RNN#

The RNN primitive computes a stack of unrolled recurrent cells, as depicted in Figure 1. \(\bias\), \(\srciter\) and \(\dstiter\) are optional parameters. If not provided, \(\bias\) and \(\srciter\) default to 0. Variable names follow the standard Conventions.

../../../../_images/unrolled_stack_rnn.jpg

The RNN primitive supports four modes for evaluation direction:

left2right will process the input data timestamps by increasing order,
right2left will process the input data timestamps by decreasing order,
bidirectional_concat will process all the stacked layers from left2right and from right2left independently, and will concatenate the output in \(\dstlayer\) over the channel dimension,
bidirectional_sum will process all the stacked layers from left2right and from right2left independently, and will sum the two outputs to \(\dstlayer\).

Even though the RNN primitive supports passing a different number of channels for \(\srclayer\), \(\srciter\), \(\dstlayer\), and \(\dstiter\), we always require the following conditions in order for the dimension to be consistent:

\(channels(\dstlayer) = channels(\dstiter)\),
when \(T > 1\), \(channels(\srciter) = channels(\dstiter)\),
when \(L > 1\), \(channels(\srclayer) = channels(\dstlayer)\),
when using the bidirectional_concat direction, \(channels(\dstlayer) = 2 * channels(\dstiter)\).

The general formula for the execution of a stack of unrolled recurrent cells depends on the current iteration of the previous layer (\(h_{t,l-1}\) and \(c_{t,l-1}\)) and the previous iteration of the current layer (\(h_{t-1, l}\)). Here is the exact equation for non-LSTM cells:

\[h_{t, l} = Cell(h_{t, l-1}, h_{t-1, l})\]

where \(t\), \(l\) are the indices of the timestamp and the layer of the cell being executed.

And here is the equation for LSTM cells:

\[(h_{t, l},c_{t,l}) = Cell(h_{t, l-1}, h_{t-1, l}, c_{t-1,l})\]

where \(t\), \(l\) are the indices of the timestamp and the layer of the cell being executed.

Cell Functions#

The RNN API provides six cell functions:

Vanilla RNN, a single-gate recurrent cell,
LSTM, a four-gate long short-term memory cell,
GRU, a three-gate gated recurrent unit cell,
Linear-before-reset GRU, a three-gate recurrent unit cell with the linear layer before the reset gate.
AUGRU, a three-gate gated recurrent unit cell with the attention update gate,
Linear-before-reset AUGRU, a three-gate recurrent unit cell with the linear layer before the reset gate and the attention update gate.

Vanilla RNN#

A single-gate recurrent cell initialized with dnnl::vanilla_rnn_forward::primitive_desc or dnnl::vanilla_rnn_forward::primitive_desc as in the following example.

auto vanilla_rnn_pd =
    dnnl::vanilla_rnn_forward::primitive_desc(engine, aprop,
    activation, direction, src_layer_desc, src_iter_desc,
    weights_layer_desc, weights_iter_desc, bias_desc,
    dst_layer_desc, dst_iter_desc, attr);

The Vanilla RNN cell should support the ReLU, Tanh and Sigmoid activation functions. The following equations defines the mathematical operation performed by the Vanilla RNN cell for the forward pass:

\[\begin{split}a_t &= W \cdot h_{t,l-1} + U \cdot h_{t-1, l} + B \\ h_t &= activation(a_t)\end{split}\]

LSTM#

LSTM (or Vanilla LSTM)#

A four-gate long short-term memory recurrent cell initialized with dnnl::lstm_forward::primitive_desc or dnnl::lstm_backward::primitive_desc as in the following example.

auto lstm_pd = dnnl::lstm_forward::primitive_desc(engine, aprop,
    direction, src_layer_desc, src_iter_h_desc, src_iter_c_desc,
    weights_layer_desc, weights_iter_desc, bias_desc,
    dst_layer_desc, dst_iter_h_desc, dst_iter_c_desc, attr);

Note that for all tensors with a dimension depending on the gates number, we implicitly require the order of these gates to be \(i\), \(f\), \(\tilde c\), and \(o\). The following equation gives the mathematical description of these gates and output for the forward pass:

\[\begin{split}i_t &= \sigma(W_i \cdot h_{t,l-1} + U_i \cdot h_{t-1, l} + B_i) \\ f_t &= \sigma(W_f \cdot h_{t,l-1} + U_f \cdot h_{t-1, l} + B_f) \\ \\ \tilde c_t &= tanh(W_{\tilde c} \cdot h_{t,l-1} + U_{\tilde c} \cdot h_{t-1, l} + B_{\tilde c}) \\ c_t &= f_t * c_{t-1} + i_t * \tilde c_t \\ \\ o_t &= \sigma(W_o \cdot h_{t,l-1} + U_o \cdot h_{t-1, l} + B_o) \\ h_t &= tanh(c_t) * o_t\end{split}\]

where \(W_*\) are stored in \(\weightslayer\), \(U_*\) are stored in \(\weightsiter\) and \(B_*\) are stored in \(\bias\).

Note

In order for the dimensions to be consistent, we require \(channels(\srciterc) = channels(\dstiterc) = channels(\dstiter)\).

LSTM with Peephole#

A four-gate long short-term memory recurrent cell with peephole initialized with dnnl::lstm_forward::primitive_desc or dnnl::lstm_backward::primitive_desc as in the following example.

auto lstm_pd = dnnl::lstm_forward::primitive_desc(engine, aprop,
    direction, src_layer_desc, src_iter_h_desc, src_iter_c_desc,
    weights_layer_desc, weights_iter_desc, weights_peephole_desc,
    bias_desc, dst_layer_desc, dst_iter_h_desc, dst_iter_c_desc,
    attr);

Similarly to vanilla LSTM, we implicitly require the order of these gates to be \(i\), \(f\), \(\tilde c\), and \(o\). For peephole weights, the gates order is:math:i, \(f\), \(o\). The following equation gives the mathematical description of these gates and output for the forward pass:

\[\begin{split}i_t &= \sigma(W_i \cdot h_{t,l-1} + U_i \cdot h_{t-1, l} + P_i \cdot c_{t-1} + B_i) \\ f_t &= \sigma(W_f \cdot h_{t,l-1} + U_f \cdot h_{t-1, l} + P_f \cdot c_{t-1} + B_f) \\ \\ \tilde c_t &= tanh(W_{\tilde c} \cdot h_{t,l-1} + U_{\tilde c} \cdot h_{t-1, l} + B_{\tilde c}) \\ c_t &= f_t * c_{t-1} + i_t * \tilde c_t \\ \\ o_t &= \sigma(W_o \cdot h_{t,l-1} + U_o \cdot h_{t-1, l} + P_o \cdot c_t + B_o) \\ h_t &= tanh(c_t) * o_t\end{split}\]

where \(P_*\) are stored in weights_peephole, and the other parameters are the same as in vanilla LSTM.

Note

If the weights_peephole_desc passed to the primitive descriptor constructor is a zero memory descriptor, the primitive will behave the same as in LSTM primitive without peephole.

LSTM with Projection#

A four-gate long short-term memory recurrent cell with projection initialized with dnnl::lstm_forward::primitive_desc or dnnl::lstm_backward::primitive_desc as in the following example.

auto lstm_pd = dnnl::lstm_forward::primitive_desc(engine, aprop,
    direction, src_layer_desc, src_iter_h_desc, src_iter_c_desc,
    weights_layer_desc, weights_iter_desc, weights_peephole_desc,
    weights_projection_desc, bias_desc, dst_layer_desc,
    dst_iter_h_desc, dst_iter_c_desc, attr);

Similarly to vanilla LSTM, we implicitly require the order of the gates to be i, \(f\), \(\tilde c\), and \(o\) for all tensors with a dimension depending on the gates. The following equation gives the mathematical description of these gates and output for the forward pass (for simplicity, LSTM without peephole is shown):

\[\begin{split}i_t &= \sigma(W_i \cdot h_{t,l-1} + U_i \cdot h_{t-1,l} + B_i) \\ f_t &= \sigma(W_f \cdot h_{t,l-1} + U_f \cdot h_{t-1,l} + B_f) \\ & \\ \tilde{c}_t &= \tanh(W_{\tilde{c}} \cdot h_{t,l-1} + U_{\tilde{c}} \cdot h_{t-1,l} + B_{\tilde{c}}) \\ c_t &= f_t * c_{t-1} + i_t * \tilde{c}_t \\ & \\ o_t &= \sigma(W_o \cdot h_{t,l-1} + U_o \cdot h_{t-1,l} + B_o) \\ h_t &= R \cdot (\tanh(c_t) * o_t)\end{split}\]

where \(R\) is stored in weights_projection, and the other parameters are the same as in vanilla LSTM.

Note

If the weights_projection_desc passed to the primitive descriptor constructor is a zero memory descriptor, the primitive will behave the same as in LSTM primitive without projection.

GRU#

A three-gate gated recurrent unit cell, initialized with dnnl::gru_forward::primitive_desc or dnnl::gru_backward::primitive_desc as in the following example.

auto gru_pd = dnnl::gru_forward::primitive_desc(engine, aprop,
    direction, src_layer_desc, src_iter_desc, weights_layer_desc,
    weights_iter_desc, bias_desc, dst_layer_desc, dst_iter_desc,
    attr);

Note that for all tensors with a dimension depending on the gates number, we implicitly require the order of these gates to be:math:u, \(r\), and \(o\). The following equation gives the mathematical definition of these gates.

\[\begin{split}u_t &= \sigma(W_u \cdot h_{t,l-1} + U_u \cdot h_{t-1, l} + B_u) \\ r_t &= \sigma(W_r \cdot h_{t,l-1} + U_r \cdot h_{t-1, l} + B_r) \\ o_t &= tanh(W_o \cdot h_{t,l-1} + U_o \cdot (r_t * h_{t-1, l}) + B_o) \\ h_t &= u_t * h_{t-1, l} + (1 - u_t) * o_t\end{split}\]

where \(W_*\) are in \(\weightslayer\), \(U_*\) are in \(\weightsiter\), and \(B_*\) are stored in \(\bias\).

