clang  8.0.0svn
__clang_cuda_device_functions.h
Go to the documentation of this file.
1 /*===---- __clang_cuda_device_functions.h - CUDA runtime support -----------===
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_DEVICE_FUNCTIONS_H__
25 #define __CLANG_CUDA_DEVICE_FUNCTIONS_H__
26 
27 #if CUDA_VERSION < 9000
28 #error This file is intended to be used with CUDA-9+ only.
29 #endif
30 
31 // __DEVICE__ is a helper macro with common set of attributes for the wrappers
32 // we implement in this file. We need static in order to avoid emitting unused
33 // functions and __forceinline__ helps inlining these wrappers at -O1.
34 #pragma push_macro("__DEVICE__")
35 #define __DEVICE__ static __device__ __forceinline__
36 
37 // libdevice provides fast low precision and slow full-recision implementations
38 // for some functions. Which one gets selected depends on
39 // __CLANG_CUDA_APPROX_TRANSCENDENTALS__ which gets defined by clang if
40 // -ffast-math or -fcuda-approx-transcendentals are in effect.
41 #pragma push_macro("__FAST_OR_SLOW")
42 #if defined(__CLANG_CUDA_APPROX_TRANSCENDENTALS__)
43 #define __FAST_OR_SLOW(fast, slow) fast
44 #else
45 #define __FAST_OR_SLOW(fast, slow) slow
46 #endif
47 
48 __DEVICE__ int __all(int __a) { return __nvvm_vote_all(__a); }
49 __DEVICE__ int __any(int __a) { return __nvvm_vote_any(__a); }
50 __DEVICE__ unsigned int __ballot(int __a) { return __nvvm_vote_ballot(__a); }
51 __DEVICE__ unsigned int __brev(unsigned int __a) { return __nv_brev(__a); }
52 __DEVICE__ unsigned long long __brevll(unsigned long long __a) {
53  return __nv_brevll(__a);
54 }
55 __DEVICE__ void __brkpt() { asm volatile("brkpt;"); }
56 __DEVICE__ void __brkpt(int __a) { __brkpt(); }
57 __DEVICE__ unsigned int __byte_perm(unsigned int __a, unsigned int __b,
58  unsigned int __c) {
59  return __nv_byte_perm(__a, __b, __c);
60 }
61 __DEVICE__ int __clz(int __a) { return __nv_clz(__a); }
62 __DEVICE__ int __clzll(long long __a) { return __nv_clzll(__a); }
63 __DEVICE__ float __cosf(float __a) { return __nv_fast_cosf(__a); }
64 __DEVICE__ double __dAtomicAdd(double *__p, double __v) {
65  return __nvvm_atom_add_gen_d(__p, __v);
66 }
67 __DEVICE__ double __dAtomicAdd_block(double *__p, double __v) {
68  return __nvvm_atom_cta_add_gen_d(__p, __v);
69 }
70 __DEVICE__ double __dAtomicAdd_system(double *__p, double __v) {
71  return __nvvm_atom_sys_add_gen_d(__p, __v);
72 }
73 __DEVICE__ double __dadd_rd(double __a, double __b) {
74  return __nv_dadd_rd(__a, __b);
75 }
76 __DEVICE__ double __dadd_rn(double __a, double __b) {
77  return __nv_dadd_rn(__a, __b);
78 }
79 __DEVICE__ double __dadd_ru(double __a, double __b) {
80  return __nv_dadd_ru(__a, __b);
81 }
82 __DEVICE__ double __dadd_rz(double __a, double __b) {
83  return __nv_dadd_rz(__a, __b);
84 }
85 __DEVICE__ double __ddiv_rd(double __a, double __b) {
86  return __nv_ddiv_rd(__a, __b);
87 }
88 __DEVICE__ double __ddiv_rn(double __a, double __b) {
89  return __nv_ddiv_rn(__a, __b);
90 }
91 __DEVICE__ double __ddiv_ru(double __a, double __b) {
92  return __nv_ddiv_ru(__a, __b);
93 }
94 __DEVICE__ double __ddiv_rz(double __a, double __b) {
95  return __nv_ddiv_rz(__a, __b);
96 }
97 __DEVICE__ double __dmul_rd(double __a, double __b) {
98  return __nv_dmul_rd(__a, __b);
99 }
100 __DEVICE__ double __dmul_rn(double __a, double __b) {
101  return __nv_dmul_rn(__a, __b);
102 }
103 __DEVICE__ double __dmul_ru(double __a, double __b) {
104  return __nv_dmul_ru(__a, __b);
105 }
106 __DEVICE__ double __dmul_rz(double __a, double __b) {
107  return __nv_dmul_rz(__a, __b);
108 }
110  return __nv_double2float_rd(__a);
111 }
113  return __nv_double2float_rn(__a);
114 }
116  return __nv_double2float_ru(__a);
117 }
119  return __nv_double2float_rz(__a);
120 }
121 __DEVICE__ int __double2hiint(double __a) { return __nv_double2hiint(__a); }
122 __DEVICE__ int __double2int_rd(double __a) { return __nv_double2int_rd(__a); }
123 __DEVICE__ int __double2int_rn(double __a) { return __nv_double2int_rn(__a); }
124 __DEVICE__ int __double2int_ru(double __a) { return __nv_double2int_ru(__a); }
125 __DEVICE__ int __double2int_rz(double __a) { return __nv_double2int_rz(__a); }
126 __DEVICE__ long long __double2ll_rd(double __a) {
127  return __nv_double2ll_rd(__a);
128 }
129 __DEVICE__ long long __double2ll_rn(double __a) {
130  return __nv_double2ll_rn(__a);
131 }
132 __DEVICE__ long long __double2ll_ru(double __a) {
133  return __nv_double2ll_ru(__a);
134 }
135 __DEVICE__ long long __double2ll_rz(double __a) {
136  return __nv_double2ll_rz(__a);
137 }
138 __DEVICE__ int __double2loint(double __a) { return __nv_double2loint(__a); }
139 __DEVICE__ unsigned int __double2uint_rd(double __a) {
140  return __nv_double2uint_rd(__a);
141 }
142 __DEVICE__ unsigned int __double2uint_rn(double __a) {
143  return __nv_double2uint_rn(__a);
144 }
145 __DEVICE__ unsigned int __double2uint_ru(double __a) {
146  return __nv_double2uint_ru(__a);
147 }
148 __DEVICE__ unsigned int __double2uint_rz(double __a) {
149  return __nv_double2uint_rz(__a);
150 }
151 __DEVICE__ unsigned long long __double2ull_rd(double __a) {
152  return __nv_double2ull_rd(__a);
153 }
154 __DEVICE__ unsigned long long __double2ull_rn(double __a) {
155  return __nv_double2ull_rn(__a);
156 }
157 __DEVICE__ unsigned long long __double2ull_ru(double __a) {
158  return __nv_double2ull_ru(__a);
159 }
160 __DEVICE__ unsigned long long __double2ull_rz(double __a) {
161  return __nv_double2ull_rz(__a);
162 }
164  return __nv_double_as_longlong(__a);
165 }
166 __DEVICE__ double __drcp_rd(double __a) { return __nv_drcp_rd(__a); }
167 __DEVICE__ double __drcp_rn(double __a) { return __nv_drcp_rn(__a); }
168 __DEVICE__ double __drcp_ru(double __a) { return __nv_drcp_ru(__a); }
169 __DEVICE__ double __drcp_rz(double __a) { return __nv_drcp_rz(__a); }
170 __DEVICE__ double __dsqrt_rd(double __a) { return __nv_dsqrt_rd(__a); }
171 __DEVICE__ double __dsqrt_rn(double __a) { return __nv_dsqrt_rn(__a); }
172 __DEVICE__ double __dsqrt_ru(double __a) { return __nv_dsqrt_ru(__a); }
173 __DEVICE__ double __dsqrt_rz(double __a) { return __nv_dsqrt_rz(__a); }
174 __DEVICE__ double __dsub_rd(double __a, double __b) {
175  return __nv_dsub_rd(__a, __b);
176 }
177 __DEVICE__ double __dsub_rn(double __a, double __b) {
178  return __nv_dsub_rn(__a, __b);
179 }
180 __DEVICE__ double __dsub_ru(double __a, double __b) {
181  return __nv_dsub_ru(__a, __b);
182 }
183 __DEVICE__ double __dsub_rz(double __a, double __b) {
184  return __nv_dsub_rz(__a, __b);
185 }
186 __DEVICE__ float __exp10f(float __a) { return __nv_fast_exp10f(__a); }
187 __DEVICE__ float __expf(float __a) { return __nv_fast_expf(__a); }
188 __DEVICE__ float __fAtomicAdd(float *__p, float __v) {
189  return __nvvm_atom_add_gen_f(__p, __v);
190 }
191 __DEVICE__ float __fAtomicAdd_block(float *__p, float __v) {
192  return __nvvm_atom_cta_add_gen_f(__p, __v);
193 }
194 __DEVICE__ float __fAtomicAdd_system(float *__p, float __v) {
195  return __nvvm_atom_sys_add_gen_f(__p, __v);
196 }
197 __DEVICE__ float __fAtomicExch(float *__p, float __v) {
198  return __nv_int_as_float(
199  __nvvm_atom_xchg_gen_i((int *)__p, __nv_float_as_int(__v)));
200 }
201 __DEVICE__ float __fAtomicExch_block(float *__p, float __v) {
202  return __nv_int_as_float(
203  __nvvm_atom_cta_xchg_gen_i((int *)__p, __nv_float_as_int(__v)));
204 }
205 __DEVICE__ float __fAtomicExch_system(float *__p, float __v) {
206  return __nv_int_as_float(
207  __nvvm_atom_sys_xchg_gen_i((int *)__p, __nv_float_as_int(__v)));
208 }
209 __DEVICE__ float __fadd_rd(float __a, float __b) {
210  return __nv_fadd_rd(__a, __b);
211 }
212 __DEVICE__ float __fadd_rn(float __a, float __b) {
213  return __nv_fadd_rn(__a, __b);
214 }
215 __DEVICE__ float __fadd_ru(float __a, float __b) {
216  return __nv_fadd_ru(__a, __b);
217 }
218 __DEVICE__ float __fadd_rz(float __a, float __b) {
219  return __nv_fadd_rz(__a, __b);
220 }
221 __DEVICE__ float __fdiv_rd(float __a, float __b) {
222  return __nv_fdiv_rd(__a, __b);
223 }
224 __DEVICE__ float __fdiv_rn(float __a, float __b) {
225  return __nv_fdiv_rn(__a, __b);
226 }
227 __DEVICE__ float __fdiv_ru(float __a, float __b) {
228  return __nv_fdiv_ru(__a, __b);
229 }
230 __DEVICE__ float __fdiv_rz(float __a, float __b) {
231  return __nv_fdiv_rz(__a, __b);
232 }
233 __DEVICE__ float __fdividef(float __a, float __b) {
234  return __nv_fast_fdividef(__a, __b);
235 }
236 __DEVICE__ int __ffs(int __a) { return __nv_ffs(__a); }
237 __DEVICE__ int __ffsll(long long __a) { return __nv_ffsll(__a); }
238 __DEVICE__ int __finite(double __a) { return __nv_isfinited(__a); }
239 __DEVICE__ int __finitef(float __a) { return __nv_finitef(__a); }
240 __DEVICE__ int __float2int_rd(float __a) { return __nv_float2int_rd(__a); }
241 __DEVICE__ int __float2int_rn(float __a) { return __nv_float2int_rn(__a); }
242 __DEVICE__ int __float2int_ru(float __a) { return __nv_float2int_ru(__a); }
243 __DEVICE__ int __float2int_rz(float __a) { return __nv_float2int_rz(__a); }
244 __DEVICE__ long long __float2ll_rd(float __a) { return __nv_float2ll_rd(__a); }
245 __DEVICE__ long long __float2ll_rn(float __a) { return __nv_float2ll_rn(__a); }
246 __DEVICE__ long long __float2ll_ru(float __a) { return __nv_float2ll_ru(__a); }
247 __DEVICE__ long long __float2ll_rz(float __a) { return __nv_float2ll_rz(__a); }
248 __DEVICE__ unsigned int __float2uint_rd(float __a) {
249  return __nv_float2uint_rd(__a);
250 }
251 __DEVICE__ unsigned int __float2uint_rn(float __a) {
252  return __nv_float2uint_rn(__a);
253 }
254 __DEVICE__ unsigned int __float2uint_ru(float __a) {
255  return __nv_float2uint_ru(__a);
256 }
257 __DEVICE__ unsigned int __float2uint_rz(float __a) {
258  return __nv_float2uint_rz(__a);
259 }
260 __DEVICE__ unsigned long long __float2ull_rd(float __a) {
261  return __nv_float2ull_rd(__a);
262 }
263 __DEVICE__ unsigned long long __float2ull_rn(float __a) {
264  return __nv_float2ull_rn(__a);
265 }
266 __DEVICE__ unsigned long long __float2ull_ru(float __a) {
267  return __nv_float2ull_ru(__a);
268 }
269 __DEVICE__ unsigned long long __float2ull_rz(float __a) {
270  return __nv_float2ull_rz(__a);
271 }
272 __DEVICE__ int __float_as_int(float __a) { return __nv_float_as_int(__a); }
273 __DEVICE__ unsigned int __float_as_uint(float __a) {
274  return __nv_float_as_uint(__a);
275 }
276 __DEVICE__ double __fma_rd(double __a, double __b, double __c) {
277  return __nv_fma_rd(__a, __b, __c);
278 }
279 __DEVICE__ double __fma_rn(double __a, double __b, double __c) {
280  return __nv_fma_rn(__a, __b, __c);
281 }
282 __DEVICE__ double __fma_ru(double __a, double __b, double __c) {
283  return __nv_fma_ru(__a, __b, __c);
284 }
285 __DEVICE__ double __fma_rz(double __a, double __b, double __c) {
286  return __nv_fma_rz(__a, __b, __c);
287 }
288 __DEVICE__ float __fmaf_ieee_rd(float __a, float __b, float __c) {
289  return __nv_fmaf_ieee_rd(__a, __b, __c);
290 }
291 __DEVICE__ float __fmaf_ieee_rn(float __a, float __b, float __c) {
292  return __nv_fmaf_ieee_rn(__a, __b, __c);
293 }
294 __DEVICE__ float __fmaf_ieee_ru(float __a, float __b, float __c) {
295  return __nv_fmaf_ieee_ru(__a, __b, __c);
296 }
297 __DEVICE__ float __fmaf_ieee_rz(float __a, float __b, float __c) {
298  return __nv_fmaf_ieee_rz(__a, __b, __c);
299 }
300 __DEVICE__ float __fmaf_rd(float __a, float __b, float __c) {
301  return __nv_fmaf_rd(__a, __b, __c);
302 }
303 __DEVICE__ float __fmaf_rn(float __a, float __b, float __c) {
304  return __nv_fmaf_rn(__a, __b, __c);
305 }
306 __DEVICE__ float __fmaf_ru(float __a, float __b, float __c) {
307  return __nv_fmaf_ru(__a, __b, __c);
308 }
309 __DEVICE__ float __fmaf_rz(float __a, float __b, float __c) {
310  return __nv_fmaf_rz(__a, __b, __c);
311 }
312 __DEVICE__ float __fmul_rd(float __a, float __b) {
313  return __nv_fmul_rd(__a, __b);
314 }
315 __DEVICE__ float __fmul_rn(float __a, float __b) {
316  return __nv_fmul_rn(__a, __b);
317 }
318 __DEVICE__ float __fmul_ru(float __a, float __b) {
319  return __nv_fmul_ru(__a, __b);
320 }
321 __DEVICE__ float __fmul_rz(float __a, float __b) {
322  return __nv_fmul_rz(__a, __b);
323 }
324 __DEVICE__ float __frcp_rd(float __a) { return __nv_frcp_rd(__a); }
325 __DEVICE__ float __frcp_rn(float __a) { return __nv_frcp_rn(__a); }
326 __DEVICE__ float __frcp_ru(float __a) { return __nv_frcp_ru(__a); }
327 __DEVICE__ float __frcp_rz(float __a) { return __nv_frcp_rz(__a); }
328 __DEVICE__ float __frsqrt_rn(float __a) { return __nv_frsqrt_rn(__a); }
329 __DEVICE__ float __fsqrt_rd(float __a) { return __nv_fsqrt_rd(__a); }
330 __DEVICE__ float __fsqrt_rn(float __a) { return __nv_fsqrt_rn(__a); }
331 __DEVICE__ float __fsqrt_ru(float __a) { return __nv_fsqrt_ru(__a); }
332 __DEVICE__ float __fsqrt_rz(float __a) { return __nv_fsqrt_rz(__a); }
333 __DEVICE__ float __fsub_rd(float __a, float __b) {
334  return __nv_fsub_rd(__a, __b);
335 }
336 __DEVICE__ float __fsub_rn(float __a, float __b) {
337  return __nv_fsub_rn(__a, __b);
338 }
339 __DEVICE__ float __fsub_ru(float __a, float __b) {
340  return __nv_fsub_ru(__a, __b);
341 }
342 __DEVICE__ float __fsub_rz(float __a, float __b) {
343  return __nv_fsub_rz(__a, __b);
344 }
345 __DEVICE__ int __hadd(int __a, int __b) { return __nv_hadd(__a, __b); }
346 __DEVICE__ double __hiloint2double(int __a, int __b) {
347  return __nv_hiloint2double(__a, __b);
348 }
349 __DEVICE__ int __iAtomicAdd(int *__p, int __v) {
350  return __nvvm_atom_add_gen_i(__p, __v);
351 }
352 __DEVICE__ int __iAtomicAdd_block(int *__p, int __v) {
353  __nvvm_atom_cta_add_gen_i(__p, __v);
354 }
355 __DEVICE__ int __iAtomicAdd_system(int *__p, int __v) {
356  __nvvm_atom_sys_add_gen_i(__p, __v);
357 }
358 __DEVICE__ int __iAtomicAnd(int *__p, int __v) {
359  return __nvvm_atom_and_gen_i(__p, __v);
360 }
361 __DEVICE__ int __iAtomicAnd_block(int *__p, int __v) {
362  return __nvvm_atom_cta_and_gen_i(__p, __v);
363 }
364 __DEVICE__ int __iAtomicAnd_system(int *__p, int __v) {
365  return __nvvm_atom_sys_and_gen_i(__p, __v);
366 }
367 __DEVICE__ int __iAtomicCAS(int *__p, int __cmp, int __v) {
368  return __nvvm_atom_cas_gen_i(__p, __cmp, __v);
369 }
370 __DEVICE__ int __iAtomicCAS_block(int *__p, int __cmp, int __v) {
371  return __nvvm_atom_cta_cas_gen_i(__p, __cmp, __v);
372 }
