| 1 |
/*===-- __clang_cuda_complex_builtins - CUDA impls of runtime complex fns ---=== |
| 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 |
|
| 10 |
#ifndef __CLANG_CUDA_COMPLEX_BUILTINS |
| 11 |
#define __CLANG_CUDA_COMPLEX_BUILTINS |
| 12 |
|
| 13 |
// This header defines __muldc3, __mulsc3, __divdc3, and __divsc3. These are |
| 14 |
// libgcc functions that clang assumes are available when compiling c99 complex |
| 15 |
// operations. (These implementations come from libc++, and have been modified |
| 16 |
// to work with CUDA and OpenMP target offloading [in C and C++ mode].) |
| 17 |
|
| 18 |
#pragma push_macro("__DEVICE__") |
| 19 |
#if defined(__OPENMP_NVPTX__) || defined(__OPENMP_AMDGCN__) |
| 20 |
#pragma omp declare target |
| 21 |
#define __DEVICE__ __attribute__((noinline, nothrow, cold, weak)) |
| 22 |
#else |
| 23 |
#define __DEVICE__ __device__ inline |
| 24 |
#endif |
| 25 |
|
| 26 |
// To make the algorithms available for C and C++ in CUDA and OpenMP we select |
| 27 |
// different but equivalent function versions. TODO: For OpenMP we currently |
| 28 |
// select the native builtins as the overload support for templates is lacking. |
| 29 |
#if !defined(__OPENMP_NVPTX__) && !defined(__OPENMP_AMDGCN__) |
| 30 |
#define _ISNANd std::isnan |
| 31 |
#define _ISNANf std::isnan |
| 32 |
#define _ISINFd std::isinf |
| 33 |
#define _ISINFf std::isinf |
| 34 |
#define _ISFINITEd std::isfinite |
| 35 |
#define _ISFINITEf std::isfinite |
| 36 |
#define _COPYSIGNd std::copysign |
| 37 |
#define _COPYSIGNf std::copysign |
| 38 |
#define _SCALBNd std::scalbn |
| 39 |
#define _SCALBNf std::scalbn |
| 40 |
#define _ABSd std::abs |
| 41 |
#define _ABSf std::abs |
| 42 |
#define _LOGBd std::logb |
| 43 |
#define _LOGBf std::logb |
| 44 |
// Rather than pulling in std::max from algorithm everytime, use available ::max. |
| 45 |
#define _fmaxd max |
| 46 |
#define _fmaxf max |
| 47 |
#else |
| 48 |
#ifdef __AMDGCN__ |
| 49 |
#define _ISNANd __ocml_isnan_f64 |
| 50 |
#define _ISNANf __ocml_isnan_f32 |
| 51 |
#define _ISINFd __ocml_isinf_f64 |
| 52 |
#define _ISINFf __ocml_isinf_f32 |
| 53 |
#define _ISFINITEd __ocml_isfinite_f64 |
| 54 |
#define _ISFINITEf __ocml_isfinite_f32 |
| 55 |
#define _COPYSIGNd __ocml_copysign_f64 |
| 56 |
#define _COPYSIGNf __ocml_copysign_f32 |
| 57 |
#define _SCALBNd __ocml_scalbn_f64 |
| 58 |
#define _SCALBNf __ocml_scalbn_f32 |
| 59 |
#define _ABSd __ocml_fabs_f64 |
| 60 |
#define _ABSf __ocml_fabs_f32 |
| 61 |
#define _LOGBd __ocml_logb_f64 |
| 62 |
#define _LOGBf __ocml_logb_f32 |
| 63 |
#define _fmaxd __ocml_fmax_f64 |
| 64 |
#define _fmaxf __ocml_fmax_f32 |
| 65 |
#else |
| 66 |
#define _ISNANd __nv_isnand |
| 67 |
#define _ISNANf __nv_isnanf |
| 68 |
#define _ISINFd __nv_isinfd |
| 69 |
#define _ISINFf __nv_isinff |
