Inline rounding.c and reduce.c by mkannwischer · Pull Request #410 · pq-code-package/mldsa-native

mkannwischer · 2025-08-04T04:41:29Z

Resolves Consider inlining reduce.c #217

With constant-time hardening in place, we should be able to safely inline reduce.c and rounding.c.
This should vastly improve non-LTO performance. As CI is using LTO, I do not expect much improvement there.

I performed benchmarks on Apple M1 (using ./scripts/tests bench -c MAC -r, clang16, i.e., no LTO):

Performance Results (no_opt)

ML-DSA-44

Version	Keypair	Sign	Verify
Main	133,231	600,444	159,982
Commit 1	130,432	595,679	157,727
Commit 2	117,094	463,152	138,418

ML-DSA-65

Version	Keypair	Sign	Verify
Main	224,854	985,587	251,728
Commit 1	219,835	975,350	248,134
Commit 2	200,026	749,140	219,666

ML-DSA-87

Version	Keypair	Sign	Verify
Main	366,192	1,208,165	403,781
Commit 1	360,544	1,195,145	398,647
Commit 2	330,288	932,454	357,744

Commit 1 vs Main (inline rounding.c)

Algorithm	Keypair	Sign	Verify
ML-DSA-44	1.02x	1.01x	1.01x
ML-DSA-65	1.02x	1.01x	1.01x
ML-DSA-87	1.02x	1.01x	1.01x

Commit 2 vs Main (inline reduce.c)

Algorithm	Keypair	Sign	Verify
ML-DSA-44	1.14x	1.30x	1.16x
ML-DSA-65	1.12x	1.32x	1.15x
ML-DSA-87	1.11x	1.30x	1.13x

Commit 2 vs Commit 1 (cumulative improvement)

Algorithm	Keypair	Sign	Verify
ML-DSA-44	1.11x	1.29x	1.14x
ML-DSA-65	1.10x	1.30x	1.13x
ML-DSA-87	1.09x	1.28x	1.11x

Performance Results (opt)

ML-DSA-44

Version	Keypair	Sign	Verify
Main	59,570	295,524	84,262
Commit 1	56,699	281,187	80,827
Commit 2	50,752	222,915	73,456

ML-DSA-65

Version	Keypair	Sign	Verify
Main	102,171	475,945	129,972
Commit 1	98,002	466,443	126,838
Commit 2	87,869	357,814	113,948

ML-DSA-87

Version	Keypair	Sign	Verify
Main	162,700	569,587	200,068
Commit 1	157,204	558,493	195,431
Commit 2	140,383	427,560	175,672

Commit 1 vs Main (inline rounding.c)

Algorithm	Keypair	Sign	Verify
ML-DSA-44	1.05x	1.05x	1.04x
ML-DSA-65	1.04x	1.02x	1.02x
ML-DSA-87	1.03x	1.02x	1.02x

Commit 2 vs Main (inline reduce.c)

Algorithm	Keypair	Sign	Verify
ML-DSA-44	1.17x	1.33x	1.15x
ML-DSA-65	1.16x	1.33x	1.14x
ML-DSA-87	1.16x	1.33x	1.14x

Commit 2 vs Commit 1 (cumulative improvement)

Algorithm	Keypair	Sign	Verify
ML-DSA-44	1.12x	1.26x	1.10x
ML-DSA-65	1.12x	1.30x	1.11x
ML-DSA-87	1.12x	1.31x	1.11x

This also closes the gap of LTO vs no-LTO - now LTO gives less than 2% speed-up.
Comparing the second commit from above with LTO on Apple M1 (using ./scripts/tests bench -c MAC -r --cflags="-flto":

Configuration	ML-DSA-44	ML-DSA-65	ML-DSA-87
Keypair cycles (avg)
No Opt
No LTO	117,094	200,026	330,288
LTO	116,422	198,518	325,527
LTO improvement	0.6% better	0.8% better	1.4% better
Opt
No LTO	50,752	87,869	140,383
LTO	50,404	87,459	139,564
LTO improvement	0.7% better	0.5% better	0.6% better
Sign cycles (avg)
No Opt
No LTO	463,152	749,140	932,454
LTO	461,629	747,895	931,065
LTO improvement	0.3% better	0.2% better	0.1% better
Opt
No LTO	222,915	357,814	427,560
LTO	222,269	357,119	427,335
LTO improvement	0.3% better	0.2% better	0.1% better
Verify cycles (avg)
No Opt
No LTO	138,418	219,666	357,744
LTO	137,450	217,326	353,520
LTO improvement	0.7% better	1.1% better	1.2% better
Opt
No LTO	73,456	113,948	175,672
LTO	72,874	112,457	172,648
LTO improvement	0.8% better	1.3% better	1.7% better

github-actions

Mac Mini (M1, 2020) benchmarks (opt)

