Supported Layers

Deep Learning HDL Toolbox™ supports the layers listed in these tables.

Input Layers

Layer	Layer Type Hardware (HW) or Software(SW)	Description and Limitations	INT8 Compatible
imageInputLayer	SW	An image input layer inputs 2-D images to a network and applies data normalization. The normalization options zero-center and zscore can run on hardware if the compile method HardwareNormalization argument is enabled and the input data is of single data type. If the HardwareNormalization option is not enabled or the input data type is int8 the normalization runs in software. Normalization specified using a function handle is not supported. See Image Input Layer Normalization Hardware Implementation. When the Normalization property is set to none the activations function cannot be used for the imageInputLayer.	Yes. Runs as single datatype in SW.
featureInputLayer	SW	A feature input layer inputs feature data to a network and applies data normalization.	Yes
sequenceInputLayer	SW	A sequence input layer inputs sequence data to a network.	Yes
wordEmbeddingLayer (Text Analytics Toolbox)	SW	A word embedding layer maps word indices to vectors.	No

Convolution and Fully Connected Layers

Layer	Layer Type Hardware (HW) or Software(SW)	Layer Output Format	Description and Limitations	INT8 Compatible
convolution2dLayer	HW	Convolution (Conv)	A 2-D convolutional layer applies sliding convolutional filters to the input. When generating code for a network using this layer, these limitations apply: Filter size must be 1-201. Stride size must be 1-100 and square. Padding size must be in the range 0-100. Dilation factor supported up to [16 16] and must be square. Padding value is not supported. When the dilation factor is a multiple of three the calculated dilated filter size must have a maximum value of the existing convolution filter size limit. In all other cases, the filter size can be as large as the maximum value of the existing convolution filter size.	Yes
groupedConvolution2dLayer	HW	Convolution (Conv)	A 2-D grouped convolutional layer separates the input channels into groups and applies sliding convolutional filters. Use grouped convolutional layers for channel-wise separable (also known as depth-wise separable) convolution. Code generation is now supported for a 2-D grouped convolution layer that has the NumGroups property set as 'channel-wise'. When generating code for a network using this layer, these limitations apply: Filter size must be 1-201 and square. For example [1 1] or [14 14]. When the NumGroups is set as 'channel-wise', filter size must be 3-14. Stride size must be 1-100 and square. Padding size must be in the range 0-100. Dilation factor must be [1 1]. When the NumGroups is not set as 'channel-wise', number of groups must be 1 or 2. The input feature number must be greater than a single multiple of the square root of the ConvThreadNumber. When the NumGroups is not set as 'channel-wise', the number of filters per group must be a multiple of the square root of the ConvThreadNumber.	Yes
convolution1dLayer	HW	Inherit from input	A 1-D convolutional layer applies sliding convolutional filters to 1-D input. The layer convolves the input by moving the filters along the input and computing the dot product of the weights and the input, then adding a bias term. When generating code for a network using this layer, these limitations apply: Dilation must be 1. Filter size must be 1-201. Stride size must be 1-100. Padding size must be in the range 0-100. Padding value is not supported.	Yes
transposedConv2dLayer	HW	Convolution (Conv)	A transposed 2-D convolution layer upsamples feature maps. When generating code for a network using this layer, these limitations apply: Filter size must be 1-64 and square. Stride size must be 1-64 and square. Padding size must be in the range 0-8. Padding value is not supported.	Yes
fullyConnectedLayer	HW	Fully Connected (FC)	A fully connected layer multiplies the input by a weight matrix, and then adds a bias vector. When generating code for a network using this layer, these limitations apply: The layer input and output size are limited by the values specified in InputMemorySize and OutputMemorySize.	Yes

