static inline int NBC_Start_round(NBC_Handle *handle) {
int num; /* number of operations */
int res;
char* ptr;
MPI_Request *tmp;
NBC_Fn_type type;
NBC_Args_send sendargs;
NBC_Args_recv recvargs;
NBC_Args_op opargs;
NBC_Args_copy copyargs;
NBC_Args_unpack unpackargs;
void *buf1, *buf2;
/* get round-schedule address */
ptr = handle->schedule->data + handle->row_offset;
NBC_GET_BYTES(ptr,num);
NBC_DEBUG(10, "start_round round at offset %d : posting %i operations\n", handle->row_offset, num);
for (int i = 0 ; i < num ; ++i) {
int offset = (intptr_t)(ptr - handle->schedule->data);
memcpy (&type, ptr, sizeof (type));
switch(type) {
case SEND:
NBC_DEBUG(5," SEND (offset %li) ", offset);
NBC_GET_BYTES(ptr,sendargs);
NBC_DEBUG(5,"*buf: %p, count: %i, type: %p, dest: %i, tag: %i)\n", sendargs.buf,
sendargs.count, sendargs.datatype, sendargs.dest, handle->tag);
/* get an additional request */
handle->req_count++;
/* get buffer */
if(sendargs.tmpbuf) {
buf1=(char*)handle->tmpbuf+(long)sendargs.buf;
} else {
buf1=(void *)sendargs.buf;
}
#ifdef NBC_TIMING
Isend_time -= MPI_Wtime();
#endif
tmp = (MPI_Request *) realloc ((void *) handle->req_array, handle->req_count * sizeof (MPI_Request));
if (NULL == tmp) {
return OMPI_ERR_OUT_OF_RESOURCE;
}
handle->req_array = tmp;
res = MCA_PML_CALL(isend(buf1, sendargs.count, sendargs.datatype, sendargs.dest, handle->tag,
MCA_PML_BASE_SEND_STANDARD, sendargs.local?handle->comm->c_local_comm:handle->comm,
handle->req_array+handle->req_count - 1));
if (OMPI_SUCCESS != res) {
NBC_Error ("Error in MPI_Isend(%lu, %i, %p, %i, %i, %lu) (%i)", (unsigned long)buf1, sendargs.count,
sendargs.datatype, sendargs.dest, handle->tag, (unsigned long)handle->comm, res);
return res;
}
#ifdef NBC_TIMING
Isend_time += MPI_Wtime();
#endif
break;
case RECV:
NBC_DEBUG(5, " RECV (offset %li) ", offset);
NBC_GET_BYTES(ptr,recvargs);
NBC_DEBUG(5, "*buf: %p, count: %i, type: %p, source: %i, tag: %i)\n", recvargs.buf, recvargs.count,
recvargs.datatype, recvargs.source, handle->tag);
/* get an additional request - TODO: req_count NOT thread safe */
handle->req_count++;
/* get buffer */
if(recvargs.tmpbuf) {
buf1=(char*)handle->tmpbuf+(long)recvargs.buf;
} else {
buf1=recvargs.buf;
}
#ifdef NBC_TIMING
Irecv_time -= MPI_Wtime();
#endif
tmp = (MPI_Request *) realloc ((void *) handle->req_array, handle->req_count * sizeof (MPI_Request));
if (NULL == tmp) {
return OMPI_ERR_OUT_OF_RESOURCE;
}
handle->req_array = tmp;
res = MCA_PML_CALL(irecv(buf1, recvargs.count, recvargs.datatype, recvargs.source, handle->tag, recvargs.local?handle->comm->c_local_comm:handle->comm,
handle->req_array+handle->req_count-1));
if (OMPI_SUCCESS != res) {
NBC_Error("Error in MPI_Irecv(%lu, %i, %p, %i, %i, %lu) (%i)", (unsigned long)buf1, recvargs.count,
recvargs.datatype, recvargs.source, handle->tag, (unsigned long)handle->comm, res);
return res;
}
#ifdef NBC_TIMING
Irecv_time += MPI_Wtime();
#endif
break;
case OP:
NBC_DEBUG(5, " OP2 (offset %li) ", offset);
NBC_GET_BYTES(ptr,opargs);
NBC_DEBUG(5, "*buf1: %p, buf2: %p, count: %i, type: %p)\n", opargs.buf1, opargs.buf2,
opargs.count, opargs.datatype);
/* get buffers */
if(opargs.tmpbuf1) {
buf1=(char*)handle->tmpbuf+(long)opargs.buf1;
} else {
buf1=(void *)opargs.buf1;
}
if(opargs.tmpbuf2) {
buf2=(char*)handle->tmpbuf+(long)opargs.buf2;
} else {
buf2=opargs.buf2;
}
ompi_op_reduce(opargs.op, buf1, buf2, opargs.count, opargs.datatype);
break;
case COPY:
NBC_DEBUG(5, " COPY (offset %li) ", offset);
NBC_GET_BYTES(ptr,copyargs);
NBC_DEBUG(5, "*src: %lu, srccount: %i, srctype: %p, *tgt: %lu, tgtcount: %i, tgttype: %p)\n",
(unsigned long) copyargs.src, copyargs.srccount, copyargs.srctype,