Details

Benchmark suite	Current: `d0da416`	Previous: `108d2d3`	Ratio
`ML-DSA-44 keypair`	`50490` cycles	`50488` cycles	`1.00`
`ML-DSA-44 sign`	`222972` cycles	`221550` cycles	`1.01`
`ML-DSA-44 verify`	`72845` cycles	`72861` cycles	`1.00`
`ML-DSA-65 keypair`	`87373` cycles	`87430` cycles	`1.00`
`ML-DSA-65 sign`	`356131` cycles	`355968` cycles	`1.00`
`ML-DSA-65 verify`	`112686` cycles	`112701` cycles	`1.00`
`ML-DSA-87 keypair`	`140128` cycles	`139983` cycles	`1.00`
`ML-DSA-87 sign`	`425731` cycles	`425589` cycles	`1.00`
`ML-DSA-87 verify`	`173211` cycles	`173161` cycles	`1.00`

This comment was automatically generated by workflow using github-action-benchmark.

github-actions

Mac Mini (M1, 2020) benchmarks (no-opt)

Details

Benchmark suite	Current: `d0da416`	Previous: `108d2d3`	Ratio
`ML-DSA-44 keypair`	`116006` cycles	`116048` cycles	`1.00`
`ML-DSA-44 sign`	`455013` cycles	`453867` cycles	`1.00`
`ML-DSA-44 verify`	`136874` cycles	`136916` cycles	`1.00`
`ML-DSA-65 keypair`	`197992` cycles	`197974` cycles	`1.00`
`ML-DSA-65 sign`	`733195` cycles	`733093` cycles	`1.00`
`ML-DSA-65 verify`	`216797` cycles	`216913` cycles	`1.00`
`ML-DSA-87 keypair`	`335075` cycles	`325071` cycles	`1.03`
`ML-DSA-87 sign`	`915069` cycles	`914947` cycles	`1.00`
`ML-DSA-87 verify`	`353281` cycles	`353151` cycles	`1.00`

This comment was automatically generated by workflow using github-action-benchmark.

github-actions

⚠️ Performance Alert ⚠️

Possible performance regression was detected for benchmark 'Mac Mini (M1, 2020) benchmarks (no-opt)'.
Benchmark result of this commit is worse than the previous benchmark result exceeding threshold 1.03.

Benchmark suite	Current: `d0da416`	Previous: `108d2d3`	Ratio
`ML-DSA-87 keypair`	`335075` cycles	`325071` cycles	`1.03`

This comment was automatically generated by workflow using github-action-benchmark.

oqs-bot

Intel Xeon 4th gen (c7i)

Details

Benchmark suite	Current: `d0da416`	Previous: `108d2d3`	Ratio
`ML-DSA-44 keypair`	`37475` cycles	`37315` cycles	`1.00`
`ML-DSA-44 sign`	`169677` cycles	`169696` cycles	`1.00`
`ML-DSA-44 verify`	`50158` cycles	`50120` cycles	`1.00`
`ML-DSA-65 keypair`	`66785` cycles	`66560` cycles	`1.00`
`ML-DSA-65 sign`	`280070` cycles	`280908` cycles	`1.00`
`ML-DSA-65 verify`	`78663` cycles	`78493` cycles	`1.00`
`ML-DSA-87 keypair`	`100811` cycles	`101215` cycles	`1.00`
`ML-DSA-87 sign`	`326631` cycles	`326489` cycles	`1.00`
`ML-DSA-87 verify`	`116185` cycles	`117806` cycles	`0.99`

This comment was automatically generated by workflow using github-action-benchmark.

oqs-bot

Intel Xeon 3rd gen (c6i)

Details

Benchmark suite	Current: `d0da416`	Previous: `108d2d3`	Ratio
`ML-DSA-44 keypair`	`61097` cycles	`61319` cycles	`1.00`
`ML-DSA-44 sign`	`264587` cycles	`265061` cycles	`1.00`
`ML-DSA-44 verify`	`80366` cycles	`80809` cycles	`0.99`
`ML-DSA-65 keypair`	`106800` cycles	`106734` cycles	`1.00`
`ML-DSA-65 sign`	`436283` cycles	`436028` cycles	`1.00`
`ML-DSA-65 verify`	`127404` cycles	`127309` cycles	`1.00`
`ML-DSA-87 keypair`	`165529` cycles	`165748` cycles	`1.00`
`ML-DSA-87 sign`	`518220` cycles	`515316` cycles	`1.01`
`ML-DSA-87 verify`	`192218` cycles	`191132` cycles	`1.01`

