24 #ifndef __CLANG_CUDA_DEVICE_FUNCTIONS_H__ 25 #define __CLANG_CUDA_DEVICE_FUNCTIONS_H__ 27 #if CUDA_VERSION < 9000 28 #error This file is intended to be used with CUDA-9+ only. 34 #pragma push_macro("__DEVICE__") 35 #define __DEVICE__ static __device__ __forceinline__ 41 #pragma push_macro("__FAST_OR_SLOW") 42 #if defined(__CLANG_CUDA_APPROX_TRANSCENDENTALS__) 43 #define __FAST_OR_SLOW(fast, slow) fast 45 #define __FAST_OR_SLOW(fast, slow) slow 65 return __nvvm_atom_add_gen_d(__p, __v);
68 return __nvvm_atom_cta_add_gen_d(__p, __v);
71 return __nvvm_atom_sys_add_gen_d(__p, __v);
189 return __nvvm_atom_add_gen_f(__p, __v);
192 return __nvvm_atom_cta_add_gen_f(__p, __v);
195 return __nvvm_atom_sys_add_gen_f(__p, __v);
350 return __nvvm_atom_add_gen_i(__p, __v);
353 __nvvm_atom_cta_add_gen_i(__p, __v);
356 __nvvm_atom_sys_add_gen_i(__p, __v);
359 return __nvvm_atom_and_gen_i(__p, __v);
362 return __nvvm_atom_cta_and_gen_i(__p, __v);
365 return __nvvm_atom_sys_and_gen_i(__p, __v);
368 return __nvvm_atom_cas_gen_i(__p, __cmp, __v);
371 return __nvvm_atom_cta_cas_gen_i(__p, __cmp, __v);
374 return __nvvm_atom_sys_cas_gen_i(__p, __cmp, __v);
377 return __nvvm_atom_xchg_gen_i(__p, __v);
380 return __nvvm_atom_cta_xchg_gen_i(__p, __v);
383 return __nvvm_atom_sys_xchg_gen_i(__p, __v);
386 return __nvvm_atom_max_gen_i(__p, __v);
389 return __nvvm_atom_cta_max_gen_i(__p, __v);
392 return __nvvm_atom_sys_max_gen_i(__p, __v);
395 return __nvvm_atom_min_gen_i(__p, __v);
398 return __nvvm_atom_cta_min_gen_i(__p, __v);
401 return __nvvm_atom_sys_min_gen_i(__p, __v);
404 return __nvvm_atom_or_gen_i(__p, __v);
407 return __nvvm_atom_cta_or_gen_i(__p, __v);
410 return __nvvm_atom_sys_or_gen_i(__p, __v);
413 return __nvvm_atom_xor_gen_i(__p, __v);
416 return __nvvm_atom_cta_xor_gen_i(__p, __v);
419 return __nvvm_atom_sys_xor_gen_i(__p, __v);
422 return __nvvm_atom_max_gen_ll(__p, __v);
425 return __nvvm_atom_cta_max_gen_ll(__p, __v);
428 return __nvvm_atom_sys_max_gen_ll(__p, __v);
431 return __nvvm_atom_min_gen_ll(__p, __v);
434 return __nvvm_atom_cta_min_gen_ll(__p, __v);
437 return __nvvm_atom_sys_min_gen_ll(__p, __v);
467 return __nvvm_atom_and_gen_ll(__p, __v);
470 return __nvvm_atom_cta_and_gen_ll(__p, __v);
473 return __nvvm_atom_sys_and_gen_ll(__p, __v);
476 return __nvvm_atom_or_gen_ll(__p, __v);
479 return __nvvm_atom_cta_or_gen_ll(__p, __v);
482 return __nvvm_atom_sys_or_gen_ll(__p, __v);
485 return __nvvm_atom_xor_gen_ll(__p, __v);
488 return __nvvm_atom_cta_xor_gen_ll(__p, __v);
491 return __nvvm_atom_sys_xor_gen_ll(__p, __v);
