performance data about ruy HOT 3 CLOSED

This spreadsheet has such data. The '8bit' sheets are comparing against gemmlowp.
https://docs.google.com/spreadsheets/d/1CB4gsI7pujNRAf5Iz5vuD783QQqO2zOu8up9IpTKdlU/edit?usp=sharing

from ruy.

This spreadsheet has such data. The '8bit' sheets are comparing against gemmlowp.
https://docs.google.com/spreadsheets/d/1CB4gsI7pujNRAf5Iz5vuD783QQqO2zOu8up9IpTKdlU/edit?usp=sharing

thanks!!
I think we will use ruy replace gemmlowp.
But another question : ( I build from source use cmake and found it generate a lot of static libs. Such as libruy_kernel_avx512.a libruy_kernel_avx.a libruy_frontend.a.
If I use ruy on arm platform, how can I only build arm related libs.

from ruy.

The architecture-specific static libraries like libruy_kernel_avx512.a are nearly empty when the actual build architecture is not the one that they are relevant to (e.g. x86 libraries are empty on ARM). They will only contain empty functions, or, when assertions are enabled, they will only contain a false-assertion ensuring that these functions are never called:

Note, anyway, that x86-specific code simply wouldn't compile when the target architecture is ARM.

So you should be able to ignore all this as an internal implementation detail of ruy. You can always assess what symbols are in your final binary by using the nm command from your ARM toolchain, e.g.

$HOME/android-ndk-r21d/toolchains/llvm/prebuilt/linux-x86_64/bin/aarch64-linux-android-nm --radix=d -S -C $HOME/ruy-build/ruy/example/parametrized_example | grep 'ruy::' | awk '{sum+=$2} END {print sum}'

(parametrized_example in the example directory is a simple way to assess overall ruy code size as it instantiates many different ruy::Mul template instantiations - actual applications would see smaller ruy code because they instantiate fewer).

from ruy.