1 /*===---- __clang_cuda_cmath.h - Device-side CUDA cmath support ------------===
2  *
3  * Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4  * See https://llvm.org/LICENSE.txt for license information.
5  * SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6  *
7  *===-----------------------------------------------------------------------===
8  */
9 #ifndef __CLANG_CUDA_CMATH_H__
10 #define __CLANG_CUDA_CMATH_H__
11 #ifndef __CUDA__
12 #error "This file is for CUDA compilation only."
13 #endif
14 
15 #include <limits>
16 
17 // CUDA lets us use various std math functions on the device side.  This file
18 // works in concert with __clang_cuda_math_forward_declares.h to make this work.
19 //
20 // Specifically, the forward-declares header declares __device__ overloads for
21 // these functions in the global namespace, then pulls them into namespace std
22 // with 'using' statements.  Then this file implements those functions, after
23 // their implementations have been pulled in.
24 //
25 // It's important that we declare the functions in the global namespace and pull
26 // them into namespace std with using statements, as opposed to simply declaring
27 // these functions in namespace std, because our device functions need to
28 // overload the standard library functions, which may be declared in the global
29 // namespace or in std, depending on the degree of conformance of the stdlib
30 // implementation.  Declaring in the global namespace and pulling into namespace
31 // std covers all of the known knowns.
32 
33 #ifdef _OPENMP
34 #define __DEVICE__ static __attribute__((always_inline))
35 #else
36 #define __DEVICE__ static __device__ __inline__ __attribute__((always_inline))
37 #endif
38 
39 // For C++ 17 we need to include noexcept attribute to be compatible
40 // with the header-defined version. This may be removed once
41 // variant is supported.
42 #if defined(_OPENMP) && defined(__cplusplus) && __cplusplus >= 201703L
43 #define __NOEXCEPT noexcept
44 #else
45 #define __NOEXCEPT
46 #endif
47 
48 #if !(defined(_OPENMP) && defined(__cplusplus))
49 __DEVICE__ long long abs(long long __n) { return ::llabs(__n); }
50 __DEVICE__ long abs(long __n) { return ::labs(__n); }
51 __DEVICE__ float abs(float __x) { return ::fabsf(__x); }
52 __DEVICE__ double abs(double __x) { return ::fabs(__x); }
53 #endif
54 // TODO: remove once variat is supported.
55 #if defined(_OPENMP) && defined(__cplusplus)
56 __DEVICE__ const float abs(const float __x) { return ::fabsf((float)__x); }
57 __DEVICE__ const double abs(const double __x) { return ::fabs((double)__x); }
58 #endif
59 __DEVICE__ float acos(float __x) { return ::acosf(__x); }
60 __DEVICE__ float asin(float __x) { return ::asinf(__x); }
61 __DEVICE__ float atan(float __x) { return ::atanf(__x); }
62 __DEVICE__ float atan2(float __x, float __y) { return ::atan2f(__x, __y); }
63 __DEVICE__ float ceil(float __x) { return ::ceilf(__x); }
64 __DEVICE__ float cos(float __x) { return ::cosf(__x); }
65 __DEVICE__ float cosh(float __x) { return ::coshf(__x); }
66 __DEVICE__ float exp(float __x) { return ::expf(__x); }
67 __DEVICE__ float fabs(float __x) __NOEXCEPT { return ::fabsf(__x); }
68 __DEVICE__ float floor(float __x) { return ::floorf(__x); }
69 __DEVICE__ float fmod(float __x, float __y) { return ::fmodf(__x, __y); }
70 // TODO: remove when variant is supported
71 #ifndef _OPENMP
72 __DEVICE__ int fpclassify(float __x) {
73   return __builtin_fpclassify(FP_NAN, FP_INFINITE, FP_NORMAL, FP_SUBNORMAL,
74                               FP_ZERO, __x);
75 }
76 __DEVICE__ int fpclassify(double __x) {
77   return __builtin_fpclassify(FP_NAN, FP_INFINITE, FP_NORMAL, FP_SUBNORMAL,
78                               FP_ZERO, __x);
79 }
80 #endif
81 __DEVICE__ float frexp(float __arg, int *__exp) {
82   return ::frexpf(__arg, __exp);
83 }
84 
85 // For inscrutable reasons, the CUDA headers define these functions for us on
86 // Windows.
87 #ifndef _MSC_VER
88 __DEVICE__ bool isinf(float __x) { return ::__isinff(__x); }
89 __DEVICE__ bool isinf(double __x) { return ::__isinf(__x); }
90 __DEVICE__ bool isfinite(float __x) { return ::__finitef(__x); }
91 // For inscrutable reasons, __finite(), the double-precision version of
92 // __finitef, does not exist when compiling for MacOS.  __isfinited is available
93 // everywhere and is just as good.
94 __DEVICE__ bool isfinite(double __x) { return ::__isfinited(__x); }
95 __DEVICE__ bool isnan(float __x) { return ::__isnanf(__x); }
96 __DEVICE__ bool isnan(double __x) { return ::__isnan(__x); }
97 #endif
98 
99 __DEVICE__ bool isgreater(float __x, float __y) {
100   return __builtin_isgreater(__x, __y);
101 }
102 __DEVICE__ bool isgreater(double __x, double __y) {
103   return __builtin_isgreater(__x, __y);
104 }
105 __DEVICE__ bool isgreaterequal(float __x, float __y) {
106   return __builtin_isgreaterequal(__x, __y);
107 }
108 __DEVICE__ bool isgreaterequal(double __x, double __y) {
109   return __builtin_isgreaterequal(__x, __y);
110 }
111 __DEVICE__ bool isless(float __x, float __y) {
112   return __builtin_isless(__x, __y);
113 }
114 __DEVICE__ bool isless(double __x, double __y) {
115   return __builtin_isless(__x, __y);
116 }
117 __DEVICE__ bool islessequal(float __x, float __y) {
118   return __builtin_islessequal(__x, __y);
119 }
120 __DEVICE__ bool islessequal(double __x, double __y) {
121   return __builtin_islessequal(__x, __y);
122 }
123 __DEVICE__ bool islessgreater(float __x, float __y) {
124   return __builtin_islessgreater(__x, __y);
125 }
126 __DEVICE__ bool islessgreater(double __x, double __y) {
127   return __builtin_islessgreater(__x, __y);
128 }
129 __DEVICE__ bool isnormal(float __x) { return __builtin_isnormal(__x); }
130 __DEVICE__ bool isnormal(double __x) { return __builtin_isnormal(__x); }
131 __DEVICE__ bool isunordered(float __x, float __y) {
132   return __builtin_isunordered(__x, __y);
133 }
134 __DEVICE__ bool isunordered(double __x, double __y) {
135   return __builtin_isunordered(__x, __y);
136 }
137 __DEVICE__ float ldexp(float __arg, int __exp) {
138   return ::ldexpf(__arg, __exp);
139 }
140 __DEVICE__ float log(float __x) { return ::logf(__x); }
141 __DEVICE__ float log10(float __x) { return ::log10f(__x); }
142 __DEVICE__ float modf(float __x, float *__iptr) { return ::modff(__x, __iptr); }
143 __DEVICE__ float pow(float __base, float __exp) {
144   return ::powf(__base, __exp);
145 }
146 __DEVICE__ float pow(float __base, int __iexp) {
147   return ::powif(__base, __iexp);
148 }
149 __DEVICE__ double pow(double __base, int __iexp) {
150   return ::powi(__base, __iexp);
151 }
152 __DEVICE__ bool signbit(float __x) { return ::__signbitf(__x); }
153 __DEVICE__ bool signbit(double __x) { return ::__signbitd(__x); }
154 __DEVICE__ float sin(float __x) { return ::sinf(__x); }
155 __DEVICE__ float sinh(float __x) { return ::sinhf(__x); }
156 __DEVICE__ float sqrt(float __x) { return ::sqrtf(__x); }
157 __DEVICE__ float tan(float __x) { return ::tanf(__x); }
158 __DEVICE__ float tanh(float __x) { return ::tanhf(__x); }
159 
160 // Notably missing above is nexttoward.  We omit it because
161 // libdevice doesn't provide an implementation, and we don't want to be in the
162 // business of implementing tricky libm functions in this header.
