def fromfilename(cls, gpu_ctx, filename, cont_write_netcdf=True):
"""
Initialize and hotstart simulation from nc-file.
cont_write_netcdf: Continue to write the results after each superstep to a new netCDF file
filename: Continue simulation based on parameters and last timestep in this file
"""
# open nc-file
sim_reader = SimReader.SimNetCDFReader(filename,
ignore_ghostcells=False)
sim_name = str(sim_reader.get('simulator_short'))
assert sim_name == cls.__name__, \
"Trying to initialize a " + \
cls.__name__ + " simulator with netCDF file based on " \
+ sim_name + " results."
# read parameters
nx = sim_reader.get("nx")
ny = sim_reader.get("ny")
dx = sim_reader.get("dx")
dy = sim_reader.get("dy")
width = nx * dx
height = ny * dy
dt = sim_reader.get("dt")
g = sim_reader.get("g")
r = sim_reader.get("bottom_friction_r")
f = sim_reader.get("coriolis_force")
beta = sim_reader.get("coriolis_beta")
minmodTheta = sim_reader.get("minmod_theta")
timeIntegrator = sim_reader.get("time_integrator")
y_zero_reference_cell = sim_reader.get("y_zero_reference_cell")
try:
wind_stress_type = sim_reader.get("wind_stress_type")
wind = Common.WindStressParams(type=wind_stress_type)
except:
wind = WindStress.WindStress()
boundaryConditions = Common.BoundaryConditions( \
sim_reader.getBC()[0], sim_reader.getBC()[1], \
sim_reader.getBC()[2], sim_reader.getBC()[3], \
sim_reader.getBCSpongeCells())
h0 = sim_reader.getH()
# get last timestep (including simulation time of last timestep)
eta0, hu0, hv0, time0 = sim_reader.getLastTimeStep()
return cls(gpu_ctx, \
h0, eta0, hu0, hv0, \
nx, ny, \
dx, dy, dt, \
g, f, r, \
t=time0, \
wind_stress=wind, \
boundary_conditions=boundaryConditions, \
write_netcdf=cont_write_netcdf)
def WindStressParams(type=99, # "no wind" \
tau0=0, rho=0, alpha=0, xm=0, Rc=0, \
x0=0, y0=0, \
u0=0, v0=0, \
wind_speed=0, wind_direction=0):
"""
Backward compatibility function to avoid rewriting old code and notebooks.
SHOULD NOT BE USED IN NEW CODE! Make WindStress object directly instead.
"""
type_ = np.int32(type)
tau0_ = np.float32(tau0)
rho_ = np.float32(rho)
rho_air_ = np.float32(1.3) # new parameter
alpha_ = np.float32(alpha)
xm_ = np.float32(xm)
Rc_ = np.float32(Rc)
x0_ = np.float32(x0)
y0_ = np.float32(y0)
u0_ = np.float32(u0)
v0_ = np.float32(v0)
wind_speed_ = np.float32(wind_speed)
wind_direction_ = np.float32(wind_direction)
if type == 0:
wind_stress = WindStress.UniformAlongShoreWindStress( \
tau0=tau0_, rho=rho_, alpha=alpha_)
elif type == 1:
wind_stress = WindStress.BellShapedAlongShoreWindStress( \
xm=xm_, tau0=tau0_, rho=rho_, alpha=alpha_)
elif type == 2:
wind_stress = WindStress.MovingCycloneWindStress( \
Rc=Rc_, x0=x0_, y0=y0_, u0=u0_, v0=v0_)
elif type == 50:
wind_stress = WindStress.GenericUniformWindStress( \
rho_air=rho_air_, wind_speed=wind_speed_, wind_direction=wind_direction_)
elif type == 99:
wind_stress = WindStress.NoWindStress()
else:
raise RuntimeError('Invalid wind stress type!')
return wind_stress
def setParameters(self,
f=0,
g=9.81,
beta=0,
r=0,
wind=WindStress.WindStress()):
self.g = g
self.f = f
self.beta = beta
self.r = r
self.wind = wind
def getWind(source_url_list, timestep_indices, timesteps, x0, x1, y0, y1):
"""
timestep_indices => index into netcdf-array, e.g. [1, 3, 5]
timestep => time at timestep, e.g. [1800, 3600, 7200]
"""
if type(source_url_list) is not list:
source_url_list = [source_url_list]
num_files = len(source_url_list)
source_url = source_url_list[0]
assert (num_files == len(timesteps)), str(num_files) + ' vs ' + str(
len(timesteps))
if (timestep_indices is None):
timestep_indices = [None] * num_files
for i in range(num_files):
timestep_indices[i] = range(len(timesteps[i]))
u_wind_list = [None] * num_files
v_wind_list = [None] * num_files
for i in range(num_files):
try:
ncfile = Dataset(source_url_list[i])
u_wind_list[i] = ncfile.variables['Uwind'][timestep_indices[i],
y0:y1, x0:x1]
v_wind_list[i] = ncfile.variables['Vwind'][timestep_indices[i],
y0:y1, x0:x1]
except Exception as e:
raise e
finally:
ncfile.close()
u_wind = u_wind_list[0].filled(0)
v_wind = v_wind_list[0].filled(0)
for i in range(1, num_files):
u_wind = np.concatenate((u_wind, u_wind_list[i].filled(0)))
v_wind = np.concatenate((v_wind, v_wind_list[i].filled(0)))
u_wind = u_wind.astype(np.float32)
v_wind = v_wind.astype(np.float32)
wind_source = WindStress.WindStress(t=np.ravel(timesteps).copy(),
X=u_wind,
Y=v_wind)
return wind_source
def init(self, driftersPerOceanModel=1):
self.windSpeed = 2.0
self.directions = np.random.rand(self.numParticles + 1) * 360
self.windX, self.windY = self.XandYfromDirections(self.directions)
#print "Directions: ", self.directions
self.driftersPerOceanModel = driftersPerOceanModel
self.windT = np.zeros((1), dtype=np.float32)
for i in range(self.numParticles + 1):
wX = [self.windX[i] * np.ones((2, 2), dtype=np.float32)]
wY = [self.windY[i] * np.ones((2, 2), dtype=np.float32)]
wind = WindStress.WindStress(self.windT, wX, wY)
#print ("Init with wind :", (wX, wY))
self.particles[i] = CDKLM16.CDKLM16(self.gpu_ctx, \
self.base_eta, self.base_hu, self.base_hv, \
self.base_H, \
self.nx, self.ny, self.dx, self.dy, self.dt, \
self.g, self.f, self.r, \
wind_stress=wind, \
boundary_conditions=self.boundaryConditions, \
write_netcdf=False)
if i == self.numParticles:
# All particles done, only the observation is left,
# and for the observation we only use one drifter, regardless of the
# number in the other particles.
driftersPerOceanModel = 1
drifters = GPUDrifterCollection.GPUDrifterCollection(
self.gpu_ctx,
driftersPerOceanModel,
observation_variance=self.observation_variance,
boundaryConditions=self.boundaryConditions,
domain_size_x=self.nx * self.dx,
domain_size_y=self.ny * self.dy)
initPos = np.random.multivariate_normal(
self.midPoint, self.initialization_cov_drifters,
driftersPerOceanModel)
drifters.setDrifterPositions(initPos)
#print "drifter particles: ", drifter.getParticlePositions()
#print "drifter observations: ", drifter.getObservationPosition()
self.particles[i].attachDrifters(drifters)
# Put the initial positions into the observation array
self._addObservation(self.observeTrueDrifters())
print("Added init to observation array")
def resample(self, newSampleIndices, reinitialization_variance):
obsTrueDrifter = self.observeTrueDrifters()
positions = self.observeDrifters()
windDirection = self.directions
newWindDirection = np.empty_like(windDirection)
newPos = np.empty((self.driftersPerOceanModel, 2))
newOceanStates = [None] * self.getNumParticles()
for i in range(self.getNumParticles()):
index = newSampleIndices[i]
#print "(particle no, position, old direction, new direction): "
newWindDirection[i] = np.random.normal(windDirection[index],
reinitialization_variance,
1)
if newWindDirection[i] > 360:
newWindDirection[i] -= 360
elif newWindDirection[i] < 0:
newWindDirection[i] += 360
newPos[:, :] = positions[index, :]
#print "\t", (index, positions[index,:], windDirection[index])
#print "\t", (index, newPos, newWindDirection[i])
wX, wY = self.XandYfromDirections(newWindDirection[i])
wX = [wX * np.ones((2, 2), dtype=np.float32)]
wY = [wY * np.ones((2, 2), dtype=np.float32)]
newWindInstance = WindStress.WindStress(self.windT, wX, wY)
# Download index's ocean state:
eta0, hu0, hv0 = self.particles[index].download()
eta1, hu1, hv1 = self.particles[index].downloadPrevTimestep()
newOceanStates[i] = (eta0, hu0, hv0, eta1, hu1, hv1)
self.particles[i].wind_stress = newWindInstance
self.particles[i].drifters.setDrifterPositions(newPos)
self.directions = newWindDirection.copy()
# New loop for transferring the correct ocean states back up to the GPU:
for i in range(self.getNumParticles()):
self.particles[i].upload(newOceanStates[i][0],
newOceanStates[i][1],
newOceanStates[i][2],
newOceanStates[i][3],
newOceanStates[i][4],
newOceanStates[i][5])
def fromfilename(cls, gpu_ctx, filename, cont_write_netcdf=True, use_lcg=False, new_netcdf_filename=None):
"""
Initialize and hotstart simulation from nc-file.
cont_write_netcdf: Continue to write the results after each superstep to a new netCDF file
filename: Continue simulation based on parameters and last timestep in this file
new_netcdf_filename: If we want to continue to write netcdf, we should use this filename. Automatically generated if None.
"""
# open nc-file
sim_reader = SimReader.SimNetCDFReader(filename, ignore_ghostcells=False)
sim_name = str(sim_reader.get('simulator_short'))
assert sim_name == cls.__name__, \
"Trying to initialize a " + \
cls.__name__ + " simulator with netCDF file based on " \
+ sim_name + " results."
