Fix RAP memory leak (#806)
This PR fixes a memory leak in the parallel RAP routine that arises when a single task is used and R contains `diagT` at input.
This commit is contained in:
parent
79e4ddb5a8
commit
41961c2935
@ -623,6 +623,7 @@ hypre_ParCSRMatrixRAPKTDevice( hypre_ParCSRMatrix *R,
|
|||||||
hypre_ParCSRMatrix *P,
|
hypre_ParCSRMatrix *P,
|
||||||
HYPRE_Int keep_transpose )
|
HYPRE_Int keep_transpose )
|
||||||
{
|
{
|
||||||
|
MPI_Comm comm = hypre_ParCSRMatrixComm(A);
|
||||||
hypre_CSRMatrix *R_diag = hypre_ParCSRMatrixDiag(R);
|
hypre_CSRMatrix *R_diag = hypre_ParCSRMatrixDiag(R);
|
||||||
hypre_CSRMatrix *R_offd = hypre_ParCSRMatrixOffd(R);
|
hypre_CSRMatrix *R_offd = hypre_ParCSRMatrixOffd(R);
|
||||||
|
|
||||||
@ -633,7 +634,7 @@ hypre_ParCSRMatrixRAPKTDevice( hypre_ParCSRMatrix *R,
|
|||||||
HYPRE_BigInt *col_map_offd_C = NULL;
|
HYPRE_BigInt *col_map_offd_C = NULL;
|
||||||
|
|
||||||
HYPRE_Int num_procs;
|
HYPRE_Int num_procs;
|
||||||
MPI_Comm comm = hypre_ParCSRMatrixComm(A);
|
|
||||||
hypre_MPI_Comm_size(comm, &num_procs);
|
hypre_MPI_Comm_size(comm, &num_procs);
|
||||||
|
|
||||||
if ( hypre_ParCSRMatrixGlobalNumRows(R) != hypre_ParCSRMatrixGlobalNumRows(A) ||
|
if ( hypre_ParCSRMatrixGlobalNumRows(R) != hypre_ParCSRMatrixGlobalNumRows(A) ||
|
||||||
@ -717,12 +718,14 @@ hypre_ParCSRMatrixRAPKTDevice( hypre_ParCSRMatrix *R,
|
|||||||
hypre_CSRMatrixDestroy(Abar);
|
hypre_CSRMatrixDestroy(Abar);
|
||||||
hypre_CSRMatrixDestroy(Pbar);
|
hypre_CSRMatrixDestroy(Pbar);
|
||||||
|
|
||||||
hypre_assert(hypre_CSRMatrixNumRows(Cbar) == hypre_ParCSRMatrixNumCols(R) + hypre_CSRMatrixNumCols(
|
hypre_assert(hypre_CSRMatrixNumRows(Cbar) ==
|
||||||
R_offd));
|
hypre_ParCSRMatrixNumCols(R) + hypre_CSRMatrixNumCols(R_offd));
|
||||||
hypre_assert(hypre_CSRMatrixNumCols(Cbar) == hypre_ParCSRMatrixNumCols(P) + num_cols_offd);
|
hypre_assert(hypre_CSRMatrixNumCols(Cbar) ==
|
||||||
|
hypre_ParCSRMatrixNumCols(P) + num_cols_offd);
|
||||||
|
|
||||||
hypre_TMemcpy(&local_nnz_Cbar, hypre_CSRMatrixI(Cbar) + hypre_ParCSRMatrixNumCols(R), HYPRE_Int, 1,
|
hypre_TMemcpy(&local_nnz_Cbar,
|
||||||
HYPRE_MEMORY_HOST, HYPRE_MEMORY_DEVICE);
|
hypre_CSRMatrixI(Cbar) + hypre_ParCSRMatrixNumCols(R),
|
||||||
|
HYPRE_Int, 1, HYPRE_MEMORY_HOST, HYPRE_MEMORY_DEVICE);
|
||||||
|
|
||||||
// Cint is the bottom part of Cbar
|
// Cint is the bottom part of Cbar
|
