Does anyone knows why this simple dot product is faster in Fortran then in Julia? (LinearAlgebra.dot is faster and with LoopVectorization the Julia code gets also faster. but this is not what I am questioning. I am curious about why without those packages the Fortran compiler is doing a better job …

The problem isn’t the type instability. It’s the promotion type. By using Float64, you lose the ability to use fma, and get halve your vector with.

Naive dot product faster in Fortran than in Juila

General Usage Performance

RoyiAvital July 24, 2021, 9:58pm 11

While you’re correct that Float64 will have half throughput of Float32 I am not sure about your statement regarding the FMA. As far as I know, FMA for Float64 is supported on AVX2 / AVX512.

Topic		Replies	Views
Why is BLAS dot product so much faster than Julia loop? Performance	18	5408	August 15, 2020
Why is dot plus faster? Performance question	6	508	May 28, 2022
Why for-loop of vector multiplication is slower that dot product? General Usage question	11	447	August 11, 2022
Fortran calling Julia. Julia 10x slower than Fortran Performance question , fortran	15	1961	May 27, 2021
Dot(x, inv.(y)) vs non-allocating one, the former is faster? Performance	5	632	March 2, 2019

Naive dot product faster in Fortran than in Juila

Related topics