on 18.08.2009 at 18:59 joel wrote :
looks smart and tricky but i don't see any advantages over computed indeces but only wasted spaces and time (for allocating additional storage for pointers) at the same time dereferencing a pointer is like computing an index - no computational advantage agai Back to this issue here is the generated code using 1/ NRC array of dimension 4 [code here] Lo and behold ;) Doing the computation of index jump int he allcoation helps tremendously. i got somewhat unexpected results
i reimplemented 'matrix' in terms of nrc here are results 1. a small mixin to vector operations disadvantage is ~10% - predictable (here by advantage i mean ratio (old_time - nrc_time)/old_time since i assume that nrc would outperform old (non-nrc) code (should be nrc_time<old_time, then ration is >0) if the ratio is negative then it is actually a disadvantage value of ratio is in % for convenience) 2. intensive matrix operations (+-* etc.) advantage is roughly zero - predictable 3. matrix multiplication advantage is roughly zero for best implementation - predictable (see mul.png) 4. matrix inversion (inplace, full pivoting) _advantage_ is ~15% - totally unpredictable (see inv.png) actually it's difficult for me to explain the reason since it implements full pivoting maybe acceleration of the search process defeats the slowness of reducing cycles or maybe i made a mistake however i can explain cases 1-3 disadvantages by the involvement of additional allocation/deallocation and setup process i'm still wondering about the latter case but i consider an invertion such an unwanted corner case and that's why i suggest to implement computed indeces for 2d case however for you, joel, there is nothing to bother since your point is perfectly motivated -- Pavel
