[Numpy-discussion] testing with amd libm/acml

[Francesc Alted]

On 11/7/12 8:41 PM, Neal Becker wrote:
> Would you expect numexpr without MKL to give a significant boost?

Yes.  Have a look at how numexpr's own multi-threaded virtual machine 
compares with numexpr using VML:

http://code.google.com/p/numexpr/wiki/NumexprVML

As it can be seen, the best results are obtained by using the 
multi-threaded VM in numexpr in combination with a single-threaded VML 
engine.  Caution: I did these benchmarks some time ago (couple of 
years?), so it might be that multi-threaded VML would have improved by 
now.  If performance is critical, some experiments should be done first 
so as to find the optimal configuration.

At any rate, VML will let you to optimally leverage the SIMD 
instructions in the cores, allowing to compute, for example, exp() in 1 
or 2 clock cycles (depending on the vector length, the number of cores 
in your system and the data precision):

http://software.intel.com/sites/products/documentation/hpc/mkl/vml/functions/exp.html

Pretty amazing.

-- 
Francesc Alted