oneAPI Deep Neural Network Library (oneDNN)  1.6.0
Performance library for Deep Learning
dnnl_types.h
Go to the documentation of this file.
1 /*******************************************************************************
2 * Copyright 2016-2020 Intel Corporation
3 *
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at
7 *
8 * http://www.apache.org/licenses/LICENSE-2.0
9 *
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
15 *******************************************************************************/
16 
19 
20 #ifndef DNNL_TYPES_H
21 #define DNNL_TYPES_H
22 
23 #ifdef __cplusplus
24 extern "C" {
25 #endif
26 
28 #include <stddef.h>
29 #include <stdint.h>
31 
34 
37 
39 typedef enum {
55 
57 
60 
62 typedef enum {
66  dnnl_f16 = 1,
68  dnnl_bf16 = 2,
70  dnnl_f32 = 3,
72  dnnl_s32 = 4,
74  dnnl_s8 = 5,
76  dnnl_u8 = 6,
78 
80 typedef enum {
95 
164 typedef enum {
170 
171  // Semantic agnostic section
172  // The physical order of dimensions is defined by the permutation of the
173  // characters, assuming that ab..z defines the natural order.
174 
175  // Plain formats
176 
183 
184  // Permuted plain formats
185 
206 
207  // Opaque blocked formats
208 
209  dnnl_Abc16a,
210  dnnl_ABc16a16b,
211  dnnl_ABc32a32b,
212  dnnl_ABc4a4b,
215  dnnl_ABc16b16a,
216  dnnl_Abc4a,
221  dnnl_ABc4b16a4b,
222  dnnl_ABc2b8a4b,
223  dnnl_ABc16b16a4b,
224  dnnl_ABc16b16a2b,
225  dnnl_ABc4b4a,
226  dnnl_ABc8a16b2a,
227  dnnl_ABc8a8b,
228  dnnl_ABc8a4b,
231  dnnl_ABc8b16a2b,
232  dnnl_BAc8a16b2a,
233  dnnl_ABc8b8a,
234  dnnl_Abcd16a,
235  dnnl_Abcd8a,
236  dnnl_ABcd16a16b,
237  dnnl_Abcd32a,
238  dnnl_ABcd32a32b,
241  dnnl_ABcd16b16a,
242  dnnl_aBCd16b16c,
243  dnnl_aBCd16c16b,
244  dnnl_Abcd4a,
249  dnnl_ABcd4b16a4b,
250  dnnl_ABcd16b16a4b,
251  dnnl_ABcd16b16a2b,
252  dnnl_ABcd4b4a,
253  dnnl_ABcd4a4b,
254  dnnl_aBCd2c4b2c,
255  dnnl_aBCd4b8c2b,
256  dnnl_aBCd4c16b4c,
257  dnnl_aBCd2c8b4c,
258  dnnl_aBCd16c16b4c,
259  dnnl_aBCd16c16b2c,
260  dnnl_aBCd4c4b,
261  dnnl_aBCd4b4c,
262  dnnl_ABcd8a16b2a,
263  dnnl_ABcd2b8a4b,
264  dnnl_ABcd8a8b,
265  dnnl_ABcd8a4b,
268  dnnl_aBCd4c8b2c,
269  dnnl_ABcd8b16a2b,
270  dnnl_aBCd8b16c2b,
271  dnnl_BAcd8a16b2a,
274  dnnl_aBCd8b8c,
275  dnnl_aBCd8b4c,
276  dnnl_aBCd8c16b2c,
277  dnnl_ABcde8a16b2a,
278  dnnl_aCBd8b16c2b,
279  dnnl_aBCd8c8b,
280  dnnl_Abcde16a,
281  dnnl_Abcde32a,
282  dnnl_ABcde16a16b,
283  dnnl_BAcde8a16b2a,
292  dnnl_ABcde16b16a,
293  dnnl_aBCde16b16c,
294  dnnl_aBCde16c16b,
295  dnnl_aBCde2c8b4c,
296  dnnl_Abcde4a,
301  dnnl_ABcde4b4a,
302  dnnl_ABcde4a4b,
303  dnnl_aBCde4b4c,
304  dnnl_aBCde2c4b2c,
305  dnnl_aBCde4b8c2b,
306  dnnl_aBCde4c16b4c,
307  dnnl_aBCde16c16b4c,
308  dnnl_aBCde16c16b2c,
309  dnnl_aBCde4c4b,
310  dnnl_Abcde8a,
311  dnnl_ABcde8a8b,
312  dnnl_ABcde8a4b,
313  dnnl_BAcde16b16a,
316  dnnl_ABcde8b16a2b,
317  dnnl_aBCde8b16c2b,
318  dnnl_aBCde4c8b2c,
319  dnnl_aCBde8b16c2b,
320  dnnl_ABcde8b8a,
321  dnnl_ABcde32a32b,
322  dnnl_aBCde8b8c,
323  dnnl_aBCde8b4c,
324  dnnl_ABc4a8b8a4b,
325  dnnl_ABcd4a8b8a4b,
326  dnnl_ABcde4a8b8a4b,
327  dnnl_BAc4b8a8b4a,
328  dnnl_BAcd4b8a8b4a,
329  dnnl_BAcde4b8a8b4a,
330  dnnl_ABcd2a8b8a2b,
331  dnnl_aBCd4b8c8b4c,
332  dnnl_aBCde4b8c8b4c,
333  dnnl_aBCde2b8c8b2c,
334  dnnl_aBCde8c16b2c,
335  dnnl_aBCde8c8b,
340  dnnl_aBCdef16b16c,
341  dnnl_aBCdef16c16b,
342  dnnl_aBCdef4c16b4c,
345  dnnl_aBCdef4c8b2c,
350  dnnl_aBCdef4c4b,
351  dnnl_aBCdef4b4c,
352  dnnl_aBCdef2c4b2c,
353  dnnl_aBCdef4b8c2b,
354  dnnl_aBCdef8b8c,
355  dnnl_aBCdef8b4c,
356  dnnl_aBCdef8c16b2c,
357  dnnl_aBCdef4b8c8b4c,
358  dnnl_aBCdef8b16c2b,
359  dnnl_aCBdef8b16c2b,
360  dnnl_aBCdef8c8b,
361  dnnl_aBdc16b,
362  dnnl_aBdC16b2c,
363  dnnl_aBdC16b4c,
364  dnnl_aBdc4b,
365  dnnl_aBdc8b,
366  dnnl_aBdec16b,
367  dnnl_aBdeC16b2c,
368  dnnl_aBdeC16b4c,
369  dnnl_aBdec32b,
370  dnnl_aBdec4b,
371  dnnl_aBdec8b,
372  dnnl_aBdefc16b,
373  dnnl_aBdefC16b2c,
374  dnnl_aCBdef16c16b,
375  dnnl_aBdefc4b,
376  dnnl_aBdefc8b,
377  dnnl_Abcdef16a,
378  dnnl_Abcdef32a,
379  dnnl_Acb16a,
380  dnnl_AcB16a2b,
381  dnnl_AcB16a4b,
382  dnnl_Acb4a,
383  dnnl_Acb8a,
384  dnnl_aCBd16b16c,
385  dnnl_aCBd16c16b,
386  dnnl_aCBde16b16c,
387  dnnl_aCBde16c16b,
388  dnnl_Acdb16a,
389  dnnl_AcdB16a2b,
390  dnnl_AcdB16a4b,
391  dnnl_Acdb32a,
392  dnnl_Acdb4a,
393  dnnl_Acdb8a,
394  dnnl_Acdeb16a,
395  dnnl_AcdeB16a2b,
396  dnnl_Acdeb4a,
397  dnnl_Acdeb8a,
398  dnnl_BAc16a16b,
399  dnnl_BAc16b16a,
400  dnnl_BAcd16a16b,
401  dnnl_BAcd16b16a,
402  dnnl_aCBd4c8b8c4b,
403  dnnl_aCBde4c8b8c4b,
404  dnnl_aCBdef4c8b8c4b,
405  dnnl_BAcde16a16b,
406  dnnl_aCBdef16b16c,
407 
411 
412  // Aliases
413 
438 
471 
488 
523 
524  // Opaque data types, are not to be used explicitly
525 
526  // data
563  dnnl_NCw16n16c = dnnl_ABc16a16b,
564  dnnl_NCdhw16n16c = dnnl_ABcde16a16b,
565  dnnl_NChw16n16c = dnnl_ABcd16a16b,
566  dnnl_NCw32n32c = dnnl_ABc32a32b,
567  dnnl_NChw32n32c = dnnl_ABcd32a32b,
568  dnnl_NCdhw32n32c = dnnl_ABcde32a32b,
569 
570  // weights, 3D
571  dnnl_IOw16o16i = dnnl_BAc16a16b,
572  dnnl_IOw16i16o = dnnl_BAc16b16a,
573  dnnl_OIw16i16o = dnnl_ABc16b16a,
574  dnnl_OIw16o16i = dnnl_ABc16a16b,
575  dnnl_Oiw16o = dnnl_Abc16a,
576  dnnl_OIw4i16o4i = dnnl_ABc4b16a4b,
577  dnnl_OIw2i8o4i = dnnl_ABc2b8a4b,
578  dnnl_OIw16i16o4i = dnnl_ABc16b16a4b,
579  dnnl_OIw16i16o2i = dnnl_ABc16b16a2b,
580  dnnl_OIw4i4o = dnnl_ABc4b4a,
581  dnnl_OIw4o4i = dnnl_ABc4a4b,
582  dnnl_Oiw4o = dnnl_Abc4a,
583  dnnl_OIw8i16o2i = dnnl_ABc8b16a2b,
584  dnnl_OIw8i8o = dnnl_ABc8b8a,
585  dnnl_OIw8o16i2o = dnnl_ABc8a16b2a,
586  dnnl_IOw8o16i2o = dnnl_BAc8a16b2a,
587  dnnl_OIw8o8i = dnnl_ABc8a8b,
588  dnnl_OIw8o4i = dnnl_ABc8a4b,
589  dnnl_Owi16o = dnnl_Acb16a,
590  dnnl_OwI16o2i = dnnl_AcB16a2b,
591  dnnl_OwI16o4i = dnnl_AcB16a4b,
592  dnnl_Owi4o = dnnl_Acb4a,
593  dnnl_Owi8o = dnnl_Acb8a,
594 
595  // weights, 4D
596  dnnl_IOhw16i16o = dnnl_BAcd16b16a,
597  dnnl_IOhw16o16i = dnnl_BAcd16a16b,
598  dnnl_Ohwi16o = dnnl_Acdb16a,
599  dnnl_OhwI16o2i = dnnl_AcdB16a2b,
600  dnnl_OhwI16o4i = dnnl_AcdB16a4b,
601  dnnl_Ohwi32o = dnnl_Acdb32a,
602  dnnl_Ohwi4o = dnnl_Acdb4a,
603  dnnl_Ohwi8o = dnnl_Acdb8a,
604  dnnl_OIhw16i16o = dnnl_ABcd16b16a,
605  dnnl_OIhw16o16i = dnnl_ABcd16a16b,
606  dnnl_Oihw16o = dnnl_Abcd16a,
607  dnnl_OIhw4i16o4i = dnnl_ABcd4b16a4b,
608  dnnl_OIhw16i16o4i = dnnl_ABcd16b16a4b,
609  dnnl_OIhw16i16o2i = dnnl_ABcd16b16a2b,
610  dnnl_OIhw4i4o = dnnl_ABcd4b4a,
611  dnnl_OIhw4o4i = dnnl_ABcd4a4b,
612  dnnl_Oihw4o = dnnl_Abcd4a,
613  dnnl_OIhw8i16o2i = dnnl_ABcd8b16a2b,
614  dnnl_OIhw8i8o = dnnl_ABcd8b8a,
615  dnnl_OIhw8o16i2o = dnnl_ABcd8a16b2a,
616  dnnl_OIhw2i8o4i = dnnl_ABcd2b8a4b,
617  dnnl_IOhw8o16i2o = dnnl_BAcd8a16b2a,
618  dnnl_OIhw8o8i = dnnl_ABcd8a8b,
619  dnnl_OIhw8o4i = dnnl_ABcd8a4b,
620 
621  // weights, 5D
622  dnnl_Odhwi16o = dnnl_Acdeb16a,
623  dnnl_OdhwI16o2i = dnnl_AcdeB16a2b,
624  dnnl_Odhwi4o = dnnl_Acdeb4a,
625  dnnl_Odhwi8o = dnnl_Acdeb8a,
626  dnnl_OIdhw16i16o = dnnl_ABcde16b16a,
627  dnnl_OIdhw16o16i = dnnl_ABcde16a16b,
628  dnnl_Oidhw16o = dnnl_Abcde16a,
629  dnnl_OIdhw4i4o = dnnl_ABcde4b4a,
630  dnnl_OIdhw4o4i = dnnl_ABcde4a4b,
631  dnnl_Oidhw4o = dnnl_Abcde4a,
632  dnnl_OIdhw8i16o2i = dnnl_ABcde8b16a2b,
633  dnnl_OIdhw8i8o = dnnl_ABcde8b8a,
634  dnnl_OIdhw8o16i2o = dnnl_ABcde8a16b2a,
635  dnnl_IOdhw8o16i2o = dnnl_BAcde8a16b2a,
636  dnnl_OIdhw4i16o4i = dnnl_ABcde4b16a4b,
637  dnnl_OIdhw2i8o4i = dnnl_ABcde2b8a4b,
638  dnnl_OIdhw8o8i = dnnl_ABcde8a8b,
639  dnnl_OIdhw8o4i = dnnl_ABcde8a4b,
640  dnnl_IOdhw16i16o = dnnl_BAcde16b16a,
641  dnnl_OIdhw4o8i8o4i = dnnl_ABcde4a8b8a4b,
642  dnnl_IOdhw16o16i = dnnl_BAcde16a16b,
643 
644  // weights w/ groups, 3D
645  dnnl_Goiw16g = dnnl_Abcd16a,
646  dnnl_Goiw8g = dnnl_Abcd8a,
647  dnnl_gIOw16o16i = dnnl_aCBd16b16c,
648  dnnl_gIOw16i16o = dnnl_aCBd16c16b,
649  dnnl_gOIw16i16o = dnnl_aBCd16c16b,
650  dnnl_gOIw16o16i = dnnl_aBCd16b16c,
651  dnnl_gOiw16o = dnnl_aBcd16b,
652  dnnl_gOIw4i16o4i = dnnl_aBCd4c16b4c,
653  dnnl_gOIw2i8o4i = dnnl_aBCd2c8b4c,
654  dnnl_gOIw16i16o4i = dnnl_aBCd16c16b4c,
655  dnnl_gOIw16i16o2i = dnnl_aBCd16c16b2c,
656  dnnl_gOIw4i4o = dnnl_aBCd4c4b,
657  dnnl_gOIw4o4i = dnnl_aBCd4b4c,
658  dnnl_gOiw4o = dnnl_aBcd4b,
659  dnnl_gOIw8i16o2i = dnnl_aBCd8c16b2c,
660  dnnl_gOIw8i8o = dnnl_aBCd8c8b,
661  dnnl_gOIw8o16i2o = dnnl_aBCd8b16c2b,
662  dnnl_gIOw8o16i2o = dnnl_aCBd8b16c2b,
663  dnnl_gOIw8o8i = dnnl_aBCd8b8c,
664  dnnl_gOIw8o4i = dnnl_aBCd8b4c,
665  dnnl_gOwi16o = dnnl_aBdc16b,
666  dnnl_gOwI16o2i = dnnl_aBdC16b2c,
667  dnnl_gOwI16o4i = dnnl_aBdC16b4c,
668  dnnl_gOwi4o = dnnl_aBdc4b,
669  dnnl_gOwi8o = dnnl_aBdc8b,
670  dnnl_Goiw32g = dnnl_Abcd32a,
671  dnnl_gOIw2i4o2i = dnnl_aBCd2c4b2c,
672  dnnl_gOIw2o4i2o = dnnl_aBCd2b4c2b,
673  dnnl_gOIw4i8o2i = dnnl_aBCd4c8b2c,
674  dnnl_gOIw4o8i2o = dnnl_aBCd4b8c2b,
675 
676  // weights w/ groups, 4D
677  dnnl_gIOhw16i16o = dnnl_aCBde16c16b,
678  dnnl_gIOhw16o16i = dnnl_aCBde16b16c,
679  dnnl_gOhwi16o = dnnl_aBdec16b,
680  dnnl_gOhwI16o2i = dnnl_aBdeC16b2c,
681  dnnl_gOhwI16o4i = dnnl_aBdeC16b4c,
682  dnnl_gOhwi32o = dnnl_aBdec32b,
683  dnnl_gOhwi4o = dnnl_aBdec4b,
684  dnnl_gOhwi8o = dnnl_aBdec8b,
685  dnnl_Goihw16g = dnnl_Abcde16a,
686  dnnl_gOIhw16i16o = dnnl_aBCde16c16b,
687  dnnl_gOIhw16o16i = dnnl_aBCde16b16c,
688  dnnl_gOihw16o = dnnl_aBcde16b,
689  dnnl_gOIhw2i8o4i = dnnl_aBCde2c8b4c,
690  dnnl_gOIhw4i16o4i = dnnl_aBCde4c16b4c,
691  dnnl_gOIhw16i16o4i = dnnl_aBCde16c16b4c,
692  dnnl_gOIhw16i16o2i = dnnl_aBCde16c16b2c,
693  dnnl_gOIhw4i4o = dnnl_aBCde4c4b,
694  dnnl_gOIhw4o4i = dnnl_aBCde4b4c,
695  dnnl_gOihw4o = dnnl_aBcde4b,
696  dnnl_Goihw8g = dnnl_Abcde8a,
697  dnnl_gOIhw8i16o2i = dnnl_aBCde8c16b2c,
698  dnnl_gOIhw8i8o = dnnl_aBCde8c8b,
699  dnnl_gOIhw8o16i2o = dnnl_aBCde8b16c2b,
700  dnnl_gIOhw8o16i2o = dnnl_aCBde8b16c2b,
701  dnnl_gOIhw8o8i = dnnl_aBCde8b8c,
702  dnnl_gOIhw8o4i = dnnl_aBCde8b4c,
703  dnnl_Goihw32g = dnnl_Abcde32a,
704 
705  dnnl_OIw4o8i8o4i = dnnl_ABc4a8b8a4b,
706  dnnl_OIhw4o8i8o4i = dnnl_ABcd4a8b8a4b,
707  dnnl_IOw4i8o8i4o = dnnl_BAc4b8a8b4a,
708  dnnl_IOhw4i8o8i4o = dnnl_BAcd4b8a8b4a,
709  dnnl_IOdhw4i8o8i4o = dnnl_BAcde4b8a8b4a,
710 
711  dnnl_OIhw2o8i8o2i = dnnl_ABcd2a8b8a2b,
712  dnnl_gOIw4o8i8o4i = dnnl_aBCd4b8c8b4c,
713  dnnl_gOIhw4o8i8o4i = dnnl_aBCde4b8c8b4c,
714  dnnl_gOIdhw4o8i8o4i = dnnl_aBCdef4b8c8b4c,
715  dnnl_gIOw4i8o8i4o = dnnl_aCBd4c8b8c4b,
716  dnnl_gIOhw4i8o8i4o = dnnl_aCBde4c8b8c4b,
717  dnnl_gIOdhw4i8o8i4o = dnnl_aCBdef4c8b8c4b,
718  dnnl_gOIhw2o8i8o2i = dnnl_aBCde2b8c8b2c,
719  dnnl_gOIhw2i4o2i = dnnl_aBCde2c4b2c,
720  dnnl_gOIhw2o4i2o = dnnl_aBCde2b4c2b,
721  dnnl_gOIhw4i8o2i = dnnl_aBCde4c8b2c,
722  dnnl_gOIhw4o8i2o = dnnl_aBCde4b8c2b,
723 
724  // weights w/ groups, 6D
725  dnnl_gIOdhw16i16o = dnnl_aCBdef16c16b,
726  dnnl_gIOdhw16o16i = dnnl_aCBdef16b16c,
727  dnnl_gOdhwi16o = dnnl_aBdefc16b,
728  dnnl_gOdhwI16o2i = dnnl_aBdefC16b2c,
729  dnnl_gOdhwi4o = dnnl_aBdefc4b,
730  dnnl_gOdhwi8o = dnnl_aBdefc8b,
731  dnnl_gOIdhw16i16o = dnnl_aBCdef16c16b,
732  dnnl_gOIdhw4i16o4i = dnnl_aBCdef4c16b4c,
733  dnnl_gOIdhw2i8o4i = dnnl_aBCdef2c8b4c,
734  dnnl_gOIdhw16o16i = dnnl_aBCdef16b16c,
735  dnnl_gOidhw16o = dnnl_aBcdef16b,
736  dnnl_gOIdhw4i4o = dnnl_aBCdef4c4b,
737  dnnl_gOIdhw4o4i = dnnl_aBCdef4b4c,
738  dnnl_gOidhw4o = dnnl_aBcdef4b,
739  dnnl_gOIdhw8i16o2i = dnnl_aBCdef8c16b2c,
740  dnnl_gOIdhw8i8o = dnnl_aBCdef8c8b,
741  dnnl_gOIdhw8o16i2o = dnnl_aBCdef8b16c2b,
742  dnnl_gIOdhw8o16i2o = dnnl_aCBdef8b16c2b,
743  dnnl_gOIdhw8o8i = dnnl_aBCdef8b8c,
744  dnnl_gOIdhw8o4i = dnnl_aBCdef8b4c,
745  dnnl_Goidhw16g = dnnl_Abcdef16a,
746  dnnl_Goidhw32g = dnnl_Abcdef32a,
747  dnnl_gOIdhw2i4o2i = dnnl_aBCdef2c4b2c,
748  dnnl_gOIdhw4i8o2i = dnnl_aBCdef4c8b2c,
749  dnnl_gOIdhw2o4i2o = dnnl_aBCdef2b4c2b,
750  dnnl_gOIdhw4o8i2o = dnnl_aBCdef4b8c2b,
752 
754 
759 
761 typedef enum {
762  // TODO: suggest renames
785 
788 typedef enum {
829 
834 
836 typedef enum {
837  dnnl_alg_kind_undef,
926  dnnl_lbr_gru = 0x4fff,
928  dnnl_binary_add = 0x1fff0,
930  dnnl_binary_mul = 0x1fff1,
932  dnnl_binary_max = 0x1fff2,
934  dnnl_binary_min = 0x1fff3,
940 
942 typedef enum {
952 
965 
978 
992 
995 
998 
1002 #define DNNL_MAX_NDIMS 12
1003 
1006 #define DNNL_RUNTIME_DIM_VAL INT64_MIN
1007 
1011 #define DNNL_RUNTIME_SIZE_VAL ((size_t)DNNL_RUNTIME_DIM_VAL)
1012 
1015 static const union {
1016  unsigned u;
1017  float f;
1018 } DNNL_RUNTIME_F32_VAL_REP = {0x7fc000d0};
1020 
