turbx 1.0.2__py3-none-any.whl

turbx/rgd_xpln_spectrum.py

@@ -0,0 +1,940 @@
+import os
+import re
+import sys
+import timeit
+from concurrent.futures import ThreadPoolExecutor
+from pathlib import Path, PurePosixPath
+
+import h5py
+import numpy as np
+import psutil
+import scipy as sp
+from mpi4py import MPI
+from scipy.signal import csd
+from tqdm import tqdm
+
+from .h5 import h5_print_contents
+from .utils import even_print, format_time_string
+
+# ======================================================================
+
+def _calc_turb_cospectrum_xpln(self, **kwargs):
+    '''
+    Calculate FFT cospectrum in [z,t] at every [x,y]
+    - Designed for analyzing unsteady, thin planes in [x]
+    - Multithreaded with ThreadPoolExecutor()
+        - scipy.signal.csd() automatically tries to run multithreaded
+        - set OMP_NUM_THREADS=1 and pass 'n_threads' to as kwarg manually
+    '''
+    
+    if (self.rank==0):
+        verbose = True
+    else:
+        verbose = False
+    
+    if verbose: print('\n'+'rgd.calc_turb_cospectrum_xpln()'+'\n'+72*'-')
+    t_start_func = timeit.default_timer()
+    
+    ## assert that the opened RGD has fsubtype 'unsteady' (i.e. is NOT a prime file)
+    if (self.fsubtype!='unsteady'):
+        raise ValueError
+    if not self.usingmpi:
+        raise NotImplementedError('function is not implemented for non-MPI usage')
+    
+    if not h5py.h5.get_config().mpi:
+        if verbose: print('h5py was not compiled for parallel usage! exiting.')
+        sys.exit(1)
+    
+    rx = kwargs.get('rx',1)
+    ry = kwargs.get('ry',1)
+    rz = kwargs.get('rz',1)
+    rt = kwargs.get('rt',1)
+    
+    # cy = kwargs.get('cy',1) ## number of subdivisions per rank [y] range
+    # if not isinstance(cy,int):
+    #     raise TypeError('cy should be an int')
+    # if (cy<1):
+    #     raise TypeError('cy should be an int')
+    
+    sy = kwargs.get('sy',1) ## number of [y] layers to read at a time
+    if not isinstance(sy,int) or (sy<1):
+        raise TypeError('sy should be a positive non-zero int')
+    
+    n_threads = kwargs.get('n_threads',1)
+    
+    ## Debug Rank:Proc Affinity
+    #pp = psutil.Process()
+    #print(f"[Rank {self.rank}] sees CPUs: {pp.cpu_affinity()}  |  n_threads={n_threads}  |  OMP_NUM_THREADS={os.environ.get('OMP_NUM_THREADS')}")
+    
+    #try:
+    #    n_threads = int(os.environ.get('OMP_NUM_THREADS'))
+    #except TypeError: ## not set
+    #    n_threads = os.cpu_count()
+    
+    fn_h5_out       = kwargs.get('fn_h5_out',None)       ## filename for output HDF5 (.h5) file
+    overlap_fac_nom = kwargs.get('overlap_fac_nom',0.50) ## nominal windows overlap factor
+    n_win           = kwargs.get('n_win',8)              ## number of segment windows for [t] PSD calc
+    #window_type     = kwargs.get('window_type','hann')   ## 'tukey','hann'
+    
+    ## only distribute data across [y]
+    if (rx!=1):
+        raise ValueError('rx!=1')
+    if (rz!=1):
+        raise ValueError('rz!=1')
+    if (rt!=1):
+        raise ValueError('rt!=1')
+    
+    if not isinstance(ry,int) or (ry<1):
+        raise ValueError('ry should be a positive non-zero int')
+    
+    ## check the choice of ranks per dimension
+    if (rx*ry*rz*rt != self.n_ranks):
+        raise ValueError('rx*ry*rz*rt != self.n_ranks')
+    if (rx>self.nx):
+        raise ValueError('rx>self.nx')
+    if (ry>self.ny):
+        raise ValueError('ry>self.ny')
+    if (rz>self.nz):
+        raise ValueError('rz>self.nz')
+    if (rt>self.nt):
+        raise ValueError('rt>self.nt')
+    
+    if (self.ny%ry!=0):
+        raise ValueError('ny not divisible by ry')
+    
+    ## distribute 4D data over ranks --> here only in [y]
+    ryl_ = np.array_split(np.arange(self.ny,dtype=np.int64),ry)
+    ryl = [[b[0],b[-1]+1] for b in ryl_ ]
+    ry1,ry2 = ryl[self.rank]
+    nyr = ry2 - ry1
+    
+    ## check all [y] ranges have same size
+    for ryl_ in ryl:
+        if not (ryl_[1]-ryl_[0]==nyr):
+            raise ValueError('[y] chunks are not even in size')
+    
+    # ## [y] sub chunk range --> cyl = list of ranges in ry1:ry2
+    # ## cy is the NUMBER of chunks for the rank sub-range
+    # cyl_ = np.array_split( np.arange(ry1,ry2) , min(cy,nyr) )
+    # cyl  = [[b[0],b[-1]+1] for b in cyl_ ]
+    # 
+    # for nyc_ in [ cyl_[1]-cyl_[0] for cyl_ in cyl ]:
+    #     if (nyc_ < 1):
+    #         #raise ValueError
+    #         print(f'rank {self.rank:d}: sub-range is <1')
+    #         self.comm.Abort(1)
+    # 
+    # if 1: ## assert that [y] sub-chunk ranges are correct
+    #     
+    #     yi = np.arange(self.ny, dtype=np.int32)
+    #     
+    #     local_indices = []
+    #     for cyl_ in cyl:
+    #         cy1, cy2 = cyl_
