Using MKL in 1.7

General Usage Performance
1

My understanding is that all you need to do is type

using MKL

in the REPL and BLAS trampoline will take care of the rest. When I do that I see no difference in performance.

What should I be doing? I’m on an Intel Mac and MKL.jl installed with no complaints.

2

What platform are you using?

3

Intel Mac

4

What does LinearAlgebra.BLAS.get_config() say for you after you load MKL?

julia> using LinearAlgebra.BLAS

julia> BLAS.get_config()
LinearAlgebra.BLAS.LBTConfig
Libraries: 
└ [ILP64] libopenblas64_.so

julia> using MKL

julia> BLAS.get_config()
LinearAlgebra.BLAS.LBTConfig
Libraries: 
└ [ILP64] libmkl_rt.so
5

Looks like it’s doing the right thing. Seems like I need more compelling benchmarks.

Thanks,

julia> BLAS.get_config()
LinearAlgebra.BLAS.LBTConfig
Libraries: 
└ [ILP64] libopenblas64_.0.3.13.dylib

julia> using MKL

julia> BLAS.get_config()
LinearAlgebra.BLAS.LBTConfig
Libraries: 
└ [ILP64] libmkl_rt.1.dylib
6

FWIW, on my Macbook Pro (Intel):

julia> using BenchmarkTools

julia> using LinearAlgebra

julia> BLAS.set_num_threads(1)

julia> A = rand(1000,1000); B = rand(1000,1000);

julia> @btime $A * $B;
  53.150 ms (2 allocations: 7.63 MiB)

julia> @btime inv($A);
  79.604 ms (5 allocations: 8.13 MiB)

julia> using MKL

julia> BLAS.set_num_threads(1)

julia> @btime $A * $B;
  51.048 ms (2 allocations: 7.63 MiB)

julia> @btime inv($A);
  62.249 ms (5 allocations: 9.35 MiB)

(UPDATED because I didn’t set the number of threads to one for MKL.)

7

Interesting. Your matrix-matrix product timings are like what I got with a different example. I did not try inv and am somewhat surprised at what you found. MKL may not be as amazing as I expected.

8

You’re setting num threads to 1 with OpenBLAS but not MKL?

julia> BLAS.get_config()
LinearAlgebra.BLAS.LBTConfig
Libraries: 
└ [ILP64] libopenblas64_.so

julia> BLAS.set_num_threads(1)

julia> A = rand(1000,1000); B = rand(1000,1000);

julia> @btime $A * $B;
  45.090 ms (2 allocations: 7.63 MiB)

julia> using MKL

julia> BLAS.get_config()
LinearAlgebra.BLAS.LBTConfig
Libraries: 
└ [ILP64] libmkl_rt.so

julia> @btime $A * $B;
  14.268 ms (2 allocations: 7.63 MiB)

julia> BLAS.set_num_threads(1)

julia> @btime $A * $B;
  45.732 ms (2 allocations: 7.63 MiB)
9

Indeed, my bad. Correct it above.

10

Is there any way I can use the output of BLAS.get_config() within a program to see if I’m using MKL or not?

11

Something along the lines of any(contains("mkl"), getfield.(BLAS.get_config().loaded_libs, :libname)) should do the job.

1 Like
12

That did it. Thanks.