#includeNot Collectivevoid PetscPrefetchBlock(const anytype *a,size_t n,int rw,int t)
| a | - pointer to first element to fetch (any type but usually PetscInt or PetscScalar) | |
| n | - number of elements to fetch | |
| rw | - 1 if the memory will be written to, otherwise 0 (ignored by many processors) | |
| t | - temporal locality (PETSC_PREFETCH_HINT_{NTA,T0,T1,T2}), see note |
Adopting Intel's x86/x86-64 conventions, there are four levels of temporal locality. Not all architectures offer equivalent locality hints, but the following macros are always defined to their closest analogue.
| PETSC_PREFETCH_HINT_NTA | - Non-temporal. Prefetches directly to L1, evicts to memory (skips higher level cache unless it was already there when prefetched). | |
| PETSC_PREFETCH_HINT_T0 | - Fetch to all levels of cache and evict to the closest level. Use this when the memory will be reused regularly despite necessary eviction from L1. | |
| PETSC_PREFETCH_HINT_T1 | - Fetch to level 2 and higher (not L1). | |
| PETSC_PREFETCH_HINT_T2 | - Fetch to high-level cache only. (On many systems, T0 and T1 are equivalent.) |
This function does nothing on architectures that do not support prefetch and never errors (even if passed an invalid address).
Level:developer
Location:include/petscsys.h
Index of all Sys routines
Table of Contents for all manual pages
Index of all manual pages