Implementation of sort/argsort and set functions per array API spec #1483

oleksandr-pavlyk · 2023-12-06T18:40:25Z

Implemented data-parallel merge-sort algorithm, which powers array API operations dpct.tensor.sort and dpctl.tensor.argsort. Implemented set functions: dpctl.tensor.unique_values, dpctl.tensor.unique_counts, dpctl.tensor.unique_inverse, and doctl.tensor.unique_all.

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?
Have you checked performance impact of proposed changes?
If this PR is a work in progress, are you opening the PR as a draft?

github-actions · 2023-12-06T19:09:40Z

View rendered docs @ https://intelpython.github.io/dpctl/pulls/1483/index.html

github-actions · 2023-12-06T19:36:08Z

Array API standard conformance tests for dpctl=0.15.1dev2=py310h15de555_20 ran successfully.
Passed: 881
Failed: 50
Skipped: 59

github-actions · 2023-12-07T01:03:10Z

Array API standard conformance tests for dpctl=0.15.1dev2=py310h15de555_23 ran successfully.
Passed: 881
Failed: 50
Skipped: 59

coveralls · 2023-12-07T17:35:59Z

coverage: 89.993% (+0.3%) from 89.693%
when pulling 5469832 on feature/sorting
into f8b2202 on master.

github-actions · 2023-12-07T18:01:27Z

Array API standard conformance tests for dpctl=0.15.1dev2=py310h15de555_27 ran successfully.
Passed: 881
Failed: 50
Skipped: 59

github-actions · 2023-12-08T20:20:48Z

Array API standard conformance tests for dpctl=0.15.1dev2=py310h15de555_32 ran successfully.
Passed: 891
Failed: 41
Skipped: 59

github-actions · 2023-12-08T20:59:44Z

Array API standard conformance tests for dpctl=0.15.1dev2=py310h15de555_33 ran successfully.
Passed: 891
Failed: 41
Skipped: 59

github-actions · 2023-12-09T00:34:02Z

Array API standard conformance tests for dpctl=0.15.1dev2=py310h15de555_35 ran successfully.
Passed: 890
Failed: 42
Skipped: 59

github-actions · 2023-12-09T04:20:49Z

Array API standard conformance tests for dpctl=0.15.1dev2=py310h15de555_40 ran successfully.
Passed: 889
Failed: 43
Skipped: 59

github-actions · 2023-12-09T05:28:15Z

Array API standard conformance tests for dpctl=0.15.1dev2=py310h15de555_41 ran successfully.
Passed: 890
Failed: 42
Skipped: 59

github-actions · 2023-12-09T14:13:04Z

Array API standard conformance tests for dpctl=0.15.1dev2=py310h15de555_43 ran successfully.
Passed: 894
Failed: 38
Skipped: 59

github-actions · 2023-12-09T17:52:26Z

Array API standard conformance tests for dpctl=0.15.1dev2=py310h15de555_44 ran successfully.
Passed: 891
Failed: 41
Skipped: 59

github-actions · 2023-12-12T04:54:00Z

Array API standard conformance tests for dpctl=0.15.1dev2=py310h15de555_49 ran successfully.
Passed: 891
Failed: 41
Skipped: 59

github-actions · 2023-12-15T16:42:13Z

Array API standard conformance tests for dpctl=0.15.1dev2=py310h15de555_67 ran successfully.
Passed: 891
Failed: 41
Skipped: 59

ndgrigorian · 2023-12-18T19:29:06Z

@oleksandr-pavlyk
Tests seem to pass using the asynchronous unique_* functions. Should I go ahead and remove set_functions_sync.py and transition to the asynchronous implementations?

For me, using the async implementation shaves off ~1ms for small arrays

In [2]: x = dpt.ones((10), dtype="c8")

In [3]: %time z = dpt.unique_values(x)
CPU times: user 267 ms, sys: 20.1 ms, total: 287 ms
Wall time: 296 ms

In [4]: %time z = dpt.unique_values(x)
CPU times: user 2.77 ms, sys: 0 ns, total: 2.77 ms
Wall time: 4.52 ms

In [5]: %time z = dpt.unique_values(x)
CPU times: user 1.51 ms, sys: 371 µs, total: 1.88 ms
Wall time: 1.8 ms

In [6]: %time z = dpt.unique_values(x)
CPU times: user 1.6 ms, sys: 0 ns, total: 1.6 ms
Wall time: 1.59 ms

In [7]: %time z = dpt.unique_values(x)
CPU times: user 1.36 ms, sys: 0 ns, total: 1.36 ms
Wall time: 1.29 ms

In [8]: %time z = dpt.unique_values(x)
CPU times: user 2.05 ms, sys: 0 ns, total: 2.05 ms
Wall time: 1.78 ms

compared to synchronized

In [2]: x = dpt.ones((10), dtype="c8")

In [3]: %time z = dpt.unique_values(x)
CPU times: user 264 ms, sys: 14.6 ms, total: 279 ms
Wall time: 287 ms

In [4]: %time z = dpt.unique_values(x)
CPU times: user 1.6 ms, sys: 1.07 ms, total: 2.67 ms
Wall time: 3.43 ms

In [5]: %time z = dpt.unique_values(x)
CPU times: user 1.62 ms, sys: 1.08 ms, total: 2.7 ms
Wall time: 3.78 ms

In [6]: %time z = dpt.unique_values(x)
CPU times: user 2.69 ms, sys: 0 ns, total: 2.69 ms
Wall time: 3.03 ms

In [7]: %time z = dpt.unique_values(x)
CPU times: user 0 ns, sys: 3.01 ms, total: 3.01 ms
Wall time: 3.33 ms

oleksandr-pavlyk · 2023-12-19T02:50:43Z

@ndgrigorian I have switched to asynchronous implementation per your feedback, but kept the synchronous one around just in case.

github-actions · 2023-12-19T03:46:19Z

Array API standard conformance tests for dpctl=0.15.1dev2=py310h15de555_74 ran successfully.
Passed: 890
Failed: 42
Skipped: 59

dpctl/tensor/_set_functions_async.py

github-actions · 2023-12-19T21:17:58Z

Array API standard conformance tests for dpctl=0.15.1dev2=py310h15de555_77 ran successfully.
Passed: 895
Failed: 37
Skipped: 59

github-actions · 2023-12-20T22:17:14Z

Array API standard conformance tests for dpctl=0.15.1dev2=py310h15de555_78 ran successfully.
Passed: 892
Failed: 40
Skipped: 59

github-actions · 2023-12-20T23:32:49Z

Array API standard conformance tests for dpctl=0.15.1dev2=py310h15de555_79 ran successfully.
Passed: 896
Failed: 36
Skipped: 59

github-actions · 2023-12-21T21:08:27Z

Array API standard conformance tests for dpctl=0.15.1dev2=py310h15de555_82 ran successfully.
Passed: 895
Failed: 37
Skipped: 59

github-actions · 2023-12-22T00:48:31Z

Array API standard conformance tests for dpctl=0.15.1dev2=py310h15de555_83 ran successfully.
Passed: 894
Failed: 38
Skipped: 59

This is `std::vector<sycl::event>` passed by reference to collect events associated with host_task submissions. The synchronizing call `mask_positions` is releasing GIL before wait on these events is called. It is either this, or accumulated host tasks must be returned to the user.

Changed hyperparameter choices to be different for CPU and GPU, resulting in 20% performance gain on GPU. The non-recursive implementation allows to avoid repeated USM allocations, resulting in performance gains for large arrays. Furthermore, corrected base step kernel to accumulate in outputT rather than in size_t, which additionally realizes savings when int32 is used as accumulator type. Using example from gh-1249, previously, on my Iris Xe laptop: ``` In [1]: import dpctl.tensor as dpt ...: ag = dpt.ones((8192, 8192), device='gpu', dtype='f4') ...: bg = dpt.ones((8192, 8192), device='gpu', dtype=bool) In [2]: cg = ag[bg] In [3]: dpt.all(cg == dpt.reshape(ag, -1)) Out[3]: usm_ndarray(True) In [4]: %timeit -n 10 -r 3 cg = ag[bg] 212 ms ± 56 ms per loop (mean ± std. dev. of 3 runs, 10 loops each) ``` while with this change: ``` In [4]: %timeit -n 10 -r 3 cg = ag[bg] 178 ms ± 24.2 ms per loop (mean ± std. dev. of 3 runs, 10 loops each) ```

As the array will be copied if it isn't C-contiguous regardless, this change does not impact performance and permits calls to set functions with strided, ND data

Set functions were making copies and then sorting the original data, resulting in incorrect results.

`unique_all` would not return a `UniqueAllResult` tuple when early exiting for an array of all unique elements Fixed problems through `unique` functions where the cumulative sum was being allocated on a separate (default) queue, causing inputs with a non-default queue to fail `unique` functions now behave correctly for 0-size array inputs

Were previously returning a 1D array of indices rather than an array with the same shape as input `x`

As mandated by the array API spec

Changes to sync implementation of set functions so that test suite passes with sync implementation as it does for async implementation.

github-actions · 2024-01-08T22:31:11Z

Array API standard conformance tests for dpctl=0.15.1dev2=py310h15de555_85 ran successfully.
Passed: 893
Failed: 16
Skipped: 82

github-actions · 2024-01-08T23:46:48Z

Array API standard conformance tests for dpctl=0.15.1dev2=py310h15de555_87 ran successfully.
Passed: 893
Failed: 16
Skipped: 82

…x FP types We use extended comparison operators compatible with NumPy's behavior: https://numpy.org/devdocs/reference/generated/numpy.sort.html Specifically, we use [R, nan] block ordering for reals, and [(R, R), (R, nan), (nan, R), (nan, nan)] for complexes.

github-actions · 2024-01-10T00:34:29Z

Array API standard conformance tests for dpctl=0.15.1dev2=py310h15de555_89 ran successfully.
Passed: 895
Failed: 14
Skipped: 82

ndgrigorian

LGTM, thank you @oleksandr-pavlyk , very close to full array API coverage.

vtavana

Thank you, works fine for zero dimensional arrays and arrays with nan.

oleksandr-pavlyk force-pushed the feature/sorting branch from 3be1f65 to 18e0db3 Compare December 9, 2023 03:23

oleksandr-pavlyk marked this pull request as ready for review December 9, 2023 13:12

oleksandr-pavlyk force-pushed the feature/sorting branch from 90578b6 to 1849fff Compare December 9, 2023 13:17

oleksandr-pavlyk requested a review from ndgrigorian December 9, 2023 18:46

oleksandr-pavlyk force-pushed the feature/sorting branch from 0660a78 to 7003c90 Compare December 12, 2023 03:50

oleksandr-pavlyk force-pushed the feature/sorting branch from 82da567 to 26d97ba Compare December 19, 2023 02:49

oleksandr-pavlyk commented Dec 19, 2023

View reviewed changes

dpctl/tensor/_set_functions_async.py Outdated Show resolved Hide resolved

ndgrigorian force-pushed the feature/sorting branch from 09620d7 to 60c6ad6 Compare December 20, 2023 22:45

oleksandr-pavlyk and others added 18 commits January 8, 2024 13:30

Use sequential sorting for small arrays

fad8470

Do not use shorthand queue::copy

fa3d8fe

Do not use queue::fill convenience shortcut functions

abfc280

Use async implementation of set functions

b149096

Permit copying when flattening for unique funcs

eeb30f6

As the array will be copied if it isn't C-contiguous regardless, this change does not impact performance and permits calls to set functions with strided, ND data

Fix typos in set functions

0035558

Set functions were making copies and then sorting the original data, resulting in incorrect results.

Sorting, unique functions properly handle 0D arrays

2fe8620

unique_all and unique_inverse inverse indices shape fixed

8a6581a

Were previously returning a 1D array of indices rather than an array with the same shape as input `x`

Set stable to True for sorting functions by default

82be69a

As mandated by the array API spec

Adds tests for 0D array cases for sorting

81610d6

Adds tests for unique functions on strided inputs

b0cc135

Tests for strided 1D data input for unique functions

8a6282f

Changes to _set_functions_sync.py for test suite to pass

dd71e4e

Changes to sync implementation of set functions so that test suite passes with sync implementation as it does for async implementation.

Removed sync implementation of set functions

b1b0af1

Renamed _set_functions_async -> _set_functions

08e5dac

oleksandr-pavlyk force-pushed the feature/sorting branch from de20631 to 08e5dac Compare January 8, 2024 22:52

oleksandr-pavlyk added 2 commits January 9, 2024 12:54

Adding tests for sorting of FP arrays with NaNs

5469832

ndgrigorian approved these changes Jan 13, 2024

View reviewed changes

vtavana approved these changes Jan 13, 2024

View reviewed changes

oleksandr-pavlyk merged commit fd84252 into master Jan 13, 2024
45 of 49 checks passed

oleksandr-pavlyk deleted the feature/sorting branch January 13, 2024 16:18

This was referenced Jan 24, 2024

Fix dpctl.tensor.put race condition #1384

Closed

Populated changelog of new features since 0.15.0 release #1510

Merged

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Implementation of sort/argsort and set functions per array API spec #1483

Implementation of sort/argsort and set functions per array API spec #1483

oleksandr-pavlyk commented Dec 6, 2023 •

edited

Loading

github-actions bot commented Dec 6, 2023

github-actions bot commented Dec 6, 2023

github-actions bot commented Dec 7, 2023

coveralls commented Dec 7, 2023 •

edited

Loading

github-actions bot commented Dec 7, 2023

github-actions bot commented Dec 8, 2023

github-actions bot commented Dec 8, 2023

github-actions bot commented Dec 9, 2023

github-actions bot commented Dec 9, 2023

github-actions bot commented Dec 9, 2023

github-actions bot commented Dec 9, 2023

github-actions bot commented Dec 9, 2023

github-actions bot commented Dec 12, 2023

github-actions bot commented Dec 15, 2023

ndgrigorian commented Dec 18, 2023 •

edited

Loading

oleksandr-pavlyk commented Dec 19, 2023

github-actions bot commented Dec 19, 2023

github-actions bot commented Dec 19, 2023

github-actions bot commented Dec 20, 2023

github-actions bot commented Dec 20, 2023

github-actions bot commented Dec 21, 2023

github-actions bot commented Dec 22, 2023

github-actions bot commented Jan 8, 2024

github-actions bot commented Jan 8, 2024

github-actions bot commented Jan 10, 2024

ndgrigorian left a comment •

edited

Loading

vtavana left a comment

Implementation of sort/argsort and set functions per array API spec #1483

Implementation of sort/argsort and set functions per array API spec #1483

Conversation

oleksandr-pavlyk commented Dec 6, 2023 • edited Loading

github-actions bot commented Dec 6, 2023

github-actions bot commented Dec 6, 2023

github-actions bot commented Dec 7, 2023

coveralls commented Dec 7, 2023 • edited Loading

github-actions bot commented Dec 7, 2023

github-actions bot commented Dec 8, 2023

github-actions bot commented Dec 8, 2023

github-actions bot commented Dec 9, 2023

github-actions bot commented Dec 9, 2023

github-actions bot commented Dec 9, 2023

github-actions bot commented Dec 9, 2023

github-actions bot commented Dec 9, 2023

github-actions bot commented Dec 12, 2023

github-actions bot commented Dec 15, 2023

ndgrigorian commented Dec 18, 2023 • edited Loading

oleksandr-pavlyk commented Dec 19, 2023

github-actions bot commented Dec 19, 2023

github-actions bot commented Dec 19, 2023

github-actions bot commented Dec 20, 2023

github-actions bot commented Dec 20, 2023

github-actions bot commented Dec 21, 2023

github-actions bot commented Dec 22, 2023

github-actions bot commented Jan 8, 2024

github-actions bot commented Jan 8, 2024

github-actions bot commented Jan 10, 2024

ndgrigorian left a comment • edited Loading

Choose a reason for hiding this comment

vtavana left a comment

Choose a reason for hiding this comment

oleksandr-pavlyk commented Dec 6, 2023 •

edited

Loading

coveralls commented Dec 7, 2023 •

edited

Loading

ndgrigorian commented Dec 18, 2023 •

edited

Loading

ndgrigorian left a comment •

edited

Loading