xref: /freebsd/contrib/arm-optimized-routines/math/test/ulp.h (revision 5a02ffc32e777041dd2dad4e651ed2a0865a0a5d)
1 /*
2  * Generic functions for ULP error estimation.
3  *
4  * Copyright (c) 2019-2023, Arm Limited.
5  * SPDX-License-Identifier: MIT OR Apache-2.0 WITH LLVM-exception
6  */
7 
8 /* For each different math function type,
9    T(x) should add a different suffix to x.
10    RT(x) should add a return type specific suffix to x. */
11 
12 #ifdef NEW_RT
13 #undef NEW_RT
14 
15 # if USE_MPFR
RT(ulpscale_mpfr)16 static int RT(ulpscale_mpfr) (mpfr_t x, int t)
17 {
18   /* TODO: pow of 2 cases.  */
19   if (mpfr_regular_p (x))
20     {
21       mpfr_exp_t e = mpfr_get_exp (x) - RT(prec);
22       if (e < RT(emin))
23 	e = RT(emin) - 1;
24       if (e > RT(emax) - RT(prec))
25 	e = RT(emax) - RT(prec);
26       return e;
27     }
28   if (mpfr_zero_p (x))
29     return RT(emin) - 1;
30   if (mpfr_inf_p (x))
31     return RT(emax) - RT(prec);
32   /* NaN.  */
33   return 0;
34 }
35 # endif
36 
37 /* Difference between exact result and closest real number that
38    gets rounded to got, i.e. error before rounding, for a correctly
39    rounded result the difference is 0.  */
RT(ulperr)40 static double RT (ulperr) (RT (float) got, const struct RT (ret) * p, int r,
41 			   int ignore_zero_sign)
42 {
43   RT(float) want = p->y;
44   RT(float) d;
45   double e;
46 
47   if (RT(asuint) (got) == RT(asuint) (want))
48     return 0.0;
49   if (isnan (got) && isnan (want))
50     /* Ignore sign of NaN.  */
51     return RT (issignaling) (got) == RT (issignaling) (want) ? 0 : INFINITY;
52   if (signbit (got) != signbit (want))
53     {
54       /* Fall through to ULP calculation if ignoring sign of zero and at
55 	 exactly one of want and got is non-zero.  */
56       if (ignore_zero_sign && want == got)
57 	return 0.0;
58       if (!ignore_zero_sign || (want != 0 && got != 0))
59 	return INFINITY;
60     }
61   if (!isfinite (want) || !isfinite (got))
62     {
63       if (isnan (got) != isnan (want))
64 	return INFINITY;
65       if (isnan (want))
66 	return 0;
67       if (isinf (got))
68 	{
69 	  got = RT(copysign) (RT(halfinf), got);
70 	  want *= 0.5f;
71 	}
72       if (isinf (want))
73 	{
74 	  want = RT(copysign) (RT(halfinf), want);
75 	  got *= 0.5f;
76 	}
77     }
78   if (r == FE_TONEAREST)
79     {
80       // TODO: incorrect when got vs want cross a powof2 boundary
81       /* error = got > want
82 	      ? got - want - tail ulp - 0.5 ulp
83 	      : got - want - tail ulp + 0.5 ulp;  */
84       d = got - want;
85       e = d > 0 ? -p->tail - 0.5 : -p->tail + 0.5;
86     }
87   else
88     {
89       if ((r == FE_DOWNWARD && got < want) || (r == FE_UPWARD && got > want)
90 	  || (r == FE_TOWARDZERO && fabs (got) < fabs (want)))
91 	got = RT(nextafter) (got, want);
92       d = got - want;
93       e = -p->tail;
94     }
95   return RT(scalbn) (d, -p->ulpexp) + e;
96 }
97 
RT(isok)98 static int RT(isok) (RT(float) ygot, int exgot, RT(float) ywant, int exwant,
