Hi,

I'd like to know what kind of performance gain I can expect by enabling sse2 with Visual Studio 2008 on the developpement branch of eigen 3?

I'm currently using the test between n and z to do some benchmarking.
For 15 itérations, I get 4 min 15 sec with only openmp.
When I activate SSE2, I get 32 sec. I just seems too good to be true.

As a more precise example, the product_symm test takes 55 secs with openmp and 5 secs with sse2.

I was wondering if those number were plausibles, or is there something else going on?

Thank you,

For 15 itérations, I get 4 min 15 sec with only openmp.
When I activate SSE2, I get 32 sec. I just seems too good to be true.

No it's normal. You get a total x8, decomposing into:
x4 because 4 floats fit in a packet
x2 because SSE addition and multiplication can run together in 1 cycle

x2 because SSE addition and multiplication can run together in 1 cycle

I am very curious how it is possible? Is it in PacketMath.h?

renorm wrote:

x2 because SSE addition and multiplication can run together in 1 cycle

I am very curious how it is possible? Is it in PacketMath.h?

No, it's in your CPU. Recent Intel CPUs (not sure about AMD) are able to execute in 1 cycle a (mulps,addps) pair, and likewise a (mulpd,addpd) pair, EVEN if operating on the same registers. It's as if there were an combined-mul-and-add instruction working in 1 cycle.

How do you get it triggered? Follow _mm_mul_ps with _mm_add_ps?

Yes: see ei_pmadd() in GenericPacketMath.h:

/** \internal \returns a * b + c (coeff-wise) */
template<typename Packet> inline Packet
ei_pmadd(const Packet& a,
const Packet& b,
const Packet& c)
{ return ei_padd(ei_pmul(a, b),c); }

Thanks for explaining.

The original question mentioned OpenMP. Does Eigen explicitly use it?

Yes, Eigen 3 uses OpenMP if it's enabled. By default it's disabled, but for example with GCC you just have to pass -fopenmp.

Hi!
I use a function of matrix-vector multiplication. This function uses only one core?
When multiplying two matrices are used all the cores.
System Configuration: Intel I7-970, Windows 7. Compiler VS2008, OpenMP enabled.

currently only matrix * matrix products is multi-threaded, not matrix * vector

Will there be implemented multithreading for matrix-vector multiplication?

sure, but I cannot say when...

If enable OpenMP, then LU decomposition is slower.
Why?

strange because I observed the opposite. Which matrix size? Do you have multi-threading enabled? if so call your executable with:

$ OMP_NUM_THREADS=number_of_real_cores ./my_app

thanks, it helped