Note

If you need to replace \(u_t\) by \((1-u_t)\) when computing \(h_t\), you can achieve this by multiplying \(W_u\), \(U_u\) and \(B_u\) by \(-1\). This is possible as \(u_t = \sigma(W_u \cdot h_{t,l-1} + U_u \cdot h_{t-1, l} + B_u)\), and \(1 – \sigma(a) = \sigma(-a)\).

Linear-Before-Reset GRU#

A three-gate gated recurrent unit cell with linear layer applied before the reset gate, initialized with dnnl::lbr_gru_forward::primitive_desc or dnnl::lbr_gru_backward::primitive_desc as in the following example.

auto lbr_gru_pd = dnnl::lbr_gru_forward::primitive_desc(engine,
    aprop, direction, src_layer_desc, src_iter_desc,
    weights_layer_desc, weights_iter_desc, bias_desc,
    dst_layer_desc, dst_iter_desc, attr);

The following equation describes the mathematical behavior of the Linear-Before-Reset GRU cell.

\[\begin{split}u_t &= \sigma(W_u \cdot h_{t,l-1} + U_u \cdot h_{t-1, l} + B_u) \\ r_t &= \sigma(W_r \cdot h_{t,l-1} + U_r \cdot h_{t-1, l} + B_r) \\ o_t &= tanh(W_o \cdot h_{t,l-1} + r_t * (U_o \cdot h_{t-1, l} + B_{u'}) + B_o) \\ h_t &= u_t * h_{t-1, l} + (1 - u_t) * o_t\end{split}\]

Note that for all tensors with a dimension depending on the gates number, except the bias, we implicitly require the order of these gates to be \(u\), \(r\), and \(o\). For the \(\bias\) tensor, we implicitly require the order of the gates to be \(u\), \(r\), \(o\), and \(u'\).

Note

AUGRU#

A three-gate gated recurrent unit cell, initialized with dnnl::augru_forward::primitive_desc or dnnl::augru_backward::primitive_desc as in the following example.

auto augru_pd = dnnl::augru_forward::primitive_desc(engine, aprop,
    direction, src_layer_desc, src_iter_desc, attention_desc,
    weights_layer_desc, weights_iter_desc, bias_desc,
    dst_layer_desc, dst_iter_desc, attr);

Note that for all tensors with a dimension depending on the gate number, we implicitly require the order of these gates to be \(u\), \(r\), and \(o\). The following equation gives the mathematical definition of these gates.

\[\begin{split}u_t &= \sigma(W_u \cdot h_{t,l-1} + U_u \cdot h_{t-1, l} + B_u) \\ r_t &= \sigma(W_r \cdot h_{t,l-1} + U_r \cdot h_{t-1, l} + B_r) \\ o_t &= \tanh(W_o \cdot h_{t,l-1} + U_o \cdot (r_t * h_{t-1, l}) + B_o) \\ \tilde u_t &= (1 - a_t) * u_t \\ h_t &= \tilde u_t * h_{t-1, l} + (1 - \tilde u_t) * o_t\end{split}\]

where \(W_*\) are in weightslayer, \(U_*\) are in weightsiter, and \(B_*\) are stored in bias.

Linear-Before-Reset AUGRU#

A three-gate gated recurrent unit cell with linear layer applied before the reset gate, initialized with dnnl::lbr_augru_forward::primitive_desc or dnnl::lbr_augru_backward::primitive_desc as in the following example.

auto lbr_augru_pd =
    dnnl::lbr_augru_forward::primitive_desc(engine, aprop,
    direction, src_layer_desc, src_iter_desc, attention_desc,
    weights_layer_desc, weights_iter_desc, bias_desc,
    dst_layer_desc, dst_iter_desc, attr);

The following equation describes the mathematical behavior of the Linear-Before-Reset AUGRU cell.

\[\begin{split}u_t &= \sigma(W_u \cdot h_{t,l-1} + U_u \cdot h_{t-1, l} + B_u) \\ r_t &= \sigma(W_r \cdot h_{t,l-1} + U_r \cdot h_{t-1, l} + B_r) \\ o_t &= \tanh(W_o \cdot h_{t,l-1} + r_t *(U_o \cdot h_{t-1, l} + B_{u'}) + B_o) \\ \tilde u_t &= (1 - a_t) * u_t \\ h_t &= \tilde u_t * h_{t-1, l} + (1 - \tilde u_t) * o_t\end{split}\]

Note that for all tensors with a dimension depending on the gate number, except the bias, we implicitly require the order of these gates to be \(u\), \(r\), and \(o\). For the bias tensor, we implicitly require the order of the gates to be \(u\), \(r\), \(o\), and \(u'\).

Execution Arguments#

When executed, the inputs and outputs should be mapped to an execution argument index as specified by the following table.

Primitive input/output	Execution argument index
\(\srclayer\)	`DNNL_ARG_SRC_LAYER`
\(\srciter\)	`DNNL_ARG_SRC_ITER`
\(\srciterc\)	`DNNL_ARG_SRC_ITER_C`
\(\weightslayer\)	`DNNL_ARG_WEIGHTS_LAYER`
\(\weightsiter\)	`DNNL_ARG_WEIGHTS_ITER`
\(\weightspeephole\)	`DNNL_ARG_WEIGHTS_PEEPHOLE`
\(\weightsprojection\)	`DNNL_ARG_WEIGHTS_PROJECTION`
\(\bias\)	`DNNL_ARG_BIAS`
\(\dstlayer\)	`DNNL_ARG_DST_LAYER`
\(\dstiter\)	`DNNL_ARG_DST_ITER`
\(\dstiterc\)	`DNNL_ARG_DST_ITER_C`
\(\workspace\)	`DNNL_ARG_WORKSPACE`
\(\diffsrclayer\)	`DNNL_ARG_DIFF_SRC_LAYER`
\(\diffsrciter\)	`DNNL_ARG_DIFF_SRC_ITER`
\(\diffsrciterc\)	`DNNL_ARG_DIFF_SRC_ITER_C`
\(\diffweightslayer\)	`DNNL_ARG_DIFF_WEIGHTS_LAYER`
\(\diffweightsiter\)	`DNNL_ARG_DIFF_WEIGHTS_ITER`
\(\diffweightspeephole\)	`DNNL_ARG_DIFF_WEIGHTS_PEEPHOLE`
\(\diffweightsprojection\)	`DNNL_ARG_DIFF_WEIGHTS_PROJECTION`
\(\diffbias\)	`DNNL_ARG_DIFF_BIAS`
\(\diffdstlayer\)	`DNNL_ARG_DIFF_DST_LAYER`
\(\diffdstiter\)	`DNNL_ARG_DIFF_DST_ITER`
\(\diffdstiterc\)	`DNNL_ARG_DIFF_DST_ITER_C`

Operation Details#

N/A

Data Types Support#

The following table lists the combination of data types that should be supported by the RNN primitive for each input and output memory object.

Note

Here we abbreviate data types names for readability. For example, dnnl::memory::data_type::f32 is abbreviated to f32.

Propagation	Cell Function	Input Data	Recurrent Data (1)	Weights	Bias	Output Data
Forward / Backward	All	`f32`	`f32`	`f32`	`f32`	`f32`
Forward / Backward (2)	All (3)	`bf16`	`bf16`	`bf16`	`f32`	`bf16`
Forward	All (3)	`f16`	`f16`	`f16`	`f16`	`f16`
Forward inference	Vanilla LSTM	`u8`	`u8`	`s8`	`f32`	`u8`, `f32`

With LSTM and Peephole LSTM cells, the cell state data type is always f32.
In backward propagation, all diff_* tensors are in f32.
Projection LSTM is not defined yet.

Data Representation#

In the oneDNN programming model, the RNN primitive is one of a few that support the placeholder memory format #dnnl::memory::format_tag::any (shortened to any from now on) and can define data and weight memory objects format based on the primitive parameters.

The following table summarizes the data layouts supported by the RNN primitive.

Input/Output Data	Recurrent Data	Layer and Iteration Weights	Peephole Weights and Bias	Projection LSTM Weights
`any`	`any`	`any`	`ldgo`	`any`, `ldio` (Forward propagation)
`ntc`, `tnc`	`ldnc`	`ldigo`, `ldgoi`	`ldgo`	`any`, `ldio` (Forward propagation)

While an RNN primitive can be created with memory formats specified explicitly, the performance is likely to be sub-optimal. When using any it is necessary to first create an RNN primitive descriptor and then query it for the actual data and weight memory objects formats.

Note

The RNN primitive should support padded tensors and views. So even if two memory descriptors share the same data layout, they might still be different.

Post-ops and Attributes#

Currently post-ops and attributes are only used by the int8 variant of LSTM.

API#

enum class dnnl::rnn_flags : unsigned#

RNN cell flags.

Values:

enumerator undef#: Undefined RNN flags.

enum class dnnl::rnn_direction#

A direction of RNN primitive execution.

Values:

enumerator undef#: Undefined RNN direction.

enumerator unidirectional_left2right#: Unidirectional execution of RNN primitive from left to right.

enumerator unidirectional_right2left#: Unidirectional execution of RNN primitive from right to left.

enumerator bidirectional_concat#: Bidirectional execution of RNN primitive with concatenation of the results.

enumerator bidirectional_sum#: Bidirectional execution of RNN primitive with summation of the results.

enumerator unidirectional#: Alias for dnnl::rnn_direction::unidirectional_left2right.

struct dnnl::vanilla_rnn_forward : public dnnl::primitive #

Vanilla RNN forward propagation primitive.

Public Functions

vanilla_rnn_forward()#: Default constructor. Produces an empty object.

vanilla_rnn_forward(const primitive_desc &pd)#

Constructs a vanilla RNN forward propagation primitive.

