Support for half_t as a scalar type by brian-kelley · Pull Request #3360 · kokkos/kokkos-kernels

brian-kelley · 2026-10-23T23:59:03Z

CMake logic to check if half_t is a natively supported type according to kokkos, and if so provide the option KokkosKernels_INST_HALF
- if not natively supported (half_t == float) then this option can't be turned on
A few fixes to library code for half_t. For example, don't pass a half_t to printf directly, and don't try to implicitly convert a bool to half_t.

But a large majority of the changes are for enabling half_t coverage in tests:

instantiate tests for half_t
increase numerical tolerances for half_t. This introduces a helper TestUtils::selectEpsilonForScalarType<Scalar>(eps_double, eps_float, eps_half) to make this cleaner
in some places, decrease the size of the matrix/vector for half_t test cases to reduce precision loss
In ode, the bdf test only runs bdf1 and skips the convergence order check

A few things need to happen, first we need to enable the logic in cmake, second we need to add some guard to only enable when half is not float otherwise we create double instantiations for the float type, thirds we will likely need to clean up some code in the library... Signed-off-by: Luc Berger-Vergiat <lberge@sandia.gov>

Now using the macro defined in Kokkos Core to signal support of half_t using native type instead of float. Adding test that exercise the half code path when the type is instantiated. Signed-off-by: Luc Berger-Vergiat <lberge@sandia.gov>

Signed-off-by: Luc Berger-Vergiat <lberge@sandia.gov>

Co-Authored-By: Laguna S2.1 Signed-off-by: Brian Kelley <bmkelle@sandia.gov>

Make sure compiler launcher from Kokkos is used with check_cxx_symbol_exists, otherwise this test doesn't work for cuda. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Signed-off-by: Brian Kelley <bmkelle@sandia.gov>

batched dense/sparse, blas serial/team, most of sparse Co-Authored-By: Laguna S 2.1 Signed-off-by: Brian Kelley <bmkelle@sandia.gov>

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Signed-off-by: Brian Kelley <bmkelle@sandia.gov>

Signed-off-by: Brian Kelley <bmkelle@sandia.gov>

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Signed-off-by: Brian Kelley <bmkelle@sandia.gov>

Signed-off-by: Brian Kelley <bmkelle@sandia.gov>

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Signed-off-by: Brian Kelley <bmkelle@sandia.gov>

Signed-off-by: Brian Kelley <bmkelle@sandia.gov>

and make nrm2w_squared test for half_t consistent with nrm2w Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Signed-off-by: Brian Kelley <bmkelle@sandia.gov>

Signed-off-by: Brian Kelley <bmkelle@sandia.gov>

lucbv

I do not think that everything needs to be addressed, but some of the comment should be turned into issues to work on later. Will do that when I get to the office...

lucbv · 2026-10-24T14:41:34Z

+      A[i * as] += ValueType(minus_abs_tiny) * ValueType(static_cast<float>(a_real < 0));
+      A[i * as] += ValueType(abs_tiny) * ValueType(static_cast<float>(a_real >= 0));


I would be curious to see if there is any performance reason not to implement this as a tertiary operator?

Suggested change

A[i * as] += ValueType(minus_abs_tiny) * ValueType(static_cast<float>(a_real < 0));

A[i * as] += ValueType(abs_tiny) * ValueType(static_cast<float>(a_real >= 0));

A[i * as] += (a_real < 0) ? ValueType(minus_abs_tiny) : ValueType(0);

A[i * as] += (a_real >= 0) ? ValueType(abs_tiny) * ValueType(0);

lucbv · 2026-10-24T14:43:43Z

 // and half_t is not an alias to float.
-#if !defined(KOKKOSKERNELS_ETI_ONLY) && !defined(KOKKOSKERNELS_IMPL_CHECK_ETI_CALLS) && \
-    defined(KOKKOS_HALF_T_IS_FLOAT) && !KOKKOS_HALF_T_IS_FLOAT
+#if !defined(KOKKOSKERNELS_ETI_ONLY) && !defined(KOKKOSKERNELS_IMPL_CHECK_ETI_CALLS) && defined(KOKKOSKERNELS_INST_HALF)


Well, we are instantiating half_t now so we could probably revisit this logic?

lucbv · 2026-10-24T19:55:05Z

For laswp that does not compute any arithmetic operations I would have expected that we can use the arithmetic traits epsilon value?

lucbv · 2026-10-24T19:59:14Z

 #pragma unroll
 #endif
-    for (int i = 0; i < m; ++i) A[i * as0] *= alpha;
+    for (int i = 0; i < m; ++i) A[i * as0] *= ValueType(alpha);


This might not be really what we want especially if ScalarType is double and ValueType is lower precision?
This goes back to discussions about the definition of our rules for type promotion...

lucbv · 2026-10-24T20:02:47Z


 template <class ScalarA, class ScalarB, class Device>
 int test_dot() {
+  const int Nbig = std::is_same<ScalarA, Kokkos::Experimental::half_t>::value ? 128 : 1024;


Is the problem that we cannot add too many random generated numbers before risking overflow with half_t?
I am suspecting that we are doing the reduction with double by default which should be fine for rounding error but might not be possible to cast without overflow. That's something that we could test for in the algorithm itself in debug mode...

ah, we don't have an InnerProductSpaceTraits or DotAccumulatingScalar specialization for half_t, so this was doing the accumulation in half_t. I'll try changing this to accumulate in float and then see if we can get away with the same vector length as other tests.

