CSR: better strategy defaults.

[tcojean]

Provide better CSR strategy defaults.

This duplicates two constructors to select a strategy based on the
executor type instead of defaulting to automatical(CUDA). This should
not break any existing behavior (the CUDA case still defaults to
automatical), but instead, fix the currently broken combinations.

Background:
Some interesting situation just happened. Due to a seemingly broken
CUDA installation on a cluster, the CUDA API call in get_num_devices()
can fail, for example, due to an insufficient driver version or something
else:

terminate called after throwing an instance of 'gko::CudaError'
  what():  ../cuda/base/executor.cpp:223: get_num_devices: cudaErrorInsufficientDriver: CUDA driver version is insufficient for CUDA runtime version

Since the current CSR strategy is to instantiate a CudaExecutor, this
means that in that case all of Ginkgo is broken at runtime, even when
not using an actual CudaExecutor. The aim of this PR is to change this
long-time problem with the default strategy type.

[tcojean]

Looks like this wouldn't be an issue with a real develop codebase, which has sparselib as a strategy everywhere. Due to relying on many commits, somehow automatical was reintroduced as a strategy in the OpenCARP branch.

This can still be a suggestion for dynamically adapting the strategy to executor type instead of only using sparselib everywhere.

[upsj]

Yes, this kind of thing has commonly lead to issues in Csr automatic tests if you forgot to pass the HIP parameter and relied on the default instead.

[upsj]

LGTM! Only minor suggestions

[tcojean]

Pinging @yhmtsai I see that this was reintroduced (on my side) in the improve_multigrid branch. It makes sense to default to automatical in the correct settings for better performance. Does this way of doing it look good? Any different default settings?

[pratikvn]

LGTM!

[yhmtsai]

default setting as automatic is okay to me.
do you think we should change the load_balance/automatic default constructor to throw error not use cuda executor?

[codecov]

Codecov Report

Merging #969 (b57f3ac) into develop (3a2c843) will increase coverage by 0.01%.
The diff coverage is 71.42%.

❗ Current head b57f3ac differs from pull request most recent head 68b41ef. Consider uploading reports for the commit 68b41ef to get more accurate results

@@             Coverage Diff             @@
##           develop     #969      +/-   ##
===========================================
+ Coverage    92.31%   92.33%   +0.01%     
===========================================
  Files          476      476              
  Lines        39390    39304      -86     
===========================================
- Hits         36364    36291      -73     
+ Misses        3026     3013      -13     

Impacted Files	Coverage Δ
include/ginkgo/core/matrix/csr.hpp	45.06% <71.42%> (-1.00%)	⬇️
common/unified/matrix/coo_kernels.cpp	36.36% <0.00%> (-63.64%)	⬇️
devices/machine_topology.cpp	77.77% <0.00%> (-4.94%)	⬇️
common/unified/matrix/diagonal_kernels.cpp	65.00% <0.00%> (-3.43%)	⬇️
common/unified/matrix/dense_kernels.cpp	94.31% <0.00%> (-1.09%)	⬇️
common/unified/matrix/ell_kernels.cpp	27.90% <0.00%> (-0.67%)	⬇️
core/matrix/dense.cpp	93.98% <0.00%> (-0.08%)	⬇️
core/matrix/hybrid.cpp	99.25% <0.00%> (-0.02%)	⬇️
reference/matrix/fbcsr_kernels.cpp	97.16% <0.00%> (-0.02%)	⬇️
... and 44 more

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 3a2c843...68b41ef. Read the comment docs.

[ginkgo-bot]

Note: This PR changes the Ginkgo ABI:

Functions changes summary: 128 Removed, 16 Changed (328 filtered out), 32 Added functions
Variables changes summary: 0 Removed, 0 Changed, 0 Added variable

For details check the full ABI diff under Artifacts here

[sonarcloud]

Kudos, SonarCloud Quality Gate passed!   

0 Bugs
0 Vulnerabilities
0 Security Hotspots
6 Code Smells

70.0% Coverage
0.0% Duplication