Runtime Flow

Inputs

The runtime side centers around two input files in the working directory:

  • param: runtime parameters and workflow switches
  • data: generated initial eigenspectrum, operators, and chain coefficients

param is parsed by Params in c++/params.hpp.

data is parsed by InputData in c++/read-input.hpp.

Entry Point

Execution starts in c++/nrg.cc:

  1. initialize MPI
  2. report OpenMP and environment information
  3. validate that param and data exist
  4. set up either a unique temporary workdir or an exact checkpoint directory
  5. run master or slave behavior depending on MPI rank

The executable-facing orchestration then moves into NRG_calculation in c++/nrg-general.hpp.

nrg.cc is intentionally thin. Nearly all interesting control flow begins once NRG_calculation is constructed.

Initialization In NRG_calculation

NRG_calculation constructs and wires together the core runtime objects:

  1. Params
  2. InputData
  3. Symmetry<S>
  4. Stats<S>
  5. ThermoStore<S>
  6. BackiterStore
  7. diagonalization engine selection

Important methods in NRG_calculation:

  • select_diag_engine(): choose MPI / OpenMP / serial diagonalization backend
  • run_phase(...): build per-phase helpers and execute the iteration loop
  • prepare_rho(...): initialize and backpropagate shell density matrices
  • prepare_rhoFDM(): initialize and backpropagate FDM density matrices
  • run_dm_phase(...): launch the second sweep when enabled

InputData performs these steps:

  1. read the data header
  2. instantiate the symmetry backend
  3. load the seed eigenspectrum and seed operators
  4. read operator blocks and coefficient tables
  5. finalize Nmax and related derived parameters

Checkpoint discovery runs after these steps because its iteration range depends on the finalized Nmax value. The discovered prefix is fully validated before phase output streams are opened.

See Input and configuration for the input-side details.

Main NRG Phase

The main run is started through run_phase(RUNTYPE::NRG, ...), which enters nrg_loop(...) in c++/core.hpp.

At a high level, each iteration does the following:

  1. build the next-shell SubspaceStructure
  2. convert that into a TaskList
  3. load a completed checkpoint or construct the Hamiltonian blocks
  4. diagonalize each uncached block through the selected DiagEngine
  5. establish energy references and truncation criteria, including the Floquet criterion when selected
  6. split eigenvectors into ancestor blocks when needed
  7. update iteration metadata and basic diagnostics
  8. for strategy=all, recalculate operators and measure before truncation
  9. truncate the eigenspectra
  10. archive truncated state into thermodynamic and backward-sweep stores
  11. recalculate irreducible operators for the next step
  12. for strategy=kept, recalculate operators and measure after truncation

The most important coordination logic lives in c++/core.hpp.

See Iteration engine for the function-level breakdown of this phase.

Optional DM-NRG And FDM Phases

After the first sweep, NRG_calculation may run density-matrix-related phases depending on Params:

  • prepare rho
  • prepare rhoFDM
  • run RUNTYPE::DMNRG

This is still driven through the same broad iteration machinery, but with different runtime mode and output semantics.

Outputs

The runtime produces several forms of output:

  • spectral and thermodynamic files in the working directory
  • stored eigenspectra and density matrices inside the temporary or checkpoint workdir
  • optional HDF5 output via Output and h5save helpers
  • logs and diagnostic dumps when enabled by parameters

The workdir is represented by Workdir in c++/workdir.hpp and is used for transient iteration-state files such as unitary matrices and density matrices.

Persistent result files are written in the calculation directory, not the workdir. See the output format reference for filenames, columns, units, and lifecycle behavior. See State and persistence for the in-memory storage model and serialization boundaries.

Mathematica Side

The C++ executable does not construct the initial problem description by itself. The Mathematica-side nrginit pipeline is responsible for:

  • model definition
  • basis construction
  • initial Hamiltonian diagonalization
  • seed operator generation
  • data file emission

That is why understanding the full project flow requires keeping both nrginit/ and c++/ in mind, even though the iterative heavy lifting happens in C++.

See nrginit workflow for the Mathematica-side entry path.