/** Perform atomic read-modify-write on the given integer or long location and
* return the location's original value.
*
* \note ARMCI RMW operations are atomic with respect to other RMW operations,
* but not with respect to other one-sided operations (get, put, acc, etc).
*
* @param[in] op Operation to be performed:
* ARMCI_FETCH_AND_ADD (int)
* ARMCI_FETCH_AND_ADD_LONG
* ARMCI_SWAP (int)
* ARMCI_SWAP_LONG
* @param[out] ploc Location to store the original value.
* @param[in] prem Location on which to perform atomic operation.
* @param[in] value Value to add to remote location (ignored for swap).
* @param[in] proc Process rank for the target buffer.
*/
int PARMCI_Rmw(int op, void *ploc, void *prem, int value, int proc) {
int is_long;
gmr_t *mreg;
mreg = gmr_lookup(prem, proc);
ARMCII_Assert_msg(mreg != NULL, "Invalid remote pointer");
if (op == ARMCI_SWAP_LONG || op == ARMCI_FETCH_AND_ADD_LONG)
is_long = 1;
else
is_long = 0;
if (op == ARMCI_SWAP || op == ARMCI_SWAP_LONG) {
long swap_val_l;
int swap_val_i;
ARMCIX_Lock_hdl(mreg->rmw_mutex, 0, proc);
PARMCI_Get(prem, is_long ? (void*) &swap_val_l : (void*) &swap_val_i,
is_long ? sizeof(long) : sizeof(int), proc);
PARMCI_Put(ploc, prem, is_long ? sizeof(long) : sizeof(int), proc);
ARMCIX_Unlock_hdl(mreg->rmw_mutex, 0, proc);
if (is_long)
*(long*) ploc = swap_val_l;
else
*(int*) ploc = swap_val_i;
}
else if (op == ARMCI_FETCH_AND_ADD || op == ARMCI_FETCH_AND_ADD_LONG) {
long fetch_val_l, new_val_l;
int fetch_val_i, new_val_i;
ARMCIX_Lock_hdl(mreg->rmw_mutex, 0, proc);
PARMCI_Get(prem, is_long ? (void*) &fetch_val_l : (void*) &fetch_val_i,
is_long ? sizeof(long) : sizeof(int), proc);
if (is_long)
new_val_l = fetch_val_l + value;
else
new_val_i = fetch_val_i + value;
PARMCI_Put(is_long ? (void*) &new_val_l : (void*) &new_val_i, prem,
is_long ? sizeof(long) : sizeof(int), proc);
ARMCIX_Unlock_hdl(mreg->rmw_mutex, 0, proc);
if (is_long)
*(long*) ploc = fetch_val_l;
else
*(int*) ploc = fetch_val_i;
}
else {
ARMCII_Error("invalid operation (%d)", op);
}
return 0;
}
int PARMCI_Acc(int datatype, void *scale,
void *src_ptr,
void *dst_ptr,
int bytes, int proc)
{
double *get_buf = (double *)l_state.acc_buf;
double *_src_buf = (double *)src_ptr;
double calc_scale = *(double *)scale;
int m, limit;
assert(bytes <= l_state.acc_buf_len);
assert(datatype == ARMCI_ACC_DBL);
assert(get_buf);
ARMCID_network_lock(proc);
PARMCI_Get(dst_ptr, get_buf, bytes, proc);
for (m=0, limit=bytes/sizeof(double); m<limit; ++m) {
if (calc_scale == 1.0) {
get_buf[m] += _src_buf[m];
}
else {
get_buf[m] += calc_scale * _src_buf[m];
}
}
PARMCI_Put(get_buf, dst_ptr, bytes, proc);
ARMCID_network_unlock(proc);
return 0;
}
static void armci_generic_lock(int mutex, int proc)
{
int i, myturn, factor=0, len=sizeof(int);
int *mutex_ticket, next_in_line;
mutex_ticket= glob_mutex[proc].turn + mutex;
myturn = register_in_mutex_queue(mutex, proc);
/* code to reduce cost of unlocking mutex on the same SMP node goes here
