conv_driver.cpp 25.5 KB
Newer Older
Chao Liu's avatar
Chao Liu committed
1
#include <iostream>
Chao Liu's avatar
Chao Liu committed
2
3
#include <numeric>
#include <initializer_list>
Chao Liu's avatar
Chao Liu committed
4
#include <cstdlib>
Chao Liu's avatar
Chao Liu committed
5
#include <stdlib.h>
Chao Liu's avatar
Chao Liu committed
6
#include <half.hpp>
Chao Liu's avatar
Chao Liu committed
7
#include "config.hpp"
Chao Liu's avatar
Chao Liu committed
8
#include "print.hpp"
Chao Liu's avatar
Chao Liu committed
9
#include "device.hpp"
Chao Liu's avatar
Chao Liu committed
10
#include "host_tensor_generator.hpp"
Chao Liu's avatar
Chao Liu committed
11
#include "conv_common.hpp"
12
#include "host_conv.hpp"
Chao Liu's avatar
Chao Liu committed
13
#include "device_tensor.hpp"
14
15
#include "device_convolution_forward_implicit_gemm_v4r1_nchw_kcyx_nkhw.hpp"
#include "device_convolution_forward_implicit_gemm_v4r4_nchw_kcyx_nkhw.hpp"
Chao Liu's avatar
Chao Liu committed
16
#include "device_convolution_forward_implicit_gemm_v4r4_nhwc_kyxc_nhwk.hpp"
17
#include "device_dynamic_convolution_forward_implicit_gemm_v4r4_nchw_kcyx_nkhw.hpp"
18
#include "device_dynamic_convolution_forward_implicit_gemm_v4r4_nhwc_kyxc_nhwk.hpp"
root's avatar
root committed
19
20
#include "device_dynamic_convolution_forward_implicit_gemm_v5r1_nchw_kcyx_nkhw.hpp"

Chao Liu's avatar
Chao Liu committed
21
int main(int argc, char* argv[])
Chao Liu's avatar
Chao Liu committed
22
{
Chao Liu's avatar
Chao Liu committed
23
24
    using namespace ck;

25
#if 0
Chao Liu's avatar
Chao Liu committed
26
27
    constexpr index_t N  = 1;
    constexpr index_t C  = 16;
Chao Liu's avatar
Chao Liu committed
28
29
30
31
32
33
34
35
36
    constexpr index_t HI = 1080;
    constexpr index_t WI = 1920;
    constexpr index_t K  = 16;
    constexpr index_t Y  = 1;
    constexpr index_t X  = 1;

    using ConvStrides   = Sequence<1, 1>;
    using ConvDilations = Sequence<1, 1>;

Chao Liu's avatar
Chao Liu committed
37
38
    using LeftPads  = Sequence<0, 0>;
    using RightPads = Sequence<0, 0>;
Jing Zhang's avatar
Jing Zhang committed
39
#elif 1
Chao Liu's avatar
Chao Liu committed
40
    constexpr index_t N  = 1;
Chao Liu's avatar
Chao Liu committed
41
    constexpr index_t C  = 16;
Jing Zhang's avatar
Jing Zhang committed
42
43
    //constexpr index_t HI = 540;
    constexpr index_t HI = 544;
Chao Liu's avatar
Chao Liu committed
44
    constexpr index_t WI = 960;
Chao Liu's avatar
Chao Liu committed
45
46
47
48
49
50
51
    constexpr index_t K  = 16;
    constexpr index_t Y  = 1;
    constexpr index_t X  = 1;

    using ConvStrides   = Sequence<1, 1>;
    using ConvDilations = Sequence<1, 1>;

Jing Zhang's avatar
Jing Zhang committed
52
53
    using LeftPads                   = Sequence<0, 0>;
    using RightPads                  = Sequence<0, 0>;
Chao Liu's avatar
Chao Liu committed
54
55
#elif 0
    constexpr index_t N  = 1;
Chao Liu's avatar
Chao Liu committed
56
    constexpr index_t C  = 16;
Chao Liu's avatar
Chao Liu committed
57
58
59
60
61
62
63
64
65
    constexpr index_t HI = 270;
    constexpr index_t WI = 480;
    constexpr index_t K  = 16;
    constexpr index_t Y  = 1;
    constexpr index_t X  = 1;

    using ConvStrides   = Sequence<1, 1>;
    using ConvDilations = Sequence<1, 1>;

Jing Zhang's avatar
Jing Zhang committed
66
67
    using LeftPads                   = Sequence<0, 0>;
    using RightPads                  = Sequence<0, 0>;
Jing Zhang's avatar
tweak  
Jing Zhang committed
68
69
70
71
72
73
74
75
76
77
78
79
80
81
#elif 0
    constexpr index_t N  = 1;
    constexpr index_t C  = 16;
    constexpr index_t HI = 2048;
    constexpr index_t WI = 2048;
    constexpr index_t K  = 16;
    constexpr index_t Y  = 3;
    constexpr index_t X  = 3;

    using ConvStrides   = Sequence<1, 1>;
    using ConvDilations = Sequence<1, 1>;

    using LeftPads  = Sequence<1, 1>;
    using RightPads = Sequence<1, 1>;
Jing Zhang's avatar
Jing Zhang committed
82
#elif 0
Chao Liu's avatar
Chao Liu committed
83
    constexpr index_t N  = 1;
root's avatar
root committed
84
    constexpr index_t C  = 16;
Chao Liu's avatar
Chao Liu committed
85
86
    constexpr index_t HI = 1080;
    constexpr index_t WI = 1920;
Jing Zhang's avatar
Jing Zhang committed
87
    constexpr index_t K  = 4;
Chao Liu's avatar
Chao Liu committed
88
89
90
91
92
93
    constexpr index_t Y  = 3;
    constexpr index_t X  = 3;

    using ConvStrides   = Sequence<1, 1>;
    using ConvDilations = Sequence<1, 1>;

