Update AVX2 decompose to use a more explainable (and very slightly faster) approach, along with bounds reasoning comments. #629

jammychiou1 · 2025-11-06T06:02:49Z

Resolves Add bounds reasoning comments to AVX2 backend #528
Separated out from Add bounds reasoning comments to AVX2 [I]NTT and basemul #560

This approach was suggested by Hanno Becker in #411 (comment), when we implemented the same function in AArch64.

The speedup is barely noticeable, less than 50 cycles per call on my laptop.

mkannwischer

Thanks @jammychiou1.
I agree with the changes you made - it's easier to explain.

Please also remove the check-magic annotations concerning your comments and instead add these constants to the whitelist.

dev/x86_64/src/poly_decompose_32_avx2.c

github-actions

Mac Mini (M1, 2020) benchmarks (opt)

Benchmark suite	Current: `fa07d1f`	Previous: `100e446`	Ratio
`ML-DSA-44 keypair`	`46416` cycles	`46421` cycles	`1.00`
`ML-DSA-44 sign`	`132718` cycles	`132738` cycles	`1.00`
`ML-DSA-44 verify`	`47837` cycles	`47840` cycles	`1.00`
`ML-DSA-65 keypair`	`81452` cycles	`81443` cycles	`1.00`
`ML-DSA-65 sign`	`219217` cycles	`219207` cycles	`1.00`
`ML-DSA-65 verify`	`80136` cycles	`80134` cycles	`1.00`
`ML-DSA-87 keypair`	`132753` cycles	`132758` cycles	`1.00`
`ML-DSA-87 sign`	`280934` cycles	`280953` cycles	`1.00`
`ML-DSA-87 verify`	`130316` cycles	`130326` cycles	`1.00`

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

github-actions

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

Benchmark suite	Current: `fa07d1f`	Previous: `100e446`	Ratio
`ML-DSA-44 keypair`	`115272` cycles	`115262` cycles	`1.00`
`ML-DSA-44 sign`	`431782` cycles	`431720` cycles	`1.00`
`ML-DSA-44 verify`	`122176` cycles	`122167` cycles	`1.00`
`ML-DSA-65 keypair`	`197436` cycles	`197490` cycles	`1.00`
`ML-DSA-65 sign`	`700971` cycles	`701274` cycles	`1.00`
`ML-DSA-65 verify`	`197693` cycles	`197702` cycles	`1.00`
`ML-DSA-87 keypair`	`325389` cycles	`325412` cycles	`1.00`
`ML-DSA-87 sign`	`884468` cycles	`884484` cycles	`1.00`
`ML-DSA-87 verify`	`328634` cycles	`328655` cycles	`1.00`

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

github-actions

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

Benchmark suite	Current: `fa07d1f`	Previous: `100e446`	Ratio
`ML-DSA-44 keypair`	`115718` cycles	`115652` cycles	`1.00`
`ML-DSA-44 sign`	`377201` cycles	`377357` cycles	`1.00`
`ML-DSA-44 verify`	`120344` cycles	`120215` cycles	`1.00`
`ML-DSA-65 keypair`	`200127` cycles	`200073` cycles	`1.00`
`ML-DSA-65 sign`	`623016` cycles	`622766` cycles	`1.00`
`ML-DSA-65 verify`	`198223` cycles	`198195` cycles	`1.00`
`ML-DSA-87 keypair`	`327615` cycles	`326756` cycles	`1.00`
`ML-DSA-87 sign`	`791103` cycles	`789971` cycles	`1.00`
`ML-DSA-87 verify`	`325264` cycles	`324409` cycles	`1.00`

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

oqs-bot

Intel Xeon 4th gen (c7i)

Benchmark suite	Current: `fa07d1f`	Previous: `100e446`	Ratio
`ML-DSA-44 keypair`	`35631` cycles	`35116` cycles	`1.01`
`ML-DSA-44 sign`	`120705` cycles	`120958` cycles	`1.00`
`ML-DSA-44 verify`	`38074` cycles	`38274` cycles	`0.99`
`ML-DSA-65 keypair`	`61818` cycles	`62757` cycles	`0.99`
`ML-DSA-65 sign`	`199188` cycles	`201252` cycles	`0.99`
`ML-DSA-65 verify`	`62198` cycles	`62387` cycles	`1.00`
`ML-DSA-87 keypair`	`94415` cycles	`94461` cycles	`1.00`
`ML-DSA-87 sign`	`230678` cycles	`230993` cycles	`1.00`
`ML-DSA-87 verify`	`94054` cycles	`95279` cycles	`0.99`

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

oqs-bot

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

Benchmark suite	Current: `fa07d1f`	Previous: `100e446`	Ratio
`ML-DSA-44 keypair`	`95091` cycles	`95412` cycles	`1.00`
`ML-DSA-44 sign`	`349043` cycles	`349579` cycles	`1.00`
`ML-DSA-44 verify`	`100848` cycles	`101012` cycles	`1.00`
`ML-DSA-65 keypair`	`165116` cycles	`165049` cycles	`1.00`
`ML-DSA-65 sign`	`566948` cycles	`567954` cycles	`1.00`
`ML-DSA-65 verify`	`165483` cycles	`165700` cycles	`1.00`
`ML-DSA-87 keypair`	`267238` cycles	`267808` cycles	`1.00`
`ML-DSA-87 sign`	`723156` cycles	`723827` cycles	`1.00`
`ML-DSA-87 verify`	`272344` cycles	`272309` cycles	`1.00`

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

github-actions

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

Benchmark suite	Current: `fa07d1f`	Previous: `100e446`	Ratio
`ML-DSA-44 keypair`	`227025` cycles	`233223` cycles	`0.97`
`ML-DSA-44 sign`	`656604` cycles	`673055` cycles	`0.98`
`ML-DSA-44 verify`	`226548` cycles	`231292` cycles	`0.98`
`ML-DSA-65 keypair`	`399858` cycles	`399604` cycles	`1.00`
`ML-DSA-65 sign`	`1093277` cycles	`1092503` cycles	`1.00`
`ML-DSA-65 verify`	`382610` cycles	`378979` cycles	`1.01`
`ML-DSA-87 keypair`	`668662` cycles	`662585` cycles	`1.01`
`ML-DSA-87 sign`	`1457596` cycles	`1442394` cycles	`1.01`
`ML-DSA-87 verify`	`632700` cycles	`631363` 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)

Benchmark suite	Current: `fa07d1f`	Previous: `100e446`	Ratio
`ML-DSA-44 keypair`	`214068` cycles	`213795` cycles	`1.00`
`ML-DSA-44 sign`	`781499` cycles	`782133` cycles	`1.00`
`ML-DSA-44 verify`	`230065` cycles	`230257` cycles	`1.00`
`ML-DSA-65 keypair`	`385084` cycles	`385239` cycles	`1.00`
`ML-DSA-65 sign`	`1326386` cycles	`1314084` cycles	`1.01`
`ML-DSA-65 verify`	`375339` cycles	`375765` cycles	`1.00`
`ML-DSA-87 keypair`	`606587` cycles	`606848` cycles	`1.00`
`ML-DSA-87 sign`	`1621233` cycles	`1623082` cycles	`1.00`
`ML-DSA-87 verify`	`617288` cycles	`617742` cycles	`1.00`

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

oqs-bot

AMD EPYC 3rd gen (c6a)

