generate_loop_schedule_v2 by kaushikcfd · Pull Request #350 · inducer/loopy

kaushikcfd · 2021-04-28T22:45:20Z

Implementation for finding loop nest around map in O(N.k), 'N' being the number of inames and 'k' being the max. loop depth.

For comparison, let's consider the kernel in #288: on main this map in computed in 5 minutes and this branch takes 30 0.4 seconds.

inducer

Some thoughts from a quick look here.

inducer · 2021-05-01T23:21:03Z

+tree = get_loop_nest_tree(kernel)

-iname_to_insns = kernel.iname_to_insns()
+loop_nest_around_map = defaultdict(frozenset)


In my view, the loop_nest_around_map should go. The loop nest tree is a strictly more informative data structure, and it is not of size O(n^2). I think we should get rid of it. I looked through the uses, and it shouldn't be hard to replace, and it'll fix another bottleneck in linearization.

In the new version, I didn't touch the old implementation (v1-scheduler) to keep things simpler in this PR. So, for such cases $O(n^2)$ stays.

I can handle it as a separate PR.

inducer · 2021-05-01T23:22:25Z

+ Updates *tree* to make *inames_to_pull_out* a loop nesting level in
+ *loop_nests*
+
+ :returns: a :class:`tuple` ``(outer_loop_nest, inner_loop_nest)``, where


Describe the types of the arguments.

Thanks, done.

inducer · 2021-05-01T23:26:14Z

+ Updates *tree* to make *inames_to_pull_out* a loop nesting level in
+ *loop_nests*


Describe which is nested "inside" and which is "outside".

inducer · 2021-05-26T23:04:45Z

Does #372 supersede this?

kaushikcfd · 2021-05-26T23:14:33Z

@inducer: Not really. #372 includes commits from this branch, so that I could do some tests on my test problems. But I propose we merge them separately. (I've updated the description of #372 to record that)

inducer · 2022-06-09T20:14:08Z

I was just rebasing this. You beat me to it! :)

inducer

Here's a first look. I haven't really gotten very far into the core algorithm, because I'm not sure I understood _pull_out_loop_nest_tree, which seems like it's a core operation.

inducer · 2022-06-09T20:21:27Z

 return wrapper

+
+# {{{ tree data structure


This is big enough to be its own file.

#694 puts the Tree class into its own file (loopy/schedule/tree.py).

inducer · 2022-06-09T20:22:21Z

 return wrapper

+
+# {{{ tree data structure


Given that this has type annotations, mypy should like them (and there should be CI saying as much).

As far as I can see, mypy checks and passes the annotations.

inducer · 2022-06-09T20:23:15Z

+@dataclass(frozen=True)
+class Tree(Generic[T]):
+ """
+ An immutable tree implementation.


Describe the role of the type variable T. Specifically describe that there's one Tree object, but many nodes T. Maybe rename T to NodeT?

c53a9af hopefully clarifies this (and renames T to NodeT).

inducer · 2022-06-09T20:24:57Z

+
+
+@dataclass(frozen=True)
+class Tree(Generic[T]):


Is this mature enough to be in pytools?

inducer · 2022-06-09T20:26:53Z

+ _parent_to_children: "PMap[T, FrozenSet[T]]"
+ _child_to_parent: "PMap[T, Optional[T]]"


In terms of data structure, this is a forest (i.e. it allows multiple trees). It wouldn't take much to allow that fully... YAGNI though.

inducer · 2022-06-09T21:34:19Z

+
+ loop_nests = sorted(loop_nests, key=lambda nest: tree.depth(nest))
+
+ for outer, inner in zip(loop_nests[:-1], loop_nests[1:]):


It seems that all you'd need to pass here is the innermost node in the loop nest tree, and tree.ancestors (at least the proposed changed version) would cheaply give you this without the need to enforce these invariants.

inducer

Thanks! Finally made it all the way through. There's a bit to digest here, but as far as I remember, it's mostly superficial. I didn't find any flaws in the thinking. Nice job!

inducer · 2022-06-10T00:02:08Z

+ if outer_loop == "":
+ continue
+
+ for child in loop_nest_tree.children(outer_loop):


Add a comment that this only works because loop_nest_tree contains a total order of loops. Otherwise it might windup with mismatched enter/leaves.

inducer · 2022-06-10T00:07:06Z

Unsubscribing... @-mention or request review once it's ready for a look or needs attention.

But I'm eager to get this in soon!

