Coverage for cuda/core/_memoryview.pyx: 77.84%

1# SPDX-FileCopyrightText: Copyright (c) 2024-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. 

2# 

3# SPDX-License-Identifier: Apache-2.0 

4  

5from __future__ import annotations 

6  

7cimport cython 

8from ._dlpack cimport * 

9from ._dlpack import classify_dl_device 

10from libc.stdint cimport intptr_t 

11from cuda.core._layout cimport _StridedLayout, get_strides_ptr 

12from cuda.core._stream import Stream 

13  

14import ctypes 

15import functools 

16import sys 

17import warnings 

18from collections.abc import Callable # no-cython-lint # used in string annotations below 

19from typing import TYPE_CHECKING, Any # no-cython-lint # used in string annotations below 

20  

21if TYPE_CHECKING: 

22 from cuda.core._tensor_map import TensorMapDescriptorOptions 

23  

24import numpy 

25  

26from cuda.bindings cimport cydriver 

27from cuda.core._rt cimport ( 

28 EventHandle, 

29 create_event_handle_for_stream, 

30 as_cu, 

31 get_last_error, 

32) 

33  

34from cuda.core._utils.cuda_utils import handle_return, driver 

35from cuda.core._utils.cuda_utils cimport HANDLE_RETURN 

36  

37  

38from cuda.core._memory import Buffer 

39from cuda.core._memory._buffer cimport Buffer as cyBuffer, Buffer_check_open 

40  

41  

42# --------------------------------------------------------------------------- 

43# Lazy tensor bridge (avoids loading _tensor_bridge.so until torch is used) 

44# --------------------------------------------------------------------------- 

45  

46cdef object _tensor_bridge = None 

47# Cache: type(obj) -> True/False for the torch tensor check. 

48# Once a type is seen, we never re-check. 

49cdef dict _torch_type_cache = {} 

50# Tri-state: None = not checked, True/False = result of version check 

51cdef object _torch_version_ok = None 

52  

53cdef inline bint _torch_version_check(): 

54 """Return True if torch >= 2.3. Memoized. 

55  

56 Lower bound: AOTI functions we use were introduced in PyTorch 2.3. 

57  

58 No upper bound. The dispatch in ``_tensor_bridge.view_as_torch_tensor`` 

59 selects between two safe paths per torch version: 

60  

61 * torch >= 2.14 -- stable PyObject->Tensor C ABI shim 

62 (``torch_tensor_from_pyobject`` + ``aoti_torch_delete_tensor_object``); 

63 part of torch's stable contract, so any 2.14+ release is covered. 

64 * torch 2.3-2.13 -- ``pyobj_to_aten_handle`` pointer-arithmetic trick, 

65 validated against the two known THPVariable layouts 

66 (``c10::MaybeOwned<at::Tensor>`` for 2.3-2.9; ``at::Tensor`` directly 

67 for 2.10-2.13; see ``_get_cdata_extra_offset`` in ``_tensor_bridge.pyx``). 

68 """ 

69 global _torch_version_ok 

70 if _torch_version_ok is not None: 

71 return <bint>_torch_version_ok 

72 torch = sys.modules.get("torch") 

73 if torch is None: 

74 _torch_version_ok = False 

75 return False 

76 try: 

77 major, minor = int(torch.__version__.split(".")[0]), \ 

78 int(torch.__version__.split(".")[1]) 

79 _torch_version_ok = (major, minor) >= (2, 3) 

80 except (ValueError, IndexError): 

81 _torch_version_ok = False 

82 return <bint>_torch_version_ok 

83  

84  

85cdef inline bint _is_torch_tensor(object obj): 

86 cdef type tp = type(obj) 125 ctx2bb{ } jb: ( 6 3 % 7 | 8 9 ! # $ + , - . / L N O M nbobxbpbybqbzbrbsbtbAbBbCbDbEbFbubGbHbIbvbJbKbwb{bMb]bNbgbhbw mb1 cb5 Z dbLbS T U V W X ; ) = ~ ebfbQ R 4 abj P Y 0 I s t x y z E F G H A B C D u K J p f 2 cce d |bb k h q l m r n o g v ~b}b[bc i a ac

87 cdef object cached = _torch_type_cache.get(tp) 125 ctx2bb{ } jb: ( 6 3 % 7 | 8 9 ! # $ + , - . / L N O M nbobxbpbybqbzbrbsbtbAbBbCbDbEbFbubGbHbIbvbJbKbwb{bMb]bNbgbhbw mb1 cb5 Z dbLbS T U V W X ; ) = ~ ebfbQ R 4 abj P Y 0 I s t x y z E F G H A B C D u K J p f 2 cce d |bb k h q l m r n o g v ~b}b[bc i a ac

88 if cached is not None: 125 ctx2bb{ } jb: ( 6 3 % 7 | 8 9 ! # $ + , - . / L N O M nbobxbpbybqbzbrbsbtbAbBbCbDbEbFbubGbHbIbvbJbKbwb{bMb]bNbgbhbw mb1 cb5 Z dbLbS T U V W X ; ) = ~ ebfbQ R 4 abj P Y 0 I s t x y z E F G H A B C D u K J p f 2 cce d |bb k h q l m r n o g v ~b}b[bc i a ac

89 return <bint>cached 111 ctx2bb} jb: ( 6 3 % 7 8 9 ! # $ + , - . / N O M nbobxbpbybqbzbrbsbtbAbBbCbDbEbFbubGbHbIbvbJbKbwb{bgbhbw 1 cb5 Z S T U V W X ; ) = ~ ebfbQ R 4 abj P Y 0 I s t x y z E F G H A B C D u K J p f 2 cce d b k h q l m r n o g v c i a

90 cdef str mod = tp.__module__ or "" 29 ctx2{ ( 3 | L Mb]bNbmbdbLb) = 4 2 e d |bb k h g v ~b}b[bi a ac

91 cdef bint result = mod.startswith("torch") and hasattr(obj, "data_ptr") \ 29 ctx2{ ( 3 | L Mb]bNbmbdbLb) = 4 2 e d |bb k h g v ~b}b[bi a ac

92 and _torch_version_check() 

93 _torch_type_cache[tp] = result # setdefault not needed for bools 29 ctx2{ ( 3 | L Mb]bNbmbdbLb) = 4 2 e d |bb k h g v ~b}b[bi a ac

94 return result 29 ctx2{ ( 3 | L Mb]bNbmbdbLb) = 4 2 e d |bb k h g v ~b}b[bi a ac

95  

96  

97cdef object _get_tensor_bridge(): 

98 """Bootstrap AOTI symbols, then import _tensor_bridge on first use.""" 

99 global _tensor_bridge 

100 if _tensor_bridge is not None: 

101 return _tensor_bridge 

102 torch_C = sys.modules.get("torch._C") 

103 if torch_C is None: 

104 raise RuntimeError( 

105 "torch._C is not loaded; cannot initialise the tensor bridge. " 

106 "Make sure PyTorch is imported before passing a torch.Tensor.") 

107 ctypes.CDLL(torch_C.__file__, mode=ctypes.RTLD_GLOBAL) 

108 from cuda.core import _tensor_bridge as tb 

109 _tensor_bridge = tb 

110 return _tensor_bridge 

111  

112  

113try: 

114 from ml_dtypes import bfloat16 

115except ImportError: 

116 bfloat16 = None 

117  

118# TODO(leofang): support NumPy structured dtypes 

119  

120  

121cdef extern from "Python.h": 

122 ctypedef struct PyTypeObject: 

123 void* tp_dict 

124 void PyType_Modified(PyTypeObject*) 

125  

126  

127cdef DLPackExchangeAPI _SMV_DLPACK_EXCHANGE_API 

128cdef bint _SMV_DLPACK_EXCHANGE_API_INITED = False 

129_SMV_DLPACK_EXCHANGE_API_CAPSULE = cpython.PyCapsule_New( 

130 <void*>&_SMV_DLPACK_EXCHANGE_API, 

131 b"dlpack_exchange_api", 

132 NULL, 

133) 

134  

135  

136cdef class StridedMemoryView: 

137 """A class holding metadata of a strided dense array/tensor. 

138  

139 A :obj:`StridedMemoryView` instance can be created in three ways: 

140  

141 1. Using the :obj:`args_viewable_as_strided_memory` decorator (recommended) 

142 2. Explicit construction relying on DLPack or CUDA Array Interface, see below. 

143 3. From :obj:`~_memory.Buffer` and shape and size tuples (see 

144 :meth:`from_buffer` classmethod) 

145  

146 ``StridedMemoryView(obj, stream_ptr)`` can be used to create a view from 

147 objects supporting either DLPack (up to v1.0) or CUDA Array Interface 

148 (CAI) v3. When wrapping an arbitrary object it will try the DLPack protocol 

149 first, then the CAI protocol. A :obj:`BufferError` is raised if neither is 

150 supported. 

151  

152 Since either way would take a consumer stream, for DLPack it is passed to 

153 ``obj.__dlpack__()`` as-is (except for :obj:`None`, see below); for CAI, a 

154 stream order will be established between the consumer stream and the 

155 producer stream (from ``obj.__cuda_array_interface__()["stream"]``), as if 

156 ``cudaStreamWaitEvent`` is called by this method. 

157  

158 To opt-out of the stream ordering operation in either DLPack or CAI, 

159 please pass ``stream_ptr=-1``. Note that this deviates (on purpose) 

160 from the semantics of ``obj.__dlpack__(stream=None, ...)`` since ``cuda.core`` 

161 does not encourage using the (legacy) default/null stream, but is 

162 consistent with the CAI's semantics. For DLPack, ``stream=-1`` will be 

163 internally passed to ``obj.__dlpack__()`` instead. 

164  

165 Parameters 

166 ---------- 

167 obj : Any 

168 Any objects that supports either DLPack (up to v1.0) or CUDA Array 

169 Interface (v3). 

170 stream_ptr: int 

171 The pointer address (as Python `int`) to the **consumer** stream. 

172 Stream ordering will be properly established unless ``-1`` is passed. 

173  

174  

175 Attributes 

176 ----------- 

177 ptr : int 

178 Pointer to the tensor buffer (as a Python `int`). 

179 device_id : int 

180 The device ID for where the tensor is located. It is -1 for CPU tensors 

181 (meaning those only accessible from the host). 

182 is_device_accessible : bool 

183 Whether the tensor data can be accessed on the GPU. 

184 readonly: bool 

185 Whether the tensor data can be modified in place. 

186 exporting_obj : Any 

187 A reference to the original tensor object that is being viewed. 

188 If the view is created with :meth:`from_buffer`, 

189 it will be the Buffer instance passed to the method. 

190  

191 """ 

192 def __init__(self, obj: object = None, stream_ptr: int | None = None) -> None: 

193 cdef str clsname = self.__class__.__name__ 9 ctx2? @ [ ] ^ ecPb_ `

194 if obj is not None: 9 ctx2? @ [ ] ^ ecPb_ `

195 # populate self's attributes 

196 if check_has_dlpack(obj): 8 ctx2? @ [ ] ^ Pb_ `

197 warnings.warn( 7 ctx1?@[]^_`

198 f"Constructing a {clsname} directly from a DLPack-supporting object is deprecated; " 7 ctx1?@[]^_`

199 "Use `StridedMemoryView.from_dlpack` or `StridedMemoryView.from_any_interface` instead.", 

200 DeprecationWarning, 7 ctx1?@[]^_`

201 stacklevel=2, 

202 ) 

203 view_as_dlpack(obj, stream_ptr, self) 7 ctx1?@[]^_`

204 else: 

205 warnings.warn( 1 ctx2Pb

206 f"Constructing a {clsname} directly from a CUDA-array-interface-supporting object is deprecated; " 1 ctx2Pb

207 "Use `StridedMemoryView.from_cuda_array_interface` or `StridedMemoryView.from_any_interface` instead.", 

208 DeprecationWarning, 1 ctx2Pb

209 stacklevel=2, 

210 ) 

211 view_as_cai(obj, stream_ptr, self) 1 ctx2Pb

212 else: 

213 warnings.warn( 1 ctx2ec

214 f"Constructing an empty {clsname} is deprecated; " 1 ctx2ec

215 "use one of the classmethods `from_dlpack`, `from_cuda_array_interface` or `from_any_interface` " 

216 "to construct a StridedMemoryView from an object", 

217 DeprecationWarning, 1 ctx2ec

218 stacklevel=2, 

219 ) 

220  

221 @classmethod 

222 def from_dlpack(cls, obj: object, stream_ptr: int | None=None) -> StridedMemoryView: 

223 """Create a view from an object supporting the `DLPack <https://dmlc.github.io/dlpack/latest/>`_ protocol. 

