dense/mpi/mmdense.c

/*
   Support for the parallel dense matrix vector multiply
*/
#include <../src/mat/impls/dense/mpi/mpidense.h>
#include <petscblaslapack.h>

PetscErrorCode MatSetUpMultiply_MPIDense(Mat mat)
{
  Mat_MPIDense *mdn = (Mat_MPIDense *)mat->data;

  PetscFunctionBegin;
  if (!mdn->Mvctx) {
    /* Create local vector that is used to scatter into */
    PetscCall(VecDestroy(&mdn->lvec));
    if (mdn->A) { PetscCall(MatCreateVecs(mdn->A, &mdn->lvec, NULL)); }
    PetscCall(PetscLayoutSetUp(mat->cmap));
    PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)mat), &mdn->Mvctx));
    PetscCall(PetscSFSetGraphWithPattern(mdn->Mvctx, mat->cmap, PETSCSF_PATTERN_ALLGATHER));
  }
  PetscFunctionReturn(PETSC_SUCCESS);
}

static PetscErrorCode MatCreateSubMatrices_MPIDense_Local(Mat, PetscInt, const IS[], const IS[], MatReuse, Mat *);

PetscErrorCode MatCreateSubMatrices_MPIDense(Mat C, PetscInt ismax, const IS isrow[], const IS iscol[], MatReuse scall, Mat *submat[])
{
  PetscInt nmax, nstages_local, nstages, i, pos, max_no;

  PetscFunctionBegin;
  /* Allocate memory to hold all the submatrices */
  if (scall != MAT_REUSE_MATRIX) PetscCall(PetscCalloc1(ismax + 1, submat));
  /* Determine the number of stages through which submatrices are done */
  nmax = 20 * 1000000 / (C->cmap->N * sizeof(PetscInt));
  if (!nmax) nmax = 1;
  nstages_local = ismax / nmax + ((ismax % nmax) ? 1 : 0);

  /* Make sure every processor loops through the nstages */
  PetscCall(MPIU_Allreduce(&nstages_local, &nstages, 1, MPIU_INT, MPI_MAX, PetscObjectComm((PetscObject)C)));

  for (i = 0, pos = 0; i < nstages; i++) {
    if (pos + nmax <= ismax) max_no = nmax;
    else if (pos == ismax) max_no = 0;
    else max_no = ismax - pos;
    PetscCall(MatCreateSubMatrices_MPIDense_Local(C, max_no, isrow + pos, iscol + pos, scall, *submat + pos));
    pos += max_no;
  }
  PetscFunctionReturn(PETSC_SUCCESS);
}

static PetscErrorCode MatCreateSubMatrices_MPIDense_Local(Mat C, PetscInt ismax, const IS isrow[], const IS iscol[], MatReuse scall, Mat *submats)
{
  Mat_MPIDense    *c = (Mat_MPIDense *)C->data;
  Mat              A = c->A;
  Mat_SeqDense    *a = (Mat_SeqDense *)A->data, *mat;
  PetscMPIInt      rank, size, tag0, tag1, idex, end, i;
  PetscInt         N = C->cmap->N, rstart = C->rmap->rstart, count;
  const PetscInt **irow, **icol, *irow_i;
  PetscInt        *nrow, *ncol, *w1, *w3, *w4, *rtable, start;
  PetscInt       **sbuf1, m, j, k, l, ct1, **rbuf1, row, proc;
  PetscInt         nrqs, msz, **ptr, *ctr, *pa, *tmp, bsz, nrqr;
  PetscInt         is_no, jmax, **rmap, *rmap_i;
  PetscInt         ctr_j, *sbuf1_j, *rbuf1_i;
  MPI_Request     *s_waits1, *r_waits1, *s_waits2, *r_waits2;
  MPI_Status      *r_status1, *r_status2, *s_status1, *s_status2;
  MPI_Comm         comm;
  PetscScalar    **rbuf2, **sbuf2;
  PetscBool        sorted;

  PetscFunctionBegin;
  PetscCall(PetscObjectGetComm((PetscObject)C, &comm));
  tag0 = ((PetscObject)C)->tag;
  PetscCallMPI(MPI_Comm_rank(comm, &rank));
  PetscCallMPI(MPI_Comm_size(comm, &size));
  m = C->rmap->N;

  /* Get some new tags to keep the communication clean */
  PetscCall(PetscObjectGetNewTag((PetscObject)C, &tag1));

  /* Check if the col indices are sorted */
  for (i = 0; i < ismax; i++) {
    PetscCall(ISSorted(isrow[i], &sorted));
    PetscCheck(sorted, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "ISrow is not sorted");
    PetscCall(ISSorted(iscol[i], &sorted));
    PetscCheck(sorted, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "IScol is not sorted");
  }