515 #define __prof_trigger(__a) asm __volatile__("pmevent \t%0;" ::"i"(__a)) 536 return __nvvm_atom_add_gen_i((
int *)__p, __v);
540 return __nvvm_atom_cta_add_gen_i((
int *)__p, __v);
544 return __nvvm_atom_sys_add_gen_i((
int *)__p, __v);
547 return __nvvm_atom_and_gen_i((
int *)__p, __v);
551 return __nvvm_atom_cta_and_gen_i((
int *)__p, __v);
555 return __nvvm_atom_sys_and_gen_i((
int *)__p, __v);
559 return __nvvm_atom_cas_gen_i((
int *)__p, __cmp, __v);
563 return __nvvm_atom_cta_cas_gen_i((
int *)__p, __cmp, __v);
567 return __nvvm_atom_sys_cas_gen_i((
int *)__p, __cmp, __v);
570 return __nvvm_atom_dec_gen_ui(__p, __v);
574 return __nvvm_atom_cta_dec_gen_ui(__p, __v);
578 return __nvvm_atom_sys_dec_gen_ui(__p, __v);
581 return __nvvm_atom_xchg_gen_i((
int *)__p, __v);
585 return __nvvm_atom_cta_xchg_gen_i((
int *)__p, __v);
589 return __nvvm_atom_sys_xchg_gen_i((
int *)__p, __v);
592 return __nvvm_atom_inc_gen_ui(__p, __v);
596 return __nvvm_atom_cta_inc_gen_ui(__p, __v);
600 return __nvvm_atom_sys_inc_gen_ui(__p, __v);
603 return __nvvm_atom_max_gen_ui(__p, __v);
607 return __nvvm_atom_cta_max_gen_ui(__p, __v);
611 return __nvvm_atom_sys_max_gen_ui(__p, __v);
614 return __nvvm_atom_min_gen_ui(__p, __v);
618 return __nvvm_atom_cta_min_gen_ui(__p, __v);
622 return __nvvm_atom_sys_min_gen_ui(__p, __v);
625 return __nvvm_atom_or_gen_i((
int *)__p, __v);
628 return __nvvm_atom_cta_or_gen_i((
int *)__p, __v);
632 return __nvvm_atom_sys_or_gen_i((
int *)__p, __v);
635 return __nvvm_atom_xor_gen_i((
int *)__p, __v);
639 return __nvvm_atom_cta_xor_gen_i((
int *)__p, __v);
643 return __nvvm_atom_sys_xor_gen_i((
int *)__p, __v);
691 unsigned long long __v) {
692 return __nvvm_atom_add_gen_ll((
long long *)__p, __v);
695 unsigned long long __v) {
696 return __nvvm_atom_cta_add_gen_ll((
long long *)__p, __v);
699 unsigned long long __v) {
700 return __nvvm_atom_sys_add_gen_ll((
long long *)__p, __v);
703 unsigned long long __v) {
704 return __nvvm_atom_and_gen_ll((
long long *)__p, __v);
707 unsigned long long __v) {
708 return __nvvm_atom_cta_and_gen_ll((
long long *)__p, __v);
711 unsigned long long __v) {
712 return __nvvm_atom_sys_and_gen_ll((
long long *)__p, __v);
715 unsigned long long __cmp,
716 unsigned long long __v) {
717 return __nvvm_atom_cas_gen_ll((
long long *)__p, __cmp, __v);
720 unsigned long long __cmp,
721 unsigned long long __v) {
722 return __nvvm_atom_cta_cas_gen_ll((
long long *)__p, __cmp, __v);
725 unsigned long long __cmp,
726 unsigned long long __v) {
727 return __nvvm_atom_sys_cas_gen_ll((
long long *)__p, __cmp, __v);
730 unsigned long long __v) {
