Implement `__cuda_array_interface__` by seibert · Pull Request #1144 · cupy/cupy

seibert · 2018-04-13T21:45:07Z

This implements a special CUDA array interface attribute, analogous to the __array_interface__ attribute on NumPy arrays. This attribute is specifically for advertising CUDA device-compatible pointers (actual device allocations, or managed memory) and associated metadata.

This allows CuPy arrays to be used with Numba-compiled CUDA functions when combined with numba/numba#2860.

seibert · 2018-05-15T18:40:21Z

(Rebasing this PR on master to see if that fixes the flake8 failures in unrelated code.)

Crissman · 2018-05-22T01:19:25Z

@seibert Thanks for the PR! If this is ready for review, please remove the [WIP] label.

Thanks!

seibert · 2018-05-22T14:20:14Z

Oh right! I forgot to change the title. Two questions I do have, now that I reread it:

This PR assumes that there is no such thing as a read-only CuPy array. Is that true?
I'm not sure where I should add a unit test for this attribute. Can you point me in the right direction?

BTW, the dev docs for Numba describe the meaning of the attributes:

http://numba.pydata.org/numba-doc/dev/cuda/cuda_array_interface.html

seibert · 2018-05-29T18:25:27Z

@Crissman: Thoughts on my questions above?

Crissman · 2018-05-31T05:06:48Z

There is no such thing as an immutable CuPy array at this time. Can't promise that won't change in the future, but we'd review the code if we added it.

Please add the unit test here: tests/cupy_tests/core_tests/test_ndarray.py

When you're ready for us to review, please remove the [WIP].

Thanks!

seibert · 2018-07-23T19:29:29Z

Sorry for the delay here. I've rewritten the patch and added tests. This is ready for review.

okuta · 2018-07-25T02:40:04Z

I will review this.

okuta · 2018-07-27T00:56:41Z

        readonly object dtype
        readonly memory.MemoryPointer data
        readonly ndarray base
+        readonly object _cuda_array_descr


Does this feature need cache? Creating dict is low cost. And, dict is a mutable object. Perhaps, user edit this value though return value of __cuda_array_interface__.

It saves to 0.3 - 0.5 microseconds per array per kernel launch, which is probably not important unless you are passing many arrays to a function. I've reverted back to the safer, uncached approach.

okuta · 2018-07-27T01:00:22Z

+                         set(['shape', 'typestr', 'data', 'version', 'descr']))
+        self.assertEqual(iface['shape'], (2,3))
+        self.assertEqual(iface['typestr'], '<f8')
+        self.assertIsInstance(iface['data'], tuple)


Please add self.assertEqual(iface['data'], arr.data.ptr`)

okuta · 2018-07-27T01:02:59Z

And please fix style.

--- original/./tests/cupy_tests/core_tests/test_ndarray.py
+++ fixed/./tests/cupy_tests/core_tests/test_ndarray.py
@@ -135,11 +135,11 @@
 class TestNdarrayCudaInterface(unittest.TestCase):
 
     def test_cuda_array_interface(self):
-        arr = cupy.zeros(shape=(2,3), dtype=cupy.float64)
+        arr = cupy.zeros(shape=(2, 3), dtype=cupy.float64)
         iface = arr.__cuda_array_interface__
         self.assertEqual(set(iface.keys()),
                          set(['shape', 'typestr', 'data', 'version', 'descr']))
-        self.assertEqual(iface['shape'], (2,3))
+        self.assertEqual(iface['shape'], (2, 3))
         self.assertEqual(iface['typestr'], '<f8')
         self.assertIsInstance(iface['data'], tuple)
         self.assertEqual(len(iface['data']), 2)
@@ -148,13 +148,13 @@
         self.assertEqual(iface['descr'], [('', '<f8')])
 
     def test_cuda_array_interface_view(self):
-        arr = cupy.zeros(shape=(10,20), dtype=cupy.float64)
-        view = arr[::2,::5]
+        arr = cupy.zeros(shape=(10, 20), dtype=cupy.float64)
+        view = arr[::2, ::5]
         iface = view.__cuda_array_interface__
         self.assertEqual(set(iface.keys()),
                          set(['shape', 'typestr', 'data', 'version',
                               'strides', 'descr']))
-        self.assertEqual(iface['shape'], (5,4))
+        self.assertEqual(iface['shape'], (5, 4))
         self.assertEqual(iface['typestr'], '<f8')
         self.assertIsInstance(iface['data'], tuple)
         self.assertEqual(len(iface['data']), 2)```

seibert · 2018-08-02T14:03:07Z

Thanks for the review! I have updated the PR with your suggestions.

okuta · 2018-08-05T10:30:52Z

+
+    def test_cuda_array_interface_view(self):
+        arr = cupy.zeros(shape=(10, 20), dtype=cupy.float64)
+        view = arr[::2,::5]


Please add a white space.

- view = arr[::2,::5] + view = arr[::2, ::5]

okuta · 2018-08-05T10:31:12Z

jenkins, test this please.

chainer-ci · 2018-08-05T11:12:04Z

Jenkins CI test (for commit 53bd090, target branch master) succeeded!

okuta · 2018-08-05T11:13:06Z

LGTM except a comment.

seibert · 2018-08-10T20:10:23Z

whitespace change has been made

okuta · 2018-08-14T04:11:56Z

jenkins, test this please.

chainer-ci · 2018-08-14T05:41:06Z

Jenkins CI test (for commit 7aebd05, target branch master) failed with status FAILURE.
(For contributors, please wait until the reviewer confirms the details of the error.)

okuta · 2018-08-14T08:14:19Z

jenkins, test this please.

chainer-ci · 2018-08-14T08:48:24Z

Jenkins CI test (for commit 7aebd05, target branch master) succeeded!

okuta · 2018-08-14T08:58:34Z

LGTM!

seibert mentioned this pull request Apr 13, 2018

__cuda_array_interface__ numba/numba#2860

Merged

seibert force-pushed the array_interface branch from f3b8234 to 37ecde8 Compare May 15, 2018 18:39

__cuda_array_interface__

20a1c8e

seibert force-pushed the array_interface branch from a7c65e0 to 20a1c8e Compare July 23, 2018 19:28

seibert changed the title ~~[WIP] __cuda_array_interface__ implementation~~ __cuda_array_interface__ implementation Jul 23, 2018

okuta added the cat:feature New features/APIs label Jul 25, 2018

okuta self-assigned this Jul 25, 2018

okuta requested changes Jul 27, 2018

View reviewed changes

update based on review suggestions

53bd090

okuta added this to the v5.0.0b4 milestone Aug 5, 2018

okuta approved these changes Aug 5, 2018

View reviewed changes

okuta requested changes Aug 5, 2018

View reviewed changes

okuta added the st:awaiting-author Awaiting response from author label Aug 5, 2018

fix whitespace

7aebd05

okuta approved these changes Aug 14, 2018

View reviewed changes

okuta merged commit cda197b into cupy:master Aug 14, 2018

kmaehashi changed the title ~~__cuda_array_interface__ implementation~~ Implement __cuda_array_interface__ Aug 22, 2018

seibert mentioned this pull request Aug 27, 2018

Argument has invalid type for CuPy array numba/numba#2786

Closed

stuartarchibald mentioned this pull request Oct 19, 2018

Compatibility with CuPy numba/numba#3422

Closed

mrocklin mentioned this pull request Dec 14, 2018

Build DataFrames from CUDA array libraries rapidsai/cudf#529

Closed

mrocklin mentioned this pull request May 26, 2019

Support the __cuda_array_interface__ protocol tensorflow/tensorflow#29039

Closed

casperdcl mentioned this pull request Feb 1, 2021

add __cuda_array_interface__ AMYPAD/CuVec#4

Closed

asi1024 removed the st:awaiting-author Awaiting response from author label Jan 19, 2022

Uh oh!

Conversation

seibert commented Apr 13, 2018 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

seibert commented May 15, 2018

Uh oh!

Crissman commented May 22, 2018

Uh oh!

seibert commented May 22, 2018

Uh oh!

seibert commented May 29, 2018

Uh oh!

Crissman commented May 31, 2018

Uh oh!

seibert commented Jul 23, 2018

Uh oh!

okuta commented Jul 25, 2018

Uh oh!

okuta Jul 27, 2018

Choose a reason for hiding this comment

Uh oh!

seibert Aug 2, 2018

Choose a reason for hiding this comment

Uh oh!

okuta Jul 27, 2018

Choose a reason for hiding this comment

Uh oh!

seibert Aug 2, 2018

Choose a reason for hiding this comment

Uh oh!

okuta commented Jul 27, 2018

Uh oh!

seibert commented Aug 2, 2018

Uh oh!

okuta Aug 5, 2018

Choose a reason for hiding this comment

Uh oh!

okuta commented Aug 5, 2018

Uh oh!

chainer-ci commented Aug 5, 2018

Uh oh!

okuta commented Aug 5, 2018

Uh oh!

seibert commented Aug 10, 2018

Uh oh!

okuta commented Aug 14, 2018

Uh oh!

chainer-ci commented Aug 14, 2018

Uh oh!

okuta commented Aug 14, 2018

Uh oh!

chainer-ci commented Aug 14, 2018

Uh oh!

okuta commented Aug 14, 2018

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants

seibert commented Apr 13, 2018 •

edited

Loading