* lockinfo_node[me].ticket = mutex_ticket;
* lockinfo_node[me].mutex = mutex;
*/
_dummy_work_ = 0.; /* must be global to fool the compiler */
do {
PARMCI_Get(mutex_ticket, &next_in_line, len, proc);
if(next_in_line > myturn)
armci_die2("armci: problem with tickets",myturn,next_in_line);
/* apply a linear backoff delay before retrying */
for(i=0; i< SPINMAX * factor; i++) _dummy_work_ += 1.;
factor += 1;
} while (myturn != next_in_line);
glob_mutex[proc].tickets[mutex] = myturn; /* save ticket value */
}
int ARMCI_Get(void *src, void *dst, int bytes, int proc)
{
int ret;
armci_profile_start_strided(&bytes, 0, proc, ARMCI_PROF_GET);
ret = PARMCI_Get(src, dst, bytes, proc);
armci_profile_stop_strided(ARMCI_PROF_GET);
return ret;
}
int ARMCI_Get(void *src, void *dst, int bytes, int proc)
{
int rval;
static double stime, etime;
stime = TIME();
rval = PARMCI_Get(src, dst, bytes, proc);
etime = TIME();
ARMCI_Get_t += etime - stime;
return rval;
}
int ARMCI_Get(void *src, void *dst, int size, int target) {
return PARMCI_Get(src, dst, size, target);
}
int ARMCI_Get(void *src, void *dst, int size, int target) {
parmci_calls++;
return PARMCI_Get(src, dst, size, target);
}
int main(int argc, char *argv[])
{
int rank, size;
int provided;
#if defined(__bgp__)
MPI_Init_thread(&argc, &argv, MPI_THREAD_MULTIPLE, &provided);
assert(provided==MPI_THREAD_MULTIPLE);
#else
MPI_Init_thread(&argc, &argv, MPI_THREAD_FUNNELED, &provided);
//assert(provided>MPI_THREAD_SINGLE);
#endif
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Comm_size(MPI_COMM_WORLD, &size);
assert( size > 1 );
PARMCI_Init_args(&argc, &argv);
int w, maxwinsize = ( argc > 1 ? atoi(argv[1]) : 1000000 );
if ( rank == 0 ) printf( "size = %d maxwinsize = %d doubles\n", size, maxwinsize );
for ( w = 1 ; w < maxwinsize ; w *= 2 )
{
double ** window;
window = (double **) PARMCI_Malloc_local( size * sizeof(double *) );
PARMCIX_Malloc_comm(MPI_COMM_WORLD, (void **) window, w * sizeof(double) );
for (int i = 0; i < w; i++) window[rank][i] = 0.0;
double * buffer;
buffer = (double *) PARMCI_Malloc_local( w * sizeof(double) );
PARMCIX_Barrier_comm(MPI_COMM_WORLD);
if (rank == 0)
for (int t=1; t<size; t+=2)
{
int bytes = w * sizeof(double);
for (int i = 0; i < w; i++) buffer[i] = (double)(t);
PARMCI_Put( buffer, window[t], bytes, t );
PARMCI_Fence( t );
for (int i = 0; i < w; i++) buffer[i] = 0.0;
PARMCI_Get( window[t], buffer, bytes, t );
int errors = 0;
for (int i = 0; i < w; i++)
if ( buffer[i] != (double)(t) ) errors++;
if ( errors > 0 )
for (int i = 0; i < w; i++)
printf("rank %d buffer[%d] = %lf \n", rank, i, buffer[i] );
}
PARMCIX_Barrier_comm(MPI_COMM_WORLD);
if (rank != 0)
{
int errors = 0;
for (int i = 0; i < w; i++)
if ( window[rank][i] != (double)(rank) ) errors++;
if ( errors > 0 )
for (int i = 0; i < w; i++)
printf("rank %d window[%d][%d] = %lf \n", rank, rank, i, window[rank][i] );
}
PARMCIX_Barrier_comm(MPI_COMM_WORLD);
if (rank == 0)
for (int t=1; t<size; t++)
{