Benchmark suite	Current: `fa07d1f`	Previous: `100e446`	Ratio
`ML-DSA-44 keypair`	`69198` cycles	`69604` cycles	`0.99`
`ML-DSA-44 sign`	`184949` cycles	`187462` cycles	`0.99`
`ML-DSA-44 verify`	`69047` cycles	`69269` cycles	`1.00`
`ML-DSA-65 keypair`	`120248` cycles	`119917` cycles	`1.00`
`ML-DSA-65 sign`	`295658` cycles	`297151` cycles	`0.99`
`ML-DSA-65 verify`	`115575` cycles	`115546` cycles	`1.00`
`ML-DSA-87 keypair`	`202548` cycles	`202342` cycles	`1.00`
`ML-DSA-87 sign`	`386766` cycles	`386965` cycles	`1.00`
`ML-DSA-87 verify`	`193569` cycles	`193643` cycles	`1.00`

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

oqs-bot

Intel Xeon 3rd gen (c6i)

Benchmark suite	Current: `fa07d1f`	Previous: `100e446`	Ratio
`ML-DSA-44 keypair`	`57336` cycles	`56956` cycles	`1.01`
`ML-DSA-44 sign`	`179376` cycles	`180499` cycles	`0.99`
`ML-DSA-44 verify`	`60900` cycles	`61247` cycles	`0.99`
`ML-DSA-65 keypair`	`99751` cycles	`99457` cycles	`1.00`
`ML-DSA-65 sign`	`296170` cycles	`296461` cycles	`1.00`
`ML-DSA-65 verify`	`99941` cycles	`100169` cycles	`1.00`
`ML-DSA-87 keypair`	`153766` cycles	`154195` cycles	`1.00`
`ML-DSA-87 sign`	`352782` cycles	`352935` cycles	`1.00`
`ML-DSA-87 verify`	`152815` cycles	`153194` cycles	`1.00`

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

oqs-bot

Graviton2

Benchmark suite	Current: `fa07d1f`	Previous: `100e446`	Ratio
`ML-DSA-44 keypair`	`116044` cycles	`116677` cycles	`0.99`
`ML-DSA-44 sign`	`377724` cycles	`379707` cycles	`0.99`
`ML-DSA-44 verify`	`120646` cycles	`121174` cycles	`1.00`
`ML-DSA-65 keypair`	`200451` cycles	`200327` cycles	`1.00`
`ML-DSA-65 sign`	`623509` cycles	`623378` cycles	`1.00`
`ML-DSA-65 verify`	`198593` cycles	`198489` cycles	`1.00`
`ML-DSA-87 keypair`	`328191` cycles	`327340` cycles	`1.00`
`ML-DSA-87 sign`	`792035` cycles	`790697` cycles	`1.00`
`ML-DSA-87 verify`	`325645` cycles	`324830` cycles	`1.00`

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

oqs-bot

AMD EPYC 4th gen (c7a)

Benchmark suite	Current: `fa07d1f`	Previous: `100e446`	Ratio
`ML-DSA-44 keypair`	`42462` cycles	`42166` cycles	`1.01`
`ML-DSA-44 sign`	`129986` cycles	`130558` cycles	`1.00`
`ML-DSA-44 verify`	`44008` cycles	`44242` cycles	`0.99`
`ML-DSA-65 keypair`	`72320` cycles	`72946` cycles	`0.99`
`ML-DSA-65 sign`	`210845` cycles	`210881` cycles	`1.00`
`ML-DSA-65 verify`	`72922` cycles	`72765` cycles	`1.00`
`ML-DSA-87 keypair`	`109431` cycles	`111282` cycles	`0.98`
`ML-DSA-87 sign`	`248355` cycles	`252306` cycles	`0.98`
`ML-DSA-87 verify`	`109568` cycles	`110921` cycles	`0.99`

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

oqs-bot

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

Benchmark suite	Current: `fa07d1f`	Previous: `100e446`	Ratio
`ML-DSA-44 keypair`	`135055` cycles	`136238` cycles	`0.99`
`ML-DSA-44 sign`	`539959` cycles	`544574` cycles	`0.99`
`ML-DSA-44 verify`	`148401` cycles	`149362` cycles	`0.99`
`ML-DSA-65 keypair`	`228535` cycles	`230475` cycles	`0.99`
`ML-DSA-65 sign`	`893053` cycles	`895163` cycles	`1.00`
`ML-DSA-65 verify`	`238247` cycles	`239847` cycles	`0.99`
`ML-DSA-87 keypair`	`373776` cycles	`376850` cycles	`0.99`
`ML-DSA-87 sign`	`1108012` cycles	`1112531` cycles	`1.00`
`ML-DSA-87 verify`	`387508` cycles	`389978` cycles	`0.99`