Parameters: pd – Primitive descriptor for a vanilla RNN forward propagation primitive.

struct primitive_desc : public dnnl::rnn_primitive_desc_base #

Primitive descriptor for a vanilla RNN forward propagation primitive.

Public Functions

primitive_desc() = default#: Default constructor. Produces an empty object.

Constructs a primitive descriptor for a vanilla RNN forward propagation primitive.

The following arguments may point to a zero memory descriptor:

src_iter_desc,
bias_desc,
dst_iter_desc.

This would then indicate that the RNN forward propagation primitive should not use them and should default to zero values instead.

Note

All memory descriptors except src_iter_desc can be initialized with an dnnl::memory::format_tag::any value of format_tag.

Parameters

aengine – Engine to use.
aprop_kind – Propagation kind. Possible values are dnnl::prop_kind::forward_training, and dnnl::prop_kind::forward_inference.
activation – Activation kind. Possible values are dnnl::algorithm::eltwise_relu, dnnl::algorithm::eltwise_tanh, or dnnl::algorithm::eltwise_logistic.
direction – RNN direction. See dnnl::rnn_direction for more info.
src_layer_desc – Memory descriptor for the input vector.
src_iter_desc – Memory descriptor for the input recurrent hidden state vector.
weights_layer_desc – Memory descriptor for the weights applied to the layer input.
weights_iter_desc – Memory descriptor for the weights applied to the recurrent input.
bias_desc – Bias memory descriptor.
dst_layer_desc – Memory descriptor for the output vector.
dst_iter_desc – Memory descriptor for the output recurrent hidden state vector.
attr – Primitive attributes to use. Attributes are optional and default to empty attributes.
allow_empty – A flag signifying whether construction is allowed to fail without throwing an exception. In this case an empty object will be produced. This flag is optional and defaults to false.

Constructs a primitive descriptor for a vanilla RNN forward propagation primitive with alpha parameter.

The following arguments may point to a zero memory descriptor:

src_iter_desc,
bias_desc,
dst_iter_desc.

This would then indicate that the RNN forward propagation primitive should not use them and should default to zero values instead.

Note

All memory descriptors except src_iter_desc can be initialized with an dnnl::memory::format_tag::any value of format_tag.

Parameters

aengine – Engine to use.
aprop_kind – Propagation kind. Possible values are dnnl::prop_kind::forward_training, and dnnl::prop_kind::forward_inference.
activation – Activation kind. Possible values are dnnl::algorithm::eltwise_relu, dnnl::algorithm::eltwise_tanh, or dnnl::algorithm::eltwise_logistic.
direction – RNN direction. See dnnl::rnn_direction for more info.
src_layer_desc – Memory descriptor for the input vector.
src_iter_desc – Memory descriptor for the input recurrent hidden state vector.
weights_layer_desc – Memory descriptor for the weights applied to the layer input.
weights_iter_desc – Memory descriptor for the weights applied to the recurrent input.
bias_desc – Bias memory descriptor.
dst_layer_desc – Memory descriptor for the output vector.
dst_iter_desc – Memory descriptor for the output recurrent hidden state vector.
alpha – Negative slope if activation is dnnl::algorithm::eltwise_relu.
attr – Primitive attributes to use. Attributes are optional and default to empty attributes.
allow_empty – A flag signifying whether construction is allowed to fail without throwing an exception. In this case an empty object will be produced. This flag is optional and defaults to false.

memory::desc src_layer_desc() const#

Returns source layer memory descriptor.

Returns: Source layer memory descriptor.

memory::desc src_iter_desc() const#

Returns source iteration memory descriptor.

Returns: Source iteration memory descriptor.
Returns: A zero memory descriptor if the primitive does not have a source iteration parameter.

memory::desc weights_layer_desc() const#

Returns weights layer memory descriptor.

Returns: Weights layer memory descriptor.

memory::desc weights_iter_desc() const#

Returns weights iteration memory descriptor.

Returns: Weights iteration memory descriptor.

memory::desc bias_desc() const#

Returns bias memory descriptor.

Returns: Bias memory descriptor.
Returns: A zero memory descriptor if the primitive does not have a bias parameter.

memory::desc dst_layer_desc() const#

Returns destination layer memory descriptor.

Returns: Destination layer memory descriptor.

memory::desc dst_iter_desc() const#

Returns destination iteration memory descriptor.

Returns: Destination iteration memory descriptor.
Returns: A zero memory descriptor if the primitive does not have a destination iteration parameter.

memory::desc workspace_desc() const#

Returns the workspace memory descriptor.

Returns: Workspace memory descriptor.
Returns: A zero memory descriptor if the primitive does not require workspace parameter.

algorithm get_cell_kind() const#

Returns an RNN cell kind parameter.

Returns: An RNN cell kind parameter.
Returns: dnnl::algorithm::undef if the primitive does not have an RNN cell kind parameter.

prop_kind get_prop_kind() const#

Returns a propagation kind.

Returns: A propagation kind.
Returns: dnnl::prop_kind::undef if the primitive does not have a propagation parameter.

algorithm get_activation_kind() const#

Returns an RNN activation kind parameter.

Returns: An RNN activation kind parameter.
Returns: dnnl::algorithm::undef if the primitive does not have an RNN activation kind parameter.

rnn_direction get_direction() const#

Returns an RNN direction parameter.

Returns: An RNN direction parameter.
Returns: dnnl::rnn_direction::undef if the primitive does not have an RNN direction parameter.

float get_alpha() const#

Returns an alpha.

Returns: An alpha.
Returns: Zero if the primitive does not have an alpha parameter.

float get_beta() const#

Returns a beta.

Returns: A beta.
Returns: Zero if the primitive does not have a beta parameter.

struct dnnl::vanilla_rnn_backward : public dnnl::primitive #

Vanilla RNN backward propagation primitive.

Public Functions

vanilla_rnn_backward()#: Default constructor. Produces an empty object.

vanilla_rnn_backward(const primitive_desc &pd)#

Constructs a vanilla RNN backward propagation primitive.

Parameters: pd – Primitive descriptor for a vanilla RNN backward propagation primitive.

struct primitive_desc : public dnnl::rnn_primitive_desc_base #

Primitive descriptor for an RNN backward propagation primitive.

Public Functions

primitive_desc() = default#: Default constructor. Produces an empty object.

primitive_desc(const engine &aengine, prop_kind aprop_kind, algorithm activation, rnn_direction direction, const memory::desc &src_layer_desc, const memory::desc &src_iter_desc, const memory::desc &weights_layer_desc, const memory::desc &weights_iter_desc, const memory::desc &bias_desc, const memory::desc &dst_layer_desc, const memory::desc &dst_iter_desc, const memory::desc &diff_src_layer_desc, const memory::desc &diff_src_iter_desc, const memory::desc &diff_weights_layer_desc, const memory::desc &diff_weights_iter_desc, const memory::desc &diff_bias_desc, const memory::desc &diff_dst_layer_desc, const memory::desc &diff_dst_iter_desc, const vanilla_rnn_forward::primitive_desc &hint_fwd_pd, const primitive_attr &attr = default_attr(), bool allow_empty = false)#

Constructs a primitive descriptor for a vanilla RNN backward propagation primitive.

The following arguments may point to a zero memory descriptor:

src_iter_desc together with diff_src_iter_desc,
bias_desc together with diff_bias_desc,
dst_iter_desc together with diff_dst_iter_desc.

This would then indicate that the RNN backward propagation primitive should not use the respective data and should use zero values instead.

Note

All the memory descriptors may be initialized with the dnnl::memory::format_tag::any value of format_tag.

Parameters

aengine – Engine to use.
aprop_kind – Propagation kind. Must be dnnl::prop_kind::backward.
activation – Activation kind. Possible values are dnnl::algorithm::eltwise_relu, dnnl::algorithm::eltwise_tanh, or dnnl::algorithm::eltwise_logistic.
direction – RNN direction. See dnnl::rnn_direction for more info.
src_layer_desc – Memory descriptor for the input vector.
src_iter_desc – Memory descriptor for the input recurrent hidden state vector.
weights_layer_desc – Memory descriptor for the weights applied to the layer input.
weights_iter_desc – Memory descriptor for the weights applied to the recurrent input.
bias_desc – Bias memory descriptor.
dst_layer_desc – Memory descriptor for the output vector.
dst_iter_desc – Memory descriptor for the output recurrent hidden state vector.
diff_src_layer_desc – Memory descriptor for the diff of input vector.
diff_src_iter_desc – Memory descriptor for the diff of input recurrent hidden state vector.
diff_weights_layer_desc – Memory descriptor for the diff of weights applied to the layer input.
diff_weights_iter_desc – Memory descriptor for the diff of weights applied to the recurrent input.
diff_bias_desc – Diff bias memory descriptor.
diff_dst_layer_desc – Memory descriptor for the diff of output vector.
diff_dst_iter_desc – Memory descriptor for the diff of output recurrent hidden state vector.
hint_fwd_pd – Primitive descriptor for a vanilla RNN forward propagation primitive. It is used as a hint for deciding which memory format to use.
attr – Primitive attributes to use. Attributes are optional and default to empty attributes.
allow_empty – A flag signifying whether construction is allowed to fail without throwing an exception. In this case an empty object will be produced. This flag is optional and defaults to false.

primitive_desc(const engine &aengine, prop_kind aprop_kind, algorithm activation, rnn_direction direction, const memory::desc &src_layer_desc, const memory::desc &src_iter_desc, const memory::desc &weights_layer_desc, const memory::desc &weights_iter_desc, const memory::desc &bias_desc, const memory::desc &dst_layer_desc, const memory::desc &dst_iter_desc, const memory::desc &diff_src_layer_desc, const memory::desc &diff_src_iter_desc, const memory::desc &diff_weights_layer_desc, const memory::desc &diff_weights_iter_desc, const memory::desc &diff_bias_desc, const memory::desc &diff_dst_layer_desc, const memory::desc &diff_dst_iter_desc, float alpha, const vanilla_rnn_forward::primitive_desc &hint_fwd_pd, const primitive_attr &attr = default_attr(), bool allow_empty = false)#

Constructs a primitive descriptor for a vanilla RNN backward propagation primitive with an alpha parameter.