163 
164 // Now we've defined everything we promised we'd define in
165 // __clang_cuda_math_forward_declares.h.  We need to do two additional things to
166 // fix up our math functions.
167 //
168 // 1) Define __device__ overloads for e.g. sin(int).  The CUDA headers define
169 //    only sin(float) and sin(double), which means that e.g. sin(0) is
170 //    ambiguous.
171 //
172 // 2) Pull the __device__ overloads of "foobarf" math functions into namespace
173 //    std.  These are defined in the CUDA headers in the global namespace,
174 //    independent of everything else we've done here.
175 
176 // We can't use std::enable_if, because we want to be pre-C++11 compatible.  But
177 // we go ahead and unconditionally define functions that are only available when
178 // compiling for C++11 to match the behavior of the CUDA headers.
179 template<bool __B, class __T = void>
180 struct __clang_cuda_enable_if {};
181 
182 template <class __T> struct __clang_cuda_enable_if<true, __T> {
183   typedef __T type;
184 };
185 
186 // Defines an overload of __fn that accepts one integral argument, calls
187 // __fn((double)x), and returns __retty.
188 #define __CUDA_CLANG_FN_INTEGER_OVERLOAD_1(__retty, __fn)                      \
189   template <typename __T>                                                      \
190   __DEVICE__                                                                   \
191       typename __clang_cuda_enable_if<std::numeric_limits<__T>::is_integer,    \
192                                       __retty>::type                           \
193       __fn(__T __x) {                                                          \
194     return ::__fn((double)__x);                                                \
195   }
196 
197 // Defines an overload of __fn that accepts one two arithmetic arguments, calls
198 // __fn((double)x, (double)y), and returns a double.
199 //
200 // Note this is different from OVERLOAD_1, which generates an overload that
201 // accepts only *integral* arguments.
202 #define __CUDA_CLANG_FN_INTEGER_OVERLOAD_2(__retty, __fn)                      \
203   template <typename __T1, typename __T2>                                      \
204   __DEVICE__ typename __clang_cuda_enable_if<                                  \
205       std::numeric_limits<__T1>::is_specialized &&                             \
206           std::numeric_limits<__T2>::is_specialized,                           \
207       __retty>::type                                                           \
208   __fn(__T1 __x, __T2 __y) {                                                   \
209     return __fn((double)__x, (double)__y);                                     \
210   }
211 
212 __CUDA_CLANG_FN_INTEGER_OVERLOAD_1(double, acos)
213 __CUDA_CLANG_FN_INTEGER_OVERLOAD_1(double, acosh)
214 __CUDA_CLANG_FN_INTEGER_OVERLOAD_1(double, asin)
215 __CUDA_CLANG_FN_INTEGER_OVERLOAD_1(double, asinh)
216 __CUDA_CLANG_FN_INTEGER_OVERLOAD_1(double, atan)
217 __CUDA_CLANG_FN_INTEGER_OVERLOAD_2(double, atan2);
218 __CUDA_CLANG_FN_INTEGER_OVERLOAD_1(double, atanh)
