Skip to content

developer_guide.md

Latest commit

 

History

History
273 lines (195 loc) · 11.9 KB

developer_guide.md

File metadata and controls

273 lines (195 loc) · 11.9 KB

KRONOS Developer Guide

1. Directory Layout and Module Organization

kronos/
├── CMakeLists.txt              # Root: project options, find_package, subdirs
├── src/
│   ├── CMakeLists.txt          # Builds kronos_lib (static) + kronos executable
│   ├── main.cpp                # Entry point: parse YAML, run SCF, write JSON
│   ├── core/                   # types.hpp, constants.hpp, crystal, element_data, spherical_harmonics
│   ├── basis/                  # plane_wave (G-vector enumeration), fft_grid (FFTW3), kpoints (MP grid)
│   ├── io/                     # input_parser (YAML), upf_parser (UPF), output_writer (JSON/HDF5)
│   ├── potential/              # hartree, xc, local_pp, nonlocal_pp, ewald, gradient, forces
│   ├── solver/                 # scf, davidson, mixing (Pulay/DIIS), fermi, bfgs
│   ├── hamiltonian/            # H|psi> application: kinetic + local (FFT) + nonlocal (GEMM)
│   ├── postprocessing/         # band_structure, dos
│   ├── gpu/                    # fft.hpp, blas.hpp, memory.hpp, gpu_stubs.cpp
│   └── utils/                  # timer (KRONOS_TIMER macro), logger (structured JSON)
├── test/                       # GoogleTest suite (15 executables)
│   ├── CMakeLists.txt          # FetchContent(googletest v1.14), kronos_add_test()
│   ├── test_helpers.hpp        # Shared crystals, pseudopotentials, tolerances
│   └── test_*.cpp              # One file per module
├── examples/                   # YAML input files
├── pseudopotentials/           # UPF files
└── docs/                       # Documentation

Each src/ subdirectory contains .hpp/.cpp pairs. Headers use module-relative includes: #include "core/types.hpp". Adding a .cpp to any src/ subdirectory is picked up automatically by GLOB_RECURSE in src/CMakeLists.txt (re-run cmake after adding files).

The dependency flow is strictly layered: core <- basis <- potential <- hamiltonian <- solver <- io/main. The gpu/ namespace is the sole bridge to vendor APIs. utils/ is used everywhere.

2. How to Add a New XC Functional

Step 1. In src/potential/xc.cpp, locate XCEvaluator::init_functional() and add a branch:

} else if (name_ == "BLYP") {
    is_gga_ = true;
#ifdef KRONOS_HAS_LIBXC
    x_func_id_ = XC_GGA_X_B88;
    c_func_id_ = XC_GGA_C_LYP;
#endif
}

Step 2. Set is_gga_ = true for GGA functionals. The SCF loop checks is_gga() to compute density gradients via GGAGradient and call evaluate_gga() instead of evaluate().

Step 3. Register the name as an allowed value in src/io/input_parser.cpp.

Step 4. (Optional) For libxc-free operation, add analytic implementations alongside builtin::lda_x and builtin::lda_c_pz in xc.cpp, routing to them in the #else path.

Step 5. Add tests:

TEST(XCFunctional, BLYPSmokeTest) {
    kronos::XCEvaluator xc("BLYP");
    EXPECT_TRUE(xc.is_gga());
    RVec rho(100, 0.01), sigma(100, 1e-4);
    auto result = xc.evaluate_gga(rho, sigma, 100.0);
    EXPECT_LT(result.energy, 0.0);
}

Currently supported: LDA_PZ, LDA_PW, PBE, PBEsol.

3. How to Add a New Potential Type

Step 1. Create src/potential/your_potential.hpp and .cpp:

#pragma once
#include "core/types.hpp"
#include "core/crystal.hpp"
#include "basis/plane_wave.hpp"
namespace kronos {
class YourPotential {
public:
    YourPotential(const Crystal& crystal, const PlaneWaveBasis& basis);
    [[nodiscard]] CVec compute(const CVec& density_g) const;   // V(G)
    [[nodiscard]] double energy(const CVec& density_g) const;   // Ry
private:
    const Crystal& crystal_;
    const PlaneWaveBasis& basis_;
};
} // namespace kronos

Step 2. Integrate into src/solver/scf.cpp:

  • Construct alongside other potentials at the start of solve().
  • Call compute() in the SCF iteration and add to V_eff.
  • Call energy() and include in the total energy decomposition.
  • Add a field to SCFResult in scf.hpp if separately reported.

