Intel(R) Math Kernel Library for Deep Neural Networks (Intel(R) MKL-DNN)
0.16
Performance library for Deep Learning
|
enum mkldnn_alg_kind_t |
Kinds of algorithms.
Enumerator | |
---|---|
mkldnn_alg_kind_undef | |
mkldnn_convolution_direct | Direct convolution. |
mkldnn_convolution_winograd | Winograd convolution. |
mkldnn_eltwise_relu | Eltwise: ReLU. |
mkldnn_eltwise_tanh | Eltwise: hyperbolic tangent non-linearity (tanh) |
mkldnn_eltwise_elu | Eltwise: parametric exponential linear unit (elu) |
mkldnn_eltwise_square | Eltwise: square. |
mkldnn_eltwise_abs | Eltwise: abs. |
mkldnn_eltwise_sqrt | Eltwise: square root. |
mkldnn_eltwise_linear | Eltwise: linear. |
mkldnn_eltwise_bounded_relu | Eltwise: bounded_relu. |
mkldnn_eltwise_soft_relu | Eltwise: soft_relu. |
mkldnn_eltwise_logistic | Eltwise: logistic. |
mkldnn_pooling_max | Max pooling. |
mkldnn_pooling_avg_include_padding | Average pooling include padding. |
mkldnn_pooling_avg_exclude_padding | Average pooling exclude padding. |
mkldnn_pooling_avg | |
mkldnn_lrn_across_channels | Local response normalization (LRN) across multiple channels. |
mkldnn_lrn_within_channel | LRN within a single channel. |
mkldnn_deconvolution_direct | Direct deconvolution. |
mkldnn_deconvolution_winograd | Winograd deconvolution. |
mkldnn_vanilla_rnn | RNN cell. |
mkldnn_vanilla_lstm | LSTM cell. |
mkldnn_vanilla_gru | GRU cell. |
mkldnn_gru_linear_before_reset | GRU cell with linear before reset. Modification of original GRU cell. Differs from mkldnn_vanilla_gru in how the new memory gate is calculated:
Primitive expects 4 biases on input: |
Flags for batch-normalization primititve.
Enumerator | |
---|---|
mkldnn_use_global_stats | Use global statistics. If specified
If not specified:
|
mkldnn_use_scaleshift | Use scale and shift parameters. If specified:
If no specified:
|
mkldnn_omit_stats | Omit statistics.
For time being had an affect on backward propagation only which allowed skipping some computations (the same semantics as mkldnn_use_global_stats) |
mkldnn_fuse_bn_relu | Fuse with ReLU. If specified:
|
enum mkldnn_data_type_t |
Memory format specification.
Intel(R) MKL-DNN uses the following notation for memory format names:
'n'
denotes the mini-batch dimension'c'
denotes a channels dimension'i'
and 'o'
denote dimensions of input and output channels'h'
and 'w'
denote spatial width and height'mkldnn_nChw8c'
describes a format where the outermost dimension is mini-batch, followed by the channel block number, followed by the spatial height and width, and finally followed by 8-element channel blocks.'mkldnn_nc'
and 'mkldnn_io'
formats can be used to describe a 2D tensor. Enumerator | |
---|---|
mkldnn_format_undef | Undefined memory format, used for empty memory descriptors. |
mkldnn_any | Unspecified format. The primitive selects a format automatically. |
mkldnn_blocked | A tensor in a generic format described by the stride and blocking values in each dimension. See mkldnn_blocking_desc_t for more information. |
mkldnn_x | 1D data tensor. |
mkldnn_nc | 2D data tensor. |
mkldnn_nchw | 4D data tensor in the |
mkldnn_nhwc | 4D data tensor in the |
mkldnn_chwn | 4D data tensor in the |
mkldnn_nChw8c | 4D data tensor in the |
mkldnn_nChw16c | 4D data tensor in the |
mkldnn_ncdhw | 5D data tensor in the |
mkldnn_ndhwc | 5D data tensor in the |
mkldnn_nCdhw16c | 5D data tensor in the |
mkldnn_oi | 2D weights tensor in the format (input channels, output channels). |
mkldnn_io | 2D weights tensor in the format (input channels, output channels). |
mkldnn_oihw | 4D weights tensor in the format (input channels, output channels, width, height). |
mkldnn_ihwo | 4D weights tensor in the format (input channels, height, width, output channels). |
mkldnn_hwio | 4D weights tensor in the format (height, width, input channels, output channels). |
mkldnn_dhwio | 5D weights tensor in the format (depth, height, width, input channels, output channels). |
mkldnn_oidhw | 5D weight tensor in the |
mkldnn_OIdhw16i16o | 6D weights tensor in the |
