Merged in dmeiser/petsc/fix-cusp-ex4_par (pull request #181)

A bug fix in `VecScatterBegin` for CUSP vectors

Merge branch 'maint'

with with -> with the the -> the

Make sure data on CPU is up to date.

If VecCUSPCopyFromGPUSome_Public cannot be called we have to call
VecCUSPCopyFromGPU. Otherwise the host data might be out of data or
unallocated, as is the case

Make sure data on CPU is up to date.

If VecCUSPCopyFromGPUSome_Public cannot be called we have to call
VecCUSPCopyFromGPU. Otherwise the host data might be out of data or
unallocated, as is the case in the new test introduced in this branch.

With this fix the new test runex4_cusp2 passes. There are some valgrind
errors due to uninitialized values. These are fixed by the additional
bug fixes in fix-vecscattercusp_stos and fix-cusp-dmda.

show more ...

Add a parallel cusp test based on ex4.c.

This test currently segfaults. Here is the valgrind stack trace:

```
==8014== Invalid read of size 8
==8014== at 0x57BCEA: Pack_1 (vpscat.c:497)
==8014=

Add a parallel cusp test based on ex4.c.

This test currently segfaults. Here is the valgrind stack trace:

```
==8014== Invalid read of size 8
==8014== at 0x57BCEA: Pack_1 (vpscat.c:497)
==8014== by 0x58A89D: VecScatterBegin_1 (vpscat.h:110)
==8014== by 0x572441: VecScatterBegin (vscat.c:1724)
==8014== by 0x407D33: main (ex4.c:35)
==8014== Address 0x0 is not stack'd, malloc'd or (recently) free'd
```

show more ...

Use VecCUSPGetArrayReadWrite to obtain scatter destination array.

VecCUSPGetArrayWrite is not sufficient because the scatter only
modifies part of the distination vector. VecCUSPGetArrayReadWrite
m

Use VecCUSPGetArrayReadWrite to obtain scatter destination array.

VecCUSPGetArrayWrite is not sufficient because the scatter only
modifies part of the distination vector. VecCUSPGetArrayReadWrite
makes sure that data on GPU has been synchronized with CPU.

show more ...

Remove tabs.

Add cusp scatter test.

This test fails.

Delete commented out code in cuda related files.

Note that there are still a fairly large number of commented out chunks
of code in the rest of the code base:

[dmeiser@ivy petsc]$ grep -R '#if\s0'

Delete commented out code in cuda related files.

Note that there are still a fairly large number of commented out chunks
of code in the rest of the code base:

[dmeiser@ivy petsc]$ grep -R '#if\s0' * |grep '^src\|include'| wc -l
87

show more ...

Fixing bug in VecCUSPAllocateCheckHost.

Bug fixes in host memory allocation/checking in VecGetLocalVector_SeqCUSP/VecRestoreLocalVector_SeqCUSP.

Add memory logging, remove debug statements.

Need to ensure that host vector is allocated.

Fix allocation of host memory in VecCUSPAllocateCheckHost.

Tweak valid_GPU_array logic.

Fix condition in VecValidValues.

Previous logic was broken in the case where vec-valid_GPU_array ==
PETSC_CUSP_UNALLOCATED.

Set getlocalvector and restorelocalvector ops entries in Vec_MPICUSP.

Need to check whether vector has a non-null data ptr before dereferencing it.

Bug fix in VecGetLocalVector and VecGetLocalVectorRead.

Adding implementations VecGetLocalVector_SeqCUSP and VecRestoreLocalVector_SeqCUSP.

Add methods for acquiring local vector views to _VecOps.

Conflicts:
include/petsc-private/vecimpl.h

Checking that the host pointer is valid should be sufficient.

Because the GPU array gets invalidated later in this function.

Ensure that host array is allocated.

This commit fixes a bug where a null pointer is returned when
((Vec_Seq*)x->data)->array == 0 and x->valid_GPU_array ==
PETSC_CUSP_UNALLOCATED. This situation a

Ensure that host array is allocated.

This commit fixes a bug where a null pointer is returned when
((Vec_Seq*)x->data)->array == 0 and x->valid_GPU_array ==
PETSC_CUSP_UNALLOCATED. This situation arises for vectors generated
with VecDuplicate_SeqCUSP.

show more ...

Replace cudaMemcpyAsync with cudaMemcpy.

Also page locked host memory registration.

Trying to fix the cudaHostRegister/cudaHostUnregister logic.