(unsigned long) copyargs.tgt, copyargs.tgtcount, copyargs.tgttype);
/* get buffers */
if(copyargs.tmpsrc) {
buf1=(char*)handle->tmpbuf+(long)copyargs.src;
} else {
buf1=copyargs.src;
}
if(copyargs.tmptgt) {
buf2=(char*)handle->tmpbuf+(long)copyargs.tgt;
} else {
buf2=copyargs.tgt;
}
res = NBC_Copy (buf1, copyargs.srccount, copyargs.srctype, buf2, copyargs.tgtcount, copyargs.tgttype,
handle->comm);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
return res;
}
break;
case UNPACK:
NBC_DEBUG(5, " UNPACK (offset %li) ", offset);
NBC_GET_BYTES(ptr,unpackargs);
NBC_DEBUG(5, "*src: %lu, srccount: %i, srctype: %p, *tgt: %lu\n", (unsigned long) unpackargs.inbuf,
unpackargs.count, unpackargs.datatype, (unsigned long) unpackargs.outbuf);
/* get buffers */
if(unpackargs.tmpinbuf) {
buf1=(char*)handle->tmpbuf+(long)unpackargs.inbuf;
} else {
buf1=unpackargs.inbuf;
}
if(unpackargs.tmpoutbuf) {
buf2=(char*)handle->tmpbuf+(long)unpackargs.outbuf;
} else {
buf2=unpackargs.outbuf;
}
res = NBC_Unpack (buf1, unpackargs.count, unpackargs.datatype, buf2, handle->comm);
if (OMPI_SUCCESS != res) {
NBC_Error ("NBC_Unpack() failed (code: %i)", res);
return res;
}
break;
default:
NBC_Error ("NBC_Start_round: bad type %li at offset %li", (long)type, offset);
return OMPI_ERROR;
}
}
/* check if we can make progress - not in the first round, this allows us to leave the
* initialization faster and to reach more overlap
*
* threaded case: calling progress in the first round can lead to a
* deadlock if NBC_Free is called in this round :-( */
if (handle->row_offset) {
res = NBC_Progress(handle);
if ((NBC_OK != res) && (NBC_CONTINUE != res)) {
return OMPI_ERROR;
}
}
return OMPI_SUCCESS;
}
static inline int NBC_Start_round(NBC_Handle *handle) {
int num; /* number of operations */
int i, res, ret=NBC_OK;
char* ptr;
NBC_Fn_type type;
NBC_Args_send sendargs;
NBC_Args_recv recvargs;
NBC_Args_op opargs;
NBC_Args_copy copyargs;
NBC_Args_unpack unpackargs;
NBC_Schedule myschedule;
void *buf1, *buf2, *buf3;
/* get round-schedule address */
myschedule = (NBC_Schedule*)((char*)*handle->schedule + handle->row_offset);
ptr = (char*) myschedule;
NBC_GET_BYTES(ptr,num);
NBC_DEBUG(10, "start_round round at address %p : posting %i operations\n", myschedule, num);
for (i=0; i<num; i++) {
NBC_GET_BYTES(ptr,type);
switch(type) {
case SEND:
NBC_DEBUG(5," SEND (offset %li) ", (long)ptr-(long)myschedule);
NBC_GET_BYTES(ptr,sendargs);
NBC_DEBUG(5,"*buf: %p, count: %i, type: %lu, dest: %i, tag: %i)\n", sendargs.buf, sendargs.count, (unsigned long)sendargs.datatype, sendargs.dest, handle->tag);
/* get an additional request */
handle->req_count++;
/* get buffer */
if(sendargs.tmpbuf) {
buf1=(char*)handle->tmpbuf+(long)sendargs.buf;
} else {
buf1=sendargs.buf;
}
#ifdef NBC_TIMING
Isend_time -= MPI_Wtime();
#endif
handle->req_array = (MPI_Request*)realloc((void*)handle->req_array, (handle->req_count)*sizeof(MPI_Request));
NBC_CHECK_NULL(handle->req_array);
res = MCA_PML_CALL(isend(buf1, sendargs.count, sendargs.datatype, sendargs.dest, handle->tag, MCA_PML_BASE_SEND_STANDARD, handle->comm, handle->req_array+handle->req_count-1));
if(OMPI_SUCCESS != res) { printf("Error in MPI_Isend(%lu, %i, %lu, %i, %i, %lu) (%i)\n", (unsigned long)buf1, sendargs.count, (unsigned long)sendargs.datatype, sendargs.dest, handle->tag, (unsigned long)handle->comm, res); ret=res; goto error; }
#ifdef NBC_TIMING
Isend_time += MPI_Wtime();
#endif
break;
case RECV:
NBC_DEBUG(5, " RECV (offset %li) ", (long)ptr-(long)myschedule);
NBC_GET_BYTES(ptr,recvargs);
NBC_DEBUG(5, "*buf: %p, count: %i, type: %lu, source: %i, tag: %i)\n", recvargs.buf, recvargs.count, (unsigned long)recvargs.datatype, recvargs.source, handle->tag);