373 __DEVICE__ int __iAtomicCAS_system(int *__p, int __cmp, int __v) {
374  return __nvvm_atom_sys_cas_gen_i(__p, __cmp, __v);
375 }
376 __DEVICE__ int __iAtomicExch(int *__p, int __v) {
377  return __nvvm_atom_xchg_gen_i(__p, __v);
378 }
379 __DEVICE__ int __iAtomicExch_block(int *__p, int __v) {
380  return __nvvm_atom_cta_xchg_gen_i(__p, __v);
381 }
382 __DEVICE__ int __iAtomicExch_system(int *__p, int __v) {
383  return __nvvm_atom_sys_xchg_gen_i(__p, __v);
384 }
385 __DEVICE__ int __iAtomicMax(int *__p, int __v) {
386  return __nvvm_atom_max_gen_i(__p, __v);
387 }
388 __DEVICE__ int __iAtomicMax_block(int *__p, int __v) {
389  return __nvvm_atom_cta_max_gen_i(__p, __v);
390 }
391 __DEVICE__ int __iAtomicMax_system(int *__p, int __v) {
392  return __nvvm_atom_sys_max_gen_i(__p, __v);
393 }
394 __DEVICE__ int __iAtomicMin(int *__p, int __v) {
395  return __nvvm_atom_min_gen_i(__p, __v);
396 }
397 __DEVICE__ int __iAtomicMin_block(int *__p, int __v) {
398  return __nvvm_atom_cta_min_gen_i(__p, __v);
399 }
400 __DEVICE__ int __iAtomicMin_system(int *__p, int __v) {
401  return __nvvm_atom_sys_min_gen_i(__p, __v);
402 }
403 __DEVICE__ int __iAtomicOr(int *__p, int __v) {
404  return __nvvm_atom_or_gen_i(__p, __v);
405 }
406 __DEVICE__ int __iAtomicOr_block(int *__p, int __v) {
407  return __nvvm_atom_cta_or_gen_i(__p, __v);
408 }
409 __DEVICE__ int __iAtomicOr_system(int *__p, int __v) {
410  return __nvvm_atom_sys_or_gen_i(__p, __v);
411 }
412 __DEVICE__ int __iAtomicXor(int *__p, int __v) {
413  return __nvvm_atom_xor_gen_i(__p, __v);
414 }
415 __DEVICE__ int __iAtomicXor_block(int *__p, int __v) {
416  return __nvvm_atom_cta_xor_gen_i(__p, __v);
417 }
418 __DEVICE__ int __iAtomicXor_system(int *__p, int __v) {
419  return __nvvm_atom_sys_xor_gen_i(__p, __v);
420 }
421 __DEVICE__ long long __illAtomicMax(long long *__p, long long __v) {
422  return __nvvm_atom_max_gen_ll(__p, __v);
423 }
424 __DEVICE__ long long __illAtomicMax_block(long long *__p, long long __v) {
425  return __nvvm_atom_cta_max_gen_ll(__p, __v);
426 }
427 __DEVICE__ long long __illAtomicMax_system(long long *__p, long long __v) {
428  return __nvvm_atom_sys_max_gen_ll(__p, __v);
429 }
430 __DEVICE__ long long __illAtomicMin(long long *__p, long long __v) {
431  return __nvvm_atom_min_gen_ll(__p, __v);
432 }
433 __DEVICE__ long long __illAtomicMin_block(long long *__p, long long __v) {
434  return __nvvm_atom_cta_min_gen_ll(__p, __v);
435 }
436 __DEVICE__ long long __illAtomicMin_system(long long *__p, long long __v) {
437  return __nvvm_atom_sys_min_gen_ll(__p, __v);
438 }
439 __DEVICE__ double __int2double_rn(int __a) { return __nv_int2double_rn(__a); }
440 __DEVICE__ float __int2float_rd(int __a) { return __nv_int2float_rd(__a); }
441 __DEVICE__ float __int2float_rn(int __a) { return __nv_int2float_rn(__a); }
442 __DEVICE__ float __int2float_ru(int __a) { return __nv_int2float_ru(__a); }
443 __DEVICE__ float __int2float_rz(int __a) { return __nv_int2float_rz(__a); }
444 __DEVICE__ float __int_as_float(int __a) { return __nv_int_as_float(__a); }
445 __DEVICE__ int __isfinited(double __a) { return __nv_isfinited(__a); }
446 __DEVICE__ int __isinf(double __a) { return __nv_isinfd(__a); }
447 __DEVICE__ int __isinff(float __a) { return __nv_isinff(__a); }
448 __DEVICE__ int __isnan(double __a) { return __nv_isnand(__a); }
449 __DEVICE__ int __isnanf(float __a) { return __nv_isnanf(__a); }
450 __DEVICE__ double __ll2double_rd(long long __a) {
451  return __nv_ll2double_rd(__a);
452 }
453 __DEVICE__ double __ll2double_rn(long long __a) {
454  return __nv_ll2double_rn(__a);
455 }
456 __DEVICE__ double __ll2double_ru(long long __a) {
457  return __nv_ll2double_ru(__a);
458 }
459 __DEVICE__ double __ll2double_rz(long long __a) {
460  return __nv_ll2double_rz(__a);
461 }
462 __DEVICE__ float __ll2float_rd(long long __a) { return __nv_ll2float_rd(__a); }
463 __DEVICE__ float __ll2float_rn(long long __a) { return __nv_ll2float_rn(__a); }
464 __DEVICE__ float __ll2float_ru(long long __a) { return __nv_ll2float_ru(__a); }
465 __DEVICE__ float __ll2float_rz(long long __a) { return __nv_ll2float_rz(__a); }
466 __DEVICE__ long long __llAtomicAnd(long long *__p, long long __v) {
467  return __nvvm_atom_and_gen_ll(__p, __v);
468 }
469 __DEVICE__ long long __llAtomicAnd_block(long long *__p, long long __v) {
470  return __nvvm_atom_cta_and_gen_ll(__p, __v);
471 }
472 __DEVICE__ long long __llAtomicAnd_system(long long *__p, long long __v) {
473  return __nvvm_atom_sys_and_gen_ll(__p, __v);
474 }
475 __DEVICE__ long long __llAtomicOr(long long *__p, long long __v) {
476  return __nvvm_atom_or_gen_ll(__p, __v);
477 }
478 __DEVICE__ long long __llAtomicOr_block(long long *__p, long long __v) {
479  return __nvvm_atom_cta_or_gen_ll(__p, __v);
480 }
481 __DEVICE__ long long __llAtomicOr_system(long long *__p, long long __v) {
482  return __nvvm_atom_sys_or_gen_ll(__p, __v);
483 }
484 __DEVICE__ long long __llAtomicXor(long long *__p, long long __v) {
485  return __nvvm_atom_xor_gen_ll(__p, __v);
486 }
487 __DEVICE__ long long __llAtomicXor_block(long long *__p, long long __v) {
488  return __nvvm_atom_cta_xor_gen_ll(__p, __v);
489 }
490 __DEVICE__ long long __llAtomicXor_system(long long *__p, long long __v) {
491  return __nvvm_atom_sys_xor_gen_ll(__p, __v);
492 }
493 __DEVICE__ float __log10f(float __a) { return __nv_fast_log10f(__a); }
494 __DEVICE__ float __log2f(float __a) { return __nv_fast_log2f(__a); }
495 __DEVICE__ float __logf(float __a) { return __nv_fast_logf(__a); }
497  return __nv_longlong_as_double(__a);
498 }
499 __DEVICE__ int __mul24(int __a, int __b) { return __nv_mul24(__a, __b); }
500 __DEVICE__ long long __mul64hi(long long __a, long long __b) {
501  return __nv_mul64hi(__a, __b);
502 }
503 __DEVICE__ int __mulhi(int __a, int __b) { return __nv_mulhi(__a, __b); }
504 __DEVICE__ unsigned int __pm0(void) { return __nvvm_read_ptx_sreg_pm0(); }
505 __DEVICE__ unsigned int __pm1(void) { return __nvvm_read_ptx_sreg_pm1(); }
506 __DEVICE__ unsigned int __pm2(void) { return __nvvm_read_ptx_sreg_pm2(); }
507 __DEVICE__ unsigned int __pm3(void) { return __nvvm_read_ptx_sreg_pm3(); }
508 __DEVICE__ int __popc(int __a) { return __nv_popc(__a); }
509 __DEVICE__ int __popcll(long long __a) { return __nv_popcll(__a); }
510 __DEVICE__ float __powf(float __a, float __b) {
511  return __nv_fast_powf(__a, __b);
512 }
513 
514 // Parameter must have a known integer value.
515 #define __prof_trigger(__a) asm __volatile__("pmevent \t%0;" ::"i"(__a))
516 __DEVICE__ int __rhadd(int __a, int __b) { return __nv_rhadd(__a, __b); }
517 __DEVICE__ unsigned int __sad(int __a, int __b, unsigned int __c) {
518  return __nv_sad(__a, __b, __c);
519 }
520 __DEVICE__ float __saturatef(float __a) { return __nv_saturatef(__a); }
521 __DEVICE__ int __signbitd(double __a) { return __nv_signbitd(__a); }
522 __DEVICE__ int __signbitf(float __a) { return __nv_signbitf(__a); }
523 __DEVICE__ void __sincosf(float __a, float *__sptr, float *__cptr) {
524  return __nv_fast_sincosf(__a, __sptr, __cptr);
525 }
526 __DEVICE__ float __sinf(float __a) { return __nv_fast_sinf(__a); }
527 __DEVICE__ int __syncthreads_and(int __a) { return __nvvm_bar0_and(__a); }
528 __DEVICE__ int __syncthreads_count(int __a) { return __nvvm_bar0_popc(__a); }
529 __DEVICE__ int __syncthreads_or(int __a) { return __nvvm_bar0_or(__a); }
530 __DEVICE__ float __tanf(float __a) { return __nv_fast_tanf(__a); }
531 __DEVICE__ void __threadfence(void) { __nvvm_membar_gl(); }
532 __DEVICE__ void __threadfence_block(void) { __nvvm_membar_cta(); };
533 __DEVICE__ void __threadfence_system(void) { __nvvm_membar_sys(); };
534 __DEVICE__ void __trap(void) { asm volatile("trap;"); }
535 __DEVICE__ unsigned int __uAtomicAdd(unsigned int *__p, unsigned int __v) {
536  return __nvvm_atom_add_gen_i((int *)__p, __v);
537 }
538 __DEVICE__ unsigned int __uAtomicAdd_block(unsigned int *__p,
539  unsigned int __v) {
540  return __nvvm_atom_cta_add_gen_i((int *)__p, __v);
541 }
542 __DEVICE__ unsigned int __uAtomicAdd_system(unsigned int *__p,
543  unsigned int __v) {
544  return __nvvm_atom_sys_add_gen_i((int *)__p, __v);
545 }
546 __DEVICE__ unsigned int __uAtomicAnd(unsigned int *__p, unsigned int __v) {
547  return __nvvm_atom_and_gen_i((int *)__p, __v);
548 }
549 __DEVICE__ unsigned int __uAtomicAnd_block(unsigned int *__p,
550  unsigned int __v) {
551  return __nvvm_atom_cta_and_gen_i((int *)__p, __v);
552 }
553 __DEVICE__ unsigned int __uAtomicAnd_system(unsigned int *__p,
554  unsigned int __v) {
555  return __nvvm_atom_sys_and_gen_i((int *)__p, __v);
556 }
557 __DEVICE__ unsigned int __uAtomicCAS(unsigned int *__p, unsigned int __cmp,
558  unsigned int __v) {
559  return __nvvm_atom_cas_gen_i((int *)__p, __cmp, __v);
560 }
561 __DEVICE__ unsigned int
562 __uAtomicCAS_block(unsigned int *__p, unsigned int __cmp, unsigned int __v) {
563  return __nvvm_atom_cta_cas_gen_i((int *)__p, __cmp, __v);
564 }
565 __DEVICE__ unsigned int
566 __uAtomicCAS_system(unsigned int *__p, unsigned int __cmp, unsigned int __v) {
567  return __nvvm_atom_sys_cas_gen_i((int *)__p, __cmp, __v);
568 }
569 __DEVICE__ unsigned int __uAtomicDec(unsigned int *__p, unsigned int __v) {
570  return __nvvm_atom_dec_gen_ui(__p, __v);
571 }
572 __DEVICE__ unsigned int __uAtomicDec_block(unsigned int *__p,
573  unsigned int __v) {
574  return __nvvm_atom_cta_dec_gen_ui(__p, __v);
575 }
576 __DEVICE__ unsigned int __uAtomicDec_system(unsigned int *__p,
577  unsigned int __v) {
578  return __nvvm_atom_sys_dec_gen_ui(__p, __v);
579 }
580 __DEVICE__ unsigned int __uAtomicExch(unsigned int *__p, unsigned int __v) {
581  return __nvvm_atom_xchg_gen_i((int *)__p, __v);
582 }
583 __DEVICE__ unsigned int __uAtomicExch_block(unsigned int *__p,
584  unsigned int __v) {
585  return __nvvm_atom_cta_xchg_gen_i((int *)__p, __v);
586 }
587 __DEVICE__ unsigned int __uAtomicExch_system(unsigned int *__p,
588  unsigned int __v) {
589  return __nvvm_atom_sys_xchg_gen_i((int *)__p, __v);
590 }
591 __DEVICE__ unsigned int __uAtomicInc(unsigned int *__p, unsigned int __v) {
592  return __nvvm_atom_inc_gen_ui(__p, __v);
593 }
594 __DEVICE__ unsigned int __uAtomicInc_block(unsigned int *__p,
595  unsigned int __v) {
596  return __nvvm_atom_cta_inc_gen_ui(__p, __v);
597 }
598 __DEVICE__ unsigned int __uAtomicInc_system(unsigned int *__p,
599  unsigned int __v) {
600  return __nvvm_atom_sys_inc_gen_ui(__p, __v);
601 }
602 __DEVICE__ unsigned int __uAtomicMax(unsigned int *__p, unsigned int __v) {
603  return __nvvm_atom_max_gen_ui(__p, __v);
604 }
605 __DEVICE__ unsigned int __uAtomicMax_block(unsigned int *__p,
606  unsigned int __v) {
607  return __nvvm_atom_cta_max_gen_ui(__p, __v);
608 }
609 __DEVICE__ unsigned int __uAtomicMax_system(unsigned int *__p,
610  unsigned int __v) {
611  return __nvvm_atom_sys_max_gen_ui(__p, __v);
612 }
613 __DEVICE__ unsigned int __uAtomicMin(unsigned int *__p, unsigned int __v) {
614  return __nvvm_atom_min_gen_ui(__p, __v);
615 }
616 __DEVICE__ unsigned int __uAtomicMin_block(unsigned int *__p,
617  unsigned int __v) {
618  return __nvvm_atom_cta_min_gen_ui(__p, __v);
619 }
620 __DEVICE__ unsigned int __uAtomicMin_system(unsigned int *__p,
621  unsigned int __v) {
622  return __nvvm_atom_sys_min_gen_ui(__p, __v);
623 }
624 __DEVICE__ unsigned int __uAtomicOr(unsigned int *__p, unsigned int __v) {
625  return __nvvm_atom_or_gen_i((int *)__p, __v);
626 }
627 __DEVICE__ unsigned int __uAtomicOr_block(unsigned int *__p, unsigned int __v) {
628  return __nvvm_atom_cta_or_gen_i((int *)__p, __v);
629 }
630 __DEVICE__ unsigned int __uAtomicOr_system(unsigned int *__p,
631  unsigned int __v) {
632  return __nvvm_atom_sys_or_gen_i((int *)__p, __v);
633 }
634 __DEVICE__ unsigned int __uAtomicXor(unsigned int *__p, unsigned int __v) {
635  return __nvvm_atom_xor_gen_i((int *)__p, __v);
636 }
637 __DEVICE__ unsigned int __uAtomicXor_block(unsigned int *__p,
638  unsigned int __v) {
639  return __nvvm_atom_cta_xor_gen_i((int *)__p, __v);
640 }
641 __DEVICE__ unsigned int __uAtomicXor_system(unsigned int *__p,
642  unsigned int __v) {
643  return __nvvm_atom_sys_xor_gen_i((int *)__p, __v);
644 }
645 __DEVICE__ unsigned int __uhadd(unsigned int __a, unsigned int __b) {
646  return __nv_uhadd(__a, __b);
647 }
648 __DEVICE__ double __uint2double_rn(unsigned int __a) {
649  return __nv_uint2double_rn(__a);
650 }
651 __DEVICE__ float __uint2float_rd(unsigned int __a) {
652  return __nv_uint2float_rd(__a);
653 }
654 __DEVICE__ float __uint2float_rn(unsigned int __a) {
655  return __nv_uint2float_rn(__a);
656 }
657 __DEVICE__ float __uint2float_ru(unsigned int __a) {
658  return __nv_uint2float_ru(__a);
659 }
660 __DEVICE__ float __uint2float_rz(unsigned int __a) {
661  return __nv_uint2float_rz(__a);
662 }
663 __DEVICE__ float __uint_as_float(unsigned int __a) {
664  return __nv_uint_as_float(__a);
665 } //
666 __DEVICE__ double __ull2double_rd(unsigned long long __a) {
667  return __nv_ull2double_rd(__a);
668 }
669 __DEVICE__ double __ull2double_rn(unsigned long long __a) {
670  return __nv_ull2double_rn(__a);
671 }
672 __DEVICE__ double __ull2double_ru(unsigned long long __a) {
673  return __nv_ull2double_ru(__a);
674 }
675 __DEVICE__ double __ull2double_rz(unsigned long long __a) {
676  return __nv_ull2double_rz(__a);
677 }
678 __DEVICE__ float __ull2float_rd(unsigned long long __a) {
679  return __nv_ull2float_rd(__a);
680 }
681 __DEVICE__ float __ull2float_rn(unsigned long long __a) {
682  return __nv_ull2float_rn(__a);
683 }
684 __DEVICE__ float __ull2float_ru(unsigned long long __a) {
685  return __nv_ull2float_ru(__a);
686 }
687 __DEVICE__ float __ull2float_rz(unsigned long long __a) {
688  return __nv_ull2float_rz(__a);
689 }
690 __DEVICE__ unsigned long long __ullAtomicAdd(unsigned long long *__p,
691  unsigned long long __v) {
692  return __nvvm_atom_add_gen_ll((long long *)__p, __v);
693 }
694 __DEVICE__ unsigned long long __ullAtomicAdd_block(unsigned long long *__p,
695  unsigned long long __v) {
696  return __nvvm_atom_cta_add_gen_ll((long long *)__p, __v);
697 }
698 __DEVICE__ unsigned long long __ullAtomicAdd_system(unsigned long long *__p,
699  unsigned long long __v) {
700  return __nvvm_atom_sys_add_gen_ll((long long *)__p, __v);
701 }
