1 /*===-- __clang_cuda_complex_builtins - CUDA impls of runtime complex fns ---===
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
7 *===-----------------------------------------------------------------------===
10 #ifndef __CLANG_CUDA_COMPLEX_BUILTINS
11 #define __CLANG_CUDA_COMPLEX_BUILTINS
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].)
18 #pragma push_macro("__DEVICE__")
19 #ifdef __OPENMP_NVPTX__
20 #pragma omp declare target
21 #define __DEVICE__ __attribute__((noinline, nothrow, cold, weak))
23 #define __DEVICE__ __device__ inline
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__)
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
45 #define _ISNANd __nv_isnand
46 #define _ISNANf __nv_isnanf
47 #define _ISINFd __nv_isinfd
48 #define _ISINFf __nv_isinff
49 #define _ISFINITEd __nv_isfinited
50 #define _ISFINITEf __nv_finitef
51 #define _COPYSIGNd __nv_copysign
52 #define _COPYSIGNf __nv_copysignf
53 #define _SCALBNd __nv_scalbn
54 #define _SCALBNf __nv_scalbnf
55 #define _ABSd __nv_fabs
56 #define _ABSf __nv_fabsf
57 #define _LOGBd __nv_logb
58 #define _LOGBf __nv_logbf
61 #if defined(__cplusplus)
65 __DEVICE__ double _Complex __muldc3(double __a, double __b, double __c,
67 double __ac = __a * __c;
68 double __bd = __b * __d;
69 double __ad = __a * __d;
70 double __bc = __b * __c;
72 __real__(z) = __ac - __bd;
73 __imag__(z) = __ad + __bc;
74 if (_ISNANd(__real__(z)) && _ISNANd(__imag__(z))) {
76 if (_ISINFd(__a) || _ISINFd(__b)) {
77 __a = _COPYSIGNd(_ISINFd(__a) ? 1 : 0, __a);
78 __b = _COPYSIGNd(_ISINFd(__b) ? 1 : 0, __b);
80 __c = _COPYSIGNd(0, __c);
82 __d = _COPYSIGNd(0, __d);
85 if (_ISINFd(__c) || _ISINFd(__d)) {
86 __c = _COPYSIGNd(_ISINFd(__c) ? 1 : 0, __c);
87 __d = _COPYSIGNd(_ISINFd(__d) ? 1 : 0, __d);
89 __a = _COPYSIGNd(0, __a);
91 __b = _COPYSIGNd(0, __b);
95 (_ISINFd(__ac) || _ISINFd(__bd) || _ISINFd(__ad) || _ISINFd(__bc))) {
97 __a = _COPYSIGNd(0, __a);
99 __b = _COPYSIGNd(0, __b);
101 __c = _COPYSIGNd(0, __c);
103 __d = _COPYSIGNd(0, __d);
107 // Can't use std::numeric_limits<double>::infinity() -- that doesn't have
108 // a device overload (and isn't constexpr before C++11, naturally).
109 __real__(z) = __builtin_huge_val() * (__a * __c - __b * __d);
110 __imag__(z) = __builtin_huge_val() * (__a * __d + __b * __c);
116 __DEVICE__ float _Complex __mulsc3(float __a, float __b, float __c, float __d) {
117 float __ac = __a * __c;
118 float __bd = __b * __d;
119 float __ad = __a * __d;
120 float __bc = __b * __c;
122 __real__(z) = __ac - __bd;
123 __imag__(z) = __ad + __bc;
124 if (_ISNANf(__real__(z)) && _ISNANf(__imag__(z))) {
126 if (_ISINFf(__a) || _ISINFf(__b)) {
127 __a = _COPYSIGNf(_ISINFf(__a) ? 1 : 0, __a);
128 __b = _COPYSIGNf(_ISINFf(__b) ? 1 : 0, __b);
130 __c = _COPYSIGNf(0, __c);
132 __d = _COPYSIGNf(0, __d);
135 if (_ISINFf(__c) || _ISINFf(__d)) {
136 __c = _COPYSIGNf(_ISINFf(__c) ? 1 : 0, __c);
137 __d = _COPYSIGNf(_ISINFf(__d) ? 1 : 0, __d);
139 __a = _COPYSIGNf(0, __a);
141 __b = _COPYSIGNf(0, __b);
145 (_ISINFf(__ac) || _ISINFf(__bd) || _ISINFf(__ad) || _ISINFf(__bc))) {
147 __a = _COPYSIGNf(0, __a);
149 __b = _COPYSIGNf(0, __b);
151 __c = _COPYSIGNf(0, __c);
153 __d = _COPYSIGNf(0, __d);
157 __real__(z) = __builtin_huge_valf() * (__a * __c - __b * __d);
158 __imag__(z) = __builtin_huge_valf() * (__a * __d + __b * __c);
164 __DEVICE__ double _Complex __divdc3(double __a, double __b, double __c,
167 // Can't use std::max, because that's defined in <algorithm>, and we don't
168 // want to pull that in for every compile. The CUDA headers define
169 // ::max(float, float) and ::max(double, double), which is sufficient for us.
