summaryrefslogtreecommitdiffstats
diff options
context:
space:
mode:
authorJieqiang Wang <jieqiang.wang@arm.com>2023-07-04 12:05:33 +0800
committerDamjan Marion <dmarion@0xa5.net>2023-10-01 21:11:32 +0000
commitbef2d6da4ea5d67647d310a3cca5ede09385cf05 (patch)
tree0da214f4f80c74a7efd66f5cb31ffb8c03b2bef3
parente7f34c93b1c928855f88896c5bdd2bc539ee77ae (diff)
build: modify N_PREFETCH on Arm N2 to achieve best perf
From GCC 12, march=armv9-a option is supported, which includes the sve and crc options needed. Furthermore, VPP L3Fwd benchmark results on N2 based servers show that N_PREFETCH set to 6 gives the best performance. Type: feature Signed-off-by: Lijian Zhang <Lijian.Zhang@arm.com> Signed-off-by: Jieqiang Wang <jieqiang.wang@arm.com> Change-Id: I9c4fcad84d4db1189d956dabab22b26d020fbfd6
-rw-r--r--src/cmake/cpu.cmake5
1 files changed, 2 insertions, 3 deletions
diff --git a/src/cmake/cpu.cmake b/src/cmake/cpu.cmake
index f683b62b8f4..18a1210f541 100644
--- a/src/cmake/cpu.cmake
+++ b/src/cmake/cpu.cmake
@@ -194,10 +194,9 @@ elseif(CMAKE_SYSTEM_PROCESSOR MATCHES "^(aarch64.*|AARCH64.*)")
N_PREFETCHES 6
CACHE_PREFETCH_BYTES 64
)
-
add_vpp_march_variant(neoversen2
- FLAGS -march=armv8.6-a+crc+crypto+sve2 -mtune=neoverse-n2
- N_PREFETCHES 8
+ FLAGS -march=armv9-a+crypto -mtune=neoverse-n2
+ N_PREFETCHES 6
CACHE_PREFETCH_BYTES 64
OFF
)