702 __DEVICE__ unsigned long long __ullAtomicAnd(unsigned long long *__p,
703  unsigned long long __v) {
704  return __nvvm_atom_and_gen_ll((long long *)__p, __v);
705 }
706 __DEVICE__ unsigned long long __ullAtomicAnd_block(unsigned long long *__p,
707  unsigned long long __v) {
708  return __nvvm_atom_cta_and_gen_ll((long long *)__p, __v);
709 }
710 __DEVICE__ unsigned long long __ullAtomicAnd_system(unsigned long long *__p,
711  unsigned long long __v) {
712  return __nvvm_atom_sys_and_gen_ll((long long *)__p, __v);
713 }
714 __DEVICE__ unsigned long long __ullAtomicCAS(unsigned long long *__p,
715  unsigned long long __cmp,
716  unsigned long long __v) {
717  return __nvvm_atom_cas_gen_ll((long long *)__p, __cmp, __v);
718 }
719 __DEVICE__ unsigned long long __ullAtomicCAS_block(unsigned long long *__p,
720  unsigned long long __cmp,
721  unsigned long long __v) {
722  return __nvvm_atom_cta_cas_gen_ll((long long *)__p, __cmp, __v);
723 }
724 __DEVICE__ unsigned long long __ullAtomicCAS_system(unsigned long long *__p,
725  unsigned long long __cmp,
726  unsigned long long __v) {
727  return __nvvm_atom_sys_cas_gen_ll((long long *)__p, __cmp, __v);
728 }
729 __DEVICE__ unsigned long long __ullAtomicExch(unsigned long long *__p,
730  unsigned long long __v) {
731  return __nvvm_atom_xchg_gen_ll((long long *)__p, __v);
732 }
733 __DEVICE__ unsigned long long __ullAtomicExch_block(unsigned long long *__p,
734  unsigned long long __v) {
735  return __nvvm_atom_cta_xchg_gen_ll((long long *)__p, __v);
736 }
737 __DEVICE__ unsigned long long __ullAtomicExch_system(unsigned long long *__p,
738  unsigned long long __v) {
739  return __nvvm_atom_sys_xchg_gen_ll((long long *)__p, __v);
740 }
741 __DEVICE__ unsigned long long __ullAtomicMax(unsigned long long *__p,
742  unsigned long long __v) {
743  return __nvvm_atom_max_gen_ull(__p, __v);
744 }
745 __DEVICE__ unsigned long long __ullAtomicMax_block(unsigned long long *__p,
746  unsigned long long __v) {
747  return __nvvm_atom_cta_max_gen_ull(__p, __v);
748 }
749 __DEVICE__ unsigned long long __ullAtomicMax_system(unsigned long long *__p,
750  unsigned long long __v) {
751  return __nvvm_atom_sys_max_gen_ull(__p, __v);
752 }
753 __DEVICE__ unsigned long long __ullAtomicMin(unsigned long long *__p,
754  unsigned long long __v) {
755  return __nvvm_atom_min_gen_ull(__p, __v);
756 }
757 __DEVICE__ unsigned long long __ullAtomicMin_block(unsigned long long *__p,
758  unsigned long long __v) {
759  return __nvvm_atom_cta_min_gen_ull(__p, __v);
760 }
761 __DEVICE__ unsigned long long __ullAtomicMin_system(unsigned long long *__p,
762  unsigned long long __v) {
763  return __nvvm_atom_sys_min_gen_ull(__p, __v);
764 }
765 __DEVICE__ unsigned long long __ullAtomicOr(unsigned long long *__p,
766  unsigned long long __v) {
767  return __nvvm_atom_or_gen_ll((long long *)__p, __v);
768 }
769 __DEVICE__ unsigned long long __ullAtomicOr_block(unsigned long long *__p,
770  unsigned long long __v) {
771  return __nvvm_atom_cta_or_gen_ll((long long *)__p, __v);
772 }
773 __DEVICE__ unsigned long long __ullAtomicOr_system(unsigned long long *__p,
774  unsigned long long __v) {
775  return __nvvm_atom_sys_or_gen_ll((long long *)__p, __v);
776 }
777 __DEVICE__ unsigned long long __ullAtomicXor(unsigned long long *__p,
778  unsigned long long __v) {
779  return __nvvm_atom_xor_gen_ll((long long *)__p, __v);
780 }
781 __DEVICE__ unsigned long long __ullAtomicXor_block(unsigned long long *__p,
782  unsigned long long __v) {
783  return __nvvm_atom_cta_xor_gen_ll((long long *)__p, __v);
784 }
785 __DEVICE__ unsigned long long __ullAtomicXor_system(unsigned long long *__p,
786  unsigned long long __v) {
787  return __nvvm_atom_sys_xor_gen_ll((long long *)__p, __v);
788 }
789 __DEVICE__ unsigned int __umul24(unsigned int __a, unsigned int __b) {
790  return __nv_umul24(__a, __b);
791 }
792 __DEVICE__ unsigned long long __umul64hi(unsigned long long __a,
793  unsigned long long __b) {
794  return __nv_umul64hi(__a, __b);
795 }
796 __DEVICE__ unsigned int __umulhi(unsigned int __a, unsigned int __b) {
797  return __nv_umulhi(__a, __b);
798 }
799 __DEVICE__ unsigned int __urhadd(unsigned int __a, unsigned int __b) {
800  return __nv_urhadd(__a, __b);
801 }
802 __DEVICE__ unsigned int __usad(unsigned int __a, unsigned int __b,
803  unsigned int __c) {
804  return __nv_usad(__a, __b, __c);
805 }
806 
807 #if CUDA_VERSION >= 9000 && CUDA_VERSION < 9020
808 __DEVICE__ unsigned int __vabs2(unsigned int __a) { return __nv_vabs2(__a); }
809 __DEVICE__ unsigned int __vabs4(unsigned int __a) { return __nv_vabs4(__a); }
810 __DEVICE__ unsigned int __vabsdiffs2(unsigned int __a, unsigned int __b) {
811  return __nv_vabsdiffs2(__a, __b);
812 }
813 __DEVICE__ unsigned int __vabsdiffs4(unsigned int __a, unsigned int __b) {
814  return __nv_vabsdiffs4(__a, __b);
815 }
816 __DEVICE__ unsigned int __vabsdiffu2(unsigned int __a, unsigned int __b) {
817  return __nv_vabsdiffu2(__a, __b);
818 }
819 __DEVICE__ unsigned int __vabsdiffu4(unsigned int __a, unsigned int __b) {
820  return __nv_vabsdiffu4(__a, __b);
821 }
822 __DEVICE__ unsigned int __vabsss2(unsigned int __a) {
823  return __nv_vabsss2(__a);
824 }
825 __DEVICE__ unsigned int __vabsss4(unsigned int __a) {
826  return __nv_vabsss4(__a);
827 }
828 __DEVICE__ unsigned int __vadd2(unsigned int __a, unsigned int __b) {
829  return __nv_vadd2(__a, __b);
830 }
831 __DEVICE__ unsigned int __vadd4(unsigned int __a, unsigned int __b) {
832  return __nv_vadd4(__a, __b);
833 }
834 __DEVICE__ unsigned int __vaddss2(unsigned int __a, unsigned int __b) {
835  return __nv_vaddss2(__a, __b);
836 }
837 __DEVICE__ unsigned int __vaddss4(unsigned int __a, unsigned int __b) {
838  return __nv_vaddss4(__a, __b);
839 }
840 __DEVICE__ unsigned int __vaddus2(unsigned int __a, unsigned int __b) {
841  return __nv_vaddus2(__a, __b);
842 }
843 __DEVICE__ unsigned int __vaddus4(unsigned int __a, unsigned int __b) {
844  return __nv_vaddus4(__a, __b);
845 }
846 __DEVICE__ unsigned int __vavgs2(unsigned int __a, unsigned int __b) {
847  return __nv_vavgs2(__a, __b);
848 }
849 __DEVICE__ unsigned int __vavgs4(unsigned int __a, unsigned int __b) {
850  return __nv_vavgs4(__a, __b);
851 }
852 __DEVICE__ unsigned int __vavgu2(unsigned int __a, unsigned int __b) {
853  return __nv_vavgu2(__a, __b);
854 }
855 __DEVICE__ unsigned int __vavgu4(unsigned int __a, unsigned int __b) {
856  return __nv_vavgu4(__a, __b);
857 }
858 __DEVICE__ unsigned int __vcmpeq2(unsigned int __a, unsigned int __b) {
859  return __nv_vcmpeq2(__a, __b);
860 }
861 __DEVICE__ unsigned int __vcmpeq4(unsigned int __a, unsigned int __b) {
862  return __nv_vcmpeq4(__a, __b);
863 }
864 __DEVICE__ unsigned int __vcmpges2(unsigned int __a, unsigned int __b) {
865  return __nv_vcmpges2(__a, __b);
866 }
867 __DEVICE__ unsigned int __vcmpges4(unsigned int __a, unsigned int __b) {
868  return __nv_vcmpges4(__a, __b);
869 }
870 __DEVICE__ unsigned int __vcmpgeu2(unsigned int __a, unsigned int __b) {
871  return __nv_vcmpgeu2(__a, __b);
872 }
873 __DEVICE__ unsigned int __vcmpgeu4(unsigned int __a, unsigned int __b) {
874  return __nv_vcmpgeu4(__a, __b);
875 }
876 __DEVICE__ unsigned int __vcmpgts2(unsigned int __a, unsigned int __b) {
877  return __nv_vcmpgts2(__a, __b);
878 }
879 __DEVICE__ unsigned int __vcmpgts4(unsigned int __a, unsigned int __b) {
880  return __nv_vcmpgts4(__a, __b);
881 }
882 __DEVICE__ unsigned int __vcmpgtu2(unsigned int __a, unsigned int __b) {
883  return __nv_vcmpgtu2(__a, __b);
884 }
885 __DEVICE__ unsigned int __vcmpgtu4(unsigned int __a, unsigned int __b) {
886  return __nv_vcmpgtu4(__a, __b);
887 }
888 __DEVICE__ unsigned int __vcmples2(unsigned int __a, unsigned int __b) {
889  return __nv_vcmples2(__a, __b);
890 }
891 __DEVICE__ unsigned int __vcmples4(unsigned int __a, unsigned int __b) {
892  return __nv_vcmples4(__a, __b);
893 }
894 __DEVICE__ unsigned int __vcmpleu2(unsigned int __a, unsigned int __b) {
895  return __nv_vcmpleu2(__a, __b);
896 }
897 __DEVICE__ unsigned int __vcmpleu4(unsigned int __a, unsigned int __b) {
898  return __nv_vcmpleu4(__a, __b);
899 }
900 __DEVICE__ unsigned int __vcmplts2(unsigned int __a, unsigned int __b) {
901  return __nv_vcmplts2(__a, __b);
902 }
903 __DEVICE__ unsigned int __vcmplts4(unsigned int __a, unsigned int __b) {
904  return __nv_vcmplts4(__a, __b);
905 }
906 __DEVICE__ unsigned int __vcmpltu2(unsigned int __a, unsigned int __b) {
907  return __nv_vcmpltu2(__a, __b);
908 }
909 __DEVICE__ unsigned int __vcmpltu4(unsigned int __a, unsigned int __b) {
910  return __nv_vcmpltu4(__a, __b);
911 }
912 __DEVICE__ unsigned int __vcmpne2(unsigned int __a, unsigned int __b) {
913  return __nv_vcmpne2(__a, __b);
914 }
915 __DEVICE__ unsigned int __vcmpne4(unsigned int __a, unsigned int __b) {
916  return __nv_vcmpne4(__a, __b);
917 }
918 __DEVICE__ unsigned int __vhaddu2(unsigned int __a, unsigned int __b) {
919  return __nv_vhaddu2(__a, __b);
920 }
921 __DEVICE__ unsigned int __vhaddu4(unsigned int __a, unsigned int __b) {
922  return __nv_vhaddu4(__a, __b);
923 }
924 __DEVICE__ unsigned int __vmaxs2(unsigned int __a, unsigned int __b) {
925  return __nv_vmaxs2(__a, __b);
926 }
927 __DEVICE__ unsigned int __vmaxs4(unsigned int __a, unsigned int __b) {
928  return __nv_vmaxs4(__a, __b);
929 }
930 __DEVICE__ unsigned int __vmaxu2(unsigned int __a, unsigned int __b) {
931  return __nv_vmaxu2(__a, __b);
932 }
933 __DEVICE__ unsigned int __vmaxu4(unsigned int __a, unsigned int __b) {
934  return __nv_vmaxu4(__a, __b);
935 }
936 __DEVICE__ unsigned int __vmins2(unsigned int __a, unsigned int __b) {
937  return __nv_vmins2(__a, __b);
938 }
939 __DEVICE__ unsigned int __vmins4(unsigned int __a, unsigned int __b) {
940  return __nv_vmins4(__a, __b);
941 }
942 __DEVICE__ unsigned int __vminu2(unsigned int __a, unsigned int __b) {
943  return __nv_vminu2(__a, __b);
944 }
945 __DEVICE__ unsigned int __vminu4(unsigned int __a, unsigned int __b) {
946  return __nv_vminu4(__a, __b);
947 }
948 __DEVICE__ unsigned int __vneg2(unsigned int __a) { return __nv_vneg2(__a); }
949 __DEVICE__ unsigned int __vneg4(unsigned int __a) { return __nv_vneg4(__a); }
950 __DEVICE__ unsigned int __vnegss2(unsigned int __a) {
951  return __nv_vnegss2(__a);
952 }
953 __DEVICE__ unsigned int __vnegss4(unsigned int __a) {
954  return __nv_vnegss4(__a);
955 }
956 __DEVICE__ unsigned int __vsads2(unsigned int __a, unsigned int __b) {
957  return __nv_vsads2(__a, __b);
958 }
959 __DEVICE__ unsigned int __vsads4(unsigned int __a, unsigned int __b) {
960  return __nv_vsads4(__a, __b);
961 }
962 __DEVICE__ unsigned int __vsadu2(unsigned int __a, unsigned int __b) {
963  return __nv_vsadu2(__a, __b);
964 }
965 __DEVICE__ unsigned int __vsadu4(unsigned int __a, unsigned int __b) {
966  return __nv_vsadu4(__a, __b);
967 }
968 __DEVICE__ unsigned int __vseteq2(unsigned int __a, unsigned int __b) {
969  return __nv_vseteq2(__a, __b);
970 }
971 __DEVICE__ unsigned int __vseteq4(unsigned int __a, unsigned int __b) {
972  return __nv_vseteq4(__a, __b);
973 }
974 __DEVICE__ unsigned int __vsetges2(unsigned int __a, unsigned int __b) {
975  return __nv_vsetges2(__a, __b);
976 }
977 __DEVICE__ unsigned int __vsetges4(unsigned int __a, unsigned int __b) {
978  return __nv_vsetges4(__a, __b);
979 }
980 __DEVICE__ unsigned int __vsetgeu2(unsigned int __a, unsigned int __b) {
981  return __nv_vsetgeu2(__a, __b);
982 }
983 __DEVICE__ unsigned int __vsetgeu4(unsigned int __a, unsigned int __b) {
984  return __nv_vsetgeu4(__a, __b);
985 }
986 __DEVICE__ unsigned int __vsetgts2(unsigned int __a, unsigned int __b) {
987  return __nv_vsetgts2(__a, __b);
988 }
989 __DEVICE__ unsigned int __vsetgts4(unsigned int __a, unsigned int __b) {
990  return __nv_vsetgts4(__a, __b);
991 }
992 __DEVICE__ unsigned int __vsetgtu2(unsigned int __a, unsigned int __b) {
993  return __nv_vsetgtu2(__a, __b);
994 }
995 __DEVICE__ unsigned int __vsetgtu4(unsigned int __a, unsigned int __b) {
996  return __nv_vsetgtu4(__a, __b);
997 }
998 __DEVICE__ unsigned int __vsetles2(unsigned int __a, unsigned int __b) {
999  return __nv_vsetles2(__a, __b);
1000 }
1001 __DEVICE__ unsigned int __vsetles4(unsigned int __a, unsigned int __b) {
1002  return __nv_vsetles4(__a, __b);
1003 }
1004 __DEVICE__ unsigned int __vsetleu2(unsigned int __a, unsigned int __b) {
1005  return __nv_vsetleu2(__a, __b);
1006 }
1007 __DEVICE__ unsigned int __vsetleu4(unsigned int __a, unsigned int __b) {
1008  return __nv_vsetleu4(__a, __b);
1009 }
1010 __DEVICE__ unsigned int __vsetlts2(unsigned int __a, unsigned int __b) {
1011  return __nv_vsetlts2(__a, __b);
1012 }
1013 __DEVICE__ unsigned int __vsetlts4(unsigned int __a, unsigned int __b) {
1014  return __nv_vsetlts4(__a, __b);
1015 }
1016 __DEVICE__ unsigned int __vsetltu2(unsigned int __a, unsigned int __b) {
1017  return __nv_vsetltu2(__a, __b);
1018 }
1019 __DEVICE__ unsigned int __vsetltu4(unsigned int __a, unsigned int __b) {
1020  return __nv_vsetltu4(__a, __b);
1021 }
1022 __DEVICE__ unsigned int __vsetne2(unsigned int __a, unsigned int __b) {
1023  return __nv_vsetne2(__a, __b);
1024 }
1025 __DEVICE__ unsigned int __vsetne4(unsigned int __a, unsigned int __b) {
1026  return __nv_vsetne4(__a, __b);
1027 }
1028 __DEVICE__ unsigned int __vsub2(unsigned int __a, unsigned int __b) {
1029  return __nv_vsub2(__a, __b);
1030 }
1031 __DEVICE__ unsigned int __vsub4(unsigned int __a, unsigned int __b) {
1032  return __nv_vsub4(__a, __b);
1033 }
1034 __DEVICE__ unsigned int __vsubss2(unsigned int __a, unsigned int __b) {
1035  return __nv_vsubss2(__a, __b);
1036 }
1037 __DEVICE__ unsigned int __vsubss4(unsigned int __a, unsigned int __b) {
1038  return __nv_vsubss4(__a, __b);
1039 }
1040 __DEVICE__ unsigned int __vsubus2(unsigned int __a, unsigned int __b) {
1041  return __nv_vsubus2(__a, __b);
1042 }
1043 __DEVICE__ unsigned int __vsubus4(unsigned int __a, unsigned int __b) {
1044  return __nv_vsubus4(__a, __b);
1045 }
1046 #else // CUDA_VERSION >= 9020
1047 // CUDA no longer provides inline assembly (or bitcode) implementation of these
1048 // functions, so we have to reimplment them. The implementation is naive and is
1049 // not optimized for performance.
1050 
1051 // Helper function to convert N-bit boolean subfields into all-0 or all-1.
