Cypress CSC-1200T Guía de usuario descargar pdf (Pagina 75)

Chapter 7. Interprocess communication 75

responding call would be

CALL SHMEM_type_op_TO_ALL(target, source, nreduce, &

pe_start, logpe_stride, pe_size, pwrk, psync)

and here type is one of {INT8, INT4, REAL8, REAL4, COMP8, COMP4},and

op is one of the operations already mentioned.

The call above applies reduction operation op on data of type type at ad-

dress source in the memories of all PEs involved. The result is stored at

address target. The argument nreduce tells on how many consecutive

data items the reduction operation is to be performed.

Let us suppose that we have two PEs both of which store a vector of 4

integer elements. If we call shmem_int8_sum_to_all with nreduce =

1, the result will be one integer which equals the sum of the ﬁrst elements

of the vectors. If nreduce equals 4, we get an array of 4 integers, and

each element in this array is the sum of the corresponding elements

in the original vectors. Thus, if the total sum of all elements in both

vectors is to be calculated, one must ﬁrst call a SHMEM routine to form

an array of partial sums, and then ﬁnish the calculation by summing up

the elements in the resulting array with, e.g., a BLAS routine.

The triple pe_start, logpe_stride, pe_size is used to deﬁne the so

called active set, which includes the PEs taking part in the reduction op-

eration. The value of pe_start is simply the number of the ﬁrst PE in

the active set. The value of logpe_stride is the logarithm (in base 2) of

the stride between the PEs, and pe_size is the number of PEs in the ac-

tive set. Thus {pe_start, logpe_stride, pe_size}={0, 1, 5} indicates

that the active set consists of the PEs 0, 2, 4, 6, 8. As another example,

{pe_start, logpe_stride, pe_size}={0, 0,n} indicates that the ac-

tive set is PEs {0, 1,...,n− 1}.

Note: all the PEs in an active set and only these should call a collective

routine!

Finally, pwrk and psync are symmetric work arrays. The argument

psync is of integer type, and of size shmem_reduce_sync_size (this

constant is deﬁned in the ﬁle mpp/shmem.h or in the ﬁle mpp/shmem.fh.

They should be included at the beginning of a code utilizing SHMEM

library). The variable psync must be initialized so that the value of each

entry is equal to shmem_sync_value. After initialization it is a good

idea to call a barrier routine to guarantee synchronization before using

psync.

The argument pwrk should be of the same type as the reduction routine

and of size max(nreduce/2 + 1, shmem_reduce_min_wrkdata_size).

1 2 ... 70 71 72 73 74 75 76 77 78 79 80 ... 123 124

Sin comentarios

Cypress CSC-1200T Guía de usuario Pagina 75