Code Generation for Deep Learning Networks Targeting ARM Mali GPUs

With GPU Coder™, you can generate optimized code for prediction of a variety of trained deep learning networks from Deep Learning Toolbox™. The generated code implements the deep convolutional neural network (CNN) by using the architecture, the layers, and parameters that you specify in the input SeriesNetwork (Deep Learning Toolbox) or DAGNetwork (Deep Learning Toolbox) object. The code generator takes advantage of the ARM® Compute Library for computer vision and machine learning. For performing deep learning on ARM Mali GPU targets, you generate code on the host development computer. Then, to build and run the executable program move the generated code to the ARM target platform. For example, HiKey960 is one of the target platforms that can execute the generated code.


  1. Deep Learning Toolbox.

  2. Deep Learning Toolbox Model for MobileNet-v2 Network support package.

  3. GPU Coder Interface for Deep Learning Libraries support package. To install the support packages, select the support package from the MATLAB® Add-Ons menu.

  4. ARM Compute Library for computer vision and machine learning must be installed on the target hardware. For information on the supported versions of the compilers and libraries, see Installing Prerequisite Products.

  5. Environment variables for the compilers and libraries. For more information, see Environment Variables.

Load Pretrained Network

  1. Load the pretrained MobileNet-v2 network. You can choose to load a different pretrained network for image classification. If you do not have the required support packages installed, the software provides a download link.

    net = mobilenetv2;

  2. The object net contains the DAGNetwork object. Use the analyzeNetwork (Deep Learning Toolbox) function to display an interactive visualization of the network architecture, to detect errors and issues in the network, and to display detailed information about the network layers. The layer information includes the sizes of layer activations and learnable parameters, the total number of learnable parameters, and the sizes of state parameters of recurrent layers.


  3. The image that you want to classify must have the same size as the input size of the network. For GoogLeNet, the size of the imageInputLayer (Deep Learning Toolbox) is 224-by-224-by-3. The Classes property of the output classificationLayer (Deep Learning Toolbox) contains the names of the classes learned by the network. View 10 random class names out of the total of 1000.

    classNames = net.Layers(end).Classes;
    numClasses = numel(classNames);
         soap dispenser 
         car wheel 

    For more information, see List of Deep Learning Layers (Deep Learning Toolbox).

Code Generation by Using cnncodegen

To generate code with the ARM Compute Library, use the targetlib option of the cnncodegen command. The cnncodegen command generates C++ code for the SeriesNetwork or DAGNetwork network object.

  1. Call cnncodegen with 'targetlib' specified as 'arm-compute-mali'. For example:

    net = googlenet;

    For 'arm-compute-mali', the value of batchsize must be 1.

    The 'targetparams' name-value pair arguments that enable you to specify Library-specific parameters for the ARM Compute Library is not applicable when targeting ARM Mali GPUs.

  2. The cnncodegen command generates code, a makefile,, and other supporting files to build the generated code on the target hardware. The command places all the generated files in the codegen folder.

  3. Write a C++ main function that calls predict. For an example main file that interfaces with the generated code, see Deep Learning Prediction on ARM Mali GPU

  4. Move the generated codegen folder and other files from the host development computer to the ARM hardware by using your preferred Secure File Copy (SCP) and Secure Shell (SSH) client. Build the executable program on the target.

Generated Code

The DAG network is generated as a C++ class (CnnMain) containing an array of 103 layer classes. The code generator reduces the number of layers is by layer fusion optimization of convolutional and batch normalization layers. A snippet of the class declaration from cnn_exec.hpp file is shown.

 cnn_exec.hpp File

  • The setup() method of the class sets up handles and allocates memory for each layer of the network object.

  • The predict() method invokes prediction for each of the 103 layers in the network.

  • The cnn_exec.cpp file contains the definitions of the object functions for the CnnMain class.

Binary files are exported for layers with parameters such as fully connected and convolution layers in the network. For instance, files cnn_CnnMain_Conv*_w and cnn_CnnMain_Conv*_b correspond to weights and bias parameters for the convolutional layers in the network. The code generator places these binary files in the codegen folder. The code generator builds the library file cnnbuild and places all the generated files in the codegen folder.


  • Code generation for the ARM Mali GPU is not supported for a 2-D grouped convolution layer that has the NumGroups property set as 'channel-wise' or a value greater than two.

See Also


Related Topics