source: codes/icosagcm/devel/Python/src/unstructured.pyx @ 640

import time
import math
import numpy as np
import dynamico.wrap as wrap
from ctypes import c_void_p, c_int, c_double, c_bool

from libc.time cimport time as ctime, time_t
cimport numpy as np

#------------- direct Cython interface to DYNAMICO routines -------------#

cdef enum: max_nb_stage=5
cdef extern :
    cdef double tauj[max_nb_stage]
    cdef double cslj[max_nb_stage][max_nb_stage]
    cdef double cflj[max_nb_stage][max_nb_stage]
    cdef int nb_stage[1]

cdef extern from "functions.h":
     cdef void dynamico_ARK_step(double *mass_col, double *rhodz, double *theta_rhodz, 
                                 double *u, double *geopot, double *w,
                                 double *theta, double *ps, double *pk, double *hflux, double *qv,
                                 double *dmass_col, double *drhodz, double *dtheta_rhodz,
                                 double *du_fast, double *du_slow,
                                 double *dPhi_fast, double *dPhi_slow, 
                                 double *dW_fast, double *dW_slow)
     cdef void dynamico_init_params()
     cpdef void dynamico_setup_xios()
     cpdef void dynamico_xios_set_timestep(double)
     cpdef void dynamico_xios_update_calendar(int)

#------------- import and wrap DYNAMICO routines -------------#

ker=wrap.Struct() # store imported fun X as funs.X

check_args = False # use True instead of False for debugging, probably with some overhead

try:    
    kernels = wrap.SharedLib(vars(ker), 'libicosa.so', check_args=check_args) 
    setvar, setvars, getvar, getvars = kernels.setvar, kernels.setvars, kernels.getvar, kernels.getvars
except OSError:
    print """
Unable to load shared library 'libkernels.so' !
    """
    raise

# providing a full prototype enables type-checking when calling
# if a number n is present in the prototype, the previous type is repeated n times
kernels.import_funs([
                     ['dynamico_setup_xios',None],
                     ['dynamico_xios_set_timestep',c_double],
                     ['dynamico_xios_update_calendar',c_int],
                     ['dynamico_init_mesh',c_void_p,13],
                     ['dynamico_init_metric', c_void_p,6],
                     ['dynamico_init_hybrid', c_void_p,3],
                     ['dynamico_caldyn_unstructured', c_double, c_void_p,20],
                     ['dynamico_partition_graph', c_int,2, c_void_p,3, c_int, c_void_p],
                     ])

# set/get global variables
eta_mass,eta_lag=(1,2)
thermo_theta,thermo_entropy,thermo_moist,thermo_boussinesq=(1,2,3,4)

kernels.addvars(
    c_bool,'hydrostatic','debug_hevi_solver',
    c_int,'llm','nqdyn','primal_num','max_primal_deg',
    'dual_num','max_dual_deg','edge_num','max_trisk_deg',
    'caldyn_thermo','caldyn_eta','nb_threads',
    c_double,'elapsed','g', 'ptop', 'cpp', 'cppv',
    'Rd', 'Rv', 'preff', 'Treff', 'pbot', 'rho_bot', 'Phi_bot')

elapsed=0.

#------------------------ Extension type performing a full ARK time step ----------------------

ctypedef double *dbl_ptr
cdef dbl_ptr ptr1(double[:] data) : return &data[0]
cdef dbl_ptr ptr2(double[:,:] data) : return &data[0,0]
cdef dbl_ptr ptr3(double[:,:,:] data) : return &data[0,0,0]
cdef dbl_ptr ptr(data):
    n=data.ndim
    if n==1 : return ptr1(data)
    if n==2 : return ptr2(data)
    if n==3 : return ptr3(data)

cdef alloc(dbl_ptr *p, allocator, n=1):
    data=allocator(n)
    p[0]=ptr(data)
    return data

cdef check_ptr(name, dbl_ptr p, np.ndarray data):
    if p != ptr(data) : print name, 'p <> ptr(data) !!'
        
cdef class Caldyn_step:
    # pointer to allocated arrays
    cdef dbl_ptr p_mass, p_theta_rhodz, p_u, p_geopot, p_w                   # prognostic
    cdef dbl_ptr p_mass_col, p_dmass_col, p_ps, p_theta, p_pk, p_hflux, p_qv # diagnostic
    cdef dbl_ptr p_drhodz, p_dtheta_rhodz, p_du_fast, p_du_slow                # tendencies
    cdef dbl_ptr p_dPhi_fast, p_dPhi_slow, p_dW_fast, p_dW_slow                # tendencies
    # allocated arrays, must remain referenced or segfault
    cdef readonly np.ndarray mass, theta_rhodz, u, geopot, w
    cdef readonly np.ndarray mass_col, dmass_col, ps, theta, pk, hflux, qv
    cdef readonly np.ndarray drhodz, dtheta_rhodz, du_fast, du_slow
    cdef readonly np.ndarray dPhi_fast, dPhi_slow, dW_fast, dW_slow

    def __init__(self,mesh,time_scheme):
        #        self.mesh=mesh
        fps, ftheta, fmass = mesh.field_ps, mesh.field_theta, mesh.field_mass
        fw, fu, fz = mesh.field_w, mesh.field_u, mesh.field_z
        # collect coefficients of time scheme
        cdef double[:]   tauj_ = time_scheme.tauj
        cdef double[:,:] cslj_ = time_scheme.csjl
        cdef double[:,:] cflj_ = time_scheme.cfjl
        ns = time_scheme.nstage
        nb_stage[0]= ns
        cdef int i,j
        for i in range(ns):
            tauj[i]=tauj_[i]
            for j in range(ns):
                cslj[i][j]=cslj_[i,j]
                cflj[i][j]=cflj_[i,j]
        # allocate arrays, store pointers to avoid overhead when calling dynamico
        #    prognostic/diagnostic
        self.ps                       = alloc(&self.p_ps, fps) 
        self.mass_col, self.dmass_col = alloc(&self.p_mass_col, fps), alloc(&self.p_dmass_col, fps), 
        self.mass, self.theta_rhodz   = alloc(&self.p_mass, fmass), alloc(&self.p_theta_rhodz, fmass),
        self.theta, self.pk           = alloc(&self.p_theta, fmass), alloc(&self.p_pk, fmass), 
        self.geopot, self.w           = alloc(&self.p_geopot, fw), alloc(&self.p_w, fw),
        self.hflux, self.u            = alloc(&self.p_hflux, fu), alloc(&self.p_u, fu) 
        self.qv                       = alloc(&self.p_qv,fz)
        #    tendencies
        self.drhodz, self.dtheta_rhodz  = alloc(&self.p_drhodz,fmass,ns), alloc(&self.p_dtheta_rhodz,fmass,ns) 
        self.du_fast, self.du_slow      = alloc(&self.p_du_fast,fu,ns), alloc(&self.p_du_slow,fu,ns)
        self.dPhi_fast, self.dPhi_slow  = alloc(&self.p_dPhi_fast,fw,ns), alloc(&self.p_dPhi_slow,fw,ns)
        self.dW_fast, self.dW_slow      = alloc(&self.p_dW_fast,fw,ns), alloc(&self.p_dW_slow,fw,ns)
    def next(self):
        #        global elapsed
        # time1=time.time()
        check_ptr('mass', self.p_mass, self.mass)
        check_ptr('u', self.p_u, self.u)
        dynamico_ARK_step(self.p_mass_col, self.p_mass, self.p_theta_rhodz,
                          self.p_u, self.p_geopot, self.p_w,
                          self.p_theta, self.p_ps, self.p_pk, self.p_hflux, self.p_qv,
                          self.p_dmass_col, self.p_drhodz, self.p_dtheta_rhodz,
                          self.p_du_fast, self.p_du_slow,
                          self.p_dPhi_fast, self.p_dPhi_slow,
                          self.p_dW_fast, self.p_dW_slow)
        #time2=time.time()
        #if time2>time1: elapsed=elapsed+time2-time1
    def setup_TRSW(self):
        setvars(('hydrostatic','caldyn_thermo','caldyn_eta'),
                (True,thermo_boussinesq,eta_lag))
        dynamico_init_params()
    def setup_HPE(self, caldyn_thermo, caldyn_eta, g, BC,thermo):
        setvars(('hydrostatic','caldyn_thermo','caldyn_eta',
                 'g','ptop','Rd','cpp','preff','Treff'),
                (True,caldyn_thermo,caldyn_eta,
                 g,BC.ptop,thermo.Rd,thermo.Cpd,thermo.p0,thermo.T0))
        dynamico_init_params()
    def setup_NH(self, caldyn_thermo, caldyn_eta, g, BC,thermo):
        setvars(('hydrostatic','caldyn_thermo','caldyn_eta',
                 'g','ptop','Rd','cpp','preff','Treff',
                 'pbot','rho_bot'),
                (False,caldyn_thermo,caldyn_eta,
                 g,BC.ptop,thermo.Rd,thermo.Cpd,thermo.p0,thermo.T0,
                 BC.pbot.max(), BC.rho_bot.max()))
        dynamico_init_params()