||||||
Cint = hypre_CSRMatrixCreate(hypre_CSRMatrixNumCols(R_offd), hypre_CSRMatrixNumCols(Cbar),
|
Cint = hypre_CSRMatrixCreate(hypre_CSRMatrixNumCols(R_offd), hypre_CSRMatrixNumCols(Cbar),
|
||||||
@ -748,10 +751,12 @@ hypre_ParCSRMatrixRAPKTDevice( hypre_ParCSRMatrix *R,
|
|||||||
|
|
||||||
// Change Cint into a BigJ matrix
|
// Change Cint into a BigJ matrix
|
||||||
// RL: TODO FIX the 'big' num of columns to global size
|
// RL: TODO FIX the 'big' num of columns to global size
|
||||||
hypre_CSRMatrixBigJ(Cint) = hypre_TAlloc(HYPRE_BigInt, hypre_CSRMatrixNumNonzeros(Cint),
|
hypre_CSRMatrixBigJ(Cint) = hypre_TAlloc(HYPRE_BigInt,
|
||||||
|
hypre_CSRMatrixNumNonzeros(Cint),
|
||||||
HYPRE_MEMORY_DEVICE);
|
HYPRE_MEMORY_DEVICE);
|
||||||
|
|
||||||
RAP_functor<1, HYPRE_BigInt> func1(hypre_ParCSRMatrixNumCols(P), hypre_ParCSRMatrixFirstColDiag(P),
|
RAP_functor<1, HYPRE_BigInt> func1(hypre_ParCSRMatrixNumCols(P),
|
||||||
|
hypre_ParCSRMatrixFirstColDiag(P),
|
||||||
col_map_offd);
|
col_map_offd);
|
||||||
#if defined(HYPRE_USING_SYCL)
|
#if defined(HYPRE_USING_SYCL)
|
||||||
HYPRE_ONEDPL_CALL( std::transform,
|
HYPRE_ONEDPL_CALL( std::transform,
|
||||||
@ -780,7 +785,8 @@ hypre_ParCSRMatrixRAPKTDevice( hypre_ParCSRMatrix *R,
|
|||||||
hypre_TFree(hypre_CSRMatrixBigJ(Cint), HYPRE_MEMORY_DEVICE);
|
hypre_TFree(hypre_CSRMatrixBigJ(Cint), HYPRE_MEMORY_DEVICE);
|
||||||
hypre_TFree(Cint, HYPRE_MEMORY_HOST);
|
hypre_TFree(Cint, HYPRE_MEMORY_HOST);
|
||||||
|
|
||||||
hypre_TMemcpy(hypre_CSRMatrixI(Cbar) + hypre_ParCSRMatrixNumCols(R), &local_nnz_Cbar, HYPRE_Int, 1,
|
hypre_TMemcpy(hypre_CSRMatrixI(Cbar) + hypre_ParCSRMatrixNumCols(R),
|
||||||
|
&local_nnz_Cbar, HYPRE_Int, 1,
|
||||||
HYPRE_MEMORY_DEVICE, HYPRE_MEMORY_HOST);
|
HYPRE_MEMORY_DEVICE, HYPRE_MEMORY_HOST);
|
||||||
|
|
||||||
/* add Cext to local part of Cbar */
|
/* add Cext to local part of Cbar */
|
||||||
@ -806,10 +812,24 @@ hypre_ParCSRMatrixRAPKTDevice( hypre_ParCSRMatrix *R,
|
|||||||
hypre_CSRMatrix *R_diagT;
|
hypre_CSRMatrix *R_diagT;
|
||||||
hypre_CSRMatrix *A_diag = hypre_ParCSRMatrixDiag(A);
|
hypre_CSRMatrix *A_diag = hypre_ParCSRMatrixDiag(A);
|
||||||
hypre_CSRMatrix *P_diag = hypre_ParCSRMatrixDiag(P);
|
hypre_CSRMatrix *P_diag = hypre_ParCSRMatrixDiag(P);
|
||||||
|
|
||||||
|
/* Recover or compute transpose of R_diag */
|
||||||
|
if (hypre_ParCSRMatrixDiagT(R))
|
||||||
