conv_driver.cpp 24.5 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"
19

Chao Liu's avatar
Chao Liu committed
20
int main(int argc, char* argv[])
Chao Liu's avatar
Chao Liu committed
21
{
Chao Liu's avatar
Chao Liu committed
22
23
    using namespace ck;

24
#if 0
Chao Liu's avatar
Chao Liu committed
25
    constexpr index_t N  = 1;
Chao Liu's avatar
Chao Liu committed
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
    constexpr index_t C  = 4;
    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>;

    using LeftPads  = Sequence<0, 0>;
    using RightPads = Sequence<0, 0>;
#elif 0
    constexpr index_t N  = 1;
    constexpr index_t C  = 4;
Chao Liu's avatar
Chao Liu committed
41
42
    constexpr index_t HI = 540;
    constexpr index_t WI = 960;
Chao Liu's avatar
Chao Liu committed
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
    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>;

    using LeftPads  = Sequence<0, 0>;
    using RightPads = Sequence<0, 0>;
#elif 0
    constexpr index_t N  = 1;
    constexpr index_t C  = 4;
    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>;

    using LeftPads  = Sequence<0, 0>;
    using RightPads = Sequence<0, 0>;
Chao Liu's avatar
Chao Liu committed
66
#elif 0
Chao Liu's avatar
Chao Liu committed
67
    constexpr index_t N  = 1;
Chao Liu's avatar
testing  
Chao Liu committed
68
    constexpr index_t C  = 4;
Chao Liu's avatar
Chao Liu committed
69
70
71
72
73
74
75
76
77
    constexpr index_t HI = 1080;
    constexpr index_t WI = 1920;
    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
78
79
    using LeftPads  = Sequence<1, 1>;
    using RightPads = Sequence<1, 1>;
Chao Liu's avatar
Chao Liu committed
80
#elif 0
Chao Liu's avatar
Chao Liu committed
81
82
83
84
85
86
87
88
89
90
91
    constexpr index_t N  = 1;
    constexpr index_t C  = 4;
    constexpr index_t HI = 1080;
    constexpr index_t WI = 1920;
    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
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
    using LeftPads  = Sequence<1, 1>;
    using RightPads = Sequence<1, 1>;
#elif 0
    constexpr index_t N  = 1;
    constexpr index_t C  = 4;
    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>;

    using LeftPads  = Sequence<1, 1>;
    using RightPads = Sequence<1, 1>;
Chao Liu's avatar
test  
Chao Liu committed
108
#elif 0
Chao Liu's avatar
Chao Liu committed
109
110
111
112
113
    constexpr index_t N  = 1;
    constexpr index_t C  = 4;
    constexpr index_t HI = 270;
    constexpr index_t WI = 480;
    constexpr index_t K  = 16;
Chao Liu's avatar
Chao Liu committed
114
115
116
117
118
119
120
121
122
    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
123
124
125
126
127
128
129
130
131
132
133
134
135
136
    // 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>;
137
#elif 0
138
139
140
141
142
143
144
145
    // 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;
146

147
    using ConvStrides   = Sequence<2, 2>;
148
149
150
151
    using ConvDilations = Sequence<1, 1>;

    using LeftPads  = Sequence<0, 0>;
    using RightPads = Sequence<0, 0>;
152
#elif 1
Chao Liu's avatar
Chao Liu committed
153
    // 3x3, 71x71
Chao Liu's avatar
Chao Liu committed
154
    constexpr index_t N  = 128;
Chao Liu's avatar
Chao Liu committed
155
156
157
158
159
160
    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
161

Chao Liu's avatar
Chao Liu committed
162
    using ConvStrides   = Sequence<2, 2>;
163
    using ConvDilations = Sequence<1, 1>;
Chao Liu's avatar
Chao Liu committed
164

Chao Liu's avatar
Chao Liu committed
165
166
    using LeftPads  = Sequence<1, 1>;
    using RightPads = Sequence<1, 1>;
Chao Liu's avatar
Chao Liu committed
167
#elif 1
Chao Liu's avatar
Chao Liu committed
168
    // 1x1, 8x8
169
    constexpr index_t N  = 128;
Chao Liu's avatar
Chao Liu committed
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
    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;
203
    constexpr index_t K  = 128;
Chao Liu's avatar
Chao Liu committed
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
    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>;
227
#elif 1
Chao Liu's avatar
Chao Liu committed
228
229
230
    // 7x1, 17x17
    constexpr index_t N  = 128;
    constexpr index_t C  = 128;
231
232
    constexpr index_t HI = 17;
    constexpr index_t WI = 17;
233
    constexpr index_t K  = 128;
Chao Liu's avatar
Chao Liu committed
234
235
236
237
238
239
240
241
    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>;
242
#elif 0
Chao Liu's avatar
Chao Liu committed
243
244
245
246
247
248
    // 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;
249
250
251
252
253
254
255
256
257
    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
258
259
260
261
262
263
    // 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;
264
265
    constexpr index_t Y  = 3;
    constexpr index_t X  = 3;
Chao Liu's avatar
Chao Liu committed
266