219 __CUDA_CLANG_FN_INTEGER_OVERLOAD_1(double, cbrt)
220 __CUDA_CLANG_FN_INTEGER_OVERLOAD_1(double, ceil)
221 __CUDA_CLANG_FN_INTEGER_OVERLOAD_2(double, copysign);
222 __CUDA_CLANG_FN_INTEGER_OVERLOAD_1(double, cos)
223 __CUDA_CLANG_FN_INTEGER_OVERLOAD_1(double, cosh)
224 __CUDA_CLANG_FN_INTEGER_OVERLOAD_1(double, erf)
225 __CUDA_CLANG_FN_INTEGER_OVERLOAD_1(double, erfc)
226 __CUDA_CLANG_FN_INTEGER_OVERLOAD_1(double, exp)
227 __CUDA_CLANG_FN_INTEGER_OVERLOAD_1(double, exp2)
228 __CUDA_CLANG_FN_INTEGER_OVERLOAD_1(double, expm1)
229 __CUDA_CLANG_FN_INTEGER_OVERLOAD_1(double, fabs)
230 __CUDA_CLANG_FN_INTEGER_OVERLOAD_2(double, fdim);
231 __CUDA_CLANG_FN_INTEGER_OVERLOAD_1(double, floor)
232 __CUDA_CLANG_FN_INTEGER_OVERLOAD_2(double, fmax);
233 __CUDA_CLANG_FN_INTEGER_OVERLOAD_2(double, fmin);
234 __CUDA_CLANG_FN_INTEGER_OVERLOAD_2(double, fmod);
235 __CUDA_CLANG_FN_INTEGER_OVERLOAD_1(int, fpclassify)
236 __CUDA_CLANG_FN_INTEGER_OVERLOAD_2(double, hypot);
237 __CUDA_CLANG_FN_INTEGER_OVERLOAD_1(int, ilogb)
238 __CUDA_CLANG_FN_INTEGER_OVERLOAD_1(bool, isfinite)
239 __CUDA_CLANG_FN_INTEGER_OVERLOAD_2(bool, isgreater);
240 __CUDA_CLANG_FN_INTEGER_OVERLOAD_2(bool, isgreaterequal);
241 __CUDA_CLANG_FN_INTEGER_OVERLOAD_1(bool, isinf);
242 __CUDA_CLANG_FN_INTEGER_OVERLOAD_2(bool, isless);
243 __CUDA_CLANG_FN_INTEGER_OVERLOAD_2(bool, islessequal);
244 __CUDA_CLANG_FN_INTEGER_OVERLOAD_2(bool, islessgreater);
245 __CUDA_CLANG_FN_INTEGER_OVERLOAD_1(bool, isnan);
246 __CUDA_CLANG_FN_INTEGER_OVERLOAD_1(bool, isnormal)
247 __CUDA_CLANG_FN_INTEGER_OVERLOAD_2(bool, isunordered);
248 __CUDA_CLANG_FN_INTEGER_OVERLOAD_1(double, lgamma)
249 __CUDA_CLANG_FN_INTEGER_OVERLOAD_1(double, log)
250 __CUDA_CLANG_FN_INTEGER_OVERLOAD_1(double, log10)
251 __CUDA_CLANG_FN_INTEGER_OVERLOAD_1(double, log1p)
252 __CUDA_CLANG_FN_INTEGER_OVERLOAD_1(double, log2)
253 __CUDA_CLANG_FN_INTEGER_OVERLOAD_1(double, logb)
254 __CUDA_CLANG_FN_INTEGER_OVERLOAD_1(long long, llrint)
255 __CUDA_CLANG_FN_INTEGER_OVERLOAD_1(long long, llround)
256 __CUDA_CLANG_FN_INTEGER_OVERLOAD_1(long, lrint)
257 __CUDA_CLANG_FN_INTEGER_OVERLOAD_1(long, lround)
258 __CUDA_CLANG_FN_INTEGER_OVERLOAD_1(double, nearbyint);
259 __CUDA_CLANG_FN_INTEGER_OVERLOAD_2(double, nextafter);
260 __CUDA_CLANG_FN_INTEGER_OVERLOAD_2(double, pow);
261 __CUDA_CLANG_FN_INTEGER_OVERLOAD_2(double, remainder);
262 __CUDA_CLANG_FN_INTEGER_OVERLOAD_1(double, rint);
263 __CUDA_CLANG_FN_INTEGER_OVERLOAD_1(double, round);
264 __CUDA_CLANG_FN_INTEGER_OVERLOAD_1(bool, signbit)
265 __CUDA_CLANG_FN_INTEGER_OVERLOAD_1(double, sin)
266 __CUDA_CLANG_FN_INTEGER_OVERLOAD_1(double, sinh)
267 __CUDA_CLANG_FN_INTEGER_OVERLOAD_1(double, sqrt)
268 __CUDA_CLANG_FN_INTEGER_OVERLOAD_1(double, tan)
269 __CUDA_CLANG_FN_INTEGER_OVERLOAD_1(double, tanh)
270 __CUDA_CLANG_FN_INTEGER_OVERLOAD_1(double, tgamma)
271 __CUDA_CLANG_FN_INTEGER_OVERLOAD_1(double, trunc);
272 
273 #undef __CUDA_CLANG_FN_INTEGER_OVERLOAD_1
274 #undef __CUDA_CLANG_FN_INTEGER_OVERLOAD_2
275 
276 // Overloads for functions that don't match the patterns expected by
277 // __CUDA_CLANG_FN_INTEGER_OVERLOAD_{1,2}.
