/mobile-ai-bench

Benchmarking Neural Network Inference on Mobile Devices

Primary LanguageC++Apache License 2.0Apache-2.0

Mobile AI Bench

License pipeline status

FAQ | 中文

In recent years, the on-device deep learning applications are getting more and more popular on mobile phones or IoT devices. It's a challenging task for the developers to deploy their deep learning models in their mobile applications or IoT devices.

They need to optionally choose a cost-effective hardware solution (i.e. chips and boards), then 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 which covers different chips and inference frameworks, with results include both speed and model accuracy, which will give insights for developers.

Daily Benchmark Results

Please check benchmark step in daily CI pipeline page, due to the lack of test devices, the CI result may not cover all hardwares and frameworks.

FAQ

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). 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 tools/benchmark.sh command.

Environment requirement

MobileAIBench supports several deep learning frameworks (called executor in this project, i.e., MACE, SNPE, ncnn, TensorFlow Lite and HIAI) 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 1: 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 --executors 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 = "8f2b92b236aa7492e4acd217a96259b0ddc1a656cbc3201c7d1c843e1f957e77",
#    strip_prefix = "snpe-1.22.2.233",
#    urls = [
#        "https://cnbj1-fds.api.xiaomi.net/aibench/third_party/snpe-1.22.2_with_libgnustl_shared.so.zip",
#    ],
#)

new_local_repository(
    name = "snpe",
    build_file = "third_party/snpe/snpe.BUILD",
    path = "/path/to/snpe",
)

Note 2: HIAI has strict license that disallows redistribution, so the default link in the Bazel WORKSPACE file is only accessible by the CI server. To benchmark HIAI in your local system (i.e. set --executors with all or HIAI explicitly), you need to login and download the SDK here, uncompress it and get the HiAI_DDK_100.200.010.011.zip file, uncompress it and modify WORKSPACE as the following:

#new_http_archive(
#    name = "hiai",
#    build_file = "third_party/hiai/hiai.BUILD",
#    sha256 = "8da8305617573bc495df8f4509fcb1655ffb073d790d9c0b6ca32ba4a4e41055",
#    strip_prefix = "HiAI_DDK_100.200.010.011",
#    type = "zip",
#    urls = [
#        "http://cnbj1.fds.api.xiaomi.com/aibench/third_party/HiAI_DDK_100.200.010.011_LITE.zip",
#    ],
#)

new_local_repository(
    name = "hiai",
    build_file = "third_party/hiai/hiai.BUILD",
    path = "/path/to/hiai",
)

Architecture

+-----------------+         +------------------+      +---------------+
|   Benchmark     |         |   BaseExecutor   | <--- | MaceExecutor  |
+-----------------+         +------------------+      +---------------+
| - executor      |-------> | - executor       |
| - model_name    |         | - device_type    |      +---------------+
| - quantize      |         |                  | <--- | SnpeExecutor  |
| - input_names   |         +------------------+      +---------------+
| - input_shapes  |         | + Init()         |
| - output_names  |         | + Prepare()      |      +---------------+
| - output_shapes |         | + Run()          | <--- | NcnnExecutor  |
| - run_interval  |         | + Finish()       |      +---------------+
| - num_threads   |         |                  |
+-----------------+         |                  |      +---------------+
| - Run()         |         |                  | <--- | TfLiteExecutor|
+-----------------+         |                  |      +---------------+
        ^     ^             |                  |
        |     |             |                  |      +---------------+
        |     |             |                  | <--- | HiaiExecutor  |
        |     |             +------------------+      +---------------+
        |     |
        |     |             +--------------------+
        |     |             |PerformanceBenchmark|
        |     --------------+--------------------+
        |                   | - Run()            |
        |                   +--------------------+
        |
        |                   +---------------+      +---------------------+                           
+--------------------+ ---> |PreProcessor   | <--- |ImageNetPreProcessor |
| PrecisionBenchmark |      +---------------+      +---------------------+
+--------------------+
| - pre_processor    |      +---------------+      +---------------------+
| - post_processor   | ---> |PostProcessor  | <--- |ImageNetPostProcessor|
| - metric_evaluator |      +---------------+      +---------------------+
+--------------------+
| - Run()            |      +---------------+
+--------------------+ ---> |MetricEvaluator|
                            +---------------+

How To Use

Benchmark Performance of all models on all executors

bash tools/benchmark.sh --benchmark_option=Performance \
                        --target_abis=armeabi-v7a,arm64-v8a,aarch64,armhf

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. Only MACE supports precision benchmark right now.

option type default explanation
--benchmark_option str Performance Benchmark options, Performance/Precision.
--output_dir str output Benchmark output directory.
--executors str all Executors(MACE/SNPE/NCNN/TFLITE/HIAI), comma separated list or all.
--device_types str all DeviceTypes(CPU/GPU/DSP/NPU), comma separated list or all.
--target_abis str armeabi-v7a Target ABIs(armeabi-v7a,arm64-v8a,aarch64,armhf), 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.
--input_dir str "" Input data directory for precision benchmark.

Configure ssh devices

For embedded ARM-Linux devices whose abi is aarch64 or armhf, ssh connection is supported. Configure ssh devices in generic-mobile-devices/devices_for_ai_bench.yml, for example:

devices:
  nanopi:
    target_abis: [aarch64, armhf]
    target_socs: RK3333
    models: Nanopi M4
    address: 10.231.46.118
    username: pi

Adding a model to run on existing executor

  • Add the new model name in aibench/proto/base.proto if not in there.

  • Configure the model info in aibench/proto/model.meta.

  • Configure the benchmark info in aibench/proto/benchmark.meta.

  • Run benchmark

    Performance benchmark.

    bash tools/benchmark.sh --benchmark_option=Performance \
                            --executors=MACE --device_types=CPU --model_names=MobileNetV1 \
                            --target_abis=armeabi-v7a,arm64-v8a,aarch64,armhf

    Precision benchmark. Only supports ImageNet images as inputs for benchmarking MACE precision.

    bash tools/benchmark.sh --benchmark_option=Precision --input_dir=/path/to/inputs \
                            --executors=MACE --device_types=CPU --model_names=MobileNetV1 \
                            --target_abis=armeabi-v7a,arm64-v8a,aarch64,armhf
  • Check benchmark result

    python report/csv_to_html.py

    Open the corresponding link in a browser to see the report.

Adding a new AI executor

  • Define executor and implement the interfaces:

    class YourExecutor : public BaseExecutor {
     public:
      YourExecutor() :
          BaseExecutor(executor_type, device_type, model_file, weight_file) {}
      
      // Init method should invoke the initializing process for your executor 
      // (e.g.  Mace needs to compile OpenCL kernel once per target). It will be
      // called only once when creating executor engine.
      virtual Status Init(int num_threads);
    
      // Load model and prepare to run. It will be called only once before 
      // benchmarking the model.
      virtual Status Prepare();
      
      // 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 executor header in aibench/benchmark/benchmark_main.cc:

    #ifdef AIBENCH_ENABLE_YOUR_EXECUTOR
    #include "aibench/executors/your_executor/your_executor.h"
    #endif
  • Add dependencies to third_party/your_executor, aibench/benchmark/BUILD and WORKSPACE. Put macro AIBENCH_ENABLE_YOUR_EXECUTOR into aibench/benchmark/BUILD at model_benchmark target.

  • Benchmark a model on existing executor

    Refer to [Adding a model to run on existing executor](#Adding a model to run on existing executor).

License

Apache License 2.0.

Notice

For third party dependencies, please refer to their licenses.