On Fri, Jan 28, 2022 at 3:27 PM Mark Adams <[email protected]> wrote:
(Junchao), Ruipeng said this was an OOM error and suggested trying the in-house with
SpGEMM: HYPRE_SetSpGemmUseCusparse(FALSE);
Should I clone a hypre argument to make a -pc_hypre_use_tpl_spgemm or -pc_hypre_use_cusparse_spgemm ?
Better use the same words as hypre, e.g., -pc_hypre_set_spgemm_use_cusparse
On Fri, Jan 28, 2022 at 1:42 PM Mark Adams <[email protected]> wrote:
Moving this to the users list (We can not talk about Crusher on public forums, but this is on Summit. I had to check this thread carefully!)
Treb is using hypre on Summit and getting this error:
CUSPARSE ERROR (code = 11, insufficient resources) at csr_spgemm_device_cusparse.c:128
This is probably from Hypre's RAP.
He has contacted OLCF, which seems like the right place to go, but does anyone have any ideas?
Treb: You might ask Hypre also. We do actually have a fair amount of experience with hypre but hypre has more!
Thanks, Mark
On Fri, Jan 28, 2022 at 1:31 PM David Trebotich <[email protected]> wrote:
Thought I sent you this...will change the order of MatSetOption to see if that helps
I ran it and get that error which I have already sent a ticket to OLCF: CUSPARSE ERROR (code = 11, insufficient resources) at csr_spgemm_device_cusparse.c:128
Here's my petscrc #do not use -mat_view with hypre-cuda if running on gpu #-mat_view :A.m:ascii_matlab -help -proj_mac_pc_type hypre -proj_mac_pc_hypre_type boomeramg -proj_mac_pc_hypre_boomeramg_no_CF -proj_mac_pc_hypre_boomeramg_agg_nl 0 -proj_mac_pc_hypre_boomeramg_coarsen_type PMIS -proj_mac_pc_hypre_boomeramg_interp_type ext+i -proj_mac_pc_hypre_boomeramg_print_statistics -proj_mac_pc_hypre_boomeramg_relax_type_all l1scaled-Jacobi -proj_mac_ksp_type gmres -proj_mac_ksp_max_it 50 -proj_mac_ksp_rtol 1.e-12 -proj_mac_ksp_atol 1.e-30 -mat_type hypre -use_gpu_aware_mpi 0 -log_view -history PETSc.history -visc_ksp_rtol 1.e-12 -visc_pc_type jacobi -visc_ksp_type gmres -visc_ksp_max_it 50 -diff_ksp_rtol 1.e-6 -diff_pc_type jacobi -diff_ksp_max_it 50 -proj_mac_ksp_converged_reason -visc_ksp_converged_reason -diff_ksp_converged_reason -proj_mac_ksp_norm_type unpreconditioned -diff_ksp_norm_type unpreconditioned -visc_ksp_norm_type unpreconditioned
And here's my code: ierr = MatSetSizes(m_mat,NN,NN,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); ierr = MatSetBlockSize(m_mat,nc);CHKERRQ(ierr); ierr = MatSetType(m_mat,MATAIJ);CHKERRQ(ierr); // ierr = MatSetOption(m_mat,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE) ;CHKERRQ(ierr); ierr = MatSetFromOptions( m_mat ); CHKERRQ(ierr); ierr = MatSeqAIJSetPreallocation(m_mat,nnzrow, d_nnz);CHKERRQ(ierr); ierr = MatMPIAIJSetPreallocation(m_mat,nnzrow, d_nnz, nnzrow/2, o_nnz);CHKERRQ(ierr); ierr = MatSetOption(m_mat,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE) ;CHKERRQ(ierr);
#if defined(PETSC_HAVE_HYPRE) ierr = MatHYPRESetPreallocation(m_mat,nnzrow, d_nnz, nnzrow/2, o_nnz);CHKERRQ(ierr); #endif
On Wed, Jan 26, 2022 at 6:18 PM Mark Adams <[email protected]> wrote:
On Wed, Jan 26, 2022 at 7:43 PM David Trebotich <[email protected]> wrote:
Can you confirm with me on the settings in .petscrc for Summit with -pc_type hypre? We were using -mat_type aijcusparse which worked a few months ago and now is not working.
I don't know the difference between cusparse, aijcusparse and hypre as -mat_type
cusparse is not a matrix type. The other two are and they both should work.
I tested the builds that I just sent in another email (eg, PETSC_DIR=/gpfs/alpine/world-shared/geo127/petsc/arch-summit-dbg-gcc-cuda PETSC_ARCH="")
$ make PETSC_DIR=$PWD PETSC_ARCH=arch-summit-opt-gnu-hypre-cuda -f gmakefile test search='ksp_ksp_tutorials-ex55_hypre_device' Using MAKEFLAGS: -- search=ksp_ksp_tutorials-ex55_hypre_device PETSC_ARCH=arch-summit-opt-gnu-hypre-cuda PETSC_DIR=/gpfs/alpine/csc314/scratch/adams/petsc2 CC arch-summit-opt-gnu-hypre-cuda/tests/ksp/ksp/tutorials/ex55.o CLINKER arch-summit-opt-gnu-hypre-cuda/tests/ksp/ksp/tutorials/ex55 TEST arch-summit-opt-gnu-hypre-cuda/tests/counts/ksp_ksp_tutorials-ex55_hypre_device.counts ok ksp_ksp_tutorials-ex55_hypre_device ok diff-ksp_ksp_tutorials-ex55_hypre_device
So this work. In this file (attached) you will see an example of a construction a matrix that we have gone over before:
/* create stiffness matrix */ ierr = MatCreate(comm,&Amat);CHKERRQ(ierr); ierr = MatSetSizes(Amat,m,m,M,M);CHKERRQ(ierr); ierr = MatSetType(Amat,MATAIJ);CHKERRQ(ierr); ierr = MatSetOption(Amat,MAT_SPD,PETSC_TRUE);CHKERRQ(ierr); ierr = MatSetFromOptions(Amat);CHKERRQ(ierr); ierr = MatSetBlockSize(Amat,2);CHKERRQ(ierr); ierr = MatSeqAIJSetPreallocation(Amat,18,NULL);CHKERRQ(ierr); ierr = MatMPIAIJSetPreallocation(Amat,18,NULL,18,NULL);CHKERRQ(ierr); #if defined(PETSC_HAVE_HYPRE) ierr = MatHYPRESetPreallocation(Amat,18,NULL,18,NULL);CHKERRQ(ierr); #endif
At the end of the file you will what is executed with this test for "hypre_device":
# command line options match GPU defaults test: suffix: hypre_device nsize: 4 requires: hypre !complex args: *-mat_type hypre* -ksp_view -ne 29 -alpha 1.e-3 -ksp_type cg *-pc_type hypre **-pc_hypre_type boomeramg *-ksp_monitor_short *-pc_hypre_boomeramg_relax_type_all l1scaled-Jacobi -pc_hypre_boomeramg_interp_type ext+i -pc_hypre_boomeramg_coarsen_type PMIS -pc_hypre_boomeramg_no_CF*
All you need is *-mat_type hypre and -pc_type hypre*. *You could also add these hypre arguments.*
If this is not working please send me a description of the problem, like any error output on your screen and your petsc.history file.
-- ---------------------- David Trebotich Lawrence Berkeley National Laboratory Computational Research Division Applied Numerical Algorithms Group [email protected] (510) 486-5984 office (510) 384-6868 mobile