# read the most recent state
H = sim_reader.getH();
# get last timestep (including simulation time of last timestep)
eta0, hu0, hv0, time0 = sim_reader.getLastTimeStep()
# For some reason, some old netcdf had 3-dimensional bathymetry.
# This fix ensures that we only use a valid H
if len(H.shape) == 3:
print("norm diff H: ", np.linalg.norm(H[0,:,:] - H[1,:,:]))
H = H[0,:,:]
# Set simulation parameters
sim_params = {
'gpu_ctx': gpu_ctx,
'eta0': eta0,
'hu0': hu0,
'hv0': hv0,
'H': H,
'nx': sim_reader.get("nx"),
'ny': sim_reader.get("ny"),
'dx': sim_reader.get("dx"),
'dy': sim_reader.get("dy"),
'dt': sim_reader.get("dt"),
'g': sim_reader.get("g"),
'f': sim_reader.get("coriolis_force"),
'r': sim_reader.get("bottom_friction_r"),
't': time0,
'theta': sim_reader.get("minmod_theta"),
'rk_order': sim_reader.get("time_integrator"),
'coriolis_beta': sim_reader.get("coriolis_beta"),
'y_zero_reference_cell': sim_reader.get("y_zero_reference_cell"),
'write_netcdf': cont_write_netcdf,
'use_lcg': use_lcg,
'netcdf_filename': new_netcdf_filename
}
# Wind stress
try:
wind_stress_type = sim_reader.get("wind_stress_type")
wind = Common.WindStressParams(type=wind_stress_type)
except:
wind = WindStress.WindStress()
sim_params['wind_stress'] = wind
# Boundary conditions
sim_params['boundary_conditions'] = Common.BoundaryConditions( \
sim_reader.getBC()[0], sim_reader.getBC()[1], \
sim_reader.getBC()[2], sim_reader.getBC()[3], \
sim_reader.getBCSpongeCells())
# Model errors
if sim_reader.has('small_scale_perturbation'):
sim_params['small_scale_perturbation'] = sim_reader.get('small_scale_perturbation') == 'True'
if sim_params['small_scale_perturbation']:
sim_params['small_scale_perturbation_amplitude'] = sim_reader.get('small_scale_perturbation_amplitude')
sim_params['small_scale_perturbation_interpolation_factor'] = sim_reader.get('small_scale_perturbation_interpolation_factor')
# Data assimilation parameters:
if sim_reader.has('model_time_step'):
sim_params['model_time_step'] = sim_reader.get('model_time_step')
return cls(**sim_params)
def __init__(self, \
gpu_ctx, \
eta0, H, hu0, hv0, \
nx, ny, \
dx, dy, dt, \
g, f=0.0, r=0.0, \
t=0.0, \
theta=1.3, use_rk2=True,
coriolis_beta=0.0, \
y_zero_reference_cell = 0, \
wind_stress=WindStress.WindStress(), \
boundary_conditions=Common.BoundaryConditions(), \
write_netcdf=False, \
comm=None, \
ignore_ghostcells=False, \
offset_x=0, offset_y=0, \
flux_slope_eps = 1.0e-1, \
depth_cutoff = 1.0e-5, \
block_width=32, block_height=16):
"""
Initialization routine
eta0: Initial deviation from mean sea level incl ghost cells, (nx+2)*(ny+2) cells
hu0: Initial momentum along x-axis incl ghost cells, (nx+1)*(ny+2) cells
hv0: Initial momentum along y-axis incl ghost cells, (nx+2)*(ny+1) cells
H: Depth from equilibrium defined on cell corners, (nx+5)*(ny+5) corners
nx: Number of cells along x-axis
ny: Number of cells along y-axis
dx: Grid cell spacing along x-axis (20 000 m)
dy: Grid cell spacing along y-axis (20 000 m)
dt: Size of each timestep (90 s)
g: Gravitational accelleration (9.81 m/s^2)
f: Coriolis parameter (1.2e-4 s^1), effectively as f = f + beta*y
r: Bottom friction coefficient (2.4e-3 m/s)
t: Start simulation at time t
theta: MINMOD theta used the reconstructions of the derivatives in the numerical scheme
use_rk2: Boolean if to use 2nd order Runge-Kutta (false -> 1st order forward Euler)
coriolis_beta: Coriolis linear factor -> f = f + beta*(y-y_0)
y_zero_reference_cell: The cell representing y_0 in the above, defined as the lower face of the cell .
wind_stress: Wind stress parameters
boundary_conditions: Boundary condition object
write_netcdf: Write the results after each superstep to a netCDF file
comm: MPI communicator
depth_cutoff: Used for defining dry cells
flux_slope_eps: Used for desingularization with dry cells
"""
ghost_cells_x = 2
ghost_cells_y = 2
y_zero_reference_cell = 2.0 + y_zero_reference_cell
# Boundary conditions
self.boundary_conditions = boundary_conditions
# Extend the computational domain if the boundary conditions
# require it
if (boundary_conditions.isSponge()):
nx = nx + boundary_conditions.spongeCells[
1] + boundary_conditions.spongeCells[3] - 2 * ghost_cells_x
ny = ny + boundary_conditions.spongeCells[
0] + boundary_conditions.spongeCells[2] - 2 * ghost_cells_y
y_zero_reference_cell = boundary_conditions.spongeCells[
2] + y_zero_reference_cell
self.use_rk2 = use_rk2
rk_order = np.int32(use_rk2 + 1)
A = None
super(KP07, self).__init__(gpu_ctx, \
nx, ny, \
ghost_cells_x, \
ghost_cells_y, \
dx, dy, dt, \
g, f, r, A, \
t, \
theta, rk_order, \
coriolis_beta, \
y_zero_reference_cell, \
wind_stress, \
write_netcdf, \
ignore_ghostcells, \
offset_x, offset_y, \
comm, \
block_width, block_height)
# Index range for interior domain (north, east, south, west)
# so that interior domain of eta is
# eta[self.interior_domain_indices[2]:self.interior_domain_indices[0], \
# self.interior_domain_indices[3]:self.interior_domain_indices[1] ]
self.interior_domain_indices = np.array([-2, -2, 2, 2])
self._set_interior_domain_from_sponge_cells()
# The ocean simulators and the swashes cases are defined on
# completely different scales. We therefore specify a different
# desingularization parameter if we run a swashes case.
# Typical values:
#ifndef SWASHES
#define KPSIMULATOR_FLUX_SLOPE_EPS 1e-1f
#define KPSIMULATOR_FLUX_SLOPE_EPS_4 1.0e-4f
#else
#define KPSIMULATOR_FLUX_SLOPE_EPS 1.0e-4f
#define KPSIMULATOR_FLUX_SLOPE_EPS_4 1.0e-16f
#endif
defines = {
'block_width': block_width,
'block_height': block_height,
'KPSIMULATOR_FLUX_SLOPE_EPS': str(flux_slope_eps) + 'f',
'KPSIMULATOR_FLUX_SLOPE_EPS_4': str(flux_slope_eps**4) + 'f',
'KPSIMULATOR_DEPTH_CUTOFF': str(depth_cutoff) + 'f'
}
#Get kernels
self.kp07_kernel = gpu_ctx.get_kernel(
"KP07_kernel.cu",
defines=defines,
compile_args={ # default, fast_math, optimal
'options': [
"--ftz=true", # false, true, true
"--prec-div=false", # true, false, false,
"--prec-sqrt=false", # true, false, false
"--fmad=false"
] # true, true, false
#'options': ["--use_fast_math"]
#'options': ["--generate-line-info"],
#nvcc_options=["--maxrregcount=39"],
#'arch': "compute_50",
#'code': "sm_50"
},
jit_compile_args={
#jit_options=[(cuda.jit_option.MAX_REGISTERS, 39)]
})
# Get CUDA functions and define data types for prepared_{async_}call()
self.swe_2D = self.kp07_kernel.get_function("swe_2D")
self.swe_2D.prepare("iifffffffffiPiPiPiPiPiPiPiPiiiiif")
self.update_wind_stress(self.kp07_kernel, self.swe_2D)
# Upload Bathymetry
self.bathymetry = Common.Bathymetry(self.gpu_ctx, self.gpu_stream, \
nx, ny, ghost_cells_x, ghost_cells_y, H, boundary_conditions)
# Adjust eta for possible dry states
Hm = self.downloadBathymetry()[1]
eta0 = np.maximum(eta0, -Hm)
#Create data by uploading to device
self.gpu_data = Common.SWEDataArakawaA(self.gpu_stream, nx, ny,
ghost_cells_x, ghost_cells_y,
eta0, hu0, hv0)
self.bc_kernel = Common.BoundaryConditionsArakawaA(gpu_ctx, \
self.nx, \
self.ny, \
ghost_cells_x, \
ghost_cells_y, \
self.boundary_conditions)
if self.write_netcdf:
self.sim_writer = SimWriter.SimNetCDFWriter(self, ignore_ghostcells=self.ignore_ghostcells, \
offset_x=self.offset_x, offset_y=self.offset_y)
def __init__(self, \
gpu_ctx, \
eta0, hu0, hv0, H, \
nx, ny, \
dx, dy, dt, \
g, f, r, \
subsample_f=10, \
angle=np.array([[0]], dtype=np.float32), \
subsample_angle=10, \
latitude=None, \
t=0.0, \
theta=1.3, rk_order=2, \
coriolis_beta=0.0, \
max_wind_direction_perturbation = 0, \
wind_stress=WindStress.WindStress(), \
boundary_conditions=Common.BoundaryConditions(), \
boundary_conditions_data=Common.BoundaryConditionsData(), \
small_scale_perturbation=False, \
small_scale_perturbation_amplitude=None, \
small_scale_perturbation_interpolation_factor = 1, \
model_time_step=None,
reportGeostrophicEquilibrium=False, \
use_lcg=False, \
write_netcdf=False, \
comm=None, \
local_particle_id=0, \
super_dir_name=None, \
netcdf_filename=None, \
ignore_ghostcells=False, \
courant_number=0.8, \
offset_x=0, offset_y=0, \
flux_slope_eps = 1.0e-1, \
desingularization_eps = 1.0e-1, \
depth_cutoff = 1.0e-5, \
block_width=12, block_height=32, num_threads_dt=256,
block_width_model_error=16, block_height_model_error=16):
"""
Initialization routine
eta0: Initial deviation from mean sea level incl ghost cells, (nx+2)*(ny+2) cells
hu0: Initial momentum along x-axis incl ghost cells, (nx+1)*(ny+2) cells
hv0: Initial momentum along y-axis incl ghost cells, (nx+2)*(ny+1) cells
H: Depth from equilibrium defined on cell corners, (nx+5)*(ny+5) corners
nx: Number of cells along x-axis
ny: Number of cells along y-axis
dx: Grid cell spacing along x-axis (20 000 m)
dy: Grid cell spacing along y-axis (20 000 m)
dt: Size of each timestep (90 s)
g: Gravitational accelleration (9.81 m/s^2)
f: Coriolis parameter (1.2e-4 s^1), effectively as f = f + beta*y
r: Bottom friction coefficient (2.4e-3 m/s)
subsample_f: Subsample the coriolis f when creating texture by factor
angle: Angle of rotation from North to y-axis as a texture (cuda.Array) or numpy array (in radians)
subsample_angle: Subsample the angles given as input when creating texture by factor
latitude: Specify latitude. This will override any f and beta plane already set (in radians)
t: Start simulation at time t
theta: MINMOD theta used the reconstructions of the derivatives in the numerical scheme
rk_order: Order of Runge Kutta method {1,2*,3}
coriolis_beta: Coriolis linear factor -> f = f + beta*(y-y_0)
max_wind_direction_perturbation: Large-scale model error emulation by per-time-step perturbation of wind direction by +/- max_wind_direction_perturbation (degrees)
wind_stress: Wind stress parameters
boundary_conditions: Boundary condition object
small_scale_perturbation: Boolean value for applying a stochastic model error
small_scale_perturbation_amplitude: Amplitude (q0 coefficient) for model error
small_scale_perturbation_interpolation_factor: Width factor for correlation in model error
model_time_step: The size of a data assimilation model step (default same as dt)
reportGeostrophicEquilibrium: Calculate the Geostrophic Equilibrium variables for each superstep
use_lcg: Use LCG as the random number generator. Default is False, which means using curand.
write_netcdf: Write the results after each superstep to a netCDF file
comm: MPI communicator
local_particle_id: Local (for each MPI process) particle id
desingularization_eps: Used for desingularizing hu/h
flux_slope_eps: Used for setting zero flux for symmetric Riemann fan
depth_cutoff: Used for defining dry cells
super_dir_name: Directory to write netcdf files to
netcdf_filename: Use this filename. (If not defined, a filename will be generated by SimWriter.)
"""
self.logger = logging.getLogger(__name__)
assert (rk_order < 4 or rk_order > 0
), "Only 1st, 2nd and 3rd order Runge Kutta supported"
if (rk_order == 3):
assert (r == 0.0
), "3rd order Runge Kutta supported only without friction"
# Sort out internally represented ghost_cells in the presence of given
# boundary conditions
ghost_cells_x = 2
ghost_cells_y = 2
#Coriolis at "first" cell
x_zero_reference_cell = ghost_cells_x
y_zero_reference_cell = ghost_cells_y # In order to pass it to the super constructor
# Boundary conditions
self.boundary_conditions = boundary_conditions
#Compensate f for reference cell (first cell in internal of domain)
north = np.array([np.sin(angle[0, 0]), np.cos(angle[0, 0])])
f = f - coriolis_beta * (x_zero_reference_cell * dx * north[0] +
y_zero_reference_cell * dy * north[1])
x_zero_reference_cell = 0
y_zero_reference_cell = 0
A = None
self.max_wind_direction_perturbation = max_wind_direction_perturbation
super(CDKLM16, self).__init__(gpu_ctx, \
nx, ny, \
ghost_cells_x, \
ghost_cells_y, \
dx, dy, dt, \
g, f, r, A, \
t, \
theta, rk_order, \
coriolis_beta, \
y_zero_reference_cell, \
wind_stress, \
write_netcdf, \
ignore_ghostcells, \
offset_x, offset_y, \
comm, \
block_width, block_height,
local_particle_id=local_particle_id)
# Index range for interior domain (north, east, south, west)
# so that interior domain of eta is
# eta[self.interior_domain_indices[2]:self.interior_domain_indices[0], \
# self.interior_domain_indices[3]:self.interior_domain_indices[1] ]
self.interior_domain_indices = np.array([-2, -2, 2, 2])
defines = {
'block_width': block_width,
'block_height': block_height,
'KPSIMULATOR_DESING_EPS': "{:.12f}f".format(desingularization_eps),
'KPSIMULATOR_FLUX_SLOPE_EPS': "{:.12f}f".format(flux_slope_eps),
'KPSIMULATOR_DEPTH_CUTOFF': "{:.12f}f".format(depth_cutoff),
'THETA': "{:.12f}f".format(self.theta),
'RK_ORDER': int(self.rk_order),
'NX': int(self.nx),
'NY': int(self.ny),
'DX': "{:.12f}f".format(self.dx),
'DY': "{:.12f}f".format(self.dy),
'GRAV': "{:.12f}f".format(self.g),
'FRIC': "{:.12f}f".format(self.r)
}
#Get kernels
self.kernel = gpu_ctx.get_kernel(
"CDKLM16_kernel.cu",
defines=defines,
compile_args={ # default, fast_math, optimal
'options': [
"--ftz=true", # false, true, true
"--prec-div=false", # true, false, false,
"--prec-sqrt=false", # true, false, false
"--fmad=false"
] # true, true, false
#'options': ["--use_fast_math"]
#'options': ["--generate-line-info"],
#nvcc_options=["--maxrregcount=39"],
#'arch': "compute_50",
#'code': "sm_50"
},
jit_compile_args={
#jit_options=[(cuda.jit_option.MAX_REGISTERS, 39)]
})
# Get CUDA functions and define data types for prepared_{async_}call()
self.cdklm_swe_2D = self.kernel.get_function("cdklm_swe_2D")
self.cdklm_swe_2D.prepare("fiPiPiPiPiPiPiPiPiffi")
self.update_wind_stress(self.kernel, self.cdklm_swe_2D)
# CUDA functions for finding max time step size:
self.num_threads_dt = num_threads_dt
self.num_blocks_dt = np.int32(self.global_size[0] *
self.global_size[1])
self.update_dt_kernels = gpu_ctx.get_kernel("max_dt.cu",
defines={
'block_width':
block_width,
'block_height':
block_height,
'NUM_THREADS':
self.num_threads_dt
})
self.per_block_max_dt_kernel = self.update_dt_kernels.get_function(
"per_block_max_dt")
self.per_block_max_dt_kernel.prepare("iifffPiPiPiPifPi")
self.max_dt_reduction_kernel = self.update_dt_kernels.get_function(
"max_dt_reduction")
self.max_dt_reduction_kernel.prepare("iPP")
# Bathymetry
self.bathymetry = Common.Bathymetry(gpu_ctx, self.gpu_stream, nx, ny,
ghost_cells_x, ghost_cells_y, H,
boundary_conditions)
# Adjust eta for possible dry states
Hm = self.downloadBathymetry()[1]
eta0 = np.maximum(eta0, -Hm)
# Create data by uploading to device
self.gpu_data = Common.SWEDataArakawaA(self.gpu_stream, nx, ny,
ghost_cells_x, ghost_cells_y,
eta0, hu0, hv0)
# Allocate memory for calculating maximum timestep
host_dt = np.zeros((self.global_size[1], self.global_size[0]),
dtype=np.float32)
self.device_dt = Common.CUDAArray2D(self.gpu_stream,
self.global_size[0],
self.global_size[1], 0, 0, host_dt)
host_max_dt_buffer = np.zeros((1, 1), dtype=np.float32)
self.max_dt_buffer = Common.CUDAArray2D(self.gpu_stream, 1, 1, 0, 0,
host_max_dt_buffer)
self.courant_number = courant_number
## Allocating memory for geostrophical equilibrium variables
self.reportGeostrophicEquilibrium = np.int32(
reportGeostrophicEquilibrium)
self.geoEq_uxpvy = None
self.geoEq_Kx = None
self.geoEq_Ly = None
if self.reportGeostrophicEquilibrium:
dummy_zero_array = np.zeros(
(ny + 2 * ghost_cells_y, nx + 2 * ghost_cells_x),
dtype=np.float32,
order='C')
self.geoEq_uxpvy = Common.CUDAArray2D(self.gpu_stream, nx, ny,
ghost_cells_x, ghost_cells_y,
dummy_zero_array)
self.geoEq_Kx = Common.CUDAArray2D(self.gpu_stream, nx, ny,
ghost_cells_x, ghost_cells_y,
dummy_zero_array)
self.geoEq_Ly = Common.CUDAArray2D(self.gpu_stream, nx, ny,
ghost_cells_x, ghost_cells_y,
dummy_zero_array)
self.constant_equilibrium_depth = np.max(H)
self.bc_kernel = Common.BoundaryConditionsArakawaA(gpu_ctx, \
self.nx, \
self.ny, \
ghost_cells_x, \
ghost_cells_y, \
self.boundary_conditions, \
boundary_conditions_data, \
)
def subsample_texture(data, factor):
ny, nx = data.shape
dx, dy = 1 / nx, 1 / ny
I = interp2d(np.linspace(0.5 * dx, 1 - 0.5 * dx, nx),
np.linspace(0.5 * dy, 1 - 0.5 * dy, ny),
data,
kind='linear')
new_nx, new_ny = max(2, nx // factor), max(2, ny // factor)
new_dx, new_dy = 1 / new_nx, 1 / new_ny
x_new = np.linspace(0.5 * new_dx, 1 - 0.5 * new_dx, new_nx)
y_new = np.linspace(0.5 * new_dy, 1 - 0.5 * new_dy, new_ny)
return I(x_new, y_new)
# Texture for angle
self.angle_texref = self.kernel.get_texref("angle_tex")
if isinstance(angle, cuda.Array):
# angle is already a texture, so we just set the texture reference
self.angle_texref.set_array(angle)
else:
#Upload data to GPU and bind to texture reference
if (subsample_angle and angle.size >= eta0.size):
self.logger.info("Subsampling angle texture by factor " +
str(subsample_angle))
self.logger.warning(
"This will give inaccurate angle along the border!")
angle = subsample_texture(angle, subsample_angle)
self.angle_texref.set_array(
cuda.np_to_array(np.ascontiguousarray(angle, dtype=np.float32),
order="C"))
# Set texture parameters
self.angle_texref.set_filter_mode(
cuda.filter_mode.LINEAR) #bilinear interpolation
self.angle_texref.set_address_mode(
0, cuda.address_mode.CLAMP) #no indexing outside domain
self.angle_texref.set_address_mode(1, cuda.address_mode.CLAMP)
self.angle_texref.set_flags(
cuda.TRSF_NORMALIZED_COORDINATES) #Use [0, 1] indexing
# Texture for coriolis f
self.coriolis_texref = self.kernel.get_texref("coriolis_f_tex")
# Create the CPU coriolis
if (latitude is not None):
if (self.f != 0.0):
raise RuntimeError(
"Cannot specify both latitude and f. Make your mind up.")
coriolis_f, _ = OceanographicUtilities.calcCoriolisParams(latitude)
coriolis_f = coriolis_f.astype(np.float32)
else:
if (self.coriolis_beta != 0.0):
if (angle.size != 1):
raise RuntimeError(
"non-constant angle cannot be combined with beta plane model (makes no sense)"
)
#Generate coordinates for all cells, including ghost cells from center to center
# [-3/2dx, nx+3/2dx] for ghost_cells_x == 2
x = np.linspace((-self.ghost_cells_x + 0.5) * self.dx,
(self.nx + self.ghost_cells_x - 0.5) * self.dx,
self.nx + 2 * self.ghost_cells_x)
y = np.linspace((-self.ghost_cells_y + 0.5) * self.dy,
(self.ny + self.ghost_cells_y - 0.5) * self.dy,
self.ny + 2 * self.ghost_cells_x)
self.logger.info(
"Using latitude to create Coriolis f texture ({:f}x{:f} cells)"
.format(x.size, y.size))
x, y = np.meshgrid(x, y)
n = x * np.sin(angle[0, 0]) + y * np.cos(
angle[0, 0]) #North vector
coriolis_f = self.f + self.coriolis_beta * n
else:
if (self.f.size == 1):
coriolis_f = np.array([[self.f]], dtype=np.float32)
elif (self.f.shape == eta0.shape):
coriolis_f = np.array(self.f, dtype=np.float32)
else:
raise RuntimeError(
"The shape of f should match up with eta or be scalar."
