int PARMCI_GetV( armci_giov_t darr[], /* descriptor array */
int len, /* length of descriptor array */
int proc /* remote process(or) ID */
)
{
int rc=0, i,direct=1;
if(len<1) return FAIL;
for(i=0;i<len;i++){
if(darr[i].src_ptr_array==NULL ||darr[i].dst_ptr_array==NULL)return FAIL2;
if(darr[i].bytes<1)return FAIL3;
if(darr[i].ptr_array_len <1) return FAIL4;
}
if(proc<0 || proc >= armci_nproc)return FAIL5;
ORDER(GET,proc); /* ensure ordering */
#ifndef QUADRICS
direct=SAMECLUSNODE(proc);
#endif
if(direct){
if(!SAMECLUSNODE(proc))DO_FENCE(proc,DIRECT_GET);
rc = armci_copy_vector(GET, darr, len, proc);
}
else{
DO_FENCE(proc,SERVER_GET);
rc = armci_pack_vector(GET, NULL, darr, len, proc,NULL);
}
if(rc) return FAIL6;
else return 0;
}
void ARMCI_Fence(int proc)
{
if (!SAMECLUSNODE(proc))
armci_profile_start(ARMCI_PROF_FENCE);
PARMCI_Fence(proc);
if (!SAMECLUSNODE(proc))
armci_profile_stop(ARMCI_PROF_FENCE);
}
/*\ Non-blocking vector API
\*/
int PARMCI_NbPutV( armci_giov_t darr[], /* descriptor array */
int len, /* length of descriptor array */
int proc, /* remote process(or) ID */
armci_hdl_t* usr_hdl /*non-blocking request handle*/
)
{
armci_ihdl_t nb_handle = (armci_ihdl_t)usr_hdl;
int rc=0, i,direct=1;
if(len<1) return FAIL;
for(i=0;i<len;i++){
if(darr[i].src_ptr_array == NULL || darr[i].dst_ptr_array ==NULL) return FAIL2;
if(darr[i].bytes<1)return FAIL3;
if(darr[i].ptr_array_len <1) return FAIL4;
}
if(proc<0 || proc >= armci_nproc)return FAIL5;
direct=SAMECLUSNODE(proc);
/* aggregate put */
if(nb_handle && nb_handle->agg_flag == SET) {
if(!direct) {
rc=armci_agg_save_giov_descriptor(darr, len, proc, PUT, nb_handle);
return rc;
}
}
else {
/*ORDER(PUT,proc); ensure ordering */
UPDATE_FENCE_INFO(proc);
/*set tag and op in the nb handle*/
if(nb_handle){
nb_handle->tag = GET_NEXT_NBTAG();
nb_handle->op = PUT;
nb_handle->proc= proc;
nb_handle->bufid=NB_NONE;
}
else
nb_handle = armci_set_implicit_handle(PUT, proc);
}
if(direct){
if(!SAMECLUSNODE(proc))DO_FENCE(proc,DIRECT_PUT);
rc = armci_copy_vector(PUT, darr, len, proc);
}
else{
DO_FENCE(proc,SERVER_NBPUT);
rc = armci_pack_vector(PUT, NULL, darr, len, proc,nb_handle);
}
if(rc) return FAIL6;
else return 0;
}
int PARMCI_AccV( int op, /* oeration code */
void *scale, /*scaling factor for accumulate */
armci_giov_t darr[], /* descriptor array */
int len, /* length of descriptor array */
int proc /* remote process(or) ID */
)
{
int rc=0, i,direct=0;
if(len<1) return FAIL;
for(i=0;i<len;i++){
if(darr[i].src_ptr_array==NULL ||darr[i].dst_ptr_array==NULL)return FAIL2;
if(darr[i].bytes<1)return FAIL3;
if(darr[i].ptr_array_len <1) return FAIL4;
}
if(proc<0 || proc >= armci_nproc)return FAIL5;
ORDER(op,proc); /* ensure ordering */
direct=SAMECLUSNODE(proc);
# if defined(ACC_COPY) && !defined(ACC_SMP)
if(armci_me != proc) direct=0;
# error "grrr"
# endif
if(direct) {
rc = armci_acc_vector( op, scale, darr, len, proc);
} else {
DO_FENCE(proc,SERVER_PUT);
rc = armci_pack_vector(op, scale, darr, len, proc,NULL);
}
if(rc) return FAIL6;
else return 0;
}
void PARMCI_Lock(int mutex, int proc)
{
#if defined(SERVER_LOCK)
int direct;
#endif
if(DEBUG)fprintf(stderr,"%d enter lock\n",armci_me);
if(!num_mutexes) armci_die("armci_lock: create mutexes first",0);
if(mutex > glob_mutex[proc].count)
armci_die2("armci_lock: mutex not allocated", mutex,
glob_mutex[proc].count);
if(armci_nproc == 1) return;
# if defined(SERVER_LOCK)
direct=SAMECLUSNODE(proc);
if(!direct)
