Skip to content

amd64: Add SHANI implementation#199

Merged
pjbgf merged 2 commits into
mainfrom
amd64-simd
Sep 5, 2025
Merged

amd64: Add SHANI implementation#199
pjbgf merged 2 commits into
mainfrom
amd64-simd

Conversation

@pjbgf

@pjbgf pjbgf commented Sep 5, 2025

Copy link
Copy Markdown
Owner

The native non-SIMD implementation provides minor improvement when compared to the Go version. Given that SHANI is widely available these days, that version is being removed to decrease codebase complexity.

This implementation is roughly 30% faster:

goos: linux
goarch: amd64
pkg: github.com/pjbgf/sha1cd/test
cpu: AMD Ryzen 7 PRO 8840HS w/ Radeon 780M Graphics
                                    │     /tmp/before      │                   /tmp/after                   │
                                    │        sec/op        │        sec/op         vs base                  │
CalculateDvMask/generic-16            0.0000002000n ± 100%
CalculateDvMask/native-16             0.0000001500n ± 167%
CalculateDvMask/cgo-16                0.0000005500n ± 100%   0.0000004500n ±  56%        ~ (p=0.586 n=10)
Hash8Bytes/sha1-16                           102.5n ±   1%          100.8n ±   3%        ~ (p=0.085 n=10)
Hash8Bytes/sha1cd_native-16                  172.7n ±   1%          122.3n ±   1%  -29.15% (p=0.000 n=10)
Hash8Bytes/sha1cd_generic-16                 206.3n ±   1%          209.0n ±   2%   +1.28% (p=0.008 n=10)
Hash8Bytes/sha1cd_cgo-16                     709.0n ±   2%          712.4n ±   1%        ~ (p=0.971 n=10)
Hash320Bytes/sha1-16                         394.9n ±   1%          399.2n ±   1%   +1.08% (p=0.034 n=10)
Hash320Bytes/sha1cd_native-16                888.7n ±   2%          563.1n ±   3%  -36.63% (p=0.000 n=10)
Hash320Bytes/sha1cd_generic-16               1.096µ ±   1%          1.065µ ±   3%   -2.83% (p=0.003 n=10)
Hash320Bytes/sha1cd_cgo-16                   1.261µ ±   2%          1.269µ ±   1%        ~ (p=0.353 n=10)
Hash1K/sha1-16                               792.9n ±   1%          798.9n ±   2%   +0.76% (p=0.007 n=10)
Hash1K/sha1cd_native-16                      2.245µ ±   3%          1.523µ ±   1%  -32.18% (p=0.000 n=10)
Hash1K/sha1cd_generic-16                     2.979µ ±   1%          2.946µ ±   2%        ~ (p=0.190 n=10)
Hash1K/sha1cd_cgo-16                         2.284µ ±   3%          2.357µ ±   2%   +3.24% (p=0.002 n=10)
Hash8K/sha1-16                               5.090µ ±   2%          5.080µ ±   1%        ~ (p=0.897 n=10)
Hash8K/sha1cd_native-16                      17.09µ ±   2%          11.43µ ±   2%  -33.09% (p=0.000 n=10)
Hash8K/sha1cd_generic-16                     22.10µ ±   1%          22.23µ ±   1%        ~ (p=0.105 n=10)
Hash8K/sha1cd_cgo-16                         12.21µ ±   1%          12.29µ ±   2%        ~ (p=0.051 n=10)
HashWithCollision/sha1cd_native-16           2.786µ ±   0%          2.194µ ±   1%  -21.25% (p=0.000 n=10)
HashWithCollision/sha1cd_generic-16          3.084µ ±   1%          3.128µ ±   3%   +1.44% (p=0.017 n=10)
HashWithCollision/sha1cd_cgo-16              2.538µ ±   4%          2.506µ ±   2%   -1.24% (p=0.016 n=10)
CalculateDvMask/go-16                                        0.0000001500n ± 100%
geomean                                      73.89n                 171.9n          -9.52%                ¹
¹ benchmark set differs from baseline; geomeans may not be comparable


                                    │ /tmp/before  │              /tmp/after              │
                                    │     B/s      │     B/s       vs base                │
Hash8Bytes/sha1-16                    74.45Mi ± 1%   75.64Mi ± 3%        ~ (p=0.089 n=10)
Hash8Bytes/sha1cd_native-16           44.18Mi ± 1%   62.35Mi ± 1%  +41.13% (p=0.000 n=10)
Hash8Bytes/sha1cd_generic-16          36.99Mi ± 1%   36.51Mi ± 2%   -1.29% (p=0.008 n=10)
Hash8Bytes/sha1cd_cgo-16              10.76Mi ± 2%   10.71Mi ± 1%        ~ (p=0.926 n=10)
Hash320Bytes/sha1-16                  772.7Mi ± 1%   764.4Mi ± 1%   -1.07% (p=0.043 n=10)
Hash320Bytes/sha1cd_native-16         343.4Mi ± 2%   542.0Mi ± 3%  +57.81% (p=0.000 n=10)
Hash320Bytes/sha1cd_generic-16        278.5Mi ± 1%   286.6Mi ± 3%   +2.90% (p=0.003 n=10)
Hash320Bytes/sha1cd_cgo-16            242.0Mi ± 2%   240.4Mi ± 2%        ~ (p=0.353 n=10)
Hash1K/sha1-16                        1.203Gi ± 1%   1.194Gi ± 1%   -0.76% (p=0.007 n=10)
Hash1K/sha1cd_native-16               435.0Mi ± 2%   641.5Mi ± 1%  +47.48% (p=0.000 n=10)
Hash1K/sha1cd_generic-16              327.8Mi ± 1%   331.5Mi ± 2%        ~ (p=0.190 n=10)
Hash1K/sha1cd_cgo-16                  427.7Mi ± 3%   414.3Mi ± 2%   -3.14% (p=0.002 n=10)
Hash8K/sha1-16                        1.499Gi ± 2%   1.502Gi ± 1%        ~ (p=0.912 n=10)
Hash8K/sha1cd_native-16               457.2Mi ± 2%   683.3Mi ± 2%  +49.45% (p=0.000 n=10)
Hash8K/sha1cd_generic-16              353.6Mi ± 1%   351.4Mi ± 1%        ~ (p=0.105 n=10)
Hash8K/sha1cd_cgo-16                  639.7Mi ± 1%   635.6Mi ± 2%        ~ (p=0.052 n=10)
HashWithCollision/sha1cd_native-16    219.1Mi ± 0%   278.3Mi ± 1%  +26.99% (p=0.000 n=10)
HashWithCollision/sha1cd_generic-16   198.0Mi ± 1%   195.1Mi ± 2%   -1.44% (p=0.023 n=10)
HashWithCollision/sha1cd_cgo-16       240.5Mi ± 4%   243.5Mi ± 2%   +1.25% (p=0.015 n=10)
geomean                               247.1Mi        271.6Mi        +9.93%

⚠️ Note that the UBC native implementations have been removed as part of this PR, and ubc.CalculateDvMaskGeneric is no longer available.

pjbgf added 2 commits September 5, 2025 17:30
@pjbgf
amd64: Introduce SHANI implementation
e239975 
The native non-SIMD implementation provides minor improvement
when compared to the Go version. Given that SHANI is widely available
these days, that version is being removed to decrease codebase complexity.

Signed-off-by: Paulo Gomes <pjbgf@linux.com>
@pjbgf
ubc: Remove native implementations
75a4461 
The initial motivation for introducing native implementation for the UBC
checks, was to act as a stepping stone on a fully native implementation of
sha1cd.

That is no longer being pursued, for two main reasons:
- The non-trivial complexity that comes with a full-on goasm version with
collision detection, which would make it harder to maintain/debug the project.
- The recently added SIMD implementations made the project reasonably fast
for the core supported architectures (arm64 and amd64).

It is important to also note that the performance gained with the native UBC
versions were minor, certainly not worth the complexity they added to the project.

Signed-off-by: Paulo Gomes <pjbgf@linux.com>
@pjbgf pjbgf added the breaking label Sep 5, 2025
@pjbgf pjbgf mentioned this pull request Sep 5, 2025
Closed
@pjbgf pjbgf merged commit 85c7a3d into main Sep 5, 2025
13 checks passed
@pjbgf pjbgf deleted the amd64-simd branch September 5, 2025 16:37
@pjbgf pjbgf mentioned this pull request May 6, 2026
Merged
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

No reviews

Assignees

No one assigned

Labels

breaking

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

1 participant

@pjbgf