xref: /freebsd/contrib/llvm-project/clang/lib/Headers/__clang_cuda_intrinsics.h (revision 06c3fb2749bda94cb5201f81ffdb8fa6c3161b2e)
1 /*===--- __clang_cuda_intrinsics.h - Device-side CUDA intrinsic wrappers ---===
2  *
3  * Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4  * See https://llvm.org/LICENSE.txt for license information.
5  * SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6  *
7  *===-----------------------------------------------------------------------===
8  */
9 #ifndef __CLANG_CUDA_INTRINSICS_H__
10 #define __CLANG_CUDA_INTRINSICS_H__
11 #ifndef __CUDA__
12 #error "This file is for CUDA compilation only."
13 #endif
14 
15 // sm_30 intrinsics: __shfl_{up,down,xor}.
16 
17 #define __SM_30_INTRINSICS_H__
18 #define __SM_30_INTRINSICS_HPP__
19 
20 #if !defined(__CUDA_ARCH__) || __CUDA_ARCH__ >= 300
21 
22 #pragma push_macro("__MAKE_SHUFFLES")
23 #define __MAKE_SHUFFLES(__FnName, __IntIntrinsic, __FloatIntrinsic, __Mask,    \
24                         __Type)                                                \
25   inline __device__ int __FnName(int __val, __Type __offset,                   \
26                                  int __width = warpSize) {                     \
27     return __IntIntrinsic(__val, __offset,                                     \
28                           ((warpSize - __width) << 8) | (__Mask));             \
29   }                                                                            \
30   inline __device__ float __FnName(float __val, __Type __offset,               \
31                                    int __width = warpSize) {                   \
32     return __FloatIntrinsic(__val, __offset,                                   \
33                             ((warpSize - __width) << 8) | (__Mask));           \
34   }                                                                            \
35   inline __device__ unsigned int __FnName(unsigned int __val, __Type __offset, \
36                                           int __width = warpSize) {            \
37     return static_cast<unsigned int>(                                          \
38         ::__FnName(static_cast<int>(__val), __offset, __width));               \
39   }                                                                            \
40   inline __device__ long long __FnName(long long __val, __Type __offset,       \
41                                        int __width = warpSize) {               \
42     struct __Bits {                                                            \
43       int __a, __b;                                                            \
44     };                                                                         \
45     _Static_assert(sizeof(__val) == sizeof(__Bits));                           \
46     _Static_assert(sizeof(__Bits) == 2 * sizeof(int));                         \
47     __Bits __tmp;                                                              \
48     memcpy(&__tmp, &__val, sizeof(__val));                                \
49     __tmp.__a = ::__FnName(__tmp.__a, __offset, __width);                      \
50     __tmp.__b = ::__FnName(__tmp.__b, __offset, __width);                      \
51     long long __ret;                                                           \
52     memcpy(&__ret, &__tmp, sizeof(__tmp));                                     \
53     return __ret;                                                              \
54   }                                                                            \
55   inline __device__ long __FnName(long __val, __Type __offset,                 \
56                                   int __width = warpSize) {                    \
57     _Static_assert(sizeof(long) == sizeof(long long) ||                        \
58                    sizeof(long) == sizeof(int));                               \
59     if (sizeof(long) == sizeof(long long)) {                                   \
60       return static_cast<long>(                                                \
61           ::__FnName(static_cast<long long>(__val), __offset, __width));       \
62     } else if (sizeof(long) == sizeof(int)) {                                  \
63       return static_cast<long>(                                                \
64           ::__FnName(static_cast<int>(__val), __offset, __width));             \
65     }                                                                          \
66   }                                                                            \
67   inline __device__ unsigned long __FnName(                                    \
68       unsigned long __val, __Type __offset, int __width = warpSize) {          \
69     return static_cast<unsigned long>(                                         \
70         ::__FnName(static_cast<long>(__val), __offset, __width));              \
71   }                                                                            \
72   inline __device__ unsigned long long __FnName(                               \
73       unsigned long long __val, __Type __offset, int __width = warpSize) {     \
74     return static_cast<unsigned long long>(                                    \
75         ::__FnName(static_cast<long long>(__val), __offset, __width));         \
76   }                                                                            \
77   inline __device__ double __FnName(double __val, __Type __offset,             \
78                                     int __width = warpSize) {                  \
79     long long __tmp;                                                           \
80     _Static_assert(sizeof(__tmp) == sizeof(__val));                            \
81     memcpy(&__tmp, &__val, sizeof(__val));                                     \
82     __tmp = ::__FnName(__tmp, __offset, __width);                              \
83     double __ret;                                                              \
84     memcpy(&__ret, &__tmp, sizeof(__ret));                                     \
85     return __ret;                                                              \
86   }
87 
88 __MAKE_SHUFFLES(__shfl, __nvvm_shfl_idx_i32, __nvvm_shfl_idx_f32, 0x1f, int);
89 // We use 0 rather than 31 as our mask, because shfl.up applies to lanes >=
90 // maxLane.
