/** One-sided get operation.
*
* @param[in] src Source address (remote)
* @param[in] dst Destination address (local)
* @param[in] size Number of bytes to transfer
* @param[in] target Process id to target
* @return 0 on success, non-zero on failure
*/
int ARMCI_Get(void *src, void *dst, int size, int target) {
gmr_t *src_mreg, *dst_mreg;
src_mreg = gmr_lookup(src, target);
dst_mreg = gmr_lookup(dst, ARMCI_GROUP_WORLD.rank);
ARMCII_Assert_msg(src_mreg != NULL, "Invalid remote pointer");
/* Local operation */
if (target == ARMCI_GROUP_WORLD.rank) {
if (ARMCII_GLOBAL_STATE.shr_buf_method != ARMCII_SHR_BUF_NOGUARD) {
if (dst_mreg) gmr_dla_lock(dst_mreg); /* FIXME: Is this a hold-while wait? Probably need an extra copy to be safe.. */
gmr_dla_lock(src_mreg);
}
ARMCI_Copy(src, dst, size);
if (ARMCII_GLOBAL_STATE.shr_buf_method != ARMCII_SHR_BUF_NOGUARD) {
if (dst_mreg) gmr_dla_unlock(dst_mreg);
gmr_dla_unlock(src_mreg);
}
}
/* Origin buffer is private */
else if (dst_mreg == NULL || ARMCII_GLOBAL_STATE.shr_buf_method == ARMCII_SHR_BUF_NOGUARD) {
gmr_lock(src_mreg, target);
gmr_get(src_mreg, src, dst, size, target);
gmr_unlock(src_mreg, target);
}
/* COPY: Either origin and target buffers are in the same window and we can't
* lock the same window twice (MPI semantics) or the user has requested
* always-copy mode. */
else {
void *dst_buf;
MPI_Alloc_mem(size, MPI_INFO_NULL, &dst_buf);
ARMCII_Assert(dst_buf != NULL);
gmr_lock(src_mreg, target);
gmr_get(src_mreg, src, dst_buf, size, target);
gmr_unlock(src_mreg, target);
gmr_dla_lock(dst_mreg);
ARMCI_Copy(dst_buf, dst, size);
MPI_Free_mem(dst_buf);
gmr_dla_unlock(dst_mreg);
}
return 0;
}
/** 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 get operation.
*
* @param[in] src Source address (remote)
* @param[in] dst Destination address (local)
* @param[in] size Number of bytes to transfer
* @param[in] target Process id to target
* @return 0 on success, non-zero on failure
*/
int PARMCI_Get(void *src, void *dst, int size, int target) {
gmr_t *src_mreg, *dst_mreg;
src_mreg = gmr_lookup(src, target);
/* If NOGUARD is set, assume the buffer is not shared */
if (ARMCII_GLOBAL_STATE.shr_buf_method != ARMCII_SHR_BUF_NOGUARD)
dst_mreg = gmr_lookup(dst, ARMCI_GROUP_WORLD.rank);
else
dst_mreg = NULL;
ARMCII_Assert_msg(src_mreg != NULL, "Invalid remote pointer");
/* Local operation */
if (target == ARMCI_GROUP_WORLD.rank && dst_mreg == NULL) {
ARMCI_Copy(src, dst, size);
}
/* Origin buffer is private */
else if (dst_mreg == NULL) {
gmr_get(src_mreg, src, dst, size, target);
gmr_flush(src_mreg, target, 0); /* it's a round trip so w.r.t. flush, local=remote */
}
/* COPY: Either origin and target buffers are in the same window and we can't
* lock the same window twice (MPI semantics) or the user has requested
* always-copy mode. */
else {
void *dst_buf;
MPI_Alloc_mem(size, MPI_INFO_NULL, &dst_buf);
ARMCII_Assert(dst_buf != NULL);
gmr_get(src_mreg, src, dst_buf, size, target);
gmr_flush(src_mreg, target, 0); /* it's a round trip so w.r.t. flush, local=remote */
ARMCI_Copy(dst_buf, dst, size);
MPI_Free_mem(dst_buf);
}
return 0;
}
