1c6db04a5SJed Brown #include <petscsys.h> /*I "petscsys.h" I*/ 276eed172SJunchao Zhang #include <petsc/private/mpiutils.h> 3e5c89e4eSSatish Balay 4e5c89e4eSSatish Balay /*@C 5811af0c4SBarry Smith PetscGatherNumberOfMessages - Computes the number of messages an MPI rank expects to receive during a neighbor communication 6e5c89e4eSSatish Balay 7d083f849SBarry Smith Collective 8e5c89e4eSSatish Balay 9e5c89e4eSSatish Balay Input Parameters: 10e5c89e4eSSatish Balay + comm - Communicator 11667f096bSBarry Smith . iflags - an array of integers of length sizeof(comm). A '1' in `ilengths`[i] represent a 12667f096bSBarry Smith message from current node to ith node. Optionally `NULL` 13667f096bSBarry Smith - ilengths - Non zero ilengths[i] represent a message to i of length `ilengths`[i]. 14667f096bSBarry Smith Optionally `NULL`. 15e5c89e4eSSatish Balay 162fe279fdSBarry Smith Output Parameter: 17e5c89e4eSSatish Balay . nrecvs - number of messages received 18e5c89e4eSSatish Balay 19e5c89e4eSSatish Balay Level: developer 20e5c89e4eSSatish Balay 21e5c89e4eSSatish Balay Notes: 22e5c89e4eSSatish Balay With this info, the correct message lengths can be determined using 23811af0c4SBarry Smith `PetscGatherMessageLengths()` 24e5c89e4eSSatish Balay 25667f096bSBarry Smith Either `iflags` or `ilengths` should be provided. If `iflags` is not 26667f096bSBarry Smith provided (`NULL`) it can be computed from `ilengths`. If `iflags` is 27667f096bSBarry Smith provided, `ilengths` is not required. 28e5c89e4eSSatish Balay 29811af0c4SBarry Smith .seealso: `PetscGatherMessageLengths()`, `PetscGatherMessageLengths2()`, `PetscCommBuildTwoSided()` 30e5c89e4eSSatish Balay @*/ 31d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscGatherNumberOfMessages(MPI_Comm comm, const PetscMPIInt iflags[], const PetscMPIInt ilengths[], PetscMPIInt *nrecvs) 32d71ae5a4SJacob Faibussowitsch { 33*8e3a54c0SPierre Jolivet PetscMPIInt size, rank, *recv_buf, i, *iflags_local = NULL, *iflags_localm; 34e5c89e4eSSatish Balay 35e5c89e4eSSatish Balay PetscFunctionBegin; 369566063dSJacob Faibussowitsch PetscCallMPI(MPI_Comm_size(comm, &size)); 379566063dSJacob Faibussowitsch PetscCallMPI(MPI_Comm_rank(comm, &rank)); 38e5c89e4eSSatish Balay 399566063dSJacob Faibussowitsch PetscCall(PetscMalloc2(size, &recv_buf, size, &iflags_localm)); 40e5c89e4eSSatish Balay 41e5c89e4eSSatish Balay /* If iflags not provided, compute iflags from ilengths */ 42e5c89e4eSSatish Balay if (!iflags) { 4328b400f6SJacob Faibussowitsch PetscCheck(ilengths, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Either iflags or ilengths should be provided"); 44e5c89e4eSSatish Balay iflags_local = iflags_localm; 45e5c89e4eSSatish Balay for (i = 0; i < size; i++) { 46e5c89e4eSSatish Balay if (ilengths[i]) iflags_local[i] = 1; 47e5c89e4eSSatish Balay else iflags_local[i] = 0; 48e5c89e4eSSatish Balay } 49a297a907SKarl Rupp } else iflags_local = (PetscMPIInt *)iflags; 50e5c89e4eSSatish Balay 51be87f6c0SPierre Jolivet /* Post an allreduce to determine the number of messages the current MPI rank will receive */ 521c2dc1cbSBarry Smith PetscCall(MPIU_Allreduce(iflags_local, recv_buf, size, MPI_INT, MPI_SUM, comm)); 53e5c89e4eSSatish Balay *nrecvs = recv_buf[rank]; 54e5c89e4eSSatish Balay 559566063dSJacob Faibussowitsch PetscCall(PetscFree2(recv_buf, iflags_localm)); 563ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 57e5c89e4eSSatish Balay } 58e5c89e4eSSatish Balay 59e5c89e4eSSatish Balay /*@C 60da81f932SPierre Jolivet PetscGatherMessageLengths - Computes information about messages that an MPI rank will receive, 61e5c89e4eSSatish Balay including (from-id,length) pairs for each message. 62e5c89e4eSSatish Balay 63d083f849SBarry Smith Collective 64e5c89e4eSSatish Balay 65e5c89e4eSSatish Balay Input Parameters: 66e5c89e4eSSatish Balay + comm - Communicator 67e5c89e4eSSatish Balay . nsends - number of messages that are to be sent. 68e5c89e4eSSatish Balay . nrecvs - number of messages being received 69e5c89e4eSSatish Balay - ilengths - an array of integers of length sizeof(comm) 70667f096bSBarry Smith a non zero `ilengths`[i] represent a message to i of length `ilengths`[i] 71e5c89e4eSSatish Balay 72e5c89e4eSSatish Balay Output Parameters: 73667f096bSBarry Smith + onodes - list of ranks from which messages are expected 74e5c89e4eSSatish Balay - olengths - corresponding message lengths 75e5c89e4eSSatish Balay 76e5c89e4eSSatish Balay Level: developer 77e5c89e4eSSatish Balay 78e5c89e4eSSatish Balay Notes: 79811af0c4SBarry Smith With this info, the correct `MPI_Irecv()` can be posted with the correct 80e5c89e4eSSatish Balay from-id, with a buffer with the right amount of memory required. 81e5c89e4eSSatish Balay 82e5c89e4eSSatish Balay The calling function deallocates the memory in onodes and olengths 83e5c89e4eSSatish Balay 84667f096bSBarry Smith To determine `nrecvs`, one can use `PetscGatherNumberOfMessages()` 85e5c89e4eSSatish Balay 86811af0c4SBarry Smith .seealso: `PetscGatherNumberOfMessages()`, `PetscGatherMessageLengths2()`, `PetscCommBuildTwoSided()` 87e5c89e4eSSatish Balay @*/ 88d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscGatherMessageLengths(MPI_Comm comm, PetscMPIInt nsends, PetscMPIInt nrecvs, const PetscMPIInt ilengths[], PetscMPIInt **onodes, PetscMPIInt **olengths) 89d71ae5a4SJacob Faibussowitsch { 906bfd7d4fSJunchao Zhang PetscMPIInt size, rank, tag, i, j; 91*8e3a54c0SPierre Jolivet MPI_Request *s_waits, *r_waits; 92*8e3a54c0SPierre Jolivet MPI_Status *w_status; 93e5c89e4eSSatish Balay 94e5c89e4eSSatish Balay PetscFunctionBegin; 959566063dSJacob Faibussowitsch PetscCallMPI(MPI_Comm_size(comm, &size)); 969566063dSJacob Faibussowitsch PetscCallMPI(MPI_Comm_rank(comm, &rank)); 979566063dSJacob Faibussowitsch PetscCall(PetscCommGetNewTag(comm, &tag)); 98e5c89e4eSSatish Balay 99e5c89e4eSSatish Balay /* cannot use PetscMalloc3() here because in the call to MPI_Waitall() they MUST be contiguous */ 1009566063dSJacob Faibussowitsch PetscCall(PetscMalloc2(nrecvs + nsends, &r_waits, nrecvs + nsends, &w_status)); 101*8e3a54c0SPierre Jolivet s_waits = PetscSafePointerPlusOffset(r_waits, nrecvs); 