Major performance regression on older Intel CPU

[greatroar]

I got the following benchmark results for sha256-simd v0.1.1 on an Intel Core i7-3770K (Go 1.16.2, Linux):

name                speed
Hash/AVX_/8Bytes-8  34.7MB/s ± 0%
Hash/AVX_/1K-8       312MB/s ± 1%
Hash/AVX_/8K-8       332MB/s ± 0%
Hash/AVX_/1M-8       336MB/s ± 0%
Hash/AVX_/5M-8       334MB/s ± 1%
Hash/AVX_/10M-8      333MB/s ± 2%
Hash/SSSE/8Bytes-8  34.6MB/s ± 1%
Hash/SSSE/1K-8       311MB/s ± 0%
Hash/SSSE/8K-8       333MB/s ± 0%
Hash/SSSE/1M-8       334MB/s ± 0%
Hash/SSSE/5M-8       334MB/s ± 1%
Hash/SSSE/10M-8      333MB/s ± 1%
Hash/GEN_/8Bytes-8  23.8MB/s ± 1%
Hash/GEN_/1K-8       210MB/s ± 1%
Hash/GEN_/8K-8       225MB/s ± 1%
Hash/GEN_/1M-8       227MB/s ± 0%
Hash/GEN_/5M-8       225MB/s ± 1%
Hash/GEN_/10M-8      226MB/s ± 1%

Performance of the GEN version is unchanged in v1.0.0, as expected, so dropping the SSSE/AVX asm version reduces throughput by a third on this CPU. sha256-simd is no longer faster than the stdlib implementation.

Would you consider resurrecting the SSSE asm?

[harshavardhana]

You should just use standard sha256 that comes with Go @greatroar this package is not useful otherwise. The purpose was to reduce the scope of this package and advise users to migrate to stdlib.

[greatroar]

You mean this package is essentially being deprecated?

[harshavardhana]

You mean this package is essentially being deprecated?

Correct @greatroar

[greatroar]

Ok, thanks for the heads up.

[AudriusButkevicius]

I don't think it makes sense to deprecate this package.

Go standard library still does not support native SHA extensions.

Best it can do is seems to be AVX2, even for modern CPUs, which results in a roughly ~5x difference.

From syncthing:

Single thread SHA256 performance is 1857 MB/s using minio/sha256-simd (431 MB/s using crypto/sha256).

[drakkan]

Hi,

can you please better clarify the package status? A pull request (#59) was just merged and this package is still used in MinIO, thank you

[klauspost]

We are reducing the maintenance cost of this package, so it will be kept as long as there are benefits. For CPUs with SHA extensions the benefit is clear, and AVX512 can have a benefit as well, but requires special implementations.

If you want a version which detects older CPUs and use the SSE/AVX code feel free to fork it.

[drakkan]

@klauspost thank you. I'm mainly interested in SHA Extensions for x86 and ARM64 for ARM

[greatroar]

On arm64, stdlib has used the SHA extensions since 2017: golang/go@7b8a7f8.