x86_64/AArch64: Add AVX2/Neon polyw1_pack to x86_64 native backend by mkannwischer · Pull Request #973 · pq-code-package/mldsa-native

mkannwischer · 2026-02-21T07:12:14Z

Integrate polyw1_pack AVX2 implementations for both GAMMA2 variants into the native backend.

`polyw1_pack` component benchmarks

Intel Xeon 8375C (c6i.metal, no Turbo Boost, no SMT)

	ML-DSA-44	ML-DSA-65	ML-DSA-87
C	500	310	311
AVX2	216	144	144
Speedup	2.3x	2.2x	2.2x

Apple M1

	ML-DSA-44	ML-DSA-65	ML-DSA-87
C	203	49	49
AArch64	32	22	22
Speedup	6.3x	2.2x	2.2x

TODO:

Add CBMC proof
Add Unit tests
Try AArch64 implementation
Create HOL-Light + conversion to asm issues

Integrate polyw1_pack AVX2 implementations for both GAMMA2 variants into the native backend. Signed-off-by: Matthias J. Kannwischer <matthias@kannwischer.eu>

github-actions

Mac Mini (M1, 2020) benchmarks (opt)

Details

Benchmark suite	Current: `885bca1`	Previous: `0b1c536`	Ratio
`ML-DSA-44 keypair`	`45691` cycles	`45683` cycles	`1.00`
`ML-DSA-44 sign`	`128992` cycles	`131163` cycles	`0.98`
`ML-DSA-44 verify`	`47012` cycles	`47529` cycles	`0.99`
`ML-DSA-65 keypair`	`80465` cycles	`80463` cycles	`1.00`
`ML-DSA-65 sign`	`214956` cycles	`215738` cycles	`1.00`
`ML-DSA-65 verify`	`79587` cycles	`79737` cycles	`1.00`
`ML-DSA-87 keypair`	`131152` cycles	`131178` cycles	`1.00`
`ML-DSA-87 sign`	`276231` cycles	`277066` cycles	`1.00`
`ML-DSA-87 verify`	`129895` cycles	`129990` 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: `885bca1`	Previous: `0b1c536`	Ratio
`ML-DSA-44 keypair`	`112049` cycles	`111974` cycles	`1.00`
`ML-DSA-44 sign`	`403823` cycles	`403601` cycles	`1.00`
`ML-DSA-44 verify`	`119939` cycles	`119892` cycles	`1.00`
`ML-DSA-65 keypair`	`192134` cycles	`192181` cycles	`1.00`
`ML-DSA-65 sign`	`657108` cycles	`657104` cycles	`1.00`
`ML-DSA-65 verify`	`193869` cycles	`193901` cycles	`1.00`
`ML-DSA-87 keypair`	`318020` cycles	`318040` cycles	`1.00`
`ML-DSA-87 sign`	`837065` cycles	`837047` cycles	`1.00`
`ML-DSA-87 verify`	`323003` cycles	`323045` cycles	`1.00`

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

oqs-bot

Intel Xeon 4th gen (c7i)

Details

Benchmark suite	Current: `885bca1`	Previous: `0b1c536`	Ratio
`ML-DSA-44 keypair`	`34648` cycles	`34818` cycles	`1.00`
`ML-DSA-44 sign`	`115881` cycles	`119743` cycles	`0.97`
`ML-DSA-44 verify`	`37118` cycles	`38134` cycles	`0.97`
`ML-DSA-65 keypair`	`60681` cycles	`60836` cycles	`1.00`
`ML-DSA-65 sign`	`198566` cycles	`200613` cycles	`0.99`
`ML-DSA-65 verify`	`62513` cycles	`62640` cycles	`1.00`
`ML-DSA-87 keypair`	`93487` cycles	`93373` cycles	`1.00`
`ML-DSA-87 sign`	`236201` cycles	`232798` cycles	`1.01`
`ML-DSA-87 verify`	`95896` cycles	`95570` cycles	`1.00`

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

oqs-bot

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

Details

Benchmark suite	Current: `885bca1`	Previous: `0b1c536`	Ratio
`ML-DSA-44 keypair`	`93841` cycles	`93709` cycles	`1.00`
`ML-DSA-44 sign`	`333807` cycles	`332609` cycles	`1.00`
`ML-DSA-44 verify`	`99869` cycles	`99635` cycles	`1.00`
`ML-DSA-65 keypair`	`159923` cycles	`160109` cycles	`1.00`
`ML-DSA-65 sign`	`543699` cycles	`544366` cycles	`1.00`
`ML-DSA-65 verify`	`160683` cycles	`160833` cycles	`1.00`
`ML-DSA-87 keypair`	`266467` cycles	`267045` cycles	`1.00`
`ML-DSA-87 sign`	`705143` cycles	`706279` cycles	`1.00`
`ML-DSA-87 verify`	`270387` cycles	`270100` cycles	`1.00`

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: `885bca1`	Previous: `0b1c536`	Ratio
`ML-DSA-44 keypair`	`113214` cycles	`113201` cycles	`1.00`
`ML-DSA-44 sign`	`350397` cycles	`355543` cycles	`0.99`
`ML-DSA-44 verify`	`116730` cycles	`117896` cycles	`0.99`
`ML-DSA-65 keypair`	`196335` cycles	`196439` cycles	`1.00`
`ML-DSA-65 sign`	`588222` cycles	`588538` cycles	`1.00`
`ML-DSA-65 verify`	`194453` cycles	`194475` cycles	`1.00`
`ML-DSA-87 keypair`	`322439` cycles	`321909` cycles	`1.00`
`ML-DSA-87 sign`	`751213` cycles	`752725` cycles	`1.00`
`ML-DSA-87 verify`	`319951` cycles	`320145` 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: `885bca1`	Previous: `0b1c536`	Ratio
`ML-DSA-44 keypair`	`68774` cycles	`69001` cycles	`1.00`
`ML-DSA-44 sign`	`179744` cycles	`187631` cycles	`0.96`
`ML-DSA-44 verify`	`67303` cycles	`69172` cycles	`0.97`
`ML-DSA-65 keypair`	`122437` cycles	`119360` cycles	`1.03`
`ML-DSA-65 sign`	`301955` cycles	`299878` cycles	`1.01`
`ML-DSA-65 verify`	`118088` cycles	`115464` cycles	`1.02`
`ML-DSA-87 keypair`	`203390` cycles	`203890` cycles	`1.00`
`ML-DSA-87 sign`	`390522` cycles	`394779` cycles	`0.99`
`ML-DSA-87 verify`	`195089` cycles	`195702` cycles	`1.00`

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

oqs-bot

Intel Xeon 3rd gen (c6i)

Details

Benchmark suite	Current: `885bca1`	Previous: `0b1c536`	Ratio
`ML-DSA-44 keypair`	`57073` cycles	`56454` cycles	`1.01`
`ML-DSA-44 sign`	`175719` cycles	`181513` cycles	`0.97`
`ML-DSA-44 verify`	`60021` cycles	`61053` cycles	`0.98`
`ML-DSA-65 keypair`	`98321` cycles	`98631` cycles	`1.00`
`ML-DSA-65 sign`	`297249` cycles	`298535` cycles	`1.00`
`ML-DSA-65 verify`	`100096` cycles	`100069` cycles	`1.00`
`ML-DSA-87 keypair`	`152050` cycles	`152650` cycles	`1.00`
`ML-DSA-87 sign`	`353023` cycles	`355109` cycles	`0.99`
`ML-DSA-87 verify`	`152239` cycles	`152994` cycles	`1.00`

