1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
|
/*************************************************************************
* Copyright (c) 2015-2019, NVIDIA CORPORATION. All rights reserved.
*
* See LICENSE.txt for license information
************************************************************************/
#ifndef NCCL_REDUCE_KERNEL_H_
#define NCCL_REDUCE_KERNEL_H_
#include "common_kernel.h"
#include <limits>
template<typename T>
struct FuncNull {
__device__ T operator()(const T x, const T y) const {
return 0;
}
};
template<typename T>
struct FuncSum {
__device__ T operator()(const T x, const T y) const {
return x + y;
}
};
template<typename T>
struct FuncProd {
__device__ T operator()(const T x, const T y) const {
return x * y;
}
};
template<typename T>
struct FuncMax {
__device__ T operator()(const T x, const T y) const {
return (x < y) ? y : x;
}
};
template<typename T>
struct FuncMin {
__device__ T operator()(const T x, const T y) const {
return (x < y) ? x : y;
}
};
#define MASK0 0x00ff00ff
#define MASK1 0xff00ff00
static __device__ uint32_t addChar4(const uint32_t x, const uint32_t y) {
/* This can be used both for signed and unsigned 8-bit addition */
const uint32_t x0 = x & MASK0;
const uint32_t x1 = x & MASK1;
const uint32_t y0 = y & MASK0;
const uint32_t y1 = y & MASK1;
const uint32_t r0 = (x0+y0);
const uint32_t r1 = (x1+y1);
return (r0 & MASK0) | (r1 & MASK1);
}
template<>
struct FuncSum<int8_t> {
__device__ uint32_t operator()(const uint32_t x, const uint32_t y) const {
#if (__CUDA_ARCH__ >= 300) && (__CUDA_ARCH__ < 500)
int32_t rv, z=0;
asm("vadd4.s32.s32.s32 %0, %1, %2, %3;" : "=r"(rv) : "r"(x), "r"(y), "r"(z));
return rv;
#else
return addChar4(x, y);
#endif
}
__device__ int8_t operator()(const int8_t x, const int8_t y) const {
return x+y;
}
};
template<>
struct FuncSum<uint8_t> {
__device__ uint32_t operator()(const uint32_t x, const uint32_t y) const {
#if (__CUDA_ARCH__ >= 300) && (__CUDA_ARCH__ < 500)
int32_t rv, z=0;
asm("vadd4.u32.u32.u32 %0, %1, %2, %3;" : "=r"(rv) : "r"(x), "r"(y), "r"(z));
return rv;
#else
return addChar4(x, y);
#endif
}
__device__ uint8_t operator()(const uint8_t x, const uint8_t y) const {
return x+y;
}
};
static __device__ uint32_t mulChar4(const uint32_t x, const uint32_t y) {
/* This can be used both for signed and unsigned 8-bit multiplication */
union converter { uint32_t storage; char4 a; };
converter cx, cy, cr;
cx.storage = x;
cy.storage = y;
cr.a.x = cx.a.x * cy.a.x;
cr.a.y = cx.a.y * cy.a.y;
cr.a.z = cx.a.z * cy.a.z;
cr.a.w = cx.a.w * cy.a.w;
return cr.storage;
}
template<>
struct FuncProd<int8_t> {
__device__ uint32_t operator()(const uint32_t x, const uint32_t y) const {
return mulChar4(x, y);
}
__device__ int8_t operator()(const int8_t x, const int8_t y) const {
return x*y;
}
};
template<>
struct FuncProd<uint8_t> {
__device__ uint32_t operator()(const uint32_t x, const uint32_t y) const {
return mulChar4(x, y);
}
__device__ uint8_t operator()(const uint8_t x, const uint8_t y) const {
return x*y;
}
};
template<>
struct FuncMax<int8_t> {
union converter { uint32_t storage; char4 a; };
__device__ uint32_t operator()(const uint32_t x, const uint32_t y) const {
#if (__CUDA_ARCH__ >= 300) && (__CUDA_ARCH__ < 500)
int32_t rv, z=0;
asm("vmax4.s32.s32.s32 %0, %1, %2, %3;" : "=r"(rv) : "r"(x), "r"(y), "r"(z));
return rv;
#else
converter cx, cy, cr;
cx.storage = x;
cy.storage = y;
cr.a.x = max(cx.a.x, cy.a.x);
cr.a.y = max(cx.a.y, cy.a.y);
cr.a.z = max(cx.a.z, cy.a.z);
cr.a.w = max(cx.a.w, cy.a.w);
return cr.storage;
#endif
}
__device__ int8_t operator()(const int8_t x, const int8_t y) const {
return (x>y) ? x : y;
}
};
template<>
struct FuncMax<uint8_t> {
union converter { uint32_t storage; uchar4 a; };