99 		      int exmay)
100 {
101   return RT(asuint) (ygot) == RT(asuint) (ywant)
102 	 && ((exgot ^ exwant) & ~exmay) == 0;
103 }
104 
RT(isok_nofenv)105 static int RT(isok_nofenv) (RT(float) ygot, RT(float) ywant)
106 {
107   return RT(asuint) (ygot) == RT(asuint) (ywant);
108 }
109 #endif
110 
T(call_fenv)111 static inline void T(call_fenv) (const struct fun *f, struct T(args) a, int r,
112 				  RT(float) * y, int *ex)
113 {
114   if (r != FE_TONEAREST)
115     fesetround (r);
116   feclearexcept (FE_ALL_EXCEPT);
117   *y = T(call) (f, a);
118   *ex = fetestexcept (FE_ALL_EXCEPT);
119   if (r != FE_TONEAREST)
120     fesetround (FE_TONEAREST);
121 }
122 
T(call_nofenv)123 static inline void T(call_nofenv) (const struct fun *f, struct T(args) a,
124 				    int r, RT(float) * y, int *ex)
125 {
126   if (r != FE_TONEAREST)
127     fesetround (r);
128   *y = T(call) (f, a);
129   *ex = 0;
130   if (r != FE_TONEAREST)
131     fesetround (FE_TONEAREST);
132 }
133 
T(call_long_fenv)134 static inline int T(call_long_fenv) (const struct fun *f, struct T(args) a,
135 				      int r, struct RT(ret) * p,
136 				      RT(float) ygot, int exgot)
137 {
138   if (r != FE_TONEAREST)
139     fesetround (r);
140   feclearexcept (FE_ALL_EXCEPT);
141   volatile struct T(args) va = a; // TODO: barrier
142   a = va;
143   RT(double) yl = T(call_long) (f, a);
144   p->y = (RT(float)) yl;
145   volatile RT(float) vy = p->y; // TODO: barrier
146   (void) vy;
147   p->ex = fetestexcept (FE_ALL_EXCEPT);
148   if (r != FE_TONEAREST)
149     fesetround (FE_TONEAREST);
150   p->ex_may = FE_INEXACT;
151   if (RT(isok) (ygot, exgot, p->y, p->ex, p->ex_may))
152     return 1;
153   p->ulpexp = RT(ulpscale) (p->y);
154   if (isinf (p->y))
155     p->tail = RT(lscalbn) (yl - (RT(double)) 2 * RT(halfinf), -p->ulpexp);
156   else
157     p->tail = RT(lscalbn) (yl - p->y, -p->ulpexp);
158   if (RT(fabs) (p->y) < RT(min_normal))
159     {
160       /* TODO: subnormal result is treated as undeflow even if it's
161 	 exact since call_long may not raise inexact correctly.  */
162       if (p->y != 0 || (p->ex & FE_INEXACT))
163 	p->ex |= FE_UNDERFLOW | FE_INEXACT;
164     }
165   return 0;
166 }
T(call_long_nofenv)167 static inline int T(call_long_nofenv) (const struct fun *f, struct T(args) a,
168 					int r, struct RT(ret) * p,
169 					RT(float) ygot, int exgot)
170 {
171   if (r != FE_TONEAREST)
172     fesetround (r);
173   RT(double) yl = T(call_long) (f, a);
174   p->y = (RT(float)) yl;
175   if (r != FE_TONEAREST)
176     fesetround (FE_TONEAREST);
177   if (RT(isok_nofenv) (ygot, p->y))
178     return 1;
179   p->ulpexp = RT(ulpscale) (p->y);
180   if (isinf (p->y))
181     p->tail = RT(lscalbn) (yl - (RT(double)) 2 * RT(halfinf), -p->ulpexp);
182   else
183     p->tail = RT(lscalbn) (yl - p->y, -p->ulpexp);
184   return 0;
185 }
186 
187 /* There are nan input args and all quiet.  */
