def test_geom(shape):
vp = np.ones(shape)
o = tuple([0] * len(shape))
d = tuple([10] * len(shape))
model = Model(o, d, shape, 4, vp, nbl=20, dt=1)
assert model.critical_dt == 1
nrec = 31
nsrc = 4
rec_coordinates = np.ones((nrec, len(shape)))
src_coordinates = np.ones((nsrc, len(shape)))
geometry = AcquisitionGeometry(model,
rec_coordinates,
src_coordinates,
t0=0.0,
tn=250)
assert geometry.grid == model.grid
assert geometry.nrec == nrec
assert geometry.nsrc == nsrc
assert geometry.src_type is None
assert geometry.rec.shape == (251, nrec)
assert norm(geometry.rec) == 0
assert geometry.src.shape == (251, nsrc)
assert norm(geometry.new_src(src_type=None)) == 0
assert norm(geometry.src) == 0
rec2 = geometry.rec.resample(num=501)
assert rec2.shape == (501, nrec)
assert rec2.grid == model.grid
assert geometry.new_rec(name="bonjour").name == "bonjour"
assert geometry.new_src(name="bonjour").name == "bonjour"
def test_default_geom(shape):
vp = np.ones(shape)
o = tuple([0] * len(shape))
d = tuple([10] * len(shape))
model = Model(o, d, shape, 4, vp, nbl=20, dt=1)
assert model.critical_dt == 1
geometry = setup_geometry(model, 250)
nrec = shape[0] * (shape[1] if len(shape) > 2 else 1)
assert geometry.grid == model.grid
assert geometry.nrec == nrec
assert geometry.nsrc == 1
assert geometry.src_type == "Ricker"
assert geometry.rec.shape == (251, nrec)
assert norm(geometry.rec) == 0
assert geometry.src.shape == (251, 1)
assert norm(geometry.new_src(src_type=None)) == 0
rec2 = geometry.rec.resample(num=501)
assert rec2.shape == (501, nrec)
assert rec2.grid == model.grid
assert geometry.new_rec(name="bonjour").name == "bonjour"
assert geometry.new_src(name="bonjour").name == "bonjour"
def test_adjoint_F(self, shape, kernel, space_order, nbpml, save,
Eu, Erec, Ev, Esrca):
"""
Unlike `test_adjoint_F` in test_adjoint.py, here we explicitly check the norms
of all Operator-evaluated Functions. The numbers we check against are derived
"manually" from sequential runs of test_adjoint::test_adjoint_F
"""
tn = 500. # Final time
nrec = 130 # Number of receivers
# Create solver from preset
solver = acoustic_setup(shape=shape, spacing=[15. for _ in shape], kernel=kernel,
nbpml=nbpml, tn=tn, space_order=space_order, nrec=nrec,
preset='layers-isotropic', dtype=np.float64)
# Run forward operator
rec, u, _ = solver.forward(save=save)
assert np.isclose(norm(u), Eu, rtol=Eu*1.e-8)
assert np.isclose(norm(rec), Erec, rtol=Erec*1.e-8)
# Run adjoint operator
srca, v, _ = solver.adjoint(rec=rec)
assert np.isclose(norm(v), Ev, rtol=Ev*1.e-8)
assert np.isclose(norm(srca), Esrca, rtol=Esrca*1.e-8)
# Adjoint test: Verify <Ax,y> matches <x, A^Ty> closely
term1 = inner(srca, solver.geometry.src)
term2 = norm(rec)**2
assert np.isclose((term1 - term2)/term1, 0., rtol=1.e-10)
def test_skewed_bounds(self, expr, expected, norm_u, norm_v):
"""Tests code generation on skewed indices."""
grid = Grid(shape=(16, 16, 16))
x, y, z = grid.dimensions
time = grid.time_dim
u = TimeFunction(name='u', grid=grid) # noqa
v = TimeFunction(name='v', grid=grid) # noqa
eqn = eval(expr)
# List comprehension would need explicit locals/globals mappings to eval
op = Operator(eqn, opt=('blocking', {'skewing': True}))
op.apply(time_M=5)
iters = FindNodes(Iteration).visit(op)
time_iter = [i for i in iters if i.dim.is_Time]
assert len(time_iter) == 1
bns, _ = assert_blocking(op, {'x0_blk0'})
iters = FindNodes(Iteration).visit(bns['x0_blk0'])
assert len(iters) == 5
assert iters[0].dim.parent is x
assert iters[1].dim.parent is y
assert iters[4].dim is z
assert iters[2].dim.parent is iters[0].dim
assert iters[3].dim.parent is iters[1].dim
assert iters[0].symbolic_min == (iters[0].dim.parent.symbolic_min +
time)
assert iters[0].symbolic_max == (iters[0].dim.parent.symbolic_max +
time)
assert iters[1].symbolic_min == (iters[1].dim.parent.symbolic_min +
time)
assert iters[1].symbolic_max == (iters[1].dim.parent.symbolic_max +
time)
assert iters[2].symbolic_min == iters[2].dim.symbolic_min
assert iters[2].symbolic_max == MIN(
iters[0].dim + iters[0].dim.symbolic_incr - 1,
iters[0].dim.symbolic_max + time)
assert iters[3].symbolic_min == iters[3].dim.symbolic_min
assert iters[3].symbolic_max == MIN(
iters[1].dim + iters[1].dim.symbolic_incr - 1,
iters[1].dim.symbolic_max + time)
assert iters[4].symbolic_min == (iters[4].dim.symbolic_min)
assert iters[4].symbolic_max == (iters[4].dim.symbolic_max)
skewed = [i.expr for i in FindNodes(Expression).visit(bns['x0_blk0'])]
assert str(skewed[0]).replace(' ', '') == expected
assert np.isclose(norm(u), norm_u, rtol=1e-5)
assert np.isclose(norm(v), norm_v, rtol=1e-5)
u.data[:] = 0
v.data[:] = 0
op2 = Operator(eqn, opt=('advanced'))
op2.apply(time_M=5)
assert np.isclose(norm(u), norm_u, rtol=1e-5)
assert np.isclose(norm(v), norm_v, rtol=1e-5)
def run(shape=(50, 50, 50),
spacing=(20.0, 20.0, 20.0),
tn=1000.0,
space_order=4,
kernel='OT2',
nbl=40,
full_run=False,
fs=False,
autotune=False,
preset='layers-isotropic',
