~siesta-maint/siesta/rel-4.0 : changes

~siesta-maint/siesta/rel-4.0 » Changes from revision 596

From Revision 577 to 558

Rev	Summary	Authors	Date
577	Fixed writing out too many wavefunctions Fixed writing out too many wavefunctions In writewave users could specify wavefunction indices beyond the number of eigenstates. This has now been fixed. Any input from the user will be truncated to the maximum number of eigenstates.	Nick Papior	6 years ago
576	Diagnostics and behavior clarifications related to H and S f... Diagnostics and behavior clarifications related to H and S folding * Minimal fix for DM initialization with severe folding When an orbital interacts with its own periodic replicas and a Gamma-point calculation is requested, the naive DM initialization with atomic information is faulty: the diagonal S entries are no longer 1.0 (but are enhanced by the contributions from the replicas), so the normalization of the DM is lost (the code will print a message to the effect). The initial charge density will be distorted with respect to its intended form (the result of a blanket normalization). Note that the program might still go on to converge to the right ground state, since the code to deal with multiple orbital interactions in the Gamma-point case is mathematically correct. There are several levels of fixes for this: 1. Use k-point sampling, as it can be argued that requesting a Gamma-point calculation for a very small cell is physically meaningless. 2. Force the use of an auxiliary supercell ('force-aux-cell T'). Internally, the code will correctly place all the different interactions in an expanded auxiliary supercell. The drawback is that complex arithmetic and diagonalization must then be used (there is no yet support in the code to avoid this). 3. Correct the initialization of the DM so that its norm is conserved and the initial charge density is more reasonable. In this patch we implement case 3. above by (always) dividing the recently-initialized diagonal DM entries by the corresponding S entry. Note that this is a zeroth-order correction, and the charge density produced by the initial DM will exhibit a deficit at the orbital position and 'ringing' enhancements at the superpositions of the orbital and its periodic images. The initial scf steps will thus show differences with a calculation done with 'force-aux-cell T', but eventually they will converge to the same ground state. Updated the manual with an explanation of diagonal-folding and its consequences. * Streamline hsparse. Kill the program if auxiliary supercell is inadequate The 'multiple-image' warnings in hsparse were not representative enough under MPI operation (only the root node got to report them). And it can be argued that they are superfluous except in a 'debug' mode of operation. The key severe diagnostic is the existence of 'folding' in a non-gamma calculation, which should lead to aborting the program and to further investigation of the internals of the code. In the gamma case, folding is reported and a further note is printed if the user is attempting to carry out a COOP/COHP analysis (which would be faulty in the presence of folding). The non-documented 'debug-folding' feature is implemented very coarsely for now. There is also a check for diagonal folding. Harris-functional calculations are stopped if found. A number of tests used the combination of very small cells (leading to 'diagonal folding') and Harris functional. The latter option has been removed. Use 'message' interface for folding warnings.	Alberto Garcia	6 years ago
575	Merged lp:~nickpapior/siesta/4.0-io-precision Merged lp:~nickpapior/siesta/4.0-io-precision This extends the precision for the EIG, KP and PDOS output.	Nick Papior	6 years ago
574	Catch charge mismatchs in occupied states Catch charge mismatchs in occupied states If the basis sets includes an orbital for a low-lying shell which was not included in the pseudopotential generation, it could happen that it is considered occupied, thus leading to a charge inconsistency. A check has bee introduced in the code to detect charge inconsistencies.	Alberto Garcia	6 years ago
573	Forgot to update Tests/Makefile with h2o NC calculation Forgot to update Tests/Makefile with h2o NC calculation	Nick Papior	6 years ago
572	Fixed bands output for NC calculations Fixed bands output for NC calculations For NC calculations the spin-components are mixed and thus there are no separate eigenvalus. This fixes the output. gnubands had a bug when explicitly specifying a non-existing spin-index. Added a bands test for NC.	Nick Papior	6 years ago
571	Fixed tests Fixed tests Now ag, batio and force_2 graphite_c6, graphite_vdw_df and pd_ice has been fixed. They run without problems and does converge. Fixed the Makefile to add missing tests. Fixed gnubands code. The emin/emax reads from the .bands file. The problem is that the emin/emax in the .bands file has a much higher precision than the eigenvalues in the file. Thus one could find values not in the correct range due to rounding errors. This has been fixed by setting emin/emax to -1e30, 1e30.	Nick Papior	6 years ago
570	Add missing 'dictRef' attributes to some CML properties Add missing 'dictRef' attributes to some CML properties The missing attributes caused problems with the AiiDA parser. As a matter of policy, all Siesta CML properties are required to have a 'dictRef' attribute. This is now enforced by the compiler after making the relevant argument non-optional in our custom version of FoX.	Alberto Garcia	6 years ago
