X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/af94b4458a6da51502a01c06756e4c11cfbe2648..db17323a11a3089f492c2cae1be7913673689d5c:/src/smpi/smpi_base.c diff --git a/src/smpi/smpi_base.c b/src/smpi/smpi_base.c index ee947cc2ef..88e93074d0 100644 --- a/src/smpi/smpi_base.c +++ b/src/smpi/smpi_base.c @@ -327,6 +327,7 @@ void smpi_mpi_start(MPI_Request request) // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later request->real_size=request->size; smpi_datatype_use(request->old_type); + smpi_comm_use(request->comm); request->action = simcall_comm_irecv(mailbox, request->buf, &request->real_size, &match_recv, request); //integrate pseudo-timing for buffering of small messages, do not bother to execute the simcall if 0 @@ -375,6 +376,7 @@ void smpi_mpi_start(MPI_Request request) // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later request->real_size=request->size; smpi_datatype_use(request->old_type); + smpi_comm_use(request->comm); //if we are giving back the control to the user without waiting for completion, we have to inject timings double sleeptime =0.0; @@ -419,7 +421,6 @@ void smpi_mpi_startall(int count, MPI_Request * requests) void smpi_mpi_request_free(MPI_Request * request) { - if((*request) != MPI_REQUEST_NULL){ (*request)->refcount--; if((*request)->refcount<0) xbt_die("wrong refcount"); @@ -582,6 +583,7 @@ static void finish_wait(MPI_Request * request, MPI_Status * status) } if(req->detached == 0) free(req->buf); } + smpi_comm_unuse(req->comm); smpi_datatype_unuse(datatype); } @@ -633,33 +635,31 @@ int smpi_mpi_testany(int count, MPI_Request requests[], int *index, *index = MPI_UNDEFINED; flag = 0; - if(count > 0) { - comms = xbt_dynar_new(sizeof(smx_action_t), NULL); - map = xbt_new(int, count); - size = 0; - for(i = 0; i < count; i++) { - if((requests[i]!=MPI_REQUEST_NULL) && requests[i]->action) { - xbt_dynar_push(comms, &requests[i]->action); - map[size] = i; - size++; - } + comms = xbt_dynar_new(sizeof(smx_action_t), NULL); + map = xbt_new(int, count); + size = 0; + for(i = 0; i < count; i++) { + if((requests[i]!=MPI_REQUEST_NULL) && requests[i]->action) { + xbt_dynar_push(comms, &requests[i]->action); + map[size] = i; + size++; } - if(size > 0) { - i = simcall_comm_testany(comms); - // not MPI_UNDEFINED, as this is a simix return code - if(i != -1) { - *index = map[i]; - finish_wait(&requests[*index], status); - flag = 1; - } - }else{ - //all requests are null or inactive, return true - flag=1; - smpi_empty_status(status); + } + if(size > 0) { + i = simcall_comm_testany(comms); + // not MPI_UNDEFINED, as this is a simix return code + if(i != -1) { + *index = map[i]; + finish_wait(&requests[*index], status); + flag = 1; } - xbt_free(map); - xbt_dynar_free(&comms); + }else{ + //all requests are null or inactive, return true + flag=1; + smpi_empty_status(status); } + xbt_free(map); + xbt_dynar_free(&comms); return flag; } @@ -1109,8 +1109,10 @@ void smpi_mpi_scatter(void *sendbuf, int sendcount, MPI_Datatype sendtype, // FIXME: check for errors smpi_datatype_extent(sendtype, &lb, &sendext); // Local copy from root - smpi_datatype_copy((char *)sendbuf + root * sendcount * sendext, - sendcount, sendtype, recvbuf, recvcount, recvtype); + if(recvbuf!=MPI_IN_PLACE){ + smpi_datatype_copy((char *)sendbuf + root * sendcount * sendext, + sendcount, sendtype, recvbuf, recvcount, recvtype); + } // Send buffers to receivers requests = xbt_new(MPI_Request, size - 1); index = 0; @@ -1148,8 +1150,10 @@ void smpi_mpi_scatterv(void *sendbuf, int *sendcounts, int *displs, // FIXME: check for errors smpi_datatype_extent(sendtype, &lb, &sendext); // Local copy from root - smpi_datatype_copy((char *)sendbuf + displs[root] * sendext, sendcounts[root], + if(recvbuf!