Merge pull request #14 from kirk0830/speedup-poisson-jac

Perf: parallelize the construction of Jacobian and RHS-vector in Poisson solving problem

Author: AsymmetryChou

dpnegf/negf/newton_raphson_speed_up.py

@@ -0,0 +1,251 @@
+import logging
+from typing import Optional, Tuple, List, Callable
+
+import numpy as np
+from scipy.sparse import lil_matrix
+
+log = logging.getLogger(__name__)
+
+def _impose_j_bound(inout, nx, ny, nz, typ, val, mask):
+    '''impose the special mask for boundary points'''
+    # Neumann boundary types
+    neumann_types_ = ['xmin', 'xmax', 'ymin', 'ymax', 'zmin', 'zmax']
+    displ_ = [+1, -1, +nx, -nx, +nx*ny, -nx*ny]
+    for t, d in zip(neumann_types_, displ_):
+        i = np.where(np.array(typ) == t)[0]
+        if len(i) == 0:
+            log.warning(f'no grid point with type {t} found')
+        # scipy sparse matrix also supports the parallelized assignment
+        inout[i, i] = val
+        inout[i, i + d] = mask
+
+    # Dirichlet boundary
+    i = np.where(typ == 'Dirichlet')[0]
+    inout[i, i] = val
+
+def _impose_b_bound(inout, nx, ny, nz, typ, phi, dirichlet_pot):
+    '''impose the special mask for boundary points in b vector'''
+    # Neumann boundary types
+    neumann_types_ = ['xmin', 'xmax', 'ymin', 'ymax', 'zmin', 'zmax']
+    displ_ = [+1, -1, +nx, -nx, +nx*ny, -nx*ny]
+    for t, d in zip(neumann_types_, displ_):
+        i = np.where(typ == t)[0]
+        if len(i) == 0:
+            log.warning(f'no grid point with type {t} found')
+        inout[i] = phi[i] - phi[i + d]
+
+    # Dirichlet boundary
+    i = np.where(typ == 'Dirichlet')[0]
+    inout[i] = phi[i] - dirichlet_pot[i]
+
+def coulomb(chr: float|np.ndarray) -> float|np.ndarray:
+    '''convert the charge to unit of Coulomb'''
+    assert isinstance(chr, (float, np.ndarray)), "chr must be a float or numpy array"
+    from dpnegf.utils.constants import elementary_charge as e
+    return chr * e
+
+def _bflux_impl(jflux, i, nx, ny, nz, phi, full_size=True) -> np.ndarray:
+    '''calculate the b vector from the jflux and phi'''
+    disp_ = [-1, +1, -nx, +nx, -nx*ny, +nx*ny]
+    if full_size:
+        assert jflux.shape == (6, len(phi))
+        bflux = np.zeros((6, len(phi)), dtype=float)
+        for j, d in enumerate(disp_):
+            bflux[j, i] = jflux[j, i] * (phi[i + d] - phi[i])
+    else:
+        assert jflux.shape == (6, len(i))
+        bflux = np.array([jf * (phi[i + d] - phi[i]) for jf, d in zip(jflux, disp_)])
+        assert bflux.shape == (6, len(i))
+    return bflux
+
+def _jflux_impl(nx, ny, nz, r, typ, sigma, eps, eps0, avgeps, with_index=False, 
+                full_size=True) -> Tuple[np.ndarray, Optional[np.ndarray]]:
+    '''kernal implementation of the Jacobian calculation.'''
