pytorch / QNNPACK

Quantized Neural Network PACKage - mobile-optimized implementation of quantized neural network operators
https://code.fb.com/ml-applications/qnnpack/
Other
1.52k stars 219 forks source link

No performance difference between ARMv7 (32-bit) Android and ARM64 (64-bit) Android #47

Closed yangwenca closed 5 years ago

yangwenca commented 5 years ago

I follow the tutorial to run speed benchmark (MobileNet v2) on ARMv7 (32-bit) and ARM64 (64-bit) Android. There is no performance difference between them.

Data from a few Android devices

device per iter (ms) iters per second
oneplus_5_32 17.27 57.92
oneplus_5_64 15.3 65.36
htc_10_32 35.56 28.12
htc_10_64 37.63 26.57
huawei_caml21_32 52.34 19.14
huawei_caml21_64 49.89 20.05
Maratyszcza commented 5 years ago

This is expected. On well-optimized code there is only a small difference between 32-bit and 64-bit ARM