94
95
    using LeftPads  = Sequence<1, 1>;
    using RightPads = Sequence<1, 1>;
Chao Liu's avatar
tweak  
Chao Liu committed
96
#elif 0
Chao Liu's avatar
Chao Liu committed
97
    constexpr index_t N  = 1;
Jing Zhang's avatar
mock up  
Jing Zhang committed
98
    constexpr index_t C  = 1;
root's avatar
root committed
99
100
    constexpr index_t HI = 1024;
    constexpr index_t WI = 2048;
Jing Zhang's avatar
mock up  
Jing Zhang committed
101
    constexpr index_t K  = 4;
Chao Liu's avatar
Chao Liu committed
102
103
104
105
106
107
    constexpr index_t Y  = 3;
    constexpr index_t X  = 3;

    using ConvStrides   = Sequence<1, 1>;
    using ConvDilations = Sequence<1, 1>;

Chao Liu's avatar
Chao Liu committed
108
109
    using LeftPads  = Sequence<1, 1>;
    using RightPads = Sequence<1, 1>;
Jing Zhang's avatar
Jing Zhang committed
110
#elif 0
Chao Liu's avatar
Chao Liu committed
111
    constexpr index_t N  = 1;
Chao Liu's avatar
Chao Liu committed
112
    constexpr index_t C  = 16;
Chao Liu's avatar
Chao Liu committed
113
114
115
116
117
118
119
120
121
    constexpr index_t HI = 540;
    constexpr index_t WI = 960;
    constexpr index_t K  = 16;
    constexpr index_t Y  = 3;
    constexpr index_t X  = 3;

    using ConvStrides   = Sequence<1, 1>;
    using ConvDilations = Sequence<1, 1>;

Jing Zhang's avatar
Jing Zhang committed
122
123
124
125
126
127
128
129
130
131
132
133
134
135
    using LeftPads  = Sequence<1, 1>;
    using RightPads = Sequence<1, 1>;
#elif 0
    constexpr index_t N  = 1;
    constexpr index_t C  = 16;
    constexpr index_t HI = 135;
    constexpr index_t WI = 240;
    constexpr index_t K  = 16;
    constexpr index_t Y  = 3;
    constexpr index_t X  = 3;

    using ConvStrides   = Sequence<1, 1>;
    using ConvDilations = Sequence<1, 1>;

Chao Liu's avatar
Chao Liu committed
136
137
    using LeftPads  = Sequence<1, 1>;
    using RightPads = Sequence<1, 1>;
Chao Liu's avatar
test  
Chao Liu committed
138
#elif 0
Chao Liu's avatar
Chao Liu committed
139
    constexpr index_t N  = 1;
Chao Liu's avatar
Chao Liu committed
140
    constexpr index_t C  = 16;
Chao Liu's avatar
Chao Liu committed
141
142
143
    constexpr index_t HI = 270;
    constexpr index_t WI = 480;
    constexpr index_t K  = 16;
Chao Liu's avatar
Chao Liu committed
144
145
146
147
148
149
150
151
152
    constexpr index_t Y  = 3;
    constexpr index_t X  = 3;

    using ConvStrides   = Sequence<1, 1>;
    using ConvDilations = Sequence<1, 1>;

    using LeftPads  = Sequence<1, 1>;
    using RightPads = Sequence<1, 1>;
#elif 0
153
154
155
156
157
158
159
160
161
162
163
164
165
166
    // 3x3, 36x36, stride 2
    constexpr index_t N  = 128;
    constexpr index_t C  = 192;
    constexpr index_t HI = 37;
    constexpr index_t WI = 37;
    constexpr index_t K  = 384;
    constexpr index_t Y  = 3;
    constexpr index_t X  = 3;

    using ConvStrides   = Sequence<2, 2>;
    using ConvDilations = Sequence<1, 1>;

    using LeftPads  = Sequence<0, 0>;
    using RightPads = Sequence<0, 0>;
167
#elif 0
168
169
170
171
172
173
174
175
    // 3x3, 35x35, stride 2
    constexpr index_t N  = 128;
    constexpr index_t C  = 192;
    constexpr index_t HI = 35;
    constexpr index_t WI = 35;
    constexpr index_t K  = 384;
    constexpr index_t Y  = 3;
    constexpr index_t X  = 3;
176

177
    using ConvStrides   = Sequence<2, 2>;
178
179
180
181
    using ConvDilations = Sequence<1, 1>;

    using LeftPads  = Sequence<0, 0>;
    using RightPads = Sequence<0, 0>;
Chao Liu's avatar
Chao Liu committed
182
#elif 0
Chao Liu's avatar
Chao Liu committed
183
    // 3x3, 71x71
Chao Liu's avatar
Chao Liu committed
184
    constexpr index_t N  = 128;
Chao Liu's avatar
Chao Liu committed
185
186
187
188
189
190
    constexpr index_t C  = 192;
    constexpr index_t HI = 71;
    constexpr index_t WI = 71;
    constexpr index_t K  = 128;
    constexpr index_t Y  = 3;
    constexpr index_t X  = 3;
Chao Liu's avatar
Chao Liu committed
191

Chao Liu's avatar
Chao Liu committed
192
    using ConvStrides   = Sequence<2, 2>;
193
    using ConvDilations = Sequence<1, 1>;
Chao Liu's avatar
Chao Liu committed
194