Chao Liu's avatar
Chao Liu committed
267
    using ConvStrides   = Sequence<2, 2>;
268
269
    using ConvDilations = Sequence<1, 1>;

270
271
272
    using LeftPads  = Sequence<0, 0>;
    using RightPads = Sequence<0, 0>;
#elif 0
Chao Liu's avatar
Chao Liu committed
273
    // 3x3, 147x147
274
    constexpr index_t N  = 128;
275
    constexpr index_t C  = 128;
Chao Liu's avatar
Chao Liu committed
276
277
    constexpr index_t HI = 147;
    constexpr index_t WI = 147;
278
    constexpr index_t K  = 128;
279
280
281
    constexpr index_t Y  = 3;
    constexpr index_t X  = 3;

Chao Liu's avatar
Chao Liu committed
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
    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>;
298
299
    using ConvDilations = Sequence<1, 1>;

300
301
    using LeftPads  = Sequence<0, 0>;
    using RightPads = Sequence<0, 0>;
Chao Liu's avatar
Chao Liu committed
302
#elif 0
Chao Liu's avatar
Chao Liu committed
303
304
305
306
307
308
309
310
    // 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
311

Chao Liu's avatar
Chao Liu committed
312
    using ConvStrides   = Sequence<2, 2>;
Chao Liu's avatar
Chao Liu committed
313
314
    using ConvDilations = Sequence<1, 1>;

Chao Liu's avatar
Chao Liu committed
315
316
    using LeftPads  = Sequence<0, 0>;
    using RightPads = Sequence<0, 0>;
Chao Liu's avatar
Chao Liu committed
317
#elif 0
Chao Liu's avatar
Chao Liu committed
318
    // 1x1, 35x35
Chao Liu's avatar
Chao Liu committed
319
    constexpr index_t N  = 128;
Chao Liu's avatar
Chao Liu committed
320
321
322
323
    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
324
325
326
327
328
329
    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
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
    // 3x3, 35x35, stride 2
Chao Liu's avatar
Chao Liu committed
334
    constexpr index_t N  = 128;
Chao Liu's avatar
Chao Liu committed
335
336
337
    constexpr index_t C  = 288;
    constexpr index_t HI = 35;
    constexpr index_t WI = 35;
Chao Liu's avatar
Chao Liu committed
338
    constexpr index_t K  = 384;
Chao Liu's avatar
Chao Liu committed
339
340
    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
    // 1x3, 8x8
Chao Liu's avatar
Chao Liu committed
349
    constexpr index_t N  = 128;
Chao Liu's avatar
Chao Liu committed
350
    constexpr index_t C  = 384;
Chao Liu's avatar
Chao Liu committed
351
352
    constexpr index_t HI = 8;
    constexpr index_t WI = 8;
Chao Liu's avatar
Chao Liu committed
353
    constexpr index_t K  = 448;
Chao Liu's avatar
Chao Liu committed
354
    constexpr index_t Y  = 1;
Chao Liu's avatar
Chao Liu committed
355
    constexpr index_t X  = 3;
Chao Liu's avatar
Chao Liu committed
356
357
358
359

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

Chao Liu's avatar
Chao Liu committed
360
361
    using LeftPads  = Sequence<0, 1>;
    using RightPads = Sequence<0, 1>;
Chao Liu's avatar
Chao Liu committed
362
#elif 0
Chao Liu's avatar
Chao Liu committed
363
    // 3x1, 8x8
Chao Liu's avatar
Chao Liu committed
364
    constexpr index_t N  = 128;
Chao Liu's avatar
Chao Liu committed
365
366
367
368
369
    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
370
371
372
373
374
    constexpr index_t X  = 1;

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

Chao Liu's avatar
Chao Liu committed
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
    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
390
391
    using LeftPads  = Sequence<0, 0>;
    using RightPads = Sequence<0, 0>;
Chao Liu's avatar
Chao Liu committed
392
#elif 0
Chao Liu's avatar
Chao Liu committed
393
    // 7x1, 73x73
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  = 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
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<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
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
    // 1x1, 14x14, stride 2
Chao Liu's avatar
Chao Liu committed
424
    constexpr index_t N  = 128;
Chao Liu's avatar
Chao Liu committed
425
426
427
428
    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
429
430
431
    constexpr index_t Y  = 1;
    constexpr index_t X  = 1;

Chao Liu's avatar
Chao Liu committed
432
    using ConvStrides   = Sequence<2, 2>;
Chao Liu's avatar
Chao Liu committed
433
434
    using ConvDilations = Sequence<1, 1>;