  PetscCall(PetscMalloc5(ismax, (PetscInt ***)&irow, ismax, (PetscInt ***)&icol, ismax, &nrow, ismax, &ncol, m, &rtable));
  for (i = 0; i < ismax; i++) {
    PetscCall(ISGetIndices(isrow[i], &irow[i]));
    PetscCall(ISGetIndices(iscol[i], &icol[i]));
    PetscCall(ISGetLocalSize(isrow[i], &nrow[i]));
    PetscCall(ISGetLocalSize(iscol[i], &ncol[i]));
  }

  /* Create hash table for the mapping :row -> proc*/
  for (i = 0, j = 0; i < size; i++) {
    jmax = C->rmap->range[i + 1];
    for (; j < jmax; j++) rtable[j] = i;
  }

  /* evaluate communication - mesg to who,length of mesg, and buffer space
     required. Based on this, buffers are allocated, and data copied into them*/
  PetscCall(PetscMalloc3(2 * size, &w1, size, &w3, size, &w4));
  PetscCall(PetscArrayzero(w1, size * 2)); /* initialize work vector*/
  PetscCall(PetscArrayzero(w3, size));     /* initialize work vector*/
  for (i = 0; i < ismax; i++) {
    PetscCall(PetscArrayzero(w4, size)); /* initialize work vector*/
    jmax   = nrow[i];
    irow_i = irow[i];
    for (j = 0; j < jmax; j++) {
      row  = irow_i[j];
      proc = rtable[row];
      w4[proc]++;
    }
    for (j = 0; j < size; j++) {
      if (w4[j]) {
        w1[2 * j] += w4[j];
        w3[j]++;
      }
    }
  }

  nrqs         = 0; /* no of outgoing messages */
  msz          = 0; /* total mesg length (for all procs) */
  w1[2 * rank] = 0; /* no mesg sent to self */
  w3[rank]     = 0;
  for (i = 0; i < size; i++) {
    if (w1[2 * i]) {
      w1[2 * i + 1] = 1;
      nrqs++;
    } /* there exists a message to proc i */
  }
  PetscCall(PetscMalloc1(nrqs + 1, &pa)); /*(proc -array)*/
  for (i = 0, j = 0; i < size; i++) {
    if (w1[2 * i]) {
      pa[j] = i;
      j++;
    }
  }

  /* Each message would have a header = 1 + 2*(no of IS) + data */
  for (i = 0; i < nrqs; i++) {
    j = pa[i];
    w1[2 * j] += w1[2 * j + 1] + 2 * w3[j];
    msz += w1[2 * j];
  }
  /* Do a global reduction to determine how many messages to expect*/
  PetscCall(PetscMaxSum(comm, w1, &bsz, &nrqr));

  /* Allocate memory for recv buffers . Make sure rbuf1[0] exists by adding 1 to the buffer length */
  PetscCall(PetscMalloc1(nrqr + 1, &rbuf1));
  PetscCall(PetscMalloc1(nrqr * bsz, &rbuf1[0]));
  for (i = 1; i < nrqr; ++i) rbuf1[i] = rbuf1[i - 1] + bsz;

  /* Post the receives */
  PetscCall(PetscMalloc1(nrqr + 1, &r_waits1));
  for (i = 0; i < nrqr; ++i) PetscCallMPI(MPI_Irecv(rbuf1[i], bsz, MPIU_INT, MPI_ANY_SOURCE, tag0, comm, r_waits1 + i));

  /* Allocate Memory for outgoing messages */
  PetscCall(PetscMalloc4(size, &sbuf1, size, &ptr, 2 * msz, &tmp, size, &ctr));
  PetscCall(PetscArrayzero(sbuf1, size));
  PetscCall(PetscArrayzero(ptr, size));
  {
    PetscInt *iptr = tmp, ict = 0;
    for (i = 0; i < nrqs; i++) {
      j = pa[i];
      iptr += ict;
      sbuf1[j] = iptr;
      ict      = w1[2 * j];
    }
  }

  /* Form the outgoing messages */
  /* Initialize the header space */
  for (i = 0; i < nrqs; i++) {
    j           = pa[i];
    sbuf1[j][0] = 0;
    PetscCall(PetscArrayzero(sbuf1[j] + 1, 2 * w3[j]));
    ptr[j] = sbuf1[j] + 2 * w3[j] + 1;
  }

  /* Parse the isrow and copy data into outbuf */
  for (i = 0; i < ismax; i++) {
    PetscCall(PetscArrayzero(ctr, size));
    irow_i = irow[i];
    jmax   = nrow[i];
    for (j = 0; j < jmax; j++) { /* parse the indices of each IS */
      row  = irow_i[j];
      proc = rtable[row];
      if (proc != rank) { /* copy to the outgoing buf*/
        ctr[proc]++;
        *ptr[proc] = row;
        ptr[proc]++;
      }
    }
    /* Update the headers for the current IS */
    for (j = 0; j < size; j++) { /* Can Optimise this loop too */
      if ((ctr_j = ctr[j])) {
        sbuf1_j            = sbuf1[j];
        k                  = ++sbuf1_j[0];
        sbuf1_j[2 * k]     = ctr_j;
        sbuf1_j[2 * k - 1] = i;
      }
    }
  }

