1e5c89e4eSSatish Balay 2c6db04a5SJed Brown #include <petscsys.h> /*I "petscsys.h" I*/ 376eed172SJunchao Zhang #include <petsc/private/mpiutils.h> 4e5c89e4eSSatish Balay 5e5c89e4eSSatish Balay /*@C 6811af0c4SBarry Smith PetscGatherNumberOfMessages - Computes the number of messages an MPI rank expects to receive during a neighbor communication 7e5c89e4eSSatish Balay 8d083f849SBarry Smith Collective 9e5c89e4eSSatish Balay 10e5c89e4eSSatish Balay Input Parameters: 11e5c89e4eSSatish Balay + comm - Communicator 12e5c89e4eSSatish Balay . iflags - an array of integers of length sizeof(comm). A '1' in ilengths[i] represent a 130298fd71SBarry Smith message from current node to ith node. Optionally NULL 14e5c89e4eSSatish Balay - ilengths - Non zero ilengths[i] represent a message to i of length ilengths[i]. 150298fd71SBarry Smith Optionally NULL. 16e5c89e4eSSatish Balay 17e5c89e4eSSatish Balay Output Parameters: 18e5c89e4eSSatish Balay . nrecvs - number of messages received 19e5c89e4eSSatish Balay 20e5c89e4eSSatish Balay Level: developer 21e5c89e4eSSatish Balay 22e5c89e4eSSatish Balay Notes: 23e5c89e4eSSatish Balay With this info, the correct message lengths can be determined using 24811af0c4SBarry Smith `PetscGatherMessageLengths()` 25e5c89e4eSSatish Balay 26e5c89e4eSSatish Balay Either iflags or ilengths should be provided. If iflags is not 270298fd71SBarry Smith provided (NULL) it can be computed from ilengths. If iflags is 28e5c89e4eSSatish Balay provided, ilengths is not required. 29e5c89e4eSSatish Balay 30811af0c4SBarry Smith .seealso: `PetscGatherMessageLengths()`, `PetscGatherMessageLengths2()`, `PetscCommBuildTwoSided()` 31e5c89e4eSSatish Balay @*/ 32d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscGatherNumberOfMessages(MPI_Comm comm, const PetscMPIInt iflags[], const PetscMPIInt ilengths[], PetscMPIInt *nrecvs) 33d71ae5a4SJacob Faibussowitsch { 340298fd71SBarry Smith PetscMPIInt size, rank, *recv_buf, i, *iflags_local = NULL, *iflags_localm = NULL; 35e5c89e4eSSatish Balay 36e5c89e4eSSatish Balay PetscFunctionBegin; 379566063dSJacob Faibussowitsch PetscCallMPI(MPI_Comm_size(comm, &size)); 389566063dSJacob Faibussowitsch PetscCallMPI(MPI_Comm_rank(comm, &rank)); 39e5c89e4eSSatish Balay 409566063dSJacob Faibussowitsch PetscCall(PetscMalloc2(size, &recv_buf, size, &iflags_localm)); 41e5c89e4eSSatish Balay 42e5c89e4eSSatish Balay /* If iflags not provided, compute iflags from ilengths */ 43e5c89e4eSSatish Balay if (!iflags) { 4428b400f6SJacob Faibussowitsch PetscCheck(ilengths, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Either iflags or ilengths should be provided"); 45e5c89e4eSSatish Balay iflags_local = iflags_localm; 46e5c89e4eSSatish Balay for (i = 0; i < size; i++) { 47e5c89e4eSSatish Balay if (ilengths[i]) iflags_local[i] = 1; 48e5c89e4eSSatish Balay else iflags_local[i] = 0; 49e5c89e4eSSatish Balay } 50a297a907SKarl Rupp } else iflags_local = (PetscMPIInt *)iflags; 51e5c89e4eSSatish Balay 52e5c89e4eSSatish Balay /* Post an allreduce to determine the numer