102e5c89e4eSSatish Balay 103e5c89e4eSSatish Balay /* Post the Irecv to get the message length-info */ 1049566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(nrecvs, olengths)); 10548a46eb9SPierre Jolivet for (i = 0; i < nrecvs; i++) PetscCallMPI(MPI_Irecv((*olengths) + i, 1, MPI_INT, MPI_ANY_SOURCE, tag, comm, r_waits + i)); 106e5c89e4eSSatish Balay 107e5c89e4eSSatish Balay /* Post the Isends with the message length-info */ 108e5c89e4eSSatish Balay for (i = 0, j = 0; i < size; ++i) { 109e5c89e4eSSatish Balay if (ilengths[i]) { 1109566063dSJacob Faibussowitsch PetscCallMPI(MPI_Isend((void *)(ilengths + i), 1, MPI_INT, i, tag, comm, s_waits + j)); 111e5c89e4eSSatish Balay j++; 112e5c89e4eSSatish Balay } 113e5c89e4eSSatish Balay } 114e5c89e4eSSatish Balay 1156aad120cSJose E. Roman /* Post waits on sends and receives */ 1169566063dSJacob Faibussowitsch if (nrecvs + nsends) PetscCallMPI(MPI_Waitall(nrecvs + nsends, r_waits, w_status)); 117e5c89e4eSSatish Balay 118e5c89e4eSSatish Balay /* Pack up the received data */ 1199566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(nrecvs, onodes)); 1206bfd7d4fSJunchao Zhang for (i = 0; i < nrecvs; ++i) { 1216bfd7d4fSJunchao Zhang (*onodes)[i] = w_status[i].MPI_SOURCE; 1226bfd7d4fSJunchao Zhang #if defined(PETSC_HAVE_OMPI_MAJOR_VERSION) 123337bb527SBarry Smith /* This line is a workaround for a bug in Open MPI 2.1.1 distributed by Ubuntu-18.04.2 LTS. 1246bfd7d4fSJunchao Zhang It happens in self-to-self MPI_Send/Recv using MPI_ANY_SOURCE for message matching. Open MPI 1256bfd7d4fSJunchao Zhang does not put correct value in recv buffer. See also 1266bfd7d4fSJunchao Zhang https://lists.mcs.anl.gov/pipermail/petsc-dev/2019-July/024803.html 1276bfd7d4fSJunchao Zhang https://www.mail-archive.com/users@lists.open-mpi.org//msg33383.html 1286bfd7d4fSJunchao Zhang */ 1296bfd7d4fSJunchao Zhang if (w_status[i].MPI_SOURCE == rank) (*olengths)[i] = ilengths[rank]; 1306bfd7d4fSJunchao Zhang #endif 1316bfd7d4fSJunchao Zhang } 1329566063dSJacob Faibussowitsch PetscCall(PetscFree2(r_waits, w_status)); 1333ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 134e5c89e4eSSatish Balay } 135dd6ea824SBarry Smith 13676eed172SJunchao Zhang /* Same as PetscGatherNumberOfMessages(), except using PetscInt for ilengths[] */ 137d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscGatherNumberOfMessages_Private(MPI_Comm comm, const PetscMPIInt iflags[], const PetscInt ilengths[], PetscMPIInt *nrecvs) 138d71ae5a4SJacob Faibussowitsch { 139*8e3a54c0SPierre Jolivet PetscMPIInt size, rank, *recv_buf, i, *iflags_local = NULL, *iflags_localm; 14076eed172SJunchao Zhang 14176eed172SJunchao Zhang PetscFunctionBegin; 1429566063dSJacob Faibussowitsch PetscCallMPI(MPI_Comm_size(comm, &size)); 1439566063dSJacob Faibussowitsch PetscCallMPI(MPI_Comm_rank(comm, &rank)); 14476eed172SJunchao Zhang 1459566063dSJacob Faibussowitsch PetscCall(PetscMalloc2(size, &recv_buf, size, &iflags_localm)); 14676eed172SJunchao Zhang 14776eed172SJunchao Zhang /* If iflags not provided, compute iflags from