Chao Liu's avatar
Chao Liu committed
195
196
    using LeftPads  = Sequence<1, 1>;
    using RightPads = Sequence<1, 1>;
Chao Liu's avatar
Chao Liu committed
197
#elif 1
Chao Liu's avatar
Chao Liu committed
198
    // 1x1, 8x8
199
    constexpr index_t N  = 128;
Chao Liu's avatar
Chao Liu committed
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
    constexpr index_t C  = 1536;
    constexpr index_t HI = 8;
    constexpr index_t WI = 8;
    constexpr index_t K  = 256;
    constexpr index_t Y  = 1;
    constexpr index_t X  = 1;

    using ConvStrides   = Sequence<1, 1>;
    using ConvDilations = Sequence<1, 1>;

    using LeftPads  = Sequence<0, 0>;
    using RightPads = Sequence<0, 0>;
#elif 0
    // 1x1, 73x73
    constexpr index_t N  = 128;
    constexpr index_t C  = 160;
    constexpr index_t HI = 73;
    constexpr index_t WI = 73;
    constexpr index_t K  = 64;
    constexpr index_t Y  = 1;
    constexpr index_t X  = 1;

    using ConvStrides   = Sequence<1, 1>;
    using ConvDilations = Sequence<1, 1>;

    using LeftPads  = Sequence<0, 0>;
    using RightPads = Sequence<0, 0>;
#elif 0
    // 3x3, 35x35
    constexpr index_t N  = 128;
    constexpr index_t C  = 96;
    constexpr index_t HI = 35;
    constexpr index_t WI = 35;
233
    constexpr index_t K  = 128;
Chao Liu's avatar
Chao Liu committed
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
    constexpr index_t Y  = 3;
    constexpr index_t X  = 3;

    using ConvStrides   = Sequence<1, 1>;
    using ConvDilations = Sequence<1, 1>;

    using LeftPads  = Sequence<1, 1>;
    using RightPads = Sequence<1, 1>;
#elif 0
    // 3x3, 71x71
    constexpr index_t N  = 128;
    constexpr index_t C  = 192;
    constexpr index_t HI = 71;
    constexpr index_t WI = 71;
    constexpr index_t K  = 192;
    constexpr index_t Y  = 3;
    constexpr index_t X  = 3;

    using ConvStrides   = Sequence<2, 2>;
    using ConvDilations = Sequence<1, 1>;

    using LeftPads  = Sequence<1, 1>;
    using RightPads = Sequence<1, 1>;
257
#elif 1
Chao Liu's avatar
Chao Liu committed
258
259
260
    // 7x1, 17x17
    constexpr index_t N  = 128;
    constexpr index_t C  = 128;
261
262
    constexpr index_t HI = 17;
    constexpr index_t WI = 17;
263
    constexpr index_t K  = 128;
Chao Liu's avatar
Chao Liu committed
264
265
266
267
268
269
270
271
    constexpr index_t Y  = 7;
    constexpr index_t X  = 1;

    using ConvStrides   = Sequence<1, 1>;
    using ConvDilations = Sequence<1, 1>;

    using LeftPads  = Sequence<3, 0>;
    using RightPads = Sequence<3, 0>;
272
#elif 0
Chao Liu's avatar
Chao Liu committed
273
274
275
276
277
278
    // 1x7, 17x17
    constexpr index_t N  = 128;
    constexpr index_t C  = 128;
    constexpr index_t HI = 17;
    constexpr index_t WI = 17;
    constexpr index_t K  = 128;
279
280
281
282
283
284
285
286
287
    constexpr index_t Y  = 1;
    constexpr index_t X  = 7;

    using ConvStrides   = Sequence<1, 1>;
    using ConvDilations = Sequence<1, 1>;

    using LeftPads  = Sequence<0, 3>;
    using RightPads = Sequence<0, 3>;
#elif 0
Chao Liu's avatar
Chao Liu committed
288
289
290
291
292
293
    // 3x3, 299x299 stride=2
    constexpr index_t N  = 128;
    constexpr index_t C  = 3;
    constexpr index_t HI = 299;
    constexpr index_t WI = 299;
    constexpr index_t K  = 32;
294
295
    constexpr index_t Y  = 3;
    constexpr index_t X  = 3;
Chao Liu's avatar
Chao Liu committed
296

Chao Liu's avatar
Chao Liu committed
297
    using ConvStrides   = Sequence<2, 2>;
298
299
    using ConvDilations = Sequence<1, 1>;

300
301
302
    using LeftPads  = Sequence<0, 0>;
    using RightPads = Sequence<0, 0>;
#elif 0
Chao Liu's avatar
Chao Liu committed
303
    // 3x3, 147x147
304
    constexpr index_t N  = 128;
305
    constexpr index_t C  = 128;
Chao Liu's avatar
Chao Liu committed
306
307
    constexpr index_t HI = 147;
    constexpr index_t WI = 147;
308
    constexpr index_t K  = 128;
309
310
311
    constexpr index_t Y  = 3;
    constexpr index_t X  = 3;

Chao Liu's avatar
Chao Liu committed
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
    using ConvStrides   = Sequence<1, 1>;
    using ConvDilations = Sequence<1, 1>;

    using LeftPads  = Sequence<1, 1>;
    using RightPads = Sequence<1, 1>;
#elif 0
    // 3x3, 149x149
    constexpr index_t N  = 128;
    constexpr index_t C  = 32;
    constexpr index_t HI = 149;
    constexpr index_t WI = 149;
    constexpr index_t K  = 32;
    constexpr index_t Y  = 3;
    constexpr index_t X  = 3;

    using ConvStrides   = Sequence<1, 1>;
328
329
    using ConvDilations = Sequence<1, 1>;