224  

225 Parameters 

226 ---------- 

227 obj : object 

228 An object implementing the `DLPack <https://dmlc.github.io/dlpack/latest/>`_ protocol 

229 (via ``__dlpack__``). 

230 stream_ptr : int, optional 

231 Stream pointer for synchronization. If ``None``, no synchronization is performed. 

232 """ 

233 cdef StridedMemoryView buf = StridedMemoryView.__new__(cls) 85 ctx2bb{ } jb: ( 6 3 8 9 ! # $ + , - . / N O M gbhbw 1 cb5 Z S T U V W X ; ) = ~ fbQ R abj I s t x y z E F G H A B C D u K J p f 2 cce d |bb k h q l m r n o g v ~b}b[bc i a ac

234 if _is_torch_tensor(obj): 85 ctx2bb{ } jb: ( 6 3 8 9 ! # $ + , - . / N O M gbhbw 1 cb5 Z S T U V W X ; ) = ~ fbQ R abj I s t x y z E F G H A B C D u K J p f 2 cce d |bb k h q l m r n o g v ~b}b[bc i a ac

235 _get_tensor_bridge().view_as_torch_tensor(obj, stream_ptr, buf) 

236 return buf 

237 view_as_dlpack(obj, stream_ptr, buf) 85 ctx2bb{ } jb: ( 6 3 8 9 ! # $ + , - . / N O M gbhbw 1 cb5 Z S T U V W X ; ) = ~ fbQ R abj I s t x y z E F G H A B C D u K J p f 2 cce d |bb k h q l m r n o g v ~b}b[bc i a ac

238 return buf 79 ctx2bb{ } jb: ( 6 3 8 9 ! # $ + , - . / N O M gbhbw 1 cb5 Z S T U V W X ; ) = ~ fbQ R abj I s t x y z E F G H A B C D u K J p f 2 e d b k h q l m r n o g v c i a

239  

240 @classmethod 

241 def from_cuda_array_interface(cls, obj: object, stream_ptr: int | None=None) -> StridedMemoryView: 

242 """Create a view from an object supporting the `__cuda_array_interface__ <https://numba.readthedocs.io/en/stable/cuda/cuda_array_interface.html>`_ protocol. 

243  

244 Parameters 

245 ---------- 

246 obj : object 

247 An object implementing the `__cuda_array_interface__ <https://numba.readthedocs.io/en/stable/cuda/cuda_array_interface.html>`_ protocol. 

248 stream_ptr : int, optional 

249 Stream pointer for synchronization. If ``None``, no synchronization is performed. 

250 """ 

251 cdef StridedMemoryView buf = StridedMemoryView.__new__(cls) 10 ctx2% 7 | L Mb]bNbmbdbLb

252 if _is_torch_tensor(obj): 10 ctx2% 7 | L Mb]bNbmbdbLb

253 _get_tensor_bridge().view_as_torch_tensor(obj, stream_ptr, buf) 

254 return buf 

255 view_as_cai(obj, stream_ptr, buf) 10 ctx2% 7 | L Mb]bNbmbdbLb

256 return buf 8 ctx2% 7 | L MbNbmbdb

257  

258 @classmethod 

259 def from_array_interface(cls, obj: object) -> StridedMemoryView: 

260 """Create a view from an object supporting the `__array_interface__ <https://numpy.org/doc/stable/reference/arrays.interface.html>`_ protocol. 

261  

262 Parameters 

263 ---------- 

264 obj : object 

265 An object implementing the `__array_interface__ <https://numpy.org/doc/stable/reference/arrays.interface.html>`_ protocol (e.g., a numpy array). 

266 """ 

267 cdef StridedMemoryView buf = StridedMemoryView.__new__(cls) 33 ctx2nbobxbpbybqbzbrbsbtbAbBbCbDbEbFbubGbHbIbvbJbKbwb{beb4 P Y 0 f e d

268 if _is_torch_tensor(obj): 33 ctx2nbobxbpbybqbzbrbsbtbAbBbCbDbEbFbubGbHbIbvbJbKbwb{beb4 P Y 0 f e d

269 _get_tensor_bridge().view_as_torch_tensor(obj, None, buf) 

270 return buf 

271 view_as_array_interface(obj, buf) 33 ctx2nbobxbpbybqbzbrbsbtbAbBbCbDbEbFbubGbHbIbvbJbKbwb{beb4 P Y 0 f e d

272 return buf 32 ctx2nbobxbpbybqbzbrbsbtbAbBbCbDbEbFbubGbHbIbvbJbKbwbeb4 P Y 0 f e d

273  

274 @classmethod 

275 def from_any_interface(cls, obj: object, stream_ptr: int | None = None) -> StridedMemoryView: 

276 """Create a view by automatically selecting the best available protocol. 

277  

278 Tries `DLPack <https://dmlc.github.io/dlpack/latest/>`_ first, then falls back to 

279 `__cuda_array_interface__ <https://numba.readthedocs.io/en/stable/cuda/cuda_array_interface.html>`_. 

280 ``torch.Tensor`` objects are transparently handled via a fast AOTI path 

281 regardless of which protocol is selected. 

282  

283 Parameters 

284 ---------- 

285 obj : object 

286 An object implementing `DLPack <https://dmlc.github.io/dlpack/latest/>`_ or 

287 `__cuda_array_interface__ <https://numba.readthedocs.io/en/stable/cuda/cuda_array_interface.html>`_. 

288 stream_ptr : int, optional 

289 Stream pointer for synchronization. If ``None``, no synchronization is performed. 

290 """ 

291 if check_has_dlpack(obj): 66 ctx2bb{ } jb: ( 6 3 % 7 | + , - . / kcN O M gbhbw 1 cb5 Z ; ) fbQ R abj I s t x y z E F G H A B C D u K J p b k h q l m r n o g v c i a

292 return cls.from_dlpack(obj, stream_ptr) 62 ctx2bb{ } jb: ( 6 3 + , - . / N O M gbhbw 1 cb5 Z ; ) fbQ R abj I s t x y z E F G H A B C D u K J p b k h q l m r n o g v c i a

293 return cls.from_cuda_array_interface(obj, stream_ptr) 3 ctx1%7|

294  

295 @classmethod 

296 def from_buffer( 

297 cls, 

298 cyBuffer buffer not None : Buffer, 

299 shape : tuple[int, ...], 

300 strides : tuple[int, ...] | None = None, 

301 *, 

302 itemsize : int | None = None, 

303 dtype : numpy.dtype | None = None, 

304 is_readonly : bool = False 

305 ) -> StridedMemoryView: 

306 """ 

307 Creates a :obj:`StridedMemoryView` instance from a :obj:`~_memory.Buffer` and shape and strides tuples. 

308 The Buffer can be either allocation coming from a :obj:`MemoryResource` or an external allocation 

309 wrapped in a :obj:`~_memory.Buffer` object with ``Buffer.from_handle(ptr, size, owner=...)``. 

310  

311 .. caution:: 

312 When creating a :obj:`StridedMemoryView` from a :obj:`~_memory.Buffer`, 

313 no synchronization is performed. It is the user's responsibility to ensure 

314 the data in ``buffer`` is properly synchronized when consuming the view. 

315  

316 Parameters 

317 ---------- 

318 buffer : :obj:`~_memory.Buffer` 

319 The buffer to create the view from. 

320 shape : :obj:`tuple` 

321 The layout describing the shape, strides and itemsize of the elements in 

322 the buffer. 

323 strides : :obj:`tuple` 

324 The layout describing the shape, strides and itemsize of the elements in 

325 the buffer. 

326 dtype : :obj:`numpy.dtype` 

327 Optional dtype. 

328 If specified, the dtype's itemsize must match the layout's itemsize. 

329 is_readonly : bool, optional 

330 Whether the mark the view as readonly. 

331 """ 

332 cdef StridedMemoryView view = StridedMemoryView.__new__(cls) 47 ctx2^b_bibkbVbWbXbYbZb0b1b2b3b4b5b6b7b8b9b!b#b$b%b'b(b)b*b+b,b-b.b/b:b;b=b?bdchcfcQbRbbc`b' lbOb*

333 if itemsize is None and dtype is None: 47 ctx2^b_bibkbVbWbXbYbZb0b1b2b3b4b5b6b7b8b9b!b#b$b%b'b(b)b*b+b,b-b.b/b:b;b=b?bdchcfcQbRbbc`b' lbOb*

334 raise ValueError("Either itemsize or dtype must be specified") 1 ctx2hc

335 if itemsize is not None and dtype is not None and itemsize != dtype.itemsize: 46 ctx2^b_bibkbVbWbXbYbZb0b1b2b3b4b5b6b7b8b9b!b#b$b%b'b(b)b*b+b,b-b.b/b:b;b=b?bdcfcQbRbbc`b' lbOb*

336 raise ValueError( 1 ctx2fc

337 f"itemsize ({itemsize}) does not match dtype.itemsize ({dtype.itemsize})" 1 ctx2fc

338 ) 

339 # (itemsize is None XOR dtype is None) OR they are equal 

340 view_buffer_strided( 45 ctx2^b_bibkbVbWbXbYbZb0b1b2b3b4b5b6b7b8b9b!b#b$b%b'b(b)b*b+b,b-b.b/b:b;b=b?bdcQbRbbc`b' lbOb*

341 view, 

342 buffer, 

343 _StridedLayout(shape=shape, strides=strides, itemsize=getattr(dtype, "itemsize", itemsize)), 45 ctx2^b_bibkbVbWbXbYbZb0b1b2b3b4b5b6b7b8b9b!b#b$b%b'b(b)b*b+b,b-b.b/b:b;b=b?bdcQbRbbc`b' lbOb*

344 dtype, 

345 is_readonly, 45 ctx2^b_bibkbVbWbXbYbZb0b1b2b3b4b5b6b7b8b9b!b#b$b%b'b(b)b*b+b,b-b.b/b:b;b=b?bdcQbRbbc`b' lbOb*

346 ) 

347 return view 43 ctx2^b_bibkbVbWbXbYbZb0b1b2b3b4b5b6b7b8b9b!b#b$b%b'b(b)b*b+b,b-b.b/b:b;b=b?bQbRb`b' lbOb*

348  

349 def __dealloc__(self) -> None: 

350 if self.dl_tensor == NULL: 184 ctx2^b_bbb{ } jb: ( ib6 3 % 7 | kb8 9 ! # $ + , - . / ? @ [ ] ^ L N O M nbobxbpbybqbzbrbsbtbAbBbCbDbEbFbubGbHbIbvbJbKbwb{bVbWbXbYbZb0b1b2b3b4b5b6b7b8b9b!b#b$b%b'b(b)b*b+b,b-b.b/b:b;b=b?bdchcfcQbRbbc`bMb]bNbgbhbw ' ecPbmblb1 cb5 Z dbLbS T U V W X _ ` ; ) = ~ ebfbQ R Ob* @b4 abSbTbUbP Y 0 I s t x y z E F G H A B C D u K J p f 2 cce d |bb k h q l m r n o g v ~b}b[bc i a ac

351 return 112 ctx2^b_bib% 7 | kbL N O M nbobxbpbybqbzbrbsbtbAbBbCbDbEbFbubGbHbIbvbJbKbwb{bVbWbXbYbZb0b1b2b3b4b5b6b7b8b9b!b#b$b%b'b(b)b*b+b,b-b.b/b:b;b=b?bdchcfcQbRbbc`bMb]bNb' ecPbmblbcb5 Z dbLbS T U V W X ebOb* 4 P Y 0 f 2 cce d |b~b}b[bc ac

352  

353 if cpython.PyCapsule_IsValid( 89 ctx2bb{ } jb: ( 6 3 8 9 ! # $ + , - . / ? @ [ ] ^ N O M gbhbw 1 cb5 Z S T U V W X _ ` ; ) = ~ fbQ R @babSbTbUbI s t x y z E F G H A B C D u K J p f 2 e d b k h q l m r n o g v c i a

