1 | /* |
2 | * ULP error checking tool for math functions. |
3 | * |
4 | * Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
5 | * See https://llvm.org/LICENSE.txt for license information. |
6 | * SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
7 | */ |
8 | |
9 | #include <ctype.h> |
10 | #include <fenv.h> |
11 | #include <float.h> |
12 | #include <math.h> |
13 | #include <stdint.h> |
14 | #include <stdio.h> |
15 | #include <stdlib.h> |
16 | #include <string.h> |
17 | #include "mathlib.h" |
18 | |
19 | /* Don't depend on mpfr by default. */ |
20 | #ifndef USE_MPFR |
21 | # define USE_MPFR 0 |
22 | #endif |
23 | #if USE_MPFR |
24 | # include <mpfr.h> |
25 | #endif |
26 | |
27 | #ifndef WANT_VMATH |
28 | /* Enable the build of vector math code. */ |
29 | # define WANT_VMATH 1 |
30 | #endif |
31 | |
32 | static inline uint64_t |
33 | asuint64 (double f) |
34 | { |
35 | union |
36 | { |
37 | double f; |
38 | uint64_t i; |
39 | } u = {f}; |
40 | return u.i; |
41 | } |
42 | |
43 | static inline double |
44 | asdouble (uint64_t i) |
45 | { |
46 | union |
47 | { |
48 | uint64_t i; |
49 | double f; |
50 | } u = {.i: i}; |
51 | return u.f; |
52 | } |
53 | |
54 | static inline uint32_t |
55 | asuint (float f) |
56 | { |
57 | union |
58 | { |
59 | float f; |
60 | uint32_t i; |
61 | } u = {f}; |
62 | return u.i; |
63 | } |
64 | |
65 | static inline float |
66 | asfloat (uint32_t i) |
67 | { |
68 | union |
69 | { |
70 | uint32_t i; |
71 | float f; |
72 | } u = {.i: i}; |
73 | return u.f; |
74 | } |
75 | |
76 | static uint64_t seed = 0x0123456789abcdef; |
77 | static uint64_t |
78 | rand64 (void) |
79 | { |
80 | seed = 6364136223846793005ull * seed + 1; |
81 | return seed ^ (seed >> 32); |
82 | } |
83 | |
84 | /* Uniform random in [0,n]. */ |
85 | static uint64_t |
86 | randn (uint64_t n) |
87 | { |
88 | uint64_t r, m; |
89 | |
90 | if (n == 0) |
91 | return 0; |
92 | n++; |
93 | if (n == 0) |
94 | return rand64 (); |
95 | for (;;) |
96 | { |
97 | r = rand64 (); |
98 | m = r % n; |
99 | if (r - m <= -n) |
100 | return m; |
101 | } |
102 | } |
103 | |
104 | struct gen |
105 | { |
106 | uint64_t start; |
107 | uint64_t len; |
108 | uint64_t start2; |
109 | uint64_t len2; |
110 | uint64_t off; |
111 | uint64_t step; |
112 | uint64_t cnt; |
113 | }; |
114 | |
115 | struct args_f1 |
116 | { |
117 | float x; |
118 | }; |
119 | |
120 | struct args_f2 |
121 | { |
122 | float x; |
123 | float x2; |
124 | }; |
125 | |
126 | struct args_d1 |
127 | { |
128 | double x; |
129 | }; |
130 | |
131 | struct args_d2 |
132 | { |
133 | double x; |
134 | double x2; |
135 | }; |
136 | |
137 | /* result = y + tail*2^ulpexp. */ |
138 | struct ret_f |
139 | { |
140 | float y; |
141 | double tail; |
142 | int ulpexp; |
143 | int ex; |
144 | int ex_may; |
145 | }; |
146 | |
147 | struct ret_d |
148 | { |
149 | double y; |
150 | double tail; |
151 | int ulpexp; |
152 | int ex; |
153 | int ex_may; |
154 | }; |
155 | |
156 | static inline uint64_t |
157 | next1 (struct gen *g) |
158 | { |
159 | /* For single argument use randomized incremental steps, |
160 | that produce dense sampling without collisions and allow |
161 | testing all inputs in a range. */ |
162 | uint64_t r = g->start + g->off; |
163 | g->off += g->step + randn (n: g->step / 2); |
164 | if (g->off > g->len) |
165 | g->off -= g->len; /* hack. */ |
166 | return r; |
167 | } |
168 | |
