1 /*
2 * Copyright (c) 2020 Raspberry Pi (Trading) Ltd.
3 *
4 * SPDX-License-Identifier: BSD-3-Clause
5 */
6
7 #include "pico/types.h"
8 #include "pico/float.h"
9 #include "pico/platform.h"
10
11 typedef uint32_t ui32;
12 typedef int32_t i32;
13
14 #define PINF ( HUGE_VAL)
15 #define MINF (-HUGE_VAL)
16 #define NANF ((float)NAN)
17 #define PZERO (+0.0)
18 #define MZERO (-0.0)
19
20 #define PI 3.14159265358979323846
21 #define LOG2 0.69314718055994530941
22 // Unfortunately in double precision ln(10) is very close to half-way between to representable numbers
23 #define LOG10 2.30258509299404568401
24 #define LOG2E 1.44269504088896340737
25 #define LOG10E 0.43429448190325182765
26 #define ONETHIRD 0.33333333333333333333
27
28 #define PIf 3.14159265358979323846f
29 #define LOG2f 0.69314718055994530941f
30 #define LOG2Ef 1.44269504088896340737f
31 #define LOG10Ef 0.43429448190325182765f
32 #define ONETHIRDf 0.33333333333333333333f
33
34 #define FUNPACK(x,e,m) e=((x)>>23)&0xff,m=((x)&0x007fffff)|0x00800000
35 #define FUNPACKS(x,s,e,m) s=((x)>>31),FUNPACK((x),(e),(m))
36
37 _Pragma("GCC diagnostic push")
38 _Pragma("GCC diagnostic ignored \"-Wstrict-aliasing\"")
39
fisnan(float x)40 static inline bool fisnan(float x) {
41 ui32 ix=*(i32*)&x;
42 return ix * 2 > 0xff000000u;
43 }
44
45 #if PICO_FLOAT_PROPAGATE_NANS
46 #define check_nan_f1(x) if (fisnan((x))) return (x)
47 #define check_nan_f2(x,y) if (fisnan((x))) return (x); else if (fisnan((y))) return (y);
48 #else
49 #define check_nan_f1(x) ((void)0)
50 #define check_nan_f2(x,y) ((void)0)
51 #endif
52
fgetsignexp(float x)53 static inline int fgetsignexp(float x) {
54 ui32 ix=*(ui32*)&x;
55 return (ix>>23)&0x1ff;
56 }
57
fgetexp(float x)58 static inline int fgetexp(float x) {
59 ui32 ix=*(ui32*)&x;
60 return (ix>>23)&0xff;
61 }
62
fldexp(float x,int de)63 static inline float fldexp(float x,int de) {
64 ui32 ix=*(ui32*)&x,iy;
65 int e;
66 e=fgetexp(x);
67 if(e==0||e==0xff) return x;
68 e+=de;
69 if(e<=0) iy=ix&0x80000000; // signed zero for underflow
70 else if(e>=0xff) iy=(ix&0x80000000)|0x7f800000ULL; // signed infinity on overflow
71 else iy=ix+((ui32)de<<23);
72 return *(float*)&iy;
73 }
74
WRAPPER_FUNC(ldexpf)75 float WRAPPER_FUNC(ldexpf)(float x, int de) {
76 check_nan_f1(x);
77 return fldexp(x, de);
78 }
79
fcopysign(float x,float y)80 static inline float fcopysign(float x,float y) {
81 ui32 ix=*(ui32*)&x,iy=*(ui32*)&y;
82 ix=((ix&0x7fffffff)|(iy&0x80000000));
83 return *(float*)&ix;
84 }
85
WRAPPER_FUNC(copysignf)86 float WRAPPER_FUNC(copysignf)(float x, float y) {
87 check_nan_f2(x,y);
88 return fcopysign(x, y);
89 }
90
fiszero(float x)91 static inline int fiszero(float x) { return fgetexp (x)==0; }
fispzero(float x)92 static inline int fispzero(float x) { return fgetsignexp(x)==0; }
fismzero(float x)93 static inline int fismzero(float x) { return fgetsignexp(x)==0x100; }
