xref: /freebsd/contrib/llvm-project/clang/lib/Headers/__clang_cuda_math.h (revision b891f61ef538a4e9b4658b4b756635c8036a5788)
1 /*===---- __clang_cuda_math.h - Device-side CUDA math 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_MATH_H__
10 #define __CLANG_CUDA_MATH_H__
11 #ifndef __CUDA__
12 #error "This file is for CUDA compilation only."
13 #endif
14 
15 #ifndef __OPENMP_NVPTX__
16 #if CUDA_VERSION < 9000
17 #error This file is intended to be used with CUDA-9+ only.
18 #endif
19 #endif
20 
21 // __DEVICE__ is a helper macro with common set of attributes for the wrappers
22 // we implement in this file. We need static in order to avoid emitting unused
23 // functions and __forceinline__ helps inlining these wrappers at -O1.
24 #pragma push_macro("__DEVICE__")
25 #ifdef __OPENMP_NVPTX__
26 #if defined(__cplusplus)
27 #define __DEVICE__ static constexpr __attribute__((always_inline, nothrow))
28 #else
29 #define __DEVICE__ static __attribute__((always_inline, nothrow))
30 #endif
31 #else
32 #define __DEVICE__ static __device__ __forceinline__
33 #endif
34 
35 // Specialized version of __DEVICE__ for functions with void return type. Needed
36 // because the OpenMP overlay requires constexpr functions here but prior to
37 // c++14 void return functions could not be constexpr.
38 #pragma push_macro("__DEVICE_VOID__")
39 #if defined(__OPENMP_NVPTX__) && defined(__cplusplus) && __cplusplus < 201402L
40 #define __DEVICE_VOID__ static __attribute__((always_inline, nothrow))
41 #else
42 #define __DEVICE_VOID__ __DEVICE__
43 #endif
44 
45 // libdevice provides fast low precision and slow full-recision implementations
46 // for some functions. Which one gets selected depends on
47 // __CLANG_CUDA_APPROX_TRANSCENDENTALS__ which gets defined by clang if
48 // -ffast-math or -fcuda-approx-transcendentals are in effect.
49 #pragma push_macro("__FAST_OR_SLOW")
50 #if defined(__CLANG_CUDA_APPROX_TRANSCENDENTALS__)
51 #define __FAST_OR_SLOW(fast, slow) fast
52 #else
53 #define __FAST_OR_SLOW(fast, slow) slow
54 #endif
55 
56 __DEVICE__ int abs(int __a) { return __nv_abs(__a); }
57 __DEVICE__ double fabs(double __a) { return __nv_fabs(__a); }
58 __DEVICE__ double acos(double __a) { return __nv_acos(__a); }
59 __DEVICE__ float acosf(float __a) { return __nv_acosf(__a); }
60 __DEVICE__ double acosh(double __a) { return __nv_acosh(__a); }
61 __DEVICE__ float acoshf(float __a) { return __nv_acoshf(__a); }
62 __DEVICE__ double asin(double __a) { return __nv_asin(__a); }
63 __DEVICE__ float asinf(float __a) { return __nv_asinf(__a); }
64 __DEVICE__ double asinh(double __a) { return __nv_asinh(__a); }
65 __DEVICE__ float asinhf(float __a) { return __nv_asinhf(__a); }
66 __DEVICE__ double atan(double __a) { return __nv_atan(__a); }
67 __DEVICE__ double atan2(double __a, double __b) { return __nv_atan2(__a, __b); }
68 __DEVICE__ float atan2f(float __a, float __b) { return __nv_atan2f(__a, __b); }
69 __DEVICE__ float atanf(float __a) { return __nv_atanf(__a); }
70 __DEVICE__ double atanh(double __a) { return __nv_atanh(__a); }
71 __DEVICE__ float atanhf(float __a) { return __nv_atanhf(__a); }
72 __DEVICE__ double cbrt(double __a) { return __nv_cbrt(__a); }
73 __DEVICE__ float cbrtf(float __a) { return __nv_cbrtf(__a); }