169 | static inline uint64_t |
170 | next2 (uint64_t *x2, struct gen *g) |
171 | { |
172 | /* For two arguments use uniform random sampling. */ |
173 | uint64_t r = g->start + randn (n: g->len); |
174 | *x2 = g->start2 + randn (n: g->len2); |
175 | return r; |
176 | } |
177 | |
178 | static struct args_f1 |
179 | next_f1 (void *g) |
180 | { |
181 | return (struct args_f1){asfloat (i: next1 (g))}; |
182 | } |
183 | |
184 | static struct args_f2 |
185 | next_f2 (void *g) |
186 | { |
187 | uint64_t x2; |
188 | uint64_t x = next2 (x2: &x2, g); |
189 | return (struct args_f2){asfloat (i: x), asfloat (i: x2)}; |
190 | } |
191 | |
192 | static struct args_d1 |
193 | next_d1 (void *g) |
194 | { |
195 | return (struct args_d1){asdouble (i: next1 (g))}; |
196 | } |
197 | |
198 | static struct args_d2 |
199 | next_d2 (void *g) |
200 | { |
201 | uint64_t x2; |
202 | uint64_t x = next2 (x2: &x2, g); |
203 | return (struct args_d2){asdouble (i: x), asdouble (i: x2)}; |
204 | } |
205 | |
206 | struct conf |
207 | { |
208 | int r; |
209 | int rc; |
210 | int quiet; |
211 | int mpfr; |
212 | int fenv; |
213 | unsigned long long n; |
214 | double softlim; |
215 | double errlim; |
216 | }; |
217 | |
218 | /* Wrappers for sincos. */ |
219 | static float sincosf_sinf(float x) {(void)cosf(x: x); return sinf(x: x);} |
220 | static float sincosf_cosf(float x) {(void)sinf(x: x); return cosf(x: x);} |
221 | static double sincos_sin(double x) {(void)cos(x: x); return sin(x: x);} |
222 | static double sincos_cos(double x) {(void)sin(x: x); return cos(x: x);} |
223 | #if USE_MPFR |
224 | static int sincos_mpfr_sin(mpfr_t y, const mpfr_t x, mpfr_rnd_t r) { mpfr_cos(y,x,r); return mpfr_sin(y,x,r); } |
225 | static int sincos_mpfr_cos(mpfr_t y, const mpfr_t x, mpfr_rnd_t r) { mpfr_sin(y,x,r); return mpfr_cos(y,x,r); } |
226 | #endif |
227 | |
228 | /* A bit of a hack: call vector functions twice with the same |
229 | input in lane 0 but a different value in other lanes: once |
230 | with an in-range value and then with a special case value. */ |
231 | static int secondcall; |
232 | |
233 | /* Wrappers for vector functions. */ |
234 | #if __aarch64__ && WANT_VMATH |
235 | typedef __f32x4_t v_float; |
236 | typedef __f64x2_t v_double; |
237 | static const float fv[2] = {1.0f, -INFINITY}; |
238 | static const double dv[2] = {1.0, -INFINITY}; |
239 | static inline v_float argf(float x) { return (v_float){x,x,x,fv[secondcall]}; } |
240 | static inline v_double argd(double x) { return (v_double){x,dv[secondcall]}; } |
241 | |
242 | static float v_sinf(float x) { return __v_sinf(argf(x))[0]; } |
243 | static float v_cosf(float x) { return __v_cosf(argf(x))[0]; } |
244 | static float v_expf_1u(float x) { return __v_expf_1u(argf(x))[0]; } |
245 | static float v_expf(float x) { return __v_expf(argf(x))[0]; } |
246 | static float v_exp2f_1u(float x) { return __v_exp2f_1u(argf(x))[0]; } |
247 | static float v_exp2f(float x) { return __v_exp2f(argf(x))[0]; } |
248 | static float v_logf(float x) { return __v_logf(argf(x))[0]; } |
249 | static float v_powf(float x, float y) { return __v_powf(argf(x),argf(y))[0]; } |
250 | static double v_sin(double x) { return __v_sin(argd(x))[0]; } |
251 | static double v_cos(double x) { return __v_cos(argd(x))[0]; } |
252 | static double v_exp(double x) { return __v_exp(argd(x))[0]; } |
253 | static double v_log(double x) { return __v_log(argd(x))[0]; } |
