def _indices(cls, **kwargs):
"""Return the default dimension indices for a given data shape
:param dimensions: Optional, list of :class:`Dimension`
objects that defines data layout.
:param shape: Optional, shape of the spatial data to
automatically infer dimension symbols.
:return: Dimension indices used for each axis.
"""
dimensions = kwargs.get('dimensions', None)
if dimensions is None:
# Infer dimensions from default and data shape
if 'shape' not in kwargs:
error("Creating symbolic data objects requries either"
"a 'shape' or 'dimensions' argument")
raise ValueError("Unknown symbol dimensions or shape")
_indices = (x, y, z)
shape = kwargs.get('shape')
if len(shape) <= 3:
dimensions = _indices[:len(shape)]
else:
dimensions = [
Dimension("x%d" % i) for i in range(1,
len(shape) + 1)
]
return dimensions
def _indices(cls, **kwargs):
"""Return the default dimension indices for a given data shape
:param grid: :class:`Grid` object from which to infer the data
shape and :class:`Dimension` indices.
:return: Dimension indices used for each axis.
"""
save = kwargs.get('save', None)
grid = kwargs.get('grid', None)
time_dim = kwargs.get('time_dim', None)
if grid is None:
error('TimeFunction objects require a grid parameter.')
raise ValueError('No grid provided for TimeFunction.')
if time_dim is None:
time_dim = grid.time_dim if save else grid.stepping_dim
elif not isinstance(time_dim, TimeDimension):
raise ValueError("time_dim must be a TimeDimension, not %s" %
type(time_dim))
assert (isinstance(time_dim, Dimension) and time_dim.is_Time)
_indices = Function._indices(**kwargs)
return tuple([time_dim] + list(_indices))
def malloc_aligned(shape, alignment=None, dtype=np.float32):
""" Allocate memory using the C function malloc_aligned
:param shape: Shape of the array to allocate
:param alignment: number of bytes to align to. Defaults to
page size if not set.
:param dtype: Numpy datatype to allocate. Default to np.float32
:returns (pointer, data_pointer) the first element of the tuple
is the reference that can be used to access the data as a ctypes
object. The second element is the low-level reference that is
needed only for the call to free.
"""
libc = ctypes.CDLL(find_library('c'))
data_pointer = ctypes.cast(ctypes.c_void_p(),
ctypes.POINTER(ctypes.c_float))
arraysize = int(reduce(mul, shape))
ctype = convert_dtype_to_ctype(dtype)
if alignment is None:
alignment = libc.getpagesize()
ret = libc.posix_memalign(ctypes.byref(data_pointer), alignment,
ctypes.c_ulong(arraysize * ctypes.sizeof(ctype)))
if not ret == 0:
error("Unable to allocate memory for shape %s", str(shape))
return None
data_pointer = ctypes.cast(
data_pointer, np.ctypeslib.ndpointer(dtype=dtype, shape=shape))
pointer = np.ctypeslib.as_array(data_pointer, shape=shape)
return (pointer, data_pointer)
def __new__(cls,
name,
grid,
ntime=None,
npoint=None,
data=None,
coordinates=None,
**kwargs):
p_dim = kwargs.get('dimension', Dimension(name='p_%s' % name))
npoint = npoint or coordinates.shape[0]
if data is None:
if ntime is None:
error('Either data or ntime are required to'
'initialise source/receiver objects')
else:
ntime = ntime or data.shape[0]
# Create the underlying SparseTimeFunction object
obj = SparseTimeFunction.__new__(cls,
name=name,
grid=grid,
dimensions=[grid.time_dim, p_dim],
npoint=npoint,
nt=ntime,
coordinates=coordinates,
**kwargs)
# If provided, copy initial data into the allocated buffer
if data is not None:
obj.data[:] = data
return obj
def __init__(self, *args, **kwargs):
if not self._cached():
self.name = kwargs.get('name')
self.grid = kwargs.get('grid', None)
if self.grid is None:
self.shape_domain = kwargs.get('shape', None)
self.dtype = kwargs.get('dtype', np.float32)
if self.shape_domain is None:
error("Creating a Function requires either 'shape'"
"or a 'grid' argument")
raise ValueError("Unknown symbol dimensions or shape")
else:
self.shape_domain = self.grid.shape_domain
self.dtype = kwargs.get('dtype', self.grid.dtype)
self.indices = self._indices(**kwargs)
self.staggered = kwargs.get('staggered',
tuple(0 for _ in self.indices))
if len(self.staggered) != len(self.indices):
error("Staggering argument needs %s entries for indices %s" %
(len(self.indices), self.indices))
raise ValueError("Insufficient staggered entries")
self.space_order = kwargs.get('space_order', 1)
self.initializer = kwargs.get('initializer', None)
if self.initializer is not None:
assert (callable(self.initializer))
self._first_touch = kwargs.get('first_touch',
configuration['first_touch'])
self._data_object = None
# Dynamically add derivative short-cuts
self._initialize_derivatives()
def _indices(cls, **kwargs):
"""Return the default dimension indices for a given data shape
:param grid: :class:`Grid` that defines the spatial domain.
:param dimensions: Optional, list of :class:`Dimension`
objects that defines data layout.
:return: Dimension indices used for each axis.
..note::
Only one of :param grid: or :param dimensions: is required.
"""
grid = kwargs.get('grid', None)
dimensions = kwargs.get('dimensions', None)
if grid is None:
if dimensions is None:
error("Creating a Function object requries either "
"a 'grid' or the 'dimensions' argument.")
raise ValueError("Unknown symbol dimensions or shape")
else:
if dimensions is not None:
warning(
"Creating Function with 'grid' and 'dimensions' "
"argument; ignoring the 'dimensions' and using 'grid'.")
dimensions = grid.dimensions
return dimensions
def coefficients(self):
"""Symbolic expression for the coefficients for sparse point
interpolation according to:
https://en.wikipedia.org/wiki/Bilinear_interpolation.
:returns: List of coefficients, eg. [b_11, b_12, b_21, b_22]
"""
# Grid indices corresponding to the corners of the cell
x1, y1, z1, x2, y2, z2 = sympy.symbols('x1, y1, z1, x2, y2, z2')
# Coordinate values of the sparse point
px, py, pz = self.point_symbols
if self.grid.dim == 2:
A = sympy.Matrix([[1, x1, y1, x1*y1],
[1, x1, y2, x1*y2],
[1, x2, y1, x2*y1],
[1, x2, y2, x2*y2]])
p = sympy.Matrix([[1],
[px],
[py],
[px*py]])
# Map to reference cell
x, y = self.grid.dimensions
reference_cell = {x1: 0, y1: 0, x2: x.spacing, y2: y.spacing}
elif self.grid.dim == 3:
A = sympy.Matrix([[1, x1, y1, z1, x1*y1, x1*z1, y1*z1, x1*y1*z1],
[1, x1, y2, z1, x1*y2, x1*z1, y2*z1, x1*y2*z1],
[1, x2, y1, z1, x2*y1, x2*z1, y2*z1, x2*y1*z1],
[1, x1, y1, z2, x1*y1, x1*z2, y1*z2, x1*y1*z2],
[1, x2, y2, z1, x2*y2, x2*z1, y2*z1, x2*y2*z1],
[1, x1, y2, z2, x1*y2, x1*z2, y2*z2, x1*y2*z2],
[1, x2, y1, z2, x2*y1, x2*z2, y1*z2, x2*y1*z2],
[1, x2, y2, z2, x2*y2, x2*z2, y2*z2, x2*y2*z2]])
p = sympy.Matrix([[1],
[px],
[py],
[pz],
[px*py],
[px*pz],
[py*pz],
[px*py*pz]])
# Map to reference cell
x, y, z = self.grid.dimensions
reference_cell = {x1: 0, y1: 0, z1: 0, x2: x.spacing,
y2: y.spacing, z2: z.spacing}
else:
error('Point interpolation only supported for 2D and 3D')
raise NotImplementedError('Interpolation coefficients not '
'implemented for %d dimensions.'
% self.grid.dim)
A = A.subs(reference_cell)
return A.inv().T.dot(p)
def sniff_compiler_version(cc):
"""
Detect the compiler version.
Adapted from: ::
https://github.com/OP2/PyOP2/
"""
try:
res = run([cc, "--version"], stdout=PIPE, stderr=DEVNULL)
ver = res.stdout.decode("utf-8")
if not ver:
return version.LooseVersion("unknown")
except UnicodeDecodeError:
return version.LooseVersion("unknown")
except FileNotFoundError:
error("The `%s` compiler isn't available on this system" % cc)
sys.exit(1)
if ver.startswith("gcc"):
compiler = "gcc"
elif ver.startswith("clang"):
compiler = "clang"
elif ver.startswith("Apple LLVM"):
compiler = "clang"
elif ver.startswith("icc"):
compiler = "icc"
elif ver.startswith("pgcc"):
compiler = "pgcc"
else:
compiler = "unknown"
ver = version.LooseVersion("unknown")
if compiler in ["gcc", "icc"]:
try:
# gcc-7 series only spits out patch level on dumpfullversion.
res = run([cc, "-dumpfullversion"], stdout=PIPE, stderr=DEVNULL)
ver = res.stdout.decode("utf-8")
ver = '.'.join(ver.strip().split('.')[:3])
if not ver:
res = run([cc, "-dumpversion"], stdout=PIPE, stderr=DEVNULL)
ver = res.stdout.decode("utf-8")
ver = '.'.join(ver.strip().split('.')[:3])
if not ver:
return version.LooseVersion("unknown")
ver = version.StrictVersion(ver)
except UnicodeDecodeError:
pass
# Pure integer versions (e.g., ggc5, rather than gcc5.0) need special handling
try:
ver = version.StrictVersion(float(ver))
except TypeError:
pass
return ver
def _write_block_report(self, times):
"""Writes auto tuning report for block sizes
:param times: sorted list - times with block sizes
:raises IOError: if fails to write report
"""
try:
full_report_text = ["%s %f\n" % (str(block), timee) for block, timee in times]
# Cache blocking dimensions
cb_dims = str(self.blocked_dims).replace(" ", '')
shape = str(self.op.shape).replace(" ", '')
# Writes all auto tuning information into full report
with open(path.join(self.report_dir,
"%s_time_o_%s_spc_bo_%s_shp_%s_b_%s_report.txt" %
(self.op.getName(), self.op.time_order,
self.op.spc_border, shape, cb_dims)), 'w') as f:
f.writelines(full_report_text)
# string that describes the model
model_desc_str = self.model_desc_template % (self.op.getName(),
self.op.time_order,
self.op.spc_border,
shape, cb_dims)
str_to_write = "%s %s\n" % (model_desc_str, str(times[0][0]).replace(" ", ''))
# initialises report file if it does not exist
if not path.isfile(self.final_report_path):
with open(self.final_report_path, 'w') as final_report:
final_report.write("f name, time o, space bo ,"
"shape, block dims, best block size\n")
final_report.write(str_to_write) # writes the string
return
else:
# reads all the contents, checks whether entry already exist and updates.
# Otherwise appends to the end of the file
with open(self.final_report_path, 'r') as final_report:
lines = final_report.readlines()
entry_found = False
for i in range(1, len(lines)):
if model_desc_str in lines[i]:
lines[i] = str_to_write
entry_found = True
break
if not entry_found: # if entry not found append string to the end of file
lines.append(str_to_write)
# writes all the contents
with open(self.final_report_path, 'w') as final_report:
final_report.writelines(lines)
except IOError as e:
error("Failed to write auto tuning report because %s" % e.message)
def GradientOperator(model,
source,
receiver,
space_order=4,
save=True,
kernel='OT2',
**kwargs):
"""
Constructor method for the gradient operator in an acoustic media
:param model: :class:`Model` object containing the physical parameters
:param source: :class:`PointData` object containing the source geometry
:param receiver: :class:`PointData` object containing the acquisition geometry
:param time_order: Time discretization order
:param space_order: Space discretization order
"""
m, damp = model.m, model.damp
# Gradient symbol and wavefield symbols
grad = Function(name='grad', grid=model.grid)
u = TimeFunction(name='u',
grid=model.grid,
save=source.nt if save else None,
time_order=2,
space_order=space_order)
v = TimeFunction(name='v',
grid=model.grid,
save=None,
time_order=2,
space_order=space_order)
rec = Receiver(name='rec',
grid=model.grid,
time_range=receiver.time_range,
npoint=receiver.npoint)
s = model.grid.stepping_dim.spacing
eqn = iso_stencil(v, m, s, damp, kernel, forward=False)
if kernel == 'OT2':
gradient_update = Inc(grad, grad - u.dt2 * v)
elif kernel == 'OT4':
gradient_update = Inc(
grad, grad - (u.dt2 + s**2 / 12.0 * u.laplace2(m**(-2))) * v)
else:
error("Unrecognized kernel, has to be OT2 or OT4")
# Add expression for receiver injection
receivers = rec.inject(field=v.backward,
expr=rec * s**2 / m,
offset=model.nbpml)
# Substitute spacing terms to reduce flops
return Operator(eqn + receivers + [gradient_update],
subs=model.spacing_map,
name='Gradient',
**kwargs)
def point_increments(self):
"""Index increments in each dimension for each point symbol"""
if self.grid.dim == 2:
return ((0, 0), (0, 1), (1, 0), (1, 1))
elif self.grid.dim == 3:
return ((0, 0, 0), (0, 1, 0), (1, 0, 0), (0, 0, 1), (1, 1, 0),
(0, 1, 1), (1, 0, 1), (1, 1, 1))
else:
error('Point interpolation only supported for 2D and 3D')
raise NotImplementedError('Point increments not defined '
'for %d dimensions.' % self.grid.dim)
def adjoint(self, matvec):
if matvec == direct:
return self
else:
if self == centered:
return centered
elif self == right:
return left
elif self == left:
return right
else:
error("Unsupported side value")
def sniff_compiler_version(cc):
"""
Detect the compiler version.
Adapted from: ::
https://github.com/OP2/PyOP2/
"""
try:
ver = check_output([cc, "--version"]).decode("utf-8")
except (CalledProcessError, UnicodeDecodeError):
return version.LooseVersion("unknown")
except FileNotFoundError:
error("The `%s` compiler isn't available on this system" % cc)
sys.exit(1)
if ver.startswith("gcc"):
compiler = "gcc"
elif ver.startswith("clang"):
compiler = "clang"
elif ver.startswith("Apple LLVM"):
compiler = "clang"
elif ver.startswith("icc"):
compiler = "icc"
else:
compiler = "unknown"
ver = version.LooseVersion("unknown")
if compiler in ["gcc", "icc"]:
try:
# gcc-7 series only spits out patch level on dumpfullversion.
ver = check_output([cc, "-dumpfullversion"], stderr=DEVNULL).decode("utf-8")
ver = '.'.join(ver.strip().split('.')[:3])
ver = version.StrictVersion(ver)
except CalledProcessError:
try:
ver = check_output([cc, "-dumpversion"], stderr=DEVNULL).decode("utf-8")
ver = '.'.join(ver.strip().split('.')[:3])
ver = version.StrictVersion(ver)
except (CalledProcessError, UnicodeDecodeError):
pass
except UnicodeDecodeError:
pass
# Pure integer versions (e.g., ggc5, rather than gcc5.0) need special handling
try:
ver = version.StrictVersion(float(ver))
except TypeError:
pass
return ver
def __indices_setup__(cls, **kwargs):
grid = kwargs.get('grid')
dimensions = kwargs.get('dimensions')
if grid is None:
if dimensions is None:
error("Creating a Function object requries either "
"a 'grid' or the 'dimensions' argument.")
raise ValueError("Unknown symbol dimensions or shape")
else:
if dimensions is not None:
warning(
"Creating Function with 'grid' and 'dimensions' "
"argument; ignoring the 'dimensions' and using 'grid'.")
dimensions = grid.dimensions
return dimensions
def __indices_setup__(cls, **kwargs):
save = kwargs.get('save')
grid = kwargs.get('grid')
time_dim = kwargs.get('time_dim')
if grid is None:
error('TimeFunction objects require a grid parameter.')