This comment was automatically generated by workflow using github-action-benchmark.

oqs-bot

Intel Xeon 4th gen (c7i) (no-opt)

Details

Benchmark suite	Current: `d0da416`	Previous: `108d2d3`	Ratio
`ML-DSA-44 keypair`	`96141` cycles	`96567` cycles	`1.00`
`ML-DSA-44 sign`	`351718` cycles	`352042` cycles	`1.00`
`ML-DSA-44 verify`	`105171` cycles	`105507` cycles	`1.00`
`ML-DSA-65 keypair`	`163839` cycles	`164849` cycles	`0.99`
`ML-DSA-65 sign`	`579619` cycles	`578557` cycles	`1.00`
`ML-DSA-65 verify`	`169522` cycles	`170920` cycles	`0.99`
`ML-DSA-87 keypair`	`273631` cycles	`275791` cycles	`0.99`
`ML-DSA-87 sign`	`734824` cycles	`732477` cycles	`1.00`
`ML-DSA-87 verify`	`281021` cycles	`281986` cycles	`1.00`

This comment was automatically generated by workflow using github-action-benchmark.

oqs-bot

AMD EPYC 3rd gen (c6a)

Details

Benchmark suite	Current: `d0da416`	Previous: `108d2d3`	Ratio
`ML-DSA-44 keypair`	`73968` cycles	`77878` cycles	`0.95`
`ML-DSA-44 sign`	`269536` cycles	`281354` cycles	`0.96`
`ML-DSA-44 verify`	`88927` cycles	`93507` cycles	`0.95`
`ML-DSA-65 keypair`	`128959` cycles	`126862` cycles	`1.02`
`ML-DSA-65 sign`	`440114` cycles	`436323` cycles	`1.01`
`ML-DSA-65 verify`	`144544` cycles	`143280` cycles	`1.01`
`ML-DSA-87 keypair`	`210813` cycles	`210856` cycles	`1.00`
`ML-DSA-87 sign`	`544841` cycles	`544492` cycles	`1.00`
`ML-DSA-87 verify`	`230027` cycles	`230070` cycles	`1.00`

This comment was automatically generated by workflow using github-action-benchmark.

oqs-bot

Intel Xeon 3rd gen (c6i) (no-opt)

Details

Benchmark suite	Current: `d0da416`	Previous: `108d2d3`	Ratio
`ML-DSA-44 keypair`	`159236` cycles	`159603` cycles	`1.00`
`ML-DSA-44 sign`	`576240` cycles	`574646` cycles	`1.00`
`ML-DSA-44 verify`	`175523` cycles	`175554` cycles	`1.00`
`ML-DSA-65 keypair`	`272628` cycles	`272349` cycles	`1.00`
`ML-DSA-65 sign`	`948903` cycles	`939166` cycles	`1.01`
`ML-DSA-65 verify`	`283749` cycles	`283581` cycles	`1.00`
`ML-DSA-87 keypair`	`453937` cycles	`452169` cycles	`1.00`
`ML-DSA-87 sign`	`1197536` cycles	`1185848` cycles	`1.01`
`ML-DSA-87 verify`	`471800` cycles	`469407` cycles	`1.01`

This comment was automatically generated by workflow using github-action-benchmark.

oqs-bot

AMD EPYC 4th gen (c7a)

Details

Benchmark suite	Current: `d0da416`	Previous: `108d2d3`	Ratio
`ML-DSA-44 keypair`	`44597` cycles	`44555` cycles	`1.00`
`ML-DSA-44 sign`	`197066` cycles	`196360` cycles	`1.00`
`ML-DSA-44 verify`	`60229` cycles	`60162` cycles	`1.00`
`ML-DSA-65 keypair`	`78871` cycles	`76125` cycles	`1.04`
`ML-DSA-65 sign`	`331848` cycles	`318170` cycles	`1.04`
`ML-DSA-65 verify`	`99320` cycles	`93458` cycles	`1.06`
`ML-DSA-87 keypair`	`115528` cycles	`115550` cycles	`1.00`
`ML-DSA-87 sign`	`368922` cycles	`367938` cycles	`1.00`
`ML-DSA-87 verify`	`138832` cycles	`136834` cycles	`1.01`

This comment was automatically generated by workflow using github-action-benchmark.

oqs-bot

⚠️ Performance Alert ⚠️

Possible performance regression was detected for benchmark 'AMD EPYC 4th gen (c7a)'.
Benchmark result of this commit is worse than the previous benchmark result exceeding threshold 1.03.

