/
cl_kernels.c
318 lines (313 loc) · 10.6 KB
/
cl_kernels.c
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
313
314
315
316
317
318
/* This file is part of rml.
rml is free software: you can redistribute it and/or modify
it under the terms of the GNU Lesser General Public License as published by
the Free Software Foundation, either version 3 of the License, or
any later version.
rml is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU Lesser General Public License for more details.
You should have received a copy of the GNU Lesser General Public License
along with rml. If not, see <https://www.gnu.org/licenses/>. */
#include "cl_kernels.h"
const char *rml_cl_program =
"__kernel void rml_matmul(__global TYPE *a, __global TYPE *b, __global TYPE *c, const unsigned int d1, const unsigned int d2, const unsigned int d3)\n"\
"{\n"\
" unsigned int id = get_global_id(0);\n"\
" c[id] = 0;\n"\
" unsigned int row = id / d3;\n"\
" unsigned int col = id % d3;\n"\
" for (unsigned int i = 0; i < d2; i++) {;\n"\
" c[id] += a[row * d2 + i] * b[i * d3 + col];\n"\
" }\n"\
"}\n"\
"\n"\
"__kernel void rml_concat(__global TYPE *a, __global TYPE *b, __global TYPE *c, __constant const unsigned int *dims_a, __constant const unsigned int *dims_b, __constant const unsigned int *dims_c, const unsigned int num_dims, const unsigned int cdim)\n"\
"{\n"\
" unsigned int id = get_global_id(0);\n"\
" unsigned int pos_workspace[MAX_ARR_SIZE];\n"\
" unsigned int i_divided = id;\n"\
" int reached_zero = 0;\n"\
" for (unsigned int d = num_dims - 1; !reached_zero; d--) {\n"\
" if (d < num_dims - 1) i_divided /= dims_c[d + 1];\n"\
" pos_workspace[d] = i_divided % dims_c[d];\n"\
" if (d == 0) reached_zero = 1;\n"\
" }\n"\
" unsigned int old_pos = 0;\n"\
" if (pos_workspace[cdim] < dims_a[cdim]) {\n"\
" for (unsigned int d = 0; d < num_dims; d++) {\n"\
" unsigned int prev_mult = 0;\n"\
" if (d > 0) prev_mult = dims_a[d];\n"\
" old_pos = old_pos * prev_mult + pos_workspace[d];\n"\
" }\n"\
" c[id] = a[old_pos];\n"\
" }\n"\
" else {\n"\
" pos_workspace[cdim] -= dims_a[cdim];\n"\
" for (unsigned int d = 0; d < num_dims; d++) {\n"\
" unsigned int prev_mult = 0;\n"\
" if (d > 0) prev_mult = dims_b[d];\n"\
" old_pos = old_pos * prev_mult + pos_workspace[d];\n"\
" }\n"\
" c[id] = b[old_pos];\n"\
" }\n"\
" \n"\
"}\n"\
"\n"\
"__kernel void rml_slice(__global TYPE *tensor, __global TYPE *result, __constant const unsigned int *lower_bound, __constant const unsigned int *upper_bound, __constant const unsigned int *tensor_dims, __constant const unsigned int *result_dims, const unsigned int num_dims)\n"\
"{\n"\
" unsigned int id = get_global_id(0);\n"\
" unsigned int pos_workspace[MAX_ARR_SIZE];\n"\
" unsigned int i_divided = id;\n"\
" int reached_zero = 0;\n"\
" for (unsigned int d = num_dims - 1; !reached_zero; d--) {\n"\
" if (d < num_dims - 1) i_divided /= result_dims[d + 1];\n"\
" pos_workspace[d] = i_divided % result_dims[d] + lower_bound[d];\n"\
" if (d == 0) reached_zero = 1;\n"\
" }\n"\
" unsigned int old_pos = 0;\n"\
" for (unsigned int d = 0; d < num_dims; d++) {\n"\
" unsigned int prev_mult = 0;\n"\
" if (d > 0) prev_mult = tensor_dims[d];\n"\
" old_pos = old_pos * prev_mult + pos_workspace[d];\n"\
" }\n"\
" result[id] = tensor[old_pos];\n"\
"}\n"\
"\n"\
"__kernel void rml_assign_slice(__global TYPE *tensor, __global TYPE *result, __constant unsigned int *lower_bound, __constant unsigned int *tensor_dims, __constant unsigned int *result_dims, const unsigned int num_dims)\n"\
"{\n"\
" unsigned int id = get_global_id(0);\n"\