254 | static double v_pow(double x, double y) { return __v_pow(argd(x),argd(y))[0]; } |
255 | #ifdef __vpcs |
256 | static float vn_sinf(float x) { return __vn_sinf(argf(x))[0]; } |
257 | static float vn_cosf(float x) { return __vn_cosf(argf(x))[0]; } |
258 | static float vn_expf_1u(float x) { return __vn_expf_1u(argf(x))[0]; } |
259 | static float vn_expf(float x) { return __vn_expf(argf(x))[0]; } |
260 | static float vn_exp2f_1u(float x) { return __vn_exp2f_1u(argf(x))[0]; } |
261 | static float vn_exp2f(float x) { return __vn_exp2f(argf(x))[0]; } |
262 | static float vn_logf(float x) { return __vn_logf(argf(x))[0]; } |
263 | static float vn_powf(float x, float y) { return __vn_powf(argf(x),argf(y))[0]; } |
264 | static double vn_sin(double x) { return __vn_sin(argd(x))[0]; } |
265 | static double vn_cos(double x) { return __vn_cos(argd(x))[0]; } |
266 | static double vn_exp(double x) { return __vn_exp(argd(x))[0]; } |
267 | static double vn_log(double x) { return __vn_log(argd(x))[0]; } |
268 | static double vn_pow(double x, double y) { return __vn_pow(argd(x),argd(y))[0]; } |
269 | static float Z_sinf(float x) { return _ZGVnN4v_sinf(argf(x))[0]; } |
270 | static float Z_cosf(float x) { return _ZGVnN4v_cosf(argf(x))[0]; } |
271 | static float Z_expf(float x) { return _ZGVnN4v_expf(argf(x))[0]; } |
272 | static float Z_exp2f(float x) { return _ZGVnN4v_exp2f(argf(x))[0]; } |
273 | static float Z_logf(float x) { return _ZGVnN4v_logf(argf(x))[0]; } |
274 | static float Z_powf(float x, float y) { return _ZGVnN4vv_powf(argf(x),argf(y))[0]; } |
275 | static double Z_sin(double x) { return _ZGVnN2v_sin(argd(x))[0]; } |
276 | static double Z_cos(double x) { return _ZGVnN2v_cos(argd(x))[0]; } |
277 | static double Z_exp(double x) { return _ZGVnN2v_exp(argd(x))[0]; } |
278 | static double Z_log(double x) { return _ZGVnN2v_log(argd(x))[0]; } |
279 | static double Z_pow(double x, double y) { return _ZGVnN2vv_pow(argd(x),argd(y))[0]; } |
280 | #endif |
281 | #endif |
282 | |
283 | struct fun |
284 | { |
285 | const char *name; |
286 | int arity; |
287 | int singleprec; |
288 | int twice; |
289 | union |
290 | { |
291 | float (*f1) (float); |
292 | float (*f2) (float, float); |
293 | double (*d1) (double); |
294 | double (*d2) (double, double); |
295 | } fun; |
296 | union |
297 | { |
298 | double (*f1) (double); |
299 | double (*f2) (double, double); |
300 | long double (*d1) (long double); |
301 | long double (*d2) (long double, long double); |
302 | } fun_long; |
303 | #if USE_MPFR |
304 | union |
305 | { |
306 | int (*f1) (mpfr_t, const mpfr_t, mpfr_rnd_t); |
307 | int (*f2) (mpfr_t, const mpfr_t, const mpfr_t, mpfr_rnd_t); |
308 | int (*d1) (mpfr_t, const mpfr_t, mpfr_rnd_t); |
309 | int (*d2) (mpfr_t, const mpfr_t, const mpfr_t, mpfr_rnd_t); |
310 | } fun_mpfr; |
311 | #endif |
312 | }; |
313 | |
314 | static const struct fun fun[] = { |
315 | #if USE_MPFR |
316 | # define F(x, x_wrap, x_long, x_mpfr, a, s, t, twice) \ |
317 | {#x, a, s, twice, {.t = x_wrap}, {.t = x_long}, {.t = x_mpfr}}, |
318 | #else |
319 | # define F(x, x_wrap, x_long, x_mpfr, a, s, t, twice) \ |
320 | {#x, a, s, twice, {.t = x_wrap}, {.t = x_long}}, |
321 | #endif |
322 | #define F1(x) F (x##f, x##f, x, mpfr_##x, 1, 1, f1, 0) |
323 | #define F2(x) F (x##f, x##f, x, mpfr_##x, 2, 1, f2, 0) |
324 | #define D1(x) F (x, x, x##l, mpfr_##x, 1, 0, d1, 0) |
