performance test of sparse-matrix vector product added

[lambday]

Tried testing mat-vec product performance with SGSparseMatrix and Eigen::SparseMatrix, used same fixed matrix for both with dim 10^6x10^6, with nnz~3^10^6, nz elements are along the diagonal, the first row and the first col.

On my machine the output is (with -O3)

time shogun (s) eigen3 (s)

0 0.940000 0.800000
1 0.940000 0.800000
2 0.940000 0.790000
3 0.940000 0.800000
4 0.940000 0.790000

without -O3, however
time shogun (s) eigen3 (s)

0 1.420000 13.450000
1 1.440000 13.460000
2 1.440000 13.480000
3 1.430000 13.470000
4 1.430000 13.470000

[lambday]

@karlnapf @sonney2k @lisitsyn please have a look :)

[lambday]

checked using valgrind, heap usage (for 1 time creation)

eigen3 : 264 allocs, 264 frees, 258,255,714 bytes allocated
shogun: 3,000,236 allocs, 3,000,236 frees, 160,038,038 bytes allocated

[lambday]

updated with dense dot

[lisitsyn]

register makes no real advantage nowadays - compiler knows how to do that

[lambday]

@lisitsyn okay, I'll remove it. But this result looks reasonable, no? Earlier I was taking num_vectors=num_cols, so got confused. I guess this is what @sonney2k was talking about.

[karlnapf]

What is the consequence of this?
shoguns sparse matrix-vector product is way slower than eigen3's

[coveralls]

Changes Unknown when pulling f114684 on lambday:feature/log_determinant into * on shogun-toolbox:develop*.