330
331
    using LeftPads  = Sequence<0, 0>;
    using RightPads = Sequence<0, 0>;
Chao Liu's avatar
Chao Liu committed
332
#elif 0
Chao Liu's avatar
Chao Liu committed
333
334
335
336
337
338
339
340
    // 3x3, 17x17, stride 2
    constexpr index_t N  = 128;
    constexpr index_t C  = 192;
    constexpr index_t HI = 17;
    constexpr index_t WI = 17;
    constexpr index_t K  = 192;
    constexpr index_t Y  = 3;
    constexpr index_t X  = 3;
Chao Liu's avatar
Chao Liu committed
341

Chao Liu's avatar
Chao Liu committed
342
    using ConvStrides   = Sequence<2, 2>;
Chao Liu's avatar
Chao Liu committed
343
344
    using ConvDilations = Sequence<1, 1>;

Chao Liu's avatar
Chao Liu committed
345
346
    using LeftPads  = Sequence<0, 0>;
    using RightPads = Sequence<0, 0>;
Chao Liu's avatar
Chao Liu committed
347
#elif 0
Chao Liu's avatar
Chao Liu committed
348
    // 1x1, 35x35
Chao Liu's avatar
Chao Liu committed
349
    constexpr index_t N  = 128;
Chao Liu's avatar
Chao Liu committed
350
351
352
353
    constexpr index_t C  = 384;
    constexpr index_t HI = 35;
    constexpr index_t WI = 35;
    constexpr index_t K  = 96;
Chao Liu's avatar
Chao Liu committed
354
355
356
357
358
359
    constexpr index_t Y  = 1;
    constexpr index_t X  = 1;

    using ConvStrides   = Sequence<1, 1>;
    using ConvDilations = Sequence<1, 1>;

Chao Liu's avatar
Chao Liu committed
360
361
    using LeftPads  = Sequence<0, 0>;
    using RightPads = Sequence<0, 0>;
Chao Liu's avatar
Chao Liu committed
362
#elif 0
Chao Liu's avatar
Chao Liu committed
363
    // 3x3, 35x35, stride 2
Chao Liu's avatar
Chao Liu committed
364
    constexpr index_t N  = 128;
Chao Liu's avatar
Chao Liu committed
365
366
367
    constexpr index_t C  = 288;
    constexpr index_t HI = 35;
    constexpr index_t WI = 35;
Chao Liu's avatar
Chao Liu committed
368
    constexpr index_t K  = 384;
Chao Liu's avatar
Chao Liu committed
369
370
    constexpr index_t Y  = 3;
    constexpr index_t X  = 3;
Chao Liu's avatar
Chao Liu committed
371

Chao Liu's avatar
Chao Liu committed
372
    using ConvStrides   = Sequence<2, 2>;
Chao Liu's avatar
Chao Liu committed
373
374
    using ConvDilations = Sequence<1, 1>;

Chao Liu's avatar
Chao Liu committed
375
376
    using LeftPads  = Sequence<0, 0>;
    using RightPads = Sequence<0, 0>;
Chao Liu's avatar
Chao Liu committed
377
#elif 0
Chao Liu's avatar
Chao Liu committed
378
    // 1x3, 8x8
Chao Liu's avatar
Chao Liu committed
379
    constexpr index_t N  = 128;
Chao Liu's avatar
Chao Liu committed
380
    constexpr index_t C  = 384;
Chao Liu's avatar
Chao Liu committed
381
382
    constexpr index_t HI = 8;
    constexpr index_t WI = 8;
Chao Liu's avatar
Chao Liu committed
383
    constexpr index_t K  = 448;
Chao Liu's avatar
Chao Liu committed
384
    constexpr index_t Y  = 1;
Chao Liu's avatar
Chao Liu committed
385
    constexpr index_t X  = 3;
Chao Liu's avatar
Chao Liu committed
386
387
388
389

    using ConvStrides   = Sequence<1, 1>;
    using ConvDilations = Sequence<1, 1>;

Chao Liu's avatar
Chao Liu committed
390
391
    using LeftPads  = Sequence<0, 1>;
    using RightPads = Sequence<0, 1>;
Chao Liu's avatar
Chao Liu committed
392
#elif 0
Chao Liu's avatar
Chao Liu committed
393
    // 3x1, 8x8
Chao Liu's avatar
Chao Liu committed
394
    constexpr index_t N  = 128;
Chao Liu's avatar
Chao Liu committed
395
396
397
398
399
    constexpr index_t C  = 448;
    constexpr index_t HI = 8;
    constexpr index_t WI = 8;
    constexpr index_t K  = 512;
    constexpr index_t Y  = 3;
Chao Liu's avatar
Chao Liu committed
400
401
402
403
404
    constexpr index_t X  = 1;

    using ConvStrides   = Sequence<1, 1>;
    using ConvDilations = Sequence<1, 1>;

Chao Liu's avatar
Chao Liu committed
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
    using LeftPads  = Sequence<1, 0>;
    using RightPads = Sequence<1, 0>;
#elif 0
    // 3x3, 147x147
    constexpr index_t N  = 128;
    constexpr index_t C  = 64;
    constexpr index_t HI = 147;
    constexpr index_t WI = 147;
    constexpr index_t K  = 96;
    constexpr index_t Y  = 3;
    constexpr index_t X  = 3;

    using ConvStrides   = Sequence<2, 2>;
    using ConvDilations = Sequence<1, 1>;

