def is_foldable(nodes):
"""
Return True if the iterable ``nodes`` consists of foldable Iterations,
False otherwise.
"""
nodes = as_tuple(nodes)
if len(nodes) <= 1 or any(not i.is_Iteration for i in nodes):
return False
if not all({PARALLEL, AFFINE}.issubset(set(i.properties)) for i in nodes):
return False
return (all_equal(i.dim for i in nodes) and
all_equal(i.limits for i in nodes) and
all_equal(i.index for i in nodes))
def config_blockshape(ctx, param, value):
if value:
# Block innermost loops if a full block shape is provided
configuration['dle-options']['blockinner'] = True
# Normalize value:
# 1. integers, not strings
# 2. sanity check the (hierarchical) blocking shape
normalized_value = []
for i, block_shape in enumerate(value):
# If hierarchical blocking is activated, say with N levels, here in
# `bs` we expect to see 3*N entries
bs = [int(x) for x in block_shape.split()]
levels = [bs[x:x + 3] for x in range(0, len(bs), 3)]
if any(len(level) != 3 for level in levels):
raise ValueError(
"Expected 3 entries per block shape level, but got "
"one level with less than 3 entries (`%s`)" % levels)
normalized_value.append(levels)
if not all_equal(len(i) for i in normalized_value):
raise ValueError(
"Found different block shapes with incompatible "
"number of levels (`%s`)" % normalized_value)
configuration['dle-options']['blocklevels'] = len(
normalized_value[0])
else:
normalized_value = []
return tuple(normalized_value)
def config_blockshape(ctx, param, value):
if value:
# Block innermost loops if a full block shape is provided
# Note: see https://github.com/devitocodes/devito/issues/320 for why
# we use blockinner=True only if the backend compiler is Intel
flag = isinstance(configuration['compiler'], IntelCompiler)
configuration['opt-options']['blockinner'] = flag
# Normalize value:
# 1. integers, not strings
# 2. sanity check the (hierarchical) blocking shape
normalized_value = []
for i, block_shape in enumerate(value):
# If hierarchical blocking is activated, say with N levels, here in
# `bs` we expect to see 3*N entries
bs = [int(x) for x in block_shape.split()]
levels = [bs[x:x + 3] for x in range(0, len(bs), 3)]
if any(len(level) != 3 for level in levels):
raise ValueError(
"Expected 3 entries per block shape level, but got "
"one level with less than 3 entries (`%s`)" % levels)
normalized_value.append(levels)
if not all_equal(len(i) for i in normalized_value):
raise ValueError(
"Found different block shapes with incompatible "
"number of levels (`%s`)" % normalized_value)
configuration['opt-options']['blocklevels'] = len(
normalized_value[0])
else:
normalized_value = []
return tuple(normalized_value)
def homogenise_gpus(gpu_infos):
if gpu_infos == []:
warning('No graphics cards detected')
return None
if all_equal(gpu_infos):
gpu_infos[0]['ncards'] = len(gpu_infos)
return gpu_infos[0]
warning('Different models of graphics cards detected')
return None
def homogenise_gpus(gpu_infos):
if gpu_infos == []:
warning('No graphics cards detected')
return {}
# Check must ignore physical IDs as they may differ
for gpu_info in gpu_infos:
gpu_info.pop('physicalid', None)
if all_equal(gpu_infos):
gpu_infos[0]['ncards'] = len(gpu_infos)
return gpu_infos[0]
warning('Different models of graphics cards detected')
return {}
def homogenise_gpus(gpu_infos):
"""
Run homogeneity checks on a list of GPUs, return GPU with count if
homogeneous, otherwise None.
"""
if gpu_infos == []:
warning('No graphics cards detected')
return {}
# Check must ignore physical IDs as they may differ
for gpu_info in gpu_infos:
gpu_info.pop('physicalid', None)
if all_equal(gpu_infos):
gpu_infos[0]['ncards'] = len(gpu_infos)
return gpu_infos[0]
warning('Different models of graphics cards detected')
return {}
