core: Add float-semantics to support low-precision by n-io · Pull Request #6254 · xdslproject/xdsl

n-io · 2026-07-14T17:43:54Z

Extends support for reduced-precision floating point values by adding support for reduced-precision literals, and a central mechanism to implement a variety of different flavours. The types itself were added in #6250. This PR adds support for literals, by introducing the FloatSemantics class based on MLIR's fltSemantics concept (alongside fltNonfiniteBehavior and fltNanEncoding).

Stack:

codecov · 2026-07-14T17:53:16Z

Codecov Report

❌ Patch coverage is 98.60140% with 2 lines in your changes missing coverage. Please review.
✅ Project coverage is 86.97%. Comparing base (999233d) to head (5c22de5).
⚠️ Report is 1 commits behind head on main.

Files with missing lines	Patch %	Lines
xdsl/dialects/builtin.py	98.60%	0 Missing and 2 partials ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #6254      +/-   ##
==========================================
+ Coverage   86.94%   86.97%   +0.02%     
==========================================
  Files         431      431              
  Lines       64924    65064     +140     
  Branches     7419     7451      +32     
==========================================
+ Hits        56450    56587     +137     
- Misses       6903     6904       +1     
- Partials     1571     1573       +2

☔ 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.

superlopuh

This is quite a lot of code to review at a time, let's split this up? I feel like the first thing to add is the overall structure, and none of the methods on _ReducedPrecisionFloatType (which I would make public). Then the next PRs can one-by-one replace the existing methods on the subclasses, like bitwidth could be one PR, and then format could be another PR. Does this sound like a reasonable plan?

n-io · 2026-07-15T09:24:47Z

Definitely happy to split off the FloatSemantics + enums, a basically empty ReducedPrecisionFloatType (now public), and the rewiring into the classes. It helps reduce the diff, but it's mostly optical and its harder to split up the logic. Separate PR raised, this is WIP PR for now.

This PR introduces the FloatSemantics class based on MLIR's `fltSemantics` concept (alongside `fltNonfiniteBehavior` and `fltNanEncoding`). This is used to calculate the `bitwidth` of reduced-precision floats. WIP: * #6254

# Conflicts: # tests/dialects/test_builtin.py # xdsl/dialects/builtin.py

Adding helpers to `FloatSemantics`. WIP: * #6254

superlopuh · 2026-07-17T08:53:08Z

-            float_type.format
+    # Reduced-precision floats pack via the reduced-float codec, not a struct
+    # format string, so (like bf16) they have no `format` attribute.
+    assert not hasattr(tf32, "format")


This feels off to me, what are we really testing here?

My gut tells me to just delete this line, I'm not sure what benefit there is to it.

It's mirroring the same (pre-existing) check for bf16 just above. Either way is fine.

Moved these next to each other, removed comment. Share the general gut feeling though.

superlopuh

TYVM

superlopuh · 2026-07-17T08:58:54Z

If you have the motivation it feels to me like it would be great to support the larger floats also

Samielakkad

Nice addition overall. One edge case I’d want covered before this lands: iter_unpack currently walks the buffer in size chunks and decodes the last slice even if it is shorter than size.

For the reduced-float types with odd byte widths, that means a truncated buffer can silently decode as if the missing high bytes were zero. That is a little different from struct.iter_unpack, which requires an exact multiple of the element size.

It may be worth either checking len(buffer) % self.size == 0 in iter_unpack / unpack, or adding an explicit test that this silent partial decode is the intended behavior. I would lean toward rejecting the partial chunk so malformed dense data does not quietly round-trip to a different value.

n-io · 2026-07-17T10:08:21Z

Thanks @Samielakkad, I think bf16 rejects too so it makes sense to replicate that.

superlopuh · 2026-07-17T10:13:36Z

+            raise ValueError(
+                f"buffer of {len(mv)} bytes is not a multiple of the {size}-byte "
+                f"element size of {self.name}"
+            )


No, this is also not the right check, as it should be OK to unpack lazily just the prefix. The error should be raised only if the last element cannot be read in full

I'm also not sure about the comment, the old code was reading potentially past the end of the buffer, not silently truncating? I would have expected an error to be raised there from an out of bounds access for the last element.