731 return __nvvm_atom_xchg_gen_ll((
long long *)__p, __v);
734 unsigned long long __v) {
735 return __nvvm_atom_cta_xchg_gen_ll((
long long *)__p, __v);
738 unsigned long long __v) {
739 return __nvvm_atom_sys_xchg_gen_ll((
long long *)__p, __v);
742 unsigned long long __v) {
743 return __nvvm_atom_max_gen_ull(__p, __v);
746 unsigned long long __v) {
747 return __nvvm_atom_cta_max_gen_ull(__p, __v);
750 unsigned long long __v) {
751 return __nvvm_atom_sys_max_gen_ull(__p, __v);
754 unsigned long long __v) {
755 return __nvvm_atom_min_gen_ull(__p, __v);
758 unsigned long long __v) {
759 return __nvvm_atom_cta_min_gen_ull(__p, __v);
762 unsigned long long __v) {
763 return __nvvm_atom_sys_min_gen_ull(__p, __v);
766 unsigned long long __v) {
767 return __nvvm_atom_or_gen_ll((
long long *)__p, __v);
770 unsigned long long __v) {
771 return __nvvm_atom_cta_or_gen_ll((
long long *)__p, __v);
774 unsigned long long __v) {
775 return __nvvm_atom_sys_or_gen_ll((
long long *)__p, __v);
778 unsigned long long __v) {
779 return __nvvm_atom_xor_gen_ll((
long long *)__p, __v);
782 unsigned long long __v) {
783 return __nvvm_atom_cta_xor_gen_ll((
long long *)__p, __v);
786 unsigned long long __v) {
787 return __nvvm_atom_sys_xor_gen_ll((
long long *)__p, __v);
793 unsigned long long __b) {
807 #if CUDA_VERSION >= 9000 && CUDA_VERSION < 9020 811 return __nv_vabsdiffs2(
__a, __b);
814 return __nv_vabsdiffs4(
__a, __b);
817 return __nv_vabsdiffu2(
__a, __b);
820 return __nv_vabsdiffu4(
__a, __b);
823 return __nv_vabsss2(__a);
826 return __nv_vabsss4(__a);
829 return __nv_vadd2(
__a, __b);
832 return __nv_vadd4(
__a, __b);
835 return __nv_vaddss2(
__a, __b);
838 return __nv_vaddss4(
__a, __b);
841 return __nv_vaddus2(
__a, __b);
844 return __nv_vaddus4(
__a, __b);
847 return __nv_vavgs2(
__a, __b);
850 return __nv_vavgs4(
__a, __b);
853 return __nv_vavgu2(
__a, __b);
856 return __nv_vavgu4(
__a, __b);
859 return __nv_vcmpeq2(
__a, __b);
862 return __nv_vcmpeq4(
__a, __b);
865 return __nv_vcmpges2(
__a, __b);
868 return __nv_vcmpges4(
__a, __b);
871 return __nv_vcmpgeu2(
__a, __b);
874 return __nv_vcmpgeu4(
__a, __b);
877 return __nv_vcmpgts2(
__a, __b);
880 return __nv_vcmpgts4(
__a, __b);
883 return __nv_vcmpgtu2(
__a, __b);
886 return __nv_vcmpgtu4(
__a, __b);
889 return __nv_vcmples2(
__a, __b);
892 return __nv_vcmples4(
__a, __b);
895 return __nv_vcmpleu2(
__a, __b);
898 return __nv_vcmpleu4(
__a, __b);
901 return __nv_vcmplts2(
__a, __b);
904 return __nv_vcmplts4(
__a, __b);
907 return __nv_vcmpltu2(
__a, __b);
910 return __nv_vcmpltu4(
