Skip to content

Introduce SIMD for arm64#198

Merged
pjbgf merged 5 commits into
mainfrom
arm
Sep 4, 2025
Merged

Introduce SIMD for arm64#198
pjbgf merged 5 commits into
mainfrom
arm

Conversation

@pjbgf

@pjbgf pjbgf commented Aug 19, 2025 •
edited
Loading

Copy link
Copy Markdown
Owner

Introduces a SIMD implementation for the arm64 architecture that is largely based on ARM64 SHA1 and aligns with upstream Go. Extensive comments were added to make it easier to maintain the code going forwards.

The previous non-SIMD implementation was removed, as the SHA1 feature is fairly well supported across modern ARM64 hardware. Keeping both versions in place is not worth the additional complexity.

The check for SHA1 feature is based on github.com/klauspost/cpuid/v2. This is a temporary dependency which will be removed once Go exports the same functionality.

The new arm64 native implementation is ~30% faster and can process ~50% more data, performing better than the cgo alternative:

pkg: github.com/pjbgf/sha1cd/test
                                   │     /tmp/before     │                 /tmp/after                 │
                                   │       sec/op        │       sec/op         vs base                │
CalculateDvMask/generic-2            0.0000005500n ± 64%   0.0000009000n ± 22%  +63.64% (p=0.003 n=10)
CalculateDvMask/native-2             0.0000008500n ± 18%   0.0000009000n ± 22%        ~ (p=1.000 n=10)
CalculateDvMask/cgo-2                 0.000001950n ± 28%    0.000002200n ± 73%        ~ (p=0.170 n=10)
Hash8Bytes/sha1-2                           105.8n ±  5%          105.2n ±  0%   -0.57% (p=0.002 n=10)
Hash8Bytes/sha1cd_native-2                  426.4n ±  0%          298.0n ±  0%  -30.12% (p=0.000 n=10)
Hash8Bytes/sha1cd_generic-2                 539.0n ±  1%          538.5n ±  1%        ~ (p=0.171 n=10)
Hash8Bytes/sha1cd_cgo-2                     1.671µ ±  3%          1.677µ ±  8%        ~ (p=0.955 n=10)
Hash320Bytes/sha1-2                         317.4n ±  1%          318.0n ±  0%        ~ (p=0.323 n=10)
Hash320Bytes/sha1cd_native-2                2.257µ ±  0%          1.483µ ±  1%  -34.29% (p=0.000 n=10)
Hash320Bytes/sha1cd_generic-2               2.938µ ±  1%          2.943µ ±  1%        ~ (p=0.238 n=10)
Hash320Bytes/sha1cd_cgo-2                   3.103µ ±  2%          3.038µ ±  2%   -2.10% (p=0.009 n=10)
Hash1K/sha1-2                               786.1n ±  1%          785.8n ±  1%        ~ (p=1.000 n=10)
Hash1K/sha1cd_native-2                      6.255µ ±  1%          4.033µ ±  0%  -35.53% (p=0.000 n=10)
Hash1K/sha1cd_generic-2                     8.194µ ±  1%          8.218µ ±  0%        ~ (p=0.119 n=10)
Hash1K/sha1cd_cgo-2                         5.856µ ±  1%          5.809µ ±  2%        ~ (p=0.089 n=10)
Hash8K/sha1-2                               5.744µ ±  1%          5.764µ ±  1%        ~ (p=0.739 n=10)
Hash8K/sha1cd_native-2                      48.23µ ±  1%          31.41µ ±  1%  -34.88% (p=0.000 n=10)
Hash8K/sha1cd_generic-2                     63.07µ ±  1%          63.05µ ±  0%        ~ (p=0.481 n=10)
Hash8K/sha1cd_cgo-2                         34.55µ ±  2%          33.83µ ±  3%        ~ (p=0.105 n=10)
HashWithCollision/sha1cd_native-2           7.518µ ±  1%          6.115µ ±  0%  -18.66% (p=0.000 n=10)
HashWithCollision/sha1cd_generic-2          8.793µ ±  1%          8.782µ ±  0%        ~ (p=0.171 n=10)
HashWithCollision/sha1cd_cgo-2              6.591µ ±  1%          6.467µ ±  1%   -1.87% (p=0.004 n=10)
geomean                                     171.9n                162.4n         -5.51%