1052 // E.g. __bool2mask(0x01000100,8) -> 0xff00ff00
1053 // __bool2mask(0x00010000,16) -> 0xffff0000
1054 __DEVICE__ unsigned int __bool2mask(unsigned int __a, int shift) {
1055  return (__a << shift) - __a;
1056 }
1057 __DEVICE__ unsigned int __vabs2(unsigned int __a) {
1058  unsigned int r;
1059  asm("vabsdiff2.s32.s32.s32 %0,%1,%2,%3;"
1060  : "=r"(r)
1061  : "r"(__a), "r"(0), "r"(0));
1062  return r;
1063 }
1064 __DEVICE__ unsigned int __vabs4(unsigned int __a) {
1065  unsigned int r;
1066  asm("vabsdiff4.s32.s32.s32 %0,%1,%2,%3;"
1067  : "=r"(r)
1068  : "r"(__a), "r"(0), "r"(0));
1069  return r;
1070 }
1071 __DEVICE__ unsigned int __vabsdiffs2(unsigned int __a, unsigned int __b) {
1072  unsigned int r;
1073  asm("vabsdiff2.s32.s32.s32 %0,%1,%2,%3;"
1074  : "=r"(r)
1075  : "r"(__a), "r"(__b), "r"(0));
1076  return r;
1077 }
1078 
1079 __DEVICE__ unsigned int __vabsdiffs4(unsigned int __a, unsigned int __b) {
1080  unsigned int r;
1081  asm("vabsdiff4.s32.s32.s32 %0,%1,%2,%3;"
1082  : "=r"(r)
1083  : "r"(__a), "r"(__b), "r"(0));
1084  return r;
1085 }
1086 __DEVICE__ unsigned int __vabsdiffu2(unsigned int __a, unsigned int __b) {
1087  unsigned int r;
1088  asm("vabsdiff2.u32.u32.u32 %0,%1,%2,%3;"
1089  : "=r"(r)
1090  : "r"(__a), "r"(__b), "r"(0));
1091  return r;
1092 }
1093 __DEVICE__ unsigned int __vabsdiffu4(unsigned int __a, unsigned int __b) {
1094  unsigned int r;
1095  asm("vabsdiff4.u32.u32.u32 %0,%1,%2,%3;"
1096  : "=r"(r)
1097  : "r"(__a), "r"(__b), "r"(0));
1098  return r;
1099 }
1100 __DEVICE__ unsigned int __vabsss2(unsigned int __a) {
1101  unsigned int r;
1102  asm("vabsdiff2.s32.s32.s32.sat %0,%1,%2,%3;"
1103  : "=r"(r)
1104  : "r"(__a), "r"(0), "r"(0));
1105  return r;
1106 }
1107 __DEVICE__ unsigned int __vabsss4(unsigned int __a) {
1108  unsigned int r;
1109  asm("vabsdiff4.s32.s32.s32.sat %0,%1,%2,%3;"
1110  : "=r"(r)
1111  : "r"(__a), "r"(0), "r"(0));
1112  return r;
1113 }
1114 __DEVICE__ unsigned int __vadd2(unsigned int __a, unsigned int __b) {
1115  unsigned int r;
1116  asm("vadd2.u32.u32.u32 %0,%1,%2,%3;" : "=r"(r) : "r"(__a), "r"(__b), "r"(0));
1117  return r;
1118 }
1119 __DEVICE__ unsigned int __vadd4(unsigned int __a, unsigned int __b) {
1120  unsigned int r;
1121  asm("vadd4.u32.u32.u32 %0,%1,%2,%3;" : "=r"(r) : "r"(__a), "r"(__b), "r"(0));
1122  return r;
1123 }
1124 __DEVICE__ unsigned int __vaddss2(unsigned int __a, unsigned int __b) {
1125  unsigned int r;
1126  asm("vadd2.s32.s32.s32.sat %0,%1,%2,%3;"
1127  : "=r"(r)
1128  : "r"(__a), "r"(__b), "r"(0));
1129  return r;
1130 }
1131 __DEVICE__ unsigned int __vaddss4(unsigned int __a, unsigned int __b) {
1132  unsigned int r;
1133  asm("vadd4.s32.s32.s32.sat %0,%1,%2,%3;"
1134  : "=r"(r)
1135  : "r"(__a), "r"(__b), "r"(0));
1136  return r;
1137 }
1138 __DEVICE__ unsigned int __vaddus2(unsigned int __a, unsigned int __b) {
1139  unsigned int r;
1140  asm("vadd2.u32.u32.u32.sat %0,%1,%2,%3;"
1141  : "=r"(r)
1142  : "r"(__a), "r"(__b), "r"(0));
1143  return r;
1144 }
1145 __DEVICE__ unsigned int __vaddus4(unsigned int __a, unsigned int __b) {
1146  unsigned int r;
1147  asm("vadd4.u32.u32.u32.sat %0,%1,%2,%3;"
1148  : "=r"(r)
1149  : "r"(__a), "r"(__b), "r"(0));
1150  return r;
1151 }
1152 __DEVICE__ unsigned int __vavgs2(unsigned int __a, unsigned int __b) {
1153  unsigned int r;
1154  asm("vavrg2.s32.s32.s32 %0,%1,%2,%3;" : "=r"(r) : "r"(__a), "r"(__b), "r"(0));
1155  return r;
1156 }
1157 __DEVICE__ unsigned int __vavgs4(unsigned int __a, unsigned int __b) {
1158  unsigned int r;
1159  asm("vavrg4.s32.s32.s32 %0,%1,%2,%3;" : "=r"(r) : "r"(__a), "r"(__b), "r"(0));
1160  return r;
1161 }
1162 __DEVICE__ unsigned int __vavgu2(unsigned int __a, unsigned int __b) {
1163  unsigned int r;
1164  asm("vavrg2.u32.u32.u32 %0,%1,%2,%3;" : "=r"(r) : "r"(__a), "r"(__b), "r"(0));
1165  return r;
1166 }
1167 __DEVICE__ unsigned int __vavgu4(unsigned int __a, unsigned int __b) {
1168  unsigned int r;
1169  asm("vavrg4.u32.u32.u32 %0,%1,%2,%3;" : "=r"(r) : "r"(__a), "r"(__b), "r"(0));
1170  return r;
1171 }
1172 __DEVICE__ unsigned int __vseteq2(unsigned int __a, unsigned int __b) {
1173  unsigned int r;
1174  asm("vset2.u32.u32.eq %0,%1,%2,%3;" : "=r"(r) : "r"(__a), "r"(__b), "r"(0));
1175  return r;
1176 }
1177 __DEVICE__ unsigned int __vcmpeq2(unsigned int __a, unsigned int __b) {
1178  return __bool2mask(__vseteq2(__a, __b), 16);
1179 }
1180 __DEVICE__ unsigned int __vseteq4(unsigned int __a, unsigned int __b) {
1181  unsigned int r;
1182  asm("vset4.u32.u32.eq %0,%1,%2,%3;" : "=r"(r) : "r"(__a), "r"(__b), "r"(0));
1183  return r;
1184 }
1185 __DEVICE__ unsigned int __vcmpeq4(unsigned int __a, unsigned int __b) {
1186  return __bool2mask(__vseteq4(__a, __b), 8);
1187 }
1188 __DEVICE__ unsigned int __vsetges2(unsigned int __a, unsigned int __b) {
1189  unsigned int r;
1190  asm("vset2.s32.s32.ge %0,%1,%2,%3;" : "=r"(r) : "r"(__a), "r"(__b), "r"(0));
1191  return r;
1192 }
1193 __DEVICE__ unsigned int __vcmpges2(unsigned int __a, unsigned int __b) {
1194  return __bool2mask(__vsetges2(__a, __b), 16);
1195 }
1196 __DEVICE__ unsigned int __vsetges4(unsigned int __a, unsigned int __b) {
1197  unsigned int r;
1198  asm("vset4.s32.s32.ge %0,%1,%2,%3;" : "=r"(r) : "r"(__a), "r"(__b), "r"(0));
1199  return r;
1200 }
1201 __DEVICE__ unsigned int __vcmpges4(unsigned int __a, unsigned int __b) {
1202  return __bool2mask(__vsetges4(__a, __b), 8);
1203 }
1204 __DEVICE__ unsigned int __vsetgeu2(unsigned int __a, unsigned int __b) {
1205  unsigned int r;
1206  asm("vset2.u32.u32.ge %0,%1,%2,%3;" : "=r"(r) : "r"(__a), "r"(__b), "r"(0));
1207  return r;
1208 }
1209 __DEVICE__ unsigned int __vcmpgeu2(unsigned int __a, unsigned int __b) {
1210  return __bool2mask(__vsetgeu2(__a, __b), 16);
1211 }
1212 __DEVICE__ unsigned int __vsetgeu4(unsigned int __a, unsigned int __b) {
1213  unsigned int r;
1214  asm("vset4.u32.u32.ge %0,%1,%2,%3;" : "=r"(r) : "r"(__a), "r"(__b), "r"(0));
1215  return r;
1216 }
1217 __DEVICE__ unsigned int __vcmpgeu4(unsigned int __a, unsigned int __b) {
1218  return __bool2mask(__vsetgeu4(__a, __b), 8);
1219 }
1220 __DEVICE__ unsigned int __vsetgts2(unsigned int __a, unsigned int __b) {
1221  unsigned int r;
1222  asm("vset2.s32.s32.gt %0,%1,%2,%3;" : "=r"(r) : "r"(__a), "r"(__b), "r"(0));
1223  return r;
1224 }
1225 __DEVICE__ unsigned int __vcmpgts2(unsigned int __a, unsigned int __b) {
1226  return __bool2mask(__vsetgts2(__a, __b), 16);
1227 }
1228 __DEVICE__ unsigned int __vsetgts4(unsigned int __a, unsigned int __b) {
1229  unsigned int r;
1230  asm("vset4.s32.s32.gt %0,%1,%2,%3;" : "=r"(r) : "r"(__a), "r"(__b), "r"(0));
1231  return r;
1232 }
1233 __DEVICE__ unsigned int __vcmpgts4(unsigned int __a, unsigned int __b) {
1234  return __bool2mask(__vsetgts4(__a, __b), 8);
1235 }
1236 __DEVICE__ unsigned int __vsetgtu2(unsigned int __a, unsigned int __b) {
1237  unsigned int r;
1238  asm("vset2.u32.u32.gt %0,%1,%2,%3;" : "=r"(r) : "r"(__a), "r"(__b), "r"(0));
1239  return r;
1240 }
1241 __DEVICE__ unsigned int __vcmpgtu2(unsigned int __a, unsigned int __b) {
1242  return __bool2mask(__vsetgtu2(__a, __b), 16);
1243 }
1244 __DEVICE__ unsigned int __vsetgtu4(unsigned int __a, unsigned int __b) {
1245  unsigned int r;
1246  asm("vset4.u32.u32.gt %0,%1,%2,%3;" : "=r"(r) : "r"(__a), "r"(__b), "r"(0));
1247  return r;
1248 }
1249 __DEVICE__ unsigned int __vcmpgtu4(unsigned int __a, unsigned int __b) {
1250  return __bool2mask(__vsetgtu4(__a, __b), 8);
1251 }
1252 __DEVICE__ unsigned int __vsetles2(unsigned int __a, unsigned int __b) {
1253  unsigned int r;
1254  asm("vset2.s32.s32.le %0,%1,%2,%3;" : "=r"(r) : "r"(__a), "r"(__b), "r"(0));
1255  return r;
1256 }
1257 __DEVICE__ unsigned int __vcmples2(unsigned int __a, unsigned int __b) {
1258  return __bool2mask(__vsetles2(__a, __b), 16);
1259 }
1260 __DEVICE__ unsigned int __vsetles4(unsigned int __a, unsigned int __b) {
1261  unsigned int r;
1262  asm("vset4.s32.s32.le %0,%1,%2,%3;" : "=r"(r) : "r"(__a), "r"(__b), "r"(0));
1263  return r;
1264 }
1265 __DEVICE__ unsigned int __vcmples4(unsigned int __a, unsigned int __b) {
1266  return __bool2mask(__vsetles4(__a, __b), 8);
1267 }
1268 __DEVICE__ unsigned int __vsetleu2(unsigned int __a, unsigned int __b) {
1269  unsigned int r;
1270  asm("vset2.u32.u32.le %0,%1,%2,%3;" : "=r"(r) : "r"(__a), "r"(__b), "r"(0));
1271  return r;
1272 }
1273 __DEVICE__ unsigned int __vcmpleu2(unsigned int __a, unsigned int __b) {
1274  return __bool2mask(__vsetleu2(__a, __b), 16);
1275 }
1276 __DEVICE__ unsigned int __vsetleu4(unsigned int __a, unsigned int __b) {
1277  unsigned int r;
1278  asm("vset4.u32.u32.le %0,%1,%2,%3;" : "=r"(r) : "r"(__a), "r"(__b), "r"(0));
1279  return r;
1280 }
1281 __DEVICE__ unsigned int __vcmpleu4(unsigned int __a, unsigned int __b) {
1282  return __bool2mask(__vsetleu4(__a, __b), 8);
1283 }
1284 __DEVICE__ unsigned int __vsetlts2(unsigned int __a, unsigned int __b) {
1285  unsigned int r;
1286  asm("vset2.s32.s32.lt %0,%1,%2,%3;" : "=r"(r) : "r"(__a), "r"(__b), "r"(0));
1287  return r;
1288 }
1289 __DEVICE__ unsigned int __vcmplts2(unsigned int __a, unsigned int __b) {
1290  return __bool2mask(__vsetlts2(__a, __b), 16);
1291 }
1292 __DEVICE__ unsigned int __vsetlts4(unsigned int __a, unsigned int __b) {
1293  unsigned int r;
1294  asm("vset4.s32.s32.lt %0,%1,%2,%3;" : "=r"(r) : "r"(__a), "r"(__b), "r"(0));
1295  return r;
1296 }
1297 __DEVICE__ unsigned int __vcmplts4(unsigned int __a, unsigned int __b) {
1298  return __bool2mask(__vsetlts4(__a, __b), 8);
1299 }
1300 __DEVICE__ unsigned int __vsetltu2(unsigned int __a, unsigned int __b) {
1301  unsigned int r;
1302  asm("vset2.u32.u32.lt %0,%1,%2,%3;" : "=r"(r) : "r"(__a), "r"(__b), "r"(0));
1303  return r;
1304 }
1305 __DEVICE__ unsigned int __vcmpltu2(unsigned int __a, unsigned int __b) {
1306  return __bool2mask(__vsetltu2(__a, __b), 16);
1307 }
1308 __DEVICE__ unsigned int __vsetltu4(unsigned int __a, unsigned int __b) {
1309  unsigned int r;
1310  asm("vset4.u32.u32.lt %0,%1,%2,%3;" : "=r"(r) : "r"(__a), "r"(__b), "r"(0));
1311  return r;
1312 }
1313 __DEVICE__ unsigned int __vcmpltu4(unsigned int __a, unsigned int __b) {
1314  return __bool2mask(__vsetltu4(__a, __b), 8);
1315 }
1316 __DEVICE__ unsigned int __vsetne2(unsigned int __a, unsigned int __b) {
1317  unsigned int r;
1318  asm("vset2.u32.u32.ne %0,%1,%2,%3;" : "=r"(r) : "r"(__a), "r"(__b), "r"(0));
1319  return r;
1320 }
1321 __DEVICE__ unsigned int __vcmpne2(unsigned int __a, unsigned int __b) {
1322  return __bool2mask(__vsetne2(__a, __b), 16);
1323 }
1324 __DEVICE__ unsigned int __vsetne4(unsigned int __a, unsigned int __b) {
1325  unsigned int r;
1326  asm("vset4.u32.u32.ne %0,%1,%2,%3;" : "=r"(r) : "r"(__a), "r"(__b), "r"(0));
1327  return r;
1328 }
1329 __DEVICE__ unsigned int __vcmpne4(unsigned int __a, unsigned int __b) {
1330  return __bool2mask(__vsetne4(__a, __b), 8);
1331 }
1332 
1333 // Based on ITEM 23 in AIM-239: http://dspace.mit.edu/handle/1721.1/6086
1334 // (a & b) + (a | b) = a + b = (a ^ b) + 2 * (a & b) =>
1335 // (a + b) / 2 = ((a ^ b) >> 1) + (a & b)
1336 // To operate on multiple sub-elements we need to make sure to mask out bits
1337 // that crossed over into adjacent elements during the shift.
1338 __DEVICE__ unsigned int __vhaddu2(unsigned int __a, unsigned int __b) {
1339  return (((__a ^ __b) >> 1) & ~0x80008000u) + (__a & __b);
1340 }
1341 __DEVICE__ unsigned int __vhaddu4(unsigned int __a, unsigned int __b) {
1342  return (((__a ^ __b) >> 1) & ~0x80808080u) + (__a & __b);
1343 }
1344 
1345 __DEVICE__ unsigned int __vmaxs2(unsigned int __a, unsigned int __b) {
1346  unsigned int r;
1347  if ((__a & 0x8000) && (__b & 0x8000)) {
1348  // Work around a bug in ptxas which produces invalid result if low element
1349  // is negative.