91 __MAKE_SHUFFLES(__shfl_up, __nvvm_shfl_up_i32, __nvvm_shfl_up_f32, 0,
92                 unsigned int);
93 __MAKE_SHUFFLES(__shfl_down, __nvvm_shfl_down_i32, __nvvm_shfl_down_f32, 0x1f,
94                 unsigned int);
95 __MAKE_SHUFFLES(__shfl_xor, __nvvm_shfl_bfly_i32, __nvvm_shfl_bfly_f32, 0x1f,
96                 int);
97 #pragma pop_macro("__MAKE_SHUFFLES")
98 
99 #endif // !defined(__CUDA_ARCH__) || __CUDA_ARCH__ >= 300
100 
101 #if CUDA_VERSION >= 9000
102 #if (!defined(__CUDA_ARCH__) || __CUDA_ARCH__ >= 300)
103 // __shfl_sync_* variants available in CUDA-9
104 #pragma push_macro("__MAKE_SYNC_SHUFFLES")
105 #define __MAKE_SYNC_SHUFFLES(__FnName, __IntIntrinsic, __FloatIntrinsic,       \
106                              __Mask, __Type)                                   \
107   inline __device__ int __FnName(unsigned int __mask, int __val,               \
108                                  __Type __offset, int __width = warpSize) {    \
109     return __IntIntrinsic(__mask, __val, __offset,                             \
110                           ((warpSize - __width) << 8) | (__Mask));             \
111   }                                                                            \
112   inline __device__ float __FnName(unsigned int __mask, float __val,           \
113                                    __Type __offset, int __width = warpSize) {  \
114     return __FloatIntrinsic(__mask, __val, __offset,                           \
115                             ((warpSize - __width) << 8) | (__Mask));           \
116   }                                                                            \
117   inline __device__ unsigned int __FnName(unsigned int __mask,                 \
118                                           unsigned int __val, __Type __offset, \
119                                           int __width = warpSize) {            \
120     return static_cast<unsigned int>(                                          \
121         ::__FnName(__mask, static_cast<int>(__val), __offset, __width));       \
122   }                                                                            \
123   inline __device__ long long __FnName(unsigned int __mask, long long __val,   \
124                                        __Type __offset,                        \
125                                        int __width = warpSize) {               \
126     struct __Bits {                                                            \
127       int __a, __b;                                                            \
128     };                                                                         \
129     _Static_assert(sizeof(__val) == sizeof(__Bits));                           \
130     _Static_assert(sizeof(__Bits) == 2 * sizeof(int));                         \
131     __Bits __tmp;                                                              \
132     memcpy(&__tmp, &__val, sizeof(__val));                                     \
133     __tmp.__a = ::__FnName(__mask, __tmp.__a, __offset, __width);              \
134     __tmp.__b = ::__FnName(__mask, __tmp.__b, __offset, __width);              \
135     long long __ret;                                                           \
136     memcpy(&__ret, &__tmp, sizeof(__tmp));                                     \
137     return __ret;                                                              \
138   }                                                                            \
139   inline __device__ unsigned long long __FnName(                               \
140       unsigned int __mask, unsigned long long __val, __Type __offset,          \
141       int __width = warpSize) {                                                \
142     return static_cast<unsigned long long>(                                    \
143         ::__FnName(__mask, static_cast<long long>(__val), __offset, __width)); \
144   }                                                                            \
145   inline __device__ long __FnName(unsigned int __mask, long __val,             \
146                                   __Type __offset, int __width = warpSize) {   \
147     _Static_assert(sizeof(long) == sizeof(long long) ||                        \
148                    sizeof(long) == sizeof(int));                               \
149     if (sizeof(long) == sizeof(long long)) {                                   \
150       return static_cast<long>(::__FnName(                                     \
151           __mask, static_cast<long long>(__val), __offset, __width));          \
152     } else if (sizeof(long) == sizeof(int)) {                                  \
153       return static_cast<long>(                                                \
154           ::__FnName(__mask, static_cast<int>(__val), __offset, __width));     \
155     }                                                                          \
156   }                                                                            \
157   inline __device__ unsigned long __FnName(                                    \
158       unsigned int __mask, unsigned long __val, __Type __offset,               \
159       int __width = warpSize) {                                                \
160     return static_cast<unsigned long>(                                         \
161         ::__FnName(__mask, static_cast<long>(__val), __offset, __width));      \
162   }                                                                            \
163   inline __device__ double __FnName(unsigned int __mask, double __val,         \
164                                     __Type __offset, int __width = warpSize) { \
165     long long __tmp;                                                           \
166     _Static_assert(sizeof(__tmp) == sizeof(__val));                            \
167     memcpy(&__tmp, &__val, sizeof(__val));                                     \
168     __tmp = ::__FnName(__mask, __tmp, __offset, __width);                      \
169     double __ret;                                                              \
170     memcpy(&__ret, &__tmp, sizeof(__ret));                                     \
171     return __ret;                                                              \
172   }
173 __MAKE_SYNC_SHUFFLES(__shfl_sync, __nvvm_shfl_sync_idx_i32,
174                      __nvvm_shfl_sync_idx_f32, 0x1f, int);
175 // We use 0 rather than 31 as our mask, because shfl.up applies to lanes >=
176 // maxLane.
