• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*===--- __clang_cuda_intrinsics.h - Device-side CUDA intrinsic wrappers ---===
2  *
3  * Permission is hereby granted, free of charge, to any person obtaining a copy
4  * of this software and associated documentation files (the "Software"), to deal
5  * in the Software without restriction, including without limitation the rights
6  * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
7  * copies of the Software, and to permit persons to whom the Software is
8  * furnished to do so, subject to the following conditions:
9  *
10  * The above copyright notice and this permission notice shall be included in
11  * all copies or substantial portions of the Software.
12  *
13  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
14  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
15  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
16  * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
17  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
18  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
19  * THE SOFTWARE.
20  *
21  *===-----------------------------------------------------------------------===
22  */
23 #ifndef __CLANG_CUDA_INTRINSICS_H__
24 #define __CLANG_CUDA_INTRINSICS_H__
25 #ifndef __CUDA__
26 #error "This file is for CUDA compilation only."
27 #endif
28 
29 // sm_30 intrinsics: __shfl_{up,down,xor}.
30 
31 #define __SM_30_INTRINSICS_H__
32 #define __SM_30_INTRINSICS_HPP__
33 
34 #if !defined(__CUDA_ARCH__) || __CUDA_ARCH__ >= 300
35 
36 #pragma push_macro("__MAKE_SHUFFLES")
37 #define __MAKE_SHUFFLES(__FnName, __IntIntrinsic, __FloatIntrinsic, __Mask)    \
38   inline __device__ int __FnName(int __in, int __offset,                       \
39                                  int __width = warpSize) {                     \
40     return __IntIntrinsic(__in, __offset,                                      \
41                           ((warpSize - __width) << 8) | (__Mask));             \
42   }                                                                            \
43   inline __device__ float __FnName(float __in, int __offset,                   \
44                                    int __width = warpSize) {                   \
45     return __FloatIntrinsic(__in, __offset,                                    \
46                             ((warpSize - __width) << 8) | (__Mask));           \
47   }                                                                            \
48   inline __device__ unsigned int __FnName(unsigned int __in, int __offset,     \
49                                           int __width = warpSize) {            \
50     return static_cast<unsigned int>(                                          \
51         ::__FnName(static_cast<int>(__in), __offset, __width));                \
52   }                                                                            \
53   inline __device__ long long __FnName(long long __in, int __offset,           \
54                                        int __width = warpSize) {               \
55     struct __Bits {                                                            \
56       int __a, __b;                                                            \
57     };                                                                         \
58     _Static_assert(sizeof(__in) == sizeof(__Bits));                            \
59     _Static_assert(sizeof(__Bits) == 2 * sizeof(int));                         \
60     __Bits __tmp;                                                              \
61     memcpy(&__in, &__tmp, sizeof(__in));                                       \
62     __tmp.__a = ::__FnName(__tmp.__a, __offset, __width);                      \
63     __tmp.__b = ::__FnName(__tmp.__b, __offset, __width);                      \
64     long long __out;                                                           \
65     memcpy(&__out, &__tmp, sizeof(__tmp));                                     \
66     return __out;                                                              \
67   }                                                                            \
68   inline __device__ unsigned long long __FnName(                               \
69       unsigned long long __in, int __offset, int __width = warpSize) {         \
70     return static_cast<unsigned long long>(                                    \
71         ::__FnName(static_cast<unsigned long long>(__in), __offset, __width)); \
72   }                                                                            \
73   inline __device__ double __FnName(double __in, int __offset,                 \
74                                     int __width = warpSize) {                  \
75     long long __tmp;                                                           \
76     _Static_assert(sizeof(__tmp) == sizeof(__in));                             \
77     memcpy(&__tmp, &__in, sizeof(__in));                                       \
78     __tmp = ::__FnName(__tmp, __offset, __width);                              \
79     double __out;                                                              \
80     memcpy(&__out, &__tmp, sizeof(__out));                                     \
81     return __out;                                                              \
82   }
83 
84 __MAKE_SHUFFLES(__shfl, __nvvm_shfl_idx_i32, __nvvm_shfl_idx_f32, 0x1f);
85 // We use 0 rather than 31 as our mask, because shfl.up applies to lanes >=
86 // maxLane.
