An important consideration when considering code offloading or parallelising is the nature of the underlying hardware configuration.

Many modern processor architectures comprise an arrangement of heterogeneous processor cores, each of which may have its own memory space and each of which may also have access to a shared main memory. Management of data flows to make best use of local memory is an important part of a code offloading/parallelisation strategy, and has a particular impact on code performance.

Analysis of memory usage and data flows between memories is complex, especially in modern heterogeneous multicore processor architectures. Tool support for memory analysis is vital.

The Codeplay Memory Analyser, illustrated below, shows which memory accesses are best suited to using local memory and which are best suited to use shared memory, through streaming DMA, for instance.

What do I do now?

Codeplay's Memory Analyser is used in conjuction with the Sieve System, specifically for  optimizing multi-core enabled code.  If you are interested in this or any of the other components the Sieve then please get in touch to discuss your needs.