The NEON unit
The NEON unit has a register bank of thirty-two 64-bit vector registers that can be operated on in parallel.
The NEON unit can view the register bank as either:
Sixteen 128-bit quadword registers,
Thirty-two 64-bit doubleword registers,
These registers can then be operated on in parallel in the NEON unit. For example, in one vector add instruction you can add eight 16-bit integers to eight other 16-bit integers to produce eight 16-bit results. This is known as vectorization (or more specifically for NEON, Single Instruction Multiple Data (SIMD) vectorization).
The NEON unit supports 8-bit, 16-bit and 32-bit integer operations,
and some 64-bit operations, in addition to single-precision (32-bit)
floating point operations. It can operate on elements in groups
of 2, 4, 8, or 16. (The Cortex-A9 processor also supports conversion
to and from 16-bit floating-point operations, which the compiler
--fp16_format is specified, from RVCT
4.0 and later, and ARM Compiler 4.1 and later.)
Vectorization of floating-point code does not always occur
automatically. For example, loops that require re-association only
vectorize when compiled with
--fpmode fast. Compiling
--fpmode fast enables the compiler to perform some
transformations that could affect the result.
The NEON unit is classified as a vector Single Instruction Multiple Data (SIMD) unit that operates on multiple elements in a vector register by using one instruction.
For example, array A is a 16-bit integer array with 8 elements.
Table 3-1 Array A
Array B has the following 8 elements:
Table 3-2 Array B
To add these arrays together, fetch each vector into a vector register and use one vector SIMD instruction to obtain the result.
Table 3-3 Result
The NEON unit can only deal with vectors that are stored consecutively in memory, so it is not possible to vectorize indirect addressing.
When writing structures, be aware that NEON structure loads require the structure to contain equal-sized members.