325 | #define D2(x) F (x, x, x##l, mpfr_##x, 2, 0, d2, 0) |
326 | F1 (sin) |
327 | F1 (cos) |
328 | F (sincosf_sinf, sincosf_sinf, sincos_sin, sincos_mpfr_sin, 1, 1, f1, 0) |
329 | F (sincosf_cosf, sincosf_cosf, sincos_cos, sincos_mpfr_cos, 1, 1, f1, 0) |
330 | F1 (exp) |
331 | F1 (exp2) |
332 | F1 (log) |
333 | F1 (log2) |
334 | F2 (pow) |
335 | D1 (exp) |
336 | D1 (exp2) |
337 | D1 (log) |
338 | D1 (log2) |
339 | D2 (pow) |
340 | #if WANT_VMATH |
341 | F (__s_sinf, __s_sinf, sin, mpfr_sin, 1, 1, f1, 0) |
342 | F (__s_cosf, __s_cosf, cos, mpfr_cos, 1, 1, f1, 0) |
343 | F (__s_expf_1u, __s_expf_1u, exp, mpfr_exp, 1, 1, f1, 0) |
344 | F (__s_expf, __s_expf, exp, mpfr_exp, 1, 1, f1, 0) |
345 | F (__s_exp2f_1u, __s_exp2f_1u, exp2, mpfr_exp2, 1, 1, f1, 0) |
346 | F (__s_exp2f, __s_exp2f, exp2, mpfr_exp2, 1, 1, f1, 0) |
347 | F (__s_powf, __s_powf, pow, mpfr_pow, 2, 1, f2, 0) |
348 | F (__s_logf, __s_logf, log, mpfr_log, 1, 1, f1, 0) |
349 | F (__s_sin, __s_sin, sinl, mpfr_sin, 1, 0, d1, 0) |
350 | F (__s_cos, __s_cos, cosl, mpfr_cos, 1, 0, d1, 0) |
351 | F (__s_exp, __s_exp, expl, mpfr_exp, 1, 0, d1, 0) |
352 | F (__s_log, __s_log, logl, mpfr_log, 1, 0, d1, 0) |
353 | F (__s_pow, __s_pow, powl, mpfr_pow, 2, 0, d2, 0) |
354 | #if __aarch64__ |
355 | F (__v_sinf, v_sinf, sin, mpfr_sin, 1, 1, f1, 1) |
356 | F (__v_cosf, v_cosf, cos, mpfr_cos, 1, 1, f1, 1) |
357 | F (__v_expf_1u, v_expf_1u, exp, mpfr_exp, 1, 1, f1, 1) |
358 | F (__v_expf, v_expf, exp, mpfr_exp, 1, 1, f1, 1) |
359 | F (__v_exp2f_1u, v_exp2f_1u, exp2, mpfr_exp2, 1, 1, f1, 1) |
360 | F (__v_exp2f, v_exp2f, exp2, mpfr_exp2, 1, 1, f1, 1) |
361 | F (__v_logf, v_logf, log, mpfr_log, 1, 1, f1, 1) |
362 | F (__v_powf, v_powf, pow, mpfr_pow, 2, 1, f2, 1) |
363 | F (__v_sin, v_sin, sinl, mpfr_sin, 1, 0, d1, 1) |
364 | F (__v_cos, v_cos, cosl, mpfr_cos, 1, 0, d1, 1) |
365 | F (__v_exp, v_exp, expl, mpfr_exp, 1, 0, d1, 1) |
366 | F (__v_log, v_log, logl, mpfr_log, 1, 0, d1, 1) |
367 | F (__v_pow, v_pow, powl, mpfr_pow, 2, 0, d2, 1) |
368 | #ifdef __vpcs |
369 | F (__vn_sinf, vn_sinf, sin, mpfr_sin, 1, 1, f1, 1) |
370 | F (__vn_cosf, vn_cosf, cos, mpfr_cos, 1, 1, f1, 1) |
371 | F (__vn_expf_1u, vn_expf_1u, exp, mpfr_exp, 1, 1, f1, 1) |
372 | F (__vn_expf, vn_expf, exp, mpfr_exp, 1, 1, f1, 1) |
373 | F (__vn_exp2f_1u, vn_exp2f_1u, exp2, mpfr_exp2, 1, 1, f1, 1) |
374 | F (__vn_exp2f, vn_exp2f, exp2, mpfr_exp2, 1, 1, f1, 1) |
375 | F (__vn_logf, vn_logf, log, mpfr_log, 1, 1, f1, 1) |
376 | F (__vn_powf, vn_powf, pow, mpfr_pow, 2, 1, f2, 1) |
377 | F (__vn_sin, vn_sin, sinl, mpfr_sin, 1, 0, d1, 1) |
378 | F (__vn_cos, vn_cos, cosl, mpfr_cos, 1, 0, d1, 1) |
379 | F (__vn_exp, vn_exp, expl, mpfr_exp, 1, 0, d1, 1) |
380 | F (__vn_log, vn_log, logl, mpfr_log, 1, 0, d1, 1) |
381 | F (__vn_pow, vn_pow, powl, mpfr_pow, 2, 0, d2, 1) |
382 | F (_ZGVnN4v_sinf, Z_sinf, sin, mpfr_sin, 1, 1, f1, 1) |
383 | F (_ZGVnN4v_cosf, Z_cosf, cos, mpfr_cos, 1, 1, f1, 1) |
384 | F (_ZGVnN4v_expf, Z_expf, exp, mpfr_exp, 1, 1, f1, 1) |
385 | F (_ZGVnN4v_exp2f, Z_exp2f, exp2, mpfr_exp2, 1, 1, f1, 1) |
386 | F (_ZGVnN4v_logf, Z_logf, log, mpfr_log, 1, 1, f1, 1) |
387 | F (_ZGVnN4vv_powf, Z_powf, pow, mpfr_pow, 2, 1, f2, 1) |
388 | F (_ZGVnN2v_sin, Z_sin, sinl, mpfr_sin, 1, 0, d1, 1) |
389 | F (_ZGVnN2v_cos, Z_cos, cosl, mpfr_cos, 1, 0, d1, 1) |
