bench_transpose.jl

import LinearAlgebra.BLAS

const libblas = Base.libblas_name
const liblapack = Base.liblapack_name

import LinearAlgebra
import LinearAlgebra: BlasReal, BlasComplex, BlasFloat, BlasInt, DimensionMismatch, checksquare, stride1, chkstride1, axpy!

import Libdl

function blas_transpose!(B,A)
    ccall((BLAS.@blasfunc("domatcopy_"), libblas), Cvoid, (Ref{UInt8}, Ref{UInt8}, Ref{BlasInt}, Ref{BlasInt}, Ref{Float64}, Ptr{Float64}, Ref{BlasInt}, Ptr{Float64},
Ref{BlasInt}),
         'C',
         'T',
         size(A,1),
         size(A,2),
         1.0,
         A,
         size(A,1),
         B,
         size(B,1))
    B
end

BLAS.set_num_threads(1) #does not seem to matter anyway

for i  in 1:4
    N = rand(500:1000)
    M = rand(500:1000)
    # M = N #uncomment for square
    println("$M $N")
    A = randn(M,N)
    B = zeros(N,M)
    @btime blas_transpose!($B,$A)
    C = copy(B)
    B .= 0
    @btime transpose!($B,$A)
    @assert norm(C-B) < 1e-10
end

On a Haswell:

887 773
  920.267 μs (0 allocations: 0 bytes)
  932.760 μs (0 allocations: 0 bytes)
748 913
  847.834 μs (0 allocations: 0 bytes)
  953.112 μs (0 allocations: 0 bytes)
952 925
  1.365 ms (0 allocations: 0 bytes)
  1.376 ms (0 allocations: 0 bytes)
835 983
  1.119 ms (0 allocations: 0 bytes)
  1.276 ms (0 allocations: 0 bytes)

On a Sandybridge:

969 803
  4.901 ms (0 allocations: 0 bytes)
  1.438 ms (0 allocations: 0 bytes)
575 900
  2.123 ms (0 allocations: 0 bytes)
  804.264 μs (0 allocations: 0 bytes)
785 625
  3.029 ms (0 allocations: 0 bytes)
  786.346 μs (0 allocations: 0 bytes)
808 940
  4.826 ms (0 allocations: 0 bytes)
  1.371 ms (0 allocations: 0 bytes)