hello

i have the same code , one using boost ublas and the same using Eigen 2
for the vector

for(int i=0;i<size_;i++)
{
T s = 0.0;

for(int k=lpos_(i);k<lpos_(i+1);k++)
{
int j = cpos_(k);
s += lo_(k) * b(j);
}

b(i) = (b(i) - s)/dia(i);
}

i have done some benchmark under windows 7 with visual c++ 2008

the code using ublast is 3 time faster as this with Eigen 2

is there some explanation ?

is it possible to optimize the eigen code using some tricks ?

thanks

Are you working in Debug or Release mode? Eigen only claims to be fast in Release mode. In particular, in addition to enabling optimization, this should disable assertions. Otherwise, Eigen will use assertions to guard every coefficient access in a vector. You can access coefficients without assertions by using vector.coeff(i) / vector.coeffRef(i) instead of vector(i).

thanks for your answers

I use eigen in release mode. but i will try the direct access function to the vector element . this slow access can have some relation with the inlining of this access function.

The other explanation could be vectorization. Make sure to enable SSE2 in your project.

There is no reason whatsoever why we should be slower than ublas
http://eigen.tuxfamily.org/index.php?ti ... August2008

Also make sure that your own project must be compiled in Release mode, since eigen is just headers.

yes the project is compiler in release mode and i have activated the sse2 optimisation.

the part of the code i have posted is slower with eigen but the dot product with eigen code is faster than ublas ex : alpha=r.dot(dc);

thanks

One never knows: you might have encountered one of these cases where MSVC faisl to inline a crucial function.

If you could either:
- send us a list of the symbols (demangled) in your executable, or
- examine the assembly and find out yourself what functions are not inlined...

that could help us fix it.

also make sure you are using initialized variables with valid values, NaN, inf and the likes might kill the performance.

replacing the cpos_ lpos with Eigen integer vector (before there were stl:vector)

and using the function .coeff(i) and .coeffref(i), the Eigen code work now at same speed as ublas , a little faster

thanks very much

that probably means that NDEBUG is not defined in your "release mode" that is weird because this should not be the default.