int tMPI_Scan(void* sendbuf, void* recvbuf, int count,
tMPI_Datatype datatype, tMPI_Op op, tMPI_Comm comm)
{
struct tmpi_thread *cur=tMPI_Get_current();
int myrank=tMPI_Comm_seek_rank(comm, cur);
int N=tMPI_Comm_N(comm);
int prev=myrank - 1; /* my previous neighbor */
int next=myrank + 1; /* my next neighbor */
#ifdef TMPI_PROFILE
tMPI_Profile_count_start(cur);
#endif
#ifdef TMPI_TRACE
tMPI_Trace_print("tMPI_Scan(%p, %p, %d, %p, %p, %p)",
sendbuf, recvbuf, count, datatype, op, comm);
#endif
if (count==0)
return TMPI_SUCCESS;
if (!recvbuf)
{
return tMPI_Error(comm, TMPI_ERR_BUF);
}
if (sendbuf==TMPI_IN_PLACE)
{
sendbuf=recvbuf;
}
/* we set our send and recv buffers */
tMPI_Atomic_ptr_set(&(comm->reduce_sendbuf[myrank]),sendbuf);
tMPI_Atomic_ptr_set(&(comm->reduce_recvbuf[myrank]),recvbuf);
/* now wait for the previous rank to finish */
if (myrank > 0)
{
void *a, *b;
int ret;
#if defined(TMPI_PROFILE) && defined(TMPI_CYCLE_COUNT)
tMPI_Profile_wait_start(cur);
#endif
/* wait for the previous neighbor's data to be ready */
tMPI_Event_wait( &(comm->csync[myrank].events[prev]) );
tMPI_Event_process( &(comm->csync[myrank].events[prev]), 1);
#if defined(TMPI_PROFILE) && defined(TMPI_CYCLE_COUNT)
tMPI_Profile_wait_stop(cur, TMPIWAIT_Reduce);
#endif
#ifdef TMPI_DEBUG
printf("%d: scanning with %d \n", myrank, prev, iteration);
fflush(stdout);
#endif
/* now do the reduction */
if (prev > 0)
{
a = (void*)tMPI_Atomic_ptr_get(&(comm->reduce_recvbuf[prev]));
}
else
{
a = (void*)tMPI_Atomic_ptr_get(&(comm->reduce_sendbuf[prev]));
}
b = sendbuf;
if ((ret=tMPI_Reduce_run_op(recvbuf, a, b, datatype,
count, op, comm)) != TMPI_SUCCESS)
{
return ret;
}
/* signal to my previous neighbor that I'm done with the data */
tMPI_Event_signal( &(comm->csync[prev].events[prev]) );
}
else
{
if (sendbuf != recvbuf)
{
/* copy the data if this is rank 0, and not MPI_IN_PLACE */
memcpy(recvbuf, sendbuf, count*datatype->size);
}
}
if (myrank < N-1)
{
/* signal to my next neighbor that I have the data */
tMPI_Event_signal( &(comm->csync[next].events[myrank]) );
/* and wait for my next neighbor to finish */
tMPI_Event_wait( &(comm->csync[myrank].events[myrank]) );
tMPI_Event_process( &(comm->csync[myrank].events[myrank]), 1);
}
#if defined(TMPI_PROFILE) && defined(TMPI_CYCLE_COUNT)
tMPI_Profile_wait_start(cur);
#endif
/*tMPI_Barrier_wait( &(comm->barrier));*/
#if defined(TMPI_PROFILE)
/*tMPI_Profile_wait_stop(cur, TMPIWAIT_Reduce);*/
tMPI_Profile_count_stop(cur, TMPIFN_Scan);
#endif
return TMPI_SUCCESS;
}
/* this is the main comm creation function. All other functions that create
comms use this*/
int tMPI_Comm_split(tMPI_Comm comm, int color, int key, tMPI_Comm *newcomm)
{
int i, j;
int N = tMPI_Comm_N(comm);
volatile tMPI_Comm *newcomm_list;
volatile int colors[MAX_PREALLOC_THREADS]; /* array with the colors
of each thread */
volatile int keys[MAX_PREALLOC_THREADS]; /* same for keys (only one of
the threads actually suplies
these arrays to the comm
structure) */
tmpi_bool i_am_first = FALSE;
int myrank = tMPI_Comm_seek_rank(comm, tMPI_Get_current());
struct tmpi_split *spl;
int ret;
#ifdef TMPI_TRACE
tMPI_Trace_print("tMPI_Comm_split(%p, %d, %d, %p)", comm, color, key,
newcomm);
#endif
if (!comm)
{
*newcomm = NULL;
return tMPI_Error(TMPI_COMM_WORLD, TMPI_ERR_COMM);
}
ret = tMPI_Thread_mutex_lock(&(comm->comm_create_lock));
if (ret != 0)
{
return tMPI_Error(TMPI_COMM_WORLD, TMPI_ERR_IO);
}
/* first get the colors */
if (!comm->new_comm)
{
/* i am apparently first */
comm->split = (struct tmpi_split*)tMPI_Malloc(sizeof(struct tmpi_split));
comm->new_comm = (tMPI_Comm*)tMPI_Malloc(N*sizeof(tMPI_Comm));
if (N <= MAX_PREALLOC_THREADS)
{
comm->split->colors = colors;
comm->split->keys = keys;
}
else
{
comm->split->colors = (int*)tMPI_Malloc(N*sizeof(int));
