Supported Deep Learning Toolbox Blocks with CMSIS Library for ARM Cortex-M Processors

R2026b

Deep learning layer blocks that can be used with the Embedded Coder® Support Package for ARM® Cortex®-M Processors require specific conditions to allow code replacement with the common microcontroller software interface standard (CMSIS) library.

You use this code replacement when generating C code from a model. The CMSIS library support these Deep Learning Toolbox™ layer blocks only when you set specific parameters, as indicated in the following table.

Deep Learning Layer Blocks

Deep Learning Layer Blocks

Deep Learning Toolbox Blocks Supported Input Data TypesParameter SpecificationsWrappers calling CMSIS function
Convolution 2D Layer (Deep Learning Toolbox)

  • fixdt(1,8,*)

  • Weights must be fixdt(1,8,*)

  • Bias must be fixdt(1,32,*)

  • Output data type must be fixdt(1,8,*)

For grouped convolution:

  • Output data type must be fixdt(1,8,*)

Note

  • Code replacement occurs only when the Layer parameter supports convolution2dLayer (Deep Learning Toolbox) objects that have DilationFactor set to the default value [1 1].

  • Padding: Supports both symmetric and asymmetric padding. Asymmetric padding allows different padding sizes for top, bottom, left, and right edges.

  • The PaddingValue property of convolution2dLayer (Deep Learning Toolbox) object must be a scalar and equal to 0.

For convolution and grouped convolution:

  • MWDLConv2DInt8InputInt8Output

For depth-wise convolution:

  • MWDLDepthwiseConv2DInt8InputInt8Output

Fully Connected Layer (Deep Learning Toolbox)

  • fixdt(1,8,*)

  • Weights must be fixdt(1,8,*)

  • Bias must be fixdt(1,32,*)

  • Output data type must be fixdt(1,8,*) or fixdt(1,32,*)

  • MWDLFCInt8InputInt8Output

  • MWDLFCInt8InputInt32Output

LSTM Layer (Deep Learning Toolbox)

  • fixdt(1,8,*)

  • Input weights must be fixdt(1,8,*)

  • Recurrent weights must be fixdt(1,8,*)

  • Bias must be fixdt(1,32,*)

  • Output data type must be fixdt(1,8,*)

  • Cell state must be fixdt(1,16,*)

  • Hidden state must be fixdt(1,8,*)

  • Stateful prediction must be disabled.

  • MWDLLSTMInt8InputInt8Output

LSTM Projected Layer (Deep Learning Toolbox)

  • fixdt(1,8,*)

  • Input weights must be fixdt(1,8,*)

  • Recurrent weights must be fixdt(1,8,*)

  • Bias must be fixdt(1,32,*)

  • Output data type must be fixdt(1,8,*)

  • Input projector weights must be fixdt(1,8,*)

  • Output projector weights must be fixdt(1,8,*)

  • MWDLPLSTMInt8InputInt8Output

See Also

Topics