[cuda.compute][cuda.coop]: Replace all usages of device arrays outside examples with new wrapper hat does not depend on cupy or numba-cuda #9653
+1,907
−1,605
There are no files selected for viewing
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,5 @@ | ||
| # Copyright (c) 2026, NVIDIA CORPORATION & AFFILIATES. ALL RIGHTS RESERVED. | ||
| # | ||
| # SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception | ||
|
|
||
| """Shared test utilities for cuda-cccl.""" |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,252 @@ | ||
| # Copyright (c) 2026, NVIDIA CORPORATION & AFFILIATES. ALL RIGHTS RESERVED. | ||
| # | ||
| # SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception | ||
|
|
||
| from __future__ import annotations | ||
|
|
||
| import math | ||
| import operator | ||
| from collections.abc import Iterable | ||
|
|
||
| import numpy as np | ||
| from numpy.typing import DTypeLike | ||
|
|
||
| from cuda.core import Buffer, Device, Stream | ||
|
|
||
|
|
||
| def get_compute_capability() -> tuple[int, int]: | ||
| return Device().compute_capability | ||
|
|
||
|
|
||
| def _normalize_shape(shape: int | Iterable[int]) -> tuple[int, ...]: | ||
| try: | ||
| dimensions = (operator.index(shape),) # type: ignore[arg-type] | ||
| except TypeError: | ||
| dimensions = tuple(operator.index(dimension) for dimension in shape) # type: ignore[union-attr] | ||
|
|
||
| if any(dimension < 0 for dimension in dimensions): | ||
| raise ValueError("negative dimensions are not allowed") | ||
|
|
||
| return dimensions | ||
|
|
||
|
|
||
| def _contiguous_strides( | ||
| shape: tuple[int, ...], itemsize: int, order: str | ||
| ) -> tuple[int, ...]: | ||
| if any(dimension == 0 for dimension in shape): | ||
| return (0,) * len(shape) | ||
|
|
||
| strides = [0] * len(shape) | ||
| stride = itemsize | ||
|
|
||
| if order == "C": | ||
| for index in range(len(shape) - 1, -1, -1): | ||
| strides[index] = stride | ||
| stride *= shape[index] | ||
| else: | ||
| for index, dimension in enumerate(shape): | ||
| strides[index] = stride | ||
| stride *= dimension | ||
|
|
||
| return tuple(strides) | ||
|
|
||
|
|
||
| def _resolve_device_and_stream( | ||
| device: Device | None, stream: Stream | None | ||
| ) -> tuple[Device, Stream]: | ||
| if device is None: | ||
| device = stream.device if stream is not None else Device() | ||
|
|
||
| if stream is not None and stream.device.device_id != device.device_id: | ||
| raise ValueError("device and stream must refer to the same device") | ||
|
|
||
| device.set_current() | ||
| return device, device.default_stream if stream is None else stream | ||
|
|
||
|
|
||
| class DeviceArray: | ||
| """A small, Buffer-backed device array for cuda-cccl tests. | ||
|
|
||
| The class intentionally provides only allocation, NumPy transfers, array | ||
| metadata, and the CUDA Array Interface. Array operations and initialization | ||
| belong on the NumPy host arrays used by the tests. | ||
| """ | ||
|
|
||
| def __init__( | ||
| self, | ||
| buffer: Buffer, | ||
| device: Device, | ||
| stream: Stream, | ||
| shape: tuple[int, ...], | ||
| dtype: np.dtype, | ||
| strides: tuple[int, ...], | ||
| order: str, | ||
| ) -> None: | ||
| self._buffer = buffer | ||
| self._device = device | ||
| self._stream = stream | ||
| self._order = order | ||
| self._shape = shape | ||
| self._dtype = dtype | ||
| self._strides = strides | ||
|
|
||
| @classmethod | ||
| def empty( | ||
| cls, | ||
| shape: int | Iterable[int], | ||
| dtype: DTypeLike, | ||
| *, | ||
| order: str = "C", | ||
| device: Device | None = None, | ||
| stream: Stream | None = None, | ||
| ) -> DeviceArray: | ||
| """Allocate an uninitialized device array.""" | ||
| shape = _normalize_shape(shape) | ||
| dtype = np.dtype(dtype) | ||
| order = order.upper() | ||
| if order not in ("C", "F"): | ||
| raise ValueError("order must be either 'C' or 'F'") | ||
| if dtype.itemsize == 0: | ||
| raise ValueError("zero-sized dtypes are not supported") | ||
|
|
||
| device, stream = _resolve_device_and_stream(device, stream) | ||
| buffer = device.allocate(math.prod(shape) * dtype.itemsize, stream=stream) | ||
| result = cls( | ||
| buffer, | ||
| device, | ||
| stream, | ||
| shape, | ||
| dtype, | ||
| _contiguous_strides(shape, dtype.itemsize, order), | ||
| order, | ||
| ) | ||
|
|
||
| # Device allocation is stream ordered. Synchronizing makes an empty array | ||
| # safe to hand to a test that subsequently uses a different stream. | ||
| stream.sync() | ||
| return result | ||
|
|
||
| @classmethod | ||
| def from_numpy( | ||
| cls, | ||
| array: np.ndarray, | ||
| *, | ||
| device: Device | None = None, | ||
| stream: Stream | None = None, | ||
| ) -> DeviceArray: | ||
| """Allocate a device array and initialize it from a NumPy array.""" | ||
| host_array = np.asarray(array) | ||
| if host_array.dtype.itemsize == 0: | ||
| raise ValueError("zero-sized dtypes are not supported") | ||
|
|
||
| if host_array.flags.c_contiguous: | ||
| order = "C" | ||
| elif host_array.flags.f_contiguous: | ||
| order = "F" | ||
| else: | ||
| host_array = np.ascontiguousarray(host_array) | ||
| order = "C" | ||
|
|
||
| device, stream = _resolve_device_and_stream(device, stream) | ||
| buffer = device.allocate(host_array.nbytes, stream=stream) | ||
| result = cls( | ||
| buffer, | ||
| device, | ||
| stream, | ||
| host_array.shape, | ||
| host_array.dtype, | ||
| host_array.strides, | ||
| order, | ||
| ) | ||
| result._copy_from_host_array(host_array, stream) | ||
| stream.sync() | ||
| return result | ||
|
|
||
| @property | ||
| def nbytes(self) -> int: | ||
| return self._buffer.size | ||
|
|
||
| @property | ||
| def dtype(self) -> np.dtype: | ||
| return self._dtype | ||
|
|
||
| def __len__(self) -> int: | ||
| if not self._shape: | ||
| raise TypeError("len() of unsized object") | ||
| return self._shape[0] | ||
|
|
||
| @property | ||
| def __cuda_array_interface__(self) -> dict[str, object]: | ||
| interface: dict[str, object] = { | ||
| "data": (0 if self.nbytes == 0 else int(self._buffer.handle), False), | ||
| "shape": self._shape, | ||
| "strides": None if self._is_c_contiguous() else self._strides, | ||
| "typestr": self._dtype.str, | ||
| "version": 3, | ||
| } | ||
| if self._dtype.fields is not None: | ||
| interface["descr"] = self._dtype.descr | ||
| return interface | ||
|
|
||
| def _is_c_contiguous(self) -> bool: | ||
| return ( | ||
| self._order == "C" | ||
| or self.nbytes == 0 | ||
| or sum(dimension > 1 for dimension in self._shape) <= 1 | ||
| ) | ||
|
|
||
| @staticmethod | ||
| def _host_buffer(array: np.ndarray) -> Buffer: | ||
| # Buffer.from_handle does not own the host memory. `owner` ties the NumPy | ||
| # allocation to this temporary Buffer; the caller also retains the array | ||
| # and synchronizes the copy stream before returning. | ||
| return Buffer.from_handle( | ||
| ptr=int(array.ctypes.data), size=array.nbytes, owner=array | ||
| ) | ||
|
|
||
| def _copy_stream(self, stream: Stream | None) -> Stream: | ||
| if stream is None: | ||
| # The allocation stream is not necessarily the last stream to have | ||
| # used the array. Synchronize the device when that stream is unknown. | ||
| self._device.sync() | ||
| return self._stream | ||
| if stream.device.device_id != self._device.device_id: | ||
| raise ValueError("copy stream must belong to the array's device") | ||
| return stream | ||
|
|
||
| def _copy_from_host_array(self, array: np.ndarray, stream: Stream) -> None: | ||
| self._buffer.copy_from(self._host_buffer(array), stream=stream) | ||
|
|
||
| def copy_from_host( | ||
| self, array: np.ndarray, *, stream: Stream | None = None | ||
| ) -> None: | ||
| """Replace the array's contents from a shape- and dtype-matched NumPy array.""" | ||
| host_array = np.asarray(array) | ||
| if host_array.shape != self._shape: | ||
| raise ValueError( | ||
| f"source shape {host_array.shape} does not match {self._shape}" | ||
| ) | ||
| if host_array.dtype != self._dtype: | ||
| raise TypeError( | ||
| f"source dtype {host_array.dtype} does not match {self._dtype}" | ||
| ) | ||
|
|
||
| if self._order == "F": | ||
| host_array = np.asfortranarray(host_array) | ||
| else: | ||
| host_array = np.ascontiguousarray(host_array) | ||
|
|
||
| self._device.set_current() | ||
| stream = self._copy_stream(stream) | ||
| self._copy_from_host_array(host_array, stream) | ||
| stream.sync() | ||
|
|
||
| def copy_to_host(self, *, stream: Stream | None = None) -> np.ndarray: | ||
| """Return an owning NumPy copy of the array.""" | ||
| self._device.set_current() | ||
| stream = self._copy_stream(stream) | ||
|
|
||
| result = np.empty(self._shape, dtype=self._dtype, order=self._order) | ||
| self._buffer.copy_to(self._host_buffer(result), stream=stream) | ||
| stream.sync() | ||
| return result |
Oops, something went wrong.
Oops, something went wrong.
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
w.r.t my comment here, can we leave the
*with-examplesextra undocumented and instead say here, "install CuPy"?