__a, __b);
913 return __nv_vcmpne2(
__a, __b);
916 return __nv_vcmpne4(
__a, __b);
919 return __nv_vhaddu2(
__a, __b);
922 return __nv_vhaddu4(
__a, __b);
925 return __nv_vmaxs2(
__a, __b);
928 return __nv_vmaxs4(
__a, __b);
931 return __nv_vmaxu2(
__a, __b);
934 return __nv_vmaxu4(
__a, __b);
937 return __nv_vmins2(
__a, __b);
940 return __nv_vmins4(
__a, __b);
943 return __nv_vminu2(
__a, __b);
946 return __nv_vminu4(
__a, __b);
951 return __nv_vnegss2(__a);
954 return __nv_vnegss4(__a);
957 return __nv_vsads2(
__a, __b);
960 return __nv_vsads4(
__a, __b);
963 return __nv_vsadu2(
__a, __b);
966 return __nv_vsadu4(
__a, __b);
969 return __nv_vseteq2(
__a, __b);
972 return __nv_vseteq4(
__a, __b);
975 return __nv_vsetges2(
__a, __b);
978 return __nv_vsetges4(
__a, __b);
981 return __nv_vsetgeu2(
__a, __b);
984 return __nv_vsetgeu4(
__a, __b);
987 return __nv_vsetgts2(
__a, __b);
990 return __nv_vsetgts4(
__a, __b);
993 return __nv_vsetgtu2(
__a, __b);
996 return __nv_vsetgtu4(
__a, __b);
999 return __nv_vsetles2(
__a, __b);
1002 return __nv_vsetles4(
__a, __b);
1005 return __nv_vsetleu2(
__a, __b);
1008 return __nv_vsetleu4(
__a, __b);
1011 return __nv_vsetlts2(
__a, __b);
1014 return __nv_vsetlts4(
__a, __b);
1017 return __nv_vsetltu2(
__a, __b);
1020 return __nv_vsetltu4(
__a, __b);
1023 return __nv_vsetne2(
__a, __b);
1026 return __nv_vsetne4(
__a, __b);
1029 return __nv_vsub2(
__a, __b);
1032 return __nv_vsub4(
__a, __b);
1035 return __nv_vsubss2(
__a, __b);
1038 return __nv_vsubss4(
__a, __b);
1041 return __nv_vsubus2(
__a, __b);
1044 return __nv_vsubus4(
__a, __b);
1046 #else // CUDA_VERSION >= 9020 1055 return (__a << shift) -
__a;
1059 asm(
"vabsdiff2.s32.s32.s32 %0,%1,%2,%3;" 1061 :
"r"(__a),
"r"(0),
"r"(0));
1066 asm(
"vabsdiff4.s32.s32.s32 %0,%1,%2,%3;" 1068 :
"r"(__a),
"r"(0),
"r"(0));
1073 asm(
"vabsdiff2.s32.s32.s32 %0,%1,%2,%3;" 1075 :
"r"(__a),
"r"(
__b),
"r"(0));
1081 asm(
"vabsdiff4.s32.s32.s32 %0,%1,%2,%3;" 1083 :
"r"(__a),
"r"(
__b),
"r"(0));
1088 asm(
"vabsdiff2.u32.u32.u32 %0,%1,%2,%3;" 1090 :
"r"(__a),
"r"(
__b),
"r"(0));
1095 asm(
"vabsdiff4.u32.u32.u32 %0,%1,%2,%3;" 1097 :
"r"(__a),
"r"(
__b),
"r"(0));
1102 asm(
"vabsdiff2.s32.s32.s32.sat %0,%1,%2,%3;" 1104 :
"r"(__a),
"r"(0),
"r"(0));
1109 asm(
"vabsdiff4.s32.s32.s32.sat %0,%1,%2,%3;" 1111 :
"r"(__a),
"r"(0),
"r"(0));
1116 asm(
"vadd2.u32.u32.u32 %0,%1,%2,%3;" :
"=r"(r) :
"r"(__a),
"r"(
__b),
"r"(0));
1121 asm(
"vadd4.u32.u32.u32 %0,%1,%2,%3;" :
"=r"(r) :
"r"(__a),
"r"(
__b),
"r"(0));
1126 asm(
"vadd2.s32.s32.s32.sat %0,%1,%2,%3;" 1128 :
"r"(__a),
"r"(
__b),
"r"(0));
