1259566567

Build Type

github

Committed by GitHub

Commit Message

Merge bdd050d8b into 8d704db75

Pull Request Pull Request #586: usm_ndarray used to be declared public api, now just api

Run Details

560 of 988 branches covered (56.68%)

Branch coverage included in aggregate %.

81 of 463 new or added lines in 4 files covered. (17.49%)

4 existing lines in 1 file now uncovered.

4012 of 5722 relevant lines covered (70.12%)

10.45 hits per line

Source File
Press 'n' to go to next uncovered line, 'b' for previous

13.78

/dpctl/tensor/_copy_utils.py

1	import numpy as np	1✔
2
3	import dpctl.memory as dpm	1✔
4	import dpctl.tensor as dpt	1✔
5
6
7	def contract_iter2(shape, strides1, strides2):	1✔
NEW 8	p = np.argsort(np.abs(strides1))[::-1]	×
NEW 9	sh = [shape[i] for i in p]	×
NEW 10	disp1 = 0	×
NEW 11	disp2 = 0	×
NEW 12	st1 = []	×
NEW 13	st2 = []	×
NEW 14	for i in p:	×
NEW 15	this_stride1 = strides1[i]	×
NEW 16	this_stride2 = strides2[i]	×
NEW 17	if this_stride1 < 0:	×
NEW 18	disp1 += this_stride1 * (shape[i] - 1)	×
NEW 19	this_stride1 = -this_stride1	×
NEW 20	if this_stride2 < 0:	×
NEW 21	disp2 += this_stride2 * (shape[i] - 1)	×
NEW 22	this_stride2 = -this_stride2	×
NEW 23	st1.append(this_stride1)	×
NEW 24	st2.append(this_stride2)	×
25	while True:
NEW 26	changed = False	×
NEW 27	k = len(sh) - 1	×
NEW 28	for i in range(k):	×
NEW 29	step1 = st1[i + 1]	×
NEW 30	jump1 = st1[i] - (sh[i + 1] - 1) * step1	×
NEW 31	step2 = st2[i + 1]	×
NEW 32	jump2 = st2[i] - (sh[i + 1] - 1) * step2	×
NEW 33	if jump1 == step1 and jump2 == step2:	×
NEW 34	changed = True	×
NEW 35	st1[i:-1] = st1[i + 1 :]	×
NEW 36	st2[i:-1] = st2[i + 1 :]	×
NEW 37	sh[i] *= sh[i + 1]	×
NEW 38	sh[i + 1 : -1] = sh[i + 2 :]	×
NEW 39	sh = sh[:-1]	×
NEW 40	st1 = st1[:-1]	×
NEW 41	st2 = st2[:-1]	×
NEW 42	break	×
NEW 43	if not changed:	×
NEW 44	break	×
NEW 45	return (sh, st1, disp1, st2, disp2)	×
46
47
48	def has_memory_overlap(x1, x2):	1✔
NEW 49	m1 = dpm.as_usm_memory(x1)	×
NEW 50	m2 = dpm.as_usm_memory(x2)	×
NEW 51	if m1.sycl_device == m2.sycl_device:	×
NEW 52	p1_beg = m1._pointer	×
NEW 53	p1_end = p1_beg + m1.nbytes	×
NEW 54	p2_beg = m2._pointer	×
NEW 55	p2_end = p2_beg + m2.nbytes	×
NEW 56	return p1_beg > p2_end or p2_beg < p1_end	×
57	else:
NEW 58	return False	×
59
60
61	def copy_to_numpy(ary):	1✔
62	if type(ary) is not dpt.usm_ndarray:	1!
NEW 63	raise TypeError	×
64	h = ary.usm_data.copy_to_host().view(ary.dtype)	1✔
65	itsz = ary.itemsize	1✔
66	strides_bytes = tuple(si * itsz for si in ary.strides)	1✔
67	offset = ary.__sycl_usm_array_interface__.get("offset", 0) * itsz	1✔
68	return np.ndarray(	1✔
69	ary.shape,
70	dtype=ary.dtype,
71	buffer=h,
72	strides=strides_bytes,
73	offset=offset,
74	)
75
76
77	def copy_from_numpy(np_ary, usm_type="device", queue=None):	1✔
78	"Copies numpy array `np_ary` into a new usm_ndarray"
79	# This may peform a copy to meet stated requirements
80	Xnp = np.require(np_ary, requirements=["A", "O", "C", "E"])	1✔
81	if queue:	1!
82	ctor_kwargs = {"queue": queue}	1✔
83	else:
NEW 84	ctor_kwargs = dict()	×
85	Xusm = dpt.usm_ndarray(	1✔
86	Xnp.shape,
87	dtype=Xnp.dtype,
88	buffer=usm_type,
89	buffer_ctor_kwargs=ctor_kwargs,
90	)
91	Xusm.usm_data.copy_from_host(Xnp.reshape((-1)).view("u1"))	1✔
92	return Xusm	1✔
93
94
95	def copy_from_numpy_into(dst, np_ary):	1✔
96	if not isinstance(np_ary, np.ndarray):	1!