569	Mark Harris-functional calculations as 'converged' Mark Harris-functional calculations as 'converged' A Harris-functional calculation should be considered 'converged' after a single iteration. Doing this explicitly avoids triggering the standard warnings.	Alberto Garcia	6 years ago
568	Update gnubands in Util/Bands Update gnubands in Util/Bands Rename new.gnubands.f90 to gnubands.f90 and update. Rename the gnubands.f to old.gnubands.f	Alberto Garcia	6 years ago
567	Enlarge number fields in _NOT_CONV messages. Document spin-... Enlarge number fields in _NOT_CONV messages. Document spin-in-scf	Alberto Garcia	6 years ago
566	Fix the computation of mulliken populations for non-collinea... Fix the computation of mulliken populations for non-collinear spin The populations were split in two contributions but one of them was not properly added in all cases.	Alberto Garcia	6 years ago
565	Fix the Siesta response to an i-PI 'EXIT' message Fix the Siesta response to an i-PI 'EXIT' message Siesta will now properly exit (by calling 'bye') when receiving an 'EXIT' message from i-PI.	Nick Papior	6 years ago
564	Remove time-reversal-symmetrization for non-collinear spin c... Remove time-reversal-symmetrization for non-collinear spin calculations While in principle TR symmetry (TRS) should be preserved in non-collinear-spin calculations, care should be taken that some subtle detail of the implementation does not break it. The appearance of spin-orbit-capable versions (>4.1) has tilted the balance towards removing the TRS symmetrization (which is not possible with spin-orbit coupling), forgoing the savings in the number of k-points in exchange for uniformity of the analysis tools. ( + Use of the same convention for phase factors and the same conjugation pattern in the building of the DM for the non-collinear case as in the spin-orbit versions.)	Alberto Garcia	6 years ago
563	Fix mulliken populations for non-collinear spin in parallel Fix mulliken populations for non-collinear spin in parallel For non-collinear spins, the mulliken routine was not parallelized correctly.	Alberto Garcia	6 years ago
562	Spin monitoring in scf cycle. Clarification of sign conventi... Spin monitoring in scf cycle. Clarification of sign convention for DM * Spin monitoring during scf cycle If the fdf variable 'Spin.In.Scf' is set to 'true', the size and components of the (total) spin polarization will be printed at every scf step. This is analogous to the 'Mulliken.In.Scf' feature. The default for the Spin.In.SCF fdf flag is 'true' for calculations involving spin. A single line of spin information is printed, prefixed by ' spin-moment:'. Note that now, for each scf step, the spin and/or mulliken analyses are printed after the line showing the energies. The initial spin polarization for every geometry iteration is also printed in 'state_init', replacing the similar calls in 'new_dm'. * Use explicitly the same DM_12 sign-convention as in SOC versions In the non-collinear case, the code is actually using internally the same sign convention for the "up-down" components of H and the DM as the spin-orbit-capable (SOC) versions (>= 4.1): The building of the dense Hamiltonian has the same form and the off-diagonal components of Vxc are identical. Routines diag2g and diag2k are actually using the same convention, but with DM* as an intermediate step. By changing the interface to the outside world, namely the correspondence between spin angles and the DM in new_dm, and the calculation of the spin (actually moment) components in spnvec, the code can now interoperate (i.e., reuse DMs) with the SOC versions. In order to reuse DM information produced by previous versions of the program in the non-collinear case, the sign of D(4) of the old DM has to be changed. This can be achieved with the dm_noncol_sign_flip4 program in Util/DensityMatrix.	Alberto Garcia	6 years ago
561	Make generation of timer report more robust. More informatio... Make generation of timer report more robust. More information in CML * Generate timer report in non-convergence and out-of-time conditions For proper accounting, the timer information should be generated even if the program stops due to non-convergence (scf, geometry) or out-of-time conditions. This is achieved by inserting appropriate calls to 'timer' before the call to 'die'. To avoid a race condition in which a non-root node might get to the 'die' call before the IO node has had time to properly write the CML, timing, and even output information, a call to 'barrier' has been inserted before the 'die' call. It might be argued that this kind of bookeeping should be left to the 'die' call itself. But it is only possible if all nodes call 'die'. * Add no_u and nnz info to CML file to ease the extracting of benchmarking information.	Alberto Garcia	6 years ago
560	Enable compatibility switch for NRTAB value Enable compatibility switch for NRTAB value Internally the two-center integrals involved in some matrix element calculations are tabulated with a preset number of elements (NRTAB). In versions 4.0.1 and prior this was 128. Since 4.0.2 the number of table elements has been increased to 1024, which translates to more accurate matrix element calculations. If preservation of the (lower accuracy) numerical results of 4.0.1 or prior versions is required for reproducibility purposes, the user can include Compat.Matel.NRTAB T in the fdf file.	Alberto Garcia	6 years ago
559	Removed unnecessary call in overlap.f to volume Removed unnecessary call in overlap.f to volume	Nick Papior	6 years ago
558	Write JSON file for tree-timer information Write JSON file for tree-timer information A more flexible output for timing information. The 'time.json' file is produced if the option use-tree-timer T is in effect.	Alberto Garcia	6 years ago

Older »