The following arguments may point to a zero memory descriptor:

src_iter_desc together with diff_src_iter_desc,
bias_desc together with diff_bias_desc,
dst_iter_desc together with diff_dst_iter_desc.

This would then indicate that the RNN backward propagation primitive should not use the respective data and should use zero values instead.

Note

All the memory descriptors may be initialized with the dnnl::memory::format_tag::any value of format_tag.

Parameters

aengine – Engine to use.
aprop_kind – Propagation kind. Must be dnnl::prop_kind::backward.
activation – Activation kind. Possible values are dnnl::algorithm::eltwise_relu, dnnl::algorithm::eltwise_tanh, or dnnl::algorithm::eltwise_logistic.
direction – RNN direction. See dnnl::rnn_direction for more info.
src_layer_desc – Memory descriptor for the input vector.
src_iter_desc – Memory descriptor for the input recurrent hidden state vector.
weights_layer_desc – Memory descriptor for the weights applied to the layer input.
weights_iter_desc – Memory descriptor for the weights applied to the recurrent input.
bias_desc – Bias memory descriptor.
dst_layer_desc – Memory descriptor for the output vector.
dst_iter_desc – Memory descriptor for the output recurrent hidden state vector.
diff_src_layer_desc – Memory descriptor for the diff of input vector.
diff_src_iter_desc – Memory descriptor for the diff of input recurrent hidden state vector.
diff_weights_layer_desc – Memory descriptor for the diff of weights applied to the layer input.
diff_weights_iter_desc – Memory descriptor for the diff of weights applied to the recurrent input.
diff_bias_desc – Diff bias memory descriptor.
diff_dst_layer_desc – Memory descriptor for the diff of output vector.
diff_dst_iter_desc – Memory descriptor for the diff of output recurrent hidden state vector.
alpha – Negative slope if activation is dnnl::algorithm::eltwise_relu.
hint_fwd_pd – Primitive descriptor for a vanilla RNN forward propagation primitive. It is used as a hint for deciding which memory format to use.
attr – Primitive attributes to use. Attributes are optional and default to empty attributes.
allow_empty – A flag signifying whether construction is allowed to fail without throwing an exception. In this case an empty object will be produced. This flag is optional and defaults to false.

memory::desc src_layer_desc() const#

Returns source layer memory descriptor.

Returns: Source layer memory descriptor.

memory::desc src_iter_desc() const#

Returns source iteration memory descriptor.

Returns: Source iteration memory descriptor.
Returns: A zero memory descriptor if the primitive does not have a source iteration parameter.

memory::desc weights_layer_desc() const#

Returns weights layer memory descriptor.

Returns: Weights layer memory descriptor.

memory::desc weights_iter_desc() const#

Returns weights iteration memory descriptor.

Returns: Weights iteration memory descriptor.

memory::desc bias_desc() const#

Returns bias memory descriptor.

Returns: Bias memory descriptor.
Returns: A zero memory descriptor if the primitive does not have a bias parameter.

memory::desc dst_layer_desc() const#

Returns destination layer memory descriptor.

Returns: Destination layer memory descriptor.

memory::desc dst_iter_desc() const#

Returns destination iteration memory descriptor.

Returns: Destination iteration memory descriptor.
Returns: A zero memory descriptor if the primitive does not have a destination iteration parameter.

memory::desc workspace_desc() const#

Returns the workspace memory descriptor.

Returns: Workspace memory descriptor.
Returns: A zero memory descriptor if the primitive does not require workspace parameter.

memory::desc diff_src_layer_desc() const#

Returns diff source layer memory descriptor.

Returns: Diff source layer memory descriptor.

memory::desc diff_src_iter_desc() const#

Returns diff source iteration memory descriptor.

Returns: Diff source iteration memory descriptor.
Returns: A zero memory descriptor if the primitive does not have a diff source iteration parameter.

memory::desc diff_weights_layer_desc() const#

Returns diff weights layer memory descriptor.

Returns: Diff weights layer memory descriptor.

memory::desc diff_weights_iter_desc() const#

Returns diff weights iteration memory descriptor.

Returns: Diff weights iteration memory descriptor.

memory::desc diff_bias_desc() const#

Returns diff bias memory descriptor.

Returns: Diff bias memory descriptor.
Returns: A zero memory descriptor if the primitive does not have a diff bias parameter.

memory::desc diff_dst_layer_desc() const#

Returns diff destination layer memory descriptor.

Returns: Diff destination layer memory descriptor.

memory::desc diff_dst_iter_desc() const#

Returns diff destination iteration memory descriptor.

Returns: Diff destination iteration memory descriptor.
Returns: A zero memory descriptor if the primitive does not have a diff destination iteration parameter.

algorithm get_cell_kind() const#

Returns an RNN cell kind parameter.

Returns: An RNN cell kind parameter.
Returns: dnnl::algorithm::undef if the primitive does not have an RNN cell kind parameter.

prop_kind get_prop_kind() const#

Returns a propagation kind.

Returns: A propagation kind.
Returns: dnnl::prop_kind::undef if the primitive does not have a propagation parameter.

algorithm get_activation_kind() const#

Returns an RNN activation kind parameter.

Returns: An RNN activation kind parameter.
Returns: dnnl::algorithm::undef if the primitive does not have an RNN activation kind parameter.

rnn_direction get_direction() const#

Returns an RNN direction parameter.

Returns: An RNN direction parameter.
Returns: dnnl::rnn_direction::undef if the primitive does not have an RNN direction parameter.

float get_alpha() const#

Returns an alpha.

Returns: An alpha.
Returns: Zero if the primitive does not have an alpha parameter.

float get_beta() const#

Returns a beta.

Returns: A beta.
Returns: Zero if the primitive does not have a beta parameter.

struct dnnl::lstm_forward : public dnnl::primitive #

LSTM forward propagation primitive.

Public Functions

lstm_forward()#: Default constructor. Produces an empty object.

lstm_forward(const primitive_desc &pd)#

Constructs an LSTM forward propagation primitive.

Parameters: pd – Primitive descriptor for an LSTM forward propagation primitive.

struct primitive_desc : public dnnl::rnn_primitive_desc_base #

Primitive descriptor for an LSTM forward propagation primitive.

Public Functions

primitive_desc() = default#: Default constructor. Produces an empty object.

Constructs a primitive descriptor for an LSTM (with or without peephole and with or without projection) forward propagation primitive.

The following arguments may point to a zero memory descriptor:

src_iter_desc together with src_iter_c_desc,
weights_peephole_desc,
bias_desc,
dst_iter_desc together with dst_iter_c_desc.

This would then indicate that the LSTM forward propagation primitive should not use them and should default to zero values instead.

The weights_projection_desc may point to a zero memory descriptor. This would then indicate that the LSTM doesn’t have recurrent projection layer.

Note

All memory descriptors can be initialized with an dnnl::memory::format_tag::any value of format_tag.

Parameters

aengine – Engine to use.
aprop_kind – Propagation kind. Possible values are dnnl::prop_kind::forward_training, and dnnl::prop_kind::forward_inference.
direction – RNN direction. See dnnl::rnn_direction for more info.
src_layer_desc – Memory descriptor for the input vector.
src_iter_desc – Memory descriptor for the input recurrent hidden state vector.
src_iter_c_desc – Memory descriptor for the input recurrent cell state vector.
weights_layer_desc – Memory descriptor for the weights applied to the layer input.
weights_iter_desc – Memory descriptor for the weights applied to the recurrent input.
weights_peephole_desc – Memory descriptor for the weights applied to the cell states (according to the Peephole LSTM formula).
weights_projection_desc – Memory descriptor for the weights applied to the hidden states to get the recurrent projection (according to the Projection LSTM formula).
bias_desc – Bias memory descriptor.
dst_layer_desc – Memory descriptor for the output vector.
dst_iter_desc – Memory descriptor for the output recurrent hidden state vector.
dst_iter_c_desc – Memory descriptor for the output recurrent cell state vector.
attr – Primitive attributes to use. Attributes are optional and default to empty attributes.
allow_empty – A flag signifying whether construction is allowed to fail without throwing an exception. In this case an empty object will be produced. This flag is optional and defaults to false.

Constructs a primitive descriptor for an LSTM (with or without peephole) forward propagation primitive.

The following arguments may point to a zero memory descriptor:

src_iter_desc together with src_iter_c_desc,
weights_peephole_desc,
bias_desc,
dst_iter_desc together with dst_iter_c_desc.

This would then indicate that the LSTM forward propagation primitive should not use them and should default to zero values instead.

Note

All memory descriptors can be initialized with an dnnl::memory::format_tag::any value of format_tag.

Parameters

aengine – Engine to use.
aprop_kind – Propagation kind. Possible values are dnnl::prop_kind::forward_training, and dnnl::prop_kind::forward_inference.
direction – RNN direction. See dnnl::rnn_direction for more info.
src_layer_desc – Memory descriptor for the input vector.
src_iter_desc – Memory descriptor for the input recurrent hidden state vector.
src_iter_c_desc – Memory descriptor for the input recurrent cell state vector.
weights_layer_desc – Memory descriptor for the weights applied to the layer input.
weights_iter_desc – Memory descriptor for the weights applied to the recurrent input.
weights_peephole_desc – Memory descriptor for the weights applied to the cell states (according to the Peephole LSTM formula).
bias_desc – Bias memory descriptor.
dst_layer_desc – Memory descriptor for the output vector.
dst_iter_desc – Memory descriptor for the output recurrent hidden state vector.
dst_iter_c_desc – Memory descriptor for the output recurrent cell state vector.
attr – Primitive attributes to use. Attributes are optional and default to empty attributes.
allow_empty – A flag signifying whether construction is allowed to fail without throwing an exception. In this case an empty object will be produced. This flag is optional and defaults to false.