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

oqs-bot

Graviton4

Details

Benchmark suite	Current: `885bca1`	Previous: `0b1c536`	Ratio
`ML-DSA-44 keypair`	`68230` cycles	`68148` cycles	`1.00`
`ML-DSA-44 sign`	`196078` cycles	`201830` cycles	`0.97`
`ML-DSA-44 verify`	`69368` cycles	`70787` cycles	`0.98`
`ML-DSA-65 keypair`	`121376` cycles	`121099` cycles	`1.00`
`ML-DSA-65 sign`	`330550` cycles	`331249` cycles	`1.00`
`ML-DSA-65 verify`	`117750` cycles	`117837` cycles	`1.00`
`ML-DSA-87 keypair`	`198142` cycles	`197912` cycles	`1.00`
`ML-DSA-87 sign`	`426496` cycles	`426817` cycles	`1.00`
`ML-DSA-87 verify`	`194325` cycles	`194367` 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: `885bca1`	Previous: `0b1c536`	Ratio
`ML-DSA-44 keypair`	`212357` cycles	`212585` cycles	`1.00`
`ML-DSA-44 sign`	`759494` cycles	`759285` cycles	`1.00`
`ML-DSA-44 verify`	`228690` cycles	`228959` cycles	`1.00`
`ML-DSA-65 keypair`	`379923` cycles	`380251` cycles	`1.00`
`ML-DSA-65 sign`	`1252035` cycles	`1251223` cycles	`1.00`
`ML-DSA-65 verify`	`371571` cycles	`372021` cycles	`1.00`
`ML-DSA-87 keypair`	`604671` cycles	`605353` cycles	`1.00`
`ML-DSA-87 sign`	`1593513` cycles	`1591234` cycles	`1.00`
`ML-DSA-87 verify`	`618457` cycles	`617441` cycles	`1.00`

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

oqs-bot

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

Details

Benchmark suite	Current: `885bca1`	Previous: `0b1c536`	Ratio
`ML-DSA-44 keypair`	`134786` cycles	`134688` cycles	`1.00`
`ML-DSA-44 sign`	`523577` cycles	`524187` cycles	`1.00`
`ML-DSA-44 verify`	`147461` cycles	`147201` cycles	`1.00`
`ML-DSA-65 keypair`	`226346` cycles	`226675` cycles	`1.00`
`ML-DSA-65 sign`	`860567` cycles	`859973` cycles	`1.00`
`ML-DSA-65 verify`	`234837` cycles	`234911` cycles	`1.00`
`ML-DSA-87 keypair`	`372003` cycles	`370452` cycles	`1.00`
`ML-DSA-87 sign`	`1083875` cycles	`1078410` cycles	`1.01`
`ML-DSA-87 verify`	`384062` cycles	`382956` cycles	`1.00`

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

oqs-bot

AMD EPYC 4th gen (c7a)

Details

Benchmark suite	Current: `885bca1`	Previous: `0b1c536`	Ratio
`ML-DSA-44 keypair`	`42148` cycles	`41562` cycles	`1.01`
`ML-DSA-44 sign`	`129288` cycles	`133651` cycles	`0.97`
`ML-DSA-44 verify`	`43759` cycles	`44169` cycles	`0.99`
`ML-DSA-65 keypair`	`71975` cycles	`72989` cycles	`0.99`
`ML-DSA-65 sign`	`214109` cycles	`220760` cycles	`0.97`
`ML-DSA-65 verify`	`73284` cycles	`74207` cycles	`0.99`
`ML-DSA-87 keypair`	`107702` cycles	`108105` cycles	`1.00`
`ML-DSA-87 sign`	`248261` cycles	`250082` cycles	`0.99`
`ML-DSA-87 verify`	`109090` cycles	`108427` cycles	`1.01`

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

oqs-bot

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

Details

Benchmark suite	Current: `885bca1`	Previous: `0b1c536`	Ratio
`ML-DSA-44 keypair`	`157594` cycles	`157373` cycles	`1.00`
`ML-DSA-44 sign`	`549971` cycles	`549788` cycles	`1.00`
`ML-DSA-44 verify`	`169054` cycles	`169220` cycles	`1.00`
`ML-DSA-65 keypair`	`267930` cycles	`267878` cycles	`1.00`
`ML-DSA-65 sign`	`903155` cycles	`903152` cycles	`1.00`
`ML-DSA-65 verify`	`274249` cycles	`274318` cycles	`1.00`
`ML-DSA-87 keypair`	`447966` cycles	`447643` cycles	`1.00`
`ML-DSA-87 sign`	`1159788` cycles	`1157310` cycles	`1.00`
`ML-DSA-87 verify`	`457774` cycles	`457942` cycles	`1.00`

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

oqs-bot

Graviton4 (no-opt)

Details

Benchmark suite	Current: `885bca1`	Previous: `0b1c536`	Ratio
`ML-DSA-44 keypair`	`128364` cycles	`128172` cycles	`1.00`
`ML-DSA-44 sign`	`447652` cycles	`447244` cycles	`1.00`
`ML-DSA-44 verify`	`138210` cycles	`142135` cycles	`0.97`
`ML-DSA-65 keypair`	`220785` cycles	`220615` cycles	`1.00`
`ML-DSA-65 sign`	`727254` cycles	`726560` cycles	`1.00`
`ML-DSA-65 verify`	`222808` cycles	`223116` cycles	`1.00`
`ML-DSA-87 keypair`	`364610` cycles	`365048` cycles	`1.00`
`ML-DSA-87 sign`	`926038` cycles	`926588` cycles	`1.00`
`ML-DSA-87 verify`	`372875` cycles	`372428` 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: `885bca1`	Previous: `0b1c536`	Ratio
`ML-DSA-44 keypair`	`120810` cycles	`120331` cycles	`1.00`
`ML-DSA-44 sign`	`447005` cycles	`447521` cycles	`1.00`
`ML-DSA-44 verify`	`130168` cycles	`132880` cycles	`0.98`
`ML-DSA-65 keypair`	`204763` cycles	`205729` cycles	`1.00`
`ML-DSA-65 sign`	`728023` cycles	`728528` cycles	`1.00`
`ML-DSA-65 verify`	`210330` cycles	`211143` cycles	`1.00`
`ML-DSA-87 keypair`	`337390` cycles	`338699` cycles	`1.00`
`ML-DSA-87 sign`	`922663` cycles	`923705` cycles	`1.00`
`ML-DSA-87 verify`	`348278` cycles	`346629` cycles	`1.00`

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

oqs-bot

Graviton3

Details

Benchmark suite	Current: `885bca1`	Previous: `0b1c536`	Ratio
`ML-DSA-44 keypair`	`72314` cycles	`72302` cycles	`1.00`
`ML-DSA-44 sign`	`205786` cycles	`211874` cycles	`0.97`
`ML-DSA-44 verify`	`73981` cycles	`75647` cycles	`0.98`
`ML-DSA-65 keypair`	`127445` cycles	`127575` cycles	`1.00`
`ML-DSA-65 sign`	`349694` cycles	`350353` cycles	`1.00`
`ML-DSA-65 verify`	`125303` cycles	`125483` cycles	`1.00`
`ML-DSA-87 keypair`	`205809` cycles	`208020` cycles	`0.99`
`ML-DSA-87 sign`	`448791` cycles	`449002` cycles	`1.00`
`ML-DSA-87 verify`	`205264` cycles	`205683` cycles	`1.00`