raise ValueError('No grid provided for TimeFunction.')
if time_dim is None:
time_dim = grid.time_dim if save else grid.stepping_dim
elif not (isinstance(time_dim, Dimension) and time_dim.is_Time):
raise ValueError("'time_dim' must be a time dimension")
return (time_dim, ) + Function.__indices_setup__(**kwargs)
def apply(self, *args, **kwargs):
"""Apply defined stencil kernel to a set of data objects"""
if len(args) <= 0:
args = self.signature
args = [a.data if isinstance(a, SymbolicData) else a for a in args]
# Check shape of argument data
for arg, v in zip(args, self.signature):
if not isinstance(arg, np.ndarray):
raise TypeError('No array data found for argument %s' % v.name)
if arg.shape != v.shape:
error('Expected argument %s with shape %s, but got shape %s' %
(v.name, v.shape, arg))
raise ValueError('Argument with wrong shape')
# Invoke kernel function with args
self.cfunction(*args)
def _indices(cls, **kwargs):
"""Return the default dimension indices for a given data shape
:param grid: :class:`Grid` object from which to infer the data
shape and :class:`Dimension` indices.
:return: Dimension indices used for each axis.
"""
save = kwargs.get('save', None)
grid = kwargs.get('grid', None)
if grid is None:
error('TimeFunction objects require a grid parameter.')
raise ValueError('No grid provided for TimeFunction.')
tidx = grid.time_dim if save else grid.stepping_dim
_indices = Function._indices(**kwargs)
return tuple([tidx] + list(_indices))
def _dle_arguments(self, dim_sizes):
# Add user-provided block sizes, if any
dle_arguments = OrderedDict()
autotune = True
for i in self.dle_arguments:
dim_size = dim_sizes.get(i.original_dim.name, i.original_dim.size)
if dim_size is None:
error('Unable to derive size of dimension %s from defaults. '
'Please provide an explicit value.' % i.original_dim.name)
raise InvalidArgument('Unknown dimension size')
if i.value:
try:
dle_arguments[i.argument.name] = i.value(dim_size)
except TypeError:
dle_arguments[i.argument.name] = i.value
autotune = False
else:
dle_arguments[i.argument.name] = dim_size
return dle_arguments, autotune
def laplacian(field, time_order, m, s):
"""
Spacial discretization for the isotropic acoustic wave equation. For a 4th
order in time formulation, the 4th order time derivative is replaced by a
double laplacian:
H = (laplacian + s**2/12 laplacian(1/m*laplacian))
:param field: Symbolic TimeFunction object, solution to be computed
:param time_order: time order
:param m: square slowness
:param s: symbol for the time-step
:return: H
"""
if time_order == 2:
biharmonic = 0
elif time_order == 4:
biharmonic = field.laplace2(1 / m)
else:
error("Unsupported time order %d, order has to be 2 or 4" % time_order)
return field.laplace + s**2 / 12 * biharmonic
def __init__(self, *args, **kwargs):
if not self._cached():
self.nt = kwargs.get('nt', 0)
self.npoint = kwargs.get('npoint')
kwargs['shape'] = (self.nt, self.npoint)
super(SparseFunction, self).__init__(self, *args, **kwargs)
if self.grid is None:
error('SparseFunction objects require a grid parameter.')
raise ValueError('No grid provided for SparseFunction.')
# Allocate and copy coordinate data
d = Dimension('d')
self.coordinates = Function(name='%s_coords' % self.name,
dimensions=[self.indices[-1], d],
shape=(self.npoint, self.grid.dim))
self._children.append(self.coordinates)
coordinates = kwargs.get('coordinates', None)
if coordinates is not None:
self.coordinates.data[:] = coordinates[:]
def __new__(cls, *args, **kwargs):
options = kwargs.get('options', {})
key = cls
obj = cls._cache_get(key)
if obj is not None:
newobj = sympy.Function.__new__(cls, *args, **options)
newobj.__init_cached__(key)
return newobj
p_dim = kwargs.get('dimension', Dimension('p_%s' % kwargs.get("name")))
npoint = kwargs.get("npoint")
coords = kwargs.get("coordinates")
if npoint is None:
if coords is None:
raise TypeError("Need either `npoint` or `coordinates`")
else:
npoint = coords.shape[0]
grid = kwargs.get("grid")
ntime = kwargs.get("ntime")
if kwargs.get("data") is None:
if ntime is None:
error('Either data or ntime are required to'
'initialise source/receiver objects')
else:
ntime = kwargs.get("ntime") or kwargs.get("data").shape[0]
# Create the underlying SparseTimeFunction object
kwargs["nt"] = ntime
kwargs['npoint'] = npoint
obj = SparseTimeFunction.__new__(cls,
dimensions=[grid.time_dim, p_dim],
**kwargs)
# If provided, copy initial data into the allocated buffer
if kwargs.get("data") is not None:
obj.data[:] = kwargs.get("data")
return obj
def __init__(self, *args, **kwargs):
if not self._cached():
super(TimeFunction, self).__init__(*args, **kwargs)
self.time_dim = kwargs.get('time_dim', None)
self.time_order = kwargs.get('time_order', 1)
self.save = kwargs.get('save', False)
if not self.save:
if self.time_dim is not None:
warning(
'Explicit time dimension size (time_dim) found for '
'TimeFunction symbol %s, despite \nusing a stepping time '
'dimension (save=False). This value will be ignored!' %
self.name)
self.time_dim = self.time_order + 1
self.indices[0].modulo = self.time_dim
else:
if self.time_dim is None:
error('Time dimension (time_dim) is required'
'to save intermediate data with save=True')
raise ValueError("Unknown time dimensions")
def unique(self, key):
"""
Returns a unique value for a given key, if such a value
exists, and raises a ``ValueError`` if it does not.