Constructs a primitive descriptor for an LSTM forward propagation primitive.

The following arguments may point to a zero memory descriptor:

src_iter_desc together with src_iter_c_desc,
bias_desc,
dst_iter_desc together with dst_iter_c_desc.

This would then indicate that the LSTM forward propagation primitive should not use them and should default to zero values instead.

Note

All memory descriptors can be initialized with an dnnl::memory::format_tag::any value of format_tag.

Parameters

aengine – Engine to use.
aprop_kind – Propagation kind. Possible values are dnnl::prop_kind::forward_training, and dnnl::prop_kind::forward_inference.
direction – RNN direction. See dnnl::rnn_direction for more info.
src_layer_desc – Memory descriptor for the input vector.
src_iter_desc – Memory descriptor for the input recurrent hidden state vector.
src_iter_c_desc – Memory descriptor for the input recurrent cell state vector.
weights_layer_desc – Memory descriptor for the weights applied to the layer input.
weights_iter_desc – Memory descriptor for the weights applied to the recurrent input.
bias_desc – Bias memory descriptor.
dst_layer_desc – Memory descriptor for the output vector.
dst_iter_desc – Memory descriptor for the output recurrent hidden state vector.
dst_iter_c_desc – Memory descriptor for the output recurrent cell state vector.
attr – Primitive attributes to use. Attributes are optional and default to empty attributes.
allow_empty – A flag signifying whether construction is allowed to fail without throwing an exception. In this case an empty object will be produced. This flag is optional and defaults to false.

memory::desc src_layer_desc() const#

Returns source layer memory descriptor.

Returns: Source layer memory descriptor.

memory::desc src_iter_desc() const#

Returns source iteration memory descriptor.

Returns: Source iteration memory descriptor.
Returns: A zero memory descriptor if the primitive does not have a source iteration parameter.

memory::desc src_iter_c_desc() const#

Returns source iteration memory descriptor.

Returns: Source iteration memory descriptor.
Returns: A zero memory descriptor if the primitive does not have a source iteration parameter.

memory::desc weights_layer_desc() const#

Returns weights layer memory descriptor.

Returns: Weights layer memory descriptor.

memory::desc weights_iter_desc() const#

Returns weights iteration memory descriptor.

Returns: Weights iteration memory descriptor.

memory::desc weights_peephole_desc() const#

Returns weights peephole memory descriptor.

Returns: Weights peephole memory descriptor.

memory::desc weights_projection_desc() const#

Returns weights projection memory descriptor.

Returns: Weights projection memory descriptor.

memory::desc bias_desc() const#

Returns bias memory descriptor.

Returns: Bias memory descriptor.
Returns: A zero memory descriptor if the primitive does not have a bias parameter.

memory::desc dst_layer_desc() const#

Returns destination layer memory descriptor.

Returns: Destination layer memory descriptor.

memory::desc dst_iter_desc() const#

Returns destination iteration memory descriptor.

Returns: Destination iteration memory descriptor.
Returns: A zero memory descriptor if the primitive does not have a destination iteration parameter.

memory::desc dst_iter_c_desc() const#

Returns source iteration memory descriptor.

Returns: Source iteration memory descriptor.
Returns: A zero memory descriptor if the primitive does not have a source iteration parameter.

memory::desc workspace_desc() const#

Returns the workspace memory descriptor.

Returns: Workspace memory descriptor.
Returns: A zero memory descriptor if the primitive does not require workspace parameter.

algorithm get_cell_kind() const#

Returns an RNN cell kind parameter.

Returns: An RNN cell kind parameter.
Returns: dnnl::algorithm::undef if the primitive does not have an RNN cell kind parameter.

prop_kind get_prop_kind() const#

Returns a propagation kind.

Returns: A propagation kind.
Returns: dnnl::prop_kind::undef if the primitive does not have a propagation parameter.

rnn_direction get_direction() const#

Returns an RNN direction parameter.

Returns: An RNN direction parameter.
Returns: dnnl::rnn_direction::undef if the primitive does not have an RNN direction parameter.

struct dnnl::lstm_backward : public dnnl::primitive #

LSTM backward propagation primitive.

Public Functions

lstm_backward()#: Default constructor. Produces an empty object.

lstm_backward(const primitive_desc &pd)#

Constructs an LSTM backward propagation primitive.

Parameters: pd – Primitive descriptor for an LSTM backward propagation primitive.

struct primitive_desc : public dnnl::rnn_primitive_desc_base #

Primitive descriptor for an LSTM backward propagation primitive.

Public Functions

primitive_desc() = default#: Default constructor. Produces an empty object.

primitive_desc(const engine &aengine, prop_kind aprop_kind, rnn_direction direction, const memory::desc &src_layer_desc, const memory::desc &src_iter_desc, const memory::desc &src_iter_c_desc, const memory::desc &weights_layer_desc, const memory::desc &weights_iter_desc, const memory::desc &weights_peephole_desc, const memory::desc &weights_projection_desc, const memory::desc &bias_desc, const memory::desc &dst_layer_desc, const memory::desc &dst_iter_desc, const memory::desc &dst_iter_c_desc, const memory::desc &diff_src_layer_desc, const memory::desc &diff_src_iter_desc, const memory::desc &diff_src_iter_c_desc, const memory::desc &diff_weights_layer_desc, const memory::desc &diff_weights_iter_desc, const memory::desc &diff_weights_peephole_desc, const memory::desc &diff_weights_projection_desc, const memory::desc &diff_bias_desc, const memory::desc &diff_dst_layer_desc, const memory::desc &diff_dst_iter_desc, const memory::desc &diff_dst_iter_c_desc, const lstm_forward::primitive_desc &hint_fwd_pd, const primitive_attr &attr = default_attr(), bool allow_empty = false)#

Constructs an LSTM (with or without peephole and with or without projection) primitive descriptor for backward propagation using prop_kind, direction, and memory descriptors.

The following arguments may point to a zero memory descriptor:

src_iter_desc together with src_iter_c_desc, diff_src_iter_desc, and diff_src_iter_c_desc,
weights_peephole_desc together with diff_weights_peephole_desc
bias_desc together with diff_bias_desc,
dst_iter_desc together with dst_iter_c_desc, diff_dst_iter_desc, and diff_dst_iter_c_desc.

This would then indicate that the LSTM backward propagation primitive should not use them and should default to zero values instead.

The weights_projection_desc together with diff_weights_projection_desc may point to a zero memory descriptor. This would then indicate that the LSTM doesn’t have recurrent projection layer.

Note

All memory descriptors can be initialized with dnnl::memory::format_tag::any value of format_tag.

Parameters

aengine – Engine to use.
aprop_kind – Propagation kind. Must be dnnl::prop_kind::backward.
direction – RNN direction. See dnnl::rnn_direction for more info.
src_layer_desc – Memory descriptor for the input vector.
src_iter_desc – Memory descriptor for the input recurrent hidden state vector.
src_iter_c_desc – Memory descriptor for the input recurrent cell state vector.
weights_layer_desc – Memory descriptor for the weights applied to the layer input.
weights_iter_desc – Memory descriptor for the weights applied to the recurrent input.
weights_peephole_desc – Memory descriptor for the weights applied to the cell states (according to the Peephole LSTM formula).
weights_projection_desc – Memory descriptor for the weights applied to the hidden states to get the recurrent projection (according to the Projection LSTM formula).
bias_desc – Bias memory descriptor.
dst_layer_desc – Memory descriptor for the output vector.
dst_iter_desc – Memory descriptor for the output recurrent hidden state vector.
dst_iter_c_desc – Memory descriptor for the output recurrent cell state vector.
diff_src_layer_desc – Memory descriptor for the diff of input vector.
diff_src_iter_desc – Memory descriptor for the diff of input recurrent hidden state vector.
diff_src_iter_c_desc – Memory descriptor for the diff of input recurrent cell state vector.
diff_weights_layer_desc – Memory descriptor for the diff of weights applied to the layer input.
diff_weights_iter_desc – Memory descriptor for the diff of weights applied to the recurrent input.
diff_weights_peephole_desc – Memory descriptor for the diff of weights applied to the cell states (according to the Peephole LSTM formula).
diff_weights_projection_desc – Memory descriptor for the diff of weights applied to the hidden states to get the recurrent projection (according to the Projection LSTM formula).
diff_bias_desc – Diff bias memory descriptor.
diff_dst_layer_desc – Memory descriptor for the diff of output vector.
diff_dst_iter_desc – Memory descriptor for the diff of output recurrent hidden state vector.
diff_dst_iter_c_desc – Memory descriptor for the diff of output recurrent cell state vector.
hint_fwd_pd – Primitive descriptor for an LSTM forward propagation primitive. It is used as a hint for deciding which memory format to use.
attr – Primitive attributes to use. Attributes are optional and default to empty attributes.
allow_empty – A flag signifying whether construction is allowed to fail without throwing an exception. In this case an empty object will be produced. This flag is optional and defaults to false.

primitive_desc(const engine &aengine, prop_kind aprop_kind, rnn_direction direction, const memory::desc &src_layer_desc, const memory::desc &src_iter_desc, const memory::desc &src_iter_c_desc, const memory::desc &weights_layer_desc, const memory::desc &weights_iter_desc, const memory::desc &weights_peephole_desc, const memory::desc &bias_desc, const memory::desc &dst_layer_desc, const memory::desc &dst_iter_desc, const memory::desc &dst_iter_c_desc, const memory::desc &diff_src_layer_desc, const memory::desc &diff_src_iter_desc, const memory::desc &diff_src_iter_c_desc, const memory::desc &diff_weights_layer_desc, const memory::desc &diff_weights_iter_desc, const memory::desc &diff_weights_peephole_desc, const memory::desc &diff_bias_desc, const memory::desc &diff_dst_layer_desc, const memory::desc &diff_dst_iter_desc, const memory::desc &diff_dst_iter_c_desc, const lstm_forward::primitive_desc &hint_fwd_pd, const primitive_attr &attr = default_attr(), bool allow_empty = false)#

Constructs an LSTM (with or without peephole) primitive descriptor for backward propagation using prop_kind, direction, and memory descriptors.