…me LoopKernel

inducer · 2024-08-24T19:40:58Z

Most review comments from here are addressed in #864.

inducer · 2024-08-25T03:04:25Z

Included in #864 via #690, closing.

kaushikcfd requested a review from inducer April 28, 2021 23:20

kaushikcfd commented Apr 29, 2021
View reviewed changes

Comment threadsetup.py Outdated

inducer reviewed May 1, 2021
View reviewed changes

kaushikcfd marked this pull request as draft May 4, 2021 15:08

kaushikcfd changed the title ~~Better loop around nest map~~generate_loop_schedule_v2May 4, 2021

inducer mentioned this pull request May 11, 2021
Add loop nest constraints v2 to better loop around nest map #360
Draft

kaushikcfd force-pushed the better_loop_around_nest_map branch 2 times, most recently from 96856f8 to 0203d09Compare May 11, 2021 13:44

kaushikcfd mentioned this pull request May 26, 2021
Move to an AST-like IR before entering codegen #372
Draft
2 tasks

kaushikcfd force-pushed the better_loop_around_nest_map branch from e14a03d to 2d526dcCompare June 8, 2021 18:51

kaushikcfd marked this pull request as ready for review June 10, 2021 06:31

kaushikcfd requested a review from inducer June 11, 2021 22:00

kaushikcfd force-pushed the better_loop_around_nest_map branch from e8b03c8 to 72334e8Compare August 16, 2021 17:23

kaushikcfd force-pushed the better_loop_around_nest_map branch from 6945d16 to d74da04Compare September 28, 2021 15:16

kaushikcfd mentioned this pull request Oct 4, 2021
Implements a Loop Fusion Transformation #493
Merged

kaushikcfd force-pushed the better_loop_around_nest_map branch from d74da04 to 1aef04fCompare October 8, 2021 21:45

kaushikcfd force-pushed the better_loop_around_nest_map branch from 1aef04f to 581e82eCompare October 22, 2021 16:20

kaushikcfd force-pushed the better_loop_around_nest_map branch 5 times, most recently from e20c6ca to 931c6e9Compare December 10, 2021 04:50

kaushikcfd force-pushed the better_loop_around_nest_map branch from 931c6e9 to 36a1a2fCompare January 12, 2022 06:59

kaushikcfd force-pushed the better_loop_around_nest_map branch from f10f24c to c1f8e0dCompare January 28, 2022 15:53

kaushikcfd force-pushed the better_loop_around_nest_map branch from c1f8e0d to 8c42d10Compare February 18, 2022 05:52

kaushikcfd force-pushed the better_loop_around_nest_map branch from 8c42d10 to 15cea20Compare March 10, 2022 15:16

inducer mentioned this pull request Apr 23, 2022
[fusion_actx] High compile times illinois-ceesd/mirgecom#644
Open

kaushikcfd force-pushed the better_loop_around_nest_map branch from 15cea20 to 5b55bcdCompare May 6, 2022 17:30

kaushikcfd force-pushed the better_loop_around_nest_map branch from 5b55bcd to f8dc959Compare May 16, 2022 13:58

kaushikcfd force-pushed the better_loop_around_nest_map branch from f8dc959 to 80425bdCompare June 9, 2022 20:09

inducer reviewed Jun 9, 2022
View reviewed changes

inducer reviewed Jun 10, 2022
View reviewed changes

kaushikcfd mentioned this pull request Jun 17, 2022
Implements loop distribution transformation #629
Draft
4 tasks

kaushikcfd force-pushed the better_loop_around_nest_map branch 2 times, most recently from 65a6b62 to c21c092Compare July 6, 2022 17:49

This was referenced Oct 11, 2022
[Update] generate_loop_schedule_v2 #690
Closed
add Tree implementation #694
Closed

kaushikcfd added 4 commits June 1, 2023 20:23

adds an immutable tree implementation
43dd810

add helpers to figure out loop nestings from a kernel
2c78a9c

adds loopy scheduler v2
0f478dd

changes in docs to account for equivalent generated codes from the sa…
fadb652
…me LoopKernel

inducer force-pushed the better_loop_around_nest_map branch from c21c092 to fadb652Compare June 2, 2023 01:23

inducer mentioned this pull request Feb 16, 2024
Add test_long_kernel#826
Merged

inducer closed this Aug 25, 2024

		Updates tree to make inames_to_pull_out a loop nesting level in
		loop_nests

		_parent_to_children: "PMap[T, FrozenSet[T]]"
		_child_to_parent: "PMap[T, Optional[T]]"


		loop_nests = sorted(loop_nests, key=lambda nest: tree.depth(nest))

		for outer, inner in zip(loop_nests[:-1], loop_nests[1:]):

Conversation

kaushikcfd commented Apr 28, 2021• edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Uh oh!

inducer left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

kaushikcfdJun 10, 2021• edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

inducer commented May 26, 2021

Uh oh!

kaushikcfd commented May 26, 2021

Uh oh!

inducer commented Jun 9, 2022

Uh oh!

inducer left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

inducer left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

inducer commented Jun 10, 2022

Uh oh!

inducer commented Aug 24, 2024

Uh oh!

inducer commented Aug 25, 2024

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

kaushikcfd commented Apr 28, 2021•
edited
Loading

kaushikcfdJun 10, 2021•
edited
Loading