Autovectorization seems like a cool way to write cross platform SIMD code. But does anyone know of solutions to the insight issue? If I were to write a function which relies on autovectorization, wouldn't I literally have to 1) compile with every compiler + compiler settings + CPU arch + platform I wanna support, 2) disassemble all resulting binaries, 3) read analyze the assembly code to verify that it's vectorized how I expect, and 4) repeat for every change?
Discussion
Loading...
This problem is what keeps me writing platform-specific functions with SIMD intrinsics every time I want vectorization. And the friction involved of writing platform-specific functions with SIMD intrinsics is what keeps me away from doing vectorization almost ever