DATA MODIFICATION(S) Sample Clauses

DATA MODIFICATION(S). The data structures used by a program depend largely on the algorithm used, but within a given algorithm there may be changes that have to be made to fit a particular programming model or in order to achieve acceptable performance. Conversely, some algorithms may be able to accept the reduced accuracy provided by half-precision floating-point or fixed point numbers which may provide better performance on a parallel processor. It may also be necessary to pay more attention to the use of unsigned data types as these may perform differently from signed data types. AOS is not so efficient where multiple threads are operating in parallel on different elements of the array. In this case the data is more efficiently stored by having separate arrays for each element of the structure. This arrangement is known as a Structure of Arrays (SOA) and is particularly effective on SIMD architectures where multiple values can be read and written from sequential memory. With very large data sets the separation between related data values in AOS can cause problems with caching and TLBs. In this case it can be effective to limit the size of an SOA to fit within the page size and then replicate this multiple times, leading to an Array of Structures of Arrays (AOSOA). In each case the most efficient partitioning between structures and arrays will depend on the particular architecture being targeted. The 2D nature of image data makes other data arrangements even more efficient for parallel processing. In particular the data can be arranged in image tiles where each tile covers a 2D rectangle of the image but is stored linearly in memory. This provides better memory locality for algorithms that operate on regions of the screen, but it increases the overhead of address calculation of a pixel. Parallel processors avoid this overhead by providing special hardware to compute these addresses if the image data is stored in a supported format.