checkpointing=False,
**kwargs):
solver = acoustic_setup(shape=shape,
spacing=spacing,
nbl=nbl,
tn=tn,
space_order=space_order,
kernel=kernel,
fs=fs,
preset=preset,
**kwargs)
info("Applying Forward")
# Whether or not we save the whole time history. We only need the full wavefield
# with 'save=True' if we compute the gradient without checkpointing, if we use
# checkpointing, PyRevolve will take care of the time history
save = full_run and not checkpointing
# Define receiver geometry (spread across x, just below surface)
rec, u, summary = solver.forward(save=save, autotune=autotune)
# print(norm(rec))
print(norm(u))
if preset == 'constant':
# With a new m as Constant
v0 = Constant(name="v", value=2.0, dtype=np.float32)
solver.forward(save=save, vp=v0)
# With a new vp as a scalar value
solver.forward(save=save, vp=2.0)
if not full_run:
return summary.gflopss, summary.oi, summary.timings, [rec, u.data]
# Smooth velocity
initial_vp = Function(name='v0',
grid=solver.model.grid,
space_order=space_order)
smooth(initial_vp, solver.model.vp)
dm = np.float32(initial_vp.data**(-2) - solver.model.vp.data**(-2))
info("Applying Adjoint")
solver.adjoint(rec, autotune=autotune)
info("Applying Born")
solver.jacobian(dm, autotune=autotune)
info("Applying Gradient")
solver.jacobian_adjoint(rec,
u,
autotune=autotune,
checkpointing=checkpointing)
return summary.gflopss, summary.oi, summary.timings, [rec, u.data]
def fwi_gradient(vp_in):
# Create symbols to hold the gradient
grad = Function(name="grad", grid=model.grid)
objective = 0.
for i in range(nshots):
# Create placeholders for the data residual and data
residual = Receiver(name='residual',
grid=model.grid,
time_range=geometry.time_axis,
coordinates=geometry.rec_positions)
d_obs = Receiver(name='d_obs',
grid=model.grid,
time_range=geometry.time_axis,
coordinates=geometry.rec_positions)
d_syn = Receiver(name='d_syn',
grid=model.grid,
time_range=geometry.time_axis,
coordinates=geometry.rec_positions)
# Update source location
solver.geometry.src_positions[0, :] = source_locations[i, :]
# Generate synthetic data from true model
solver.forward(vp=model.vp, rec=d_obs)
# Compute smooth data and full forward wavefield u0
_, u0, _ = solver.forward(vp=vp_in, save=True, rec=d_syn)
# Compute gradient from data residual and update objective function
residual = compute_residual(residual, d_obs, d_syn)
objective += .5 * norm(residual)**2
solver.gradient(rec=residual, u=u0, vp=vp_in, grad=grad)
return objective, grad
def test_adjoint_F(self, mkey, shape, kernel, space_order):
"""
Adjoint test for the forward modeling operator.
The forward modeling operator F generates a shot record (measurements)
from a source while the adjoint of F generates measurments at the source
location from data. This test uses the conventional dot test:
< Fx, y> = <x, F^T y>
"""
tn = 500. # Final time
# Create solver from preset
solver = acoustic_setup(shape=shape, spacing=[15. for _ in shape], kernel=kernel,
nbl=10, tn=tn, space_order=space_order,
**(presets[mkey]), dtype=np.float64)
# Create adjoint receiver symbol
srca = Receiver(name='srca', grid=solver.model.grid,
time_range=solver.geometry.time_axis,
coordinates=solver.geometry.src_positions)
# Run forward and adjoint operators
rec, _, _ = solver.forward(save=False)
solver.adjoint(rec=rec, srca=srca)
# Adjoint test: Verify <Ax,y> matches <x, A^Ty> closely
term1 = np.dot(srca.data.reshape(-1), solver.geometry.src.data)
term2 = norm(rec) ** 2
info('<Ax,y>: %f, <x, A^Ty>: %f, difference: %4.4e, ratio: %f'
% (term1, term2, (term1 - term2)/term1, term1 / term2))
assert np.isclose((term1 - term2)/term1, 0., atol=1.e-12)
def test_elastic_stability(shape):
spacing = tuple([20] * len(shape))
_, _, _, [rec1, rec2, v, tau] = run(shape=shape,
spacing=spacing,
tn=20000.0,
nbl=0)
assert np.isfinite(norm(rec1))
def test_isoacoustic_stability(shape, k):
spacing = tuple([20] * len(shape))
_, _, _, [rec, _] = run(shape=shape,
spacing=spacing,
tn=20000.0,
nbl=0,
kernel=k)
assert np.isfinite(norm(rec))
def test_tti_stability(shape, kernel):
spacing = tuple([20] * len(shape))
_, _, _, [rec, _, _] = run(shape=shape,
spacing=spacing,
kernel=kernel,
tn=16000.0,
nbl=0)
assert np.isfinite(norm(rec))
def test_viscoelastic_stability(ndim):
shape = tuple([11] * ndim)
spacing = tuple([20] * ndim)
_, _, _, [rec1, rec2, v, tau] = run(shape=shape,
spacing=spacing,
tn=20000.0,
nbl=0)
assert np.isfinite(norm(rec1))
def test_tti_stability(kernel, ndim):
shape = tuple([11] * ndim)
spacing = tuple([20] * ndim)
_, _, _, [rec, _, _] = run(shape=shape,
spacing=spacing,
kernel=kernel,
tn=16000.0,
nbl=0)
assert np.isfinite(norm(rec))
def test_viscoacoustic_stability(ndim, kernel):
shape = tuple([11] * ndim)
spacing = tuple([20] * ndim)
_, _, _, [rec] = run(shape=shape,
spacing=spacing,
tn=20000.0,
nbl=0,
kernel=kernel)
assert np.isfinite(norm(rec))
def iso_acoustic(self, opt):
shape = (101, 101)
extent = (1000, 1000)
origin = (0., 0.)