mkldnn_OIdhw16o16i | 6D weights tensor in the |
mkldnn_Oidhw16o | 5D weights tensor in the blocked version of |
mkldnn_Odhwi16o | 5D weights tensor in the blocked version of |
mkldnn_OIhw8i8o | 4D weights tensor in the |
mkldnn_OIhw16i16o | 4D weights tensor in the |
mkldnn_OIhw4i16o4i | 4D weights tensor in the |
mkldnn_OIhw8i16o2i | 4D weights tensor in the |
mkldnn_OIdhw8i16o2i | 5D weights tensor in the |
mkldnn_OIhw8o16i2o | 4D weights tensor in the |
mkldnn_OIhw8o8i | 4D weights tensor in the |
mkldnn_OIhw16o16i | 4D weights tensor in the |
mkldnn_IOhw16o16i | 4D weights tensor in the |
mkldnn_Oihw8o | 4D weights tensor in the format (output channels, input channels, height, width) with output channels data laid out in memory in 8-element blocks. |
mkldnn_Oihw16o | 4D weights tensor in the format (output channels, input channels, height, width) with output channels data laid out in memory in 16-element blocks. |
mkldnn_Ohwi8o | 4D weights tensor in the format (output channels, width, height, input channels) with output channels data laid out in memory in 8-element blocks. |
mkldnn_Ohwi16o | 4D weights tensor in the format (output channels, width, height, input channels) with output channels data laid out in memory in 16-element blocks. |
mkldnn_OhIw16o4i | 4D weights tensor in the |
mkldnn_goihw | 5D weights tensor in the |
mkldnn_hwigo | 5D weights tensor in the |
mkldnn_gOIhw8i8o | 5D weights tensor in the blocked version of |
mkldnn_gOIhw16i16o | 5D weights tensor in the blocked version of |
mkldnn_gOIhw4i16o4i | 5D weights tensor in the |
mkldnn_gOIhw8i16o2i | 5D weights tensor in the |
mkldnn_gOIdhw8i16o2i | 6D weights tensor in the |
mkldnn_gOIhw8o16i2o | 5D weights tensor in the |
mkldnn_gOIhw8o8i | 5D weights tensor in the blocked version of |
mkldnn_gOIhw16o16i | 5D weights tensor in the blocked version of |
mkldnn_gIOhw16o16i | 5D weights tensor in the blocked version of |
mkldnn_gOihw8o | 5D weights tensor in the blocked version of |
mkldnn_gOihw16o | 5D weights tensor in the blocked version of |
mkldnn_gOhwi8o | 5D weights tensor in the blocked version of |
mkldnn_gOhwi16o | 5D weights tensor in the blocked version of |
mkldnn_Goihw8g | 5D weights tensor in the blocked version of |
mkldnn_Goihw16g | 5D weights tensor in the blocked version of |
mkldnn_gOhIw16o4i | 5D weights tensor in the |
mkldnn_goidhw | 6D weight tensor in the |
mkldnn_gOIdhw16i16o | 6D weights tensor in the |
mkldnn_gOIdhw16o16i | 6D weights tensor in the blocked version of |
mkldnn_gOidhw16o | 6D weights tensor in the blocked version of |
mkldnn_gOdhwi16o | 6D weights tensor in the blocked version of |
mkldnn_ntc | 3D data tensor in the format (batch, seq_length, input channels). |
mkldnn_tnc | 3D data tensor in the format (seq_length, batch, input channels). |
mkldnn_ldsnc | 5D states tensor in the format (num_layers, num_directions, num_states, batch, state channels). |
mkldnn_ldigo | 5D weights tensor in the format (num_layers, num_directions, input_chanels, num_gates, output_channels). |
mkldnn_ldigo_p | 5D weights tensor in the blocked format. |
mkldnn_ldgoi | 5D weights tensor in the format (num_layers, num_directions, num_gates, output_channels, input_chanels). |
mkldnn_ldgoi_p | 5D weights tensor in the blocked format. |
mkldnn_ldgo | 4D bias tensor in the format (num_layers, num_directions, num_gates, output_channels). |
mkldnn_wino_fmt | General tensor format for integer 8bit winograd convolution. |
mkldnn_format_last | Just a sentinel, not real memory format. Must be changed after new format is added. |
mkldnn_oIhw8i | 4D weights tensor in the oihw format with input channels data laid out in memory in 8-element blocks. |
mkldnn_oIhw16i | 4D weights tensor in the oihw format with input channels data laid out in memory in 16-element blocks. |
Kinds of primitives.
Used to implement a way to extend the library with new primitives without changing the ABI.
enum mkldnn_prop_kind_t |
Kinds of propagation.
enum mkldnn_round_mode_t |
enum mkldnn_status_t |
Status values returned by Intel(R) MKL-DNN functions.