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 -fgpu-approx-transcendentals are in effect. 49#pragma push_macro("__FAST_OR_SLOW") 50#if defined(__CLANG_GPU_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