)
if (subsample_f and coriolis_f.size >= eta0.size):
self.logger.info("Subsampling coriolis texture by factor " +
str(subsample_f))
self.logger.warning(
"This will give inaccurate coriolis along the border!")
coriolis_f = subsample_texture(coriolis_f, subsample_f)
#Upload data to GPU and bind to texture reference
self.coriolis_texref.set_array(
cuda.np_to_array(np.ascontiguousarray(coriolis_f,
dtype=np.float32),
order="C"))
# Set texture parameters
self.coriolis_texref.set_filter_mode(
cuda.filter_mode.LINEAR) #bilinear interpolation
self.coriolis_texref.set_address_mode(
0, cuda.address_mode.CLAMP) #no indexing outside domain
self.coriolis_texref.set_address_mode(1, cuda.address_mode.CLAMP)
self.coriolis_texref.set_flags(
cuda.TRSF_NORMALIZED_COORDINATES) #Use [0, 1] indexing
# Small scale perturbation:
self.small_scale_perturbation = small_scale_perturbation
self.small_scale_model_error = None
self.small_scale_perturbation_interpolation_factor = small_scale_perturbation_interpolation_factor
if small_scale_perturbation:
self.small_scale_model_error = OceanStateNoise.OceanStateNoise.fromsim(
self,
soar_q0=small_scale_perturbation_amplitude,
interpolation_factor=
small_scale_perturbation_interpolation_factor,
use_lcg=use_lcg,
block_width=block_width_model_error,
block_height=block_height_model_error)
# Data assimilation model step size
self.model_time_step = model_time_step
self.total_time_steps = 0
if model_time_step is None:
self.model_time_step = self.dt
if self.write_netcdf:
self.sim_writer = SimWriter.SimNetCDFWriter(self, super_dir_name=super_dir_name, filename=netcdf_filename, \
ignore_ghostcells=self.ignore_ghostcells, offset_x=self.offset_x, offset_y=self.offset_y)
# Update timestep if dt is given as zero
if self.dt <= 0:
self.updateDt()
def __init__(self, \
gpu_ctx, \
H, eta0, hu0, hv0, \
nx, ny, \
dx, dy, dt, \
g, f, r, \
t=0.0, \
coriolis_beta=0.0, \
y_zero_reference_cell = 1, \
wind_stress=WindStress.WindStress(), \
boundary_conditions=Common.BoundaryConditions(), \
write_netcdf=False, \
comm=None, \
ignore_ghostcells=False, \
offset_x=0, offset_y=0, \
block_width=16, block_height=16):
"""
Initialization routine
H: Water depth incl ghost cells, (nx+2)*(ny+2) cells
eta0: Initial deviation from mean sea level incl ghost cells, (nx+2)*(ny+2) cells
hu0: Initial momentum along x-axis incl ghost cells, (nx+1)*(ny+2) cells
hv0: Initial momentum along y-axis incl ghost cells, (nx+2)*(ny+3) cells
nx: Number of cells along x-axis
ny: Number of cells along y-axis
dx: Grid cell spacing along x-axis (20 000 m)
dy: Grid cell spacing along y-axis (20 000 m)
dt: Size of each timestep (90 s)
g: Gravitational accelleration (9.81 m/s^2)
f: Coriolis parameter (1.2e-4 s^1), effectively as f = f + beta*y
r: Bottom friction coefficient (2.4e-3 m/s)
coriolis_beta: Coriolis linear factor -> f = f + beta*y
y_zero_reference_cell: The cell representing y_0 in the above, defined as the lower face of the cell .
wind_stress: Wind stress parameters
boundary_conditions: Boundary condition object
write_netcdf: Write the results after each superstep to a netCDF file
comm: MPI communicator
"""
#### THIS ALLOWS MAKES IT POSSIBLE TO GIVE THE OLD INPUT SHAPES TO NEW GHOST CELL REGIME: Only valid for benchmarking!
if (eta0.shape == (ny, nx)):
new_eta = np.zeros((ny+2, nx+2), dtype=np.float32)
new_eta[:ny, :nx] = eta0.copy()
eta0 = new_eta.copy()
if (H.shape == (ny, nx)):
new_H = np.ones((ny+2, nx+2), dtype=np.float32)*np.max(H)
new_H[:ny,:nx] = H.copy()
H = new_H.copy()
if (hu0.shape == (ny, nx+1)):
new_hu = np.zeros((ny+2, nx+1), dtype=np.float32)
new_hu[:ny, :nx+1] = hu0.copy()
hu0 = new_hu.copy()
if (hv0.shape == (ny+1, nx)):
new_hv = np.zeros((ny+3, nx+2), dtype=np.float32)
new_hv[:ny+1,:nx] = hv0.copy()
hv0 = new_hv.copy()
#Create data by uploading to device
ghost_cells_x = 1
ghost_cells_y = 1
y_zero_reference_cell = y_zero_reference_cell
# Index range for interior domain (north, east, south, west)
# so that interior domain of eta is
# eta[self.interior_domain_indices[2]:self.interior_domain_indices[0], \
# self.interior_domain_indices[3]:self.interior_domain_indices[1] ]
self.interior_domain_indices = np.array([-1, -1, 1, 1])
self.boundary_conditions = boundary_conditions
if boundary_conditions.isSponge():
nx = nx - 2 + boundary_conditions.spongeCells[1] + boundary_conditions.spongeCells[3]
ny = ny - 2 + boundary_conditions.spongeCells[0] + boundary_conditions.spongeCells[2]
y_zero_reference_cell = y_zero_reference_cell + boundary_conditions.spongeCells[2]
rk_order = None
theta = None
A = None
super(FBL, self).__init__(gpu_ctx, \
nx, ny, \
ghost_cells_x, \
ghost_cells_y, \
dx, dy, dt, \
g, f, r, A, \
t, \
theta, rk_order, \
coriolis_beta, \
y_zero_reference_cell, \
wind_stress, \
write_netcdf, \
ignore_ghostcells, \
offset_x, offset_y, \
comm, \
block_width, block_height)
self._set_interior_domain_from_sponge_cells()
#Get kernels
self.step_kernel = gpu_ctx.get_kernel("FBL_step_kernel.cu",
defines={'block_width': block_width, 'block_height': block_height},
compile_args={
'no_extern_c': True,
'options': ["--use_fast_math"],
#'options': ["--generate-line-info"],
#'options': ["--maxrregcount=32"]
#'arch': "compute_50",
#'code': "sm_50"
},
jit_compile_args={
#jit_options=[(cuda.jit_option.MAX_REGISTERS, 39)]
}
)
# Get CUDA functions
self.fblStepKernel = self.step_kernel.get_function("fblStepKernel")
# Prepare kernel lauches
self.fblStepKernel.prepare("iiffffffffPiPiPiPiif")
# Set up textures
self.update_wind_stress(self.step_kernel, self.fblStepKernel)
self.H = Common.CUDAArray2D(self.gpu_stream, nx, ny, ghost_cells_x, ghost_cells_y, H)
self.gpu_data = Common.SWEDataArakawaC(self.gpu_stream, nx, ny, ghost_cells_x, ghost_cells_y, eta0, hu0, hv0, fbl=True)
# Domain including ghost cells
self.nx_halo = np.int32(nx + 2)
self.ny_halo = np.int32(ny + 2)
self.bc_kernel = FBL_boundary_conditions(self.gpu_ctx, \
self.nx, \
self.ny, \
self.boundary_conditions
)
# Bit-wise boolean for wall boundary conditions
self.wall_bc = np.int32(0)
if (self.boundary_conditions.north == 1):
self.wall_bc = self.wall_bc | 0x01
if (self.boundary_conditions.east == 1):
self.wall_bc = self.wall_bc | 0x02
if (self.boundary_conditions.south == 1):
self.wall_bc = self.wall_bc | 0x04
if (self.boundary_conditions.west == 1):
self.wall_bc = self.wall_bc | 0x08
if self.write_netcdf:
self.sim_writer = SimWriter.SimNetCDFWriter(self, ignore_ghostcells=self.ignore_ghostcells, \
staggered_grid=True, \
offset_x=self.offset_x, offset_y=self.offset_y)
def getWindSourceterm(source_url_list, timestep_indices, timesteps, x0, x1, y0,
y1):
"""
timestep_indices => index into netcdf-array, e.g. [1, 3, 5]
timestep => time at timestep, e.g. [1800, 3600, 7200]
"""
if type(source_url_list) is not list:
source_url_list = [source_url_list]
num_files = len(source_url_list)
source_url = source_url_list[0]
assert (num_files == len(timesteps)), str(num_files) + ' vs ' + str(
len(timesteps))
if (timestep_indices is None):
timestep_indices = [None] * num_files
for i in range(num_files):
timestep_indices[i] = range(len(timesteps[i]))
u_wind_list = [None] * num_files
v_wind_list = [None] * num_files
for i in range(num_files):
try:
ncfile = Dataset(source_url_list[i])
u_wind_list[i] = ncfile.variables['Uwind'][timestep_indices[i],
y0:y1, x0:x1]
v_wind_list[i] = ncfile.variables['Vwind'][timestep_indices[i],
y0:y1, x0:x1]
except Exception as e:
raise e
finally:
ncfile.close()
u_wind = u_wind_list[0].filled(0)
v_wind = v_wind_list[0].filled(0)
for i in range(1, num_files):
u_wind = np.concatenate((u_wind, u_wind_list[i].filled(0)))
v_wind = np.concatenate((v_wind, v_wind_list[i].filled(0)))
wind_speed = np.sqrt(np.power(u_wind, 2) + np.power(v_wind, 2))
# C_drag as defined by Engedahl (1995)
#(See "Documentation of simple ocean models for use in ensemble predictions. Part II: Benchmark cases"
#at https://www.met.no/publikasjoner/met-report/met-report-2012 for details.) /
def computeDrag(wind_speed):
C_drag = np.where(wind_speed < 11, 0.0012,
0.00049 + 0.000065 * wind_speed)
return C_drag
C_drag = computeDrag(wind_speed)
rho_a = 1.225 # Density of air
rho_w = 1025 # Density of water
#Wind stress is then
# tau_s = rho_a * C_drag * |W|W
wind_stress = C_drag * wind_speed * rho_a / rho_w
wind_stress_u = wind_stress * u_wind
wind_stress_v = wind_stress * v_wind
wind_source = WindStress.WindStress(t=np.ravel(timesteps).copy(),
X=wind_stress_u,
Y=wind_stress_v)
return wind_source
def __init__(self, \
gpu_ctx, \
H, eta0, hu0, hv0, \
nx, ny, \
dx, dy, dt, \
g, f, r, A=0.0, \
t=0.0, \
coriolis_beta=0.0, \
y_zero_reference_cell = 0, \
wind_stress=WindStress.WindStress(), \
boundary_conditions=Common.BoundaryConditions(), \
write_netcdf=False, \
ignore_ghostcells=False, \
offset_x=0, offset_y=0, \
block_width=16, block_height=16):
"""
Initialization routine
H: Water depth incl ghost cells, (nx+2)*(ny+2) cells
eta0: Initial deviation from mean sea level incl ghost cells, (nx+2)*(ny+2) cells
hu0: Initial momentum along x-axis incl ghost cells, (nx+1)*(ny+2) cells
hv0: Initial momentum along y-axis incl ghost cells, (nx+2)*(ny+1) cells
nx: Number of cells along x-axis
ny: Number of cells along y-axis
dx: Grid cell spacing along x-axis (20 000 m)
dy: Grid cell spacing along y-axis (20 000 m)
dt: Size of each timestep (90 s)
g: Gravitational accelleration (9.81 m/s^2)
f: Coriolis parameter (1.2e-4 s^1), effectively as f = f + beta*y
r: Bottom friction coefficient (2.4e-3 m/s)
A: Eddy viscosity coefficient (O(dx))
t: Start simulation at time t
coriolis_beta: Coriolis linear factor -> f = f + beta*(y-y_0)
y_zero_reference_cell: The cell representing y_0 in the above, defined as the lower face of the cell .
