Memory bandwidth benchmark

Build

Run these commands in terminal:

It is going to create a directory named membw where you are

git clone https://github.com/doug65536/membw.git
cd membw
make

To run it

make run

(OPTIONAL) To produce a windows executable:

sudo apt install mingw-w64-common mingw-w64-tools binutils-mingw-w64 g++-mingw-w64-x86-64

(OPTIONAL) To produce an aarch64/RPI4 executable:

sudo apt install gcc-aarch64-linux-gnu \
    g++-aarch64-linux-gnu binutils-aarch64-linux-gnu

(OPTIONAL) Produce membw.exe

make build-windows

(OPTIONAL) Produce a specific platform's executable

make build-arm
make build-linux
make build-windows

(OPTIONAL) Try a horrible compile with no vectors to see how bad it is:

make VECFLAGS=-mno-sse -B run

Example

You can build for linux, windows, x86_64 or arm (NEON). It probably builds on almost any architecture, but it will use horrible scalar instructions for measurement. It really does make a big difference.

This code allows the CPU to fully take advantage of the prefetcher, the memory access pattern is very predictable and the memory controller has a lot of information to work with.