Benchmark suite	Current: `d0da416`	Previous: `108d2d3`	Ratio
`ML-DSA-65 keypair`	`78871` cycles	`76125` cycles	`1.04`
`ML-DSA-65 sign`	`331848` cycles	`318170` cycles	`1.04`
`ML-DSA-65 verify`	`99320` cycles	`93458` cycles	`1.06`

This comment was automatically generated by workflow using github-action-benchmark.

oqs-bot

AMD EPYC 3rd gen (c6a) (no-opt)

Details

Benchmark suite	Current: `d0da416`	Previous: `108d2d3`	Ratio
`ML-DSA-44 keypair`	`136592` cycles	`136387` cycles	`1.00`
`ML-DSA-44 sign`	`553396` cycles	`552602` cycles	`1.00`
`ML-DSA-44 verify`	`154957` cycles	`154968` cycles	`1.00`
`ML-DSA-65 keypair`	`227998` cycles	`228515` cycles	`1.00`
`ML-DSA-65 sign`	`896453` cycles	`892614` cycles	`1.00`
`ML-DSA-65 verify`	`243881` cycles	`243755` cycles	`1.00`
`ML-DSA-87 keypair`	`376788` cycles	`376213` cycles	`1.00`
`ML-DSA-87 sign`	`1116492` cycles	`1119589` cycles	`1.00`
`ML-DSA-87 verify`	`398522` cycles	`397698` cycles	`1.00`

This comment was automatically generated by workflow using github-action-benchmark.

oqs-bot

Graviton4

Details

Benchmark suite	Current: `d0da416`	Previous: `108d2d3`	Ratio
`ML-DSA-44 keypair`	`73176` cycles	`73120` cycles	`1.00`
`ML-DSA-44 sign`	`282505` cycles	`282739` cycles	`1.00`
`ML-DSA-44 verify`	`87069` cycles	`87104` cycles	`1.00`
`ML-DSA-65 keypair`	`128592` cycles	`128241` cycles	`1.00`
`ML-DSA-65 sign`	`460852` cycles	`460971` cycles	`1.00`
`ML-DSA-65 verify`	`139113` cycles	`138989` cycles	`1.00`
`ML-DSA-87 keypair`	`208290` cycles	`207784` cycles	`1.00`
`ML-DSA-87 sign`	`564264` cycles	`563219` cycles	`1.00`
`ML-DSA-87 verify`	`222990` cycles	`222452` cycles	`1.00`

This comment was automatically generated by workflow using github-action-benchmark.

oqs-bot

AMD EPYC 4th gen (c7a) (no-opt)

Details

Benchmark suite	Current: `d0da416`	Previous: `108d2d3`	Ratio
`ML-DSA-44 keypair`	`121280` cycles	`121795` cycles	`1.00`
`ML-DSA-44 sign`	`463697` cycles	`463017` cycles	`1.00`
`ML-DSA-44 verify`	`137500` cycles	`137430` cycles	`1.00`
`ML-DSA-65 keypair`	`206358` cycles	`206690` cycles	`1.00`
`ML-DSA-65 sign`	`753174` cycles	`745788` cycles	`1.01`
`ML-DSA-65 verify`	`216373` cycles	`216570` cycles	`1.00`
`ML-DSA-87 keypair`	`340934` cycles	`339701` cycles	`1.00`
`ML-DSA-87 sign`	`956008` cycles	`948690` cycles	`1.01`
`ML-DSA-87 verify`	`357881` cycles	`355383` cycles	`1.01`

This comment was automatically generated by workflow using github-action-benchmark.

oqs-bot

Graviton3

Details

Benchmark suite	Current: `d0da416`	Previous: `108d2d3`	Ratio
`ML-DSA-44 keypair`	`78065` cycles	`77998` cycles	`1.00`
`ML-DSA-44 sign`	`303838` cycles	`304214` cycles	`1.00`
`ML-DSA-44 verify`	`95596` cycles	`95550` cycles	`1.00`
`ML-DSA-65 keypair`	`135065` cycles	`134987` cycles	`1.00`
`ML-DSA-65 sign`	`496942` cycles	`496900` cycles	`1.00`
`ML-DSA-65 verify`	`151295` cycles	`151237` cycles	`1.00`
`ML-DSA-87 keypair`	`217777` cycles	`217475` cycles	`1.00`
`ML-DSA-87 sign`	`606834` cycles	`606659` cycles	`1.00`
`ML-DSA-87 verify`	`239581` cycles	`239653` cycles	`1.00`

This comment was automatically generated by workflow using github-action-benchmark.

oqs-bot

Graviton4 (no-opt)