The following arguments may point to a zero memory descriptor:

src_iter_desc together with src_iter_c_desc, diff_src_iter_desc, and diff_src_iter_c_desc,
weights_peephole_desc together with diff_weights_peephole_desc
bias_desc together with diff_bias_desc,
dst_iter_desc together with dst_iter_c_desc, diff_dst_iter_desc, and diff_dst_iter_c_desc.

This would then indicate that the LSTM backward propagation primitive should not use them and should default to zero values instead.

Note

All memory descriptors may be initialized with dnnl::memory::format_tag::any value of format_tag.

Parameters

aengine – Engine to use.
aprop_kind – Propagation kind. Must be dnnl::prop_kind::backward.
direction – RNN direction. See dnnl::rnn_direction for more info.
src_layer_desc – Memory descriptor for the input vector.
src_iter_desc – Memory descriptor for the input recurrent hidden state vector.
src_iter_c_desc – Memory descriptor for the input recurrent cell state vector.
weights_layer_desc – Memory descriptor for the weights applied to the layer input.
weights_iter_desc – Memory descriptor for the weights applied to the recurrent input.
weights_peephole_desc – Memory descriptor for the weights applied to the cell states (according to the Peephole LSTM formula).
bias_desc – Bias memory descriptor.
dst_layer_desc – Memory descriptor for the output vector.
dst_iter_desc – Memory descriptor for the output recurrent hidden state vector.
dst_iter_c_desc – Memory descriptor for the output recurrent cell state vector.
diff_src_layer_desc – Memory descriptor for the diff of input vector.
diff_src_iter_desc – Memory descriptor for the diff of input recurrent hidden state vector.
diff_src_iter_c_desc – Memory descriptor for the diff of input recurrent cell state vector.
diff_weights_layer_desc – Memory descriptor for the diff of weights applied to the layer input.
diff_weights_iter_desc – Memory descriptor for the diff of weights applied to the recurrent input.
diff_weights_peephole_desc – Memory descriptor for the diff of weights applied to the cell states (according to the Peephole LSTM formula).
diff_bias_desc – Diff bias memory descriptor.
diff_dst_layer_desc – Memory descriptor for the diff of output vector.
diff_dst_iter_desc – Memory descriptor for the diff of output recurrent hidden state vector.
diff_dst_iter_c_desc – Memory descriptor for the diff of output recurrent cell state vector.
hint_fwd_pd – Primitive descriptor for an LSTM forward propagation primitive. It is used as a hint for deciding which memory format to use.
attr – Primitive attributes to use. Attributes are optional and default to empty attributes.
allow_empty – A flag signifying whether construction is allowed to fail without throwing an exception. In this case an empty object will be produced. This flag is optional and defaults to false.

primitive_desc(const engine &aengine, prop_kind aprop_kind, rnn_direction direction, const memory::desc &src_layer_desc, const memory::desc &src_iter_desc, const memory::desc &src_iter_c_desc, const memory::desc &weights_layer_desc, const memory::desc &weights_iter_desc, const memory::desc &bias_desc, const memory::desc &dst_layer_desc, const memory::desc &dst_iter_desc, const memory::desc &dst_iter_c_desc, const memory::desc &diff_src_layer_desc, const memory::desc &diff_src_iter_desc, const memory::desc &diff_src_iter_c_desc, const memory::desc &diff_weights_layer_desc, const memory::desc &diff_weights_iter_desc, const memory::desc &diff_bias_desc, const memory::desc &diff_dst_layer_desc, const memory::desc &diff_dst_iter_desc, const memory::desc &diff_dst_iter_c_desc, const lstm_forward::primitive_desc &hint_fwd_pd, const primitive_attr &attr = default_attr(), bool allow_empty = false)#

Constructs an LSTM primitive descriptor for backward propagation using prop_kind, direction, and memory descriptors.

The following arguments may point to a zero memory descriptor:

src_iter_desc together with src_iter_c_desc, diff_src_iter_desc, and diff_src_iter_c_desc,
bias_desc together with diff_bias_desc,
dst_iter_desc together with dst_iter_c_desc, diff_dst_iter_desc, and diff_dst_iter_c_desc.

This would then indicate that the LSTM backward propagation primitive should not use them and should default to zero values instead.

Note

All memory descriptors may be initialized with dnnl::memory::format_tag::any value of format_tag.

Parameters

aengine – Engine to use.
aprop_kind – Propagation kind. Must be dnnl::prop_kind::backward.
direction – RNN direction. See dnnl::rnn_direction for more info.
src_layer_desc – Memory descriptor for the input vector.
src_iter_desc – Memory descriptor for the input recurrent hidden state vector.
src_iter_c_desc – Memory descriptor for the input recurrent cell state vector.
weights_layer_desc – Memory descriptor for the weights applied to the layer input.
weights_iter_desc – Memory descriptor for the weights applied to the recurrent input.
bias_desc – Bias memory descriptor.
dst_layer_desc – Memory descriptor for the output vector.
dst_iter_desc – Memory descriptor for the output recurrent hidden state vector.
dst_iter_c_desc – Memory descriptor for the output recurrent cell state vector.
diff_src_layer_desc – Memory descriptor for the diff of input vector.
diff_src_iter_desc – Memory descriptor for the diff of input recurrent hidden state vector.
diff_src_iter_c_desc – Memory descriptor for the diff of input recurrent cell state vector.
diff_weights_layer_desc – Memory descriptor for the diff of weights applied to the layer input.
diff_weights_iter_desc – Memory descriptor for the diff of weights applied to the recurrent input.
diff_bias_desc – Diff bias memory descriptor.
diff_dst_layer_desc – Memory descriptor for the diff of output vector.
diff_dst_iter_desc – Memory descriptor for the diff of output recurrent hidden state vector.
diff_dst_iter_c_desc – Memory descriptor for the diff of output recurrent cell state vector.
hint_fwd_pd – Primitive descriptor for a convolution forward propagation primitive. It is used as a hint for deciding which memory format to use.
attr – Primitive attributes to use. Attributes are optional and default to empty attributes.
allow_empty – A flag signifying whether construction is allowed to fail without throwing an exception. In this case an empty object will be produced. This flag is optional and defaults to false.

memory::desc src_layer_desc() const#

Returns source layer memory descriptor.

Returns: Source layer memory descriptor.

memory::desc src_iter_desc() const#

Returns source iteration memory descriptor.

Returns: Source iteration memory descriptor.
Returns: A zero memory descriptor if the primitive does not have a source iteration parameter.

memory::desc src_iter_c_desc() const#

Returns source iteration memory descriptor.

Returns: Source iteration memory descriptor.
Returns: A zero memory descriptor if the primitive does not have a source iteration parameter.

memory::desc weights_layer_desc() const#

Returns weights layer memory descriptor.

Returns: Weights layer memory descriptor.

memory::desc weights_iter_desc() const#

Returns weights iteration memory descriptor.

Returns: Weights iteration memory descriptor.

memory::desc weights_peephole_desc() const#

Returns weights peephole memory descriptor.

Returns: Weights peephole memory descriptor.

memory::desc weights_projection_desc() const#

Returns weights projection memory descriptor.

Returns: Weights projection memory descriptor.

memory::desc bias_desc() const#

Returns bias memory descriptor.

Returns: Bias memory descriptor.
Returns: A zero memory descriptor if the primitive does not have a bias parameter.

memory::desc dst_layer_desc() const#

Returns destination layer memory descriptor.

Returns: Destination layer memory descriptor.

memory::desc dst_iter_desc() const#

Returns destination iteration memory descriptor.

Returns: Destination iteration memory descriptor.
Returns: A zero memory descriptor if the primitive does not have a destination iteration parameter.

memory::desc dst_iter_c_desc() const#

Returns source iteration memory descriptor.

Returns: Source iteration memory descriptor.
Returns: A zero memory descriptor if the primitive does not have a source iteration parameter.

memory::desc workspace_desc() const#

Returns the workspace memory descriptor.

Returns: Workspace memory descriptor.
Returns: A zero memory descriptor if the primitive does not require workspace parameter.

memory::desc diff_src_layer_desc() const#

Returns diff source layer memory descriptor.

Returns: Diff source layer memory descriptor.

memory::desc diff_src_iter_desc() const#

Returns diff source iteration memory descriptor.

Returns: Diff source iteration memory descriptor.
Returns: A zero memory descriptor if the primitive does not have a diff source iteration parameter.

memory::desc diff_src_iter_c_desc() const#

Returns diff source recurrent cell state memory descriptor.

Returns: Diff source recurrent cell state memory descriptor.

memory::desc diff_weights_layer_desc() const#

Returns diff weights layer memory descriptor.

Returns: Diff weights layer memory descriptor.

memory::desc diff_weights_iter_desc() const#

Returns diff weights iteration memory descriptor.

Returns: Diff weights iteration memory descriptor.

memory::desc diff_weights_peephole_desc() const#

Returns diff weights peephole memory descriptor.

Returns: Diff weights peephole memory descriptor.

memory::desc diff_weights_projection_desc() const#

Returns diff weights projection memory descriptor.

Returns: Diff weights projection memory descriptor.

memory::desc diff_bias_desc() const#

Returns diff bias memory descriptor.

Returns: Diff bias memory descriptor.
Returns: A zero memory descriptor if the primitive does not have a diff bias parameter.

memory::desc diff_dst_layer_desc() const#

Returns diff destination layer memory descriptor.

Returns: Diff destination layer memory descriptor.

memory::desc diff_dst_iter_desc() const#

Returns diff destination iteration memory descriptor.