armci_rem_lock(mutex,proc, glob_mutex[proc].tickets + mutex );
else
# endif
armci_generic_lock(mutex,proc);
if(DEBUG)fprintf(stderr,"%d leave lock\n",armci_me);
}
void ARMCI_DoFence(int proc)
{
int i;
if(!SAMECLUSNODE(proc) && (armci_nclus >1)){
int cluster = armci_clus_id(proc);
armci_rem_ack(cluster);
}
}
/* Check whether the oricess is in the same domain */
int armci_domain_same_id (armci_domain_t domain, int proc)
{
#if 0
int rc = SAMECLUSNODE(proc);
return rc;
#else
assert(0);
#endif
}
int PARMCI_NbAccV( int op, /* oeration code */
void *scale, /*scaling factor for accumulate */
armci_giov_t darr[], /* descriptor array */
int len, /* length of descriptor array */
int proc, /* remote process(or) ID */
armci_hdl_t* usr_hdl /*non-blocking request handle*/
)
{
armci_ihdl_t nb_handle = (armci_ihdl_t)usr_hdl;
int rc=0, i,direct=1;
if(len<1) return FAIL;
for(i=0;i<len;i++)
{
if(darr[i].src_ptr_array==NULL ||darr[i].dst_ptr_array==NULL)return FAIL2;
if(darr[i].bytes<1)return FAIL3;
if(darr[i].ptr_array_len <1) return FAIL4;
}
if(proc<0 || proc >= armci_nproc)return FAIL5;
UPDATE_FENCE_INFO(proc);
direct=SAMECLUSNODE(proc);
if(nb_handle){
nb_handle->tag = GET_NEXT_NBTAG();
nb_handle->op = op;
nb_handle->proc= proc;
nb_handle->bufid=NB_NONE;
}
else
nb_handle = armci_set_implicit_handle(op, proc);
# if defined(ACC_COPY) && !defined(ACC_SMP)
if(armci_me != proc) direct=0;
# endif
if(direct)
rc = armci_acc_vector( op, scale, darr, len, proc);
else{
DO_FENCE(proc,SERVER_NBPUT);
rc = armci_pack_vector(op, scale, darr, len, proc,nb_handle);
}
if(rc) return FAIL6;
else return 0;
}
int armci_copy_vector(int op, /* operation code */
armci_giov_t darr[], /* descriptor array */
int len, /* length of descriptor array */
int proc /* remote process(or) ID */
)
{
int i,s,shmem= SAMECLUSNODE(proc);
int armci_th_idx = ARMCI_THREAD_IDX;
if(shmem){
/* local/shared memory copy */
for(i = 0; i< len; i++){
for( s=0; s< darr[i].ptr_array_len; s++){
armci_copy(darr[i].src_ptr_array[s],darr[i].dst_ptr_array[s],darr[i].bytes);
}
}
}else {
switch(op){
case PUT:
for(i = 0; i< len; i++){
UPDATE_FENCE_STATE(proc, PUT, darr[i].ptr_array_len);
for( s=0; s< darr[i].ptr_array_len; s++){
armci_put(darr[i].src_ptr_array[s],darr[i].dst_ptr_array[s],
darr[i].bytes, proc);
}
}
break;
case GET:
for(i = 0; i< len; i++){
for( s=0; s< darr[i].ptr_array_len; s++){
armci_get(darr[i].src_ptr_array[s],darr[i].dst_ptr_array[s],
darr[i].bytes,proc);
}
}
break;
default:
armci_die("armci_copy_vector: wrong optype",op);
}
}
return 0;
}
int ARMCI_Rmw(int op, int *ploc, int *prem, int extra, int proc)
{
if(!SAMECLUSNODE(proc)){
armci_rem_rmw(op, ploc, prem, extra, proc);
return 0;
}
switch (op) {
case ARMCI_FETCH_AND_ADD:
case ARMCI_FETCH_AND_ADD_LONG:
case ARMCI_SWAP:
case ARMCI_SWAP_LONG:
armci_generic_rmw(op, ploc, prem, extra, proc);
break;
default: armci_die("rmw: operation not supported",op);
}
return 0;
}
int armci_gpc_local_exec(int h, int to, int from, void *hdr, int hlen,
void *data, int dlen,
void *rhdr, int rhlen,
void *rdata, int rdlen, int rtype) {
int rhsize, rdsize;
int (*func)();
int hnd = -h + GPC_OFFSET;
if(hnd <0 || hnd>= GPC_SLOTS)
armci_die2("armci_gpc_local_exec: bad callback handle",hnd,GPC_SLOTS);
if(!_table[hnd]) armci_die("armci_gpc_local_exec: NULL function",hnd);
func = _table[hnd];
if(!SAMECLUSNODE(to))
armci_die("armci_gpc_local_exec: GPC call to a different node received!",
armci_me);
/* func(to, from, hdr, hlen, data, dlen, rhdr, rhlen, &rhsize, */
