Support negative preceding/following for ROW window functions #14093

mythrocks · 2023-09-12T17:31:28Z

Description

This commit adds support for "offset" ROW windows, where the preceding and following
window bounds are allowed to have negative values. This allows window definitions to
exclude the current row entirely.

Prior to this change, ROW-based windows had to include the current row, causing
preceding and following to support only non-negative values. Additionally, the
inclusion of the current row would count against the min_periods check.

The following is an example of the new "negative" semantics. Consider the input:

auto const row = ints_column{1, 2, 3, 4};

If the window bounds are specified as (preceding=3, following=-1), then the window
for the third row (3) is {1, 2}.
following=-1 indicates a "following" row before the current row.

A negative value for preceding follows the existing convention of including the
current row. This makes it slightly more involved:

preceding=2 indicates one row before the current row.
preceding=1 indicates the current row.
preceding=0 indicates one row past (i.e. after) the current row.
preceding=-1 indicates two rows after the current row.
Et cetera.

min_periods checks continue to be honoured as before, but the requirement for
positive min_periods is dropped. min_periods only need be non-negative.

Checklist

I am familiar with the Contributing Guidelines.
New or existing tests cover these changes.
The documentation is up to date with these changes.

This commit adds support for "offset" ROW windows, where the preceding and following window bounds are allowed to have negative values. This allows window definitions to exclude the current row entirely. Prior to this change, ROW-based windows *had* to include the current row, causing `preceding` and `following` to support only non-negative values. Additionally, the inclusion of the current row would count against the `min_periods` check. The following is an example of the new "negative" semantics. Consider the input: ```c++ auto const row = ints_column{1, 2, 3, 4}; ``` If the window bounds are specified as (preceding=3, following=-1), then the window for the third row (`3`) is `{1, 2}`. `following=-1` indicates a "following" row *before* the current row. A negative value for `preceding` follows the existing convention of including the current row. This makes it slightly more involved: 1. `preceding=2` indicates *one* row before the current row. 2. `preceding=1` indicates the current row. 3. `preceding=0` indicates one row past (i.e. after) the current row. 4. `preceding=-1` indicates two rows after the current row. Et cetera. `min_periods` checks continue to be honoured as before, but the requirement for positive `min_periods` is dropped. `min_periods` only need be non-negative. Signed-off-by: MithunR <[email protected]>

Also, removed prints.

…ndows-wip

copy-pr-bot · 2023-09-12T17:31:30Z

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

mythrocks · 2023-09-12T18:50:31Z

/ok to test

mythrocks · 2023-09-18T18:21:48Z

~~Moved this back to draft. There are a couple of things pending:~~
1. A Spark test had revealed a possible bug at the group boundaries. That test needs to be formalized in the CUDF test.
2. One if-check can be moved out of the kernel for fixed window bounds.
3. Must merge latest from origin/branch-23.10.

~~I'll update these very shortly.~~

Update: This PR is ready for review.

…ndows-wip

mythrocks · 2023-09-18T21:15:35Z

A note to the reviewer:

Most of the bulk seems to be the tests.
In grouped_rolling.cu: The new row_based_preceding_calc and row_based_following_calc are a result of refactoring what used to be a simpler lambda. This refactor allows the if (following < 0) check to be constexpr.
Here, an additional change was made to materialize the preceding_column and following_column, so that the rolling kernels need not be instantiated with complicated iterators.

robertmaynard

We need to update https://github.com/rapidsai/cudf/blob/branch-23.10/cpp/include/cudf/rolling.hpp#L181 to document the relaxation of preceding_window and following_window

cpp/tests/rolling/offset_row_window_test.cpp

…ndows-wip

divyegala · 2023-09-19T21:35:44Z

cpp/src/rolling/grouped_rolling.cu

@@ -94,6 +93,109 @@ std::unique_ptr<column> grouped_rolling_window(table_view const& group_keys,

 namespace detail {

+/// Preceding window calculation functor.
+template <bool preceding_less_than_1>


What is the advantage and disadvantage of having this template?

Ah, right. This is to enable the if constexpr() in operator():

if constexpr (preceding_less_than_1) { // where 1 indicates only the current row. auto group_end = _group_offsets_begin[group_label + 1]; return thrust::maximum{}(_preceding_window, -(group_end - 1 - idx)); } else { ... }

When this used to be a lambda, this used to be:

if (preceding < 1) { ... } else { ... }

That if check would have run once per row in the window. Or worst case N**2 times, over the column. The template was so that this is checked once.

Has there been a tangible benefit to runtime perf with this change?

Sorry if the refactor isn't immediately obvious. This commit shows the refactor: 589f2a5

Ah, I don't have a benchmark in place for this.

I did assume that lifting this if check out would be beneficial.

(Sorry for the delayed response. GitHub reply notifications seem to be delayed. )

Would you rather I postponed this refactor until I have confirmed the advantage in a benchmark?

I'd be happy to do that if it helps get the main feature in.

No, it's totally fine to get this in as-is as long as there's no significant change to compile times.

Thanks for accommodating, @divyegala.

I've measured how much longer grouped_rolling.cu takes to compile, with all the changes in this PR thrown in.
Old: 98s
New: 93s

It's taking less time to compile now. This might have to do with my materializing the preceding/following columns, to simplify the iterators passed to cudf::detail::rolling_window.

As an aside, the compile time for rolling_fixed_window.cu has reduced from 80s to 65s.

So @divyegala's instinct was right, to keep an eye on the compile times. I think I've mitigated it for now.

mythrocks · 2023-09-20T00:06:32Z

Ah, I haven't updated the header yet. Will do shortly.

This now explains the semantics for negative window bounds.

…ndows-wip

mythrocks · 2023-09-20T19:49:13Z

We need to update https://github.com/rapidsai/cudf/blob/branch-23.10/cpp/include/cudf/rolling.hpp#L181 to document the relaxation of preceding_window and following_window

I've updated the docs in rolling.hpp.

cpp/include/cudf/rolling.hpp

…ndows-wip

mythrocks · 2023-09-20T22:13:05Z

/ok to test

mythrocks · 2023-09-21T18:45:09Z

/merge

mythrocks added 9 commits September 5, 2023 14:46

Minor refactor to shorten agg operators.

7e4ae8b

Refactor to support ungrouped case.

a577d2e

Tests for ungrouped cases.

f2a9e1e

Removed unused includes.

e616ecc

Tentative fix for off-by-1 reads of the next group.

6d65898

Materialized preceding/following offsets for fixed windows.

eaff7dd

Also, removed prints.

Formatting.

c3e7f8b

Merge remote-tracking branch 'origin/branch-23.10' into offset-row-wi…

4aca25e

…ndows-wip

mythrocks self-assigned this Sep 12, 2023

mythrocks requested a review from a team as a code owner September 12, 2023 17:31

mythrocks requested review from robertmaynard and divyegala September 12, 2023 17:31

mythrocks marked this pull request as draft September 12, 2023 17:31

github-actions bot added libcudf Affects libcudf (C++/CUDA) code. CMake CMake build issue labels Sep 12, 2023

mythrocks added feature request New feature or request Spark Functionality that helps Spark RAPIDS non-breaking Non-breaking change labels Sep 12, 2023

mythrocks mentioned this pull request Sep 12, 2023

Support negative preceding/following for ROW-based window functions NVIDIA/spark-rapids#9229

Merged

mythrocks marked this pull request as ready for review September 18, 2023 15:51

mythrocks marked this pull request as draft September 18, 2023 18:19

mythrocks added 3 commits September 18, 2023 13:13

Moved if-check for negative windows to constexpr.

589f2a5

Fixed up test.

5f310a2

Merge remote-tracking branch 'origin/branch-23.10' into offset-row-wi…

4d4683c

…ndows-wip

mythrocks marked this pull request as ready for review September 18, 2023 20:58

robertmaynard requested changes Sep 19, 2023

View reviewed changes

cpp/tests/rolling/offset_row_window_test.cpp Show resolved Hide resolved

cpp/tests/rolling/offset_row_window_test.cpp Show resolved Hide resolved

Proper test for clamping window bounds at group boundaries.

ada90c2

mythrocks requested a review from robertmaynard September 19, 2023 19:26

Merge remote-tracking branch 'origin/branch-23.10' into offset-row-wi…

f7c504b

…ndows-wip

divyegala reviewed Sep 19, 2023

View reviewed changes

divyegala approved these changes Sep 20, 2023

View reviewed changes

mythrocks added 2 commits September 20, 2023 12:26

Updated grouped_rolling_window documentation:

da78c8e

This now explains the semantics for negative window bounds.

Merge remote-tracking branch 'origin/branch-23.10' into offset-row-wi…

1821eaf

…ndows-wip

robertmaynard requested changes Sep 20, 2023

View reviewed changes

cpp/include/cudf/rolling.hpp Show resolved Hide resolved

mythrocks added 2 commits September 20, 2023 15:12

Adjusted description for preceding/following.

46de250

Merge remote-tracking branch 'origin/branch-23.10' into offset-row-wi…

b66fd14

…ndows-wip

mythrocks requested a review from robertmaynard September 20, 2023 22:13

robertmaynard approved these changes Sep 21, 2023

View reviewed changes

rapids-bot bot merged commit ec744de into rapidsai:branch-23.10 Sep 21, 2023
54 checks passed

mythrocks mentioned this pull request Jul 19, 2024

Introduce utility function to calculate inputs to rolling_window #16305

Draft

3 tasks

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Support negative preceding/following for ROW window functions #14093

Support negative preceding/following for ROW window functions #14093

mythrocks commented Sep 12, 2023 •

edited

Loading

copy-pr-bot bot commented Sep 12, 2023

mythrocks commented Sep 12, 2023

mythrocks commented Sep 18, 2023 •

edited

Loading

mythrocks commented Sep 18, 2023

robertmaynard left a comment

divyegala Sep 19, 2023

mythrocks Sep 19, 2023

divyegala Sep 19, 2023

mythrocks Sep 19, 2023

mythrocks Sep 20, 2023 •

edited

Loading

mythrocks Sep 20, 2023

divyegala Sep 20, 2023

mythrocks Sep 20, 2023

mythrocks Sep 20, 2023

mythrocks commented Sep 20, 2023

mythrocks commented Sep 20, 2023

mythrocks commented Sep 20, 2023

mythrocks commented Sep 21, 2023

Support negative preceding/following for ROW window functions #14093

Support negative preceding/following for ROW window functions #14093

Conversation

mythrocks commented Sep 12, 2023 • edited Loading

Description

Checklist

copy-pr-bot bot commented Sep 12, 2023

mythrocks commented Sep 12, 2023

mythrocks commented Sep 18, 2023 • edited Loading

mythrocks commented Sep 18, 2023

robertmaynard left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

mythrocks Sep 20, 2023 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

mythrocks commented Sep 20, 2023

mythrocks commented Sep 20, 2023

mythrocks commented Sep 20, 2023

mythrocks commented Sep 21, 2023

mythrocks commented Sep 12, 2023 •

edited

Loading

mythrocks commented Sep 18, 2023 •

edited

Loading

mythrocks Sep 20, 2023 •

edited

Loading