390 | F (_ZGVnN2v_exp, Z_exp, expl, mpfr_exp, 1, 0, d1, 1) |
391 | F (_ZGVnN2v_log, Z_log, logl, mpfr_log, 1, 0, d1, 1) |
392 | F (_ZGVnN2vv_pow, Z_pow, powl, mpfr_pow, 2, 0, d2, 1) |
393 | #endif |
394 | #endif |
395 | #endif |
396 | #undef F |
397 | #undef F1 |
398 | #undef F2 |
399 | #undef D1 |
400 | #undef D2 |
401 | {0}}; |
402 | |
403 | /* Boilerplate for generic calls. */ |
404 | |
405 | static inline int |
406 | ulpscale_f (float x) |
407 | { |
408 | int e = asuint (f: x) >> 23 & 0xff; |
409 | if (!e) |
410 | e++; |
411 | return e - 0x7f - 23; |
412 | } |
413 | static inline int |
414 | ulpscale_d (double x) |
415 | { |
416 | int e = asuint64 (f: x) >> 52 & 0x7ff; |
417 | if (!e) |
418 | e++; |
419 | return e - 0x3ff - 52; |
420 | } |
421 | static inline float |
422 | call_f1 (const struct fun *f, struct args_f1 a) |
423 | { |
424 | return f->fun.f1 (a.x); |
425 | } |
426 | static inline float |
427 | call_f2 (const struct fun *f, struct args_f2 a) |
428 | { |
429 | return f->fun.f2 (a.x, a.x2); |
430 | } |
431 | |
432 | static inline double |
433 | call_d1 (const struct fun *f, struct args_d1 a) |
434 | { |
435 | return f->fun.d1 (a.x); |
436 | } |
437 | static inline double |
438 | call_d2 (const struct fun *f, struct args_d2 a) |
439 | { |
440 | return f->fun.d2 (a.x, a.x2); |
441 | } |
442 | static inline double |
443 | call_long_f1 (const struct fun *f, struct args_f1 a) |
444 | { |
445 | return f->fun_long.f1 (a.x); |
446 | } |
447 | static inline double |
448 | call_long_f2 (const struct fun *f, struct args_f2 a) |
449 | { |
450 | return f->fun_long.f2 (a.x, a.x2); |
451 | } |
452 | static inline long double |
453 | call_long_d1 (const struct fun *f, struct args_d1 a) |
454 | { |
455 | return f->fun_long.d1 (a.x); |
456 | } |
457 | static inline long double |
458 | call_long_d2 (const struct fun *f, struct args_d2 a) |
459 | { |
460 | return f->fun_long.d2 (a.x, a.x2); |
461 | } |
462 | static inline void |
463 | printcall_f1 (const struct fun *f, struct args_f1 a) |
464 | { |
465 | printf (format: "%s(%a)" , f->name, a.x); |
466 | } |
467 | static inline void |
468 | printcall_f2 (const struct fun *f, struct args_f2 a) |
469 | { |
470 | printf (format: "%s(%a, %a)" , f->name, a.x, a.x2); |
471 | } |
472 | static inline void |
473 | printcall_d1 (const struct fun *f, struct args_d1 a) |
474 | { |
475 | printf (format: "%s(%a)" , f->name, a.x); |
476 | } |
477 | static inline void |
478 | printcall_d2 (const struct fun *f, struct args_d2 a) |
479 | { |
480 | printf (format: "%s(%a, %a)" , f->name, a.x, a.x2); |
481 | } |
482 | static inline void |
483 | printgen_f1 (const struct fun *f, struct gen *gen) |
484 | { |
485 | printf (format: "%s in [%a;%a]" , f->name, asfloat (i: gen->start), |
486 | asfloat (i: gen->start + gen->len)); |
487 | } |
488 | static inline void |
489 | printgen_f2 (const struct fun *f, struct gen *gen) |
490 | { |
491 | printf (format: "%s in [%a;%a] x [%a;%a]" , f->name, asfloat (i: gen->start), |
492 | asfloat (i: gen->start + gen->len), asfloat (i: gen->start2), |
493 | asfloat (i: gen->start2 + gen->len2)); |
494 | } |
495 | static inline void |
496 | printgen_d1 (const struct fun *f, struct gen *gen) |
497 | { |
498 | printf (format: "%s in [%a;%a]" , f->name, asdouble (i: gen->start), |
499 | asdouble (i: gen->start + gen->len)); |
500 | } |
501 | static inline void |
502 | printgen_d2 (const struct fun *f, struct gen *gen) |