T(qnanpropagation)188 static inline int T(qnanpropagation) (struct T(args) a)
189 {
190   return T(reduce) (a, isnan, ||) && !T(reduce) (a, RT(issignaling), ||);
191 }
T(sum)192 static inline RT(float) T(sum) (struct T(args) a)
193 {
194   return T(reduce) (a, , +);
195 }
196 
197 /* returns 1 if the got result is ok.  */
T(call_mpfr_fix)198 static inline int T(call_mpfr_fix) (const struct fun *f, struct T(args) a,
199 				     int r_fenv, struct RT(ret) * p,
200 				     RT(float) ygot, int exgot)
201 {
202 #if USE_MPFR
203   int t, t2;
204   mpfr_rnd_t r = rmap (r_fenv);
205   MPFR_DECL_INIT(my, RT(prec_mpfr));
206   MPFR_DECL_INIT(mr, RT(prec));
207   MPFR_DECL_INIT(me, RT(prec_mpfr));
208   mpfr_clear_flags ();
209   t = T(call_mpfr) (my, f, a, r);
210   /* Double rounding.  */
211   t2 = mpfr_set (mr, my, r);
212   if (t2)
213     t = t2;
214   mpfr_set_emin (RT(emin));
215   mpfr_set_emax (RT(emax));
216   t = mpfr_check_range (mr, t, r);
217   t = mpfr_subnormalize (mr, t, r);
218   mpfr_set_emax (MPFR_EMAX_DEFAULT);
219   mpfr_set_emin (MPFR_EMIN_DEFAULT);
220   p->y = mpfr_get_d (mr, r);
221   p->ex = t ? FE_INEXACT : 0;
222   p->ex_may = FE_INEXACT;
223   if (mpfr_underflow_p () && (p->ex & FE_INEXACT))
224     /* TODO: handle before and after rounding uflow cases.  */
225     p->ex |= FE_UNDERFLOW;
226   if (mpfr_overflow_p ())
227     p->ex |= FE_OVERFLOW | FE_INEXACT;
228   if (mpfr_divby0_p ())
229     p->ex |= FE_DIVBYZERO;
230   //if (mpfr_erangeflag_p ())
231   //  p->ex |= FE_INVALID;
232   if (!mpfr_nanflag_p () && RT(isok) (ygot, exgot, p->y, p->ex, p->ex_may))
233     return 1;
234   if (mpfr_nanflag_p () && !T(qnanpropagation) (a))
235     p->ex |= FE_INVALID;
236   p->ulpexp = RT(ulpscale_mpfr) (my, t);
237   if (!isfinite (p->y))
238     {
239       p->tail = 0;
240       if (isnan (p->y))
241 	{
242 	  /* If an input was nan keep its sign.  */
243 	  p->y = T(sum) (a);
244 	  if (!isnan (p->y))
245 	    p->y = (p->y - p->y) / (p->y - p->y);
246 	  return RT(isok) (ygot, exgot, p->y, p->ex, p->ex_may);
247 	}
248       mpfr_set_si_2exp (mr, signbit (p->y) ? -1 : 1, 1024, MPFR_RNDN);
249       if (mpfr_cmpabs (my, mr) >= 0)
250 	return RT(isok) (ygot, exgot, p->y, p->ex, p->ex_may);
251     }
252   mpfr_sub (me, my, mr, MPFR_RNDN);
253   mpfr_mul_2si (me, me, -p->ulpexp, MPFR_RNDN);
254   p->tail = mpfr_get_d (me, MPFR_RNDN);
255   return 0;
256 #else
257   abort ();
258 #endif
259 }
260 
T(cmp)261 static int T(cmp) (const struct fun *f, struct gen *gen,
262 		     const struct conf *conf)
263 {
264   double maxerr = 0;
265   uint64_t cnt = 0;
266   uint64_t cnt1 = 0;
267   uint64_t cnt2 = 0;
268   uint64_t cntfail = 0;
269   int r = conf->r;
270   int use_mpfr = conf->mpfr;
271   int fenv = conf->fenv;
272   for (;;)
273     {
274       struct RT(ret) want;
275       struct T(args) a = T(next) (gen);
276       int exgot;
277       int exgot2;
278       RT(float) ygot;
279       RT(float) ygot2;
280       int fail = 0;
281       if (fenv)
282 	T(call_fenv) (f, a, r, &ygot, &exgot);
283       else
284 	T(call_nofenv) (f, a, r, &ygot, &exgot);
285       if (f->twice) {
286 	secondcall = 1;
287 	if (fenv)
288 	  T(call_fenv) (f, a, r, &ygot2, &exgot2);
289 	else
290 	  T(call_nofenv) (f, a, r, &ygot2, &exgot2);
291 	secondcall = 0;
292 	if (RT(asuint) (ygot) != RT(asuint) (ygot2))
293 	  {
294 	    fail = 1;
295 	    cntfail++;
296 	    T(printcall) (f, a);
297 	    printf (" got %a then %a for same input\n", ygot, ygot2);
298 	  }
299       }
300       cnt++;
301       int ok = use_mpfr
302 		 ? T(call_mpfr_fix) (f, a, r, &want, ygot, exgot)
303 		 : (fenv ? T(call_long_fenv) (f, a, r, &want, ygot, exgot)
304 			 : T(call_long_nofenv) (f, a, r, &want, ygot, exgot));
305       if (!ok)
306 	{
307 	  int print = 0;
308 	  double err = RT (ulperr) (ygot, &want, r, conf->ignore_zero_sign);
309 	  double abserr = fabs (err);
310 	  // TODO: count errors below accuracy limit.
311 	  if (abserr > 0)
312 	    cnt1++;
313 	  if (abserr > 1)
314 	    cnt2++;
315 	  if (abserr > conf->errlim)
316 	    {
317 	      print = 1;
318 	      if (!fail)
319 		{
320 		  fail = 1;
321 		  cntfail++;
322 		}
323 	    }
324 	  if (abserr > maxerr)
325 	    {
326 	      maxerr = abserr;
327 	      if (!conf->quiet && abserr > conf->softlim)
328 		print = 1;
329 	    }
330 	  if (print)
331 	    {
332 	      T(printcall) (f, a);
333 	      // TODO: inf ulp handling
334 	      printf (" got %a want %a %+g ulp err %g\n", ygot, want.y,
335 		      want.tail, err);
336 	    }
337 	  int diff = fenv ? exgot ^ want.ex : 0;
338 	  if (fenv && (diff & ~want.ex_may))
339 	    {
340 	      if (!fail)
341 		{
342 		  fail = 1;
343 		  cntfail++;
344 		}
345 	      T(printcall) (f, a);
346 	      printf (" is %a %+g ulp, got except 0x%0x", want.y, want.tail,
347 		      exgot);
348 	      if (diff & exgot)
349 		printf (" wrongly set: 0x%x", diff & exgot);
350 	      if (diff & ~exgot)
351 		printf (" wrongly clear: 0x%x", diff & ~exgot);
352 	      putchar ('\n');
353 	    }
354 	}
355       if (cnt >= conf->n)
356 	break;
357       if (!conf->quiet && cnt % 0x100000 == 0)
358 	printf ("progress: %6.3f%% cnt %llu cnt1 %llu cnt2 %llu cntfail %llu "
359 		"maxerr %g\n",
360 		100.0 * cnt / conf->n, (unsigned long long) cnt,
361 		(unsigned long long) cnt1, (unsigned long long) cnt2,
362 		(unsigned long long) cntfail, maxerr);
363     }
364   double cc = cnt;
365   if (cntfail)
366     printf ("FAIL ");
367   else
368     printf ("PASS ");
369   T(printgen) (f, gen);
370   printf (" round %c errlim %g maxerr %g %s cnt %llu cnt1 %llu %g%% cnt2 %llu "
371 	  "%g%% cntfail %llu %g%%\n",
372 	  conf->rc, conf->errlim,
373 	  maxerr, conf->r == FE_TONEAREST ? "+0.5" : "+1.0",
374 	  (unsigned long long) cnt,
375 	  (unsigned long long) cnt1, 100.0 * cnt1 / cc,
376 	  (unsigned long long) cnt2, 100.0 * cnt2 / cc,
377 	  (unsigned long long) cntfail, 100.0 * cntfail / cc);
378   return !!cntfail;
379 }
380