1# Unit tests for CMSIS-NN 2 3Unit test CMSIS-NN functions on any [Arm Mbed OS](https://os.mbed.com/mbed-os/) supported HW or using a fixed virtual platform (FVP) based on [Arm Corstone-300 software](https://developer.arm.com/ip-products/subsystem/corstone/corstone-300). 4 5The [Unity test framework](http://www.throwtheswitch.org/unity) is used for running the actual unit tests. 6 7## Requirements 8 9Python3 is required. 10It has been tested with Python 3.6 and it has been tested on Ubuntu 16, 18 and 20. 11 12Make sure to use the latest pip version before starting. 13If in a virtual environment just start by upgrading pip. 14 15``` 16pip install --upgrade pip 17``` 18 19See below for what pip packages are needed. 20 21### Executing unit tests 22 23If using the script unittest_targets.py for executing unit tests, the following packages are needed. 24 25``` 26pip install pyserial mbed-ls mbed-cli termcolor mercurial pyelftools==0.29 pyyaml jsonschema jinja2 mbed_host_tests mbed_greentea pycryptodome pyusb cmsis_pack_manager psutil cryptography click cbor 27``` 28 29Python packages mbed-cli and and mbed-ls are command line tools so it should not matter if those are installed under Python2 or Python3. These packages have been tested for Python2, with the following versions: mbed-ls(1.7.9) and mbed-cli(1.10.1). They have also been tested for Python3, with the following versions: mbed-ls(1.7.12) and mbed-cli(1.10.5). Package mercurial is needed for package mbed-cli. 30 31### Generating new test data 32 33For generating new test data, the following packages are needed. 34 35``` 36pip install numpy packaging tensorflow tf-keras~=2.16 37``` 38 39 40For generating new data, the python3 packages tensorflow, numpy and packaging are required. Most unit tests use a Keras generated model for reference. The SVDF unit test use a json template as input for generating a model. To do so flatc compiler is needed and it requires a schema file. 41 42#### Get flatc and schema 43 44Note this is only needed for generating SVDF unit tests. 45 46For flatc compiler clone this [repo](https://github.com/google/flatbuffers) and build: 47``` 48cd flatbuffers 49cmake -G "Unix Makefiles" -DCMAKE_BUILD_TYPE=Release 50make 51``` 52Remember to add the built flatc binary to the path. 53 54For schema file download [schema.fbs](https://raw.githubusercontent.com/tensorflow/tensorflow/master/tensorflow/lite/schema/schema.fbs). 55 56#### Interpreter for generating reference output 57Python package tensorflow is always needed however the script has the option to use other interpreters than the tensorflow default, which will generate the actual reference output. 58 59##### tflite_runtime 60Python package tflite_runtime can be installed with pip and it can also be built locally. Check this [link](https://www.tensorflow.org/lite/guide/build_cmake_pip) on how to do that. 61Use the -h flag to get more info on supported interpreters. 62 63##### tflite_micro 64Python package tflite_micro can be installed with pip and it can also be built locally. See this comment for more info: https://github.com/tensorflow/tflite-micro/issues/1484#issuecomment-1677842603. This interpreter is only partially supported, see *Tests depending on TFLM interpreter*. 65 66## Getting started 67 68### Using Arm Mbed OS supported hardware 69 70Connect any HW (e.g. NUCLEO_F746ZG) that is supported by Arm Mbed OS. Multiple boards are supported. Note that board must be mounted, which may or may not be done automatically depending on OS. If all requirements are satisfied you can just run: 71 72``` 73./unittest_targets.py 74``` 75 76Use the -h flag to get more info. 77 78### Using FVP based on Arm Corstone-300 software 79 80The build for unit tests differs from the build of CMSIS-NN as a [standalone library](https://github.com/ARM-software/CMSIS-NN/blob/main/README.md#building-cmsis-nn-as-a-library) in that, there is a dependency to [CMSIS](https://github.com/ARM-software/CMSIS_5) project for the startup files from CMSIS-Core. This is specified by the mandatory CMSIS_PATH CMake argument. 81 82 83Here is an example for testing on Arm Cortex-M55: 84 85``` 86cd </path/to/CMSIS_NN/Tests/Unittest> 87mkdir build 88cd build 89cmake .. -DCMAKE_TOOLCHAIN_FILE=</path/to/ethos-u-core-platform>/cmake/toolchain/arm-none-eabi-gcc.cmake -DTARGET_CPU=cortex-m55 -DCMSIS_PATH=</path/to/CMSIS> 90make 91``` 92 93This will build all unit tests. You can also just build a specific unit test only, for example: 94 95``` 96make test_arm_depthwise_conv_s8_opt 97``` 98 99Then you need to download and install the FVP based Arm Corstone-300 software, for example: 100 101``` 102mkdir -p /home/$user/FVP 103wget https://developer.arm.com/-/media/Arm%20Developer%20Community/Downloads/OSS/FVP/Corstone-300/FVP_Corstone_SSE-300_Ethos-U55_11.12_57.tgz 104tar -xvzf FVP_Corstone_SSE-300_Ethos-U55_11.12_57.tgz 105./FVP_Corstone_SSE-300_Ethos-U55.sh --i-agree-to-the-contained-eula --no-interactive -d /home/$user/FVP 106export PATH="/home/$user/FVP/models/Linux64_GCC-6.4:$PATH" 107``` 108 109Finally you can run the unit tests. For example: 110 111``` 112FVP_Corstone_SSE-300_Ethos-U55 --cpulimit 2 -C mps3_board.visualisation.disable-visualisation=1 -C mps3_board.telnetterminal0.start_telnet=0 -C mps3_board.uart0.out_file="-" -C mps3_board.uart0.unbuffered_output=1 ./TestCases/test_arm_depthwise_conv_s8_opt/test_arm_depthwise_conv_s8_opt.elf 113``` 114 115## Generating new test data 116**NOTE:** The data generation scrips are being reworked, see *Refactoring of generate_test_data* 117 118Generating new test data is done with the following script. Use the -h flag to get more info. 119 120``` 121./generate_test_data.py -h 122 123``` 124 125The script use a concept of test data sets, i.e. it need a test set data name as input. It will then generate files with that name as prefix. Multiple header files of different test sets can then be included in the actual unit test files. 126When adding a new test data set, new c files should be added or existing c files should be updated to use the new data set. See overview of the folders on how/where to add new c files. 127 128The steps to add a new unit test are as follows. Add a new test test in the load_all_testdatasets() function. Run the generate script with that new test set as input. Add the new generated header files to an existing or new unit test. 129 130### Tests depending on TFLM interpreter 131 132If TFL and TFLM reference kernels differ, CMSIS-NN aims to be bit-exact to TFLM reference kernels. Hence those operators depends on tflite_micro interpreter. 133 134Operator bit-exactness compability: 135 136| Operator | TFL bit-exact | TFLM bit-exact | Notes 137| --- | --- | --- | --- 138| convolution | x | x | 139| fully_connected | x | x | 140| lstm | | x | 141| svdf | | x | Operator is only fully supported by TFLM. 142| softmax | x | x | 143| avgpool | x | x | 144| maxpool | x | x | 145| add | x | x | 146| mul | x | x | 147 148### Refactoring of generate_test_data.py 149Test data generation is in progress of incrementally moving over to the cleaned up scripts placed in `RefactoredTestGen`. 150 151To try out the new scripts, use 152``` 153./RefactoredTestGen/generate_test_data.py --help 154``` 155 156The previous generate_test_data will remain as the main data generator until all functionality is replicated with the new scripts. 157 158Current progress: 159 160| Operator | Old | New | Notes 161| --- | --- | --- | --- 162| convolution | x | x | New version only supports 16x8 163| fully_connected | x | x | New version only supports int4 packed weights 164| lstm | x | x | Only new version supporting 16x8 165| svdf | x | | 166| softmax | x | | 167| avgpool | x | | 168| maxpool | x | | 169| add | x | | 170| mul | x | | 171 172 173## Overview of the Folders 174 175- `Corstone-300` - These are dependencies, like linker files etc, needed when building binaries targeting the FVP based on Arm Corstone-300 software. This is mostly taken from Arm Ethos-U Core Platform project. 176- `Mbed` - These are the Arm Mbed OS settings that are used. See Mbed/README.md. 177- `Output` - This will be created when building. 178- `PregeneratedData` - Host local(Not part of GitHub) test data for model creation using Keras in unit tests. It can be used for debug purposes when 179 adding new operators or debugging existing ones. 180- `TestCases` - Here are the actual unit tests. For each function under test there is a folder under here. 181- `TestCases/<cmsis-nn function name>` - For each function under test there is a folder with the same name with test_ prepended to the name and it contains a c-file with the actual unit tests. For example for arm_convolve_s8() the file is called test_arm_convolve_s8.c 182- `TestCases/<cmsis-nn function name>/Unity` - This folder contains a Unity file that calls the actual unit tests. For example for arm_convolve_s8() the file is called unity_test_arm_convolve_s8.c. 183- `TestCases/<cmsis-nn function name>/Unity/TestRunner` - This folder will contain the autogenerated Unity test runner. 184- `TestCases/TestData` - This is auto generated test data in .h files that the unit tests are using. The data in PregenrateData folder has fp32 data of a network whereas here it is the quantized equivalent of the same. They are not the same. All data can regenerated or only parts of it (e.g. only bias data). Of course even the config can be regenerated. This partial/full regeneration is useful during debugging. 185- `TestCases/Common` - Common files used in test data generation is placed here. 186- `RefactoredTestGen` - Temporary location for new test generation scripts 187 188## Formatting 189 190The python test scripts should be formatted with yapf. 191 192``` 193pip install --upgrade yapf 194``` 195 196The following settings are used. 197 198``` 199python -m yapf --in-place --style='{based_on_style:pep8,column_limit:120,indent_width:4}' *.py 200``` 201