Chao Liu's avatar
Chao Liu committed
420
421
    using LeftPads  = Sequence<0, 0>;
    using RightPads = Sequence<0, 0>;
Chao Liu's avatar
Chao Liu committed
422
#elif 0
Chao Liu's avatar
Chao Liu committed
423
    // 7x1, 73x73
Chao Liu's avatar
Chao Liu committed
424
    constexpr index_t N  = 128;
Chao Liu's avatar
Chao Liu committed
425
426
427
428
429
    constexpr index_t C  = 64;
    constexpr index_t HI = 73;
    constexpr index_t WI = 73;
    constexpr index_t K  = 64;
    constexpr index_t Y  = 7;
Chao Liu's avatar
Chao Liu committed
430
431
432
433
434
    constexpr index_t X  = 1;

    using ConvStrides   = Sequence<1, 1>;
    using ConvDilations = Sequence<1, 1>;

Chao Liu's avatar
Chao Liu committed
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
    using LeftPads  = Sequence<3, 0>;
    using RightPads = Sequence<3, 0>;
#elif 0
    // 3x3, 73x73
    constexpr index_t N  = 128;
    constexpr index_t C  = 64;
    constexpr index_t HI = 73;
    constexpr index_t WI = 73;
    constexpr index_t K  = 96;
    constexpr index_t Y  = 3;
    constexpr index_t X  = 3;

    using ConvStrides   = Sequence<1, 1>;
    using ConvDilations = Sequence<1, 1>;

Chao Liu's avatar
Chao Liu committed
450
451
    using LeftPads  = Sequence<0, 0>;
    using RightPads = Sequence<0, 0>;
Chao Liu's avatar
Chao Liu committed
452
#elif 0
Chao Liu's avatar
Chao Liu committed
453
    // 1x1, 14x14, stride 2
Chao Liu's avatar
Chao Liu committed
454
    constexpr index_t N  = 128;
Chao Liu's avatar
Chao Liu committed
455
456
457
458
    constexpr index_t C  = 1024;
    constexpr index_t HI = 14;
    constexpr index_t WI = 14;
    constexpr index_t K  = 2048;
Chao Liu's avatar
Chao Liu committed
459
460
461
    constexpr index_t Y  = 1;
    constexpr index_t X  = 1;

Chao Liu's avatar
Chao Liu committed
462
    using ConvStrides   = Sequence<2, 2>;
Chao Liu's avatar
Chao Liu committed
463
464
    using ConvDilations = Sequence<1, 1>;

Chao Liu's avatar
Chao Liu committed
465
466
    using LeftPads  = Sequence<0, 0>;
    using RightPads = Sequence<0, 0>;
467
#elif 0
Chao Liu's avatar
Chao Liu committed
468
    // 1x1, 14x14
Chao Liu's avatar
Chao Liu committed
469
    constexpr index_t N  = 128;
Chao Liu's avatar
Chao Liu committed
470
471
472
473
    constexpr index_t C  = 1024;
    constexpr index_t HI = 14;
    constexpr index_t WI = 14;
    constexpr index_t K  = 256;
Chao Liu's avatar
Chao Liu committed
474
475
476
477
478
479
    constexpr index_t Y  = 1;
    constexpr index_t X  = 1;

    using ConvStrides   = Sequence<1, 1>;
    using ConvDilations = Sequence<1, 1>;

Chao Liu's avatar
Chao Liu committed
480
481
    using LeftPads  = Sequence<0, 0>;
    using RightPads = Sequence<0, 0>;
Chao Liu's avatar
Chao Liu committed
482
#elif 0
Chao Liu's avatar
Chao Liu committed
483
    // 1x1, 14x14, stride 2
Chao Liu's avatar
Chao Liu committed
484
    constexpr index_t N  = 128;
Chao Liu's avatar
Chao Liu committed
485
    constexpr index_t C  = 1024;
Chao Liu's avatar
Chao Liu committed
486
487
    constexpr index_t HI = 14;
    constexpr index_t WI = 14;
Chao Liu's avatar
Chao Liu committed
488
    constexpr index_t K  = 512;
Chao Liu's avatar
Chao Liu committed
489
490
491
    constexpr index_t Y  = 1;
    constexpr index_t X  = 1;

Chao Liu's avatar
Chao Liu committed
492
    using ConvStrides   = Sequence<2, 2>;
Chao Liu's avatar
Chao Liu committed
493
494
    using ConvDilations = Sequence<1, 1>;

Chao Liu's avatar
Chao Liu committed
495
496
    using LeftPads  = Sequence<0, 0>;
    using RightPads = Sequence<0, 0>;
Chao Liu's avatar
Chao Liu committed
497
#elif 1
Chao Liu's avatar
Chao Liu committed
498
499
    // 3x3, 28x28
    constexpr index_t N  = 128;
Chao Liu's avatar
Chao Liu committed
500
    constexpr index_t C  = 128;
Chao Liu's avatar
Chao Liu committed
501
502
503
504
505
506
507
508
509
510
511
    constexpr index_t HI = 28;
    constexpr index_t WI = 28;
    constexpr index_t K  = 128;
    constexpr index_t Y  = 3;
    constexpr index_t X  = 3;

    using ConvStrides   = Sequence<1, 1>;
    using ConvDilations = Sequence<1, 1>;

    using LeftPads  = Sequence<1, 1>;
    using RightPads = Sequence<1, 1>;
512
#elif 1
Chao Liu's avatar
Chao Liu committed
513
    // 3x3, 14x14
Chao Liu's avatar
Chao Liu committed
514
    constexpr index_t N  = 128;
Chao Liu's avatar
Chao Liu committed
515
    constexpr index_t C  = 256;
Chao Liu's avatar
Chao Liu committed
516
517
518
    constexpr index_t HI = 14;
    constexpr index_t WI = 14;
    constexpr index_t K  = 256;
Chao Liu's avatar
Chao Liu committed
519
520
521
522
523
524
525
526
    constexpr index_t Y  = 3;
    constexpr index_t X  = 3;

    using ConvStrides   = Sequence<1, 1>;
    using ConvDilations = Sequence<1, 1>;

    using LeftPads  = Sequence<1, 1>;
    using RightPads = Sequence<1, 1>;
Chao Liu's avatar
Chao Liu committed
527
#elif 0
Chao Liu's avatar
Chao Liu committed
528
529
530
531
532
533
    // 1x1, 56x56, stride 2
    constexpr index_t N  = 128;
    constexpr index_t C  = 256;
    constexpr index_t HI = 56;
    constexpr index_t WI = 56;
    constexpr index_t K  = 128;
Chao Liu's avatar
Chao Liu committed
534
535
536
    constexpr index_t Y  = 1;
    constexpr index_t X  = 1;