1350  unsigned mask = __vcmpgts2(__a, __b);
1351  r = (__a & mask) | (__b & ~mask);
1352  } else {
1353  asm("vmax2.s32.s32.s32 %0,%1,%2,%3;"
1354  : "=r"(r)
1355  : "r"(__a), "r"(__b), "r"(0));
1356  }
1357  return r;
1358 }
1359 __DEVICE__ unsigned int __vmaxs4(unsigned int __a, unsigned int __b) {
1360  unsigned int r;
1361  asm("vmax4.s32.s32.s32 %0,%1,%2,%3;" : "=r"(r) : "r"(__a), "r"(__b), "r"(0));
1362  return r;
1363 }
1364 __DEVICE__ unsigned int __vmaxu2(unsigned int __a, unsigned int __b) {
1365  unsigned int r;
1366  asm("vmax2.u32.u32.u32 %0,%1,%2,%3;" : "=r"(r) : "r"(__a), "r"(__b), "r"(0));
1367  return r;
1368 }
1369 __DEVICE__ unsigned int __vmaxu4(unsigned int __a, unsigned int __b) {
1370  unsigned int r;
1371  asm("vmax4.u32.u32.u32 %0,%1,%2,%3;" : "=r"(r) : "r"(__a), "r"(__b), "r"(0));
1372  return r;
1373 }
1374 __DEVICE__ unsigned int __vmins2(unsigned int __a, unsigned int __b) {
1375  unsigned int r;
1376  asm("vmin2.s32.s32.s32 %0,%1,%2,%3;" : "=r"(r) : "r"(__a), "r"(__b), "r"(0));
1377  return r;
1378 }
1379 __DEVICE__ unsigned int __vmins4(unsigned int __a, unsigned int __b) {
1380  unsigned int r;
1381  asm("vmin4.s32.s32.s32 %0,%1,%2,%3;" : "=r"(r) : "r"(__a), "r"(__b), "r"(0));
1382  return r;
1383 }
1384 __DEVICE__ unsigned int __vminu2(unsigned int __a, unsigned int __b) {
1385  unsigned int r;
1386  asm("vmin2.u32.u32.u32 %0,%1,%2,%3;" : "=r"(r) : "r"(__a), "r"(__b), "r"(0));
1387  return r;
1388 }
1389 __DEVICE__ unsigned int __vminu4(unsigned int __a, unsigned int __b) {
1390  unsigned int r;
1391  asm("vmin4.u32.u32.u32 %0,%1,%2,%3;" : "=r"(r) : "r"(__a), "r"(__b), "r"(0));
1392  return r;
1393 }
1394 __DEVICE__ unsigned int __vsads2(unsigned int __a, unsigned int __b) {
1395  unsigned int r;
1396  asm("vabsdiff2.s32.s32.s32.add %0,%1,%2,%3;"
1397  : "=r"(r)
1398  : "r"(__a), "r"(__b), "r"(0));
1399  return r;
1400 }
1401 __DEVICE__ unsigned int __vsads4(unsigned int __a, unsigned int __b) {
1402  unsigned int r;
1403  asm("vabsdiff4.s32.s32.s32.add %0,%1,%2,%3;"
1404  : "=r"(r)
1405  : "r"(__a), "r"(__b), "r"(0));
1406  return r;
1407 }
1408 __DEVICE__ unsigned int __vsadu2(unsigned int __a, unsigned int __b) {
1409  unsigned int r;
1410  asm("vabsdiff2.u32.u32.u32.add %0,%1,%2,%3;"
1411  : "=r"(r)
1412  : "r"(__a), "r"(__b), "r"(0));
1413  return r;
1414 }
1415 __DEVICE__ unsigned int __vsadu4(unsigned int __a, unsigned int __b) {
1416  unsigned int r;
1417  asm("vabsdiff4.u32.u32.u32.add %0,%1,%2,%3;"
1418  : "=r"(r)
1419  : "r"(__a), "r"(__b), "r"(0));
1420  return r;
1421 }
1422 
1423 __DEVICE__ unsigned int __vsub2(unsigned int __a, unsigned int __b) {
1424  unsigned int r;
1425  asm("vsub2.u32.u32.u32 %0,%1,%2,%3;" : "=r"(r) : "r"(__a), "r"(__b), "r"(0));
1426  return r;
1427 }
1428 __DEVICE__ unsigned int __vneg2(unsigned int __a) { return __vsub2(0, __a); }
1429 
1430 __DEVICE__ unsigned int __vsub4(unsigned int __a, unsigned int __b) {
1431  unsigned int r;
1432  asm("vsub4.u32.u32.u32 %0,%1,%2,%3;" : "=r"(r) : "r"(__a), "r"(__b), "r"(0));
1433  return r;
1434 }
1435 __DEVICE__ unsigned int __vneg4(unsigned int __a) { return __vsub4(0, __a); }
1436 __DEVICE__ unsigned int __vsubss2(unsigned int __a, unsigned int __b) {
1437  unsigned int r;
1438  asm("vsub2.s32.s32.s32.sat %0,%1,%2,%3;"
1439  : "=r"(r)
1440  : "r"(__a), "r"(__b), "r"(0));
1441  return r;
1442 }
1443 __DEVICE__ unsigned int __vnegss2(unsigned int __a) {
1444  return __vsubss2(0, __a);
1445 }
1446 __DEVICE__ unsigned int __vsubss4(unsigned int __a, unsigned int __b) {
1447  unsigned int r;
1448  asm("vsub4.s32.s32.s32.sat %0,%1,%2,%3;"
1449  : "=r"(r)
1450  : "r"(__a), "r"(__b), "r"(0));
1451  return r;
1452 }
1453 __DEVICE__ unsigned int __vnegss4(unsigned int __a) {
1454  return __vsubss4(0, __a);
1455 }
1456 __DEVICE__ unsigned int __vsubus2(unsigned int __a, unsigned int __b) {
1457  unsigned int r;
1458  asm("vsub2.u32.u32.u32.sat %0,%1,%2,%3;"
1459  : "=r"(r)
1460  : "r"(__a), "r"(__b), "r"(0));
1461  return r;
1462 }
1463 __DEVICE__ unsigned int __vsubus4(unsigned int __a, unsigned int __b) {
1464  unsigned int r;
1465  asm("vsub4.u32.u32.u32.sat %0,%1,%2,%3;"
1466  : "=r"(r)
1467  : "r"(__a), "r"(__b), "r"(0));
1468  return r;
1469 }
1470 #endif // CUDA_VERSION >= 9020
1471 __DEVICE__ int abs(int __a) { return __nv_abs(__a); }
1472 __DEVICE__ double acos(double __a) { return __nv_acos(__a); }
1473 __DEVICE__ float acosf(float __a) { return __nv_acosf(__a); }
1474 __DEVICE__ double acosh(double __a) { return __nv_acosh(__a); }
1475 __DEVICE__ float acoshf(float __a) { return __nv_acoshf(__a); }
1476 __DEVICE__ double asin(double __a) { return __nv_asin(__a); }
1477 __DEVICE__ float asinf(float __a) { return __nv_asinf(__a); }
1478 __DEVICE__ double asinh(double __a) { return __nv_asinh(__a); }
1479 __DEVICE__ float asinhf(float __a) { return __nv_asinhf(__a); }
1480 __DEVICE__ double atan(double __a) { return __nv_atan(__a); }
1481 __DEVICE__ double atan2(double __a, double __b) { return __nv_atan2(__a, __b); }
1482 __DEVICE__ float atan2f(float __a, float __b) { return __nv_atan2f(__a, __b); }
1483 __DEVICE__ float atanf(float __a) { return __nv_atanf(__a); }
1484 __DEVICE__ double atanh(double __a) { return __nv_atanh(__a); }
1485 __DEVICE__ float atanhf(float __a) { return __nv_atanhf(__a); }
1486 __DEVICE__ double cbrt(double __a) { return __nv_cbrt(__a); }
1487 __DEVICE__ float cbrtf(float __a) { return __nv_cbrtf(__a); }
1488 __DEVICE__ double ceil(double __a) { return __nv_ceil(__a); }
1489 __DEVICE__ float ceilf(float __a) { return __nv_ceilf(__a); }
1490 __DEVICE__ int clock() { return __nvvm_read_ptx_sreg_clock(); }
1491 __DEVICE__ long long clock64() { return __nvvm_read_ptx_sreg_clock64(); }
1492 __DEVICE__ double copysign(double __a, double __b) {
1493  return __nv_copysign(__a, __b);
1494 }
1495 __DEVICE__ float copysignf(float __a, float __b) {
1496  return __nv_copysignf(__a, __b);
1497 }
1498 __DEVICE__ double cos(double __a) { return __nv_cos(__a); }
1499 __DEVICE__ float cosf(float __a) {
1501 }
1502 __DEVICE__ double cosh(double __a) { return __nv_cosh(__a); }
1503 __DEVICE__ float coshf(float __a) { return __nv_coshf(__a); }
1504 __DEVICE__ double cospi(double __a) { return __nv_cospi(__a); }
1505 __DEVICE__ float cospif(float __a) { return __nv_cospif(__a); }
1506 __DEVICE__ double cyl_bessel_i0(double __a) { return __nv_cyl_bessel_i0(__a); }
1507 __DEVICE__ float cyl_bessel_i0f(float __a) { return __nv_cyl_bessel_i0f(__a); }
1508 __DEVICE__ double cyl_bessel_i1(double __a) { return __nv_cyl_bessel_i1(__a); }
1509 __DEVICE__ float cyl_bessel_i1f(float __a) { return __nv_cyl_bessel_i1f(__a); }
1510 __DEVICE__ double erf(double __a) { return __nv_erf(__a); }
1511 __DEVICE__ double erfc(double __a) { return __nv_erfc(__a); }
1512 __DEVICE__ float erfcf(float __a) { return __nv_erfcf(__a); }
1513 __DEVICE__ double erfcinv(double __a) { return __nv_erfcinv(__a); }
1514 __DEVICE__ float erfcinvf(float __a) { return __nv_erfcinvf(__a); }
1515 __DEVICE__ double erfcx(double __a) { return __nv_erfcx(__a); }
1516 __DEVICE__ float erfcxf(float __a) { return __nv_erfcxf(__a); }
1517 __DEVICE__ float erff(float __a) { return __nv_erff(__a); }
1518 __DEVICE__ double erfinv(double __a) { return __nv_erfinv(__a); }
1519 __DEVICE__ float erfinvf(float __a) { return __nv_erfinvf(__a); }
1520 __DEVICE__ double exp(double __a) { return __nv_exp(__a); }
1521 __DEVICE__ double exp10(double __a) { return __nv_exp10(__a); }
1522 __DEVICE__ float exp10f(float __a) { return __nv_exp10f(__a); }
1523 __DEVICE__ double exp2(double __a) { return __nv_exp2(__a); }
1524 __DEVICE__ float exp2f(float __a) { return __nv_exp2f(__a); }
1525 __DEVICE__ float expf(float __a) { return __nv_expf(__a); }
1526 __DEVICE__ double expm1(double __a) { return __nv_expm1(__a); }
1527 __DEVICE__ float expm1f(float __a) { return __nv_expm1f(__a); }
1528 __DEVICE__ double fabs(double __a) { return __nv_fabs(__a); }
1529 __DEVICE__ float fabsf(float __a) { return __nv_fabsf(__a); }
1530 __DEVICE__ double fdim(double __a, double __b) { return __nv_fdim(__a, __b); }
1531 __DEVICE__ float fdimf(float __a, float __b) { return __nv_fdimf(__a, __b); }
1532 __DEVICE__ double fdivide(double __a, double __b) { return __a / __b; }
1533 __DEVICE__ float fdividef(float __a, float __b) {
1534 #if __FAST_MATH__ && !__CUDA_PREC_DIV
1535  return __nv_fast_fdividef(__a, __b);
1536 #else
1537  return __a / __b;
1538 #endif
1539 }
1540 __DEVICE__ double floor(double __f) { return __nv_floor(__f); }
1541 __DEVICE__ float floorf(float __f) { return __nv_floorf(__f); }
1542 __DEVICE__ double fma(double __a, double __b, double __c) {
1543  return __nv_fma(__a, __b, __c);
1544 }
1545 __DEVICE__ float fmaf(float __a, float __b, float __c) {
1546  return __nv_fmaf(__a, __b, __c);
1547 }
1548 __DEVICE__ double fmax(double __a, double __b) { return __nv_fmax(__a, __b); }
1549 __DEVICE__ float fmaxf(float __a, float __b) { return __nv_fmaxf(__a, __b); }
1550 __DEVICE__ double fmin(double __a, double __b) { return __nv_fmin(__a, __b); }
1551 __DEVICE__ float fminf(float __a, float __b) { return __nv_fminf(__a, __b); }
1552 __DEVICE__ double fmod(double __a, double __b) { return __nv_fmod(__a, __b); }
1553 __DEVICE__ float fmodf(float __a, float __b) { return __nv_fmodf(__a, __b); }
1554 __DEVICE__ double frexp(double __a, int *__b) { return __nv_frexp(__a, __b); }
1555 __DEVICE__ float frexpf(float __a, int *__b) { return __nv_frexpf(__a, __b); }
1556 __DEVICE__ double hypot(double __a, double __b) { return __nv_hypot(__a, __b); }
1557 __DEVICE__ float hypotf(float __a, float __b) { return __nv_hypotf(__a, __b); }
1558 __DEVICE__ int ilogb(double __a) { return __nv_ilogb(__a); }
1559 __DEVICE__ int ilogbf(float __a) { return __nv_ilogbf(__a); }
1560 __DEVICE__ double j0(double __a) { return __nv_j0(__a); }
1561 __DEVICE__ float j0f(float __a) { return __nv_j0f(__a); }
1562 __DEVICE__ double j1(double __a) { return __nv_j1(__a); }
1563 __DEVICE__ float j1f(float __a) { return __nv_j1f(__a); }
1564 __DEVICE__ double jn(int __n, double __a) { return __nv_jn(__n, __a); }
1565 __DEVICE__ float jnf(int __n, float __a) { return __nv_jnf(__n, __a); }
1566 #if defined(__LP64__)
1567 __DEVICE__ long labs(long __a) { return llabs(__a); };
1568 #else
1569 __DEVICE__ long labs(long __a) { return __nv_abs(__a); };
1570 #endif
1571 __DEVICE__ double ldexp(double __a, int __b) { return __nv_ldexp(__a, __b); }
1572 __DEVICE__ float ldexpf(float __a, int __b) { return __nv_ldexpf(__a, __b); }
1573 __DEVICE__ double lgamma(double __a) { return __nv_lgamma(__a); }
1574 __DEVICE__ float lgammaf(float __a) { return __nv_lgammaf(__a); }
1575 __DEVICE__ long long llabs(long long __a) { return __nv_llabs(__a); }
1576 __DEVICE__ long long llmax(long long __a, long long __b) {
1577  return __nv_llmax(__a, __b);
1578 }
1579 __DEVICE__ long long llmin(long long __a, long long __b) {
1580  return __nv_llmin(__a, __b);
1581 }
1582 __DEVICE__ long long llrint(double __a) { return __nv_llrint(__a); }
1583 __DEVICE__ long long llrintf(float __a) { return __nv_llrintf(__a); }
1584 __DEVICE__ long long llround(double __a) { return __nv_llround(__a); }
1585 __DEVICE__ long long llroundf(float __a) { return __nv_llroundf(__a); }
1586 __DEVICE__ double log(double __a) { return __nv_log(__a); }
1587 __DEVICE__ double log10(double __a) { return __nv_log10(__a); }
1588 __DEVICE__ float log10f(float __a) { return __nv_log10f(__a); }
1589 __DEVICE__ double log1p(double __a) { return __nv_log1p(__a); }
1590 __DEVICE__ float log1pf(float __a) { return __nv_log1pf(__a); }
1591 __DEVICE__ double log2(double __a) { return __nv_log2(__a); }
1592 __DEVICE__ float log2f(float __a) {
1594 }
1595 __DEVICE__ double logb(double __a) { return __nv_logb(__a); }
1596 __DEVICE__ float logbf(float __a) { return __nv_logbf(__a); }
1597 __DEVICE__ float logf(float __a) {
1599 }
1600 #if defined(__LP64__)
1601 __DEVICE__ long lrint(double __a) { return llrint(__a); }
1602 __DEVICE__ long lrintf(float __a) { return __float2ll_rn(__a); }
1603 __DEVICE__ long lround(double __a) { return llround(__a); }
1604 __DEVICE__ long lroundf(float __a) { return llroundf(__a); }
1605 #else
1606 __DEVICE__ long lrint(double __a) { return (long)rint(__a); }
1607 __DEVICE__ long lrintf(float __a) { return __float2int_rn(__a); }
1608 __DEVICE__ long lround(double __a) { return round(__a); }
1609 __DEVICE__ long lroundf(float __a) { return roundf(__a); }
1610 #endif
1611 __DEVICE__ int max(int __a, int __b) { return __nv_max(__a, __b); }
1612 __DEVICE__ void *memcpy(void *__a, const void *__b, size_t __c) {
1613  return __builtin_memcpy(__a, __b, __c);
1614 }
1615 __DEVICE__ void *memset(void *__a, int __b, size_t __c) {
1616  return __builtin_memset(__a, __b, __c);
1617 }
1618 __DEVICE__ int min(int __a, int __b) { return __nv_min(__a, __b); }
1619 __DEVICE__ double modf(double __a, double *__b) { return __nv_modf(__a, __b); }
1620 __DEVICE__ float modff(float __a, float *__b) { return __nv_modff(__a, __b); }
1621 __DEVICE__ double nearbyint(double __a) { return __nv_nearbyint(__a); }
1622 __DEVICE__ float nearbyintf(float __a) { return __nv_nearbyintf(__a); }
1623 __DEVICE__ double nextafter(double __a, double __b) {
1624  return __nv_nextafter(__a, __b);
1625 }
1626 __DEVICE__ float nextafterf(float __a, float __b) {
1627  return __nv_nextafterf(__a, __b);
1628 }
1629 __DEVICE__ double norm(int __dim, const double *__t) {
1630  return __nv_norm(__dim, __t);
1631 }
1632 __DEVICE__ double norm3d(double __a, double __b, double __c) {
1633  return __nv_norm3d(__a, __b, __c);
1634 }
1635 __DEVICE__ float norm3df(float __a, float __b, float __c) {
1636  return __nv_norm3df(__a, __b, __c);
1637 }
1638 __DEVICE__ double norm4d(double __a, double __b, double __c, double __d) {
1639  return __nv_norm4d(__a, __b, __c, __d);
1640 }
1641 __DEVICE__ float norm4df(float __a, float __b, float __c, float __d) {
1642  return __nv_norm4df(__a, __b, __c, __d);
1643 }
1644 __DEVICE__ double normcdf(double __a) { return __nv_normcdf(__a); }
1645 __DEVICE__ float normcdff(float __a) { return __nv_normcdff(__a); }
1646 __DEVICE__ double normcdfinv(double __a) { return __nv_normcdfinv(__a); }
1647 __DEVICE__ float normcdfinvf(float __a) { return __nv_normcdfinvf(__a); }
1648 __DEVICE__ float normf(int __dim, const float *__t) {
1649  return __nv_normf(__dim, __t);
1650 }
1651 __DEVICE__ double pow(double __a, double __b) { return __nv_pow(__a, __b); }
1652 __DEVICE__ float powf(float __a, float __b) { return __nv_powf(__a, __b); }
1653 __DEVICE__ double powi(double __a, int __b) { return __nv_powi(__a, __b); }
1654 __DEVICE__ float powif(float __a, int __b) { return __nv_powif(__a, __b); }
1655 __DEVICE__ double rcbrt(double __a) { return __nv_rcbrt(__a); }
1656 __DEVICE__ float rcbrtf(float __a) { return __nv_rcbrtf(__a); }
1657 __DEVICE__ double remainder(double __a, double __b) {
1658  return __nv_remainder(__a, __b);
1659 }
1660 __DEVICE__ float remainderf(float __a, float __b) {
1661  return __nv_remainderf(__a, __b);
1662 }
1663 __DEVICE__ double remquo(double __a, double __b, int *__c) {
1664  return __nv_remquo(__a, __b, __c);
1665 }
1666 __DEVICE__ float remquof(float __a, float __b, int *__c) {
1667  return __nv_remquof(__a, __b, __c);
1668 }
1669 __DEVICE__ double rhypot(double __a, double __b) {
1670  return __nv_rhypot(__a, __b);
1671 }
1672 __DEVICE__ float rhypotf(float __a, float __b) {
1673  return __nv_rhypotf(__a, __b);
1674 }
1675 __DEVICE__ double rint(double __a) { return __nv_rint(__a); }
1676 __DEVICE__ float rintf(float __a) { return __nv_rintf(__a); }
1677 __DEVICE__ double rnorm(int __a, const double *__b) {
1678  return __nv_rnorm(__a, __b);
1679 }
1680 __DEVICE__ double rnorm3d(double __a, double __b, double __c) {
1681  return __nv_rnorm3d(__a, __b, __c);
1682 }
1683 __DEVICE__ float rnorm3df(float __a, float __b, float __c) {
1684  return __nv_rnorm3df(__a, __b, __c);
1685 }
1686 __DEVICE__ double rnorm4d(double __a, double __b, double __c, double __d) {
1687  return __nv_rnorm4d(__a, __b, __c, __d);
1688 }
1689 __DEVICE__ float rnorm4df(float __a, float __b, float __c, float __d) {
1690  return __nv_rnorm4df(__a, __b, __c, __d);
1691 }
1692 __DEVICE__ float rnormf(int __dim, const float *__t) {
1693  return __nv_rnormf(__dim, __t);
1694 }
1695 __DEVICE__ double round(double __a) { return __nv_round(__a); }
1696 __DEVICE__ float roundf(float __a) { return __nv_roundf(__a); }
1697 __DEVICE__ double rsqrt(double __a) { return __nv_rsqrt(__a); }
1698 __DEVICE__ float rsqrtf(float __a) { return __nv_rsqrtf(__a); }
1699 __DEVICE__ double scalbn(double __a, int __b) { return __nv_scalbn(__a, __b); }
1700 __DEVICE__ float scalbnf(float __a, int __b) { return __nv_scalbnf(__a, __b); }
1701 __DEVICE__ double scalbln(double __a, long __b) {
1702  if (__b > INT_MAX)
1703  return __a > 0 ? HUGE_VAL : -HUGE_VAL;
1704  if (__b < INT_MIN)
1705  return __a > 0 ? 0.0 : -0.0;
1706  return scalbn(__a, (int)__b);
1707 }
1708 __DEVICE__ float scalblnf(float __a, long __b) {
1709  if (__b > INT_MAX)
1710  return __a > 0 ? HUGE_VALF : -HUGE_VALF;
1711  if (__b < INT_MIN)
1712  return __a > 0 ? 0.f : -0.f;
1713  return scalbnf(__a, (int)__b);
1714 }
1715 __DEVICE__ double sin(double __a) { return __nv_sin(__a); }
1716 __DEVICE__ void sincos(double __a, double *__sptr, double *__cptr) {