Chao Liu's avatar
Chao Liu committed
435
436
    using LeftPads  = Sequence<0, 0>;
    using RightPads = Sequence<0, 0>;
437
#elif 0
Chao Liu's avatar
Chao Liu committed
438
    // 1x1, 14x14
Chao Liu's avatar
Chao Liu committed
439
    constexpr index_t N  = 128;
Chao Liu's avatar
Chao Liu committed
440
441
442
443
    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
444
445
446
447
448
449
    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
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
    constexpr index_t C  = 1024;
Chao Liu's avatar
Chao Liu committed
456
457
    constexpr index_t HI = 14;
    constexpr index_t WI = 14;
Chao Liu's avatar
Chao Liu committed
458
    constexpr index_t K  = 512;
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>;
Chao Liu's avatar
Chao Liu committed
467
#elif 1
Chao Liu's avatar
Chao Liu committed
468
469
    // 3x3, 28x28
    constexpr index_t N  = 128;
Chao Liu's avatar
Chao Liu committed
470
    constexpr index_t C  = 128;
Chao Liu's avatar
Chao Liu committed
471
472
473
474
475
476
477
478
479
480
481
    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>;
482
#elif 1
Chao Liu's avatar
Chao Liu committed
483
    // 3x3, 14x14
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  = 256;
Chao Liu's avatar
Chao Liu committed
486
487
488
    constexpr index_t HI = 14;
    constexpr index_t WI = 14;
    constexpr index_t K  = 256;
Chao Liu's avatar
Chao Liu committed
489
490
491
492
493
494
495
496
    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
497
#elif 0
Chao Liu's avatar
Chao Liu committed
498
499
500
501
502
503
    // 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
504
505
506
    constexpr index_t Y  = 1;
    constexpr index_t X  = 1;

Chao Liu's avatar
Chao Liu committed
507
    using ConvStrides   = Sequence<2, 2>;
Chao Liu's avatar
Chao Liu committed
508
509
    using ConvDilations = Sequence<1, 1>;

Chao Liu's avatar
Chao Liu committed
510
511
    using LeftPads  = Sequence<0, 0>;
    using RightPads = Sequence<0, 0>;
Chao Liu's avatar
Chao Liu committed
512
#elif 0
Chao Liu's avatar
Chao Liu committed
513
    // 7x7, 230x230 stride=2
Chao Liu's avatar
Chao Liu committed
514
    constexpr index_t N  = 128;
Chao Liu's avatar
Chao Liu committed
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
    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
534
535
    constexpr index_t Y  = 1;
    constexpr index_t X  = 1;
Chao Liu's avatar
Chao Liu committed
536

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>;
542
#elif 0
Chao Liu's avatar
Chao Liu committed
543
    // 1x1, 28x28, 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
    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
551
552
553
554
555
556

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

    using LeftPads  = Sequence<0, 0>;
    using RightPads = Sequence<0, 0>;
557
#elif 1
Chao Liu's avatar
Chao Liu committed
558
    // 1x1, 7x7
559
    constexpr index_t N  = 128;
Chao Liu's avatar
Chao Liu committed
560
    constexpr index_t C  = 512;
561
562
    constexpr index_t HI = 7;
    constexpr index_t WI = 7;
Chao Liu's avatar
Chao Liu committed
563
564
565
    constexpr index_t K  = 2048;
    constexpr index_t Y  = 1;
    constexpr index_t X  = 1;
566
567
568
569

    using ConvStrides   = Sequence<1, 1>;
    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
    // 3x3, 7x7
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 = 7;
    constexpr index_t WI = 7;
    constexpr index_t K  = 512;
    constexpr index_t Y  = 3;
    constexpr index_t X  = 3;
581
582
583
584

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

Chao Liu's avatar
Chao Liu committed
585
586
    using LeftPads  = Sequence<1, 1>;
    using RightPads = Sequence<1, 1>;
Chao Liu's avatar
Chao Liu committed
587
#elif 0
Chao Liu's avatar
Chao Liu committed
588
    // 1x1, 56x56
589
    constexpr index_t N  = 128;
Chao Liu's avatar
Chao Liu committed
590
591
592
593
594
    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;
595
    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>;
Chao Liu's avatar
Chao Liu committed
602
#elif 0
Chao Liu's avatar
Chao Liu committed
603
604
605
606
607
608
609
610
611
612
613
614
615
616
    // 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
617
#endif
Chao Liu's avatar
Chao Liu committed
618

Chao Liu's avatar
Chao Liu committed
619
620
621
    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
622
        in_nchw_desc, wei_kcyx_desc, ConvStrides{}, ConvDilations{}, LeftPads{}, RightPads{});
Chao Liu's avatar
Chao Liu committed
623

Chao Liu's avatar
Chao Liu committed
624
625
626
    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
627
628
629
630
    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
631