177 __MAKE_SYNC_SHUFFLES(__shfl_up_sync, __nvvm_shfl_sync_up_i32,
178                      __nvvm_shfl_sync_up_f32, 0, unsigned int);
179 __MAKE_SYNC_SHUFFLES(__shfl_down_sync, __nvvm_shfl_sync_down_i32,
180                      __nvvm_shfl_sync_down_f32, 0x1f, unsigned int);
181 __MAKE_SYNC_SHUFFLES(__shfl_xor_sync, __nvvm_shfl_sync_bfly_i32,
182                      __nvvm_shfl_sync_bfly_f32, 0x1f, int);
183 #pragma pop_macro("__MAKE_SYNC_SHUFFLES")
184 
185 inline __device__ void __syncwarp(unsigned int mask = 0xffffffff) {
186   return __nvvm_bar_warp_sync(mask);
187 }
188 
189 inline __device__ void __barrier_sync(unsigned int id) {
190   __nvvm_barrier_sync(id);
191 }
192 
193 inline __device__ void __barrier_sync_count(unsigned int id,
194                                             unsigned int count) {
195   __nvvm_barrier_sync_cnt(id, count);
196 }
197 
198 inline __device__ int __all_sync(unsigned int mask, int pred) {
199   return __nvvm_vote_all_sync(mask, pred);
200 }
201 
202 inline __device__ int __any_sync(unsigned int mask, int pred) {
203   return __nvvm_vote_any_sync(mask, pred);
204 }
205 
206 inline __device__ int __uni_sync(unsigned int mask, int pred) {
207   return __nvvm_vote_uni_sync(mask, pred);
208 }
209 
210 inline __device__ unsigned int __ballot_sync(unsigned int mask, int pred) {
211   return __nvvm_vote_ballot_sync(mask, pred);
212 }
213 
214 inline __device__ unsigned int __activemask() {
215 #if CUDA_VERSION < 9020
216   return __nvvm_vote_ballot(1);
217 #else
218   unsigned int mask;
219   asm volatile("activemask.b32 %0;" : "=r"(mask));
220   return mask;
221 #endif
222 }
223 
224 inline __device__ unsigned int __fns(unsigned mask, unsigned base, int offset) {
225   return __nvvm_fns(mask, base, offset);
226 }
227 
228 #endif // !defined(__CUDA_ARCH__) || __CUDA_ARCH__ >= 300
229 
230 // Define __match* builtins CUDA-9 headers expect to see.
231 #if !defined(__CUDA_ARCH__) || __CUDA_ARCH__ >= 700
232 inline __device__ unsigned int __match32_any_sync(unsigned int mask,
233                                                   unsigned int value) {
234   return __nvvm_match_any_sync_i32(mask, value);
235 }
236 
237 inline __device__ unsigned int
238 __match64_any_sync(unsigned int mask, unsigned long long value) {
239   return __nvvm_match_any_sync_i64(mask, value);
240 }
241 
242 inline __device__ unsigned int
243 __match32_all_sync(unsigned int mask, unsigned int value, int *pred) {
244   return __nvvm_match_all_sync_i32p(mask, value, pred);
245 }
246 
247 inline __device__ unsigned int
248 __match64_all_sync(unsigned int mask, unsigned long long value, int *pred) {
249   return __nvvm_match_all_sync_i64p(mask, value, pred);
250 }
251 #include "crt/sm_70_rt.hpp"
252 
253 #endif // !defined(__CUDA_ARCH__) || __CUDA_ARCH__ >= 700
254 #endif // __CUDA_VERSION >= 9000
255 
256 // sm_32 intrinsics: __ldg and __funnelshift_{l,lc,r,rc}.
