Deep Neural Network Library (DNNL)
1.90.1
Performance library for Deep Learning
Go to the documentation of this file.
464 dnnl_NCw16n16c = dnnl_ABc16a16b,
465 dnnl_NCdhw16n16c = dnnl_ABcde16a16b,
466 dnnl_NChw16n16c = dnnl_ABcd16a16b,
467 dnnl_NChw32n32c = dnnl_ABcd32a32b,
470 dnnl_IOw16o16i = dnnl_BAc16a16b,
471 dnnl_IOw16i16o = dnnl_BAc16b16a,
472 dnnl_OIw16i16o = dnnl_ABc16b16a,
473 dnnl_OIw16o16i = dnnl_ABc16a16b,
474 dnnl_Oiw16o = dnnl_Abc16a,
475 dnnl_OIw4i16o4i = dnnl_ABc4b16a4b,
476 dnnl_OIw4i4o = dnnl_ABc4b4a,
477 dnnl_Oiw4o = dnnl_Abc4a,
478 dnnl_OIw8i16o2i = dnnl_ABc8b16a2b,
479 dnnl_OIw8i8o = dnnl_ABc8b8a,
480 dnnl_OIw8o16i2o = dnnl_ABc8a16b2a,
481 dnnl_IOw8o16i2o = dnnl_BAc8a16b2a,
482 dnnl_OIw8o8i = dnnl_ABc8a8b,
483 dnnl_Owi16o = dnnl_Acb16a,
484 dnnl_Owi4o = dnnl_Acb4a,
485 dnnl_Owi8o = dnnl_Acb8a,
488 dnnl_IOhw16i16o = dnnl_BAcd16b16a,
489 dnnl_IOhw16o16i = dnnl_BAcd16a16b,
490 dnnl_Ohwi16o = dnnl_Acdb16a,
491 dnnl_Ohwi32o = dnnl_Acdb32a,
492 dnnl_Ohwi4o = dnnl_Acdb4a,
493 dnnl_Ohwi8o = dnnl_Acdb8a,
494 dnnl_OIhw16i16o = dnnl_ABcd16b16a,
495 dnnl_OIhw16o16i = dnnl_ABcd16a16b,
496 dnnl_Oihw16o = dnnl_Abcd16a,
497 dnnl_OIhw4i16o4i = dnnl_ABcd4b16a4b,
498 dnnl_OIhw4i4o = dnnl_ABcd4b4a,
499 dnnl_Oihw4o = dnnl_Abcd4a,
500 dnnl_OIhw8i16o2i = dnnl_ABcd8b16a2b,
502 dnnl_OIhw8o16i2o = dnnl_ABcd8a16b2a,
503 dnnl_IOhw8o16i2o = dnnl_BAcd8a16b2a,
504 dnnl_OIhw8o8i = dnnl_ABcd8a8b,
507 dnnl_Odhwi16o = dnnl_Acdeb16a,
508 dnnl_Odhwi4o = dnnl_Acdeb4a,
509 dnnl_Odhwi8o = dnnl_Acdeb8a,
510 dnnl_OIdhw16i16o = dnnl_ABcde16b16a,
511 dnnl_OIdhw16o16i = dnnl_ABcde16a16b,
512 dnnl_Oidhw16o = dnnl_Abcde16a,
513 dnnl_OIdhw4i4o = dnnl_ABcde4b4a,
514 dnnl_Oidhw4o = dnnl_Abcde4a,
515 dnnl_OIdhw8i16o2i = dnnl_ABcde8b16a2b,
516 dnnl_OIdhw8i8o = dnnl_ABcde8b8a,
517 dnnl_OIdhw8o16i2o = dnnl_ABcde8a16b2a,
518 dnnl_IOdhw8o16i2o = dnnl_BAcde8a16b2a,
519 dnnl_OIdhw8o8i = dnnl_ABcde8a8b,
520 dnnl_IOdhw16i16o = dnnl_BAcde16b16a,
523 dnnl_Goiw16g = dnnl_Abcd16a,
524 dnnl_gIOw16o16i = dnnl_aCBd16b16c,
525 dnnl_gIOw16i16o = dnnl_aCBd16c16b,
526 dnnl_gOIw16i16o = dnnl_aBCd16c16b,
527 dnnl_gOIw16o16i = dnnl_aBCd16b16c,
529 dnnl_gOIw4i16o4i = dnnl_aBCd4c16b4c,
530 dnnl_gOIw4i4o = dnnl_aBCd4c4b,
532 dnnl_gOIw8i16o2i = dnnl_aBCd8c16b2c,
533 dnnl_gOIw8i8o = dnnl_aBCd8c8b,
534 dnnl_gOIw8o16i2o = dnnl_aBCd8b16c2b,
535 dnnl_gIOw8o16i2o = dnnl_aCBd8b16c2b,
