GEMM slower than GEMV slower than AXPY equivalent on Intel i5 CPU

Consider the following Julia code (using Julia 0.4-dev):

let
       B = randn(1000, 1000)
       v = randn(1000)
       y = randn(1000); sB = [slice(B, :, j) for j = 1:size(B, 2)]

       @time for i = 1:1000;BLAS.gemm!('N', 'N', 1.0, B, v, 1.0, y);end;

       @time for i = 1:1000;BLAS.gemv!('N', 1.0, B, v, 1.0, y);end;

       @time for i = 1:1000;
           for j = 1:size(B,2)
             BLAS.axpy!(v[j], sB[j], y)
           end
       end
end

On @andreasnoack's machine, a Macbook Pro with i7-4870HQ CPU, GEMM is 4 times slower than GEMV:

elapsed time: 1.084909686 seconds (0 bytes allocated)
elapsed time: 0.2644927 seconds (0 bytes allocated)
elapsed time: 0.321705553 seconds (0 bytes allocated)

On my machine, a Macbook Pro with i5-4258U, I get similar behavior, but also that the AXPY equivalent is the fastest of the 3 computations:

elapsed time: 1.693223657 seconds (0 bytes allocated)
elapsed time: 0.818590556 seconds (0 bytes allocated)
elapsed time: 0.715702898 seconds (0 bytes allocated)

I find the relative performance behaviors surprising.

Commencez par reproduire le benchmark Julia avec BLAS.gemm!, BLAS.gemv! et BLAS.axpy! sur les CPU Intel indiqués, puis comparez les temps relatifs. L’issue ne nomme ni fichiers source, ni tests, ni changement spécifique ; précisez si l’objectif est d’expliquer la disparité ou d’améliorer une opération avant de définir ce qui constitue l’achèvement.