Lite Architecture Benchmark

Benchmark v0.1

mobile net V1
	1 thread	2 thread	4 thread
Qualcomm 435	243.05 ms	123.647 ms	84.673 ms
Qualcomm 625	183.97 ms	94.495 ms	54.256 ms
Qualcomm 835	91.071 ms	48.809 ms	27.581 ms
Qualcomm 845	64.3953 ms	33.816 ms	20.131 ms
MTK Helio P60	237.793 ms	121.988 ms	68.516 ms
arm linux(rk3399)	120.819 ms	65.0624 ms	--

mobile net v2
	1 thread	2 thread	4 thread
Qualcomm 435	175.27 ms	89.214 ms	61.344 ms
Qualcomm 625	137.464 ms	73.499 ms	46.679 ms
Qualcomm 835	60.135 ms	38.197 ms	21.134 ms
Qualcomm 845	42.81 ms	23.581 ms	14.114 ms
MTK Helio P60	184.172 ms	97.292 ms	56.531 ms
arm linux(rk3399)	95.654 ms	51.881 ms	--

inception v4
	1 thread	2 thread	4 thread
Qualcomm 435	2545.87 ms	1325.26 ms	925.012 ms
Qualcomm 625	2021.72 ms	1042.22 ms	614.647 ms
Qualcomm 835	895.008 ms	527.226 ms	326.85 ms
Qualcomm 845	647.563 ms	390.266 ms	259.983 ms
MTK Helio P60	2393.99 ms	1284.46 ms	745.351 ms
arm linux(rk3399)	1141.35 ms	659.916 ms	--

Resnet50
	1 thread	2 thread	4 thread
Qualcomm 435	1779.96 ms	907.695 ms	594.149 ms
Qualcomm 625	1401.76 ms	714.77 ms	402.68 ms
Qualcomm 835	656.405 ms	374.671 ms	217.913 ms
Qualcomm 845	454.95 ms	254.511 ms	170.833 ms
MTK Helio P60	1696.32 ms	886 ms	485.072 ms
arm linux(rk3399)	834.659 ms	461.797 ms	--

In order to reproduce the benchmark test results, follow these steps:

1. Checkout the latest paddle lite mobile code.

2. Go to paddle lite source directory and build docker image

cd <paddle-lite-repo>
docker build --file paddle/fluid/lite/tools/Dockerfile.mobile --tag paddle-lite-mobile:latest .

3. Start the docker container. This will start the newly created docker image, and mount the paddle lite source directory to /paddle-lite path.

docker run -it --name <your_container_name> --net=host --privileged -v $(pwd):/paddle-lite paddle-lite-mobile bash

4. Paddle lite makes use of cmake to build benchmark program. In docker shell, go to paddle lite source directory and create a build directory.

cd /paddle-lite && mkdir build.armv8 && cd build.armv8

5. Run cmake command:

this will create the make file for armv8

cmake .. \
    -DWITH_GPU=OFF \
    -DWITH_LITE=ON \
    -DLITE_WITH_CUDA=OFF \
    -DLITE_WITH_X86=OFF \
    -DLITE_WITH_ARM=ON \
    -DLITE_WITH_LIGHT_WEIGHT_FRAMEWORK=ON \
    -DWITH_TESTING=ON \
    -DWITH_MKL=OFF \
-DARM_TARGET_OS="android" -DARM_TARGET_ARCH_ABI="arm64-v8a"

6. Run make command:

make -j 4

7. The built benchmark binary is <paddle-lite-mobile-source-root>/build.armv8/paddle/fluid/lite/api/test_model_bin. Use android adb to push the binary into your android phone. After connecting your phone to your computer, please execute the following command outside docker shell in paddle lite source directory:

adb push ./build.armv8/paddle/fluid/lite/api/test_model_bin /data/local/tmp/
adb shell chmod +x /data/local/tmp/test_model_bin

8. Now push a test model into your phone. Contemporarily we support 4 models:

model name	model
mobilenet v1	http://paddle-inference-dist.bj.bcebos.com/mobilenet_v1.tar.gz
mobilenet v2	http://paddle-inference-dist.bj.bcebos.com/mobilenet_v2_relu.tar.gz
inception v4	http://paddle-inference-dist.bj.bcebos.com/inception_v4.tar.gz
resnet50	http://paddle-inference-dist.bj.bcebos.com/resnet50.tar.gz

download the models you need. Here I take mobilenet v1 for an example, please extract and push it into your phone.

tar xzvf mobilenet_v1.tar.gz
adb push mobilenet_v1 /data/local/tmp/mobilenet_v1

9. Start the benchmark program

adb shell /data/local/tmp/test_model_bin /data/local/tmp/mobilenet_v1 50 1