Step 3. Re-run cmake (auto-globbed). Write tests, register with kronos_add_test().

If the potential contributes to forces, add derivatives in src/potential/forces.cpp following the compute_local_forces() / compute_nonlocal_forces() pattern.

4. Test Conventions and How to Add Tests

Shared Helpers (test/test_helpers.hpp)

kronos::test provides factories and tolerance constants:

Helper Description
make_si_diamond_crystal() Si diamond FCC, a=5.43 A, 2 atoms
make_nacl_crystal() NaCl rocksalt, 8 atoms
make_cscl_crystal() CsCl primitive, 2 atoms
make_si_pp_map() Local-only Gaussian Si pseudopotential
make_si_pp_map_nonlocal() Si PP with p-type KB projector, D=-2.0 Ry
make_si_diamond_displaced(delta, atom, dir) Displaced crystal for force finite-difference tests

Tolerances: ENERGY_TOL (1e-6), FORCE_TOL (1e-4), FFT_TOL (1e-10), TIGHT_TOL (1e-12), LOOSE_TOL (1e-3).

Test Categories

  • Regression (test_regression.cpp): Frozen baseline energies as constexpr double. SetUpTestSuite() runs SCF once and shares the SCFResult.
  • Physics (test_physics.cpp): Physical invariants -- Hermiticity, force sum rules, sign of energy components.
  • Convergence (test_convergence.cpp): Energy vs cutoff monotonicity, k-grid convergence, SCF iteration limits.
  • Validation (test_validation.cpp): Comparison against Quantum ESPRESSO references in test/reference/.

Adding a New Test

  1. Create test/test_yourmodule.cpp.
  2. Add kronos_add_test(test_yourmodule) to test/CMakeLists.txt.
  3. Build and run: cmake --build build -j$(nproc) && ./build/test_yourmodule
  4. Use GTEST_SKIP() if SCF convergence fails non-deterministically.
  5. Keep cutoffs low (8-15 Ry) and use Gamma-only for speed.
  6. Full suite: cd build && ctest --output-on-failure -E test_utils

5. Build System Details

CMake Options

Option Default Description
KRONOS_GPU_BACKEND none GPU backend: none, cuda, hip
KRONOS_BUILD_TESTS ON Build GoogleTest suite
KRONOS_BUILD_PYTHON OFF Build pybind11 Python bindings

Dependencies

Required: FFTW3, BLAS, LAPACK, yaml-cpp.

Optional (with compile definitions): HDF5 (KRONOS_HAS_HDF5), MPI (KRONOS_HAS_MPI), libxc (KRONOS_HAS_LIBXC), spglib (KRONOS_HAS_SPGLIB).

GPU: CUDA toolkit defines KRONOS_GPU_CUDA; ROCm defines KRONOS_GPU_HIP.

Build Commands

cmake -B build -S .                              # configure (CPU-only)
cmake --build build -j$(nproc)                   # compile
cd build && ctest --output-on-failure            # all tests
./build/src/kronos examples/si_bulk.yaml         # run (note: build/src/kronos)
./build/test_basis --gtest_filter='*BasisSize'   # single test

6. Coding Style and Conventions

C++20 Features Used

  • std::numbers::pi, std::numbers::sqrt2 from <numbers>
  • Structured bindings: auto [crystal, input] = parse_input(file);
  • constexpr for all physical/math constants
  • [[nodiscard]] on query methods and pure functions
  • std::map::contains() (C++20)
  • No compiler extensions (CMAKE_CXX_EXTENSIONS OFF)

Precision

double and std::complex<double> only for physics. The aliases in core/types.hpp enforce this: real_t = double, complex_t = std::complex<double>, CVec = std::vector<complex_t>, RVec = std::vector<double>.

Naming

  • Classes: PascalCase -- PlaneWaveBasis, SCFSolver, XCEvaluator
  • Functions: snake_case -- compute_forces(), num_pw(), is_gga()
  • Constants: snake_case in kronos::constants -- constants::pi, constants::rydberg_to_ev
  • Members: trailing underscore -- crystal_, is_gga_, name_
  • Enums: PascalCase values -- CalculationType::SCF, SmearingType::Gaussian
  • Macros: KRONOS_UPPER_CASE -- KRONOS_TIMER, KRONOS_HAS_LIBXC

Unit System (Rydberg Atomic Units)

  • Kinetic: T = |k+G|^2 (not /2 as in Hartree)
  • Hartree: V_H(G) = 8*pi*n(G)/|G|^2 (not 4*pi)
  • Coulomb tail: V_loc(r) -> -2Z/r (factor of 2 vs Hartree)
  • Lattice: angstrom on input, bohr internally

Profiling

Wrap expensive functions with KRONOS_TIMER("name"). The macro creates a ScopedTimer that records wall time in the global TimerRegistry singleton. Results appear in SCFResult::timing and JSON output.