503 | { |
504 | printf (format: "%s in [%a;%a] x [%a;%a]" , f->name, asdouble (i: gen->start), |
505 | asdouble (i: gen->start + gen->len), asdouble (i: gen->start2), |
506 | asdouble (i: gen->start2 + gen->len2)); |
507 | } |
508 | |
509 | #define reduce_f1(a, f, op) (f (a.x)) |
510 | #define reduce_f2(a, f, op) (f (a.x) op f (a.x2)) |
511 | #define reduce_d1(a, f, op) (f (a.x)) |
512 | #define reduce_d2(a, f, op) (f (a.x) op f (a.x2)) |
513 | |
514 | #ifndef IEEE_754_2008_SNAN |
515 | # define IEEE_754_2008_SNAN 1 |
516 | #endif |
517 | static inline int |
518 | issignaling_f (float x) |
519 | { |
520 | uint32_t ix = asuint (f: x); |
521 | if (!IEEE_754_2008_SNAN) |
522 | return (ix & 0x7fc00000) == 0x7fc00000; |
523 | return 2 * (ix ^ 0x00400000) > 2u * 0x7fc00000; |
524 | } |
525 | static inline int |
526 | issignaling_d (double x) |
527 | { |
528 | uint64_t ix = asuint64 (f: x); |
529 | if (!IEEE_754_2008_SNAN) |
530 | return (ix & 0x7ff8000000000000) == 0x7ff8000000000000; |
531 | return 2 * (ix ^ 0x0008000000000000) > 2 * 0x7ff8000000000000ULL; |
532 | } |
533 | |
534 | #if USE_MPFR |
535 | static mpfr_rnd_t |
536 | rmap (int r) |
537 | { |
538 | switch (r) |
539 | { |
540 | case FE_TONEAREST: |
541 | return MPFR_RNDN; |
542 | case FE_TOWARDZERO: |
543 | return MPFR_RNDZ; |
544 | case FE_UPWARD: |
545 | return MPFR_RNDU; |
546 | case FE_DOWNWARD: |
547 | return MPFR_RNDD; |
548 | } |
549 | return -1; |
550 | } |
551 | |
552 | #define prec_mpfr_f 50 |
553 | #define prec_mpfr_d 80 |
554 | #define prec_f 24 |
555 | #define prec_d 53 |
556 | #define emin_f -148 |
557 | #define emin_d -1073 |
558 | #define emax_f 128 |
559 | #define emax_d 1024 |
560 | static inline int |
561 | call_mpfr_f1 (mpfr_t y, const struct fun *f, struct args_f1 a, mpfr_rnd_t r) |
562 | { |
563 | MPFR_DECL_INIT (x, prec_f); |
564 | mpfr_set_flt (x, a.x, MPFR_RNDN); |
565 | return f->fun_mpfr.f1 (y, x, r); |
566 | } |
567 | static inline int |
568 | call_mpfr_f2 (mpfr_t y, const struct fun *f, struct args_f2 a, mpfr_rnd_t r) |
569 | { |
570 | MPFR_DECL_INIT (x, prec_f); |
571 | MPFR_DECL_INIT (x2, prec_f); |
572 | mpfr_set_flt (x, a.x, MPFR_RNDN); |
573 | mpfr_set_flt (x2, a.x2, MPFR_RNDN); |
574 | return f->fun_mpfr.f2 (y, x, x2, r); |
575 | } |
576 | static inline int |
577 | call_mpfr_d1 (mpfr_t y, const struct fun *f, struct args_d1 a, mpfr_rnd_t r) |
578 | { |
579 | MPFR_DECL_INIT (x, prec_d); |
580 | mpfr_set_d (x, a.x, MPFR_RNDN); |
581 | return f->fun_mpfr.d1 (y, x, r); |
582 | } |
583 | static inline int |
584 | call_mpfr_d2 (mpfr_t y, const struct fun *f, struct args_d2 a, mpfr_rnd_t r) |
585 | { |
586 | MPFR_DECL_INIT (x, prec_d); |
587 | MPFR_DECL_INIT (x2, prec_d); |
588 | mpfr_set_d (x, a.x, MPFR_RNDN); |
589 | mpfr_set_d (x2, a.x2, MPFR_RNDN); |
590 | return f->fun_mpfr.d2 (y, x, x2, r); |
591 | } |
592 | #endif |
593 | |
594 | #define float_f float |
595 | #define double_f double |
596 | #define copysign_f copysignf |
597 | #define nextafter_f nextafterf |
598 | #define fabs_f fabsf |
599 | #define asuint_f asuint |
600 | #define asfloat_f asfloat |
601 | #define scalbn_f scalbnf |
602 | #define lscalbn_f scalbn |
603 | #define halfinf_f 0x1p127f |
604 | #define min_normal_f 0x1p-126f |
605 | |
606 | #define float_d double |
607 | #define double_d long double |
608 | #define copysign_d copysign |