#----------------------------- Base class for dynamics ------------------------

class Caldyn:
    def __init__(self,mesh):
        self.mesh=mesh
        fps, ftheta, fmass = mesh.field_ps, mesh.field_theta, mesh.field_mass
        fw, fu, fz = mesh.field_w, mesh.field_u, mesh.field_z
        self.ps, self.ms, self.dms       = fps(), fps(), fps()
        self.s, self.hs, self.dhs        = ftheta(), ftheta(), ftheta()
        self.pk, self.berni, self.geopot, self.hflux = fmass(),fmass(),fw(),fu()
        self.qu, self.qv                 = fu(),fz()
        self.fmass, self.ftheta, self.fu, self.fw = fmass, ftheta, fu, fw
    def bwd_fast_slow(self, flow, tau):
        global elapsed
        time1=time.time()
        flow,fast,slow = self._bwd_fast_slow_(flow,tau)
        time2=time.time()
        elapsed=elapsed+time2-time1
        return flow,fast,slow

# when calling caldyn_unstructured, arrays for tendencies must be re-created each time
# to avoid overwriting in the same memory space when time scheme is multi-stage

#-------------------------- Shallow-water dynamics ---------------------

class Caldyn_RSW(Caldyn):
    def __init__(self,mesh):
        Caldyn.__init__(self,mesh)
        setvars(('hydrostatic','caldyn_thermo','caldyn_eta'),
                (True,thermo_boussinesq,eta_lag))
        self.dhs = self.fmass()
        dynamico_init_params()
    def _bwd_fast_slow_(self, flow, tau):
        h,u = flow
        # h*s = h => uniform buoyancy s=1 => shallow-water
        dh, du_slow, du_fast, hs, buf = self.fmass(), self.fu(), self.fu(), h.copy(), self.geopot
        ker.dynamico_caldyn_unstructured(tau, self.ms, h, hs, u, self.geopot, buf,
                  self.s, self.ps, self.pk, self.hflux, self.qv,
                  self.dms, dh, self.dhs, du_fast, du_slow,
                  buf, buf, buf, buf)
        return (h,u), (0.,du_fast), (dh,du_slow)

#----------------------------------- HPE ------------------------------------

class Caldyn_HPE(Caldyn):
    def __init__(self,caldyn_thermo,caldyn_eta, mesh,metric,thermo,BC,g):
        Caldyn.__init__(self,mesh)
        setvars(('hydrostatic','caldyn_thermo','caldyn_eta',
                 'g','ptop','Rd','cpp','preff','Treff'),
                (True,caldyn_thermo,caldyn_eta,
                 g,BC.ptop,thermo.Rd,thermo.Cpd,thermo.p0,thermo.T0))
        dynamico_init_params()
    def _bwd_fast_slow_(self, flow, tau):
        dm, dS, du_slow, du_fast, buf = self.fmass(), self.ftheta(), self.fu(), self.fu(), self.geopot
        m,S,u = flow
        ker.dynamico_caldyn_unstructured(tau, self.ms, m, S, u, self.geopot, buf,
                  self.s, self.ps, self.pk, self.hflux, self.qv,
                  self.dms, dm, dS, du_fast, du_slow,
                  buf, buf, buf, buf)
        return (m,S,u), (0.,0.,du_fast), (dm,dS,du_slow)