Chao Liu's avatar
Chao Liu committed
537
    using ConvStrides   = Sequence<2, 2>;
Chao Liu's avatar
Chao Liu committed
538
539
    using ConvDilations = Sequence<1, 1>;

Chao Liu's avatar
Chao Liu committed
540
541
    using LeftPads  = Sequence<0, 0>;
    using RightPads = Sequence<0, 0>;
Chao Liu's avatar
Chao Liu committed
542
#elif 0
Chao Liu's avatar
Chao Liu committed
543
    // 7x7, 230x230 stride=2
Chao Liu's avatar
Chao Liu committed
544
    constexpr index_t N  = 128;
Chao Liu's avatar
Chao Liu committed
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
    constexpr index_t C  = 3;
    constexpr index_t HI = 230;
    constexpr index_t WI = 230;
    constexpr index_t K  = 64;
    constexpr index_t Y  = 7;
    constexpr index_t X  = 7;

    using ConvStrides   = Sequence<2, 2>;
    using ConvDilations = Sequence<1, 1>;

    using LeftPads  = Sequence<0, 0>;
    using RightPads = Sequence<0, 0>;
#elif 0
    // 1x1, 28x28, stride = 2
    constexpr index_t N  = 128;
    constexpr index_t C  = 512;
    constexpr index_t HI = 28;
    constexpr index_t WI = 28;
    constexpr index_t K  = 1024;
Chao Liu's avatar
Chao Liu committed
564
565
    constexpr index_t Y  = 1;
    constexpr index_t X  = 1;
Chao Liu's avatar
Chao Liu committed
566

Chao Liu's avatar
Chao Liu committed
567
    using ConvStrides   = Sequence<2, 2>;
Chao Liu's avatar
Chao Liu committed
568
569
    using ConvDilations = Sequence<1, 1>;

Chao Liu's avatar
Chao Liu committed
570
571
    using LeftPads  = Sequence<0, 0>;
    using RightPads = Sequence<0, 0>;
572
#elif 0
Chao Liu's avatar
Chao Liu committed
573
    // 1x1, 28x28, stride 2
Chao Liu's avatar
Chao Liu committed
574
    constexpr index_t N  = 128;
Chao Liu's avatar
Chao Liu committed
575
576
577
578
579
580
    constexpr index_t C  = 512;
    constexpr index_t HI = 28;
    constexpr index_t WI = 28;
    constexpr index_t K  = 256;
    constexpr index_t Y  = 1;
    constexpr index_t X  = 1;
Chao Liu's avatar
Chao Liu committed
581
582
583
584
585
586

    using ConvStrides   = Sequence<2, 2>;
    using ConvDilations = Sequence<1, 1>;

    using LeftPads  = Sequence<0, 0>;
    using RightPads = Sequence<0, 0>;
587
#elif 1
Chao Liu's avatar
Chao Liu committed
588
    // 1x1, 7x7
589
    constexpr index_t N  = 128;
Chao Liu's avatar
Chao Liu committed
590
    constexpr index_t C  = 512;
591
592
    constexpr index_t HI = 7;
    constexpr index_t WI = 7;
Chao Liu's avatar
Chao Liu committed
593
594
595
    constexpr index_t K  = 2048;
    constexpr index_t Y  = 1;
    constexpr index_t X  = 1;
596
597
598
599

    using ConvStrides   = Sequence<1, 1>;
    using ConvDilations = Sequence<1, 1>;

Chao Liu's avatar
Chao Liu committed
600
601
    using LeftPads  = Sequence<0, 0>;
    using RightPads = Sequence<0, 0>;
602
#elif 0
Chao Liu's avatar
Chao Liu committed
603
    // 3x3, 7x7
604
    constexpr index_t N  = 128;
Chao Liu's avatar
Chao Liu committed
605
606
607
608
609
610
    constexpr index_t C  = 512;
    constexpr index_t HI = 7;
    constexpr index_t WI = 7;
    constexpr index_t K  = 512;
    constexpr index_t Y  = 3;
    constexpr index_t X  = 3;
611
612
613
614

    using ConvStrides   = Sequence<1, 1>;
    using ConvDilations = Sequence<1, 1>;

Chao Liu's avatar
Chao Liu committed
615
616
    using LeftPads  = Sequence<1, 1>;
    using RightPads = Sequence<1, 1>;
Chao Liu's avatar
Chao Liu committed
617
#elif 0
Chao Liu's avatar
Chao Liu committed
618
    // 1x1, 56x56
619
    constexpr index_t N  = 128;
Chao Liu's avatar
Chao Liu committed
620
621
622
623
624
    constexpr index_t C  = 64;
    constexpr index_t HI = 56;
    constexpr index_t WI = 56;
    constexpr index_t K  = 64;
    constexpr index_t Y  = 1;
625
    constexpr index_t X  = 1;
626
627
628
629

    using ConvStrides   = Sequence<1, 1>;
    using ConvDilations = Sequence<1, 1>;