1717  return __nv_sincos(__a, __sptr, __cptr);
1718 }
1719 __DEVICE__ void sincosf(float __a, float *__sptr, float *__cptr) {
1720  return __FAST_OR_SLOW(__nv_fast_sincosf, __nv_sincosf)(__a, __sptr, __cptr);
1721 }
1722 __DEVICE__ void sincospi(double __a, double *__sptr, double *__cptr) {
1723  return __nv_sincospi(__a, __sptr, __cptr);
1724 }
1725 __DEVICE__ void sincospif(float __a, float *__sptr, float *__cptr) {
1726  return __nv_sincospif(__a, __sptr, __cptr);
1727 }
1728 __DEVICE__ float sinf(float __a) {
1730 }
1731 __DEVICE__ double sinh(double __a) { return __nv_sinh(__a); }
1732 __DEVICE__ float sinhf(float __a) { return __nv_sinhf(__a); }
1733 __DEVICE__ double sinpi(double __a) { return __nv_sinpi(__a); }
1734 __DEVICE__ float sinpif(float __a) { return __nv_sinpif(__a); }
1735 __DEVICE__ double sqrt(double __a) { return __nv_sqrt(__a); }
1736 __DEVICE__ float sqrtf(float __a) { return __nv_sqrtf(__a); }
1737 __DEVICE__ double tan(double __a) { return __nv_tan(__a); }
1738 __DEVICE__ float tanf(float __a) { return __nv_tanf(__a); }
1739 __DEVICE__ double tanh(double __a) { return __nv_tanh(__a); }
1740 __DEVICE__ float tanhf(float __a) { return __nv_tanhf(__a); }
1741 __DEVICE__ double tgamma(double __a) { return __nv_tgamma(__a); }
1742 __DEVICE__ float tgammaf(float __a) { return __nv_tgammaf(__a); }
1743 __DEVICE__ double trunc(double __a) { return __nv_trunc(__a); }
1744 __DEVICE__ float truncf(float __a) { return __nv_truncf(__a); }
1745 __DEVICE__ unsigned long long ullmax(unsigned long long __a,
1746  unsigned long long __b) {
1747  return __nv_ullmax(__a, __b);
1748 }
1749 __DEVICE__ unsigned long long ullmin(unsigned long long __a,
1750  unsigned long long __b) {
1751  return __nv_ullmin(__a, __b);
1752 }
1753 __DEVICE__ unsigned int umax(unsigned int __a, unsigned int __b) {
1754  return __nv_umax(__a, __b);
1755 }
1756 __DEVICE__ unsigned int umin(unsigned int __a, unsigned int __b) {
1757  return __nv_umin(__a, __b);
1758 }
1759 __DEVICE__ double y0(double __a) { return __nv_y0(__a); }
1760 __DEVICE__ float y0f(float __a) { return __nv_y0f(__a); }
1761 __DEVICE__ double y1(double __a) { return __nv_y1(__a); }
1762 __DEVICE__ float y1f(float __a) { return __nv_y1f(__a); }
1763 __DEVICE__ double yn(int __a, double __b) { return __nv_yn(__a, __b); }
1764 __DEVICE__ float ynf(int __a, float __b) { return __nv_ynf(__a, __b); }
1765 
1766 #pragma pop_macro("__DEVICE__")
1767 #pragma pop_macro("__FAST_OR_SLOW")
1768 #endif // __CLANG_CUDA_DEVICE_FUNCTIONS_H__
__device__ int __nv_ilogb(double __a)
__device__ double __nv_acos(double __a)
__DEVICE__ unsigned int __vhaddu2(unsigned int __a, unsigned int __b)
__device__ unsigned int __nv_umul24(unsigned int __a, unsigned int __b)
__DEVICE__ unsigned long long ullmax(unsigned long long __a, unsigned long long __b)
__device__ float __nv_fmaf_rd(float __a, float __b, float __c)
__DEVICE__ float __ull2float_ru(unsigned long long __a)
__DEVICE__ unsigned int __vavgs4(unsigned int __a, unsigned int __b)
__DEVICE__ void __threadfence_system(void)
__DEVICE__ unsigned int __vnegss4(unsigned int __a)
__device__ double __nv_expm1(double __a)
__DEVICE__ float tanf(float __a)
__device__ double __nv_rnorm4d(double __a, double __b, double __c, double __d)
__DEVICE__ int __isfinited(double __a)
__DEVICE__ int __iAtomicMin_system(int *__p, int __v)
__DEVICE__ unsigned int __vcmpne2(unsigned int __a, unsigned int __b)
__DEVICE__ unsigned int __uAtomicCAS_system(unsigned int *__p, unsigned int __cmp, unsigned int __v)
__DEVICE__ unsigned int __pm2(void)
__DEVICE__ unsigned int __vcmpne4(unsigned int __a, unsigned int __b)
__DEVICE__ float fmodf(float __a, float __b)
__device__ float __nv_floorf(float __a)
__DEVICE__ double __ull2double_rd(unsigned long long __a)
__DEVICE__ float rnorm4df(float __a, float __b, float __c, float __d)
__DEVICE__ unsigned int __vcmpleu4(unsigned int __a, unsigned int __b)
__device__ unsigned int __nv_urhadd(unsigned int __a, unsigned int __b)
__DEVICE__ unsigned int __vsetlts4(unsigned int __a, unsigned int __b)
__DEVICE__ float __fmaf_ru(float __a, float __b, float __c)
__DEVICE__ float __fmul_ru(float __a, float __b)
__device__ double __nv_dsub_rn(double __a, double __b)
__device__ float __nv_roundf(float __a)
__DEVICE__ unsigned long long __ullAtomicXor(unsigned long long *__p, unsigned long long __v)
__device__ void __nv_sincospif(float __a, float *__b, float *__c)
__device__ float __nv_fast_powf(float __a, float __b)
__DEVICE__ float __frcp_rn(float __a)
__device__ double __nv_fmod(double __a, double __b)
__DEVICE__ unsigned int __vneg4(unsigned int __a)
__device__ unsigned int __nv_double2uint_rz(double __a)
__device__ float __nv_fast_fdividef(float __a, float __b)
__DEVICE__ unsigned int __uAtomicMin_block(unsigned int *__p, unsigned int __v)
__DEVICE__ int ilogb(double __a)
__device__ int __nv_signbitf(float __a)
__device__ float __nv_int2float_ru(int __a)
__device__ float __nv_j1f(float __a)
__DEVICE__ double tan(double __a)
__device__ float __nv_double2float_rz(double __a)
__device__ float __nv_powif(float __a, int __b)
__device__ long long __nv_double2ll_rd(double __a)
__DEVICE__ float __uint2float_rz(unsigned int __a)
__DEVICE__ int __double2hiint(double __a)
__DEVICE__ float __ll2float_rn(long long __a)
__DEVICE__ unsigned long long __ullAtomicExch_system(unsigned long long *__p, unsigned long long __v)
__DEVICE__ double __ull2double_rn(unsigned long long __a)
__device__ double __nv_log(double __a)
__device__ double __nv_floor(double __a)
__DEVICE__ float acoshf(float __a)
__DEVICE__ double cosh(double __a)
__DEVICE__ float coshf(float __a)
__DEVICE__ long labs(long __a)
__device__ double __nv_atan2(double __a, double __b)
__device__ double __nv_hiloint2double(int __a, int __b)
__device__ double __nv_nextafter(double __a, double __b)
__DEVICE__ int __float2int_ru(float __a)
__DEVICE__ unsigned int __umulhi(unsigned int __a, unsigned int __b)
__DEVICE__ unsigned int __vsetne2(unsigned int __a, unsigned int __b)
__DEVICE__ int __iAtomicAdd(int *__p, int __v)
__device__ long long __nv_llround(double __a)
__DEVICE__ float __fsub_rz(float __a, float __b)
__DEVICE__ float __frcp_rd(float __a)
__device__ int __nv_finitef(float __a)
__DEVICE__ double __fma_rz(double __a, double __b, double __c)
__DEVICE__ int __iAtomicAnd_block(int *__p, int __v)
__DEVICE__ int __iAtomicXor(int *__p, int __v)
__DEVICE__ double norm3d(double __a, double __b, double __c)
__device__ double __nv_cosh(double __a)
__device__ double __nv_drcp_rz(double __a)
__DEVICE__ double pow(double __a, double __b)
__device__ float __nv_frsqrt_rn(float __a)
__device__ double __nv_fdim(double __a, double __b)
__device__ double __nv_ull2double_rd(unsigned long long __a)
__DEVICE__ unsigned int __vsadu4(unsigned int __a, unsigned int __b)
__DEVICE__ double exp2(double __a)
__DEVICE__ unsigned int __vsetges2(unsigned int __a, unsigned int __b)
__device__ float __nv_ynf(int __a, float __b)
__device__ double __nv_remainder(double __a, double __b)
__DEVICE__ int __syncthreads_and(int __a)
__DEVICE__ float nextafterf(float __a, float __b)
__DEVICE__ unsigned int __vneg2(unsigned int __a)
__device__ float __nv_fsqrt_rz(float __a)
__device__ float __nv_log10f(float __a)
__device__ double __nv_log1p(double __a)
__device__ int __nv_isfinited(double __a)
__DEVICE__ long long __llAtomicXor(long long *__p, long long __v)
__device__ float __nv_ull2float_rd(unsigned long long __a)
#define HUGE_VAL
A positive double constant expression.
Definition: opencl-c.h:171
__device__ void __nv_sincos(double __a, double *__b, double *__c)
__DEVICE__ unsigned int __float2uint_rn(float __a)
__DEVICE__ long long __llAtomicOr(long long *__p, long long __v)
__DEVICE__ int __clz(int __a)
__device__ double __nv_dadd_rd(double __a, double __b)
__DEVICE__ double lgamma(double __a)
__device__ double __nv_dmul_rz(double __a, double __b)
__device__ int __nv_float2int_rn(float __a)
__DEVICE__ unsigned int __vcmpeq2(unsigned int __a, unsigned int __b)
__device__ double __nv_fma_rn(double __a, double __b, double __c)
__DEVICE__ int __iAtomicOr_block(int *__p, int __v)
__DEVICE__ int abs(int __a)
__device__ float __nv_tanhf(float __a)
__DEVICE__ float cyl_bessel_i0f(float __a)
__DEVICE__ int __iAtomicXor_block(int *__p, int __v)
__DEVICE__ double log(double __a)
__device__ float __nv_fmodf(float __a, float __b)
__DEVICE__ unsigned int __uAtomicInc(unsigned int *__p, unsigned int __v)
__DEVICE__ unsigned int __uAtomicOr_system(unsigned int *__p, unsigned int __v)
__DEVICE__ float normcdff(float __a)
__device__ double __nv_fmin(double __a, double __b)
__device__ float __nv_fadd_rd(float __a, float __b)
__DEVICE__ int ilogbf(float __a)
__device__ float __nv_frcp_ru(float __a)
__DEVICE__ float frexpf(float __a, int *__b)
__DEVICE__ double __ll2double_rz(long long __a)
__DEVICE__ unsigned int __uAtomicInc_block(unsigned int *__p, unsigned int __v)
__DEVICE__ unsigned int __vsetlts2(unsigned int __a, unsigned int __b)
__device__ long long __nv_double2ll_rn(double __a)
__DEVICE__ double __dsqrt_ru(double __a)
__DEVICE__ float ceilf(float __a)
__DEVICE__ unsigned int __vmins4(unsigned int __a, unsigned int __b)
__DEVICE__ unsigned int __vsadu2(unsigned int __a, unsigned int __b)
__device__ float __nv_cyl_bessel_i1f(float __a)
__device__ float __nv_int_as_float(int __a)
__device__ double __nv_drcp_rn(double __a)
__DEVICE__ unsigned int __uAtomicXor(unsigned int *__p, unsigned int __v)
__device__ long long __nv_float2ll_rz(float __a)
__device__ double __nv_drcp_ru(double __a)
__device__ float __nv_remainderf(float __a, float __b)
__DEVICE__ double __ll2double_rn(long long __a)
__device__ int __nv_clzll(long long __a)
__DEVICE__ float __fAtomicAdd(float *__p, float __v)
__device__ unsigned long long __nv_ullmax(unsigned long long __a, unsigned long long __b)
__device__ float __nv_expm1f(float __a)
__DEVICE__ double erf(double __a)
__device__ double __nv_sin(double __a)
__DEVICE__ unsigned long long __double2ull_rd(double __a)
__DEVICE__ void sincospif(float __a, float *__sptr, float *__cptr)
__DEVICE__ double __dAtomicAdd_block(double *__p, double __v)
__DEVICE__ int __ffsll(long long __a)
__device__ float __nv_fast_log2f(float __a)
__device__ long long __nv_llabs(long long __a)
__device__ float __nv_j0f(float __a)
__DEVICE__ double fmin(double __a, double __b)
__DEVICE__ double y0(double __a)
__DEVICE__ double cyl_bessel_i0(double __a)
__device__ double __nv_frexp(double __a, int *__b)
__device__ unsigned long long __nv_double2ull_rz(double __a)
__DEVICE__ unsigned long long __double2ull_rn(double __a)
__DEVICE__ double rhypot(double __a, double __b)
__DEVICE__ float __fmaf_ieee_rn(float __a, float __b, float __c)
__DEVICE__ unsigned long long __ullAtomicMax_system(unsigned long long *__p, unsigned long long __v)
__DEVICE__ float __ull2float_rz(unsigned long long __a)
__DEVICE__ unsigned int __uAtomicCAS_block(unsigned int *__p, unsigned int __cmp, unsigned int __v)
__device__ float __nv_int2float_rn(int __a)
__DEVICE__ double cospi(double __a)
__device__ float __nv_fdimf(float __a, float __b)
__DEVICE__ double __fma_rn(double __a, double __b, double __c)
#define __FAST_OR_SLOW(fast, slow)
__device__ float __nv_fast_expf(float __a)
__device__ float __nv_scalbnf(float __a, int __b)
__DEVICE__ unsigned int __vsubus4(unsigned int __a, unsigned int __b)
__device__ float __nv_remquof(float __a, float __b, int *__c)
__device__ int __nv_ffsll(long long __a)
__DEVICE__ float __fAtomicExch_system(float *__p, float __v)
__DEVICE__ float __expf(float __a)
__DEVICE__ double log1p(double __a)
__device__ double __nv_jn(int __a, double __b)
__DEVICE__ float __fmaf_ieee_ru(float __a, float __b, float __c)
__device__ double __nv_dsqrt_rn(double __a)
__device__ void __nv_sincospi(double __a, double *__b, double *__c)
__device__ float __nv_normcdfinvf(float __a)
__DEVICE__ float __ull2float_rd(unsigned long long __a)
__DEVICE__ float exp10f(float __a)
__device__ double __nv_norm4d(double __a, double __b, double __c, double __d)
__DEVICE__ double sinpi(double __a)
__device__ float __nv_y0f(float __a)
__DEVICE__ int __isinff(float __a)
__device__ unsigned int __nv_double2uint_rn(double __a)
__DEVICE__ unsigned int __pm3(void)
__DEVICE__ float __fsub_ru(float __a, float __b)
__DEVICE__ double rnorm(int __a, const double *__b)
__DEVICE__ unsigned int __vsetltu4(unsigned int __a, unsigned int __b)
__DEVICE__ float powif(float __a, int __b)
__device__ double __nv_dmul_rn(double __a, double __b)
__device__ float __nv_rintf(float __a)
__DEVICE__ float log2f(float __a)
__DEVICE__ float __fadd_rd(float __a, float __b)
__DEVICE__ unsigned int __uAtomicXor_system(unsigned int *__p, unsigned int __v)
__device__ float __nv_cosf(float __a)
__device__ int __nv_popc(int __a)
__DEVICE__ double log2(double __a)
__DEVICE__ double normcdfinv(double __a)
__DEVICE__ double j0(double __a)
__DEVICE__ double __dmul_rd(double __a, double __b)
__DEVICE__ float remquof(float __a, float __b, int *__c)
__device__ unsigned long long __nv_float2ull_ru(float __a)
__DEVICE__ unsigned int __vabs4(unsigned int __a)
__device__ float __nv_erfcinvf(float __a)
__DEVICE__ double __dsub_ru(double __a, double __b)
__DEVICE__ float __fdividef(float __a, float __b)
__DEVICE__ float y0f(float __a)
__DEVICE__ void __threadfence(void)
__DEVICE__ float __saturatef(float __a)
__device__ int __nv_brev(int __a)
__DEVICE__ float __int2float_rz(int __a)
__device__ float __nv_fdiv_rn(float __a, float __b)
__DEVICE__ float rsqrtf(float __a)
__DEVICE__ float __logf(float __a)
__DEVICE__ double copysign(double __a, double __b)
__device__ int __nv_hadd(int __a, int __b)
__DEVICE__ double cyl_bessel_i1(double __a)
__device__ float __nv_asinhf(float __a)
__DEVICE__ unsigned int __vseteq2(unsigned int __a, unsigned int __b)
__DEVICE__ float expm1f(float __a)
__DEVICE__ double fma(double __a, double __b, double __c)
__device__ float __nv_ll2float_rn(long long __a)
__DEVICE__ double sqrt(double __a)
__DEVICE__ float __fmaf_rd(float __a, float __b, float __c)
__device__ float __nv_frexpf(float __a, int *__b)
__device__ float __nv_fmul_rd(float __a, float __b)
__DEVICE__ float rhypotf(float __a, float __b)
__device__ int __nv_sad(int __a, int __b, int __c)
__device__ double __nv_atan(double __a)
__device__ long long __nv_mul64hi(long long __a, long long __b)
__device__ double __nv_exp10(double __a)
__DEVICE__ double erfinv(double __a)
__DEVICE__ unsigned int __bool2mask(unsigned int __a, int shift)
__DEVICE__ long long __llAtomicXor_block(long long *__p, long long __v)
__DEVICE__ int __iAtomicAnd(int *__p, int __v)
__DEVICE__ int __mulhi(int __a, int __b)
__device__ double __nv_y0(double __a)
__DEVICE__ int __popc(int __a)
__DEVICE__ float asinhf(float __a)
__DEVICE__ float __ll2float_ru(long long __a)
__DEVICE__ float __uint2float_ru(unsigned int __a)
__device__ int __nv_min(int __a, int __b)
__DEVICE__ double exp(double __a)
__device__ long long __nv_llmin(long long __a, long long __b)
__device__ double __nv_lgamma(double __a)
__DEVICE__ int __float2int_rd(float __a)
__DEVICE__ unsigned int __vcmples2(unsigned int __a, unsigned int __b)
__DEVICE__ unsigned int __float2uint_rz(float __a)
__device__ unsigned int __nv_float2uint_rn(float __a)
__device__ float __nv_rnormf(int __a, const float *__b)
__device__ float __nv_fsqrt_rd(float __a)
__DEVICE__ unsigned int __vcmpeq4(unsigned int __a, unsigned int __b)
__DEVICE__ float __int2float_ru(int __a)
__DEVICE__ unsigned int __vminu4(unsigned int __a, unsigned int __b)
__DEVICE__ unsigned int __vavgu4(unsigned int __a, unsigned int __b)
__DEVICE__ float __fmaf_rz(float __a, float __b, float __c)
__device__ unsigned int __nv_usad(unsigned int __a, unsigned int __b, unsigned int __c)
__DEVICE__ float erfinvf(float __a)
__device__ double __nv_ddiv_ru(double __a, double __b)
__DEVICE__ double j1(double __a)
__DEVICE__ long long __llAtomicAnd_block(long long *__p, long long __v)
__device__ double __nv_dsub_rz(double __a, double __b)
__DEVICE__ int __hadd(int __a, int __b)
__device__ unsigned long long __nv_float2ull_rd(float __a)
__DEVICE__ int __iAtomicMin_block(int *__p, int __v)
__device__ unsigned int __nv_float2uint_rd(float __a)
__DEVICE__ float __fdiv_rd(float __a, float __b)
__DEVICE__ unsigned int __vsetgtu2(unsigned int __a, unsigned int __b)
__device__ double __nv_ll2double_rn(long long __a)
__device__ int __nv_ffs(int __a)
__DEVICE__ double jn(int __n, double __a)
__device__ float __nv_int2float_rz(int __a)
__DEVICE__ unsigned int __vadd4(unsigned int __a, unsigned int __b)
__device__ float __nv_uint2float_rn(unsigned int __a)
__DEVICE__ float __fsqrt_rd(float __a)
__device__ float __nv_nextafterf(float __a, float __b)
__DEVICE__ double atan(double __a)
__DEVICE__ unsigned int __vsub2(unsigned int __a, unsigned int __b)
__DEVICE__ float __fmul_rd(float __a, float __b)
__DEVICE__ int __iAtomicAdd_system(int *__p, int __v)
__device__ unsigned long long __nv_umul64hi(unsigned long long __a, unsigned long long __b)
__DEVICE__ double rsqrt(double __a)
__device__ double __nv_cos(double __a)
__DEVICE__ double tanh(double __a)
__device__ int __nv_isinfd(double __a)
__device__ double __nv_rhypot(double __a, double __b)
__DEVICE__ unsigned int __vsetgtu4(unsigned int __a, unsigned int __b)
__DEVICE__ double powi(double __a, int __b)
__DEVICE__ unsigned int __vcmpgeu4(unsigned int __a, unsigned int __b)
__device__ float __nv_fsub_ru(float __a, float __b)
__DEVICE__ unsigned long long __ullAtomicMin_system(unsigned long long *__p, unsigned long long __v)
static __inline__ void int __a
