def _allocate(self,
where,
obj=None,
dtype=None,
shape=None,
strides=None,
copy=True):
if dtype:
dtype = np.dtype(dtype)
if where == 'host':
if obj is not None:
self._host = np.array(obj, dtype, copy=copy)
else:
self._host = np.empty(shape, dtype,
_s2o(dtype, shape, strides))
else:
# Don't import this at module-scope as it may not be available
# in all environments (e.g., CUDASIM)
from numba.cuda.cudadrv import devicearray as da
if obj is not None:
# If 'obj' is an array-like object but not an ndarray,
# construct an ndarray first to extract all the parameters we need.
if not isinstance(obj, np.ndarray):
obj = np.array(obj, copy=False)
self._gpu = da.from_array_like(obj)
else:
if strides is None:
strides = _o2s(dtype, shape, 'C')
self._gpu = da.DeviceNDArray(shape, strides, dtype)
def _broadcast_add_axis(self, ary, newshape):
newax = len(newshape) - len(ary.shape)
# Add 0 strides for missing dimension
newstrides = (0,) * newax + ary.strides
return devicearray.DeviceNDArray(shape=newshape,
strides=newstrides,
dtype=ary.dtype,
gpu_data=ary.gpu_data)
def test_devicearray_relaxed_strides(self):
# From the reproducer in Issue #6824.
# Construct a device array that is contiguous even though
# the strides for the first axis (800) are not equal to
# the strides * size (10 * 8 = 80) for the previous axis,
# because the first axis size is 1.
arr = devicearray.DeviceNDArray((1, 10), (800, 8), np.float64)
# Ensure we still believe the array to be contiguous because
# strides checking is relaxed.
self.assertTrue(arr.flags['C_CONTIGUOUS'])
self.assertTrue(arr.flags['F_CONTIGUOUS'])
def test_devicearray_strict_strides(self):
# From the reproducer in Issue #6824.
with override_config('NPY_RELAXED_STRIDES_CHECKING', 0):
# Construct a device array that is not contiguous because
# the strides for the first axis (800) are not equal to
# the strides * size (10 * 8 = 80) for the previous axis.
arr = devicearray.DeviceNDArray((1, 10), (800, 8), np.float64)
# Ensure we don't believe the array to be contiguous becase strides
# checking is strict.
self.assertFalse(arr.flags['C_CONTIGUOUS'])
self.assertFalse(arr.flags['F_CONTIGUOUS'])
def broadcast_device(self, ary, shape):
ax_differs = [ax for ax in range(len(shape))
if ax >= ary.ndim
or ary.shape[ax] != shape[ax]]
missingdim = len(shape) - len(ary.shape)
strides = [0] * missingdim + list(ary.strides)
for ax in ax_differs:
strides[ax] = 0
return devicearray.DeviceNDArray(shape=shape,
strides=strides,
dtype=ary.dtype,
gpu_data=ary.gpu_data)
def _broadcast_scalar_input(self, ary, shape):
return devicearray.DeviceNDArray(shape=shape,
strides=(0,),
dtype=ary.dtype,
gpu_data=ary.gpu_data)