/* get an additional request - TODO: req_count NOT thread safe */
handle->req_count++;
/* get buffer */
if(recvargs.tmpbuf) {
buf1=(char*)handle->tmpbuf+(long)recvargs.buf;
} else {
buf1=recvargs.buf;
}
#ifdef NBC_TIMING
Irecv_time -= MPI_Wtime();
#endif
handle->req_array = (MPI_Request*)realloc((void*)handle->req_array, (handle->req_count)*sizeof(MPI_Request));
NBC_CHECK_NULL(handle->req_array);
res = MCA_PML_CALL(irecv(buf1, recvargs.count, recvargs.datatype, recvargs.source, handle->tag, handle->comm, handle->req_array+handle->req_count-1));
if(OMPI_SUCCESS != res) { printf("Error in MPI_Irecv(%lu, %i, %lu, %i, %i, %lu) (%i)\n", (unsigned long)buf1, recvargs.count, (unsigned long)recvargs.datatype, recvargs.source, handle->tag, (unsigned long)handle->comm, res); ret=res; goto error; }
#ifdef NBC_TIMING
Irecv_time += MPI_Wtime();
#endif
break;
case OP:
NBC_DEBUG(5, " OP (offset %li) ", (long)ptr-(long)myschedule);
NBC_GET_BYTES(ptr,opargs);
NBC_DEBUG(5, "*buf1: %p, buf2: %p, buf3: %p, count: %i, type: %lu)\n", opargs.buf1, opargs.buf2, opargs.buf3, opargs.count, (unsigned long)opargs.datatype);
/* get buffers */
if(opargs.tmpbuf1) {
buf1=(char*)handle->tmpbuf+(long)opargs.buf1;
} else {
buf1=opargs.buf1;
}
if(opargs.tmpbuf2) {
buf2=(char*)handle->tmpbuf+(long)opargs.buf2;
} else {
buf2=opargs.buf2;
}
if(opargs.tmpbuf3) {
buf3=(char*)handle->tmpbuf+(long)opargs.buf3;
} else {
buf3=opargs.buf3;
}
ompi_3buff_op_reduce(opargs.op, buf1, buf2, buf3, opargs.count, opargs.datatype);
break;
case COPY:
NBC_DEBUG(5, " COPY (offset %li) ", (long)ptr-(long)myschedule);
NBC_GET_BYTES(ptr,copyargs);
NBC_DEBUG(5, "*src: %lu, srccount: %i, srctype: %lu, *tgt: %lu, tgtcount: %i, tgttype: %lu)\n", (unsigned long)copyargs.src, copyargs.srccount, (unsigned long)copyargs.srctype, (unsigned long)copyargs.tgt, copyargs.tgtcount, (unsigned long)copyargs.tgttype);
/* get buffers */
if(copyargs.tmpsrc) {
buf1=(char*)handle->tmpbuf+(long)copyargs.src;
} else {
buf1=copyargs.src;
}
if(copyargs.tmptgt) {
buf2=(char*)handle->tmpbuf+(long)copyargs.tgt;
} else {
buf2=copyargs.tgt;
}
res = NBC_Copy(buf1, copyargs.srccount, copyargs.srctype, buf2, copyargs.tgtcount, copyargs.tgttype, handle->comm);
if(res != NBC_OK) { printf("NBC_Copy() failed (code: %i)\n", res); ret=res; goto error; }
break;
case UNPACK:
NBC_DEBUG(5, " UNPACK (offset %li) ", (long)ptr-(long)myschedule);
NBC_GET_BYTES(ptr,unpackargs);
NBC_DEBUG(5, "*src: %lu, srccount: %i, srctype: %lu, *tgt: %lu\n", (unsigned long)unpackargs.inbuf, unpackargs.count, (unsigned long)unpackargs.datatype, (unsigned long)unpackargs.outbuf);
/* get buffers */
if(unpackargs.tmpinbuf) {
buf1=(char*)handle->tmpbuf+(long)unpackargs.inbuf;
} else {
buf1=unpackargs.outbuf;
}
if(unpackargs.tmpoutbuf) {
buf2=(char*)handle->tmpbuf+(long)unpackargs.outbuf;
} else {
buf2=unpackargs.outbuf;
}
res = NBC_Unpack(buf1, unpackargs.count, unpackargs.datatype, buf2, handle->comm);
if(res != NBC_OK) { printf("NBC_Unpack() failed (code: %i)\n", res); ret=res; goto error; }
break;
default:
printf("NBC_Start_round: bad type %li at offset %li\n", (long)type, (long)ptr-(long)myschedule);
ret=NBC_BAD_SCHED;
goto error;
}
}
/* check if we can make progress - not in the first round, this allows us to leave the
* initialization faster and to reach more overlap
*
* threaded case: calling progress in the first round can lead to a
* deadlock if NBC_Free is called in this round :-( */
if(handle->row_offset != sizeof(int)) {
res = NBC_Progress(handle);
if((NBC_OK != res) && (NBC_CONTINUE != res)) { printf("Error in NBC_Progress() (%i)\n", res); ret=res; goto error; }
}
error:
return ret;
}
