What's the problem with this simple multi-thread code?

[tkf]

@photor

function loop_dot_auto(x, A, y; ntasks = Threads.nthreads())
    s = zeros(ntasks)
    @inbounds Threads.@threads for i = eachindex(y)

@inbounds needs to go inside @threads What is the cause of this performance difference between Julia and Cython? - #4 by tkf

In general, try minimizing the scope of @inbounds.

@lmiq

julia> function loop_dot_serial(x, A, y)
           s = 0.
           for i = eachindex(y)
               for k = eachindex(x)
                   s += x[k]*A[k,i]*y[i]

I think this needs @inbounds to compare this properly with LoopVectorization.jl and Tullio.jl. This also applies to @threads and @floop.

@lmiq

I am trying to avoid the hot cache by re-initializing the arrays every time.

GC may re-use the same memory region. Maybe a more robust approach is to allocate a lot of arrays such that the total number of bytes is at least as large as (say) a double of the L3 cache size and use them one by one.