+    #         local_indices += [ yi_ for yi_ in yi[cy1:cy2] ]
+    #     
+    #     G = self.comm.gather([ self.rank , local_indices ], root=0)
+    #     G = self.comm.bcast(G, root=0)
+    #     
+    #     all_indices = []
+    #     for G_ in G:
+    #         all_indices += G_[1]
+    #     all_indices = np.array( sorted(all_indices), dtype=np.int32 )
+    #     
+    #     if not np.array_equal( all_indices , yi ):
+    #         raise AssertionError
+    
+    if (nyr%sy!=0):
+        raise ValueError('nyr not divisible by sy')
+    
+    ## output filename : HDF5 (.h5)
+    if (fn_h5_out is None): ## automatically determine name
+        fname_path = os.path.dirname(self.fname)
+        fname_base = os.path.basename(self.fname)
+        fname_root, fname_ext = os.path.splitext(fname_base)
+        fname_root = re.findall(r'io\S+_mpi_[0-9]+', fname_root)[0]
+        fname_fft_h5_base = fname_root+'_fft.h5'
+        fn_h5_out = str(PurePosixPath(fname_path, fname_fft_h5_base))
+    if (Path(fn_h5_out).suffix != '.h5'):
+        raise ValueError(f"fn_h5_out='{str(fn_h5_out)}' must end in .h5")
+    if os.path.isfile(fn_h5_out):
+        #if (os.path.getsize(fn_h5_out) > 8*1024**3):
+        #    raise ValueError(f"fn_h5_out='{str(fn_h5_out)}' exists and is >8 [GB]. exiting for your own safety.")
+        if (fn_h5_out == self.fname):
+            raise ValueError(f"fn_h5_out='{str(fn_h5_out)}' cannot be same as input filename.")
+    
+    if verbose: even_print( 'fn_h5'      , self.fname )
+    if verbose: even_print( 'fn_h5_out'  , fn_h5_out  )
+    if verbose: print(72*'-')
+    self.comm.Barrier()
+    
+    ## the data dictionary to be written to .h5 later
+    data = {}
+    
+    ## infile
+    fsize = os.path.getsize(self.fname)/1024**3
+    if verbose: even_print(os.path.basename(self.fname),'%0.1f [GB]'%fsize)
+    if verbose: even_print('nx',f'{self.nx:d}')
+    if verbose: even_print('ny',f'{self.ny:d}')
+    if verbose: even_print('nz',f'{self.nz:d}')
+    if verbose: even_print('nt',f'{self.nt:d}')
+    if verbose: even_print('ngp',f'{self.ngp/1e6:0.1f} [M]')
+    #if verbose: even_print('cy',f'{cy:d}')
+    if verbose: even_print('sy',f'{sy:d}')
+    if verbose: even_print('n_ranks',f'{self.n_ranks:d}')
+    if verbose: even_print('n_threads',f'{n_threads:d}')
+    if verbose: print(72*'-')
+    
+    ## 0D freestream scalars
+    lchar   = self.lchar   ; data['lchar']   = lchar
+    U_inf   = self.U_inf   ; data['U_inf']   = U_inf
+    rho_inf = self.rho_inf ; data['rho_inf'] = rho_inf
+    T_inf   = self.T_inf   ; data['T_inf']   = T_inf
+    
+    #data['M_inf'] = self.M_inf
+    data['Ma'] = self.Ma
+    data['Pr'] = self.Pr
+    
+    ## read in 1D coordinate arrays & re-dimensionalize
+    x = np.copy( self['dims/x'][()] * self.lchar )
+    y = np.copy( self['dims/y'][()] * self.lchar )
+    z = np.copy( self['dims/z'][()] * self.lchar )
+    t = np.copy( self['dims/t'][()] * self.tchar )
+    
+    nx = self.nx ; data['nx'] = nx
+    ny = self.ny ; data['ny'] = ny
+    nz = self.nz ; data['nz'] = nz
+    nt = self.nt ; data['nt'] = nt
+    
+    ## assert constant Δz
+    dz0 = np.diff(z)[0]
+    if not np.all(np.isclose(np.diff(z), dz0, rtol=1e-6)):
+        raise NotImplementedError('Δz not constant')
+    dz = np.diff(z)[0]
+    
+    ## dimensional [s]
+    dt = self.dt * self.tchar
+    np.testing.assert_allclose(dt, t[1]-t[0], rtol=1e-12, atol=1e-12)
+    
+    t_meas = self.duration * self.tchar
+    np.testing.assert_allclose(t_meas, t.max()-t.min(), rtol=1e-12, atol=1e-12)
+    
+    zrange = z.max() - z.min()
+    
+    data['x'] = x
+    data['y'] = y
+    data['z'] = z
+    
+    data['t']      = t
+    data['t_meas'] = t_meas
+    data['dt']     = dt
+    data['dz']     = dz
+    data['zrange'] = zrange
+    
+    if verbose: even_print( 'Δt/tchar'       , f'{dt/self.tchar:0.8f}' )
+    if verbose: even_print( 'Δt'             , f'{dt:0.3e} [s]'        )
+    if verbose: even_print( 'duration/tchar' , f'{self.duration:0.1f}' )
+    if verbose: even_print( 'duration'       , f'{self.duration*self.tchar:0.3e} [s]' )
+    if verbose: print(72*'-')
+    
+    ## report
+    if verbose:
+        even_print('Δt'     , f'{dt    :0.5e} [s]' )
+        even_print('t_meas' , f'{t_meas:0.5e} [s]' )
+        even_print('Δz'     , f'{dz0   :0.5e} [m]' )
+        even_print('zrange' , f'{zrange:0.5e} [m]' )
+        print(72*'-')
+    
+    ## establish [t] windowing (old)
+    #win_len, overlap = get_overlapping_window_size(nt, n_win, overlap_fac_nom)
+    #overlap_fac = overlap / win_len
+    #tw, n_win, n_pad = get_overlapping_windows(t, win_len, overlap)
+    #t_meas_per_win = (win_len-1)*dt
+    #data['win_len']        = win_len
+    #data['overlap_fac']    = overlap_fac