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

oqs-bot

Graviton3 (no-opt)

Details

Benchmark suite	Current: `885bca1`	Previous: `0b1c536`	Ratio
`ML-DSA-44 keypair`	`138660` cycles	`138464` cycles	`1.00`
`ML-DSA-44 sign`	`484010` cycles	`484091` cycles	`1.00`
`ML-DSA-44 verify`	`148504` cycles	`156396` cycles	`0.95`
`ML-DSA-65 keypair`	`241327` cycles	`241147` cycles	`1.00`
`ML-DSA-65 sign`	`792592` cycles	`792223` cycles	`1.00`
`ML-DSA-65 verify`	`240723` cycles	`241092` cycles	`1.00`
`ML-DSA-87 keypair`	`395470` cycles	`396403` cycles	`1.00`
`ML-DSA-87 sign`	`1013125` cycles	`1012979` cycles	`1.00`
`ML-DSA-87 verify`	`402895` cycles	`402335` cycles	`1.00`

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

oqs-bot

Graviton2

Details

Benchmark suite	Current: `885bca1`	Previous: `0b1c536`	Ratio
`ML-DSA-44 keypair`	`113609` cycles	`113289` cycles	`1.00`
`ML-DSA-44 sign`	`350629` cycles	`355883` cycles	`0.99`
`ML-DSA-44 verify`	`117092` cycles	`117973` cycles	`0.99`
`ML-DSA-65 keypair`	`196545` cycles	`196446` cycles	`1.00`
`ML-DSA-65 sign`	`587254` cycles	`589191` cycles	`1.00`
`ML-DSA-65 verify`	`194237` cycles	`194679` cycles	`1.00`
`ML-DSA-87 keypair`	`322301` cycles	`322682` cycles	`1.00`
`ML-DSA-87 sign`	`752294` cycles	`752805` cycles	`1.00`
`ML-DSA-87 verify`	`320094` cycles	`320327` cycles	`1.00`

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

oqs-bot

Graviton2 (no-opt)

Details

Benchmark suite	Current: `885bca1`	Previous: `0b1c536`	Ratio
`ML-DSA-44 keypair`	`212986` cycles	`213066` cycles	`1.00`
`ML-DSA-44 sign`	`760407` cycles	`760558` cycles	`1.00`
`ML-DSA-44 verify`	`241293` cycles	`233103` cycles	`1.04`
`ML-DSA-65 keypair`	`380806` cycles	`381034` cycles	`1.00`
`ML-DSA-65 sign`	`1252121` cycles	`1252511` cycles	`1.00`
`ML-DSA-65 verify`	`372320` cycles	`372570` cycles	`1.00`
`ML-DSA-87 keypair`	`606317` cycles	`606046` cycles	`1.00`
`ML-DSA-87 sign`	`1593381` cycles	`1593756` cycles	`1.00`
`ML-DSA-87 verify`	`618121` cycles	`617945` cycles	`1.00`

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

oqs-bot

⚠️ Performance Alert ⚠️

Possible performance regression was detected for benchmark 'Graviton2 (no-opt)'.
Benchmark result of this commit is worse than the previous benchmark result exceeding threshold 1.03.

Benchmark suite	Current: `885bca1`	Previous: `0b1c536`	Ratio
`ML-DSA-44 verify`	`241293` cycles	`233103` cycles	`1.04`

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: `885bca1`	Previous: `0b1c536`	Ratio
`ML-DSA-44 keypair`	`827435` cycles	`828343` cycles	`1.00`
`ML-DSA-44 sign`	`3238214` cycles	`3235012` cycles	`1.00`
`ML-DSA-44 verify`	`921978` cycles	`920749` cycles	`1.00`
`ML-DSA-65 keypair`	`1412999` cycles	`1413905` cycles	`1.00`
`ML-DSA-65 sign`	`5347624` cycles	`5341776` cycles	`1.00`
`ML-DSA-65 verify`	`1477830` cycles	`1478062` cycles	`1.00`
`ML-DSA-87 keypair`	`2312761` cycles	`2313582` cycles	`1.00`
`ML-DSA-87 sign`	`6663968` cycles	`6664057` cycles	`1.00`
`ML-DSA-87 verify`	`2410302` cycles	`2412445` 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: `885bca1`	Previous: `0b1c536`	Ratio
`ML-DSA-44 keypair`	`278362` cycles	`277272` cycles	`1.00`
`ML-DSA-44 sign`	`798797` cycles	`822468` cycles	`0.97`
`ML-DSA-44 verify`	`277666` cycles	`277832` cycles	`1.00`
`ML-DSA-65 keypair`	`480516` cycles	`475993` cycles	`1.01`
`ML-DSA-65 sign`	`1349323` cycles	`1333415` cycles	`1.01`
`ML-DSA-65 verify`	`456855` cycles	`458979` cycles	`1.00`
`ML-DSA-87 keypair`	`817862` cycles	`817627` cycles	`1.00`
`ML-DSA-87 sign`	`1841380` cycles	`1833605` cycles	`1.00`
`ML-DSA-87 verify`	`788029` cycles	`798022` cycles	`0.99`

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: `885bca1`	Previous: `0b1c536`	Ratio
`ML-DSA-44 keypair`	`464640` cycles	`465283` cycles	`1.00`
`ML-DSA-44 sign`	`2150220` cycles	`2151007` cycles	`1.00`
`ML-DSA-44 verify`	`550751` cycles	`550792` cycles	`1.00`
`ML-DSA-65 keypair`	`779027` cycles	`780624` cycles	`1.00`
`ML-DSA-65 sign`	`3514123` cycles	`3517857` cycles	`1.00`
`ML-DSA-65 verify`	`856201` cycles	`854537` cycles	`1.00`
`ML-DSA-87 keypair`	`1261706` cycles	`1268967` cycles	`0.99`
`ML-DSA-87 sign`	`4350624` cycles	`4402745` cycles	`0.99`
`ML-DSA-87 verify`	`1373405` cycles	`1380067` cycles	`1.00`

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

oqs-bot · 2026-02-21T07:38:13Z

CBMC Results (ML-DSA-65)

Full Results (175 proofs)