| 70 |
#define _ISFINITEd __nv_isfinited |
| 71 |
#define _ISFINITEf __nv_finitef |
| 72 |
#define _COPYSIGNd __nv_copysign |
| 73 |
#define _COPYSIGNf __nv_copysignf |
| 74 |
#define _SCALBNd __nv_scalbn |
| 75 |
#define _SCALBNf __nv_scalbnf |
| 76 |
#define _ABSd __nv_fabs |
| 77 |
#define _ABSf __nv_fabsf |
| 78 |
#define _LOGBd __nv_logb |
| 79 |
#define _LOGBf __nv_logbf |
| 80 |
#define _fmaxd __nv_fmax |
| 81 |
#define _fmaxf __nv_fmaxf |
| 82 |
#endif |
| 83 |
#endif |
| 84 |
|
| 85 |
#if defined(__cplusplus) |
| 86 |
extern "C" { |
| 87 |
#endif |
| 88 |
|
| 89 |
__DEVICE__ double _Complex __muldc3(double __a, double __b, double __c, |
| 90 |
double __d) { |
| 91 |
double __ac = __a * __c; |
| 92 |
double __bd = __b * __d; |
| 93 |
double __ad = __a * __d; |
| 94 |
double __bc = __b * __c; |
| 95 |
double _Complex z; |
| 96 |
__real__(z) = __ac - __bd; |
| 97 |
__imag__(z) = __ad + __bc; |
| 98 |
if (_ISNANd(__real__(z)) && _ISNANd(__imag__(z))) { |
| 99 |
int __recalc = 0; |
| 100 |
if (_ISINFd(__a) || _ISINFd(__b)) { |
| 101 |
__a = _COPYSIGNd(_ISINFd(__a) ? 1 : 0, __a); |
| 102 |
__b = _COPYSIGNd(_ISINFd(__b) ? 1 : 0, __b); |
| 103 |
if (_ISNANd(__c)) |
| 104 |
__c = _COPYSIGNd(0, __c); |
| 105 |
if (_ISNANd(__d)) |
| 106 |
__d = _COPYSIGNd(0, __d); |
| 107 |
__recalc = 1; |
| 108 |
} |
| 109 |
if (_ISINFd(__c) || _ISINFd(__d)) { |
| 110 |
__c = _COPYSIGNd(_ISINFd(__c) ? 1 : 0, __c); |
| 111 |
__d = _COPYSIGNd(_ISINFd(__d) ? 1 : 0, __d); |
| 112 |
if (_ISNANd(__a)) |
| 113 |
__a = _COPYSIGNd(0, __a); |
| 114 |
if (_ISNANd(__b)) |
| 115 |
__b = _COPYSIGNd(0, __b); |
| 116 |
__recalc = 1; |
| 117 |
} |
| 118 |
if (!__recalc && |
| 119 |
(_ISINFd(__ac) || _ISINFd(__bd) || _ISINFd(__ad) || _ISINFd(__bc))) { |
| 120 |
if (_ISNANd(__a)) |
| 121 |
__a = _COPYSIGNd(0, __a); |
| 122 |
if (_ISNANd(__b)) |
| 123 |
__b = _COPYSIGNd(0, __b); |
| 124 |
if (_ISNANd(__c)) |
| 125 |
__c = _COPYSIGNd(0, __c); |
| 126 |
if (_ISNANd(__d)) |
| 127 |
__d = _COPYSIGNd(0, __d); |
| 128 |
__recalc = 1; |
| 129 |
} |
| 130 |
if (__recalc) { |
| 131 |
// Can't use std::numeric_limits<double>::infinity() -- that doesn't have |
| 132 |
// a device overload (and isn't constexpr before C++11, naturally). |
| 133 |
__real__(z) = __builtin_huge_val() * (__a * __c - __b * __d); |
| 134 |
__imag__(z) = __builtin_huge_val() * (__a * __d + __b * __c); |
| 135 |
} |
| 136 |
} |
| 137 |
return z; |
| 138 |
} |
| 139 |
|
| 140 |
__DEVICE__ float _Complex __mulsc3(float __a, float __b, float __c, float __d) { |
| 141 |
float __ac = __a * __c; |
| 142 |
float __bd = __b * __d; |
| 143 |
float __ad = __a * __d; |
| 144 |
float __bc = __b * __c; |
| 145 |
float _Complex z; |
| 146 |
__real__(z) = __ac - __bd; |
| 147 |
__imag__(z) = __ad + __bc; |
| 148 |