/* rdata, rdlen, &rdsize); */
/* return 0; */
return func(to, from, hdr, hlen, data, dlen, rhdr, rhlen, &rhsize,
rdata, rdlen, &rdsize, rtype);
}
/*\ Send Request to Execute callback function in a global address space
* Arguments:
* f - handle to the callback function
* p - remote processor
* hdr - header data - used to pack extra args for callback (local buffer)
* hlen - size of header data < ARMCI_GPC_HLEN
* data - bulk data passed to callback (local buffer)
* dlen - length of bulk data
* rhdr - ptr to reply header (return args from callback)
* rhlen - length of buffer to store reply header < ARMCI_GPC_HLEN
* rdata - ptr to where reply data from callback should be stored (local buf)
* rdlen - size of the buffer to store reply data
* nbh - nonblocking handle
*
\*/
int ARMCI_Gpc_exec(int h, int p, void *hdr, int hlen, void *data, int dlen,
void *rhdr, int rhlen, void *rdata, int rdlen, gpc_hdl_t* nbh)
{
int hnd = -h + GPC_OFFSET;
int err = 0;
armci_hdl_t *ahdl = (nbh ? &(nbh->ahdl): NULL);
if(hnd <0 || hnd>= GPC_SLOTS)
err += fprintf(stderr, "ARMCI_Gpc_exec: bad callback handle %d: %d\n",hnd,GPC_SLOTS);
if(!_table[hnd])
err += fprintf(stderr, "ARMCI_Gpc_exec: NULL function %d",hnd);
if(hlen<0 || hlen>=ARMCI_Gpc_get_hlen())
err += fprintf(stderr, "ARMCI_Gpc_exec: Invalid send header size %d %d\n", hlen, ARMCI_Gpc_get_hlen());
if(rhlen<0 || rhlen>=ARMCI_Gpc_get_hlen())
err += fprintf(stderr, "ARMCI_Gpc_exec: Invalid recv header size %d %d\n", rhlen, ARMCI_Gpc_get_hlen());
if(dlen<0 || dlen>=ARMCI_Gpc_get_dlen())
err += fprintf(stderr, "ARMCI_Gpc_exec: Invalid send data size %d %d\n", dlen, ARMCI_Gpc_get_dlen());
if(rdlen<0 || rdlen>=ARMCI_Gpc_get_dlen())
err += fprintf(stderr, "ARMCI_Gpc_exec: Invalid recv data size %d %d\n", rdlen, ARMCI_Gpc_get_dlen());
if(hlen>0 && hdr==NULL)
err += fprintf(stderr, "ARMCI_Gpc_exec: Null send header for non-zero header size %d\n", hlen);
if(rhlen>0 && rhdr==NULL)
err += fprintf(stderr, "ARMCI_Gpc_exec: Null recv header for non-zero header size %d\n", rhlen);
if(dlen>0 && data==NULL)
err += fprintf(stderr, "ARMCI_Gpc_exec: Null send data for non-zero data size %d\n", dlen);
if(rdlen>0 && rdata==NULL)
err += fprintf(stderr, "ARMCI_Gpc_exec: Null recv data for non-zero header size %d\n", rdlen);
if(p<0 || p >= armci_nproc)
err += fprintf(stderr, "ARMCI_Gpc_exec: Invalid target processor id %d\n", p, armci_nproc);
if(err)
return FAIL;
if(rhlen + rdlen == 0)
armci_die("Zero reply header + data length not yet supported", 0);
if(nbh)
nbh->proc = p;
#if 1
if(SAMECLUSNODE(p) && armci_nproc==1) {
int rhsize, rdsize;
int (*func)();
/* fprintf(stderr, "%d:: armci gpc exec. SAMECLUSNODE\n", armci_me); */
func = _table[hnd];
if(func(p, armci_me, hdr, hlen, data, dlen, rhdr, rhlen, &rhsize,
rdata, rdlen, &rdsize, GPC_INIT) != GPC_DONE) {
func(p, armci_me, hdr, hlen, data, dlen, rhdr, rhlen, &rhsize,
rdata, rdlen, &rdsize, GPC_WAIT);
}
#ifndef VAPI
PARMCI_Fence(p);
#endif
return 0;
}
#endif
/* fprintf(stderr, "%d:: armci gpc exec. invoking armci gpc\n", armci_me); */
return armci_gpc(h, p, hdr, hlen, data, dlen,
rhdr, rhlen, rdata, rdlen, ahdl);
}
void ARMCI_Gpc_test(gpc_hdl_t *nbh) {
if(SAMECLUSNODE(nbh->proc))
return;
PARMCI_Test(&nbh->ahdl);
}
int armci_domain_same_id (armci_domain_t domain, int proc)
{
int rc = SAMECLUSNODE(proc);
return(rc);
}
void ARMCI_Gpc_wait(gpc_hdl_t *nbh) {
if(SAMECLUSNODE(nbh->proc))
return;
ARMCI_Wait(&nbh->ahdl);
}