Proof	Status	Current	Previous	Change
`TOTAL`	✅	2347s	2286s	+2.7%
`polyvecl_pointwise_acc_montgomery_c`	✅	246s	226s	+9%
`mld_attempt_signature_generation`	✅	206s	197s	+5%
`sign_verify_internal`	✅	183s	177s	+3%
`polyvec_matrix_expand`	✅	151s	145s	+4%
`rej_uniform_native`	✅	149s	144s	+3%
`poly_pointwise_montgomery_c`	✅	145s	138s	+5%
`mld_invntt_layer`	✅	124s	117s	+6%
`mld_ct_memcmp`	✅	86s	79s	+9%
`polyvec_matrix_expand_serial`	✅	66s	65s	+2%
`sign_signature_internal`	✅	53s	50s	+6%
`mld_ntt_layer`	✅	50s	44s	+14%
`keccak_squeezeblocks_x4`	✅	42s	42s	+0%
`mld_compute_t0_t1_tr_from_sk_components`	✅	27s	27s	+0%
`poly_chknorm_c`	✅	21s	16s	+31%
`rej_uniform`	✅	21s	21s	+0%
`fqmul`	✅	19s	18s	+6%
`polymat_permute_bitrev_to_custom`	✅	19s	18s	+6%
`polyveck_decompose`	✅	18s	16s	+12%
`rej_uniform_c`	✅	18s	19s	-5%
`polyt0_unpack`	✅	17s	17s	+0%
`poly_uniform_4x`	✅	15s	13s	+15%
`poly_uniform_eta_4x`	✅	15s	17s	-12%
`mld_ntt_butterfly_block`	✅	14s	13s	+8%
`mld_polyvecl_permute_bitrev_to_custom_native`	✅	13s	14s	-7%
`keccakf1600x4_permute_native`	✅	12s	14s	-14%
`polyvec_matrix_pointwise_montgomery`	✅	12s	14s	-14%
`polyveck_use_hint`	✅	12s	14s	-14%
`sign`	✅	12s	9s	+33%
`keccak_absorb_once_x4`	✅	11s	10s	+10%
`polyveck_invntt_tomont`	✅	11s	8s	+38%
`keccakf1600_permute_native`	✅	10s	9s	+11%
`mld_check_pct`	✅	10s	9s	+11%
`polyveck_ntt`	✅	10s	6s	+67%
`polyveck_power2round`	✅	9s	11s	-18%
`polyveck_reduce`	✅	9s	9s	+0%
`sign_pk_from_sk`	✅	9s	8s	+12%
`mld_sample_s1_s2`	✅	8s	4s	+100%
`poly_invntt_tomont_c`	✅	8s	8s	+0%
`polyveck_add`	✅	8s	9s	-11%
`polyveck_caddq`	✅	8s	8s	+0%
`polyveck_shiftl`	✅	8s	8s	+0%
`polyveck_sub`	✅	8s	7s	+14%
`keccakf1600_permute`	✅	7s	9s	-22%
`mld_prepare_domain_separation_prefix`	✅	7s	5s	+40%
`poly_decompose_c`	✅	7s	7s	+0%
`poly_ntt_c`	✅	7s	2s	+250%
`polyvecl_ntt`	✅	7s	8s	-12%
`mld_compute_pack_z`	✅	6s	7s	-14%
`poly_use_hint_native`	✅	6s	4s	+50%
`polyeta_unpack`	✅	6s	7s	-14%
`polyveck_unpack_t0`	✅	6s	3s	+100%
`polyvecl_uniform_gamma1_serial`	✅	6s	4s	+50%
`sign_open`	✅	6s	5s	+20%
`keccakf1600_extract_bytes (big endian)`	✅	5s	3s	+67%
`mld_h`	✅	5s	3s	+67%
`mld_sample_s1_s2_serial`	✅	5s	6s	-17%
`poly_add`	✅	5s	4s	+25%
`poly_caddq_c`	✅	5s	3s	+67%
`polyveck_make_hint`	✅	5s	5s	+0%
`polyveck_pack_t0`	✅	5s	3s	+67%
`polyveck_pointwise_poly_montgomery`	✅	5s	6s	-17%
`polyvecl_chknorm`	✅	5s	5s	+0%
`rej_eta_native`	✅	5s	4s	+25%
`sign_signature`	✅	5s	4s	+25%
`sign_signature_pre_hash_internal`	✅	5s	6s	-17%
`sign_verify`	✅	5s	4s	+25%
`unpack_sk`	✅	5s	5s	+0%
`fqscale`	✅	4s	4s	+0%
`intt_native_x86_64`	✅	4s	3s	+33%
`keccak_absorb`	✅	4s	5s	-20%
`keccak_init`	✅	4s	3s	+33%
`make_hint`	✅	4s	3s	+33%
`poly_caddq_native`	✅	4s	4s	+0%
`poly_chknorm`	✅	4s	3s	+33%
`poly_decompose_native`	✅	4s	3s	+33%
`poly_pointwise_montgomery`	✅	4s	5s	-20%
`poly_pointwise_montgomery_native`	✅	4s	2s	+100%
`poly_shiftl`	✅	4s	2s	+100%
`poly_uniform`	✅	4s	4s	+0%
`poly_uniform_eta`	✅	4s	3s	+33%
`poly_use_hint_c`	✅	4s	7s	-43%
`polyveck_chknorm`	✅	4s	5s	-20%
`polyveck_pack_w1`	✅	4s	6s	-33%
`polyveck_unpack_eta`	✅	4s	4s	+0%
`polyvecl_pointwise_acc_montgomery`	✅	4s	5s	-20%
`polyvecl_pointwise_acc_montgomery_native`	✅	4s	5s	-20%
`polyz_unpack`	✅	4s	2s	+100%
`polyz_unpack_c`	✅	4s	5s	-20%
`polyz_unpack_native`	✅	4s	3s	+33%
`power2round`	✅	4s	2s	+100%
`rej_eta_c`	✅	4s	3s	+33%
`sign_keypair`	✅	4s	3s	+33%
`sign_keypair_internal`	✅	4s	6s	-33%
`sign_signature_extmu`	✅	4s	4s	+0%
`sign_verify_extmu`	✅	4s	4s	+0%
`sign_verify_pre_hash_internal`	✅	4s	3s	+33%
`sign_verify_pre_hash_shake256`	✅	4s	6s	-33%
`unpack_hints`	✅	4s	5s	-20%
`unpack_sig`	✅	4s	4s	+0%
`keccak_finalize`	✅	3s	2s	+50%
`keccak_squeeze`	✅	3s	2s	+50%
`keccakf1600_xor_bytes`	✅	3s	3s	+0%
`keccakf1600x4_permute`	✅	3s	1s	+200%
`keccakf1600x4_xor_bytes`	✅	3s	2s	+50%
`mld_ct_cmask_neg_i32`	✅	3s	2s	+50%
`mld_ct_cmask_nonzero_u32`	✅	3s	2s	+50%
`mld_ct_cmask_nonzero_u8`	✅	3s	2s	+50%
`mld_ct_get_optblocker_u32`	✅	3s	4s	-25%
`mld_ct_get_optblocker_u8`	✅	3s	2s	+50%
`mld_value_barrier_u8`	✅	3s	3s	+0%
`montgomery_reduce`	✅	3s	2s	+50%
`pack_pk`	✅	3s	3s	+0%
`pack_sig_c_h`	✅	3s	2s	+50%
`poly_caddq`	✅	3s	3s	+0%
`poly_challenge`	✅	3s	4s	-25%
`poly_chknorm_native`	✅	3s	4s	-25%
`poly_decompose`	✅	3s	3s	+0%
`poly_invntt_tomont`	✅	3s	6s	-50%
`poly_invntt_tomont_native`	✅	3s	7s	-57%
`poly_ntt`	✅	3s	5s	-40%
`poly_power2round`	✅	3s	5s	-40%
`poly_sub`	✅	3s	5s	-40%
`poly_uniform_gamma1`	✅	3s	5s	-40%
`polyeta_pack`	✅	3s	2s	+50%
`polyt0_pack`	✅	3s	4s	-25%
`polyveck_pack_eta`	✅	3s	4s	-25%
`polyvecl_uniform_gamma1`	✅	3s	5s	-40%
`polyvecl_unpack_z`	✅	3s	2s	+50%
`polyw1_pack`	✅	3s	1s	+200%
`polyz_pack`	✅	3s	4s	-25%
`reduce32`	✅	3s	3s	+0%
`rej_eta`	✅	3s	4s	-25%
`shake128_absorb`	✅	3s	2s	+50%
`shake128_init`	✅	3s	1s	+200%
`shake128_release`	✅	3s	3s	+0%
`shake256x4_absorb_once`	✅	3s	4s	-25%
`sign_signature_pre_hash_shake256`	✅	3s	4s	-25%
`unpack_pk`	✅	3s	3s	+0%
`caddq`	✅	2s	3s	-33%
`decompose`	✅	2s	5s	-60%
`keccakf1600_xor_bytes (big endian)`	✅	2s	3s	-33%
`keccakf1600x4_extract_bytes`	✅	2s	1s	+100%
`mld_ct_abs_i32`	✅	2s	4s	-50%
`mld_ct_sel_int32`	✅	2s	3s	-33%
`mld_keccakf1600_extract_bytes`	✅	2s	2s	+0%
`ntt_native_x86_64`	✅	2s	3s	-33%
`pack_sig_z`	✅	2s	2s	+0%
`pack_sk`	✅	2s	2s	+0%
`poly_caddq_native_aarch64`	✅	2s	6s	-67%
`poly_make_hint`	✅	2s	3s	-33%
`poly_ntt_native`	✅	2s	3s	-33%
`poly_reduce`	✅	2s	5s	-60%
`poly_uniform_gamma1_4x`	✅	2s	6s	-67%
`polyt1_pack`	✅	2s	2s	+0%
`polyt1_unpack`	✅	2s	6s	-67%
`polyvecl_pack_eta`	✅	2s	4s	-50%
`polyvecl_permute_bitrev_to_custom`	✅	2s	3s	-33%
`polyvecl_unpack_eta`	✅	2s	5s	-60%
`shake128_finalize`	✅	2s	3s	-33%
`shake128x4_absorb_once`	✅	2s	4s	-50%
`shake128x4_squeezeblocks`	✅	2s	1s	+100%
`shake256`	✅	2s	2s	+0%
`shake256_absorb`	✅	2s	5s	-60%
`shake256_finalize`	✅	2s	5s	-60%
`shake256_init`	✅	2s	2s	+0%
`shake256_release`	✅	2s	1s	+100%
`shake256_squeeze`	✅	2s	2s	+0%
`sys_check_capability`	✅	2s	3s	-33%
`use_hint`	✅	2s	2s	+0%
`mld_ct_get_optblocker_i64`	✅	1s	1s	+0%
`mld_value_barrier_i64`	✅	1s	2s	-50%
`mld_value_barrier_u32`	✅	1s	3s	-67%
`poly_use_hint`	✅	1s	3s	-67%
`shake128_squeeze`	✅	1s	2s	-50%
`shake256x4_squeezeblocks`	✅	1s	3s	-67%