257 
258 // Prevent the vanilla sm_32 intrinsics header from being included.
259 #define __SM_32_INTRINSICS_H__
260 #define __SM_32_INTRINSICS_HPP__
261 
262 #if !defined(__CUDA_ARCH__) || __CUDA_ARCH__ >= 320
263 
264 inline __device__ char __ldg(const char *ptr) { return __nvvm_ldg_c(ptr); }
265 inline __device__ short __ldg(const short *ptr) { return __nvvm_ldg_s(ptr); }
266 inline __device__ int __ldg(const int *ptr) { return __nvvm_ldg_i(ptr); }
267 inline __device__ long __ldg(const long *ptr) { return __nvvm_ldg_l(ptr); }
268 inline __device__ long long __ldg(const long long *ptr) {
269   return __nvvm_ldg_ll(ptr);
270 }
271 inline __device__ unsigned char __ldg(const unsigned char *ptr) {
272   return __nvvm_ldg_uc(ptr);
273 }
274 inline __device__ signed char __ldg(const signed char *ptr) {
275   return __nvvm_ldg_uc((const unsigned char *)ptr);
276 }
277 inline __device__ unsigned short __ldg(const unsigned short *ptr) {
278   return __nvvm_ldg_us(ptr);
279 }
280 inline __device__ unsigned int __ldg(const unsigned int *ptr) {
281   return __nvvm_ldg_ui(ptr);
282 }
283 inline __device__ unsigned long __ldg(const unsigned long *ptr) {
284   return __nvvm_ldg_ul(ptr);
285 }
286 inline __device__ unsigned long long __ldg(const unsigned long long *ptr) {
287   return __nvvm_ldg_ull(ptr);
288 }
289 inline __device__ float __ldg(const float *ptr) { return __nvvm_ldg_f(ptr); }
290 inline __device__ double __ldg(const double *ptr) { return __nvvm_ldg_d(ptr); }
291 
292 inline __device__ char2 __ldg(const char2 *ptr) {
293   typedef char c2 __attribute__((ext_vector_type(2)));
294   // We can assume that ptr is aligned at least to char2's alignment, but the
295   // load will assume that ptr is aligned to char2's alignment.  This is only
296   // safe if alignof(c2) <= alignof(char2).
297   c2 rv = __nvvm_ldg_c2(reinterpret_cast<const c2 *>(ptr));
298   char2 ret;
299   ret.x = rv[0];
300   ret.y = rv[1];
301   return ret;
302 }
303 inline __device__ char4 __ldg(const char4 *ptr) {
304   typedef char c4 __attribute__((ext_vector_type(4)));
305   c4 rv = __nvvm_ldg_c4(reinterpret_cast<const c4 *>(ptr));
306   char4 ret;
307   ret.x = rv[0];
308   ret.y = rv[1];
309   ret.z = rv[2];
310   ret.w = rv[3];
311   return ret;
312 }
313 inline __device__ short2 __ldg(const short2 *ptr) {
314   typedef short s2 __attribute__((ext_vector_type(2)));
315   s2 rv = __nvvm_ldg_s2(reinterpret_cast<const s2 *>(ptr));
316   short2 ret;
317   ret.x = rv[0];
318   ret.y = rv[1];
319   return ret;
320 }
321 inline __device__ short4 __ldg(const short4 *ptr) {
322   typedef short s4 __attribute__((ext_vector_type(4)));
323   s4 rv = __nvvm_ldg_s4(reinterpret_cast<const s4 *>(ptr));
324   short4 ret;
325   ret.x = rv[0];
326   ret.y = rv[1];
327   ret.z = rv[2];
328   ret.w = rv[3];
329   return ret;
330 }
331 inline __device__ int2 __ldg(const int2 *ptr) {
332   typedef int i2 __attribute__((ext_vector_type(2)));
333   i2 rv = __nvvm_ldg_i2(reinterpret_cast<const i2 *>(ptr));
334   int2 ret;
335   ret.x = rv[0];
336   ret.y = rv[1];
337   return ret;
338 }
339 inline __device__ int4 __ldg(const int4 *ptr) {
340   typedef int i4 __attribute__((ext_vector_type(4)));
341   i4 rv = __nvvm_ldg_i4(reinterpret_cast<const i4 *>(ptr));
342   int4 ret;
343   ret.x = rv[0];
344   ret.y = rv[1];
345   ret.z = rv[2];
346   ret.w = rv[3];
347   return ret;
348 }