170 double __logbw = _LOGBd(max(_ABSd(__c), _ABSd(__d)));
171 if (_ISFINITEd(__logbw)) {
172 __ilogbw = (int)__logbw;
173 __c = _SCALBNd(__c, -__ilogbw);
174 __d = _SCALBNd(__d, -__ilogbw);
176 double __denom = __c * __c + __d * __d;
178 __real__(z) = _SCALBNd((__a * __c + __b * __d) / __denom, -__ilogbw);
179 __imag__(z) = _SCALBNd((__b * __c - __a * __d) / __denom, -__ilogbw);
180 if (_ISNANd(__real__(z)) && _ISNANd(__imag__(z))) {
181 if ((__denom == 0.0) && (!_ISNANd(__a) || !_ISNANd(__b))) {
182 __real__(z) = _COPYSIGNd(__builtin_huge_val(), __c) * __a;
183 __imag__(z) = _COPYSIGNd(__builtin_huge_val(), __c) * __b;
184 } else if ((_ISINFd(__a) || _ISINFd(__b)) && _ISFINITEd(__c) &&
186 __a = _COPYSIGNd(_ISINFd(__a) ? 1.0 : 0.0, __a);
187 __b = _COPYSIGNd(_ISINFd(__b) ? 1.0 : 0.0, __b);
188 __real__(z) = __builtin_huge_val() * (__a * __c + __b * __d);
189 __imag__(z) = __builtin_huge_val() * (__b * __c - __a * __d);
190 } else if (_ISINFd(__logbw) && __logbw > 0.0 && _ISFINITEd(__a) &&
192 __c = _COPYSIGNd(_ISINFd(__c) ? 1.0 : 0.0, __c);
193 __d = _COPYSIGNd(_ISINFd(__d) ? 1.0 : 0.0, __d);
194 __real__(z) = 0.0 * (__a * __c + __b * __d);
195 __imag__(z) = 0.0 * (__b * __c - __a * __d);
201 __DEVICE__ float _Complex __divsc3(float __a, float __b, float __c, float __d) {
203 float __logbw = _LOGBf(max(_ABSf(__c), _ABSf(__d)));
204 if (_ISFINITEf(__logbw)) {
205 __ilogbw = (int)__logbw;
206 __c = _SCALBNf(__c, -__ilogbw);
207 __d = _SCALBNf(__d, -__ilogbw);
209 float __denom = __c * __c + __d * __d;
211 __real__(z) = _SCALBNf((__a * __c + __b * __d) / __denom, -__ilogbw);
212 __imag__(z) = _SCALBNf((__b * __c - __a * __d) / __denom, -__ilogbw);
213 if (_ISNANf(__real__(z)) && _ISNANf(__imag__(z))) {
214 if ((__denom == 0) && (!_ISNANf(__a) || !_ISNANf(__b))) {
215 __real__(z) = _COPYSIGNf(__builtin_huge_valf(), __c) * __a;
216 __imag__(z) = _COPYSIGNf(__builtin_huge_valf(), __c) * __b;
217 } else if ((_ISINFf(__a) || _ISINFf(__b)) && _ISFINITEf(__c) &&
219 __a = _COPYSIGNf(_ISINFf(__a) ? 1 : 0, __a);
220 __b = _COPYSIGNf(_ISINFf(__b) ? 1 : 0, __b);
221 __real__(z) = __builtin_huge_valf() * (__a * __c + __b * __d);
222 __imag__(z) = __builtin_huge_valf() * (__b * __c - __a * __d);
223 } else if (_ISINFf(__logbw) && __logbw > 0 && _ISFINITEf(__a) &&
225 __c = _COPYSIGNf(_ISINFf(__c) ? 1 : 0, __c);
226 __d = _COPYSIGNf(_ISINFf(__d) ? 1 : 0, __d);
227 __real__(z) = 0 * (__a * __c + __b * __d);
228 __imag__(z) = 0 * (__b * __c - __a * __d);
234 #if defined(__cplusplus)
253 #ifdef __OPENMP_NVPTX__
254 #pragma omp end declare target
257 #pragma pop_macro("__DEVICE__")
259 #endif // __CLANG_CUDA_COMPLEX_BUILTINS