Skip to content

Map Cortex-A320 to A510 kernel selection path#10173

Open
npitre wants to merge 2 commits into
google:masterfrom
npitre:xnnpack-a320
Open

Map Cortex-A320 to A510 kernel selection path#10173
npitre wants to merge 2 commits into
google:masterfrom
npitre:xnnpack-a320

Conversation

@npitre

@npitre npitre commented May 5, 2026

Copy link
Copy Markdown
Contributor

Follow-up to #10060 (Zephyr platform support).

Bumps the cpuinfo dependency to include the Cortex-A320 MIDR decode
(pytorch/cpuinfo#384, merged) and adds a one-line mapping in
cpuinfo_to_xnn_uarch() so XNNPACK selects the optimised kernel
path on Cortex-A320 cores.

The A320 is an ARMv9.2-A efficiency core in the same family as
the A510/A520. We map it to xnn_uarch_cortex_a510 since A520 is
not exposed in xnn_uarch either, and A510 is the closest existing
entry; the mapping can be revisited once silicon is available for
benchmarking.

npitre added 2 commits May 5, 2026 14:11
Bring in pytorch/cpuinfo commit 3681f0c which adds the Cortex-A320
(MIDR part 0xD8F) entry to the MIDR decode table.  This makes
cpuinfo_uarch_cortex_a320 available for use in
src/configs/hardware-config.c.

Signed-off-by: Nicolas Pitre <npitre@baylibre.com>
The Cortex-A320 is an ARMv9.2-A efficiency core in the same family
as the A510/A520.  Map it to xnn_uarch_cortex_a510 (the closest
existing entry; A520 is not exposed in xnn_uarch) so XNNPACK selects
the optimised GEMM and depthwise convolution kernels tuned for this
class of core rather than falling back to the generic path.

The mapping can be revisited once silicon is available for
benchmarking.

Signed-off-by: Nicolas Pitre <npitre@baylibre.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants