Because it is a linker flag and clang prints a warning otherwise.

In most cases it just makes the error messages longer without helping
much (you would need to use clang if you actually want more useful error
messages). And nvcc prints some useless warning because of it...

It is not needed, since CUSPARSE is included in all versions of CUDA
since at least 7.0.

Also note that CUSPARSE is needed only for some tests and legacy code
(tnl-benchmark-spmv and wrappers in Matrices::CSR) and experimental code
(Preconditioners::ILU0). Everything is still guarded by the
HAVE_CUSPARSE macro, so the user can pass -DHAVE_CUSPARSE -lcusparse to
the compiler to enable the features.

This way they can use either yes/no, on/off, or true/false.

It is not used in current code, it was used only in some legacy benchmarks.

And of course our implementation is faster ;-)

There is shor test of this stuff for 1D 2D and 3D.

Relpaced tabs with 4spaces.

Standard CMAKE_CXX_FLAGS are a space-delimited list, but our
CXX_TESTS_FLAGS are manually passed to CUDA_ADD_EXECUTABLE, which
expects a semicolon-delimited list. Blame cmake for such a crazy
language...