def _gen_basic(self, ls, nd):
src = basic_kernel.render(preamble=self.preamble,
reduce_expr=self.reduce_expr,
name="reduk",
out_arg=self.out_arg,
nd=nd,
arguments=self.arguments,
local_size=ls,
redux=self.redux,
neutral=self.neutral,
map_expr=self.expression)
spec = ['uint32', gpuarray.GpuArray]
spec.extend('uint32' for _ in range(nd))
for i, arg in enumerate(self.arguments):
spec.append(arg.spec())
if arg.isarray():
spec.append('uint32')
spec.extend('int32' for _ in range(nd))
k = gpuarray.GpuKernel(src,
"reduk",
spec,
context=self.context,
cluda=True,
**self.flags)
return k, src, spec
def _make_basic(self, nd):
name = "elem_" + str(nd)
src = self.render_basic(nd, name=name)
return gpuarray.GpuKernel(src,
name,
self.argspec_basic(nd),
context=self.context,
cluda=True,
**self.flags)
def __init__(self,
context,
arguments,
operation,
preamble="",
dimspec_limit=2,
spec_limit=10):
if isinstance(arguments, str):
self.arguments = parse_c_args(arguments)
else:
self.arguments = tuple(arguments)
self.operation = operation
self.expression = massage_op(operation)
self.context = context
self._spec_limit = spec_limit
self._dimspec_limit = dimspec_limit
if not any(arg.isarray() for arg in self.arguments):
raise RuntimeError("ElemwiseKernel can only be used with "
"functions that have at least one "
"vector argument.")
have_small = False
have_double = False
have_complex = False
for arg in self.arguments:
if arg.dtype.itemsize < 4 and type(arg) == ArrayArg:
have_small = True
if arg.dtype in [numpy.float64, numpy.complex128]:
have_double = True
if arg.dtype in [numpy.complex64, numpy.complex128]:
have_complex = True
self.flags = dict(have_small=have_small,
have_double=have_double,
have_complex=have_complex)
self.preamble = preamble
self.contig_src = contiguous_kernel.render(preamble=self.preamble,
name="elem_contig",
arguments=self.arguments,
expression=self.operation)
self.contig_k = gpuarray.GpuKernel(self.contig_src,
"elem_contig",
self.argspec_contig(),
context=self.context,
cluda=True,
**self.flags)
self._speckey = None
self._dims = None
def _make_dimspec(self, n, nd, dims):
src = dimspec_kernel.render(preamble=self.preamble,
name="elemk",
n=n,
nd=nd,
dims=dims,
arguments=self.arguments,
expression=self.expression)
return gpuarray.GpuKernel(src,
"elemk",
self.argspec_dimspec(nd),
context=self.context,
cluda=True,
**self.flags)
def _make_specialized(self, n, nd, dims, strs, offsets):
src = specialized_kernel.render(preamble=self.preamble,
name="elemk",
n=n,
nd=nd,
dim=dims,
strs=strs,
arguments=self.arguments,
expression=self.expression,
offsets=offsets)
return gpuarray.GpuKernel(src,
"elemk",
self.argspec_specialized(),
context=self.context,
cluda=True,
**self.flags)