  /*  Now  post the sends */
  PetscCall(PetscMalloc1(nrqs + 1, &s_waits1));
  for (i = 0; i < nrqs; ++i) {
    j = pa[i];
    PetscCallMPI(MPI_Isend(sbuf1[j], w1[2 * j], MPIU_INT, j, tag0, comm, s_waits1 + i));
  }

  /* Post receives to capture the row_data from other procs */
  PetscCall(PetscMalloc1(nrqs + 1, &r_waits2));
  PetscCall(PetscMalloc1(nrqs + 1, &rbuf2));
  for (i = 0; i < nrqs; i++) {
    j     = pa[i];
    count = (w1[2 * j] - (2 * sbuf1[j][0] + 1)) * N;
    PetscCall(PetscMalloc1(count + 1, &rbuf2[i]));
    PetscCallMPI(MPI_Irecv(rbuf2[i], count, MPIU_SCALAR, j, tag1, comm, r_waits2 + i));
  }

  /* Receive messages(row_nos) and then, pack and send off the rowvalues
     to the correct processors */

  PetscCall(PetscMalloc1(nrqr + 1, &s_waits2));
  PetscCall(PetscMalloc1(nrqr + 1, &r_status1));
  PetscCall(PetscMalloc1(nrqr + 1, &sbuf2));

  {
    PetscScalar *sbuf2_i, *v_start;
    PetscInt     s_proc;
    for (i = 0; i < nrqr; ++i) {
      PetscCallMPI(MPI_Waitany(nrqr, r_waits1, &idex, r_status1 + i));
      s_proc  = r_status1[i].MPI_SOURCE; /* send processor */
      rbuf1_i = rbuf1[idex];             /* Actual message from s_proc */
      /* no of rows = end - start; since start is array idex[], 0idex, whel end
         is length of the buffer - which is 1idex */
      start = 2 * rbuf1_i[0] + 1;
      PetscCallMPI(MPI_Get_count(r_status1 + i, MPIU_INT, &end));
      /* allocate memory sufficinet to hold all the row values */
      PetscCall(PetscMalloc1((end - start) * N, &sbuf2[idex]));
      sbuf2_i = sbuf2[idex];
      /* Now pack the data */
      for (j = start; j < end; j++) {
        row     = rbuf1_i[j] - rstart;
        v_start = a->v + row;
        for (k = 0; k < N; k++) {
          sbuf2_i[0] = v_start[0];
          sbuf2_i++;
          v_start += a->lda;
        }
      }
      /* Now send off the data */
      PetscCallMPI(MPI_Isend(sbuf2[idex], (end - start) * N, MPIU_SCALAR, s_proc, tag1, comm, s_waits2 + i));
    }
  }
  /* End Send-Recv of IS + row_numbers */
  PetscCall(PetscFree(r_status1));
  PetscCall(PetscFree(r_waits1));
  PetscCall(PetscMalloc1(nrqs + 1, &s_status1));
  if (nrqs) PetscCallMPI(MPI_Waitall(nrqs, s_waits1, s_status1));
  PetscCall(PetscFree(s_status1));
  PetscCall(PetscFree(s_waits1));

  /* Create the submatrices */
  if (scall == MAT_REUSE_MATRIX) {
    for (i = 0; i < ismax; i++) {
      mat = (Mat_SeqDense *)submats[i]->data;
      PetscCheck(!(submats[i]->rmap->n != nrow[i]) && !(submats[i]->cmap->n != ncol[i]), PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Cannot reuse matrix. wrong size");
      PetscCall(PetscArrayzero(mat->v, submats[i]->rmap->n * submats[i]->cmap->n));

      submats[i]->factortype = C->factortype;
    }
  } else {
    for (i = 0; i < ismax; i++) {
      PetscCall(MatCreate(PETSC_COMM_SELF, submats + i));
      PetscCall(MatSetSizes(submats[i], nrow[i], ncol[i], nrow[i], ncol[i]));
      PetscCall(MatSetType(submats[i], ((PetscObject)A)->type_name));
      PetscCall(MatSeqDenseSetPreallocation(submats[i], NULL));
    }
  }