+    #data['overlap']        = overlap
+    #data['n_win']          = n_win
+    #data['t_meas_per_win'] = t_meas_per_win
+    
+    # ## temporal [t] frequency (f) vector (Short Time Fourier Transform)
+    # freq_full = sp.fft.fftfreq(n=win_len, d=dt)
+    # fp        = np.where(freq_full>0) ## indices of positive values
+    # freq      = np.copy(freq_full[fp])
+    # df        = freq[1] - freq[0]
+    # nf        = freq.size
+    
+    ## establish [t] windowing & get frequency
+    nperseg     = nt // n_win
+    noverlap    = int(round(nperseg*overlap_fac_nom))
+    overlap_fac = noverlap / nperseg
+    fs          = 1./dt ## dimensional [1/s]
+    
+    ## get [freq] vector
+    freq,_ = csd(
+            np.zeros((nt,),dtype=np.float64),
+            np.zeros((nt,),dtype=np.float64),
+            fs=fs,
+            nperseg=nperseg,
+            noverlap=noverlap,
+            window='hann',
+            detrend='constant',
+            scaling='density',
+            return_onesided=True,
+            )
+    
+    nf = freq.shape[0]
+    df = np.diff(freq)[0]
+    
+    data['nperseg']  = nperseg
+    data['noverlap'] = noverlap
+    data['freq']     = freq
+    data['df']       = df
+    data['nf']       = nf
+    
+    if verbose:
+        even_print('overlap_fac (nominal)' , f'{overlap_fac_nom:0.5f}'  )
+        even_print('n_win'                 , f'{n_win:d}'               )
+        even_print('nperseg'               , f'{nperseg:d}'             )
+        even_print('noverlap'              , f'{noverlap:d}'            )
+        even_print('overlap_fac'           , f'{overlap_fac:0.5f}'      )
+        print(72*'-')
+    
+    if verbose:
+        even_print('freq min',f'{freq.min():0.1f} [Hz]')
+        even_print('freq max',f'{freq.max():0.1f} [Hz]')
+        even_print('df',f'{df:0.1f} [Hz]')
+        even_print('nf',f'{nf:d}')
+        print(72*'-')
+    
+    ## spatial [x] wavenumber (kx) and wavelength (λx)
+    # λx  = u/f
+    # kx  = 2·π·f/u
+    # λx+ = λx/(ν/uτ)
+    # kx+ = (2·π·f/u)·(ν/uτ)
+    
+    ## kx = 2·π·f/u --> [y,f]
+    #kx = np.copy( 2*np.pi*freq[na,:] / u_avg[:,na] )
+    
+    ## λx = u/f --> [y,f]
+    #lx = np.copy( u_avg[:,na] / freq[na,:] )
+    
+    #data['kx'] = kx
+    #data['lx'] = lx
+    
+    ## spatial [z] wavenumber (kz) vector
+    kz_full = sp.fft.fftfreq(n=nz, d=dz0) * ( 2 * np.pi )
+    kzp     = np.where(kz_full>0) ## indices of positive values
+    kz      = np.copy(kz_full[kzp])
+    dkz     = kz[1] - kz[0]
+    nkz     = kz.shape[0]
+    
+    ## wavenumber vector should be size nz//2-1
+    if (nkz!=nz//2-1):
+        raise ValueError
+    
+    data['kz']  = kz
+    data['dkz'] = dkz
+    data['nkz'] = nkz
+    
+    if verbose:
+        even_print('kz min',f'{kz.min():0.1f} [1/m]')
+        even_print('kz max',f'{kz.max():0.1f} [1/m]')
+        even_print('dkz',f'{dkz:0.1f} [1/m]')
+        even_print('nkz',f'{nkz:d}')
+        print(72*'-')
+    
+    ## Wavelength λz = (2·π)/kz
+    lz = np.copy( 2 * np.pi / kz )
+    data['lz'] = lz
+    
+    # ===
+    
+    ## cospectrum pairs
+    ## [ (str:var1, bool:ρ_weighting) , (str:var2, bool:ρ_weighting) ]
+    fft_combis = [
+    
+    [ ('u',True) , ('v',True) ], ## [ ρ·u″ , ρ·v″ ]
+    [ ('u',True) , ('u',True) ], ## [ ρ·u″ , ρ·u″ ]
+    [ ('v',True) , ('v',True) ], ## [ ρ·v″ , ρ·v″ ]
+    [ ('w',True) , ('w',True) ], ## [ ρ·w″ , ρ·w″ ]
+    
+    [ ('u',False) , ('v',False) ], ## [ u′ , v′ ]
+    [ ('u',False) , ('u',False) ], ## [ u′ , u′ ]
+    [ ('v',False) , ('v',False) ], ## [ v′ , v′ ]
+    [ ('w',False) , ('w',False) ], ## [ w′ , w′ ]
+    
+    [ ('p',False) , ('u',False) ], ## [ p′ , u′ ]
+    [ ('p',False) , ('v',False) ], ## [ p′ , v′ ]
+    [ ('p',False) , ('w',False) ], ## [ p′ , w′ ]
+    
+    [ ('p',False) , ('u',True) ], ## [ p′ , ρ·u″ ]
+    [ ('p',False) , ('v',True) ], ## [ p′ , ρ·v″ ]
+    [ ('p',False) , ('w',True) ], ## [ p′ , ρ·w″ ]
+    
+    [ ('p',False)   , ('p',False)   ], ## [ p′ , p′ ]
+    [ ('T',False)   , ('T',False)   ], ## [ T′ , T′ ]
+    [ ('rho',False) , ('rho',False) ], ## [ ρ′ , ρ′ ]
+    
+    ]
+    
+    ## generate FFT cospectrum scalar names
+    scalars = []
+    for fft_combi in fft_combis:
+        if not isinstance(fft_combi,list):
+            raise RuntimeError
+        if not len(fft_combi)==2:
+            raise RuntimeError
+        
+        sL = fft_combi[0][0]
+        do_density_weighting_L = fft_combi[0][1]
+        if do_density_weighting_L:
+            sLs = f'r{sL}II'
+        else:
+            sLs = f'{sL}I'
+        
+        sR = fft_combi[1][0]
+        do_density_weighting_R = fft_combi[1][1]
+        if do_density_weighting_R:
+            sRs = f'r{sR}II'
+        else:
+            sRs = f'{sR}I'
+        
+        scalars.append(f'{sLs}_{sRs}')
+    
+    ## generate avg scalar names
+    scalars_Re_avg = []
+    scalars_Fv_avg = []
+    for fft_combi in fft_combis:
+        sL = fft_combi[0][0]
+        do_density_weighting_L = fft_combi[0][1]
+        sR = fft_combi[1][0]
+        do_density_weighting_R = fft_combi[1][1]
+        
+        if do_density_weighting_L or do_density_weighting_R:
+            if 'rho' not in scalars_Re_avg:
+                scalars_Re_avg.append('rho')
+        
+        if do_density_weighting_L:
+            if sL not in scalars_Fv_avg:
+                scalars_Fv_avg.append(sL)
+        else:
+            if sL not in scalars_Re_avg:
+                scalars_Re_avg.append(sL)
+        
+        if do_density_weighting_R:
+            if sR not in scalars_Fv_avg:
+                scalars_Fv_avg.append(sR)
+        else:
+            if sR not in scalars_Re_avg:
+                scalars_Re_avg.append(sR)
+    
+    ## numpy formatted arrays: buffers for PSD & other data (rank-local)
+    Ekz        = np.zeros(shape=(nyr,nkz ) , dtype={'names':scalars        , 'formats':[ np.dtype(np.complex128) for s in scalars        ] })
+    Ef         = np.zeros(shape=(nyr,nf  ) , dtype={'names':scalars        , 'formats':[ np.dtype(np.complex128) for s in scalars        ] })
+    covariance = np.zeros(shape=(nyr,    ) , dtype={'names':scalars        , 'formats':[ np.dtype(np.float64)    for s in scalars        ] })
+    avg_Re     = np.zeros(shape=(nyr,    ) , dtype={'names':scalars_Re_avg , 'formats':[ np.dtype(np.float64)    for s in scalars_Re_avg ] })
+    avg_Fv     = np.zeros(shape=(nyr,    ) , dtype={'names':scalars_Fv_avg , 'formats':[ np.dtype(np.float64)    for s in scalars_Fv_avg ] })
+    
+    if verbose:
+        even_print('n turb spectrum scalar combinations' , '%i'%(len(fft_combis),))
+        print(72*'-')
+    
+    ## window for [z] -- rectangular because [z] is assumed periodic already
+    window_z       = np.ones(nz,dtype=np.float64)
+    sum_sqrt_win_z = np.sum(np.sqrt(window_z))
+    mean_sq_win_z  = np.mean(window_z**2)
+    if verbose:
+        #even_print('sum(sqrt(window_z))'      , '%0.5f'%(sum_sqrt_win_z,))
+        even_print('sum(sqrt(window_z)) / nz' , '%0.5f'%(sum_sqrt_win_z/nz,))
+        even_print('mean(window_z**2)'        , '%0.5f'%(mean_sq_win_z,))
+    
+    # ## window function for [t]
+    # if (window_type=='tukey'):
+    #     window_t = sp.signal.windows.tukey(win_len,alpha=0.5,sym=False) ## α=0:rectangular, α=1:Hann
+    # elif (window_type=='hann'):
+    #     window_t = sp.signal.windows.hann(win_len,sym=False)
+    # elif (window_type is None):
+    #     window_t = np.ones(win_len, dtype=np.float64)
+    # else:
+    #     raise ValueError
+    # 
+    # if verbose:
+    #     even_print('window type [t]', '\'%s\''%str(window_type))
+    # 
+    # ## sum of sqrt of window: needed for normalization
+    # sum_sqrt_win_t = np.sum(np.sqrt(window_t))
+    # if verbose:
+    #     #even_print('sum(sqrt(window_t))'          , '%0.5f'%(sum_sqrt_win_t,))
+    #     even_print('sum(sqrt(window_t)) / win_len', '%0.5f'%(sum_sqrt_win_t/win_len,))
+    
+    #if verbose: print(72*'-')
+    
+    # ==============================================================
+    # check memory
+    # ==============================================================
+    
+    hostname = MPI.Get_processor_name()
+    mem_free_gb = psutil.virtual_memory().free / 1024**3
+    G = self.comm.gather([ self.rank , hostname , mem_free_gb ], root=0)
+    G = self.comm.bcast(G, root=0)
+    
+    host_mem = {}
+    for rank, host, mem in G:
+        if host not in host_mem or mem < host_mem[host]:
+            host_mem[host] = mem
+    total_free = sum(host_mem.values())
+    
+    if verbose:
+        print(72*'-')
+        for key,value in host_mem.items():
+            even_print(f'RAM free {key}', f'{int(np.floor(value)):d} [GB]')
+        even_print('RAM free (local,min)', f'{int(np.floor(min(host_mem.values()))):d} [GB]')
+        even_print('RAM free (global)', f'{int(np.floor(total_free)):d} [GB]')
+    
+    shape_read = (nx,sy,nz,nt) ## local
+    if verbose: even_print('read shape (local)', f'[{nx:d},{sy:d},{nz:d},{nt:d}]')
+    data_gb = np.dtype(np.float64).itemsize * np.prod(shape_read) / 1024**3
+    if verbose: even_print('read size (global)', f'{int(np.ceil(data_gb*ry)):d} [GB]')
+    
+    if verbose: even_print('read size (global) ×8', f'{int(np.ceil(data_gb*ry*8)):d} [GB]')
+    ram_usage_est = data_gb*ry*8/total_free
+    if verbose: even_print('RAM usage estimate', f'{100*ram_usage_est:0.1f} [%]')
+    
+    self.comm.Barrier()
+    if (ram_usage_est>0.80):
+        print('RAM consumption might be too high. exiting.')