int bytes = w * sizeof(double);
double t0, t1, t2, dt1, dt2, bw1, bw2;
for (int i = 0; i < w; i++) buffer[i] = (double)(-1);
t0 = MPI_Wtime();
PARMCI_Put( buffer, window[t], bytes, t );
t1 = MPI_Wtime();
PARMCI_Fence( t );
t2 = MPI_Wtime();
dt1 = t1 - t0;
dt2 = t2 - t0;
bw1 = bytes / dt1;
bw2 = bytes / dt2;
bw1 /= 1000000.0;
bw2 /= 1000000.0;
printf("PARMCI_Put of from rank %4d to rank %4d of %9d bytes - local: %lf s (%lf MB/s) remote: %lf s (%lf MB/s) \n",
t, 0, bytes, dt1, bw1, dt2, bw2);
fflush(stdout);
}
PARMCIX_Barrier_comm(MPI_COMM_WORLD);
PARMCI_Free_local( (void *) buffer );
PARMCIX_Free_comm(MPI_COMM_WORLD, (void *) window[rank] );
PARMCI_Free_local( (void *) window );
}
PARMCI_Finalize();
printf("%d: all done \n", rank );
fflush(stdout);
MPI_Finalize();
return 0;
}
int PARMCI_NbGet(void *src, void *dst, int bytes, int proc, armci_hdl_t *hdl)
{
int rc;
rc = PARMCI_Get(src, dst, bytes, proc);
return 0;
}
int armci_pack_vector(int op, void *scale, armci_giov_t darr[],int len,
int proc,armci_ihdl_t nb_handle)
{
armci_giov_t extra; /* keeps data remainder of set to be processed in chunks */
armci_giov_t save; /* keeps original value of set to be processed in chunks */
armci_giov_t *ndarr; /* points to first array element to be processed now */
int rc=0, nlen, count=0;
armcip_init_giov_t(&extra);
armcip_init_giov_t(&save);
ndarr = darr;
save.src_ptr_array=NULL; /* indicates that save slot is empty */
while(len){
armci_split_dscr_array(ndarr, len, &extra, &nlen, &save);
# if defined(REMOTE_OP)
/* A problem will occur if len is 1 and nlen is 0. This corresponds to a
* situation where the size of an individual element is found to exceed
* BUFSIZE1. Treat this as a single transfer of contiguous data using
* the standard PARMCI_Get/Put/Acc call */
if (len == 1 && nlen == 0) {
if(ARMCI_ACC(op))rc=PARMCI_Acc(op, scale, ndarr[0].src_ptr_array[0],
ndarr[0].dst_ptr_array[0],ndarr[0].bytes, proc);
else if(op == GET)rc=PARMCI_Get(ndarr[0].src_ptr_array[0],
ndarr[0].dst_ptr_array[0],ndarr[0].bytes, proc);
else if(op == PUT)rc=PARMCI_Put(ndarr[0].src_ptr_array[0],
ndarr[0].dst_ptr_array[0],ndarr[0].bytes, proc);
else armci_die("Unknown op in armci_pack_vector",op);
nlen = 1;
} else {
rc = armci_rem_vector(op, scale, ndarr,nlen,proc,0,nb_handle);
}
# else
if(ARMCI_ACC(op))rc=armci_acc_vector(op,scale,ndarr,nlen,proc);
else rc = armci_copy_vector(op,ndarr,nlen,proc);
# endif
if(rc) break;
/* non-NULL pointer indicates that set was split */
if(extra.src_ptr_array){
if(nb_handle) {
nb_handle->bufid = NB_MULTI; /*can be set multiple times here; but not reset here*/
}
ndarr[nlen-1]=extra; /* set the pointer to remainder of last set */
nlen--; /* since last set not done in full need to process it again */
}else{
if(save.src_ptr_array){
ndarr[0]=save;
save.src_ptr_array=NULL; /* indicates that save slot is empty */
}
if(nlen == 0)
armci_die("vector packetization problem:buffer too small",BUFSIZE1);
}
len -=nlen;
ndarr +=nlen;
count ++;
}
return rc;
}