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

oqs-bot

Graviton3

Benchmark suite	Current: `fa07d1f`	Previous: `100e446`	Ratio
`ML-DSA-44 keypair`	`74308` cycles	`74255` cycles	`1.00`
`ML-DSA-44 sign`	`228603` cycles	`228755` cycles	`1.00`
`ML-DSA-44 verify`	`78250` cycles	`78127` cycles	`1.00`
`ML-DSA-65 keypair`	`130496` cycles	`130420` cycles	`1.00`
`ML-DSA-65 sign`	`378316` cycles	`378291` cycles	`1.00`
`ML-DSA-65 verify`	`129294` cycles	`129164` cycles	`1.00`
`ML-DSA-87 keypair`	`209590` cycles	`211688` cycles	`0.99`
`ML-DSA-87 sign`	`479315` cycles	`479661` cycles	`1.00`
`ML-DSA-87 verify`	`208641` cycles	`210182` cycles	`0.99`

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

oqs-bot

Graviton4

Benchmark suite	Current: `fa07d1f`	Previous: `100e446`	Ratio
`ML-DSA-44 keypair`	`70000` cycles	`69864` cycles	`1.00`
`ML-DSA-44 sign`	`214918` cycles	`215244` cycles	`1.00`
`ML-DSA-44 verify`	`72777` cycles	`72692` cycles	`1.00`
`ML-DSA-65 keypair`	`124033` cycles	`123579` cycles	`1.00`
`ML-DSA-65 sign`	`353286` cycles	`353468` cycles	`1.00`
`ML-DSA-65 verify`	`120824` cycles	`120718` cycles	`1.00`
`ML-DSA-87 keypair`	`202214` cycles	`201648` cycles	`1.00`
`ML-DSA-87 sign`	`451358` cycles	`451997` cycles	`1.00`
`ML-DSA-87 verify`	`198404` cycles	`198649` cycles	`1.00`

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

oqs-bot

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

Benchmark suite	Current: `fa07d1f`	Previous: `100e446`	Ratio
`ML-DSA-44 keypair`	`159164` cycles	`157982` cycles	`1.01`
`ML-DSA-44 sign`	`569001` cycles	`567777` cycles	`1.00`
`ML-DSA-44 verify`	`170753` cycles	`169763` cycles	`1.01`
`ML-DSA-65 keypair`	`271354` cycles	`271455` cycles	`1.00`
`ML-DSA-65 sign`	`926525` cycles	`925734` cycles	`1.00`
`ML-DSA-65 verify`	`275640` cycles	`275498` cycles	`1.00`
`ML-DSA-87 keypair`	`451014` cycles	`451543` cycles	`1.00`
`ML-DSA-87 sign`	`1182715` cycles	`1183249` cycles	`1.00`
`ML-DSA-87 verify`	`460835` cycles	`460624` cycles	`1.00`

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

oqs-bot

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

Benchmark suite	Current: `fa07d1f`	Previous: `100e446`	Ratio
`ML-DSA-44 keypair`	`120143` cycles	`120533` cycles	`1.00`
`ML-DSA-44 sign`	`453777` cycles	`456002` cycles	`1.00`
`ML-DSA-44 verify`	`130320` cycles	`132129` cycles	`0.99`
`ML-DSA-65 keypair`	`204830` cycles	`207895` cycles	`0.99`
`ML-DSA-65 sign`	`732904` cycles	`742729` cycles	`0.99`
`ML-DSA-65 verify`	`209363` cycles	`211005` cycles	`0.99`
`ML-DSA-87 keypair`	`337665` cycles	`337927` cycles	`1.00`
`ML-DSA-87 sign`	`923416` cycles	`923041` cycles	`1.00`
`ML-DSA-87 verify`	`344913` cycles	`345844` cycles	`1.00`

