性能数据

大家可以参考测试方法文档对模型进行测试。

ARM测试环境

  • 测试模型

    • fp32模型

      • mobilenet_v1

      • mobilenet_v2

      • squeezenet_v1.1

      • mnasnet

      • shufflenet_v1

    • int8模型

      • mobilenet_v1

      • mobilenet_v2

  • 测试机器(android ndk ndk-r17c)

    • 骁龙855

      • xiaomi mi9, snapdragon 855 (enable sdot instruction)

      • 4xA76(1@2.84GHz + 3@2.4GHz) + 4xA55@1.78GHz

    • 骁龙845

      • xiaomi mi8, 845

      • 2.8GHz(大四核),1.7GHz(小四核)

    • 骁龙835

      • xiaomi mix2, snapdragon 835

      • 2.45GHz(大四核),1.9GHz(小四核)

  • 测试说明

    • branch: release/v2.8

    • warmup=10, repeats=100,统计平均时间,单位是ms

    • 当线程数为1时,DeviceInfo::Global().SetRunMode设置LITE_POWER_HIGH,否者设置LITE_POWER_NO_BIND

    • 模型的输入图像的维度是{1, 3, 224, 224},输入图像的每一位数值是1

ARM测试数据

fp32模型测试数据

paddlepaddle model

骁龙855armv7armv7armv7armv8armv8armv8
threads num124124
mobilenet_v134.2119.7811.5329.9317.3410.04
mobilenet_v223.5914.078.4721.3012.897.81
shufflenet_v14.092.882.043.962.672.08
squeezenet_v1.118.9812.508.1816.6311.497.48
mnasnet29.4712.757.2622.9211.856.71
骁龙845armv7armv7armv7armv8armv8armv8
threads num124124
mobilenet_v164.2636.7118.3262.1932.0816.89
mobilenet_v243.2824.4813.6940.3122.4312.72
shufflenet_v17.394.563.187.184.633.24
squeezenet_v1.135.2122.3812.9132.7120.4112.07
mnasnet38.3326.2612.2137.4220.6111.57
骁龙835armv7armv7armv7armv8armv8armv8
threads num124124
mobilenet_v191.6350.3629.9486.8646.3926.43
mobilenet_v262.335.2922.0157.6432.8319.25
shufflenet_v19.815.994.199.205.774.05
squeezenet_v1.151.2232.7019.8647.2330.5918.11
mnasnet57.1732.6019.6753.7430.0217.74

caffe model

骁龙855armv7armv7armv7armv8armv8armv8
threads num124124
mobilenet_v132.2318.6010.6130.9418.199.94
mobilenet_v229.8917.4610.8127.0316.309.73
shufflenet_v14.862.942.103.892.822.11
骁龙845armv7armv7armv7armv8armv8armv8
threads num124124
mobilenet_v165.2035.1118.9261.2532.1517.32
mobilenet_v255.5330.8317.5651.6228.9215.95
shufflenet_v17.384.553.197.164.353.30
骁龙835armv7armv7armv7armv8armv8armv8
threads num124124
mobilenet_v192.3150.9430.7287.4746.4126.19
mobilenet_v281.3245.1028.1275.5742.4725.71
shufflenet_v29.915.984.209.595.764.06

int8量化模型测试数据

骁龙855armv7armv7armv7armv8armv8armv8
threads num124124
mobilenet_v121.2510.885.4313.197.663.95
mobilenet_v216.9910.235.6812.637.594.34
骁龙845armv7armv7armv7armv8armv8armv8
threads num124124
mobilenet_v151.4328.1414.3745.1733.1212.60
mobilenet_v238.9821.6411.8033.1218.4410.02
骁龙835armv7armv7armv7armv8armv8armv8
threads num124124
mobilenet_v161.9132.7516.6057.4630.0315.37
mobilenet_v248.8726.1513.7442.6122.6311.79

华为麒麟NPU的性能数据

请参考PaddleLite使用华为麒麟NPU预测部署的最新性能数据

瑞芯微NPU的性能数据

请参考PaddleLite使用瑞芯微NPU预测部署的最新性能数据

联发科APU的性能数据

请参考PaddleLite使用联发科APU预测部署的最新性能数据

颖脉NNA的性能数据

请参考PaddleLite使用颖脉NNA预测部署的最新性能数据