def copy_to_device(self, ary, stream=0):
"""Copy `ary` to `self`.
If `ary` is a CUDA memory, perform a device-to-device transfer.
Otherwise, perform a a host-to-device transfer.
"""
if ary.size == 0:
# Nothing to do
return
sentry_contiguous(self)
stream = self._default_stream(stream)
self_core, ary_core = array_core(self), array_core(ary)
if _driver.is_device_memory(ary):
sentry_contiguous(ary)
check_array_compatibility(self_core, ary_core)
_driver.device_to_device(self, ary, self.alloc_size, stream=stream)
else:
# Ensure same contiguity. Only makes a host-side copy if necessary
# (i.e., in order to materialize a writable strided view)
ary_core = np.array(
ary_core,
order='C' if self_core.flags['C_CONTIGUOUS'] else 'F',
subok=True,
copy=not ary_core.flags['WRITEABLE'])
check_array_compatibility(self_core, ary_core)
_driver.host_to_device(self,
ary_core,
self.alloc_size,
stream=stream)
def _do_setitem(self, key, value, stream=0):
stream = self._default_stream(stream)
# If the record didn't have a default stream, and the user didn't
# provide a stream, then we will use the default stream for the
# assignment kernel and synchronize on it.
synchronous = not stream
if synchronous:
ctx = devices.get_context()
stream = ctx.get_default_stream()
# (1) prepare LHS
typ, offset = self.dtype.fields[key]
newdata = self.gpu_data.view(offset)
lhs = type(self)(dtype=typ, stream=stream, gpu_data=newdata)
# (2) prepare RHS
rhs, _ = auto_device(lhs.dtype.type(value), stream=stream)
# (3) do the copy
_driver.device_to_device(lhs, rhs, rhs.dtype.itemsize, stream)
if synchronous:
stream.synchronize()
def test_d2d(self):
hst = np.arange(100, dtype=np.uint32)
hst2 = np.empty_like(hst)
sz = hst.size * hst.dtype.itemsize
dev1 = self.context.memalloc(sz)
dev2 = self.context.memalloc(sz)
driver.host_to_device(dev1, hst, sz)
driver.device_to_device(dev2, dev1, sz)
driver.device_to_host(hst2, dev2, sz)
self.assertTrue(np.all(hst == hst2))
def test_d2d(self):
hst = np.arange(100, dtype=np.uint32)
hst2 = np.empty_like(hst)
sz = hst.size * hst.dtype.itemsize
dev1 = self.context.memalloc(sz)
dev2 = self.context.memalloc(sz)
driver.host_to_device(dev1, hst, sz)
driver.device_to_device(dev2, dev1, sz)
driver.device_to_host(hst2, dev2, sz)
self.assertTrue(np.all(hst == hst2))