" unsigned int pos_workspace[MAX_ARR_SIZE];\n"\
" unsigned int i_divided = id;\n"\
" int reached_zero = 0;\n"\
" for (unsigned int d = num_dims - 1; !reached_zero; d--) {\n"\
" if (d < num_dims - 1) i_divided /= tensor_dims[d + 1];\n"\
" pos_workspace[d] = i_divided % tensor_dims[d] + lower_bound[d];\n"\
" if (d == 0) reached_zero = 1;\n"\
" }\n"\
" unsigned int new_pos = 0;\n"\
" for (unsigned int d = 0; d < num_dims; d++) {\n"\
" unsigned int prev_mult = 0;\n"\
" if (d > 0) prev_mult = result_dims[d];\n"\
" new_pos = new_pos * prev_mult + pos_workspace[d];\n"\
" }\n"\
" result[new_pos] = tensor[id];\n"\
"}\n"\
"\n"\
"__kernel void rml_transpose(__global TYPE *tensor, __global TYPE *result, const unsigned int in_r, const unsigned int in_c)\n"\
"{\n"\
" unsigned int id = get_global_id(0);\n"\
" unsigned int new_r = id / in_r;\n"\
" unsigned int new_c = id % in_r;\n"\
" result[id] = tensor[new_c * in_c + new_r];\n"\
"}\n"\
"\n"\
"__kernel void rml_permute(__global TYPE *tensor, __global TYPE *result, __constant unsigned int *perms, __constant unsigned int *dims, const unsigned int num_dims)\n"\
"{\n"\
" unsigned int id = get_global_id(0);\n"\
" unsigned int pos_workspace[MAX_ARR_SIZE];\n"\
" unsigned int i_divided = id;\n"\
" int reached_zero = 0;\n"\
" for (unsigned int d = num_dims - 1; !reached_zero; d--) {\n"\
" if (d < num_dims - 1) i_divided /= dims[d + 1];\n"\
" pos_workspace[d] = i_divided % dims[d];\n"\
" if (d == 0) reached_zero = 1;\n"\
" }\n"\
" unsigned int new_pos = 0;\n"\
" for (unsigned int d = 0; d < num_dims; d++) {\n"\
" unsigned int prev_mult = 0;\n"\
" if (d > 0) prev_mult = dims[perms[d]];\n"\
" new_pos = new_pos * prev_mult + pos_workspace[perms[d]];\n"\
" }\n"\
" result[new_pos] = tensor[id];\n"\
"}\n"\
"\n"\
"__kernel void rml_cast_float(__global TYPE *a, __global float *b)\n"\
"{\n"\
" unsigned int id = get_global_id(0);\n"\
" b[id] = (float) a[id];\n"\
"}\n"\
"\n"\
"#ifdef cl_khr_fp64\n"\
"__kernel void rml_cast_double(__global TYPE *a, __global double *b)\n"\
"{\n"\
" unsigned int id = get_global_id(0);\n"\
" b[id] = (double) a[id];\n"\
"}\n"\
"#endif\n"\
"\n"\
"__kernel void rml_add(__global TYPE *a, __global TYPE *b, __global TYPE *c)\n"\
"{\n"\
" unsigned int id = get_global_id(0);\n"\
" c[id] = a[id] + b[id];\n"\
"}\n"\
"\n"\
"__kernel void rml_sub(__global TYPE *a, __global TYPE *b, __global TYPE *c)\n"\
"{\n"\
" unsigned int id = get_global_id(0);\n"\
" c[id] = a[id] - b[id];\n"\
"}\n"\
"__kernel void rml_mul(__global TYPE *a, __global TYPE *b, __global TYPE *c)\n"\
"{\n"\
" unsigned int id = get_global_id(0);\n"\
" c[id] = a[id] * b[id];\n"\
"}\n"\
"\n"\
"__kernel void rml_div(__global TYPE *a, __global TYPE *b, __global TYPE *c)\n"\
"{\n"\
" unsigned int id = get_global_id(0);\n"\
" c[id] = a[id] / b[id];\n"\
"}\n"\
"\n"\
"__kernel void rml_increment(__global TYPE *a, const TYPE scalar)\n"\
"{\n"\
" unsigned int id = get_global_id(0);\n"\
" a[id] += scalar;\n"\
"}\n"\
"\n"\
"__kernel void rml_scale(__global TYPE *a, const TYPE scalar)\n"\
"{\n"\
" unsigned int id = get_global_id(0);\n"\
" a[id] *= scalar;\n"\
"}\n"\
"\n"\
"__kernel void rml_exp(__global TYPE *a)\n"\
"{\n"\
" unsigned int id = get_global_id(0);\n"\
" a[id] = exp(a[id]);\n"\
"}\n"\
"\n"\
"__kernel void rml_log(__global TYPE *a)\n"\
"{\n"\
" unsigned int id = get_global_id(0);\n"\
" a[id] = log(a[id]);\n"\
"}\n"\
"\n"\
"__kernel void rml_pow(__global TYPE *a, const TYPE scalar)\n"\
"{\n"\
" unsigned int id = get_global_id(0);\n"\
" a[id] = pow(a[id], scalar);\n"\
