AppleAccelerate SVD is slower than libopenblas64

[OliverDudgeon]

I was benchmarking this on my M4 Pro Macbook Pro (10 core) and I noticed that although I get benefits for some BLAS operations (matrix multiplications etc) SVD is a fair bit slower than libopenblas64. Here's the results I get on my machine:

• AppleAccelerate

In [1]: using LinearAlgebra, BenchmarkTools, AppleAccelerate

In [2]: A = rand(ComplexF32, 2000, 2000);

In [3]: @btime svd(A);
  1.548 s (21 allocations: 169.04 MiB)

• 10 BLAS threads

In [1]: using LinearAlgebra, BenchmarkTools

In [2]: BLAS.get_num_threads()
Out[2]: 10

In [3]: A = rand(ComplexF32, 2000, 2000);

In [4]: @btime svd(A);
  841.862 ms (21 allocations: 169.04 MiB)

[ViralBShah]

This is a vendor BLAS - so not much we can do about this. Probably should be reported to Apple.