| 1 | /*===-- __clang_cuda_complex_builtins - CUDA impls of runtime complex fns ---=== |
| 2 | * |
| 3 | * Permission is hereby granted, free of charge, to any person obtaining a copy |
| 4 | * of this software and associated documentation files (the "Software"), to deal |
| 5 | * in the Software without restriction, including without limitation the rights |
| 6 | * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell |
| 7 | * copies of the Software, and to permit persons to whom the Software is |
| 8 | * furnished to do so, subject to the following conditions: |
| 9 | * |
| 10 | * The above copyright notice and this permission notice shall be included in |
| 11 | * all copies or substantial portions of the Software. |
| 12 | * |
| 13 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR |
| 14 | * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, |
| 15 | * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE |
| 16 | * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER |
| 17 | * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, |
| 18 | * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN |
| 19 | * THE SOFTWARE. |
| 20 | * |
| 21 | *===-----------------------------------------------------------------------=== |
| 22 | */ |
| 23 | |
| 24 | #ifndef __CLANG_CUDA_COMPLEX_BUILTINS |
| 25 | #define __CLANG_CUDA_COMPLEX_BUILTINS |
| 26 | |
| 27 | // This header defines __muldc3, __mulsc3, __divdc3, and __divsc3. These are |
| 28 | // libgcc functions that clang assumes are available when compiling c99 complex |
| 29 | // operations. (These implementations come from libc++, and have been modified |
| 30 | // to work with CUDA.) |
| 31 | |
| 32 | extern "C" inline __device__ double _Complex __muldc3(double __a, double __b, |
| 33 | double __c, double __d) { |
| 34 | double __ac = __a * __c; |
| 35 | double __bd = __b * __d; |
| 36 | double __ad = __a * __d; |
| 37 | double __bc = __b * __c; |
| 38 | double _Complex z; |
| 39 | __real__(z) = __ac - __bd; |
| 40 | __imag__(z) = __ad + __bc; |
| 41 | if (std::isnan(__real__(z)) && std::isnan(__imag__(z))) { |
| 42 | int __recalc = 0; |
| 43 | if (std::isinf(__a) || std::isinf(__b)) { |
| 44 | __a = std::copysign(std::isinf(__a) ? 1 : 0, __a); |
| 45 | __b = std::copysign(std::isinf(__b) ? 1 : 0, __b); |
| 46 | if (std::isnan(__c)) |
| 47 | __c = std::copysign(0, __c); |
| 48 | if (std::isnan(__d)) |
| 49 | __d = std::copysign(0, __d); |
| 50 | __recalc = 1; |
| 51 | } |
| 52 | if (std::isinf(__c) || std::isinf(__d)) { |
| 53 | __c = std::copysign(std::isinf(__c) ? 1 : 0, __c); |
| 54 | __d = std::copysign(std::isinf(__d) ? 1 : 0, __d); |
| 55 | if (std::isnan(__a)) |
| 56 | __a = std::copysign(0, __a); |
| 57 | if (std::isnan(__b)) |
| 58 | __b = std::copysign(0, __b); |
| 59 | __recalc = 1; |
| 60 | } |
| 61 | if (!__recalc && (std::isinf(__ac) || std::isinf(__bd) || |
| 62 | std::isinf(__ad) || std::isinf(__bc))) { |
| 63 | if (std::isnan(__a)) |
| 64 | __a = std::copysign(0, __a); |
| 65 | if (std::isnan(__b)) |
| 66 | __b = std::copysign(0, __b); |
| 67 | if (std::isnan(__c)) |
| 68 | __c = std::copysign(0, __c); |
| 69 | if (std::isnan(__d)) |
| 70 | __d = std::copysign(0, __d); |
| 71 | __recalc = 1; |
| 72 | } |
| 73 | if (__recalc) { |
| 74 | // Can't use std::numeric_limits<double>::infinity() -- that doesn't have |
| 75 | // a device overload (and isn't constexpr before C++11, naturally). |
| 76 | __real__(z) = __builtin_huge_valf() * (__a * __c - __b * __d); |
| 77 | __imag__(z) = __builtin_huge_valf() * (__a * __d + __b * __c); |
| 78 | } |
| 79 | } |
| 80 | return z; |
| 81 | } |
| 82 | |
| 83 | extern "C" inline __device__ float _Complex __mulsc3(float __a, float __b, |
| 84 | float __c, float __d) { |
| 85 | float __ac = __a * __c; |
| 86 | float __bd = __b * __d; |
| 87 | float __ad = __a * __d; |
| 88 | float __bc = __b * __c; |
| 89 | float _Complex z; |
| 90 | __real__(z) = __ac - __bd; |
| 91 | __imag__(z) = __ad + __bc; |
| 92 | if (std::isnan(__real__(z)) && std::isnan(__imag__(z))) { |
| 93 | int __recalc = 0; |
| 94 | if (std::isinf(__a) || std::isinf(__b)) { |
| 95 | __a = std::copysign(std::isinf(__a) ? 1 : 0, __a); |
| 96 | __b = std::copysign(std::isinf(__b) ? 1 : 0, __b); |
| 97 | if (std::isnan(__c)) |
| 98 | __c = std::copysign(0, __c); |
| 99 | if (std::isnan(__d)) |
| 100 | __d = std::copysign(0, __d); |
| 101 | __recalc = 1; |
| 102 | } |
| 103 | if (std::isinf(__c) || std::isinf(__d)) { |
| 104 | __c = std::copysign(std::isinf(__c) ? 1 : 0, __c); |
| 105 | __d = std::copysign(std::isinf(__d) ? 1 : 0, __d); |
| 106 | if (std::isnan(__a)) |
| 107 | __a = std::copysign(0, __a); |
| 108 | if (std::isnan(__b)) |
| 109 | __b = std::copysign(0, __b); |
| 110 | __recalc = 1; |
| 111 | } |
| 112 | if (!__recalc && (std::isinf(__ac) || std::isinf(__bd) || |
| 113 | std::isinf(__ad) || std::isinf(__bc))) { |
| 114 | if (std::isnan(__a)) |
| 115 | __a = std::copysign(0, __a); |
| 116 | if (std::isnan(__b)) |
| 117 | __b = std::copysign(0, __b); |
| 118 | if (std::isnan(__c)) |
| 119 | __c = std::copysign(0, __c); |
| 120 | if (std::isnan(__d)) |
| 121 | __d = std::copysign(0, __d); |
| 122 | __recalc = 1; |
| 123 | } |
| 124 | if (__recalc) { |
| 125 | __real__(z) = __builtin_huge_valf() * (__a * __c - __b * __d); |
| 126 | __imag__(z) = __builtin_huge_valf() * (__a * __d + __b * __c); |
| 127 | } |
| 128 | } |
| 129 | return z; |
| 130 | } |
| 131 | |
| 132 | extern "C" inline __device__ double _Complex __divdc3(double __a, double __b, |
| 133 | double __c, double __d) { |
| 134 | int __ilogbw = 0; |
| 135 | // Can't use std::max, because that's defined in <algorithm>, and we don't |
| 136 | // want to pull that in for every compile. The CUDA headers define |
| 137 | // ::max(float, float) and ::max(double, double), which is sufficient for us. |
| 138 | double __logbw = std::logb(max(std::abs(__c), std::abs(__d))); |
| 139 | if (std::isfinite(__logbw)) { |
| 140 | __ilogbw = (int)__logbw; |
| 141 | __c = std::scalbn(__c, -__ilogbw); |
| 142 | __d = std::scalbn(__d, -__ilogbw); |
| 143 | } |
| 144 | double __denom = __c * __c + __d * __d; |