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

oqs-bot

Graviton2 (no-opt)

Benchmark suite	Current: `fa07d1f`	Previous: `100e446`	Ratio
`ML-DSA-44 keypair`	`214842` cycles	`214202` cycles	`1.00`
`ML-DSA-44 sign`	`782554` cycles	`794999` cycles	`0.98`
`ML-DSA-44 verify`	`230631` cycles	`229962` cycles	`1.00`
`ML-DSA-65 keypair`	`385817` cycles	`385876` cycles	`1.00`
`ML-DSA-65 sign`	`1310148` cycles	`1307768` cycles	`1.00`
`ML-DSA-65 verify`	`376009` cycles	`376256` cycles	`1.00`
`ML-DSA-87 keypair`	`607294` cycles	`607000` cycles	`1.00`
`ML-DSA-87 sign`	`1624685` cycles	`1625772` cycles	`1.00`
`ML-DSA-87 verify`	`617770` cycles	`617491` cycles	`1.00`

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

oqs-bot

Graviton3 (no-opt)

Benchmark suite	Current: `fa07d1f`	Previous: `100e446`	Ratio
`ML-DSA-44 keypair`	`138816` cycles	`138783` cycles	`1.00`
`ML-DSA-44 sign`	`493083` cycles	`493854` cycles	`1.00`
`ML-DSA-44 verify`	`148367` cycles	`148389` cycles	`1.00`
`ML-DSA-65 keypair`	`242529` cycles	`242264` cycles	`1.00`
`ML-DSA-65 sign`	`809972` cycles	`809969` cycles	`1.00`
`ML-DSA-65 verify`	`240719` cycles	`240614` cycles	`1.00`
`ML-DSA-87 keypair`	`396675` cycles	`396621` cycles	`1.00`
`ML-DSA-87 sign`	`1027482` cycles	`1027277` cycles	`1.00`
`ML-DSA-87 verify`	`401597` cycles	`401369` cycles	`1.00`

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

oqs-bot

Graviton4 (no-opt)

Benchmark suite	Current: `fa07d1f`	Previous: `100e446`	Ratio
`ML-DSA-44 keypair`	`133144` cycles	`133258` cycles	`1.00`
`ML-DSA-44 sign`	`498479` cycles	`498179` cycles	`1.00`
`ML-DSA-44 verify`	`144897` cycles	`144918` cycles	`1.00`
`ML-DSA-65 keypair`	`227070` cycles	`226755` cycles	`1.00`
`ML-DSA-65 sign`	`812705` cycles	`812078` cycles	`1.00`
`ML-DSA-65 verify`	`231517` cycles	`231580` cycles	`1.00`
`ML-DSA-87 keypair`	`374798` cycles	`375108` cycles	`1.00`
`ML-DSA-87 sign`	`1020759` cycles	`1020839` cycles	`1.00`
`ML-DSA-87 verify`	`383690` cycles	`383524` cycles	`1.00`

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

github-actions

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

Benchmark suite	Current: `fa07d1f`	Previous: `100e446`	Ratio
`ML-DSA-44 keypair`	`289514` cycles	`290746` cycles	`1.00`
`ML-DSA-44 sign`	`930458` cycles	`937533` cycles	`0.99`
`ML-DSA-44 verify`	`291385` cycles	`291943` cycles	`1.00`
`ML-DSA-65 keypair`	`491840` cycles	`493090` cycles	`1.00`
`ML-DSA-65 sign`	`1538201` cycles	`1526359` cycles	`1.01`
`ML-DSA-65 verify`	`477106` cycles	`476058` cycles	`1.00`
`ML-DSA-87 keypair`	`833556` cycles	`843754` cycles	`0.99`
`ML-DSA-87 sign`	`2048886` cycles	`2088455` cycles	`0.98`
`ML-DSA-87 verify`	`813904` cycles	`818519` cycles	`0.99`