=MPI_IN_PLACE){ + smpi_datatype_copy((char *)sendbuf + displs[root] * sendext, sendcounts[root], sendtype, recvbuf, recvcount, recvtype); + } // Send buffers to receivers requests = xbt_new(MPI_Request, size - 1); index = 0; @@ -1178,9 +1182,11 @@ void smpi_mpi_reduce(void *sendbuf, void *recvbuf, int count, MPI_Request *requests; void **tmpbufs; + char* sendtmpbuf = (char*) sendbuf; if( sendbuf == MPI_IN_PLACE ) { - sendtmpbuf = (char *)recvbuf; + sendtmpbuf = (char *)xbt_malloc(count*smpi_datatype_get_extent(datatype)); + smpi_datatype_copy(recvbuf, count, datatype,sendtmpbuf, count, datatype); } rank = smpi_comm_rank(comm); @@ -1233,6 +1239,10 @@ void smpi_mpi_reduce(void *sendbuf, void *recvbuf, int count, } xbt_free(tmpbufs); xbt_free(requests); + + if( sendbuf == MPI_IN_PLACE ) { + xbt_free(sendtmpbuf); + } } } @@ -1246,7 +1256,7 @@ void smpi_mpi_allreduce(void *sendbuf, void *recvbuf, int count, void smpi_mpi_scan(void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm) { - int system_tag = 888; + int system_tag = -888; int rank, size, other, index; MPI_Aint lb = 0, dataext = 0; MPI_Request *requests; @@ -1281,14 +1291,94 @@ void smpi_mpi_scan(void *sendbuf, void *recvbuf, int count, } // Wait for completion of all comms. smpi_mpi_startall(size - 1, requests); - for(other = 0; other < size - 1; other++) { - index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE); - if(index == MPI_UNDEFINED) { - break; + + if(smpi_op_is_commute(op)){ + for(other = 0; other < size - 1; other++) { + index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE); + if(index == MPI_UNDEFINED) { + break; + } + if(index < rank) { + // #Request is below rank: it's a irecv + smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype); + } + } + }else{ + //non commutative case, wait in order + for(other = 0; other < size - 1; other++) { + smpi_mpi_wait(&(requests[other]), MPI_STATUS_IGNORE); + if(index < rank) { + smpi_op_apply(op, tmpbufs[other], recvbuf, &count, &datatype); + } + } + } + for(index = 0; index < rank; index++) { + xbt_free(tmpbufs[index]); + } + xbt_free(tmpbufs); + xbt_free(requests); +} + +void smpi_mpi_exscan(void *sendbuf, void *recvbuf, int count, + MPI_Datatype datatype, MPI_Op op, MPI_Comm comm) +{ + int system_tag = -888; + int rank, size, other, index; + MPI_Aint lb = 0, dataext = 0; + MPI_Request *requests; + void **tmpbufs; + int recvbuf_is_empty=1; + rank = smpi_comm_rank(comm); + size = smpi_comm_size(comm); + + // FIXME: check for errors + smpi_datatype_extent(datatype, &lb, &dataext); + + // Send/Recv buffers to/from others; + requests = xbt_new(MPI_Request, size - 1); + tmpbufs = xbt_new(void *, rank); + index = 0; + for(other = 0; other < rank; other++) { + // FIXME: possibly overkill we we have contiguous/noncontiguous data + // mapping... + tmpbufs[index] = xbt_malloc(count * dataext); + requests[index] = + smpi_irecv_init(tmpbufs[index], count, datatype, other, system_tag, + comm); + index++; + } + for(other = rank + 1; other < size; other++) { + requests[index] = + smpi_isend_init(sendbuf, count, datatype, other, system_tag, comm); + index++; + } + // Wait for completion of all comms. + smpi_mpi_startall(size - 1, requests); + if(smpi_op_is_commute(op)){ + for(other = 0; other < size - 1; other++) { + index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE); + if(index == MPI_UNDEFINED) { + break; + } + if(index < rank) { + if(recvbuf_is_empty){ + smpi_datatype_copy(tmpbufs[index], count, datatype, recvbuf, count, datatype); + recvbuf_is_empty=0; + }else + // #Request is below rank: it's a irecv + smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype); + } } - if(index < rank) { - // #Request is below rank: it's a irecv - smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype); + }else{ + //non commutative case, wait in order + for(other = 0; other < size - 1; other++) { + smpi_mpi_wait(&(requests[other]), MPI_STATUS_IGNORE); + if(index < rank) { + if(recvbuf_is_empty){ + smpi_datatype_copy(tmpbufs[other], count, datatype, recvbuf, count, datatype); + recvbuf_is_empty=0; + }else smpi_op_apply(op, tmpbufs[other], recvbuf, &count, &datatype); + } } } for(index = 0; index < rank; index++) {