Details

Benchmark suite	Current: `d0da416`	Previous: `108d2d3`	Ratio
`ML-DSA-44 keypair`	`134654` cycles	`134693` cycles	`1.00`
`ML-DSA-44 sign`	`520087` cycles	`508601` cycles	`1.02`
`ML-DSA-44 verify`	`149526` cycles	`149568` cycles	`1.00`
`ML-DSA-65 keypair`	`228421` cycles	`228621` cycles	`1.00`
`ML-DSA-65 sign`	`824504` cycles	`819792` cycles	`1.01`
`ML-DSA-65 verify`	`237394` cycles	`237589` cycles	`1.00`
`ML-DSA-87 keypair`	`377029` cycles	`377513` cycles	`1.00`
`ML-DSA-87 sign`	`1029694` cycles	`1030352` cycles	`1.00`
`ML-DSA-87 verify`	`391025` cycles	`390819` cycles	`1.00`

This comment was automatically generated by workflow using github-action-benchmark.

oqs-bot

Graviton2

Details

Benchmark suite	Current: `d0da416`	Previous: `108d2d3`	Ratio
`ML-DSA-44 keypair`	`120565` cycles	`120764` cycles	`1.00`
`ML-DSA-44 sign`	`488358` cycles	`488941` cycles	`1.00`
`ML-DSA-44 verify`	`145830` cycles	`145977` cycles	`1.00`
`ML-DSA-65 keypair`	`207541` cycles	`207520` cycles	`1.00`
`ML-DSA-65 sign`	`801841` cycles	`802514` cycles	`1.00`
`ML-DSA-65 verify`	`232008` cycles	`231762` cycles	`1.00`
`ML-DSA-87 keypair`	`336864` cycles	`336645` cycles	`1.00`
`ML-DSA-87 sign`	`986735` cycles	`986284` cycles	`1.00`
`ML-DSA-87 verify`	`369607` cycles	`370182` cycles	`1.00`

This comment was automatically generated by workflow using github-action-benchmark.

oqs-bot

Graviton3 (no-opt)

Details

Benchmark suite	Current: `d0da416`	Previous: `108d2d3`	Ratio
`ML-DSA-44 keypair`	`139834` cycles	`139752` cycles	`1.00`
`ML-DSA-44 sign`	`513319` cycles	`505561` cycles	`1.02`
`ML-DSA-44 verify`	`154555` cycles	`154400` cycles	`1.00`
`ML-DSA-65 keypair`	`245231` cycles	`244774` cycles	`1.00`
`ML-DSA-65 sign`	`824646` cycles	`822040` cycles	`1.00`
`ML-DSA-65 verify`	`248436` cycles	`248601` cycles	`1.00`
`ML-DSA-87 keypair`	`397826` cycles	`397508` cycles	`1.00`
`ML-DSA-87 sign`	`1043598` cycles	`1043535` cycles	`1.00`
`ML-DSA-87 verify`	`411641` cycles	`411198` cycles	`1.00`

This comment was automatically generated by workflow using github-action-benchmark.

oqs-bot

Graviton2 (no-opt)

Details

Benchmark suite	Current: `d0da416`	Previous: `108d2d3`	Ratio
`ML-DSA-44 keypair`	`215285` cycles	`215420` cycles	`1.00`
`ML-DSA-44 sign`	`798500` cycles	`810106` cycles	`0.99`
`ML-DSA-44 verify`	`239636` cycles	`239763` cycles	`1.00`
`ML-DSA-65 keypair`	`383817` cycles	`384187` cycles	`1.00`
`ML-DSA-65 sign`	`1313303` cycles	`1308318` cycles	`1.00`
`ML-DSA-65 verify`	`385056` cycles	`385474` cycles	`1.00`
`ML-DSA-87 keypair`	`611765` cycles	`613297` cycles	`1.00`
`ML-DSA-87 sign`	`1666280` cycles	`1667822` cycles	`1.00`
`ML-DSA-87 verify`	`637589` cycles	`638454` cycles	`1.00`

This comment was automatically generated by workflow using github-action-benchmark.

github-actions

Arm Cortex-A55 (Snapdragon 888) benchmarks (opt)

Details

Benchmark suite	Current: `d0da416`	Previous: `108d2d3`	Ratio
`ML-DSA-44 keypair`	`294317` cycles	`293781` cycles	`1.00`
`ML-DSA-44 sign`	`1225926` cycles	`1224715` cycles	`1.00`
`ML-DSA-44 verify`	`356148` cycles	`356492` cycles	`1.00`
`ML-DSA-65 keypair`	`498809` cycles	`499030` cycles	`1.00`
`ML-DSA-65 sign`	`1966811` cycles	`1968154` cycles	`1.00`
`ML-DSA-65 verify`	`554922` cycles	`555182` cycles	`1.00`
`ML-DSA-87 keypair`	`846548` cycles	`859424` cycles	`0.99`
`ML-DSA-87 sign`	`2536389` cycles	`2621075` cycles	`0.97`
`ML-DSA-87 verify`	`920026` cycles	`926631` cycles	`0.99`