+        self.comm.Abort(1)
+    
+    # ==============================================================
+    # main loop
+    # ==============================================================
+    
+    if verbose:
+        progress_bar = tqdm(
+            #total=len(fft_combis)*cy,
+            total=len(fft_combis)*(nyr//sy),
+            ncols=100,
+            desc='fft',
+            leave=True,
+            file=sys.stdout,
+            mininterval=0.1,
+            smoothing=0.,
+            #bar_format="\033[B{l_bar}{bar}| {n}/{total} [{percentage:.1f}%] {elapsed}/{remaining}\033[A\n\b",
+            bar_format="{l_bar}{bar}| {n}/{total} [{percentage:.1f}%] {elapsed}/{remaining}",
+            ascii="░█",
+            colour='#FF6600',
+            )
+    
+    for cci,cc in enumerate(fft_combis): ## fft pairs
+        
+        if verbose: tqdm.write(72*'-')
+        
+        scalar_L = cc[0][0]
+        do_density_weighting_L = cc[0][1]
+        scalar_R = cc[1][0]
+        do_density_weighting_R = cc[1][1]
+        
+        if do_density_weighting_L:
+            sLs  = f'r{scalar_L}II'
+            sLsF = f'ρ·{scalar_L}″'
+        else:
+            sLs  = f'{scalar_L}I'
+            sLsF = f'{scalar_L}′'
+        
+        if do_density_weighting_R:
+            sRs  = f'r{scalar_R}II'
+            sRsF = f'ρ·{scalar_R}″'
+        else:
+            sRs  = f'{scalar_R}I'
+            sRsF = f'{scalar_R}′'
+        
+        msg = f'[{sLsF},{sRsF}]'
+        
+        dset_L   = self[f'data/{scalar_L}']
+        dset_R   = self[f'data/{scalar_R}']
+        dset_rho = self['data/rho']
+        
+        scalar = scalars[cci]
+        
+        ## assert scalar name
+        if (f'{sLs}_{sRs}' != scalar):
+            raise RuntimeError(f"'{sLs}_{sRs}' != '{scalar}'")
+        
+        ## [y] loop outer (grid subchunks within rank)
+        # for cyl_ in cyl:
+        #     cy1, cy2 = cyl_
+        #     nyc = cy2 - cy1
+        
+        for ci in range(nyr//sy):
+            
+            ## buffers
+            data_L = np.zeros((nx,sy,nz,nt), dtype=np.float64)
+            data_R = np.zeros((nx,sy,nz,nt), dtype=np.float64)
+            if do_density_weighting_L or do_density_weighting_R:
+                rho = np.zeros((nx,sy,nz,nt), dtype=np.float64)
+            else:
+                rho = None
+            
+            cy1 = ry1 + ci*sy
+            cy2 = cy1 + sy
+            nyc = cy2 - cy1
+            
+            self.comm.Barrier()
+            t_start = timeit.default_timer()
+            
+            ## read data L
+            n_scalars_read = 1 ## initialize
+            scalar_str = scalar_L
+            with dset_L.collective:
+                data_L[:,:,:,:] = dset_L[:,:,cy1:cy2,:].T
+            
+            ## read data R (if != data L)
+            if (scalar_L==scalar_R):
+                data_R[:,:,:,:] = data_L[:,:,:,:]
+            else:
+                n_scalars_read += 1
+                scalar_str += f',{scalar_R}'
+                with dset_R.collective:
+                    data_R[:,:,:,:] = dset_R[:,:,cy1:cy2,:].T
+            
+            ## read ρ
+            if do_density_weighting_L or do_density_weighting_R:
+                n_scalars_read += 1
+                scalar_str += ',ρ'
+                with dset_rho.collective:
+                    rho[:,:,:,:] = dset_rho[:,:,cy1:cy2,:].T
+            else:
+                rho = None
+            
+            self.comm.Barrier()
+            t_delta = timeit.default_timer() - t_start
+            data_gb = n_scalars_read * ( nx * ry * (cy2-cy1) * nz * nt * dset_L.dtype.itemsize ) / 1024**3
+            if verbose:
+                tqdm.write(even_print(f'read: {scalar_str}', '%0.3f [GB]  %0.3f [s]  %0.3f [GB/s]'%(data_gb,t_delta,(data_gb/t_delta)), s=True))
+            
+            ## data_L and data_R should be [nx,nyc,nz,nt] where nyc is the chunk [y] range
+            if ( data_L.shape != (nx,nyc,nz,nt) ) or ( data_R.shape != (nx,nyc,nz,nt) ):
+                print(f'rank {self.rank:d}: shape violation')
+                self.comm.Abort(1)
+            if (rho is not None) and ( rho.shape != (nx,nyc,nz,nt) ):
+                print(f'rank {self.rank:d}: shape violation')
+                self.comm.Abort(1)
+            
+            # === redimensionalize
+            
+            if scalar_L in ['u','v','w',]:
+                data_L *= U_inf
+            elif scalar_L in ['p',]:
+                data_L *= rho_inf * U_inf**2
+            elif scalar_L in ['T',]:
+                data_L *= T_inf
+            elif scalar_L in ['rho',]:
+                data_L *= rho_inf
+            else:
+                raise RuntimeError
+            
+            if scalar_R in ['u','v','w',]:
+                data_R *= U_inf
