{math}[GCCcore/13.2.0] ArmComputeLibrary v23.08

[migueldiascosta]

(created using eb --new-pr)

The motivation for this easyconfig was that on Arm (at least on a64fx, but probably also applies to other Arm processors) a pip-installed PyTorch was multiple times faster than an easybuilt one, and an analysis with perf showed that ACL was being used (also a recent OpenBLAS with support for ARM_SVE, but should be taken care by using PyTorch with a more recent toolchain and OpenBLAS, e.g. #20489)

This is not the most recent version of ACL, but PyTorch 2.3 (the one in #20489) says that the maximum supported version is 23.08

Using this with PyTorch 2.3 requires setting USE_MKLDNN=ON, USE_MKLDNN_ACL=ON, USE_MKLDNN_CBLAS=ON, and a patch derived from Ryo-not-rio/oneDNN@ca60ff4 to the bundled oneDNN

[migueldiascosta]

from https://github.com/pytorch/pytorch/blob/main/.ci/docker/common/install_acl.sh#L13-L16

in particular, arch=armv8a multi_isa=1 should be more generic without loosing functionality/performance:

https://github.com/ARM-software/ComputeLibrary/blob/de7288cb71e6b9190f52e50a44ed68c309e4a041/docs/user_guide/library.dox#L567-L578

benchmarks on a64fx compared to arch=armv8.2-a-sve didn't show any difference

[migueldiascosta]

Test report by @migueldiascosta
SUCCESS
Build succeeded for 1 out of 1 (1 easyconfigs in total)
cna0003.deucalion.macc.fccn.pt - Linux Rocky Linux 8.5, AArch64, UNKNOWN, Python 3.6.8
See https://gist.github.com/migueldiascosta/4f83eebbd8ba97bd57069d1d5a16be30 for a full test report.

[Flamefire]

When we set $ACL_ROOT_DIR we do not need the part of the patch where the FindACL is changed, see easybuilders/easybuild-easyblocks#4096

[boegel]

@migueldiascosta Is it worth still merging this now that we have a newer version merged? See

• {math}[GCCcore/13.3.0] ArmComputeLibrary v25.02 #25566

[Flamefire]

@boegel That one is for a different toolchain. Do you suggest to update the version for the toolchain used in this PR?

[migueldiascosta]

I'm ok with closing this PR, since we're not likely to enable ACL for PyTorch/2.3-foss-2023b (the one this was originally targeted at)

[Flamefire]

Why not? Given the huge performance difference I'd actually update all PyTorch easyconfigs to use imkl on x68 and ACL for Arm maybe starting at 2023a, as 2022b is the oldest active one

As for versions I'd use the PYPI PyTorch packages as reference

[migueldiascosta]

@Flamefire just thought we would likely not bother. let me fix the shared library extension in this PR then, same as in the merged one

[github-actions]

Updated software ArmComputeLibrary-23.08-GCCcore-13.2.0.eb

Diff against ArmComputeLibrary-25.02-GCCcore-13.3.0.eb

easybuild/easyconfigs/a/ArmComputeLibrary/ArmComputeLibrary-25.02-GCCcore-13.3.0.eb

diff --git a/easybuild/easyconfigs/a/ArmComputeLibrary/ArmComputeLibrary-25.02-GCCcore-13.3.0.eb b/easybuild/easyconfigs/a/ArmComputeLibrary/ArmComputeLibrary-23.08-GCCcore-13.2.0.eb
index 0c178b7c0b..8a0c0b24dd 100644
--- a/easybuild/easyconfigs/a/ArmComputeLibrary/ArmComputeLibrary-25.02-GCCcore-13.3.0.eb
+++ b/easybuild/easyconfigs/a/ArmComputeLibrary/ArmComputeLibrary-23.08-GCCcore-13.2.0.eb
@@ -1,21 +1,21 @@
 easyblock = 'SCons'
 
 name = 'ArmComputeLibrary'
-version = '25.02'
+version = '23.08'
 
 homepage = 'https://github.com/ARM-software/ComputeLibrary'
 description = """The Arm Compute Library is a collection of low-level machine learning functions optimized for
  Arm® Cortex®-A, Arm® Neoverse® and Arm® Mali™ GPUs architectures."""
 
-toolchain = {'name': 'GCCcore', 'version': '13.3.0'}
+toolchain = {'name': 'GCCcore', 'version': '13.2.0'}
 
 source_urls = ['https://github.com/ARM-software/ComputeLibrary/archive/refs/tags/']
 sources = ['v%(version)s.tar.gz']
-checksums = ['339376cd05b5efe83a3909333956d7663022f0dd8c7977a35e04b35551546be6']
+checksums = ['62f514a555409d4401e5250b290cdf8cf1676e4eb775e5bd61ea6a740a8ce24f']
 
 builddependencies = [
-    ('binutils', '2.42'),
-    ('SCons', '4.9.0'),
+    ('binutils', '2.40'),
+    ('SCons', '4.6.0'),
 ]
 
 prefix_arg = 'install_dir='

[Flamefire]

I expect that less tests rather than more will fail so changing those ECs will be little work with high gain, which makes it worth going back as far as reasonably possible.
Are you going to create PRs for ACL in the other toolchains? I.e. 12.3, 14.2, 14.3. Then I'll open the PRs to add them to PyTorch

[migueldiascosta]

I can create those ACL PRs, yes. For PyTorch-2.1.2-foss-2023a.eb (GCCcore 12.3.0) though, not sure which ACL version to use, there was no .ci/docker/common/install_acl.sh on that version of PyTorch, any suggestions?

[migueldiascosta]

from https://github.com/pytorch/pytorch/blob/v2.1.2/cmake/public/ComputeLibrary.cmake looks like anything higher than ACL 21.02 should be ok, but probably safer to use exactly ACL 21.02 for PyTorch 2.1.2

[Flamefire]

I found a way: Extract the wheel and run strings torch.libs/libarm_compute-*.so | grep arm_compute_version

• 2.3.0: arm_compute_version=v23.08
• 2.1.2: arm_compute_version=v23.05.1