def test_scheduling_sparse_functions(self):
"""Tests loop scheduling in presence of sparse functions."""
grid = Grid((10, 10))
time = grid.time_dim
u1 = TimeFunction(name="u1", grid=grid, save=10, time_order=2)
u2 = TimeFunction(name="u2", grid=grid, time_order=2)
sf1 = SparseFunction(name='sf1', grid=grid, npoint=1, ntime=10)
sf2 = SparseFunction(name='sf2', grid=grid, npoint=1, ntime=10)
# Deliberately inject into u1, rather than u1.forward, to create a WAR w/ eqn3
eqn1 = Eq(u1.forward, u1 + 2.0 - u1.backward)
eqn2 = sf1.inject(u1, expr=sf1)
eqn3 = Eq(u2.forward, u2 + 2*u2.backward - u1.dt2)
eqn4 = sf2.interpolate(u2)
op = Operator([eqn1] + eqn2 + [eqn3] + eqn4)
trees = retrieve_iteration_tree(op)
assert len(trees) == 4
# Time loop not shared due to the WAR
assert trees[0][0].dim is time and trees[0][0] is trees[1][0] # this IS shared
assert trees[1][0] is not trees[2][0]
assert trees[2][0].dim is time and trees[2][0] is trees[3][0] # this IS shared
# Now single, shared time loop expected
eqn2 = sf1.inject(u1.forward, expr=sf1)
op = Operator([eqn1] + eqn2 + [eqn3] + eqn4)
trees = retrieve_iteration_tree(op)
assert len(trees) == 4
assert all(trees[0][0] is i[0] for i in trees)
def test_adjoint_inject_interpolate(self, shape, coords, npoints=19):
"""
Verify that p.inject is the adjoint of p.interpolate for a
devito SparseFunction p
"""
grid = Grid(shape)
a = Function(name="a", grid=grid)
a.data[:] = 0.
c = Function(name='c', grid=grid)
c.data[:] = 27.
assert a.grid == c.grid
# Inject receiver
p = SparseFunction(name="p", grid=grid, npoint=npoints)
for i, r in enumerate(coords):
p.coordinates.data[:, i] = np.linspace(r[0], r[1], npoints)
p.data[:] = 1.2
expr = p.inject(field=a, expr=p)
# Read receiver
p2 = SparseFunction(name="p2", grid=grid, npoint=npoints)
for i, r in enumerate(coords):
p2.coordinates.data[:, i] = np.linspace(r[0], r[1], npoints)
expr2 = p2.interpolate(expr=c)
Operator(expr + expr2)(a=a, c=c)
# < P x, y > - < x, P^T y>
# Px => p2
# y => p
# x => c
# P^T y => a
term1 = np.dot(p2.data.reshape(-1), p.data.reshape(-1))
term2 = np.dot(c.data.reshape(-1), a.data.reshape(-1))
assert np.isclose((term1-term2) / term1, 0., atol=1.e-6)
def test_override_composite_data(self):
i, j = dimify('i j')
grid = Grid(shape=(10, 10), dimensions=(i, j))
original_coords = (1., 1.)
new_coords = (2., 2.)
p_dim = Dimension('p_src')
u = TimeFunction(name='u', grid=grid, time_order=2, space_order=2)
src1 = SparseFunction(name='src1',
grid=grid,
dimensions=[time, p_dim],
npoint=1,
nt=10,
coordinates=original_coords)
src2 = SparseFunction(name='src1',
grid=grid,
dimensions=[time, p_dim],
npoint=1,
nt=10,
coordinates=new_coords)
op = Operator(src1.inject(u, src1))
# Move the source from the location where the setup put it so we can test
# whether the override picks up the original coordinates or the changed ones
# Operator.arguments() returns a tuple of (data, dimension_sizes)
args = op.arguments(src1=src2)[0]
arg_name = src1.name + "_coords"
assert (np.array_equal(args[arg_name], np.asarray((new_coords, ))))
def test_injection_wodup(self):
"""
Test injection operator when the sparse points don't need to be replicated
("wodup" -> w/o duplication) over multiple MPI ranks.
"""
grid = Grid(shape=(4, 4), extent=(3.0, 3.0))
f = Function(name='f', grid=grid, space_order=0)
f.data[:] = 0.
if grid.distributor.myrank == 0:
coords = [(0.5, 0.5), (0.5, 2.5), (2.5, 0.5), (2.5, 2.5)]
else:
coords = []
sf = SparseFunction(name='sf',
grid=grid,
npoint=len(coords),
coordinates=coords)
sf.data[:] = 4.
# This is the situation at this point
# O is a grid point
# * is a sparse point
#
# O --- O --- O --- O
# | * | | * |
# O --- O --- O --- O
# | | | |
# O --- O --- O --- O
# | * | | * |
# O --- O --- O --- O
op = Operator(sf.inject(field=f, expr=sf + 1))
op.apply()
assert np.all(f.data == 1.25)
def test_injection_dup(self):
"""
Test injection operator when the sparse points are replicated over
multiple MPI ranks.