Returns: Diff destination iteration memory descriptor.
Returns: A zero memory descriptor if the primitive does not have a diff destination iteration parameter.

memory::desc diff_dst_iter_c_desc() const#

Returns diff destination recurrent cell state memory descriptor.

Returns: Diff destination recurrent cell state memory descriptor.

algorithm get_cell_kind() const#

Returns an RNN cell kind parameter.

Returns: An RNN cell kind parameter.
Returns: dnnl::algorithm::undef if the primitive does not have an RNN cell kind parameter.

prop_kind get_prop_kind() const#

Returns a propagation kind.

Returns: A propagation kind.
Returns: dnnl::prop_kind::undef if the primitive does not have a propagation parameter.

rnn_direction get_direction() const#

Returns an RNN direction parameter.

Returns: An RNN direction parameter.
Returns: dnnl::rnn_direction::undef if the primitive does not have an RNN direction parameter.

struct dnnl::gru_forward : public dnnl::primitive #

GRU forward propagation primitive.

Public Functions

gru_forward()#: Default constructor. Produces an empty object.

gru_forward(const primitive_desc &pd)#

Constructs a GRU forward propagation primitive.

Parameters: pd – Primitive descriptor for a GRU forward propagation primitive.

struct primitive_desc : public dnnl::rnn_primitive_desc_base #

Primitive descriptor for a GRU forward propagation primitive.

Public Functions

primitive_desc() = default#: Default constructor. Produces an empty object.

Constructs a primitive descriptor for a GRU forward propagation primitive.

The following arguments may point to a zero memory descriptor:

src_iter_desc,
bias_desc,
dst_iter_desc.

This would then indicate that the GRU forward propagation primitive should not use them and should default to zero values instead.

Note

All memory descriptors except src_iter_desc may be initialized with an dnnl::memory::format_tag::any value of format_tag.

Parameters

aengine – Engine to use.
aprop_kind – Propagation kind. Possible values are dnnl::prop_kind::forward_training, and dnnl::prop_kind::forward_inference.
direction – RNN direction. See dnnl::rnn_direction for more info.
src_layer_desc – Memory descriptor for the input vector.
src_iter_desc – Memory descriptor for the input recurrent hidden state vector.
weights_layer_desc – Memory descriptor for the weights applied to the layer input.
weights_iter_desc – Memory descriptor for the weights applied to the recurrent input.
bias_desc – Bias memory descriptor.
dst_layer_desc – Memory descriptor for the output vector.
dst_iter_desc – Memory descriptor for the output recurrent hidden state vector.
attr – Primitive attributes to use. Attributes are optional and default to empty attributes.
allow_empty – A flag signifying whether construction is allowed to fail without throwing an exception. In this case an empty object will be produced. This flag is optional and defaults to false.

memory::desc src_layer_desc() const#

Returns source layer memory descriptor.

Returns: Source layer memory descriptor.

memory::desc src_iter_desc() const#

Returns source iteration memory descriptor.

Returns: Source iteration memory descriptor.
Returns: A zero memory descriptor if the primitive does not have a source iteration parameter.

memory::desc weights_layer_desc() const#

Returns weights layer memory descriptor.

Returns: Weights layer memory descriptor.

memory::desc weights_iter_desc() const#

Returns weights iteration memory descriptor.

Returns: Weights iteration memory descriptor.

memory::desc bias_desc() const#

Returns bias memory descriptor.

Returns: Bias memory descriptor.
Returns: A zero memory descriptor if the primitive does not have a bias parameter.

memory::desc dst_layer_desc() const#

Returns destination layer memory descriptor.

Returns: Destination layer memory descriptor.

memory::desc dst_iter_desc() const#

Returns destination iteration memory descriptor.

Returns: Destination iteration memory descriptor.
Returns: A zero memory descriptor if the primitive does not have a destination iteration parameter.

memory::desc workspace_desc() const#

Returns the workspace memory descriptor.

Returns: Workspace memory descriptor.
Returns: A zero memory descriptor if the primitive does not require workspace parameter.

algorithm get_cell_kind() const#

Returns an RNN cell kind parameter.

Returns: An RNN cell kind parameter.
Returns: dnnl::algorithm::undef if the primitive does not have an RNN cell kind parameter.

prop_kind get_prop_kind() const#

Returns a propagation kind.

Returns: A propagation kind.
Returns: dnnl::prop_kind::undef if the primitive does not have a propagation parameter.

rnn_direction get_direction() const#

Returns an RNN direction parameter.

Returns: An RNN direction parameter.
Returns: dnnl::rnn_direction::undef if the primitive does not have an RNN direction parameter.

struct dnnl::gru_backward : public dnnl::primitive #

GRU backward propagation primitive.

Public Functions

gru_backward()#: Default constructor. Produces an empty object.

gru_backward(const primitive_desc &pd)#

Constructs a GRU backward propagation primitive.

Parameters: pd – Primitive descriptor for a GRU backward propagation primitive.

struct primitive_desc : public dnnl::rnn_primitive_desc_base #

Primitive descriptor for a GRU backward propagation primitive.

Public Functions

primitive_desc() = default#: Default constructor. Produces an empty object.

Constructs a primitive descriptor for a GRU backward propagation primitive.

The following arguments may point to a zero memory descriptor:

src_iter_desc together with diff_src_iter_desc,
bias_desc together with diff_bias_desc,
dst_iter_desc together with diff_dst_iter_desc.

This would then indicate that the GRU backward propagation primitive should not use them and should default to zero values instead.

Note

All memory descriptors may be initialized with dnnl::memory::format_tag::any value of format_tag.

Parameters

aengine – Engine to use.
aprop_kind – Propagation kind. Must be dnnl::prop_kind::backward.
direction – RNN direction. See dnnl::rnn_direction for more info.
src_layer_desc – Memory descriptor for the input vector.
src_iter_desc – Memory descriptor for the input recurrent hidden state vector.
weights_layer_desc – Memory descriptor for the weights applied to the layer input.
weights_iter_desc – Memory descriptor for the weights applied to the recurrent input.
bias_desc – Bias memory descriptor.
dst_layer_desc – Memory descriptor for the output vector.
dst_iter_desc – Memory descriptor for the output recurrent hidden state vector.
diff_src_layer_desc – Memory descriptor for the diff of input vector.
diff_src_iter_desc – Memory descriptor for the diff of input recurrent hidden state vector.
diff_weights_layer_desc – Memory descriptor for the diff of weights applied to the layer input.
diff_weights_iter_desc – Memory descriptor for the diff of weights applied to the recurrent input.
diff_bias_desc – Diff bias memory descriptor.
diff_dst_layer_desc – Memory descriptor for the diff of output vector.
diff_dst_iter_desc – Memory descriptor for the diff of output recurrent hidden state vector.
hint_fwd_pd – Primitive descriptor for a GRU forward propagation primitive. It is used as a hint for deciding which memory format to use.
attr – Primitive attributes to use. Attributes are optional and default to empty attributes.
allow_empty – A flag signifying whether construction is allowed to fail without throwing an exception. In this case an empty object will be produced. This flag is optional and defaults to false.

memory::desc src_layer_desc() const#

Returns source layer memory descriptor.

Returns: Source layer memory descriptor.

memory::desc src_iter_desc() const#

Returns source iteration memory descriptor.

Returns: Source iteration memory descriptor.
Returns: A zero memory descriptor if the primitive does not have a source iteration parameter.

memory::desc weights_layer_desc() const#

Returns weights layer memory descriptor.

Returns: Weights layer memory descriptor.

memory::desc weights_iter_desc() const#

Returns weights iteration memory descriptor.

Returns: Weights iteration memory descriptor.

memory::desc bias_desc() const#

Returns bias memory descriptor.

Returns: Bias memory descriptor.
Returns: A zero memory descriptor if the primitive does not have a bias parameter.

memory::desc dst_layer_desc() const#

Returns destination layer memory descriptor.

Returns: Destination layer memory descriptor.

memory::desc dst_iter_desc() const#

Returns destination iteration memory descriptor.

Returns: Destination iteration memory descriptor.
Returns: A zero memory descriptor if the primitive does not have a destination iteration parameter.

memory::desc workspace_desc() const#

Returns the workspace memory descriptor.

Returns: Workspace memory descriptor.
Returns: A zero memory descriptor if the primitive does not require workspace parameter.

memory::desc diff_src_layer_desc() const#

Returns diff source layer memory descriptor.

Returns: Diff source layer memory descriptor.

memory::desc diff_src_iter_desc() const#

Returns diff source iteration memory descriptor.

Returns: Diff source iteration memory descriptor.
Returns: A zero memory descriptor if the primitive does not have a diff source iteration parameter.

memory::desc diff_weights_layer_desc() const#

Returns diff weights layer memory descriptor.

Returns: Diff weights layer memory descriptor.

memory::desc diff_weights_iter_desc() const#

Returns diff weights iteration memory descriptor.

Returns: Diff weights iteration memory descriptor.

memory::desc diff_bias_desc() const#

Returns diff bias memory descriptor.

Returns: Diff bias memory descriptor.
Returns: A zero memory descriptor if the primitive does not have a diff bias parameter.

memory::desc diff_dst_layer_desc() const#

Returns diff destination layer memory descriptor.

Returns: Diff destination layer memory descriptor.

memory::desc diff_dst_iter_desc() const#

Returns diff destination iteration memory descriptor.

Returns: Diff destination iteration memory descriptor.
Returns: A zero memory descriptor if the primitive does not have a diff destination iteration parameter.

algorithm get_cell_kind() const#

Returns an RNN cell kind parameter.

Returns: An RNN cell kind parameter.
Returns: dnnl::algorithm::undef if the primitive does not have an RNN cell kind parameter.

prop_kind get_prop_kind() const#

Returns a propagation kind.

Returns: A propagation kind.
Returns: dnnl::prop_kind::undef if the primitive does not have a propagation parameter.

rnn_direction get_direction() const#

Returns an RNN direction parameter.

