1Frequently asked questions 2========================== 3 4Problems seen when trying to build armnn and ComputeLibrary obtained from GitHub 5----------------------------------------------------------------------------- 6 7Some users have encountered difficulties when attempting to build armnn and ComputeLibrary obtained from GitHub. The build generally fails reporting missing dependencies or fields in aclCommon, backendsCommon, cl or neon. These errors can look like this: 8 9error: ‘HARD_SWISH’ is not a member of ‘AclActivationFunction {aka arm_compute::ActivationLayerInfo::ActivationFunction}’ 10 11The most common reason for these errors are a mismatch between armnn and clframework revisions. For any version of Arm NN the coresponding version of ComputeLibrary is detailed in scripts/get_compute_library.sh as DEFAULT_CLFRAMEWORKREVISION 12 13On *nix like systems running this script will checkout ComputeLibrary, with the current default SHA, into ../../clframework/ relative to the location of the script. 14 15Segmentation fault following a failed call to armnn::Optimize using CpuRef backend. 16--------------------------------------------------------- 17 18In some error scenarios of calls to armnn::Optimize a null pointer may be returned. This contravenes the function documentation however, it can happen. Users are advised to check the value returned from the function as a precaution. 19 20If you encounter this problem and are able to isolate it consider contributing a solution. 21 22Adding or removing -Dxxx options when building Arm NN does not always work. 23--------------------------------------------------------- 24 25Arm NN uses CMake for build configuration. CMake uses a cumulative cache of user options. That is, setting a value once on a cmake command line will be persisted until either you explicitly change the value or delete the cache. To delete the cache in Arm NN you must delete the build directory. 26 27Many DynamicBackendTests fail with "Base path for shared objects does not exist". 28--------------------------------------------------------- 29This problem most commonly occurs when the compile and runtime environments for the unit tests differ. These dynamic backend tests rely on a set of test files and directories at runtime. These files are created by default during the cmake build. At runtime the tests will look for these files in src/backends/backendsCommon/test/ relative to where the Unittests executable was built. The usual solution to to copy these files and directories into the new unit test execution environment. You can also specify a new root path for these files by adding a command line parameter to the Unittests executable: Unittests -- --dynamic-backend-build-dir "new path" 30 31 32Tensorflow Lite benchmarking tool fails with segmentation fault when using the Arm NN delegate. 33--------------------------------------------------------- 34There are occaisional problems using native build versions of the Tensorflow Lite benchmarking tool. It can be sensitive to errors in command line parameter usage. A simple misspelling of a delegate name will result in a bus error. Here is a sample command line usage that is known to work for the Arm NN delegate. 35 36This example is for: 37 38* Execution on Android using a native binary downloaded from [Tensorflow Lite performance measurment](https://www.tensorflow.org/lite/performance/measurement#native_benchmark_binary). 39* Uses a TF Lite model that has been downloaded from the ML-zoo. In this case [MobileNet v2 1.0 224 UINT8](https://github.com/ARM-software/ML-zoo/tree/master/models/image_classification/mobilenet_v2_1.0_224/tflite_uint8). 40* Arm NN and its dependent libraries are in the current directory, /data/local/tmp. 41 42~~~ 43LD_LIBRARY_PATH=/vendor/lib64/egl:/vendor/lib/egl/:. ./android_aarch64_benchmark_model --num_threads=4 --graph=/data/local/tmp/mobilenet_v2_1.0_224_quantized_1_default_1.tflite --external_delegate_path="libarmnnDelegate.so" --external_delegate_options="backends:GpuAcc" 44~~~ 45 46Arm NN fails to build intermittently on 18.04 Ubuntu 47--------------------------------------------------------- 48Building Arm NN fails intermittently with error: 49 50c++: internal compiler error: Killed (program cc1plus) 51 52This errors appears to be related to the number of cmake jobs used to build Arm NN.Try limiting the jobs to 2 by modifying the make to: 53 54make -j2 55 56Arm NN UnitTests fails intermittently with segmentation fault on aarch64. 57---------------------------------------------------------- 58The DefaultAsyncExeuteWithThreads test seems to be throwing intermittent segmentation fault while running Arm NN Unittest in aarch64 architecture. This test will pass if you run the Unittest again. 59 60Arm NN delegate build fails with "undefined reference to `absl::lts_20220623::raw_logging_internal::RawLog" 61---------------------------------------------------------- 62This build failure occurs because Tensorflow 2.10 has been built with GCC version older than 9.3.1. The solution is to rebuild with 9.3.1 or later. 63 64 65How can I run a pretrained model using Arm NN?. 66----------------------------------------------- 67The easiest way is to use Arm NN's tool ExecuteNetwork, the source code for this tool is located in armnn/tests/ExecuteNetwork/. This tool is highly configurable allowing multiple options to specify different parameters and it has profiling capabilities which can be enabled by specifying the option --event-based-profiling. 68 69An example of running a model with ExecuteNetwork: 70~~~ 71LD_LIBRARY_PATH=.:$LD_LIBRARY_PATH ./ExecuteNetwork -f tflite-binary -m ./my_test_model.tflite -c CpuAcc,CpuRef 72~~~ 73In the command above we specify two backends, if the an operator required by the model is not supported in the first backend CpuAcc Arm NN will try to use the second backend CpuRef. 74 75Arm NN Multithreading support. 76------------------------------ 77Running multiple inferences in multiple threads concurrently is not supported, but concurrent inferences can be executed in multiple processes and a simple way to do this is by using Arm NN's ExecuteNetwork. 78ArmNN supports multithreading at kernel level and this is implemented in Arm Compute Library (ACL) (https://github.com/ARM-software/ComputeLibrary/). 79During inference, at the operator level, the main thread will create multiple threads and execute the same kernel on different parts of the data. At runtime ACL will detect the number of CPU cores in the system and use one thread per cpu core for each kernel. 80Multithreading at operator level is not supported due to limitations in ACL, for more information please refer to https://arm-software.github.io/ComputeLibrary/latest/architecture.xhtml#architecture_thread_safety 81