Question about fp32/int8

ermig1979 / Synet

A small framework to infer neural network

MIT License

137 stars 26 forks source link

Closed teor292 closed 3 years ago

teor292 commented 3 years ago

Hi. We run converted from OpenVino framework 'face-detection-retail-0005' network and get next inference results on aarch64:

For fp32: 243,4ms For int8: 856ms.

It is strange that for int8 the calculation is slower. Is that how it should be?

ermig1979 commented 3 years ago

Unfortunately INT8 fused multiply add operation is not good implemented on ARM platform. So it is a feature but not a bug.

teor292 commented 3 years ago

Ok, thanks.