v = np.empty(shape, dtype=np.float32)
v[:, :51] = 1.5
v[:, 51:] = 2.5
grid = Grid(shape=shape, extent=extent, origin=origin)
t0 = 0.
tn = 1000.
dt = 1.6
time_range = TimeAxis(start=t0, stop=tn, step=dt)
f0 = 0.010
src = RickerSource(name='src',
grid=grid,
f0=f0,
npoint=1,
time_range=time_range)
domain_size = np.array(extent)
src.coordinates.data[0, :] = domain_size * .5
src.coordinates.data[0, -1] = 20.
rec = Receiver(name='rec',
grid=grid,
npoint=101,
time_range=time_range)
rec.coordinates.data[:, 0] = np.linspace(0, domain_size[0], num=101)
rec.coordinates.data[:, 1] = 20.
u = TimeFunction(name="u", grid=grid, time_order=2, space_order=2)
m = Function(name='m', grid=grid)
m.data[:] = 1. / (v * v)
pde = m * u.dt2 - u.laplace
stencil = Eq(u.forward, solve(pde, u.forward))
src_term = src.inject(field=u.forward, expr=src * dt**2 / m)
rec_term = rec.interpolate(expr=u.forward)
op = Operator([stencil] + src_term + rec_term,
opt=opt,
language='openmp')
# Make sure we've indeed generated OpenMP offloading code
assert 'omp target' in str(op)
op(time=time_range.num - 1, dt=dt)
assert np.isclose(norm(rec), 490.55, atol=1e-2, rtol=0)
def run(problem, **kwargs):
"""
A single run with a specific set of performance parameters.
"""
setup = model_type[problem]['setup']
options = {}
time_order = kwargs.pop('time_order')[0]
space_order = kwargs.pop('space_order')[0]
autotune = kwargs.pop('autotune')
options['autotune'] = autotune
block_shapes = as_tuple(kwargs.pop('block_shape'))
operator = kwargs.pop('operator', 'forward')
# Should a specific block-shape be used? Useful if one wants to skip
# the autotuning pass as a good block-shape is already known
# Note: the following piece of code is horribly *hacky*, but it works for now
for i, block_shape in enumerate(block_shapes):
for n, level in enumerate(block_shape):
for d, s in zip(['x', 'y', 'z'], level):
options['%s%d_blk%d_size' % (d, i, n)] = s
solver = setup(space_order=space_order, time_order=time_order, **kwargs)
retval = run_op(solver, operator, **options)
try:
rank = MPI.COMM_WORLD.rank
except AttributeError:
# MPI not available
rank = 0
dumpfile = kwargs.pop('dump_summary')
if dumpfile:
if configuration['profiling'] != 'advanced':
raise RuntimeError(
"Must set DEVITO_PROFILING=advanced (or, alternatively, "
"DEVITO_LOGGING=PERF) with --dump-summary")
if rank == 0:
with open(dumpfile, 'w') as f:
summary = retval[-1]
assert isinstance(summary, PerformanceSummary)
f.write(str(summary.globals['fdlike']))
dumpfile = kwargs.pop('dump_norms')
if dumpfile:
norms = [
"'%s': %f" % (i.name, norm(i)) for i in retval[:-1]
if isinstance(i, DiscreteFunction)
]
if rank == 0:
with open(dumpfile, 'w') as f:
f.write("{%s}" % ', '.join(norms))
return retval
def test_adjoint_J(self, mkey, shape, kernel, space_order, setup_func):
"""
Adjoint test for the FWI Jacobian operator.
