SIMD instruction sets

This article describes the main features of the SIMD extensions to modern microprocessor instruction sets that have been introduced and developed since the 1990s. The arithmetic and logic operations performed by scalar instruction set instructions are performed on the maximum size of processor registers, 32 or 64 bits. However, programs can work on data of smaller size, such as bytes (8 bits) or 16-bit words, whether signed or unsigned. This is the case for image processing, signal processing and many other applications. The principle of SIMD instructions is therefore to use larger registers (128, 256 or 512 bits) and to perform the same operation on vectors containing several elements 8 bits, 16 or 32 bits for integers, 32 or 64 bits for floating-point numbers.

The characteristics of these extensions are illustrated from the most widely used: SSE to SSE4.2, AVX and AVX2 for Intel's IA-32 and Intel 64 instruction sets, ARM's Neon and Neon2, Altivec and its various IBM variants.

Floating arithmetic instructions pose no problems. The problem of carry for integer arithmetic instructions is discussed in detail. As SIMD instructions perform the same operation on all elements of a vector, if-then-else conditional structures require special handling. Memory instructions must access elements located at successive memory addresses, which implies special processing when this is not the case. The classic example is data stored in memory as array structures (AoS), which must be transformed into array structures (SoA) to enable SIMD calculations.

Most SIMD instructions are natural extensions of scalar instruction sets, accompanied by data manipulation instructions to facilitate SIMD processing. These instructions are accompanied by ad hoc instructions for specific applications. SIMD extensions have also been called multimedia extensions, as they were originally intended to make general-purpose processors competitive in multimedia, signal processing and security applications. Typical examples of ad hoc instructions include motion detection, complex number calculations, cryptography, etc.

This article also explains how to use these instructions. One possibility is to help the compiler to "vectorize", i.e. to use them. The other approach is to use intrinsics to be inserted into C or C++ code: these are function calls to SIMD instructions to be used in particular for integer arithmetic or when high-level transformations are required that are not within the compiler "s reach.

Recent 512-bit extensions remove some of the limitations of SIMD extensions by allowing partial processing of vector elements according to a mask, and by enabling memory accesses at non-consecutive addresses. These developments bring...