87 __MAKE_SHUFFLES(__shfl_up, __nvvm_shfl_up_i32, __nvvm_shfl_up_f32, 0);
88 __MAKE_SHUFFLES(__shfl_down, __nvvm_shfl_down_i32, __nvvm_shfl_down_f32, 0x1f);
89 __MAKE_SHUFFLES(__shfl_xor, __nvvm_shfl_bfly_i32, __nvvm_shfl_bfly_f32, 0x1f);
90 
91 #pragma pop_macro("__MAKE_SHUFFLES")
92 
93 #endif // !defined(__CUDA_ARCH__) || __CUDA_ARCH__ >= 300
94 
95 // sm_32 intrinsics: __ldg and __funnelshift_{l,lc,r,rc}.
96 
97 // Prevent the vanilla sm_32 intrinsics header from being included.
98 #define __SM_32_INTRINSICS_H__
99 #define __SM_32_INTRINSICS_HPP__
100 
101 #if !defined(__CUDA_ARCH__) || __CUDA_ARCH__ >= 320
102 
__ldg(const char * ptr)103 inline __device__ char __ldg(const char *ptr) { return __nvvm_ldg_c(ptr); }
__ldg(const short * ptr)104 inline __device__ short __ldg(const short *ptr) { return __nvvm_ldg_s(ptr); }
__ldg(const int * ptr)105 inline __device__ int __ldg(const int *ptr) { return __nvvm_ldg_i(ptr); }
__ldg(const long * ptr)106 inline __device__ long __ldg(const long *ptr) { return __nvvm_ldg_l(ptr); }
__ldg(const long long * ptr)107 inline __device__ long long __ldg(const long long *ptr) {
108   return __nvvm_ldg_ll(ptr);
109 }
__ldg(const unsigned char * ptr)110 inline __device__ unsigned char __ldg(const unsigned char *ptr) {
111   return __nvvm_ldg_uc(ptr);
112 }
__ldg(const unsigned short * ptr)113 inline __device__ unsigned short __ldg(const unsigned short *ptr) {
114   return __nvvm_ldg_us(ptr);
115 }
__ldg(const unsigned int * ptr)116 inline __device__ unsigned int __ldg(const unsigned int *ptr) {
117   return __nvvm_ldg_ui(ptr);
118 }
__ldg(const unsigned long * ptr)119 inline __device__ unsigned long __ldg(const unsigned long *ptr) {
120   return __nvvm_ldg_ul(ptr);
121 }
__ldg(const unsigned long long * ptr)122 inline __device__ unsigned long long __ldg(const unsigned long long *ptr) {
123   return __nvvm_ldg_ull(ptr);
124 }
__ldg(const float * ptr)125 inline __device__ float __ldg(const float *ptr) { return __nvvm_ldg_f(ptr); }
__ldg(const double * ptr)126 inline __device__ double __ldg(const double *ptr) { return __nvvm_ldg_d(ptr); }
127 
__ldg(const char2 * ptr)128 inline __device__ char2 __ldg(const char2 *ptr) {
129   typedef char c2 __attribute__((ext_vector_type(2)));
130   // We can assume that ptr is aligned at least to char2's alignment, but the
131   // load will assume that ptr is aligned to char2's alignment.  This is only
132   // safe if alignof(c2) <= alignof(char2).
