#include <stdio.h>#include <Padico/MPCircuit.h>#include "Schedule.h"#include "Internal.h"#include "DistributionBloc.h"
Include dependency graph for PC/Schedule.cc:
Go to the source code of this file.
Defines | |
| #define | DEBUG_COMM |
| #define | DEBUG_INTERNAL |
Functions | |
| void | computeReceiveBlock1D (const GlobalData_t &gd, const LocalData_t &dd, const Topology_t &stopo, const Topology_t &dtopo, vector< LocalData_t > &vOut) |
| void | computeSendBlock1D (const GlobalData_t &gd, const LocalData_t &sd, const Topology_t &stopo, const Topology_t &dtopo, vector< LocalData_t > &vOut) |
| void | doSchedule (const GlobalData_t &gd, const LocalData_t &ld, const Topology_t &ctopo, vector< LocalData_t > &sched_send, vector< LocalData_t > &sched_recv, void *comm) |
Define Documentation
| #define DEBUG_COMM |
Definition at line 17 of file PC/Schedule.cc.
| #define DEBUG_INTERNAL |
Definition at line 16 of file PC/Schedule.cc.
Function Documentation
| void computeReceiveBlock1D | ( | const GlobalData_t & | gd, |
| const LocalData_t & | dd, | ||
| const Topology_t & | stopo, | ||
| const Topology_t & | dtopo, | ||
| vector< LocalData_t > & | vOut | ||
| ) |
Definition at line 105 of file PC/Schedule.cc.
References blockSize(), getProcRangeInf(), and getProcRangeSup().
{
#ifdef DEBUG_INTERNAL
cerr << "\nIn compute Receive Schedule--------------------\n";
fprintf(stderr, "stopo: %ld\tdtopo: %ld\n",stopo.total, dtopo.total);
fprintf(stderr, "gd.len %ld\tdd.start %d\tdd.len %d\n", gd.len, dd.start, dd.len);
#endif
if (stopo.total == dtopo.total) {
vOut.push_back(dd);
#ifdef DEBUG_INTERNAL
fprintf(stderr, " rank:%d start:%d len:%d base:%p\n", dd.rank, dd.start, dd.len, dd.base);
#endif
} else {
// Apend mode
vOut.reserve(vOut.size()+stopo.total); // in bloc mode, at most one msg from each src node
unsigned sbsz = blockSize(gd.len, stopo.total);
unsigned long dlow = dd.start;
unsigned long dhigh = dlow + dd.len;
unsigned fpid, lpid;
fpid = getProcRangeInf(dlow, sbsz);
lpid = getProcRangeSup(dhigh, sbsz);
#ifdef DEBUG_INTERNAL
fprintf(stderr, " loop from %d to %d width stotal: %ld\n", fpid, lpid, stopo.total);
#endif
// for each dest bloc
for(unsigned i=fpid; i <= lpid; i++) {
vOut.resize(vOut.size()+1);
LocalData_t& s = vOut[vOut.size()-1];
s.rank = i;
unsigned tmp = i*sbsz;
s.start = ( dlow >= tmp)?dlow:tmp; // max
tmp = (i+1)*sbsz;
unsigned end = ( dhigh <= tmp)?dhigh:tmp; // min
s.len = end - s.start;
s.base = dd.base + ((s.start - dd.start) * gd.unit_size);
#ifdef DEBUG_INTERNAL
fprintf(stderr, " r: from:%d start:%d len:%d base:%p\n", s.rank, s.start, s.len, s.base);
#endif
}
}
}
Here is the call graph for this function:
| void computeSendBlock1D | ( | const GlobalData_t & | gd, |
| const LocalData_t & | sd, | ||
| const Topology_t & | stopo, | ||
| const Topology_t & | dtopo, | ||
| vector< LocalData_t > & | vOut | ||
| ) |
Definition at line 39 of file PC/Schedule.cc.
References blockSize(), getProcRangeInf(), and getProcRangeSup().