1023 #define DNNL_RUNTIME_F32_VAL (DNNL_RUNTIME_F32_VAL_REP.f)
1024 
1026 static const int DNNL_RUNTIME_S32_VAL_REP = INT32_MIN;
1028 
1031 #define DNNL_RUNTIME_S32_VAL DNNL_RUNTIME_S32_VAL_REP
1032 
1034 typedef int64_t dnnl_dim_t;
1035 
1038 
1042 typedef struct {
1046  // Innermost section
1047  // ASSUMPTION: the innermost blocks are always dense
1056 
1058 typedef enum {
1061  // Tensors of weights for 2x3 winograd convolutions.
1065  // Tensor of weights for 4x3 convolution.
1068 
1070 typedef struct {
1071  dnnl_wino_memory_format_t wino_format;
1072  int r;
1073  int alpha;
1074  int ic;
1075  int oc;
1076  int ic_block;
1077  int oc_block;
1078  int ic2_block;
1079  int oc2_block;
1080  float adj_scale;
1081  size_t size;
1083 
1084 typedef enum {
1085  dnnl_packed_format_undef = 0,
1086  dnnl_ldigo_p,
1087  dnnl_ldgoi_p
1088 } dnnl_rnn_packed_memory_format_t;
1089 
1092 #define DNNL_RNN_MAX_N_PARTS 4
1093 
1095 typedef struct {
1096  dnnl_rnn_packed_memory_format_t format;
1097  int n_parts;
1098  int n;
1099  int ldb;
1100  int parts[DNNL_RNN_MAX_N_PARTS];
1101  size_t part_pack_size[DNNL_RNN_MAX_N_PARTS];
1102  unsigned pack_part[DNNL_RNN_MAX_N_PARTS];
1103  size_t offset_compensation;
1104  size_t size;
1105  char reserved[200];
1107 
1109 typedef enum {
1110  dnnl_memory_extra_flag_none = 0x0U,
1119  dnnl_memory_extra_flag_scale_adjust = 0x2U,
1120  dnnl_memory_extra_flag_gpu_rnn_u8s8_compensation = 0x4U,
1122 
1124 typedef struct {
1127  uint64_t flags;
1133  char reserved[64];
1135 
1140 typedef struct {
1142  int ndims;
1158 
1161 
1164 
1168 
1172 
1175  union {
1183  // ... other descriptions possible
1184  } format_desc;
1185 
1188 
1191 struct dnnl_memory;
1192 
1194 typedef struct dnnl_memory *dnnl_memory_t;
1195 
1197 typedef const struct dnnl_memory *const_dnnl_memory_t;
1198 
1199 #define DNNL_MEMORY_NONE (NULL)
1200 #define DNNL_MEMORY_ALLOCATE ((void *)(size_t)-1)
1201 
1203 
1208 
1210 typedef void *dnnl_op_desc_t;
1212 typedef const void *const_dnnl_op_desc_t;
1213 
1216 
1219 
1222 
1224 typedef struct {
1258  dnnl_dims_t padding[2];
1262 
1264 
1267 
1270 
1272 
1275 
1277 typedef struct {
1288  int axis;
1292 
1294 
1297 
1299 typedef struct {
1343  float alpha, beta;
1345 
1347 
1350 
1352 typedef struct {
1366 
1368 
1371 
1375 
1377 
1380 
1382 typedef struct {
1409  dnnl_dims_t padding[2];
1413 
1415 
1418 
1420 typedef struct {
1438  float lrn_alpha;
1440  float lrn_beta;
1442  float lrn_k;
1443 } dnnl_lrn_desc_t;
1444 
1446 
1449 
1451 typedef struct {
1468  dnnl_memory_desc_t diff_data_scaleshift_desc;
1475  unsigned flags;
1477 
1479 
1482 
1484 typedef struct {
1503  dnnl_memory_desc_t diff_data_scaleshift_desc;
1512  unsigned flags;
1514 
1516 
1519 
1521 typedef struct {
1548 
1550 
1553 
1555 typedef enum {
1559 
1561 typedef enum {
1575 
1577 typedef struct {
1615 
1642 
1644  unsigned int flags;
1648  float alpha;
1649  float beta;
1650 
1651 } dnnl_rnn_desc_t;
1652 
1654 
1657 
1659 typedef struct {
1668  dnnl_memory_desc_t src_desc[2];
1672 
1674 
1677 
1685 typedef struct {
1700 
1702 
1705 
1707 typedef struct {
1726  float factors[DNNL_MAX_NDIMS];
1728 
1730 
1732 
1735 
1737 typedef enum {
1745 
1748 struct dnnl_engine;
1750 typedef struct dnnl_engine *dnnl_engine_t;
1751 #if 0
1752 // FIXME: looks like this never happens
1754 typedef const struct dnnl_engine *const_dnnl_engine_t;
1755 #endif
1756 
1758 
1763 
1767 
1770 
1772 typedef const struct dnnl_primitive_desc_iterator
1774 
1777 struct dnnl_primitive_desc;
1778 
1781 
1784 
1786 
1789 
1791 typedef enum {
1815 
1821 struct dnnl_primitive_attr;
1822 
1826 
1829 
1848 struct dnnl_post_ops;
1849 
1852 
1854 typedef const struct dnnl_post_ops *const_dnnl_post_ops_t;
1855 
1857 
1860 
1863 struct dnnl_primitive;
1868 
1870 #define DNNL_ARG_SRC_0 1
1871 #define DNNL_ARG_SRC DNNL_ARG_SRC_0
1874 #define DNNL_ARG_SRC_LAYER DNNL_ARG_SRC_0
1877 #define DNNL_ARG_FROM DNNL_ARG_SRC_0
1880 
1882 #define DNNL_ARG_SRC_1 2
1883 #define DNNL_ARG_SRC_ITER DNNL_ARG_SRC_1
1886 
1888 #define DNNL_ARG_SRC_2 3
1889 #define DNNL_ARG_SRC_ITER_C DNNL_ARG_SRC_2
1892 
1894 #define DNNL_ARG_DST_0 17
1895 #define DNNL_ARG_DST DNNL_ARG_DST_0
1898 #define DNNL_ARG_TO DNNL_ARG_DST_0
1901 #define DNNL_ARG_DST_LAYER DNNL_ARG_DST_0
1903 
1905 #define DNNL_ARG_DST_1 18
1906 #define DNNL_ARG_DST_ITER DNNL_ARG_DST_1
1909 
1911 #define DNNL_ARG_DST_2 19
1912 #define DNNL_ARG_DST_ITER_C DNNL_ARG_DST_2
1915 
1917 #define DNNL_ARG_WEIGHTS_0 33
1918 #define DNNL_ARG_WEIGHTS DNNL_ARG_WEIGHTS_0
1921 #define DNNL_ARG_SCALE_SHIFT DNNL_ARG_WEIGHTS_0
1924 #define DNNL_ARG_WEIGHTS_LAYER DNNL_ARG_WEIGHTS_0
1927 
1929 #define DNNL_ARG_WEIGHTS_1 34
1930 #define DNNL_ARG_WEIGHTS_ITER DNNL_ARG_WEIGHTS_1
1933 
1935 #define DNNL_ARG_WEIGHTS_2 35
1936 #define DNNL_ARG_WEIGHTS_PEEPHOLE DNNL_ARG_WEIGHTS_2
1939 
1941 #define DNNL_ARG_WEIGHTS_3 36
1942 #define DNNL_ARG_WEIGHTS_PROJECTION DNNL_ARG_WEIGHTS_3
1945 
1947 #define DNNL_ARG_BIAS 41
1948 
1950 #define DNNL_ARG_MEAN 49
1951 #define DNNL_ARG_VARIANCE 50
1953 
1956 #define DNNL_ARG_WORKSPACE 64
1957 #define DNNL_ARG_SCRATCHPAD 80
1959 
1961 #define DNNL_ARG_DIFF_SRC_0 129
1962 #define DNNL_ARG_DIFF_SRC DNNL_ARG_DIFF_SRC_0
1965 #define DNNL_ARG_DIFF_SRC_LAYER DNNL_ARG_DIFF_SRC_0
1968 
1970 #define DNNL_ARG_DIFF_SRC_1 130
1971 #define DNNL_ARG_DIFF_SRC_ITER DNNL_ARG_DIFF_SRC_1
1974 
1976 #define DNNL_ARG_DIFF_SRC_2 131
1977 #define DNNL_ARG_DIFF_SRC_ITER_C DNNL_ARG_DIFF_SRC_2
1980 
1982 #define DNNL_ARG_DIFF_DST_0 145
1983 #define DNNL_ARG_DIFF_DST DNNL_ARG_DIFF_DST_0
1986 #define DNNL_ARG_DIFF_DST_LAYER DNNL_ARG_DIFF_DST_0
1989 
1991 #define DNNL_ARG_DIFF_DST_1 146
1992 #define DNNL_ARG_DIFF_DST_ITER DNNL_ARG_DIFF_DST_1
1995 
1997 #define DNNL_ARG_DIFF_DST_2 147
1998 #define DNNL_ARG_DIFF_DST_ITER_C DNNL_ARG_DIFF_DST_2
2001 
2003 #define DNNL_ARG_DIFF_WEIGHTS_0 161
2004 #define DNNL_ARG_DIFF_WEIGHTS DNNL_ARG_DIFF_WEIGHTS_0
2007 #define DNNL_ARG_DIFF_SCALE_SHIFT DNNL_ARG_DIFF_WEIGHTS_0
2010 #define DNNL_ARG_DIFF_WEIGHTS_LAYER DNNL_ARG_DIFF_WEIGHTS_0
2013 
2015 #define DNNL_ARG_DIFF_WEIGHTS_1 162
2016 #define DNNL_ARG_DIFF_WEIGHTS_ITER DNNL_ARG_DIFF_WEIGHTS_1
2019 