+            elif scalar_R in ['p',]:
+                data_R *= rho_inf * U_inf**2
+            elif scalar_R in ['T',]:
+                data_R *= T_inf
+            elif scalar_R in ['rho',]:
+                data_R *= rho_inf
+            else:
+                raise RuntimeError
+            
+            if (rho is not None):
+                rho *= rho_inf
+            
+            # === compute mean-removed data
+            
+            ## Reynolds avg(□) or Favre avg(ρ·□)/avg(ρ) in [t]
+            if do_density_weighting_L:
+                rho_avg     = np.mean(        rho , axis=3, dtype=np.float64, keepdims=True) ## [x,y,z,1]
+                data_L_avg  = np.mean( rho*data_L , axis=3, dtype=np.float64, keepdims=True) ## [x,y,z,1]
+                data_L_avg /= rho_avg
+            else:
+                data_L_avg = np.mean( data_L , axis=3, dtype=np.float64, keepdims=True) ## [x,y,z,1]
+            
+            ## Reynolds avg(□) or Favre avg(ρ·□)/avg(ρ) in [t]
+            if do_density_weighting_R:
+                rho_avg     = np.mean(        rho , axis=3, dtype=np.float64, keepdims=True) ## [x,y,z,1]
+                data_R_avg  = np.mean( rho*data_R , axis=3, dtype=np.float64, keepdims=True) ## [x,y,z,1]
+                data_R_avg /= rho_avg
+            else:
+                data_R_avg = np.mean( data_R , axis=3, dtype=np.float64, keepdims=True) ## [x,y,z,1]
+            
+            ## Reynolds prime □′ or Favre prime □″
+            data_L -= data_L_avg
+            data_R -= data_R_avg
+            
+            ## LEFT
+            ## Assert stationarity + definition Re/Fv averaging
+            ## avg(□′)==0 or avg(ρ·□″)==0
+            if do_density_weighting_L:
+                a_ = np.mean(rho*data_L, axis=3, dtype=np.float64, keepdims=True)
+            else:
+                a_ = np.mean(data_L, axis=3, dtype=np.float64, keepdims=True)
+            if not np.allclose( a_, np.zeros_like(a_), atol=1e-6 ):
+                print(f'rank {self.rank:d}: avg(□′)!=0 or avg(ρ·□″)!=0')
+                self.comm.Abort(1)
+            
+            ## RIGHT
+            ## Assert stationarity + definition Re/Fv averaging
+            ## avg(□′)==0 or avg(ρ·□″)==0
+            if do_density_weighting_R:
+                a_ = np.mean(rho*data_R, axis=3, dtype=np.float64, keepdims=True)
+            else:
+                a_ = np.mean(data_R, axis=3, dtype=np.float64, keepdims=True)
+            if not np.allclose( a_, np.zeros_like(a_), atol=1e-6 ):
+                print(f'rank {self.rank:d}: avg(□′)!=0 or avg(ρ·□″)!=0')
+                self.comm.Abort(1)
+            
+            ## covariance: <□′·□′> OR <ρ□″·ρ□″> etc. --> note that this is NOT the typical Favre <ρ·□″□″> except in special cases
+            if do_density_weighting_L and do_density_weighting_R:
+                covariance_ = np.mean( rho*data_L * rho*data_R , axis=3 , dtype=np.float64, keepdims=True)
+            elif do_density_weighting_L and not do_density_weighting_R:
+                covariance_ = np.mean( rho*data_L * data_R , axis=3 , dtype=np.float64, keepdims=True)
+            elif not do_density_weighting_L and do_density_weighting_R:
+                covariance_ = np.mean( data_L * rho*data_R , axis=3 , dtype=np.float64, keepdims=True)
+            elif not do_density_weighting_L and not do_density_weighting_R:
+                covariance_ = np.mean( data_L * data_R , axis=3 , dtype=np.float64, keepdims=True)
+            else:
+                raise RuntimeError
+            
+            ## write this chunk/scalar's covariance to covariance buffer
+            ## avg over [x,z] : [x,y,z,1] --> [y]
+            yiA = cy1 - ry1
+            yiB = cy2 - ry1
+            covariance[scalar][yiA:yiB] = np.squeeze( np.mean( covariance_ , axis=(0,2,3) , dtype=np.float64) )
+            
+            ## write (rank-local) 1D [y] averages
+            if do_density_weighting_L or do_density_weighting_R:
+                avg_Re['rho'][yiA:yiB] = np.squeeze( np.mean( rho_avg    , axis=(0,2,3) , dtype=np.float64) )
+            
+            if do_density_weighting_L:
+                avg_Fv[scalar_L][yiA:yiB] = np.squeeze( np.mean( data_L_avg , axis=(0,2,3) , dtype=np.float64) )
+            else:
+                avg_Re[scalar_L][yiA:yiB] = np.squeeze( np.mean( data_L_avg , axis=(0,2,3) , dtype=np.float64) )
+            
+            if do_density_weighting_R:
+                avg_Fv[scalar_R][yiA:yiB] = np.squeeze( np.mean( data_R_avg , axis=(0,2,3) , dtype=np.float64) )
+            else:
+                avg_Re[scalar_R][yiA:yiB] = np.squeeze( np.mean( data_R_avg , axis=(0,2,3) , dtype=np.float64) )
+            
+            # ===============================================================================
