transforms: lower polynomials to arithmetic operations by szerdick · Pull Request #5951 · xdslproject/xdsl

szerdick · 2026-04-27T16:15:03Z

Based on the specified polynomial family, evaluation interval, and scheme lower the polynomial approximation to arithmetic operations. Only supports Clenshaw's recurrence evaluation so far.

codecov · 2026-04-27T16:23:04Z

Codecov Report

❌ Patch coverage is 92.42424% with 5 lines in your changes missing coverage. Please review.
✅ Project coverage is 86.74%. Comparing base (88cc66f) to head (49cfbf0).
⚠️ Report is 73 commits behind head on main.

Files with missing lines	Patch %	Lines
xdsl/transforms/polynomial_eval_to_arith.py	92.06%	3 Missing and 2 partials ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #5951      +/-   ##
==========================================
+ Coverage   86.73%   86.74%   +0.01%     
==========================================
  Files         425      426       +1     
  Lines       63539    63628      +89     
  Branches     7289     7293       +4     
==========================================
+ Hits        55108    55196      +88     
- Misses       6865     6866       +1     
  Partials     1566     1566

☔ View full report in Codecov by Sentry.
📢 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 · 2026-04-27T16:43:57Z

We can take a look at this once the polynomial dialect is merged

superlopuh · 2026-05-14T12:26:51Z

+            raise NotImplementedError(
+                f"polynomial.eval scheme {scheme.value!r} has no lowering"
+            )


Can you please raise a PassFailedException instead and add a test for it so that we detect what to update when new cases are added?

superlopuh · 2026-05-14T12:27:10Z

+        upper = op.domain_upper.value.data if op.domain_upper is not None else None
+
+        scheme = op.eval_scheme
+        if scheme is polynomial.EvalScheme.CLENSHAW:


feels like a good place to use a match statement instead

superlopuh · 2026-05-14T12:30:04Z

+def _float_constant(
+    value: float,
+    tp: AnyFloat | VectorType[AnyFloat] | TensorType[AnyFloat],
+    rewriter: PatternRewriter,
+) -> arith.ConstantOp:
+    """Create and insert a float constant, handling scalar/vector/tensor types."""
+    if isa(tp, VectorType[AnyFloat]):
+        attr = DenseIntOrFPElementsAttr.from_list(tp, [value])
+    elif isa(tp, TensorType[AnyFloat]):
+        attr = DenseIntOrFPElementsAttr.from_list(tp, [value])
+    elif isa(tp, AnyFloat):
+        attr = FloatAttr(value, tp)
+    else:
+        raise TypeError(f"Unsupported type for float constant: {tp}")
+    return rewriter.insert(arith.ConstantOp(attr))
+


this feels familiar, do we have code like this in the framework already? I think there was something similar in your exp to polynomial pass, is that right? I think we might want a helper for this in builtin with dedicated tests.

Hm, what is kind of suboptimal is that the function returns an arithmetic.ConstantOp but the arithmetic dialect is lower than the builtin dialect in the dialect stack (arith imports builtin).

But I could make a function in builtin dialect that returns FloatAttr | DenseIntOrFPElementsAttr and then the function in each of the passes would just be this:

return rewriter.insert(arith.ConstantOp(float_constant_attr(value, tp)))

but it feels a bit awkward to define another function then in each pass

Yeah sorry that's what I meant

I still think we should do this first before merging this PR

alexarice

I haven't checked the mathematics, but overall looks fine to me.

alexarice · 2026-06-01T08:37:48Z

@@ -0,0 +1,111 @@
+// RUN: xdsl-opt -p polynomial-eval-to-arith %s | filecheck %s
+
+builtin.module {


Nit: There's no need to have the builtin.module op explicitly, and not having it means you don't have to indent the entire file.

alexarice · 2026-06-01T08:41:55Z

+        offset = -(upper + lower) / (upper - lower)
+        scale_op = _float_constant(scale, tp, rewriter)
+        offset_op = _float_constant(offset, tp, rewriter)
+        scaled = rewriter.insert(arith.MulfOp(x, scale_op.result))


Not essential, but I wonder if this would all be a lot cleaner with an implicit builder?

alexarice · 2026-06-01T08:43:53Z

    lower_affine.LowerAffinePass(),
    convert_scf_to_riscv_scf.ConvertScfToRiscvPass(),
    expand_math_to_polynomials.ExpandMathToPolynomialsPass(),
+    polynomial_eval_to_arith.PolynomialEvalToArithPass(),


This feels like it should be a separate change?

superlopuh · 2026-06-01T09:02:38Z

+        pytest.fail(
+            f"EvalScheme.{scheme.name} has no dispatch branch in "
+            f"polynomial-eval-to-arith. Add a case for it in "
+            f"PolynomialEvalToArith.match_and_rewrite. ({e})"
+        )


let's test this with lit/filecheck instead

szerdick requested a review from superlopuh April 27, 2026 16:15

superlopuh marked this pull request as draft April 27, 2026 16:43

szerdick force-pushed the szerdick/add-polynomial-eval-pass branch from 7dae066 to 58ae869 Compare May 6, 2026 14:09

szerdick mentioned this pull request May 7, 2026

transforms: (exp) lower exp to polynomial dialect #6044

Draft

szerdick marked this pull request as ready for review May 8, 2026 14:32

szerdick added 9 commits May 13, 2026 10:23

add polynomial ops and heir compat tests

5323843

make default polynomialtype macro

6eee718

get method

594db9d

edit test to match numbers

e31e996

add generic format and heir-opt first tests

0c2d643

resolve pyright issues

a7c6d8d

pyrithgt issue resolve

73054d3

add lowering pass from polynomial dialect to arithmetic operations

fd27cf2

add to pipeline

e6f7a53

szerdick force-pushed the szerdick/add-polynomial-eval-pass branch from 2674c91 to e6f7a53 Compare May 13, 2026 12:11

szerdick added 4 commits May 14, 2026 13:55

weird comments

2add8e6

Merge branch 'main' into szerdick/add-polynomial-eval-pass

bb053d9

pyritht fix

ff69cee

pyritht fix2

80923ba

superlopuh reviewed May 14, 2026

View reviewed changes

Comment thread xdsl/transforms/expand_polynomial_eval.py Outdated

superlopuh reviewed May 14, 2026

View reviewed changes

szerdick mentioned this pull request May 14, 2026

kernels: (exp_micro) add linalg variant that expands exp to polynomial approximations based on acc_bound opencompl/riscv-paper-experiments#357

Open

rename pass and add test

49cfbf0

szerdick requested a review from superlopuh May 15, 2026 09:36

alexarice reviewed Jun 1, 2026

View reviewed changes

alexarice assigned szerdick Jun 1, 2026

alexarice added the transformations Passes, rewrites label Jun 1, 2026

superlopuh reviewed Jun 1, 2026

View reviewed changes

		@@ -0,0 +1,111 @@
		// RUN: xdsl-opt -p polynomial-eval-to-arith %s \| filecheck %s

		builtin.module {

Uh oh!

Conversation

szerdick commented Apr 27, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

codecov Bot commented Apr 27, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Codecov Report

Uh oh!

superlopuh commented Apr 27, 2026

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

szerdick May 14, 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!

superlopuh Jun 1, 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!

alexarice 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!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

szerdick commented Apr 27, 2026 •

edited

Loading

codecov Bot commented Apr 27, 2026 •

edited

Loading

szerdick May 14, 2026 •

edited

Loading

superlopuh Jun 1, 2026 •

edited

Loading