133   c2 rv = __nvvm_ldg_c2(reinterpret_cast<const c2 *>(ptr));
134   char2 ret;
135   ret.x = rv[0];
136   ret.y = rv[1];
137   return ret;
138 }
__ldg(const char4 * ptr)139 inline __device__ char4 __ldg(const char4 *ptr) {
140   typedef char c4 __attribute__((ext_vector_type(4)));
141   c4 rv = __nvvm_ldg_c4(reinterpret_cast<const c4 *>(ptr));
142   char4 ret;
143   ret.x = rv[0];
144   ret.y = rv[1];
145   ret.z = rv[2];
146   ret.w = rv[3];
147   return ret;
148 }
__ldg(const short2 * ptr)149 inline __device__ short2 __ldg(const short2 *ptr) {
150   typedef short s2 __attribute__((ext_vector_type(2)));
151   s2 rv = __nvvm_ldg_s2(reinterpret_cast<const s2 *>(ptr));
152   short2 ret;
153   ret.x = rv[0];
154   ret.y = rv[1];
155   return ret;
156 }
__ldg(const short4 * ptr)157 inline __device__ short4 __ldg(const short4 *ptr) {
158   typedef short s4 __attribute__((ext_vector_type(4)));
159   s4 rv = __nvvm_ldg_s4(reinterpret_cast<const s4 *>(ptr));
160   short4 ret;
161   ret.x = rv[0];
162   ret.y = rv[1];
163   ret.z = rv[2];
164   ret.w = rv[3];
165   return ret;
166 }
__ldg(const int2 * ptr)167 inline __device__ int2 __ldg(const int2 *ptr) {
168   typedef int i2 __attribute__((ext_vector_type(2)));
169   i2 rv = __nvvm_ldg_i2(reinterpret_cast<const i2 *>(ptr));
170   int2 ret;
171   ret.x = rv[0];
172   ret.y = rv[1];
173   return ret;
174 }
__ldg(const int4 * ptr)175 inline __device__ int4 __ldg(const int4 *ptr) {
176   typedef int i4 __attribute__((ext_vector_type(4)));
177   i4 rv = __nvvm_ldg_i4(reinterpret_cast<const i4 *>(ptr));
178   int4 ret;
179   ret.x = rv[0];
180   ret.y = rv[1];
181   ret.z = rv[2];
182   ret.w = rv[3];
183   return ret;
184 }
__ldg(const longlong2 * ptr)185 inline __device__ longlong2 __ldg(const longlong2 *ptr) {
186   typedef long long ll2 __attribute__((ext_vector_type(2)));
187   ll2 rv = __nvvm_ldg_ll2(reinterpret_cast<const ll2 *>(ptr));
188   longlong2 ret;
189   ret.x = rv[0];
190   ret.y = rv[1];
191   return ret;
192 }
193 
__ldg(const uchar2 * ptr)194 inline __device__ uchar2 __ldg(const uchar2 *ptr) {
195   typedef unsigned char uc2 __attribute__((ext_vector_type(2)));
196   uc2 rv = __nvvm_ldg_uc2(reinterpret_cast<const uc2 *>(ptr));
197   uchar2 ret;
198   ret.x = rv[0];
199   ret.y = rv[1];
200   return ret;
201 }
__ldg(const uchar4 * ptr)202 inline __device__ uchar4 __ldg(const uchar4 *ptr) {
203   typedef unsigned char uc4 __attribute__((ext_vector_type(4)));
204   uc4 rv = __nvvm_ldg_uc4(reinterpret_cast<const uc4 *>(ptr));
205   uchar4 ret;
206   ret.x = rv[0];
207   ret.y = rv[1];
208   ret.z = rv[2];
209   ret.w = rv[3];
210   return ret;
211 }
__ldg(const ushort2 * ptr)212 inline __device__ ushort2 __ldg(const ushort2 *ptr) {
213   typedef unsigned short us2 __attribute__((ext_vector_type(2)));
214   us2 rv = __nvvm_ldg_us2(reinterpret_cast<const us2 *>(ptr));
215   ushort2 ret;
216   ret.x = rv[0];
217   ret.y = rv[1];
218   return ret;
219 }
__ldg(const ushort4 * ptr)220 inline __device__ ushort4 __ldg(const ushort4 *ptr) {
221   typedef unsigned short us4 __attribute__((ext_vector_type(4)));
222   us4 rv = __nvvm_ldg_us4(reinterpret_cast<const us4 *>(ptr));
223   ushort4 ret;
224   ret.x = rv[0];
225   ret.y = rv[1];
226   ret.z = rv[2];
227   ret.w = rv[3];
228   return ret;
229 }
__ldg(const uint2 * ptr)230 inline __device__ uint2 __ldg(const uint2 *ptr) {
231   typedef unsigned int ui2 __attribute__((ext_vector_type(2)));
232   ui2 rv = __nvvm_ldg_ui2(reinterpret_cast<const ui2 *>(ptr));
233   uint2 ret;
