Concluding remarks

SIMD extensions, particularly those from Intel, have become indispensable for high-performance computing (HPC). Compilers have become much more efficient at taking them into account, and manual programming via "intrinsics" offers additional gains when the compiler is unable to "vectorize".

In Intel's case, we've moved on from "pure" SIMD, where operations are performed on all SIMD register elements, to a SIMD with certain vector characteristics (mask instructions, gather-scatter memory access). While the characteristics of SIMD and vector extensions are similar, there is one fundamental difference: there is one SIMD instruction for each operation on each SIMD register size, which leads to a proliferation in the number of operation codes. In contrast, in the vector approach, specific operations are used to configure the size of vector registers and their contents,...