| 145 | double _Complex z; |
| 146 | __real__(z) = std::scalbn((__a * __c + __b * __d) / __denom, -__ilogbw); |
| 147 | __imag__(z) = std::scalbn((__b * __c - __a * __d) / __denom, -__ilogbw); |
| 148 | if (std::isnan(__real__(z)) && std::isnan(__imag__(z))) { |
| 149 | if ((__denom == 0.0) && (!std::isnan(__a) || !std::isnan(__b))) { |
| 150 | __real__(z) = std::copysign(__builtin_huge_valf(), __c) * __a; |
| 151 | __imag__(z) = std::copysign(__builtin_huge_valf(), __c) * __b; |
| 152 | } else if ((std::isinf(__a) || std::isinf(__b)) && std::isfinite(__c) && |
| 153 | std::isfinite(__d)) { |
| 154 | __a = std::copysign(std::isinf(__a) ? 1.0 : 0.0, __a); |
| 155 | __b = std::copysign(std::isinf(__b) ? 1.0 : 0.0, __b); |
| 156 | __real__(z) = __builtin_huge_valf() * (__a * __c + __b * __d); |
| 157 | __imag__(z) = __builtin_huge_valf() * (__b * __c - __a * __d); |
| 158 | } else if (std::isinf(__logbw) && __logbw > 0.0 && std::isfinite(__a) && |
| 159 | std::isfinite(__b)) { |
| 160 | __c = std::copysign(std::isinf(__c) ? 1.0 : 0.0, __c); |
| 161 | __d = std::copysign(std::isinf(__d) ? 1.0 : 0.0, __d); |
| 162 | __real__(z) = 0.0 * (__a * __c + __b * __d); |
| 163 | __imag__(z) = 0.0 * (__b * __c - __a * __d); |
| 164 | } |
| 165 | } |
| 166 | return z; |
| 167 | } |
| 168 | |
| 169 | extern "C" inline __device__ float _Complex __divsc3(float __a, float __b, |
| 170 | float __c, float __d) { |
| 171 | int __ilogbw = 0; |
| 172 | float __logbw = std::logb(max(std::abs(__c), std::abs(__d))); |
| 173 | if (std::isfinite(__logbw)) { |
| 174 | __ilogbw = (int)__logbw; |
| 175 | __c = std::scalbn(__c, -__ilogbw); |
| 176 | __d = std::scalbn(__d, -__ilogbw); |
| 177 | } |
| 178 | float __denom = __c * __c + __d * __d; |
| 179 | float _Complex z; |
| 180 | __real__(z) = std::scalbn((__a * __c + __b * __d) / __denom, -__ilogbw); |
| 181 | __imag__(z) = std::scalbn((__b * __c - __a * __d) / __denom, -__ilogbw); |
| 182 | if (std::isnan(__real__(z)) && std::isnan(__imag__(z))) { |
| 183 | if ((__denom == 0) && (!std::isnan(__a) || !std::isnan(__b))) { |
| 184 | __real__(z) = std::copysign(__builtin_huge_valf(), __c) * __a; |
| 185 | __imag__(z) = std::copysign(__builtin_huge_valf(), __c) * __b; |
| 186 | } else if ((std::isinf(__a) || std::isinf(__b)) && std::isfinite(__c) && |
| 187 | std::isfinite(__d)) { |
| 188 | __a = std::copysign(std::isinf(__a) ? 1 : 0, __a); |
| 189 | __b = std::copysign(std::isinf(__b) ? 1 : 0, __b); |
| 190 | __real__(z) = __builtin_huge_valf() * (__a * __c + __b * __d); |
| 191 | __imag__(z) = __builtin_huge_valf() * (__b * __c - __a * __d); |
| 192 | } else if (std::isinf(__logbw) && __logbw > 0 && std::isfinite(__a) && |
| 193 | std::isfinite(__b)) { |
| 194 | __c = std::copysign(std::isinf(__c) ? 1 : 0, __c); |
| 195 | __d = std::copysign(std::isinf(__d) ? 1 : 0, __d); |
| 196 | __real__(z) = 0 * (__a * __c + __b * __d); |
| 197 | __imag__(z) = 0 * (__b * __c - __a * __d); |
| 198 | } |
| 199 | } |
| 200 | return z; |
| 201 | } |
| 202 | |
| 203 | #endif // __CLANG_CUDA_COMPLEX_BUILTINS |
| 204 | |