74 __DEVICE__ double ceil(double __a) { return __nv_ceil(__a); }
75 __DEVICE__ float ceilf(float __a) { return __nv_ceilf(__a); }
76 __DEVICE__ double copysign(double __a, double __b) {
77   return __nv_copysign(__a, __b);
78 }
79 __DEVICE__ float copysignf(float __a, float __b) {
80   return __nv_copysignf(__a, __b);
81 }
82 __DEVICE__ double cos(double __a) { return __nv_cos(__a); }
83 __DEVICE__ float cosf(float __a) {
84   return __FAST_OR_SLOW(__nv_fast_cosf, __nv_cosf)(__a);
85 }
86 __DEVICE__ double cosh(double __a) { return __nv_cosh(__a); }
87 __DEVICE__ float coshf(float __a) { return __nv_coshf(__a); }
88 __DEVICE__ double cospi(double __a) { return __nv_cospi(__a); }
89 __DEVICE__ float cospif(float __a) { return __nv_cospif(__a); }
90 __DEVICE__ double cyl_bessel_i0(double __a) { return __nv_cyl_bessel_i0(__a); }
91 __DEVICE__ float cyl_bessel_i0f(float __a) { return __nv_cyl_bessel_i0f(__a); }
92 __DEVICE__ double cyl_bessel_i1(double __a) { return __nv_cyl_bessel_i1(__a); }
93 __DEVICE__ float cyl_bessel_i1f(float __a) { return __nv_cyl_bessel_i1f(__a); }
94 __DEVICE__ double erf(double __a) { return __nv_erf(__a); }
95 __DEVICE__ double erfc(double __a) { return __nv_erfc(__a); }
96 __DEVICE__ float erfcf(float __a) { return __nv_erfcf(__a); }
97 __DEVICE__ double erfcinv(double __a) { return __nv_erfcinv(__a); }
98 __DEVICE__ float erfcinvf(float __a) { return __nv_erfcinvf(__a); }
99 __DEVICE__ double erfcx(double __a) { return __nv_erfcx(__a); }
100 __DEVICE__ float erfcxf(float __a) { return __nv_erfcxf(__a); }
101 __DEVICE__ float erff(float __a) { return __nv_erff(__a); }
102 __DEVICE__ double erfinv(double __a) { return __nv_erfinv(__a); }
103 __DEVICE__ float erfinvf(float __a) { return __nv_erfinvf(__a); }
104 __DEVICE__ double exp(double __a) { return __nv_exp(__a); }
105 __DEVICE__ double exp10(double __a) { return __nv_exp10(__a); }
106 __DEVICE__ float exp10f(float __a) { return __nv_exp10f(__a); }
107 __DEVICE__ double exp2(double __a) { return __nv_exp2(__a); }
108 __DEVICE__ float exp2f(float __a) { return __nv_exp2f(__a); }
109 __DEVICE__ float expf(float __a) { return __nv_expf(__a); }
110 __DEVICE__ double expm1(double __a) { return __nv_expm1(__a); }
111 __DEVICE__ float expm1f(float __a) { return __nv_expm1f(__a); }
112 __DEVICE__ float fabsf(float __a) { return __nv_fabsf(__a); }
113 __DEVICE__ double fdim(double __a, double __b) { return __nv_fdim(__a, __b); }
114 __DEVICE__ float fdimf(float __a, float __b) { return __nv_fdimf(__a, __b); }
115 __DEVICE__ double fdivide(double __a, double __b) { return __a / __b; }
116 __DEVICE__ float fdividef(float __a, float __b) {
117 #if __FAST_MATH__ && !__CUDA_PREC_DIV
118   return __nv_fast_fdividef(__a, __b);
119 #else
120   return __a / __b;
121 #endif
122 }
123 __DEVICE__ double floor(double __f) { return __nv_floor(__f); }
124 __DEVICE__ float floorf(float __f) { return __nv_floorf(__f); }
125 __DEVICE__ double fma(double __a, double __b, double __c) {
126   return __nv_fma(__a, __b, __c);
127 }
128 __DEVICE__ float fmaf(float __a, float __b, float __c) {
129   return __nv_fmaf(__a, __b, __c);
130 }
131 __DEVICE__ double fmax(double __a, double __b) { return __nv_fmax(__a, __b); }
132 __DEVICE__ float fmaxf(float __a, float __b) { return __nv_fmaxf(__a, __b); }
