def _prepare_args(self, ty, val, stream, retr, kernelargs):
"""
Convert arguments to ctypes and append to kernelargs
"""
if isinstance(ty, types.Array):
devary, conv = devicearray.auto_device(val, stream=stream)
if conv:
retr.append(lambda: devary.copy_to_host(val, stream=stream))
c_intp = ctypes.c_ssize_t
meminfo = ctypes.c_void_p(0)
parent = ctypes.c_void_p(0)
nitems = c_intp(devary.size)
itemsize = c_intp(devary.dtype.itemsize)
data = ctypes.c_void_p(driver.device_pointer(devary))
kernelargs.append(meminfo)
kernelargs.append(parent)
kernelargs.append(nitems)
kernelargs.append(itemsize)
kernelargs.append(data)
for ax in range(devary.ndim):
kernelargs.append(c_intp(devary.shape[ax]))
for ax in range(devary.ndim):
kernelargs.append(c_intp(devary.strides[ax]))
elif isinstance(ty, types.Integer):
cval = getattr(ctypes, "c_%s" % ty)(val)
kernelargs.append(cval)
elif ty == types.float64:
cval = ctypes.c_double(val)
kernelargs.append(cval)
elif ty == types.float32:
cval = ctypes.c_float(val)
kernelargs.append(cval)
elif ty == types.boolean:
cval = ctypes.c_uint8(int(val))
kernelargs.append(cval)
elif ty == types.complex64:
kernelargs.append(ctypes.c_float(val.real))
kernelargs.append(ctypes.c_float(val.imag))
elif ty == types.complex128:
kernelargs.append(ctypes.c_double(val.real))
kernelargs.append(ctypes.c_double(val.imag))
elif isinstance(ty, types.Record):
devrec, conv = devicearray.auto_device(val, stream=stream)
if conv:
retr.append(lambda: devrec.copy_to_host(val, stream=stream))
kernelargs.append(devrec)
else:
raise NotImplementedError(ty, val)
def _prepare_args(self, ty, val, stream, retr, kernelargs):
"""
Convert arguments to ctypes and append to kernelargs
"""
if isinstance(ty, types.Array):
devary, conv = devicearray.auto_device(val, stream=stream)
if conv:
retr.append(lambda: devary.copy_to_host(val, stream=stream))
c_intp = ctypes.c_ssize_t
meminfo = ctypes.c_void_p(0)
parent = ctypes.c_void_p(0)
nitems = c_intp(devary.size)
itemsize = c_intp(devary.dtype.itemsize)
data = ctypes.c_void_p(driver.device_pointer(devary))
kernelargs.append(meminfo)
kernelargs.append(parent)
kernelargs.append(nitems)
kernelargs.append(itemsize)
kernelargs.append(data)
for ax in range(devary.ndim):
kernelargs.append(c_intp(devary.shape[ax]))
for ax in range(devary.ndim):
kernelargs.append(c_intp(devary.strides[ax]))
elif isinstance(ty, types.Integer):
cval = getattr(ctypes, "c_%s" % ty)(val)
kernelargs.append(cval)
elif ty == types.float64:
cval = ctypes.c_double(val)
kernelargs.append(cval)
elif ty == types.float32:
cval = ctypes.c_float(val)
kernelargs.append(cval)
elif ty == types.boolean:
cval = ctypes.c_uint8(int(val))
kernelargs.append(cval)
elif ty == types.complex64:
kernelargs.append(ctypes.c_float(val.real))
kernelargs.append(ctypes.c_float(val.imag))
elif ty == types.complex128:
kernelargs.append(ctypes.c_double(val.real))
kernelargs.append(ctypes.c_double(val.imag))
elif isinstance(ty, types.Record):
devrec, conv = devicearray.auto_device(val, stream=stream)
if conv:
retr.append(lambda: devrec.copy_to_host(val, stream=stream))
kernelargs.append(devrec)
else:
raise NotImplementedError(ty, val)
def _unpack_argument(ty, val, kernelargs, retr):
"""
Convert arguments to ctypes and append to kernelargs
"""
if isinstance(ty, types.Array):