ilengths */ 14876eed172SJunchao Zhang if (!iflags) { 14928b400f6SJacob Faibussowitsch PetscCheck(ilengths, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Either iflags or ilengths should be provided"); 15076eed172SJunchao Zhang iflags_local = iflags_localm; 15176eed172SJunchao Zhang for (i = 0; i < size; i++) { 15276eed172SJunchao Zhang if (ilengths[i]) iflags_local[i] = 1; 15376eed172SJunchao Zhang else iflags_local[i] = 0; 15476eed172SJunchao Zhang } 15576eed172SJunchao Zhang } else iflags_local = (PetscMPIInt *)iflags; 15676eed172SJunchao Zhang 157be87f6c0SPierre Jolivet /* Post an allreduce to determine the number of messages the current MPI rank will receive */ 1581c2dc1cbSBarry Smith PetscCall(MPIU_Allreduce(iflags_local, recv_buf, size, MPI_INT, MPI_SUM, comm)); 15976eed172SJunchao Zhang *nrecvs = recv_buf[rank]; 16076eed172SJunchao Zhang 1619566063dSJacob Faibussowitsch PetscCall(PetscFree2(recv_buf, iflags_localm)); 1623ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 16376eed172SJunchao Zhang } 16476eed172SJunchao Zhang 16576eed172SJunchao Zhang /* Same as PetscGatherMessageLengths(), except using PetscInt for message lengths */ 166d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscGatherMessageLengths_Private(MPI_Comm comm, PetscMPIInt nsends, PetscMPIInt nrecvs, const PetscInt ilengths[], PetscMPIInt **onodes, PetscInt **olengths) 167d71ae5a4SJacob Faibussowitsch { 16876eed172SJunchao Zhang PetscMPIInt size, rank, tag, i, j; 169*8e3a54c0SPierre Jolivet MPI_Request *s_waits, *r_waits; 170*8e3a54c0SPierre Jolivet MPI_Status *w_status; 17176eed172SJunchao Zhang 17276eed172SJunchao Zhang PetscFunctionBegin; 1739566063dSJacob Faibussowitsch PetscCallMPI(MPI_Comm_size(comm, &size)); 1749566063dSJacob Faibussowitsch PetscCallMPI(MPI_Comm_rank(comm, &rank)); 1759566063dSJacob Faibussowitsch PetscCall(PetscCommGetNewTag(comm, &tag)); 17676eed172SJunchao Zhang 17776eed172SJunchao Zhang /* cannot use PetscMalloc3() here because in the call to MPI_Waitall() they MUST be contiguous */ 1789566063dSJacob Faibussowitsch PetscCall(PetscMalloc2(nrecvs + nsends, &r_waits, nrecvs + nsends, &w_status)); 179*8e3a54c0SPierre Jolivet s_waits = PetscSafePointerPlusOffset(r_waits, nrecvs); 18076eed172SJunchao Zhang 18176eed172SJunchao Zhang /* Post the Irecv to get the message length-info */ 1829566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(nrecvs, olengths)); 18348a46eb9SPierre Jolivet for (i = 0; i < nrecvs; i++) PetscCallMPI(MPI_Irecv((*olengths) + i, 1, MPIU_INT, MPI_ANY_SOURCE, tag, comm, r_waits + i)); 18476eed172SJunchao Zhang 18576eed172SJunchao Zhang /* Post the Isends with the message length-info */ 18676eed172SJunchao Zhang for (i = 0, j = 0; i < size; ++i) { 18776eed172SJunchao Zhang if (ilengths[i]) { 1889566063dSJacob Faibussowitsch PetscCallMPI(MPI_Isend((void *)(ilengths + i), 1, MPIU_INT, i, tag, comm, s_waits + j)); 18976eed172SJunchao Zhang j++; 19076eed172SJunchao Zhang } 19176eed172SJunchao Zhang } 19276eed172SJunchao Zhang 1936aad120cSJose E. Roman /* Post waits on