234   ret.x = rv[0];
235   ret.y = rv[1];
236   return ret;
237 }
__ldg(const uint4 * ptr)238 inline __device__ uint4 __ldg(const uint4 *ptr) {
239   typedef unsigned int ui4 __attribute__((ext_vector_type(4)));
240   ui4 rv = __nvvm_ldg_ui4(reinterpret_cast<const ui4 *>(ptr));
241   uint4 ret;
242   ret.x = rv[0];
243   ret.y = rv[1];
244   ret.z = rv[2];
245   ret.w = rv[3];
246   return ret;
247 }
__ldg(const ulonglong2 * ptr)248 inline __device__ ulonglong2 __ldg(const ulonglong2 *ptr) {
249   typedef unsigned long long ull2 __attribute__((ext_vector_type(2)));
250   ull2 rv = __nvvm_ldg_ull2(reinterpret_cast<const ull2 *>(ptr));
251   ulonglong2 ret;
252   ret.x = rv[0];
253   ret.y = rv[1];
254   return ret;
255 }
256 
__ldg(const float2 * ptr)257 inline __device__ float2 __ldg(const float2 *ptr) {
258   typedef float f2 __attribute__((ext_vector_type(2)));
259   f2 rv = __nvvm_ldg_f2(reinterpret_cast<const f2 *>(ptr));
260   float2 ret;
261   ret.x = rv[0];
262   ret.y = rv[1];
263   return ret;
264 }
__ldg(const float4 * ptr)265 inline __device__ float4 __ldg(const float4 *ptr) {
266   typedef float f4 __attribute__((ext_vector_type(4)));
267   f4 rv = __nvvm_ldg_f4(reinterpret_cast<const f4 *>(ptr));
268   float4 ret;
269   ret.x = rv[0];
270   ret.y = rv[1];
271   ret.z = rv[2];
272   ret.w = rv[3];
273   return ret;
274 }
__ldg(const double2 * ptr)275 inline __device__ double2 __ldg(const double2 *ptr) {
276   typedef double d2 __attribute__((ext_vector_type(2)));
277   d2 rv = __nvvm_ldg_d2(reinterpret_cast<const d2 *>(ptr));
278   double2 ret;
279   ret.x = rv[0];
280   ret.y = rv[1];
281   return ret;
282 }
283 
284 // TODO: Implement these as intrinsics, so the backend can work its magic on
285 // these.  Alternatively, we could implement these as plain C and try to get
286 // llvm to recognize the relevant patterns.
__funnelshift_l(unsigned low32,unsigned high32,unsigned shiftWidth)287 inline __device__ unsigned __funnelshift_l(unsigned low32, unsigned high32,
288                                            unsigned shiftWidth) {
289   unsigned result;
290   asm("shf.l.wrap.b32 %0, %1, %2, %3;"
291       : "=r"(result)
292       : "r"(low32), "r"(high32), "r"(shiftWidth));
293   return result;
294 }
__funnelshift_lc(unsigned low32,unsigned high32,unsigned shiftWidth)295 inline __device__ unsigned __funnelshift_lc(unsigned low32, unsigned high32,
296                                             unsigned shiftWidth) {
297   unsigned result;
298   asm("shf.l.clamp.b32 %0, %1, %2, %3;"
299       : "=r"(result)
300       : "r"(low32), "r"(high32), "r"(shiftWidth));
301   return result;
302 }
__funnelshift_r(unsigned low32,unsigned high32,unsigned shiftWidth)303 inline __device__ unsigned __funnelshift_r(unsigned low32, unsigned high32,
304                                            unsigned shiftWidth) {
305   unsigned result;
306   asm("shf.r.wrap.b32 %0, %1, %2, %3;"
307       : "=r"(result)
308       : "r"(low32), "r"(high32), "r"(shiftWidth));
309   return result;
310 }
__funnelshift_rc(unsigned low32,unsigned high32,unsigned shiftWidth)311 inline __device__ unsigned __funnelshift_rc(unsigned low32, unsigned high32,
312                                             unsigned shiftWidth) {
313   unsigned ret;
314   asm("shf.r.clamp.b32 %0, %1, %2, %3;"
315       : "=r"(ret)
316       : "r"(low32), "r"(high32), "r"(shiftWidth));
317   return ret;
318 }
319 
320 #endif // !defined(__CUDA_ARCH__) || __CUDA_ARCH__ >= 320
321 
322 #endif // defined(__CLANG_CUDA_INTRINSICS_H__)
323