wind_stress: Wind stress parameters
boundary_conditions: Boundary condition object
write_netcdf: Write the results after each superstep to a netCDF file
"""
# Sort out internally represented ghost_cells in the presence of given
# boundary conditions
halo_x = 1
halo_y = 1
ghost_cells_x = 1
ghost_cells_y = 1
y_zero_reference_cell = y_zero_reference_cell + 1
self.boundary_conditions = boundary_conditions
if boundary_conditions.isSponge():
nx = nx + boundary_conditions.spongeCells[
1] + boundary_conditions.spongeCells[3] - 2 * ghost_cells_x
ny = ny + boundary_conditions.spongeCells[
0] + boundary_conditions.spongeCells[2] - 2 * ghost_cells_y
y_zero_reference_cell = y_zero_reference_cell + boundary_conditions.spongeCells[
2]
# self.<parameters> are sat in parent constructor:
rk_order = None
theta = None
super(CTCS, self).__init__(gpu_ctx, \
nx, ny, \
ghost_cells_x, \
ghost_cells_y, \
dx, dy, dt, \
g, f, r, A, \
t, \
theta, rk_order, \
coriolis_beta, \
y_zero_reference_cell, \
wind_stress, \
write_netcdf, \
ignore_ghostcells, \
offset_x, offset_y, \
block_width, block_height)
# Index range for interior domain (north, east, south, west)
# so that interior domain of eta is
# eta[self.interior_domain_indices[2]:self.interior_domain_indices[0], \
# self.interior_domain_indices[3]:self.interior_domain_indices[1] ]
self.interior_domain_indices = np.array([-1, -1, 1, 1])
self._set_interior_domain_from_sponge_cells()
#Get kernels
self.u_kernel = gpu_ctx.get_kernel("CTCS_U_kernel.cu",
defines={
'block_width': block_width,
'block_height': block_height
})
self.v_kernel = gpu_ctx.get_kernel("CTCS_V_kernel.cu",
defines={
'block_width': block_width,
'block_height': block_height
})
self.eta_kernel = gpu_ctx.get_kernel("CTCS_eta_kernel.cu",
defines={
'block_width': block_width,
'block_height': block_height
})
# Get CUDA functions
self.computeUKernel = self.u_kernel.get_function("computeUKernel")
self.computeVKernel = self.v_kernel.get_function("computeVKernel")
self.computeEtaKernel = self.eta_kernel.get_function(
"computeEtaKernel")
# Prepare kernel lauches
self.computeUKernel.prepare("iiiifffffffffPiPiPiPiPif")
self.computeVKernel.prepare("iiiifffffffffPiPiPiPiPif")
self.computeEtaKernel.prepare("iiffffffffPiPiPi")
# Set up textures
self.update_wind_stress(self.u_kernel, self.computeUKernel)
self.update_wind_stress(self.v_kernel, self.computeVKernel)
#Create data by uploading to device
self.H = Common.CUDAArray2D(self.gpu_stream, nx, ny, halo_x, halo_y, H)
self.gpu_data = Common.SWEDataArakawaC(self.gpu_stream, nx, ny, halo_x,
halo_y, eta0, hu0, hv0)
# Global size needs to be larger than the default from parent.__init__
self.global_size = ( \
int(np.ceil((self.nx+2*halo_x) / float(self.local_size[0]))), \
int(np.ceil((self.ny+2*halo_y) / float(self.local_size[1]))) \
)
self.bc_kernel = CTCS_boundary_condition(gpu_ctx, \
self.nx, \
self.ny, \
self.boundary_conditions, \
halo_x, halo_y \
)
if self.write_netcdf:
self.sim_writer = SimWriter.SimNetCDFWriter(self, ignore_ghostcells=self.ignore_ghostcells, \
staggered_grid=True, offset_x=self.offset_x, offset_y=self.offset_y)
def __init__(self, \
gpu_ctx, \
eta0, Hi, hu0, hv0, \
nx, ny, \
dx, dy, dt, \
g, f=0.0, r=0.0, \
t=0.0, \
theta=1.3, use_rk2=True,
coriolis_beta=0.0, \
y_zero_reference_cell = 0, \
wind_stress=WindStress.WindStress(), \
boundary_conditions=Common.BoundaryConditions(), \
write_netcdf=False, \
ignore_ghostcells=False, \
offset_x=0, offset_y=0, \
block_width=32, block_height=16):
"""
Initialization routine
eta0: Initial deviation from mean sea level incl ghost cells, (nx+2)*(ny+2) cells
hu0: Initial momentum along x-axis incl ghost cells, (nx+1)*(ny+2) cells
hv0: Initial momentum along y-axis incl ghost cells, (nx+2)*(ny+1) cells
Hi: Depth from equilibrium defined on cell corners, (nx+5)*(ny+5) corners
nx: Number of cells along x-axis
ny: Number of cells along y-axis
dx: Grid cell spacing along x-axis (20 000 m)
dy: Grid cell spacing along y-axis (20 000 m)
dt: Size of each timestep (90 s)
g: Gravitational accelleration (9.81 m/s^2)
f: Coriolis parameter (1.2e-4 s^1), effectively as f = f + beta*y
r: Bottom friction coefficient (2.4e-3 m/s)
t: Start simulation at time t
theta: MINMOD theta used the reconstructions of the derivatives in the numerical scheme
use_rk2: Boolean if to use 2nd order Runge-Kutta (false -> 1st order forward Euler)
coriolis_beta: Coriolis linear factor -> f = f + beta*(y-y_0)
y_zero_reference_cell: The cell representing y_0 in the above, defined as the lower face of the cell .
wind_stress: Wind stress parameters
boundary_conditions: Boundary condition object
write_netcdf: Write the results after each superstep to a netCDF file
"""
## After changing from (h, B) to (eta, H), several of the simulator settings used are wrong. This check will help detect that.
if ( np.sum(eta0 - Hi[:-1, :-1] > 0) > nx):
assert(False), "It seems you are using water depth/elevation h and bottom topography B, while you should use water level eta and equillibrium depth H."
ghost_cells_x = 2
ghost_cells_y = 2
y_zero_reference_cell = 2.0 + y_zero_reference_cell
# Boundary conditions
self.boundary_conditions = boundary_conditions
# Extend the computational domain if the boundary conditions
# require it
if (boundary_conditions.isSponge()):
nx = nx + boundary_conditions.spongeCells[1] + boundary_conditions.spongeCells[3] - 2*ghost_cells_x
ny = ny + boundary_conditions.spongeCells[0] + boundary_conditions.spongeCells[2] - 2*ghost_cells_y
y_zero_reference_cell = boundary_conditions.spongeCells[2] + y_zero_reference_cell
self.use_rk2 = use_rk2
rk_order = np.int32(use_rk2 + 1)
A = None
super(KP07, self).__init__(gpu_ctx, \
nx, ny, \
ghost_cells_x, \
ghost_cells_y, \
dx, dy, dt, \
g, f, r, A, \
t, \
theta, rk_order, \
coriolis_beta, \
y_zero_reference_cell, \
wind_stress, \
write_netcdf, \
ignore_ghostcells, \
offset_x, offset_y, \
block_width, block_height)
# Index range for interior domain (north, east, south, west)
# so that interior domain of eta is
# eta[self.interior_domain_indices[2]:self.interior_domain_indices[0], \
# self.interior_domain_indices[3]:self.interior_domain_indices[1] ]
self.interior_domain_indices = np.array([-2,-2,2,2])
self._set_interior_domain_from_sponge_cells()
#Get kernels
self.kp07_kernel = gpu_ctx.get_kernel("KP07_kernel.cu", defines={'block_width': block_width, 'block_height': block_height})
# Get CUDA functions and define data types for prepared_{async_}call()
self.swe_2D = self.kp07_kernel.get_function("swe_2D")
self.swe_2D.prepare("iifffffffffiPiPiPiPiPiPiPiPiiiiif")
self.update_wind_stress(self.kp07_kernel, self.swe_2D)
#Create data by uploading to device
self.gpu_data = Common.SWEDataArakawaA(self.gpu_stream, nx, ny, ghost_cells_x, ghost_cells_y, eta0, hu0, hv0)
#Bathymetry
self.bathymetry = Common.Bathymetry(self.gpu_ctx, self.gpu_stream, \
nx, ny, ghost_cells_x, ghost_cells_y, Hi, boundary_conditions)
self.bc_kernel = Common.BoundaryConditionsArakawaA(gpu_ctx, \
self.nx, \
self.ny, \
ghost_cells_x, \
ghost_cells_y, \
self.boundary_conditions)
if self.write_netcdf:
self.sim_writer = SimWriter.SimNetCDFWriter(self, ignore_ghostcells=self.ignore_ghostcells, \
offset_x=self.offset_x, offset_y=self.offset_y)
def __init__(self, \
gpu_ctx, \
eta0, hu0, hv0, Hi, \
nx, ny, \
dx, dy, dt, \
g, f, r, \
t=0.0, \
theta=1.3, rk_order=2, \
coriolis_beta=0.0, \
y_zero_reference_cell = 0, \
max_wind_direction_perturbation = 0, \
wind_stress=WindStress.WindStress(), \
boundary_conditions=Common.BoundaryConditions(), \
small_scale_perturbation=False, \
small_scale_perturbation_amplitude=None, \
h0AsWaterElevation=False, \
reportGeostrophicEquilibrium=False, \
write_netcdf=False, \
ignore_ghostcells=False, \
offset_x=0, offset_y=0, \
block_width=32, block_height=4):
"""
Initialization routine
eta0: Initial deviation from mean sea level incl ghost cells, (nx+2)*(ny+2) cells
hu0: Initial momentum along x-axis incl ghost cells, (nx+1)*(ny+2) cells
hv0: Initial momentum along y-axis incl ghost cells, (nx+2)*(ny+1) cells
Hi: Depth from equilibrium defined on cell corners, (nx+5)*(ny+5) corners
nx: Number of cells along x-axis
ny: Number of cells along y-axis
dx: Grid cell spacing along x-axis (20 000 m)
dy: Grid cell spacing along y-axis (20 000 m)
dt: Size of each timestep (90 s)
g: Gravitational accelleration (9.81 m/s^2)
f: Coriolis parameter (1.2e-4 s^1), effectively as f = f + beta*y
r: Bottom friction coefficient (2.4e-3 m/s)
t: Start simulation at time t
theta: MINMOD theta used the reconstructions of the derivatives in the numerical scheme
rk_order: Order of Runge Kutta method {1,2*,3}
coriolis_beta: Coriolis linear factor -> f = f + beta*(y-y_0)
y_zero_reference_cell: The cell representing y_0 in the above, defined as the lower face of the cell .