349 inline __device__ longlong2 __ldg(const longlong2 *ptr) {
350   typedef long long ll2 __attribute__((ext_vector_type(2)));
351   ll2 rv = __nvvm_ldg_ll2(reinterpret_cast<const ll2 *>(ptr));
352   longlong2 ret;
353   ret.x = rv[0];
354   ret.y = rv[1];
355   return ret;
356 }
357 
358 inline __device__ uchar2 __ldg(const uchar2 *ptr) {
359   typedef unsigned char uc2 __attribute__((ext_vector_type(2)));
360   uc2 rv = __nvvm_ldg_uc2(reinterpret_cast<const uc2 *>(ptr));
361   uchar2 ret;
362   ret.x = rv[0];
363   ret.y = rv[1];
364   return ret;
365 }
366 inline __device__ uchar4 __ldg(const uchar4 *ptr) {
367   typedef unsigned char uc4 __attribute__((ext_vector_type(4)));
368   uc4 rv = __nvvm_ldg_uc4(reinterpret_cast<const uc4 *>(ptr));
369   uchar4 ret;
370   ret.x = rv[0];
371   ret.y = rv[1];
372   ret.z = rv[2];
373   ret.w = rv[3];
374   return ret;
375 }
376 inline __device__ ushort2 __ldg(const ushort2 *ptr) {
377   typedef unsigned short us2 __attribute__((ext_vector_type(2)));
378   us2 rv = __nvvm_ldg_us2(reinterpret_cast<const us2 *>(ptr));
379   ushort2 ret;
380   ret.x = rv[0];
381   ret.y = rv[1];
382   return ret;
383 }
384 inline __device__ ushort4 __ldg(const ushort4 *ptr) {
385   typedef unsigned short us4 __attribute__((ext_vector_type(4)));
386   us4 rv = __nvvm_ldg_us4(reinterpret_cast<const us4 *>(ptr));
387   ushort4 ret;
388   ret.x = rv[0];
389   ret.y = rv[1];
390   ret.z = rv[2];
391   ret.w = rv[3];
392   return ret;
393 }
394 inline __device__ uint2 __ldg(const uint2 *ptr) {
395   typedef unsigned int ui2 __attribute__((ext_vector_type(2)));
396   ui2 rv = __nvvm_ldg_ui2(reinterpret_cast<const ui2 *>(ptr));
397   uint2 ret;
398   ret.x = rv[0];
399   ret.y = rv[1];
400   return ret;
401 }
402 inline __device__ uint4 __ldg(const uint4 *ptr) {
403   typedef unsigned int ui4 __attribute__((ext_vector_type(4)));
404   ui4 rv = __nvvm_ldg_ui4(reinterpret_cast<const ui4 *>(ptr));
405   uint4 ret;
406   ret.x = rv[0];
407   ret.y = rv[1];
408   ret.z = rv[2];
409   ret.w = rv[3];
410   return ret;
411 }
412 inline __device__ ulonglong2 __ldg(const ulonglong2 *ptr) {
413   typedef unsigned long long ull2 __attribute__((ext_vector_type(2)));
414   ull2 rv = __nvvm_ldg_ull2(reinterpret_cast<const ull2 *>(ptr));
415   ulonglong2 ret;
416   ret.x = rv[0];
417   ret.y = rv[1];
418   return ret;
419 }
420 
421 inline __device__ float2 __ldg(const float2 *ptr) {
422   typedef float f2 __attribute__((ext_vector_type(2)));
423   f2 rv = __nvvm_ldg_f2(reinterpret_cast<const f2 *>(ptr));
424   float2 ret;
425   ret.x = rv[0];
426   ret.y = rv[1];
427   return ret;
428 }
429 inline __device__ float4 __ldg(const float4 *ptr) {
430   typedef float f4 __attribute__((ext_vector_type(4)));
431   f4 rv = __nvvm_ldg_f4(reinterpret_cast<const f4 *>(ptr));
432   float4 ret;
433   ret.x = rv[0];
434   ret.y = rv[1];
435   ret.z = rv[2];
436   ret.w = rv[3];
437   return ret;
438 }
439 inline __device__ double2 __ldg(const double2 *ptr) {
440   typedef double d2 __attribute__((ext_vector_type(2)));
441   d2 rv = __nvvm_ldg_d2(reinterpret_cast<const d2 *>(ptr));
442   double2 ret;
443   ret.x = rv[0];
444   ret.y = rv[1];
445   return ret;
446 }
447 
448 // TODO: Implement these as intrinsics, so the backend can work its magic on
449 // these.  Alternatively, we could implement these as plain C and try to get
450 // llvm to recognize the relevant patterns.