fisinf(float x)94 static inline int fisinf(float x) { return fgetexp (x)==0xff; }
fispinf(float x)95 static inline int fispinf(float x) { return fgetsignexp(x)==0xff; }
fisminf(float x)96 static inline int fisminf(float x) { return fgetsignexp(x)==0x1ff; }
97
fisint(float x)98 static inline int fisint(float x) {
99 ui32 ix=*(ui32*)&x,m;
100 int e=fgetexp(x);
101 if(e==0) return 1; // 0 is an integer
102 e-=0x7f; // remove exponent bias
103 if(e<0) return 0; // |x|<1
104 e=23-e; // bit position in mantissa with significance 1
105 if(e<=0) return 1; // |x| large, so must be an integer
106 m=(1<<e)-1; // mask for bits of significance <1
107 if(ix&m) return 0; // not an integer
108 return 1;
109 }
110
fisoddint(float x)111 static inline int fisoddint(float x) {
112 ui32 ix=*(ui32*)&x,m;
113 int e=fgetexp(x);
114 e-=0x7f; // remove exponent bias
115 if(e<0) return 0; // |x|<1; 0 is not odd
116 e=23-e; // bit position in mantissa with significance 1
117 if(e<0) return 0; // |x| large, so must be even
118 m=(1<<e)-1; // mask for bits of significance <1 (if any)
119 if(ix&m) return 0; // not an integer
120 if(e==23) return 1; // value is exactly 1
121 return (ix>>e)&1;
122 }
123
fisstrictneg(float x)124 static inline int fisstrictneg(float x) {
125 ui32 ix=*(ui32*)&x;
126 if(fiszero(x)) return 0;
127 return ix>>31;
128 }
129
fisneg(float x)130 static inline int fisneg(float x) {
131 ui32 ix=*(ui32*)&x;
132 return ix>>31;
133 }
134
fneg(float x)135 static inline float fneg(float x) {
136 ui32 ix=*(ui32*)&x;
137 ix^=0x80000000;
138 return *(float*)&ix;
139 }
140
fispo2(float x)141 static inline int fispo2(float x) {
142 ui32 ix=*(ui32*)&x;
143 if(fiszero(x)) return 0;
144 if(fisinf(x)) return 0;
145 ix&=0x007fffff;
146 return ix==0;
147 }
148
fnan_or(float x)149 static inline float fnan_or(float x) {
150 #if PICO_FLOAT_PROPAGATE_NANS
151 return NANF;
152 #else
153 return x;
154 #endif
155 }
156
WRAPPER_FUNC(truncf)157 float WRAPPER_FUNC(truncf)(float x) {
158 check_nan_f1(x);
159 ui32 ix=*(ui32*)&x,m;
160 int e=fgetexp(x);
161 e-=0x7f; // remove exponent bias
162 if(e<0) { // |x|<1
163 ix&=0x80000000;
164 return *(float*)&ix;
165 }
166 e=23-e; // bit position in mantissa with significance 1
167 if(e<=0) return x; // |x| large, so must be an integer
168 m=(1<<e)-1; // mask for bits of significance <1
169 ix&=~m;
170 return *(float*)&ix;
171 }
172
WRAPPER_FUNC(roundf)173 float WRAPPER_FUNC(roundf)(float x) {
174 check_nan_f1(x);
175 ui32 ix=*(ui32*)&x,m;
176 int e=fgetexp(x);
177 e-=0x7f; // remove exponent bias
178 if(e<-1) { // |x|<0.5
179 ix&=0x80000000;
180 return *(float*)&ix;
181 }
182 if(e==-1) { // 0.5<=|x|<1
183 ix&=0x80000000;
184 ix|=0x3f800000; // ±1
185 return *(float*)&ix;
186 }
187 e=23-e; // bit position in mantissa with significance 1, <=23