NEW 97	raise TypeError("Expected numpy.ndarray, got {}".format(type(np_ary)))	×
98	for i in range(dst.size):	1✔
99	mi = np.unravel_index(i, dst.shape)	1✔
100	host_buf = np.array(np_ary[mi], ndmin=1).view("u1")	1✔
101	usm_mem = dpm.as_usm_memory(dst[mi])	1✔
102	usm_mem.copy_from_host(host_buf)	1✔
103
104
105	class Dummy:	1✔
106	def __init__(self, iface):	1✔
NEW 107	self.__sycl_usm_array_interface__ = iface	×
108
109
110	def copy_same_dtype(dst, src):	1✔
NEW 111	if type(dst) is not dpt.usm_ndarray or type(src) is not dpt.usm_ndarray:	×
NEW 112	raise TypeError	×
113
NEW 114	if dst.shape != src.shape:	×
NEW 115	raise ValueError	×
116
NEW 117	if dst.dtype != src.dtype:	×
NEW 118	raise ValueError	×
119
120	# check that memory regions do not overlap
NEW 121	if has_memory_overlap(dst, src):	×
NEW 122	tmp = copy_to_numpy(src)	×
NEW 123	copy_from_numpy_into(dst, tmp)	×
NEW 124	return	×
125
NEW 126	if (dst.flags & 1) and (src.flags & 1):	×
NEW 127	dst.usm_data.copy_from_device(src.usm_data)	×
NEW 128	return	×
129
130	# simplify strides
NEW 131	sh_i, dst_st, dst_disp, src_st, src_disp = contract_iter2(	×
132	dst.shape, dst.strides, src.strides
133	)
NEW 134	src_iface = src.__sycl_usm_array_interface__	×
NEW 135	dst_iface = dst.__sycl_usm_array_interface__	×
NEW 136	src_iface["shape"] = tuple()	×
NEW 137	src_iface.pop("strides", None)	×
NEW 138	dst_iface["shape"] = tuple()	×
NEW 139	dst_iface.pop("strides", None)	×
NEW 140	for i in range(dst.size):	×
NEW 141	mi = np.unravel_index(i, sh_i)	×
NEW 142	dst_offset = dst_disp	×
NEW 143	src_offset = src_disp	×
NEW 144	for j, dst_stj, src_stj in zip(mi, dst_st, src_st):	×
NEW 145	dst_offset = dst_offset + j * dst_stj	×
NEW 146	src_offset = src_offset + j * src_stj	×
NEW 147	dst_iface["offset"] = dst_offset	×
NEW 148	src_iface["offset"] = src_offset	×
NEW 149	msrc = dpm.as_usm_memory(Dummy(src_iface))	×
NEW 150	mdst = dpm.as_usm_memory(Dummy(dst_iface))	×
NEW 151	mdst.copy_from_device(msrc)	×
152
153
154	def copy_same_shape(dst, src):	1✔
NEW 155	if src.dtype == dst.dtype:	×
NEW 156	copy_same_dtype(dst, src)	×
157
158	# check that memory regions do not overlap
NEW 159	if has_memory_overlap(dst, src):	×
NEW 160	tmp = copy_to_numpy(src)	×
NEW 161	tmp = tmp.astype(dst.dtype)	×
NEW 162	copy_from_numpy_into(dst, tmp)	×
NEW 163	return	×
164
165	# simplify strides
NEW 166	sh_i, dst_st, dst_disp, src_st, src_disp = contract_iter2(	×
167	dst.shape, dst.strides, src.strides
168	)
NEW 169	src_iface = src.__sycl_usm_array_interface__	×
NEW 170	dst_iface = dst.__sycl_usm_array_interface__	×
NEW 171	src_iface["shape"] = tuple()	×
NEW 172	src_iface.pop("strides", None)	×
NEW 173	dst_iface["shape"] = tuple()	×
NEW 174	dst_iface.pop("strides", None)	×
NEW 175	for i in range(dst.size):	×
NEW 176	mi = np.unravel_index(i, sh_i)	×
NEW 177	dst_offset = dst_disp	×
NEW 178	src_offset = src_disp	×
NEW 179	for j, dst_stj, src_stj in zip(mi, dst_st, src_st):	×
NEW 180	dst_offset = dst_offset + j * dst_stj	×
NEW 181	src_offset = src_offset + j * src_stj	×
NEW 182	dst_iface["offset"] = dst_offset	×
NEW 183	src_iface["offset"] = src_offset	×
NEW 184	msrc = dpm.as_usm_memory(Dummy(src_iface))	×
NEW 185	mdst = dpm.as_usm_memory(Dummy(dst_iface))	×
NEW 186	tmp = msrc.copy_to_host().view(src.dtype)	×
NEW 187	tmp = tmp.astype(dst.dtype)	×
NEW 188	mdst.copy_from_host(tmp.view("u1"))	×
189
190