oqs-bot · 2026-02-21T07:38:22Z

CBMC Results (ML-DSA-44)

Full Results (175 proofs)

Proof	Status	Current	Previous	Change
`TOTAL`	✅	2048s	2055s	-0.3%
`sign_verify_internal`	✅	253s	254s	-0%
`mld_attempt_signature_generation`	✅	228s	221s	+3%
`polyvecl_pointwise_acc_montgomery_c`	✅	211s	208s	+1%
`rej_uniform_native`	✅	142s	144s	-1%
`poly_pointwise_montgomery_c`	✅	130s	143s	-9%
`mld_ct_memcmp`	✅	80s	82s	-2%
`mld_invntt_layer`	✅	50s	50s	+0%
`sign_signature_internal`	✅	45s	45s	+0%
`keccak_squeezeblocks_x4`	✅	43s	44s	-2%
`mld_ntt_layer`	✅	43s	44s	-2%
`poly_invntt_tomont_c`	✅	40s	39s	+3%
`rej_uniform`	✅	20s	20s	+0%
`fqmul`	✅	18s	19s	-5%
`rej_uniform_c`	✅	17s	20s	-15%
`poly_uniform_eta_4x`	✅	16s	19s	-16%
`mld_ntt_butterfly_block`	✅	15s	13s	+15%
`poly_uniform_4x`	✅	15s	13s	+15%
`polymat_permute_bitrev_to_custom`	✅	15s	16s	-6%
`polyt0_unpack`	✅	15s	15s	+0%
`mld_compute_t0_t1_tr_from_sk_components`	✅	14s	15s	-7%
`polyvec_matrix_expand`	✅	14s	16s	-12%
`keccakf1600x4_permute_native`	✅	13s	14s	-7%
`mld_polyvecl_permute_bitrev_to_custom_native`	✅	13s	14s	-7%
`polyeta_unpack`	✅	13s	12s	+8%
`poly_chknorm_c`	✅	12s	12s	+0%
`polyz_unpack_c`	✅	12s	12s	+0%
`keccak_absorb_once_x4`	✅	11s	9s	+22%
`keccakf1600_permute`	✅	9s	8s	+12%
`mld_check_pct`	✅	9s	6s	+50%
`polyveck_add`	✅	9s	7s	+29%
`keccakf1600_permute_native`	✅	8s	6s	+33%
`polyveck_decompose`	✅	8s	7s	+14%
`polyvec_matrix_expand_serial`	✅	7s	7s	+0%
`polyvec_matrix_pointwise_montgomery`	✅	7s	6s	+17%
`polyveck_pointwise_poly_montgomery`	✅	7s	5s	+40%
`sign_keypair`	✅	7s	2s	+250%
`sign_verify_pre_hash_internal`	✅	7s	4s	+75%
`mld_prepare_domain_separation_prefix`	✅	6s	5s	+20%
`poly_uniform_eta`	✅	6s	5s	+20%
`polyt0_pack`	✅	6s	4s	+50%
`polyveck_caddq`	✅	6s	8s	-25%
`polyveck_invntt_tomont`	✅	6s	6s	+0%
`polyveck_ntt`	✅	6s	4s	+50%
`polyveck_reduce`	✅	6s	5s	+20%
`polyvecl_chknorm`	✅	6s	6s	+0%
`sign`	✅	6s	5s	+20%
`sign_verify`	✅	6s	3s	+100%
`unpack_hints`	✅	6s	5s	+20%
`decompose`	✅	5s	3s	+67%
`fqscale`	✅	5s	1s	+400%
`keccak_absorb`	✅	5s	8s	-38%
`keccakf1600x4_extract_bytes`	✅	5s	3s	+67%
`mld_compute_pack_z`	✅	5s	6s	-17%
`mld_h`	✅	5s	6s	-17%
`mld_sample_s1_s2_serial`	✅	5s	3s	+67%
`poly_chknorm_native`	✅	5s	6s	-17%
`poly_decompose_c`	✅	5s	5s	+0%
`poly_ntt_c`	✅	5s	2s	+150%
`poly_use_hint_c`	✅	5s	5s	+0%
`polyeta_pack`	✅	5s	3s	+67%
`polyveck_chknorm`	✅	5s	3s	+67%
`polyveck_power2round`	✅	5s	6s	-17%
`polyveck_use_hint`	✅	5s	5s	+0%
`sign_verify_extmu`	✅	5s	2s	+150%
`sign_verify_pre_hash_shake256`	✅	5s	3s	+67%
`keccak_finalize`	✅	4s	4s	+0%
`keccak_init`	✅	4s	3s	+33%
`keccak_squeeze`	✅	4s	5s	-20%
`keccakf1600_xor_bytes (big endian)`	✅	4s	2s	+100%
`keccakf1600x4_permute`	✅	4s	3s	+33%
`mld_ct_abs_i32`	✅	4s	2s	+100%
`mld_sample_s1_s2`	✅	4s	4s	+0%
`mld_value_barrier_u8`	✅	4s	1s	+300%
`montgomery_reduce`	✅	4s	2s	+100%
`ntt_native_x86_64`	✅	4s	3s	+33%
`poly_challenge`	✅	4s	3s	+33%
`poly_pointwise_montgomery_native`	✅	4s	3s	+33%
`poly_power2round`	✅	4s	5s	-20%
`poly_uniform`	✅	4s	5s	-20%
`polyt1_unpack`	✅	4s	4s	+0%
`polyveck_sub`	✅	4s	4s	+0%
`polyveck_unpack_eta`	✅	4s	4s	+0%
`polyvecl_ntt`	✅	4s	6s	-33%
`polyvecl_pack_eta`	✅	4s	3s	+33%
`polyvecl_permute_bitrev_to_custom`	✅	4s	2s	+100%
`polyvecl_pointwise_acc_montgomery`	✅	4s	4s	+0%
`polyvecl_pointwise_acc_montgomery_native`	✅	4s	3s	+33%
`polyz_unpack_native`	✅	4s	3s	+33%
`reduce32`	✅	4s	3s	+33%
`rej_eta_c`	✅	4s	4s	+0%
`rej_eta_native`	✅	4s	4s	+0%
`shake128x4_squeezeblocks`	✅	4s	2s	+100%
`shake256`	✅	4s	3s	+33%
`shake256_release`	✅	4s	3s	+33%
