Cypress CSC-1200T Guía de usuario descargar pdf (Pagina 74)

74 Cray T3E User’s Guide

Other point-to-point routines

The routines shmem_put and shmem_get operate correctly when the data

being transferred consists of items with a size of 8 bytes. If the size

of a single data item is only 4 bytes (32 bits), one must call instead

the routines shmem_put4 or shmem_get4 in Fortran or shmem_put32,

shmem_get32 in C and C++.

Another crucial restriction for the basic versions of shmem_put and

shmem_get is that they accept only consecutive data items. To move

data with a non-unit stride from a PE to another (or to itself, for that

matter), there are extended versions shmem_iget and shmem_iput.

The C language syntax for shmem_iget is

void shmem_iget(void *target, void *source, int target_inc,

int source_inc, int len, int pe)

For Fortran 90 this is

CALL SHMEM_IGET(target, source, target_inc, source_inc, &

len, pe)

The syntax is similar for the routine shmem_iput and the correspond-

ing 4-byte versions. The integer valued arguments target_inc and

source_inc contain the strides between consecutive data items in the

target and the source data, respectively.

Atomic operations

SHMEM library contains several fast routines for updating or checking

the value of a single variable on a remote PE. These routines are called

atomic operations. These include, e.g., shmem_inc for incrementing the

value of a remote data object and shmem_swap for exchanging the values

of a single remote and local variable.

7.4.4 Reduction routines

A reduction routine computes a single value from a data object dis-

tributed over several PEs’ memories, e.g., the sum of the elements of a

vector. SHMEM contains the following reduction routines: shmem_and,

shmem_max, shmem_min, shmem_or, shmem_prod, shmem_sum and

shmem_xor. A general call to these routines in C is of the form

void shmem_type_op_to_all(type *target, type *source,

int nreduce, int PE_start, int logPE_stride, int PE_size,

type *pWrk, long *pSync);

where type is one of {short, int, float, double, complex, complexd}

and op is one of {sum, prod, min, max, and, or, xor}. In Fortran, the cor-

1 2 ... 69 70 71 72 73 74 75 76 77 78 79 ... 123 124

Sin comentarios

Cypress CSC-1200T Guía de usuario Pagina 74