"}\n"\
"\n"\
"__kernel void rml_sin(__global TYPE *a)\n"\
"{\n"\
" unsigned int id = get_global_id(0);\n"\
" a[id] = sin(a[id]);\n"\
"}\n"\
"\n"\
"__kernel void rml_cos(__global TYPE *a)\n"\
"{\n"\
" unsigned int id = get_global_id(0);\n"\
" a[id] = cos(a[id]);\n"\
"}\n"\
"\n"\
"__kernel void rml_tan(__global TYPE *a)\n"\
"{\n"\
" unsigned int id = get_global_id(0);\n"\
" a[id] = tan(a[id]);\n"\
"}\n"\
"\n"\
"__kernel void rml_sinh(__global TYPE *a)\n"\
"{\n"\
" unsigned int id = get_global_id(0);\n"\
" a[id] = sinh(a[id]);\n"\
"}\n"\
"\n"\
"__kernel void rml_cosh(__global TYPE *a)\n"\
"{\n"\
" unsigned int id = get_global_id(0);\n"\
" a[id] = cosh(a[id]);\n"\
"}\n"\
"\n"\
"__kernel void rml_tanh(__global TYPE *a)\n"\
"{\n"\
" unsigned int id = get_global_id(0);\n"\
" a[id] = tanh(a[id]);\n"\
"}\n"\
"\n"\
"__kernel void rml_asin(__global TYPE *a)\n"\
"{\n"\
" unsigned int id = get_global_id(0);\n"\
" a[id] = asin(a[id]);\n"\
"}\n"\
"\n"\
"__kernel void rml_acos(__global TYPE *a)\n"\
"{\n"\
" unsigned int id = get_global_id(0);\n"\
" a[id] = acos(a[id]);\n"\
"}\n"\
"\n"\
"__kernel void rml_atan(__global TYPE *a)\n"\
"{\n"\
" unsigned int id = get_global_id(0);\n"\
" a[id] = atan(a[id]);\n"\
"}\n"\
"\n"\
"__kernel void rml_asinh(__global TYPE *a)\n"\
"{\n"\
" unsigned int id = get_global_id(0);\n"\
" a[id] = asinh(a[id]);\n"\
"}\n"\
"\n"\
"__kernel void rml_acosh(__global TYPE *a)\n"\
"{\n"\
" unsigned int id = get_global_id(0);\n"\
" a[id] = acosh(a[id]);\n"\
"}\n"\
"\n"\
"__kernel void rml_atanh(__global TYPE *a)\n"\
"{\n"\
" unsigned int id = get_global_id(0);\n"\
" a[id] = atanh(a[id]);\n"\
"}\n"\
"\n"\
"__kernel void rml_abs(__global TYPE *a)\n"\
"{\n"\
" unsigned int id = get_global_id(0);\n"\
" a[id] = a[id] >= 0 ? a[id] : -a[id];\n"\
"}\n"\
"\n"\
"__kernel void rml_clamp(__global TYPE *a, const TYPE min, const TYPE max, const unsigned int code)\n"\
"{\n"\
" unsigned int id = get_global_id(0);\n"\
" if (code % 2 == 0 && a[id] < min) a[id] = min;\n"\
" if (code > 0 && a[id] > max) a[id] = max;\n"\
"}\n"\
"\n"\
"__kernel void rml_max(__global TYPE *tensor, __global TYPE *result, const unsigned int pool_size, const unsigned int in_size)\n"\
"{\n"\
" unsigned int id = get_global_id(0);\n"\
" for (unsigned int i = 0; i < pool_size && id * pool_size + i < in_size; i++) {\n"\
" if (i == 0 || result[id] < tensor[id * pool_size + i]) result[id] = tensor[id * pool_size + i];\n"\
" }\n"\
" \n"\
"}\n"\
"\n"\
"__kernel void rml_min(__global TYPE *tensor, __global TYPE *result, const unsigned int pool_size, const unsigned int in_size)\n"\
"{\n"\
" unsigned int id = get_global_id(0);\n"\
" for (unsigned int i = 0; i < pool_size && id * pool_size + i < in_size; i++) {\n"\
" if (i == 0 || result[id] > tensor[id * pool_size + i]) result[id] = tensor[id * pool_size + i];\n"\
" }\n"\
"}\n"\
"\n"\
"__kernel void rml_sum(__global TYPE *tensor, __global TYPE *result, const unsigned int pool_size, const unsigned int in_size)\n"\
"{\n"\
" unsigned int id = get_global_id(0);\n"\
" for (unsigned int i = 0; i < pool_size && id * pool_size + i < in_size; i++) {\n"\
" if (i == 0) result[id] = tensor[id * pool_size + i];\n"\
" else result[id] += tensor[id * pool_size + i];\n"\
" }\n"\
"}\n"\
"\n"\
"__kernel void rml_diag(__global TYPE *tensor, __global TYPE *result, const unsigned int inc_amount)\n"\
"{\n"\
" unsigned int id = get_global_id(0);\n"\
" if (id % inc_amount == 0) {\n"\
" result[id] = tensor[id / inc_amount];\n"\
" }\n"\
" else {\n"\
" result[id] = 0;\n"\
" }\n"\
"}\n"\
"\n";