if (_ISNANf(__real__(z)) && _ISNANf(__imag__(z))) { |
| 149 |
int __recalc = 0; |
| 150 |
if (_ISINFf(__a) || _ISINFf(__b)) { |
| 151 |
__a = _COPYSIGNf(_ISINFf(__a) ? 1 : 0, __a); |
| 152 |
__b = _COPYSIGNf(_ISINFf(__b) ? 1 : 0, __b); |
| 153 |
if (_ISNANf(__c)) |
| 154 |
__c = _COPYSIGNf(0, __c); |
| 155 |
if (_ISNANf(__d)) |
| 156 |
__d = _COPYSIGNf(0, __d); |
| 157 |
__recalc = 1; |
| 158 |
} |
| 159 |
if (_ISINFf(__c) || _ISINFf(__d)) { |
| 160 |
__c = _COPYSIGNf(_ISINFf(__c) ? 1 : 0, __c); |
| 161 |
__d = _COPYSIGNf(_ISINFf(__d) ? 1 : 0, __d); |
| 162 |
if (_ISNANf(__a)) |
| 163 |
__a = _COPYSIGNf(0, __a); |
| 164 |
if (_ISNANf(__b)) |
| 165 |
__b = _COPYSIGNf(0, __b); |
| 166 |
__recalc = 1; |
| 167 |
} |
| 168 |
if (!__recalc && |
| 169 |
(_ISINFf(__ac) || _ISINFf(__bd) || _ISINFf(__ad) || _ISINFf(__bc))) { |
| 170 |
if (_ISNANf(__a)) |
| 171 |
__a = _COPYSIGNf(0, __a); |
| 172 |
if (_ISNANf(__b)) |
| 173 |
__b = _COPYSIGNf(0, __b); |
| 174 |
if (_ISNANf(__c)) |
| 175 |
__c = _COPYSIGNf(0, __c); |
| 176 |
if (_ISNANf(__d)) |
| 177 |
__d = _COPYSIGNf(0, __d); |
| 178 |
__recalc = 1; |
| 179 |
} |
| 180 |
if (__recalc) { |
| 181 |
__real__(z) = __builtin_huge_valf() * (__a * __c - __b * __d); |
| 182 |
__imag__(z) = __builtin_huge_valf() * (__a * __d + __b * __c); |
| 183 |
} |
| 184 |
} |
| 185 |
return z; |
| 186 |
} |
| 187 |
|
| 188 |
__DEVICE__ double _Complex __divdc3(double __a, double __b, double __c, |
| 189 |
double __d) { |
| 190 |
int __ilogbw = 0; |
| 191 |
// Can't use std::max, because that's defined in <algorithm>, and we don't |
| 192 |
// want to pull that in for every compile. The CUDA headers define |
| 193 |
// ::max(float, float) and ::max(double, double), which is sufficient for us. |
| 194 |
double __logbw = _LOGBd(_fmaxd(_ABSd(__c), _ABSd(__d))); |
| 195 |
if (_ISFINITEd(__logbw)) { |
| 196 |
__ilogbw = (int)__logbw; |
| 197 |
__c = _SCALBNd(__c, -__ilogbw); |
| 198 |
__d = _SCALBNd(__d, -__ilogbw); |
| 199 |
} |
| 200 |
double __denom = __c * __c + __d * __d; |
| 201 |
double _Complex z; |
| 202 |
__real__(z) = _SCALBNd((__a * __c + __b * __d) / __denom, -__ilogbw); |
| 203 |
__imag__(z) = _SCALBNd((__b * __c - __a * __d) / __denom, -__ilogbw); |
| 204 |
if (_ISNANd(__real__(z)) && _ISNANd(__imag__(z))) { |
| 205 |
if ((__denom == 0.0) && (!_ISNANd(__a) || !_ISNANd(__b))) { |
| 206 |
__real__(z) = _COPYSIGNd(__builtin_huge_val(), __c) * __a; |
| 207 |
__imag__(z) = _COPYSIGNd(__builtin_huge_val(), __c) * __b; |
| 208 |
} else if ((_ISINFd(__a) || _ISINFd(__b)) && _ISFINITEd(__c) && |
| 209 |
_ISFINITEd(__d)) { |
| 210 |
__a = _COPYSIGNd(_ISINFd(__a) ? 1.0 : 0.0, __a); |
| 211 |
__b = _COPYSIGNd(_ISINFd(__b) ? 1.0 : 0.0, __b); |
| 212 |
__real__(z) = __builtin_huge_val() * (__a * __c + __b * __d); |
| 213 |