This comment was automatically generated by workflow using github-action-benchmark.

github-actions

SpacemiT K1 8 (Banana Pi F3) benchmarks (no-opt)

Details

Benchmark suite	Current: `d0da416`	Previous: `108d2d3`	Ratio
`ML-DSA-44 keypair`	`828262` cycles	`946682` cycles	`0.87`
`ML-DSA-44 sign`	`3363987` cycles	`4434339` cycles	`0.76`
`ML-DSA-44 verify`	`931672` cycles	`1090557` cycles	`0.85`
`ML-DSA-65 keypair`	`1395687` cycles	`1578765` cycles	`0.88`
`ML-DSA-65 sign`	`5489358` cycles	`7340298` cycles	`0.75`
`ML-DSA-65 verify`	`1481977` cycles	`1723027` cycles	`0.86`
`ML-DSA-87 keypair`	`2314070` cycles	`2543418` cycles	`0.91`
`ML-DSA-87 sign`	`6937123` cycles	`8975747` cycles	`0.77`
`ML-DSA-87 verify`	`2444260` cycles	`2750040` cycles	`0.89`

This comment was automatically generated by workflow using github-action-benchmark.

github-actions

Arm Cortex-A76 (Raspberry Pi 5) benchmarks (opt)

Details

Benchmark suite	Current: `d0da416`	Previous: `108d2d3`	Ratio
`ML-DSA-44 keypair`	`120435` cycles	`120254` cycles	`1.00`
`ML-DSA-44 sign`	`488092` cycles	`488304` cycles	`1.00`
`ML-DSA-44 verify`	`145648` cycles	`145589` cycles	`1.00`
`ML-DSA-65 keypair`	`207583` cycles	`207315` cycles	`1.00`
`ML-DSA-65 sign`	`801742` cycles	`802722` cycles	`1.00`
`ML-DSA-65 verify`	`232036` cycles	`231653` cycles	`1.00`
`ML-DSA-87 keypair`	`336621` cycles	`336116` cycles	`1.00`
`ML-DSA-87 sign`	`985366` cycles	`985370` cycles	`1.00`
`ML-DSA-87 verify`	`369932` cycles	`370098` cycles	`1.00`

This comment was automatically generated by workflow using github-action-benchmark.

github-actions

Arm Cortex-A76 (Raspberry Pi 5) benchmarks (no-opt)

Details

Benchmark suite	Current: `d0da416`	Previous: `108d2d3`	Ratio
`ML-DSA-44 keypair`	`214982` cycles	`214931` cycles	`1.00`
`ML-DSA-44 sign`	`797243` cycles	`797190` cycles	`1.00`
`ML-DSA-44 verify`	`239466` cycles	`239916` cycles	`1.00`
`ML-DSA-65 keypair`	`383206` cycles	`384108` cycles	`1.00`
`ML-DSA-65 sign`	`1320407` cycles	`1313611` cycles	`1.01`
`ML-DSA-65 verify`	`384384` cycles	`385437` cycles	`1.00`
`ML-DSA-87 keypair`	`611054` cycles	`612519` cycles	`1.00`
`ML-DSA-87 sign`	`1663336` cycles	`1665118` cycles	`1.00`
`ML-DSA-87 verify`	`637137` cycles	`638211` cycles	`1.00`

This comment was automatically generated by workflow using github-action-benchmark.

github-actions

Arm Cortex-A55 (Snapdragon 888) benchmarks (no-opt)

Details

Benchmark suite	Current: `d0da416`	Previous: `108d2d3`	Ratio
`ML-DSA-44 keypair`	`465292` cycles	`464533` cycles	`1.00`
`ML-DSA-44 sign`	`2235730` cycles	`2239528` cycles	`1.00`
`ML-DSA-44 verify`	`560391` cycles	`560538` cycles	`1.00`
`ML-DSA-65 keypair`	`778045` cycles	`778157` cycles	`1.00`
`ML-DSA-65 sign`	`3647080` cycles	`3662068` cycles	`1.00`
`ML-DSA-65 verify`	`865702` cycles	`866024` cycles	`1.00`
`ML-DSA-87 keypair`	`1254784` cycles	`1257170` cycles	`1.00`
`ML-DSA-87 sign`	`4531602` cycles	`4516332` cycles	`1.00`
`ML-DSA-87 verify`	`1384203` cycles	`1387643` cycles	`1.00`

This comment was automatically generated by workflow using github-action-benchmark.

mkannwischer · 2025-08-04T05:11:40Z

SpacemiT K1 8 (Banana Pi F3) benchmarks (no-opt)

