Popularity
8 Stars
Updated Last
2 Years Ago
Started In
March 2016

SIMDVectors

Join the chat at https://gitter.im/KristofferC/SIMDVectors.jl

Build Status

This is an experimental package that uses the PR #15244 to create a stack allocated fixed size vector which supports SIMD operations. It is very similar in spirit to the SIMD.jl package excpet this is written in pure julia. It also supports type promotions and should cleanly work with "exotic" number types like BigFloat.

For this package to work, the branch above needs to be used and to actually get SIMD operations, julia needs to be started with the -O3 flag.

There are currently a few ambiguity warnings when the package is loaded, This is annoying but should not cause any real problems.

Loading and storing SIMDVectors

A SIMDVector can be created by for example using load(SIMDVector{N}, v, offset=0) where N is the length of the vector, v is vector to load data from and offset is an offset into v where to start loading data:

julia> v = load(SIMDVector{7}, rand(12))
7-element SIMDVectors.SIMDVector{3,2,1,Float64}:
 0.0333167
 0.52255
 0.171032
 0.667967
 0.832219
 0.586471

A SIMDVector can be stored back in a normal Vector with the store! function:

julia> vec_store = similar(v)

julia> store!(vec_store, v)
7-element Array{Float64, 1}:
 0.0333167
 0.52255
 0.171032
 0.667967
 0.832219
 0.586471

Operations on and between SIMDVectors.

A SIMDVector looks like a normal Vector but internally the data is packed such that, when possible, vectorized instructions are used when operators are performed on and between SIMDVector's. If the length of the vector are such that not all numbers fit in vector registers, scalar operations are performed on the rest.

julia> va = load(SIMDVector{9}, rand(Float32, 12));

julia> vb = load(SIMDVector{9}, rand(Float32, 12));


julia> @code_native va + vb
...
    vaddps  (%rdx), %xmm0, %xmm0   # One packed add for the first set of four Float32s
    vaddps  16(%rdx), %xmm1, %xmm1 # Second packed add for second set of four Float32s
    vmovss  32(%rsi), %xmm2
    vaddss  32(%rdx), %xmm2, %xmm2 # One scalar add for the rest
...

Reduction (sum, prod, maximum, minimum) are also available:

julia> sum(va)
4.901259f0

julia> maximum(va)
0.93982494f0

Promotions

Operators between two different types will convert like normal vectors:

julia> va = load(SIMDVector{9}, rand(Float64, 12));

julia> vb = load(SIMDVector{9}, rand(Float32, 12));

julia> va + vb
9-element SIMDVectors.SIMDVector{4,2,1,Float64}:
 0.648343
 1.02155
 0.676522
 0.92291
 1.14035
 1.46949
 0.599293
 1.1952
 1.02997

User defined number types

SIMDVector's' should gracefully handle arbitrary julia number types. This makes it so that a SIMDVector can be used even if you are unsure what data it will hold.

julia> a = load(SIMDVector{4}, big(rand(12))); # Load Big floats into a SIMDVector

julia> a+a # Works fine
4-element SIMDVectors.SIMDVector{0,0,4,BigFloat}:
 2.531343636343290626200541737489402294158935546875000000000000000000000000000000e-01
 3.366090705330369026171410951064899563789367675781250000000000000000000000000000e-01
 1.697265196033196144043131425860337913036346435546875000000000000000000000000000
 1.206431829930139532081057041068561375141143798828125000000000000000000000000000