MATSEQAIJCUSPARSE: add support for sparse MatMat operations

CUDA: add a couple of missing Lazy initializations

checkbadSource (master) fixes

Merge remote-tracking branch 'origin/release'

checkbadSource: apply rules to *.cu *.cpp sources, and expand CHKERRQ check to CHKERR(Q|MPI|CUDA|CUBLAS|CUSPARSE)

Merge remote-tracking branch 'origin/release'

call PetscCUDAInitializeCheck() in MatConvert_SeqAIJ_SeqAIJCUSPARSE since it uses cuSparse()

Missing this sometimes produces errors in the CU tests, but other times it passes. I don't understand why

call PetscCUDAInitializeCheck() in MatConvert_SeqAIJ_SeqAIJCUSPARSE since it uses cuSparse()

Missing this sometimes produces errors in the CU tests, but other times it passes. I don't understand why it is not needed.

Commit-type: error-checking, Bug-fix, portability-fix, testing-fix, CUDA
/spend 10m
Reported-by: GitLab CI

show more ...

Merge branch 'adams/cusparse-host-buffer' into 'master'

Adams/cusparse host buffer

See merge request petsc/petsc!3502

added cache of buffer to avoid managed memory mallocs

Adding Cuda and Kokkos assembly. Added Device assembly to Landau operator. Added Kokkos test mat/ex5k.

MATSEQAJICUSPARSE: perform analysis only once

CUSPARSE: use csrsv2 interface starting from CUDA 9.0

MATSEQAIJCUSPARSE: fix bug with all zero matrix

MATCUSPARSE: Implement fast assembly from COO data

Merge remote-tracking branch 'origin/release' into master

bug fix

Reducing first coarse grid to one process/device in Cuda and manual override of this and other heuristics for process reduction in GAMG coarse grids.

Report all GPU errors as PETSC_ERR_GPU so that the test harness can retry any GPU failure

We still get occassional errors:

ot ok ksp_ksp_tutorials-ex71_bddc_cusparse # Error code: 180

That occur a

Report all GPU errors as PETSC_ERR_GPU so that the test harness can retry any GPU failure

We still get occassional errors:

ot ok ksp_ksp_tutorials-ex71_bddc_cusparse # Error code: 180

That occur after the cuSPARSE library must have been loaded since the load command already succeeded

Commit-type: error-checking, CI, enhancement
/spend 15m

show more ...

MATMPIAIJCUSPARSE: fix conversions

MatInode: set it to false when using device matrices

Fix building PETSC 3.14 with cuda 10.0

add cuda matrix for meta data method, assembly

Do not log transfer time as GPU time. GPU time is used to get performance ON the GPU (for example flop rates)
logging transfer times distorts these numbers since they would include the time to copy d

Do not log transfer time as GPU time. GPU time is used to get performance ON the GPU (for example flop rates)
logging transfer times distorts these numbers since they would include the time to copy down the matrix.
The transfer times are already logged in the copy to and from events and the performance flop rates including the
copy times are already logged in the overall flop rate.

Commit-type: bug-fix, style-fix
/spend 5m

show more ...

Add support of CUDA-11

Revert "CUDA: Get PETSc to compile with CUDA 11."

This reverts commit 21a6055d1ede0202436f221d06f48a7b9836052d.