On Mon, May 1, 2023 at 8:14 AM Andrew Ho wrote:
> Hi,
>
> I noticed that when I compile PETSc/HYPRE with GPU support, it demands
> that I use GPU vectors/matrices (in the form of either
> VECCUDA/MATMPIAIJCUSPARSE, VECHIP/MATMPIAIJHIPSPARSE, or
> VECKOKKOS/MATMPIAIJKOKKOS). However, I would like to do some
> comparisons/tuning vs. CPU solvers for my particular application, and when
> I try to pass in regular CPU vectors/sparse matrices, PETSc complains:
>
> [0]PETSC ERROR: Invalid argument
> [0]PETSC ERROR: HYPRE_MEMORY_DEVICE expects a device vector. You need to
> enable PETSc device support, for example, in some cases, -vec_type cuda
> [0]PETSC ERROR: See https://petsc.org/release/faq/ for trouble shooting.
> [0]PETSC ERROR: Petsc Development GIT revision: v3.19.1-126-g02e62876438
> GIT Date: 2023-04-30 09:10:41 -0500
> [0]PETSC ERROR: ./myprog on a named butter by pb Sun Apr 30 23:07:05 2023
> [0]PETSC ERROR: Configure options --prefix=/home/pb/opt/petsc/3.19.1
> --with-debugging=0 --with-large-file-io=1 --with-hdf5-dir=/home/pb/.local
> --download-hypre --download-metis --download-parmetis
> --download-superlu_dist --COPTFLAGS=-O3 --FOPTFLAGS=-O3 --CXXOPTFLAGS=-O3
> --CUDAOPTFLAGS=-O3 --with-64-bit-indices=1 --download-slepc
> --download-fblaslapack --with-cuda --with-kokkos-dir=/home/pb/.local
> --with-kokkos-kernels-dir=/home/pb/.local --with-openmp
> [0]PETSC ERROR: #1 VecGetArrayForHYPRE() at
> /scratch/pb/code/third_party/petsc/petsc/src/vec/vec/impls/hypre/vhyp.c:95
> [0]PETSC ERROR: #2 VecHYPRE_IJVectorPushVecRead() at
> /scratch/pb/code/third_party/petsc/petsc/src/vec/vec/impls/hypre/vhyp.c:138
> [0]PETSC ERROR: #3 PCApply_HYPRE() at
> /scratch/pb/code/third_party/petsc/petsc/src/ksp/pc/impls/hypre/hypre.c:433
> [0]PETSC ERROR: #4 PCApply() at
> /scratch/pb/code/third_party/petsc/petsc/src/ksp/pc/interface/precon.c:441
>
> [0]PETSC ERROR: #5 KSP_PCApply() at
> /scratch/pb/code/third_party/petsc/petsc/include/petsc/private/kspimpl.h:381
> [0]PETSC ERROR: #6 KSPInitialResidual() at
> /scratch/pb/code/third_party/petsc/petsc/src/ksp/ksp/interface/itres.c:64
> [0]PETSC ERROR: #7 KSPSolve_GMRES() at
> /scratch/pb/code/third_party/petsc/petsc/src/ksp/ksp/impls/gmres/gmres.c:226
> [0]PETSC ERROR: #8 KSPSolve_Private() at
> /scratch/pb/code/third_party/petsc/petsc/src/ksp/ksp/interface/itfunc.c:898
> [0]PETSC ERROR: #9 KSPSolve() at
> /scratch/pb/code/third_party/petsc/petsc/src/ksp/ksp/interface/itfunc.c:1070
> [0]PETSC ERROR: #10 main() at
> /home/pb/code/misc/kokkos_test/src/main.cpp:257
> [0]PETSC ERROR: PETSc Option Table entries:
> [0]PETSC ERROR: -pc_hypre_type boomeramg (source: command line)
> [0]PETSC ERROR: -pc_type hypre (source: command line)
> [0]PETSC ERROR: -use_gpu_aware_mpi 0 (source: command line)
>
> Do I have to use a separate PETSc/HYPRE build to do the CPU/GPU
> comparisons?
>
Yes, that is the case for now. Hypre recently added preliminary support
for running CPU/GPU tests with a single build. I did some experiments in
petsc with it. I got rid of the error you showed, but failed to have the
expected convergence rate. So, there are still bugs.
>
> I was also wondering if the base PETSc KSP solvers use any kind of GPU
> acceleration, and if this is configurable at runtime.
>
Unlike Hypre, if petsc is configured with GPU support, it can run CPU-only
tests (even on machines with GPUs)
>
> I tried solving a 2D poisson system, and whether I allocated CPU or GPU
> vectors the KSP solver will complete in roughly the same non-trivial amount
> of time, which makes me suspect that either it is never being solved on the
> GPU or is always solved on the GPU since I'm only using a single CPU core.
> This is with PETSC compiled with CUDA/Kokkos support.
>
You can add -log_view -log_view_gpu_time to compare the profiling result.