Definition: emmintrin.h:4099
__device__ long long __nv_double2ll_rz(double __a)
__DEVICE__ unsigned int __vcmpgeu2(unsigned int __a, unsigned int __b)
__device__ int __nv_signbitd(double __a)
__device__ float __nv_copysignf(float __a, float __b)
__DEVICE__ float cospif(float __a)
__DEVICE__ double __dsub_rd(double __a, double __b)
__DEVICE__ float __int2float_rn(int __a)
__DEVICE__ long long llmin(long long __a, long long __b)
__DEVICE__ float fminf(float __a, float __b)
__DEVICE__ int __clzll(long long __a)
__DEVICE__ int __float2int_rn(float __a)
__DEVICE__ unsigned int __vabsss4(unsigned int __a)
__device__ float __nv_expf(float __a)
__device__ unsigned long long __nv_double2ull_ru(double __a)
__DEVICE__ float __fsqrt_rz(float __a)
__DEVICE__ unsigned int __uAtomicAnd(unsigned int *__p, unsigned int __v)
__device__ float __nv_coshf(float __a)
__device__ float __nv_powf(float __a, float __b)
__DEVICE__ double __dadd_rd(double __a, double __b)
__DEVICE__ unsigned int __uAtomicExch_block(unsigned int *__p, unsigned int __v)
__DEVICE__ double __ddiv_rn(double __a, double __b)
__device__ float __nv_atanf(float __a)
__DEVICE__ double __dsqrt_rd(double __a)
__DEVICE__ float __tanf(float __a)
__device__ double __nv_fma_ru(double __a, double __b, double __c)
__DEVICE__ float sinhf(float __a)
__DEVICE__ long long __llAtomicAnd_system(long long *__p, long long __v)
__DEVICE__ double __ull2double_rz(unsigned long long __a)
__DEVICE__ void * memset(void *__a, int __b, size_t __c)
__DEVICE__ unsigned int __vmaxu4(unsigned int __a, unsigned int __b)
__device__ float __nv_cospif(float __a)
__device__ float __nv_fast_sinf(float __a)
__DEVICE__ float __fAtomicAdd_block(float *__p, float __v)
__device__ double __nv_fabs(double __a)
__device__ float __nv_ll2float_rz(long long __a)
__DEVICE__ unsigned long long __float2ull_rn(float __a)
__DEVICE__ unsigned long long __ullAtomicAdd_block(unsigned long long *__p, unsigned long long __v)
__DEVICE__ unsigned int __uAtomicMax(unsigned int *__p, unsigned int __v)
__DEVICE__ unsigned long long __ullAtomicXor_block(unsigned long long *__p, unsigned long long __v)
__DEVICE__ double erfc(double __a)
__device__ float __nv_normf(int __a, const float *__b)
__DEVICE__ float erfcinvf(float __a)
__device__ float __nv_exp2f(float __a)
__DEVICE__ float acosf(float __a)
__DEVICE__ int __iAtomicXor_system(int *__p, int __v)
__device__ float __nv_rcbrtf(float __a)
__DEVICE__ unsigned long long ullmin(unsigned long long __a, unsigned long long __b)
__DEVICE__ double __ddiv_rd(double __a, double __b)
__DEVICE__ float fmaxf(float __a, float __b)
__DEVICE__ unsigned int __vcmpges4(unsigned int __a, unsigned int __b)
__device__ float __nv_ll2float_ru(long long __a)
__DEVICE__ float fdividef(float __a, float __b)
__device__ unsigned int __nv_double2uint_ru(double __a)
__DEVICE__ unsigned int __vmaxs2(unsigned int __a, unsigned int __b)
__DEVICE__ unsigned int __vcmplts4(unsigned int __a, unsigned int __b)
__DEVICE__ long long __double2ll_rn(double __a)
__DEVICE__ float nearbyintf(float __a)
__DEVICE__ double round(double __a)
__device__ int __nv_popcll(long long __a)
__DEVICE__ int __iAtomicMax(int *__p, int __v)
__device__ double __nv_ll2double_rz(long long __a)
__DEVICE__ int __iAtomicCAS_block(int *__p, int __cmp, int __v)
static __inline unsigned char unsigned int unsigned int unsigned int * __p
Definition: adxintrin.h:38
__DEVICE__ int __double2int_rd(double __a)
__device__ double __nv_tan(double __a)
__DEVICE__ int clock()
__device__ float __nv_truncf(float __a)
__DEVICE__ float atanf(float __a)
__device__ double __nv_sinh(double __a)
__DEVICE__ float __fAtomicAdd_system(float *__p, float __v)
__DEVICE__ unsigned int __vcmpgtu2(unsigned int __a, unsigned int __b)
__DEVICE__ unsigned int __float2uint_ru(float __a)
__device__ double __nv_round(double __a)
__DEVICE__ double nearbyint(double __a)
__DEVICE__ float __fAtomicExch_block(float *__p, float __v)
__DEVICE__ double atanh(double __a)
__device__ double __nv_rsqrt(double __a)
__device__ float __nv_norm3df(float __a, float __b, float __c)
__DEVICE__ float __fsqrt_rn(float __a)
__DEVICE__ float logf(float __a)
__device__ double __nv_ldexp(double __a, int __b)
__DEVICE__ unsigned int __vavgu2(unsigned int __a, unsigned int __b)
__DEVICE__ float __fdiv_rz(float __a, float __b)
__DEVICE__ float fmaf(float __a, float __b, float __c)
__device__ long long __nv_llmax(long long __a, long long __b)
__DEVICE__ unsigned int __vseteq4(unsigned int __a, unsigned int __b)
__DEVICE__ long long __float2ll_rz(float __a)
__DEVICE__ unsigned int __vsetles4(unsigned int __a, unsigned int __b)
__DEVICE__ float __fadd_rn(float __a, float __b)
__device__ int __nv_float2int_rd(float __a)
__DEVICE__ void sincosf(float __a, float *__sptr, float *__cptr)
__DEVICE__ float rnorm3df(float __a, float __b, float __c)
__DEVICE__ void __threadfence_block(void)
__DEVICE__ unsigned int umin(unsigned int __a, unsigned int __b)
__DEVICE__ float __sinf(float __a)
__DEVICE__ float atanhf(float __a)
__DEVICE__ unsigned long long __ullAtomicOr(unsigned long long *__p, unsigned long long __v)
__DEVICE__ float roundf(float __a)
__device__ long long __nv_float2ll_ru(float __a)
__DEVICE__ double __dsqrt_rn(double __a)
__DEVICE__ void sincos(double __a, double *__sptr, double *__cptr)
__device__ float __nv_norm4df(float __a, float __b, float __c, float __d)
__device__ int __nv_abs(int __a)
__DEVICE__ int __iAtomicExch(int *__p, int __v)
__device__ float __nv_y1f(float __a)
__device__ double __nv_ull2double_rz(unsigned long long __a)
__DEVICE__ float y1f(float __a)
__device__ unsigned int __nv_umulhi(unsigned int __a, unsigned int __b)
__DEVICE__ float __fsqrt_ru(float __a)
__DEVICE__ float ldexpf(float __a, int __b)
__DEVICE__ double log10(double __a)
__DEVICE__ unsigned int __brev(unsigned int __a)
__DEVICE__ unsigned int __vsetleu2(unsigned int __a, unsigned int __b)
__device__ float __nv_fadd_ru(float __a, float __b)
__device__ double __nv_ddiv_rz(double __a, double __b)
__device__ double __nv_normcdfinv(double __a)
__DEVICE__ double cbrt(double __a)
__device__ float __nv_jnf(int __a, float __b)
__DEVICE__ float rcbrtf(float __a)
__device__ float __nv_normcdff(float __a)
__DEVICE__ float __uint2float_rn(unsigned int __a)
__device__ float __nv_fast_tanf(float __a)
__DEVICE__ unsigned int __uAtomicMin_system(unsigned int *__p, unsigned int __v)
__DEVICE__ void __trap(void)
__DEVICE__ unsigned long long __float2ull_ru(float __a)
__DEVICE__ long lroundf(float __a)
__DEVICE__ float __fadd_ru(float __a, float __b)
__DEVICE__ void sincospi(double __a, double *__sptr, double *__cptr)
__DEVICE__ double modf(double __a, double *__b)
__device__ double __nv_dsub_rd(double __a, double __b)
__device__ int __nv_double2int_rz(double __a)
__DEVICE__ double nextafter(double __a, double __b)
__DEVICE__ double erfcinv(double __a)
__device__ float __nv_uint_as_float(unsigned int __a)
__DEVICE__ float __double2float_rn(double __a)
__device__ double __nv_normcdf(double __a)
__device__ int __nv_rhadd(int __a, int __b)
__DEVICE__ int __iAtomicExch_system(int *__p, int __v)
__DEVICE__ double __drcp_ru(double __a)
__DEVICE__ unsigned int __vmins2(unsigned int __a, unsigned int __b)
__device__ float __nv_fminf(float __a, float __b)
__device__ double __nv_nearbyint(double __a)
__DEVICE__ unsigned int __double2uint_rn(double __a)
__DEVICE__ unsigned int __vabsdiffu2(unsigned int __a, unsigned int __b)
__device__ double __nv_uint2double_rn(unsigned int __i)
#define __DEVICE__
__device__ double __nv_rint(double __a)
__DEVICE__ unsigned long long __ullAtomicAnd_block(unsigned long long *__p, unsigned long long __v)
__DEVICE__ unsigned int __byte_perm(unsigned int __a, unsigned int __b, unsigned int __c)
__DEVICE__ unsigned long long __ullAtomicCAS_system(unsigned long long *__p, unsigned long long __cmp, unsigned long long __v)
__DEVICE__ unsigned int __uAtomicAdd(unsigned int *__p, unsigned int __v)
__DEVICE__ unsigned long long __ullAtomicMin_block(unsigned long long *__p, unsigned long long __v)
__DEVICE__ unsigned int __vcmpges2(unsigned int __a, unsigned int __b)
__device__ double __nv_log10(double __a)
#define INT_MIN
Definition: limits.h:67
__DEVICE__ unsigned int __vcmpleu2(unsigned int __a, unsigned int __b)
__DEVICE__ float __log10f(float __a)
__DEVICE__ double scalbln(double __a, long __b)
__DEVICE__ unsigned long long __brevll(unsigned long long __a)
__device__ float __nv_fdiv_ru(float __a, float __b)
__DEVICE__ int __syncthreads_or(int __a)
__DEVICE__ unsigned int __uAtomicOr_block(unsigned int *__p, unsigned int __v)
__DEVICE__ float erff(float __a)
__device__ double __nv_rnorm3d(double __a, double __b, double __c)
__DEVICE__ float __fmul_rn(float __a, float __b)
__DEVICE__ unsigned long long __ullAtomicXor_system(unsigned long long *__p, unsigned long long __v)
__DEVICE__ float tanhf(float __a)
__DEVICE__ float exp2f(float __a)
__device__ unsigned long long __nv_double_as_longlong(double __a)
__device__ unsigned long long __nv_ullmin(unsigned long long __a, unsigned long long __b)
__DEVICE__ double remainder(double __a, double __b)
static __inline__ vector float vector float __b
Definition: altivec.h:534
__DEVICE__ long long __double2ll_ru(double __a)
__DEVICE__ float fabsf(float __a)
__DEVICE__ unsigned int __vabsdiffs2(unsigned int __a, unsigned int __b)
__device__ double __nv_dsqrt_rd(double __a)
__DEVICE__ float atan2f(float __a, float __b)
__device__ int __nv_float_as_int(float __a)
__DEVICE__ unsigned long long __ullAtomicMax(unsigned long long *__p, unsigned long long __v)
__device__ double __nv_scalbn(double __a, int __b)
__DEVICE__ unsigned int __vcmpgts2(unsigned int __a, unsigned int __b)
__device__ double __nv_sinpi(double __a)
__DEVICE__ unsigned int __vsetgts2(unsigned int __a, unsigned int __b)
__device__ long long __nv_brevll(long long __a)
__device__ double __nv_remquo(double __a, double __b, int *__c)
__DEVICE__ unsigned long long __double2ull_ru(double __a)
__device__ double __nv_j0(double __a)
__DEVICE__ unsigned int __vaddss4(unsigned int __a, unsigned int __b)
__DEVICE__ float __uint_as_float(unsigned int __a)
__DEVICE__ float ynf(int __a, float __b)
__DEVICE__ double __int2double_rn(int __a)
__device__ float __nv_log2f(float __a)
__DEVICE__ unsigned int __ballot(int __a)
__device__ int __nv_double2int_ru(double __a)
__DEVICE__ float __fmaf_rn(float __a, float __b, float __c)
__device__ float __nv_fast_log10f(float __a)
__DEVICE__ int __isinf(double __a)
__DEVICE__ float cbrtf(float __a)
__device__ int __nv_float2int_rz(float __a)
__device__ double __nv_dmul_rd(double __a, double __b)
__DEVICE__ float __fmaf_ieee_rd(float __a, float __b, float __c)
__DEVICE__ int __rhadd(int __a, int __b)
__DEVICE__ unsigned int __vsads2(unsigned int __a, unsigned int __b)
__device__ float __nv_logf(float __a)
__DEVICE__ double __ll2double_ru(long long __a)
__device__ double __nv_fma_rz(double __a, double __b, double __c)
__device__ float __nv_cyl_bessel_i0f(float __a)
__device__ int __nv_double2hiint(double __a)
__DEVICE__ double __fma_ru(double __a, double __b, double __c)
__DEVICE__ void __brkpt()
__DEVICE__ int __syncthreads_count(int __a)
__device__ float __nv_logbf(float __a)
__DEVICE__ double __ll2double_rd(long long __a)
__DEVICE__ double asin(double __a)
__DEVICE__ long long __mul64hi(long long __a, long long __b)
__DEVICE__ double acos(double __a)
__device__ float __nv_lgammaf(float __a)
__DEVICE__ unsigned long long __ullAtomicOr_block(unsigned long long *__p, unsigned long long __v)
__DEVICE__ unsigned int __vcmpltu2(unsigned int __a, unsigned int __b)
__DEVICE__ int __double2int_rn(double __a)
__DEVICE__ long long __double2ll_rz(double __a)
__device__ float __nv_fmaf_ieee_rd(float __a, float __b, float __c)
__DEVICE__ unsigned int __vsads4(unsigned int __a, unsigned int __b)
__DEVICE__ unsigned long long __double2ull_rz(double __a)
__DEVICE__ unsigned int __vsetgeu2(unsigned int __a, unsigned int __b)
__device__ float __nv_fmaf_rn(float __a, float __b, float __c)
__DEVICE__ long long llabs(long long __a)
__DEVICE__ unsigned int __vmaxu2(unsigned int __a, unsigned int __b)
__device__ double __nv_tgamma(double __a)
__DEVICE__ int __iAtomicExch_block(int *__p, int __v)
__device__ unsigned long long __nv_double2ull_rn(double __a)
__DEVICE__ int __iAtomicMax_block(int *__p, int __v)
__DEVICE__ float expf(float __a)
__device__ double __nv_norm(int __a, const double *__b)
__DEVICE__ float powf(float __a, float __b)
__DEVICE__ float rnormf(int __dim, const float *__t)
__device__ int __nv_isinff(float __a)
__device__ float __nv_asinf(float __a)
__DEVICE__ float __frcp_ru(float __a)
__DEVICE__ int __ffs(int __a)
__DEVICE__ int __double2int_rz(double __a)
__DEVICE__ unsigned int __uAtomicMax_block(unsigned int *__p, unsigned int __v)
__device__ unsigned int __nv_float2uint_ru(float __a)
__DEVICE__ float tgammaf(float __a)
__DEVICE__ unsigned int __uAtomicInc_system(unsigned int *__p, unsigned int __v)
__device__ float __nv_double2float_rd(double __a)
__DEVICE__ double __dadd_ru(double __a, double __b)
__device__ double __nv_dsqrt_ru(double __a)
__device__ double __nv_j1(double __a)
__DEVICE__ double frexp(double __a, int *__b)
__DEVICE__ float remainderf(float __a, float __b)
__DEVICE__ double __dadd_rz(double __a, double __b)
__DEVICE__ unsigned int __uAtomicExch_system(unsigned int *__p, unsigned int __v)
__DEVICE__ unsigned int __vabsss2(unsigned int __a)
__device__ int __nv_float2int_ru(float __a)
__DEVICE__ float __ull2float_rn(unsigned long long __a)
__device__ float __nv_ull2float_ru(unsigned long long __a)
__DEVICE__ int __iAtomicMin(int *__p, int __v)
__DEVICE__ double __dsub_rn(double __a, double __b)
__device__ double __nv_hypot(double __a, double __b)
__DEVICE__ int __iAtomicOr(int *__p, int __v)
__DEVICE__ long long __llAtomicOr_block(long long *__p, long long __v)
__DEVICE__ double ldexp(double __a, int __b)
__device__ double __nv_drcp_rd(double __a)
__DEVICE__ unsigned int __vmaxs4(unsigned int __a, unsigned int __b)
__device__ float __nv_uint2float_ru(unsigned int __a)
__DEVICE__ unsigned int __uAtomicMax_system(unsigned int *__p, unsigned int __v)
__device__ float __nv_fmaf_rz(float __a, float __b, float __c)
__device__ float __nv_fmul_rz(float __a, float __b)
__DEVICE__ unsigned int __vadd2(unsigned int __a, unsigned int __b)
__DEVICE__ float truncf(float __a)
__DEVICE__ unsigned int __double2uint_rd(double __a)
__device__ float __nv_ceilf(float __a)
__DEVICE__ unsigned int __vcmples4(unsigned int __a, unsigned int __b)
__DEVICE__ double __ull2double_ru(unsigned long long __a)
__device__ long long __nv_float2ll_rn(float __a)
__device__ float __nv_double2float_ru(double __a)
__device__ float __nv_double2float_rn(double __a)
__device__ float __nv_int2float_rd(int __a)
__DEVICE__ int __all(int __a)
__DEVICE__ unsigned int umax(unsigned int __a, unsigned int __b)
__DEVICE__ double __ddiv_rz(double __a, double __b)
__device__ float __nv_exp10f(float __a)
__DEVICE__ unsigned int __vcmplts2(unsigned int __a, unsigned int __b)
__DEVICE__ float asinf(float __a)
__DEVICE__ double __dAtomicAdd_system(double *__p, double __v)
__DEVICE__ long long __illAtomicMax_system(long long *__p, long long __v)
__DEVICE__ unsigned int __uAtomicDec(unsigned int *__p, unsigned int __v)
__DEVICE__ float __ll2float_rz(long long __a)
__DEVICE__ unsigned int __uAtomicAdd_system(unsigned int *__p, unsigned int __v)
__DEVICE__ double sin(double __a)
__DEVICE__ unsigned long long __ullAtomicAnd_system(unsigned long long *__p, unsigned long long __v)
__DEVICE__ double exp10(double __a)
__DEVICE__ unsigned long long __float2ull_rd(float __a)
__device__ float __nv_acoshf(float __a)
__DEVICE__ double tgamma(double __a)
__DEVICE__ double norm(int __dim, const double *__t)
__DEVICE__ unsigned int __uhadd(unsigned int __a, unsigned int __b)
__DEVICE__ float __fdiv_ru(float __a, float __b)
__device__ float __nv_sinhf(float __a)
__device__ double __nv_erf(double __a)
__device__ double __nv_pow(double __a, double __b)
__DEVICE__ unsigned int __double2uint_rz(double __a)
__DEVICE__ float erfcf(float __a)
__device__ double __nv_ull2double_rn(unsigned long long __a)
__device__ double __nv_acosh(double __a)
__device__ float __nv_fmaf_ieee_ru(float __a, float __b, float __c)
__device__ int __nv_mul24(int __a, int __b)
__device__ double __nv_dmul_ru(double __a, double __b)
__device__ unsigned long long __nv_float2ull_rz(float __a)
__DEVICE__ unsigned int __usad(unsigned int __a, unsigned int __b, unsigned int __c)
__device__ unsigned int __nv_umax(unsigned int __a, unsigned int __b)
__device__ float __nv_ll2float_rd(long long __a)
__DEVICE__ unsigned int __vsubss2(unsigned int __a, unsigned int __b)
__device__ float __nv_fabsf(float __a)
__device__ long long __nv_llrint(double __a)
__device__ float __nv_fast_cosf(float __a)
__device__ float __nv_fmaf_ieee_rn(float __a, float __b, float __c)
__device__ int __nv_uhadd(unsigned int __a, unsigned int __b)
__device__ double __nv_dadd_rz(double __a, double __b)
__DEVICE__ double acosh(double __a)
__device__ double __nv_ll2double_rd(long long __a)
__device__ double __nv_dsub_ru(double __a, double __b)
__DEVICE__ float scalbnf(float __a, int __b)
#define HUGE_VALF
A positive float constant expression.