`shake256x4_absorb_once`	✅	4s	4s	+0%
`sign_open`	✅	4s	6s	-33%
`sign_pk_from_sk`	✅	4s	6s	-33%
`sign_signature_extmu`	✅	4s	7s	-43%
`sign_signature_pre_hash_internal`	✅	4s	4s	+0%
`unpack_sk`	✅	4s	4s	+0%
`caddq`	✅	3s	3s	+0%
`intt_native_x86_64`	✅	3s	3s	+0%
`keccakf1600_extract_bytes (big endian)`	✅	3s	3s	+0%
`mld_ct_cmask_neg_i32`	✅	3s	2s	+50%
`mld_ct_cmask_nonzero_u32`	✅	3s	2s	+50%
`mld_ct_get_optblocker_i64`	✅	3s	2s	+50%
`mld_ct_get_optblocker_u8`	✅	3s	1s	+200%
`mld_keccakf1600_extract_bytes`	✅	3s	6s	-50%
`mld_value_barrier_i64`	✅	3s	4s	-25%
`pack_sig_c_h`	✅	3s	4s	-25%
`pack_sig_z`	✅	3s	2s	+50%
`pack_sk`	✅	3s	3s	+0%
`poly_add`	✅	3s	4s	-25%
`poly_caddq`	✅	3s	3s	+0%
`poly_caddq_native_aarch64`	✅	3s	3s	+0%
`poly_decompose_native`	✅	3s	3s	+0%
`poly_invntt_tomont`	✅	3s	2s	+50%
`poly_invntt_tomont_native`	✅	3s	4s	-25%
`poly_pointwise_montgomery`	✅	3s	2s	+50%
`poly_shiftl`	✅	3s	4s	-25%
`poly_uniform_gamma1_4x`	✅	3s	4s	-25%
`poly_use_hint_native`	✅	3s	3s	+0%
`polyt1_pack`	✅	3s	1s	+200%
`polyveck_make_hint`	✅	3s	4s	-25%
`polyveck_pack_t0`	✅	3s	3s	+0%
`polyveck_pack_w1`	✅	3s	4s	-25%
`polyveck_shiftl`	✅	3s	5s	-40%
`polyveck_unpack_t0`	✅	3s	4s	-25%
`polyvecl_uniform_gamma1`	✅	3s	2s	+50%
`polyvecl_uniform_gamma1_serial`	✅	3s	3s	+0%
`polyw1_pack`	✅	3s	2s	+50%
`polyz_unpack`	✅	3s	5s	-40%
`rej_eta`	✅	3s	1s	+200%
`shake128_init`	✅	3s	2s	+50%
`shake256_absorb`	✅	3s	2s	+50%
`shake256_init`	✅	3s	3s	+0%
`shake256_squeeze`	✅	3s	5s	-40%
`sign_keypair_internal`	✅	3s	4s	-25%
`sign_signature`	✅	3s	6s	-50%
`sys_check_capability`	✅	3s	4s	-25%
`unpack_pk`	✅	3s	3s	+0%
`use_hint`	✅	3s	2s	+50%
`make_hint`	✅	2s	3s	-33%
`mld_ct_cmask_nonzero_u8`	✅	2s	4s	-50%
`mld_value_barrier_u32`	✅	2s	4s	-50%
`pack_pk`	✅	2s	3s	-33%
`poly_caddq_c`	✅	2s	3s	-33%
`poly_caddq_native`	✅	2s	2s	+0%
`poly_chknorm`	✅	2s	4s	-50%
`poly_decompose`	✅	2s	4s	-50%
`poly_make_hint`	✅	2s	4s	-50%
`poly_ntt_native`	✅	2s	2s	+0%
`poly_sub`	✅	2s	3s	-33%
`poly_uniform_gamma1`	✅	2s	4s	-50%
`poly_use_hint`	✅	2s	2s	+0%
`polyvecl_unpack_z`	✅	2s	5s	-60%
`polyz_pack`	✅	2s	2s	+0%
`shake128_absorb`	✅	2s	1s	+100%
`shake128_release`	✅	2s	3s	-33%
`shake128_squeeze`	✅	2s	2s	+0%
`shake256_finalize`	✅	2s	4s	-50%
`shake256x4_squeezeblocks`	✅	2s	2s	+0%
`sign_signature_pre_hash_shake256`	✅	2s	5s	-60%
`unpack_sig`	✅	2s	3s	-33%
`keccakf1600_xor_bytes`	✅	1s	2s	-50%
`keccakf1600x4_xor_bytes`	✅	1s	2s	-50%
`mld_ct_get_optblocker_u32`	✅	1s	1s	+0%
`mld_ct_sel_int32`	✅	1s	4s	-75%
`poly_ntt`	✅	1s	3s	-67%
`poly_reduce`	✅	1s	4s	-75%
`polyveck_pack_eta`	✅	1s	3s	-67%
`polyvecl_unpack_eta`	✅	1s	2s	-50%
`power2round`	✅	1s	2s	-50%
`shake128_finalize`	✅	1s	2s	-50%
`shake128x4_absorb_once`	✅	1s	4s	-75%

github-actions

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

Details

Benchmark suite	Current: `885bca1`	Previous: `0b1c536`	Ratio
`ML-DSA-44 keypair`	`231675` cycles	`223693` cycles	`1.04`
`ML-DSA-44 sign`	`637708` cycles	`608242` cycles	`1.05`
`ML-DSA-44 verify`	`228579` cycles	`221112` cycles	`1.03`
`ML-DSA-65 keypair`	`412694` cycles	`394259` cycles	`1.05`
`ML-DSA-65 sign`	`1064124` cycles	`1015180` cycles	`1.05`
`ML-DSA-65 verify`	`390621` cycles	`372405` cycles	`1.05`
`ML-DSA-87 keypair`	`682821` cycles	`653922` cycles	`1.04`
`ML-DSA-87 sign`	`1412962` cycles	`1363561` cycles	`1.04`
`ML-DSA-87 verify`	`668826` cycles	`637673` 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 (opt)'.
Benchmark result of this commit is worse than the previous benchmark result exceeding threshold 1.03.