Chao Liu's avatar
Chao Liu committed
630
631
    using LeftPads  = Sequence<0, 0>;
    using RightPads = Sequence<0, 0>;
Chao Liu's avatar
Chao Liu committed
632
#elif 0
Chao Liu's avatar
Chao Liu committed
633
634
635
636
637
638
639
640
641
642
643
644
645
646
    // 3x3, 56x56
    constexpr index_t N  = 128;
    constexpr index_t C  = 64;
    constexpr index_t HI = 56;
    constexpr index_t WI = 56;
    constexpr index_t K  = 64;
    constexpr index_t Y  = 3;
    constexpr index_t X  = 3;

    using ConvStrides   = Sequence<1, 1>;
    using ConvDilations = Sequence<1, 1>;

    using LeftPads  = Sequence<1, 1>;
    using RightPads = Sequence<1, 1>;
Chao Liu's avatar
Chao Liu committed
647
#endif
Chao Liu's avatar
Chao Liu committed
648

Chao Liu's avatar
Chao Liu committed
649
650
651
    auto in_nchw_desc  = make_native_tensor_descriptor_packed(Sequence<N, C, HI, WI>{});
    auto wei_kcyx_desc = make_native_tensor_descriptor_packed(Sequence<K, C, Y, X>{});
    auto out_nkhw_desc = get_convolution_output_default_4d_tensor_descriptor(
Chao Liu's avatar
Chao Liu committed
652
        in_nchw_desc, wei_kcyx_desc, ConvStrides{}, ConvDilations{}, LeftPads{}, RightPads{});
Chao Liu's avatar
Chao Liu committed
653

Chao Liu's avatar
Chao Liu committed
654
655
656
    ostream_tensor_descriptor(in_nchw_desc, std::cout << "in_nchw_desc: ");
    ostream_tensor_descriptor(wei_kcyx_desc, std::cout << "wei_kcyx_desc: ");
    ostream_tensor_descriptor(out_nkhw_desc, std::cout << "out_nkhw_desc: ");
Chao Liu's avatar
Chao Liu committed
657
658
659
660
    print_array("LeftPads", to_multi_index(LeftPads{}));
    print_array("RightPads", to_multi_index(RightPads{}));
    print_array("ConvStrides", to_multi_index(ConvStrides{}));
    print_array("ConvDilations", to_multi_index(ConvDilations{}));
Chao Liu's avatar
Chao Liu committed
661

root's avatar
root committed
662
#if 0
root's avatar
root committed
663
    using in_data_t                  = float;
Chao Liu's avatar
Chao Liu committed
664
    constexpr index_t in_vector_size = 1;
root's avatar
root committed
665
    using acc_data_t                 = float;
Chao Liu's avatar
Chao Liu committed
666
    using out_data_t                 = float;
root's avatar
root committed
667
#elif 0
Chao Liu's avatar
tweak  
Chao Liu committed
668
669
670
671
    using in_data_t                  = float;
    constexpr index_t in_vector_size = 1;
    using acc_data_t                 = float;
    using out_data_t                 = int8_t;
Chao Liu's avatar
Chao Liu committed
672
#elif 1
Chao Liu's avatar
Chao Liu committed
673
    using in_data_t                  = int8_t;
Jing Zhang's avatar
Jing Zhang committed
674
    constexpr index_t in_vector_size = 16;
Chao Liu's avatar
Chao Liu committed
675
676
    using acc_data_t                 = int32_t;
    using out_data_t                 = int8_t;
Chao Liu's avatar
Chao Liu committed
677
678
679
680
681
682
#endif

    Tensor<in_data_t> in_nchw(make_HostTensorDescriptor(in_nchw_desc));
    Tensor<in_data_t> wei_kcyx(make_HostTensorDescriptor(wei_kcyx_desc));
    Tensor<out_data_t> out_nkhw_host(make_HostTensorDescriptor(out_nkhw_desc));
    Tensor<out_data_t> out_nkhw_device(make_HostTensorDescriptor(out_nkhw_desc));
Chao Liu's avatar
Chao Liu committed
683

Chao Liu's avatar
Chao Liu committed
684
    std::size_t num_thread = std::thread::hardware_concurrency();
Chao Liu's avatar
Chao Liu committed
685

Chao Liu's avatar
Chao Liu committed
686
    if(argc != 4)
Chao Liu's avatar
Chao Liu committed
687
    {
Chao Liu's avatar
Chao Liu committed
688
        printf("arg1: do_verification, arg2: do_log, arg3: nrepeat\n");
Chao Liu's avatar
Chao Liu committed
689
690
691
692
        exit(1);
    }

    bool do_verification = atoi(argv[1]);
Chao Liu's avatar
Chao Liu committed
693
694
    bool do_log          = atoi(argv[2]);
    index_t nrepeat      = atoi(argv[3]);
695
696
697

    if(do_verification)
    {
root's avatar
root committed
698
#if 0
699
        in_nchw.GenerateTensorValue(GeneratorTensor_1{}, num_thread);
Chao Liu's avatar
Chao Liu committed
700
        wei_kcyx.GenerateTensorValue(GeneratorTensor_1{}, num_thread);
Chao Liu's avatar
Chao Liu committed
701
702
#elif 0
        in_nchw.GenerateTensorValue(GeneratorTensor_1{}, num_thread);
Chao Liu's avatar
Chao Liu committed
703
        wei_kcyx.GenerateTensorValue(GeneratorTensor_2{-5, 5}, num_thread);
704
#elif 0
705
        in_nchw.GenerateTensorValue(GeneratorTensor_2{-5, 5}, num_thread);
706
        wei_kcyx.GenerateTensorValue(GeneratorTensor_1{}, num_thread);
Chao Liu's avatar
Chao Liu committed
707
#elif 1
708
        in_nchw.GenerateTensorValue(GeneratorTensor_2{-5, 5}, num_thread);
Chao Liu's avatar
Chao Liu committed
709
        wei_kcyx.GenerateTensorValue(GeneratorTensor_2{-5, 5}, num_thread);
Chao Liu's avatar
Chao Liu committed
710
#elif 0
711
712
713
714
715
716
        in_nchw.GenerateTensorValue(GeneratorTensor_2{1, 5}, num_thread);

        auto gen_wei = [](auto... is) {
            return GeneratorTensor_2{1, 5}(is...) * GeneratorTensor_Checkboard{}(is...);
        };
        wei_kcyx.GenerateTensorValue(gen_wei, num_thread);
Chao Liu's avatar
Chao Liu committed
717
#endif
718
    }
Chao Liu's avatar
Chao Liu committed
719