536 dnnl_gOIw8o8i = dnnl_aBCd8b8c,
537 dnnl_gOwi16o = dnnl_aBdc16b,
538 dnnl_gOwi4o = dnnl_aBdc4b,
539 dnnl_gOwi8o = dnnl_aBdc8b,
542 dnnl_gIOhw16i16o = dnnl_aCBde16c16b,
543 dnnl_gIOhw16o16i = dnnl_aCBde16b16c,
544 dnnl_gOhwi16o = dnnl_aBdec16b,
545 dnnl_gOhwi32o = dnnl_aBdec32b,
546 dnnl_gOhwi4o = dnnl_aBdec4b,
547 dnnl_gOhwi8o = dnnl_aBdec8b,
548 dnnl_Goihw16g = dnnl_Abcde16a,
549 dnnl_gOIhw16i16o = dnnl_aBCde16c16b,
550 dnnl_gOIhw16o16i = dnnl_aBCde16b16c,
552 dnnl_gOIhw2i8o4i = dnnl_aBCde2c8b4c,
553 dnnl_gOIhw4i16o4i = dnnl_aBCde4c16b4c,
554 dnnl_gOIhw4i4o = dnnl_aBCde4c4b,
555 dnnl_gOIhw4o4i = dnnl_aBCde4b4c,
557 dnnl_Goihw8g = dnnl_Abcde8a,
558 dnnl_gOIhw8i16o2i = dnnl_aBCde8c16b2c,
559 dnnl_gOIhw8i8o = dnnl_aBCde8c8b,
560 dnnl_gOIhw8o16i2o = dnnl_aBCde8b16c2b,
561 dnnl_gIOhw8o16i2o = dnnl_aCBde8b16c2b,
562 dnnl_gOIhw8o8i = dnnl_aBCde8b8c,
564 dnnl_OIhw4o8i8o4i = dnnl_ABcd4a8b8a4b,
565 dnnl_OIhw2o8i8o2i = dnnl_ABcd2a8b8a2b,
566 dnnl_gOIhw4o8i8o4i = dnnl_aBCde4b8c8b4c,
567 dnnl_gOIhw2o8i8o2i = dnnl_aBCde2b8c8b2c,
570 dnnl_gIOdhw16i16o = dnnl_aCBdef16c16b,
571 dnnl_gOdhwi16o = dnnl_aBdefc16b,
572 dnnl_gOdhwi4o = dnnl_aBdefc4b,
573 dnnl_gOdhwi8o = dnnl_aBdefc8b,
574 dnnl_gOIdhw16i16o = dnnl_aBCdef16c16b,
575 dnnl_gOIdhw16o16i = dnnl_aBCdef16b16c,
577 dnnl_gOIdhw4i4o = dnnl_aBCdef4c4b,
579 dnnl_gOIdhw8i16o2i = dnnl_aBCdef8c16b2c,
580 dnnl_gOIdhw8i8o = dnnl_aBCdef8c8b,
581 dnnl_gOIdhw8o16i2o = dnnl_aBCdef8b16c2b,
582 dnnl_gIOdhw8o16i2o = dnnl_aCBdef8b16c2b,
583 dnnl_gOIdhw8o8i = dnnl_aBCdef8b8c,
584 dnnl_Goidhw16g = dnnl_Abcdef16a,
775 #define DNNL_MAX_NDIMS 12
829 dnnl_packed_format_undef = 0,
832 } dnnl_rnn_packed_memory_format_t;
836 #define DNNL_RNN_MAX_N_PARTS 4
840 dnnl_rnn_packed_memory_format_t format;
847 size_t offset_compensation;
854 dnnl_memory_extra_flag_none = 0x0U,
863 dnnl_memory_extra_flag_scale_adjust = 0x2U,
864 dnnl_memory_extra_flag_gpu_rnn_u8s8_compensation = 0x4U,
873 int compensation_mask;
943 #define DNNL_MEMORY_NONE (NULL)
944 #define DNNL_MEMORY_ALLOCATE ((void *)(size_t)-1)
1151 float batch_norm_epsilon;
1183 float layer_norm_epsilon;
1338 typedef const struct dnnl_engine *const_dnnl_engine_t;
1441 #define DNNL_ARG_SRC_0 1
1442 #define DNNL_ARG_SRC DNNL_ARG_SRC_0
1443 #define DNNL_ARG_SRC_LAYER DNNL_ARG_SRC_0
1444 #define DNNL_ARG_FROM DNNL_ARG_SRC_0
1446 #define DNNL_ARG_SRC_1 2