+            # At this point you have 4D [x,y,z,t] □′ or □″ data
+            # ===============================================================================
+            
+            def __fft_z_thread_kernel(xi,ti,yii,do_density_weighting_L,do_density_weighting_R):
+                
+                ## 1D [z] □′ or ρ·□″ vectors
+                if do_density_weighting_L:
+                    uL = rho[xi,yii,:,ti] * data_L[xi,yii,:,ti]
+                else:
+                    uL = data_L[xi,yii,:,ti]
+                
+                if do_density_weighting_R:
+                    uR = rho[xi,yii,:,ti] * data_R[xi,yii,:,ti]
+                else:
+                    uR = data_R[xi,yii,:,ti]
+                
+                ## One-sided amplitude spectra
+                A1 = sp.fft.fft( uL * window_z )[kzp] / nz
+                A2 = sp.fft.fft( uR * window_z )[kzp] / nz
+                
+                #P = 2. * np.real(A1*np.conj(A2)) / ( dkz * mean_sq_win_z )
+                
+                ## One-sided complex cross-spectral density in [kz]
+                P = 2. * A1 * np.conj(A2) / ( dkz * mean_sq_win_z )
+                
+                return xi,ti,P
+            
+            def __fft_t_thread_kernel(xi,zi,yii,do_density_weighting_L,do_density_weighting_R):
+                
+                ## 1D [t] □′ or ρ·□″ vectors
+                if do_density_weighting_L:
+                    uL = rho[xi,yii,zi,:] * data_L[xi,yii,zi,:]
+                else:
+                    uL = data_L[xi,yii,zi,:]
+                
+                if do_density_weighting_R:
+                    uR = rho[xi,yii,zi,:] * data_R[xi,yii,zi,:]
+                else:
+                    uR = data_R[xi,yii,zi,:]
+                
+                ## ## OLD with manual windowing
+                ## uL_, nw, n_pad = get_overlapping_windows(uL, win_len, overlap)
+                ## uR_, nw, n_pad = get_overlapping_windows(uR, win_len, overlap)
+                ## 
+                ## ## STFT buffer
+                ## E_ijk_buf = np.zeros((nw,nf), dtype=np.float64)
+                ## 
+                ## ## compute fft for each overlapped window segment
+                ## for wi in range(nw):
+                ##     A1 = sp.fft.fft( uL_[wi,:] * window_t )[fp] / sum_sqrt_win_t
+                ##     A2 = sp.fft.fft( uR_[wi,:] * window_t )[fp] / sum_sqrt_win_t
+                ##     E_ijk_buf[wi,:] = 2. * np.real(A1*np.conj(A2)) / df
+                ## 
+                ## ## mean across short time FFT (STFT) segments
+                ## E_ijk = np.mean(E_ijk_buf, axis=0, dtype=np.float64)
+                
+                ## One-sided complex cross-spectral density in [f] (using Welch's method)
+                _,P = csd(
+                        uL,uR,
+                        fs=fs,
+                        nperseg=nperseg,
+                        noverlap=noverlap,
+                        window='hann',
+                        detrend='constant',
+                        scaling='density',
+                        return_onesided=True,
+                        )
+                
+                ### Real part = co-spectral density (in-phase contribution to covariance)
+                #P = np.real(P)
+                #return xi,zi,P
+                
+                return xi,zi,P
+            
+            # ===============================================================================
+            
+            ## [y] loop inner ([y] indices within subdivision within rank)
+            for yi in range(cy1,cy2):
+                
+                yii  = yi - cy1 ## chunk local
+                yiii = yi - ry1 ## rank local
+                
+                ## PSD buffers for [y] loop inner
+                E_xt = np.zeros((nx,nt,nkz) , dtype=np.complex128) ## [x,t] range for FFT(z)
+                E_xz = np.zeros((nx,nz,nf ) , dtype=np.complex128) ## [x,z] range for FFT(t)
+                
+                # ===========================================================================
+                # FFT(z) : loop over [x,t]
+                # ===========================================================================
+                
+                ## concurrent/threaded execution for fft(z)
+                tasks = [(xi,ti,yii,do_density_weighting_L,do_density_weighting_R) for xi in range(nx) for ti in range(nt)]
+                with ThreadPoolExecutor(max_workers=n_threads) as executor:
+                    results = executor.map(lambda task: __fft_z_thread_kernel(*task,), tasks)
+                    for xi,ti,P in results:
+                        E_xt[xi,ti,:] = P
+                
+                ## for xi in range(nx):
+                ##     for ti in range(nt):
+                ##         ...