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

github-actions

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

Benchmark suite	Current: `fa07d1f`	Previous: `100e446`	Ratio
`ML-DSA-44 keypair`	`309445` cycles	`304061` cycles	`1.02`
`ML-DSA-44 sign`	`1226595` cycles	`1204370` cycles	`1.02`
`ML-DSA-44 verify`	`347072` cycles	`331394` cycles	`1.05`
`ML-DSA-65 keypair`	`574923` cycles	`577955` cycles	`0.99`
`ML-DSA-65 sign`	`2020807` cycles	`1998603` cycles	`1.01`
`ML-DSA-65 verify`	`550026` cycles	`552006` cycles	`1.00`
`ML-DSA-87 keypair`	`869120` cycles	`870486` cycles	`1.00`
`ML-DSA-87 sign`	`2508517` cycles	`2493534` cycles	`1.01`
`ML-DSA-87 verify`	`894364` cycles	`896546` 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 '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: `fa07d1f`	Previous: `100e446`	Ratio
`ML-DSA-44 verify`	`347072` cycles	`331394` cycles	`1.05`

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

mkannwischer

Thanks @jammychiou1. I am good with the changes now.

WDYT @hanno-becker?

dev/x86_64/src/poly_decompose_32_avx2.c

dev/x86_64/src/poly_decompose_88_avx2.c

hanno-becker

Overall LGTM and really helps understanding the code, thank you @jammychiou1. A few smaller change requests, see comments.

jammychiou1 · 2025-11-11T06:28:24Z

Thank you @hanno-becker for your suggestions (and your wonderful work on #659)!

Please let me know if there's more things to change. If not, I'll clean up the commit history to prepare for the merge into main.

mldsa/src/rounding.h

The new approach is adapted from our Neon implementation. See <#411 (comment)> for more information on the idea. Bounds reasoning comments are also added. Signed-off-by: jammychiou1 <[email protected]>

Edit some comments while we're at it. Signed-off-by: jammychiou1 <[email protected]>

Signed-off-by: jammychiou1 <[email protected]>

hanno-becker

LGTM, Thank you @jammychiou1

jammychiou1 mentioned this pull request Nov 6, 2025

Add bounds reasoning comments to AVX2 [I]NTT and basemul #560

Merged

jammychiou1 marked this pull request as ready for review November 7, 2025 01:53

jammychiou1 requested a review from a team as a code owner November 7, 2025 01:53

mkannwischer requested changes Nov 7, 2025

View reviewed changes

dev/x86_64/src/poly_decompose_32_avx2.c Show resolved Hide resolved

dev/x86_64/src/poly_decompose_32_avx2.c Outdated Show resolved Hide resolved

jammychiou1 force-pushed the decompose-explanation branch 2 times, most recently from 5082756 to fa07d1f Compare November 8, 2025 08:13

jammychiou1 added the benchmark label Nov 8, 2025

github-actions bot reviewed Nov 8, 2025

View reviewed changes

oqs-bot reviewed Nov 8, 2025

View reviewed changes

jammychiou1 changed the title ~~Add bounds reasoning comments to AVX2 decompose~~ Update AVX2 decompose to use a more explainable (and very slightly faster) approach, along with bounds reasoning comments. Nov 8, 2025

github-actions bot reviewed Nov 8, 2025

View reviewed changes

oqs-bot reviewed Nov 8, 2025

View reviewed changes

github-actions bot reviewed Nov 8, 2025

View reviewed changes

mkannwischer approved these changes Nov 8, 2025

View reviewed changes

mkannwischer mentioned this pull request Nov 10, 2025

Perform liboqs benchmarks #546

Closed