Activation Layers

Layer	Layer Type Hardware (HW) or Software(SW)	Layer Output Format	Description and Limitations	INT8 Compatible
reluLayer	HW	Layer is fused.	A ReLU layer performs a threshold operation to each element of the input where any value less than zero is set to zero. A ReLU layer is supported only when it is preceded by any of these layers: Convolution Fully Connected Adder	Yes
leakyReluLayer	HW	Layer is fused.	A leaky ReLU layer performs a threshold operation where any input value less than zero is multiplied by a fixed scalar. A leaky ReLU layer is supported only when it is preceded by any of these layers: Convolution Fully Connected Adder	Yes
clippedReluLayer	HW	Layer is fused.	A clipped ReLU layer performs a threshold operation where any input value less than zero is set to zero and any value above the clipping ceiling is set to that clipping ceiling value. A clipped ReLU layer is supported only when it is preceded by any of these layers: Convolution Fully Connected Adder	Yes
tanhLayer	HW	Inherit from input	A hyperbolic tangent (tanh) activation layer applies the tanh function on the layer inputs.	Yes. Runs as single datatype in HW.
swishLayer	HW	Inherit from input	A swish layer applies the swish activation function on layer inputs.	No
dlhdl.layer.mishLayer	HW	Inherit from input	A mish layer applies the mish activation function on layer inputs.	No
geluLayer	HW	Inherit from input	A Gaussian error linear unit (GELU) layer weighs the inputs by its probability under a Gaussian distribution. A GELU layer is supported only when it is preceded by any of these layers: Convolution Fully Connected When generating code for a network using this layer, these limitations apply: The Approximation property must be set to tanh. The Aprroximation value of none is not supported.	No

Normalization, Dropout, and Cropping Layers

Layer	Layer Type Hardware (HW) or Software(SW)	Layer Output Format	Description and Limitations	INT8 Compatible
batchNormalizationLayer	HW	Layer is fused.	A batch normalization layer normalizes each input channel across a mini-batch. A batch normalization layer is supported when preceded by an image input layer, convolution layer, or as a standalone layer.	Yes
crossChannelNormalizationLayer	HW	Convolution (Conv)	A channel-wise local response (cross-channel) normalization layer carries out channel-wise normalization. The WindowChannelSize must be in the range of 3-9 for code generation.	Yes. Runs as single datatype in HW.
dropoutLayer	NoOP on inference	NoOP on inference	A dropout layer randomly sets input elements to zero within a given probability.	Yes
resize2dLayer (Image Processing Toolbox)	HW	Inherit from input	A 2-D resize layer resizes 2-D input by a scale factor, to a specified height and width, or to the size of a reference input feature map. When generating code for a network using this layer, these limitations apply: The Method property must be set to nearest. The GeometricTransformationMode property must be set to half-pixel. The NearestRoundingMode property must be set to round. The ratio of the output size to input size must be an integer and in the range between two and 256. The resize2DLayer is not supported for Intel® FPGA and SoC boards.	Yes
crop2dLayer	HW	Inherit from input	A 2-D crop layer applies 2-D cropping to the input. When generating code for a network using this layer, these limitations apply: The maximum number of rows or columns in the input to the layer must be 512. The minimum number of columns in the output of the layer must be three.	Yes
dlhdl.layer.reshapeLayer	HW	Inherit from input	A reshape layer changes the shape of the layer activation data.	Yes
dlhdl.layer.sliceLayer	HW	Inherit from input	A slice layer divides the input to the layer into an equal number of groups along the channel dimension of the image. When generating code for a network using this layer, these limitations apply: Number of input channels to the layer must be a multiple of the Groups property of the layer. The group size should be a multiple of the convolution thread number. The size of the input to the layer must be a multiple of the data transfer number. Mathematically the data transfer number is calculated as power(2,nextpow2(sqrt(ConvThreadNumber))).	Yes
upsample2DLayer	HW	Inherit from input	During compiler optimization Deep Learning HDL Toolbox replaces this layer with a resize2-D layer.	Yes
layerNormalizationLayer	HW	Inherit from input	A layer normalization layer normalizes a mini-batch of data across all channels for each observation independently. During compiler optimization Deep Learning HDL Toolbox replaces this layer with a fully connected layer, and maps the scale to fully connected layer weights and offset to the fully connected layer bias. When generating code for a network using this layer, these limitations apply: The value of the Epsilon property must be 1e-5.	No

