Re: [petsc-dev] Faulty logic in MatHYPRESetPreallocation?
Stefano,Please find attach a trivial example that will segfault from within hypre if you put back on the (IMHO unnecessary) hypre_AuxParCSRMatrixNeedAux(aux_matrix) = 1 in PR #1305.You need to use two processes or more.Thanks,Pierre ex15.c Description: Binary data On 29 Dec 2018, at 10:13 AM, Stefano Zampiniwrote:Barry,Probably this is an oversight from me. When I coded the MATHYPRE class, I did not make much attention to the preallocation. Also, hypre did not seem very robust to matrix reusage at the time. Take a look at the comments at line 85 in ex115 from mat tests. Probably Pierre's fix will curate that problem too. StefanoIl giorno Sab 22 Dic 2018, 06:30 Smith, Barry F. ha scritto: Stefeno, Doe this look right to you? Thanks Barry > On Dec 17, 2018, at 3:18 AM, Pierre Jolivet via petsc-dev wrote: > > Hello, > I am not sure the hypre routines are called in the right order in MatHYPRESetPreallocation. > Indeed, attached, you’ll find two Massif logs, one with the current order (massif.out.bad), and the other (massif.out.better) with the attached patch applied. > In the former, you can see at the snapshot #31 that there is a call to hypre_CTAlloc in hypre_AuxParCSRMatrixInitialize (aux_parcsr_matrix.c:163). > This means that hypre is allocating memory as if it was not aware of the number of nonzero per row, even though I supply in my application the two arrays dnnz and onnz. > > Am I missing something? > > Thanks, > Pierre > >
Re: [petsc-dev] Faulty logic in MatHYPRESetPreallocation?
Pierre, Could you make a pull request on this? Thanks Barry > On Dec 17, 2018, at 3:18 AM, Pierre Jolivet via petsc-dev > wrote: > > Hello, > I am not sure the hypre routines are called in the right order in > MatHYPRESetPreallocation. > Indeed, attached, you’ll find two Massif logs, one with the current order > (massif.out.bad), and the other (massif.out.better) with the attached patch > applied. > In the former, you can see at the snapshot #31 that there is a call to > hypre_CTAlloc in hypre_AuxParCSRMatrixInitialize (aux_parcsr_matrix.c:163). > This means that hypre is allocating memory as if it was not aware of the > number of nonzero per row, even though I supply in my application the two > arrays dnnz and onnz. > > Am I missing something? > > Thanks, > Pierre > >
Re: [petsc-dev] Faulty logic in MatHYPRESetPreallocation?
Stefeno, Doe this look right to you? Thanks Barry > On Dec 17, 2018, at 3:18 AM, Pierre Jolivet via petsc-dev > wrote: > > Hello, > I am not sure the hypre routines are called in the right order in > MatHYPRESetPreallocation. > Indeed, attached, you’ll find two Massif logs, one with the current order > (massif.out.bad), and the other (massif.out.better) with the attached patch > applied. > In the former, you can see at the snapshot #31 that there is a call to > hypre_CTAlloc in hypre_AuxParCSRMatrixInitialize (aux_parcsr_matrix.c:163). > This means that hypre is allocating memory as if it was not aware of the > number of nonzero per row, even though I supply in my application the two > arrays dnnz and onnz. > > Am I missing something? > > Thanks, > Pierre > >