Move related routines together

Update pack/unpack routines to do packing/unpacking for all neighbors in at most two routines

One is used to pack data in self to self communication; The second is used for remote communication.
So

Update pack/unpack routines to do packing/unpacking for all neighbors in at most two routines

One is used to pack data in self to self communication; The second is used for remote communication.
So that on GPU, we can use at most two kernels to do packing/packing for all neighbors instead of multiple kernels

show more ...

Code style change

Merge remote-tracking branch 'origin/jczhang/fix-reinstate-some-mpi-check'

Reinstate MPI_Win_create check and revise MPI shared memory check

See merge request petsc/petsc!2197

Rename PETSC_HAVE_MPI_WIN_CREATE_FEATURE to PETSC_HAVE_MPI_PROCESS_SHARED_MEMORY

To have clear distinction from PETSC_HAVE_MPI_WIN_CREATE

Added test for VecGhostUpdate with MIN_VALUES

Updated document and manual

Added MAX_VALUES and MIN_VALUES support in VecScatter

Requested-by: "Aulisa, Eugenio" <Eugenio.Aulisa@ttu.edu>

Remove wrong usage of int datatype from PETSc code

Merge branch 'jed/fix-vecscatter-overflow' into 'master'

Jed/fix vecscatter overflow

See merge request petsc/petsc!2093

VecScatter: fix premature overflow caused by PetscInt indexing by bytes

When sizeof(PetscInt) == 4, we can store all array indices, but not
necessarily all offsets in bytes. This effectively reduce

VecScatter: fix premature overflow caused by PetscInt indexing by bytes

When sizeof(PetscInt) == 4, we can store all array indices, but not
necessarily all offsets in bytes. This effectively reduces the max
problem size (in case of large messages) to 2B / sizeof(PetscScalar),
which is an easy limit to hit.

The copy_lengths field could be changed to size_t to accommodate, but
I think the code is more clear to change it to a count in units of
PetscScalar and then use PetscArraycpy instead of PetscMemcpy, which
is the approach taken in this commit.

Reported-by: Paolo Orsini <paolo.orsini@opengosim.com>

show more ...

CUDA: Fixed incorrect use of separate streams.

This solves synchronization problems that have arisen due to the incorrect use of multiple CUDA streams for vector and matrix operations (without using

CUDA: Fixed incorrect use of separate streams.

This solves synchronization problems that have arisen due to the incorrect use of multiple CUDA streams for vector and matrix operations (without using proper synchronization mechanisms).
In particular, SNES ex19 and ex56 now run reliably (no failure after 20+ reruns).
Instead, the default stream (NULL pointer) is now used for all CUDA operations.
I don't have performance comparisons at hand for the performance implications in this commit, but expect any changes to be small.
Correctness first :-)

show more ...

Merge branch 'balay/fix-PetscCUBLASGetHandle-export-symbol' into 'master'

cuda: remove __CUDACC__ usage - as code is inconstant with it - and...

See merge request petsc/petsc!2085

Merge branch 'master' into 'master'

Add --download-hpddm

See merge request petsc/petsc!1994

cuda: remove __CUDACC__ usage - as code is inconstant with it - and PETSC_HAVE_CUDA already protects its usage.

This fixes the issue where PetscCUBLASGetHandle was mangled incorrectly as __CUDACC__

cuda: remove __CUDACC__ usage - as code is inconstant with it - and PETSC_HAVE_CUDA already protects its usage.

This fixes the issue where PetscCUBLASGetHandle was mangled incorrectly as __CUDACC__ prevented petsccublas.h from getting include

Reported-by: Jose E. Roman <jroman@dsic.upv.es>
Suggested-by: Stefano Zampini <stefano.zampini@gmail.com>

show more ...

Fixes and improvements for PCHPDDM and KSPHPDDM

{Vec|Mat}PinToCPU: cannot call return if PetscFunctionBegin has been called

VecScatterCreate: use optimized VECSCATTERSEQ for seq transfers

VecScatter_Seq: both vectors need to be CUDA vectors

Vec{Seq|MPI}{CUDA|VIENNACL}: allocate host memory in default constructors

this prevents from initial memory transfers of zeroed memory

VecCUDA: support VecCopy(veccuda,vecstandard)

minor

CUDA: VecCUDAGetArray and friends now always part of the interface.

The availability of these functions no longer depends on CUDA compiled into PETSc.
This simplifies user code built on top: No more

CUDA: VecCUDAGetArray and friends now always part of the interface.

The availability of these functions no longer depends on CUDA compiled into PETSc.
This simplifies user code built on top: No more recompilations required.

show more ...

CUDA: Moved PetscCUBLASGetHandle() to separate header petsccublas.h

Previously users had to provide a path to cublas_v2.h when using
VecCUDAGetArray() & friends.
By moving PetscCUBLASGetHandle() t

CUDA: Moved PetscCUBLASGetHandle() to separate header petsccublas.h

Previously users had to provide a path to cublas_v2.h when using
VecCUDAGetArray() & friends.
By moving PetscCUBLASGetHandle() to a separate header, one only needs to provide include paths for cublas_v2.h if using PetscCUBLASGetHandle().

Reference:
https://bitbucket.org/petsc/petsc-pre-gitlab/commits/5ca5d0dde916a923823f0864fc8674697d2df730#comment-7766631

Thanks-to: Jed Brown <jed@jedbrown.org>

show more ...

Section: Put header in the "right place"

and fix compile warnings