133 __DEVICE__ double fmin(double __a, double __b) { return __nv_fmin(__a, __b); }
134 __DEVICE__ float fminf(float __a, float __b) { return __nv_fminf(__a, __b); }
135 __DEVICE__ double fmod(double __a, double __b) { return __nv_fmod(__a, __b); }
136 __DEVICE__ float fmodf(float __a, float __b) { return __nv_fmodf(__a, __b); }
137 __DEVICE__ double frexp(double __a, int *__b) { return __nv_frexp(__a, __b); }
138 __DEVICE__ float frexpf(float __a, int *__b) { return __nv_frexpf(__a, __b); }
139 __DEVICE__ double hypot(double __a, double __b) { return __nv_hypot(__a, __b); }
140 __DEVICE__ float hypotf(float __a, float __b) { return __nv_hypotf(__a, __b); }
141 __DEVICE__ int ilogb(double __a) { return __nv_ilogb(__a); }
142 __DEVICE__ int ilogbf(float __a) { return __nv_ilogbf(__a); }
143 __DEVICE__ double j0(double __a) { return __nv_j0(__a); }
144 __DEVICE__ float j0f(float __a) { return __nv_j0f(__a); }
145 __DEVICE__ double j1(double __a) { return __nv_j1(__a); }
146 __DEVICE__ float j1f(float __a) { return __nv_j1f(__a); }
147 __DEVICE__ double jn(int __n, double __a) { return __nv_jn(__n, __a); }
148 __DEVICE__ float jnf(int __n, float __a) { return __nv_jnf(__n, __a); }
149 #if defined(__LP64__) || defined(_WIN64)
150 __DEVICE__ long labs(long __a) { return __nv_llabs(__a); };
151 #else
152 __DEVICE__ long labs(long __a) { return __nv_abs(__a); };
153 #endif
154 __DEVICE__ double ldexp(double __a, int __b) { return __nv_ldexp(__a, __b); }
155 __DEVICE__ float ldexpf(float __a, int __b) { return __nv_ldexpf(__a, __b); }
156 __DEVICE__ double lgamma(double __a) { return __nv_lgamma(__a); }
157 __DEVICE__ float lgammaf(float __a) { return __nv_lgammaf(__a); }
158 __DEVICE__ long long llabs(long long __a) { return __nv_llabs(__a); }
159 __DEVICE__ long long llmax(long long __a, long long __b) {
160   return __nv_llmax(__a, __b);
161 }
162 __DEVICE__ long long llmin(long long __a, long long __b) {
163   return __nv_llmin(__a, __b);
164 }
165 __DEVICE__ long long llrint(double __a) { return __nv_llrint(__a); }
166 __DEVICE__ long long llrintf(float __a) { return __nv_llrintf(__a); }
167 __DEVICE__ long long llround(double __a) { return __nv_llround(__a); }
168 __DEVICE__ long long llroundf(float __a) { return __nv_llroundf(__a); }
169 __DEVICE__ double round(double __a) { return __nv_round(__a); }
170 __DEVICE__ float roundf(float __a) { return __nv_roundf(__a); }
171 __DEVICE__ double log(double __a) { return __nv_log(__a); }
172 __DEVICE__ double log10(double __a) { return __nv_log10(__a); }
173 __DEVICE__ float log10f(float __a) { return __nv_log10f(__a); }
174 __DEVICE__ double log1p(double __a) { return __nv_log1p(__a); }
175 __DEVICE__ float log1pf(float __a) { return __nv_log1pf(__a); }
176 __DEVICE__ double log2(double __a) { return __nv_log2(__a); }
177 __DEVICE__ float log2f(float __a) {
178   return __FAST_OR_SLOW(__nv_fast_log2f, __nv_log2f)(__a);
179 }
180 __DEVICE__ double logb(double __a) { return __nv_logb(__a); }
181 __DEVICE__ float logbf(float __a) { return __nv_logbf(__a); }
182 __DEVICE__ float logf(float __a) {
183   return __FAST_OR_SLOW(__nv_fast_logf, __nv_logf)(__a);
184 }
185 #if defined(__LP64__) || defined(_WIN64)
186 __DEVICE__ long lrint(double __a) { return llrint(__a); }