max_wind_direction_perturbation: Large-scale model error emulation by per-time-step perturbation of wind direction by +/- max_wind_direction_perturbation (degrees)
wind_stress: Wind stress parameters
boundary_conditions: Boundary condition object
h0AsWaterElevation: True if h0 is described by the surface elevation, and false if h0 is described by water depth
reportGeostrophicEquilibrium: Calculate the Geostrophic Equilibrium variables for each superstep
write_netcdf: Write the results after each superstep to a netCDF file
"""
## After changing from (h, B) to (eta, H), several of the simulator settings used are wrong. This check will help detect that.
if ( np.sum(eta0 - Hi[:-1, :-1] > 0) > nx):
assert(False), "It seems you are using water depth/elevation h and bottom topography B, while you should use water level eta and equillibrium depth H."
assert( rk_order < 4 or rk_order > 0 ), "Only 1st, 2nd and 3rd order Runge Kutta supported"
if (rk_order == 3):
assert(r == 0.0), "3rd order Runge Kutta supported only without friction"
# Sort out internally represented ghost_cells in the presence of given
# boundary conditions
ghost_cells_x = 2
ghost_cells_y = 2
y_zero_reference_cell = 2 + y_zero_reference_cell
# Boundary conditions
self.boundary_conditions = boundary_conditions
if (boundary_conditions.isSponge()):
nx = nx + boundary_conditions.spongeCells[1] + boundary_conditions.spongeCells[3] - 2*ghost_cells_x
ny = ny + boundary_conditions.spongeCells[0] + boundary_conditions.spongeCells[2] - 2*ghost_cells_y
y_zero_reference_cell = boundary_conditions.spongeCells[2] + y_zero_reference_cell
A = None
self.max_wind_direction_perturbation = max_wind_direction_perturbation
super(CDKLM16, self).__init__(gpu_ctx, \
nx, ny, \
ghost_cells_x, \
ghost_cells_y, \
dx, dy, dt, \
g, f, r, A, \
t, \
theta, rk_order, \
coriolis_beta, \
y_zero_reference_cell, \
wind_stress, \
write_netcdf, \
ignore_ghostcells, \
offset_x, offset_y, \
block_width, block_height)
# Index range for interior domain (north, east, south, west)
# so that interior domain of eta is
# eta[self.interior_domain_indices[2]:self.interior_domain_indices[0], \
# self.interior_domain_indices[3]:self.interior_domain_indices[1] ]
self.interior_domain_indices = np.array([-2,-2,2,2])
self._set_interior_domain_from_sponge_cells()
#Get kernels
self.kernel = gpu_ctx.get_kernel("CDKLM16_kernel.cu", defines={'block_width': block_width, 'block_height': block_height})
# Get CUDA functions and define data types for prepared_{async_}call()
self.swe_2D = self.kernel.get_function("swe_2D")
self.swe_2D.prepare("iifffffffffiiPiPiPiPiPiPiPiPifiiiiiPiPiPi")
self.update_wind_stress(self.kernel, self.swe_2D)
#Create data by uploading to device
self.gpu_data = Common.SWEDataArakawaA(self.gpu_stream, nx, ny, ghost_cells_x, ghost_cells_y, eta0, hu0, hv0)
## Allocating memory for geostrophical equilibrium variables
self.reportGeostrophicEquilibrium = np.int32(reportGeostrophicEquilibrium)
dummy_zero_array = np.zeros((ny+2*ghost_cells_y, nx+2*ghost_cells_x), dtype=np.float32, order='C')
self.geoEq_uxpvy = Common.CUDAArray2D(self.gpu_stream, nx, ny, ghost_cells_x, ghost_cells_y, dummy_zero_array)
self.geoEq_Kx = Common.CUDAArray2D(self.gpu_stream, nx, ny, ghost_cells_x, ghost_cells_y, dummy_zero_array)
self.geoEq_Ly = Common.CUDAArray2D(self.gpu_stream, nx, ny, ghost_cells_x, ghost_cells_y, dummy_zero_array)
#Bathymetry
self.bathymetry = Common.Bathymetry(gpu_ctx, self.gpu_stream, nx, ny, ghost_cells_x, ghost_cells_y, Hi, boundary_conditions)
self.h0AsWaterElevation = h0AsWaterElevation
if self.h0AsWaterElevation:
self.bathymetry.waterElevationToDepth(self.gpu_data.h0)
self.constant_equilibrium_depth = np.max(Hi)
self.bc_kernel = Common.BoundaryConditionsArakawaA(gpu_ctx, \
self.nx, \
self.ny, \
ghost_cells_x, \
ghost_cells_y, \
self.boundary_conditions, \
)
# Small scale perturbation:
self.small_scale_perturbation = small_scale_perturbation
self.small_scale_model_error = None
if small_scale_perturbation:
if small_scale_perturbation_amplitude is None:
self.small_scale_model_error = OceanStateNoise.OceanStateNoise.fromsim(self)
else:
self.small_scale_model_error = OceanStateNoise.OceanStateNoise.fromsim(self, soar_q0=small_scale_perturbation_amplitude)
if self.write_netcdf:
self.sim_writer = SimWriter.SimNetCDFWriter(self, ignore_ghostcells=self.ignore_ghostcells, \
offset_x=self.offset_x, offset_y=self.offset_y)
def step(self, t_end=0.0, apply_stochastic_term=True, write_now=True, update_dt=False):
"""
Function which steps n timesteps.
apply_stochastic_term: Boolean value for whether the stochastic
perturbation (if any) should be applied.
"""
if self.t == 0:
self.bc_kernel.update_bc_values(self.gpu_stream, self.t)
self.bc_kernel.boundaryCondition(self.gpu_stream, \
self.gpu_data.h0, self.gpu_data.hu0, self.gpu_data.hv0)
t_now = 0.0
while (t_now < t_end):
#for i in range(0, n):
# Get new random wind direction (emulationg large-scale model error)
if(self.max_wind_direction_perturbation > 0.0 and self.wind_stress.type() == 1):
# max perturbation +/- max_wind_direction_perturbation deg within original wind direction (at t=0)
perturbation = 2.0*(np.random.rand()-0.5) * self.max_wind_direction_perturbation;
new_wind_stress = WindStress.GenericUniformWindStress( \
rho_air=self.wind_stress.rho_air, \
wind_speed=self.wind_stress.wind_speed, \
wind_direction=self.wind_stress.wind_direction + perturbation)
# Upload new wind stress params to device
cuda.memcpy_htod_async(int(self.wind_stress_dev), new_wind_stress.tostruct(), stream=self.gpu_stream)
# Calculate dt if using automatic dt
if (self.dt <= 0 or update_dt):
self.updateDt()
local_dt = np.float32(min(self.dt, np.float32(t_end - t_now)))
wind_stress_t = np.float32(self.update_wind_stress(self.kernel, self.cdklm_swe_2D))
self.bc_kernel.update_bc_values(self.gpu_stream, self.t)
#self.bc_kernel.boundaryCondition(self.cl_queue, \
# self.gpu_data.h1, self.gpu_data.hu1, self.gpu_data.hv1)
# 2nd order Runge Kutta
if (self.rk_order == 2):
self.callKernel(self.gpu_data.h0, self.gpu_data.hu0, self.gpu_data.hv0, \
self.gpu_data.h1, self.gpu_data.hu1, self.gpu_data.hv1, \
local_dt, wind_stress_t, 0)
self.bc_kernel.boundaryCondition(self.gpu_stream, \
self.gpu_data.h1, self.gpu_data.hu1, self.gpu_data.hv1)
self.callKernel(self.gpu_data.h1, self.gpu_data.hu1, self.gpu_data.hv1, \
self.gpu_data.h0, self.gpu_data.hu0, self.gpu_data.hv0, \
local_dt, wind_stress_t, 1)
# Applying final boundary conditions after perturbation (if applicable)
elif (self.rk_order == 1):
self.callKernel(self.gpu_data.h0, self.gpu_data.hu0, self.gpu_data.hv0, \
self.gpu_data.h1, self.gpu_data.hu1, self.gpu_data.hv1, \
local_dt, wind_stress_t, 0)
self.gpu_data.swap()
# Applying boundary conditions after perturbation (if applicable)
# 3rd order RK method:
elif (self.rk_order == 3):
self.callKernel(self.gpu_data.h0, self.gpu_data.hu0, self.gpu_data.hv0, \
self.gpu_data.h1, self.gpu_data.hu1, self.gpu_data.hv1, \
local_dt, wind_stress_t, 0)
self.bc_kernel.boundaryCondition(self.gpu_stream, \
self.gpu_data.h1, self.gpu_data.hu1, self.gpu_data.hv1)
self.callKernel(self.gpu_data.h1, self.gpu_data.hu1, self.gpu_data.hv1, \
self.gpu_data.h0, self.gpu_data.hu0, self.gpu_data.hv0, \
local_dt, wind_stress_t, 1)
self.bc_kernel.boundaryCondition(self.gpu_stream, \
self.gpu_data.h1, self.gpu_data.hu1, self.gpu_data.hv1)
self.callKernel(self.gpu_data.h1, self.gpu_data.hu1, self.gpu_data.hv1, \
self.gpu_data.h0, self.gpu_data.hu0, self.gpu_data.hv0, \
local_dt, wind_stress_t, 2)
# Applying final boundary conditions after perturbation (if applicable)
# Perturb ocean state with model error
if self.small_scale_perturbation and apply_stochastic_term:
self.small_scale_model_error.perturbSim(self)
# Apply boundary conditions
self.bc_kernel.boundaryCondition(self.gpu_stream, \
self.gpu_data.h0, self.gpu_data.hu0, self.gpu_data.hv0)
# Evolve drifters
if self.hasDrifters:
self.drifters.drift(self.gpu_data.h0, self.gpu_data.hu0, \
self.gpu_data.hv0, \
np.float32(self.constant_equilibrium_depth), \
self.nx, self.ny, self.dx, self.dy, \
local_dt, \
np.int32(2), np.int32(2))
self.t += np.float64(local_dt)
t_now += np.float64(local_dt)
self.num_iterations += 1
if self.write_netcdf and write_now:
self.sim_writer.writeTimestep(self)
return self.t
def __init__(self, \
gpu_ctx, \
eta0, hu0, hv0, H, \
nx, ny, \
dx, dy, dt, \
g, f, r, \
angle=np.array([[0]], dtype=np.float32), \
t=0.0, \
theta=1.3, rk_order=2, \
coriolis_beta=0.0, \
max_wind_direction_perturbation = 0, \
wind_stress=WindStress.WindStress(), \
boundary_conditions=Common.BoundaryConditions(), \
boundary_conditions_data=Common.BoundaryConditionsData(), \
small_scale_perturbation=False, \
small_scale_perturbation_amplitude=None, \
small_scale_perturbation_interpolation_factor = 1, \
model_time_step=None,
reportGeostrophicEquilibrium=False, \
use_lcg=False, \
write_netcdf=False, \
comm=None, \
netcdf_filename=None, \
ignore_ghostcells=False, \
courant_number=0.8, \
offset_x=0, offset_y=0, \
flux_slope_eps = 1.0e-1, \
desingularization_eps = 1.0e-1, \
depth_cutoff = 1.0e-5, \
block_width=32, block_height=8, num_threads_dt=256,
block_width_model_error=16, block_height_model_error=16):
"""
Initialization routine
eta0: Initial deviation from mean sea level incl ghost cells, (nx+2)*(ny+2) cells
hu0: Initial momentum along x-axis incl ghost cells, (nx+1)*(ny+2) cells
hv0: Initial momentum along y-axis incl ghost cells, (nx+2)*(ny+1) cells
H: Depth from equilibrium defined on cell corners, (nx+5)*(ny+5) corners
nx: Number of cells along x-axis
ny: Number of cells along y-axis
dx: Grid cell spacing along x-axis (20 000 m)
dy: Grid cell spacing along y-axis (20 000 m)
dt: Size of each timestep (90 s)
g: Gravitational accelleration (9.81 m/s^2)
f: Coriolis parameter (1.2e-4 s^1), effectively as f = f + beta*y
r: Bottom friction coefficient (2.4e-3 m/s)
angle: Angle of rotation from North to y-axis
t: Start simulation at time t
theta: MINMOD theta used the reconstructions of the derivatives in the numerical scheme
rk_order: Order of Runge Kutta method {1,2*,3}
coriolis_beta: Coriolis linear factor -> f = f + beta*(y-y_0)
max_wind_direction_perturbation: Large-scale model error emulation by per-time-step perturbation of wind direction by +/- max_wind_direction_perturbation (degrees)
wind_stress: Wind stress parameters
boundary_conditions: Boundary condition object
small_scale_perturbation: Boolean value for applying a stochastic model error
small_scale_perturbation_amplitude: Amplitude (q0 coefficient) for model error
small_scale_perturbation_interpolation_factor: Width factor for correlation in model error
model_time_step: The size of a data assimilation model step (default same as dt)
reportGeostrophicEquilibrium: Calculate the Geostrophic Equilibrium variables for each superstep
use_lcg: Use LCG as the random number generator. Default is False, which means using curand.
