search

ermig1979 / Simd

C++ image processing and machine learning library with using of SIMD: SSE, AVX, AVX-512, AMX for x86/x64, VMX(Altivec) and VSX(Power7) for PowerPC, NEON for ARM.
http://ermig1979.github.io/Simd
MIT License
2.04k stars 407 forks source link

Int8/Uint8 convolution support #79

Closed christ115 closed 5 years ago

christ115 commented 5 years ago

Just for your information.

In the near two years, model quantization is becoming popular in deep learning, especially for mobile devices. The core of quantization is low-bits (usually in 8-bits) convolution, which can reduce the memory bandwidth and run-time of deep network significantly. Many deep learning frameworks have implemented related features, e.g.,

Do you have plan to add this feature into SimdLib?

ermig1979 commented 5 years ago

Recently I spend main time to optimize the inference of neural network on the base of Synet and Simd Library. Of course I know about model quantization and have corresponding plans. But these are medium term plans. In the nearest future I'm going to add support of NHWC data format (native data format for Tensorflow) to Synet and optimizations for ARM NEON.