Chao Liu's avatar
Chao Liu committed
720
#if 0
721
    device_convolution_forward_implicit_gemm_v4r1_nchw_kcyx_nkhw(in_nchw_desc,
722
723
724
725
726
727
728
729
730
731
                                                                 in_nchw,
                                                                 wei_kcyx_desc,
                                                                 wei_kcyx,
                                                                 out_nkhw_desc,
                                                                 out_nkhw_device,
                                                                 ConvStrides{},
                                                                 ConvDilations{},
                                                                 LeftPads{},
                                                                 RightPads{},
                                                                 nrepeat);
Chao Liu's avatar
Chao Liu committed
732
#elif 0
733
734
735
736
737
738
739
740
741
742
743
    device_convolution_forward_implicit_gemm_v4r4_nchw_kcyx_nkhw(in_nchw_desc,
                                                                 in_nchw,
                                                                 wei_kcyx_desc,
                                                                 wei_kcyx,
                                                                 out_nkhw_desc,
                                                                 out_nkhw_device,
                                                                 ConvStrides{},
                                                                 ConvDilations{},
                                                                 LeftPads{},
                                                                 RightPads{},
                                                                 nrepeat);
Chao Liu's avatar
Chao Liu committed
744
#elif 0
Chao Liu's avatar
Chao Liu committed
745
746
747
748
749
750
751
752
753
754
755
    device_convolution_forward_implicit_gemm_v4r4_nhwc_kyxc_nhwk(in_nchw_desc,
                                                                 in_nchw,
                                                                 wei_kcyx_desc,
                                                                 wei_kcyx,
                                                                 out_nkhw_desc,
                                                                 out_nkhw_device,
                                                                 ConvStrides{},
                                                                 ConvDilations{},
                                                                 LeftPads{},
                                                                 RightPads{},
                                                                 nrepeat);
root's avatar
tweak  
root committed
756
#elif 0
Chao Liu's avatar
Chao Liu committed
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
    device_dynamic_convolution_forward_implicit_gemm_v4r4_nchw_kcyx_nkhw<in_data_t,
                                                                         in_vector_size,
                                                                         acc_data_t,
                                                                         out_data_t>

        (in_nchw_desc,
         in_nchw,
         wei_kcyx_desc,
         wei_kcyx,
         out_nkhw_desc,
         out_nkhw_device,
         ConvStrides{},
         ConvDilations{},
         LeftPads{},
         RightPads{},
         nrepeat);
root's avatar
root committed
773
#elif 0
Chao Liu's avatar
Chao Liu committed
774
775
776
    device_dynamic_convolution_forward_implicit_gemm_v4r4_nhwc_kyxc_nhwk<in_data_t,
                                                                         in_vector_size,
                                                                         acc_data_t,
777
                                                                         out_data_t>
Chao Liu's avatar
Chao Liu committed
778
779
780
781
782
783
784
785
786
787
788
789

        (in_nchw_desc,
         in_nchw,
         wei_kcyx_desc,
         wei_kcyx,
         out_nkhw_desc,
         out_nkhw_device,
         ConvStrides{},
         ConvDilations{},
         LeftPads{},
         RightPads{},
         nrepeat);
root's avatar
root committed
790
#elif 1
Jing Zhang's avatar
mock up  
Jing Zhang committed
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
    device_dynamic_convolution_forward_implicit_gemm_v5r1_nchw_kcyx_nkhw<in_data_t,
                                                                         in_vector_size,
                                                                         acc_data_t,
                                                                         out_data_t>(
        in_nchw_desc,
        in_nchw,
        wei_kcyx_desc,
        wei_kcyx,
        out_nkhw_desc,
        out_nkhw_device,
        ConvStrides{},
        ConvDilations{},
        LeftPads{},
        RightPads{},
        nrepeat);
806
#endif
Chao Liu's avatar
Chao Liu committed
807

808
    if(do_verification)
809
    {
Chao Liu's avatar
Chao Liu committed
810
811
812
813
814
815
816
817
        host_direct_convolution(in_nchw,
                                wei_kcyx,
                                out_nkhw_host,
                                ConvStrides{},
                                ConvDilations{},
                                LeftPads{},
                                RightPads{});

818
        check_error(out_nkhw_host, out_nkhw_device);
Chao Liu's avatar
Chao Liu committed
819

Chao Liu's avatar
Chao Liu committed
820
821
822
823
824
825
826
        if(do_log)
        {
            LogRange(std::cout << "in_nchw : ", in_nchw.mData, ",") << std::endl;
            LogRange(std::cout << "wei_kcyx: ", wei_kcyx.mData, ",") << std::endl;
            LogRange(std::cout << "out_nkhw_host  : ", out_nkhw_host.mData, ",") << std::endl;
            LogRange(std::cout << "out_nkhw_device: ", out_nkhw_device.mData, ",") << std::endl;
        }
827
    }
828
}