+                
+                ## avg in [x,t] & write in rank context
+                Ekz[scalar][yiii,:] = np.mean(E_xt, axis=(0,1))
+                
+                # ===========================================================================
+                # FFT(t) : loop over [x,z], use windows
+                # ===========================================================================
+                
+                ## concurrent/threaded execution for fft(t)
+                tasks = [(xi,zi,yii,do_density_weighting_L,do_density_weighting_R) for xi in range(nx) for zi in range(nz)]
+                with ThreadPoolExecutor(max_workers=n_threads) as executor:
+                    results = executor.map(lambda task: __fft_t_thread_kernel(*task,), tasks)
+                    for xi,zi,P in results:
+                        E_xz[xi,zi,:] = P
+                
+                ## for xi in range(nx):
+                ##     for zi in range(nz):
+                ##         ...
+                
+                ## avg in [x,z] & write in rank context
+                Ef[scalar][yiii,:] = np.mean(E_xz, axis=(0,1))
+            
+            self.comm.Barrier()
+            t_delta = timeit.default_timer() - t_start
+            if verbose: tqdm.write(even_print(msg, format_time_string(t_delta), s=True))
+            if verbose: progress_bar.update() ## (scalar, [y] chunk) progress
+            
+            #break ## debug --> only do one round in [y] sub-chunk loop
+    
+    if verbose: progress_bar.close()
+    self.comm.Barrier()
+    if verbose: print(72*'-')
+    
+    # ==============================================================
+    # write HDF5 (.h5) file
+    # ==============================================================
+    
+    ## overwrite outfile!
+    ## open on rank 0 and write attributes, dimensions, etc.
+    if (self.rank==0):
+        with h5py.File(fn_h5_out, 'w') as hfw:
+            
+            ## write floats,ints as top-level attributes
+            for key,val in data.items():
+                if isinstance(data[key], (int,np.int32,np.int64)):
+                    hfw.attrs[key] = val
+                elif isinstance(data[key], (float,np.float32,np.float64)):
+                    hfw.attrs[key] = val
+                elif isinstance(data[key], np.ndarray):
+                    pass
+                else:
+                    print(f'key {key} is type {str(type(data[key]))}')
+                    self.comm.Abort(1)
+            
+            ## write numpy arrays
+            hfw.create_dataset( 'dims/x'    , data=x    ) ## [m]
+            hfw.create_dataset( 'dims/y'    , data=y    ) ## [m]
+            hfw.create_dataset( 'dims/z'    , data=z    ) ## [m]
+            hfw.create_dataset( 'dims/t'    , data=t    ) ## [s]
+            hfw.create_dataset( 'dims/freq' , data=freq ) ## [1/s] | [Hz]
+            hfw.create_dataset( 'dims/kz'   , data=kz   ) ## [1/m]
+            hfw.create_dataset( 'dims/lz'   , data=lz   ) ## [m]
+            
+            ## initialize datasets
+            for scalar in scalars:
+                hfw.create_dataset( f'covariance/{scalar}' , shape=(ny,)    , dtype=np.float64    , chunks=None    , data=np.full((ny,),0.,np.float64)       )
+                hfw.create_dataset( f'Ekz/{scalar}'        , shape=(ny,nkz) , dtype=np.complex128 , chunks=(1,nkz) , data=np.full((ny,nkz),0.,np.complex128) )
+                hfw.create_dataset( f'Ef/{scalar}'         , shape=(ny,nf)  , dtype=np.complex128 , chunks=(1,nf)  , data=np.full((ny,nf),0.,np.complex128)  )
+            
+            ## initialize datasets 1D [y] mean
+            for scalar in avg_Re.dtype.names:
+                hfw.create_dataset( f'avg/Re/{scalar}', shape=(ny,), dtype=np.float64, chunks=None, data=np.full((ny,),0.,np.float64) )
+            for scalar in avg_Fv.dtype.names:
+                hfw.create_dataset( f'avg/Fv/{scalar}', shape=(ny,), dtype=np.float64, chunks=None, data=np.full((ny,),0.,np.float64) )
+    
+    self.comm.Barrier()
+    
+    with h5py.File(fn_h5_out, 'a', driver='mpio', comm=self.comm) as hfw:
+        
+        ## collectively write covariance,Ekz,Ef
+        for scalar in scalars:
+            dset = hfw[f'covariance/{scalar}']
+            with dset.collective:
+                dset[ry1:ry2] = covariance[scalar][:]
+            dset = hfw[f'Ekz/{scalar}']
+            with dset.collective:
+                dset[ry1:ry2,:] = Ekz[scalar][:,:]
+            dset = hfw[f'Ef/{scalar}']
+            with dset.collective:
+                dset[ry1:ry2,:] = Ef[scalar][:,:]
+        
+        ## collectively write 1D [y] avgs (Reynolds,Favre)
+        for scalar in avg_Re.dtype.names:
+            dset = hfw[f'avg/Re/{scalar}']
+            with dset.collective:
+                dset[ry1:ry2] = avg_Re[scalar][:]
+        for scalar in avg_Fv.dtype.names:
+            dset = hfw[f'avg/Fv/{scalar}']
+            with dset.collective:
+                dset[ry1:ry2] = avg_Fv[scalar][:]
+    
+    ## report file contents
+    self.comm.Barrier()
+    if (self.rank==0):
+        even_print( os.path.basename(fn_h5_out) , f'{(os.path.getsize(fn_h5_out)/1024**2):0.1f} [MB]' )
+        print(72*'-')
+        with h5py.File(fn_h5_out,'r') as hfr:
+            h5_print_contents(hfr)
+    self.comm.Barrier()
+    
+    if verbose: print(72*'-')
+    if verbose: print('total time : rgd.calc_turb_cospectrum_xpln() : %s'%format_time_string((timeit.default_timer() - t_start_func)))
+    if verbose: print(72*'-')
+    return