Reading the struct documentation, it says that the bytes buffer should be a multiple of the struct size but not what error is raised if it isn't...

I think the OOB error didn't happen because of the slicing.

superlopuh · 2026-07-17T10:41:33Z

+                    f"buffer ends with a partial {self.name} value: "
+                    f"{len(chunk)} of {size} bytes"
+                )
+            yield self.decode_bits(int.from_bytes(chunk, "little"))


Thanks for updating this, but I still think it would be good to be consistent across implementations. Could you please add f64 to your test for this, and check that the error type is the same for both the struct and this manual implementation, and that the errors are consistent on whether the whole buffer needs to have size multiple of this element's size or just the last element?

If we're going for consistency, we should throw a struct.error. However, to do so I'd like to inherit from StructPackableType. Although we're not using any of its code, we do provide the same methods, so it's somewhat cleaner. I've also parameterized

@pytest.mark.parametrize("type_", [f64, bf16, tf32]) def test_float_rejects_truncated_buffer(type_: AnyFloat):

to demonstrate uniform rejection errors across these types.

Let's iterate on the PR I just opened (#6273) to decide what errors we want, feels like a good order of solving these issues to me.

superlopuh · 2026-07-17T15:00:40Z

+    assert type_.unpack(packed, 2) == (1.5, 2.0)
+    truncated = packed[:-1]  # ends part-way through the final element
+    with pytest.raises(struct.error):
+        list(type_.iter_unpack(truncated))


This seems important to me, does it raise when you fetch the first element or all the elements

Suggested change

list(type_.iter_unpack(truncated))

next(type_.iter_unpack(truncated))

Yes, with a small fix for bf16 to match others.

In the PR I just opened it actually raises before you fetch the first element

Samielakkad

Found one edge case in f8E8M0FNU: negative inputs lose their sign before rounding, so -4.0 is encoded as +4.0 (0x81). LLVM rejects signed values for this format, while ml_dtypes maps them to 0xff/NaN.

We should pick one of those behaviors rather than silently taking the absolute value. Tests for -0.0 and one negative finite value would catch it.

superlopuh

LGTM!

n-io · 2026-07-21T08:54:22Z

Thanks @superlopuh, and good spot @Samielakkad! It's what one gets for sweeping the hexadecimal spectrum (:

n-io added 2 commits July 14, 2026 19:31

core: add FloatSemantics to support low-precision

90601f3

Add filecheck

526c1ce

n-io self-assigned this Jul 14, 2026

n-io added the core xDSL core (ir, textual format, ...) label Jul 14, 2026

Move code around to reduce diff

4e26ebe

n-io requested a review from superlopuh July 14, 2026 18:02

Coverage

357ad82

superlopuh reviewed Jul 15, 2026

View reviewed changes

Comment thread tests/dialects/test_builtin.py Outdated

superlopuh reviewed Jul 15, 2026

View reviewed changes

Iterating

8885b6d

n-io changed the title ~~core: Add float-semantics to support low-precision~~ [WIP] core: Add float-semantics to support low-precision Jul 15, 2026

n-io mentioned this pull request Jul 15, 2026

core: Derive low-precision fp bitwidth from semantics #6256

Merged

superlopuh marked this pull request as draft July 15, 2026 10:49

n-io added 3 commits July 15, 2026 18:03

Merge branch 'main' into nicolai/float-semantics

f927a40

# Conflicts: # tests/dialects/test_builtin.py # xdsl/dialects/builtin.py

comment

138c864

comment

b11425f

n-io marked this pull request as ready for review July 15, 2026 16:48

n-io requested a review from superlopuh July 15, 2026 16:49

superlopuh reviewed Jul 15, 2026

View reviewed changes

Comment thread xdsl/dialects/builtin.py

n-io mentioned this pull request Jul 16, 2026

core: add FloatSemantics helpers #6263

Merged

n-io added a commit that referenced this pull request Jul 16, 2026

core: add FloatSemantics helpers (#6263)

