CMSIS NN software library is a collection of efficient neural network kernels developed to maximize the performance and minimize the memory footprint of neural networks on Arm Cortex-M processors.
The library follows the int8 and int16 quantization specification of TensorFlow Lite for Microcontrollers. This means CMSIS-NN is bit-exact with Tensorflow Lite reference kernels. In some cases TFL and TFLM reference kernels may not be bit-exact. In that case CMSIS-NN follows TFLM reference kernels. The unit test readme provides an overview.
There is a single branch called 'main'. Tags are created during a release. Two releases are planned to be done in a year. The releases can be found here .
In general optimizations are written for an architecture feature. This falls into one of the following categories. Based on feature flags for a processor or architecture provided to the compiler, the right implementation is picked.
There is always a pure C implementation for an operator. This is used for processors like Arm Cortex-M0 or Cortex-M3.
Processors with DSP extension uses Single Instruction Multiple Data(SIMD) instructions for optimization. Examples of processors here are Cortex-M4 or a Cortex-M33 configured with optional DSP extension.
Processors with Arm Helium Technology use the Arm M-profile Vector Extension(MVE) instructions for optimization. Examples are Cortex-M55 or Cortex-M85 configured with MVE.
Operator | C int8 |
C int16 |
C int4* |
DSP int8 |
DSP int16 |
DSP int4* |
MVE int8 |
MVE int16 |
MVE int4* |
---|---|---|---|---|---|---|---|---|---|
Conv2D | Yes | Yes | Yes | Yes | Yes | Yes | Yes | Yes | Yes |
DepthwiseConv2D | Yes | Yes | Yes | Yes | Yes | Yes | Yes | Yes | Yes |
TransposeConv2D | Yes | No | No | Yes | No | No | Yes | No | No |
Fully Connected | Yes | Yes | Yes | Yes | Yes | Yes | Yes | Yes | Yes |
Batch Matmul | Yes | Yes | No | Yes | Yes | No | Yes | Yes | No |
Add | Yes | Yes | N/A | Yes | Yes | N/A | Yes | Yes | N/A |
Mul | Yes | Yes | N/A | Yes | Yes | N/A | Yes | Yes | N/A |
MaxPooling | Yes | Yes | N/A | Yes | Yes | N/A | Yes | Yes | N/A |
AvgPooling | Yes | Yes | N/A | Yes | Yes | N/A | Yes | Yes | N/A |
Softmax | Yes | Yes | N/A | Yes | Yes | N/A | Yes | No | N/A |
LSTM | Yes | Yes | No | Yes | Yes | No | Yes | Yes | No |
SVDF | Yes | No | No | Yes | No | No | Yes | No | No |
- int4 weights + int8 activations
First, a thank you for the contribution. Here are some guidelines and good to know information to get started.
By default, follow the style used in the file. You'll soon start noticing a pattern like
- Variable and function names are lower case with an underscore separator.
- Hungarian notation is not used. Well, almost.
- If the variable names don't convey the action, then add comments.
One function per file is followed in most places. In those cases, the file name must match the function name. Connect the function to an appropriate Doxygen group as well.
Function prototypes must have a detailed comment header in Doxygen format. You can execute the doxygen document generation script in the Documentation/Doxygen folder to check that no errors are introduced.
For any new features and bug fixes, new unit tests are needed. Improvements have to be verifed by unit tests. If you do not have the means to execute the tests, you can still make the PR and comment that you need help in completing/executing the unit tests.
Each File has a version number and a date field that must be updated when making any change to that file. The versioning follows Semantic Versioning 2.0.0 format. For details check: https://semver.org/
It is recommended to use toolchain files from Arm Ethos-U Core Platform project. These are supporting TARGET_CPU, which is a required argument. Note that if not specifying TARGET_CPU, these toolchains will set some default. The format must be TARGET_CPU=cortex-mXX, see examples below.
Here is an example:
cd </path/to/CMSIS_NN>
mkdir build
cd build
cmake .. -DCMAKE_TOOLCHAIN_FILE=</path/to/ethos-u-core-platform>/cmake/toolchain/arm-none-eabi-gcc.cmake -DTARGET_CPU=cortex-m55
make
Some more examples:
cmake .. -DCMAKE_TOOLCHAIN_FILE=</path/to/ethos-u-core-platform>/cmake/toolchain/armclang.cmake -DTARGET_CPU=cortex-m55
cmake .. -DCMAKE_TOOLCHAIN_FILE=</path/to/ethos-u-core-platform>/cmake/toolchain/arm-none-eabi-gcc.cmake -DTARGET_CPU=cortex-m7
cmake .. -DCMAKE_TOOLCHAIN_FILE=</path/to/ethos-u-core-platform>/cmake/toolchain/armclang.cmake -DTARGET_CPU=cortex-m3
Default optimization level is set at Ofast. This can be overwritten with CMake on command line by using "-DCMSIS_OPTIMIZATION_LEVEL". Please change according to project needs. Just bear in mind this can impact performance. With only optimization level -O0, ARM_MATH_AUTOVECTORIZE needs to be defined for processors with Helium Technology.
The compiler option '-fomit-frame-pointer' is enabled by default at -O and higher. When no optimization level is specified, you may need to specify '-fomit-frame-pointer'.
The compiler option '-fno-builtin' does not utilize optimized implementations of e.g. memcpy and memset, which are heavily used by CMSIS-NN. It can significantly downgrade performance. So this should be avoided. The compiler option '-ffreestanding' should also be avoided as it enables '-fno-builtin' implicitly.
Another option is to enable CMSIS_NN_USE_SINGLE_ROUNDING. This may affect the output. If enabling this the equivalent flag should be enabled in TFL/TFLM.
- CMSIS-NN is tested on Arm Compiler 6 and on Arm GNU Toolchain.
- IAR compiler is not tested and there can be compilation and/or performance issues.
- Compilation for Host is not supported out of the box. It should be possible to use the C implementation and compile for host with minor stubbing effort.
This product confirms to Arm’s inclusive language policy and, to the best of our knowledge, does not contain any non-inclusive language. If you find something that concerns you, email [email protected].
For any questions or to reach the CMSIS-NN team, please create a new issue in https://github.com/ARM-software/CMSIS-NN/issues