In recent years, the on-device deep learning applications are getting more and more popular. It's a challenging task for application developers to deploy their deep learning models in their applications. They need to choose a proper inference framework, optionally utilizing quantization or compression techniques regarding the precision-performance trade-off, and finally run the model on one or more of heterogeneous computing devices. How to make an appropriate decision among these choices is a tedious and time-consuming task.
Mobile AI Benchmark (i.e. MobileAIBench) is an end-to-end benchmark tool to test the models' runtime in the different neural network frameworks on mobile devices, which hopefully can provide insights for the developers.
Please check benchmark step in daily CI pipeline page.
Q: Why are benchmark results not stable on my device?
A: Due to power save considerations, some SoCs have aggressive and advanced
power control scheduling to reduce power consumption which make performance
quite unstable (especially CPU runtime). Benchmark results highly depend on
states of devices, e.g., running processes, temperature, power control policy.
It is recommended to disable power control policy (as shown in tools/power.sh) if possible (e.g., rooted phone).
Otherwise, keep your device at idle state with low temperature, and benchmark one model on one framework each time.
Q: Why do some devices run faster (or slower) than expected in the CI benchmark result?
A: Some devices is rooted and has some specialized performance tuning while some others is not rooted and failed to make such tuning (see the code for more details).
Q: Why is ncnn initialization time much less than others?
A: ncnn benchmark uses fake model parameters and skips loading weights from filesystem.
Q: Does benchmark use all available cores of devices?
A: Most modern Android phones use ARM big.LITTLE architecture which can lead to significant variance between different runs of the benchmark, we use only available big cores to reduce this variance by taskset command for MACE/NCNN/TFLITE benchmark.
Moreover, there are no well-defined APIs for SNPE to bind to big cores and set thread count.
Thread count can be set by adding --num_threads to benchmark.py command.
MobileAIBench supports several deep learning frameworks (MACE, SNPE, ncnn and TensorFlow Lite) currently, which may require the following dependencies:
| Software | Installation command | Tested version |
|---|---|---|
| Python | 2.7 | |
| ADB | apt-get install android-tools-adb | Required by Android run, >= 1.0.32 |
| Android NDK | NDK installation guide | Required by Android build, r15c |
| Bazel | bazel installation guide | 0.13.0 |
| CMake | apt-get install cmake | >= 3.11.3 |
| FileLock | pip install -I filelock==3.0.0 | Required by Android run |
| PyYaml | pip install -I pyyaml==3.12 | 3.12.0 |
| sh | pip install -I sh==1.12.14 | 1.12.14 |
| SNPE (optional) | download and uncompress | 1.18.0 |
Note: SNPE
has strict license that disallows redistribution, so the default link in the
Bazel WORKSPACE file is only accessible by the CI server. To benchmark SNPE
in your local system (i.e. set --frameworks with all or SNPE explicitly),
you need to download the SDK here,
uncompress it, copy libgnustl_shared.so
and modify WORKSPACE as the following:
#new_http_archive(
# name = "snpe",
# build_file = "third_party/snpe/snpe.BUILD",
# sha256 = "820dda1eaa5d36f7548fc803122c2c119f669a905cca03349f0480d023f7ed17",
# strip_prefix = "snpe-1.18.0",
# urls = [
# "https://cnbj1-fds.api.xiaomi.net/aibench/third_party/snpe-1.18.0.zip",
# ],
#)
new_local_repository(
name = "snpe",
build_file = "third_party/snpe/snpe.BUILD",
path = "/path/to/snpe-1.18.0",
)+-----------------+ +------------------+ +---------------+
| Benchmark | | BaseExecutor | <--- | MaceExecutor |
+-----------------+ +------------------+ +---------------+
| - executor |-------> | - framework |
| - model_name | | - runtime | +---------------+
| - model_file | | | <--- | SnpeExecutor |
| - input_names | +------------------+ +---------------+
| - input_files | | + Init() |
| - input_shapes | | + Prepare() | +---------------+
| - output_names | | + Run() | <--- | NcnnExecutor |
| - output_shapes | | + Finish() | +---------------+
+-----------------+ +------------------+
| - Register() | +---------------+
| - Run() | <--- | TfLiteExecutor|
+-----------------+ +---------------+
python tools/benchmark.py --output_dir=output --frameworks=all \
--runtimes=all --model_names=all \
--target_abis=armeabi-v7a,arm64-v8a
The whole benchmark may take a few time, and continuous benchmarking may heat the device very quickly, so you may set the following arguments according to your interests.