187 __DEVICE__ long lrintf(float __a) { return __float2ll_rn(__a); }
188 __DEVICE__ long lround(double __a) { return llround(__a); }
189 __DEVICE__ long lroundf(float __a) { return llroundf(__a); }
190 #else
191 __DEVICE__ long lrint(double __a) { return (long)rint(__a); }
192 __DEVICE__ long lrintf(float __a) { return __float2int_rn(__a); }
193 __DEVICE__ long lround(double __a) { return round(__a); }
194 __DEVICE__ long lroundf(float __a) { return roundf(__a); }
195 #endif
196 __DEVICE__ int max(int __a, int __b) { return __nv_max(__a, __b); }
197 __DEVICE__ int min(int __a, int __b) { return __nv_min(__a, __b); }
198 __DEVICE__ double modf(double __a, double *__b) { return __nv_modf(__a, __b); }
199 __DEVICE__ float modff(float __a, float *__b) { return __nv_modff(__a, __b); }
200 __DEVICE__ double nearbyint(double __a) { return __builtin_nearbyint(__a); }
201 __DEVICE__ float nearbyintf(float __a) { return __builtin_nearbyintf(__a); }
202 __DEVICE__ double nextafter(double __a, double __b) {
203   return __nv_nextafter(__a, __b);
204 }
205 __DEVICE__ float nextafterf(float __a, float __b) {
206   return __nv_nextafterf(__a, __b);
207 }
208 __DEVICE__ double norm(int __dim, const double *__t) {
209   return __nv_norm(__dim, __t);
210 }
211 __DEVICE__ double norm3d(double __a, double __b, double __c) {
212   return __nv_norm3d(__a, __b, __c);
213 }
214 __DEVICE__ float norm3df(float __a, float __b, float __c) {
215   return __nv_norm3df(__a, __b, __c);
216 }
217 __DEVICE__ double norm4d(double __a, double __b, double __c, double __d) {
218   return __nv_norm4d(__a, __b, __c, __d);
219 }
220 __DEVICE__ float norm4df(float __a, float __b, float __c, float __d) {
221   return __nv_norm4df(__a, __b, __c, __d);
222 }
223 __DEVICE__ double normcdf(double __a) { return __nv_normcdf(__a); }
224 __DEVICE__ float normcdff(float __a) { return __nv_normcdff(__a); }
225 __DEVICE__ double normcdfinv(double __a) { return __nv_normcdfinv(__a); }
226 __DEVICE__ float normcdfinvf(float __a) { return __nv_normcdfinvf(__a); }
227 __DEVICE__ float normf(int __dim, const float *__t) {
228   return __nv_normf(__dim, __t);
229 }
230 __DEVICE__ double pow(double __a, double __b) { return __nv_pow(__a, __b); }
231 __DEVICE__ float powf(float __a, float __b) { return __nv_powf(__a, __b); }
232 __DEVICE__ double powi(double __a, int __b) { return __nv_powi(__a, __b); }
233 __DEVICE__ float powif(float __a, int __b) { return __nv_powif(__a, __b); }
234 __DEVICE__ double rcbrt(double __a) { return __nv_rcbrt(__a); }
235 __DEVICE__ float rcbrtf(float __a) { return __nv_rcbrtf(__a); }
236 __DEVICE__ double remainder(double __a, double __b) {
237   return __nv_remainder(__a, __b);
238 }
239 __DEVICE__ float remainderf(float __a, float __b) {
240   return __nv_remainderf(__a, __b);
241 }
242 __DEVICE__ double remquo(double __a, double __b, int *__c) {
243   return __nv_remquo(__a, __b, __c);
244 }
245 __DEVICE__ float remquof(float __a, float __b, int *__c) {
246   return __nv_remquof(__a, __b, __c);
247 }
248 __DEVICE__ double rhypot(double __a, double __b) {
249   return __nv_rhypot(__a, __b);
250 }
251 __DEVICE__ float rhypotf(float __a, float __b) {
252   return __nv_rhypotf(__a, __b);
253 }
254 // __nv_rint* in libdevice is buggy and produces incorrect results.