451 inline __device__ unsigned __funnelshift_l(unsigned low32, unsigned high32,
452                                            unsigned shiftWidth) {
453   unsigned result;
454   asm("shf.l.wrap.b32 %0, %1, %2, %3;"
455       : "=r"(result)
456       : "r"(low32), "r"(high32), "r"(shiftWidth));
457   return result;
458 }
459 inline __device__ unsigned __funnelshift_lc(unsigned low32, unsigned high32,
460                                             unsigned shiftWidth) {
461   unsigned result;
462   asm("shf.l.clamp.b32 %0, %1, %2, %3;"
463       : "=r"(result)
464       : "r"(low32), "r"(high32), "r"(shiftWidth));
465   return result;
466 }
467 inline __device__ unsigned __funnelshift_r(unsigned low32, unsigned high32,
468                                            unsigned shiftWidth) {
469   unsigned result;
470   asm("shf.r.wrap.b32 %0, %1, %2, %3;"
471       : "=r"(result)
472       : "r"(low32), "r"(high32), "r"(shiftWidth));
473   return result;
474 }
475 inline __device__ unsigned __funnelshift_rc(unsigned low32, unsigned high32,
476                                             unsigned shiftWidth) {
477   unsigned ret;
478   asm("shf.r.clamp.b32 %0, %1, %2, %3;"
479       : "=r"(ret)
480       : "r"(low32), "r"(high32), "r"(shiftWidth));
481   return ret;
482 }
483 
484 #endif // !defined(__CUDA_ARCH__) || __CUDA_ARCH__ >= 320
485 
486 #if CUDA_VERSION >= 11000
487 extern "C" {
488 __device__ inline size_t __nv_cvta_generic_to_global_impl(const void *__ptr) {
489   return (size_t)(void __attribute__((address_space(1))) *)__ptr;
490 }
491 __device__ inline size_t __nv_cvta_generic_to_shared_impl(const void *__ptr) {
492   return (size_t)(void __attribute__((address_space(3))) *)__ptr;
493 }
494 __device__ inline size_t __nv_cvta_generic_to_constant_impl(const void *__ptr) {
495   return (size_t)(void __attribute__((address_space(4))) *)__ptr;
496 }
497 __device__ inline size_t __nv_cvta_generic_to_local_impl(const void *__ptr) {
498   return (size_t)(void __attribute__((address_space(5))) *)__ptr;
499 }
500 __device__ inline void *__nv_cvta_global_to_generic_impl(size_t __ptr) {
501   return (void *)(void __attribute__((address_space(1))) *)__ptr;
502 }
503 __device__ inline void *__nv_cvta_shared_to_generic_impl(size_t __ptr) {
504   return (void *)(void __attribute__((address_space(3))) *)__ptr;
505 }
506 __device__ inline void *__nv_cvta_constant_to_generic_impl(size_t __ptr) {
507   return (void *)(void __attribute__((address_space(4))) *)__ptr;
508 }
509 __device__ inline void *__nv_cvta_local_to_generic_impl(size_t __ptr) {
510   return (void *)(void __attribute__((address_space(5))) *)__ptr;
511 }
512 __device__ inline cuuint32_t __nvvm_get_smem_pointer(void *__ptr) {
513   return __nv_cvta_generic_to_shared_impl(__ptr);
514 }
515 } // extern "C"
516 
517 #if !defined(__CUDA_ARCH__) || __CUDA_ARCH__ >= 800
518 __device__ inline unsigned __reduce_add_sync(unsigned __mask,
519                                              unsigned __value) {
520   return __nvvm_redux_sync_add(__mask, __value);
521 }
522 __device__ inline unsigned __reduce_min_sync(unsigned __mask,
523                                              unsigned __value) {
524   return __nvvm_redux_sync_umin(__mask, __value);
525 }
526 __device__ inline unsigned __reduce_max_sync(unsigned __mask,
527                                              unsigned __value) {
528   return __nvvm_redux_sync_umax(__mask, __value);
529 }
530 __device__ inline int __reduce_min_sync(unsigned __mask, int __value) {
531   return __nvvm_redux_sync_min(__mask, __value);
532 }