354 self.metadata, DLPACK_VERSIONED_TENSOR_USED_NAME): 89 ctx2bb{ } jb: ( 6 3 8 9 ! # $ + , - . / ? @ [ ] ^ N O M gbhbw 1 cb5 Z S T U V W X _ ` ; ) = ~ fbQ R @babSbTbUbI s t x y z E F G H A B C D u K J p f 2 e d b k h q l m r n o g v c i a

355 data = cpython.PyCapsule_GetPointer( 87 ctx2bb{ } jb: ( 6 3 8 9 ! # $ + , - . / ? @ [ ] ^ N O M gbhbw 1 cb5 Z S T U V W X _ ` ; ) = ~ fbQ R @babSbTbUbI s t x y z E F G H A B C D u K J p f 2 e d k h q l m r n o v c i a

356 self.metadata, DLPACK_VERSIONED_TENSOR_USED_NAME) 87 ctx2bb{ } jb: ( 6 3 8 9 ! # $ + , - . / ? @ [ ] ^ N O M gbhbw 1 cb5 Z S T U V W X _ ` ; ) = ~ fbQ R @babSbTbUbI s t x y z E F G H A B C D u K J p f 2 e d k h q l m r n o v c i a

357 dlm_tensor_ver = <DLManagedTensorVersioned*>data 87 ctx2bb{ } jb: ( 6 3 8 9 ! # $ + , - . / ? @ [ ] ^ N O M gbhbw 1 cb5 Z S T U V W X _ ` ; ) = ~ fbQ R @babSbTbUbI s t x y z E F G H A B C D u K J p f 2 e d k h q l m r n o v c i a

358 if dlm_tensor_ver.deleter != NULL: 87 ctx2bb{ } jb: ( 6 3 8 9 ! # $ + , - . / ? @ [ ] ^ N O M gbhbw 1 cb5 Z S T U V W X _ ` ; ) = ~ fbQ R @babSbTbUbI s t x y z E F G H A B C D u K J p f 2 e d k h q l m r n o v c i a

359 dlm_tensor_ver.deleter(dlm_tensor_ver) 82 ctx2bb{ } jb: ( 6 3 8 9 ! # $ + , - . / ? @ [ ] ^ N O M gbhbw 1 cb5 Z S T U V W X _ ` ; ) = ~ fbQ R abI s t x y z E F G H A B C D u K J p f 2 e d k h q l m r n o c i a

360 elif cpython.PyCapsule_IsValid( 3 ctx1bga

361 self.metadata, DLPACK_TENSOR_USED_NAME): 3 ctx1bga

362 data = cpython.PyCapsule_GetPointer( 3 ctx1bga

363 self.metadata, DLPACK_TENSOR_USED_NAME) 3 ctx1bga

364 dlm_tensor = <DLManagedTensor*>data 3 ctx1bga

365 if dlm_tensor.deleter != NULL: 3 ctx1bga

366 dlm_tensor.deleter(dlm_tensor) 2 ctx1ba

367  

368 def view( 

369 self, layout : _StridedLayout | None = None, dtype : numpy.dtype | None = None 

370 ) -> StridedMemoryView: 

371 """ 

372 Creates a new view with adjusted layout and dtype. 

373 Same as calling :meth:`from_buffer` with the current buffer. 

374 """ 

375 cdef StridedMemoryView view = StridedMemoryView.__new__(self.__class__) 16 ctx2N O M QbRbcb5 Z S T U V W X 2 c

376 if layout is None and dtype is None: 16 ctx2N O M QbRbcb5 Z S T U V W X 2 c

377 return self 1 ctx2cb

378 if layout is None: 15 ctx2N O M QbRb5 Z S T U V W X 2 c

379 layout = self.get_layout() 7 ctx1NOM5Z2c

380 if dtype is None: 15 ctx2N O M QbRb5 Z S T U V W X 2 c

381 dtype = self.get_dtype() 8 ctx2QbRbS T U V W X

382 view_buffer_strided(view, self.get_buffer(), layout, dtype, self.readonly) 15 ctx2N O M QbRb5 Z S T U V W X 2 c

383 return view 14 ctx2N O M QbRbZ S T U V W X 2 c

384  

385 def as_tensor_map( 

386 self, 

387 box_dim: tuple[int, ...] | None = None, 

388 *, 

389 options: TensorMapDescriptorOptions | None = None, 

390 element_strides: tuple[int, ...] | None = None, 

391 data_type: object = None, 

392 interleave: object = None, 

393 swizzle: object = None, 

394 l2_promotion: object = None, 

395 oob_fill: object = None, 

396 ) -> object: 

397 """Create a tiled :obj:`TensorMapDescriptor` from this view. 

398  

399 This is the public entry point for creating tiled tensor map 

400 descriptors in ``cuda.core``. Pass either ``box_dim`` and the 

401 individual keyword arguments directly, or provide bundled tiled 

402 options via ``options=``. 

403 """ 

404 from cuda.core._tensor_map import TensorMapDescriptor 8 ctx2( ib6 3 % 7 | kb

405  

406 kwargs = {} 8 ctx2( ib6 3 % 7 | kb

407 if options is not None: 8 ctx2( ib6 3 % 7 | kb

408 kwargs["options"] = options 

409 if element_strides is not None: 8 ctx2( ib6 3 % 7 | kb

410 kwargs["element_strides"] = element_strides 3 ctx1(63

411 if data_type is not None: 8 ctx2( ib6 3 % 7 | kb

412 kwargs["data_type"] = data_type 4 ctx1(63%

413 if interleave is not None: 8 ctx2( ib6 3 % 7 | kb

414 kwargs["interleave"] = interleave 1 ctx1(

415 if swizzle is not None: 8 ctx2( ib6 3 % 7 | kb

416 kwargs["swizzle"] = swizzle 1 ctx1(

417 if l2_promotion is not None: 8 ctx2( ib6 3 % 7 | kb

418 kwargs["l2_promotion"] = l2_promotion 1 ctx1(

419 if oob_fill is not None: 8 ctx2( ib6 3 % 7 | kb

420 kwargs["oob_fill"] = oob_fill 1 ctx1(

421 return TensorMapDescriptor._from_tiled(self, box_dim, **kwargs) 8 ctx2( ib6 3 % 7 | kb

422  

423 def copy_from( 

424 self, 

425 StridedMemoryView other not None: StridedMemoryView, 

426 stream: Stream | None, 

427 allocator: object = None, 

428 blocking: bool | None = None, 

429 ) -> None: 

430 """ 

431 Copies the data from the other view into this view. 

432  

433 The copy can be performed between following memory spaces: 

434 host-to-device, device-to-host, device-to-device (on the same device). 

435  

436 Parameters 

437 ---------- 

438 other : StridedMemoryView 

439 The view to copy data from. 

440 stream : Stream | None, optional 

441 The stream to schedule the copy on. 

442 allocator : MemoryResource | None, optional 

443 If temporary buffers are needed, the specified memory resources 

444 will be used to allocate the memory. If not specified, default 

445 resources will be used. 

446 blocking : bool | None, optional 

447 Whether the call should block until the copy is complete. 

448 * ``True``: the ``stream`` is synchronized with the host at the end of the call, 

449 blocking until the copy is complete. 

450 * ``False``: if possible, the call returns immediately once the copy is scheduled. 

451 However, in some cases of host-to-device or device-to-host copies, the call may 

452 still synchronize with the host if necessary. 

453 * ``None`` (default): 

454 * for device-to-device, it defaults to ``False`` (non-blocking), 

455 * for host-to-device or device-to-host, it defaults to ``True`` (blocking). 

456 """ 

457 raise NotImplementedError("Sorry, not supported: copy_from") 1 ctx2gb

458  

459 def copy_to( 

460 self, 

461 StridedMemoryView other not None: StridedMemoryView, 

462 stream: Stream | None = None, 

463 allocator: object = None, 

464 blocking: bool | None = None, 

465 ) -> None: 

466 """ 

467 Copies the data from this view into the ``other`` view. 

468  

469 For details, see :meth:`copy_from`. 

470 """ 

471 raise NotImplementedError("Sorry, not supported: copy_to") 1 ctx2hb

472  

473 def __dlpack__( 

474 self, 

475 *, 

476 stream: int | None = None, 

477 max_version: tuple[int, int] | None = None, 

478 dl_device: tuple[int, int] | None = None, 

479 copy: bool | None = None, 

480 ) -> object: 

481 # Similar to Buffer.__dlpack__: no implicit synchronization is performed. 

482 if dl_device is not None: 45 ctx1}LNOMw'PY0IstxyzEFGHABCDuKJpfedbkhqlmrnogvcia

483 raise BufferError("Sorry, not supported: dl_device other than None") 1 ctx1}

484 if copy is True: 45 ctx1}LNOMw'PY0IstxyzEFGHABCDuKJpfedbkhqlmrnogvcia

485 raise BufferError("Sorry, not supported: copy=True") 1 ctx1}

486  

487 cdef bint versioned 

488 if max_version is None: 45 ctx1}LNOMw'PY0IstxyzEFGHABCDuKJpfedbkhqlmrnogvcia

489 versioned = False 9 ctx1LNOM'Jbga

490 else: 

491 if not isinstance(max_version, tuple) or len(max_version) != 2: 38 ctx1}wPY0IstxyzEFGHABCDuKpfedbkhqlmrnovcia

492 raise BufferError(f"Expected max_version tuple[int, int], got {max_version}") 1 ctx1}

493 versioned = max_version >= (1, 0) 37 ctx1wPY0IstxyzEFGHABCDuKpfedbkhqlmrnovcia

494  

495 # NOTE: stream is accepted for protocol compatibility but not used. 

496 cdef object capsule = _smv_make_py_capsule(self, versioned) 44 ctx1LNOMw'PY0IstxyzEFGHABCDuKJpfedbkhqlmrnogvcia

497 return capsule 38 ctx1LwPIstxyzEFGHABCDuKJpfedbkhqlmrnogvcia

498  

499 def __dlpack_device__(self) -> tuple[int, int]: 

500 cdef _DLDeviceType device_type 

501 cdef int32_t device_id 

502 _smv_get_dl_device(self, &device_type, &device_id) 28 ctx2bb{ ib6 3 % 7 kbL w P f e d b k h q l m r n o g v c i a

503 return (<int>device_type, int(device_id)) 28 ctx2bb{ ib6 3 % 7 kbL w P f e d b k h q l m r n o g v c i a

504  

505 @property 

506 def _layout(self) -> _StridedLayout: 

507 """ 

508 The layout of the tensor. For StridedMemoryView created from DLPack or CAI, 

509 the layout is inferred from the tensor object's metadata. 

510 """ 

511 return self.get_layout() 42 ctx2VbWbXbYbZb0b1b2b3b4b5b6b7b8b9b!b#b$b%b'b(b)b*b+b,b-b.b/b:b;b=b?bQbRbecZ S T U V W X

512  

513 @property 

514 def size(self) -> int: 

515 return self.get_layout().get_volume() 40 ctx28 9 ! # $ + , - . / ? @ [ ] ^ nbobxbpbybqbzbrbsbtbAbBbCbDbEbFbubGbHbIbvbJbKbwbMb

516  

517 @property 

518 def shape(self) -> tuple[int, ...]: 

519 """ 

520 Shape of the tensor. 

521 """ 

522 return self.get_layout().get_shape_tuple() 100 ctx2^b_b: 8 9 ! # $ + , - . / ? @ [ ] ^ nbobxbpbybqbzbrbsbtbAbBbCbDbEbFbubGbHbIbvbJbKbwbVbWbXbYbZb0b1b2b3b4b5b6b7b8b9b!b#b$b%b'b(b)b*b+b,b-b.b/b:b;b=b?bQbRbMbNbPbmblb1 dbS T U V W X _ ` ; ) = ~ j b k c i

523  

524 @property 

525 def strides(self) -> tuple[int, ...] | None: 

526 """ 

527 Strides of the tensor (in **counts**, not bytes). 

528 """ 

529 return self.get_layout().get_strides_tuple() 76 ctx2: 8 9 ! # $ + , - . / ? @ [ ] ^ nbobxbpbybqbzbrbsbtbAbBbCbDbEbFbubGbHbIbvbJbKbwbVbWbXbYbZb0b1b2b3b4b5b6b7b8b9b!b#b$b%b'b(b)b*b+b,b-b.b/b:b;b=b?bMbNblb1

530  

531 @property 

532 def dtype(self) -> numpy.dtype | None: 

533 """ 

534 Data type of the tensor. 

535  

536 Supports standard NumPy dtypes as well as narrow data types (e.g., ``bfloat16``) 

537 when the optional `ml_dtypes <https://github.com/jax-ml/ml_dtypes>`_ package is 

538 installed. If ``ml_dtypes`` is not available and such a tensor is encountered, 

539 a :obj:`NotImplementedError` will be raised. 

540 """ 

541 return self.get_dtype() 75 ctx2nbobxbpbybqbzbrbsbtbAbBbCbDbEbFbubGbHbIbvbJbKbwbVbWbXbYbZb0b1b2b3b4b5b6b7b8b9b!b#b$b%b'b(b)b*b+b,b-b.b/b:b;b=b?b`blb1 Z S T U V W X Y 0 h q l m r n o

542  

543 def __repr__(self) -> str: 

544 return (f"StridedMemoryView(ptr={self.ptr},\n" 2 ctx2lb1

545 + f" shape={self.shape},\n" 2 ctx2lb1

546 + f" strides={self.strides},\n" 2 ctx2lb1

547 + f" itemsize={self._layout.itemsize},\n" 2 ctx2lb1

548 + f" dtype={get_simple_repr(self.dtype)},\n" 2 ctx2lb1

549 + f" device_id={self.device_id},\n" 2 ctx2lb1

550 + f" is_device_accessible={self.is_device_accessible},\n" 2 ctx2lb1

551 + f" readonly={self.readonly},\n" 2 ctx2lb1

552 + f" exporting_obj={get_simple_repr(self.exporting_obj)})") 2 ctx2lb1

553  

554 @cython.critical_section 

555 cdef inline _StridedLayout get_layout(self): 

556 cdef _StridedLayout layout 

557 if self._layout is None: 159 ctx2^b_b: 6 3 % 7 | kb8 9 ! # $ + , - . / ? @ [ ] ^ L N O M nbobxbpbybqbzbrbsbtbAbBbCbDbEbFbubGbHbIbvbJbKbwb{bVbWbXbYbZb0b1b2b3b4b5b6b7b8b9b!b#b$b%b'b(b)b*b+b,b-b.b/b:b;b=b?bQbRbMb]bNbw ' ecPbmblb1 5 Z dbLbS T U V W X _ ` ; ) = ~ ebQ R Ob* 4 j P Y 0 I s t x y z E F G H A B C D u K J p f 2 e d b k h q l m r n o g v c i a

558 if self.dl_tensor: 118 ctx2: 6 3 % 7 | 8 9 ! # $ + , - . / ? @ [ ] ^ L N O M nbobxbpbybqbzbrbsbtbAbBbCbDbEbFbubGbHbIbvbJbKbwb{bMb]bNbw ecPbmb1 5 Z dbLbS T U V W X _ ` ; ) = ~ ebQ R 4 j P Y 0 I s t x y z E F G H A B C D u K J p f 2 e d b k h q l m r n o g v c i a

559 layout = layout_from_dlpack(self.dl_tensor) 75 ctx1:6389!#$+,-./?@[]^NOMw15ZSTUVWX_`;)=~QRjIstxyzEFGHABCDuKJp2edbkhqlmrnogvcia

560 elif self.metadata is not None: 45 ctx2% 7 | L nbobxbpbybqbzbrbsbtbAbBbCbDbEbFbubGbHbIbvbJbKbwb{bMb]bNbecPbmbdbLbeb4 P Y 0 f e d

561 layout = layout_from_cai(self.metadata) 44 ctx2% 7 | L nbobxbpbybqbzbrbsbtbAbBbCbDbEbFbubGbHbIbvbJbKbwb{bMb]bNbPbmbdbLbeb4 P Y 0 f e d

562 else: 

563 raise ValueError("Cannot infer layout from the exporting object") 1 ctx2ec

564 if self._layout is None: 115 ctx2: 6 3 % 7 | 8 9 ! # $ + , - . / ? @ [ ] ^ L N O M nbobxbpbybqbzbrbsbtbAbBbCbDbEbFbubGbHbIbvbJbKbwbMbNbw Pbmb1 5 Z dbLbS T U V W X _ ` ; ) = ~ ebQ R 4 j P Y 0 I s t x y z E F G H A B C D u K J p f 2 e d b k h q l m r n o g v c i a

565 self._layout = layout 115 ctx2: 6 3 % 7 | 8 9 ! # $ + , - . / ? @ [ ] ^ L N O M nbobxbpbybqbzbrbsbtbAbBbCbDbEbFbubGbHbIbvbJbKbwbMbNbw Pbmb1 5 Z dbLbS T U V W X _ ` ; ) = ~ ebQ R 4 j P Y 0 I s t x y z E F G H A B C D u K J p f 2 e d b k h q l m r n o g v c i a

566 return self._layout 156 ctx2^b_b: 6 3 % 7 | kb8 9 ! # $ + , - . / ? @ [ ] ^ L N O M nbobxbpbybqbzbrbsbtbAbBbCbDbEbFbubGbHbIbvbJbKbwbVbWbXbYbZb0b1b2b3b4b5b6b7b8b9b!b#b$b%b'b(b)b*b+b,b-b.b/b:b;b=b?bQbRbMbNbw ' Pbmblb1 5 Z dbLbS T U V W X _ ` ; ) = ~ ebQ R Ob* 4 j P Y 0 I s t x y z E F G H A B C D u K J p f 2 e d b k h q l m r n o g v c i a

567  

568 @cython.critical_section 

569 cdef inline object get_buffer(self): 

570 """ 

571 Returns Buffer instance with the underlying data. 

572 If the SMV was created from a Buffer, it will return the same Buffer instance. 

573 Otherwise, it will create a new instance with owner set to the exporting object. 

574 """ 

575 cdef object buffer 

576 if self._buffer is None: 20 ctx2ib% 7 kbL N O M QbRb5 Z S T U V W X 2 c

577 if isinstance(self.exporting_obj, Buffer): 16 ctx1%7LNOM5ZSTUVWX2c

578 buffer = self.exporting_obj 1 ctx12

579 else: 

580 buffer = Buffer.from_handle(self.ptr, 0, owner=self.exporting_obj) 15 ctx1%7LNOM5ZSTUVWXc

581 if self._buffer is None: 16 ctx1%7LNOM5ZSTUVWX2c

582 self._buffer = buffer 16 ctx1%7LNOM5ZSTUVWX2c

583 return self._buffer 20 ctx2ib% 7 kbL N O M QbRb5 Z S T U V W X 2 c

584  

585 @cython.critical_section 

586 cdef inline object get_dtype(self): 

587 cdef object dtype 

588 if self._dtype is None: 122 ctx2ib7 kbL N O M nbobxbpbybqbzbrbsbtbAbBbCbDbEbFbubGbHbIbvbJbKbwbVbWbXbYbZb0b1b2b3b4b5b6b7b8b9b!b#b$b%b'b(b)b*b+b,b-b.b/b:b;b=b?bQbRb`bw ' lb1 Z S T U V W X ebQ R Ob* 4 j P Y 0 I s t x y z E F G H A B C D u K J p f e d b k h q l m r n o g v c i a

589 dtype = None 82 ctx2ib7 L nbobxbpbybqbzbrbsbtbAbBbCbDbEbFbubGbHbIbvbJbKbwbw ' lb1 S T U V W X ebQ R Ob* 4 j P Y 0 I s t x y z E F G H A B C D u K J p f e d b k h q l m r n o g v c i a

590 if self.dl_tensor != NULL: 82 ctx2ib7 L nbobxbpbybqbzbrbsbtbAbBbCbDbEbFbubGbHbIbvbJbKbwbw ' lb1 S T U V W X ebQ R Ob* 4 j P Y 0 I s t x y z E F G H A B C D u K J p f e d b k h q l m r n o g v c i a

591 dtype = dtype_dlpack_to_numpy(&self.dl_tensor.dtype) 45 ctx1w1STUVWXQRjIstxyzEFGHABCDuKJpedbkhqlmrnogvcia

592 elif isinstance(self.metadata, int): 39 ctx2ib7 L nbobxbpbybqbzbrbsbtbAbBbCbDbEbFbubGbHbIbvbJbKbwb' lbebOb* 4 P Y 0 f e d

593 # AOTI dtype code stored by the torch tensor bridge 

594 dtype = _get_tensor_bridge().resolve_aoti_dtype( 

595 self.metadata) 

596 elif self.metadata is not None: 39 ctx2ib7 L nbobxbpbybqbzbrbsbtbAbBbCbDbEbFbubGbHbIbvbJbKbwb' lbebOb* 4 P Y 0 f e d

597 dtype = _typestr2dtype(self.metadata["typestr"]) 34 ctx27 L nbobxbpbybqbzbrbsbtbAbBbCbDbEbFbubGbHbIbvbJbKbwbeb4 P Y 0 f e d

598 if self._dtype is None: 82 ctx2ib7 L nbobxbpbybqbzbrbsbtbAbBbCbDbEbFbubGbHbIbvbJbKbwbw ' lb1 S T U V W X ebQ R Ob* 4 j P Y 0 I s t x y z E F G H A B C D u K J p f e d b k h q l m r n o g v c i a

599 self._dtype = dtype 82 ctx2ib7 L nbobxbpbybqbzbrbsbtbAbBbCbDbEbFbubGbHbIbvbJbKbwbw ' lb1 S T U V W X ebQ R Ob* 4 j P Y 0 I s t x y z E F G H A B C D u K J p f e d b k h q l m r n o g v c i a

600 return self._dtype 122 ctx2ib7 kbL N O M nbobxbpbybqbzbrbsbtbAbBbCbDbEbFbubGbHbIbvbJbKbwbVbWbXbYbZb0b1b2b3b4b5b6b7b8b9b!b#b$b%b'b(b)b*b+b,b-b.b/b:b;b=b?bQbRb`bw ' lb1 Z S T U V W X ebQ R Ob* 4 j P Y 0 I s t x y z E F G H A B C D u K J p f e d b k h q l m r n o g v c i a

601  

602  

603cdef void _smv_pycapsule_deleter(object capsule) noexcept: 

604 cdef DLManagedTensor* dlm_tensor 

605 cdef DLManagedTensorVersioned* dlm_tensor_ver 

606 # Do not invoke the deleter on a used capsule. 