+
+    # get all indices of the grid points of type "in"
+    i = np.where(typ == "in")[0]
+    # j stands for flux, <listcomp> faster than for loop
+    disp_ = [-1, +1, -nx, +nx, -nx*ny, +nx*ny]
+    if not full_size:
+        jflux = np.array([avgeps(eps[i + d], eps[i]) * eps0 * sigma[i, j // 2] \
+                          / np.abs(r[i + d, j // 2] - r[i, j // 2])
+                          for j, d in enumerate(disp_)]).reshape((6, -1))
+        assert jflux.shape == (6, len(i))
+    else:
+        jflux = np.zeros((6, len(typ)), dtype=float)
+        for j, d in enumerate(disp_):
+            jflux[j, i] =   avgeps(eps[i + d], eps[i]) * eps0 * sigma[i, j // 2] \
+                          / np.abs(r[i + d, j // 2] - r[i, j // 2])
+
+    return (jflux, i if with_index else None)
+
+def _jacobian(inout, flux, i, nx, ny, nz, typ, zfree, phi, phi_, beta):
+    '''calculate the Jacobian matrix for the Poisson equation'''
+    disp_ = [-1, +1, -nx, +nx, -nx*ny, +nx*ny]
+    # diagonal elements
+    assert flux.shape == (6, len(i))
+    inout[i, i]  = -np.sum(flux, axis=0)
+    inout[i, i] -= beta * \
+        np.abs(coulomb(zfree[i])) * np.exp(-beta * np.sign(zfree[i]) * (phi[i] - phi_[i]))
+    # non-diagonal elements
+    for jf, d in zip(flux, disp_):
+        inout[i, i + d] = jf
+    # impose the boundary conditions
+    _impose_j_bound(inout, nx, ny, nz, typ, val=1.0, mask=-1.0)
+
+def _rhsvec(inout, flux, i, nx, ny, nz, typ, zfree, zfixed, phi, phi_, dirichlet_pot, beta):
+    '''calculate the right-hand side vector for the Poisson equation'''
+    # calculate the b vector
+    assert flux.shape == (6, len(i))
+    inout[i]  = np.sum(flux, axis=0)
+    inout[i] += coulomb(zfree[i]) * \
+        np.exp(-beta * np.sign(zfree[i]) * (phi[i] - phi_[i]))
+    inout[i] += coulomb(zfixed[i])
+    # impose the boundary conditions
+    _impose_b_bound(inout, nx, ny, nz, typ, phi, dirichlet_pot)
+    
+    inout *= -1 # for convenience change the sign of B in later NR iteration
+
+def nr_construct(jinout: Optional[lil_matrix], 
+              binout: Optional[np.ndarray],
+              grid_dim: Tuple[int, int, int]|List[int],
+              gridpoint_coords: np.ndarray, 
+              gridpoint_typ: List[str]|np.ndarray,
+              gridpoint_surfarea: np.ndarray, 
+              eps: np.ndarray, 
+              phi: np.ndarray, 
+              phi_: np.ndarray, 
+              free_chr: np.ndarray,
+              fixed_chr: Optional[np.ndarray] = None,
+              dirichlet_pot: Optional[np.ndarray] = None,
+              eps0: float = 8.854187817e-12, 
+              beta: float = 1/(1.380649e-23 * 300), 
+              feps: Callable[[float, float], float] = lambda eps1, eps2: (eps1 + eps2) / 2.0) -> None:
+    '''
+    refactored version of the implementation to calculate the Jacobian for the Poisson equation
+    using the Newton-Raphson method (finit difference method).
+    
+    NOTE: the input would be modified in-place
+    
+    Parameters
+    ----------
+    jinout : scipy.sparse.lil_matrix
+        The Jacobian matrix to be calculated. It is a sparse matrix in LIL format. If is None,
+        then the Jacobian will not be calculated.
+    binout : np.ndarray
+        The b vector to be calculated, which is the right-hand side of the linear system.
+        It is a dense numpy array. If is None, then the b vector will not be calculated.
+    grid_dim : tuple or list of int
+        The dimensions of the grid as (nx, ny, nz), where nx, ny, and nz are the number of grid 
+        points in the x, y, and z directions, respectively.
+    gridpoint_coords : np.ndarray
+        The coordinates of the grid points, shape (nr, 3), where nr is the total number of grid
+        points.
+    gridpoint_typ : list of str
+        The type of grid point, with length nr
+    gridpoint_surfarea : np.ndarray
+        The area of Voronoi surface in three dimension, shape (nr, 3). For example the gridpoint_
+        area[ir, 0] indicates the area of surface whose norm vector is along x-axis
+    eps : np.ndarray
+        The relative permittivity of each grid point
+    phi : np.ndarray
+        The electrostatic potential mapped on real-space grid point
+    phi\_ : np.ndarray
+        The "old" electrostatic potential mapped on real-space grid point
+    free_chr : np.ndarray
+        The free charge density mapped on real-space grid point
+    fixed_chr : np.ndarray, optional
+        The fixed charge density mapped on real-space grid point, shape (nr,). 
+    dirichlet_pot : np.ndarray, optional
+        The Dirichlet potential mapped on real-space grid point, shape (nr,). 
+    eps0 : float, optional
+        The vacuum permittivity, a constant value. Default is 8.854187817e-12 F/m.
+    beta : float, optional
+        The inverse temperature (1/kBT), used in the Boltzmann factor for charge density.
+        Default is 1/(1.380649e-23 * 300), which corresponds to room temperature (300 K). 
+    feps : callable, optional
+        the functor to calculate the average permittivity on a pair of grid points.
+        It should take two arguments, eps1 and eps2, and return the average permittivity.
+        If not provided, it defaults to the arithmetic mean.
