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
303
304
305
306
307
308
309
310
311
312
|
/* OpenCL built-in library: subroups functionality
Copyright (c) 2022-2023 Pekka Jääskeläinen / Intel Finland Oy
Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to
deal in the Software without restriction, including without limitation the
rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
sell copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in
all copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
IN THE SOFTWARE.
*/
/* The default implementation of subgroups for CPU drivers. It uses work-group
sized local buffers for exchanging the data. The subgroup size is by default
the local X dimension side, unless restricted with the
intel_reqd_sub_group_size metadata.
*/
#include <math.h>
#include "work_group_alloca.h"
size_t _CL_OVERLOADABLE get_local_id (unsigned int dimindx);
size_t _CL_OVERLOADABLE get_local_linear_id (void);
size_t _CL_OVERLOADABLE get_local_size (unsigned int dimindx);
/* Magic variable that is expanded in Workgroup.cc */
extern uint _pocl_sub_group_size;
uint _CL_OVERLOADABLE
get_sub_group_size (void)
{
return _pocl_sub_group_size;
}
uint _CL_OVERLOADABLE
get_max_sub_group_size (void)
{
return get_sub_group_size ();
}
uint _CL_OVERLOADABLE
get_num_sub_groups (void)
{
return (uint)get_local_size (0) * get_local_size (1) * get_local_size (2)
/ get_max_sub_group_size ();
}
uint _CL_OVERLOADABLE
get_enqueued_num_sub_groups (void)
{
return 1;
}
uint _CL_OVERLOADABLE
get_sub_group_id (void)
{
return (uint)get_local_linear_id () / get_max_sub_group_size ();
}
uint _CL_OVERLOADABLE
get_sub_group_local_id (void)
{
return (uint)get_local_linear_id () % get_max_sub_group_size ();
}
static size_t _CL_OVERLOADABLE
get_first_llid (void)
{
return get_sub_group_id () * get_max_sub_group_size ();
}
void _CL_OVERLOADABLE sub_group_barrier (cl_mem_fence_flags flags);
#define SUB_GROUP_SHUFFLE_PT(PREFIX, TYPE) \
__attribute__ ((always_inline)) \
TYPE _CL_OVERLOADABLE PREFIX##sub_group_shuffle (TYPE val, uint index) \
{ \
volatile TYPE *temp_storage \
= __pocl_work_group_alloca (sizeof (TYPE), sizeof (TYPE), 0); \
temp_storage[get_local_linear_id ()] = val; \
sub_group_barrier (CLK_LOCAL_MEM_FENCE); \
return temp_storage[get_first_llid () + index % get_sub_group_size ()]; \
}
/* Define both the non-prefixed (khr) and Intel-prefixed shuffles. */
#define SUB_GROUP_SHUFFLE_T(TYPE) \
SUB_GROUP_SHUFFLE_PT (, TYPE) \
SUB_GROUP_SHUFFLE_PT (intel_, TYPE)
SUB_GROUP_SHUFFLE_T (char)
SUB_GROUP_SHUFFLE_T (uchar)
SUB_GROUP_SHUFFLE_T (short)
SUB_GROUP_SHUFFLE_T (ushort)
SUB_GROUP_SHUFFLE_T (int)
SUB_GROUP_SHUFFLE_T (uint)
SUB_GROUP_SHUFFLE_T (long)
SUB_GROUP_SHUFFLE_T (ulong)
#ifdef cl_khr_fp16
SUB_GROUP_SHUFFLE_T (half)
/* OpenCL C mangles half 'h' whereas C (clang) mangles it 'fp16'.
