Single-precision (fp32) build support by hardik-corintis · Pull Request #5033 · firedrakeproject/firedrake

hardik-corintis · 2026-04-15T14:04:27Z

Description

Adds single-precision (fp32) build support. Firedrake can now run on a PETSc installation compiled with --with-precision=single. The approach mirrors complex mode: precision is detected at import time from PETSc's build variables and flows through from there.

AI disclosure: Parts of this PR were developed with assistance from Claude (Anthropic). All changes have been reviewed, tested locally, and are fully understood by the author.

Prerequisite

Requires https://gitlab.com/petsc/petsc/-/merge_requests/9272 (petsc4py: handle PETSC_DOUBLE in DMSwarm.getField). Without it, DMSwarm.getField() raises AssertionError on fp32 builds because PETSC_DOUBLE is not mapped to a numpy dtype in the single-precision case where PETSC_REAL != PETSC_DOUBLE.

Changes

scripts/firedrake-configure

Adds --arch single / ScalarType.SINGLE; passes --with-precision=single to PETSc configure; excludes fftw and suitesparse (no fp32 support in those libraries)

tsfc/parameters.py, tsfc/loopy.py, tsfc/kernel_interface/common.py, tsfc/ufl_utils.py

scalar_type / scalar_type_c derived from PETSc precision at import time; constant initializers cast to the kernel scalar dtype

Core (evaluate.h, locate.c, pointquery_utils.py, pointeval_utils.py, mg/kernels.py)

Replace hardcoded double / int with PetscReal / PetscInt in generated C code
Convergence epsilon in point query tightened to 1e-6 in fp32 mode vs 1e-12 in fp64, to stay within single-precision range

firedrake/mesh.py, firedrake/utility_meshes.py

Vertex coordinates and reference-cell distances use PETSc.RealType; physical coordinate arrays for rtree and DMSwarmPIC_coor remain float64 (required by the rtree C API and PETSc swarm internals)

firedrake/function.py

Point evaluation coerces coordinates to float64 regardless of ScalarType, for geometric robustness in cell location

firedrake/assemble.py, firedrake/functionspaceimpl.py, pyop2/codegen/builder.py

Replace dtype=int with dtype=IntType in numpy.prod and array allocation calls

firedrake/utils.py

Adds single_mode boolean flag (mirrors complex_mode)

Tests

Adds @pytest.mark.skipsingle marker for tests incompatible with fp32
tests/firedrake/conftest.py: registers the marker and wires it up
Skips a small set of tests that require double-precision accuracy (test_locate_cell, test_interpolate_cross_mesh[extrudedcube], test_parallel_high_order_location)

.github/workflows/core.yml

Adds single to the CI matrix alongside default and complex

Known limitations

test_parallel_high_order_location is skipped in fp32: high-order cell location in a warped mesh requires double-precision accuracy that fp32 cannot provide at tolerance=0.0001.

connorjward · 2026-04-15T14:19:56Z

Thanks for this.

adds --arch single

What about single+complex? Isn't that a valid configuration?

PETSc version bump (v3.24.5 → v3.25.0)

This isn't necessary. That's all going to be taken care of when I release the next major version in the next 24 hours.

Fix needed upstream in petsc4py: if ctype == PETSC_DOUBLE: typenum = NPY_DOUBLE in petsc4py/PETSc/DMSwarm.pyx.

Can you get this fixed upstream? Clearly Claude already knows what to do.

hardik-corintis · 2026-05-10T00:21:26Z

Thanks for this.

adds --arch single

What about single+complex? Isn't that a valid configuration?

Added ScalarType.SINGLE_COMPLEX (--arch single-complex) to firedrake-configure for all four platform targets: single_mode (from PETSC_PRECISION) and complex_mode (from PETSC_SCALAR) are detected independently, and ScalarType already resolves to numpy.complex64 for a single+complex build.

PETSc version bump (v3.24.5 → v3.25.0)

This isn't necessary. That's all going to be taken care of when I release the next major version in the next 24 hours.

Removed from the PR description — thanks.

Fix needed upstream in petsc4py: if ctype == PETSC_DOUBLE: typenum = NPY_DOUBLE in petsc4py/PETSc/DMSwarm.pyx.

Can you get this fixed upstream? Clearly Claude already knows what to do.

Opened a PR here: https://gitlab.com/petsc/petsc/-/merge_requests/9272

connorjward

I've just looked at the CI/install related side of this (at a glance everything else seems pretty good).

We just have to be careful about adding new test builds to Firedrake. I will try and figure out a solution soon.

connorjward · 2026-05-20T14:36:42Z

We just have to be careful about adding new test builds to Firedrake. I will try and figure out a solution soon.

I have a solution for this in #5117 (merged). Once we update main (#5134) then you can use it in this pull request. You basically need to follow the same procedure that we have for testing complex or CUDA builds.

…spatialindex float64 coercion

- mesh.py: clarify why PETSc.RealType is correct for plex_from_cell_list - mg/kernels.py: fix to_reference_coords_kernel signature (PetscScalar *X -> %(RealType)s *X) and add RealType to the template substitution dict - evaluate.h: comment explaining why int/double is intentional for evaluate() - test_stokes_mini.py: replace fp32 solver workaround with @pytest.mark.skipsingle; fieldsplit path needs a dedicated fp32 test - conftest.py: remove trailing blank line

…nd skip VertexOnlyMesh tests in fp32

…ests

… markers

…p32 skip refinements

Co-authored-by: Connor Ward <c.ward20@imperial.ac.uk>

connorjward

In general I think this is good - thank you!

We have to think about testing this. I don't mind having partial test coverage and skipping a lot of things provided that it is clearly recorded that that is only happening because we haven't gotten around to fixing things. For example an issue discussing this should definitely be opened. It might even be a "good first issue".

connorjward · 2026-05-21T14:54:58Z


 @pytest.mark.skipcomplex
 @pytest.mark.parallel([1, 2])
+@pytest.mark.skipsingle  # VertexOnlyMesh point location has fp32 precision issues


Is this still true? You look to have done some work on point location

connorjward · 2026-05-21T14:56:40Z

    assert errornorm(w, wcheck) < tol


+@pytest.mark.skipsingle  # asserts ksp.its == 1 (exact algebraic inverse); not achievable in fp32


@JHopeCollins is there a way we can run this test in single precision? What is the right way to loosen things?

connorjward · 2026-05-21T14:57:26Z

        "unitsquare_from_high_order",
        "unitsquare_to_high_order",
-        "extrudedcube",
+        "extrudedcube",  # petsc/petsc!9272 fixes fp32 DMSwarm.getField; re-add skipsingle if that MR is not in the shipped PETSc


Can this comment go? Has the PETSc MR been merged?

connorjward · 2026-05-21T14:58:57Z

    return errornorm(uexact, u, degree_rise=0), errornorm(pexact, p, degree_rise=0)


+@pytest.mark.skipsingle  # Schur complement fieldsplit does not converge in fp32; tracked in https://github.com/firedrakeproject/firedrake/pull/5033


What does it mean "tracked in 5033", that's this PR

connorjward · 2026-05-21T15:00:49Z

 from firedrake import *
 import pytest

+pytestmark = pytest.mark.skipsingle


I think it would be helpful wherever we have this to give a reason for the skip. I don't want confusion between "this is genuinely impossible" (e.g. skipnogpu) and "TODO: this just hasn't been worked out".

connorjward · 2026-05-21T15:02:08Z

        if isinstance(temp, gem.Constant):
-            data.append(lp.TemporaryVariable(name, shape=temp.shape, dtype=dtype, initializer=temp.array, address_space=lp.AddressSpace.LOCAL, read_only=True))
+            # loopy raises if initializer.dtype != declared dtype (e.g. float64 GEM constant in fp32 build).
+            initializer = temp.array.astype(dtype) if temp.array.dtype != dtype else temp.array


I feel like we should really track down where we're inserting float64s and make them the right scalar/real type instead

connorjward · 2026-05-21T15:02:37Z

+    in single- and double-precision builds. The threshold is loose enough that
+    discretization error dominates round-off for both fp32 and fp64."""
+    err = helmholtz(4)[0]   # 16x16 mesh, degree 2 — expected L2 error ~2e-4
+    assert float(err) < 1e-2


Why this new test? What is it testing that the other tests aren't?

connorjward · 2026-05-21T15:06:42Z

+        maintainer="the Firedrake team",
+        contact="on Slack",


I'm not thrilled about taking on the maintenance burden of these additional configurations given that we have never tested them. We're happy to maintain previous Ubuntus because we used to maintain them. The idea behind COMMUNITY_ARCHS is that we could make these other people's responsibilities - like yours or Corintis'.

connorjward · 2026-05-21T15:11:24Z

            return cache[tolerance]
        except KeyError:
            IntTypeC = as_cstr(IntType)
+            RealTypeC = RealType_c


