search

ENCP / CNNdroid

Open Source Library for GPU-Accelerated Execution of Trained Deep Convolutional Neural Networks on Android
MIT License
539 stars 181 forks source link

what 's the time cost in your device?(CaffeNet demo) #31

Open wangkaizjuzju opened 7 years ago

wangkaizjuzju commented 7 years ago

i test the CaffeNet demo in my phone(samsung Galaxy s8 ), it make take 400ms per image. Is it abnormal? If the test result is right , i don't think this CNNdroid is fast enough. Thank you very much.

akbarxie commented 6 years ago

if you run the src code it may be not very fast,but if you can chang the code according to your models,it can be twice or tripple faster than the src code . 1、remove the data copy from the GPU memory to the CPU 2、give up all if and else code in the kernel code