of messages the current node will receive */ 531c2dc1cbSBarry Smith PetscCall(MPIU_Allreduce(iflags_local, recv_buf, size, MPI_INT, MPI_SUM, comm)); 54e5c89e4eSSatish Balay *nrecvs = recv_buf[rank]; 55e5c89e4eSSatish Balay 569566063dSJacob Faibussowitsch PetscCall(PetscFree2(recv_buf, iflags_localm)); 57*3ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 58e5c89e4eSSatish Balay } 59e5c89e4eSSatish Balay 60e5c89e4eSSatish Balay /*@C 61811af0c4SBarry Smith PetscGatherMessageLengths - Computes infomation about messages that an MPI rank will receive, 62e5c89e4eSSatish Balay including (from-id,length) pairs for each message. 63e5c89e4eSSatish Balay 64d083f849SBarry Smith Collective 65e5c89e4eSSatish Balay 66e5c89e4eSSatish Balay Input Parameters: 67e5c89e4eSSatish Balay + comm - Communicator 68e5c89e4eSSatish Balay . nsends - number of messages that are to be sent. 69e5c89e4eSSatish Balay . nrecvs - number of messages being received 70e5c89e4eSSatish Balay - ilengths - an array of integers of length sizeof(comm) 71e5c89e4eSSatish Balay a non zero ilengths[i] represent a message to i of length ilengths[i] 72e5c89e4eSSatish Balay 73e5c89e4eSSatish Balay Output Parameters: 74e5c89e4eSSatish Balay + onodes - list of node-ids from which messages are expected 75e5c89e4eSSatish Balay - olengths - corresponding message lengths 76e5c89e4eSSatish Balay 77e5c89e4eSSatish Balay Level: developer 78e5c89e4eSSatish Balay 79e5c89e4eSSatish Balay Notes: 80811af0c4SBarry Smith With this info, the correct `MPI_Irecv()` can be posted with the correct 81e5c89e4eSSatish Balay from-id, with a buffer with the right amount of memory required. 82e5c89e4eSSatish Balay 83e5c89e4eSSatish Balay The calling function deallocates the memory in onodes and olengths 84e5c89e4eSSatish Balay 85811af0c4SBarry Smith To determine nrecvs, one can use `PetscGatherNumberOfMessages()` 86e5c89e4eSSatish Balay 87811af0c4SBarry Smith .seealso: `PetscGatherNumberOfMessages()`, `PetscGatherMessageLengths2()`, `PetscCommBuildTwoSided()` 88e5c89e4eSSatish Balay @*/ 89d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscGatherMessageLengths(MPI_Comm comm, PetscMPIInt nsends, PetscMPIInt nrecvs, const PetscMPIInt ilengths[], PetscMPIInt **onodes, PetscMPIInt **olengths) 90d71ae5a4SJacob Faibussowitsch { 916bfd7d4fSJunchao Zhang PetscMPIInt size, rank, tag, i, j; 920298fd71SBarry Smith MPI_Request *s_waits = NULL, *r_waits = NULL; 930298fd71SBarry Smith MPI_Status *w_status = NULL; 94e5c89e4eSSatish Balay 95e5c89e4eSSatish Balay PetscFunctionBegin; 969566063dSJacob Faibussowitsch PetscCallMPI(MPI_Comm_size(comm, &size)); 979566063dSJacob Faibussowitsch PetscCallMPI(MPI_Comm_rank(comm, &rank)); 989566063dSJacob Faibussowitsch PetscCall(PetscCommGetNewTag(comm, &tag)); 99e5c89e4eSSatish Balay 100e5c89e4eSSatish Balay /* cannot use PetscMalloc3() here because in the call to MPI_Waitall() they MUST be contiguous */ 1019566063dSJacob Faibussowitsch PetscCall(PetscMalloc2(nrecvs + nsends, &r_waits, nrecvs + nsends, &w_status)); 