write_netcdf: Write the results after each superstep to a netCDF file
comm: MPI communicator
desingularization_eps: Used for desingularizing hu/h
flux_slope_eps: Used for setting zero flux for symmetric Riemann fan
depth_cutoff: Used for defining dry cells
netcdf_filename: Use this filename. (If not defined, a filename will be generated by SimWriter.)
"""
self.logger = logging.getLogger(__name__)
assert( rk_order < 4 or rk_order > 0 ), "Only 1st, 2nd and 3rd order Runge Kutta supported"
if (rk_order == 3):
assert(r == 0.0), "3rd order Runge Kutta supported only without friction"
# Sort out internally represented ghost_cells in the presence of given
# boundary conditions
ghost_cells_x = 2
ghost_cells_y = 2
#Coriolis at "first" cell
x_zero_reference_cell = ghost_cells_x
y_zero_reference_cell = ghost_cells_y # In order to pass it to the super constructor
# Boundary conditions
self.boundary_conditions = boundary_conditions
if (boundary_conditions.isSponge()):
nx = nx + boundary_conditions.spongeCells[1] + boundary_conditions.spongeCells[3] - 2*ghost_cells_x
ny = ny + boundary_conditions.spongeCells[0] + boundary_conditions.spongeCells[2] - 2*ghost_cells_y
x_zero_reference_cell += boundary_conditions.spongeCells[3]
y_zero_reference_cell += boundary_conditions.spongeCells[2]
#Compensate f for reference cell (first cell in internal of domain)
north = np.array([np.sin(angle[0,0]), np.cos(angle[0,0])])
f = f - coriolis_beta * (x_zero_reference_cell*dx*north[0] + y_zero_reference_cell*dy*north[1])
x_zero_reference_cell = 0
y_zero_reference_cell = 0
A = None
self.max_wind_direction_perturbation = max_wind_direction_perturbation
super(CDKLM16, self).__init__(gpu_ctx, \
nx, ny, \
ghost_cells_x, \
ghost_cells_y, \
dx, dy, dt, \
g, f, r, A, \
t, \
theta, rk_order, \
coriolis_beta, \
y_zero_reference_cell, \
wind_stress, \
write_netcdf, \
ignore_ghostcells, \
offset_x, offset_y, \
comm, \
block_width, block_height)
# Index range for interior domain (north, east, south, west)
# so that interior domain of eta is
# eta[self.interior_domain_indices[2]:self.interior_domain_indices[0], \
# self.interior_domain_indices[3]:self.interior_domain_indices[1] ]
self.interior_domain_indices = np.array([-2,-2,2,2])
self._set_interior_domain_from_sponge_cells()
defines={'block_width': block_width, 'block_height': block_height,
'KPSIMULATOR_DESING_EPS': str(desingularization_eps)+'f',
'KPSIMULATOR_FLUX_SLOPE_EPS': str(flux_slope_eps)+'f',
'KPSIMULATOR_DEPTH_CUTOFF': str(depth_cutoff)+'f'}
#Get kernels
self.kernel = gpu_ctx.get_kernel("CDKLM16_kernel.cu",
defines=defines,
compile_args={ # default, fast_math, optimal
'options' : ["--ftz=true", # false, true, true
"--prec-div=false", # true, false, false,
"--prec-sqrt=false", # true, false, false
"--fmad=false"] # true, true, false
#'options': ["--use_fast_math"]
#'options': ["--generate-line-info"],
#nvcc_options=["--maxrregcount=39"],
#'arch': "compute_50",
#'code': "sm_50"
},
jit_compile_args={
#jit_options=[(cuda.jit_option.MAX_REGISTERS, 39)]
}
)
# Get CUDA functions and define data types for prepared_{async_}call()
self.cdklm_swe_2D = self.kernel.get_function("cdklm_swe_2D")
self.cdklm_swe_2D.prepare("iiffffffffiiPiPiPiPiPiPiPiPiffi")
self.update_wind_stress(self.kernel, self.cdklm_swe_2D)
# CUDA functions for finding max time step size:
self.num_threads_dt = num_threads_dt
self.num_blocks_dt = np.int32(self.global_size[0]*self.global_size[1])
self.update_dt_kernels = gpu_ctx.get_kernel("max_dt.cu",
defines={'block_width': block_width,
'block_height': block_height,
'NUM_THREADS': self.num_threads_dt})
self.per_block_max_dt_kernel = self.update_dt_kernels.get_function("per_block_max_dt")
self.per_block_max_dt_kernel.prepare("iifffPiPiPiPifPi")
self.max_dt_reduction_kernel = self.update_dt_kernels.get_function("max_dt_reduction")
self.max_dt_reduction_kernel.prepare("iPP")
# Bathymetry
self.bathymetry = Common.Bathymetry(gpu_ctx, self.gpu_stream, nx, ny, ghost_cells_x, ghost_cells_y, H, boundary_conditions)
# Adjust eta for possible dry states
Hm = self.downloadBathymetry()[1]
eta0 = np.maximum(eta0, -Hm)
# Create data by uploading to device
self.gpu_data = Common.SWEDataArakawaA(self.gpu_stream, nx, ny, ghost_cells_x, ghost_cells_y, eta0, hu0, hv0)
# Allocate memory for calculating maximum timestep
host_dt = np.zeros((self.global_size[1], self.global_size[0]), dtype=np.float32)
self.device_dt = Common.CUDAArray2D(self.gpu_stream, self.global_size[0], self.global_size[1],
0, 0, host_dt)
host_max_dt_buffer = np.zeros((1,1), dtype=np.float32)
self.max_dt_buffer = Common.CUDAArray2D(self.gpu_stream, 1, 1, 0, 0, host_max_dt_buffer)
self.courant_number = courant_number
## Allocating memory for geostrophical equilibrium variables
self.reportGeostrophicEquilibrium = np.int32(reportGeostrophicEquilibrium)
self.geoEq_uxpvy = None
self.geoEq_Kx = None
self.geoEq_Ly = None
if self.reportGeostrophicEquilibrium:
dummy_zero_array = np.zeros((ny+2*ghost_cells_y, nx+2*ghost_cells_x), dtype=np.float32, order='C')
self.geoEq_uxpvy = Common.CUDAArray2D(self.gpu_stream, nx, ny, ghost_cells_x, ghost_cells_y, dummy_zero_array)
self.geoEq_Kx = Common.CUDAArray2D(self.gpu_stream, nx, ny, ghost_cells_x, ghost_cells_y, dummy_zero_array)
self.geoEq_Ly = Common.CUDAArray2D(self.gpu_stream, nx, ny, ghost_cells_x, ghost_cells_y, dummy_zero_array)
self.constant_equilibrium_depth = np.max(H)
self.bc_kernel = Common.BoundaryConditionsArakawaA(gpu_ctx, \
self.nx, \
self.ny, \
ghost_cells_x, \
ghost_cells_y, \
self.boundary_conditions, \
boundary_conditions_data, \
)
# Small scale perturbation:
self.small_scale_perturbation = small_scale_perturbation
self.small_scale_model_error = None
self.small_scale_perturbation_interpolation_factor = small_scale_perturbation_interpolation_factor
if small_scale_perturbation:
if small_scale_perturbation_amplitude is None:
self.small_scale_model_error = OceanStateNoise.OceanStateNoise.fromsim(self,
interpolation_factor=small_scale_perturbation_interpolation_factor,
use_lcg=use_lcg,
block_width=block_width_model_error,
block_height=block_height_model_error)
else:
self.small_scale_model_error = OceanStateNoise.OceanStateNoise.fromsim(self,
soar_q0=small_scale_perturbation_amplitude,
interpolation_factor=small_scale_perturbation_interpolation_factor,
use_lcg=use_lcg,
block_width=block_width_model_error,
block_height=block_height_model_error)
# Data assimilation model step size
self.model_time_step = model_time_step
if model_time_step is None:
self.model_time_step = self.dt
self.total_time_steps = 0
if self.write_netcdf:
self.sim_writer = SimWriter.SimNetCDFWriter(self, filename=netcdf_filename, ignore_ghostcells=self.ignore_ghostcells, \
offset_x=self.offset_x, offset_y=self.offset_y)
#Upload data to GPU and bind to texture reference
self.angle_texref = self.kernel.get_texref("angle_tex")
self.angle_texref.set_array(cuda.np_to_array(np.ascontiguousarray(angle, dtype=np.float32), order="C"))
# Set texture parameters
self.angle_texref.set_filter_mode(cuda.filter_mode.LINEAR) #bilinear interpolation
self.angle_texref.set_address_mode(0, cuda.address_mode.CLAMP) #no indexing outside domain
self.angle_texref.set_address_mode(1, cuda.address_mode.CLAMP)
self.angle_texref.set_flags(cuda.TRSF_NORMALIZED_COORDINATES) #Use [0, 1] indexing
def __init__(self, \
gpu_ctx, \
H, eta0, hu0, hv0, \
nx, ny, \
dx, dy, dt, \
g, f, r, \
t=0.0, \
coriolis_beta=0.0, \
y_zero_reference_cell = 0, \
wind_stress=WindStress.WindStress(), \
boundary_conditions=Common.BoundaryConditions(), \
write_netcdf=False, \
ignore_ghostcells=False, \
offset_x=0, offset_y=0, \
block_width=16, block_height=16):
"""
Initialization routine
H: Water depth incl ghost cells, (nx+2)*(ny+2) cells
eta0: Initial deviation from mean sea level incl ghost cells, (nx+2)*(ny+2) cells
hu0: Initial momentum along x-axis incl ghost cells, (nx+1)*(ny+2) cells
hv0: Initial momentum along y-axis incl ghost cells, (nx+2)*(ny+1) cells
nx: Number of cells along x-axis
ny: Number of cells along y-axis
dx: Grid cell spacing along x-axis (20 000 m)
dy: Grid cell spacing along y-axis (20 000 m)
dt: Size of each timestep (90 s)
g: Gravitational accelleration (9.81 m/s^2)
f: Coriolis parameter (1.2e-4 s^1), effectively as f = f + beta*y
r: Bottom friction coefficient (2.4e-3 m/s)
coriolis_beta: Coriolis linear factor -> f = f + beta*y
y_zero_reference_cell: The cell representing y_0 in the above, defined as the lower face of the cell .