1133 asm(
"vadd4.s32.s32.s32.sat %0,%1,%2,%3;" 1135 :
"r"(__a),
"r"(
__b),
"r"(0));
1140 asm(
"vadd2.u32.u32.u32.sat %0,%1,%2,%3;" 1142 :
"r"(__a),
"r"(
__b),
"r"(0));
1147 asm(
"vadd4.u32.u32.u32.sat %0,%1,%2,%3;" 1149 :
"r"(__a),
"r"(
__b),
"r"(0));
1154 asm(
"vavrg2.s32.s32.s32 %0,%1,%2,%3;" :
"=r"(r) :
"r"(__a),
"r"(
__b),
"r"(0));
1159 asm(
"vavrg4.s32.s32.s32 %0,%1,%2,%3;" :
"=r"(r) :
"r"(__a),
"r"(
__b),
"r"(0));
1164 asm(
"vavrg2.u32.u32.u32 %0,%1,%2,%3;" :
"=r"(r) :
"r"(__a),
"r"(
__b),
"r"(0));
1169 asm(
"vavrg4.u32.u32.u32 %0,%1,%2,%3;" :
"=r"(r) :
"r"(__a),
"r"(
__b),
"r"(0));
1174 asm(
"vset2.u32.u32.eq %0,%1,%2,%3;" :
"=r"(r) :
"r"(__a),
"r"(
__b),
"r"(0));
1182 asm(
"vset4.u32.u32.eq %0,%1,%2,%3;" :
"=r"(r) :
"r"(__a),
"r"(
__b),
"r"(0));
1190 asm(
"vset2.s32.s32.ge %0,%1,%2,%3;" :
"=r"(r) :
"r"(__a),
"r"(
__b),
"r"(0));
1198 asm(
"vset4.s32.s32.ge %0,%1,%2,%3;" :
"=r"(r) :
"r"(__a),
"r"(
__b),
"r"(0));
1206 asm(
"vset2.u32.u32.ge %0,%1,%2,%3;" :
"=r"(r) :
"r"(__a),
"r"(
__b),
"r"(0));
1214 asm(
"vset4.u32.u32.ge %0,%1,%2,%3;" :
"=r"(r) :
"r"(__a),
"r"(
__b),
"r"(0));
1222 asm(
"vset2.s32.s32.gt %0,%1,%2,%3;" :
"=r"(r) :
"r"(__a),
"r"(
__b),
"r"(0));
1230 asm(
"vset4.s32.s32.gt %0,%1,%2,%3;" :
"=r"(r) :
"r"(__a),
"r"(
__b),
"r"(0));
1238 asm(
"vset2.u32.u32.gt %0,%1,%2,%3;" :
"=r"(r) :
"r"(__a),
"r"(
__b),
"r"(0));
1246 asm(
"vset4.u32.u32.gt %0,%1,%2,%3;" :
"=r"(r) :
"r"(__a),
"r"(
__b),
"r"(0));
1254 asm(
"vset2.s32.s32.le %0,%1,%2,%3;" :
"=r"(r) :
"r"(__a),
"r"(
__b),
"r"(0));
1262 asm(
"vset4.s32.s32.le %0,%1,%2,%3;" :
"=r"(r) :
"r"(__a),
"r"(
__b),
"r"(0));
1270 asm(
"vset2.u32.u32.le %0,%1,%2,%3;" :
"=r"(r) :
"r"(__a),
"r"(
__b),
"r"(0));
1278 asm(
"vset4.u32.u32.le %0,%1,%2,%3;" :
"=r"(r) :
"r"(__a),
"r"(
__b),
"r"(0));
1286 asm(
"vset2.s32.s32.lt %0,%1,%2,%3;" :
"=r"(r) :
"r"(__a),
"r"(
__b),
"r"(0));
1294 asm(
"vset4.s32.s32.lt %0,%1,%2,%3;" :
"=r"(r) :
"r"(__a),
"r"(
__b),
"r"(0));
1302 asm(
"vset2.u32.u32.lt %0,%1,%2,%3;" :
"=r"(r) :
"r"(__a),
"r"(
__b),
"r"(0));
1310 asm(
"vset4.u32.u32.lt %0,%1,%2,%3;" :
"=r"(r) :
"r"(__a),
"r"(
__b),
"r"(0));
1318 asm(
"vset2.u32.u32.ne %0,%1,%2,%3;" :
"=r"(r) :
"r"(__a),
"r"(
__b),
"r"(0));
1326 asm(
"vset4.u32.u32.ne %0,%1,%2,%3;" :
"=r"(r) :
"r"(__a),
"r"(
__b),
"r"(0));
1339 return (((__a ^ __b) >> 1) & ~0x80008000u) + (__a &
__b);
1342 return (((__a ^ __b) >> 1) & ~0x80808080u) + (__a &
__b);
1347 if ((__a & 0x8000) && (__b & 0x8000)) {
1351 r = (__a & mask) | (__b & ~mask);
1353 asm(
"vmax2.s32.s32.s32 %0,%1,%2,%3;" 1355 :
"r"(__a),
"r"(
__b),
"r"(0));
1361 asm(