Benchmark suite	Current: `885bca1`	Previous: `0b1c536`	Ratio
`ML-DSA-44 keypair`	`231675` cycles	`223693` cycles	`1.04`
`ML-DSA-44 sign`	`637708` cycles	`608242` cycles	`1.05`
`ML-DSA-44 verify`	`228579` cycles	`221112` cycles	`1.03`
`ML-DSA-65 keypair`	`412694` cycles	`394259` cycles	`1.05`
`ML-DSA-65 sign`	`1064124` cycles	`1015180` cycles	`1.05`
`ML-DSA-65 verify`	`390621` cycles	`372405` cycles	`1.05`
`ML-DSA-87 keypair`	`682821` cycles	`653922` cycles	`1.04`
`ML-DSA-87 sign`	`1412962` cycles	`1363561` cycles	`1.04`
`ML-DSA-87 verify`	`668826` cycles	`637673` cycles	`1.05`

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

oqs-bot · 2026-02-21T07:40:14Z

CBMC Results (ML-DSA-87)

Full Results (175 proofs)

Proof	Status	Current	Previous	Change
`TOTAL`	✅	2628s	2449s	+7.3%
`sign_verify_internal`	✅	376s	353s	+7%
`mld_attempt_signature_generation`	✅	255s	227s	+12%
`polyvecl_pointwise_acc_montgomery_c`	✅	195s	165s	+18%
`polyvec_matrix_expand`	✅	164s	153s	+7%
`poly_pointwise_montgomery_c`	✅	155s	128s	+21%
`rej_uniform_native`	✅	151s	139s	+9%
`mld_invntt_layer`	✅	121s	114s	+6%
`polyvec_matrix_expand_serial`	✅	114s	110s	+4%
`mld_ct_memcmp`	✅	86s	74s	+16%
`mld_ntt_layer`	✅	47s	44s	+7%
`sign_signature_internal`	✅	47s	46s	+2%
`keccak_squeezeblocks_x4`	✅	42s	42s	+0%
`mld_compute_t0_t1_tr_from_sk_components`	✅	26s	25s	+4%
`polymat_permute_bitrev_to_custom`	✅	26s	24s	+8%
`fqmul`	✅	23s	18s	+28%
`rej_uniform`	✅	22s	21s	+5%
`poly_uniform_eta_4x`	✅	20s	17s	+18%
`rej_uniform_c`	✅	18s	16s	+12%
`poly_chknorm_c`	✅	17s	17s	+0%
`poly_uniform_4x`	✅	16s	17s	-6%
`polyveck_add`	✅	15s	13s	+15%
`polyt0_unpack`	✅	14s	17s	-18%
`polyvec_matrix_pointwise_montgomery`	✅	14s	12s	+17%
`polyveck_power2round`	✅	14s	14s	+0%
`keccakf1600x4_permute_native`	✅	13s	12s	+8%
`polyeta_unpack`	✅	13s	13s	+0%
`mld_ntt_butterfly_block`	✅	12s	13s	-8%
`keccak_absorb_once_x4`	✅	11s	10s	+10%
`polyveck_reduce`	✅	11s	13s	-15%
`polyveck_use_hint`	✅	11s	9s	+22%
`sign`	✅	10s	7s	+43%
`mld_check_pct`	✅	9s	7s	+29%
`poly_invntt_tomont_c`	✅	9s	9s	+0%
`polyveck_pointwise_poly_montgomery`	✅	9s	8s	+12%
`mld_compute_pack_z`	✅	8s	7s	+14%
`mld_sample_s1_s2`	✅	8s	6s	+33%
`mld_sample_s1_s2_serial`	✅	8s	6s	+33%
`poly_decompose_c`	✅	8s	7s	+14%
`polyveck_chknorm`	✅	8s	6s	+33%
`polyveck_ntt`	✅	8s	8s	+0%
`polyveck_shiftl`	✅	8s	6s	+33%
`polyvecl_ntt`	✅	8s	11s	-27%
`keccakf1600_permute`	✅	7s	7s	+0%
`keccakf1600_permute_native`	✅	7s	10s	-30%
`mld_polyvecl_permute_bitrev_to_custom_native`	✅	7s	8s	-12%
`polyveck_caddq`	✅	7s	7s	+0%
`polyveck_decompose`	✅	7s	6s	+17%
`polyveck_invntt_tomont`	✅	7s	10s	-30%
`polyveck_sub`	✅	7s	7s	+0%
`sign_pk_from_sk`	✅	7s	9s	-22%
`sign_signature`	✅	7s	5s	+40%
`sign_signature_pre_hash_internal`	✅	7s	5s	+40%
`unpack_sig`	✅	7s	5s	+40%
`keccak_absorb`	✅	6s	7s	-14%
`polyz_unpack_native`	✅	6s	4s	+50%
`sign_keypair_internal`	✅	6s	6s	+0%
`mld_ct_sel_int32`	✅	5s	2s	+150%
`mld_prepare_domain_separation_prefix`	✅	5s	3s	+67%
`ntt_native_x86_64`	✅	5s	4s	+25%
`poly_caddq_c`	✅	5s	3s	+67%
`poly_challenge`	✅	5s	5s	+0%
`poly_uniform_eta`	✅	5s	4s	+25%
`poly_uniform_gamma1`	✅	5s	3s	+67%
`polyt0_pack`	✅	5s	5s	+0%
`polyveck_make_hint`	✅	5s	5s	+0%
`polyveck_pack_t0`	✅	5s	2s	+150%
`polyveck_unpack_t0`	✅	5s	6s	-17%
`polyvecl_unpack_eta`	✅	5s	3s	+67%
`polyvecl_unpack_z`	✅	5s	3s	+67%
`rej_eta`	✅	5s	2s	+150%
`rej_eta_c`	✅	5s	4s	+25%
`shake128_release`	✅	5s	3s	+67%
`sign_keypair`	✅	5s	3s	+67%
`sign_signature_pre_hash_shake256`	✅	5s	4s	+25%
`sign_verify_extmu`	✅	5s	3s	+67%
`unpack_hints`	✅	5s	4s	+25%
`unpack_sk`	✅	5s	6s	-17%
`keccak_init`	✅	4s	3s	+33%
`keccakf1600_xor_bytes (big endian)`	✅	4s	3s	+33%
`keccakf1600x4_extract_bytes`	✅	4s	2s	+100%
`mld_ct_abs_i32`	✅	4s	2s	+100%
`mld_h`	✅	4s	2s	+100%
`pack_sig_z`	✅	4s	3s	+33%
`pack_sk`	✅	4s	3s	+33%
`poly_add`	✅	4s	4s	+0%
`poly_caddq`	✅	4s	2s	+100%
`poly_decompose`	✅	4s	3s	+33%
`poly_decompose_native`	✅	4s	5s	-20%
`poly_make_hint`	✅	4s	2s	+100%
`poly_uniform`	✅	4s	6s	-33%
`polyt1_pack`	✅	4s	2s	+100%
`polyveck_unpack_eta`	✅	4s	3s	+33%
`polyvecl_chknorm`	✅	4s	4s	+0%
`polyvecl_pack_eta`	✅	4s	3s	+33%
`polyvecl_pointwise_acc_montgomery`	✅	4s	4s	+0%
`polyvecl_uniform_gamma1`	✅	4s	4s	+0%
`polyvecl_uniform_gamma1_serial`	✅	4s	7s	-43%
`polyz_unpack`	✅	4s	3s	+33%
`polyz_unpack_c`	✅	4s	5s	-20%
`rej_eta_native`	✅	4s	5s	-20%
`shake256x4_absorb_once`	✅	4s	2s	+100%
`sign_open`	✅	4s	6s	-33%
`sign_signature_extmu`	✅	4s	5s	-20%
`sign_verify_pre_hash_internal`	✅	4s	5s	-20%
`sign_verify_pre_hash_shake256`	✅	4s	4s	+0%
`caddq`	✅	3s	4s	-25%
`intt_native_x86_64`	✅	3s	5s	-40%
`keccakf1600_xor_bytes`	✅	3s	2s	+50%
`mld_ct_cmask_neg_i32`	✅	3s	1s	+200%
`mld_ct_cmask_nonzero_u8`	✅	3s	3s	+0%
`mld_ct_get_optblocker_i64`	✅	3s	3s	+0%
`mld_ct_get_optblocker_u32`	✅	3s	2s	+50%
`pack_sig_c_h`	✅	3s	3s	+0%
`poly_caddq_native`	✅	3s	2s	+50%
`poly_caddq_native_aarch64`	✅	3s	3s	+0%
`poly_chknorm_native`	✅	3s	2s	+50%
`poly_invntt_tomont`	✅	3s	2s	+50%
`poly_ntt_c`	✅	3s	1s	+200%
`poly_ntt_native`	✅	3s	6s	-50%
`poly_pointwise_montgomery`	✅	3s	4s	-25%
`poly_power2round`	✅	3s	3s	+0%
`poly_reduce`	✅	3s	4s	-25%
`poly_shiftl`	✅	3s	2s	+50%
`poly_sub`	✅	3s	3s	+0%
`poly_uniform_gamma1_4x`	✅	3s	6s	-50%
`polyt1_unpack`	✅	3s	3s	+0%
`polyveck_pack_eta`	✅	3s	2s	+50%
`polyvecl_pointwise_acc_montgomery_native`	✅	3s	3s	+0%
`polyw1_pack`	✅	3s	2s	+50%
`reduce32`	✅	3s	4s	-25%
`shake128_finalize`	✅	3s	3s	+0%
`shake256`	✅	3s	4s	-25%
`shake256_finalize`	✅	3s	3s	+0%
`shake256_release`	✅	3s	3s	+0%
`shake256_squeeze`	✅	3s	2s	+50%
`sign_verify`	✅	3s	2s	+50%
`unpack_pk`	✅	3s	6s	-50%
`use_hint`	✅	3s	3s	+0%
`decompose`	✅	2s	4s	-50%
`fqscale`	✅	2s	5s	-60%
`keccak_finalize`	✅	2s	2s	+0%
`keccak_squeeze`	✅	2s	4s	-50%
`keccakf1600x4_permute`	✅	2s	2s	+0%
`keccakf1600x4_xor_bytes`	✅	2s	2s	+0%
`mld_ct_get_optblocker_u8`	✅	2s	2s	+0%
`mld_value_barrier_i64`	✅	2s	3s	-33%
`mld_value_barrier_u32`	✅	2s	2s	+0%
`mld_value_barrier_u8`	✅	2s	2s	+0%
`montgomery_reduce`	✅	2s	3s	-33%
`pack_pk`	✅	2s	4s	-50%
`poly_chknorm`	✅	2s	2s	+0%
`poly_ntt`	✅	2s	3s	-33%
`poly_pointwise_montgomery_native`	✅	2s	3s	-33%
`poly_use_hint`	✅	2s	3s	-33%
`poly_use_hint_native`	✅	2s	4s	-50%
`polyeta_pack`	✅	2s	4s	-50%
`polyveck_pack_w1`	✅	2s	4s	-50%
`polyvecl_permute_bitrev_to_custom`	✅	2s	2s	+0%
`polyz_pack`	✅	2s	2s	+0%
`power2round`	✅	2s	4s	-50%
`shake128_absorb`	✅	2s	2s	+0%
`shake128_init`	✅	2s	2s	+0%
`shake128_squeeze`	✅	2s	5s	-60%
`shake128x4_absorb_once`	✅	2s	4s	-50%
`shake128x4_squeezeblocks`	✅	2s	2s	+0%
`shake256_absorb`	✅	2s	3s	-33%
`shake256x4_squeezeblocks`	✅	2s	1s	+100%
`sys_check_capability`	✅	2s	3s	-33%
`keccakf1600_extract_bytes (big endian)`	✅	1s	3s	-67%
`make_hint`	✅	1s	5s	-80%
`mld_ct_cmask_nonzero_u32`	✅	1s	4s	-75%
`mld_keccakf1600_extract_bytes`	✅	1s	1s	+0%
`poly_invntt_tomont_native`	✅	1s	2s	-50%
`poly_use_hint_c`	✅	1s	4s	-75%
`shake256_init`	✅	1s	2s	-50%

