Re: [ublas] Bad prod efficiency of sparse/symmetry matrix ?

From: Gunter Winkler (guwi17_at_[hidden])
Date: 2006-12-06 03:41:03

On Wednesday 06 December 2006 05:10, Seabook wrote:
> There are some data about normal/sparse/symmetry matrix.
>                      Debug(ms)Release(msec)
> Sparse           139094       46390
> Symmetry      139094       46390
> Normal             2578         1329
> The efficiency are really terrible?!
> Did I forget something ? (such as declare or method of prod or assign data
> )

Reading your code I noticed that you do matrix-matrix-multiplications. It is 
no surprise that multiplication of full sparse matrices is much slower than 
multiplication of dense matrices, because the latter can be extremely 
optimized. The sparse algorithm wins over dense if the number of nonzeroes 
per row is small compared to the size. Although the overall times is quite 
high. On the other hand the orientation (row major vs. column major) is very 
important for sparse products

Please try these operations and report the times.
compressed_matrix<float, column_major> * compressed_matrix<float, 
column_major>
compressed_matrix<float, column_major> * matrix<float, column_major>
compressed_matrix<float, column_major> * matrix<float, row_major>
compressed_matrix<float, column_major> * symmetric_matrix<float, column_major>
compressed_matrix<float, column_major> * symmetric_matrix<float, row_major>

please check whether C=prod(A,B) or axpy_prod(A,B,X,true) is faster.

PS: you initialize sy1 twice (and do nothing with sy2)

mfg
Gunter