Pooling and Unpooling Layers

Layer	Layer Type Hardware (HW) or Software(SW)	Layer Output Format	Description and Limitations	INT8 Compatible
maxPooling2dLayer	HW	Convolution (Conv)	A max pooling layer performs downsampling by dividing the layer input into rectangular pooling regions and computing the maximum of each region. When generating code for a network using this layer, these limitations apply: Pool size must be 1-201. Stride size must be 1-100 and square. Padding size must be in the range 0-100. HasUnpoolingOutputs is supported. When this parameter is enabled, these limitations apply for code generation for this layer: Pool size must be 2-by-2 or 3-by-3. The stride size must be the same as the filter size. Padding size is not supported. Pool size and stride size must be square. For example, [2 2].	Yes No, when HasUnpoolingOutputs is enabled.
maxUnpooling2dLayer	HW	Convolution (Conv)	A max unpooling layer unpools the output of a max pooling layer.	No
averagePooling2dLayer	HW	Convolution (Conv)	An average pooling layer performs downsampling by dividing the layer input into rectangular pooling regions and computing the average values of each region. When generating code for a network using this layer, these limitations apply: Pool size must be 1-201. Stride size must be 1-100 and square. Padding size must be in the range 0-100.	Yes
globalAveragePooling2dLayer	HW	Convolution (Conv)	A global average pooling layer performs downsampling by computing the mean of the height and width dimensions of the input. When generating code for a network using this layer, these limitations apply: The pool size must be 1-201 and square.	Yes

Combination Layers

Layer	Layer Type Hardware (HW) or Software(SW)	Layer Output Format	Description and Limitations	INT8 Compatible
additionLayer	HW	Inherit from input.	An addition layer adds inputs from multiple neural network layers element-wise. You can now generated code for this layer with int8 data type when the layer is combined with a Leaky ReLU or Clipped ReLU layer. When generating code for a network using this layer, these limitations apply: Both input layers must have the same output layer format. For example, both layers must have conv output format or fc output format.	Yes
depthConcatenationLayer	HW	Inherit from input.	A depth concatenation layer takes inputs that have the same height and width and concatenates them along the third dimension (the channel dimension). When generating code for a network using this layer, these limitations apply: The input activation feature number must be a multiple of the data transfer number. Mathematically the data transfer number is calculated as power(2,nextpow2(sqrt(ConvThreadNumber))). See, External Memory Data Format. Layers that have a conv output format and layers that have an FC output format cannot be concatenated together.	Yes
multiplicationLayer	HW	Inherit from input	A multiplication layer multiplies inputs from multiple neural network layers element-wise.	Yes
dlhdl.layer.splitLayer	HW	Inherit from input	A split layer divides a layer output into equal parts along the channel dimension of the image.	Yes

Sequence Layers

Layer	Layer Type Hardware (HW) or Software(SW)	Description and Limitations	INT8 Compatible
lstmLayer	HW	An LSTM layer learns long-term dependencies between time steps in time series and sequence data. The layer performs additive interactions, which can help improve gradient flow over long sequences during training. When generating code for a network using this layer, these limitations apply: The input must be of single data type. The OutputMode property must be set to sequence.	No
gruLayer	HW	A GRU layer is an RNN layer that learns dependencies between time steps in time series and sequence data. When generating code for a network using this layer, these limitations apply: Inputs must be of single data type. You must set the GRU layer OutputMode to sequence.	No
lstmProjectedLayer	HW	A projected LSTM layer is a type of deep learning layer that enables compression by reducing the number of stored learnable parameters. When generating code for a network using this layer, these limitations apply: The input must be of single data type. The OutputMode property must be set to sequence.	No
gruProjectedLayer	HW	A projected GRU layer is a type of deep learning layer that enables compression by reducing the number of stored learnable parameters. When generating code for a network using this layer, these limitations apply: Inputs must be of single data type. The OutputMode property must be set to sequence.	No