:param key: Key for which to retrieve a unique value
"""
candidates = self.getall(key)
def compare_to_first(v):
first = candidates[0]
if isinstance(first, np.ndarray) or isinstance(v, np.ndarray):
return (first == v).all()
else:
return first == v
if len(candidates) == 1:
return candidates[0]
elif all(map(compare_to_first, candidates)):
return candidates[0]
else:
error("Unable to find unique value for key %s, candidates: %s" %
(key, candidates))
raise ValueError('Inconsistent values for key reduction')
def locate_intel_advisor():
"""
Detect if Intel Advisor is installed on the machine and return
its location if it is.
"""
path = None
try:
# Check if the directory to Intel Advisor is specified
path = Path(os.environ['DEVITO_ADVISOR_DIR'])
except KeyError:
# Otherwise, 'sniff' the location of Advisor's directory
error_msg = 'Intel Advisor cannot be found on your system, consider if you'\
' have sourced its environment variables correctly. Information can'\
' be found at https://software.intel.com/content/www/us/en/develop/'\
'documentation/advisor-user-guide/top/launch-the-intel-advisor/'\
'intel-advisor-cli/setting-and-using-intel-advisor-environment'\
'-variables.html'
try:
res = run(["advixe-cl", "--version"], stdout=PIPE, stderr=DEVNULL)
ver = res.stdout.decode("utf-8")
if not ver:
error(error_msg)
return None
except (UnicodeDecodeError, FileNotFoundError):
error(error_msg)
return None
env_path = os.environ["PATH"]
env_path_dirs = env_path.split(":")
for env_path_dir in env_path_dirs:
# intel/advisor is the advisor directory for Intel Parallel Studio,
# intel/oneapi/advisor is the directory for Intel oneAPI
if "intel/advisor" in env_path_dir or "intel/oneapi/advisor" in env_path_dir:
path = Path(env_path_dir)
if path.name.startswith('bin'):
path = path.parent
if not path:
error(error_msg)
return None
if path.joinpath('bin64').joinpath('advixe-cl').is_file():
return path
else:
warning("Requested `advisor` profiler, but couldn't locate executable"
"in advisor directory")
return None
def staggered_diff(f, dim, order, stagger=centered, theta=0, phi=0):
"""
Utility function to generate staggered derivatives
:param f: function objects, eg. `f(x, y)` or `g(t, x, y, z)`.
:param dims: symbol defining the dimension wrt. which
to differentiate, eg. `x`, `y`, `z` or `t`.
:param order: Order of the coefficient discretization and thus
the width of the resulting stencil expression.
:param stagger: Shift for the FD, `left`, `right` or `centered`
:param theta: Dip (or polar) angle for rotated FD
:param phi: Azimuth angle for rotated FD
"""
ndim = len(f.space_dimensions)
off = dict([(d, 0) for d in f.space_dimensions])
if stagger == left:
off[dim] = -.5
elif stagger == right:
off[dim] = .5
else:
off[dim] = 0
if theta == 0 and phi == 0:
diff = dim.spacing
idx = list(
set([(dim + int(i + .5 + off[dim]) * diff)
for i in range(-int(order / 2), int(order / 2))]))
if int(order / 2) == 1:
idx = [dim - diff, dim]
deriv = f.diff(dim).as_finite_difference(idx,
x0=dim +
off[dim] * dim.spacing)
return deriv.evalf(_PRECISION)
else:
if ndim < 2 or ndim > 3:
error("Only 2D and 3D supports rotated finite difference,"
" %d dimensions as input" % ndim)
raise ValueError("Non supported number of dimensions")
x = f.space_dimensions[0]
z = f.space_dimensions[-1]
idxx = list(
set([(x + int(i + .5 + off[x]) * x.spacing)
for i in range(-int(order / 2), int(order / 2))]))
dx = f.diff(x).as_finite_difference(idxx, x0=x + off[x] * x.spacing)
dx = dx.evalf(_PRECISION)
idxz = list(
set([(z + int(i + .5 + off[z]) * z.spacing)
for i in range(-int(order / 2), int(order / 2))]))
dz = f.diff(z).as_finite_difference(idxz, x0=z + off[z] * z.spacing)
dz = dx.evalf(_PRECISION)
dy = 0
is_y = False
if ndim == 3:
y = f.space_dimensions[1]
idxy = list(
set([(y + int(i + .5 + off[y]) * y.spacing)
for i in range(-int(order / 2), int(order / 2))]))
dy = f.diff(y).as_finite_difference(idxy,
x0=y + off[y] * y.spacing)
dy = dy.evalf(_PRECISION)
is_y = (dim == y)
if dim == x:
return cos(theta) * cos(phi) * dx + sin(phi) * cos(theta) * dy -\
sin(theta) * dz
elif dim == z:
return sin(theta) * cos(phi) * dx + sin(phi) * sin(theta) * dy +\
cos(theta) * dz
elif is_y:
return -sin(phi) * dx + cos(phi) * dy
else:
return 0
def fwi_gradient_shot(vp_in, i, solver_params):
error("Initialising solver")
tn = solver_params['tn']
nbl = solver_params['nbl']
space_order = solver_params['space_order']
dtype = solver_params['dtype']
origin = solver_params['origin']
spacing = solver_params['spacing']
shots_container = solver_params['shots_container']
true_d, source_location, old_dt = load_shot(i, container=shots_container)
model = Model(vp=vp_in,
nbl=nbl,
space_order=space_order,
dtype=dtype,
shape=vp_in.shape,
origin=origin,
spacing=spacing,
bcs="damp")
geometry = create_geometry(model, tn, source_location)
error("tn: %d, nt: %d, dt: %f" % (geometry.tn, geometry.nt, geometry.dt))
error("Reinterpolate shot from %d samples to %d samples" %
(true_d.shape[0], geometry.nt))
true_d = reinterpolate(true_d, geometry.nt, old_dt)
solver = AcousticWaveSolver(model,
geometry,
kernel='OT2',
nbl=nbl,
space_order=space_order,
dtype=dtype)
grad = Function(name="grad", grid=model.grid)
residual = Receiver(name='rec',
grid=model.grid,
time_range=geometry.time_axis,
coordinates=geometry.rec_positions)
u0 = TimeFunction(name='u',
grid=model.grid,
time_order=2,
space_order=solver.space_order,
save=geometry.nt)
error("Forward prop")
smooth_d, _, _ = solver.forward(save=True, u=u0)
error("Misfit")
residual.data[:] = smooth_d.data[:] - true_d[:]
objective = .5 * np.linalg.norm(residual.data.ravel())**2
error("Gradient")
solver.gradient(rec=residual, u=u0, grad=grad)
grad = clip_boundary_and_numpy(grad.data, model.nbl)
return objective, -grad
def demo_model(preset, **kwargs):
"""
Utility function to create preset :class:`Model` objects for
demonstration and testing purposes. The particular presets are ::
* 'layer2D': Simple two-layer model with velocities 1.5 km/s
and 2.5 km/s in the top and bottom layer respectively.
* 'marmousi2D': Loads the 2D Marmousi data set from the given
filepath. Requires the ``opesci/data`` repository
to be available on your machine.
"""
if preset.lower() in ['constant-isotropic']:
# A constant single-layer model in a 2D or 3D domain
# with velocity 1.5km/s.
shape = kwargs.pop('shape', (101, 101))
spacing = kwargs.pop('spacing', tuple([10. for _ in shape]))
origin = kwargs.pop('origin', tuple([0. for _ in shape]))
nbpml = kwargs.pop('nbpml', 10)
vp = kwargs.pop('vp', 1.5)
return Model(vp=vp,
origin=origin,
shape=shape,
spacing=spacing,
nbpml=nbpml,
**kwargs)
elif preset.lower() in ['constant-tti']:
# A constant single-layer model in a 2D or 3D domain
# with velocity 1.5km/s.
shape = kwargs.pop('shape', (101, 101))
spacing = kwargs.pop('spacing', tuple([10. for _ in shape]))
origin = kwargs.pop('origin', tuple([0. for _ in shape]))
nbpml = kwargs.pop('nbpml', 10)
v = np.empty(shape, dtype=np.float32)
v[:] = 1.5
epsilon = .3 * np.ones(shape)
delta = .2 * np.ones(shape)
theta = .7 * np.ones(shape)
phi = None
if len(shape) > 2:
phi = .35 * np.ones(shape)
return Model(vp=v,
origin=origin,
shape=shape,
spacing=spacing,
nbpml=nbpml,
epsilon=epsilon,
delta=delta,
theta=theta,
phi=phi,
**kwargs)
elif preset.lower() in [
'layers-isotropic', 'twolayer-isotropic', '2layer-isotropic'
]:
# A two-layer model in a 2D or 3D domain with two different
# velocities split across the height dimension:
# By default, the top part of the domain has 1.5 km/s,
# and the bottom part of the domain has 2.5 km/s.
shape = kwargs.pop('shape', (101, 101))
spacing = kwargs.pop('spacing', tuple([10. for _ in shape]))
origin = kwargs.pop('origin', tuple([0. for _ in shape]))
nbpml = kwargs.pop('nbpml', 10)
ratio = kwargs.pop('ratio', 2)
vp_top = kwargs.pop('vp_top', 1.5)
vp_bottom = kwargs.pop('vp_bottom', 2.5)
# Define a velocity profile in km/s
v = np.empty(shape, dtype=np.float32)
v[:] = vp_top # Top velocity (background)
v[..., int(shape[-1] / ratio):] = vp_bottom # Bottom velocity
return Model(vp=v,
origin=origin,
shape=shape,
spacing=spacing,
nbpml=nbpml,
**kwargs)
elif preset.lower() in ['layers-tti', 'twolayer-tti', '2layer-tti']:
# A two-layer model in a 2D or 3D domain with two different
# velocities split across the height dimension:
# By default, the top part of the domain has 1.5 km/s,
# and the bottom part of the domain has 2.5 km/s.\
shape = kwargs.pop('shape', (101, 101))
spacing = kwargs.pop('spacing', tuple([10. for _ in shape]))
origin = kwargs.pop('origin', tuple([0. for _ in shape]))
nbpml = kwargs.pop('nbpml', 10)
ratio = kwargs.pop('ratio', 2)
vp_top = kwargs.pop('vp_top', 1.5)
vp_bottom = kwargs.pop('vp_bottom', 2.5)
# Define a velocity profile in km/s
v = np.empty(shape, dtype=np.float32)
v[:] = vp_top # Top velocity (background)
v[..., int(shape[-1] / ratio):] = vp_bottom # Bottom velocity
epsilon = .3 * (v - 1.5)
delta = .2 * (v - 1.5)
theta = .5 * (v - 1.5)
phi = None
if len(shape) > 2:
phi = .1 * (v - 1.5)
return Model(vp=v,
origin=origin,
shape=shape,
spacing=spacing,
nbpml=nbpml,
epsilon=epsilon,
delta=delta,
theta=theta,
phi=phi,
**kwargs)
elif preset.lower() in ['circle-isotropic']:
# A simple circle in a 2D domain with a background velocity.
# By default, the circle velocity is 2.5 km/s,
# and the background veloity is 3.0 km/s.
shape = kwargs.pop('shape', (101, 101))
spacing = kwargs.pop('spacing', tuple([10. for _ in shape]))
origin = kwargs.pop('origin', tuple([0. for _ in shape]))
nbpml = kwargs.pop('nbpml', 10)
vp = kwargs.pop('vp', 3.0)
vp_background = kwargs.pop('vp_background', 2.5)
r = kwargs.pop('r', 15)
# Only a 2D preset is available currently
assert (len(shape) == 2)
v = np.empty(shape, dtype=np.float32)
v[:] = vp_background
a, b = shape[0] / 2, shape[1] / 2
y, x = np.ogrid[-a:shape[0] - a, -b:shape[1] - b]
v[x * x + y * y <= r * r] = vp
return Model(vp=v,
origin=origin,
shape=shape,
spacing=spacing,
nbpml=nbpml)
elif preset.lower() in ['marmousi-isotropic', 'marmousi2d-isotropic']:
shape = (1601, 401)
spacing = (7.5, 7.5)
origin = (0., 0.)
# Read 2D Marmousi model from opesc/data repo
data_path = kwargs.get('data_path', None)
if data_path is None:
error("Path to opesci/data not found! Please specify with "
"'data_path=<path/to/opesci/data>'")
raise ValueError("Path to model data unspecified")
path = os.path.join(data_path, 'Simple2D/vp_marmousi_bi')
v = np.fromfile(path, dtype='float32', sep="")
v = v.reshape(shape)
# Cut the model to make it slightly cheaper
v = v[301:-300, :]
return Model(vp=v,
origin=origin,
shape=v.shape,
spacing=spacing,
nbpml=20)
elif preset.lower() in ['marmousi-tti2d', 'marmousi2d-tti']:
shape_full = (201, 201, 70)
shape = (201, 70)
spacing = (10., 10.)
origin = (0., 0.)
nbpml = kwargs.pop('nbpml', 20)
# Read 2D Marmousi model from opesc/data repo
data_path = kwargs.pop('data_path', None)
if data_path is None:
error("Path to opesci/data not found! Please specify with "
"'data_path=<path/to/opesci/data>'")
raise ValueError("Path to model data unspecified")
path = os.path.join(data_path, 'marmousi3D/vp_marmousi_bi')
# velocity
vp = 1e-3 * np.fromfile(os.path.join(data_path,
'marmousi3D/MarmousiVP.raw'),
dtype='float32',
sep="")
vp = vp.reshape(shape_full)
vp = vp[101, :, :]
# Epsilon, in % in file, resale between 0 and 1
epsilon = np.fromfile(os.path.join(data_path,
'marmousi3D/MarmousiEps.raw'),
dtype='float32',
sep="") * 1e-2
epsilon = epsilon.reshape(shape_full)
epsilon = epsilon[101, :, :]
# Delta, in % in file, resale between 0 and 1
delta = np.fromfile(os.path.join(data_path,
'marmousi3D/MarmousiDelta.raw'),
dtype='float32',
sep="") * 1e-2
delta = delta.reshape(shape_full)
delta = delta[101, :, :]
# Theta, in degrees in file, resale in radian
theta = np.fromfile(os.path.join(data_path,
'marmousi3D/MarmousiTilt.raw'),
dtype='float32',
sep="")
theta = np.float32(np.pi / 180 * theta.reshape(shape_full))
theta = theta[101, :, :]
return Model(vp=vp,
origin=origin,
shape=shape,
spacing=spacing,
nbpml=nbpml,
epsilon=epsilon,
delta=delta,
theta=theta,
**kwargs)
elif preset.lower() in ['marmousi-tti3d', 'marmousi3d-tti']:
shape = (201, 201, 70)
spacing = (10., 10., 10.)
origin = (0., 0., 0.)
nbpml = kwargs.pop('nbpml', 20)
# Read 2D Marmousi model from opesc/data repo
data_path = kwargs.pop('data_path', None)
if data_path is None:
error("Path to opesci/data not found! Please specify with "
"'data_path=<path/to/opesci/data>'")
raise ValueError("Path to model data unspecified")
path = os.path.join(data_path, 'marmousi3D/vp_marmousi_bi')
# Velcoity
vp = 1e-3 * np.fromfile(os.path.join(data_path,
'marmousi3D/MarmousiVP.raw'),
dtype='float32',
sep="")
vp = vp.reshape(shape)
# Epsilon, in % in file, resale between 0 and 1
epsilon = np.fromfile(os.path.join(data_path,
'marmousi3D/MarmousiEps.raw'),
dtype='float32',
sep="") * 1e-2
epsilon = epsilon.reshape(shape)
# Delta, in % in file, resale between 0 and 1
delta = np.fromfile(os.path.join(data_path,
'marmousi3D/MarmousiDelta.raw'),
dtype='float32',
sep="") * 1e-2
delta = delta.reshape(shape)
# Theta, in degrees in file, resale in radian
theta = np.fromfile(os.path.join(data_path,
'marmousi3D/MarmousiTilt.raw'),
dtype='float32',
sep="")
theta = np.float32(np.pi / 180 * theta.reshape(shape))
# Phi, in degrees in file, resale in radian
phi = np.fromfile(os.path.join(data_path, 'marmousi3D/Azimuth.raw'),
dtype='float32',
sep="")
phi = np.float32(np.pi / 180 * phi.reshape(shape))
return Model(vp=vp,
origin=origin,
shape=shape,
spacing=spacing,
nbpml=nbpml,
epsilon=epsilon,
delta=delta,
theta=theta,
phi=phi,
**kwargs)
else:
error('Unknown model preset name %s' % preset)
def arguments(self, *args, **kwargs):
"""
Return the arguments necessary to apply the Operator.
"""
if len(args) == 0:
args = self.parameters
# Will perform auto-tuning if the user requested it and loop blocking was used
maybe_autotune = kwargs.get('autotune', False)
# Initialise argument map and a map of dimension names to values
arguments = OrderedDict([(arg.name, arg) for arg in self.parameters])
dim_sizes = dict([(arg.name, arg.size) for arg in self.parameters
if isinstance(arg, Dimension)])
o_vals = {}
for name, arg in kwargs.items():
# Override explicitly provided dim sizes from **kwargs
if name in dim_sizes:
dim_sizes[name] = arg
# Override explicitly provided SymbolicData
if name in arguments and isinstance(arguments[name], SymbolicData):
# Override the original symbol
o_vals[name] = arg
original = arguments[name]
if original.is_CompositeData:
for orig, child in zip(original.children, arg.children):
o_vals[orig.name] = child
# Replace the overridden values with the provided ones
for argname in o_vals.keys():
arguments[argname] = o_vals[argname]
# Traverse positional args and infer loop sizes for open dimensions
f_args = [(name, f) for name, f in arguments.items()
if isinstance(f, SymbolicData)]
for fname, f in f_args:
arguments[fname] = self._arg_data(f)
shape = self._arg_shape(f)
# Ensure data dimensions match symbol dimensions
for i, dim in enumerate(f.indices):
# We don't need to check sizes for buffered dimensions
# against data shapes, all we need is the size of the parent.
if dim.is_Buffered:
continue
# Check data sizes for dimensions with a fixed size
if dim.size is not None:
if not shape[i] <= dim.size:
error(
'Size of data argument for %s is greater than the size '
'of dimension %s: %d' %
(fname, dim.name, dim.size))
raise InvalidArgument('Wrong data shape encountered')
else:
continue
if dim_sizes[dim.name] is None:
# We haven't determined the size of this dimension yet,
# try to infer it from the data shape.
dim_sizes[dim.name] = shape[i]
else:
# We know the dimension size, check if data shape agrees
if not dim_sizes[dim.name] <= shape[i]:
error('Size of dimension %s was determined to be %d, '
'but data for symbol %s has shape %d.' %
(dim.name, dim_sizes[dim.name], fname, shape[i]))
raise InvalidArgument('Wrong data shape encountered')
# Make sure we have defined all buffered dimensions and their parents,
# even if they are not explicitly given or used.
d_args = [d for d in arguments.values() if isinstance(d, Dimension)]
for d in d_args:
if d.is_Buffered:
if dim_sizes[d.parent.name] is None:
dim_sizes[d.parent.name] = dim_sizes[d.name]
if dim_sizes[d.name] is None:
dim_sizes[d.name] = dim_sizes[d.parent.name]
# Add user-provided block sizes, if any
dle_arguments = OrderedDict()
for i in self._dle_state.arguments:
dim_size = dim_sizes.get(i.original_dim.name, i.original_dim.size)
if dim_size is None:
error('Unable to derive size of dimension %s from defaults. '
'Please provide an explicit value.' %
i.original_dim.name)
raise InvalidArgument('Unknown dimension size')
if i.value:
try:
dle_arguments[i.argument.name] = i.value(dim_size)
except TypeError:
dle_arguments[i.argument.name] = i.value
# User-provided block size available, do not autotune
maybe_autotune = False
else:
dle_arguments[i.argument.name] = dim_size
dim_sizes.update(dle_arguments)
# Insert loop size arguments from dimension values
d_args = [d for d in arguments.values() if isinstance(d, Dimension)]
for d in d_args:
arguments[d.name] = dim_sizes[d.name]
# Might have been asked to auto-tune the block size
if maybe_autotune:
arguments = self._autotune(arguments)
# Add profiler structs
arguments.update(self._extra_arguments())
# Sanity check argument derivation
for name, arg in arguments.items():
if isinstance(arg, SymbolicData) or isinstance(arg, Dimension):
raise ValueError('Runtime argument %s not defined' % arg)
return arguments, dim_sizes
def _arg_data(self, argument):
# Ensure we're dealing or deriving numpy arrays
data = argument.data
if not isinstance(data, np.ndarray):
error('No array data found for argument %s' % argument.name)
return data