sends and receives */ 1949566063dSJacob Faibussowitsch if (nrecvs + nsends) PetscCallMPI(MPI_Waitall(nrecvs + nsends, r_waits, w_status)); 19576eed172SJunchao Zhang 19676eed172SJunchao Zhang /* Pack up the received data */ 1979566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(nrecvs, onodes)); 19876eed172SJunchao Zhang for (i = 0; i < nrecvs; ++i) { 19976eed172SJunchao Zhang (*onodes)[i] = w_status[i].MPI_SOURCE; 20076eed172SJunchao Zhang if (w_status[i].MPI_SOURCE == rank) (*olengths)[i] = ilengths[rank]; /* See comments in PetscGatherMessageLengths */ 20176eed172SJunchao Zhang } 2029566063dSJacob Faibussowitsch PetscCall(PetscFree2(r_waits, w_status)); 2033ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 20476eed172SJunchao Zhang } 20576eed172SJunchao Zhang 206e5c89e4eSSatish Balay /*@C 207811af0c4SBarry Smith PetscGatherMessageLengths2 - Computes info about messages that a MPI rank will receive, 208811af0c4SBarry Smith including (from-id,length) pairs for each message. Same functionality as `PetscGatherMessageLengths()` 209e5c89e4eSSatish Balay except it takes TWO ilenths and output TWO olengths. 210e5c89e4eSSatish Balay 211d083f849SBarry Smith Collective 212e5c89e4eSSatish Balay 213e5c89e4eSSatish Balay Input Parameters: 214e5c89e4eSSatish Balay + comm - Communicator 215e5c89e4eSSatish Balay . nsends - number of messages that are to be sent. 216e5c89e4eSSatish Balay . nrecvs - number of messages being received 2176b867d5aSJose E. Roman . ilengths1 - first array of integers of length sizeof(comm) 2186b867d5aSJose E. Roman - ilengths2 - second array of integers of length sizeof(comm) 219e5c89e4eSSatish Balay 220e5c89e4eSSatish Balay Output Parameters: 221667f096bSBarry Smith + onodes - list of ranks from which messages are expected 2226b867d5aSJose E. Roman . olengths1 - first corresponding message lengths 2236b867d5aSJose E. Roman - olengths2 - second message lengths 224e5c89e4eSSatish Balay 225e5c89e4eSSatish Balay Level: developer 226e5c89e4eSSatish Balay 227e5c89e4eSSatish Balay Notes: 228811af0c4SBarry Smith With this info, the correct `MPI_Irecv()` can be posted with the correct 229e5c89e4eSSatish Balay from-id, with a buffer with the right amount of memory required. 230e5c89e4eSSatish Balay 231667f096bSBarry Smith The calling function should `PetscFree()` the memory in `onodes` and `olengths` 232e5c89e4eSSatish Balay 233667f096bSBarry Smith To determine `nrecvs`, one can use `PetscGatherNumberOfMessages()` 234e5c89e4eSSatish Balay 235811af0c4SBarry Smith .seealso: `PetscGatherMessageLengths()`, `PetscGatherNumberOfMessages()`, `PetscCommBuildTwoSided()` 236e5c89e4eSSatish Balay @*/ 237d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscGatherMessageLengths2(MPI_Comm comm, PetscMPIInt nsends, PetscMPIInt nrecvs, const PetscMPIInt ilengths1[], const PetscMPIInt ilengths2[], PetscMPIInt **onodes, PetscMPIInt **olengths1, PetscMPIInt **olengths2) 238d71ae5a4SJacob Faibussowitsch { 239*8e3a54c0SPierre Jolivet PetscMPIInt size, tag, i, j, *buf_s, *buf_r, *buf_j = NULL; 240*8e3a54c0SPierre