c_intp = ctypes.c_ssize_t
# if a dgpu is present, move the data to the device.
if dgpu_present:
devary, conv = devicearray.auto_device(val, devices.get_context())
if conv:
retr.append(lambda: devary.copy_to_host(val))
data = devary.device_ctypes_pointer
else:
data = ctypes.c_void_p(val.ctypes.data)
meminfo = parent = ctypes.c_void_p(0)
nitems = c_intp(val.size)
itemsize = c_intp(val.dtype.itemsize)
kernelargs.append(meminfo)
kernelargs.append(parent)
kernelargs.append(nitems)
kernelargs.append(itemsize)
kernelargs.append(data)
for ax in range(val.ndim):
kernelargs.append(c_intp(val.shape[ax]))
for ax in range(val.ndim):
kernelargs.append(c_intp(val.strides[ax]))
elif isinstance(ty, types.Integer):
cval = getattr(ctypes, "c_%s" % ty)(val)
kernelargs.append(cval)
elif ty == types.float64:
cval = ctypes.c_double(val)
kernelargs.append(cval)
elif ty == types.float32:
cval = ctypes.c_float(val)
kernelargs.append(cval)
elif ty == types.boolean:
cval = ctypes.c_uint8(int(val))
kernelargs.append(cval)
elif ty == types.complex64:
kernelargs.append(ctypes.c_float(val.real))
kernelargs.append(ctypes.c_float(val.imag))
elif ty == types.complex128:
kernelargs.append(ctypes.c_double(val.real))
kernelargs.append(ctypes.c_double(val.imag))
else:
raise NotImplementedError(ty, val)
def _unpack_argument(ty, val, kernelargs, retr):
"""
Convert arguments to ctypes and append to kernelargs
"""
if isinstance(ty, types.Array):
c_intp = ctypes.c_ssize_t
# if a dgpu is present, move the data to the device.
if dgpu_present:
devary, conv = devicearray.auto_device(val, devices.get_context())
if conv:
retr.append(lambda: devary.copy_to_host(val))
data = devary.device_ctypes_pointer
else:
data = ctypes.c_void_p(val.ctypes.data)
meminfo = parent = ctypes.c_void_p(0)
nitems = c_intp(val.size)
itemsize = c_intp(val.dtype.itemsize)
kernelargs.append(meminfo)
kernelargs.append(parent)
kernelargs.append(nitems)
kernelargs.append(itemsize)
kernelargs.append(data)
for ax in range(val.ndim):
kernelargs.append(c_intp(val.shape[ax]))
for ax in range(val.ndim):
kernelargs.append(c_intp(val.strides[ax]))
elif isinstance(ty, types.Integer):
cval = getattr(ctypes, "c_%s" % ty)(val)
kernelargs.append(cval)
elif ty == types.float64:
cval = ctypes.c_double(val)
kernelargs.append(cval)
elif ty == types.float32:
cval = ctypes.c_float(val)
kernelargs.append(cval)
elif ty == types.boolean:
cval = ctypes.c_uint8(int(val))
kernelargs.append(cval)
elif ty == types.complex64:
kernelargs.append(ctypes.c_float(val.real))
kernelargs.append(ctypes.c_float(val.imag))
elif ty == types.complex128:
kernelargs.append(ctypes.c_double(val.real))
kernelargs.append(ctypes.c_double(val.imag))
else:
raise NotImplementedError(ty, val)
def _unpack_argument(ty, val, kernelargs):
"""
Convert arguments to ctypes and append to kernelargs
"""
if isinstance(ty, types.Array):
c_intp = ctypes.c_ssize_t
meminfo = parent = ctypes.c_void_p(0)
nitems = c_intp(val.size)
itemsize = c_intp(val.dtype.itemsize)
data = ctypes.c_void_p(val.ctypes.data)
kernelargs.append(meminfo)
kernelargs.append(parent)
kernelargs.append(nitems)
kernelargs.append(itemsize)
kernelargs.append(data)
for ax in range(val.ndim):
kernelargs.append(c_intp(val.shape[ax]))
for ax in range(val.ndim):
kernelargs.append(c_intp(val.strides[ax]))
elif isinstance(ty, types.Integer):
cval = getattr(ctypes, "c_%s" % ty)(val)
kernelargs.append(cval)
elif ty == types.float64:
cval = ctypes.c_double(val)
kernelargs.append(cval)
elif ty == types.float32:
cval = ctypes.c_float(val)
kernelargs.append(cval)
elif ty == types.boolean:
cval = ctypes.c_uint8(int(val))
kernelargs.append(cval)
elif ty == types.complex64:
kernelargs.append(ctypes.c_float(val.real))
kernelargs.append(ctypes.c_float(val.imag))
elif ty == types.complex128:
kernelargs.append(ctypes.c_double(val.real))
kernelargs.append(ctypes.c_double(val.imag))
else:
raise NotImplementedError(ty, val)
def _unpack_argument(ty, val, kernelargs):
"""
Convert arguments to ctypes and append to kernelargs
"""
if isinstance(ty, types.Array):
c_intp = ctypes.c_ssize_t
meminfo = parent = ctypes.c_void_p(0)
nitems = c_intp(val.size)
itemsize = c_intp(val.dtype.itemsize)
data = ctypes.c_void_p(val.ctypes.data)
kernelargs.append(meminfo)
kernelargs.append(parent)
kernelargs.append(nitems)
kernelargs.append(itemsize)
kernelargs.append(data)