7. Debugging Tips

Structured JSON Logs

KRONOS emits structured JSON lines to stderr via Logger::instance(). Each line has timestamp, level, event, message, and custom fields. Capture with: ./build/src/kronos input.yaml 2>kronos.log

Logger::instance().info("scf_step", "SCF iteration complete",
    {{"step", std::to_string(step)}, {"energy_ry", std::to_string(energy)}});

SCF Diagnostics

Each SCF step prints: step number, total energy, |dE|, density residual |dn|, wall time.

Symptom Likely Cause Fix
Energy oscillates Mixing too aggressive Reduce PulayMixer alpha (default 0.3)
Energy increases Wrong sign in potential Check V_loc(G=0), Hartree prefactor
Energy jump > 1 Ry Numerical overflow or bug Auto-aborts; check logs
Density residual stalls Poor initial density Check atomic density superposition
Davidson diverges Near-degenerate states Auto-switches to LOBPCG

Energy Decomposition

SCFResult provides: kinetic_energy, hartree_energy, xc_energy, local_pp_energy, nonlocal_pp_energy, ewald_energy, smearing_energy. Compare individual components against QE to isolate discrepancies. Ewald energy is the easiest to validate (no SCF dependence) -- if it matches QE to 5+ figures, the crystal setup is correct.

Force Debugging

Validate forces via finite differences using make_si_diamond_displaced(delta, atom_idx, dir):

double F_numerical = -(E_plus - E_minus) / (2.0 * delta_bohr);
EXPECT_NEAR(F_numerical, F_analytic, FORCE_TOL);

Forces decompose into Ewald, local PP, and nonlocal PP components, all stored separately in SCFResult.

Common Pitfalls

  1. Rydberg vs Hartree. Factor-of-2 differences everywhere: kinetic |k+G|^2 not /2, Hartree 8*pi not 4*pi, Coulomb -2Z/r not -Z/r.
  2. UPF conventions. UPF stores r*beta(r) for projectors and 4*pi*r^2*rho(r) for atomic density. Divide by appropriate powers of r when loading.
  3. G=0 terms. Hartree sets V_H(G=0) = 0. Local PP has a finite V_loc(G=0). Wrong sign shifts total energy by a constant.
  4. K-point formula. KRONOS uses k = (2n-N-1)/(2N) + s/(2N). For even N with shift=0, the grid is off-Gamma.
  5. FFT grid. ecutrho >= 4 * ecutwfc for norm-conserving PPs. Grid dimensions come from ecutrho.

8. GPU Development

The kronos::gpu namespace is the sole vendor API boundary:

  • gpu/fft.hpp -- cuFFT/rocFFT wrapper
  • gpu/blas.hpp -- cuBLAS/rocBLAS wrapper
  • gpu/memory.hpp -- device memory management
  • gpu_stubs.cpp -- throws GPUNotAvailableError in CPU-only builds

Adding a kernel: (1) declare in src/gpu/*.hpp, (2) add stub in gpu_stubs.cpp, (3) implement in .cu (CUDA) or .cpp with #ifdef KRONOS_GPU_HIP (HIP), (4) call from physics code via gpu::your_function(). For deterministic results: CUBLAS_WORKSPACE_CONFIG=:4096:8.

9. Error Handling

Situation Action
Invalid YAML input throw std::invalid_argument with field name and allowed values
Unknown YAML key Hard abort (strict schema)
UPF parse failure throw UPFParseError with file path and line number
Negative density Clamp to 0; abort if magnitude > 1e-6
Energy oscillation > 1 Ry Abort with diagnostic
SCF non-convergence Return SCFResult{converged=false} (not an exception)
Davidson divergence Auto-switch to LOBPCG
GPU OOM Auto-fallback to CPU

Exit codes from main.cpp: 0 (success), 1 (SCF not converged), 2 (input error), 3 (PP error), 4 (other).