Jolivet MPI_Request *s_waits, *r_waits; 241*8e3a54c0SPierre Jolivet MPI_Status *w_status; 242e5c89e4eSSatish Balay 243e5c89e4eSSatish Balay PetscFunctionBegin; 2449566063dSJacob Faibussowitsch PetscCallMPI(MPI_Comm_size(comm, &size)); 2459566063dSJacob Faibussowitsch PetscCall(PetscCommGetNewTag(comm, &tag)); 246e5c89e4eSSatish Balay 2473bf92927SBarry Smith /* cannot use PetscMalloc5() because r_waits and s_waits must be contiguous for the call to MPI_Waitall() */ 2489566063dSJacob Faibussowitsch PetscCall(PetscMalloc4(nrecvs + nsends, &r_waits, 2 * nrecvs, &buf_r, 2 * nsends, &buf_s, nrecvs + nsends, &w_status)); 249*8e3a54c0SPierre Jolivet s_waits = PetscSafePointerPlusOffset(r_waits, nrecvs); 250e5c89e4eSSatish Balay 251e5c89e4eSSatish Balay /* Post the Irecv to get the message length-info */ 2529566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(nrecvs + 1, olengths1)); 2539566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(nrecvs + 1, olengths2)); 254e5c89e4eSSatish Balay for (i = 0; i < nrecvs; i++) { 255e5c89e4eSSatish Balay buf_j = buf_r + (2 * i); 2569566063dSJacob Faibussowitsch PetscCallMPI(MPI_Irecv(buf_j, 2, MPI_INT, MPI_ANY_SOURCE, tag, comm, r_waits + i)); 257e5c89e4eSSatish Balay } 258e5c89e4eSSatish Balay 259e5c89e4eSSatish Balay /* Post the Isends with the message length-info */ 260e5c89e4eSSatish Balay for (i = 0, j = 0; i < size; ++i) { 261e5c89e4eSSatish Balay if (ilengths1[i]) { 262e5c89e4eSSatish Balay buf_j = buf_s + (2 * j); 263e5c89e4eSSatish Balay buf_j[0] = *(ilengths1 + i); 264e5c89e4eSSatish Balay buf_j[1] = *(ilengths2 + i); 2659566063dSJacob Faibussowitsch PetscCallMPI(MPI_Isend(buf_j, 2, MPI_INT, i, tag, comm, s_waits + j)); 266e5c89e4eSSatish Balay j++; 267e5c89e4eSSatish Balay } 268e5c89e4eSSatish Balay } 26908401ef6SPierre Jolivet PetscCheck(j == nsends, PETSC_COMM_SELF, PETSC_ERR_PLIB, "j %d not equal to expected number of sends %d", j, nsends); 270e5c89e4eSSatish Balay 2716aad120cSJose E. Roman /* Post waits on sends and receives */ 2729566063dSJacob Faibussowitsch if (nrecvs + nsends) PetscCallMPI(MPI_Waitall(nrecvs + nsends, r_waits, w_status)); 273e5c89e4eSSatish Balay 274e5c89e4eSSatish Balay /* Pack up the received data */ 2759566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(nrecvs + 1, onodes)); 276e5c89e4eSSatish Balay for (i = 0; i < nrecvs; ++i) { 277e5c89e4eSSatish Balay (*onodes)[i] = w_status[i].MPI_SOURCE; 278e5c89e4eSSatish Balay buf_j = buf_r + (2 * i); 279e5c89e4eSSatish Balay (*olengths1)[i] = buf_j[0]; 280e5c89e4eSSatish Balay (*olengths2)[i] = buf_j[1]; 281e5c89e4eSSatish Balay } 282e5c89e4eSSatish Balay 2839566063dSJacob Faibussowitsch PetscCall(PetscFree4(r_waits, buf_r, buf_s, w_status)); 2843ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 285e5c89e4eSSatish Balay } 286e5c89e4eSSatish Balay 287e5c89e4eSSatish Balay /* 288a5b23f4aSJose E. Roman Allocate a buffer sufficient to hold messages of size specified in olengths. 