1447 #define DNNL_ARG_SRC_ITER DNNL_ARG_SRC_1
1449 #define DNNL_ARG_SRC_2 3
1450 #define DNNL_ARG_SRC_ITER_C DNNL_ARG_SRC_2
1452 #define DNNL_ARG_DST_0 17
1453 #define DNNL_ARG_DST DNNL_ARG_DST_0
1454 #define DNNL_ARG_TO DNNL_ARG_DST_0
1455 #define DNNL_ARG_DST_LAYER DNNL_ARG_DST_0
1457 #define DNNL_ARG_DST_1 18
1458 #define DNNL_ARG_DST_ITER DNNL_ARG_DST_1
1460 #define DNNL_ARG_DST_2 19
1461 #define DNNL_ARG_DST_ITER_C DNNL_ARG_DST_2
1463 #define DNNL_ARG_WEIGHTS_0 33
1464 #define DNNL_ARG_WEIGHTS DNNL_ARG_WEIGHTS_0
1465 #define DNNL_ARG_SCALE_SHIFT DNNL_ARG_WEIGHTS_0
1466 #define DNNL_ARG_WEIGHTS_LAYER DNNL_ARG_WEIGHTS_0
1468 #define DNNL_ARG_WEIGHTS_1 34
1469 #define DNNL_ARG_WEIGHTS_ITER DNNL_ARG_WEIGHTS_1
1471 #define DNNL_ARG_BIAS 41
1473 #define DNNL_ARG_MEAN 49
1474 #define DNNL_ARG_VARIANCE 50
1476 #define DNNL_ARG_WORKSPACE 64
1477 #define DNNL_ARG_SCRATCHPAD 80
1479 #define DNNL_ARG_DIFF_SRC_0 129
1480 #define DNNL_ARG_DIFF_SRC DNNL_ARG_DIFF_SRC_0
1481 #define DNNL_ARG_DIFF_SRC_LAYER DNNL_ARG_DIFF_SRC_0
1483 #define DNNL_ARG_DIFF_SRC_1 130
1484 #define DNNL_ARG_DIFF_SRC_ITER DNNL_ARG_DIFF_SRC_1
1486 #define DNNL_ARG_DIFF_SRC_2 131
1487 #define DNNL_ARG_DIFF_SRC_ITER_C DNNL_ARG_DIFF_SRC_2
1489 #define DNNL_ARG_DIFF_DST_0 145
1490 #define DNNL_ARG_DIFF_DST DNNL_ARG_DIFF_DST_0
1491 #define DNNL_ARG_DIFF_DST_LAYER DNNL_ARG_DIFF_DST_0
1493 #define DNNL_ARG_DIFF_DST_1 146
1494 #define DNNL_ARG_DIFF_DST_ITER DNNL_ARG_DIFF_DST_1
1496 #define DNNL_ARG_DIFF_DST_2 147
1497 #define DNNL_ARG_DIFF_DST_ITER_C DNNL_ARG_DIFF_DST_2
1499 #define DNNL_ARG_DIFF_WEIGHTS_0 161
1500 #define DNNL_ARG_DIFF_WEIGHTS DNNL_ARG_DIFF_WEIGHTS_0
1501 #define DNNL_ARG_DIFF_SCALE_SHIFT DNNL_ARG_DIFF_WEIGHTS_0
1502 #define DNNL_ARG_DIFF_WEIGHTS_LAYER DNNL_ARG_DIFF_WEIGHTS_0
1504 #define DNNL_ARG_DIFF_WEIGHTS_1 162
1505 #define DNNL_ARG_DIFF_WEIGHTS_ITER DNNL_ARG_DIFF_WEIGHTS_1
1507 #define DNNL_ARG_DIFF_BIAS 169
1509 #define DNNL_ARG_MULTIPLE_SRC 1024
1510 #define DNNL_ARG_MULTIPLE_DST 2048
number of inputs expected
Definition: dnnl_types.h:1563
2D RNN statistics tensor, an alias to dnnl_ba
Definition: dnnl_types.h:345
Default order execution.
Definition: dnnl_types.h:1620
3D CNN activations tensor blocked by channels with block size 16, an alias to dnnl_aBc16b
Definition: dnnl_types.h:456
Primitive or engine failed on execution.