+    '''
+    # shortcut: nothing to do
+    if jinout is None and binout is None:
+        return
+    
+    # check dimensions
+    if not isinstance(grid_dim, (tuple, list)) or len(grid_dim) != 3:
+        raise ValueError("grid_dim must be a tuple or list of three integers")
+    if any(not isinstance(d, int) or d <= 0 for d in grid_dim):
+        raise ValueError("grid_dim must contain positive integers")
+    nx, ny, nz = grid_dim; nr = nx * ny * nz
+    
+    # simple sanity check
+    if jinout is not None:
+        if not isinstance(jinout, lil_matrix):
+            raise TypeError("jinout (jacobian) must be a scipy.sparse.lil_matrix")
+        if jinout.shape != (nr, nr):
+            raise ValueError(f"jinout (jacobian) must have shape ({nr}, {nr}), but got {jinout.shape}")
+
+    if binout is not None:
+        if not isinstance(binout, np.ndarray):
+            raise TypeError("binout must be a numpy array")
+        if binout.shape != (nr,):
+            raise ValueError(f"binout must have shape ({nr},), but got {binout.shape}")
+    
+    # check the common input parameters
+    if not isinstance(gridpoint_coords, np.ndarray) or gridpoint_coords.shape != (nr, 3):
+        raise ValueError(f"gridpoint_coords must be a numpy array of shape ({nr}, 3)")
+    gridpoint_typ = np.array(list(gridpoint_typ.values()), dtype=str) \
+        if isinstance(gridpoint_typ, dict) else np.array(gridpoint_typ, dtype=str)
+    if not isinstance(gridpoint_typ, np.ndarray) or len(gridpoint_typ) != nr:
+        raise ValueError(f"gridpoint_typ must be a list of length {nr}")
+    if not isinstance(gridpoint_surfarea, np.ndarray) or gridpoint_surfarea.shape != (nr, 3):
+        raise ValueError(f"gridpoint_surfarea must be a numpy array of shape ({nr}, 3)")
+    if not isinstance(eps, np.ndarray) or eps.shape != (nr,):
+        raise ValueError(f"eps must be a numpy array of shape ({nr},)")
+    if not isinstance(phi, np.ndarray) or phi.shape != (nr,):
+        raise ValueError(f"phi must be a numpy array of shape ({nr},)")
+    if not isinstance(phi_, np.ndarray) or phi_.shape != (nr,):
+        raise ValueError(f"phi_ must be a numpy array of shape ({nr},)")
+    if not isinstance(free_chr, np.ndarray) or free_chr.shape != (nr,):
+        raise ValueError(f"free_chr must be a numpy array of shape ({nr},)")
+    
+    if binout is not None:
+        if not isinstance(fixed_chr, np.ndarray) or fixed_chr.shape != (nr,):
+            raise ValueError(f"fixed_chr must be a numpy array of shape ({nr},)")
+        if not isinstance(dirichlet_pot, np.ndarray) or dirichlet_pot.shape != (nr,):
+            raise ValueError(f"dirichlet_pot must be a numpy array of shape ({nr},)")
+
+    if not isinstance(eps0, (float, int)):
+        raise TypeError("eps0 must be a float or int")
+    if not isinstance(beta, (float, int)):
+        raise TypeError("beta must be a float or int")
+    if not callable(feps):
+        raise TypeError("feps must be a callable function with strictly two arguments: "
+                        "eps1 and eps2 to calculate the average permittivity")
+    
+    # ================================ implementation =======================================
+    # making alias for more succint code
+    r      = gridpoint_coords
+    typ    = np.array(gridpoint_typ, dtype=str)
+    sigma  = gridpoint_surfarea
+    zfree  = free_chr
+    zfixed = fixed_chr
+    
+    # calculate the flux of jacobian for all grid points of type "in"
+    jflux, ind = _jflux_impl(nx, ny, nz, r, typ, sigma, eps, eps0, feps, 
+                             with_index=True, full_size=False)
+    # the "ind" is the indices of the grid points of type "in"
+    assert jflux.shape == (6, len(ind)) # ensure the full_size is disabled
+    
+    if jinout is not None: # only calculate jacobian when jinout is provided (allocated)
+        _jacobian(jinout, jflux, ind, nx, ny, nz, typ, zfree, phi, phi_, beta)
+        assert jinout.shape == (nr, nr)
+
+    if binout is not None: # only calculate b vector when binout is provided (allocated)
+        # calculate bflux with jflux
+        bflux = _bflux_impl(jflux, ind, nx, ny, nz, phi, full_size=False)
+        assert bflux.shape == (6, len(ind))
+        _rhsvec(binout, bflux, ind, nx, ny, nz, typ, zfree, zfixed, phi, phi_, 
+                dirichlet_pot, beta)
+        assert binout.shape == (nr,)