Content deleted Content added
No edit summary |
|||
Line 87:
<ref>{{cite web|title=Multi-Target Vectorization With MTPS C++ Generic Library|url=http://www.metz.supelec.fr/metz/recherche/publis_pdf/Supelec753.pdf}}a C++ template library for producing optimised memory access patterns</ref>
Different considerations for memory access patterns appear in parallelism beyond locality of reference, namely the separation of reads and writes. E.g: even if the reads and writes are 'perfectly' local, it can be impossible to parallelise due to dependancies; separating the reads and writes into separate areas yields a ''different'' memory access pattern, maybe initially appear worse in pure ''locality'' terms, but desirable to leverage modern parallel hardware.<ref name="gpu gems2"/>
Locality of reference may also refer to individual variables (e.g. the ability of a compiler to cache them in registers), whist the term memory access pattern only refers to data held in an indexable memory (especially [[main memory]]).
| |||