"vmax4.s32.s32.s32 %0,%1,%2,%3;" :
"=r"(r) :
"r"(__a),
"r"(
__b),
"r"(0));
1366 asm(
"vmax2.u32.u32.u32 %0,%1,%2,%3;" :
"=r"(r) :
"r"(__a),
"r"(
__b),
"r"(0));
1371 asm(
"vmax4.u32.u32.u32 %0,%1,%2,%3;" :
"=r"(r) :
"r"(__a),
"r"(
__b),
"r"(0));
1376 asm(
"vmin2.s32.s32.s32 %0,%1,%2,%3;" :
"=r"(r) :
"r"(__a),
"r"(
__b),
"r"(0));
1381 asm(
"vmin4.s32.s32.s32 %0,%1,%2,%3;" :
"=r"(r) :
"r"(__a),
"r"(
__b),
"r"(0));
1386 asm(
"vmin2.u32.u32.u32 %0,%1,%2,%3;" :
"=r"(r) :
"r"(__a),
"r"(
__b),
"r"(0));
1391 asm(
"vmin4.u32.u32.u32 %0,%1,%2,%3;" :
"=r"(r) :
"r"(__a),
"r"(
__b),
"r"(0));
1396 asm(
"vabsdiff2.s32.s32.s32.add %0,%1,%2,%3;" 1398 :
"r"(__a),
"r"(
__b),
"r"(0));
1403 asm(
"vabsdiff4.s32.s32.s32.add %0,%1,%2,%3;" 1405 :
"r"(__a),
"r"(
__b),
"r"(0));
1410 asm(
"vabsdiff2.u32.u32.u32.add %0,%1,%2,%3;" 1412 :
"r"(__a),
"r"(
__b),
"r"(0));
1417 asm(
"vabsdiff4.u32.u32.u32.add %0,%1,%2,%3;" 1419 :
"r"(__a),
"r"(
__b),
"r"(0));
1425 asm(
"vsub2.u32.u32.u32 %0,%1,%2,%3;" :
"=r"(r) :
"r"(__a),
"r"(
__b),
"r"(0));
1432 asm(
"vsub4.u32.u32.u32 %0,%1,%2,%3;" :
"=r"(r) :
"r"(__a),
"r"(
__b),
"r"(0));
1438 asm(
"vsub2.s32.s32.s32.sat %0,%1,%2,%3;" 1440 :
"r"(__a),
"r"(
__b),
"r"(0));
1448 asm(
"vsub4.s32.s32.s32.sat %0,%1,%2,%3;" 1450 :
"r"(__a),
"r"(
__b),
"r"(0));
1458 asm(
"vsub2.u32.u32.u32.sat %0,%1,%2,%3;" 1460 :
"r"(__a),
"r"(
__b),
"r"(0));
1465 asm(
"vsub4.u32.u32.u32.sat %0,%1,%2,%3;" 1467 :
"r"(__a),
"r"(
__b),
"r"(0));
1470 #endif // CUDA_VERSION >= 9020 1534 #if __FAST_MATH__ && !__CUDA_PREC_DIV 1566 #if defined(__LP64__) 1600 #if defined(__LP64__) 1613 return __builtin_memcpy(__a, __b, __c);
1616 return __builtin_memset(__a, __b, __c);
1705 return __a > 0 ? 0.0 : -0.0;
1706 return scalbn(__a, (
int)__b);
1712 return __a > 0 ? 0.f : -0.f;
1713 return scalbnf(__a, (
int)__b);
1746 unsigned long long __b) {
1750 unsigned long long __b) {
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)
struct __storeu_i16 *__P __v
__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.
__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
__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
__DEVICE__ int __double2int_rd(double __a)
__device__ double __nv_tan(double __a)
__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)
__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)
__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
__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.
__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)
__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
__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)