The Nvidia Grace platform has a cacheline size with 64B, unlike the 128B size on other arm64 generic platforms. In the meson file, the cacheline size is missing to be defined explicitly. So it will choose 128 for building and it will impact the performance.
In the ACM whitepaper, "The NVIDIA Grace CPU implements 72 Neoverse V2 Cores [16] with a 64-B cache line size". Reference link: https://dl.acm.org/doi/10.1145/3723851.3723853 Fixes: d007038c0121 ("config/arm: add NVIDIA Grace CPU") Cc: [email protected] Cc: [email protected] Signed-off-by: Bing Zhao <[email protected]> --- config/arm/meson.build | 3 +++ 1 file changed, 3 insertions(+) diff --git a/config/arm/meson.build b/config/arm/meson.build index 523b0fc0ed..f7eac7da0a 100644 --- a/config/arm/meson.build +++ b/config/arm/meson.build @@ -535,6 +535,9 @@ soc_grace = { 'implementer': '0x41', 'part_number': '0xd4f', 'extra_march_features': ['crypto'], + 'flags': [ + ['RTE_CACHE_LINE_SIZE', '"64"'], + ], 'numa': true } -- 2.34.1
