RNN

Intel® oneAPI Deep Neural Network Developer Guide and Reference

Download PDF

ID 768875

Date 9/17/2024

Version

Public

Visible to Intel only — GUID: GUID-99EA5458-AA2C-4A5E-93EA-9BCCE1424639

View Details

RNN

API Reference

General

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

Figure 1: Example of stacked recurrent cells unrolled over the time dimension and executed with the `left2right` direction. Dashed lines represent optional parameters.

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 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 .

Even though the RNN primitive supports passing a different number of channels for , , , and , we always require the following conditions in order for the dimension to be consistent:

,
when , ,
when , ,
when using the bidirectional_concat direction, .

The general formula for the execution of a stack of unrolled recurrent cells depends on the current iteration of the previous layer ( and ) and the previous iteration of the current layer (). Here is the exact equation for non-LSTM cells:

where are the indices of the timestamp and the layer of the cell being executed.

And here is the equation for LSTM cells:

where are the indices of the timestamp and the layer of the cell being executed.

Cell Functions

The RNN API provides four 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::primitive_desc() or dnnl::vanilla_rnn_backward::primitive_desc::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);

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

LSTM

LSTM (or Vanilla LSTM)

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

auto lstm_pd = 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);

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

where are stored in , are stored in and are stored in .

NOTE:

In order for the dimensions to be consistent, we require

LSTM with Peephole

A four-gate long short-term memory recurrent cell with peephole initialized with dnnl::lstm_forward::primitive_desc::primitive_desc() or dnnl::lstm_backward::primitive_desc::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);

Similarly to vanilla LSTM, we implicitly require the order of the gates to be i, f, , and o for all tensors with a dimension depending on the gates. For peephole weights, the gates order is i, f, o. The following equation gives the mathematical description of these gates and output for the forward pass:

where 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 (or LSTMP)

A four-gate long short-term memory recurrent cell with projection initialized with dnnl::lstm_forward::primitive_desc::primitive_desc() or dnnl::lstm_backward::primitive_desc::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);

Similarly to vanilla LSTM, we implicitly require the order of the gates to be i, f, , 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):

where 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::primitive_desc() or dnnl::gru_backward::primitive_desc::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);

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.

where are in , are in , and are stored in .

NOTE:

If you need to replace u_t by (1-u_t) when computing h_t, you can achieve this by multiplying

and

. This is possible as

, and

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::primitive_desc() or dnnl::lbr_gru_backward::primitive_desc::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);

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

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 tensor, we implicitly require the order of the gates to be u, r, o, and u’`.

NOTE:

If you need to replace u_t by (1-u_t) when computing h_t, you can achieve this by multiplying

and

. This is possible as

, and

AUGRU

A three-gate gated recurrent unit cell, initialized with dnnl::augru_forward::primitive_desc::primitive_desc() or dnnl::augru_backward::primitive_desc::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);

where are in , are in , and are stored in .

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::primitive_desc() or dnnl::lbr_augru_backward::primitive_desc::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);

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

Considerations for Training

When using the RNN API for training, the forward pass should use the forward_training propagation kind, and a workspace should be passed to both the forward pass and the backward pass. Note that after executing the backward pass, the workspace is no more valid and should be populated once again by another forward pass.

The RNN primitive backward pass accumulates gradients to its weight outputs (namely , , , , ). Hence, these tensors should be properly initialized to zero before their first use, and can be reused across calls to accumulate gradients if need be. This behavior can be altered by the RNN flag diff_weights_overwrite. If this flag is set weight gradients will be initialized by zeros by the RNN primitive.