2021 #define DNNL_ARG_DIFF_WEIGHTS_2 163
2022 #define DNNL_ARG_DIFF_WEIGHTS_PEEPHOLE DNNL_ARG_DIFF_WEIGHTS_2
2025 
2027 #define DNNL_ARG_DIFF_WEIGHTS_3 164
2028 #define DNNL_ARG_DIFF_WEIGHTS_PROJECTION DNNL_ARG_DIFF_WEIGHTS_3
2031 
2033 #define DNNL_ARG_DIFF_BIAS 169
2034 
2036 #define DNNL_ARG_ATTR_OUTPUT_SCALES 513
2037 
2040 #define DNNL_ARG_MULTIPLE_SRC 1024
2041 #define DNNL_ARG_MULTIPLE_DST 2048
2044 
2046 #define DNNL_ARG_ATTR_ZERO_POINTS 4096
2047 
2050 #define DNNL_ARG_ATTR_POST_OP_DW 8192
2051 
2054 typedef struct {
2055  int arg;
2057 } dnnl_exec_arg_t;
2058 
2060 
2063 
2093 typedef enum {
2095 
2098 
2101 
2104 
2109 
2112 
2115 
2117 
2118  // memory and op descriptor section
2137 
2138  // memory descriptor section
2149 
2150  // Max value to prevent UB for internal use only dnnl_query_t
2151  dnnl_query_max = 0x7fff,
2152 } dnnl_query_t;
2153 
2155 
2157 
2160 
2162 typedef enum {
2173 
2176 struct dnnl_stream;
2178 typedef struct dnnl_stream *dnnl_stream_t;
2180 typedef const struct dnnl_stream *const_dnnl_stream_t;
2181 
2183 struct dnnl_stream_attr;
2185 typedef struct dnnl_stream_attr *dnnl_stream_attr_t;
2187 typedef const struct dnnl_stream_attr *const_dnnl_stream_attr_t;
2188 
2190 
2193 
2195 #define DNNL_RUNTIME_NONE 0u
2196 
2198 #define DNNL_RUNTIME_SEQ 1u
2199 
2201 #define DNNL_RUNTIME_OMP 2u
2202 
2204 #define DNNL_RUNTIME_TBB 4u
2205 
2207 #define DNNL_RUNTIME_THREADPOOL 8u
2208 
2210 #define DNNL_RUNTIME_OCL 256u
2211 
2214 typedef struct {
2215  int major;
2216  int minor;
2217  int patch;
2218  const char *hash;
2219  unsigned cpu_runtime;
2220  unsigned gpu_runtime;
2221 } dnnl_version_t;
2222 
2224 #define DNNL_JIT_PROFILE_NONE 0u
2225 
2227 #define DNNL_JIT_PROFILE_VTUNE 1u
2228 
2230 #define DNNL_JIT_PROFILE_LINUX_PERFMAP 2u
2231 
2233 #define DNNL_JIT_PROFILE_LINUX_JITDUMP 4u
2234 
2237 #define DNNL_JIT_PROFILE_LINUX_JITDUMP_USE_TSC 8u
2238 
2240 #define DNNL_JIT_PROFILE_LINUX_PERF \
2241  (DNNL_JIT_PROFILE_LINUX_JITDUMP | DNNL_JIT_PROFILE_LINUX_PERFMAP)
2242 
2244 typedef enum {
2247 
2250 
2253 
2256 
2260 
2264 
2268 
2273 
2278 
2283 } dnnl_cpu_isa_t;
2284 
2286 
2288 
2289 #ifdef __cplusplus
2290 }
2291 #endif
2292 
2293 #endif
dnnl_memory_desc_t diff_dst_layer_desc
Destination gradient layer memory descriptor.
Definition: dnnl_types.h:1629
A layer normalization primitive.
Definition: dnnl_types.h:814
destination grad. memory desc
Definition: dnnl_types.h:2145
An element-wise primitive.
Definition: dnnl_types.h:804
dnnl_alg_kind_t activation_kind
Activation function used for vanilla_rnn cell kind.
Definition: dnnl_types.h:1647
dnnl_memory_desc_t weights_layer_desc
Weights layer memory descriptor.
Definition: dnnl_types.h:1596
dnnl_alg_kind_t alg_kind
The kind of pooling algorithm.
Definition: dnnl_types.h:1393
6D tensor blocked by 3rd dimension with block size 4
Definition: dnnl_types.h:347
execution engine
Definition: dnnl_types.h:2096
dnnl_memory_desc_t weights_desc
Weights memory descriptor.
Definition: dnnl_types.h:1534
dnnl_memory_desc_t diff_dst_iter_c_desc
Destination gradient iteration memory descriptor for cell state.
Definition: dnnl_types.h:1633
A batch normalization primitive.
Definition: dnnl_types.h:812
Eltwise: bounded_relu.
Definition: dnnl_types.h:863
Undefined memory format tag.
Definition: dnnl_types.h:169
3D tensor blocked by 2nd dimension with block size 16
Definition: dnnl_types.h:214
A tensor in a generic format described by the stride and blocking values in each dimension.
Definition: dnnl_types.h:89
4D LSTM projection tensor in the format (num_layers, num_directions, num_channels_in_hidden_state, num_channels_in_recurrent_projection).
Definition: dnnl_types.h:512
CPU engine.
Definition: dnnl_types.h:1741
dnnl_primitive_kind_t primitive_kind
The kind of primitive.
Definition: dnnl_types.h:1454
Eltwise: ReLU (dst for backward)
Definition: dnnl_types.h:890
destination memory desc
Definition: dnnl_types.h:2144
Direct deconvolution.
Definition: dnnl_types.h:845
struct dnnl_memory * dnnl_memory_t
A memory handle.
Definition: dnnl_types.h:1194
A descriptor for an RNN operation.
Definition: dnnl_types.h:1577
The user manages the scratchpad allocation by querying and providing the scratchpad memory to primiti...
Definition: dnnl_types.h:1813
layer normalization descriptor
Definition: dnnl_types.h:2129
memory consumption – extra
Definition: dnnl_types.h:2103
dnnl_primitive_kind_t
Kinds of primitives.
Definition: dnnl_types.h:788
permuted 3D tensor
Definition: dnnl_types.h:194
Eltwise: linear.
Definition: dnnl_types.h:861
dnnl_memory_desc_t bias_desc
Bias memory descriptor.
Definition: dnnl_types.h:1538
dnnl_dim_t local_size
The number of channels to sum over (for cross-channel LRN) or the side length of the square region to...
Definition: dnnl_types.h:1436
Internal weights format for 2x3 Winograd.
Definition: dnnl_types.h:1063
dnnl_memory_desc_t data_scaleshift_desc
Scale and shift data and gradient memory descriptors.
Definition: dnnl_types.h:1502
void * dnnl_op_desc_t
A pointer to any of the operation descriptors.
Definition: dnnl_types.h:1210
dnnl_alg_kind_t alg_kind
LRN algorithm.
Definition: dnnl_types.h:1429
A resampling primitive.
Definition: dnnl_types.h:828
dnnl_format_kind_t format_kind
Memory format kind.
Definition: dnnl_types.h:1174
An opaque structure to describe a primitive.
dnnl_prop_kind_t prop_kind
The kind of propagation.
Definition: dnnl_types.h:1457
GRU cell with linear before reset.
Definition: dnnl_types.h:926
Any ISA (excepting those listed as initial support)
Definition: dnnl_types.h:2246
dnnl_dims_t strides
Pooling kernel strides for spatial dimensions.
Definition: dnnl_types.h:1403
4D CNN weights tensor, an alias to dnnl_acdb
Definition: dnnl_types.h:456
Undefined data type, used for empty memory descriptors.
Definition: dnnl_types.h:64
dnnl_memory_desc_t dst_desc
Destination memory descriptor.
Definition: dnnl_types.h:1696
dnnl_dim_t group_size
Number of groups.
Definition: dnnl_types.h:1290
dnnl_memory_desc_t bias_desc
Bias memory descriptor.
Definition: dnnl_types.h:1244
5D CNN weights tensor, an alias to dnnl_acdeb
Definition: dnnl_types.h:468
float scale_adjust
Scale applied to the data.