d1b98b5

Adding helpers to `FloatSemantics`. WIP: * #6254

merging main

9c16dfa

n-io requested a review from superlopuh July 16, 2026 09:42

n-io changed the title ~~[WIP] core: Add float-semantics to support low-precision~~ core: Add float-semantics to support low-precision Jul 16, 2026

superlopuh reviewed Jul 16, 2026

View reviewed changes

Comment thread xdsl/dialects/builtin.py Outdated

superlopuh reviewed Jul 16, 2026

View reviewed changes

Comment thread xdsl/dialects/builtin.py Outdated

superlopuh reviewed Jul 16, 2026

View reviewed changes

Comment thread xdsl/dialects/builtin.py Outdated

dense

6d1456b

n-io requested a review from superlopuh July 16, 2026 13:13

Iterating

b86ec37

superlopuh reviewed Jul 17, 2026

View reviewed changes

Comment thread xdsl/dialects/builtin.py Outdated

superlopuh approved these changes Jul 17, 2026

View reviewed changes

Samielakkad reviewed Jul 17, 2026

View reviewed changes

n-io added 2 commits July 17, 2026 11:40

cleanup

e871964

reduced float rejects nonmatching buf size

3224963

n-io requested a review from Samielakkad July 17, 2026 10:08

superlopuh requested changes Jul 17, 2026

View reviewed changes

iterating

90a6952

n-io requested a review from superlopuh July 17, 2026 10:35

superlopuh reviewed Jul 17, 2026

View reviewed changes

n-io added 2 commits July 17, 2026 14:09

test for uniformity

ebd606a

Inherit from StructPackableType and override all methods

33e61df

n-io requested a review from superlopuh July 17, 2026 12:44

superlopuh reviewed Jul 17, 2026

View reviewed changes

Changing to next

a0c4a94

n-io requested a review from superlopuh July 17, 2026 15:14

Samielakkad reviewed Jul 19, 2026

View reviewed changes

n-io added 3 commits July 21, 2026 10:28

reject negative values on unsigned low-precision FP types

29ecdab

Merge branch 'main' into nicolai/float-semantics

8da2a8d

Cleanup

5c22de5

superlopuh approved these changes Jul 21, 2026

View reviewed changes

n-io merged commit 701c481 into main Jul 21, 2026
24 checks passed

n-io deleted the nicolai/float-semantics branch July 21, 2026 08:55

	list(type_.iter_unpack(truncated))
	next(type_.iter_unpack(truncated))

Uh oh!

Conversation

n-io commented Jul 14, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

codecov Bot commented Jul 14, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Codecov Report

Uh oh!

Uh oh!

superlopuh left a comment

Choose a reason for hiding this comment

Uh oh!

n-io commented Jul 15, 2026

Uh oh!

Uh oh!

Uh oh!

Uh oh!

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!

superlopuh left a comment

Choose a reason for hiding this comment

Uh oh!

superlopuh commented Jul 17, 2026

Uh oh!

Samielakkad left a comment

Choose a reason for hiding this comment

Uh oh!

n-io commented Jul 17, 2026

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!

n-io Jul 17, 2026 • 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!

Samielakkad left a comment • 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!

superlopuh left a comment

Choose a reason for hiding this comment

Uh oh!

n-io commented Jul 21, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

n-io commented Jul 14, 2026 •

edited

Loading

codecov Bot commented Jul 14, 2026 •

edited

Loading

n-io Jul 17, 2026 •

edited

Loading

Samielakkad left a comment •

edited

Loading

n-io commented Jul 21, 2026 •

edited

Loading