/** Optimized implementation of the ARMCI IOV operation that uses a single
* lock/unlock pair.
*/
int ARMCII_Iov_op_batched(enum ARMCII_Op_e op, void **src, void **dst, int count, int elem_count,
MPI_Datatype type, int proc) {
int i;
gmr_t *mreg;
void *shr_ptr;
switch(op) {
case ARMCII_OP_PUT:
shr_ptr = dst[0];
break;
case ARMCII_OP_GET:
shr_ptr = src[0];
break;
case ARMCII_OP_ACC:
shr_ptr = dst[0];
break;
default:
ARMCII_Error("unknown operation (%d)", op);
return 1;
}
mreg = gmr_lookup(shr_ptr, proc);
ARMCII_Assert_msg(mreg != NULL, "Invalid remote pointer");
gmr_lock(mreg, proc);
for (i = 0; i < count; i++) {
if ( ARMCII_GLOBAL_STATE.iov_batched_limit > 0
&& i % ARMCII_GLOBAL_STATE.iov_batched_limit == 0
&& i > 0 )
{
gmr_unlock(mreg, proc);
gmr_lock(mreg, proc);
}
switch(op) {
case ARMCII_OP_PUT:
gmr_put(mreg, src[i], dst[i], elem_count, proc);
break;
case ARMCII_OP_GET:
gmr_get(mreg, src[i], dst[i], elem_count, proc);
break;
case ARMCII_OP_ACC:
gmr_accumulate(mreg, src[i], dst[i], elem_count, type, proc);
break;
default:
ARMCII_Error("unknown operation (%d)", op);
return 1;
}
}
gmr_unlock(mreg, proc);
return 0;
}
/** One-sided accumulate operation.
*
* @param[in] datatype ARMCI data type for the accumulate operation (see armci.h)
* @param[in] scale Pointer for a scalar of type datatype that will be used to
* scale values in the source buffer
* @param[in] src Source address (remote)
* @param[in] dst Destination address (local)
* @param[in] bytes Number of bytes to transfer
* @param[in] proc Process id to target
* @return 0 on success, non-zero on failure
*/
int PARMCI_Acc(int datatype, void *scale, void *src, void *dst, int bytes, int proc) {
void *src_buf;
int count, type_size, scaled;
MPI_Datatype type;
gmr_t *src_mreg, *dst_mreg;
/* If NOGUARD is set, assume the buffer is not shared */
if (ARMCII_GLOBAL_STATE.shr_buf_method != ARMCII_SHR_BUF_NOGUARD)
src_mreg = gmr_lookup(src, ARMCI_GROUP_WORLD.rank);
else
src_mreg = NULL;
dst_mreg = gmr_lookup(dst, proc);
ARMCII_Assert_msg(dst_mreg != NULL, "Invalid remote pointer");
/* Prepare the input data: Apply scaling if needed and acquire the DLA lock if
* needed. We hold the DLA lock if (src_buf == src && src_mreg != NULL). */
scaled = ARMCII_Buf_acc_is_scaled(datatype, scale);
if (scaled) {
MPI_Alloc_mem(bytes, MPI_INFO_NULL, &src_buf);
ARMCII_Assert(src_buf != NULL);
ARMCII_Buf_acc_scale(src, src_buf, bytes, datatype, scale);
} else {
src_buf = src;
}
/* Check if we need to copy: user requested it or same mem region */
if ( (src_buf == src) /* buf_prepare didn't make a copy */
&& (ARMCII_GLOBAL_STATE.shr_buf_method == ARMCII_SHR_BUF_COPY || src_mreg == dst_mreg) )
{
MPI_Alloc_mem(bytes, MPI_INFO_NULL, &src_buf);
ARMCII_Assert(src_buf != NULL);
ARMCI_Copy(src, src_buf, bytes);
}
ARMCII_Acc_type_translate(datatype, &type, &type_size);
count = bytes/type_size;
ARMCII_Assert_msg(bytes % type_size == 0,
"Transfer size is not a multiple of the datatype size");
/* TODO: Support a local accumulate operation more efficiently */
gmr_accumulate(dst_mreg, src_buf, dst, count, type, proc);
gmr_flush(dst_mreg, proc, 1); /* flush_local */
if (src_buf != src)
MPI_Free_mem(src_buf);
return 0;
}