for ax in range(val.ndim):
kernelargs.append(c_intp(val.shape[ax]))
for ax in range(val.ndim):
kernelargs.append(c_intp(val.strides[ax]))
elif isinstance(ty, types.Integer):
cval = getattr(ctypes, "c_%s" % ty)(val)
kernelargs.append(cval)
elif ty == types.float64:
cval = ctypes.c_double(val)
kernelargs.append(cval)
elif ty == types.float32:
cval = ctypes.c_float(val)
kernelargs.append(cval)
elif ty == types.boolean:
cval = ctypes.c_uint8(int(val))
kernelargs.append(cval)
elif ty == types.complex64:
kernelargs.append(ctypes.c_float(val.real))
kernelargs.append(ctypes.c_float(val.imag))
elif ty == types.complex128:
kernelargs.append(ctypes.c_double(val.real))
kernelargs.append(ctypes.c_double(val.imag))
else:
raise NotImplementedError(ty, val)
def _prepare_args(self, ty, val, stream, retr, kernelargs):
"""
Convert arguments to ctypes and append to kernelargs
"""
# map the arguments using any extension you've registered
for extension in reversed(self.extensions):
ty, val = extension.prepare_args(ty, val, stream=stream, retr=retr)
if isinstance(ty, types.Array):
devary = wrap_arg(val).to_device(retr, stream)
c_intp = ctypes.c_ssize_t
meminfo = ctypes.c_void_p(0)
parent = ctypes.c_void_p(0)
nitems = c_intp(devary.size)
itemsize = c_intp(devary.dtype.itemsize)
data = ctypes.c_void_p(driver.device_pointer(devary))
kernelargs.append(meminfo)
kernelargs.append(parent)
kernelargs.append(nitems)
kernelargs.append(itemsize)
kernelargs.append(data)
for ax in range(devary.ndim):
kernelargs.append(c_intp(devary.shape[ax]))
for ax in range(devary.ndim):
kernelargs.append(c_intp(devary.strides[ax]))
elif isinstance(ty, types.Integer):
cval = getattr(ctypes, "c_%s" % ty)(val)
kernelargs.append(cval)
elif ty == types.float64:
cval = ctypes.c_double(val)
kernelargs.append(cval)
elif ty == types.float32:
cval = ctypes.c_float(val)
kernelargs.append(cval)
elif ty == types.boolean:
cval = ctypes.c_uint8(int(val))
kernelargs.append(cval)
elif ty == types.complex64:
kernelargs.append(ctypes.c_float(val.real))
kernelargs.append(ctypes.c_float(val.imag))
elif ty == types.complex128:
kernelargs.append(ctypes.c_double(val.real))
kernelargs.append(ctypes.c_double(val.imag))
elif isinstance(ty, (types.NPDatetime, types.NPTimedelta)):
kernelargs.append(ctypes.c_int64(val.view(np.int64)))
elif isinstance(ty, types.Record):
devrec = wrap_arg(val).to_device(retr, stream)
kernelargs.append(devrec)
else:
raise NotImplementedError(ty, val)
def _prepare_args(self, ty, val, stream, retr, kernelargs):
"""
Convert arguments to ctypes and append to kernelargs
"""
# map the arguments using any extension you've registered
for extension in reversed(self.extensions):
ty, val = extension.prepare_args(
ty,
val,
stream=stream,
retr=retr)
if isinstance(ty, types.Array):
if isinstance(ty, types.SmartArrayType):
devary = val.get('gpu')
retr.append(lambda: val.mark_changed('gpu'))
outer_parent = ctypes.c_void_p(0)
kernelargs.append(outer_parent)
else:
devary = wrap_arg(val).to_device(retr, stream)
c_intp = ctypes.c_ssize_t
meminfo = ctypes.c_void_p(0)
parent = ctypes.c_void_p(0)
nitems = c_intp(devary.size)
itemsize = c_intp(devary.dtype.itemsize)
data = ctypes.c_void_p(driver.device_pointer(devary))
kernelargs.append(meminfo)
kernelargs.append(parent)
kernelargs.append(nitems)
kernelargs.append(itemsize)
kernelargs.append(data)
for ax in range(devary.ndim):
kernelargs.append(c_intp(devary.shape[ax]))
for ax in range(devary.ndim):
kernelargs.append(c_intp(devary.strides[ax]))
elif isinstance(ty, types.Integer):
cval = getattr(ctypes, "c_%s" % ty)(val)
kernelargs.append(cval)
elif ty == types.float64:
cval = ctypes.c_double(val)
kernelargs.append(cval)
elif ty == types.float32:
cval = ctypes.c_float(val)
kernelargs.append(cval)
elif ty == types.boolean:
cval = ctypes.c_uint8(int(val))
kernelargs.append(cval)
elif ty == types.complex64:
kernelargs.append(ctypes.c_float(val.real))
kernelargs.append(ctypes.c_float(val.imag))
elif ty == types.complex128:
kernelargs.append(ctypes.c_double(val.real))
kernelargs.append(ctypes.c_double(val.imag))
elif isinstance(ty, (types.NPDatetime, types.NPTimedelta)):
kernelargs.append(ctypes.c_int64(val.view(np.int64)))
elif isinstance(ty, types.Record):
devrec = wrap_arg(val).to_device(retr, stream)
kernelargs.append(devrec)
else:
raise NotImplementedError(ty, val)