533 __device__ inline int __reduce_max_sync(unsigned __mask, int __value) {
534   return __nvvm_redux_sync_max(__mask, __value);
535 }
536 __device__ inline unsigned __reduce_or_sync(unsigned __mask, unsigned __value) {
537   return __nvvm_redux_sync_or(__mask, __value);
538 }
539 __device__ inline unsigned __reduce_and_sync(unsigned __mask,
540                                              unsigned __value) {
541   return __nvvm_redux_sync_and(__mask, __value);
542 }
543 __device__ inline unsigned __reduce_xor_sync(unsigned __mask,
544                                              unsigned __value) {
545   return __nvvm_redux_sync_xor(__mask, __value);
546 }
547 
548 __device__ inline void __nv_memcpy_async_shared_global_4(void *__dst,
549                                                          const void *__src,
550                                                          unsigned __src_size) {
551   __nvvm_cp_async_ca_shared_global_4(
552       (void __attribute__((address_space(3))) *)__dst,
553       (const void __attribute__((address_space(1))) *)__src, __src_size);
554 }
555 __device__ inline void __nv_memcpy_async_shared_global_8(void *__dst,
556                                                          const void *__src,
557                                                          unsigned __src_size) {
558   __nvvm_cp_async_ca_shared_global_8(
559       (void __attribute__((address_space(3))) *)__dst,
560       (const void __attribute__((address_space(1))) *)__src, __src_size);
561 }
562 __device__ inline void __nv_memcpy_async_shared_global_16(void *__dst,
563                                                           const void *__src,
564                                                           unsigned __src_size) {
565   __nvvm_cp_async_ca_shared_global_16(
566       (void __attribute__((address_space(3))) *)__dst,
567       (const void __attribute__((address_space(1))) *)__src, __src_size);
568 }
569 
570 __device__ inline void *
571 __nv_associate_access_property(const void *__ptr, unsigned long long __prop) {
572   // TODO: it appears to provide compiler with some sort of a hint. We do not
573   // know what exactly it is supposed to do. However, CUDA headers suggest that
574   // just passing through __ptr should not affect correctness. They do so on
575   // pre-sm80 GPUs where this builtin is not available.
576   return (void*)__ptr;
577 }
578 #endif // !defined(__CUDA_ARCH__) || __CUDA_ARCH__ >= 800
579 
580 #if !defined(__CUDA_ARCH__) || __CUDA_ARCH__ >= 900
581 __device__ inline unsigned __isCtaShared(const void *ptr) {
582   return __isShared(ptr);
583 }
584 
585 __device__ inline unsigned __isClusterShared(const void *__ptr) {
586   return __nvvm_isspacep_shared_cluster(__ptr);
587 }
588 
589 __device__ inline void *__cluster_map_shared_rank(const void *__ptr,
590                                                   unsigned __rank) {
591   return __nvvm_mapa((void *)__ptr, __rank);
592 }
593 
594 __device__ inline unsigned __cluster_query_shared_rank(const void *__ptr) {
595   return __nvvm_getctarank((void *)__ptr);
596 }
597 
598 __device__ inline uint2
599 __cluster_map_shared_multicast(const void *__ptr,
600                                unsigned int __cluster_cta_mask) {
601   return make_uint2((unsigned)__cvta_generic_to_shared(__ptr),
602                     __cluster_cta_mask);
603 }
604 
605 __device__ inline unsigned __clusterDimIsSpecified() {
606   return __nvvm_is_explicit_cluster();
607 }
608 
609 __device__ inline dim3 __clusterDim() {
610   return dim3(__nvvm_read_ptx_sreg_cluster_nctaid_x(),
611               __nvvm_read_ptx_sreg_cluster_nctaid_y(),
