SIMD Support

Type VecElement{T} is intended for building libraries of SIMD operations. Practical use of it requires using llvmcall. The type is defined as:

  1. struct VecElement{T}
  2.     value::T
  3. end

It has a special compilation rule: a homogeneous tuple of VecElement{T} maps to an LLVM vector type when T is a primitive bits type and the tuple length is in the set {2-6,8-10,16}.

At -O3, the compiler might automatically vectorize operations on such tuples. For example, the following program, when compiled with julia -O3 generates two SIMD addition instructions (addps) on x86 systems:

  1. const m128 = NTuple{4,VecElement{Float32}}
  3. function add(a::m128, b::m128)
  4.     (VecElement(a[1].value+b[1].value),
  5.      VecElement(a[2].value+b[2].value),
  6.      VecElement(a[3].value+b[3].value),
  7.      VecElement(a[4].value+b[4].value))
  8. end
  10. triple(c::m128) = add(add(c,c),c)
  12. code_native(triple,(m128,))

However, since the automatic vectorization cannot be relied upon, future use will mostly be via libraries that use llvmcall.