Chao Liu's avatar
Chao Liu committed
632
#if 1
Chao Liu's avatar
Chao Liu committed
633
634
    using in_data_t  = float;
    using out_data_t = float;
Chao Liu's avatar
Chao Liu committed
635
636
637
638
639
640
641
642
643
#else
    using in_data_t  = half_float::half;
    using out_data_t = half_float::half;
#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
644

Chao Liu's avatar
Chao Liu committed
645
    std::size_t num_thread = std::thread::hardware_concurrency();
Chao Liu's avatar
Chao Liu committed
646

Chao Liu's avatar
Chao Liu committed
647
648
649
650
651
652
653
    if(argc != 3)
    {
        printf("arg1: do_verification, arg2: nrepeat\n");
        exit(1);
    }

    bool do_verification = atoi(argv[1]);
Chao Liu's avatar
Chao Liu committed
654
    index_t nrepeat      = atoi(argv[2]);
655
656
657

    if(do_verification)
    {
Chao Liu's avatar
Chao Liu committed
658
#if 0
659
        in_nchw.GenerateTensorValue(GeneratorTensor_1{}, num_thread);
Chao Liu's avatar
Chao Liu committed
660
        wei_kcyx.GenerateTensorValue(GeneratorTensor_1{}, num_thread);
Chao Liu's avatar
Chao Liu committed
661
662
#elif 0
        in_nchw.GenerateTensorValue(GeneratorTensor_1{}, num_thread);
Chao Liu's avatar
bug fix  
Chao Liu committed
663
        wei_kcyx.GenerateTensorValue(GeneratorTensor_3{}, num_thread);
664
#elif 0
665
        in_nchw.GenerateTensorValue(GeneratorTensor_2{-5, 5}, num_thread);
666
        wei_kcyx.GenerateTensorValue(GeneratorTensor_1{}, num_thread);
Chao Liu's avatar
Chao Liu committed
667
#elif 1
668
        in_nchw.GenerateTensorValue(GeneratorTensor_2{-5, 5}, num_thread);
Chao Liu's avatar
Chao Liu committed
669
        wei_kcyx.GenerateTensorValue(GeneratorTensor_2{-5, 5}, num_thread);
Chao Liu's avatar
Chao Liu committed
670
#elif 0
671
672
673
674
675
676
        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
677
#endif
678
    }
Chao Liu's avatar
Chao Liu committed
679

Chao Liu's avatar
Chao Liu committed
680
#if 0
681
    device_convolution_forward_implicit_gemm_v4r1_nchw_kcyx_nkhw(in_nchw_desc,
682
683
684
685
686
687
688
689
690
691
                                                                 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
692
#elif 0
693
694
695
696
697
698
699
700
701
702
703
    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
704
705
706
707
708
709
710
711
712
713
714
715
#elif 0
    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);
716
#elif 1
717
718
719
720
721
722
723
724
725
726
727
    device_dynamic_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);
728
729
730
731
732
733
734
735
736
737
738
739
#elif 1
    device_dynamic_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);
740
#endif
Chao Liu's avatar
Chao Liu committed
741

742
    if(do_verification)
743
    {
Chao Liu's avatar
Chao Liu committed
744
#if 0
745
746
        if(Y == 3 && X == 3 && ConvStrides{}[0] == 1 && ConvStrides{}[1] == 1 &&
           ConvDilations{}[0] == 1 && ConvDilations{}[1] == 1)
747
        {
Chao Liu's avatar
Chao Liu committed
748
749
            host_winograd_3x3_convolution(
                in_nchw, wei_kcyx, out_nkhw_host, LeftPads{}, RightPads{});
750
751
        }
        else
Chao Liu's avatar
Chao Liu committed
752
#endif
753
        {
754
755
756
757
758
            host_direct_convolution(in_nchw,
                                    wei_kcyx,
                                    out_nkhw_host,
                                    ConvStrides{},
                                    ConvDilations{},
Chao Liu's avatar
Chao Liu committed
759
760
                                    LeftPads{},
                                    RightPads{});
761
762
        }
        check_error(out_nkhw_host, out_nkhw_device);
Chao Liu's avatar
Chao Liu committed
763

Chao Liu's avatar
Chao Liu committed
764
#if 0
765
        LogRange(std::cout << "in_nchw : ", in_nchw.mData, ",") << std::endl;
Chao Liu's avatar
Chao Liu committed
766
        LogRange(std::cout << "wei_kcyx: ", wei_kcyx.mData, ",") << std::endl;
767
768
        LogRange(std::cout << "out_nkhw_host  : ", out_nkhw_host.mData, ",") << std::endl;
        LogRange(std::cout << "out_nkhw_device: ", out_nkhw_device.mData, ",") << std::endl;
Chao Liu's avatar
Chao Liu committed
769
#endif
770
    }
771
}