278 template <typename __T1, typename __T2, typename __T3>
279 __DEVICE__ typename __clang_cuda_enable_if<
280     std::numeric_limits<__T1>::is_specialized &&
281         std::numeric_limits<__T2>::is_specialized &&
282         std::numeric_limits<__T3>::is_specialized,
283     double>::type
284 fma(__T1 __x, __T2 __y, __T3 __z) {
285   return std::fma((double)__x, (double)__y, (double)__z);
286 }
287 
288 template <typename __T>
289 __DEVICE__ typename __clang_cuda_enable_if<std::numeric_limits<__T>::is_integer,
290                                            double>::type
291 frexp(__T __x, int *__exp) {
292   return std::frexp((double)__x, __exp);
293 }
294 
295 template <typename __T>
296 __DEVICE__ typename __clang_cuda_enable_if<std::numeric_limits<__T>::is_integer,
297                                            double>::type
298 ldexp(__T __x, int __exp) {
299   return std::ldexp((double)__x, __exp);
300 }
301 
302 template <typename __T1, typename __T2>
303 __DEVICE__ typename __clang_cuda_enable_if<
304     std::numeric_limits<__T1>::is_specialized &&
305         std::numeric_limits<__T2>::is_specialized,
306     double>::type
307 remquo(__T1 __x, __T2 __y, int *__quo) {
308   return std::remquo((double)__x, (double)__y, __quo);
309 }
310 
311 template <typename __T>
312 __DEVICE__ typename __clang_cuda_enable_if<std::numeric_limits<__T>::is_integer,
313                                            double>::type
314 scalbln(__T __x, long __exp) {
315   return std::scalbln((double)__x, __exp);
316 }
317 
318 template <typename __T>
319 __DEVICE__ typename __clang_cuda_enable_if<std::numeric_limits<__T>::is_integer,
320                                            double>::type
321 scalbn(__T __x, int __exp) {
322   return std::scalbn((double)__x, __exp);
323 }
324 
325 // We need to define these overloads in exactly the namespace our standard
326 // library uses (including the right inline namespace), otherwise they won't be
327 // picked up by other functions in the standard library (e.g. functions in
328 // <complex>).  Thus the ugliness below.
329 #ifdef _LIBCPP_BEGIN_NAMESPACE_STD
330 _LIBCPP_BEGIN_NAMESPACE_STD
331 #else
332 namespace std {
333 #ifdef _GLIBCXX_BEGIN_NAMESPACE_VERSION
334 _GLIBCXX_BEGIN_NAMESPACE_VERSION
335 #endif
336 #endif
337 
338 // Pull the new overloads we defined above into namespace std.
