Re: [eigen] Re: [Blitz-devel] Fork Blitz++?

[Gael Guennebaud <gael.guennebaud@xxxxxxxxx>]

On Wed, Jan 20, 2016 at 10:32 AM, Christoph Hertzberg <chtz@xxxxxxxxxxxxxxxxxxxxxxxx> wrote:

I was not surprised that Eigen is faster -- mostly as it uses SIMD.
But I think for a fairer comparison, you should at least initialize the values in the Blitz++ implementation (working on uninitialized values tends to be slow, if there is a number of NaNs and Infs in the data), and probably also compare how big the difference without SIMD would be (or with just SSE2/3/4, instead of AVX+FMA).

Initializing the Blitz tensors did not make any difference here (I disabled SSE exception anyways).

So to complete the bench:

Eigen with no vectorization at all (-DEIGEN_DONT_VECTORIZE  -fno-vectorize) : 3.5s

Eigen with SSE2 only: 1s

Eigen with AVX: 0.57s

Eigen with FMA : 0.35s

Eigen with FMA+OpenMP: 0.1s (4 threads)

Blitz: 15s

 

SIMD is not all, cache friendliness is also very important!

gael