void F77_FUNC(mpix_comm_rank2global,MPIX_COMM_RANK2GLOBAL)(MPI_Fint * comm, MPI_Fint * crank, MPI_Fint * grank, MPI_Fint * ierr)
{
MPI_Comm mycomm = MPI_Comm_f2c(*comm);
int mycrank = -1, mygrank = -1;
safecast(*crank, &mycrank);
*ierr = (MPI_Fint)MPIX_Comm_rank2global(mycomm, mycrank, &mygrank);
*grank = (MPI_Fint)mygrank;
return;
}
static void
QMP_declare_multiple_bgspi(QMP_msghandle_t mh)
{
if(mh->paired) {
int npaired = 0;
int src0 = -1;
while(1) {
int src = -1;
for(QMP_msghandle_t m=mh->next; m; m=m->next) {
if(src<=src0 || (m->srce_node>src0 && m->srce_node<src)) src = m->srce_node;
}
if(src<=src0) break;
for(QMP_msghandle_t m=mh->next; m; m=m->next) {
if(m->srce_node==src) {
int nmsg = 0, nmsgo;
switch(m->mm->type) {
case MM_user_buf: nmsg = 1; break;
case MM_strided_buf:
if(m->type==MH_send) nmsg = m->mm->mm.st.nblocks;
break;
case MM_strided_array_buf:
if(m->type==MH_send) {
for(int i=0; i<m->mm->mm.sa.narray; i++) {
nmsg += m->mm->mm.sa.nblocks[i];
}
}
break;
case MM_indexed_buf:
if(m->type==MH_send) nmsg = m->mm->mm.in.count;
break;
}
//printf("%i pairing %i -> %i\n", QMP_get_node_number(), m->srce_node, m->dest_node);
if(m->type==MH_send) {
MPI_Status st;
MPI_Send(&nmsg, sizeof(nmsg), MPI_BYTE, m->dest_node, 0, m->comm->mpicomm);
MPI_Recv(&nmsgo, sizeof(nmsgo), MPI_BYTE, m->dest_node, 0, m->comm->mpicomm, &st);
} else if(m->type==MH_recv) {
MPI_Status st;
MPI_Recv(&nmsgo, sizeof(nmsgo), MPI_BYTE, m->srce_node, 0, m->comm->mpicomm, &st);
MPI_Send(&nmsg, sizeof(nmsg), MPI_BYTE, m->srce_node, 0, m->comm->mpicomm);
} else {
QMP_FATAL("error: bad msg type");
}
m->paired = 0;
if(nmsg>0 && nmsgo>0) {
void *base[nmsg];
size_t size[nmsg];
switch(m->mm->type) {
case MM_user_buf: base[0]=m->base; size[0]=m->mm->nbytes; break;
case MM_strided_buf:
{
int offset = 0;
for(int i=0; i<nmsg; i++) {
base[i] = (char*)m->base + offset;
size[i] = m->mm->mm.st.blksize;
offset += m->mm->mm.st.stride;
}
}
break;
case MM_strided_array_buf:
{
int k = 0;
for(int i=0; i<m->mm->mm.sa.narray; i++) {
int offset = 0;
char *basei = (char*)m->base + m->mm->mm.sa.disp[i];
for(int j=0; j<m->mm->mm.sa.nblocks[i]; j++) {
base[k] = basei + offset;
size[k] = m->mm->mm.sa.blksize[i];
k++;
offset += m->mm->mm.sa.stride[i];
}
}
}
break;
case MM_indexed_buf:
{
double avgbl = 0;
for(int i=0; i<nmsg; i++) {
base[i] = (char*)m->base + m->mm->mm.in.elemsize*m->mm->mm.in.index[i];
size[i] = m->mm->mm.in.elemsize*m->mm->mm.in.blocklen[i];
avgbl += m->mm->mm.in.blocklen[i];
}
avgbl = (m->mm->mm.in.elemsize*avgbl)/nmsg;
//if(QMP_get_node_number()==0) printf("indexed: n: %6i e: %6i a: %9.2f\n", nmsg, m->mm->mm.in.elemsize, avgbl);
}
break;
}
if(m->type==MH_send) {
int grank;
MPIX_Comm_rank2global(m->comm->mpicomm, m->dest_node, &grank);
m->qspimsg = qspi_create_msg();
qspi_set_send_multi(grank, base, size, nmsg, m->qspimsg);
qspi_prepare(&m->qspimsg, 1);
//printf("%i testing msg send to %i\n", m->srce_node, m->dest_node);
//qspi_start(m->qspimsg);
//qspi_wait(m->qspimsg);
//printf("%i testing msg send to %i done\n", m->srce_node, m->dest_node);
m->qspicts = qspi_create_msg();
qspi_set_recv(grank, &m->clear_to_send, sizeof(m->clear_to_send), m->qspicts);
qspi_prepare(&m->qspicts, 1);
//printf("%i testing cts send to %i\n", m->srce_node, m->dest_node);
//qspi_start(m->qspicts);
//qspi_wait(m->qspicts);
//printf("%i testing cts send to %i done\n", m->srce_node, m->dest_node);
} else { // recv
int grank;
MPIX_Comm_rank2global(m->comm->mpicomm, m->srce_node, &grank);
m->qspimsg = qspi_create_msg();
//qspi_set_recv_multi(grank, base, size, nmsg, m->qspimsg);
qspi_set_recv(grank, base[0], size[0], m->qspimsg);
qspi_prepare(&m->qspimsg, 1);
//printf("%i testing msg recv from %i\n", m->dest_node, m->srce_node);
//qspi_start(m->qspimsg);
//qspi_wait(m->qspimsg);
//printf("%i testing msg recv from %i done\n", m->dest_node, m->srce_node);
m->qspicts = qspi_create_msg();
qspi_set_send(grank, &m->clear_to_send, sizeof(m->clear_to_send), m->qspicts);
qspi_prepare(&m->qspicts, 1);
//printf("%i testing cts recv from %i\n", m->dest_node, m->srce_node);
//qspi_start(m->qspicts);
//qspi_wait(m->qspicts);
//printf("%i testing cts recv from %i done\n", m->dest_node, m->srce_node);
}
npaired++;
m->paired = 1;
m->useSPI = 1;
//printf("%i paired %i -> %i\n", QMP_get_node_number(), m->srce_node, m->dest_node);
} else {
//printf("%i unpaired %i -> %i type: %i\n", QMP_get_node_number(), m->srce_node, m->dest_node, m->mm->type);
}
}
}
src0 = src;
}
mh->useSPI = npaired;
}
//QMP_declare_multiple_mpi(mh);
}