255 __DEVICE__ double rint(double __a) { return __builtin_rint(__a); }
256 __DEVICE__ float rintf(float __a) { return __builtin_rintf(__a); }
257 __DEVICE__ double rnorm(int __a, const double *__b) {
258   return __nv_rnorm(__a, __b);
259 }
260 __DEVICE__ double rnorm3d(double __a, double __b, double __c) {
261   return __nv_rnorm3d(__a, __b, __c);
262 }
263 __DEVICE__ float rnorm3df(float __a, float __b, float __c) {
264   return __nv_rnorm3df(__a, __b, __c);
265 }
266 __DEVICE__ double rnorm4d(double __a, double __b, double __c, double __d) {
267   return __nv_rnorm4d(__a, __b, __c, __d);
268 }
269 __DEVICE__ float rnorm4df(float __a, float __b, float __c, float __d) {
270   return __nv_rnorm4df(__a, __b, __c, __d);
271 }
272 __DEVICE__ float rnormf(int __dim, const float *__t) {
273   return __nv_rnormf(__dim, __t);
274 }
275 __DEVICE__ double rsqrt(double __a) { return __nv_rsqrt(__a); }
276 __DEVICE__ float rsqrtf(float __a) { return __nv_rsqrtf(__a); }
277 __DEVICE__ double scalbn(double __a, int __b) { return __nv_scalbn(__a, __b); }
278 __DEVICE__ float scalbnf(float __a, int __b) { return __nv_scalbnf(__a, __b); }
279 __DEVICE__ double scalbln(double __a, long __b) {
280   if (__b > INT_MAX)
281     return __a > 0 ? HUGE_VAL : -HUGE_VAL;
282   if (__b < INT_MIN)
283     return __a > 0 ? 0.0 : -0.0;
284   return scalbn(__a, (int)__b);
285 }
286 __DEVICE__ float scalblnf(float __a, long __b) {
287   if (__b > INT_MAX)
288     return __a > 0 ? HUGE_VALF : -HUGE_VALF;
289   if (__b < INT_MIN)
290     return __a > 0 ? 0.f : -0.f;
291   return scalbnf(__a, (int)__b);
292 }
293 __DEVICE__ double sin(double __a) { return __nv_sin(__a); }
294 __DEVICE_VOID__ void sincos(double __a, double *__s, double *__c) {
295   return __nv_sincos(__a, __s, __c);
296 }
297 __DEVICE_VOID__ void sincosf(float __a, float *__s, float *__c) {
298   return __FAST_OR_SLOW(__nv_fast_sincosf, __nv_sincosf)(__a, __s, __c);
299 }
300 __DEVICE_VOID__ void sincospi(double __a, double *__s, double *__c) {
301   return __nv_sincospi(__a, __s, __c);
302 }
303 __DEVICE_VOID__ void sincospif(float __a, float *__s, float *__c) {
304   return __nv_sincospif(__a, __s, __c);
305 }
306 __DEVICE__ float sinf(float __a) {
307   return __FAST_OR_SLOW(__nv_fast_sinf, __nv_sinf)(__a);
308 }
309 __DEVICE__ double sinh(double __a) { return __nv_sinh(__a); }
310 __DEVICE__ float sinhf(float __a) { return __nv_sinhf(__a); }
311 __DEVICE__ double sinpi(double __a) { return __nv_sinpi(__a); }
312 __DEVICE__ float sinpif(float __a) { return __nv_sinpif(__a); }
313 __DEVICE__ double sqrt(double __a) { return __nv_sqrt(__a); }
314 __DEVICE__ float sqrtf(float __a) { return __nv_sqrtf(__a); }
315 __DEVICE__ double tan(double __a) { return __nv_tan(__a); }
316 __DEVICE__ float tanf(float __a) { return __nv_tanf(__a); }
317 __DEVICE__ double tanh(double __a) { return __nv_tanh(__a); }
318 __DEVICE__ float tanhf(float __a) { return __nv_tanhf(__a); }
319 __DEVICE__ double tgamma(double __a) { return __nv_tgamma(__a); }
320 __DEVICE__ float tgammaf(float __a) { return __nv_tgammaf(__a); }
321 __DEVICE__ double trunc(double __a) { return __nv_trunc(__a); }
322 __DEVICE__ float truncf(float __a) { return __nv_truncf(__a); }
323 __DEVICE__ unsigned long long ullmax(unsigned long long __a,
324                                      unsigned long long __b) {
325   return __nv_ullmax(__a, __b);
326 }
327 __DEVICE__ unsigned long long ullmin(unsigned long long __a,
328                                      unsigned long long __b) {
329   return __nv_ullmin(__a, __b);
330 }
331 __DEVICE__ unsigned int umax(unsigned int __a, unsigned int __b) {
332   return __nv_umax(__a, __b);
333 }
334 __DEVICE__ unsigned int umin(unsigned int __a, unsigned int __b) {
335   return __nv_umin(__a, __b);
336 }
337 __DEVICE__ double y0(double __a) { return __nv_y0(__a); }
338 __DEVICE__ float y0f(float __a) { return __nv_y0f(__a); }
339 __DEVICE__ double y1(double __a) { return __nv_y1(__a); }
340 __DEVICE__ float y1f(float __a) { return __nv_y1f(__a); }
341 __DEVICE__ double yn(int __a, double __b) { return __nv_yn(__a, __b); }
342 __DEVICE__ float ynf(int __a, float __b) { return __nv_ynf(__a, __b); }
343 
344 #pragma pop_macro("__DEVICE__")
345 #pragma pop_macro("__DEVICE_VOID__")
346 #pragma pop_macro("__FAST_OR_SLOW")
347 
348 #endif // __CLANG_CUDA_MATH_H__
349