"""
grid = Grid(shape=(4, 4), extent=(3.0, 3.0))
x, y = grid.dimensions
f = Function(name='f', grid=grid)
f.data[:] = 0.
if grid.distributor.myrank == 0:
coords = [(0.5, 0.5), (1.5, 2.5), (1.5, 1.5), (2.5, 1.5)]
else:
coords = []
sf = SparseFunction(name='sf',
grid=grid,
npoint=len(coords),
coordinates=coords)
sf.data[:] = 4.
# Global view (left) and local view (right, after domain decomposition)
# O is a grid point
# x is a halo point
# A, B, C, D are sparse points
# Rank0 Rank1
# O --- O --- O --- O O --- O --- x x --- O --- O
# | A | | | | A | | | | |
# O --- O --- O --- O O --- O --- x x --- O --- O
# | | C | B | --> | | C | | C | B |
# O --- O --- O --- O x --- x --- x x --- x --- x
# | | D | | Rank2 Rank3
# O --- O --- O --- O x --- x --- x x --- x --- x
# | | C | | C | B |
# O --- O --- x x --- O --- O
# | | D | | D | |
# O --- O --- x x --- O --- O
#
# Expected `f.data` (global view)
#
# 1.25 --- 1.25 --- 0.00 --- 0.00
# | | | |
# 1.25 --- 2.50 --- 2.50 --- 1.25
# | | | |
# 0.00 --- 2.50 --- 3.75 --- 1.25
# | | | |
# 0.00 --- 1.25 --- 1.25 --- 0.00
op = Operator(sf.inject(field=f, expr=sf + 1))
op.apply()
glb_pos_map = grid.distributor.glb_pos_map
if LEFT in glb_pos_map[x] and LEFT in glb_pos_map[y]: # rank0
assert np.all(f.data_ro_domain == [[1.25, 1.25], [1.25, 2.5]])
elif LEFT in glb_pos_map[x] and RIGHT in glb_pos_map[y]: # rank1
assert np.all(f.data_ro_domain == [[0., 0.], [2.5, 1.25]])
elif RIGHT in glb_pos_map[x] and LEFT in glb_pos_map[y]:
assert np.all(f.data_ro_domain == [[0., 2.5], [0., 1.25]])
elif RIGHT in glb_pos_map[x] and RIGHT in glb_pos_map[y]:
assert np.all(f.data_ro_domain == [[3.75, 1.25], [1.25, 0.]])
def test_sparse_function(self, operate_on_empty_cache):
"""Test caching of SparseFunctions and children objects."""
grid = Grid(shape=(3, 3))
init_cache_size = len(_SymbolCache)
cur_cache_size = len(_SymbolCache)
u = SparseFunction(name='u', grid=grid, npoint=1, nt=10)
# created: u, u(inds), p_u, h_p_u, u_coords, u_coords(inds), d, h_d
ncreated = 8
assert len(_SymbolCache) == cur_cache_size + ncreated
cur_cache_size = len(_SymbolCache)
i = u.inject(expr=u, field=u)
# created: ii_u_0*2 (Symbol and ConditionalDimension), ii_u_1*2, ii_u_2*2,
# ii_u_3*2, px, py, posx, posy, u_coords (as indexified),
ncreated = 2+2+2+2+2+1+1+1
# Note that injection is now lazy so no new symbols should be created
assert len(_SymbolCache) == cur_cache_size
i.evaluate
assert len(_SymbolCache) == cur_cache_size + ncreated
# No new symbolic obejcts are created
u.inject(expr=u, field=u)
assert len(_SymbolCache) == cur_cache_size + ncreated
# Let's look at clear_cache now
del u
del i
clear_cache()
# At this point, not all children objects have been cleared. In particular, the
# ii_u_* Symbols are still alive, as well as p_u and h_p_u. This is because
# in the first clear_cache they were still referenced by their "parent" objects
# (e.g., ii_u_* by ConditionalDimensions, through `condition`)
assert len(_SymbolCache) == init_cache_size + 8
clear_cache()
# Now we should be back to the original state
assert len(_SymbolCache) == init_cache_size
def test_scheduling_after_rewrite():
"""Tests loop scheduling after DSE-induced expression hoisting."""
grid = Grid((10, 10))
u1 = TimeFunction(name="u1", grid=grid, save=10, time_order=2)
u2 = TimeFunction(name="u2", grid=grid, time_order=2)
sf1 = SparseFunction(name='sf1', grid=grid, npoint=1, ntime=10)
const = Function(name="const", grid=grid, space_order=2)
# Deliberately inject into u1, rather than u1.forward, to create a WAR
eqn1 = Eq(u1.forward, u1 + sin(const))
eqn2 = sf1.inject(u1.forward, expr=sf1)
eqn3 = Eq(u2.forward, u2 - u1.dt2 + sin(const))
op = Operator([eqn1] + eqn2 + [eqn3])
trees = retrieve_iteration_tree(op)
# Check loop nest structure
assert len(trees) == 4
assert all(i.dim == j for i, j in zip(trees[0], grid.dimensions)) # time invariant
assert trees[1][0].dim == trees[2][0].dim == trees[3][0].dim == grid.time_dim