SIMD backends for math functions by tretre91 · Pull Request #3 · CExA-project/CExA-Experimental

tretre91 · 2025-04-17T08:08:33Z

This PR adds Kokkos SIMD compatible wrappers for simd math libraries. A user can use this as a cmake project, and include CEXA_SIMD_Backends.hpp in their program, or directly include CEXA_SIMD_{SLEEF,SVML,AOCLLIBM}.hpp.

Currently, only sleef and svml are supported in the CMake, I will add aocl libm later as the support for vectorized functions is not complete at the moment (users can still use the standalone header)

I also added a basic accuracy test to ensure that the functions give a reasonable (<=4ulp) difference from the exact value

- don't install gtest when tests are enabled - Bump cmake min version to 3.23 - Use file sets - Use configure_file to pass the macros for selecting sleef or svml

PaulGannay

Why are some of the file name in all caps (CEXA_SIMD_AOCLLIBM.hpp, ...)?

PaulGannay · 2026-03-30T11:07:18Z

+#define CEXA_IMPL_SIMD_AOCL_AVX2_UNARY_FUNCTION(func) \
+  CEXA_IMPL_SIMD_AOCL_AVX2_S4_UNARY_FUNCTION(func)    \
+  CEXA_IMPL_SIMD_AOCL_AVX2_S8_UNARY_FUNCTION(func)    \
+  CEXA_IMPL_SIMD_AOCL_AVX2_D4_UNARY_FUNCTION(func)


Is there a reason why unary macro are defined as 3 separate macro merged in a fourth one while the binary macro are all defined in a single macro?

Because AOCL LibM doesn't support every function for every vector type, CEXA_IMPL_SIMD_AOCL_AVX2_UNARY_FUNCTION is used when the function is implemented for fp32x8, fp32x4 and fp64x4, otherwise I use one or more of the three separate macros, for example cbrt is only defined for fp32x4, so I use CEXA_IMPL_SIMD_AOCL_AVX2_S4_UNARY_FUNCTION(cbrt).

We only use the binary macro for pow which is supported on all vector types, which is why there is a single macro for binary functions.

PaulGannay · 2026-03-30T11:11:35Z

+#cmakedefine CEXA_SIMD_ENABLE_SLEEF
+#cmakedefine CEXA_SIMD_ENABLE_SVML


Shouldn't AOCL-libm also be in this file?

I haven't figured out yet how to link AOCL cleanly through CMake, the libm examples rely on a user provided variable with the path to an aocl-libm install instead of doing a find_package or using pkg-config, they mention .deb/.rpm packages which don't seem to exist so I cannot look at this either to see if they export CMake files which could be used with a find_package

PaulGannay · 2026-03-30T11:37:09Z

+option(
+  CEXA_SIMD_ENABLE_SLEEF
+  "Use the sleef library to implement the simd math functions"
+  OFF
+)
+option(CEXA_SIMD_ENABLE_SVML "Use the svml to implement the simd math functions" OFF)
+option(CEXA_SIMD_ENABLE_TESTS "Enable accuracy tests" OFF)
+option(CEXA_SIMD_ENABLE_INSTALL "Enable installation of the library" ON)
+
+if(CEXA_SIMD_ENABLE_SLEEF AND CEXA_SIMD_ENABLE_SVML)
+  message(FATAL_ERROR "Only one backend can be enabled at a time")
+endif()
+


No AOCL-libm?

Co-authored-by: Paul Gannay <paul.gannay@cea.fr>

- update the CI to test with kokkos 4.6 and 5.1

tretre91 · 2026-04-01T13:15:21Z

How do we make sure that the new implementation of these simd math functions get selected over the generic ones in Kokkos ?

I can try to setup a CI job running the Kokkos simd perf tests with and without the sleef backend and compare the results, assuming that running inside a github runner doesn't introduce too much noise

tretre91 · 2026-04-01T13:19:36Z

No particular reason except for the convenience of not having to remember which letters are lowercase in "AOCLLibM". I can change it if needed

I would prefer it to be small cap only, but I have no particular rational, I'll let you decide.

By "small cap only" do you mean something like cexa_simd_sleef or CEXA_SIMD_sleef ?

tpadioleau · 2026-04-01T17:16:29Z

What about other headers CEXA_SIMD_SVML, CEXA_SIMD_AOCLLIBM ?

Good catch, I didn't notice that sleef was duplicated. For AOCLLIBM, I don't support it through CMake yet, they don't seem to export files to enable find_package, so I will probably have to implement a FindAOCLLIBM file myself.

Co-authored-by: Thomas Padioleau <thomas.padioleau@cea.fr>

tretre91 · 2026-04-03T13:15:09Z

Removed the AOCL Libm header for now, I will add it back in another PR once I figure out how to include it with CMake

tpadioleau

Looks ok to me

tretre91 self-assigned this Apr 17, 2025

tretre91 and others added 7 commits March 3, 2026 11:05

simd-backends: inital commit, added sleef and svml

951cac7

simd-backends: Make the library installable

c4cb086

simd-backends: add license headers

5fe50d1

simd-backends: add a README

e1811bb

Fix a typo in the README

45fd12d

Update README.md

70c71cc

Rename the Kokkos_* macros and options to CEXA_*

1e4e43c

tretre91 force-pushed the simd-backends branch from 505bc25 to 1e4e43c Compare March 3, 2026 16:28

tretre91 and others added 17 commits March 6, 2026 10:30

Update the README

4d6af24

Add comments explaining why some functions are ccommented out

44a28f2

Update the copyright notice

444f3f1

Overhaul the CMake files

a3476a4

- don't install gtest when tests are enabled - Bump cmake min version to 3.23 - Use file sets - Use configure_file to pass the macros for selecting sleef or svml

Include Kokkos_Macros instead of Kokkos_Core in the sleef wrapper

00da541

Update the README

001246e

Add a CI

1a6aef7

ci: fix some paths

90875f0

ci: Install sleef from source as the version from apt is too old

a8b494e

Mention the sleef version requirement in the README

73813ed

ci: cache sleef

54e28f6

Link to the right sleef target

0a84af9

ci: get the sleef libdir using find

7e9148d

ci: use quotes in the find argument

ff65bf5

ci: use Ninja as CMake generator

2c395c9

ci: add a find_package call when patching kokkos

88a80db

Don't use the CONFIG mode for find_package

264f451

PaulGannay reviewed Mar 30, 2026

View reviewed changes

Adrien-Tab reviewed Mar 30, 2026

View reviewed changes

Comment thread simd-backends/CMakeLists.txt Outdated

tretre91 and others added 2 commits March 31, 2026 14:17

Undef internal macros after their use

07550ac

Update the accuracy test based on suggestions from review

dba1595

Co-authored-by: Paul Gannay <paul.gannay@cea.fr>

tretre91 added 4 commits April 1, 2026 14:31

Remove the CEXA_SIMD_ENABLE_INSTALL option

571a28f

Use a more explicit ci step name than "Patch Kokkos"

812e0af

Set the min kokkos version to 4.6

cf4b06f

- update the CI to test with kokkos 4.6 and 5.1

ci: checkout code before restoring the cache

e52c9eb

Update the README

6a42646

tretre91 added 4 commits April 1, 2026 16:40

ci: add a job running the Kokkos SIMD perf tests

f9a1342

ci: fix a path in the perf test job

213e49b

ci: activate the venv before running the comapre script

3bcf138

ci: filter benchmarks to only run math functions

7602685

tpadioleau reviewed Apr 1, 2026

View reviewed changes

Comment thread simd-backends/README.md Outdated

tpadioleau reviewed Apr 1, 2026

View reviewed changes

Comment thread simd-backends/src/CMakeLists.txt Outdated

tretre91 added 5 commits April 2, 2026 11:09

Update the README

be742a8

ci: exclude scalar benchmarks and increase the number of iterations

c22bf00

Fix a typo in th src CMakeLists

a906d59

ci: reorder the cmake configure arguments

ea5b136

ci: limit the number of parallel cmake jobs to nproc

ddbc37b

tpadioleau reviewed Apr 3, 2026

View reviewed changes

Comment thread simd-backends/src/CMakeLists.txt Outdated

tpadioleau reviewed Apr 3, 2026

View reviewed changes

Comment thread simd-backends/src/CEXA_SIMD_SVML.hpp

tpadioleau reviewed Apr 3, 2026

View reviewed changes

Comment thread simd-backends/README.md Outdated

tpadioleau reviewed Apr 3, 2026

View reviewed changes

Comment thread simd-backends/README.md Outdated

tretre91 and others added 3 commits April 3, 2026 15:01

Guard exp, log and cbrt if using an intel compiler

0a664ae

Apply suggestions from review

879bc5a

Co-authored-by: Thomas Padioleau <thomas.padioleau@cea.fr>

Remove the AOCL header

8b7bf1d

tpadioleau approved these changes Apr 3, 2026

View reviewed changes

tpadioleau merged commit 2cf6968 into main Apr 7, 2026
3 checks passed

tretre91 deleted the simd-backends branch May 7, 2026 16:22

		#cmakedefine CEXA_SIMD_ENABLE_SLEEF
		#cmakedefine CEXA_SIMD_ENABLE_SVML

Uh oh!

Conversation

tretre91 commented Apr 17, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

PaulGannay left a comment

Choose a reason for hiding this comment

Uh oh!

PaulGannay Mar 30, 2026

Choose a reason for hiding this comment

Uh oh!

tretre91 Mar 31, 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!

Uh oh!

PaulGannay Mar 30, 2026

Choose a reason for hiding this comment

Uh oh!

tretre91 Mar 31, 2026

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

PaulGannay Mar 30, 2026

Choose a reason for hiding this comment

Uh oh!

Uh oh!

tretre91 commented Apr 1, 2026

Uh oh!

tretre91 commented Apr 1, 2026

Uh oh!

Uh oh!

tpadioleau Apr 1, 2026

Choose a reason for hiding this comment

Uh oh!

tretre91 Apr 2, 2026

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

tretre91 commented Apr 3, 2026

Uh oh!

tpadioleau left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

tretre91 commented Apr 17, 2025 •

edited

Loading

tretre91 Mar 31, 2026 •

edited

Loading