__imag__(z) = __builtin_huge_val() * (__b * __c - __a * __d); |
| 214 |
} else if (_ISINFd(__logbw) && __logbw > 0.0 && _ISFINITEd(__a) && |
| 215 |
_ISFINITEd(__b)) { |
| 216 |
__c = _COPYSIGNd(_ISINFd(__c) ? 1.0 : 0.0, __c); |
| 217 |
__d = _COPYSIGNd(_ISINFd(__d) ? 1.0 : 0.0, __d); |
| 218 |
__real__(z) = 0.0 * (__a * __c + __b * __d); |
| 219 |
__imag__(z) = 0.0 * (__b * __c - __a * __d); |
| 220 |
} |
| 221 |
} |
| 222 |
return z; |
| 223 |
} |
| 224 |
|
| 225 |
__DEVICE__ float _Complex __divsc3(float __a, float __b, float __c, float __d) { |
| 226 |
int __ilogbw = 0; |
| 227 |
float __logbw = _LOGBf(_fmaxf(_ABSf(__c), _ABSf(__d))); |
| 228 |
if (_ISFINITEf(__logbw)) { |
| 229 |
__ilogbw = (int)__logbw; |
| 230 |
__c = _SCALBNf(__c, -__ilogbw); |
| 231 |
__d = _SCALBNf(__d, -__ilogbw); |
| 232 |
} |
| 233 |
float __denom = __c * __c + __d * __d; |
| 234 |
float _Complex z; |
| 235 |
__real__(z) = _SCALBNf((__a * __c + __b * __d) / __denom, -__ilogbw); |
| 236 |
__imag__(z) = _SCALBNf((__b * __c - __a * __d) / __denom, -__ilogbw); |
| 237 |
if (_ISNANf(__real__(z)) && _ISNANf(__imag__(z))) { |
| 238 |
if ((__denom == 0) && (!_ISNANf(__a) || !_ISNANf(__b))) { |
| 239 |
__real__(z) = _COPYSIGNf(__builtin_huge_valf(), __c) * __a; |
| 240 |
__imag__(z) = _COPYSIGNf(__builtin_huge_valf(), __c) * __b; |
| 241 |
} else if ((_ISINFf(__a) || _ISINFf(__b)) && _ISFINITEf(__c) && |
| 242 |
_ISFINITEf(__d)) { |
| 243 |
__a = _COPYSIGNf(_ISINFf(__a) ? 1 : 0, __a); |
| 244 |
__b = _COPYSIGNf(_ISINFf(__b) ? 1 : 0, __b); |
| 245 |
__real__(z) = __builtin_huge_valf() * (__a * __c + __b * __d); |
| 246 |
__imag__(z) = __builtin_huge_valf() * (__b * __c - __a * __d); |
| 247 |
} else if (_ISINFf(__logbw) && __logbw > 0 && _ISFINITEf(__a) && |
| 248 |
_ISFINITEf(__b)) { |
| 249 |
__c = _COPYSIGNf(_ISINFf(__c) ? 1 : 0, __c); |
| 250 |
__d = _COPYSIGNf(_ISINFf(__d) ? 1 : 0, __d); |
| 251 |
__real__(z) = 0 * (__a * __c + __b * __d); |
| 252 |
__imag__(z) = 0 * (__b * __c - __a * __d); |
| 253 |
} |
| 254 |
} |
| 255 |
return z; |
| 256 |
} |
| 257 |
|
| 258 |
#if defined(__cplusplus) |
| 259 |
} // extern "C" |
| 260 |
#endif |
| 261 |
|
| 262 |
#undef _ISNANd |
| 263 |
#undef _ISNANf |
| 264 |
#undef _ISINFd |
| 265 |
#undef _ISINFf |
| 266 |
#undef _COPYSIGNd |
| 267 |
#undef _COPYSIGNf |
| 268 |
#undef _ISFINITEd |
| 269 |
#undef _ISFINITEf |
| 270 |
#undef _SCALBNd |
| 271 |
#undef _SCALBNf |
| 272 |
#undef _ABSd |
| 273 |
#undef _ABSf |
| 274 |
#undef _LOGBd |
| 275 |
#undef _LOGBf |
| 276 |
#undef _fmaxd |
| 277 |
#undef _fmaxf |
| 278 |
|
| 279 |
#if defined(__OPENMP_NVPTX__) || defined(__OPENMP_AMDGCN__) |
| 280 |
#pragma omp end declare target |
| 281 |
#endif |
| 282 |
|
| 283 |
#pragma pop_macro("__DEVICE__") |
| 284 |
|
| 285 |
#endif // __CLANG_CUDA_COMPLEX_BUILTINS |