Flux and Octavian

qwerty · November 4, 2022, 9:50am

Does Flux use Octavian when run on cpu?

baggepinnen · November 4, 2022, 10:06am

No, but SimpleChains.jl uses LoopVectorization

PumasAI/SimpleChains.jl/blob/main/src/dense.jl#L182


      
          @inline function dense!(
            f::F,
            C::AbstractVecOrMat{<:Base.HWReal},
            A::AbstractMatrix{<:Base.HWReal},
            B::AbstractVecOrMat{<:Base.HWReal},
            ::True,
            ::True,
          ) where {F}
            Kp1 = ArrayInterface.size(A, StaticInt(2))
            K = Kp1 - StaticInt(1)
            @turbo for n ∈ indices((B, C), 2), m ∈ indices((A, C), 1)
              Cmn = zero(eltype(C))
              for k ∈ 1:K
                Cmn += A[m, k] * B[k, n]
              end
              C[m, n] = f(Cmn + A[m, Kp1])
            end
          end
          @inline function dense!(
            f::F,
            C::AbstractVecOrMat{<:Base.HWReal},

Topic		Replies	Views
Octavian Matmul slow in for loop Performance matrices	0	230	April 28, 2023
ANN: LoopVectorization 0.12: multithreading and better handling of discontiguous memory accesses Performance	16	2184	March 17, 2021
[ANN] LoopVectorization Package Announcements	157	23393	May 27, 2020
Julia matrix-multiplication performance Performance linearalgebra	20	8719	October 30, 2022
Implementing a ConvNet that doesn't allocate during inference (SimpleChains.jl?) Machine Learning	3	139	February 27, 2025

Flux and Octavian

Related topics