The Jacobian operator J generates a linearized shot record (measurements)
from a model perturbation dm while the adjoint of J generates the FWI gradient
from an adjoint source (usually data residual). This test uses the conventional
dot test:
< Jx, y> = <x ,J^T y>
"""
tn = 500. # Final time
nbl = 10 + space_order / 2
spacing = tuple([10.] * len(shape))
# Create solver from preset
solver = setup_func(shape=shape,
spacing=spacing,
vp_bottom=2,
kernel=kernel,
nbl=nbl,
tn=tn,
space_order=space_order,
**(presets[mkey]),
dtype=np.float64)
# Create initial model (m0) with a constant velocity throughout
model0 = demo_model(**(presets[mkey]),
vp_top=1.5,
vp_bottom=1.5,
spacing=spacing,
space_order=space_order,
shape=shape,
nbl=nbl,
dtype=np.float64,
grid=solver.model.grid)
# Compute initial born perturbation from m - m0
dm = (solver.model.vp.data**(-2) - model0.vp.data**(-2))
du = solver.jacobian(dm, model=model0)[0]
# Compute the full bg field(s) & gradient from initial perturbation
if setup_func is tti_setup:
u0, v0 = solver.forward(save=True, model=model0)[1:-1]
im, _ = solver.jacobian_adjoint(du, u0, v0, model=model0)
else:
u0 = solver.forward(save=True, model=model0)[1]
im, _ = solver.jacobian_adjoint(du, u0, model=model0)
# Adjoint test: Verify <Ax,y> matches <x, A^Ty> closely
term1 = np.dot(im.data.reshape(-1), dm.reshape(-1))
term2 = norm(du)**2
info('<x, J^Ty>: %f, <Jx,y>: %f, difference: %4.4e, ratio: %f' %
(term1, term2, (term1 - term2) / term1, term1 / term2))
assert np.isclose((term1 - term2) / term1, 0., atol=1.e-12)
def test_adjoint_J(self, shape, space_order):
"""
Adjoint test for the FWI Jacobian operator.
The Jacobian operator J generates a linearized shot record (measurements)
from a model perturbation dm while the adjoint of J generates the FWI gradient
from an adjoint source (usually data residual). This test uses the conventional
dot test:
< Jx, y> = <x ,J^T y>
"""
tn = 500. # Final time
nbl = 10 + space_order / 2
spacing = tuple([10.] * len(shape))
# Create solver from preset
solver = acoustic_setup(shape=shape,
spacing=spacing,
nlayers=2,
vp_bottom=2,
nbl=nbl,
tn=tn,
space_order=space_order,
preset='layers-isotropic',
dtype=np.float64)
# Create initial model (m0) with a constant velocity throughout
model0 = demo_model('layers-isotropic',
vp_top=1.5,
vp_bottom=1.5,
spacing=spacing,
space_order=space_order,
shape=shape,
nbl=nbl,
dtype=np.float64,
grid=solver.model.grid)
# Compute the full wavefield u0
_, u0, _ = solver.forward(save=True, vp=model0.vp)
# Compute initial born perturbation from m - m0
dm = (solver.model.vp.data**(-2) - model0.vp.data**(-2))
du, _, _, _ = solver.born(dm, vp=model0.vp)
# Compute gradientfrom initial perturbation
im, _ = solver.gradient(du, u0, vp=model0.vp)
# Adjoint test: Verify <Ax,y> matches <x, A^Ty> closely
term1 = np.dot(im.data.reshape(-1), dm.reshape(-1))
term2 = norm(du)**2
info('<Jx,y>: %f, <x, J^Ty>: %f, difference: %4.4e, ratio: %f' %
(term1, term2, (term1 - term2) / term1, term1 / term2))
assert np.isclose((term1 - term2) / term1, 0., atol=1.e-12)
def run(problem, **kwargs):
"""
A single run with a specific set of performance parameters.
"""
setup = model_type[problem]['setup']
options = {}
time_order = kwargs.pop('time_order')[0]
space_order = kwargs.pop('space_order')[0]
autotune = kwargs.pop('autotune')
block_shapes = as_tuple(kwargs.pop('block_shape'))
# Should a specific block-shape be used? Useful if one wants to skip
# the autotuning pass as a good block-shape is already known
# Note: the following piece of code is horribly *hacky*, but it works for now
for i, block_shape in enumerate(block_shapes):
for n, level in enumerate(block_shape):
for d, s in zip(['x', 'y', 'z'], level):
options['%s%d_blk%d_size' % (d, i, n)] = s
solver = setup(space_order=space_order, time_order=time_order, **kwargs)
retval = solver.forward(autotune=autotune, **options)
# With MPI, only rank0 writes to disk
try:
rank = MPI.COMM_WORLD.rank
except TypeError:
# MPI not available
rank = 0
if rank == 0:
dumpfile = kwargs.pop('dump_summary')
if dumpfile:
with open(dumpfile, 'w') as f:
summary = retval[-1]
assert isinstance(summary, PerformanceSummary)
f.write(str(summary.globals['fdlike']))
dumpfile = kwargs.pop('dump_norms')
if dumpfile:
norms = [
"'%s': %f" % (i.name, norm(i)) for i in retval[:-1]
if isinstance(i, DiscreteFunction)
]
with open(dumpfile, 'w') as f:
f.write("{%s}" % ', '.join(norms))
return retval
def cli_run_jit_backdoor(problem, **kwargs):
"""`click` interface for the `run_jit_backdoor` mode in `benchmark.py`."""
# Preset shared parameters
kwargs['space_order'] = [12]
kwargs['time_order'] = [2]
kwargs['nbl'] = 10
kwargs['spacing'] = (20.0, 20.0, 20.0)
# Preset problem-specific parameters
if problem == 'tti':
kwargs['shape'] = (350, 350, 350)
kwargs['tn'] = 50 # End time of the simulation in ms
kwargs['dse'] = 'aggressive'
# Reference norms for the output fields
reference = {'rec': 66.417102, 'u': 30.707737, 'v': 30.707728}
elif problem == 'acoustic':
kwargs['shape'] = (492, 492, 492)
kwargs['tn'] = 100 # End time of the simulation in ms
kwargs['dse'] = 'advanced'
# Reference norms for the output fields
reference = {'rec': 184.526400, 'u': 151.545837}
else:
assert False
# Dummy values as they will be unused
kwargs['block_shape'] = []
kwargs['autotune'] = 'off'
retval = run_jit_backdoor(problem, **kwargs)
if retval is not None:
for i in retval:
if isinstance(i, DiscreteFunction):
v = norm(i)
info(
"norm(%s) = %f (expected = %f, delta = %f)" %
(i.name, v, reference[i.name], abs(v - reference[i.name])))
# Record DEVITO_ environment
env = [(k, v) for k, v in os.environ.items() if k.startswith('DEVITO_')]
content = "{%s}" % ", ".join("'%s': '%s'" % (k, v) for k, v in env)
with open('env.py', 'w') as f:
f.write(content)
def run_jit_backdoor(problem, **kwargs):
"""
A single run using the DEVITO_JIT_BACKDOOR to test kernel customization.
"""
configuration['develop-mode'] = False
setup = model_type[problem]['setup']
time_order = kwargs.pop('time_order')[0]
space_order = kwargs.pop('space_order')[0]
autotune = kwargs.pop('autotune')
info("Preparing simulation...")
solver = setup(space_order=space_order, time_order=time_order, **kwargs)
# Generate code (but do not JIT yet)
op = solver.op_fwd()
# Get the filename in the JIT cache
cfile = "%s.c" % str(op._compiler.get_jit_dir().joinpath(op._soname))
if not os.path.exists(cfile):
# First time we run this problem, let's generate and jit-compile code
op.cfunction
info("You may now edit the generated code in `%s`. "
"Then save the file, and re-run this benchmark." % cfile)
return
info("Running wave propagation Operator...")
@switchconfig(jit_backdoor=True)
def _run_jit_backdoor():
return run_op(solver, 'forward', autotune=autotune)
retval = _run_jit_backdoor()
dumpnorms = kwargs.pop('dump_norms')
if dumpnorms:
for i in retval[:-1]:
if isinstance(i, DiscreteFunction):
info("'%s': %f" % (i.name, norm(i)))
return retval
def forward(self,
src=None,
rec=None,
u=None,
v=None,
vp=None,
epsilon=None,
delta=None,
theta=None,
phi=None,
save=False,
kernel='centered',
**kwargs):
"""
Forward modelling function that creates the necessary
data objects for running a forward modelling operator.
Parameters
----------
geometry : AcquisitionGeometry
Geometry object that contains the source (SparseTimeFunction) and
receivers (SparseTimeFunction) and their position.
u : TimeFunction, optional
The computed wavefield first component.
v : TimeFunction, optional
The computed wavefield second component.
vp : Function or float, optional
The time-constant velocity.
epsilon : Function or float, optional
The time-constant first Thomsen parameter.
delta : Function or float, optional
The time-constant second Thomsen parameter.
theta : Function or float, optional
The time-constant Dip angle (radians).
phi : Function or float, optional
The time-constant Azimuth angle (radians).
save : bool, optional
Whether or not to save the entire (unrolled) wavefield.
kernel : str, optional
Type of discretization, centered or shifted.
Returns
-------
Receiver, wavefield and performance summary.
"""
if kernel == 'staggered':
time_order = 1
dims = self.model.space_dimensions
stagg_u = (-dims[-1])
stagg_v = (-dims[0],
-dims[1]) if self.model.grid.dim == 3 else (-dims[0])
else:
time_order = 2
stagg_u = stagg_v = None
# Source term is read-only, so re-use the default
src = src or self.geometry.src
# Create a new receiver object to store the result
rec = rec or Receiver(name='rec',
grid=self.model.grid,
time_range=self.geometry.time_axis,
coordinates=self.geometry.rec_positions)
# Create the forward wavefield if not provided
if u is None:
u = TimeFunction(name='u',
grid=self.model.grid,
staggered=stagg_u,
save=self.geometry.nt if save else None,
time_order=time_order,
space_order=self.space_order)
# Create the forward wavefield if not provided
if v is None:
v = TimeFunction(name='v',
grid=self.model.grid,
staggered=stagg_v,
save=self.geometry.nt if save else None,
time_order=time_order,
space_order=self.space_order)
print("Initial Norm u", norm(u))
print("Initial Norm v", norm(v))
if kernel == 'staggered':
vx, vz, vy = particle_velocity_fields(self.model, self.space_order)
kwargs["vx"] = vx
kwargs["vz"] = vz
if vy is not None:
kwargs["vy"] = vy
# Pick vp and Thomsen parameters from model unless explicitly provided
kwargs.update(
self.model.physical_params(vp=vp,
epsilon=epsilon,
delta=delta,
theta=theta,
phi=phi))
if self.model.dim < 3:
kwargs.pop('phi', None)
# Execute operator and return wavefield and receiver data
op = self.op_fwd(kernel, save)
print(kwargs)
summary = op.apply(src=src,
u=u,
v=v,
dt=kwargs.pop('dt', self.dt),
**kwargs)
regnormu = norm(u)
regnormv = norm(v)
print("Norm u:", regnormu)
print("Norm v:", regnormv)
if 0:
cmap = plt.cm.get_cmap("viridis")
values = u.data[0, :, :, :]
vistagrid = pv.UniformGrid()
vistagrid.dimensions = np.array(values.shape) + 1
vistagrid.spacing = (1, 1, 1)
vistagrid.origin = (0, 0, 0
) # The bottom left corner of the data set
vistagrid.cell_arrays["values"] = values.flatten(order="F")
vistaslices = vistagrid.slice_orthogonal()
vistagrid.plot(show_edges=True)
vistaslices.plot(cmap=cmap)
print("=========================================")
s_u = TimeFunction(name='s_u',
grid=self.model.grid,
space_order=self.space_order,
time_order=1)
s_v = TimeFunction(name='s_v',
grid=self.model.grid,
space_order=self.space_order,
time_order=1)
src_u = src.inject(field=s_u.forward,
expr=src * self.model.grid.time_dim.spacing**2 /
self.model.m)
src_v = src.inject(field=s_v.forward,
expr=src * self.model.grid.time_dim.spacing**2 /
self.model.m)
op_f = Operator([src_u, src_v])
op_f.apply(src=src, dt=kwargs.pop('dt', self.dt))
print("Norm s_u", norm(s_u))
print("Norm s_v", norm(s_v))
# Get the nonzero indices
nzinds = np.nonzero(s_u.data[0]) # nzinds is a tuple
assert len(nzinds) == len(self.model.grid.shape)
shape = self.model.grid.shape
x, y, z = self.model.grid.dimensions
time = self.model.grid.time_dim
t = self.model.grid.stepping_dim
source_mask = Function(name='source_mask',
shape=self.model.grid.shape,
dimensions=(x, y, z),
space_order=0,
dtype=np.int32)
source_id = Function(name='source_id',
shape=shape,
dimensions=(x, y, z),
space_order=0,
dtype=np.int32)
print("source_id data indexes start from 0 now !!!")
# source_id.data[nzinds[0], nzinds[1], nzinds[2]] = tuple(np.arange(1, len(nzinds[0])+1))
source_id.data[nzinds[0], nzinds[1],
nzinds[2]] = tuple(np.arange(len(nzinds[0])))
source_mask.data[nzinds[0], nzinds[1], nzinds[2]] = 1
# plot3d(source_mask.data, model)
# import pdb; pdb.set_trace()
print("Number of unique affected points is: %d", len(nzinds[0]))
# Assert that first and last index are as expected
assert (source_id.data[nzinds[0][0], nzinds[1][0], nzinds[2][0]] == 0)
assert (source_id.data[nzinds[0][-1], nzinds[1][-1],
nzinds[2][-1]] == len(nzinds[0]) - 1)
assert (source_id.data[nzinds[0][len(nzinds[0]) - 1],
nzinds[1][len(nzinds[0]) - 1],
nzinds[2][len(nzinds[0]) -
1]] == len(nzinds[0]) - 1)
assert (np.all(np.nonzero(source_id.data)) == np.all(
np.nonzero(source_mask.data)))
assert (np.all(np.nonzero(source_id.data)) == np.all(
np.nonzero(s_u.data[0])))
print(
"-At this point source_mask and source_id have been popoulated correctly-"
)
nnz_shape = (self.model.grid.shape[0], self.model.grid.shape[1])
nnz_sp_source_mask = Function(name='nnz_sp_source_mask',
shape=(list(nnz_shape)),
dimensions=(x, y),
space_order=0,
dtype=np.int32)
nnz_sp_source_mask.data[:, :] = source_mask.data[:, :, :].sum(2)
inds = np.where(source_mask.data == 1.)
print("Grid - source positions:", inds)
maxz = len(np.unique(inds[-1]))
# Change only 3rd dim
sparse_shape = (self.model.grid.shape[0], self.model.grid.shape[1],
maxz)
assert (len(
nnz_sp_source_mask.dimensions) == (len(source_mask.dimensions) -
1))
# Note : sparse_source_id is not needed as long as sparse info is kept in mask
# sp_source_id.data[inds[0],inds[1],:] = inds[2][:maxz]
id_dim = Dimension(name='id_dim')
b_dim = Dimension(name='b_dim')
save_src_u = TimeFunction(name='save_src_u',
shape=(src.shape[0], nzinds[1].shape[0]),
dimensions=(src.dimensions[0], id_dim))
save_src_v = TimeFunction(name='save_src_v',
shape=(src.shape[0], nzinds[1].shape[0]),
dimensions=(src.dimensions[0], id_dim))
save_src_u_term = src.inject(
field=save_src_u[src.dimensions[0], source_id],
expr=src * self.model.grid.time_dim.spacing**2 / self.model.m)
save_src_v_term = src.inject(
field=save_src_v[src.dimensions[0], source_id],
expr=src * self.model.grid.time_dim.spacing**2 / self.model.m)
print("Injecting to empty grids")
op1 = Operator([save_src_u_term, save_src_v_term])
op1.apply(src=src, dt=kwargs.pop('dt', self.dt))
print("Injecting to empty grids finished")
sp_zi = Dimension(name='sp_zi')
sp_source_mask = Function(name='sp_source_mask',
shape=(list(sparse_shape)),
dimensions=(x, y, sp_zi),
space_order=0,
dtype=np.int32)
# Now holds IDs
sp_source_mask.data[inds[0], inds[1], :] = tuple(
inds[-1][:len(np.unique(inds[-1]))])
assert (np.count_nonzero(sp_source_mask.data) == len(nzinds[0]))
assert (len(sp_source_mask.dimensions) == 3)
# import pdb; pdb.set_trace() .
zind = Scalar(name='zind', dtype=np.int32)
xb_size = Scalar(name='xb_size', dtype=np.int32)
yb_size = Scalar(name='yb_size', dtype=np.int32)
x0_blk0_size = Scalar(name='x0_blk0_size', dtype=np.int32)
y0_blk0_size = Scalar(name='y0_blk0_size', dtype=np.int32)
block_sizes = Function(name='block_sizes',
shape=(4, ),
dimensions=(b_dim, ),
space_order=0,
dtype=np.int32)
bsizes = (8, 8, 32, 32)
block_sizes.data[:] = bsizes
# eqxb = Eq(xb_size, block_sizes[0])
# eqyb = Eq(yb_size, block_sizes[1])
# eqxb2 = Eq(x0_blk0_size, block_sizes[2])
# eqyb2 = Eq(y0_blk0_size, block_sizes[3])
eq0 = Eq(sp_zi.symbolic_max,
nnz_sp_source_mask[x, y] - 1,
implicit_dims=(time, x, y))
# eq1 = Eq(zind, sp_source_mask[x, sp_zi], implicit_dims=(time, x, sp_zi))
eq1 = Eq(zind,
sp_source_mask[x, y, sp_zi],
implicit_dims=(time, x, y, sp_zi))
inj_u = source_mask[x, y, zind] * save_src_u[time, source_id[x, y,
zind]]
inj_v = source_mask[x, y, zind] * save_src_v[time, source_id[x, y,
zind]]
eq_u = Inc(u.forward[t + 1, x, y, zind],
inj_u,
implicit_dims=(time, x, y, sp_zi))
eq_v = Inc(v.forward[t + 1, x, y, zind],
inj_v,
implicit_dims=(time, x, y, sp_zi))
# The additional time-tiling equations
# tteqs = (eqxb, eqyb, eqxb2, eqyb2, eq0, eq1, eq_u, eq_v)
performance_map = np.array([[0, 0, 0, 0, 0]])
bxstart = 4
bxend = 17
bystart = 4
byend = 17
bstep = 16
txstart = 8
txend = 9
tystart = 8
tyend = 9
tstep = 16
# Temporal autotuning
for tx in range(txstart, txend, tstep):
# import pdb; pdb.set_trace()
for ty in range(tystart, tyend, tstep):
for bx in range(bxstart, bxend, bstep):
for by in range(bystart, byend, bstep):
block_sizes.data[:] = [tx, ty, bx, by]
eqxb = Eq(xb_size, block_sizes[0])
eqyb = Eq(yb_size, block_sizes[1])
eqxb2 = Eq(x0_blk0_size, block_sizes[2])
eqyb2 = Eq(y0_blk0_size, block_sizes[3])
u.data[:] = 0
v.data[:] = 0
print("-----")
tteqs = (eqxb, eqyb, eqxb2, eqyb2, eq0, eq1, eq_u,
eq_v)
op_tt = self.op_fwd(kernel, save, tteqs)
summary_tt = op_tt.apply(u=u,
v=v,
dt=kwargs.pop('dt', self.dt),
**kwargs)
norm_tt_u = norm(u)
norm_tt_v = norm(v)
print("Norm u:", regnormu)
print("Norm v:", regnormv)
print("Norm(tt_u):", norm_tt_u)
print("Norm(tt_v):", norm_tt_v)
print(
"===Temporal blocking======================================"
)
performance_map = np.append(performance_map, [[
tx, ty, bx, by,
summary_tt.globals['fdlike'].gflopss
]], 0)
print(performance_map)
# tids = np.unique(performance_map[:, 0])
#for tid in tids:
bids = np.where((performance_map[:, 0] == tx)
& (performance_map[:, 1] == ty))
bx_data = np.unique(performance_map[bids, 2])
by_data = np.unique(performance_map[bids, 3])
gptss_data = performance_map[bids, 4]
gptss_data = gptss_data.reshape(len(bx_data), len(by_data))
fig, ax = plt.subplots()
im = ax.imshow(gptss_data)
pause(2)
# We want to show all ticks...
ax.set_xticks(np.arange(len(bx_data)))
ax.set_yticks(np.arange(len(by_data)))
# ... and label them with the respective list entries
ax.set_xticklabels(bx_data)
ax.set_yticklabels(by_data)
ax.set_title(
"Gpts/s for fixed tile size. (Sweeping block sizes)")
fig.tight_layout()
fig.colorbar(im, ax=ax)
# ax = sns.heatmap(gptss_data, linewidth=0.5)
plt.savefig(
str(shape[0]) + str(np.int32(tx)) + str(np.int32(ty)) +
".pdf")
if 0:
cmap = plt.cm.get_cmap("viridis")
values = u.data[0, :, :, :]
vistagrid = pv.UniformGrid()
vistagrid.dimensions = np.array(values.shape) + 1
vistagrid.spacing = (1, 1, 1)
vistagrid.origin = (0, 0, 0
) # The bottom left corner of the data set
vistagrid.cell_arrays["values"] = values.flatten(order="F")
vistaslices = vistagrid.slice_orthogonal()
vistagrid.plot(show_edges=True)
vistaslices.plot(cmap=cmap)
return rec, u, v, summary
ste = 0.9
stepx = (ste - stx) / int(np.sqrt(src.npoint))
src.coordinates.data[:, :2] = np.array(
np.meshgrid(np.arange(stx, ste, stepx), np.arange(
stx, ste, stepx))).T.reshape(-1, 2) * np.array(model.domain_size[:1])
src.coordinates.data[:, -1] = 20 # Depth is 20m
# f : perform source injection on an empty grid
f = TimeFunction(name="f", grid=model.grid, space_order=so, time_order=2)
src_f = src.inject(field=f.forward, expr=src * dt**2 / model.m)
# op_f = Operator([src_f], opt=('advanced', {'openmp': True}))
op_f = Operator([src_f])
op_f.apply(time=time_range.num - 1)
normf = norm(f)
print("==========")
print(normf)
print("===========")
# uref : reference solution
# uref = TimeFunction(name="uref", grid=model.grid, space_order=so, time_order=2)
# src_term_ref = src.inject(field=uref.forward, expr=src * dt**2 / model.m)
# pde_ref = model.m * uref.dt2 - uref.laplace + model.damp * uref.dt
# stencil_ref = Eq(uref.forward, solve(pde_ref, uref.forward))
#Get the nonzero indices
nzinds = np.nonzero(f.data[0]) # nzinds is a tuple
assert len(nzinds) == len(shape)
shape = model.grid.shape
def test_viscoacoustic(dtype):
_, _, _, [rec] = run(dtype=dtype)
assert np.isclose(norm(rec), 18.7749, atol=1e-3, rtol=0)
def test_viscoelastic():
_, _, _, [rec1, rec2, v, tau] = run()
assert np.isclose(norm(rec1), 12.28040, atol=1e-3, rtol=0)
assert np.isclose(norm(rec2), 0.312461, atol=1e-3, rtol=0)
def test_isoacoustic_stability(ndim):
shape = tuple([11]*ndim)
spacing = tuple([20]*ndim)
_, _, _, [rec, _] = run(shape=shape, spacing=spacing, tn=20000.0, nbl=0)
assert np.isfinite(norm(rec))
def J_adjoint_checkpointing(model,
src_coords,
wavelet,
rec_coords,
recin,
space_order=8,
is_residual=False,
n_checkpoints=None,
maxmem=None,
return_obj=False,
isic=False,
ws=None,
t_sub=1):
"""
Jacobian (adjoint fo born modeling operator) operator on a shot record
as a source (i.e data residual). Outputs the gradient with Checkpointing.
Parameters
----------
model: Model
Physical model
src_coords: Array
Coordiantes of the source(s)
wavelet: Array
Source signature
rec_coords: Array
Coordiantes of the receiver(s)
recin: Array
Receiver data
space_order: Int (optional)
Spatial discretization order, defaults to 8
checkpointing: Bool
Whether or not to use checkpointing
n_checkpoints: Int
Number of checkpoints for checkpointing
maxmem: Float
Maximum memory to use for checkpointing
isic : Bool
Whether or not to use ISIC imaging condition
ws : Array
Extended source spatial distribution
is_residual: Bool
Whether to treat the input as the residual or as the observed data
Returns
----------
Array
Adjoint jacobian on the input data (gradient)
"""
# Optimal checkpointing
op_f, u, rec_g = forward(model,
src_coords,
rec_coords,
wavelet,
space_order=space_order,
return_op=True,
ws=ws)
op, g, v = gradient(model,
recin,
rec_coords,
u,
space_order=space_order,
return_op=True,
isic=isic)
nt = wavelet.shape[0]
rec = Receiver(name='rec',
grid=model.grid,
ntime=nt,
coordinates=rec_coords)
cp = DevitoCheckpoint([uu for uu in as_tuple(u)])
if maxmem is not None:
memsize = (cp.size * u.data.itemsize)
n_checkpoints = int(np.floor(maxmem * 10**6 / memsize))
# Op arguments
uk = {uu.name: uu for uu in as_tuple(u)}
vk = {**uk, **{vv.name: vv for vv in as_tuple(v)}}
uk.update({'rcv%s' % as_tuple(u)[0].name: rec_g})
vk.update({'src%s' % as_tuple(v)[0].name: rec})
# Wrapped ops
wrap_fw = CheckpointOperator(op_f, vp=model.vp, **uk)
wrap_rev = CheckpointOperator(op, vp=model.vp, **vk)
# Run forward
wrp = Revolver(cp, wrap_fw, wrap_rev, n_checkpoints, nt - 2)
wrp.apply_forward()
# Residual and gradient
if is_residual is True: # input data is already the residual
rec.data[:] = recin[:]
else:
rec.data[:] = rec.data[:] - recin[:] # input is observed data
wrp.apply_reverse()
if return_obj:
return .5 * model.critical_dt * norm(rec)**2, g.data
return g.data
def test_viscoacoustic(kernel, time_order, normrec, atol):
_, _, _, [rec, _, _] = run(kernel=kernel, time_order=time_order)
assert np.isclose(norm(rec), normrec, atol=atol, rtol=0)
dm=dm, rho=rho, dt=dt_full)
# Time axis
t0 = 0.
tn = 300.
dt = model.critical_dt
time = TimeAxis(start=t0, step=dt_full, stop=tn)
print(time.num)
src = RickerSource(name='src', grid=model.grid, f0=0.020, time_range=time, npoint=1)
src.coordinates.data[:, 0] = 5000.
src.coordinates.data[:, 1] = 350.
nrec = 501
rec_coords = np.empty((nrec, 2))
rec_coords[:, 0] = np.linspace(0., 10000., nrec)
rec_coords[:, 1] = 6.
####### RUN #########
# Forward
op = TTIPropagators(model, space_order=so)
rec, u, v = op.forward(src, rec_coords, save=True)
grad = op.gradient(rec, u, v, isic=True)
linD, u, v, summary = op.born(src, rec_coords, sub=(4, 1), autotune=('aggressive', 'runtime'))
grad2 = op.gradient(linD, u, v, sub=(4, 1), isic=True, autotune=('basic', 'runtime'))
print(norm(grad))
print(norm(grad2))
print(norm(rec))
print(norm(linD))
def test_isoacoustic(fs, normrec, dtype):
_, _, _, [rec, _] = run(fs=fs, dtype=dtype)
assert np.isclose(norm(rec), normrec, rtol=1e-3, atol=0)
def test_elastic(dtype):
_, _, _, [rec1, rec2, v, tau] = run(dtype=dtype)
assert np.isclose(norm(rec1), 19.25636, atol=1e-3, rtol=0)
assert np.isclose(norm(rec2), 0.627606, atol=1e-3, rtol=0)
