Hello,

In the attached file, I compared numpy and scipy svd function for types np.float32, np.float64, np.complex64, np.complex128. Surprisingly, under numpy, np.float32 and np.conplex64 times are similar to np.float64 and np.complex128, respectively. However, under scipy, results are doubled as expected when passing from single to double precision.

I suspect numpy to force double precision, while scipy is not, which is the correct result. As SVD is a computing intensive function, using single precision can be crucial to limit computation time.

This has been tested on numpy 1.12.1, either under windows and linux.

Thanking you,
GuL

test_svd_cpu_gul.py.txt