This seems pretty redundant

connorjward · 2026-05-21T15:12:39Z

            self.tolerance = tolerance
-        xs = np.asarray(xs, dtype=utils.ScalarType)
+        # Physical coordinates: always float64 (libspatialindex requires double).
+        xs = np.asarray(xs, dtype=np.float64)


This is going to confuse people in future. Can you put comments in some of the places where we end up seeing double when you would otherwise expect to see RealType. E.g. above

{IntTypeC} locator(struct Function *f, double *x, {RealTypeC} *X, {RealTypeC} *ref_cell_dists_l1, {IntTypeC} *cells, {IntTypeC} npoints, size_t ncells_ignore, {IntTypeC}* cells_ignore)

hardik-corintis marked this pull request as draft April 15, 2026 14:13

Olender mentioned this pull request Apr 23, 2026

Add support for single precision NDF-Poli-USP/spyro#281

Open

hardik-corintis force-pushed the fp32-support branch 2 times, most recently from 0dc7472 to dd517a9 Compare May 9, 2026 23:10

hardik-corintis marked this pull request as ready for review May 10, 2026 00:18

connorjward requested changes May 11, 2026

View reviewed changes

Comment thread .github/workflows/core.yml Outdated

Comment thread scripts/firedrake-configure Outdated

hardik-corintis added 20 commits May 21, 2026 11:35

Fix single-precision (fp32) support: TSFC initializer dtype cast and …

fac7012

…spatialindex float64 coercion

Replacing double with Petsc.RealType

fcf7b57

Use PETSc types for precision and integer width in C code

561feed

Replace hardcoded dtypes with PETSc-derived types

47a1ad2

Add single-precision (fp32) build support and detection

e38b9a7

Fix sparsity.pyx: revert broken PetscScalar cimport

39f82af

skipping tests for single precision

fa4b924

DBL_MAX to PETSC_MAX_REAL

d3d8aec

fp32 detection in tests

2a0b767

ensuring that coordinates are real

b0bbca6

Fix PetscScalar/PetscReal type mismatch in prolong/restrict kernels a…

a986f20

…nd skip VertexOnlyMesh tests in fp32

restoring tests/test_durations.json

6ddb6ae

reverting solver parameters back to original form

3076a4d

Port missing dtype=int -> IntType fixes

d65cd77

adding dtypes for int

64cce3f

Update skipsingle comment for test_parallel_high_order_location

21337e9

removed alias fp32

15fd4cc

Add single-complex arch to firedrake-configure; drop _fp32 alias in t…

b8a679f

…ests

Fix duplicate SINGLE arch keys and missing CC in UNKNOWN SINGLE_COMPLEX

3cd08b8

hardik-corintis force-pushed the fp32-support branch from cf73186 to 3cd08b8 Compare May 21, 2026 11:22

hardik-corintis added 5 commits May 21, 2026 13:22

Add comments explaining Cython typedef nominal hint for fp32 builds

088e87e

convergence eps constant, PETSC_MAX_REAL, tsfc type guard, skipsingle…

af8f153

… markers

IntType cleanup, arch moves to CommunityArch, test tolerance fixes, f…

57be852

…p32 skip refinements

Restore blank lines

326b0d9

cleaning up comments

e13b4e1

connorjward requested changes May 21, 2026

View reviewed changes

Comment thread .github/workflows/push.yml Outdated

connorjward added the ci:single Run the test suite in single precision label May 21, 2026

Update .github/workflows/push.yml

40ab2d8

Co-authored-by: Connor Ward <c.ward20@imperial.ac.uk>

hardik-corintis requested a review from connorjward May 21, 2026 14:36

fixing linting

2af6283

connorjward requested changes May 21, 2026

View reviewed changes

		assert errornorm(w, wcheck) < tol


		@pytest.mark.skipsingle # asserts ksp.its == 1 (exact algebraic inverse); not achievable in fp32

		return errornorm(uexact, u, degree_rise=0), errornorm(pexact, p, degree_rise=0)


		@pytest.mark.skipsingle # Schur complement fieldsplit does not converge in fp32; tracked in https://github.com/firedrakeproject/firedrake/pull/5033

Conversation

hardik-corintis commented Apr 15, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Description

Prerequisite

Changes

Known limitations

Uh oh!

connorjward commented Apr 15, 2026

Uh oh!

hardik-corintis commented May 10, 2026

Uh oh!

connorjward left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

connorjward commented May 20, 2026

Uh oh!

Uh oh!

connorjward 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!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

hardik-corintis commented Apr 15, 2026 •

edited

Loading