188 if(e<=0) return x; // |x| large, so must be an integer
189 m=1<<(e-1); // mask for bit of significance 0.5
190 ix+=m;
191 m=m+m-1; // mask for bits of significance <1
192 ix&=~m;
193 return *(float*)&ix;
194 }
195
WRAPPER_FUNC(floorf)196 float WRAPPER_FUNC(floorf)(float x) {
197 check_nan_f1(x);
198 ui32 ix=*(ui32*)&x,m;
199 int e=fgetexp(x);
200 if(e==0) { // x==0
201 ix&=0x80000000;
202 return *(float*)&ix;
203 }
204 e-=0x7f; // remove exponent bias
205 if(e<0) { // |x|<1, not zero
206 if(fisneg(x)) return -1;
207 return PZERO;
208 }
209 e=23-e; // bit position in mantissa with significance 1
210 if(e<=0) return x; // |x| large, so must be an integer
211 m=(1<<e)-1; // mask for bit of significance <1
212 if(fisneg(x)) ix+=m; // add 1-ε to magnitude if negative
213 ix&=~m; // truncate
214 return *(float*)&ix;
215 }
216
WRAPPER_FUNC(ceilf)217 float WRAPPER_FUNC(ceilf)(float x) {
218 check_nan_f1(x);
219 ui32 ix=*(ui32*)&x,m;
220 int e=fgetexp(x);
221 if(e==0) { // x==0
222 ix&=0x80000000;
223 return *(float*)&ix;
224 }
225 e-=0x7f; // remove exponent bias
226 if(e<0) { // |x|<1, not zero
227 if(fisneg(x)) return MZERO;
228 return 1;
229 }
230 e=23-e; // bit position in mantissa with significance 1
231 if(e<=0) return x; // |x| large, so must be an integer
232 m=(1<<e)-1; // mask for bit of significance <1
233 if(!fisneg(x)) ix+=m; // add 1-ε to magnitude if positive
234 ix&=~m; // truncate
235 return *(float*)&ix;
236 }
237
WRAPPER_FUNC(asinf)238 float WRAPPER_FUNC(asinf)(float x) {
239 check_nan_f1(x);
240 float u;
241 u=(1.0f-x)*(1.0f+x);
242 if(fisstrictneg(u)) return fnan_or(PINF);
243 return atan2f(x,sqrtf(u));
244 }
245
WRAPPER_FUNC(acosf)246 float WRAPPER_FUNC(acosf)(float x) {
247 check_nan_f1(x);
248 float u;
249 u=(1.0f-x)*(1.0f+x);
250 if(fisstrictneg(u)) return fnan_or(PINF);
251 return atan2f(sqrtf(u),x);
252 }
253
WRAPPER_FUNC(atanf)254 float WRAPPER_FUNC(atanf)(float x) {
255 check_nan_f1(x);
256 if(fispinf(x)) return (float)( PIf/2);
257 if(fisminf(x)) return (float)(-PIf/2);
258 return atan2f(x,1.0f);
259 }
260
WRAPPER_FUNC(sinhf)261 float WRAPPER_FUNC(sinhf)(float x) {
262 check_nan_f1(x);
263 return fldexp((expf(x)-expf(fneg(x))),-1);
264 }
265
WRAPPER_FUNC(coshf)266 float WRAPPER_FUNC(coshf)(float x) {
267 check_nan_f1(x);
268 return fldexp((expf(x)+expf(fneg(x))),-1);
269 }
270
WRAPPER_FUNC(tanhf)271 float WRAPPER_FUNC(tanhf)(float x) {
272 check_nan_f1(x);
273 float u;
274 int e;
275 e=fgetexp(x);
276 if(e>=4+0x7f) { // |x|>=16?
277 if(!fisneg(x)) return 1; // 1 << exp 2x; avoid generating infinities later
278 else return -1; // 1 >> exp 2x
279 }
280 u=expf(fldexp(x,1));
281 return (u-1.0f)/(u+1.0f);
282 }
283
WRAPPER_FUNC(asinhf)284 float WRAPPER_FUNC(asinhf)(float x) {
285 check_nan_f1(x);
286 int e;
287 e=fgetexp(x);
288 if(e>=16+0x7f) { // |x|>=2^16?
289 if(!fisneg(x)) return logf( x )+LOG2f; // 1/x^2 << 1
290 else return fneg(logf(fneg(x))+LOG2f); // 1/x^2 << 1
291 }
292 if(x>0) return (float)log(sqrt((double)x*(double)x+1.0)+(double)x);
293 else return fneg((float)log(sqrt((double)x*(double)x+1.0)-(double)x));
294 }
295
WRAPPER_FUNC(acoshf)296 float WRAPPER_FUNC(acoshf)(float x) {
297 check_nan_f1(x);
298 int e;
299 if(fisneg(x)) x=fneg(x);
300 e=fgetexp(x);
301 if(e>=16+0x7f) return logf(x)+LOG2f; // |x|>=2^16?
302 return (float)log(sqrt(((double)x+1.0)*((double)x-1.0))+(double)x);
303 }
304
WRAPPER_FUNC(atanhf)305 float WRAPPER_FUNC(atanhf)(float x) {
306 check_nan_f1(x);
307 return fldexp(logf((1.0f+x)/(1.0f-x)),-1);
308 }
309
WRAPPER_FUNC(exp2f)310 float WRAPPER_FUNC(exp2f)(float x) { check_nan_f1(x); return (float)exp((double)x*LOG2); }
WRAPPER_FUNC(log2f)311 float WRAPPER_FUNC(log2f)(float x) { check_nan_f1(x); return logf(x)*LOG2Ef; }
WRAPPER_FUNC(exp10f)312 float WRAPPER_FUNC(exp10f)(float x) { check_nan_f1(x); return (float)exp((double)x*LOG10); }
WRAPPER_FUNC(log10f)313 float WRAPPER_FUNC(log10f)(float x) { check_nan_f1(x); return logf(x)*LOG10Ef; }
314
WRAPPER_FUNC(expm1f)315 float WRAPPER_FUNC(expm1f)(float x) { check_nan_f1(x); return (float)(exp((double)x)-1); }
WRAPPER_FUNC(log1pf)316 float WRAPPER_FUNC(log1pf)(float x) { check_nan_f1(x); return (float)(log(1+(double)x)); }
WRAPPER_FUNC(fmaf)317 float WRAPPER_FUNC(fmaf)(float x,float y,float z) {
318 check_nan_f2(x,y);
319 check_nan_f1(z);
320 return (float)((double)x*(double)y+(double)z);
321 } // has double rounding so not exact
322
323 // general power, x>0
fpow_1(float x,float y)324 static inline float fpow_1(float x,float y) {
325 return (float)exp(log((double)x)*(double)y); // using double-precision intermediates for better accuracy
326 }
327
fpow_int2(float x,int y)328 static float fpow_int2(float x,int y) {
329 float u;
330 if(y==1) return x;
331 u=fpow_int2(x,y/2);
332 u*=u;
333 if(y&1) u*=x;
334 return u;
335 }
336
337 // for the case where x not zero or infinity, y small and not zero
fpowint_1(float x,int y)338 static inline float fpowint_1(float x,int y) {
339 if(y<0) x=1.0f/x,y=-y;
340 return fpow_int2(x,y);
341 }
342
343 // for the case where x not zero or infinity
fpowint_0(float x,int y)344 static float fpowint_0(float x,int y) {
345 int e;
346 if(fisneg(x)) {
347 if(fisoddint(y)) return fneg(fpowint_0(fneg(x),y));
348 else return fpowint_0(fneg(x),y);
349 }
350 if(fispo2(x)) {
351 e=fgetexp(x)-0x7f;
352 if(y>=256) y= 255; // avoid overflow
353 if(y<-256) y=-256;
354 y*=e;
355 return fldexp(1,y);
356 }
357 if(y==0) return 1;
358 if(y>=-32&&y<=32) return fpowint_1(x,y);
359 return fpow_1(x,y);
360 }
361
WRAPPER_FUNC(powintf)362 float WRAPPER_FUNC(powintf)(float x,int y) {
363 _Pragma("GCC diagnostic push")
364 _Pragma("GCC diagnostic ignored \"-Wfloat-equal\"")
365 if(x==1.0f||y==0) return 1;
366 if(x==0.0f) {
367 if(y>0) {
368 if(y&1) return x;
369 else return 0;
370 }
371 if((y&1)) return fcopysign(PINF,x);
372 return PINF;
373 }
374 _Pragma("GCC diagnostic pop")
375 check_nan_f1(x);
376 if(fispinf(x)) {
377 if(y<0) return 0;
378 else return PINF;
379 }
380 if(fisminf(x)) {
381 if(y>0) {
382 if((y&1)) return MINF;
383 else return PINF;
384 }
385 if((y&1)) return MZERO;
386 else return PZERO;
387 }
388 return fpowint_0(x,y);
389 }
390
391 // for the case where y is guaranteed a finite integer, x not zero or infinity
fpow_0(float x,float y)392 static float fpow_0(float x,float y) {
393 int e,p;
394 if(fisneg(x)) {
395 if(fisoddint(y)) return fneg(fpow_0(fneg(x),y));
396 else return fpow_0(fneg(x),y);
397 }
398 p=(int)y;
399 if(fispo2(x)) {
400 e=fgetexp(x)-0x7f;
401 if(p>=256) p= 255; // avoid overflow
402 if(p<-256) p=-256;
403 p*=e;
404 return fldexp(1,p);
405 }
406 if(p==0) return 1;
407 if(p>=-32&&p<=32) return fpowint_1(x,p);
408 return fpow_1(x,y);
409 }
410
WRAPPER_FUNC(powf)411 float WRAPPER_FUNC(powf)(float x,float y) {
412 _Pragma("GCC diagnostic push")
413 _Pragma("GCC diagnostic ignored \"-Wfloat-equal\"")
414 if(x==1.0f||fiszero(y)) return 1;
415 check_nan_f2(x,y);
416 if(x==-1.0f&&fisinf(y)) return 1;
417 _Pragma("GCC diagnostic pop")
418 if(fiszero(x)) {
419 if(!fisneg(y)) {
420 if(fisoddint(y)) return x;
421 else return 0;
422 }
423 if(fisoddint(y)) return fcopysign(PINF,x);
424 return PINF;
425 }
426 if(fispinf(x)) {
427 if(fisneg(y)) return 0;
428 else return PINF;
429 }
430 if(fisminf(x)) {
431 if(!fisneg(y)) {
432 if(fisoddint(y)) return MINF;
433 else return PINF;
434 }
435 if(fisoddint(y)) return MZERO;
436 else return PZERO;
437 }
438 if(fispinf(y)) {
439 if(fgetexp(x)<0x7f) return PZERO;
440 else return PINF;
441 }
442 if(fisminf(y)) {
443 if(fgetexp(x)<0x7f) return PINF;
444 else return PZERO;
445 }
446 if(fisint(y)) return fpow_0(x,y);
447 if(fisneg(x)) return PINF;
448 return fpow_1(x,y);
449 }
450
WRAPPER_FUNC(hypotf)451 float WRAPPER_FUNC(hypotf)(float x,float y) {
452 check_nan_f2(x,y);
453 int ex,ey;
454 ex=fgetexp(x); ey=fgetexp(y);
455 if(ex>=0x7f+50||ey>=0x7f+50) { // overflow, or nearly so
456 x=fldexp(x,-70),y=fldexp(y,-70);
457 return fldexp(sqrtf(x*x+y*y), 70);
458 }
459 else if(ex<=0x7f-50&&ey<=0x7f-50) { // underflow, or nearly so
460 x=fldexp(x, 70),y=fldexp(y, 70);
461 return fldexp(sqrtf(x*x+y*y),-70);
462 }
463 return sqrtf(x*x+y*y);
464 }
465
WRAPPER_FUNC(cbrtf)466 float WRAPPER_FUNC(cbrtf)(float x) {
467 check_nan_f1(x);
468 int e;
469 if(fisneg(x)) return fneg(cbrtf(fneg(x)));
470 if(fiszero(x)) return fcopysign(PZERO,x);
471 e=fgetexp(x)-0x7f;
472 e=(e*0x5555+0x8000)>>16; // ~e/3, rounded
473 x=fldexp(x,-e*3);
474 x=expf(logf(x)*ONETHIRDf);
475 return fldexp(x,e);
476 }
477
478 // reduces mx*2^e modulo my, returning bottom bits of quotient at *pquo
479 // 2^23<=|mx|,my<2^24, e>=0; 0<=result<my
frem_0(i32 mx,i32 my,int e,int * pquo)480 static i32 frem_0(i32 mx,i32 my,int e,int*pquo) {
481 int quo=0,q,r=0,s;
482 if(e>0) {
483 r=0xffffffffU/(ui32)(my>>7); // reciprocal estimate Q16
484 }
485 while(e>0) {
486 s=e; if(s>12) s=12; // gain up to 12 bits on each iteration
487 q=(mx>>9)*r; // Q30
488 q=((q>>(29-s))+1)>>1; // Q(s), rounded
489 mx=(mx<<s)-my*q;
490 quo=(quo<<s)+q;
491 e-=s;
492 }
493 if(mx>=my) mx-=my,quo++; // when e==0 mx can be nearly as big as 2my
494 if(mx>=my) mx-=my,quo++;
495 if(mx<0) mx+=my,quo--;
496 if(mx<0) mx+=my,quo--;
497 if(pquo) *pquo=quo;
498 return mx;
499 }
500
WRAPPER_FUNC(fmodf)501 float WRAPPER_FUNC(fmodf)(float x,float y) {
502 check_nan_f2(x,y);
503 ui32 ix=*(ui32*)&x,iy=*(ui32*)&y;
504 int sx,ex,ey;
505 i32 mx,my;
506 FUNPACKS(ix,sx,ex,mx);
507 FUNPACK(iy,ey,my);
508 if(ex==0xff) {
509 return fnan_or(PINF);
510 }
511 if(ey==0) return PINF;
512 if(ex==0) {
513 if(!fisneg(x)) return PZERO;
514 return MZERO;
515 }
516 if(ex<ey) return x; // |x|<|y|, including case x=±0
517 mx=frem_0(mx,my,ex-ey,0);
518 if(sx) mx=-mx;
519 return fix2float(mx,0x7f-ey+23);
520 }
521
WRAPPER_FUNC(remquof)522 float WRAPPER_FUNC(remquof)(float x,float y,int*quo) {
523 check_nan_f2(x,y);
524 ui32 ix=*(ui32*)&x,iy=*(ui32*)&y;
525 int sx,sy,ex,ey,q;
526 i32 mx,my;
527 FUNPACKS(ix,sx,ex,mx);
528 FUNPACKS(iy,sy,ey,my);
529 if(quo) *quo=0;
530 if(ex==0xff) return PINF;
531 if(ey==0) return PINF;
532 if(ex==0) return PZERO;
533 if(ey==0xff) return x;
534 if(ex<ey-1) return x; // |x|<|y|/2
535 if(ex==ey-1) {
536 if(mx<=my) return x; // |x|<=|y|/2, even quotient
537 // here |y|/2<|x|<|y|
538 if(!sx) { // x>|y|/2
539 mx-=my+my;
540 ey--;
541 q=1;
542 } else { // x<-|y|/2
543 mx=my+my-mx;
544 ey--;
545 q=-1;
546 }
547 }
548 else {
549 if(sx) mx=-mx;
550 mx=frem_0(mx,my,ex-ey,&q);
551 if(mx+mx>my || (mx+mx==my&&(q&1)) ) { // |x|>|y|/2, or equality and an odd quotient?
552 mx-=my;
553 q++;
554 }
555 }
556 if(sy) q=-q;
557 if(quo) *quo=q;
558 return fix2float(mx,0x7f-ey+23);
559 }
560
WRAPPER_FUNC(dremf)561 float WRAPPER_FUNC(dremf)(float x,float y) { check_nan_f2(x,y); return remquof(x,y,0); }
562
WRAPPER_FUNC(remainderf)563 float WRAPPER_FUNC(remainderf)(float x,float y) { check_nan_f2(x,y); return remquof(x,y,0); }
564
565 _Pragma("GCC diagnostic pop") // strict-aliasing
566