Like we talked about yesterday, native dot was not accumulating in double (even for float vectors). I opened #3365 to fix this. If possible I'd like to get that merged first

lucbv · 2026-10-24T20:33:14Z

+  const int max_x   = 1;
+  const int max_y   = 1;
+  const int max_val = 1;


Same as above

lucbv · 2026-10-24T20:33:22Z

+  const int max_x   = 10;
+  const int max_y   = 10;
+  const int max_val = 10;


Same as above

lucbv · 2026-10-24T20:33:33Z

+  const int max_x   = 10;
+  const int max_y   = 10;
+  const int max_val = 10;


Same as above

lucbv · 2026-10-24T20:33:51Z

+  const int max_x     = 1;
+  const int max_y     = 1;
+  const mag_t max_val = static_cast<mag_t>(2);


Even more suspicious...

lucbv · 2026-10-24T20:34:45Z

We might want to promote the use of KokkosKernels::ArithTraits<>::one()/zero() to our LLM friends

Signed-off-by: Brian Kelley <bmkelle@sandia.gov>

- Fix selectEpsilonForScalarType; if KOKKOS_HALF_T_IS_FLOAT it should use the float value (2nd arg) when scalar_t is either half_t or float, not the half value (3rd arg). - Fix tolerances in ODE Newton tests (were wrong when KOKKOS_HALF_T_IS_FLOAT for same reason). Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Signed-off-by: Brian Kelley <bmkelle@sandia.gov>

lucbv and others added 16 commits October 19, 2026 13:50

Fixing the check logic to bring in headers from Kokkos

d505092

Signed-off-by: Luc Berger-Vergiat <lberge@sandia.gov>

removing debug output

f50d7fe

Signed-off-by: Luc Berger-Vergiat <lberge@sandia.gov>

Type handling and test tolerance changes for half_t

26e37ea

Co-Authored-By: Laguna S2.1 Signed-off-by: Brian Kelley <bmkelle@sandia.gov>

Fix cmake logic to find KOKKOS_CORE_HAS_HALF

933d32d

Make sure compiler launcher from Kokkos is used with check_cxx_symbol_exists, otherwise this test doesn't work for cuda. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Signed-off-by: Brian Kelley <bmkelle@sandia.gov>

Increase half_t test coverage

93368f7

batched dense/sparse, blas serial/team, most of sparse Co-Authored-By: Laguna S 2.1 Signed-off-by: Brian Kelley <bmkelle@sandia.gov>

Finish final past for half_t test coverage

2f26aff

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Signed-off-by: Brian Kelley <bmkelle@sandia.gov>

Add missing include

c87a6c0

Signed-off-by: Brian Kelley <bmkelle@sandia.gov>

Fix half_t as printf argument

176d92c

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Signed-off-by: Brian Kelley <bmkelle@sandia.gov>

Avoid half_t(bool) construction; cast to float first

868c57a

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Signed-off-by: Brian Kelley <bmkelle@sandia.gov>

Simplify choosing tolerance based on scalar type

f100e41

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Signed-off-by: Brian Kelley <bmkelle@sandia.gov>

Increase tolerance for trsv_mv

02e45fe

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Signed-off-by: Brian Kelley <bmkelle@sandia.gov>

Formatting

395b005

Signed-off-by: Brian Kelley <bmkelle@sandia.gov>

Use easier matrix for half_t trsv test

255724f

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Signed-off-by: Brian Kelley <bmkelle@sandia.gov>

Formatting

8a40aeb

Signed-off-by: Brian Kelley <bmkelle@sandia.gov>

brian-kelley changed the title ~~Full support for half_t as a scalar type~~ Support for half_t as a scalar type Oct 23, 2026

brian-kelley and others added 2 commits October 23, 2026 18:47

Revert scal float test being removed

997ea6d

and make nrm2w_squared test for half_t consistent with nrm2w Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Signed-off-by: Brian Kelley <bmkelle@sandia.gov>

Remove TODO that is now completed

6939579

Signed-off-by: Brian Kelley <bmkelle@sandia.gov>

lucbv requested changes Oct 24, 2026

View reviewed changes

brian-kelley and others added 3 commits October 26, 2026 10:43

Simplify nrm2w, nrm2w_squared test tolerances

dda89dd

Signed-off-by: Brian Kelley <bmkelle@sandia.gov>

Formatting

506c987

Signed-off-by: Brian Kelley <bmkelle@sandia.gov>

		A[i * as] += ValueType(minus_abs_tiny) * ValueType(static_cast<float>(a_real < 0));
		A[i * as] += ValueType(abs_tiny) * ValueType(static_cast<float>(a_real >= 0));

Conversation

brian-kelley commented Oct 23, 2026

Uh oh!

lucbv 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!

brian-kelley Oct 27, 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!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

This branch has not been deployed

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

brian-kelley Oct 27, 2026 •

edited

Loading