Definition: dnnl_types.h:61
Undefined data type, used for empty memory descriptors.
Definition: dnnl_types.h:69
GRU cell with linear before reset.
Definition: dnnl_types.h:717
Internal weights format for 4x3 Winograd.
Definition: dnnl_types.h:809
Backward bias propagation.
Definition: dnnl_types.h:609
2D CNN activations tensor, an alias to dnnl_ba
Definition: dnnl_types.h:341
permuted 3D tensor
Definition: dnnl_types.h:197
Description of tensor of packed weights for rnn.
Definition: dnnl_types.h:838
(scratch) memory, additional to all inputs and outputs memory (bytes)
Definition: dnnl_types.h:1572
inner product descriptor
Definition: dnnl_types.h:1594
A sum primitive.
Definition: dnnl_types.h:624
4D tensor blocked by 2nd dimension with block size 8
Definition: dnnl_types.h:244
plain 6D tensor
Definition: dnnl_types.h:187
2D CNN activations tensor, an alias to dnnl_ab
Definition: dnnl_types.h:339
A pooling primitive.
Definition: dnnl_types.h:634
dnnl_query_t
Primitive descriptor query specification.
Definition: dnnl_types.h:1557
32-bit/single-precision floating point.
Definition: dnnl_types.h:75
2D CNN weights tensor, an alias to dnnl_ba
Definition: dnnl_types.h:364
32-bit signed integer.
Definition: dnnl_types.h:77
An opaque structure to describe a primitive descriptor iterator.
dnnl_dim_t dnnl_dims_t[DNNL_MAX_NDIMS]
A type to describe tensor dimensions.
Definition: dnnl_types.h:780
scratchpad memory desc
Definition: dnnl_types.h:1608
Eltwise: abs.
Definition: dnnl_types.h:673
convolution descriptor
Definition: dnnl_types.h:1585
const struct dnnl_primitive_attr * const_dnnl_primitive_attr_t
A constant primitive descriptor attributes handle.
Definition: dnnl_types.h:1396
Forward data propagation (training mode).
Definition: dnnl_types.h:593
CPU engine.
Definition: dnnl_types.h:1324
permuted 5D tensor
Definition: dnnl_types.h:204
8-bit signed integer.
Definition: dnnl_types.h:79
4D CNN activations tensor blocked by channels with block size 8, an alias to dnnl_aBcd8b
Definition: dnnl_types.h:453
plain 4D tensor
Definition: dnnl_types.h:185
3D CNN weights tensor, an alias to dnnl_cba
Definition: dnnl_types.h:370
4D tensor blocked by 1st and 2nd dimension with block size 8
Definition: dnnl_types.h:249
An opaque structure to describe an engine.
5D tensor blocked by 2nd dimension with block size 16
Definition: dnnl_types.h:259
permuted 3D tensor
Definition: dnnl_types.h:192
binary descriptor
Definition: dnnl_types.h:1597
Primitive iterator passed over last primitive descriptor.
Definition: dnnl_types.h:59
primitive kind
Definition: dnnl_types.h:1561
Winograd deconvolution.
Definition: dnnl_types.h:663
dnnl_engine_kind_t
Kinds of engines.
Definition: dnnl_types.h:1320
A descriptor of a convolution operation.
Definition: dnnl_types.h:956
Eltwise: soft_relu.
Definition: dnnl_types.h:681
#define DNNL_MAX_NDIMS
Maximum number of dimensions a tensor can have.
Definition: dnnl_types.h:774
dnnl_status_t
Status values returned by the library functions.
Definition: dnnl_types.h:49
In-order execution.
Definition: dnnl_types.h:1622
An inner product primitive.
Definition: dnnl_types.h:642
5D tensor blocked by 2nd dimension with block size 8
Definition: dnnl_types.h:275
An auxiliary structure to specify primitive's inputs/outputs at execution.
Definition: dnnl_types.h:1519
A (out-of-place) concat primitive.
Definition: dnnl_types.h:622
4D CNN weights tensor, an alias to dnnl_bcda
Definition: dnnl_types.h:380
dnnl_rnn_flags_t
Flags for RNN cell.
Definition: dnnl_types.h:1216
A descriptor for an RNN operation.
Definition: dnnl_types.h:1234
Winograd convolution.
Definition: dnnl_types.h:657
Undefined memory format, used for empty memory descriptors.
Definition: dnnl_types.h:803
A layer normalization primitive.
Definition: dnnl_types.h:640
permuted 5D tensor
Definition: dnnl_types.h:191
Eltwise: parametric exponential linear unit (elu)
Definition: dnnl_types.h:669
A descriptor of an inner product operation.
Definition: dnnl_types.h:1187
stub
Definition: dnnl_types.h:1583
struct dnnl_stream * dnnl_stream_t
An execution stream handle.
Definition: dnnl_types.h:1633
4D CNN weights tensor, an alias to dnnl_abcd
Definition: dnnl_types.h:374
Max pooling.
Definition: dnnl_types.h:694
Binary add.
Definition: dnnl_types.h:719
destination grad. memory desc
Definition: dnnl_types.h:1606
plain 1D tensor
Definition: dnnl_types.h:182
plain 2D tensor
Definition: dnnl_types.h:183
no query
Definition: dnnl_types.h:1558
Eltwise: bounded_relu.
Definition: dnnl_types.h:679
void * dnnl_op_desc_t
A pointer to any of the operation descriptors.
Definition: dnnl_types.h:951
Undefined primitive.
Definition: dnnl_types.h:616
5D CNN activations tensor, an alias to dnnl_acdeb
Definition: dnnl_types.h:359
dnnl_rnn_direction_t
A direction of RNN primitive execution.
Definition: dnnl_types.h:1219
const struct dnnl_primitive_desc * const_dnnl_primitive_desc_t
A constant primitive descriptor handle.
Definition: dnnl_types.h:1369
workspace memory desc
Definition: dnnl_types.h:1607
op descriptor
Definition: dnnl_types.h:1584
5D CNN weights tensor, an alias to dnnl_abcde
Definition: dnnl_types.h:384
A descriptor of a pooling operation.
Definition: dnnl_types.h:1068
5D CNN weights tensor (incl. groups), an alias to dnnl_acbde
Definition: dnnl_types.h:399
batch normalization descriptor
Definition: dnnl_types.h:1592
dnnl_format_kind_t
Memory format kind.
Definition: dnnl_types.h:85
Local response normalization (LRN) across multiple channels.
Definition: dnnl_types.h:701
permuted 5D tensor
Definition: dnnl_types.h:201
for creating scratchpad memory
Definition: dnnl_types.h:1575
Undefined memory format tag.
Definition: dnnl_types.h:171
dnnl_wino_memory_format_t
Winograd-specific formats.
Definition: dnnl_types.h:801
A batch normalization primitive.
Definition: dnnl_types.h:638
A rnn primitive.
Definition: dnnl_types.h:644
4D CNN weights tensor, an alias to dnnl_cdba
Definition: dnnl_types.h:376
Average pooling include padding.
Definition: dnnl_types.h:696
permuted 4D tensor
Definition: dnnl_types.h:203
The operation was successful.
Definition: dnnl_types.h:51
Use scale and shift parameters.
Definition: dnnl_types.h:750
16-bit/half-precision floating point.
Definition: dnnl_types.h:71
4D tensor blocked by 2nd dimension with block size 4
Definition: dnnl_types.h:236
The operation failed due to an out-of-memory condition.
Definition: dnnl_types.h:53
Eltwise: gelu.
Definition: dnnl_types.h:690
3D CNN weights tensor, an alias to dnnl_acb
Definition: dnnl_types.h:368
The operation failed because of incorrect function arguments.
Definition: dnnl_types.h:55
Backward propagation (with respect to all parameters).
Definition: dnnl_types.h:603
3D CNN activations tensor, an alias to dnnl_acb
Definition: dnnl_types.h:349
A descriptor of a Softmax operation.
Definition: dnnl_types.h:1052
Direct deconvolution.
Definition: dnnl_types.h:661
#define DNNL_RNN_MAX_N_PARTS
Maximum number of parts of RNN weights tensor that require separate computation.
Definition: dnnl_types.h:835
source gradient memory desc
Definition: dnnl_types.h:1602
const struct dnnl_primitive_desc_iterator * const_dnnl_primitive_desc_iterator_t
A constant primitive descriptor iterator handle.
Definition: dnnl_types.h:1353
Eltwise: swish.
Definition: dnnl_types.h:692
An opaque structure to describe a primitive descriptor.
Out-of-order execution.
Definition: dnnl_types.h:1624
A deconvolution primitive.
Definition: dnnl_types.h:628
source memory desc
Definition: dnnl_types.h:1601
Backward weights propagation.
Definition: dnnl_types.h:607
An LRN primitive.
Definition: dnnl_types.h:636
execution engine
Definition: dnnl_types.h:1560
4D CNN weights tensor, an alias to dnnl_bacd
Definition: dnnl_types.h:382
dnnl_memory_extra_flags_t
Flags for memory special features.
Definition: dnnl_types.h:852
source engine
Definition: dnnl_types.h:1577
Backward data propagation.
Definition: dnnl_types.h:605
weights memory descriptor desc
Definition: dnnl_types.h:1603
Eltwise: linear.
Definition: dnnl_types.h:677
4D RNN states tensor in the format (num_layers, num_directions, batch, state channels).
Definition: dnnl_types.h:409
propagation kind
Definition: dnnl_types.h:1580
4D tensor blocked by 2nd dimension with block size 16
Definition: dnnl_types.h:230
number of outputs expected
Definition: dnnl_types.h:1564
Packed weights format used in RNN.
Definition: dnnl_types.h:98
4D CNN weights tensor, an alias to dnnl_acdb
Definition: dnnl_types.h:378
Undefined memory format kind, used for empty memory descriptors.
Definition: dnnl_types.h:87
Use global statistics.
Definition: dnnl_types.h:737
permuted 5D tensor
Definition: dnnl_types.h:193
A descriptor of a Local Response Normalization (LRN) operation.
Definition: dnnl_types.h:1101
deconvolution descriptor
Definition: dnnl_types.h:1586
A descriptor of a shuffle operation.
Definition: dnnl_types.h:999
rnn descriptor
Definition: dnnl_types.h:1595
dnnl_normalization_flags_t
Flags for batch normalization primitive.
Definition: dnnl_types.h:725
Version type.
Definition: dnnl_types.h:41
A descriptor of a Batch Normalization operation.
Definition: dnnl_types.h:1127
Internal weights format for 2x3 Winograd.
Definition: dnnl_types.h:806
memory consumption – extra
Definition: dnnl_types.h:1567
An unspecified engine.
Definition: dnnl_types.h:1322
A descriptor of a element-wise operation.
Definition: dnnl_types.h:1016
struct dnnl_primitive_attr * dnnl_primitive_attr_t
A primitive descriptor attributes handle that controls primitive behavior.
Definition: dnnl_types.h:1393
GPU engine.
Definition: dnnl_types.h:1326
shuffle descriptor
Definition: dnnl_types.h:1587
dnnl_convolution_desc_t dnnl_deconvolution_desc_t
A descriptor of a deconvolution operation.
Definition: dnnl_types.h:996
Generic description of blocked data layout for most memory formats.
Definition: dnnl_types.h:785
8-bit unsigned integer.
Definition: dnnl_types.h:81
5D RNN weights tensor in the format (num_layers, num_directions, num_gates, output_channels,...
Definition: dnnl_types.h:423
Internal weights format for 2x3 Winograd.
Definition: dnnl_types.h:805
A descriptor of a Layer Normalization operation.
Definition: dnnl_types.h:1155
2D RNN statistics tensor, an alias to dnnl_ab
Definition: dnnl_types.h:343
5D CNN activations tensor blocked by channels with block size 16, an alias to dnnl_aBcde16b
Definition: dnnl_types.h:438
Description of tensor of weights for winograd 2x3 convolution.
Definition: dnnl_types.h:813
RNN cell.
Definition: dnnl_types.h:705
Binary mul.
Definition: dnnl_types.h:721
struct dnnl_engine * dnnl_engine_t
An engine handle.
Definition: dnnl_types.h:1333
6D tensor blocked by 2nd dimension with block size 4
Definition: dnnl_types.h:292
dnnl_alg_kind_t
Kinds of algorithms.
Definition: dnnl_types.h:652
4D CNN activations tensor, an alias to dnnl_abcd
Definition: dnnl_types.h:351
const struct dnnl_post_ops * const_dnnl_post_ops_t
A constant post operation chain handle.
Definition: dnnl_types.h:1422
Eltwise: hyperbolic tangent non-linearity (tanh)
Definition: dnnl_types.h:667
LRN within a single channel.
Definition: dnnl_types.h:703
5D tensor blocked by 2nd dimension with block size 4
Definition: dnnl_types.h:266
struct dnnl_primitive * dnnl_primitive_t
A primitive handle.
Definition: dnnl_types.h:1433
5D RNN weights tensor in the format (num_layers, num_directions, input_channels, num_gates,...
Definition: dnnl_types.h:416
3D CNN activations tensor blocked by channels with block size 8, an alias to dnnl_aBc8b
Definition: dnnl_types.h:462
An opaque structure for primitive descriptor attributes.
runtime estimation (seconds)
Definition: dnnl_types.h:1566
dnnl_data_type_t
Data type specification.
Definition: dnnl_types.h:67
destination engine
Definition: dnnl_types.h:1578
Queried element is not required for given primitive.
Definition: dnnl_types.h:63
Unidirectional execution of RNN primitive from left to right.
Definition: dnnl_types.h:1221
Bidirectional execution of RNN primitive with concatenation of the results.
Definition: dnnl_types.h:1226
5D CNN activations tensor, an alias to dnnl_abcde
Definition: dnnl_types.h:357
Average pooling exclude padding.
Definition: dnnl_types.h:698
struct dnnl_primitive_desc_iterator * dnnl_primitive_desc_iterator_t
A primitive descriptor iterator handle.
Definition: dnnl_types.h:1350
4D CNN activations tensor, an alias to dnnl_bcda
Definition: dnnl_types.h:355
stub
Definition: dnnl_types.h:1600
5D CNN weights tensor, an alias to dnnl_acdeb
Definition: dnnl_types.h:388
permuted 3D tensor
Definition: dnnl_types.h:202
An opaque structure for a chain of post operations.
A shuffle primitive.
Definition: dnnl_types.h:620
3D CNN weights tensor, an alias to dnnl_abc
Definition: dnnl_types.h:366
permuted 4D tensor
Definition: dnnl_types.h:200
Eltwise: square root.
Definition: dnnl_types.h:675
An element-wise primitive.
Definition: dnnl_types.h:630
plain 3D tensor
Definition: dnnl_types.h:184
destination memory desc
Definition: dnnl_types.h:1605
permuted 5D tensor
Definition: dnnl_types.h:195
Indicates the weights have an additional buffer, that depends on the compensation_mask.
Definition: dnnl_types.h:861
4D CNN weights tensor (incl. groups), an alias to dnnl_abcd
Definition: dnnl_types.h:393
Internal weights format for 2x3 Winograd.
Definition: dnnl_types.h:807
Undefined memory format tag.
Definition: dnnl_types.h:174
Memory descriptor.
Definition: dnnl_types.h:883
GRU cell.
Definition: dnnl_types.h:709
Convolution algorithm(either direct or Winograd) is chosen just in time.
Definition: dnnl_types.h:659
Weights format used in 8bit Winograd convolution.
Definition: dnnl_types.h:96
weights grad. memory desc
Definition: dnnl_types.h:1604
const struct dnnl_memory * const_dnnl_memory_t
A constant memory handle.
Definition: dnnl_types.h:940
int64_t dnnl_dim_t
A type to describe tensor dimension.
Definition: dnnl_types.h:777
Eltwise: ReLU.
Definition: dnnl_types.h:665
5D CNN weights tensor (incl. groups), an alias to dnnl_abcde
Definition: dnnl_types.h:395
A binary primitive.
Definition: dnnl_types.h:648
2D CNN weights tensor, an alias to dnnl_ab
Definition: dnnl_types.h:362
3D tensor blocked by 2nd dimension with block size 16
Definition: dnnl_types.h:212
dnnl_stream_flags_t
Stream flags.
Definition: dnnl_types.h:1617
A softmax primitive.
Definition: dnnl_types.h:632
3D RNN data tensor in the format (batch, seq_length, input channels).
Definition: dnnl_types.h:406
Default stream configuration.
Definition: dnnl_types.h:1626
struct dnnl_post_ops * dnnl_post_ops_t
A post operation chain handle.
Definition: dnnl_types.h:1419
4D CNN activations tensor blocked by channels with block size 16, an alias to dnnl_aBcd16b
Definition: dnnl_types.h:447
dnnl_format_tag_t
Memory format tag specification.
Definition: dnnl_types.h:169
struct dnnl_primitive_desc * dnnl_primitive_desc_t
A primitive descriptor handle.
Definition: dnnl_types.h:1366
The library manages scratchpad (default)
Definition: dnnl_types.h:1379
4D RNN bias tensor in the format (num_layers, num_directions, num_gates, output_channels).
Definition: dnnl_types.h:430
5D CNN weights tensor (incl. groups), an alias to dnnl_decab
Definition: dnnl_types.h:397
Eltwise: logistic.
Definition: dnnl_types.h:683
3D tensor blocked by 2nd dimension with block size 4
Definition: dnnl_types.h:216
3D CNN weights tensor, an alias to dnnl_bca
Definition: dnnl_types.h:372
Forward data propagation (alias for dnnl_forward_inference).
Definition: dnnl_types.h:599
A descriptor of a binary operation.
Definition: dnnl_types.h:1301
permuted 4D tensor
Definition: dnnl_types.h:198
6D tensor blocked by 2nd dimension with block size 16
Definition: dnnl_types.h:288
permuted 3D tensor
Definition: dnnl_types.h:199
plain 5D tensor
Definition: dnnl_types.h:186
const struct dnnl_primitive * const_dnnl_primitive_t
A constant primitive handle.
Definition: dnnl_types.h:1435
Bidirectional execution of RNN primitive with summation of the results.
Definition: dnnl_types.h:1229
dnnl_prop_kind_t
Kinds of propagation.
Definition: dnnl_types.h:587
Direct convolution.
Definition: dnnl_types.h:655
lrn descriptor
Definition: dnnl_types.h:1591
Forward data propagation (inference mode).
Definition: dnnl_types.h:597
A matrix multiplication primitive.
Definition: dnnl_types.h:646
dnnl_primitive_kind_t
Kinds of primitives.
Definition: dnnl_types.h:614
Unidirectional execution of RNN primitive from right to left.
Definition: dnnl_types.h:1223
Undefined propagation type.
Definition: dnnl_types.h:590
dnnl_scratchpad_mode_t
Scratchpad mode.
Definition: dnnl_types.h:1377
prop_kind
Propagation kind.
Definition: dnnl.hpp:275
Forward data propagation (alias for dnnl_forward_training).
Definition: dnnl_types.h:601
3D tensor blocked by 2nd dimension with block size 8
Definition: dnnl_types.h:222
6D CNN weights tensor (incl. groups), an alias to dnnl_abcdef
Definition: dnnl_types.h:401
A reorder primitive.
Definition: dnnl_types.h:618
layer normalization descriptor
Definition: dnnl_types.h:1593
softmax descriptor
Definition: dnnl_types.h:1589
5D CNN weights tensor, an alias to dnnl_bcdea
Definition: dnnl_types.h:390
Eltwise: square.
Definition: dnnl_types.h:671
permuted 2D tensor
Definition: dnnl_types.h:196
3D CNN activations tensor, an alias to dnnl_abc
Definition: dnnl_types.h:347
permuted 4D tensor
Definition: dnnl_types.h:194
pooling descriptor
Definition: dnnl_types.h:1590
struct dnnl_memory * dnnl_memory_t
A memory handle.
Definition: dnnl_types.h:937
4D CNN activations tensor, an alias to dnnl_acdb
Definition: dnnl_types.h:353
3D CNN activations tensor blocked by channels with block size 4, an alias to dnnl_aBc4b
Definition: dnnl_types.h:459
A user shall query and provide the scratchpad memory to primitives.
Definition: dnnl_types.h:1381
GEMM descriptor.
Definition: dnnl_types.h:1596
5D CNN activations tensor blocked by channels with block size 8, an alias to dnnl_aBcde8b
Definition: dnnl_types.h:444
Fuse with ReLU.
Definition: dnnl_types.h:763
Unspecified format kind.
Definition: dnnl_types.h:90
Eltwise: exponent.
Definition: dnnl_types.h:685
const struct dnnl_stream * const_dnnl_stream_t
A constant execution stream handle.
Definition: dnnl_types.h:1635
const typedef void * const_dnnl_op_desc_t
A pointer to any of the operation descriptors (constant variant).
Definition: dnnl_types.h:953
3D RNN data tensor in the format (seq_length, batch, input channels).
Definition: dnnl_types.h:404
eltwise descriptor
Definition: dnnl_types.h:1588
5D CNN activations tensor blocked by channels with block size 4, an alias to dnnl_aBcde4b
Definition: dnnl_types.h:441
non-standard 16-bit (bfloat16 w/ 7 bit mantissa) floating point.
Definition: dnnl_types.h:73
A tensor in a generic format described by the stride and blocking values in each dimension.
Definition: dnnl_types.h:94
1D tensor, an alias to dnnl_a
Definition: dnnl_types.h:337
Just a sentinel, not real memory format tag.
Definition: dnnl_types.h:332
5D CNN weights tensor, an alias to dnnl_cdeba
Definition: dnnl_types.h:386
The operation failed because requested functionality is not implemented.
Definition: dnnl_types.h:57
permuted 5D tensor
Definition: dnnl_types.h:205
LSTM cell.
Definition: dnnl_types.h:707
4D CNN activations tensor blocked by channels with block size 4, an alias to dnnl_aBcd4b
Definition: dnnl_types.h:450
A convolution primitive.
Definition: dnnl_types.h:626