MATLAB Coder: How do I build the Arm CMSIS-NN library for Deep Learning C code generation and deployment?

I see a few deep learning networks supported for code generation using MATLAB Coder:

• Deep Learning Networks and Layers Supported for C++ Code Generation

I'm looking to generate code from my deep learning network to run on Arm Cortex-M CPUs using MATLAB Coder and the Arm CMSIS-NN library. What are the steps to do this?

Answer 1

[Bill Chou]

To generate and run C code for deep neural networks on Cortex-M hardware, you must have the CMSIS-NN library. See the post below for instructions on building the library:

• Build Steps for CMSIS-NN Library

Answer 2

[Bill Chou]

This is an old answer that has since been updated, and is kept here for reference only. Please see this post for up-to-date instructions.

To generate and run C code for deep neural networks on Cortex-M hardware, you must have the CMSIS-NN (Cortex Microcontroller Software Interface Standard - Neural Networks) library. This following describes the build steps for CMSIS-NN on Windows and Linux platforms that use a cross-compiler toolchain.

Requirements

To build the CMSIS-NN static library, you must first create a makefile. Copy the following code snippet into a file and save it as Makefile.mk. This sample makefile builds the CMSIS-NN library for the Cortex-M7 core. To build the library for other targets, modify the CFLAGS variable as appropriate for your target hardware.

STATIC_LIB_DIR=./lib 
STATIC_LIB_FILE_NAME = libcmsisnn.a 

CFLAGS= -fPIC -c -mcpu=cortex-m7 -Ofast -DNDEBUG 

INC = -I../Core/Include -I../DSP/PrivateInclude -I../DSP/Include -I../NN/Include  

rwildcard=$(foreach d,$(wildcard $(1:=/*)),$(call rwildcard,$d,$2) $(filter $(subst *,%,$2),$d)) 

SOURCE_FILES = $(call rwildcard,Source,*.c) 
OBJECT_FILES = $(patsubst %.c,%.o,$(wildcard $(SOURCE_FILES))) 

$(STATIC_LIB_FILE_NAME): $(OBJECT_FILES) 
ar -r -o $(STATIC_LIB_DIR)/$@ $^ 

#Compiling every *.c to *.o 
%.o: %.c dirmake 
arm-none-eabi-gcc -c $(INC) $(CFLAGS) -o $@  $< 

dirmake: 
@mkdir -p $(STATIC_LIB_DIR) 

clean: 
rm -f $(OBJECT_FILES) $(STATIC_LIB_DIR)/$(STATIC_LIB_FILE_NAME) 

Note: Make sure that Makefile was written properly, using the above code, and following makefile indentation rules.

We have validated the CMSIS-NN library build process below using:

• Windows platform: GNU Arm Embedded Toolchain version 10.3.1
• Linux platform: GNU Arm Embedded Toolchain version 8.3.0

Linux Platform

1. Install the open-source GNU Arm Embedded Toolchain provided by ARM (See https://developer.arm.com/tools-and-software/open-source-software/developer-tools/gnu-toolchain/gnu-rm for more information).

2. Download the source code for CMSIS version 5.7.0 (available at https://github.com/ARM-software/CMSIS_5/archive/refs/tags/5.7.0.zip).

3. Unzip the source code to a folder and follow these steps to build and generate the static library:

• Open a Linux terminal.
• Change directory to the CMSIS-NN source folder by running this command. Here, <CMSIS Root folder> points to the extracted CMSIS folder.

cd <CMSIS Root folder>/CMSIS/NN 

• Copy the Makefile.mk file that you created to the current directory.
• Run the makefile by using the make command.

make -f Makefile.mk 

• Running the makefile creates the static library libcmsisnn.a in the ./lib folder.

4. Configure the MATLAB environment to generate code that uses the CMSIS-NN library:

• At /usr/local, create a folder named cmsisnn.
• Copy the header files located at <CMSIS Root folder>/CMSIS/DSP/Include and <CMSIS Root folder>/CMSIS/NN/Include to the location /usr/local/cmsisnn/include.
• Copy the generated static library located at <CMSIS Root folder>/CMSIS/NN/Include to the location /usr/local/cmsisnn/lib.
• At the MATLAB command window, set the MATLAB environment variable CMSISNN_PATH to the location of the CMSIS-NN install folder.

setenv('CMSISNN_PATH’, ‘/usr/local/cmsisnn’) 

Windows Platform

1. Install open-source GNU Arm Embedded Toolchain provided by ARM (available at https://developer.arm.com/-/media/Files/downloads/gnu-rm/10.3-2021.10/gcc-arm-none-eabi-10.3-2021.10-win32.exe).

For more information, see https://developer.arm.com/tools-and-software/open-source-software/developer-tools/gnu-toolchain/gnu-rm.

2. Download and install the 'build system(make.exe)' and 'archiving(ar.exe)' tools. We have tested this step by using Cygwin (available at https://www.cygwin.com/) to run the makefile.

3. Download the source code for CMSIS version 5.7.0 (available at https://github.com/ARM-software/CMSIS_5/archive/refs/tags/5.7.0.zip).

4. Unzip the source code to a folder, and follow these steps to build and generate the static library:

• Open a Windows command prompt.
• Change directory to the CMSIS-NN source folder by running this command. Here, <CMSIS Root folder> points to the extracted CMSIS folder.

cd <CMSIS Root folder>\CMSIS\NN 

• Copy the Makefile.mk file that you created to the current directory.
• Run the makefile by using the make command.

make -f Makefile.mk 

• Running the makefile creates the static library libcmsisnn.a in the .\lib folder.

5. Configure the MATLAB environment to generate code that uses the CMSIS-NN library:

• At C:\Program Files, create a folder named cmsisnn.
• Copy the header files located at <CMSIS Root folder>\CMSIS\DSP\Include and <CMSIS Root folder>\CMSIS\NN\Include to the location C:\Program Files\cmsisnn\include.
• Copy the generated static library located at <CMSIS Root folder>\CMSIS\NN\Include to the location C:\Program Files\cmsisnn\lib.
• At the MATLAB command window, set the MATLAB environment variable CMSISNN_PATH to the location of the CMSIS-NN install folder.

setenv('CMSISNN_PATH’, ‘C:\Program Files\cmsisnn’)