607 if cpython.PyCapsule_IsValid(capsule, DLPACK_TENSOR_UNUSED_NAME): 40 ctx2L w @bSbTbUbP I s t x y z E F G H A B C D u K J p f e d b k h q l m r n o c i a

608 dlm_tensor = <DLManagedTensor*>( 2 ctx1LJ

609 cpython.PyCapsule_GetPointer(capsule, DLPACK_TENSOR_UNUSED_NAME) 2 ctx1LJ

610 ) 

611 if dlm_tensor.deleter: 2 ctx1LJ

612 dlm_tensor.deleter(dlm_tensor) 2 ctx1LJ

613 elif cpython.PyCapsule_IsValid(capsule, DLPACK_VERSIONED_TENSOR_UNUSED_NAME): 38 ctx2w @bSbTbUbP I s t x y z E F G H A B C D u K p f e d b k h q l m r n o c i a

614 dlm_tensor_ver = <DLManagedTensorVersioned*>( 1 ctx1p

615 cpython.PyCapsule_GetPointer(capsule, DLPACK_VERSIONED_TENSOR_UNUSED_NAME) 1 ctx1p

616 ) 

617 if dlm_tensor_ver.deleter: 1 ctx1p

618 dlm_tensor_ver.deleter(dlm_tensor_ver) 1 ctx1p

619  

620  

621cdef inline void _smv_release_export_resources(void* manager_ctx, int64_t* shape_ptr) noexcept with gil: 

622 if shape_ptr: 46 ctx1LNOMw'*4PY0IstxyzEFGHABCDuKJpfedbkhqlmrnogvcia

623 stdlib.free(shape_ptr) 37 ctx1LwPIstxyzEFGHABCDuJpfedbkhqlmrnogvcia

624 if manager_ctx: 46 ctx1LNOMw'*4PY0IstxyzEFGHABCDuKJpfedbkhqlmrnogvcia

625 cpython.Py_DECREF(<object>manager_ctx) 46 ctx1LNOMw'*4PY0IstxyzEFGHABCDuKJpfedbkhqlmrnogvcia

626  

627  

628cdef void _smv_deleter(DLManagedTensor* tensor) noexcept with gil: 

629 if tensor: 11 ctx1LNOM'Y0Jbga

630 _smv_release_export_resources(tensor.manager_ctx, tensor.dl_tensor.shape) 9 ctx1LNOM'Jbga

631 tensor.manager_ctx = NULL 9 ctx1LNOM'Jbga

632 stdlib.free(tensor) 9 ctx1LNOM'Jbga

633  

634  

635cdef void _smv_versioned_deleter(DLManagedTensorVersioned* tensor) noexcept with gil: 

636 if tensor: 43 ctx1NOMw'*4PY0IstxyzEFGHABCDuKpfedbkhqlmrnovcia

637 _smv_release_export_resources(tensor.manager_ctx, tensor.dl_tensor.shape) 39 ctx1w*4PY0IstxyzEFGHABCDuKpfedbkhqlmrnovcia

638 tensor.manager_ctx = NULL 39 ctx1w*4PY0IstxyzEFGHABCDuKpfedbkhqlmrnovcia

639 stdlib.free(tensor) 39 ctx1w*4PY0IstxyzEFGHABCDuKpfedbkhqlmrnovcia

640  

641  

642cdef inline DLManagedTensorVersioned* _smv_allocate_dlm_tensor_versioned() except? NULL: 

643 cdef DLManagedTensorVersioned* dlm_tensor_ver = NULL 40 ctx1w*4jPY0IstxyzEFGHABCDuKpfedbkhqlmrnovcia

644 dlm_tensor_ver = <DLManagedTensorVersioned*>stdlib.malloc(sizeof(DLManagedTensorVersioned)) 40 ctx1w*4jPY0IstxyzEFGHABCDuKpfedbkhqlmrnovcia

645 if dlm_tensor_ver == NULL: 40 ctx1w*4jPY0IstxyzEFGHABCDuKpfedbkhqlmrnovcia

646 raise MemoryError() 

647 dlm_tensor_ver.dl_tensor.shape = NULL 40 ctx1w*4jPY0IstxyzEFGHABCDuKpfedbkhqlmrnovcia

648 dlm_tensor_ver.manager_ctx = NULL 40 ctx1w*4jPY0IstxyzEFGHABCDuKpfedbkhqlmrnovcia

649 return dlm_tensor_ver 40 ctx1w*4jPY0IstxyzEFGHABCDuKpfedbkhqlmrnovcia

650  

651  

652cdef inline DLManagedTensor* _smv_allocate_dlm_tensor() except? NULL: 

653 cdef DLManagedTensor* dlm_tensor = NULL 9 ctx1LNOM'Jbga

654 dlm_tensor = <DLManagedTensor*>stdlib.malloc(sizeof(DLManagedTensor)) 9 ctx1LNOM'Jbga

655 if dlm_tensor == NULL: 9 ctx1LNOM'Jbga

656 raise MemoryError() 

657 dlm_tensor.dl_tensor.shape = NULL 9 ctx1LNOM'Jbga

658 dlm_tensor.manager_ctx = NULL 9 ctx1LNOM'Jbga

659 return dlm_tensor 9 ctx1LNOM'Jbga

660  

661  

662cdef inline int _smv_dtype_numpy_to_dlpack(object dtype_obj, DLDataType* out_dtype) except -1: 

663 cdef object np_dtype = numpy.dtype(dtype_obj) 48 ctx2L N O M w ebQ R 4 j P Y 0 I s t x y z E F G H A B C D u K J p f e d b k h q l m r n o g v c i a

664 if np_dtype.fields is not None: 48 ctx2L N O M w ebQ R 4 j P Y 0 I s t x y z E F G H A B C D u K J p f e d b k h q l m r n o g v c i a

665 raise BufferError("Structured dtypes are not supported for DLPack export") 1 ctx1O

666 if not np_dtype.isnative and np_dtype.byteorder not in ("=", "|"): 47 ctx2L N M w ebQ R 4 j P Y 0 I s t x y z E F G H A B C D u K J p f e d b k h q l m r n o g v c i a

667 raise BufferError("Non-native-endian dtypes are not supported for DLPack export") 3 ctx2N eb4

668  

669 cdef str kind = np_dtype.kind 44 ctx1LMwQRjPY0IstxyzEFGHABCDuKJpfedbkhqlmrnogvcia

670 cdef int bits = np_dtype.itemsize * 8 44 ctx1LMwQRjPY0IstxyzEFGHABCDuKJpfedbkhqlmrnogvcia

671 cdef uint8_t code 

672 if kind == "b": 44 ctx1LMwQRjPY0IstxyzEFGHABCDuKJpfedbkhqlmrnogvcia

673 if bits != 8: 1 ctx1I

674 raise BufferError(f"Unsupported bool dtype itemsize: {np_dtype.itemsize}") 

675 code = <uint8_t>kDLBool 1 ctx1I

676 elif kind == "i": 43 ctx1LMwQRjPY0stxyzEFGHABCDuKJpfedbkhqlmrnogvcia

677 if bits not in (8, 16, 32, 64): 25 ctx1wQRPEFGHJpfbkhqlmrnogvcia

678 raise BufferError(f"Unsupported signed integer dtype: {np_dtype}") 

679 code = <uint8_t>kDLInt 25 ctx1wQRPEFGHJpfbkhqlmrnogvcia

680 elif kind == "u": 18 ctx1LMjY0stxyzABCDuKed

681 if bits not in (8, 16, 32, 64): 4 ctx1ABCD

682 raise BufferError(f"Unsupported unsigned integer dtype: {np_dtype}") 

683 code = <uint8_t>kDLUInt 4 ctx1ABCD

684 elif kind == "f": 14 ctx1LMjY0stxyzuKed

685 if bits not in (16, 32, 64): 8 ctx1Lj0xyzed

686 raise BufferError(f"Unsupported floating dtype: {np_dtype}") 1 ctx10

687 code = <uint8_t>kDLFloat 7 ctx1Ljxyzed

688 elif kind == "c": 6 ctx1MYstuK

689 if bits not in (64, 128): 5 ctx1YstuK

690 raise BufferError(f"Unsupported complex dtype: {np_dtype}") 1 ctx1Y

691 code = <uint8_t>kDLComplex 4 ctx1stuK

692 else: 

693 raise BufferError(f"Unsupported dtype for DLPack export: {np_dtype}") 1 ctx1M

694  

695 out_dtype.code = code 41 ctx1LwQRjPIstxyzEFGHABCDuKJpfedbkhqlmrnogvcia

696 out_dtype.bits = <uint8_t>bits 41 ctx1LwQRjPIstxyzEFGHABCDuKJpfedbkhqlmrnogvcia

697 out_dtype.lanes = <uint16_t>1 41 ctx1LwQRjPIstxyzEFGHABCDuKJpfedbkhqlmrnogvcia

698 return 0 41 ctx1LwQRjPIstxyzEFGHABCDuKJpfedbkhqlmrnogvcia

699  

700  

701cdef inline int _smv_get_dl_device( 

702 StridedMemoryView view, 

703 _DLDeviceType* out_device_type, 

704 int32_t* out_device_id, 

705) except -1: 

706 cdef _DLDeviceType device_type 

707 cdef int32_t device_id 

708 cdef object buf 

709 if view.dl_tensor != NULL: 49 ctx2bb{ ib6 3 % 7 kbL w Q R j P I s t x y z E F G H A B C D u K J p f e d b k h q l m r n o g v c i a

710 device_type = view.dl_tensor.device.device_type 42 ctx2bb{ 6 3 w Q R j I s t x y z E F G H A B C D u K J p e d b k h q l m r n o g v c i a

711 if device_type == _kDLCUDA: 42 ctx2bb{ 6 3 w Q R j I s t x y z E F G H A B C D u K J p e d b k h q l m r n o g v c i a

712 device_id = view.dl_tensor.device.device_id 2 ctx163

713 else: 

714 # CPU, CUDAHost, and CUDAManaged use device_id=0 in DLPack. 

715 device_id = 0 40 ctx2bb{ w Q R j I s t x y z E F G H A B C D u K J p e d b k h q l m r n o g v c i a

716 elif view.is_device_accessible: 9 ctx2ib% 7 kbL P f e d

717 buf = view.get_buffer() 5 ctx2ib% 7 kbL

718 dev_type, dev_id = classify_dl_device(buf) 5 ctx2ib% 7 kbL

719 device_type = <_DLDeviceType>dev_type 5 ctx2ib% 7 kbL

720 device_id = <int32_t>dev_id 5 ctx2ib% 7 kbL

721 else: 

722 device_type = _kDLCPU 4 ctx1Pfed

723 device_id = 0 4 ctx1Pfed

724  

725 out_device_type[0] = device_type 49 ctx2bb{ ib6 3 % 7 kbL w Q R j P I s t x y z E F G H A B C D u K J p f e d b k h q l m r n o g v c i a

726 out_device_id[0] = device_id 49 ctx2bb{ ib6 3 % 7 kbL w Q R j P I s t x y z E F G H A B C D u K J p f e d b k h q l m r n o g v c i a

727 return 0 49 ctx2bb{ ib6 3 % 7 kbL w Q R j P I s t x y z E F G H A B C D u K J p f e d b k h q l m r n o g v c i a

728  

729  

730cdef inline int _smv_setup_dl_tensor_common( 

731 DLTensor* dl_tensor, 

732 StridedMemoryView view, 

733 _StridedLayout layout, 

734) except -1: 

735 cdef object dtype_obj = view.get_dtype() 51 ctx2L N O M w ' ebQ R Ob* 4 j P Y 0 I s t x y z E F G H A B C D u K J p f e d b k h q l m r n o g v c i a

736 if dtype_obj is None: 51 ctx2L N O M w ' ebQ R Ob* 4 j P Y 0 I s t x y z E F G H A B C D u K J p f e d b k h q l m r n o g v c i a

737 raise BufferError( 3 ctx2' Ob*

738 "Cannot export StridedMemoryView via DLPack without dtype information; " 

739 "create the view with dtype specified." 

740 ) 

741 _smv_dtype_numpy_to_dlpack(dtype_obj, &dl_tensor.dtype) 48 ctx2L N O M w ebQ R 4 j P Y 0 I s t x y z E F G H A B C D u K J p f e d b k h q l m r n o g v c i a

742 _smv_get_dl_device(view, &dl_tensor.device.device_type, &dl_tensor.device.device_id) 41 ctx1LwQRjPIstxyzEFGHABCDuKJpfedbkhqlmrnogvcia

743  

744 cdef int ndim = layout.base.ndim 41 ctx1LwQRjPIstxyzEFGHABCDuKJpfedbkhqlmrnogvcia

745 dl_tensor.ndim = ndim 41 ctx1LwQRjPIstxyzEFGHABCDuKJpfedbkhqlmrnogvcia

746 if layout.get_volume() == 0: 41 ctx1LwQRjPIstxyzEFGHABCDuKJpfedbkhqlmrnogvcia

747 dl_tensor.data = NULL 1 ctx1u

748 else: 

749 dl_tensor.data = <void*><intptr_t>view.ptr 40 ctx1LwQRjPIstxyzEFGHABCDKJpfedbkhqlmrnogvcia

750 dl_tensor.byte_offset = 0 41 ctx1LwQRjPIstxyzEFGHABCDuKJpfedbkhqlmrnogvcia

751 return 0 41 ctx1LwQRjPIstxyzEFGHABCDuKJpfedbkhqlmrnogvcia

752  

753  

754cdef inline int _smv_setup_dl_tensor(DLTensor* dl_tensor, StridedMemoryView view) except -1: 

755 cdef _StridedLayout layout = view.get_layout() 47 ctx1LNOMw'*4jPY0IstxyzEFGHABCDuKJpfedbkhqlmrnogvcia

756 _smv_setup_dl_tensor_common(dl_tensor, view, layout) 47 ctx1LNOMw'*4jPY0IstxyzEFGHABCDuKJpfedbkhqlmrnogvcia

757  

758 cdef int i 

759 cdef int64_t* shape_strides = NULL 39 ctx1LwjPIstxyzEFGHABCDuKJpfedbkhqlmrnogvcia

760 cdef int64_t* strides_src = NULL 39 ctx1LwjPIstxyzEFGHABCDuKJpfedbkhqlmrnogvcia

761 cdef int ndim = dl_tensor.ndim 39 ctx1LwjPIstxyzEFGHABCDuKJpfedbkhqlmrnogvcia

762 if ndim == 0: 39 ctx1LwjPIstxyzEFGHABCDuKJpfedbkhqlmrnogvcia

763 dl_tensor.shape = NULL 1 ctx1K

764 dl_tensor.strides = NULL 1 ctx1K

765 else: 

766 # DLPack v1.2+ requires non-NULL strides for ndim != 0. 