612               __nvvm_read_ptx_sreg_cluster_nctaid_z());
613 }
614 
615 __device__ inline dim3 __clusterRelativeBlockIdx() {
616   return dim3(__nvvm_read_ptx_sreg_cluster_ctaid_x(),
617               __nvvm_read_ptx_sreg_cluster_ctaid_y(),
618               __nvvm_read_ptx_sreg_cluster_ctaid_z());
619 }
620 
621 __device__ inline dim3 __clusterGridDimInClusters() {
622   return dim3(__nvvm_read_ptx_sreg_nclusterid_x(),
623               __nvvm_read_ptx_sreg_nclusterid_y(),
624               __nvvm_read_ptx_sreg_nclusterid_z());
625 }
626 
627 __device__ inline dim3 __clusterIdx() {
628   return dim3(__nvvm_read_ptx_sreg_clusterid_x(),
629               __nvvm_read_ptx_sreg_clusterid_y(),
630               __nvvm_read_ptx_sreg_clusterid_z());
631 }
632 
633 __device__ inline unsigned __clusterRelativeBlockRank() {
634   return __nvvm_read_ptx_sreg_cluster_ctarank();
635 }
636 
637 __device__ inline unsigned __clusterSizeInBlocks() {
638   return __nvvm_read_ptx_sreg_cluster_nctarank();
639 }
640 
641 __device__ inline void __cluster_barrier_arrive() {
642   __nvvm_barrier_cluster_arrive();
643 }
644 
645 __device__ inline void __cluster_barrier_arrive_relaxed() {
646   __nvvm_barrier_cluster_arrive_relaxed();
647 }
648 
649 __device__ inline void __cluster_barrier_wait() {
650   __nvvm_barrier_cluster_wait();
651 }
652 
653 __device__ inline void __threadfence_cluster() { __nvvm_fence_sc_cluster(); }
654 
655 __device__ inline float2 atomicAdd(float2 *__ptr, float2 __val) {
656   float2 __ret;
657   __asm__("atom.add.v2.f32         {%0, %1}, [%2], {%3, %4};"
658           : "=f"(__ret.x), "=f"(__ret.y)
659           : "l"(__ptr), "f"(__val.x), "f"(__val.y));
660   return __ret;
661 }
662 
663 __device__ inline float2 atomicAdd_block(float2 *__ptr, float2 __val) {
664   float2 __ret;
665   __asm__("atom.cta.add.v2.f32         {%0, %1}, [%2], {%3, %4};"
666           : "=f"(__ret.x), "=f"(__ret.y)
667           : "l"(__ptr), "f"(__val.x), "f"(__val.y));
668   return __ret;
669 }
670 
671 __device__ inline float2 atomicAdd_system(float2 *__ptr, float2 __val) {
672   float2 __ret;
673   __asm__("atom.sys.add.v2.f32         {%0, %1}, [%2], {%3, %4};"
674           : "=f"(__ret.x), "=f"(__ret.y)
675           : "l"(__ptr), "f"(__val.x), "f"(__val.y));
676   return __ret;
677 }
678 
679 __device__ inline float4 atomicAdd(float4 *__ptr, float4 __val) {
680   float4 __ret;
681   __asm__("atom.add.v4.f32         {%0, %1, %2, %3}, [%4], {%5, %6, %7, %8};"
682           : "=f"(__ret.x), "=f"(__ret.y), "=f"(__ret.z), "=f"(__ret.w)
683           : "l"(__ptr), "f"(__val.x), "f"(__val.y), "f"(__val.z), "f"(__val.w));
684   return __ret;
685 }
686 
687 __device__ inline float4 atomicAdd_block(float4 *__ptr, float4 __val) {
688   float4 __ret;
689   __asm__(
690       "atom.cta.add.v4.f32         {%0, %1, %2, %3}, [%4], {%5, %6, %7, %8};"
691       : "=f"(__ret.x), "=f"(__ret.y), "=f"(__ret.z), "=f"(__ret.w)
692       : "l"(__ptr), "f"(__val.x), "f"(__val.y), "f"(__val.z), "f"(__val.w));
693   return __ret;
694 }
695 
696 __device__ inline float4 atomicAdd_system(float4 *__ptr, float4 __val) {
697   float4 __ret;
698   __asm__(
699       "atom.sys.add.v4.f32         {%0, %1, %2, %3}, [%4], {%5, %6, %7, %8};"
700       : "=f"(__ret.x), "=f"(__ret.y), "=f"(__ret.z), "=f"(__ret.w)
701       : "l"(__ptr), "f"(__val.x), "f"(__val.y), "f"(__val.z), "f"(__val.w)
702       :);
703   return __ret;
704 }
705 
706 #endif // !defined(__CUDA_ARCH__) || __CUDA_ARCH__ >= 900
707 #endif // CUDA_VERSION >= 11000
708 
709 #endif // defined(__CLANG_CUDA_INTRINSICS_H__)
710