Generalize the temperature scalar type for forward-mode AD by Samm-Py · Pull Request #35 · ORNL-MDF/Finch

Samm-Py · 2026-08-05T04:10:58Z

Summary

Make the temperature field scalar configurable in Grid, Boundary,
Layer, and SolidificationData, while retaining double as the default.
Derive the solver scalar from its temperature view and add scalar-valued
material/source properties plus ADL-based math and value customization
points. Finch core remains independent of any AD library.
Add an optional finch_sensitivity application using cpp_oti_lib to
compute first-order sensitivities for six material and source parameters in
one solve and compare them with central finite differences.

Compatibility

Existing callers continue to use double without source changes.
Mesh geometry, phase-change thresholds, and downstream solidification-event
storage remain double.
The sensitivity application is built only when cpp_oti_lib is found; the
standard Finch build has no new dependency.

Validation

clang-format 14.0.6 check passed for every changed C++ file.
GCC 9.4 Release builds passed with -Wall -Wextra -pedantic -Werror, both
with and without cpp_oti_lib available.
The single_line temperature outputs are bit-for-bit identical to upstream
main; solidification records are identical (their parallel emission order
is nondeterministic).
The sensitivity application completed on one and four MPI ranks. OTI values
matched the ordinary double solve, and all six derivatives were validated
against central finite differences using FINCH_FD_STEP=1e-8.

Finch's field type was fixed to double. This threads a Scalar template parameter (defaulting to double) through Grid, Boundary, Solver, Layer and SolidificationData, so the temperature field can carry a user-supplied arithmetic type. With double the generated code and temperature fields are unchanged; verified bit-identical temperature output and identical solidification records on the single_line case. The differentiated material and source inputs move into a MaterialProperties struct so a caller can supply values of the field scalar type rather than plain doubles read from the input deck. Quantities that only select a branch (solidus, liquidus) and all mesh geometry stay double. Finch_Scalar.hpp adds exp/fmin/fmax dispatch: arithmetic types forward to Kokkos as before, anything else resolves by ADL into the scalar type's own namespace. Finch therefore carries no dependency on any AD library. The optional finch_sensitivity application demonstrates the result with cpp_oti_lib, obtaining d(QoI)/dp for six material and source parameters from a single solve and validating each against central finite differences. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01PNkQTS2g7P6dNDybhWRkwz

Samm-Py · 2026-08-05T04:19:31Z

@Rombur The PR is open and ready for review.

streeve · 2026-08-05T17:12:19Z

+/****************************************************************************
+ * OTI (order-truncated imaginary) scalar type glue for Finch.
+ *
+ * This header is the ONLY place where Finch and cpp_oti_lib meet. It lives in


Change: I would pick a third place rather than put it here

Moved the Finch/Sparrow adapter out of applications into integrations/sparrow/Finch_Sparrow.hpp.

streeve · 2026-08-05T17:12:59Z

+
+# Optional forward-mode AD sensitivity application. Built only when the
+# header-only cpp_oti_lib is available; Finch itself never depends on it.
+find_path(CPP_OTI_LIB_INCLUDE_DIR otinum/otinum.hpp


Clarification: where is the OTI lib? And can we make it a real target instead? What's the plan there

The library is header only. Including this file into your path is enough.

Sure, but why is Sam hiding it from me

We are going to move it to a private repo inside https://github.com/ORNL-MDF and then open source it.

streeve · 2026-08-05T17:14:07Z

Suggestion: I think SingleLayer_Sensitivity.cpp is more clear

streeve · 2026-08-05T17:16:14Z

+    KOKKOS_INLINE_FUNCTION static double
+    value( const oti::otinum<M, N, Coeff>& x )
+    {
+        return static_cast<double>( x.real() );


Clarification: if we're going to all the trouble of propagating these AD types, is it worth also exposing the underlying scalar type as double/float?

Changes made to preserve underlying float/double coefficient type.

Rombur · 2026-08-06T12:56:46Z

+
+# Optional forward-mode AD sensitivity application. Built only when the
+# header-only cpp_oti_lib is available; Finch itself never depends on it.
+find_path(CPP_OTI_LIB_INCLUDE_DIR otinum/otinum.hpp


The library is header only. Including this file into your path is enough.

Rombur · 2026-08-06T13:18:42Z

+    case TwoSigma:
+        return "two_sigma";
+    default:
+        return "unknown";


You can assert here or does the code work with unknown?

Replaced silent unknown parameter labels with exceptions to prevent masked invalid-index errors.

Rombur · 2026-08-06T13:21:28Z

@@ -0,0 +1,161 @@
+/****************************************************************************


Here and in the new files, you forgot the copyright for Finch

Added Finch’s standard copyright and BSD-3-Clause license header to every new file.

Rombur · 2026-08-06T13:22:29Z

+
+constexpr int NP = Finch::Sensitivity::NumParameters;
+
+// First-order jets: one variable per differentiated parameter, derivatives


Please add a comment what jets means.

Changed Jet to OTINumber

Rombur · 2026-08-06T13:23:46Z

+{
+    std::array<std::string, 6> bc_types = { "adiabatic", "adiabatic",
+                                            "adiabatic", "adiabatic",
+                                            "adiabatic", "adiabatic" };


Should these boundary conditions come from the input file @streeve ?

Yeah, that would be a nice improvement

Rombur · 2026-08-06T13:27:39Z

+
+    // Evaluate the QoIs on the host. Copying the field to a host mirror keeps
+    // this independent of Kokkos reducer support for compound scalar types,
+    // which is not needed for a once-per-run diagnostic.


Does this mean that parallel_reduce does not work out of the box with otinum?

parallel_reduce works with a Kokkos-enabled otinum. The original version conservatively copied the temperature field to the CPU and summed it serially. We have since replaced that with Cabana’s grid_parallel_reduce and validated it on OpenMP and CUDA.

Rombur · 2026-08-06T13:43:33Z

+// solver only ever compares against -- solidus and liquidus -- stay double:
+// they select a branch rather than entering the arithmetic.
+template <typename Scalar>
+struct MaterialProperties


I am not sure about the name because absorption and two_sigma are source properties.

Renamed the bundle to SolverParameters because it contains both material and source parameters.

Co-authored-by: Bruno Turcksin <bruno.turcksin@gmail.com>

streeve reviewed Aug 5, 2026

View reviewed changes

Rombur reviewed Aug 6, 2026

View reviewed changes

Samm-Py and others added 2 commits August 7, 2026 11:22

Update applications/Finch_OTI.hpp

d087a00

Co-authored-by: Bruno Turcksin <bruno.turcksin@gmail.com>

Update src/Finch_Grid.hpp

f0a907d

Co-authored-by: Bruno Turcksin <bruno.turcksin@gmail.com>

Samm-Py requested review from Rombur and streeve August 7, 2026 19:22

MythicManSam added 2 commits August 7, 2026 14:30

Address Sparrow integration review comments

cad3a85

Clarify scalar template comment

8ce26e5

Samm-Py mentioned this pull request Aug 7, 2026

Oti spike #36

Closed

Avoid private Sparrow checkout in Finch CI

470c38a

		@@ -0,0 +1,161 @@
		/****************************************************************************


		constexpr int NP = Finch::Sensitivity::NumParameters;

		// First-order jets: one variable per differentiated parameter, derivatives

Uh oh!

Conversation

Samm-Py commented Aug 5, 2026

Summary

Compatibility

Validation

Uh oh!

Samm-Py commented Aug 5, 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!

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!

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!

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!

4 participants