Changelog
Berkeley 2.1 (Jul/2019)
BerkeleyGW 2.1 is the first version based on a new and more modern coding infrastructure, with better support for new compilers and improved consistency checks. For the end users, the main noticeable features include increased performance and bug fixes for I/O operations involving the new HDF5 file format and for the subspace code. We also improved considerably the documentation of the code, with a new and expanded user manual.
Some highlights and features for the end user:

New user manual for the code, which comes bundled with the code, and which is also available online: http://manual.berkeleygw.org

New wrapper for the StochasticGW code.

Bug fixes when writing HDF5 wavefunctions in parallel which could cause the code to hang (relevant to the ParaBands code).

Improved error checking for operations involving HDF5 files.

Improved support for compilers, including PGI and NAG.

Improved performance and stability for building BerkeleyGW in parallel with new dependency system.

Improved performance of HDF5 routines for the subspace code and for reading
chimat.h5
files.
Berkeley 2.0 (May/2018)
BerkeleyGW 2.0 represents the culmination of nearly two years of development effort, and this release contains a number of important new features and capabilities including:
Main new features:

Initial release of ParaBands: a new tool for efficiently generating wavefunction files including many empty orbitals required for BerkeleyGW calculations.

Full BSE calculations that do not employ the TammDancoff approximation.

Improved algorithms for kpoint sampling in 2D, which include the newly proposed nonuniform neck subsampling (NNS) and the cluster sampling interpolation (CSI) algorithms.

Accelerated fullfrequency GW calculations through the use of a lowrank subspace approximation for expressing the dielectric matrix. In fact, largescale fullfrequency GW calculations are now faster than calculations using plasmonpole models!

Significant performance improvements throughout, but particularly in the calculation of the fullfrequency dielectric matrix and evaluation of the fullfrequency Sigma operator. Continued optimizations were made throughout the package for multi and manycore architectures including Intel XeonPhi, which allows BerkeleyGW to scale half a million cores on Cori 2 for largescale calculations!

Improved user and developer documentation, as well as a new quick reference guide (see the link on the top of the page).
Berkeley 1.2 (Aug/2016)
F. H. da Jornada, J. Deslippe, D. VigilFowler, J. I. Mustafa, T. Rangel, F. Bruneval, F. Liu, D. Y. Qiu, D. A. Strubbe, G. Samsonidze, J. Lischner.
Features marked with [*] change the default behavior of the code relative to BGW1.1 and may cause a small change in the numbers produced by the code.
New features

Added new methods to deal with the frequencydependence of the polarizability:

Added GodbyNeeds (GN) plasmonpole model.
 Added spectral method for realaxis (RA) fullfrequency (FF) calculations. The spectral method allows one to compute the dielectric matrix much faster for many frequency points on the real axis. The old method for RAFF calculations is now refered to as the AdlerWiser method. Still, RA is no longer the default FF method (see next item).

Added Contourdeformation (CD) fullfrequency (FF) formalism. The CD is the recommended (and default) scheme for calculations performing an explicit evaluation of the dynamical effects of the dielectric matrix without plasmonpole models. It requires the evaluation of the dielectric matrix on both real and imaginary frequencies, but with typically a much smaller number of frequencies. The previous formalism is now refered to as realaxis (RA) fullfrequency (FF) formalism. CD is now the default method for FF calculations.

Released nonlinear optics postprocessing utility: This codes allows one to compute the interexciton transitions in nonlinearoptics experiments.

Improved usability in the code output and added dynamic verbosity switch: BerkeleyGW has now a much simpler and neater output. The code gives time estimates for most tasks it performs  at least for those that are more timeconsuming. There is also a runtime flag supported by all codes to switch the amount of verbosity the code produces. The old compilation flag
DVERBOSE
is now deprecated. 
Improved usability in the code input: Several input parameters, such as band occupations, are automatically detected. If not set, parameters such as the cutoff for the bare and screened Coulomb potential are also read from the wave function and dielectric matrix, respectively.

Automatic solution of Dyson's equation in FF calculations: BerkeleyGW now automatically solves Dyson's equation for FF calculations. It uses a varying number of frequencies to find the graphical solution for the quasiparticle energies, and perform extrapolation (with appropriate warning messages) if no intersection could be found. The code will also output the files
eqp0.dat
andeqp1.dat
directly, containing the offshell and onshell quasiparticle energies. 
Added ABINIT wrapper: It is possible to interface the ABINIT code with BerkeleyGW. For now, the wrapper can only output complexvalued wavefunctions, even for systems with inversion symmetry.
Improvements

Added support for XeonPhi Knight's Landing (KNL)based systems.

Improved OpenMP scaling throughout. Using OpenMP is now recommended for largescale computations.

Added new and more performant HDF5 file formats for
epsmat
andbsemat
matrices. This is recommend as default for all builds. Note that files are generally not compatible with BerkeleyGW 1.1 and earlier. Binaryepsmat
files can be converted to the HDF5 format with theepsmat_old2hdf5
utility. Olderepsmat.h5
files from BerkeleyGW1.1beta2 can be converted to the current format with theepsmat_hdf5_upgrade
utility. 
Added parallelization over frequencies in the epsilon code. This allows for better scalability in fullfrequency calculations. This scheme is supported in the old AdlerWiser and in the new ContourDeformation formalisms.

Improved the MonteCarlo average scheme used to compute the Coulomb potential. [*] For bulk systems, the new default scheme is to compute the average of Coulomb potential v(q+G) for all qpoints and Gvectors. We use a hybrid scheme to make this evaluation faster. While the code is slightly slower for small calculations, results should converge much faster with kpoint sampling. One can still use the old defaults with the input option
cell_average_cutoff 1.0d12
. Note that no change was performed for 2D, 1D and 0D systems, or 3D metals. 
Rewrote kpoint interpolation engine. [*] The previous kpoint interpolation scheme used in the
haydock
,absorption
, andinteqp
codes used a greedy algorithm to search for the closest coarsegrid points around each finegrid points, which lead to discontinuities of the interpolands. The new interpolation engine uses QHull, which is bundled with BerkeleyGW, to perform a Delaunay tessellation of the coarsegrid kpoints. The new interpolation scheme is much more robust. You can use the old interpolation scheme with thegreedy_interpolation
flag. 
Generalized scheme for kpoint interpolation. The kpoint interpolation is used in the haydock, absorption, and inteqp codes. The new scheme is useful for metals and systems with valenceconduction band character mixing in the BSE.
Misc

Changed logic to deal with invalid GPP frequency modes. [*] When the code performs a HLGPP or a GNGPP calculation and finds an invalid mode with frequency with \omega_{G,G'}^2 < 0, the code will now treat that mode within the static COHSEX approximation (i.e., move frequency to infinity). The previous behavior was to "find" a purely complex mode frequency and relax the causality constraint. One can switch which strategy to use with the
invalid_gpp_mode
input flag. 
Many bug fixes and performance improvement.