767 shape_strides = <int64_t*>stdlib.malloc(sizeof(int64_t) * 2 * ndim) 38 ctx1LwjPIstxyzEFGHABCDuJpfedbkhqlmrnogvcia

768 if shape_strides == NULL: 38 ctx1LwjPIstxyzEFGHABCDuJpfedbkhqlmrnogvcia

769 raise MemoryError() 

770 try: 38 ctx1LwjPIstxyzEFGHABCDuJpfedbkhqlmrnogvcia

771 strides_src = get_strides_ptr(layout.base) 38 ctx1LwjPIstxyzEFGHABCDuJpfedbkhqlmrnogvcia

772 for i in range(ndim): 38 ctx1LwjPIstxyzEFGHABCDuJpfedbkhqlmrnogvcia

773 shape_strides[i] = layout.base.shape[i] 38 ctx1LwjPIstxyzEFGHABCDuJpfedbkhqlmrnogvcia

774 shape_strides[i + ndim] = strides_src[i] 38 ctx1LwjPIstxyzEFGHABCDuJpfedbkhqlmrnogvcia

775 except Exception: 

776 stdlib.free(shape_strides) 

777 raise 

778 dl_tensor.shape = shape_strides 38 ctx1LwjPIstxyzEFGHABCDuJpfedbkhqlmrnogvcia

779 dl_tensor.strides = shape_strides + ndim 38 ctx1LwjPIstxyzEFGHABCDuJpfedbkhqlmrnogvcia

780 return 0 39 ctx1LwjPIstxyzEFGHABCDuKJpfedbkhqlmrnogvcia

781  

782  

783cdef inline int _smv_setup_dltensor_borrowed(DLTensor* dl_tensor, StridedMemoryView view) except -1: 

784 cdef _StridedLayout layout = view.get_layout() 4 ctx2ebQ R Ob

785 _smv_setup_dl_tensor_common(dl_tensor, view, layout) 4 ctx2ebQ R Ob

786  

787 if dl_tensor.ndim == 0: 2 ctx1QR

788 dl_tensor.shape = NULL 1 ctx1Q

789 dl_tensor.strides = NULL 1 ctx1Q

790 else: 

791 dl_tensor.shape = layout.base.shape 1 ctx1R

792 # For temporary/non-owning exchange we provide explicit strides. 

793 dl_tensor.strides = get_strides_ptr(layout.base) 1 ctx1R

794 return 0 2 ctx1QR

795  

796  

797cdef inline int _smv_fill_managed_tensor_versioned( 

798 DLManagedTensorVersioned* dlm_tensor_ver, 

799 StridedMemoryView view, 

800) except -1: 

801 cpython.Py_INCREF(view) 40 ctx1w*4jPY0IstxyzEFGHABCDuKpfedbkhqlmrnovcia

802 dlm_tensor_ver.manager_ctx = <void*>view 40 ctx1w*4jPY0IstxyzEFGHABCDuKpfedbkhqlmrnovcia

803 dlm_tensor_ver.deleter = _smv_versioned_deleter 40 ctx1w*4jPY0IstxyzEFGHABCDuKpfedbkhqlmrnovcia

804 dlm_tensor_ver.version.major = DLPACK_MAJOR_VERSION 40 ctx1w*4jPY0IstxyzEFGHABCDuKpfedbkhqlmrnovcia

805 dlm_tensor_ver.version.minor = DLPACK_MINOR_VERSION 40 ctx1w*4jPY0IstxyzEFGHABCDuKpfedbkhqlmrnovcia

806 dlm_tensor_ver.flags = DLPACK_FLAG_BITMASK_READ_ONLY if view.readonly else 0 40 ctx1w*4jPY0IstxyzEFGHABCDuKpfedbkhqlmrnovcia

807 _smv_setup_dl_tensor(&dlm_tensor_ver.dl_tensor, view) 40 ctx1w*4jPY0IstxyzEFGHABCDuKpfedbkhqlmrnovcia

808 return 0 36 ctx1wjPIstxyzEFGHABCDuKpfedbkhqlmrnovcia

809  

810  

811cdef inline int _smv_fill_managed_tensor( 

812 DLManagedTensor* dlm_tensor, 

813 StridedMemoryView view, 

814) except -1: 

815 cpython.Py_INCREF(view) 9 ctx1LNOM'Jbga

816 dlm_tensor.manager_ctx = <void*>view 9 ctx1LNOM'Jbga

817 dlm_tensor.deleter = _smv_deleter 9 ctx1LNOM'Jbga

818 _smv_setup_dl_tensor(&dlm_tensor.dl_tensor, view) 9 ctx1LNOM'Jbga

819 return 0 5 ctx1LJbga

820  

821  

822cdef object _smv_make_py_capsule(StridedMemoryView view, bint versioned): 

823 cdef DLManagedTensor* dlm_tensor = NULL 44 ctx1LNOMw'PY0IstxyzEFGHABCDuKJpfedbkhqlmrnogvcia

824 cdef DLManagedTensorVersioned* dlm_tensor_ver = NULL 44 ctx1LNOMw'PY0IstxyzEFGHABCDuKJpfedbkhqlmrnogvcia

825 cdef object capsule = None 44 ctx1LNOMw'PY0IstxyzEFGHABCDuKJpfedbkhqlmrnogvcia

826 cdef void* tensor_ptr = NULL 44 ctx1LNOMw'PY0IstxyzEFGHABCDuKJpfedbkhqlmrnogvcia

827 cdef const char* capsule_name 

828 try: 44 ctx1LNOMw'PY0IstxyzEFGHABCDuKJpfedbkhqlmrnogvcia

829 if versioned: 44 ctx1LNOMw'PY0IstxyzEFGHABCDuKJpfedbkhqlmrnogvcia

830 dlm_tensor_ver = _smv_allocate_dlm_tensor_versioned() 37 ctx1wPY0IstxyzEFGHABCDuKpfedbkhqlmrnovcia

831 _smv_fill_managed_tensor_versioned(dlm_tensor_ver, view) 37 ctx1wPY0IstxyzEFGHABCDuKpfedbkhqlmrnovcia

832 tensor_ptr = <void*>dlm_tensor_ver 35 ctx1wPIstxyzEFGHABCDuKpfedbkhqlmrnovcia

833 capsule_name = DLPACK_VERSIONED_TENSOR_UNUSED_NAME 35 ctx1wPIstxyzEFGHABCDuKpfedbkhqlmrnovcia

834 else: 

835 dlm_tensor = _smv_allocate_dlm_tensor() 9 ctx1LNOM'Jbga

836 _smv_fill_managed_tensor(dlm_tensor, view) 9 ctx1LNOM'Jbga

837 tensor_ptr = <void*>dlm_tensor 5 ctx1LJbga

838 capsule_name = DLPACK_TENSOR_UNUSED_NAME 5 ctx1LJbga

839 capsule = cpython.PyCapsule_New(tensor_ptr, capsule_name, _smv_pycapsule_deleter) 38 ctx1LwPIstxyzEFGHABCDuKJpfedbkhqlmrnogvcia

840 except Exception: 6 ctx1NOM'Y0

841 if capsule is None: 6 ctx1NOM'Y0

842 _smv_deleter(dlm_tensor) 6 ctx1NOM'Y0

843 _smv_versioned_deleter(dlm_tensor_ver) 6 ctx1NOM'Y0

844 raise 6 ctx1NOM'Y0

845 return capsule 38 ctx1LwPIstxyzEFGHABCDuKJpfedbkhqlmrnogvcia

846  

847  

848cdef inline StridedMemoryView _smv_from_dlpack_capsule(object capsule, object exporting_obj): 

849 cdef void* data = NULL 5 ctx2@bj SbTbUb

850 cdef DLTensor* dl_tensor = NULL 5 ctx2@bj SbTbUb

851 cdef DLManagedTensorVersioned* dlm_tensor_ver = NULL 5 ctx2@bj SbTbUb

852 cdef DLManagedTensor* dlm_tensor = NULL 5 ctx2@bj SbTbUb

853 cdef bint is_readonly = False 5 ctx2@bj SbTbUb

854 cdef const char* used_name = NULL 5 ctx2@bj SbTbUb

855 if cpython.PyCapsule_IsValid(capsule, DLPACK_VERSIONED_TENSOR_UNUSED_NAME): 5 ctx2@bj SbTbUb

856 data = cpython.PyCapsule_GetPointer(capsule, DLPACK_VERSIONED_TENSOR_UNUSED_NAME) 5 ctx2@bj SbTbUb

857 dlm_tensor_ver = <DLManagedTensorVersioned*>data 5 ctx2@bj SbTbUb

858 dl_tensor = &dlm_tensor_ver.dl_tensor 5 ctx2@bj SbTbUb

859 is_readonly = bool((dlm_tensor_ver.flags & DLPACK_FLAG_BITMASK_READ_ONLY) != 0) 5 ctx2@bj SbTbUb

860 used_name = DLPACK_VERSIONED_TENSOR_USED_NAME 5 ctx2@bj SbTbUb

861 elif cpython.PyCapsule_IsValid(capsule, DLPACK_TENSOR_UNUSED_NAME): 

862 data = cpython.PyCapsule_GetPointer(capsule, DLPACK_TENSOR_UNUSED_NAME) 

863 dlm_tensor = <DLManagedTensor*>data 

864 dl_tensor = &dlm_tensor.dl_tensor 

865 is_readonly = False 

866 used_name = DLPACK_TENSOR_USED_NAME 

867 else: 

868 raise BufferError("Invalid DLPack capsule") 

869  

870 cpython.PyCapsule_SetName(capsule, used_name) 5 ctx2@bj SbTbUb

871  

872 cdef StridedMemoryView view = StridedMemoryView.__new__(StridedMemoryView) 5 ctx2@bj SbTbUb

873 view.dl_tensor = dl_tensor 5 ctx2@bj SbTbUb

874 view.metadata = capsule 5 ctx2@bj SbTbUb

875 view.ptr = <intptr_t>(dl_tensor.data) + <intptr_t>(dl_tensor.byte_offset) 5 ctx2@bj SbTbUb

876 view.readonly = is_readonly 5 ctx2@bj SbTbUb

877 view.exporting_obj = exporting_obj 5 ctx2@bj SbTbUb

878 if dl_tensor.device.device_type == _kDLCPU: 5 ctx2@bj SbTbUb

879 view.device_id = -1 1 ctx1j

880 view.is_device_accessible = False 1 ctx1j

881 elif dl_tensor.device.device_type in (_kDLCUDA, _kDLCUDAHost, _kDLCUDAManaged): 1 ctx2Sb

882 view.device_id = dl_tensor.device.device_id 3 ctx2SbTbUb

883 view.is_device_accessible = True 3 ctx2SbTbUb

884 else: 

885 raise BufferError("device not supported") 1 ctx2@b

886 return view 4 ctx2j SbTbUb

887  

888  

889cdef int _smv_managed_tensor_allocator( 

890 DLTensor* prototype, 

891 DLManagedTensorVersioned** out, 

892 void* error_ctx, 

893 void (*SetError)(void* error_ctx, const char* kind, const char* message) noexcept, 

894) noexcept with gil: 

895 if out != NULL: 1 ctx2ic

896 out[0] = NULL 1 ctx2ic

897 if SetError != NULL: 1 ctx2ic

898 SetError(error_ctx, b"NotImplementedError", b"managed_tensor_allocator is not supported by StridedMemoryView") 

899 cpython.PyErr_SetString(NotImplementedError, b"managed_tensor_allocator is not supported by StridedMemoryView") 1 ctx2ic

900 return -1 1 ctx2ic

901  

902  

903cdef int _smv_managed_tensor_from_py_object_no_sync( 

904 void* py_object, 

905 DLManagedTensorVersioned** out, 

906) except -1 with gil: 

907 cdef DLManagedTensorVersioned* dlm_tensor_ver = NULL 4 ctx2* 4 abj

908 if out == NULL: 4 ctx2* 4 abj

909 raise RuntimeError("out cannot be NULL") 1 ctx2ab

910 out[0] = NULL 4 ctx2* 4 abj

911 cdef object obj = <object>py_object 4 ctx2* 4 abj

912 if not isinstance(obj, StridedMemoryView): 4 ctx2* 4 abj

913 raise TypeError("py_object must be a StridedMemoryView") 1 ctx2ab

914 try: 3 ctx1*4j

915 dlm_tensor_ver = _smv_allocate_dlm_tensor_versioned() 3 ctx1*4j

916 _smv_fill_managed_tensor_versioned(dlm_tensor_ver, <StridedMemoryView>obj) 3 ctx1*4j

917 except BaseException: 2 ctx1*4

918 # Rollback must also run for KeyboardInterrupt. 