|
{
|
||||||
|
R_diagT = hypre_ParCSRMatrixDiagT(R);
|
||||||
|
}
|
||||||
|
else
|
||||||
|
{
|
||||||
hypre_CSRMatrixTransposeDevice(R_diag, &R_diagT, 1);
|
hypre_CSRMatrixTransposeDevice(R_diag, &R_diagT, 1);
|
||||||
|
}
|
||||||
|
|
||||||
C_diag = hypre_CSRMatrixTripleMultiplyDevice(R_diagT, A_diag, P_diag);
|
C_diag = hypre_CSRMatrixTripleMultiplyDevice(R_diagT, A_diag, P_diag);
|
||||||
C_offd = hypre_CSRMatrixCreate(hypre_ParCSRMatrixNumCols(R), 0, 0);
|
C_offd = hypre_CSRMatrixCreate(hypre_ParCSRMatrixNumCols(R), 0, 0);
|
||||||
hypre_CSRMatrixInitialize_v2(C_offd, 0, HYPRE_MEMORY_DEVICE);
|
hypre_CSRMatrixInitialize_v2(C_offd, 0, HYPRE_MEMORY_DEVICE);
|
||||||
|
|
||||||
|
/* Keep or destroy transpose of R_diag */
|
||||||
|
if (!hypre_ParCSRMatrixDiagT(R))
|
||||||
|
{
|
||||||
if (keep_transpose)
|
if (keep_transpose)
|
||||||
{
|
{
|
||||||
hypre_ParCSRMatrixDiagT(R) = R_diagT;
|
hypre_ParCSRMatrixDiagT(R) = R_diagT;
|
||||||
@ -819,6 +839,7 @@ hypre_ParCSRMatrixRAPKTDevice( hypre_ParCSRMatrix *R,
|
|||||||
hypre_CSRMatrixDestroy(R_diagT);
|
hypre_CSRMatrixDestroy(R_diagT);
|
||||||
}
|
}
|
||||||
}
|
}
|
||||||
|
}
|
||||||
|
|
||||||
/* Move the diagonal entry to the first of each row */
|
/* Move the diagonal entry to the first of each row */
|
||||||
hypre_CSRMatrixMoveDiagFirstDevice(C_diag);
|
hypre_CSRMatrixMoveDiagFirstDevice(C_diag);
|
||||||
@ -842,9 +863,11 @@ hypre_ParCSRMatrixRAPKTDevice( hypre_ParCSRMatrix *R,
|
|||||||
{
|
{
|
||||||
hypre_ParCSRMatrixDeviceColMapOffd(C) = col_map_offd_C;
|
hypre_ParCSRMatrixDeviceColMapOffd(C) = col_map_offd_C;
|
||||||
|
|
||||||
hypre_ParCSRMatrixColMapOffd(C) = hypre_TAlloc(HYPRE_BigInt, num_cols_offd_C, HYPRE_MEMORY_HOST);
|
hypre_ParCSRMatrixColMapOffd(C) = hypre_TAlloc(HYPRE_BigInt,
|
||||||
hypre_TMemcpy(hypre_ParCSRMatrixColMapOffd(C), col_map_offd_C, HYPRE_BigInt, num_cols_offd_C,
|
num_cols_offd_C,
|
||||||
HYPRE_MEMORY_HOST, HYPRE_MEMORY_DEVICE);
|
HYPRE_MEMORY_HOST);
|
||||||
|
hypre_TMemcpy(hypre_ParCSRMatrixColMapOffd(C), col_map_offd_C, HYPRE_BigInt,
|
||||||
|
num_cols_offd_C, HYPRE_MEMORY_HOST, HYPRE_MEMORY_DEVICE);
|
||||||
}
|
}
|
||||||
|
|
||||||
hypre_assert(!hypre_CSRMatrixCheckDiagFirstDevice(hypre_ParCSRMatrixDiag(C)));
|
hypre_assert(!hypre_CSRMatrixCheckDiagFirstDevice(hypre_ParCSRMatrixDiag(C)));
|
||||||
|
|||||||
Loading…
Reference in New Issue
Block a user