def __init__(self,
tpname,
fields,
declname=None,
align_bytes=None,
aligned_prime_to=[]):
"""Initialize a structure declarator.
*tpname* is the name of the structure, while *declname* is the
name used for the declarator. *pad_bytes* is the number of
padding bytes added at the end of the structure.
*fields* is a list of :class:`Declarator` instances.
*align_bytes* is an integer that causes the structure to be
padded to an integer multiple of itself.
*aligned_prime_to* is a list of integers.
If the resulting structure's size
is ``s``, then ``s//align_bytes`` will be made prime to all
numbers in *aligned_prime_to*. (Sounds obscure? It's needed
for avoiding bank conflicts in CUDA programming.)
"""
format = "".join(f.struct_format() for f in fields)
bytes = _struct.calcsize(format)
natural_align_bytes = max(f.alignment_requirement() for f in fields)
if align_bytes is None:
align_bytes = natural_align_bytes
elif align_bytes < natural_align_bytes:
from warnings import warn
warn("requested struct alignment smaller than natural alignment")
self.align_bytes = align_bytes
padded_bytes = ((bytes + align_bytes - 1) // align_bytes) * align_bytes
def satisfies_primality(n):
from pymbolic.algorithm import gcd
for p in aligned_prime_to:
if gcd(n, p) != 1:
return False
return True
while not satisfies_primality(padded_bytes // align_bytes):
padded_bytes += align_bytes
self.get_init().__init__(self, tpname, fields, declname,
padded_bytes - bytes)
if self.pad_bytes:
self.format = format + "%dx" % self.pad_bytes
self.bytes = padded_bytes
else:
self.format = format
self.bytes = bytes
assert _struct.calcsize(self.format) == self.bytes
def __init__(self, tpname, fields, declname=None,
align_bytes=None, aligned_prime_to=[]):
"""Initialize a structure declarator.
*tpname* is the name of the structure, while *declname* is the
name used for the declarator. *pad_bytes* is the number of
padding bytes added at the end of the structure.
*fields* is a list of :class:`Declarator` instances.
*align_bytes* is an integer that causes the structure to be
padded to an integer multiple of itself.
*aligned_prime_to* is a list of integers.
If the resulting structure's size
is ``s``, then ``s//align_bytes`` will be made prime to all
numbers in *aligned_prime_to*. (Sounds obscure? It's needed
for avoiding bank conflicts in CUDA programming.)
"""
format = "".join(f.struct_format() for f in fields)
bytes = _struct.calcsize(format)
natural_align_bytes = max(f.alignment_requirement() for f in fields)
if align_bytes is None:
align_bytes = natural_align_bytes
elif align_bytes < natural_align_bytes:
from warnings import warn
warn("requested struct alignment smaller than natural alignment")
self.align_bytes = align_bytes
padded_bytes = ((bytes + align_bytes - 1) // align_bytes) * align_bytes
def satisfies_primality(n):
from pymbolic.algorithm import gcd
for p in aligned_prime_to:
if gcd(n, p) != 1:
return False
return True
while not satisfies_primality(padded_bytes // align_bytes):
padded_bytes += align_bytes
Struct.__init__(self, tpname, fields, declname, padded_bytes - bytes)
if self.pad_bytes:
self.format = format + "%dx" % self.pad_bytes
self.bytes = padded_bytes
else:
self.format = format
self.bytes = bytes
assert _struct.calcsize(self.format) == self.bytes
def function_prepare_pre_v4(func, arg_types, block=None, shared=None, texrefs=[]):
from warnings import warn
if block is not None:
warn("setting the block size in Function.prepare is deprecated",
DeprecationWarning, stacklevel=2)
func._set_block_shape(*block)
if shared is not None:
warn("setting the shared memory size in Function.prepare is deprecated",
DeprecationWarning, stacklevel=2)
func._set_shared_size(shared)
func.texrefs = texrefs
func.arg_format = ""
param_size = 0
for i, arg_type in enumerate(arg_types):
if isinstance(arg_type, type) and np is not None and np.number in arg_type.__mro__:
func.arg_format += np.dtype(arg_type).char
elif isinstance(arg_type, str):
func.arg_format += arg_type
else:
func.arg_format += np.dtype(np.intp).char
from pycuda._pvt_struct import calcsize
func._param_set_size(calcsize(func.arg_format))
return func
def function_prepare(func, arg_types, block, shared=None, texrefs=[]):
func.set_block_shape(*block)
if shared is not None:
func.set_shared_size(shared)
func.texrefs = texrefs
try:
import numpy
except ImportError:
numpy = None
func.arg_format = ""
param_size = 0
for i, arg_type in enumerate(arg_types):
if isinstance(
arg_type, type
) and numpy is not None and numpy.number in arg_type.__mro__:
func.arg_format += numpy.dtype(arg_type).char
elif isinstance(arg_type, str):
func.arg_format += arg_type
else:
func.arg_format += numpy.dtype(numpy.intp).char
from pycuda._pvt_struct import calcsize
func.param_set_size(calcsize(func.arg_format))
return func
def function_prepare_pre_v4(func, arg_types, block=None, shared=None, texrefs=[]):
from warnings import warn
if block is not None:
warn("setting the block size in Function.prepare is deprecated",
DeprecationWarning, stacklevel=2)
func._set_block_shape(*block)
if shared is not None:
warn("setting the shared memory size in Function.prepare is deprecated",
DeprecationWarning, stacklevel=2)
func._set_shared_size(shared)
func.texrefs = texrefs
func.arg_format = ""
param_size = 0
for i, arg_type in enumerate(arg_types):
if isinstance(arg_type, type) and np is not None and np.number in arg_type.__mro__:
func.arg_format += np.dtype(arg_type).char
elif isinstance(arg_type, str):
func.arg_format += arg_type
else:
func.arg_format += np.dtype(np.intp).char
from pycuda._pvt_struct import calcsize
func._param_set_size(calcsize(func.arg_format))
return func
def function_prepare(func, arg_types, block, shared=None, texrefs=[]):
func.set_block_shape(*block)
if shared is not None:
func.set_shared_size(shared)
func.texrefs = texrefs
try:
import numpy
except ImportError:
numpy = None
func.arg_format = ""
param_size = 0
for i, arg_type in enumerate(arg_types):
if isinstance(arg_type, type) and numpy is not None and numpy.number in arg_type.__mro__:
func.arg_format += numpy.dtype(arg_type).char
elif isinstance(arg_type, str):
func.arg_format += arg_type
else:
func.arg_format += numpy.dtype(numpy.intp).char
from pycuda._pvt_struct import calcsize
func.param_set_size(calcsize(func.arg_format))
return func
def is_64_bit_platform():
return _struct.calcsize('l') == 8
def alignment_requirement(self):
return _struct.calcsize(self.struct_format())
def is64bit():
return _struct.calcsize('1') == 8
def is_long_64_bit():
return _struct.calcsize('l') == 8