{
#ifdef DEBUG_INTERNAL
cerr << "\nIn compute Send Schedule--------------------\n";
fprintf(stderr, "stopo: %ld\tdtopo: %ld\n",stopo.total, dtopo.total);
fprintf(stderr, "gd.len %ld\tsd.start %d\tsd.len %d\n",gd.len, sd.start, sd.len);
#endif
if (stopo.total == dtopo.total) {
// vOut.reserve(vOut.size()+dtopo.total); // in bloc mode, at most one msg to each dest node
vOut.push_back(sd);
#ifdef DEBUG_INTERNAL
fprintf(stderr, " rank:%d start:%d len:%d base:%p\n", sd.rank, sd.start, sd.len, sd.base);
#endif
} else {
// Append mode
vOut.reserve(vOut.size()+dtopo.total); // in bloc mode, at most one msg to each dest node
unsigned dbsz = blockSize(gd.len, dtopo.total);
unsigned long slow = sd.start;
unsigned long shigh = slow + sd.len;
unsigned fpid, lpid;
fpid = getProcRangeInf(slow, dbsz);
lpid = getProcRangeSup(shigh, dbsz);
#ifdef DEBUG_INTERNAL
fprintf(stderr, " loop from %d to %d width dtotal: %ld\n", fpid, lpid, dtopo.total);
#endif
// for each dest bloc
for(unsigned i=fpid; i <= lpid; i++) {
vOut.resize(vOut.size()+1);
LocalData_t& s = vOut[vOut.size()-1];
s.rank = i;
unsigned tmp = i*dbsz;
s.start = ( slow >= tmp)?slow:tmp; // max
tmp = (i+1)*dbsz;
unsigned end = ( shigh <= tmp)?shigh:tmp; // min
s.len = end - s.start;
s.base = sd.base + ((s.start - sd.start) * gd.unit_size);
#ifdef DEBUG_INTERNAL
fprintf(stderr, " s: to:%d start:%d len:%d base:%p\n", s.rank, s.start, s.len, s.base);
#endif
}
}
#ifdef DEBUG_INTERNAL
cerr << "\nIn compute Send Schedule-------------------- done\n";
#endif
}
Here is the call graph for this function:
| void doSchedule | ( | const GlobalData_t & | gd, |
| const LocalData_t & | ld, | ||
| const Topology_t & | ctopo, | ||
| vector< LocalData_t > & | sched_send, | ||
| vector< LocalData_t > & | sched_recv, | ||
| void * | comm | ||
| ) |
Definition at line 170 of file PC/Schedule.cc.
References cmp_rank(), and getProcId().
{
cerr << "\nIn doSchedule--------------------\n";
padico_mpcircuit_t schd_mpc = (padico_mpcircuit_t) comm;
if (sched_send.size() || sched_recv.size()) {
#ifndef NO_COM
void* rreq[sched_recv.size()];
unsigned ri;
ri=0;
#endif
vector<LocalData_t*> local_recv;
vector<LocalData_t*> local_send;
local_recv.clear();
local_send.clear();
// Sending data
// Post Asynchronous MPCircuit receive
#ifdef DEBUG_COM
cerr << " #sched_recv: " << sched_recv.size() << endl;
#endif
for(unsigned i=0; i < sched_recv.size(); i++) {
unsigned from = getProcId(sched_recv[i].rank, ctopo);
if (from == ld.rank) {
#ifdef DEBUG_COMM
fprintf(stderr, " recv: schedr no=%d start=%d len=%d from=%d LOCAL\n", i,
sched_recv[i].start, sched_recv[i].len, from);
#endif
local_recv.push_back(&sched_recv[i]);
} else {
#ifdef DEBUG_COMM
fprintf(stderr, " recv: schedr no=%d start=%d len=%d from=%d base=%p\n", i,
sched_recv[i].start, sched_recv[i].len, from, sched_recv[i].base);
#endif
#ifndef NO_COM
rreq[ri++] = padico_mpcircuit_Irecv(sched_recv[i].base, sched_recv[i].len*gd.unit_size,
from, 51, schd_mpc );
#endif
}
}
// Send data via MPCircuit
#ifdef DEBUG_COMM
cerr << " #sched_send: " << sched_send.size() << endl;
#endif
for(unsigned i=0; i < sched_send.size(); i++) {
unsigned to = getProcId(sched_send[i].rank, ctopo);
if (to == ld.rank) {
#ifdef DEBUG_COMM
fprintf(stderr, " send: scheds no=%d start=%d len=%d to=%d LOCAL\n", i,
sched_send[i].start, sched_send[i].len, to);
#endif
local_send.push_back(&sched_send[i]);
} else {
#ifdef DEBUG_COMM
fprintf(stderr, " send: scheds no=%d start=%d len=%d to=%d base=%p\n", i,
sched_send[i].start, sched_send[i].len, to, sched_send[i].base);
#endif
#ifndef NO_COM
padico_mpcircuit_send(sched_send[i].base, sched_send[i].len*gd.unit_size,
to, 51, schd_mpc);
#endif
}
}
// Do local communication vie memcpy
if (local_recv.size() != local_send.size()) {
cerr << "Error: local recv & send have different size: " << local_recv.size() << " " << local_send.size() << endl;
}
for(unsigned i=0; i < local_recv.size(); i++) {
if (local_recv[i]->len != local_send[i]->len) {
cerr << "Error: local recv & send have different len for i= "<<i<< " :" << local_recv[i]->len << " " << local_send[i]->len << endl;
}
#ifdef DEBUG_COMM
fprintf(stderr, " local: scheds no=%d start=%d len=%d\n", i,
sched_send[i].start, sched_send[i].len);
#endif
#ifndef NO_COM
memcpy(local_recv[i]->base, local_send[i]->base, local_send[i]->len*gd.unit_size);
#endif
}
// Wait all receive & send
#ifndef NO_COM
#ifdef DEBUG_INTERNAL
cerr << "WAITING local communications to end...\n";
#endif
padico_mpcircuit_waitAll(rreq, ri);
#ifdef DEBUG_INTERNAL
cerr << "WAITING local communications to end...ok \n";
#endif
#endif
}
}
Here is the call graph for this function:
