With this changes Octopus appears to run correctly on the Intel Xeon Phi.

Adding tool to look up the values reported by the buildslaves for a particular match. This can be helpful, for example, in choosing an appropriate tolerance.

since this is not a recommended value and in fact the code emits a warning if
it finds it.

* Bugfix: condition for when to read TDProjStateStart was accidentally reversed in r3597 (6 yrs ago!).
* Improvement of documentation regarding TDOutput = td_occup, populations.

* Comments on how to implement OutputMatrixElements for parallel in states.

For TDOutput = td_occup or populations, we copy the whole states object, then deallocate a lot of it, then refill it. It would be more efficient to just refrain from copying the parts we don't need. In particular, before we needed memory available for all the states again, whereas now, if TDProjStart > 1, we will only need memory for the number of states we will actually use.

* Adding a warning if KPointsGrid is not set, when writing BerkeleyGW.
* Allowing KPointsGrid to be set even if KPoints or KPointsReduced is used to specify the k-points. This has no effect (I hope!) except that BerkeleyGW output needs the values from there.

* Removing debugging output left in r10909.

Fix to r10873 (job parallelism in testsuite): the check for the status of a test via the return code of its pid could fail, because it was possible for a new process with the same pid to have started before the check occurred. This happened on the chum_x86_64_pgi_test slave, producing error message:

./oct-run_testsuite.sh: line 123: wait: pid 23407 is not a child of this shell

The new approach is to capture the return code into a variable named after the test, which will then be read when the job has finished.

Correction to generation of Fourier shell: in r8674 I added a "convention" as to whether to include -N/2 or N/2 in the shell. Well, to follow how plane-wave codes do it, we should actually include neither of them.

ifort and other compilers objected to previous interface form for zdgemm. Now, unified into blas_gemm, which seems satisfactory.

Creating C wrappers for calling dgemm with complex values from hamiltonian_base, so we can actually use blas_m and have interfaces. This is along the same lines as the existing "dablas" wrappers for scaling a complex vector by a real scalar.

sgfft specifies the maximal size for argument 'trig' which means uninitialized or out-of-bounds data might be copied when the size is actually smaller. Since this is a module, there is no need to specify the size anyway.

* Using float constant for float arithmetic.

* Using optional_default.
* Adding extra checks in states_orthogonalize_single.

* Fixing intent(out) :: y to intent(inout) in bpdn interface for dgemv, as we correctly have in the main code.
* Doxygen comments.
* Replacing non-standard extension 'dfloat' with equivalent 'dble' in newuoa.

* Fixing format statement to avoid ****** output.
* Unused variable.
* More precision in output of mesh cutoff.

Bugfix: OutputMatrixElements = momentum or angular_momentum suffered from a wide array of problems, too many to bother listing. I hope no one was using them, because any results were surely wrong. Now they work, and are tested.

* Blocking some OutputMatrixElements calculations that are incompatible with states parallelization.

'input_error' should be implemented via 'messages_fatal', for consistency, and to take advantage of messages_fatal's writing of the stack trace and new provision for making sure message gets written before abort.

When messages_fatal is called with only_root_writes = .true., it is possible for a non-root process to call MPI_Abort before root has written the error message, so it is never printed. This seems to happen particularly when running directly on the command line. To avoid this, the other processes will wait, and then write it themselves if root does not. Also this will avoid missing messages if only_root_writes is used mistakenly for an error case that does not happen on root.

Notes and error message regarding pseudopotentials only being usable in 3D. Also, adding an assertion that will be triggered if double grid is called other than in 3D, which is assumed extensively in those routines.

The derivatives tests do not use the Coordinates and Species, but something must be present anyway for the initialization to proceed without an error. However, it would be better to provide something that at least would be acceptable for an ordinary run. Thus, I am putting a dummy user-defined species instead of the Na pseudopotential for the 1D and 2D runs.