wind_stress: Wind stress parameters
boundary_conditions: Boundary condition object
write_netcdf: Write the results after each superstep to a netCDF file
"""
#Create data by uploading to device
ghost_cells_x = 0
ghost_cells_y = 0
y_zero_reference_cell = y_zero_reference_cell
self.asym_ghost_cells = [0, 0, 0, 0] # [N, E, S, W]
# Index range for interior domain (north, east, south, west)
# so that interior domain of eta is
# eta[self.interior_domain_indices[2]:self.interior_domain_indices[0], \
# self.interior_domain_indices[3]:self.interior_domain_indices[1] ]
self.interior_domain_indices = np.array([None, None, 0, 0])
self.boundary_conditions = boundary_conditions
# Add asym ghost cell if periodic boundary condition:
if (self.boundary_conditions.north == 2) or \
(self.boundary_conditions.south == 2):
self.asym_ghost_cells[0] = 1
self.interior_domain_indices[0] = -1
if (self.boundary_conditions.east == 2) or \
(self.boundary_conditions.west == 2):
self.asym_ghost_cells[1] = 1
self.interior_domain_indices[1] = -1
if boundary_conditions.isSponge():
nx = nx + boundary_conditions.spongeCells[
1] + boundary_conditions.spongeCells[
3] # - self.asym_ghost_cells[1] - self.asym_ghost_cells[3]
ny = ny + boundary_conditions.spongeCells[
0] + boundary_conditions.spongeCells[
2] # - self.asym_ghost_cells[0] - self.asym_ghost_cells[2]
y_zero_reference_cell = y_zero_reference_cell + boundary_conditions.spongeCells[
2]
rk_order = None
theta = None
A = None
super(FBL, self).__init__(gpu_ctx, \
nx, ny, \
ghost_cells_x, \
ghost_cells_y, \
dx, dy, dt, \
g, f, r, A, \
t, \
theta, rk_order, \
coriolis_beta, \
y_zero_reference_cell, \
wind_stress, \
write_netcdf, \
ignore_ghostcells, \
offset_x, offset_y, \
block_width, block_height)
self._set_interior_domain_from_sponge_cells()
#Get kernels
self.u_kernel = gpu_ctx.get_kernel("FBL_U_kernel.cu",
defines={
'block_width': block_width,
'block_height': block_height
})
self.v_kernel = gpu_ctx.get_kernel("FBL_V_kernel.cu",
defines={
'block_width': block_width,
'block_height': block_height
})
self.eta_kernel = gpu_ctx.get_kernel("FBL_eta_kernel.cu",
defines={
'block_width': block_width,
'block_height': block_height
})
# Get CUDA functions
self.computeUKernel = self.u_kernel.get_function("computeUKernel")
self.computeVKernel = self.v_kernel.get_function("computeVKernel")
self.computeEtaKernel = self.eta_kernel.get_function(
"computeEtaKernel")
# Prepare kernel lauches
self.computeUKernel.prepare("iiffffffffPiPiPiPif")
self.computeVKernel.prepare("iiffffffffPiPiPiPif")
self.computeEtaKernel.prepare("iiffffffffPiPiPiPi")
# Set up textures
self.update_wind_stress(self.u_kernel, self.computeUKernel)
self.update_wind_stress(self.v_kernel, self.computeVKernel)
self.H = Common.CUDAArray2D(self.gpu_stream, nx, ny, ghost_cells_x,
ghost_cells_y, H, self.asym_ghost_cells)
self.gpu_data = Common.SWEDataArakawaC(self.gpu_stream, nx, ny,
ghost_cells_x, ghost_cells_y,
eta0, hu0, hv0,
self.asym_ghost_cells)
# Overwrite halo with asymetric ghost cells
self.nx_halo = np.int32(nx + self.asym_ghost_cells[1] +
self.asym_ghost_cells[3])
self.ny_halo = np.int32(ny + self.asym_ghost_cells[0] +
self.asym_ghost_cells[2])
self.bc_kernel = FBL_periodic_boundary(self.gpu_ctx, \
self.nx, \
self.ny, \
self.boundary_conditions, \
self.asym_ghost_cells
)
self.totalNumIterations = 0
if self.write_netcdf:
self.sim_writer = SimWriter.SimNetCDFWriter(self, ignore_ghostcells=self.ignore_ghostcells, \
staggered_grid=True, offset_x=self.offset_x, offset_y=self.offset_y)
def __init__(self, \
cl_ctx, \
h0, hu0, hv0, \
Bi, \
nx, ny, \
dx, dy, dt, \
g, f, r, \
theta=1.3, use_rk2=True, \
wind_stress=WindStress.NoWindStress(), \
boundary_conditions=Common.BoundaryConditions(), \
h0AsWaterElevation=True, \
block_width=16, block_height=16):
print("Using RECURSIVE CDKLM scheme!")
self.cl_ctx = cl_ctx
#Create an OpenCL command queue
self.cl_queue = cl.CommandQueue(self.cl_ctx)
#Get kernels
self.kernel = Common.get_kernel(self.cl_ctx, "recursiveCDKLM16_kernel.opencl", defines={'block_width': block_width, 'block_height': block_height})
# Boundary Conditions
self.boundary_conditions = boundary_conditions
self.boundaryType = np.int32(1)
if (boundary_conditions.north == 2 and boundary_conditions.east == 2):
self.boundaryType = np.int32(2)
elif (boundary_conditions.north == 2):
self.boundaryType = np.int32(3)
elif (boundary_conditions.east == 2):
self.boundaryType = np.int32(4)
#Create data by uploading to device
ghost_cells_x = 3
ghost_cells_y = 3
self.ghost_cells_x = 3
self.ghost_cells_y = 3
if (boundary_conditions.isSponge()):
nx = nx + boundary_conditions.spongeCells[1] + boundary_conditions.spongeCells[3] - 2*self.ghost_cells_x
ny = ny + boundary_conditions.spongeCells[0] + boundary_conditions.spongeCells[2] - 2*self.ghost_cells_y
y_zero_reference = boundary_conditions.spongeCells[2]
#Create data by uploading to device
self.cl_data = Common.SWEDataArakawaA(self.cl_ctx, nx, ny, ghost_cells_x, ghost_cells_y, h0, hu0, hv0)
#Bathymetry
self.bathymetry = Common.Bathymetry(self.cl_ctx, self.cl_queue, nx, ny, ghost_cells_x, ghost_cells_y, Bi, boundary_conditions)
#Save input parameters
#Notice that we need to specify them in the correct dataformat for the
#OpenCL kernel
self.nx = np.int32(nx)
self.ny = np.int32(ny)
self.dx = np.float32(dx)
self.dy = np.float32(dy)
self.dt = np.float32(dt)
self.g = np.float32(g)
self.f = np.float32(f)
self.r = np.float32(r)
self.theta = np.float32(theta)
self.use_rk2 = use_rk2
self.wind_stress = wind_stress
self.h0AsWaterElevation = h0AsWaterElevation
#Initialize time
self.t = np.float32(0.0)
#Compute kernel launch parameters
self.local_size = (block_width, block_height)
self.global_size = ( \
int(np.ceil(self.nx / float(self.local_size[0])) * self.local_size[0]), \
int(np.ceil(self.ny / float(self.local_size[1])) * self.local_size[1]) \
)
self.bc_kernel = Common.BoundaryConditionsArakawaA(self.cl_ctx, \
self.nx, \
self.ny, \
ghost_cells_x, \
ghost_cells_y, \
self.boundary_conditions, \
)
if self.h0AsWaterElevation:
self.bathymetry.waterElevationToDepth(self.cl_data.h0)
def __init__(self, gpu_ctx, numDrifters, \
observation_variance=0.01, wind = WindStress.WindStress(),
wind_drift_factor = 0.0,\
boundaryConditions=Common.BoundaryConditions(), \
initialization_cov_drifters=None, \
domain_size_x=1.0, domain_size_y=1.0, \
gpu_stream=None, \
block_width = 64):
super(GPUDrifterCollection,
self).__init__(numDrifters,
observation_variance=observation_variance,
boundaryConditions=boundaryConditions,
domain_size_x=domain_size_x,
domain_size_y=domain_size_y)
# Define CUDA environment:
self.gpu_ctx = gpu_ctx
self.block_width = block_width
self.block_height = 1
self.wind = wind
self.wind_drift_factor = np.float32(wind_drift_factor)
# TODO: Where should the cl_queue come from?
# For sure, the drifter and the ocean simulator should use
# the same queue...
self.gpu_stream = gpu_stream
if self.gpu_stream is None:
self.gpu_stream = cuda.Stream()
self.sensitivity = 1.0
self.driftersHost = np.zeros(
(self.getNumDrifters() + 1, 2)).astype(np.float32, order='C')
self.driftersDevice = Common.CUDAArray2D(self.gpu_stream, \
2, self.getNumDrifters()+1, 0, 0, \
self.driftersHost)
self.drift_kernels = gpu_ctx.get_kernel("driftKernels.cu", \
defines={'block_width': self.block_width, 'block_height': self.block_height
})
# Get CUDA functions and define data types for prepared_{async_}call()
self.passiveDrifterKernel = self.drift_kernels.get_function(
"passiveDrifterKernel")
self.passiveDrifterKernel.prepare("iifffiiPiPiPiPiiiiPifff")
self.enforceBoundaryConditionsKernel = self.drift_kernels.get_function(
"enforceBoundaryConditions")
self.enforceBoundaryConditionsKernel.prepare("ffiiiPi")
if self.wind_drift_factor:
#Initialize wind parameters
self.wind_textures = {}
self.wind_timestamps = {}
self.update_wind(self.drift_kernels, self.passiveDrifterKernel,
0.0)
self.local_size = (self.block_width, self.block_height, 1)
self.global_size = (\
int(np.ceil((self.getNumDrifters() + 2)/float(self.block_width))), \
1)
# Initialize drifters:
self.uniformly_distribute_drifters(
initialization_cov_drifters=initialization_cov_drifters)
