SIMD struggles, seeking solutions (with KangarooTwelve.jl)

kristoffer.carlsson · November 7, 2023, 9:36am

SIMD.jl has some checks that are supposed to fix the undefined behavior:

eschnett/SIMD.jl/blob/0334b82341981a72821d3268cfe079ac0843e6ac/src/simdvec.jl#L319-L339C4


      
          # Bitshifts
          # See https://github.com/JuliaLang/julia/blob/7426625b5c07b0d93110293246089a259a0a677d/src/intrinsics.cpp#L1179-L1196
          # Shifting with a value larger than the number of bits in the type is undefined behavior
          # so set to zero in those cases.
          @inline function shl_int(x::Vec{N, T1}, y::Vec{N, T2}) where {N, T1<:IntegerTypes, T2<:IntegerTypes}
              vifelse(y > sizeof(T1) * 8,
                  zero(Vec{N, T1}),
                  Vec(Intrinsics.shl(x.data, convert(Vec{N,T1}, y).data)))
          end
          
          @inline function lshr_int(x::Vec{N, T1}, y::Vec{N, T2}) where {N, T1<:IntegerTypes, T2<:IntegerTypes}
              vifelse(y > sizeof(T1) * 8,
                  zero(Vec{N, T1}),
                  Vec(Intrinsics.lshr(x.data, convert(Vec{N,T1}, y).data)))
          end
          
          @inline function ashr_int(x::Vec{N, T1}, y::Vec{N, T2}) where {N, T1<:IntegerTypes, T2<:IntegerTypes}
              vifelse(y > sizeof(T1) * 8,
                      Vec(Intrinsics.ashr(x.data, Vec{N,T1}(sizeof(T1)*8-1).data)),
                      Vec(Intrinsics.ashr(x.data, Vec{N,T1}(y).data)))
          end

But perhaps it is buggy?

tecosaur · November 7, 2023, 10:21am

That’s interesting Julius. Thanks for that analysis.

I might consider that hacky approach in future, though I’m also interested to see if https://github.com/JuliaLang/julia/pull/44186 (kindly pointed out to me by Sukera) might remove the overhead that’s being such a pain here?

Sukera · November 7, 2023, 10:32am

Likely! Shouldn’t these be y >= sizeof(T1)*8?

kristoffer.carlsson · November 7, 2023, 11:38am

Indeed, fix some shift operations to agree with Base (and prevent undefined behavior) by KristofferC · Pull Request #119 · eschnett/SIMD.jl · GitHub should fix that I think