compiler: Keep the outer Dimension parallel on a short par-tile by mloubout · Pull Request #3021 · devitocodes/devito

mloubout · 2026-09-08T20:26:49Z

When a multi par-tile entry is shorter than the blocked nest, the innermost Dimensions consume the available block sizes and the outermost one runs into a StopIteration. It was then dropped into compact, which promotes it back to its root Dimension, discarding its BlockDimension.

Tested in pro

codecov · 2026-09-08T20:41:36Z

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 83.77%. Comparing base (33c3b12) to head (18681a8).

Additional details and impacted files

@@           Coverage Diff           @@
##             main    #3021   +/-   ##
=======================================
  Coverage   83.77%   83.77%           
=======================================
  Files         257      257           
  Lines       55007    55028   +21     
  Branches     4708     4710    +2     
=======================================
+ Hits        46083    46102   +19     
- Misses       8109     8111    +2     
  Partials      815      815

Flag	Coverage Δ
pytest-gpu-aomp-amdgpuX	`68.58% <20.00%> (-0.01%)`	⬇️
pytest-gpu-gcc-	`78.40% <33.33%> (-0.01%)`	⬇️
pytest-gpu-icx-	`78.32% <33.33%> (-0.02%)`	⬇️
pytest-gpu-nvc-nvidiaX	`69.11% <95.83%> (+<0.01%)`	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

FabioLuporini

this change is pretty deep and regardless of the tweaks we're going to make it wants some tests across OSS and PRO (tests that are not simply tuner tests)

FabioLuporini

much better now than the previous iteration, but still a couple of questions

FabioLuporini · 2026-09-09T12:24:21Z

-    def __new__(cls, items, default=None, sparse=None, reduce=None):
+    # A short entry is padded to the nest it lands on rather than losing its
+    # outermost BlockDimensions. Only meaningful on a device
+    stretch = False


this shouldn't be a class attribute

It also gets set in separately in the __new__...

that's why it shouldn't be a class attribute

FabioLuporini · 2026-09-09T12:25:45Z

+        """
+        Pad `umt` with its innermost size until it covers `dims`.
+
+        A user-supplied par-tile entry may be shorter than the nest it lands on.


essentially this (x loop on the host) would occur for unoptimized kernels, or kernels that are not in normal form (2+1 loops) -- do I get it right?

however, and regardless, UnboundTuple should keep returning the last entry upon repeated next() calls, see here:

https://github.com/devitocodes/devito/blob/main/devito/tools/data_structures.py#L778

so I'm not sure I understand why any of this would be necessary

essentially this (x loop on the host) would occur for unoptimized kernels, or kernels that are not in normal form (2+1 loops) -- do I get it right?

yes

EdCaunt · 2026-09-09T13:05:01Z

-    def __new__(cls, items, default=None, sparse=None, reduce=None):
+    # A short entry is padded to the nest it lands on rather than losing its
+    # outermost BlockDimensions. Only meaningful on a device
+    stretch = False


It also gets set in separately in the __new__...

EdCaunt · 2026-09-09T13:06:02Z

                                sparse=oo.pop('par-tile-sparse', None),
-                                reduce=oo.pop('par-tile-reduce', None))
+                                reduce=oo.pop('par-tile-reduce', None),
+                                stretch=True)


A more descriptive name (par-tile-stretch or par-tile-extend) might be good?

EdCaunt · 2026-09-09T13:07:39Z

+            return obj
        elif isinstance(items, bool):
            if not default:
                raise ValueError("Expected `default` value, got None")


f"Expected default value, got {default}"?

EdCaunt · 2026-09-09T13:08:47Z

+                stretch=False):
        if not items:
-            return UnboundedMultiTuple()
+            obj = UnboundedMultiTuple()


I know this is pre-existing, but maybe a brief comment explaining what this short-circuit is for wouldn't go amiss

EdCaunt · 2026-09-09T13:35:06Z


        umt.iter()

+        if stretch:


Given stretch relies on self.stretch and self.umt, is it worth having this logic within self.stretched?

I think this comment is now obsolete

FabioLuporini · 2026-09-09T14:30:10Z

+        if stretch:
+            # Walk this nest's own entry, which repeats its innermost size if
+            # it is shorter than the nest, rather than running out of sizes
+            return umt.curitem().reset()


why the reset() at the end ?

When a multi `par-tile` entry is shorter than the blocked nest, the innermost Dimensions consume the available block sizes and the outermost one runs into a StopIteration, raised by the top-level tuple once the entry is exhausted. It was then dropped into `compact`, which promotes it back to its root Dimension, discarding its BlockDimension. On a device that is a serialization: the Dimension ends up outside the blocked nest, `filter_iterations` rejects it as non-parallel, and the kernel is launched over a 2D grid from a host loop iterating the outer Dimension one slice at a time. Walk this nest's own entry instead, which repeats its innermost size for as long as the nest needs it, so the nest stays fully blocked. Whether to do so is decided when the par-tile is built, where the target is already known, rather than by inspecting a cluster back in the pass. It applies to multi par-tiles on a device only: a single user-supplied par-tile defines the block rank on purpose, and on a host a short entry is the documented way to ask for 2.5D blocking.

mloubout added the compiler label Sep 8, 2026

mloubout force-pushed the blocking-keep-outer-dim-parallel branch from b77ed90 to 173ac59 Compare September 9, 2026 01:03

FabioLuporini reviewed Sep 9, 2026

View reviewed changes

Comment thread devito/passes/clusters/blocking.py Outdated

Comment thread devito/passes/clusters/blocking.py Outdated

Comment thread devito/passes/clusters/blocking.py Outdated

Comment thread devito/passes/clusters/blocking.py Outdated

mloubout force-pushed the blocking-keep-outer-dim-parallel branch 2 times, most recently from 0c1f941 to 8962a2a Compare September 9, 2026 12:06

FabioLuporini reviewed Sep 9, 2026

View reviewed changes

mloubout force-pushed the blocking-keep-outer-dim-parallel branch from 8962a2a to 94b8008 Compare September 9, 2026 13:19

EdCaunt reviewed Sep 9, 2026

View reviewed changes

FabioLuporini reviewed Sep 9, 2026

View reviewed changes

mloubout force-pushed the blocking-keep-outer-dim-parallel branch from 94b8008 to 18681a8 Compare September 9, 2026 14:56

Uh oh!

Conversation

mloubout commented Sep 8, 2026

Uh oh!

codecov Bot commented Sep 8, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Codecov Report

Uh oh!

FabioLuporini left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

FabioLuporini 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!

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!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

codecov Bot commented Sep 8, 2026 •

edited

Loading