Returns: An RNN direction parameter.
Returns: dnnl::rnn_direction::undef if the primitive does not have an RNN direction parameter.

struct dnnl::lbr_gru_forward : public dnnl::primitive #

LBR GRU forward propagation primitive.

Public Functions

lbr_gru_forward()#: Default constructor. Produces an empty object.

lbr_gru_forward(const primitive_desc &pd)#

Constructs an LBR GRU forward propagation primitive.

Parameters: pd – Primitive descriptor for an LBR GRU forward propagation primitive.

struct primitive_desc : public dnnl::rnn_primitive_desc_base #

Primitive descriptor for an LBR GRU forward propagation primitive.

Public Functions

primitive_desc() = default#: Default constructor. Produces an empty object.

Constructs a primitive descriptor for LBR GRU forward propagation primitive.

The following arguments may point to a zero memory descriptor:

src_iter_desc,
bias_desc,
dst_iter_desc.

This would then indicate that the LBR GRU forward propagation primitive should not use them and should default to zero values instead.

Note

All memory descriptors except src_iter_desc may be initialized with an dnnl::memory::format_tag::any value of format_tag.

Parameters

aengine – Engine to use.
aprop_kind – Propagation kind. Possible values are dnnl::prop_kind::forward_training, and dnnl::prop_kind::forward_inference.
direction – RNN direction. See dnnl::rnn_direction for more info.
src_layer_desc – Memory descriptor for the input vector.
src_iter_desc – Memory descriptor for the input recurrent hidden state vector.
weights_layer_desc – Memory descriptor for the weights applied to the layer input.
weights_iter_desc – Memory descriptor for the weights applied to the recurrent input.
bias_desc – Bias memory descriptor.
dst_layer_desc – Memory descriptor for the output vector.
dst_iter_desc – Memory descriptor for the output recurrent hidden state vector.
attr – Primitive attributes to use. Attributes are optional and default to empty attributes.
allow_empty – A flag signifying whether construction is allowed to fail without throwing an exception. In this case an empty object will be produced. This flag is optional and defaults to false.

memory::desc src_layer_desc() const#

Returns source layer memory descriptor.

Returns: Source layer memory descriptor.

memory::desc src_iter_desc() const#

Returns source iteration memory descriptor.

Returns: Source iteration memory descriptor.
Returns: A zero memory descriptor if the primitive does not have a source iteration parameter.

memory::desc weights_layer_desc() const#

Returns weights layer memory descriptor.

Returns: Weights layer memory descriptor.

memory::desc weights_iter_desc() const#

Returns weights iteration memory descriptor.

Returns: Weights iteration memory descriptor.

memory::desc bias_desc() const#

Returns bias memory descriptor.

Returns: Bias memory descriptor.
Returns: A zero memory descriptor if the primitive does not have a bias parameter.

memory::desc dst_layer_desc() const#

Returns destination layer memory descriptor.

Returns: Destination layer memory descriptor.

memory::desc dst_iter_desc() const#

Returns destination iteration memory descriptor.

Returns: Destination iteration memory descriptor.
Returns: A zero memory descriptor if the primitive does not have a destination iteration parameter.

memory::desc workspace_desc() const#

Returns the workspace memory descriptor.

Returns: Workspace memory descriptor.
Returns: A zero memory descriptor if the primitive does not require workspace parameter.

algorithm get_cell_kind() const#

Returns an RNN cell kind parameter.

Returns: An RNN cell kind parameter.
Returns: dnnl::algorithm::undef if the primitive does not have an RNN cell kind parameter.

prop_kind get_prop_kind() const#

Returns a propagation kind.

Returns: A propagation kind.
Returns: dnnl::prop_kind::undef if the primitive does not have a propagation parameter.

rnn_direction get_direction() const#

Returns an RNN direction parameter.

Returns: An RNN direction parameter.
Returns: dnnl::rnn_direction::undef if the primitive does not have an RNN direction parameter.

struct dnnl::lbr_gru_backward : public dnnl::primitive #

LBR GRU backward propagation primitive.

Public Functions

lbr_gru_backward()#: Default constructor. Produces an empty object.

lbr_gru_backward(const primitive_desc &pd)#

Constructs an LBR GRU backward propagation primitive.

Parameters: pd – Primitive descriptor for an LBR GRU backward propagation primitive.

struct primitive_desc : public dnnl::rnn_primitive_desc_base #

Primitive descriptor for an LBR GRU backward propagation primitive.

Public Functions

primitive_desc() = default#: Default constructor. Produces an empty object.

Constructs a primitive descriptor for LBR GRU backward propagation primitive.

The following arguments may point to a zero memory descriptor:

src_iter_desc together with diff_src_iter_desc,
bias_desc together with diff_bias_desc,
dst_iter_desc together with diff_dst_iter_desc.

This would then indicate that the LBR GRU backward propagation primitive should not use them and should default to zero values instead.

Note

All memory descriptors may be initialized with dnnl::memory::format_tag::any value of format_tag.

Parameters

aengine – Engine to use.
aprop_kind – Propagation kind. Must be dnnl::prop_kind::backward.
direction – RNN direction. See dnnl::rnn_direction for more info.
src_layer_desc – Memory descriptor for the input vector.
src_iter_desc – Memory descriptor for the input recurrent hidden state vector.
weights_layer_desc – Memory descriptor for the weights applied to the layer input.
weights_iter_desc – Memory descriptor for the weights applied to the recurrent input.
bias_desc – Bias memory descriptor.
dst_layer_desc – Memory descriptor for the output vector.
dst_iter_desc – Memory descriptor for the output recurrent hidden state vector.
diff_src_layer_desc – Memory descriptor for the diff of input vector.
diff_src_iter_desc – Memory descriptor for the diff of input recurrent hidden state vector.
diff_weights_layer_desc – Memory descriptor for the diff of weights applied to the layer input.
diff_weights_iter_desc – Memory descriptor for the diff of weights applied to the recurrent input.
diff_bias_desc – Diff bias memory descriptor.
diff_dst_layer_desc – Memory descriptor for the diff of output vector.
diff_dst_iter_desc – Memory descriptor for the diff of output recurrent hidden state vector.
hint_fwd_pd – Primitive descriptor for an LBR GRU forward propagation primitive. It is used as a hint for deciding which memory format to use.
attr – Primitive attributes to use. Attributes are optional and default to empty attributes.
allow_empty – A flag signifying whether construction is allowed to fail without throwing an exception. In this case an empty object will be produced. This flag is optional and defaults to false.

memory::desc src_layer_desc() const#

Returns source layer memory descriptor.

Returns: Source layer memory descriptor.

memory::desc src_iter_desc() const#

Returns source iteration memory descriptor.

Returns: Source iteration memory descriptor.
Returns: A zero memory descriptor if the primitive does not have a source iteration parameter.

memory::desc weights_layer_desc() const#

Returns weights layer memory descriptor.

Returns: Weights layer memory descriptor.

memory::desc weights_iter_desc() const#

Returns weights iteration memory descriptor.

Returns: Weights iteration memory descriptor.

memory::desc bias_desc() const#

Returns bias memory descriptor.

Returns: Bias memory descriptor.
Returns: A zero memory descriptor if the primitive does not have a bias parameter.

memory::desc dst_layer_desc() const#

Returns destination layer memory descriptor.

Returns: Destination layer memory descriptor.

memory::desc dst_iter_desc() const#

Returns destination iteration memory descriptor.

Returns: Destination iteration memory descriptor.
Returns: A zero memory descriptor if the primitive does not have a destination iteration parameter.

memory::desc workspace_desc() const#

Returns the workspace memory descriptor.

Returns: Workspace memory descriptor.
Returns: A zero memory descriptor if the primitive does not require workspace parameter.

memory::desc diff_src_layer_desc() const#

Returns diff source layer memory descriptor.

Returns: Diff source layer memory descriptor.

memory::desc diff_src_iter_desc() const#

Returns diff source iteration memory descriptor.

Returns: Diff source iteration memory descriptor.
Returns: A zero memory descriptor if the primitive does not have a diff source iteration parameter.

memory::desc diff_weights_layer_desc() const#

Returns diff weights layer memory descriptor.

Returns: Diff weights layer memory descriptor.

memory::desc diff_weights_iter_desc() const#

Returns diff weights iteration memory descriptor.

Returns: Diff weights iteration memory descriptor.

memory::desc diff_bias_desc() const#

Returns diff bias memory descriptor.

Returns: Diff bias memory descriptor.
Returns: A zero memory descriptor if the primitive does not have a diff bias parameter.

memory::desc diff_dst_layer_desc() const#

Returns diff destination layer memory descriptor.

Returns: Diff destination layer memory descriptor.

memory::desc diff_dst_iter_desc() const#

Returns diff destination iteration memory descriptor.

Returns: Diff destination iteration memory descriptor.
Returns: A zero memory descriptor if the primitive does not have a diff destination iteration parameter.

algorithm get_cell_kind() const#

Returns an RNN cell kind parameter.

Returns: An RNN cell kind parameter.
Returns: dnnl::algorithm::undef if the primitive does not have an RNN cell kind parameter.

prop_kind get_prop_kind() const#

Returns a propagation kind.

Returns: A propagation kind.
Returns: dnnl::prop_kind::undef if the primitive does not have a propagation parameter.

rnn_direction get_direction() const#

Returns an RNN direction parameter.

Returns: An RNN direction parameter.
Returns: dnnl::rnn_direction::undef if the primitive does not have an RNN direction parameter.

oneAPI Specification 1.2-rev-1 documentation

RNN

Contents

RNN#

Cell Functions#

Vanilla RNN#

LSTM#

LSTM (or Vanilla LSTM)#

LSTM with Peephole#

LSTM with Projection#

GRU#

Linear-Before-Reset GRU#

AUGRU#

Linear-Before-Reset AUGRU#

Execution Arguments#

Operation Details#

Data Types Support#

Data Representation#

Post-ops and Attributes#

API#