                                   │ /tmp/before  │              /tmp/after              │
                                   │     B/s      │      B/s       vs base                │
Hash8Bytes/sha1-2                    72.13Mi ± 5%    72.54Mi ± 0%   +0.56% (p=0.001 n=10)
Hash8Bytes/sha1cd_native-2           17.89Mi ± 0%    25.60Mi ± 0%  +43.10% (p=0.000 n=10)
Hash8Bytes/sha1cd_generic-2          14.15Mi ± 1%    14.17Mi ± 1%        ~ (p=0.202 n=10)
Hash8Bytes/sha1cd_cgo-2              4.568Mi ± 3%    4.554Mi ± 8%        ~ (p=0.926 n=10)
Hash320Bytes/sha1-2                  961.3Mi ± 1%    959.6Mi ± 0%        ~ (p=0.247 n=10)
Hash320Bytes/sha1cd_native-2         135.2Mi ± 0%    205.8Mi ± 1%  +52.23% (p=0.000 n=10)
Hash320Bytes/sha1cd_generic-2        103.8Mi ± 1%    103.7Mi ± 1%        ~ (p=0.255 n=10)
Hash320Bytes/sha1cd_cgo-2            98.35Mi ± 2%   100.46Mi ± 2%   +2.15% (p=0.010 n=10)
Hash1K/sha1-2                        1.213Gi ± 1%    1.214Gi ± 1%        ~ (p=1.000 n=10)
Hash1K/sha1cd_native-2               156.1Mi ± 1%    242.2Mi ± 0%  +55.12% (p=0.000 n=10)
Hash1K/sha1cd_generic-2              119.2Mi ± 1%    118.8Mi ± 0%        ~ (p=0.117 n=10)
Hash1K/sha1cd_cgo-2                  166.8Mi ± 1%    168.1Mi ± 2%        ~ (p=0.089 n=10)
Hash8K/sha1-2                        1.328Gi ± 1%    1.324Gi ± 1%        ~ (p=0.739 n=10)
Hash8K/sha1cd_native-2               162.0Mi ± 1%    248.8Mi ± 1%  +53.56% (p=0.000 n=10)
Hash8K/sha1cd_generic-2              123.9Mi ± 1%    123.9Mi ± 0%        ~ (p=0.481 n=10)
Hash8K/sha1cd_cgo-2                  226.1Mi ± 2%    230.9Mi ± 3%        ~ (p=0.109 n=10)
HashWithCollision/sha1cd_native-2    81.19Mi ± 1%    99.83Mi ± 0%  +22.96% (p=0.000 n=10)
HashWithCollision/sha1cd_generic-2   69.42Mi ± 1%    69.50Mi ± 0%        ~ (p=0.196 n=10)
HashWithCollision/sha1cd_cgo-2       92.60Mi ± 1%    94.38Mi ± 1%   +1.92% (p=0.004 n=10)
geomean                              114.8Mi         127.0Mi       +10.62%

pjbgf added 3 commits August 19, 2025 13:45
@pjbgf
cgo: Align digest creation with purego implementation
2692cee 
Signed-off-by: Paulo Gomes <[email protected]>
@pjbgf
ubc: Add noescape and nosplit directives
88099dc 
The native implementation for the DV mask calculation were missing
the noescape directives. For further optimisation, the wrapping funcs
are now marked with nosplit.

Signed-off-by: Paulo Gomes <[email protected]>
@pjbgf
Run golden tests on generic and native implementations
6fb6277 
Signed-off-by: Paulo Gomes <[email protected]>
@pjbgf pjbgf mentioned this pull request Aug 19, 2025
Closed
@pjbgf
arm64: Introduce SIMD implementation
8792028 
Introduces a SIMD implementation for the arm64 architecture that is largely
based on ARM64 SHA1 and aligns with upstream Go. Extensive comments were added
to make it easier to maintain the code going forwards.

The previous non-SIMD implementation was removed, as the SHA1 feature is fairly
well supported across modern ARM64 hardware. Keeping both versions in place
is not worth the additional complexity.

The check for SHA1 feature is based on github.com/klauspost/cpuid/v2. This is a
temporary dependency which will be removed once Go exports the same functionality.

Signed-off-by: Paulo Gomes <[email protected]>
@pjbgf pjbgf merged commit f052d33 into main Sep 4, 2025
13 checks passed
@pjbgf pjbgf deleted the arm branch September 4, 2025 08:44
@pjbgf pjbgf mentioned this pull request Sep 5, 2025
Closed
@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

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

1 participant

@pjbgf