Definition: opencl-c.h:164
__device__ float __nv_fsqrt_rn(float __a)
__DEVICE__ unsigned long long __ullAtomicCAS_block(unsigned long long *__p, unsigned long long __cmp, unsigned long long __v)
__DEVICE__ long long __llAtomicXor_system(long long *__p, long long __v)
__device__ double __nv_cyl_bessel_i1(double __a)
__device__ float __nv_erfcf(float __a)
__DEVICE__ double fmod(double __a, double __b)
__DEVICE__ void * memcpy(void *__a, const void *__b, size_t __c)
__DEVICE__ float __fsub_rn(float __a, float __b)
__device__ float __nv_fmaxf(float __a, float __b)
__device__ double __nv_atanh(double __a)
__device__ int __nv_ilogbf(float __a)
__device__ int __nv_double2int_rd(double __a)
__device__ double __nv_asinh(double __a)
__DEVICE__ float jnf(int __n, float __a)
__DEVICE__ long long __float2ll_rn(float __a)
__device__ float __nv_nearbyintf(float __a)
__device__ float __nv_erfcxf(float __a)
__device__ float __nv_frcp_rn(float __a)
__device__ double __nv_ll2double_ru(long long __a)
__DEVICE__ long long __float2ll_ru(float __a)
__device__ float __nv_fdiv_rd(float __a, float __b)
__DEVICE__ float sinf(float __a)
__DEVICE__ int __iAtomicCAS(int *__p, int __cmp, int __v)
__DEVICE__ double __dsqrt_rz(double __a)
__device__ float __nv_erff(float __a)
__DEVICE__ int __float2int_rz(float __a)
__DEVICE__ unsigned int __uAtomicDec_block(unsigned int *__p, unsigned int __v)
__DEVICE__ unsigned int __urhadd(unsigned int __a, unsigned int __b)
__DEVICE__ unsigned int __uAtomicExch(unsigned int *__p, unsigned int __v)
__DEVICE__ int __any(int __a)
__DEVICE__ float fdimf(float __a, float __b)
__DEVICE__ double logb(double __a)
__DEVICE__ int __signbitd(double __a)
__device__ float __nv_rsqrtf(float __a)
__device__ int __nv_max(int __a, int __b)
__DEVICE__ unsigned int __vavgs2(unsigned int __a, unsigned int __b)
__DEVICE__ int __iAtomicOr_system(int *__p, int __v)
__device__ int __nv_clz(int __a)
__device__ float __nv_fast_logf(float __a)
__DEVICE__ unsigned int __vsetleu4(unsigned int __a, unsigned int __b)
__device__ void __nv_sincosf(float __a, float *__b, float *__c)
__device__ double __nv_yn(int __a, double __b)
__DEVICE__ double fdim(double __a, double __b)
__device__ float __nv_fmul_rn(float __a, float __b)
__DEVICE__ unsigned int __vsetles2(unsigned int __a, unsigned int __b)
__device__ float __nv_ull2float_rn(unsigned long long __a)
__DEVICE__ double __fma_rd(double __a, double __b, double __c)
__DEVICE__ long long llrintf(float __a)
__device__ double __nv_ddiv_rd(double __a, double __b)
__DEVICE__ unsigned int __vsubss4(unsigned int __a, unsigned int __b)
__device__ int __nv_mulhi(int __a, int __b)
__DEVICE__ float norm3df(float __a, float __b, float __c)
__DEVICE__ unsigned int __double2uint_ru(double __a)
__DEVICE__ float norm4df(float __a, float __b, float __c, float __d)
__DEVICE__ unsigned int __pm1(void)
__DEVICE__ double rint(double __a)
__device__ double __nv_trunc(double __a)
__DEVICE__ int __double2loint(double __a)
__device__ float __nv_sinf(float __a)
__DEVICE__ double __hiloint2double(int __a, int __b)
__DEVICE__ float __double2float_rz(double __a)
__DEVICE__ float __double2float_rd(double __a)
__DEVICE__ double cos(double __a)
__device__ double __nv_dadd_ru(double __a, double __b)
__DEVICE__ float __exp10f(float __a)
__device__ float __nv_ldexpf(float __a, int __b)
__device__ float __nv_ull2float_rz(unsigned long long __a)
__device__ double __nv_fma_rd(double __a, double __b, double __c)
__DEVICE__ float __fadd_rz(float __a, float __b)
__device__ double __nv_logb(double __a)
__DEVICE__ long lrintf(float __a)
__DEVICE__ unsigned int __uAtomicDec_system(unsigned int *__p, unsigned int __v)
__DEVICE__ float __fdiv_rn(float __a, float __b)
__device__ double __nv_copysign(double __a, double __b)
__device__ double __nv_erfcx(double __a)
__DEVICE__ long lrint(double __a)
__device__ long long __nv_float2ll_rd(float __a)
__DEVICE__ long long __illAtomicMax_block(long long *__p, long long __v)
__DEVICE__ long long __illAtomicMax(long long *__p, long long __v)
__DEVICE__ unsigned int __vaddss2(unsigned int __a, unsigned int __b)
__DEVICE__ unsigned int __uAtomicAdd_block(unsigned int *__p, unsigned int __v)
__DEVICE__ double __dmul_rn(double __a, double __b)
__device__ long long __nv_double2ll_ru(double __a)
__DEVICE__ double yn(int __a, double __b)
__DEVICE__ unsigned int __vsetltu2(unsigned int __a, unsigned int __b)
__DEVICE__ long long llroundf(float __a)
__DEVICE__ long long __illAtomicMin_block(long long *__p, long long __v)
__DEVICE__ double erfcx(double __a)
__DEVICE__ unsigned long long __ullAtomicMax_block(unsigned long long *__p, unsigned long long __v)
__device__ float __nv_modff(float __a, float *__b)
__device__ unsigned long long __nv_double2ull_rd(double __a)
__DEVICE__ float __cosf(float __a)
__DEVICE__ unsigned int __umul24(unsigned int __a, unsigned int __b)
__device__ double __nv_erfcinv(double __a)
__DEVICE__ float __uint2float_rd(unsigned int __a)
__device__ double __nv_rnorm(int __a, const double *__b)
__DEVICE__ float sinpif(float __a)
__DEVICE__ unsigned long long __ullAtomicAdd_system(unsigned long long *__p, unsigned long long __v)
__DEVICE__ double fabs(double __a)
__DEVICE__ float cyl_bessel_i1f(float __a)
__DEVICE__ long long __double2ll_rd(double __a)
__device__ float __nv_sinpif(float __a)
__DEVICE__ float normcdfinvf(float __a)
__DEVICE__ double __uint2double_rn(unsigned int __a)
__DEVICE__ unsigned long long __ullAtomicOr_system(unsigned long long *__p, unsigned long long __v)
__device__ double __nv_int2double_rn(int __a)
__device__ float __nv_acosf(float __a)
__device__ double __nv_ddiv_rn(double __a, double __b)
__DEVICE__ float scalblnf(float __a, long __b)
__device__ float __nv_fmaf_ru(float __a, float __b, float __c)
__DEVICE__ float cosf(float __a)
__DEVICE__ double normcdf(double __a)
__device__ double __nv_powi(double __a, int __b)
__DEVICE__ double hypot(double __a, double __b)
__device__ double __nv_rcbrt(double __a)
__DEVICE__ long long __illAtomicMin_system(long long *__p, long long __v)
__device__ int __nv_isnanf(float __a)
__DEVICE__ unsigned int __vnegss2(unsigned int __a)
__device__ float __nv_fmul_ru(float __a, float __b)
__device__ float __nv_fadd_rz(float __a, float __b)
__device__ float __nv_fmaf_ieee_rz(float __a, float __b, float __c)
__DEVICE__ int __popcll(long long __a)
__device__ float __nv_frcp_rz(float __a)
__DEVICE__ unsigned int __uAtomicAnd_system(unsigned int *__p, unsigned int __v)
__DEVICE__ int __signbitf(float __a)
__DEVICE__ unsigned int __pm0(void)
__device__ double __nv_longlong_as_double(long long __a)
__DEVICE__ long long llround(double __a)
__DEVICE__ int __finite(double __a)
__device__ double __nv_ceil(double __a)
__device__ long long __nv_llroundf(float __a)
__DEVICE__ unsigned int __vaddus4(unsigned int __a, unsigned int __b)
__DEVICE__ double asinh(double __a)
__DEVICE__ double remquo(double __a, double __b, int *__c)
__DEVICE__ unsigned int __vminu2(unsigned int __a, unsigned int __b)
__device__ float __nv_fmaf(float __a, float __b, float __c)
__DEVICE__ void __sincosf(float __a, float *__sptr, float *__cptr)
__DEVICE__ unsigned long long __ullAtomicExch(unsigned long long *__p, unsigned long long __v)
__DEVICE__ float __fsub_rd(float __a, float __b)
__DEVICE__ float logbf(float __a)
__DEVICE__ double ceil(double __a)
__DEVICE__ unsigned int __vsubus2(unsigned int __a, unsigned int __b)
__DEVICE__ double y1(double __a)
__DEVICE__ double fmax(double __a, double __b)
__device__ float __nv_fast_exp10f(float __a)
__DEVICE__ unsigned int __sad(int __a, int __b, unsigned int __c)
__DEVICE__ unsigned int __vsub4(unsigned int __a, unsigned int __b)
__device__ double __nv_dadd_rn(double __a, double __b)
__device__ double __nv_fmax(double __a, double __b)
__DEVICE__ unsigned long long __ullAtomicAdd(unsigned long long *__p, unsigned long long __v)
__DEVICE__ double expm1(double __a)
__device__ double __nv_sqrt(double __a)
__device__ float __nv_fsub_rz(float __a, float __b)
__device__ int __nv_byte_perm(int __a, int __b, int __c)
__DEVICE__ float j1f(float __a)
__device__ float __nv_atanhf(float __a)
__device__ double __nv_y1(double __a)
__DEVICE__ int __double2int_ru(double __a)
__DEVICE__ float hypotf(float __a, float __b)
__DEVICE__ float __powf(float __a, float __b)
__DEVICE__ double rcbrt(double __a)
__DEVICE__ float sqrtf(float __a)
__DEVICE__ long long __illAtomicMin(long long *__p, long long __v)
__device__ double __nv_exp(double __a)
__DEVICE__ long long llrint(double __a)
__device__ double __nv_asin(double __a)
__device__ double __nv_dsqrt_rz(double __a)
__DEVICE__ unsigned long long __ullAtomicCAS(unsigned long long *__p, unsigned long long __cmp, unsigned long long __v)
__device__ double __nv_modf(double __a, double *__b)
__DEVICE__ int __float_as_int(float __a)
__DEVICE__ int __iAtomicAdd_block(int *__p, int __v)
__DEVICE__ unsigned int __vcmpgts4(unsigned int __a, unsigned int __b)
__device__ float __nv_rnorm4df(float __a, float __b, float __c, float __d)
__device__ double __nv_ull2double_ru(unsigned long long __a)
__device__ long long __nv_llrintf(float __a)
__DEVICE__ float log10f(float __a)
__DEVICE__ double __dsub_rz(double __a, double __b)
__DEVICE__ long long llmax(long long __a, long long __b)
__DEVICE__ long long __llAtomicOr_system(long long *__p, long long __v)
__DEVICE__ unsigned int __uAtomicXor_block(unsigned int *__p, unsigned int __v)
__DEVICE__ double atan2(double __a, double __b)
__DEVICE__ float __fmaf_ieee_rz(float __a, float __b, float __c)
__device__ float __nv_fsqrt_ru(float __a)
__device__ float __nv_fdiv_rz(float __a, float __b)
__device__ float __nv_saturatef(float __a)
#define INT_MAX
Definition: limits.h:62
__DEVICE__ float __frcp_rz(float __a)
__DEVICE__ float __int_as_float(int __a)
__DEVICE__ unsigned int __vsetne4(unsigned int __a, unsigned int __b)
__DEVICE__ long long clock64()
__device__ double __nv_log2(double __a)
__DEVICE__ unsigned int __uAtomicCAS(unsigned int *__p, unsigned int __cmp, unsigned int __v)
__DEVICE__ unsigned int __vcmpgtu4(unsigned int __a, unsigned int __b)
__DEVICE__ double __ddiv_ru(double __a, double __b)
__DEVICE__ double __dAtomicAdd(double *__p, double __v)
__device__ unsigned int __nv_double2uint_rd(double __a)
__DEVICE__ unsigned long long __umul64hi(unsigned long long __a, unsigned long long __b)
__DEVICE__ unsigned int __uAtomicOr(unsigned int *__p, unsigned int __v)
__DEVICE__ int __iAtomicAnd_system(int *__p, int __v)
__device__ double __nv_norm3d(double __a, double __b, double __c)
__DEVICE__ double rnorm4d(double __a, double __b, double __c, double __d)
__DEVICE__ double scalbn(double __a, int __b)
__device__ float __nv_log1pf(float __a)
__device__ double __nv_cospi(double __a)
__device__ float __nv_uint2float_rz(unsigned int __a)
__device__ void __nv_fast_sincosf(float __a, float *__sptr, float *__cptr)
__DEVICE__ float lgammaf(float __a)
__DEVICE__ float __ll2float_rd(long long __a)
__DEVICE__ float erfcxf(float __a)
__device__ unsigned int __nv_umin(unsigned int __a, unsigned int __b)
__device__ float __nv_cbrtf(float __a)
__DEVICE__ unsigned int __uAtomicMin(unsigned int *__p, unsigned int __v)
__device__ int __nv_double2int_rn(double __a)
__DEVICE__ unsigned int __vsetgts4(unsigned int __a, unsigned int __b)
__DEVICE__ unsigned int __float2uint_rd(float __a)
__DEVICE__ int max(int __a, int __b)
__device__ float __nv_frcp_rd(float __a)
__DEVICE__ long long __double_as_longlong(double __a)
__DEVICE__ double __dmul_rz(double __a, double __b)
__DEVICE__ float __int2float_rd(int __a)
__DEVICE__ unsigned int __vabsdiffs4(unsigned int __a, unsigned int __b)
__DEVICE__ unsigned long long __ullAtomicAnd(unsigned long long *__p, unsigned long long __v)
__DEVICE__ float normf(int __dim, const float *__t)
__DEVICE__ int __mul24(int __a, int __b)
__DEVICE__ double __drcp_rn(double __a)
__DEVICE__ double floor(double __f)
__DEVICE__ unsigned int __vsetgeu4(unsigned int __a, unsigned int __b)
__device__ float __nv_uint2float_rd(unsigned int __a)
__DEVICE__ int min(int __a, int __b)
__device__ float __nv_sqrtf(float __a)
__DEVICE__ long long __float2ll_rd(float __a)
__DEVICE__ float floorf(float __f)
__device__ double __nv_tanh(double __a)
__device__ double __nv_cyl_bessel_i0(double __a)
__DEVICE__ int __isnan(double __a)
__DEVICE__ long lround(double __a)
__device__ float __nv_tgammaf(float __a)
__DEVICE__ unsigned int __vabs2(unsigned int __a)
__device__ float __nv_fadd_rn(float __a, float __b)
__DEVICE__ unsigned long long __ullAtomicMin(unsigned long long *__p, unsigned long long __v)
__DEVICE__ double __dadd_rn(double __a, double __b)
__device__ unsigned int __nv_float_as_uint(float __a)
__DEVICE__ float modff(float __a, float *__b)
__DEVICE__ double rnorm3d(double __a, double __b, double __c)
__DEVICE__ float __fAtomicExch(float *__p, float __v)
__device__ int __nv_double2loint(double __a)
__DEVICE__ unsigned int __uAtomicAnd_block(unsigned int *__p, unsigned int __v)
__DEVICE__ double norm4d(double __a, double __b, double __c, double __d)
__DEVICE__ float j0f(float __a)
__DEVICE__ int __iAtomicCAS_system(int *__p, int __cmp, int __v)
__device__ float __nv_erfinvf(float __a)
__device__ unsigned long long __nv_float2ull_rn(float __a)
__device__ int __nv_isnand(double __a)
__DEVICE__ double __drcp_rd(double __a)
static __inline__ vector float vector float vector float __c
Definition: altivec.h:4199
__DEVICE__ float log1pf(float __a)
__DEVICE__ unsigned int __vaddus2(unsigned int __a, unsigned int __b)
__DEVICE__ unsigned int __vcmpltu4(unsigned int __a, unsigned int __b)
__device__ float __nv_fsub_rd(float __a, float __b)
__DEVICE__ double __dmul_ru(double __a, double __b)
__DEVICE__ unsigned int __float_as_uint(float __a)
__DEVICE__ double sinh(double __a)
__DEVICE__ float __frsqrt_rn(float __a)
__DEVICE__ double fdivide(double __a, double __b)
__device__ double __nv_cbrt(double __a)
__device__ double __nv_fma(double __a, double __b, double __c)
__device__ double __nv_erfc(double __a)
__device__ float __nv_atan2f(float __a, float __b)
__DEVICE__ double __longlong_as_double(long long __a)
__device__ float __nv_hypotf(float __a, float __b)
__device__ float __nv_fsub_rn(float __a, float __b)
__DEVICE__ double __drcp_rz(double __a)
__DEVICE__ double trunc(double __a)
__DEVICE__ float __double2float_ru(double __a)
__DEVICE__ unsigned long long __ullAtomicExch_block(unsigned long long *__p, unsigned long long __v)
__device__ unsigned int __nv_float2uint_rz(float __a)
__DEVICE__ float __log2f(float __a)
__DEVICE__ int __isnanf(float __a)
__DEVICE__ unsigned int __vabsdiffu4(unsigned int __a, unsigned int __b)
__DEVICE__ unsigned int __vhaddu4(unsigned int __a, unsigned int __b)
__device__ double __nv_exp2(double __a)
__DEVICE__ int __finitef(float __a)
__DEVICE__ float copysignf(float __a, float __b)
__DEVICE__ float rintf(float __a)
__DEVICE__ long long __llAtomicAnd(long long *__p, long long __v)
__device__ float __nv_tanf(float __a)
__device__ double __nv_erfinv(double __a)
__device__ float __nv_rnorm3df(float __a, float __b, float __c)
__device__ float __nv_rhypotf(float __a, float __b)
__DEVICE__ unsigned long long __float2ull_rz(float __a)
__DEVICE__ float __fmul_rz(float __a, float __b)
__DEVICE__ int __iAtomicMax_system(int *__p, int __v)
__DEVICE__ unsigned int __vsetges4(unsigned int __a, unsigned int __b)