comm->split->keys = (int*)tMPI_Malloc(N*sizeof(int));
}
comm->split->Ncol_init = tMPI_Comm_N(comm);
comm->split->can_finish = FALSE;
i_am_first = TRUE;
/* the main communicator contains a list the size of grp.N */
}
newcomm_list = comm->new_comm; /* we copy it to the local stacks because
we can later erase comm->new_comm safely */
spl = comm->split; /* we do the same for spl */
spl->colors[myrank] = color;
spl->keys[myrank] = key;
spl->Ncol_init--;
if (spl->Ncol_init == 0)
{
ret = tMPI_Thread_cond_signal(&(comm->comm_create_prep));
if (ret != 0)
{
return tMPI_Error(TMPI_COMM_WORLD, TMPI_ERR_IO);
}
}
if (!i_am_first)
{
/* all other threads can just wait until the creator thread is
finished */
while (!spl->can_finish)
{
ret = tMPI_Thread_cond_wait(&(comm->comm_create_finish),
&(comm->comm_create_lock) );
if (ret != 0)
{
return tMPI_Error(TMPI_COMM_WORLD, TMPI_ERR_IO);
}
}
}
else
{
int Ncomms = 0;
int comm_color_[MAX_PREALLOC_THREADS];
int comm_N_[MAX_PREALLOC_THREADS];
int *comm_color = comm_color_; /* there can't be more comms than N*/
int *comm_N = comm_N_; /* the number of procs in a group */
int *comm_groups; /* the groups */
tMPI_Comm *comms; /* the communicators */
/* wait for the colors to be done */
/*if (N>1)*/
while (spl->Ncol_init > 0)
{
ret = tMPI_Thread_cond_wait(&(comm->comm_create_prep),
&(comm->comm_create_lock));
if (ret != 0)
{
return tMPI_Error(TMPI_COMM_WORLD, TMPI_ERR_IO);
}
}
/* reset the state so that a new comm creating function can run */
spl->Ncol_destroy = N;
comm->new_comm = 0;
comm->split = 0;
comm_groups = (int*)tMPI_Malloc(N*N*sizeof(int));
if (N > MAX_PREALLOC_THREADS)
{
comm_color = (int*)tMPI_Malloc(N*sizeof(int));
comm_N = (int*)tMPI_Malloc(N*sizeof(int));
}
/* count colors, allocate and split up communicators */
tMPI_Split_colors(N, (int*)spl->colors,
(int*)spl->keys,
&Ncomms,
comm_N, comm_color, comm_groups);
/* allocate a bunch of communicators */
comms = (tMPI_Comm*)tMPI_Malloc(Ncomms*sizeof(tMPI_Comm));
for (i = 0; i < Ncomms; i++)
{
ret = tMPI_Comm_alloc(&(comms[i]), comm, comm_N[i]);
if (ret != TMPI_SUCCESS)
{
return ret;
}
}
/* now distribute the comms */
for (i = 0; i < Ncomms; i++)
{
comms[i]->grp.N = comm_N[i];
for (j = 0; j < comm_N[i]; j++)
{
comms[i]->grp.peers[j] =
comm->grp.peers[comm_groups[i*comm->grp.N + j]];
}
}
/* and put them into the newcomm_list */
for (i = 0; i < N; i++)
{
newcomm_list[i] = TMPI_COMM_NULL;
for (j = 0; j < Ncomms; j++)
{
if (spl->colors[i] == comm_color[j])
{
newcomm_list[i] = comms[j];
break;
}
}
}
#ifdef TMPI_DEBUG
/* output */
for (i = 0; i < Ncomms; i++)
{
printf("Group %d (color %d) has %d members: ",
i, comm_color[i], comm_N[i]);
for (j = 0; j < comm_N[i]; j++)
{
printf(" %d ", comm_groups[comm->grp.N*i + j]);
}
printf(" rank: ");
for (j = 0; j < comm_N[i]; j++)
{
printf(" %d ", spl->keys[comm_groups[N*i + j]]);
}
printf(" color: ");
for (j = 0; j < comm_N[i]; j++)
{
printf(" %d ", spl->colors[comm_groups[N*i + j]]);
}
printf("\n");
}
#endif
if (N > MAX_PREALLOC_THREADS)
{
free((int*)spl->colors);
free((int*)spl->keys);
free(comm_color);
free(comm_N);
}
free(comm_groups);
free(comms);
spl->can_finish = TRUE;
/* tell the waiting threads that there's a comm ready */
ret = tMPI_Thread_cond_broadcast(&(comm->comm_create_finish));
if (ret != 0)
{
return tMPI_Error(TMPI_COMM_WORLD, TMPI_ERR_IO);
}
}
/* here the individual threads get their comm object */
*newcomm = newcomm_list[myrank];
/* free when we have assigned them all, so we can reuse the object*/
spl->Ncol_destroy--;
if (spl->Ncol_destroy == 0)
{
free((void*)newcomm_list);
free(spl);
}
ret = tMPI_Thread_mutex_unlock(&(comm->comm_create_lock));
if (ret != 0)
{
return tMPI_Error(TMPI_COMM_WORLD, TMPI_ERR_IO);
}
return TMPI_SUCCESS;
}