102e5c89e4eSSatish Balay s_waits = r_waits + nrecvs; 103e5c89e4eSSatish Balay 104e5c89e4eSSatish Balay /* Post the Irecv to get the message length-info */ 1059566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(nrecvs, olengths)); 10648a46eb9SPierre Jolivet for (i = 0; i < nrecvs; i++) PetscCallMPI(MPI_Irecv((*olengths) + i, 1, MPI_INT, MPI_ANY_SOURCE, tag, comm, r_waits + i)); 107e5c89e4eSSatish Balay 108e5c89e4eSSatish Balay /* Post the Isends with the message length-info */ 109e5c89e4eSSatish Balay for (i = 0, j = 0; i < size; ++i) { 110e5c89e4eSSatish Balay if (ilengths[i]) { 1119566063dSJacob Faibussowitsch PetscCallMPI(MPI_Isend((void *)(ilengths + i), 1, MPI_INT, i, tag, comm, s_waits + j)); 112e5c89e4eSSatish Balay j++; 113e5c89e4eSSatish Balay } 114e5c89e4eSSatish Balay } 115e5c89e4eSSatish Balay 1166aad120cSJose E. Roman /* Post waits on sends and receives */ 1179566063dSJacob Faibussowitsch if (nrecvs + nsends) PetscCallMPI(MPI_Waitall(nrecvs + nsends, r_waits, w_status)); 118e5c89e4eSSatish Balay 119e5c89e4eSSatish Balay /* Pack up the received data */ 1209566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(nrecvs, onodes)); 1216bfd7d4fSJunchao Zhang for (i = 0; i < nrecvs; ++i) { 1226bfd7d4fSJunchao Zhang (*onodes)[i] = w_status[i].MPI_SOURCE; 1236bfd7d4fSJunchao Zhang #if defined(PETSC_HAVE_OMPI_MAJOR_VERSION) 1246bfd7d4fSJunchao Zhang /* This line is a workaround for a bug in OpenMPI-2.1.1 distributed by Ubuntu-18.04.2 LTS. 1256bfd7d4fSJunchao Zhang It happens in self-to-self MPI_Send/Recv using MPI_ANY_SOURCE for message matching. OpenMPI 1266bfd7d4fSJunchao Zhang does not put correct value in recv buffer. See also 1276bfd7d4fSJunchao Zhang https://lists.mcs.anl.gov/pipermail/petsc-dev/2019-July/024803.html 1286bfd7d4fSJunchao Zhang https://www.mail-archive.com/users@lists.open-mpi.org//msg33383.html 1296bfd7d4fSJunchao Zhang */ 1306bfd7d4fSJunchao Zhang if (w_status[i].MPI_SOURCE == rank) (*olengths)[i] = ilengths[rank]; 1316bfd7d4fSJunchao Zhang #endif 1326bfd7d4fSJunchao Zhang } 1339566063dSJacob Faibussowitsch PetscCall(PetscFree2(r_waits, w_status)); 134*3ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 135e5c89e4eSSatish Balay } 136dd6ea824SBarry Smith 13776eed172SJunchao Zhang /* Same as PetscGatherNumberOfMessages(), except using PetscInt for ilengths[] */ 138d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscGatherNumberOfMessages_Private(MPI_Comm comm, const PetscMPIInt iflags[], const PetscInt ilengths[], PetscMPIInt *nrecvs) 139d71ae5a4SJacob Faibussowitsch { 14076eed172SJunchao Zhang PetscMPIInt size, rank, *recv_buf, i, *iflags_local = NULL, *iflags_localm = NULL; 14176eed172SJunchao Zhang 14276eed172SJunchao Zhang PetscFunctionBegin; 1439566063dSJacob Faibussowitsch PetscCallMPI(MPI_Comm_size(comm, &size)); 1449566063dSJacob Faibussowitsch PetscCallMPI(MPI_Comm_rank(comm, &rank)); 14576eed172SJunchao Zhang 1469566063dSJacob Faibussowitsch PetscCall(PetscMalloc2(size, &recv_buf, size, &iflags_localm)); 14776eed172SJunchao Zhang 14876eed172SJunchao Zhang /* If iflags not provided, compute iflags from ilengths */ 14976eed172SJunchao Zhang if (!iflags) { 15028b400f6SJacob Faibussowitsch PetscCheck(ilengths, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Either iflags or ilengths should be provided"); 15176eed172SJunchao Zhang iflags_local = iflags_localm; 15276eed172SJunchao Zhang for (i = 0; i < size; i++) { 15376eed172SJunchao Zhang if (ilengths[i]) iflags_local[i] = 1; 15476eed172SJunchao Zhang else iflags_local[i] = 0; 15576eed172SJunchao Zhang } 15676eed172SJunchao Zhang } else iflags_local = (PetscMPIInt *)iflags; 15776eed172SJunchao Zhang 15876eed172SJunchao Zhang /* Post an allreduce to determine the numer of messages the current node will receive */ 1591c2dc1cbSBarry Smith PetscCall(MPIU_Allreduce(iflags_local, recv_buf, size, MPI_INT, MPI_SUM, comm)); 16076eed172SJunchao Zhang *nrecvs = recv_buf[rank]; 16176eed172SJunchao Zhang 1629566063dSJacob Faibussowitsch PetscCall(PetscFree2(recv_buf, iflags_localm)); 163*3ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 16476eed172SJunchao Zhang } 16576eed172SJunchao Zhang 16676eed172SJunchao Zhang /* Same as PetscGatherMessageLengths(), except using PetscInt for message lengths */ 167d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscGatherMessageLengths_Private(MPI_Comm comm, PetscMPIInt nsends, PetscMPIInt nrecvs, const PetscInt ilengths[], PetscMPIInt **onodes, PetscInt **olengths) 168d71ae5a4SJacob Faibussowitsch { 16976eed172SJunchao Zhang PetscMPIInt size, rank, tag, i, j; 17076eed172SJunchao Zhang MPI_Request *s_waits = NULL, *r_waits = NULL; 17176eed172SJunchao Zhang MPI_Status *w_status = NULL; 17276eed172SJunchao Zhang 17376eed172SJunchao Zhang PetscFunctionBegin; 1749566063dSJacob Faibussowitsch PetscCallMPI(MPI_Comm_size(comm, &size)); 1759566063dSJacob Faibussowitsch PetscCallMPI(MPI_Comm_rank(comm, &rank)); 1769566063dSJacob Faibussowitsch PetscCall(PetscCommGetNewTag(comm, &tag)); 17776eed172SJunchao Zhang 17876eed172SJunchao Zhang /* cannot use PetscMalloc3() here because in the call to MPI_Waitall() they MUST be contiguous */ 1799566063dSJacob Faibussowitsch PetscCall(PetscMalloc2(nrecvs + nsends, &r_waits, nrecvs + nsends, &w_status)); 18076eed172SJunchao Zhang s_waits = r_waits + nrecvs; 18176eed172SJunchao Zhang 18276eed172SJunchao Zhang /* Post the Irecv to get the message length-info */ 1839566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(nrecvs, olengths)); 18448a46eb9SPierre Jolivet for (i = 0; i < nrecvs; i++) PetscCallMPI(MPI_Irecv((*olengths) + i, 1, MPIU_INT, MPI_ANY_SOURCE, tag, comm, r_waits + i)); 18576eed172SJunchao Zhang 18676eed172SJunchao Zhang /* Post the Isends with the message length-info */ 18776eed172SJunchao Zhang for (i = 0, j = 0; i < size; ++i) { 18876eed172SJunchao Zhang if (ilengths[i]) { 1899566063dSJacob Faibussowitsch PetscCallMPI(MPI_Isend((void *)(ilengths + i), 1, MPIU_INT, i, tag, comm, s_waits + j)); 19076eed172SJunchao Zhang j++; 19176eed172SJunchao Zhang } 19276eed172SJunchao Zhang } 19376eed172SJunchao Zhang 1946aad120cSJose E. Roman /* Post waits on sends and receives */ 1959566063dSJacob Faibussowitsch if (nrecvs + nsends) PetscCallMPI(MPI_Waitall(nrecvs + nsends, r_waits, w_status)); 19676eed172SJunchao Zhang 19776eed172SJunchao Zhang /* Pack up the received data */ 1989566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(nrecvs, onodes)); 19976eed172SJunchao Zhang for (i = 0; i < nrecvs; ++i) { 20076eed172SJunchao Zhang (*onodes)[i] = w_status[i].MPI_SOURCE; 20176eed172SJunchao Zhang if (w_status[i].MPI_SOURCE == rank) (*olengths)[i] = ilengths[rank]; /* See comments in PetscGatherMessageLengths */ 20276eed172SJunchao Zhang } 2039566063dSJacob Faibussowitsch PetscCall(PetscFree2(r_waits, w_status)); 204*3ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 20576eed172SJunchao Zhang } 20676eed172SJunchao Zhang 207e5c89e4eSSatish Balay /*@C 208811af0c4SBarry Smith PetscGatherMessageLengths2 - Computes info about messages that a MPI rank will receive, 209811af0c4SBarry Smith including (from-id,length) pairs for each message. Same functionality as `PetscGatherMessageLengths()` 210e5c89e4eSSatish Balay except it takes TWO ilenths and output TWO olengths. 211e5c89e4eSSatish Balay 212d083f849SBarry Smith Collective 213e5c89e4eSSatish Balay 214e5c89e4eSSatish Balay Input Parameters: 215e5c89e4eSSatish Balay + comm - Communicator 216e5c89e4eSSatish Balay . nsends - number of messages that are to be sent. 217e5c89e4eSSatish Balay . nrecvs - number of messages being received 2186b867d5aSJose E. Roman . ilengths1 - first array of integers of length sizeof(comm) 2196b867d5aSJose E. Roman - ilengths2 - second array of integers of length sizeof(comm) 220e5c89e4eSSatish Balay 221e5c89e4eSSatish Balay Output Parameters: 222e5c89e4eSSatish Balay + onodes - list of node-ids from which messages are expected 2236b867d5aSJose E. Roman . olengths1 - first corresponding message lengths 2246b867d5aSJose E. Roman - olengths2 - second message lengths 225e5c89e4eSSatish Balay 226e5c89e4eSSatish Balay Level: developer 227e5c89e4eSSatish Balay 228e5c89e4eSSatish Balay Notes: 229811af0c4SBarry Smith With this info, the correct `MPI_Irecv()` can be posted with the correct 230e5c89e4eSSatish Balay from-id, with a buffer with the right amount of memory required. 231e5c89e4eSSatish Balay 232e5c89e4eSSatish Balay The calling function deallocates the memory in onodes and olengths 233e5c89e4eSSatish Balay 234c2916339SPierre Jolivet To determine nrecvs, one can use PetscGatherNumberOfMessages() 235e5c89e4eSSatish Balay 236811af0c4SBarry Smith .seealso: `PetscGatherMessageLengths()`, `PetscGatherNumberOfMessages()`, `PetscCommBuildTwoSided()` 237e5c89e4eSSatish Balay @*/ 238d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscGatherMessageLengths2(MPI_Comm comm, PetscMPIInt nsends, PetscMPIInt nrecvs, const PetscMPIInt ilengths1[], const PetscMPIInt ilengths2[], PetscMPIInt **onodes, PetscMPIInt **olengths1, PetscMPIInt **olengths2) 239d71ae5a4SJacob Faibussowitsch { 2400298fd71SBarry Smith PetscMPIInt size, tag, i, j, *buf_s = NULL, *buf_r = NULL, *buf_j = NULL; 2410298fd71SBarry Smith MPI_Request *s_waits = NULL, *r_waits = NULL; 2420298fd71SBarry Smith MPI_Status *w_status = NULL; 243e5c89e4eSSatish Balay 244e5c89e4eSSatish Balay PetscFunctionBegin; 2459566063dSJacob Faibussowitsch PetscCallMPI(MPI_Comm_size(comm, &size)); 2469566063dSJacob Faibussowitsch PetscCall(PetscCommGetNewTag(comm, &tag)); 247e5c89e4eSSatish Balay 2483bf92927SBarry Smith /* cannot use PetscMalloc5() because r_waits and s_waits must be contiguous for the call to MPI_Waitall() */ 2499566063dSJacob Faibussowitsch PetscCall(PetscMalloc4(nrecvs + nsends, &r_waits, 2 * nrecvs, &buf_r, 2 * nsends, &buf_s, nrecvs + nsends, &w_status)); 250e5c89e4eSSatish Balay s_waits = r_waits + nrecvs; 251e5c89e4eSSatish Balay 252e5c89e4eSSatish Balay /* Post the Irecv to get the message length-info */ 2539566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(nrecvs + 1, olengths1)); 2549566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(nrecvs + 1, olengths2)); 255e5c89e4eSSatish Balay for (i = 0; i < nrecvs; i++) { 256e5c89e4eSSatish Balay buf_j = buf_r + (2 * i); 2579566063dSJacob Faibussowitsch PetscCallMPI(MPI_Irecv(buf_j, 2, MPI_INT, MPI_ANY_SOURCE, tag, comm, r_waits + i)); 258e5c89e4eSSatish Balay } 259e5c89e4eSSatish Balay 260e5c89e4eSSatish Balay /* Post the Isends with the message length-info */ 261e5c89e4eSSatish Balay for (i = 0, j = 0; i < size; ++i) { 262e5c89e4eSSatish Balay if (ilengths1[i]) { 263e5c89e4eSSatish Balay buf_j = buf_s + (2 * j); 264e5c89e4eSSatish Balay buf_j[0] = *(ilengths1 + i); 265e5c89e4eSSatish Balay buf_j[1] = *(ilengths2 + i); 2669566063dSJacob Faibussowitsch PetscCallMPI(MPI_Isend(buf_j, 2, MPI_INT, i, tag, comm, s_waits + j)); 267e5c89e4eSSatish Balay j++; 268e5c89e4eSSatish Balay } 269e5c89e4eSSatish Balay } 27008401ef6SPierre Jolivet PetscCheck(j == nsends, PETSC_COMM_SELF, PETSC_ERR_PLIB, "j %d not equal to expected number of sends %d", j, nsends); 271e5c89e4eSSatish Balay 2726aad120cSJose E. Roman /* Post waits on sends and receives */ 2739566063dSJacob Faibussowitsch if (nrecvs + nsends) PetscCallMPI(MPI_Waitall(nrecvs + nsends, r_waits, w_status)); 274e5c89e4eSSatish Balay 275e5c89e4eSSatish Balay /* Pack up the received data */ 2769566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(nrecvs + 1, onodes)); 277e5c89e4eSSatish Balay for (i = 0; i < nrecvs; ++i) { 278e5c89e4eSSatish Balay (*onodes)[i] = w_status[i].MPI_SOURCE; 279e5c89e4eSSatish Balay buf_j = buf_r + (2 * i); 280e5c89e4eSSatish Balay (*olengths1)[i] = buf_j[0]; 281e5c89e4eSSatish Balay (*olengths2)[i] = buf_j[1]; 282e5c89e4eSSatish Balay } 283e5c89e4eSSatish Balay 2849566063dSJacob Faibussowitsch PetscCall(PetscFree4(r_waits, buf_r, buf_s, w_status)); 285*3ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 286e5c89e4eSSatish Balay } 287e5c89e4eSSatish Balay 288e5c89e4eSSatish Balay /* 289a5b23f4aSJose E. Roman Allocate a buffer sufficient to hold messages of size specified in olengths. 290e5c89e4eSSatish Balay And post Irecvs on these buffers using node info from onodes 291e5c89e4eSSatish Balay */ 292d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscPostIrecvInt(MPI_Comm comm, PetscMPIInt tag, PetscMPIInt nrecvs, const PetscMPIInt onodes[], const PetscMPIInt olengths[], PetscInt ***rbuf, MPI_Request **r_waits) 293d71ae5a4SJacob Faibussowitsch { 294c05d87d6SBarry Smith PetscInt **rbuf_t, i, len = 0; 295e5c89e4eSSatish Balay MPI_Request *r_waits_t; 296e5c89e4eSSatish Balay 297e5c89e4eSSatish Balay PetscFunctionBegin; 298e5c89e4eSSatish Balay /* compute memory required for recv buffers */ 299e5c89e4eSSatish Balay for (i = 0; i < nrecvs; i++) len += olengths[i]; /* each message length */ 300e5c89e4eSSatish Balay 301e5c89e4eSSatish Balay /* allocate memory for recv buffers */ 3029566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(nrecvs + 1, &rbuf_t)); 3039566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(len, &rbuf_t[0])); 304e5c89e4eSSatish Balay for (i = 1; i < nrecvs; ++i) rbuf_t[i] = rbuf_t[i - 1] + olengths[i - 1]; 305e5c89e4eSSatish Balay 306e5c89e4eSSatish Balay /* Post the receives */ 3079566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(nrecvs, &r_waits_t)); 30848a46eb9SPierre 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)); 309e5c89e4eSSatish Balay 310e5c89e4eSSatish Balay *rbuf = rbuf_t; 311e5c89e4eSSatish Balay *r_waits = r_waits_t; 312*3ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 313e5c89e4eSSatish Balay } 314e5c89e4eSSatish Balay 315d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscPostIrecvScalar(MPI_Comm comm, PetscMPIInt tag, PetscMPIInt nrecvs, const PetscMPIInt onodes[], const PetscMPIInt olengths[], PetscScalar ***rbuf, MPI_Request **r_waits) 316d71ae5a4SJacob Faibussowitsch { 317052f0c41SBarry Smith PetscMPIInt i; 318e5c89e4eSSatish Balay PetscScalar **rbuf_t; 319e5c89e4eSSatish Balay MPI_Request *r_waits_t; 320c05d87d6SBarry Smith PetscInt len = 0; 321e5c89e4eSSatish Balay 322fe28d99cSBarry Smith PetscFunctionBegin; 323e5c89e4eSSatish Balay /* compute memory required for recv buffers */ 324e5c89e4eSSatish Balay for (i = 0; i < nrecvs; i++) len += olengths[i]; /* each message length */ 325e5c89e4eSSatish Balay 326e5c89e4eSSatish Balay /* allocate memory for recv buffers */ 3279566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(nrecvs + 1, &rbuf_t)); 3289566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(len, &rbuf_t[0])); 329e5c89e4eSSatish Balay for (i = 1; i < nrecvs; ++i) rbuf_t[i] = rbuf_t[i - 1] + olengths[i - 1]; 330e5c89e4eSSatish Balay 331e5c89e4eSSatish Balay /* Post the receives */ 3329566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(nrecvs, &r_waits_t)); 33348a46eb9SPierre 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)); 334e5c89e4eSSatish Balay 335e5c89e4eSSatish Balay *rbuf = rbuf_t; 336e5c89e4eSSatish Balay *r_waits = r_waits_t; 337*3ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 338e5c89e4eSSatish Balay } 339