__device__ uint32_t operator()(const uint32_t x, const uint32_t y) const {
#if (__CUDA_ARCH__ >= 300) && (__CUDA_ARCH__ < 500)
int32_t rv, z=0;
asm("vmax4.u32.u32.u32 %0, %1, %2, %3;" : "=r"(rv) : "r"(x), "r"(y), "r"(z));
return rv;
#else
converter cx, cy, cr;
cx.storage = x;
cy.storage = y;
cr.a.x = max(cx.a.x, cy.a.x);
cr.a.y = max(cx.a.y, cy.a.y);
cr.a.z = max(cx.a.z, cy.a.z);
cr.a.w = max(cx.a.w, cy.a.w);
return cr.storage;
#endif
}
__device__ uint8_t operator()(const uint8_t x, const uint8_t y) const {
return (x>y) ? x : y;
}
};
template<>
struct FuncMin<int8_t> {
union converter { uint32_t storage; char4 a; };
__device__ uint32_t operator()(const uint32_t x, const uint32_t y) const {
#if (__CUDA_ARCH__ >= 300) && (__CUDA_ARCH__ < 500)
int32_t rv, z=0;
asm("vmin4.s32.s32.s32 %0, %1, %2, %3;" : "=r"(rv) : "r"(x), "r"(y), "r"(z));
return rv;
#else
converter cx, cy, cr;
cx.storage = x;
cy.storage = y;
cr.a.x = min(cx.a.x, cy.a.x);
cr.a.y = min(cx.a.y, cy.a.y);
cr.a.z = min(cx.a.z, cy.a.z);
cr.a.w = min(cx.a.w, cy.a.w);
return cr.storage;
#endif
}
__device__ int8_t operator()(const int8_t x, const int8_t y) const {
return (x<y) ? x : y;
}
};
template<>
struct FuncMin<uint8_t> {
union converter { uint32_t storage; uchar4 a; };
__device__ uint32_t operator()(const uint32_t x, const uint32_t y) const {
#if (__CUDA_ARCH__ >= 300) && (__CUDA_ARCH__ < 500)
int32_t rv, z=0;
asm("vmin4.u32.u32.u32 %0, %1, %2, %3;" : "=r"(rv) : "r"(x), "r"(y), "r"(z));
return rv;
#else
converter cx, cy, cr;
cx.storage = x;
cy.storage = y;
cr.a.x = min(cx.a.x, cy.a.x);
cr.a.y = min(cx.a.y, cy.a.y);
cr.a.z = min(cx.a.z, cy.a.z);
cr.a.w = min(cx.a.w, cy.a.w);
return cr.storage;
#endif
}
__device__ uint8_t operator()(const uint8_t x, const uint8_t y) const {
return (x<y) ? x : y;
}
};
template<>
struct FuncSum<half> {
__device__ half2 operator()(const half2 x, const half2 y) const {
#if __CUDA_ARCH__ >= 530 && __CUDA_ARCH__ != 610
return __hadd2(x, y);
#else
float2 fx, fy, fr;
fx = __half22float2(x);
fy = __half22float2(y);
fr.x = fx.x + fy.x;
fr.y = fx.y + fy.y;
return __float22half2_rn(fr);
#endif
}
__device__ half operator()(const half x, const half y) const {
#if __CUDA_ARCH__ >= 530 && __CUDA_ARCH__ != 610
return __hadd(x, y);
#else
return __float2half( __half2float(x) + __half2float(y) );
#endif
}
};
template<>
struct FuncProd<half> {
__device__ half2 operator()(const half2 x, const half2 y) const {
#if __CUDA_ARCH__ >= 530 && __CUDA_ARCH__ != 610
return __hmul2(x, y);
#else
float2 fx, fy, fr;
fx = __half22float2(x);
fy = __half22float2(y);
fr.x = fx.x * fy.x;
fr.y = fx.y * fy.y;
return __float22half2_rn(fr);
#endif
}
__device__ half operator()(const half x, const half y) const {
#if __CUDA_ARCH__ >= 530 && __CUDA_ARCH__ != 610
return __hmul(x, y);
#else
return __float2half( __half2float(x) * __half2float(y) );
#endif
}
};
template<>
struct FuncMax<half> {
__device__ half2 operator()(const half2 x, const half2 y) const {
float2 fx, fy, fr;
fx = __half22float2(x);
fy = __half22float2(y);
fr.x = fmaxf(fx.x, fy.x);
fr.y = fmaxf(fx.y, fy.y);
return __float22half2_rn(fr);
}
__device__ half operator()(const half x, const half y) const {
float fx, fy, fm;
fx = __half2float(x);
fy = __half2float(y);
fm = fmaxf(fx, fy);
return __float2half(fm);
}
};
template<>
struct FuncMin<half> {
__device__ half2 operator()(const half2 x, const half2 y) const {
float2 fx, fy, fr;
fx = __half22float2(x);
fy = __half22float2(y);
fr.x = fminf(fx.x, fy.x);
fr.y = fminf(fx.y, fy.y);
return __float22half2_rn(fr);
}
__device__ half operator()(const half x, const half y) const {
float fx, fy, fm;
fx = __half2float(x);
fy = __half2float(y);
fm = fminf(fx, fy);
return __float2half(fm);
}
};
#endif // REDUCE_KERNEL_H_
|