609 | #define nextafter_d nextafter |
610 | #define fabs_d fabs |
611 | #define asuint_d asuint64 |
612 | #define asfloat_d asdouble |
613 | #define scalbn_d scalbn |
614 | #define lscalbn_d scalbnl |
615 | #define halfinf_d 0x1p1023 |
616 | #define min_normal_d 0x1p-1022 |
617 | |
618 | #define NEW_RT |
619 | #define RT(x) x##_f |
620 | #define T(x) x##_f1 |
621 | #include "ulp.h" |
622 | #undef T |
623 | #define T(x) x##_f2 |
624 | #include "ulp.h" |
625 | #undef T |
626 | #undef RT |
627 | |
628 | #define NEW_RT |
629 | #define RT(x) x##_d |
630 | #define T(x) x##_d1 |
631 | #include "ulp.h" |
632 | #undef T |
633 | #define T(x) x##_d2 |
634 | #include "ulp.h" |
635 | #undef T |
636 | #undef RT |
637 | |
638 | static void |
639 | usage (void) |
640 | { |
641 | puts (s: "./ulp [-q] [-m] [-f] [-r nudz] [-l soft-ulplimit] [-e ulplimit] func " |
642 | "lo [hi [x lo2 hi2] [count]]" ); |
643 | puts (s: "Compares func against a higher precision implementation in [lo; hi]." ); |
644 | puts (s: "-q: quiet." ); |
645 | puts (s: "-m: use mpfr even if faster method is available." ); |
646 | puts (s: "-f: disable fenv testing (rounding modes and exceptions)." ); |
647 | puts (s: "Supported func:" ); |
648 | for (const struct fun *f = fun; f->name; f++) |
649 | printf (format: "\t%s\n" , f->name); |
650 | exit (status: 1); |
651 | } |
652 | |
653 | static int |
654 | cmp (const struct fun *f, struct gen *gen, const struct conf *conf) |
655 | { |
656 | int r = 1; |
657 | if (f->arity == 1 && f->singleprec) |
658 | r = cmp_f1 (f, gen, conf); |
659 | else if (f->arity == 2 && f->singleprec) |
660 | r = cmp_f2 (f, gen, conf); |
661 | else if (f->arity == 1 && !f->singleprec) |
662 | r = cmp_d1 (f, gen, conf); |
663 | else if (f->arity == 2 && !f->singleprec) |
664 | r = cmp_d2 (f, gen, conf); |
665 | else |
666 | usage (); |
667 | return r; |
668 | } |
669 | |
670 | static uint64_t |
671 | getnum (const char *s, int singleprec) |
672 | { |
673 | // int i; |
674 | uint64_t sign = 0; |
675 | // char buf[12]; |
676 | |
677 | if (s[0] == '+') |
678 | s++; |
679 | else if (s[0] == '-') |
680 | { |
681 | sign = singleprec ? 1ULL << 31 : 1ULL << 63; |
682 | s++; |
683 | } |
684 | /* 0xXXXX is treated as bit representation, '-' flips the sign bit. */ |
685 | if (s[0] == '0' && tolower (c: s[1]) == 'x' && strchr (s: s, c: 'p') == 0) |
686 | return sign ^ strtoull (nptr: s, endptr: 0, base: 0); |
687 | // /* SNaN, QNaN, NaN, Inf. */ |
688 | // for (i=0; s[i] && i < sizeof buf; i++) |
689 | // buf[i] = tolower(s[i]); |
690 | // buf[i] = 0; |
691 | // if (strcmp(buf, "snan") == 0) |
692 | // return sign | (singleprec ? 0x7fa00000 : 0x7ff4000000000000); |
693 | // if (strcmp(buf, "qnan") == 0 || strcmp(buf, "nan") == 0) |
694 | // return sign | (singleprec ? 0x7fc00000 : 0x7ff8000000000000); |
695 | // if (strcmp(buf, "inf") == 0 || strcmp(buf, "infinity") == 0) |
696 | // return sign | (singleprec ? 0x7f800000 : 0x7ff0000000000000); |
697 | /* Otherwise assume it's a floating-point literal. */ |
698 | return sign |
699 | | (singleprec ? asuint (f: strtof (nptr: s, endptr: 0)) : asuint64 (f: strtod (nptr: s, endptr: 0))); |
700 | } |
701 | |
702 | static void |
703 | parsegen (struct gen *g, int argc, char *argv[], const struct fun *f) |
704 | { |
705 | int singleprec = f->singleprec; |
706 | int arity = f->arity; |
707 | uint64_t a, b, a2, b2, n; |