Benchmark suite Current: d0da416 Previous: 108d2d3 Ratio
ML-DSA-44 keypair 828262 cycles 946682 cycles 0.87
ML-DSA-44 sign 3363987 cycles 4434339 cycles 0.76
ML-DSA-44 verify 931672 cycles 1090557 cycles 0.85
ML-DSA-65 keypair 1395687 cycles 1578765 cycles 0.88
ML-DSA-65 sign 5489358 cycles 7340298 cycles 0.75
ML-DSA-65 verify 1481977 cycles 1723027 cycles 0.86
ML-DSA-87 keypair 2314070 cycles 2543418 cycles 0.91
ML-DSA-87 sign 6937123 cycles 8975747 cycles 0.77
ML-DSA-87 verify 2444260 cycles 2750040 cycles 0.89
This comment was automatically generated by workflow using github-action-benchmark.

This is very extreme. All else looks as expected.

github-actions

Arm Cortex-A72 (Raspberry Pi 4) benchmarks (opt)

Details

Benchmark suite	Current: `d0da416`	Previous: `faefbbe`	Ratio
`ML-DSA-44 keypair`	`243965` cycles	`240968` cycles	`1.01`
`ML-DSA-44 sign`	`848049` cycles	`830594` cycles	`1.02`
`ML-DSA-44 verify`	`273856` cycles	`262743` cycles	`1.04`
`ML-DSA-65 keypair`	`419643` cycles	`408068` cycles	`1.03`
`ML-DSA-65 sign`	`1354778` cycles	`1340034` cycles	`1.01`
`ML-DSA-65 verify`	`435934` cycles	`427802` cycles	`1.02`
`ML-DSA-87 keypair`	`687064` cycles	`683402` cycles	`1.01`
`ML-DSA-87 sign`	`1727385` cycles	`1740210` cycles	`0.99`
`ML-DSA-87 verify`	`696746` cycles	`702708` cycles	`0.99`

This comment was automatically generated by workflow using github-action-benchmark.

github-actions

⚠️ Performance Alert ⚠️

Possible performance regression was detected for benchmark 'Arm Cortex-A72 (Raspberry Pi 4) benchmarks (opt)'.
Benchmark result of this commit is worse than the previous benchmark result exceeding threshold 1.03.

Benchmark suite	Current: `d0da416`	Previous: `faefbbe`	Ratio
`ML-DSA-44 verify`	`273856` cycles	`262743` cycles	`1.04`

This comment was automatically generated by workflow using github-action-benchmark.

github-actions

Arm Cortex-A72 (Raspberry Pi 4) benchmarks (no-opt)

Details

Benchmark suite	Current: `d0da416`	Previous: `faefbbe`	Ratio
`ML-DSA-44 keypair`	`322104` cycles	`306602` cycles	`1.05`
`ML-DSA-44 sign`	`1259169` cycles	`1197826` cycles	`1.05`
`ML-DSA-44 verify`	`353015` cycles	`342184` cycles	`1.03`
`ML-DSA-65 keypair`	`593365` cycles	`560307` cycles	`1.06`
`ML-DSA-65 sign`	`1997127` cycles	`1972800` cycles	`1.01`
`ML-DSA-65 verify`	`560706` cycles	`546049` cycles	`1.03`
`ML-DSA-87 keypair`	`890525` cycles	`869087` cycles	`1.02`
`ML-DSA-87 sign`	`2589746` cycles	`2463008` cycles	`1.05`
`ML-DSA-87 verify`	`934872` cycles	`890530` cycles	`1.05`

This comment was automatically generated by workflow using github-action-benchmark.

github-actions

⚠️ Performance Alert ⚠️

Possible performance regression was detected for benchmark 'Arm Cortex-A72 (Raspberry Pi 4) benchmarks (no-opt)'.
Benchmark result of this commit is worse than the previous benchmark result exceeding threshold 1.03.

Benchmark suite	Current: `d0da416`	Previous: `faefbbe`	Ratio
`ML-DSA-44 keypair`	`322104` cycles	`306602` cycles	`1.05`
`ML-DSA-44 sign`	`1259169` cycles	`1197826` cycles	`1.05`
`ML-DSA-44 verify`	`353015` cycles	`342184` cycles	`1.03`
`ML-DSA-65 keypair`	`593365` cycles	`560307` cycles	`1.06`
`ML-DSA-87 sign`	`2589746` cycles	`2463008` cycles	`1.05`
`ML-DSA-87 verify`	`934872` cycles	`890530` cycles	`1.05`

This comment was automatically generated by workflow using github-action-benchmark.

mldsa/reduce.h

hanno-becker

Noted some pre-existing copy-pasta you can fix along the way if you like. Otherwise LGTM.

After the constant-time hardening has been applied, we can simply inline all functions. This should dramatically improve non-LTO performance. Signed-off-by: Matthias J. Kannwischer <matthias@kannwischer.eu>

mkannwischer mentioned this pull request Aug 4, 2025

Inline reduce.c #363

Closed

mkannwischer force-pushed the inline2 branch from 2baa7fa to d0da416 Compare August 4, 2025 04:59

mkannwischer added the benchmark label Aug 4, 2025

github-actions bot reviewed Aug 4, 2025

View reviewed changes

oqs-bot reviewed Aug 4, 2025

View reviewed changes

github-actions bot reviewed Aug 4, 2025

View reviewed changes

mkannwischer marked this pull request as ready for review August 4, 2025 05:11

mkannwischer requested a review from a team as a code owner August 4, 2025 05:11

github-actions bot reviewed Aug 4, 2025

View reviewed changes

This was referenced Aug 4, 2025

Add native implementation of poly_decompose #411

Merged

Add native implementation for poly_caddq #414

Merged

hanno-becker reviewed Aug 7, 2025

View reviewed changes

mldsa/reduce.h Outdated Show resolved Hide resolved

hanno-becker approved these changes Aug 7, 2025

View reviewed changes

mkannwischer added 2 commits August 7, 2025 14:13

Inline rounding.c

7250ed3

After the constant-time hardening has been applied, we can simply inline all functions. This should dramatically improve non-LTO performance. Signed-off-by: Matthias J. Kannwischer <matthias@kannwischer.eu>

Inline reduce.c

b945b53

After the constant-time hardening has been applied, we can simply inline all functions. This should dramatically improve non-LTO performance. Signed-off-by: Matthias J. Kannwischer <matthias@kannwischer.eu>

mkannwischer force-pushed the inline2 branch from d0da416 to b945b53 Compare August 7, 2025 06:13

mkannwischer merged commit cede4fa into main Aug 7, 2025
209 checks passed

mkannwischer deleted the inline2 branch August 7, 2025 08:11

Conversation

mkannwischer commented Aug 4, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Performance Results (no_opt)

ML-DSA-44

ML-DSA-65

ML-DSA-87

Commit 1 vs Main (inline rounding.c)

Commit 2 vs Main (inline reduce.c)

Commit 2 vs Commit 1 (cumulative improvement)

Performance Results (opt)

ML-DSA-44

ML-DSA-65

ML-DSA-87

Commit 1 vs Main (inline rounding.c)

Commit 2 vs Main (inline reduce.c)

Commit 2 vs Commit 1 (cumulative improvement)

Uh oh!

github-actions bot left a comment

Choose a reason for hiding this comment

Mac Mini (M1, 2020) benchmarks (opt)

Uh oh!

github-actions bot left a comment

Choose a reason for hiding this comment

Mac Mini (M1, 2020) benchmarks (no-opt)

Uh oh!

github-actions bot left a comment

Choose a reason for hiding this comment

⚠️ Performance Alert ⚠️

Uh oh!

oqs-bot left a comment

Choose a reason for hiding this comment

Intel Xeon 4th gen (c7i)

Uh oh!

oqs-bot left a comment

Choose a reason for hiding this comment

Intel Xeon 3rd gen (c6i)

Uh oh!

oqs-bot left a comment

Choose a reason for hiding this comment

Intel Xeon 4th gen (c7i) (no-opt)

Uh oh!

oqs-bot left a comment

Choose a reason for hiding this comment

AMD EPYC 3rd gen (c6a)

Uh oh!

oqs-bot left a comment

Choose a reason for hiding this comment

Intel Xeon 3rd gen (c6i) (no-opt)

Uh oh!

oqs-bot left a comment

Choose a reason for hiding this comment

AMD EPYC 4th gen (c7a)

Uh oh!

oqs-bot left a comment

Choose a reason for hiding this comment

⚠️ Performance Alert ⚠️

Uh oh!

oqs-bot left a comment

Choose a reason for hiding this comment

AMD EPYC 3rd gen (c6a) (no-opt)

Uh oh!

oqs-bot left a comment

Choose a reason for hiding this comment

Graviton4

Uh oh!

oqs-bot left a comment

Choose a reason for hiding this comment

AMD EPYC 4th gen (c7a) (no-opt)

Uh oh!

oqs-bot left a comment

Choose a reason for hiding this comment

Graviton3

Uh oh!

oqs-bot left a comment

Choose a reason for hiding this comment

Graviton4 (no-opt)

Uh oh!

oqs-bot left a comment

Choose a reason for hiding this comment

mkannwischer commented Aug 4, 2025 •

edited

Loading