We need to provide a wrapper for the OpenCL C compatible mangling. */
half
_Z23intel_sub_group_shuffleDhj (half val, uint mask)
{
return intel_sub_group_shuffle (val, mask);
}
#endif
SUB_GROUP_SHUFFLE_T (float)
SUB_GROUP_SHUFFLE_T (double)
#define SUB_GROUP_SHUFFLE_XOR_PT(PREFIX, TYPE) \
__attribute__ ((always_inline)) \
TYPE _CL_OVERLOADABLE PREFIX##sub_group_shuffle_xor (TYPE val, uint mask) \
{ \
volatile TYPE *temp_storage \
= __pocl_work_group_alloca (sizeof (TYPE), sizeof (TYPE), 0); \
temp_storage[get_local_linear_id ()] = val; \
sub_group_barrier (CLK_LOCAL_MEM_FENCE); \
return temp_storage[get_first_llid () \
+ (get_sub_group_local_id () ^ mask) \
% get_sub_group_size ()]; \
}
/* Define both the non-prefixed (khr) and Intel-prefixed shuffles. */
#define SUB_GROUP_SHUFFLE_XOR_T(TYPE) \
SUB_GROUP_SHUFFLE_XOR_PT (, TYPE) \
SUB_GROUP_SHUFFLE_XOR_PT (intel_, TYPE)
SUB_GROUP_SHUFFLE_XOR_T (char)
SUB_GROUP_SHUFFLE_XOR_T (uchar)
SUB_GROUP_SHUFFLE_XOR_T (short)
SUB_GROUP_SHUFFLE_XOR_T (ushort)
SUB_GROUP_SHUFFLE_XOR_T (int)
SUB_GROUP_SHUFFLE_XOR_T (uint)
SUB_GROUP_SHUFFLE_XOR_T (long)
SUB_GROUP_SHUFFLE_XOR_T (ulong)
SUB_GROUP_SHUFFLE_XOR_T (float)
SUB_GROUP_SHUFFLE_XOR_T (double)
#define SUB_GROUP_BROADCAST_T(TYPE) \
__attribute__ ((always_inline)) TYPE _CL_OVERLOADABLE sub_group_broadcast ( \
TYPE val, uint id) \
{ \
return sub_group_shuffle (val, id); \
}
SUB_GROUP_BROADCAST_T (int)
SUB_GROUP_BROADCAST_T (uint)
SUB_GROUP_BROADCAST_T (long)
SUB_GROUP_BROADCAST_T (ulong)
SUB_GROUP_BROADCAST_T (float)
SUB_GROUP_BROADCAST_T (double)
#define SUB_GROUP_REDUCE_OT(OPNAME, OPERATION, TYPE) \
__attribute__ ((always_inline)) \
TYPE _CL_OVERLOADABLE sub_group_reduce_##OPNAME (TYPE val) \
{ \
volatile TYPE *temp_storage \
= __pocl_work_group_alloca (sizeof (TYPE), sizeof (TYPE), 0); \
temp_storage[get_local_linear_id ()] = val; \
sub_group_barrier (CLK_LOCAL_MEM_FENCE); \
if (get_sub_group_local_id () == 0) \
{ \
for (uint i = 1; i < get_sub_group_size (); ++i) \
{ \
TYPE a = temp_storage[get_first_llid ()], \
b = temp_storage[get_first_llid () + i]; \
temp_storage[get_first_llid ()] = OPERATION; \
} \
} \
sub_group_barrier (CLK_LOCAL_MEM_FENCE); \
return temp_storage[get_first_llid ()]; \
}
#define SUB_GROUP_REDUCE_T(OPNAME, OPERATION) \
SUB_GROUP_REDUCE_OT (OPNAME, OPERATION, int) \
SUB_GROUP_REDUCE_OT (OPNAME, OPERATION, uint) \
SUB_GROUP_REDUCE_OT (OPNAME, OPERATION, long) \
SUB_GROUP_REDUCE_OT (OPNAME, OPERATION, ulong) \
SUB_GROUP_REDUCE_OT (OPNAME, OPERATION, float) \
SUB_GROUP_REDUCE_OT (OPNAME, OPERATION, double)
SUB_GROUP_REDUCE_T (add, a + b)
SUB_GROUP_REDUCE_T (min, a > b ? b : a)
SUB_GROUP_REDUCE_T (max, a > b ? a : b)
#ifdef cl_khr_fp16
SUB_GROUP_REDUCE_OT (add, a + b, half)
SUB_GROUP_REDUCE_OT (max, a > b ? a : b, half)
half
_Z20sub_group_reduce_maxDh (half val)
{
return sub_group_reduce_max (val);
}
half
_Z20sub_group_reduce_addDh (half val)
{
return sub_group_reduce_add (val);
}
#endif
#define SUB_GROUP_SCAN_INCLUSIVE_OT(OPNAME, OPERATION, TYPE) \
__attribute__ ((always_inline)) \
TYPE _CL_OVERLOADABLE sub_group_scan_inclusive_##OPNAME (TYPE val) \
{ \
volatile TYPE *data \
= __pocl_work_group_alloca (sizeof (TYPE), sizeof (TYPE), 0); \
data[get_local_linear_id ()] = val; \
sub_group_barrier (CLK_LOCAL_MEM_FENCE); \
if (get_sub_group_local_id () == 0) \
{ \