191	def copy_from_usm_ndarray_to_usm_ndarray(dst, src):	1✔
NEW 192	if type(dst) is not dpt.usm_ndarray or type(src) is not dpt.usm_ndarray:	×
NEW 193	raise TypeError	×
194
NEW 195	if dst.ndim == src.ndim and dst.shape == src.shape:	×
NEW 196	copy_same_shape(dst, src)	×
197
NEW 198	try:	×
NEW 199	common_shape = np.broadcast_shapes(dst.shape, src.shape)	×
NEW 200	except ValueError:	×
NEW 201	raise ValueError	×
202
NEW 203	if dst.size < src.size:	×
NEW 204	raise ValueError	×
205
NEW 206	if len(common_shape) > dst.ndim:	×
NEW 207	ones_count = len(common_shape) - dst.ndim	×
NEW 208	for k in range(ones_count):	×
NEW 209	if common_shape[k] != 1:	×
NEW 210	raise ValueError	×
NEW 211	common_shape = common_shape[ones_count:]	×
212
NEW 213	if src.ndim < len(common_shape):	×
NEW 214	new_src_strides = (0,) * (len(common_shape) - src.ndim) + src.strides	×
NEW 215	src_same_shape = dpt.usm_ndarray(	×
216	common_shape, dtype=src.dtype, buffer=src, strides=new_src_strides
217	)
218	else:
NEW 219	src_same_shape = src	×
220
NEW 221	copy_same_shape(dst, src_same_shape)	×
222
223
224	def astype(usm_ary, newdtype, order="K", casting="unsafe", copy=True):	1✔
225	"""
226	astype(usm_array, new_dtype, order="K", casting="unsafe", copy=True)
227
228	Returns a copy of the array, cast to a specified type.
229
230	A view can be returned, if possible, when `copy=False` is used.
231	"""
NEW 232	if not isinstance(usm_ary, dpt.usm_ndarray):	×
NEW 233	return TypeError(	×
234	"Expected object of type dpt.usm_ndarray, got {}".format(
235	type(usm_ary)
236	)
237	)
NEW 238	ary_dtype = usm_ary.dtype	×
NEW 239	target_dtype = np.dtype(newdtype)	×
NEW 240	if not np.can_cast(ary_dtype, target_dtype, casting=casting):	×
NEW 241	raise TypeError(	×
242	"Can not cast from {} to {} according to rule {}".format(
243	ary_dtype, newdtype, casting
244	)
245	)
NEW 246	c_contig = usm_ary.flags & 1	×
NEW 247	f_contig = usm_ary.flags & 2	×
NEW 248	needs_copy = copy or not (ary_dtype == target_dtype)	×
NEW 249	if not needs_copy and (order != "K"):	×
NEW 250	needs_copy = (c_contig and order not in ["A", "C"]) or (	×
251	f_contig and order not in ["A", "F"]
252	)
NEW 253	if needs_copy:	×
NEW 254	copy_order = "C"	×
NEW 255	if order == "C":	×
NEW 256	pass	×
NEW 257	elif order == "F":	×
NEW 258	copy_order = order	×
NEW 259	elif order == "A":	×
NEW 260	if usm_ary.flags & 2:	×
NEW 261	copy_order = "F"	×
NEW 262	elif order == "K":	×
NEW 263	if usm_ary.flags & 2:	×
NEW 264	copy_order = "F"	×
NEW 265	R = dpt.usm_ndarray(	×
266	usm_ary.shape,
267	dtype=target_dtype,
268	buffer=usm_ary.usm_type,
269	order=copy_order,
270	buffer_ctor_kwargs={"queue": usm_ary.sycl_queue},
271	)
NEW 272	if order == "K" and (not c_contig and not f_contig):	×
NEW 273	original_strides = usm_ary.strides	×
NEW 274	ind = sorted(	×
275	range(usm_ary.ndim),
276	key=lambda i: abs(original_strides[i]),
277	reverse=True,
278	)
NEW 279	new_strides = tuple(R.strides[ind[i]] for i in ind)	×
NEW 280	R = dpt.usm_ndarray(	×
281	usm_ary.shape,
282	dtype=target_dtype,
283	buffer=R.usm_data,
284	strides=new_strides,
285	)
NEW 286	copy_from_usm_ndarray_to_usm_ndarray(R, usm_ary)	×
NEW 287	return R	×
288	else:
NEW 289	return usm_ary	×

IntelPython / dpctl / 1259566567

Source File Press 'n' to go to next uncovered line, 'b' for previous

Source File
Press 'n' to go to next uncovered line, 'b' for previous