Twice slower AES-XTS performance (compared to OpenSSL) on AMD Ryzen 7 7840HS processor

[socketpair]

I asked AI to write a performance benchamrk comparing in same conditions OpenSSL and WolfSSL.

12th Gen Intel(R) Core(TM) i7-1260P:

./bench_wolfssl 2000000
throughput: 6.810 GiB/s (7.312 GB/s)

./bench_openssl 2000000
throughput: 7.178 GiB/s (7.707 GB/s)

AMD Ryzen 7 7840HS:

./bench_wolfssl 1000000
throughput: 6.627 GiB/s (7.115 GB/s)

./bench_openssl 1000000
throughput: 12.360 GiB/s (13.271 GB/s)

Benchmark code: https://github.com/socketpair/AES-XTS-bench

[LinuxJedi]

Hi @socketpair, can you please give us the compile options you used for wolfSSL and OpenSSL? I can only see the compile options for your benchmark demo in the link.

[SparkiDev]

Hi @socketpair,

I believe that OpenSSL has AES-XTS implemented using AVX512 instructions.
We don't currently have this but it is on our roadmap for this year.

Sean

[socketpair]

OPENSSL_ia32cap="~0x200000000000" your_program gives the same speed on this AMD (as WolfSSL). So, yes, it's effect of not implemented AVX512.

[socketpair]

https://download.copr.fedorainfracloud.org/results/kni/wolfssl/fedora-43-x86_64/08818627-wolfssl/wolfssl.spec

           --disable-static              \
           --enable-all                  \
           --enable-all-crypto           \
           --disable-qt                  \
           --with-sys-crypto-policy      \
           --disable-armasm              \
           --enable-aesni

Is it enough to achieve maximum performance on x86_64?

[SparkiDev]

Hi @socketpair,

You need: --enable-intelasm to enable all Intel x64 assembly.
--enable-aesni should enable some optimisations but it is best to use the one above.
--disable-static is to enable shared library build only.
--disable-armasm is not needed as it is off by default.

The rest are for enabling and disabling features pretty much and don't matter.
--enable-aesxts will have AES-XTS built.
Disable other algorithms if you don't need them.

Sean