919 _smv_versioned_deleter(dlm_tensor_ver) 2 ctx1*4

920 raise 2 ctx1*4

921 out[0] = dlm_tensor_ver 1 ctx1j

922 return 0 1 ctx1j

923  

924  

925cdef int _smv_managed_tensor_to_py_object_no_sync( 

926 DLManagedTensorVersioned* tensor, 

927 void** out_py_object, 

928) except -1 with gil: 

929 if out_py_object == NULL: 7 ctx2@bj SbTbUbuclc

930 raise RuntimeError("out_py_object cannot be NULL") 1 ctx2uc

931 out_py_object[0] = NULL 6 ctx2@bj SbTbUblc

932 if tensor == NULL: 6 ctx2@bj SbTbUblc

933 raise RuntimeError("tensor cannot be NULL") 1 ctx2lc

934 cdef object capsule = cpython.PyCapsule_New( 5 ctx2@bj SbTbUb

935 <void*>tensor, 

936 DLPACK_VERSIONED_TENSOR_UNUSED_NAME, 

937 _smv_pycapsule_deleter, 

938 ) 

939 cdef object py_view = _smv_from_dlpack_capsule(capsule, capsule) 5 ctx2@bj SbTbUb

940 cpython.Py_INCREF(py_view) 4 ctx2j SbTbUb

941 out_py_object[0] = <void*>py_view 4 ctx2j SbTbUb

942 return 0 4 ctx2j SbTbUb

943  

944  

945cdef int _smv_dltensor_from_py_object_no_sync( 

946 void* py_object, 

947 DLTensor* out, 

948) except -1 with gil: 

949 if out == NULL: 6 ctx2ebfbQ mcR Ob

950 raise RuntimeError("out cannot be NULL") 1 ctx2fb

951 cdef object obj = <object>py_object 5 ctx2ebQ mcR Ob

952 if not isinstance(obj, StridedMemoryView): 5 ctx2ebQ mcR Ob

953 raise TypeError("py_object must be a StridedMemoryView") 1 ctx2mc

954 _smv_setup_dltensor_borrowed(out, <StridedMemoryView>obj) 4 ctx2ebQ R Ob

955 return 0 2 ctx1QR

956  

957  

958cdef int _smv_current_work_stream( 

959 _DLDeviceType device_type, 

960 int32_t device_id, 

961 void** out_current_stream, 

962) except -1 with gil: 

963 if out_current_stream == NULL: 2 ctx2vcpc

964 raise RuntimeError("out_current_stream cannot be NULL") 1 ctx2vc

965 # cuda.core has no global/current stream state today. 

966 out_current_stream[0] = NULL 1 ctx2pc

967 return 0 1 ctx2pc

968  

969  

970cdef void _init_smv_dlpack_exchange_api(): 

971 global _SMV_DLPACK_EXCHANGE_API_INITED 

972 if _SMV_DLPACK_EXCHANGE_API_INITED: 

973 return 

974 _SMV_DLPACK_EXCHANGE_API.header.version.major = DLPACK_MAJOR_VERSION 

975 _SMV_DLPACK_EXCHANGE_API.header.version.minor = DLPACK_MINOR_VERSION 

976 _SMV_DLPACK_EXCHANGE_API.header.prev_api = NULL 

977 _SMV_DLPACK_EXCHANGE_API.managed_tensor_allocator = _smv_managed_tensor_allocator 

978 _SMV_DLPACK_EXCHANGE_API.managed_tensor_from_py_object_no_sync = _smv_managed_tensor_from_py_object_no_sync 

979 _SMV_DLPACK_EXCHANGE_API.managed_tensor_to_py_object_no_sync = _smv_managed_tensor_to_py_object_no_sync 

980 _SMV_DLPACK_EXCHANGE_API.dltensor_from_py_object_no_sync = _smv_dltensor_from_py_object_no_sync 

981 _SMV_DLPACK_EXCHANGE_API.current_work_stream = _smv_current_work_stream 

982 _SMV_DLPACK_EXCHANGE_API_INITED = True 

983  

984  

985_init_smv_dlpack_exchange_api() 

986# cdef classes are immutable types in Cython 3, so inject these attributes 

987# directly into the type dict. 

988(<dict>(<PyTypeObject*>StridedMemoryView).tp_dict)["__dlpack_c_exchange_api__"] = _SMV_DLPACK_EXCHANGE_API_CAPSULE 

989(<dict>(<PyTypeObject*>StridedMemoryView).tp_dict)["__c_dlpack_exchange_api__"] = _SMV_DLPACK_EXCHANGE_API_CAPSULE 

990PyType_Modified(<PyTypeObject*>StridedMemoryView) 

991  

992  

993cdef str get_simple_repr(obj): 

994 # TODO: better handling in np.dtype objects 

995 cdef object obj_class 

996 cdef str obj_repr 

997 if isinstance(obj, type): 2 ctx2lb1

998 obj_class = obj 

999 else: 

1000 obj_class = obj.__class__ 2 ctx2lb1

1001 if obj_class.__module__ in (None, "builtins"): 2 ctx2lb1

1002 obj_repr = obj_class.__name__ 1 ctx2lb

1003 else: 

1004 obj_repr = f"{obj_class.__module__}.{obj_class.__name__}" 2 ctx2lb1

1005 return obj_repr 2 ctx2lb1

1006  

1007  

1008  

1009cdef bint check_has_dlpack(obj) except*: 

1010 cdef bint has_dlpack 

1011 if hasattr(obj, "__dlpack__") and hasattr(obj, "__dlpack_device__"): 81 ctx2bb{ } jb: ( 6 3 % 7 | 8 9 ! # $ + , - . / ? @ [ ] ^ kcN O M gbhbw Pbgcmb1 cb5 Z _ ` ; ) fbQ R abj I s t x y z E F G H A B C D u K J p b k h q l m r n o g v c i a

1012 has_dlpack = True 74 ctx2bb{ } jb: ( 6 3 8 9 ! # $ + , - . / ? @ [ ] ^ N O M gbhbw 1 cb5 Z _ ` ; ) fbQ R abj I s t x y z E F G H A B C D u K J p b k h q l m r n o g v c i a

1013 elif hasattr(obj, "__cuda_array_interface__"): 7 ctx2% 7 | kcPbgcmb

1014 has_dlpack = False 6 ctx2% 7 | Pbgcmb

1015 else: 

1016 raise BufferError( 1 ctx2kc

1017 "the input object does not support any data exchange protocol") 

1018 return has_dlpack 80 ctx2bb{ } jb: ( 6 3 % 7 | 8 9 ! # $ + , - . / ? @ [ ] ^ N O M gbhbw Pbgcmb1 cb5 Z _ ` ; ) fbQ R abj I s t x y z E F G H A B C D u K J p b k h q l m r n o g v c i a

1019  

1020  

1021cdef class _StridedMemoryViewProxy: 

1022 cdef readonly: 

1023 object obj 

1024 bint has_dlpack 

1025  

1026 def __init__(self, obj: object) -> None: 

1027 self.obj = obj 7 ctx28 9 ! # $ gcmb

1028 self.has_dlpack = check_has_dlpack(obj) 7 ctx28 9 ! # $ gcmb

1029  

1030 cpdef StridedMemoryView view(self, stream_ptr=None): 

1031 if self.has_dlpack: 6 ctx28 9 ! # $ mb

1032 return StridedMemoryView.from_dlpack(self.obj, stream_ptr) 5 ctx189!#$

1033 else: 

1034 return StridedMemoryView.from_cuda_array_interface(self.obj, stream_ptr) 1 ctx2mb

1035  

1036  

1037cdef StridedMemoryView view_as_dlpack(obj, stream_ptr, view=None): 

1038 cdef int dldevice, device_id 

1039 cdef bint is_device_accessible, is_readonly = False 92 ctx2bb{ } jb: ( 6 3 8 9 ! # $ + , - . / ? @ [ ] ^ N O M gbhbw 1 cb5 Z S T U V W X _ ` ; ) = ~ fbQ R abj I s t x y z E F G H A B C D u K J p f 2 cce d |bb k h q l m r n o g v ~b}b[bc i a ac

1040 is_device_accessible = False 92 ctx2bb{ } jb: ( 6 3 8 9 ! # $ + , - . / ? @ [ ] ^ N O M gbhbw 1 cb5 Z S T U V W X _ ` ; ) = ~ fbQ R abj I s t x y z E F G H A B C D u K J p f 2 cce d |bb k h q l m r n o g v ~b}b[bc i a ac

1041 # Callers may pass an int-convertible stream wrapper (e.g. numba-cuda's 

1042 # CUstream). Normalize it so the -1 and truthiness checks below see an int. 

1043 if stream_ptr is not None: 92 ctx2bb{ } jb: ( 6 3 8 9 ! # $ + , - . / ? @ [ ] ^ N O M gbhbw 1 cb5 Z S T U V W X _ ` ; ) = ~ fbQ R abj I s t x y z E F G H A B C D u K J p f 2 cce d |bb k h q l m r n o g v ~b}b[bc i a ac

1044 stream_ptr = int(stream_ptr) 88 ctx2bb{ } jb: ( 6 3 8 9 ! # $ + , - . / ? @ [ ] ^ N O M gbhbw 1 cb5 Z S T U V W X _ ` ; ) = ~ fbQ R abj I s t x y z E F G H A B C D u K J p f 2 b k h q l m r n o g v ~b}b[bc i a ac

1045 dldevice, device_id = obj.__dlpack_device__() 92 ctx2bb{ } jb: ( 6 3 8 9 ! # $ + , - . / ? @ [ ] ^ N O M gbhbw 1 cb5 Z S T U V W X _ ` ; ) = ~ fbQ R abj I s t x y z E F G H A B C D u K J p f 2 cce d |bb k h q l m r n o g v ~b}b[bc i a ac

1046 if dldevice == _kDLCPU: 92 ctx2bb{ } jb: ( 6 3 8 9 ! # $ + , - . / ? @ [ ] ^ N O M gbhbw 1 cb5 Z S T U V W X _ ` ; ) = ~ fbQ R abj I s t x y z E F G H A B C D u K J p f 2 cce d |bb k h q l m r n o g v ~b}b[bc i a ac

1047 if device_id != 0: 82 ctx2} jb: ( 8 9 ! # $ + , - . / ? @ [ ] ^ N O M gbhbw 1 cb5 Z S T U V W X _ ` ; ) = ~ fbQ R abj I s t x y z E F G H A B C D u K J p f ccb k h q l m r n o g v ~b[bc i a

1048 raise ValueError(f"__dlpack_device__() reported device_id {device_id} for a kDLCPU tensor; expected 0") 1 ctx2~b

1049 device_id = -1 81 ctx2} jb: ( 8 9 ! # $ + , - . / ? @ [ ] ^ N O M gbhbw 1 cb5 Z S T U V W X _ ` ; ) = ~ fbQ R abj I s t x y z E F G H A B C D u K J p f ccb k h q l m r n o g v [bc i a

1050 if stream_ptr is None: 81 ctx2} jb: ( 8 9 ! # $ + , - . / ? @ [ ] ^ N O M gbhbw 1 cb5 Z S T U V W X _ ` ; ) = ~ fbQ R abj I s t x y z E F G H A B C D u K J p f ccb k h q l m r n o g v [bc i a

1051 raise BufferError("stream=None is ambiguous with view()") 1 ctx2cc

1052 elif stream_ptr == -1: 80 ctx2} jb: ( 8 9 ! # $ + , - . / ? @ [ ] ^ N O M gbhbw 1 cb5 Z S T U V W X _ ` ; ) = ~ fbQ R abj I s t x y z E F G H A B C D u K J p f b k h q l m r n o g v [bc i a

1053 stream_ptr = None 77 ctx2} jb: ( 8 9 ! # $ + , - . / ? @ [ ] ^ N O M gbhbw 1 cb5 Z S T U V W X _ ; = fbQ R abj I s t x y z E F G H A B C D u K J p f b k h q l m r n o g v [bc i a

1054 elif dldevice == _kDLCUDA: 

1055 if device_id < 0: 5 ctx26 3 2 |b}b

1056 raise ValueError( 1 ctx2}b

1057 f"__dlpack_device__() reported device_id {device_id} for a kDLCUDA tensor; expected 0 or more") 1 ctx2}b

1058 is_device_accessible = True 4 ctx26 3 2 |b

1059 # no need to check other stream values, it's a pass-through 

1060 if stream_ptr is None: 4 ctx26 3 2 |b

1061 raise BufferError("stream=None is ambiguous with view()") 1 ctx2|b

1062 elif dldevice in (_kDLCUDAHost, _kDLCUDAManaged): 1 ctx1e

1063 is_device_accessible = True 4 ctx2bb{ e d

1064 # just do a pass-through without any checks, as pinned/managed memory can be 

1065 # accessed on both host and device 

1066 else: 

1067 raise BufferError("device not supported") 1 ctx2ac

1068  

1069 cdef object capsule 

1070 try: 87 ctx2bb{ } jb: ( 6 3 8 9 ! # $ + , - . / ? @ [ ] ^ N O M gbhbw 1 cb5 Z S T U V W X _ ` ; ) = ~ fbQ R abj I s t x y z E F G H A B C D u K J p f 2 e d b k h q l m r n o g v [bc i a

1071 capsule = obj.__dlpack__( 87 ctx2bb{ } jb: ( 6 3 8 9 ! # $ + , - . / ? @ [ ] ^ N O M gbhbw 1 cb5 Z S T U V W X _ ` ; ) = ~ fbQ R abj I s t x y z E F G H A B C D u K J p f 2 e d b k h q l m r n o g v [bc i a

1072 stream=int(stream_ptr) if stream_ptr else None, 87 ctx2bb{ } jb: ( 6 3 8 9 ! # $ + , - . / ? @ [ ] ^ N O M gbhbw 1 cb5 Z S T U V W X _ ` ; ) = ~ fbQ R abj I s t x y z E F G H A B C D u K J p f 2 e d b k h q l m r n o g v [bc i a

1073 max_version=(DLPACK_MAJOR_VERSION, DLPACK_MINOR_VERSION)) 87 ctx2bb{ } jb: ( 6 3 8 9 ! # $ + , - . / ? @ [ ] ^ N O M gbhbw 1 cb5 Z S T U V W X _ ` ; ) = ~ fbQ R abj I s t x y z E F G H A B C D u K J p f 2 e d b k h q l m r n o g v [bc i a

1074 except TypeError: 3 ctx1bga

1075 capsule = obj.__dlpack__( 3 ctx1bga

1076 stream=int(stream_ptr) if stream_ptr else None) 3 ctx1bga

1077  

1078 cdef void* data = NULL 87 ctx2bb{ } jb: ( 6 3 8 9 ! # $ + , - . / ? @ [ ] ^ N O M gbhbw 1 cb5 Z S T U V W X _ ` ; ) = ~ fbQ R abj I s t x y z E F G H A B C D u K J p f 2 e d b k h q l m r n o g v [bc i a

1079 cdef DLTensor* dl_tensor = NULL 87 ctx2bb{ } jb: ( 6 3 8 9 ! # $ + , - . / ? @ [ ] ^ N O M gbhbw 1 cb5 Z S T U V W X _ ` ; ) = ~ fbQ R abj I s t x y z E F G H A B C D u K J p f 2 e d b k h q l m r n o g v [bc i a

1080 cdef DLManagedTensorVersioned* dlm_tensor_ver 

1081 cdef DLManagedTensor* dlm_tensor 

1082 cdef const char *used_name = NULL 87 ctx2bb{ } jb: ( 6 3 8 9 ! # $ + , - . / ? @ [ ] ^ N O M gbhbw 1 cb5 Z S T U V W X _ ` ; ) = ~ fbQ R abj I s t x y z E F G H A B C D u K J p f 2 e d b k h q l m r n o g v [bc i a

1083 if cpython.PyCapsule_IsValid( 87 ctx2bb{ } jb: ( 6 3 8 9 ! # $ + , - . / ? @ [ ] ^ N O M gbhbw 1 cb5 Z S T U V W X _ ` ; ) = ~ fbQ R abj I s t x y z E F G H A B C D u K J p f 2 e d b k h q l m r n o g v [bc i a

1084 capsule, DLPACK_VERSIONED_TENSOR_UNUSED_NAME): 

1085 data = cpython.PyCapsule_GetPointer( 86 ctx2bb{ } jb: ( 6 3 8 9 ! # $ + , - . / ? @ [ ] ^ N O M gbhbw 1 cb5 Z S T U V W X _ ` ; ) = ~ fbQ R abj I s t x y z E F G H A B C D u K J p f 2 e d b k h q l m r n o g v c i a

1086 capsule, DLPACK_VERSIONED_TENSOR_UNUSED_NAME) 

1087 dlm_tensor_ver = <DLManagedTensorVersioned*>data 86 ctx2bb{ } jb: ( 6 3 8 9 ! # $ + , - . / ? @ [ ] ^ N O M gbhbw 1 cb5 Z S T U V W X _ ` ; ) = ~ fbQ R abj I s t x y z E F G H A B C D u K J p f 2 e d b k h q l m r n o g v c i a

1088 dl_tensor = &dlm_tensor_ver.dl_tensor 86 ctx2bb{ } jb: ( 6 3 8 9 ! # $ + , - . / ? @ [ ] ^ N O M gbhbw 1 cb5 Z S T U V W X _ ` ; ) = ~ fbQ R abj I s t x y z E F G H A B C D u K J p f 2 e d b k h q l m r n o g v c i a

1089 is_readonly = bool((dlm_tensor_ver.flags & DLPACK_FLAG_BITMASK_READ_ONLY) != 0) 86 ctx2bb{ } jb: ( 6 3 8 9 ! # $ + , - . / ? @ [ ] ^ N O M gbhbw 1 cb5 Z S T U V W X _ ` ; ) = ~ fbQ R abj I s t x y z E F G H A B C D u K J p f 2 e d b k h q l m r n o g v c i a

1090 used_name = DLPACK_VERSIONED_TENSOR_USED_NAME 86 ctx2bb{ } jb: ( 6 3 8 9 ! # $ + , - . / ? @ [ ] ^ N O M gbhbw 1 cb5 Z S T U V W X _ ` ; ) = ~ fbQ R abj I s t x y z E F G H A B C D u K J p f 2 e d b k h q l m r n o g v c i a

1091 elif cpython.PyCapsule_IsValid( 4 ctx2b g [ba

1092 capsule, DLPACK_TENSOR_UNUSED_NAME): 

1093 data = cpython.PyCapsule_GetPointer( 3 ctx1bga

1094 capsule, DLPACK_TENSOR_UNUSED_NAME) 

1095 dlm_tensor = <DLManagedTensor*>data 3 ctx1bga

1096 dl_tensor = &dlm_tensor.dl_tensor 3 ctx1bga

1097 is_readonly = False 3 ctx1bga

1098 used_name = DLPACK_TENSOR_USED_NAME 3 ctx1bga

1099 else: 

1100 raise ValueError(f"__dlpack__() returned {capsule!r}, not an unconsumed DLPack capsule") 1 ctx2[b

1101  

1102 cpython.PyCapsule_SetName(capsule, used_name) 86 ctx2bb{ } jb: ( 6 3 8 9 ! # $ + , - . / ? @ [ ] ^ N O M gbhbw 1 cb5 Z S T U V W X _ ` ; ) = ~ fbQ R abj I s t x y z E F G H A B C D u K J p f 2 e d b k h q l m r n o g v c i a

1103  

1104 cdef StridedMemoryView buf = StridedMemoryView() if view is None else view 86 ctx2bb{ } jb: ( 6 3 8 9 ! # $ + , - . / ? @ [ ] ^ N O M gbhbw 1 cb5 Z S T U V W X _ ` ; ) = ~ fbQ R abj I s t x y z E F G H A B C D u K J p f 2 e d b k h q l m r n o g v c i a

1105 buf.dl_tensor = dl_tensor 86 ctx2bb{ } jb: ( 6 3 8 9 ! # $ + , - . / ? @ [ ] ^ N O M gbhbw 1 cb5 Z S T U V W X _ ` ; ) = ~ fbQ R abj I s t x y z E F G H A B C D u K J p f 2 e d b k h q l m r n o g v c i a

1106 buf.metadata = capsule 86 ctx2bb{ } jb: ( 6 3 8 9 ! # $ + , - . / ? @ [ ] ^ N O M gbhbw 1 cb5 Z S T U V W X _ ` ; ) = ~ fbQ R abj I s t x y z E F G H A B C D u K J p f 2 e d b k h q l m r n o g v c i a

1107 buf.ptr = <intptr_t>(dl_tensor.data) + <intptr_t>(dl_tensor.byte_offset) 86 ctx2bb{ } jb: ( 6 3 8 9 ! # $ + , - . / ? @ [ ] ^ N O M gbhbw 1 cb5 Z S T U V W X _ ` ; ) = ~ fbQ R abj I s t x y z E F G H A B C D u K J p f 2 e d b k h q l m r n o g v c i a

1108 buf.device_id = device_id 86 ctx2bb{ } jb: ( 6 3 8 9 ! # $ + , - . / ? @ [ ] ^ N O M gbhbw 1 cb5 Z S T U V W X _ ` ; ) = ~ fbQ R abj I s t x y z E F G H A B C D u K J p f 2 e d b k h q l m r n o g v c i a

1109 buf.is_device_accessible = is_device_accessible 86 ctx2bb{ } jb: ( 6 3 8 9 ! # $ + , - . / ? @ [ ] ^ N O M gbhbw 1 cb5 Z S T U V W X _ ` ; ) = ~ fbQ R abj I s t x y z E F G H A B C D u K J p f 2 e d b k h q l m r n o g v c i a

1110 buf.readonly = is_readonly 86 ctx2bb{ } jb: ( 6 3 8 9 ! # $ + , - . / ? @ [ ] ^ N O M gbhbw 1 cb5 Z S T U V W X _ ` ; ) = ~ fbQ R abj I s t x y z E F G H A B C D u K J p f 2 e d b k h q l m r n o g v c i a

1111 buf.exporting_obj = obj 86 ctx2bb{ } jb: ( 6 3 8 9 ! # $ + , - . / ? @ [ ] ^ N O M gbhbw 1 cb5 Z S T U V W X _ ` ; ) = ~ fbQ R abj I s t x y z E F G H A B C D u K J p f 2 e d b k h q l m r n o g v c i a

1112  

1113 return buf 86 ctx2bb{ } jb: ( 6 3 8 9 ! # $ + , - . / ? @ [ ] ^ N O M gbhbw 1 cb5 Z S T U V W X _ ` ; ) = ~ fbQ R abj I s t x y z E F G H A B C D u K J p f 2 e d b k h q l m r n o g v c i a

1114  

1115  

1116@functools.lru_cache 

1117def _typestr2dtype(str typestr) -> numpy.dtype: