Allow asarray to work on sequences of usm_array by oleksandr-pavlyk · Pull Request #1139 · IntelPython/dpctl

oleksandr-pavlyk · 2023-03-27T10:57:25Z

In [1]: import dpctl.tensor as dpt, dpctl

In [2]: m = dpt.ones((2,4), dtype='i4')

In [3]: w = dpt.full(4, -1)

In [4]: v = dpt.zeros(4)

In [5]: dpt.asarray([m, [w, v]])
Out[5]:
usm_ndarray([[[ 1.,  1.,  1.,  1.],
              [ 1.,  1.,  1.,  1.]],

             [[-1., -1., -1., -1.],
              [ 0.,  0.,  0.,  0.]]], dtype=float32)

Have you provided a meaningful PR description?
Have you added a test, reproducer or referred to an issue with a reproducer?
Have you tested your changes locally for CPU and GPU devices?
Have you made sure that new changes do not introduce compiler warnings?
If this PR is a work in progress, are you filing the PR as a draft?

github-actions · 2023-03-27T11:15:41Z

View rendered docs @ https://intelpython.github.io/dpctl/pulls/1139/index.html

coveralls · 2023-03-27T11:17:08Z

Coverage: 82.857% (+0.05%) from 82.802% when pulling d294fb6 on fix-gh-1134-asarray into 068f65c on master.

github-actions · 2023-03-27T11:29:22Z

Array API standard conformance tests for dpctl=0.14.3dev0=py310h76be34b_24 ran successfully.
Passed: 46
Failed: 788
Skipped: 280

npolina4 · 2023-03-27T20:28:00Z

import dpctl.tensor as dpt
import dpnp

dpt.asarray([dpnp.ones(3)])

Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/localdisk/work/npolina/ws/dpctl/dpctl/tensor/_ctors.py", line 543, in asarray
    seq_shape, seq_dt, devs = _array_info_sequence(obj)
  File "/localdisk/work/npolina/ws/dpctl/dpctl/tensor/_ctors.py", line 85, in _array_info_sequence
    el_dim, el_dt, el_dev = _array_info_dispatch(el)
  File "/localdisk/work/npolina/ws/dpctl/dpctl/tensor/_ctors.py", line 75, in _array_info_dispatch
    raise ValueError(type(obj))
ValueError: <class 'dpnp.dpnp_array.dpnp_array'>

oleksandr-pavlyk · 2023-03-28T21:31:58Z

import dpctl.tensor as dpt
import dpnp

dpt.asarray([dpnp.ones(3)])

Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/localdisk/work/npolina/ws/dpctl/dpctl/tensor/_ctors.py", line 543, in asarray
    seq_shape, seq_dt, devs = _array_info_sequence(obj)
  File "/localdisk/work/npolina/ws/dpctl/dpctl/tensor/_ctors.py", line 85, in _array_info_sequence
    el_dim, el_dt, el_dev = _array_info_dispatch(el)
  File "/localdisk/work/npolina/ws/dpctl/dpctl/tensor/_ctors.py", line 75, in _array_info_dispatch
    raise ValueError(type(obj))
ValueError: <class 'dpnp.dpnp_array.dpnp_array'>

This should be fixed now:

In [1]: import dpctl.tensor as dpt, dpnp

In [2]: dpt.asarray([dpnp.ones(3)])
Out[2]: usm_ndarray([[1., 1., 1.]], dtype=float32)

In [3]: dpnp.asarray([dpnp.ones(3)])
Out[3]: array([[1., 1., 1.]], dtype=float32)

github-actions · 2023-03-28T21:52:35Z

Array API standard conformance tests for dpctl=0.14.3dev0=py310h76be34b_33 ran successfully.
Passed: 46
Failed: 788
Skipped: 280

npolina4 · 2023-03-28T22:45:36Z

>>> dpnp.array(dpnp.ones(0))
array([])
>>> dpnp.array([dpnp.ones(0)])
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/nfs/site/home/npolina/work/ws/dpnp_bench/dpnp/dpnp/dpnp_iface_arraycreation.py", line 208, in array
    return dpnp_container.asarray(x1,
  File "/nfs/site/home/npolina/work/ws/dpnp_bench/dpnp/dpnp/dpnp_container.py", line 100, in asarray
    array_obj = dpt.asarray(x1_obj,
  File "/nfs/site/home/npolina/work/ws/dpctl/dpctl/tensor/_ctors.py", line 561, in asarray
    seq_shape, seq_dt, devs = _array_info_sequence(obj)
  File "/nfs/site/home/npolina/work/ws/dpctl/dpctl/tensor/_ctors.py", line 86, in _array_info_sequence
    el_dim, el_dt, el_dev = _array_info_dispatch(el)
  File "/nfs/site/home/npolina/work/ws/dpctl/dpctl/tensor/_ctors.py", line 74, in _array_info_dispatch
    usm_ar = _usm_ndarray_from_suai(obj)
  File "/nfs/site/home/npolina/work/ws/dpctl/dpctl/tensor/_ctors.py", line 226, in _usm_ndarray_from_suai
    membuf = dpm.as_usm_memory(obj)
  File "dpctl/memory/_memory.pyx", line 823, in dpctl.memory._memory.as_usm_memory
  File "dpctl/memory/_memory.pyx", line 208, in dpctl.memory._memory._Memory._cinit_other
  File "dpctl/memory/_sycl_usm_array_interface_utils.pxi", line 199, in dpctl.memory._memory._USMBufferData.from_sycl_usm_ary_iface
  File "dpctl/memory/_sycl_usm_array_interface_utils.pxi", line 97, in dpctl.memory._memory._pointers_from_shape_and_stride
ValueError: Array shape elements need to be positive

oleksandr-pavlyk · 2023-03-29T03:35:03Z

Thank for tracing this down, @npolina4 ! A simpler reproducer:

In [7]: dpt.asarray([dpt.ones(0)])
Out[7]: usm_ndarray([], dtype=float32)

In [8]: try:
   ...:    dpt.asarray([dpnp.ones(0)])
   ...: except ValueError as e:
   ...:     print(e)
   ...:
Array shape elements need to be positive

github-actions · 2023-03-29T17:31:38Z

Array API standard conformance tests for dpctl=0.14.3dev0=py310h76be34b_37 ran successfully.
Passed: 46
Failed: 788
Skipped: 280

npolina4 · 2023-03-29T18:18:55Z

In [1]: import dpnp

In [2]: x = dpnp.array([1, 2, 3], device='cpu')

In [3]: dpnp.asarray(x)
Out[3]: array([1, 2, 3])

In [4]: dpnp.asarray([x])

RuntimeError: Unrecogized typenum 17 encountered.

oleksandr-pavlyk · 2023-03-29T18:46:40Z

@npolina4 This must be a bug in dpnp, since it works as expected in dpctl.tensor:


In [1]: import dpctl.tensor as dpt

In [2]: x = dpt.asarray([1, 2, 3], device='cpu')

In [3]: class MockArray:
   ...:     def __init__(self, arr):
   ...:         self.arr_ = arr
   ...:     @property
   ...:     def __sycl_usm_array_interface__(self):
   ...:         return self.arr_.__sycl_usm_array_interface__
   ...:

In [4]: m = MockArray(x)

In [5]: dpt.asarray(m)
Out[5]: usm_ndarray([1, 2, 3])

In [6]: dpt.asarray([m])
Out[6]: usm_ndarray([[1, 2, 3]])

In [7]: dpt.asarray([m, m])
Out[7]:
usm_ndarray([[1, 2, 3],
             [1, 2, 3]])

dpnp array is being treated by dpctl.tensor as an object with __sycl_usm_array_interface__, and the above example works.

The type number 17 corresponds to type np.object_:

In [13]: np.dtype(np.object_).num
Out[13]: 17

Examples: ``` import dpctl.tensor as dpt m = dpt.ones((2,4), dtype='i4') w = dpt.zeros(4) v = dpt.full(4, -1) ar = dpt.asarray([m, [w, v]]) ar2 = dpt.asarray([m, [w, v]], device='cpu') ```

``` In [1]: import dpctl.tensor._tensor_impl as ti, dpctl.tensor as dpt, dpnp In [2]: dpt.asarray([dpnp.ones(3)]) Out[2]: usm_ndarray([[1., 1., 1.]], dtype=float32) In [3]: dpnp.asarray([dpnp.ones(3)]) Out[3]: array([[1., 1., 1.]], dtype=float32) ```

npolina4 · 2023-03-29T19:12:23Z

+                f"does not support {dtype} natively."
+            )
+        dtype = _mapped_dt
+    if order in "KA":


Suggested change

if order in "KA":

if order in ("K", "A"):

Using str appears faster overall:

In [1]: t = "C" In [2]: %timeit t in ("K", "A") 40.3 ns ± 0.768 ns per loop (mean ± std. dev. of 7 runs, 10,000,000 loops each) In [3]: %timeit t in "KA" 26.9 ns ± 0.518 ns per loop (mean ± std. dev. of 7 runs, 10,000,000 loops each) In [4]: %timeit t in ("K", "A") 40.3 ns ± 0.486 ns per loop (mean ± std. dev. of 7 runs, 10,000,000 loops each) In [5]: %timeit t in "KA" 27.6 ns ± 1.17 ns per loop (mean ± std. dev. of 7 runs, 10,000,000 loops each) In [6]: t = "K" In [7]: %timeit t in ("K", "A") 25.7 ns ± 0.39 ns per loop (mean ± std. dev. of 7 runs, 10,000,000 loops each) In [8]: %timeit t in "KA" 26.2 ns ± 0.488 ns per loop (mean ± std. dev. of 7 runs, 10,000,000 loops each)

github-actions · 2023-03-29T19:23:09Z

Array API standard conformance tests for dpctl=0.14.3dev0=py310h76be34b_39 ran successfully.
Passed: 46
Failed: 788
Skipped: 280

github-actions · 2023-03-29T19:25:28Z

Array API standard conformance tests for dpctl=0.14.3dev0=py310h76be34b_39 ran successfully.
Passed: 45
Failed: 789
Skipped: 280

npolina4 · 2023-03-29T20:10:12Z

import dpnp
import dpctl.tensor as dpt

x = dpnp.ones((3,3), device='cpu')
dpt.asarray([x], device='gpu')
RuntimeError: Unrecogized typenum 17 encountered.

…polina4

…polina4

oleksandr-pavlyk · 2023-03-29T20:47:26Z

Thank you @npolina4 for thorough review. I have fixed the underlying issue.

The reported error can be reproduced without involvement of dpnp as follows:

import dpctl.tensor as dpt, numpy as np
u = dpt.empty(5)
n = np.empty(5, dtype="O")
u[...] = n   # raises  RuntimeError: Unrecogized typenum 17 encountered. which trips cascade of other exceptions

github-actions · 2023-03-29T21:08:10Z

Array API standard conformance tests for dpctl=0.14.3dev0=py310h76be34b_41 ran successfully.
Passed: 46
Failed: 788
Skipped: 280

github-actions · 2023-03-29T23:33:54Z

Deleted rendered PR docs from intelpython.github.com/dpctl, latest should be updated shortly. 🤞

github-actions · 2023-03-29T23:58:54Z

Array API standard conformance tests for dpctl=0.14.3dev0=py310h76be34b_46 ran successfully.
Passed: 46
Failed: 788
Skipped: 280

oleksandr-pavlyk requested review from antonwolfy and npolina4 March 27, 2023 18:18

oleksandr-pavlyk force-pushed the fix-gh-1134-asarray branch from b9e3ea7 to a0a341c Compare March 29, 2023 18:54

oleksandr-pavlyk added 7 commits March 29, 2023 13:57

First implementation to resolve gh-1134

7f0f48a

Examples: ``` import dpctl.tensor as dpt m = dpt.ones((2,4), dtype='i4') w = dpt.zeros(4) v = dpt.full(4, -1) ar = dpt.asarray([m, [w, v]]) ar2 = dpt.asarray([m, [w, v]], device='cpu') ```

Added tests for asarray of sequence of usm_arrays

5ef475e

Fixed a type in array dispatcher caught by a test

86cf7d8

Updated docstrings for MemoryUSM* classes

89eed94

Change to allow as_usm_memory to consume SUAI with zero dimensions

9150790

Added a test based on example found by @npolina4

3036a67

oleksandr-pavlyk force-pushed the fix-gh-1134-asarray branch from a0a341c to 3036a67 Compare March 29, 2023 18:57

npolina4 previously approved these changes Mar 29, 2023

View reviewed changes

oleksandr-pavlyk added 2 commits March 29, 2023 15:30

Fixed issue with sequence of arrays from different device found by @n…

2a90321

…polina4

Added tests based on example found by @npolina4

d294fb6

oleksandr-pavlyk dismissed npolina4’s stale review via d294fb6 March 29, 2023 20:37

oleksandr-pavlyk mentioned this pull request Mar 29, 2023

Improved error handling for usm_ndarray.__setitem__ #1146

Merged

6 tasks

oleksandr-pavlyk merged commit 3352c3d into master Mar 29, 2023

oleksandr-pavlyk deleted the fix-gh-1134-asarray branch March 29, 2023 23:33

Conversation

oleksandr-pavlyk commented Mar 27, 2023

Uh oh!

github-actions Bot commented Mar 27, 2023

Uh oh!

coveralls commented Mar 27, 2023 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

github-actions Bot commented Mar 27, 2023

Uh oh!

npolina4 commented Mar 27, 2023

Uh oh!

oleksandr-pavlyk commented Mar 28, 2023

Uh oh!

github-actions Bot commented Mar 28, 2023

Uh oh!

npolina4 commented Mar 28, 2023

Uh oh!

oleksandr-pavlyk commented Mar 29, 2023

Uh oh!

github-actions Bot commented Mar 29, 2023

Uh oh!

npolina4 commented Mar 29, 2023

Uh oh!

oleksandr-pavlyk commented Mar 29, 2023

Uh oh!

npolina4 Mar 29, 2023

Choose a reason for hiding this comment

Uh oh!

oleksandr-pavlyk Mar 29, 2023

Choose a reason for hiding this comment

Uh oh!

github-actions Bot commented Mar 29, 2023

Uh oh!

github-actions Bot commented Mar 29, 2023

Uh oh!

npolina4 commented Mar 29, 2023

Uh oh!

oleksandr-pavlyk commented Mar 29, 2023

Uh oh!

github-actions Bot commented Mar 29, 2023

Uh oh!

github-actions Bot commented Mar 29, 2023

Uh oh!

github-actions Bot commented Mar 29, 2023

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

coveralls commented Mar 27, 2023 •

edited

Loading