Definition: dnnl_types.h:1131
dnnl_convolution_desc_t dnnl_deconvolution_desc_t
A descriptor of a deconvolution operation.
Definition: dnnl_types.h:1269
4D CNN weights tensor (incl. groups), an alias to dnnl_abcd
Definition: dnnl_types.h:473
dnnl_primitive_kind_t primitive_kind
The kind of primitive.
Definition: dnnl_types.h:1355
dnnl_alg_kind_t alg_kind
The kind of the convolution algorithm.
Definition: dnnl_types.h:1234
In-order execution.
Definition: dnnl_types.h:2167
dnnl_data_type_t accum_data_type
The accumulator data type. Initialized automatically.
Definition: dnnl_types.h:1260
Use no normalization flags.
Definition: dnnl_types.h:951
scratchpad memory desc
Definition: dnnl_types.h:2147
dnnl_memory_desc_t src_desc
Source memory descriptor.
Definition: dnnl_types.h:1718
dnnl_prop_kind_t prop_kind
The kind of propagation.
Definition: dnnl_types.h:1388
dnnl_memory_t memory
Input/output memory.
Definition: dnnl_types.h:2056
const void * const_dnnl_op_desc_t
A pointer to any of the operation descriptors (constant variant).
Definition: dnnl_types.h:1212
Intel Streaming SIMD Extensions 4.1 (Intel SSE4.1)
Definition: dnnl_types.h:2249
Eltwise: clip.
Definition: dnnl_types.h:882
Intel AVX-512, Intel DL Boost and bfloat16 support and Intel AMX with 8-bit integer and bfloat16 supp...
Definition: dnnl_types.h:2282
An opaque structure for primitive descriptor attributes.
dnnl_memory_desc_t dst_desc
Destination memory descriptor.
Definition: dnnl_types.h:1248
logsoftmax descriptor
Definition: dnnl_types.h:2134
permuted 4D tensor
Definition: dnnl_types.h:191
dnnl_memory_desc_t src_desc
Source memory descriptor.
Definition: dnnl_types.h:1395
float lrn_alpha
LRN alpha parameter.
Definition: dnnl_types.h:1438
dnnl_memory_desc_t diff_weights_desc
Weights gradient memory descriptor.
Definition: dnnl_types.h:1536
dnnl_memory_desc_t dst_iter_c_desc
Destination iter memory descriptor for cell state.
Definition: dnnl_types.h:1606
Primitive iterator passed over last primitive descriptor.
Definition: dnnl_types.h:49
int minor
Minor version.
Definition: dnnl_types.h:2216
2D CNN weights tensor, an alias to dnnl_ab
Definition: dnnl_types.h:440
An opaque structure to describe a memory.
dnnl_memory_desc_t dst_layer_desc
Destination layer memory descriptor.
Definition: dnnl_types.h:1602
permuted 5D tensor
Definition: dnnl_types.h:192
dnnl_engine_kind_t
Kinds of engines.
Definition: dnnl_types.h:1737
struct dnnl_primitive_desc * dnnl_primitive_desc_t
A primitive descriptor handle.
Definition: dnnl_types.h:1780
Eltwise: square root (dst for backward)
Definition: dnnl_types.h:896
non-standard 16-bit (bfloat16 w/ 7 bit mantissa) floating point.
Definition: dnnl_types.h:68
Parameter to allow internal only primitives without undefined behavior.
Definition: dnnl_types.h:832
Undefined primitive.
Definition: dnnl_types.h:790
dnnl_dims_t strides
Convolution strides in each spatial dimension.
Definition: dnnl_types.h:1252
int softmax_axis
The axis along which to perform the softmax.
Definition: dnnl_types.h:1364
Unidirectional execution of RNN primitive from left to right.
Definition: dnnl_types.h:1563
The library manages the scratchpad allocation according to the policy specified by the DNNL_ENABLE_CO...
Definition: dnnl_types.h:1808
5D CNN weights tensor (incl. groups), an alias to dnnl_acbde
Definition: dnnl_types.h:481
3D tensor blocked by 2nd dimension with block size 4
Definition: dnnl_types.h:220
3D CNN activations tensor blocked by channels with block size 8, an alias to dnnl_aBc8b ...
Definition: dnnl_types.h:562
dnnl_dims_t dims
Dimensions in the following order:
Definition: dnnl_types.h:1157
const struct dnnl_memory * const_dnnl_memory_t
A constant memory handle.
Definition: dnnl_types.h:1197
4D CNN activations tensor blocked by channels with block size 16, an alias to dnnl_aBcd16b ...
Definition: dnnl_types.h:544
binary descriptor
Definition: dnnl_types.h:2133
dnnl_memory_desc_t diff_weights_iter_desc
Weights gradient iter memory descriptor.
Definition: dnnl_types.h:1625
permuted 4D tensor
Definition: dnnl_types.h:186
Description of tensor of packed weights for rnn.
Definition: dnnl_types.h:1095
A descriptor of a pooling operation.
Definition: dnnl_types.h:1382
2D CNN activations tensor, an alias to dnnl_ab
Definition: dnnl_types.h:417
plain 2D tensor
Definition: dnnl_types.h:178
4D CNN activations tensor, an alias to dnnl_bcda
Definition: dnnl_types.h:433
4D RNN bias tensor in the format (num_layers, num_directions, num_gates, output_channels).
Definition: dnnl_types.h:522
permuted 5D tensor
Definition: dnnl_types.h:199
Undefined memory format kind, used for empty memory descriptors.
Definition: dnnl_types.h:82
dnnl_memory_desc_t bias_desc
Bias memory descriptor.
Definition: dnnl_types.h:1694
Bidirectional execution of RNN primitive with summation of the results.
Definition: dnnl_types.h:1571
32-bit signed integer.
Definition: dnnl_types.h:72
dnnl_memory_desc_t diff_src_iter_c_desc
Source gradient iter memory descriptor for cell state.
Definition: dnnl_types.h:1621
int inner_nblks
The number of innermost blocks, e.g. 3 in case of OIhw_4i16o4i_
Definition: dnnl_types.h:1049
Direct convolution.
Definition: dnnl_types.h:839
int major
Major version.
Definition: dnnl_types.h:2215
Just a sentinel, not real memory format tag.
Definition: dnnl_types.h:410
5D tensor blocked by 1st dimension with block size 8
Definition: dnnl_types.h:289
An opaque structure to describe a primitive descriptor iterator.
pooling descriptor
Definition: dnnl_types.h:2126
Convolution algorithm(either direct or Winograd) is chosen just in time.
Definition: dnnl_types.h:843
float lrn_beta
LRN beta parameter.
Definition: dnnl_types.h:1440
dnnl_memory_desc_t src_desc
Source memory descriptor.
Definition: dnnl_types.h:1530
Unidirectional execution of RNN primitive from right to left.
Definition: dnnl_types.h:1565
A deconvolution primitive.
Definition: dnnl_types.h:802
dnnl_memory_desc_t src_iter_desc
Source iteration memory descriptor for hidden state.
Definition: dnnl_types.h:1592
5D tensor blocked by 2nd dimension with block size 8
Definition: dnnl_types.h:315
8-bit unsigned integer.
Definition: dnnl_types.h:76
dnnl_alg_kind_t alg_kind
The kind of the binary algorithm.
Definition: dnnl_types.h:1666
A descriptor of a matrix multiplication operation.
Definition: dnnl_types.h:1685
permuted 6D tensor
Definition: dnnl_types.h:205
3D RNN data tensor in the format (batch, seq_length, input channels).
Definition: dnnl_types.h:492
dnnl_alg_kind_t cell_kind
RNN cell kind.
Definition: dnnl_types.h:1586
const char * hash
Git hash of the sources (may be absent)
Definition: dnnl_types.h:2218
5D CNN weights tensor, an alias to dnnl_cdeba
Definition: dnnl_types.h:466
4D CNN activations tensor, an alias to dnnl_acdb
Definition: dnnl_types.h:431
dnnl_query_t
Primitive descriptor query specification.
Definition: dnnl_types.h:2093
Intel Advanced Vector Extensions (Intel AVX)
Definition: dnnl_types.h:2252
Intel AVX-512 subset for Intel Xeon Phi processors 7235, 7285, 7295 Series.
Definition: dnnl_types.h:2263
dnnl_memory_desc_t weights_desc
Weights memory descriptor.
Definition: dnnl_types.h:1692
Backward data propagation.
Definition: dnnl_types.h:779
dnnl_memory_desc_t stat_desc
Statistics memory descriptor.
Definition: dnnl_types.h:1472
A descriptor of a binary operation.
Definition: dnnl_types.h:1659
source gradient memory desc
Definition: dnnl_types.h:2141
A binary primitive.
Definition: dnnl_types.h:822
Structure containing version information as per Semantic Versioning
Definition: dnnl_types.h:2214
4D tensor blocked by 2nd dimension with block size 32
Definition: dnnl_types.h:246
int arg
An argument index, e.g. DNNL_ARG_SRC.
Definition: dnnl_types.h:2055
dnnl_memory_desc_t diff_bias_desc
Bias gradient memory descriptor.
Definition: dnnl_types.h:1540
LSTM cell.
Definition: dnnl_types.h:916
Packed weights format used in RNN.
Definition: dnnl_types.h:93
4D CNN weights tensor, an alias to dnnl_bcda
Definition: dnnl_types.h:458
4D tensor blocked by 2nd dimension with block size 16
Definition: dnnl_types.h:240
A reorder primitive.
Definition: dnnl_types.h:792
dnnl_memory_desc_t weights_desc
Weights memory descriptor.
Definition: dnnl_types.h:1240
A descriptor of a convolution operation.
Definition: dnnl_types.h:1224
2D CNN activations tensor, an alias to dnnl_ba
Definition: dnnl_types.h:419
dnnl_data_type_t accum_data_type
The accumulator data type. Initialized automatically.
Definition: dnnl_types.h:1546
softmax descriptor
Definition: dnnl_types.h:2125
Intel AVX-512 subset for Intel Xeon Scalable processor family and Intel Core processor family...
Definition: dnnl_types.h:2267
no query
Definition: dnnl_types.h:2094
dnnl_scratchpad_mode_t
Scratchpad mode.
Definition: dnnl_types.h:1791
Fuse with ReLU.
Definition: dnnl_types.h:990
batch normalization descriptor
Definition: dnnl_types.h:2128
dnnl_dims_t padded_dims
Size of the data including padding in each dimension.
Definition: dnnl_types.h:1163
dnnl_dim_t offset0
Offset from memory origin to the current block, non-zero only in a description of a memory sub-block...
Definition: dnnl_types.h:1171
dnnl_primitive_kind_t primitive_kind
The kind of primitive.
Definition: dnnl_types.h:1710
5D CNN weights tensor, an alias to dnnl_abcde
Definition: dnnl_types.h:462
dnnl_dims_t padded_offsets
Per-dimension offset from the padding to actual data, the top-level tensor with offsets applied must ...
Definition: dnnl_types.h:1167
Undefined memory format, used for empty memory descriptors.
Definition: dnnl_types.h:1060
runtime estimation (seconds)
Definition: dnnl_types.h:2102
Backward propagation (with respect to all parameters).
Definition: dnnl_types.h:777
An unspecified engine.
Definition: dnnl_types.h:1739
5D tensor blocked by 1st dimension with block size 16
Definition: dnnl_types.h:287
Eltwise: ReLU.
Definition: dnnl_types.h:849
GPU engine.
Definition: dnnl_types.h:1743
dnnl_memory_desc_t dst_desc
Destination memory descriptor.
Definition: dnnl_types.h:1670
dnnl_primitive_kind_t primitive_kind
The kind of primitive.
Definition: dnnl_types.h:1662
4D RNN states tensor in the format (num_layers, num_directions, batch, state channels).
Definition: dnnl_types.h:495
5D RNN weights tensor in the format (num_layers, num_directions, input_channels, num_gates, output_channels).
Definition: dnnl_types.h:502
6D CNN weights tensor (incl. groups), an alias to dnnl_acbdef
Definition: dnnl_types.h:485
3D tensor blocked by 2nd dimension with block size 32
Definition: dnnl_types.h:218
(scratch) memory, additional to all inputs and outputs memory (bytes)
Definition: dnnl_types.h:2108
dnnl_prop_kind_t prop_kind
The kind of propagation.
Definition: dnnl_types.h:1583
Eltwise: pow.
Definition: dnnl_types.h:884
5D tensor blocked by 2nd dimension with block size 4
Definition: dnnl_types.h:300
dnnl_memory_extra_flags_t
Flags for memory special features.
Definition: dnnl_types.h:1109
5D RNN weights tensor in the format (num_layers, num_directions, num_gates, output_channels, input_channels).
Definition: dnnl_types.h:509
permuted 4D tensor
Definition: dnnl_types.h:201
An opaque structure to describe an engine.
dnnl_memory_desc_t data_desc
Source and destination memory descriptor.
Definition: dnnl_types.h:1431
dnnl_memory_desc_t src_desc
Source memory descriptor.
Definition: dnnl_types.h:1690
dnnl_memory_desc_t src_desc
Source memory descriptor.
Definition: dnnl_types.h:1236
6D CNN weights tensor (incl. groups), an alias to dnnl_abcdef
Definition: dnnl_types.h:483
Forward data propagation (inference mode).
Definition: dnnl_types.h:771
dnnl_memory_desc_t data_desc
Source and destination memory descriptor.
Definition: dnnl_types.h:1492
Undefined RNN flags.
Definition: dnnl_types.h:1557
A sum primitive.
Definition: dnnl_types.h:798
unsigned cpu_runtime
CPU runtime.
Definition: dnnl_types.h:2219
5D CNN weights tensor (incl. groups), an alias to dnnl_decab
Definition: dnnl_types.h:479
dnnl_prop_kind_t prop_kind
The kind of propagation.
Definition: dnnl_types.h:1305
Intel Advanced Vector Extensions 512 (Intel AVX-512) subset for Intel Xeon Phi processors x200 Series...
Definition: dnnl_types.h:2259
dnnl_memory_desc_t data_desc
Source and destination memory descriptor.
Definition: dnnl_types.h:1360
dnnl_memory_desc_t diff_data_desc
Source and destination gradient memory descriptor.
Definition: dnnl_types.h:1494
A descriptor of a element-wise operation.
Definition: dnnl_types.h:1299
dnnl_dims_t inner_blks
The size of the blocks, e.g. {4, 16, 4} in case of OIhw_4i16o4i
Definition: dnnl_types.h:1051
A descriptor of a Softmax operation.
Definition: dnnl_types.h:1352
An opaque structure for a chain of post operations.
A descriptor of an inner product operation.
Definition: dnnl_types.h:1521
Eltwise: round.
Definition: dnnl_types.h:888
The operation failed because of incorrect function arguments.
Definition: dnnl_types.h:45
3D RNN data tensor in the format (seq_length, batch, input channels).
Definition: dnnl_types.h:490
dnnl_memory_desc_t diff_weights_projection_desc
Weights gradient projection memory descriptor.
Definition: dnnl_types.h:1641
dnnl_rnn_direction_t
A direction of RNN primitive execution.
Definition: dnnl_types.h:1561
Internal weights format for 2x3 Winograd.
Definition: dnnl_types.h:1062
inner product descriptor
Definition: dnnl_types.h:2130
5D tensor blocked by 2nd dimension with block size 16
Definition: dnnl_types.h:291
6D tensor blocked by 2nd dimension with block size 4
Definition: dnnl_types.h:349
dnnl_status_t
Status values returned by the library functions.
Definition: dnnl_types.h:39
dnnl_memory_desc_t dst_desc
Destination memory descriptor.
Definition: dnnl_types.h:1399
Binary mul.
Definition: dnnl_types.h:930
dnnl_memory_desc_t dst_iter_desc
Destination iter memory descriptor for hidden state.
Definition: dnnl_types.h:1604
A softmax primitive.
Definition: dnnl_types.h:806
const struct dnnl_primitive_desc * const_dnnl_primitive_desc_t
A constant primitive descriptor handle.
Definition: dnnl_types.h:1783
int64_t dnnl_dim_t
A type to describe tensor dimension.
Definition: dnnl_types.h:1034
Forward data propagation (alias for dnnl_forward_training).
Definition: dnnl_types.h:775
primitive kind
Definition: dnnl_types.h:2097
matrix multiplication (matmul) descriptor
Definition: dnnl_types.h:2135
Default stream configuration.
Definition: dnnl_types.h:2171
dnnl_memory_desc_t dst_desc
Destination memory descriptor.
Definition: dnnl_types.h:1542
dnnl_memory_desc_t stat_desc
Mean and variance data memory descriptors.
Definition: dnnl_types.h:1509
const struct dnnl_primitive * const_dnnl_primitive_t
A constant primitive handle.
Definition: dnnl_types.h:1867
LRN within a single channel.
Definition: dnnl_types.h:912
dnnl_data_type_t
Data type specification.
Definition: dnnl_types.h:62
plain 4D tensor
Definition: dnnl_types.h:180
Generic description of blocked data layout for most memory formats.
Definition: dnnl_types.h:1042
plain 6D tensor
Definition: dnnl_types.h:182
Eltwise: exp (dst for backward)
Definition: dnnl_types.h:900
Winograd convolution.
Definition: dnnl_types.h:841
5D tensor blocked by 3rd dimension with block size 4
Definition: dnnl_types.h:337
dnnl_wino_desc_t wino_desc
Tensor of weights for integer 8bit winograd convolution.
Definition: dnnl_types.h:1180
dnnl_alg_kind_t alg_kind
The kind of eltwise algorithm.
Definition: dnnl_types.h:1318
dnnl_format_tag_t
Memory format tag specification.
Definition: dnnl_types.h:164
4D LSTM projection tensor in the format (num_layers, num_directions, num_channels_in_recurrent_projec...
Definition: dnnl_types.h:515
Max pooling.
Definition: dnnl_types.h:902
Eltwise: natural logarithm.
Definition: dnnl_types.h:880
dnnl_memory_desc_t dst_desc
Destination memory descriptor.
Definition: dnnl_types.h:1722
float batch_norm_epsilon
Batch normalization epsilon parameter.
Definition: dnnl_types.h:1474
dnnl_memory_desc_t weights_projection_desc
Weights projection memory descriptor.
Definition: dnnl_types.h:1614
dnnl_prop_kind_t prop_kind
The kind of propagation.
Definition: dnnl_types.h:1490
Description of tensor of weights for winograd 2x3 convolution.
Definition: dnnl_types.h:1070
dnnl_prop_kind_t prop_kind
The kind of propagation.
Definition: dnnl_types.h:1358
4D CNN activations tensor blocked by channels with block size 4, an alias to dnnl_aBcd4b ...
Definition: dnnl_types.h:547
dnnl_memory_desc_t bias_desc
Bias memory descriptor.
Definition: dnnl_types.h:1600
3D CNN weights tensor, an alias to dnnl_cba
Definition: dnnl_types.h:448
4D tensor blocked by 2nd dimension with block size 8
Definition: dnnl_types.h:267
Out-of-order execution.
Definition: dnnl_types.h:2169
Binary min.
Definition: dnnl_types.h:934
dnnl_primitive_kind_t primitive_kind
The kind of primitive.
Definition: dnnl_types.h:1423
dnnl_format_kind_t
Memory format kind.
Definition: dnnl_types.h:80
permuted 3D tensor
Definition: dnnl_types.h:188
4D tensor blocked by 3rd dimension with block size 4
Definition: dnnl_types.h:285
5D CNN activations tensor blocked by channels with block size 4, an alias to dnnl_aBcde4b ...
Definition: dnnl_types.h:535
dnnl_primitive_kind_t primitive_kind
The kind of primitive.
Definition: dnnl_types.h:1580
Average pooling (alias for dnnl_pooling_avg_exclude_padding)
Definition: dnnl_types.h:908
dnnl_memory_desc_t diff_src_layer_desc
Source gradient layer memory descriptor.
Definition: dnnl_types.h:1617
int axis
Axis for shuffling.
Definition: dnnl_types.h:1288
dnnl_rnn_packed_desc_t rnn_packed_desc
Tensor of packed weights for RNN.
Definition: dnnl_types.h:1182
3D tensor blocked by 2nd dimension with block size 8
Definition: dnnl_types.h:230
Binary add.
Definition: dnnl_types.h:928
3D CNN activations tensor, an alias to dnnl_abc
Definition: dnnl_types.h:425
deconvolution descriptor
Definition: dnnl_types.h:2122
A pooling primitive.
Definition: dnnl_types.h:808
dnnl_dims_t strides
The strides between the outermost blocks.
Definition: dnnl_types.h:1045
rnn descriptor
Definition: dnnl_types.h:2131
dnnl_primitive_kind_t primitive_kind
The kind of primitive.
Definition: dnnl_types.h:1227
Eltwise: logistic.
Definition: dnnl_types.h:867
A descriptor of a shuffle operation.
Definition: dnnl_types.h:1277
const struct dnnl_primitive_desc_iterator * const_dnnl_primitive_desc_iterator_t
A constant primitive descriptor iterator handle.
Definition: dnnl_types.h:1772
dnnl_dims_t kernel
Pooling kernel spatial dimensions.
Definition: dnnl_types.h:1405
Eltwise: hyperbolic tangent non-linearity (tanh) (dst for backward)
Definition: dnnl_types.h:892
Winograd deconvolution.
Definition: dnnl_types.h:847
permuted 4D tensor
Definition: dnnl_types.h:202
number of outputs expected
Definition: dnnl_types.h:2100
#define DNNL_MAX_NDIMS
Maximum number of dimensions a tensor can have.
Definition: dnnl_types.h:1002
5D CNN weights tensor (incl. groups), an alias to dnnl_abcde
Definition: dnnl_types.h:477
dnnl_memory_desc_t diff_src_desc
Source gradient memory descriptor.
Definition: dnnl_types.h:1532
float lrn_k
LRN k parameter.
Definition: dnnl_types.h:1442
dnnl_rnn_direction_t direction
The direction of RNN primitive execution.
Definition: dnnl_types.h:1588
6D CNN weights tensor (incl. groups), an alias to dnnl_defcab
Definition: dnnl_types.h:487
GEMM descriptor (internal)
Definition: dnnl_types.h:2132
plain 1D tensor
Definition: dnnl_types.h:177
Bidirectional execution of RNN primitive with concatenation of the results.
Definition: dnnl_types.h:1568
permuted 2D tensor
Definition: dnnl_types.h:193
permuted 5D tensor
Definition: dnnl_types.h:204
permuted 6D tensor
Definition: dnnl_types.h:190
dnnl_memory_desc_t src_iter_c_desc
Source iteration memory descriptor for cell state.
Definition: dnnl_types.h:1594
stub
Definition: dnnl_types.h:2119
dnnl_memory_desc_t diff_desc
Source and Destination of gradient memory descriptor.
Definition: dnnl_types.h:1362
3D CNN weights tensor, an alias to dnnl_bca
Definition: dnnl_types.h:450
5D CNN activations tensor blocked by channels with block size 8, an alias to dnnl_aBcde8b ...
Definition: dnnl_types.h:538
propagation kind
Definition: dnnl_types.h:2116
An inner product primitive.
Definition: dnnl_types.h:816
Use global statistics.
Definition: dnnl_types.h:964
5D CNN activations tensor blocked by channels with block size 32, an alias to dnnl_aBcde32b ...
Definition: dnnl_types.h:529
GRU cell.
Definition: dnnl_types.h:918
3D CNN activations tensor, an alias to dnnl_acb
Definition: dnnl_types.h:427
The operation was successful.
Definition: dnnl_types.h:41
A descriptor of a Layer Normalization operation.
Definition: dnnl_types.h:1484
dnnl_primitive_kind_t primitive_kind
The kind of primitive.
Definition: dnnl_types.h:1280
8-bit signed integer.
Definition: dnnl_types.h:74
convolution descriptor
Definition: dnnl_types.h:2121
dnnl_memory_desc_t weights_peephole_desc
Weights peephole memory descriptor.
Definition: dnnl_types.h:1610
RNN cell.
Definition: dnnl_types.h:914
Alias for dnnl_unidirectional_left2right.
Definition: dnnl_types.h:1573
A (out-of-place) concat primitive.
Definition: dnnl_types.h:796
Internal weights format for 2x3 Winograd.
Definition: dnnl_types.h:1064
dnnl_dims_t dilates
Convolution dilates in each spatial dimension.
Definition: dnnl_types.h:1254
const struct dnnl_stream * const_dnnl_stream_t
A constant execution stream handle.
Definition: dnnl_types.h:2180
2D RNN statistics tensor, an alias to dnnl_ba
Definition: dnnl_types.h:423
5D CNN activations tensor, an alias to dnnl_acdeb
Definition: dnnl_types.h:437
Intel AVX-512 and Intel Deep Learning Boost (Intel DL Boost) support for Intel Xeon Scalable processo...
Definition: dnnl_types.h:2272
Undefined memory format tag.
Definition: dnnl_types.h:166
permuted 5D tensor
Definition: dnnl_types.h:189
Eltwise: square root.
Definition: dnnl_types.h:859
dnnl_memory_desc_t data_desc
Source and destination memory descriptor.
Definition: dnnl_types.h:1459
int patch
Patch version.
Definition: dnnl_types.h:2217
4D CNN activations tensor blocked by channels with block size 32, an alias to dnnl_aBcd32b ...
Definition: dnnl_types.h:541
permuted 3D tensor
Definition: dnnl_types.h:197
dnnl_memory_desc_t diff_data_desc
Source and destination gradient memory descriptor.
Definition: dnnl_types.h:1433
source memory desc
Definition: dnnl_types.h:2140
Eltwise: swish.
Definition: dnnl_types.h:878
dnnl_prop_kind_t prop_kind
The kind of propagation.
Definition: dnnl_types.h:1426
#define DNNL_RNN_MAX_N_PARTS
Maximum number of parts of RNN weights tensor that require separate computation.
Definition: dnnl_types.h:1092
struct dnnl_stream_attr * dnnl_stream_attr_t
An execution stream attributes handle.
Definition: dnnl_types.h:2185
Memory descriptor.
Definition: dnnl_types.h:1140
dnnl_wino_memory_format_t
Winograd-specific formats.
Definition: dnnl_types.h:1058
2D RNN statistics tensor, an alias to dnnl_ab
Definition: dnnl_types.h:421
dnnl_primitive_kind_t primitive_kind
The kind of primitive.
Definition: dnnl_types.h:1487
dnnl_data_type_t data_type
Data type of the tensor elements.
Definition: dnnl_types.h:1160
A matrix multiplication primitive.
Definition: dnnl_types.h:826
Queried element is not required for given primitive.
Definition: dnnl_types.h:53
dnnl_data_type_t accum_data_type
The accumulator data type. Initialized automatically.
Definition: dnnl_types.h:1411
Eltwise: erf-based gelu.
Definition: dnnl_types.h:886
Indicates the weights have an additional buffer, that depends on the compensation_mask.
Definition: dnnl_types.h:1118
dnnl_blocking_desc_t blocking
Description of the data layout for memory formats that use blocking.
Definition: dnnl_types.h:1178
dnnl_dims_t inner_idxs
The logical indices of the blocks, e.g.
Definition: dnnl_types.h:1054
The operation failed due to an out-of-memory condition.
Definition: dnnl_types.h:43
Backward weights propagation.
Definition: dnnl_types.h:781
5D tensor blocked by 2nd dimension with block size 32
Definition: dnnl_types.h:298
3D CNN weights tensor, an alias to dnnl_acb
Definition: dnnl_types.h:446
dnnl_memory_desc_t diff_weights_desc
Weights gradient memory descriptor.
Definition: dnnl_types.h:1242
Default order execution.
Definition: dnnl_types.h:2165
dnnl_memory_desc_t data_scaleshift_desc
Scale and shift data and gradient memory descriptors.
Definition: dnnl_types.h:1467
workspace memory desc
Definition: dnnl_types.h:2146
dnnl_primitive_kind_t primitive_kind
The kind of primitive.
Definition: dnnl_types.h:1688
dnnl_memory_desc_t diff_dst_desc
Destination gradient memory descriptor.
Definition: dnnl_types.h:1250
eltwise descriptor
Definition: dnnl_types.h:2124
number of inputs expected
Definition: dnnl_types.h:2099
const struct dnnl_stream_attr * const_dnnl_stream_attr_t
A constant execution stream attributes handle.
Definition: dnnl_types.h:2187
shuffle descriptor
Definition: dnnl_types.h:2123
Average pooling include padding.
Definition: dnnl_types.h:904
dnnl_memory_desc_t diff_data_desc
Source and destination gradient memory descriptor.
Definition: dnnl_types.h:1461
Weights format used in 8bit Winograd convolution.
Definition: dnnl_types.h:91
dnnl_memory_desc_t diff_src_desc
Source gradient memory descriptor.
Definition: dnnl_types.h:1720
32-bit/single-precision floating point.
Definition: dnnl_types.h:70
permuted 5D tensor
Definition: dnnl_types.h:203
A structure that contains an index and a memory object, and is used to pass arguments to dnnl_primiti...
Definition: dnnl_types.h:2054
Intel AVX-512, Intel DL Boost and bfloat16 support for Intel Xeon Scalable processor family and Intel...
Definition: dnnl_types.h:2277
dnnl_prop_kind_t prop_kind
The kind of propagation.
Definition: dnnl_types.h:1713
4D CNN activations tensor blocked by channels with block size 8, an alias to dnnl_aBcd8b ...
Definition: dnnl_types.h:550
struct dnnl_primitive_attr * dnnl_primitive_attr_t
A primitive descriptor attributes handle that controls primitive behavior.
Definition: dnnl_types.h:1825
lrn descriptor
Definition: dnnl_types.h:2127
dnnl_memory_desc_t src_layer_desc
Source layer memory descriptor.
Definition: dnnl_types.h:1590
dnnl_memory_desc_t diff_dst_desc
Destination gradient memory descriptor.
Definition: dnnl_types.h:1724
4D tensor blocked by 1st and 2nd dimension with block size 8
Definition: dnnl_types.h:273
A shuffle primitive.
Definition: dnnl_types.h:794
for creating scratchpad memory
Definition: dnnl_types.h:2111
float layer_norm_epsilon
Layer normalization epsilon parameter.
Definition: dnnl_types.h:1511
dnnl_memory_desc_t diff_weights_peephole_desc
Weights gradient peephole memory descriptor.
Definition: dnnl_types.h:1637
struct dnnl_engine * dnnl_engine_t
An engine handle.
Definition: dnnl_types.h:1750
unsigned int flags
RNN cell flags.
Definition: dnnl_types.h:1644
permuted 4D tensor
Definition: dnnl_types.h:198
4D tensor blocked by 2nd dimension with block size 4
Definition: dnnl_types.h:248
Binary max.
Definition: dnnl_types.h:932
Unspecified format kind.
Definition: dnnl_types.h:85
dnnl_prop_kind_t prop_kind
The kind of propagation.
Definition: dnnl_types.h:1283
2D CNN weights tensor, an alias to dnnl_ba
Definition: dnnl_types.h:442
Average pooling exclude padding.
Definition: dnnl_types.h:906
3D CNN activations tensor blocked by channels with block size 16, an alias to dnnl_aBc16b ...
Definition: dnnl_types.h:556
dnnl_alg_kind_t
Kinds of algorithms.
Definition: dnnl_types.h:836
4D CNN weights tensor, an alias to dnnl_cdba
Definition: dnnl_types.h:454
5D CNN weights tensor, an alias to dnnl_bacde
Definition: dnnl_types.h:464
permuted 3D tensor
Definition: dnnl_types.h:200
dnnl_rnn_flags_t
Flags for RNN cell.
Definition: dnnl_types.h:1555
uint64_t flags
The flags contain arbitrary extra information, such as compensation.
Definition: dnnl_types.h:1127
dnnl_memory_desc_t diff_dst_iter_desc
Destination gradient iteration memory descriptor for hidden state.
Definition: dnnl_types.h:1631
Description of extra information stored in memory.
Definition: dnnl_types.h:1124
4D CNN weights tensor (incl. groups), an alias to dnnl_dcab
Definition: dnnl_types.h:475
dnnl_prop_kind_t prop_kind
The kind of propagation.
Definition: dnnl_types.h:1528
An LRN primitive.
Definition: dnnl_types.h:810
A descriptor of a Local Response Normalization (LRN) operation.
Definition: dnnl_types.h:1420
dnnl_softmax_desc_t dnnl_logsoftmax_desc_t
A descriptor of a LogSoftmax operation.
Definition: dnnl_types.h:1374
int ndims
Number of dimensions.
Definition: dnnl_types.h:1142
dnnl_stream_flags_t
Stream flags.
Definition: dnnl_types.h:2162
dnnl_dim_t dnnl_dims_t[DNNL_MAX_NDIMS]
A type to describe tensor dimensions.
Definition: dnnl_types.h:1037
Undefined propagation type.
Definition: dnnl_types.h:764
dnnl_memory_desc_t diff_data_desc
Source and destination gradient memory descriptor.
Definition: dnnl_types.h:1322
Primitive or engine failed on execution.
Definition: dnnl_types.h:51
op descriptor
Definition: dnnl_types.h:2120
6D tensor blocked by 2nd dimension with block size 8
Definition: dnnl_types.h:344
dnnl_memory_desc_t weights_iter_desc
Weights iteration memory descriptor.
Definition: dnnl_types.h:1598
const struct dnnl_post_ops * const_dnnl_post_ops_t
A constant post operation chain handle.
Definition: dnnl_types.h:1854
Eltwise: exponent.
Definition: dnnl_types.h:869
1D tensor, an alias to dnnl_a
Definition: dnnl_types.h:415
Local response normalization (LRN) across multiple channels.
Definition: dnnl_types.h:910
dnnl_primitive_kind_t primitive_kind
The kind of primitive.
Definition: dnnl_types.h:1385
resampling descriptor
Definition: dnnl_types.h:2136
const struct dnnl_primitive_attr * const_dnnl_primitive_attr_t
A constant primitive descriptor attributes handle.
Definition: dnnl_types.h:1828
A matrix multiplication primitive (internal).
Definition: dnnl_types.h:820
dnnl_memory_desc_t diff_bias_desc
Bias gradient memory descriptor.
Definition: dnnl_types.h:1627
plain 3D tensor
Definition: dnnl_types.h:179
Use scale and shift parameters.
Definition: dnnl_types.h:977
3D CNN activations tensor blocked by channels with block size 32, an alias to dnnl_aBc32b ...
Definition: dnnl_types.h:553
A descriptor of a Batch Normalization operation.
Definition: dnnl_types.h:1451
5D CNN weights tensor, an alias to dnnl_bcdea
Definition: dnnl_types.h:470
struct dnnl_primitive * dnnl_primitive_t
A primitive handle.
Definition: dnnl_types.h:1865
Intel Advanced Vector Extensions 2 (Intel AVX2)
Definition: dnnl_types.h:2255
3D CNN activations tensor blocked by channels with block size 4, an alias to dnnl_aBc4b ...
Definition: dnnl_types.h:559
struct dnnl_stream * dnnl_stream_t
An execution stream handle.
Definition: dnnl_types.h:2178
4D CNN weights tensor, an alias to dnnl_bacd
Definition: dnnl_types.h:460
Eltwise: exponential linear unit (elu) (dst for backward)
Definition: dnnl_types.h:894
Eltwise: exponential linear unit (elu)
Definition: dnnl_types.h:853
Forward data propagation (alias for dnnl_forward_inference).
Definition: dnnl_types.h:773
6D tensor blocked by 2nd dimension with block size 16
Definition: dnnl_types.h:339
5D CNN activations tensor blocked by channels with block size 16, an alias to dnnl_aBcde16b ...
Definition: dnnl_types.h:532
An opaque structure to describe a primitive descriptor.
dnnl_prop_kind_t prop_kind
The kind of propagation.
Definition: dnnl_types.h:1231
struct dnnl_primitive_desc_iterator * dnnl_primitive_desc_iterator_t
A primitive descriptor iterator handle.
Definition: dnnl_types.h:1769
Eltwise: abs.
Definition: dnnl_types.h:857
dnnl_memory_desc_t data_desc
Source and destination memory descriptor.
Definition: dnnl_types.h:1320
Internal weights format for 4x3 Winograd.
Definition: dnnl_types.h:1066
Forward data propagation (training mode).
Definition: dnnl_types.h:767
permuted 5D tensor
Definition: dnnl_types.h:187
dnnl_memory_desc_t data_desc
Source and destination memory descriptor, and source and destination gradient memory descriptor...
Definition: dnnl_types.h:1286
Eltwise: tanh-based gelu (alias for dnnl_eltwise_gelu_tanh)
Definition: dnnl_types.h:876
dnnl_memory_desc_t diff_src_desc
Source gradient memory descriptor.
Definition: dnnl_types.h:1238
A rnn primitive.
Definition: dnnl_types.h:818
An opaque structure to describe an execution stream.
4D CNN activations tensor, an alias to dnnl_abcd
Definition: dnnl_types.h:429
A logsoftmax primitive.
Definition: dnnl_types.h:824
permuted 5D tensor
Definition: dnnl_types.h:196
The operation failed because requested functionality is not implemented.
Definition: dnnl_types.h:47
Eltwise: gelu.
Definition: dnnl_types.h:874
dnnl_data_type_t accum_data_type
The accumulator data type. Initialized automatically.
Definition: dnnl_types.h:1698
dnnl_memory_desc_t diff_src_iter_desc
Source gradient iter memory descriptor for hidden state.
Definition: dnnl_types.h:1619
dnnl_memory_desc_t diff_src_desc
Source gradient memory descriptor.
Definition: dnnl_types.h:1397
Eltwise: hyperbolic tangent non-linearity (tanh)
Definition: dnnl_types.h:851
struct dnnl_post_ops * dnnl_post_ops_t
A post operation chain handle.
Definition: dnnl_types.h:1851
A descriptor of resampling operation.
Definition: dnnl_types.h:1707
dnnl_memory_desc_t diff_dst_desc
Destination gradient memory descriptor.
Definition: dnnl_types.h:1401
dnnl_prop_kind_t
Kinds of propagation.
Definition: dnnl_types.h:761
stub
Definition: dnnl_types.h:2139
Nearest Neighbor Resampling Method.
Definition: dnnl_types.h:936
permuted 4D tensor
Definition: dnnl_types.h:195
dnnl_memory_desc_t diff_dst_desc
Destination gradient memory descriptor.
Definition: dnnl_types.h:1544
16-bit/half-precision floating point.
Definition: dnnl_types.h:66
source engine
Definition: dnnl_types.h:2113
dnnl_memory_desc_t diff_bias_desc
Bias gradient memory descriptor.
Definition: dnnl_types.h:1246
dnnl_alg_kind_t alg_kind
The kind of the resampling algorithm.
Definition: dnnl_types.h:1716
dnnl_normalization_flags_t
Flags for normalization primitives.
Definition: dnnl_types.h:942
weights grad. memory desc
Definition: dnnl_types.h:2143
A convolution primitive.
Definition: dnnl_types.h:800
dnnl_primitive_kind_t primitive_kind
The kind of primitive.
Definition: dnnl_types.h:1302
memory desc of an execute argument
Definition: dnnl_types.h:2148
Backward bias propagation.
Definition: dnnl_types.h:783
5D CNN activations tensor, an alias to dnnl_abcde
Definition: dnnl_types.h:435
4D CNN weights tensor, an alias to dnnl_abcd
Definition: dnnl_types.h:452
Eltwise: logistic (dst for backward)
Definition: dnnl_types.h:898
dnnl_memory_desc_t diff_weights_layer_desc
Weights gradient layer memory descriptor.
Definition: dnnl_types.h:1623
weights memory descriptor desc
Definition: dnnl_types.h:2142
unsigned gpu_runtime
GPU runtime.
Definition: dnnl_types.h:2220
Linear Resampling Method.
Definition: dnnl_types.h:938
3D CNN weights tensor, an alias to dnnl_abc
Definition: dnnl_types.h:444
dnnl_cpu_isa_t
CPU instruction set flags.
Definition: dnnl_types.h:2244
int compensation_mask
Compensation mask.
Definition: dnnl_types.h:1129
Eltwise: soft_relu.
Definition: dnnl_types.h:865
plain 5D tensor
Definition: dnnl_types.h:181
destination engine
Definition: dnnl_types.h:2114
dnnl_primitive_kind_t primitive_kind
The kind of primitive.
Definition: dnnl_types.h:1524
Eltwise: square.
Definition: dnnl_types.h:855
float alpha
Algorithm specific parameter.
Definition: dnnl_types.h:1343