708 | if (argc < 1) |
709 | usage (); |
710 | b = a = getnum (s: argv[0], singleprec); |
711 | n = 0; |
712 | if (argc > 1 && strcmp (s1: argv[1], s2: "x" ) == 0) |
713 | { |
714 | argc -= 2; |
715 | argv += 2; |
716 | } |
717 | else if (argc > 1) |
718 | { |
719 | b = getnum (s: argv[1], singleprec); |
720 | if (argc > 2 && strcmp (s1: argv[2], s2: "x" ) == 0) |
721 | { |
722 | argc -= 3; |
723 | argv += 3; |
724 | } |
725 | } |
726 | b2 = a2 = getnum (s: argv[0], singleprec); |
727 | if (argc > 1) |
728 | b2 = getnum (s: argv[1], singleprec); |
729 | if (argc > 2) |
730 | n = strtoull (nptr: argv[2], endptr: 0, base: 0); |
731 | if (argc > 3) |
732 | usage (); |
733 | //printf("ab %lx %lx ab2 %lx %lx n %lu\n", a, b, a2, b2, n); |
734 | if (arity == 1) |
735 | { |
736 | g->start = a; |
737 | g->len = b - a; |
738 | if (n - 1 > b - a) |
739 | n = b - a + 1; |
740 | g->off = 0; |
741 | g->step = n ? (g->len + 1) / n : 1; |
742 | g->start2 = g->len2 = 0; |
743 | g->cnt = n; |
744 | } |
745 | else if (arity == 2) |
746 | { |
747 | g->start = a; |
748 | g->len = b - a; |
749 | g->off = g->step = 0; |
750 | g->start2 = a2; |
751 | g->len2 = b2 - a2; |
752 | g->cnt = n; |
753 | } |
754 | else |
755 | usage (); |
756 | } |
757 | |
758 | int |
759 | main (int argc, char *argv[]) |
760 | { |
761 | const struct fun *f; |
762 | struct gen gen; |
763 | struct conf conf; |
764 | conf.rc = 'n'; |
765 | conf.quiet = 0; |
766 | conf.mpfr = 0; |
767 | conf.fenv = 1; |
768 | conf.softlim = 0; |
769 | conf.errlim = INFINITY; |
770 | for (;;) |
771 | { |
772 | argc--; |
773 | argv++; |
774 | if (argc < 1) |
775 | usage (); |
776 | if (argv[0][0] != '-') |
777 | break; |
778 | switch (argv[0][1]) |
779 | { |
780 | case 'e': |
781 | argc--; |
782 | argv++; |
783 | if (argc < 1) |
784 | usage (); |
785 | conf.errlim = strtod (nptr: argv[0], endptr: 0); |
786 | break; |
787 | case 'f': |
788 | conf.fenv = 0; |
789 | break; |
790 | case 'l': |
791 | argc--; |
792 | argv++; |
793 | if (argc < 1) |
794 | usage (); |
795 | conf.softlim = strtod (nptr: argv[0], endptr: 0); |
796 | break; |
797 | case 'm': |
798 | conf.mpfr = 1; |
799 | break; |
800 | case 'q': |
801 | conf.quiet = 1; |
802 | break; |
803 | case 'r': |
804 | conf.rc = argv[0][2]; |
805 | if (!conf.rc) |
806 | { |
807 | argc--; |
808 | argv++; |
809 | if (argc < 1) |
810 | usage (); |
811 | conf.rc = argv[0][0]; |
812 | } |
813 | break; |
814 | default: |
815 | usage (); |
816 | } |
817 | } |
818 | switch (conf.rc) |
819 | { |
820 | case 'n': |
821 | conf.r = FE_TONEAREST; |
822 | break; |
823 | case 'u': |
824 | conf.r = FE_UPWARD; |
825 | break; |
826 | case 'd': |
827 | conf.r = FE_DOWNWARD; |
828 | break; |
829 | case 'z': |
830 | conf.r = FE_TOWARDZERO; |
831 | break; |
832 | default: |
833 | usage (); |
834 | } |
835 | for (f = fun; f->name; f++) |
836 | if (strcmp (s1: argv[0], s2: f->name) == 0) |
837 | break; |
838 | if (!f->name) |
839 | usage (); |
840 | if (!f->singleprec && LDBL_MANT_DIG == DBL_MANT_DIG) |
841 | conf.mpfr = 1; /* Use mpfr if long double has no extra precision. */ |
842 | if (!USE_MPFR && conf.mpfr) |
843 | { |
844 | puts (s: "mpfr is not available." ); |
845 | return 0; |
846 | } |
847 | argc--; |
848 | argv++; |
849 | parsegen (g: &gen, argc, argv, f); |
850 | conf.n = gen.cnt; |
851 | return cmp (f, gen: &gen, conf: &conf); |
852 | } |
853 | |