  /* Assemble the matrices */
  {
    PetscInt     col;
    PetscScalar *imat_v, *mat_v, *imat_vi, *mat_vi;

    for (i = 0; i < ismax; i++) {
      mat    = (Mat_SeqDense *)submats[i]->data;
      mat_v  = a->v;
      imat_v = mat->v;
      irow_i = irow[i];
      m      = nrow[i];
      for (j = 0; j < m; j++) {
        row  = irow_i[j];
        proc = rtable[row];
        if (proc == rank) {
          row     = row - rstart;
          mat_vi  = mat_v + row;
          imat_vi = imat_v + j;
          for (k = 0; k < ncol[i]; k++) {
            col            = icol[i][k];
            imat_vi[k * m] = mat_vi[col * a->lda];
          }
        }
      }
    }
  }

  /* Create row map-> This maps c->row to submat->row for each submat*/
  /* this is a very expensive operation wrt memory usage */
  PetscCall(PetscMalloc1(ismax, &rmap));
  PetscCall(PetscCalloc1(ismax * C->rmap->N, &rmap[0]));
  for (i = 1; i < ismax; i++) rmap[i] = rmap[i - 1] + C->rmap->N;
  for (i = 0; i < ismax; i++) {
    rmap_i = rmap[i];
    irow_i = irow[i];
    jmax   = nrow[i];
    for (j = 0; j < jmax; j++) rmap_i[irow_i[j]] = j;
  }

  /* Now Receive the row_values and assemble the rest of the matrix */
  PetscCall(PetscMalloc1(nrqs + 1, &r_status2));
  {
    PetscInt     is_max, tmp1, col, *sbuf1_i, is_sz;
    PetscScalar *rbuf2_i, *imat_v, *imat_vi;

    for (tmp1 = 0; tmp1 < nrqs; tmp1++) { /* For each message */
      PetscCallMPI(MPI_Waitany(nrqs, r_waits2, &i, r_status2 + tmp1));
      /* Now dig out the corresponding sbuf1, which contains the IS data_structure */
      sbuf1_i = sbuf1[pa[i]];
      is_max  = sbuf1_i[0];
      ct1     = 2 * is_max + 1;
      rbuf2_i = rbuf2[i];
      for (j = 1; j <= is_max; j++) { /* For each IS belonging to the message */
        is_no  = sbuf1_i[2 * j - 1];
        is_sz  = sbuf1_i[2 * j];
        mat    = (Mat_SeqDense *)submats[is_no]->data;
        imat_v = mat->v;
        rmap_i = rmap[is_no];
        m      = nrow[is_no];
        for (k = 0; k < is_sz; k++, rbuf2_i += N) { /* For each row */
          row = sbuf1_i[ct1];
          ct1++;
          row     = rmap_i[row];
          imat_vi = imat_v + row;
          for (l = 0; l < ncol[is_no]; l++) { /* For each col */
            col            = icol[is_no][l];
            imat_vi[l * m] = rbuf2_i[col];
          }
        }
      }
    }
  }
  /* End Send-Recv of row_values */
  PetscCall(PetscFree(r_status2));
  PetscCall(PetscFree(r_waits2));
  PetscCall(PetscMalloc1(nrqr + 1, &s_status2));
  if (nrqr) PetscCallMPI(MPI_Waitall(nrqr, s_waits2, s_status2));
  PetscCall(PetscFree(s_status2));
  PetscCall(PetscFree(s_waits2));

  /* Restore the indices */
  for (i = 0; i < ismax; i++) {
    PetscCall(ISRestoreIndices(isrow[i], irow + i));
    PetscCall(ISRestoreIndices(iscol[i], icol + i));
  }

  PetscCall(PetscFree5(*(PetscInt ***)&irow, *(PetscInt ***)&icol, nrow, ncol, rtable));
  PetscCall(PetscFree3(w1, w3, w4));
  PetscCall(PetscFree(pa));

  for (i = 0; i < nrqs; ++i) PetscCall(PetscFree(rbuf2[i]));
  PetscCall(PetscFree(rbuf2));
  PetscCall(PetscFree4(sbuf1, ptr, tmp, ctr));
  PetscCall(PetscFree(rbuf1[0]));
  PetscCall(PetscFree(rbuf1));

  for (i = 0; i < nrqr; ++i) PetscCall(PetscFree(sbuf2[i]));

  PetscCall(PetscFree(sbuf2));
  PetscCall(PetscFree(rmap[0]));
  PetscCall(PetscFree(rmap));

  for (i = 0; i < ismax; i++) {
    PetscCall(MatAssemblyBegin(submats[i], MAT_FINAL_ASSEMBLY));
    PetscCall(MatAssemblyEnd(submats[i], MAT_FINAL_ASSEMBLY));
  }
  PetscFunctionReturn(PETSC_SUCCESS);
}

PETSC_INTERN PetscErrorCode MatScale_MPIDense(Mat inA, PetscScalar alpha)
{
  Mat_MPIDense *A = (Mat_MPIDense *)inA->data;

  PetscFunctionBegin;
  PetscCall(MatScale(A->A, alpha));
  PetscFunctionReturn(PETSC_SUCCESS);
}