219E-7    , Hex  2^-21   * -.93F27BB593E805 */
364 /*C4     =  2.0875623401082232009E-9    , Hex  2^-28   *  .8F74C8FA1E3FF0 */
365 /*C5     = -1.1355178117642986178E-11   ; Hex  2^-36   * -.C7C32D0A5C5A63 */
366 static long        C0x[] = { 0xaaaa3e2a, 0xa9f0aaaa};
367 #define       C0    (*(double*)C0x)
368 static long        C1x[] = { 0x0b60bbb6, 0x0ccab60a};
369 #define       C1    (*(double*)C1x)
370 static long        C2x[] = { 0x0d0038d0, 0x098fcdcd};
371 #define       C2    (*(double*)C2x)
372 static long        C3x[] = { 0xf27bb593, 0xe805b593};
373 #define       C3    (*(double*)C3x)
374 static long        C4x[] = { 0x74c8320f, 0x3ff0fa1e};
375 #define       C4    (*(double*)C4x)
376 static long        C5x[] = { 0xc32dae47, 0x5a630a5c};
377 #define       C5    (*(double*)C5x)
378 #else	/* IEEE double  */
379 static double
380 C0     =  4.1666666666666504759E-2    , /*Hex  2^ -5   *  1.555555555553E */
381 C1     = -1.3888888888865301516E-3    , /*Hex  2^-10   * -1.6C16C16C14199 */
382 C2     =  2.4801587269650015769E-5    , /*Hex  2^-16   *  1.A01A01971CAEB */
383 C3     = -2.7557304623183959811E-7    , /*Hex  2^-22   * -1.27E4F1314AD1A */
384 C4     =  2.0873958177697780076E-9    , /*Hex  2^-29   *  1.1EE3B60DDDC8C */
385 C5     = -1.1250289076471311557E-11   ; /*Hex  2^-37   * -1.8BD5986B2A52E */
386 #endif
387 
388 static double cos__C(z)
389 double z;
390 {
391 	return(z*z*(C0+z*(C1+z*(C2+z*(C3+z*(C4+z*C5))))));
392 }
393