#----------------------------------- NH ------------------------------------

class Caldyn_NH(Caldyn):
    def __init__(self,caldyn_thermo,caldyn_eta, mesh,metric,thermo,BC,g):
        Caldyn.__init__(self,mesh)
        setvars(('hydrostatic','caldyn_thermo','caldyn_eta',
                 'g','ptop','Rd','cpp','preff','Treff',
                 'pbot','rho_bot'),
                (False,caldyn_thermo,caldyn_eta,
                 g,BC.ptop,thermo.Rd,thermo.Cpd,thermo.p0,thermo.T0,
                 BC.pbot.max(), BC.rho_bot.max()))
        dynamico_init_params()
    def bwd_fast_slow(self, flow, tau):
        ftheta, fmass, fu, fw = self.ftheta, self.fmass, self.fu, self.fw 
        dm, dS, du_slow, du_fast = fmass(), ftheta(), fu(), fu()
        dPhi_slow, dPhi_fast, dW_slow, dW_fast = fw(), fw(), fw(), fw()
        m,S,u,Phi,W = flow
        ker.dynamico_caldyn_unstructured(tau, self.ms, m, S, u, Phi, W,
                  self.s, self.ps, self.pk, self.hflux, self.qv,
                  self.dms, dm, dS, du_fast, du_slow,
                  dPhi_fast, dPhi_slow, dW_fast, dW_slow)
        return ((m,S,u,Phi,W), (0.,0.,du_fast,dPhi_fast,dW_fast), 
                (dm,dS,du_slow,dPhi_slow,dW_slow))

#------------------------ Copy mesh info to Fortran side -------------------

def init_mesh(llm, nqdyn, edge_num, primal_num, dual_num,
              max_trisk_deg, max_primal_deg, max_dual_deg,
              primal_nb, primal_edge, primal_ne,
              dual_nb,dual_edge,dual_ne,dual_vertex, 
              left,right,down,up,trisk_deg,trisk,
              Ai, Av, fv, le_de, Riv2, wee):
    setvars( ('llm','nqdyn','edge_num','primal_num','dual_num',
              'max_trisk_deg','max_primal_deg','max_dual_deg'),
             (llm, nqdyn, edge_num, primal_num, dual_num, 
              max_trisk_deg, max_primal_deg, max_dual_deg) )        
    print('init_mesh ...')
    ker.dynamico_init_mesh(primal_nb,primal_edge,primal_ne,
              dual_nb,dual_edge,dual_ne,dual_vertex,
              left,right,down,up,trisk_deg,trisk)
    print ('...done')
    print('init_metric ...')
    ker.dynamico_init_metric(Ai,Av,fv,le_de,Riv2,wee)
    print ('...done')

#------------------------ Mesh partitioning ------------------------

# Helper functions and interface to ParMETIS
# list_stencil converts an adjacency graph from array format index[num_cells, MAX_EDGES] to compressed format
# loc_stencil returns the start/end indices (vtxdist) expected by ParMETIS
# i.e. index[start:end] with start=vtxdist[cell], end=vtxdist[cell+1] lists the edges of cell 'cell'

def list_stencil(degree, stencil, cond=lambda x:True):
    for i in range(degree.size):
        for j in range(degree[i]):
            s=stencil[i,j]
            if cond(s): yield stencil[i,j]
                
def loc_stencil(degree, stencil):
    loc=0
    for i in range(degree.size):
        yield loc
        loc=loc+degree[i]
    yield loc

def partition_mesh(degree, stencil, nparts): 
    # arguments : PArray1D and PArray2D describing mesh, number of desired partitions
    dim_cell, degree, stencil = degree.dim, degree.data, stencil.data
    comm, vtxdist, idx_start, idx_end = dim_cell.comm, dim_cell.vtxdist, dim_cell.start, dim_cell.end
    mpi_rank, mpi_size = comm.Get_rank(), comm.Get_size()
    adjncy_loc, xadj_loc = list_stencil(degree, stencil), loc_stencil(degree, stencil)
    adjncy_loc, xadj_loc = [np.asarray(list(x), dtype=np.int32) for x in (adjncy_loc, xadj_loc)]
    owner = np.zeros(idx_end-idx_start, dtype=np.int32);
    ker.dynamico_partition_graph(mpi_rank, mpi_size, vtxdist, xadj_loc, adjncy_loc, nparts, owner)
    return owner