github-actions

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

Details

Benchmark suite	Current: `885bca1`	Previous: `0b1c536`	Ratio
`ML-DSA-44 keypair`	`313397` cycles	`321433` cycles	`0.97`
`ML-DSA-44 sign`	`1215690` cycles	`1202861` cycles	`1.01`
`ML-DSA-44 verify`	`343963` cycles	`340204` cycles	`1.01`
`ML-DSA-65 keypair`	`572366` cycles	`569351` cycles	`1.01`
`ML-DSA-65 sign`	`2038364` cycles	`1955934` cycles	`1.04`
`ML-DSA-65 verify`	`548523` cycles	`548845` cycles	`1.00`
`ML-DSA-87 keypair`	`908267` cycles	`885828` cycles	`1.03`
`ML-DSA-87 sign`	`2517161` cycles	`2512147` cycles	`1.00`
`ML-DSA-87 verify`	`925719` cycles	`902578` cycles	`1.03`

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: `885bca1`	Previous: `0b1c536`	Ratio
`ML-DSA-65 sign`	`2038364` cycles	`1955934` cycles	`1.04`

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

github-actions

⚠️ Performance Alert ⚠️

Possible performance regression was detected for benchmark 'Arm Cortex-A55 (Snapdragon 888) benchmarks (opt)'.
Benchmark result of this commit is worse than the previous benchmark result exceeding threshold 1.03.

Benchmark suite	Current: `4b4c152`	Previous: `0b1c536`	Ratio
`ML-DSA-87 sign`	`1900701` cycles	`1833605` cycles	`1.04`

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

Add AArch64 assembly implementations of polyw1_pack for both GAMMA2 variants using TBL-based byte extraction from 32-bit coefficient lanes. Signed-off-by: Matthias J. Kannwischer <matthias@kannwischer.eu>

mkannwischer · 2026-02-21T13:07:55Z

The scheme benchmarks suggest we should implement the 88 variant (ML-DSA-44), but not the 32 variant (ML-DSA-65/ML-DSA-87). It makes sense that the 88 one is harder to auto-vectorize, but it feels a little inconsistent to only implement one.

WDYT @hanno-becker?

x86_64: Add AVX2 polyw1_pack to x86_64 native backend

1bb503b

Integrate polyw1_pack AVX2 implementations for both GAMMA2 variants into the native backend. Signed-off-by: Matthias J. Kannwischer <matthias@kannwischer.eu>

mkannwischer force-pushed the x86-native-w1-pack branch from 0b22412 to 1bb503b Compare February 21, 2026 07:19

mkannwischer added the benchmark label Feb 21, 2026