Output Layer

Layer	Layer Type Hardware (HW) or Software(SW)	Description and Limitations	INT8 Compatible
softmaxLayer	SW and HW	A softmax layer applies a softmax function to the input. If the softmax layer is implemented in hardware: The inputs must be in the range -87 to 88. Softmax layer followed by adder layer or depth concatenation layer is not supported. The inputs to this layer must have the format 1-by-N, N-by-1, 1-by-1-by-N, N-by-1-by-1, and 1-by-N-by-1. If the convolution module of the deep learning processor is enabled the square root of the convolution thread number must be an integral power of two. If not, the layer is implemented in software.	Yes. Runs as single datatype in SW.
classificationLayer	SW	A classification layer computes the cross-entropy loss for multiclass classification issues that have mutually exclusive classes.	Yes
regressionLayer	SW	A regression layer computes the half mean squared error loss for regression problems.	Yes
sigmoidLayer	HW	A sigmoid layer applies a sigmoid function to the input. The sigmoid layer is implemented in the custom module of the deep learning processor configuration and runs as single datatype in HW.	Yes. Runs as single datatype in HW.

Keras and ONNX Layers

Layer	Layer Type Hardware (HW) or Software(SW)	Layer Output Format	Description and Limitations	INT8 Compatible
nnet.keras.layer.FlattenCStyleLayer	HW	Layer will be fused	Flatten activations into 1-D layers assuming C-style (row-major) order. A nnet.keras.layer.FlattenCStyleLayer is supported only when it is followed by a fully connected layer.	Yes
nnet.keras.layer.ZeroPadding2dLayer	HW	Layer will be fused.	Zero padding layer for 2-D input. A nnet.keras.layer.ZeroPadding2dLayer is supported only when it is followed by a convolution layer, maxpool layer, or a grouped convolution layer.	Yes
nnet.onnx.layer.FlattenInto2dLayer	HW	Layer will be fused	Flattens a MATLAB® 2D image batch in the way ONNX does, producing a 2D output array with CB format. A nnet.onnx.layer.FlattenInto2dLayer layer is fused with the following fully connected layer.	Yes
nnet.onnx.layer.FlattenLayer	HW	Layer will be fused	Flatten layer for ONNX™ network. A nnet.onnx.layer.FlattenLayer layer must be followed by a fully connected layer or a depth concatenation layer. If the layer following the nnet.onnx.layer.FlattenLayer layer is a depth concatenation layer: All inputs to the depth concatenation layer must flatten to layers of the same type. The following layer must be a fully connected layer. Flatten layers that are inputs to depth concatenation layers must not share inputs.	Yes
flattenLayer	HW/SW	Layer will be fused	A flatten layer collapses the spatial dimensions of the input into the channel dimension. A flatten layer should be followed by a fully connected layer. Starting in R2024b, you can use a flatten layer as the last layer in a network when you implement it as a SW layer.	Yes

Custom Layers

Layer	Layer Type Hardware (HW) or Software(SW)	Layer Output Format	Description and Limitations	INT8 Compatible
Custom Layers	HW	Inherit from input	Custom layers, with or without learnable parameters, that you define for your problem. To learn how to define your custom deep learning layers, see Create Deep Learning Processor Configuration for Custom Layers.	No

Image Input Layer Normalization Hardware Implementation

To enable hardware implementation of the normalization functions for the image input layer, set the HardwareNormalization argument of the compile method to auto or on. When HardwareNormalization is set to auto, the compile method looks for the presence of addition and multiplication layers to implement the normalization function on hardware. The normalization is implemented on hardware by:

• Creating a new constant layer, This layer holds the value which is to be subtracted.

• Using existing addition and multiplication layers. The layers to be used depends on the normalization function being implemented.

Constant Layer Buffer Content

This table describes the value stored in the constant layer buffer.

Normalization Function	Number of Constants	Constant Layer Buffer Value
zerocenter	1	- Mean
zscore	2	The first constant value is -Mean. The second constant value is 1/StandardDeviation