289e5c89e4eSSatish Balay And post Irecvs on these buffers using node info from onodes 290e5c89e4eSSatish Balay */ 291d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscPostIrecvInt(MPI_Comm comm, PetscMPIInt tag, PetscMPIInt nrecvs, const PetscMPIInt onodes[], const PetscMPIInt olengths[], PetscInt ***rbuf, MPI_Request **r_waits) 292d71ae5a4SJacob Faibussowitsch { 293c05d87d6SBarry Smith PetscInt **rbuf_t, i, len = 0; 294e5c89e4eSSatish Balay MPI_Request *r_waits_t; 295e5c89e4eSSatish Balay 296e5c89e4eSSatish Balay PetscFunctionBegin; 297e5c89e4eSSatish Balay /* compute memory required for recv buffers */ 298e5c89e4eSSatish Balay for (i = 0; i < nrecvs; i++) len += olengths[i]; /* each message length */ 299e5c89e4eSSatish Balay 300e5c89e4eSSatish Balay /* allocate memory for recv buffers */ 3019566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(nrecvs + 1, &rbuf_t)); 3029566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(len, &rbuf_t[0])); 303e5c89e4eSSatish Balay for (i = 1; i < nrecvs; ++i) rbuf_t[i] = rbuf_t[i - 1] + olengths[i - 1]; 304e5c89e4eSSatish Balay 305e5c89e4eSSatish Balay /* Post the receives */ 3069566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(nrecvs, &r_waits_t)); 30748a46eb9SPierre Jolivet for (i = 0; i < nrecvs; ++i) PetscCallMPI(MPI_Irecv(rbuf_t[i], olengths[i], MPIU_INT, onodes[i], tag, comm, r_waits_t + i)); 308e5c89e4eSSatish Balay 309e5c89e4eSSatish Balay *rbuf = rbuf_t; 310e5c89e4eSSatish Balay *r_waits = r_waits_t; 3113ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 312e5c89e4eSSatish Balay } 313e5c89e4eSSatish Balay 314d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscPostIrecvScalar(MPI_Comm comm, PetscMPIInt tag, PetscMPIInt nrecvs, const PetscMPIInt onodes[], const PetscMPIInt olengths[], PetscScalar ***rbuf, MPI_Request **r_waits) 315d71ae5a4SJacob Faibussowitsch { 316052f0c41SBarry Smith PetscMPIInt i; 317e5c89e4eSSatish Balay PetscScalar **rbuf_t; 318e5c89e4eSSatish Balay MPI_Request *r_waits_t; 319c05d87d6SBarry Smith PetscInt len = 0; 320e5c89e4eSSatish Balay 321fe28d99cSBarry Smith PetscFunctionBegin; 322e5c89e4eSSatish Balay /* compute memory required for recv buffers */ 323e5c89e4eSSatish Balay for (i = 0; i < nrecvs; i++) len += olengths[i]; /* each message length */ 324e5c89e4eSSatish Balay 325e5c89e4eSSatish Balay /* allocate memory for recv buffers */ 3269566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(nrecvs + 1, &rbuf_t)); 3279566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(len, &rbuf_t[0])); 328e5c89e4eSSatish Balay for (i = 1; i < nrecvs; ++i) rbuf_t[i] = rbuf_t[i - 1] + olengths[i - 1]; 329e5c89e4eSSatish Balay 330e5c89e4eSSatish Balay /* Post the receives */ 3319566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(nrecvs, &r_waits_t)); 33248a46eb9SPierre Jolivet for (i = 0; i < nrecvs; ++i) PetscCallMPI(MPI_Irecv(rbuf_t[i], olengths[i], MPIU_SCALAR, onodes[i], tag, comm, r_waits_t + i)); 333e5c89e4eSSatish Balay 334e5c89e4eSSatish Balay *rbuf = rbuf_t; 335e5c89e4eSSatish Balay *r_waits = r_waits_t; 3363ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 337e5c89e4eSSatish Balay } 338