for (uint i = 1; i < get_sub_group_size (); ++i) \
{ \
TYPE a = data[get_first_llid () + i - 1], \
b = data[get_first_llid () + i]; \
data[get_first_llid () + i] = OPERATION; \
} \
} \
sub_group_barrier (CLK_LOCAL_MEM_FENCE); \
return data[get_local_linear_id ()]; \
}
#define SUB_GROUP_SCAN_INCLUSIVE_T(OPNAME, OPERATION) \
SUB_GROUP_SCAN_INCLUSIVE_OT (OPNAME, OPERATION, int) \
SUB_GROUP_SCAN_INCLUSIVE_OT (OPNAME, OPERATION, uint) \
SUB_GROUP_SCAN_INCLUSIVE_OT (OPNAME, OPERATION, long) \
SUB_GROUP_SCAN_INCLUSIVE_OT (OPNAME, OPERATION, ulong) \
SUB_GROUP_SCAN_INCLUSIVE_OT (OPNAME, OPERATION, float) \
SUB_GROUP_SCAN_INCLUSIVE_OT (OPNAME, OPERATION, double)
SUB_GROUP_SCAN_INCLUSIVE_T (add, a + b)
SUB_GROUP_SCAN_INCLUSIVE_T (min, a > b ? b : a)
SUB_GROUP_SCAN_INCLUSIVE_T (max, a > b ? a : b)
#define SUB_GROUP_SCAN_EXCLUSIVE_OT(OPNAME, OPERATION, TYPE, ID) \
__attribute__ ((always_inline)) \
TYPE _CL_OVERLOADABLE sub_group_scan_exclusive_##OPNAME (TYPE val) \
{ \
volatile TYPE *data = __pocl_work_group_alloca ( \
sizeof (TYPE), sizeof (TYPE), sizeof (TYPE)); \
data[get_local_linear_id () + 1] = val; \
data[get_first_llid ()] = ID; \
sub_group_barrier (CLK_LOCAL_MEM_FENCE); \
if (get_sub_group_local_id () == 0) \
{ \
for (uint i = 1; i < get_sub_group_size (); ++i) \
{ \
TYPE a = data[get_first_llid () + i - 1], \
b = data[get_first_llid () + i]; \
data[get_first_llid () + i] = OPERATION; \
} \
} \
sub_group_barrier (CLK_LOCAL_MEM_FENCE); \
return data[get_local_linear_id ()]; \
}
SUB_GROUP_SCAN_EXCLUSIVE_OT (add, a + b, int, 0)
SUB_GROUP_SCAN_EXCLUSIVE_OT (add, a + b, uint, 0)
SUB_GROUP_SCAN_EXCLUSIVE_OT (add, a + b, long, 0)
SUB_GROUP_SCAN_EXCLUSIVE_OT (add, a + b, ulong, 0)
SUB_GROUP_SCAN_EXCLUSIVE_OT (add, a + b, float, 0.0f)
SUB_GROUP_SCAN_EXCLUSIVE_OT (add, a + b, double, 0.0)
SUB_GROUP_SCAN_EXCLUSIVE_OT (min, a > b ? b : a, int, INT_MAX)
SUB_GROUP_SCAN_EXCLUSIVE_OT (min, a > b ? b : a, uint, UINT_MAX)
SUB_GROUP_SCAN_EXCLUSIVE_OT (min, a > b ? b : a, long, LONG_MAX)
SUB_GROUP_SCAN_EXCLUSIVE_OT (min, a > b ? b : a, ulong, ULONG_MAX)
SUB_GROUP_SCAN_EXCLUSIVE_OT (min, a > b ? b : a, float, +INFINITY)
SUB_GROUP_SCAN_EXCLUSIVE_OT (min, a > b ? b : a, double, +INFINITY)
SUB_GROUP_SCAN_EXCLUSIVE_OT (max, a > b ? a : b, int, INT_MIN)
SUB_GROUP_SCAN_EXCLUSIVE_OT (max, a > b ? a : b, uint, 0)
SUB_GROUP_SCAN_EXCLUSIVE_OT (max, a > b ? a : b, long, LONG_MIN)
SUB_GROUP_SCAN_EXCLUSIVE_OT (max, a > b ? a : b, ulong, 0)
SUB_GROUP_SCAN_EXCLUSIVE_OT (max, a > b ? a : b, float, -INFINITY)
SUB_GROUP_SCAN_EXCLUSIVE_OT (max, a > b ? a : b, double, -INFINITY)
__attribute__ ((always_inline)) uint4 _CL_OVERLOADABLE
sub_group_ballot (int predicate)
{
/* The results for the ballot for all of the WG's SGs. */
uint4 *flags = __pocl_work_group_alloca (sizeof (uint4), sizeof (uint4), 0);
/* Temporary storage for the predicate flags of all WIs in the WG. */
char *res = __pocl_work_group_alloca (sizeof (char), 4, 0);
res[get_local_linear_id ()] = !!predicate;
sub_group_barrier (CLK_LOCAL_MEM_FENCE);
if (get_sub_group_local_id () == 0)
{
flags[get_sub_group_id ()] = 0;
uint *f = (uint*) (flags + get_sub_group_id ());
for (uint i = 0; i < get_sub_group_size () && i < 128; ++i)
{
f[i / 32] |= res[get_first_llid () + i] << (i % 32);
}
}
sub_group_barrier (CLK_LOCAL_MEM_FENCE);
return flags[get_sub_group_id ()];
}
|