339 using ::acos;
340 using ::acosh;
341 using ::asin;
342 using ::asinh;
343 using ::atan;
344 using ::atan2;
345 using ::atanh;
346 using ::cbrt;
347 using ::ceil;
348 using ::copysign;
349 using ::cos;
350 using ::cosh;
351 using ::erf;
352 using ::erfc;
353 using ::exp;
354 using ::exp2;
355 using ::expm1;
356 using ::fabs;
357 using ::fdim;
358 using ::floor;
359 using ::fma;
360 using ::fmax;
361 using ::fmin;
362 using ::fmod;
363 using ::fpclassify;
364 using ::frexp;
365 using ::hypot;
366 using ::ilogb;
367 using ::isfinite;
368 using ::isgreater;
369 using ::isgreaterequal;
370 using ::isless;
371 using ::islessequal;
372 using ::islessgreater;
373 using ::isnormal;
374 using ::isunordered;
375 using ::ldexp;
376 using ::lgamma;
377 using ::llrint;
378 using ::llround;
379 using ::log;
380 using ::log10;
381 using ::log1p;
382 using ::log2;
383 using ::logb;
384 using ::lrint;
385 using ::lround;
386 using ::nearbyint;
387 using ::nextafter;
388 using ::pow;
389 using ::remainder;
390 using ::remquo;
391 using ::rint;
392 using ::round;
393 using ::scalbln;
394 using ::scalbn;
395 using ::signbit;
396 using ::sin;
397 using ::sinh;
398 using ::sqrt;
399 using ::tan;
400 using ::tanh;
401 using ::tgamma;
402 using ::trunc;
403 
404 // Well this is fun: We need to pull these symbols in for libc++, but we can't
405 // pull them in with libstdc++, because its ::isinf and ::isnan are different
406 // than its std::isinf and std::isnan.
407 #ifndef __GLIBCXX__
408 using ::isinf;
409 using ::isnan;
410 #endif
411 
412 // Finally, pull the "foobarf" functions that CUDA defines in its headers into
413 // namespace std.
414 using ::acosf;
415 using ::acoshf;
416 using ::asinf;
417 using ::asinhf;
418 using ::atan2f;
419 using ::atanf;
420 using ::atanhf;
421 using ::cbrtf;
422 using ::ceilf;
423 using ::copysignf;
424 using ::cosf;
425 using ::coshf;
426 using ::erfcf;
427 using ::erff;
428 using ::exp2f;
429 using ::expf;
430 using ::expm1f;
431 using ::fabsf;
432 using ::fdimf;
433 using ::floorf;
434 using ::fmaf;
435 using ::fmaxf;
436 using ::fminf;
437 using ::fmodf;
438 using ::frexpf;
439 using ::hypotf;
440 using ::ilogbf;
441 using ::ldexpf;
442 using ::lgammaf;
443 using ::llrintf;
444 using ::llroundf;
445 using ::log10f;
446 using ::log1pf;
447 using ::log2f;
448 using ::logbf;
449 using ::logf;
450 using ::lrintf;
451 using ::lroundf;
452 using ::modff;
453 using ::nearbyintf;
454 using ::nextafterf;
455 using ::powf;
456 using ::remainderf;
457 using ::remquof;
458 using ::rintf;
459 using ::roundf;
460 // TODO: remove once variant is supported
461 #ifndef _OPENMP
462 using ::scalblnf;
463 #endif
464 using ::scalbnf;
465 using ::sinf;
466 using ::sinhf;
467 using ::sqrtf;
468 using ::tanf;
469 using ::tanhf;
470 using ::tgammaf;
471 using ::truncf;
472 
473 #ifdef _LIBCPP_END_NAMESPACE_STD
474 _LIBCPP_END_NAMESPACE_STD
475 #else
476 #ifdef _GLIBCXX_BEGIN_NAMESPACE_VERSION
477 _GLIBCXX_END_NAMESPACE_VERSION
478 #endif
479 } // namespace std
480 #endif
481 
482 #undef __NOEXCEPT
483 #undef __DEVICE__
484 
485 #endif
486