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
	DNNL_ARG_SRC_LAYER
	DNNL_ARG_SRC_LAYER_ATTENTION
	DNNL_ARG_SRC_ITER
	DNNL_ARG_SRC_ITER_C
	DNNL_ARG_WEIGHTS_LAYER
	DNNL_ARG_WEIGHTS_ITER
	DNNL_ARG_WEIGHTS_PEEPHOLE
	DNNL_ARG_WEIGHTS_PROJECTION
	DNNL_ARG_BIAS
	DNNL_ARG_DST_LAYER
	DNNL_ARG_DST_ITER
	DNNL_ARG_DST_ITER_C
	DNNL_WORKSPACE
	DNNL_ARG_DIFF_SRC_LAYER
	DNNL_ARG_DIFF_SRC_LAYER_ATTENTION
	DNNL_ARG_DIFF_SRC_ITER
	DNNL_ARG_DIFF_SRC_ITER_C
	DNNL_ARG_DIFF_WEIGHTS_LAYER
	DNNL_ARG_DIFF_WEIGHTS_ITER
	DNNL_ARG_DIFF_WEIGHTS_PEEPHOLE
	DNNL_ARG_DIFF_WEIGHTS_PROJECTION
	DNNL_ARG_DIFF_BIAS
	DNNL_ARG_DIFF_DST_LAYER
	DNNL_ARG_DIFF_DST_ITER
	DNNL_ARG_DIFF_DST_ITER_C

Implementation Details

Data Type Support

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

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, LSTMP and GRU	u8	u8	s8	f32	u8, f32
Forward inference	Vanilla LSTM, LSTMP	s8	s8	s8	f32	s8, f32

With LSTM and Peephole LSTM cells, the cell state datatype is f32, except for the f16 configuration.
In backward propagation, all diff_* tensors are in f32.
Projection LSTM is not supported.

WARNING:

There might be hardware and/or implementation specific restrictions. Check Implementation Limitations section below.

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.

Propagation	Input/Output Data	Recurrent Data	Layer and Iteration Weights	Peephole Weights and Bias	Projection LSTM Weights
Forward / Backward	dnnl_format_tag_any	dnnl_format_tag_any	dnnl_format_tag_any	dnnl_ldgo	dnnl_format_tag_any
Forward	dnnl_ntc , dnnl_tnc	dnnl_ldnc	dnnl_ldigo	dnnl_ldgo	dnnl_ldio
Backward	dnnl_ntc , dnnl_tnc	dnnl_ldnc	dnnl_ldigo , dnnl_ldgoi (GPU)	dnnl_ldgo	dnnl_ldoi

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 supports 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 variants of LSTM and GRU. See the markdown RNN int8 inference example for more details on how to use and set these quantization parameters.

Implementation Limitations

Refer to Data Types for limitations related to data types support.
Bias must always be present (that is, the corresponding memory descriptor argument cannot be zero memory descriptor when the RNN primitive descriptor is initialized).
CPU
- oneDNN supports s8 as input data only on systems with Advanced Matrix Extension(AMX) support.
- Projection LSTM for bf16 data type is not supported.
- f16 data type is not supported.
GPU
- No support for AUGRU.
- No support for Peephole LSTM and Projection LSTM.
- Int8 support is provided for LSTM only.
- Int8 workloads require weights layouts to be dnnl_format_tag_any.
- Bias and cell state of bf16 data type is not supported.

Example

LSTM RNN Primitive Example

This C++ API example demonstrates how to create and execute an LSTM RNN primitive in forward training propagation mode.

Key optimizations included in this example:

Creation of optimized memory format from the primitive descriptor.

Vanilla RNN Primitive Example

This C++ API example demonstrates how to create and execute a Vanilla RNN primitive in forward training propagation mode.

Key optimizations included in this example:

Creation of optimized memory format from the primitive descriptor.

AUGRU RNN Primitive Example

This C++ API example demonstrates how to create and execute an AUGRU RNN primitive in forward training propagation mode.

Key optimizations included in this example:

Creation of optimized memory format from the primitive descriptor.

Linear-Before-Reset GRU RNN Primitive Example

This C++ API example demonstrates how to create and execute a Linear-Before-Reset GRU RNN primitive in forward training propagation mode.

Key optimizations included in this example:

Creation of optimized memory format from the primitive descriptor.

Select Your Language

Using Intel.com Search

Quick Links

Recent Searches

Advanced Search

Only search in

Intel® oneAPI Deep Neural Network Developer Guide and Reference

RNN

General

Cell Functions

Considerations for Training

Execution Arguments

Implementation Details

Implementation Limitations

Example