CPU Dispatching: Make your code both portable and fast (2020)

The article discusses CPU dispatching as a method to enhance software performance while maintaining portability. When a function is critical for performance, compiling it with the `-march=native` flag optimizes it for the local CPU, but this limits its distribution. CPU dispatching allows the binary to detect the CPU features at runtime and execute the appropriate version of the code. The article outlines two approaches: manual CPU dispatching and compiler-assisted dispatching. Manual dispatching involves creating multiple implementations of a function and using a dispatcher to select the correct one based on the detected CPU type. Compiler-assisted dispatching, available in GCC and CLANG, simplifies this process by allowing the compiler to handle function selection based on CPU capabilities, using attributes like `__target__` and `__ifunc`. The article also highlights the importance of SIMD (Single Instruction Multiple Data) for optimizing performance in data-intensive applications. The authors conducted tests comparing different implementations of a summation function using AVX and SSE instructions, demonstrating significant performance improvements. Ultimately, CPU dispatching, especially when combined with vectorization, can maximize CPU resource utilization without sacrificing code portability.

- CPU dispatching enhances performance while maintaining portability.

- It allows binaries to select code versions based on CPU features at runtime.

- Manual and compiler-assisted dispatching are two main approaches.

- SIMD instructions can significantly improve performance in data processing.

- Testing showed substantial speed improvements using AVX and SSE implementations.