| option | type | default | explanation |
|---|---|---|---|
| --output_dir | str | output | Benchmark output directory. |
| --frameworks | str | all | Frameworks(MACE/SNPE/NCNN/TFLITE), comma separated list or all. |
| --runtimes | str | all | Runtimes(CPU/GPU/DSP), comma separated list or all. |
| --target_abis | str | armeabi-v7a | Target ABIs(armeabi-v7a,arm64-v8a), comma separated list. |
| --model_names | str | all | Model names(InceptionV3,MobileNetV1...), comma separated list or all. |
| --run_interval | int | 10 | Run interval between benchmarks, seconds. |
| --num_threads | int | 4 | The number of threads. |
-
Register model benchmark
Register benchmark in
aibench/benchmark/benchmark_main.cc:#ifdef AIBENCH_ENABLE_YOUR_FRAMEWORK std::unique_ptr<aibench::YourFrameworkExecutor> your_framework_executor(new aibench::YourFrameworkExecutor()); AIBENCH_BENCHMARK(your_framework_executor.get(), MODEL_NAME, FRAMEWORK_NAME, RUNTIME, MODEL_FILE, (std::vector<std::string>{INPUT_NAME}), (std::vector<std::string>{INPUT_FILE}), (std::vector<std::vector<int64_t>>{INPUT_SHAPE}), (std::vector<std::string>{OUTPUT_NAME}), (std::vector<std::vector<int64_t>>{OUTPUT_SHAPE})); #endif
e.g.
AIBENCH_BENCHMARK(mobilenetv1_mace_cpu_executor.get(), MobileNetV1, MACE, CPU, mobilenet_v1, (std::vector<std::string>{"input"}), (std::vector<std::string>{"dog.npy"}), (std::vector<std::vector<int64_t>>{{1, 224, 224, 3}}), (std::vector<std::string>{ "MobilenetV1/Predictions/Reshape_1"}), (std::vector<std::vector<int64_t>>{{1, 1001}}));
-
Register model in
tools/model_list.py. -
Configure model file and input file
Configure
MODEL_FILEandINPUT_FILEintools/model_and_input.yml. -
Run benchmark
python tools/benchmark.py --output_dir=output --frameworks=MACE \ --runtimes=CPU --model_names=MobileNetV1 \ --target_abis=armeabi-v7a,arm64-v8a -
Check benchmark result
cat output/report.csv
-
Define
executorand implement the interfaces:class YourFrameworkExecutor : public BaseExecutor { public: YourFrameworkExecutor() : BaseExecutor(FRAMEWORK_NAME, RUNTIME) {} // Init method should invoke the initializing process for your framework // (e.g. Mace needs to compile OpenCL kernel once per target). It will be // called only once when creating framework engine. virtual Status Init(const char *model_name, int num_threads); // Load model and prepare to run. It will be called only once before // benchmarking the model. virtual Status Prepare(const char *model_name); // Run the model. It will be called more than once. virtual Status Run(const std::map<std::string, BaseTensor> &inputs, std::map<std::string, BaseTensor> *outputs); // Unload model and free the memory after benchmarking. It will be called // only once. virtual void Finish(); };
-
Include your framework header in
aibench/benchmark/benchmark_main.cc:#ifdef AIBENCH_ENABLE_YOUR_FRAMEWORK #include "aibench/executors/your_framework/your_framework_executor.h" #endif
-
Add dependencies to
third_party/your_framework,aibench/benchmark/BUILDandWORKSPACE. Put macroAIBENCH_ENABLE_YOUR_FRAMEWORKintoaibench/benchmark/BUILDatmodel_benchmarktarget. -
Benchmark a model on existing framework
Refer to [Adding a model to run on existing framework](#Adding a model to run on existing framework).
For third party dependencies, please refer to their licenses.