Operations with saturation/ceil

[davidbp]

Hello,

I could not find saturation operations implemented (maybe they are not there).

The standard + operator does

x = Vec(UInt8.((255,255,255,255)))
y = Vec(UInt8.((1,0,0,0)))
x + y
<4 x UInt8>[0, 255, 255, 255]

I would like to do a sum_ceil that keeps the maximum value of the datatype in case of overflow or the min in case of underflow.

x = Vec(UInt8.((255,255,255,255)))
y = Vec(UInt8.((1,0,0,0)))
sum_ceil(x,y)  
<4 x UInt8>[255, 255, 255, 255]

According to intel_X86 there are specific instructions that do this (no need to check manually overflow etc..) in a single instruction call. Do you have plans to add them? Or are they there but I could not see them.

[KristofferC]

Yes, these exist in LLVM (https://llvm.org/docs/LangRef.html#saturation-arithmetic-intrinsics).

They could easily be added to https://github.com/eschnett/SIMD.jl/pull/63/files#diff-04c8d1e34d3b52bf0e6db708969277ecR127. I might do it soon unless I get beaten to it.

[davidbp]

Thanks for your fast answer Kristoffer, I hope that you can manage to do it. This is beyond the current knowledge that I have about the low level parts interfacing LLVM and Julia. Nevertheles I hope to be able to build upon this features once they are ready !.

[KristofferC]

Added in e3321e5 (you need to use the branch for that PR for now).

julia> x = Vec(UInt8.((255,255,255,255)))
<4 x UInt8>[255, 255, 255, 255]

julia> y = Vec(UInt8.((1,0,0,0)))
<4 x UInt8>[1, 0, 0, 0]

julia> SIMD.add_saturate(x, y)
<4 x UInt8>[255, 255, 255, 255]