Add bounds reasoning comments to AArch64 backend #667

jammychiou1 · 2025-11-11T09:05:49Z

~~Towards resolving Add bounds reasoning comments to AArch64 backend #529. Still need to do decompose.~~
Resolves Add bounds reasoning comments to AArch64 backend #529

Note that the Montgomery reduction within the basemuls is changed to the positive/subtractive variant, in order to match mld_montgomery_reduce() and reuse the same bound (and reasoning, when we add it later in #602). The performance is expected to be identical.

github-actions

Mac Mini (M1, 2020) benchmarks (opt)

Benchmark suite	Current: `37d7c09`	Previous: `314ad8e`	Ratio
`ML-DSA-44 keypair`	`46421` cycles	`46419` cycles	`1.00`
`ML-DSA-44 sign`	`132731` cycles	`132731` cycles	`1`
`ML-DSA-44 verify`	`47842` cycles	`47839` cycles	`1.00`
`ML-DSA-65 keypair`	`81446` cycles	`81443` cycles	`1.00`
`ML-DSA-65 sign`	`219217` cycles	`219226` cycles	`1.00`
`ML-DSA-65 verify`	`80140` cycles	`80140` cycles	`1`
`ML-DSA-87 keypair`	`132750` cycles	`132771` cycles	`1.00`
`ML-DSA-87 sign`	`280896` cycles	`280924` cycles	`1.00`
`ML-DSA-87 verify`	`130322` cycles	`130330` 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: `37d7c09`	Previous: `314ad8e`	Ratio
`ML-DSA-44 keypair`	`115261` cycles	`115269` cycles	`1.00`
`ML-DSA-44 sign`	`431714` cycles	`431703` cycles	`1.00`
`ML-DSA-44 verify`	`122155` cycles	`122175` cycles	`1.00`
`ML-DSA-65 keypair`	`197428` cycles	`197429` cycles	`1.00`
`ML-DSA-65 sign`	`700994` cycles	`700966` cycles	`1.00`
`ML-DSA-65 verify`	`197676` cycles	`197672` cycles	`1.00`
`ML-DSA-87 keypair`	`325419` cycles	`325394` cycles	`1.00`
`ML-DSA-87 sign`	`884517` cycles	`884436` cycles	`1.00`
`ML-DSA-87 verify`	`328665` cycles	`328656` cycles	`1.00`

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

oqs-bot

Intel Xeon 4th gen (c7i)

Benchmark suite	Current: `37d7c09`	Previous: `314ad8e`	Ratio
`ML-DSA-44 keypair`	`35434` cycles	`35082` cycles	`1.01`
`ML-DSA-44 sign`	`120642` cycles	`120829` cycles	`1.00`
`ML-DSA-44 verify`	`38292` cycles	`38217` cycles	`1.00`
`ML-DSA-65 keypair`	`63150` cycles	`62792` cycles	`1.01`
`ML-DSA-65 sign`	`201518` cycles	`201578` cycles	`1.00`
`ML-DSA-65 verify`	`63025` cycles	`62619` cycles	`1.01`
`ML-DSA-87 keypair`	`95825` cycles	`94417` cycles	`1.01`
`ML-DSA-87 sign`	`235564` cycles	`230458` cycles	`1.02`
`ML-DSA-87 verify`	`94899` cycles	`93737` cycles	`1.01`

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

oqs-bot

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

Benchmark suite	Current: `37d7c09`	Previous: `314ad8e`	Ratio
`ML-DSA-44 keypair`	`95193` cycles	`95251` cycles	`1.00`
`ML-DSA-44 sign`	`349301` cycles	`349405` cycles	`1.00`
`ML-DSA-44 verify`	`100948` cycles	`100993` cycles	`1.00`
`ML-DSA-65 keypair`	`165304` cycles	`164441` cycles	`1.01`
`ML-DSA-65 sign`	`567264` cycles	`567124` cycles	`1.00`
`ML-DSA-65 verify`	`165488` cycles	`165226` cycles	`1.00`
`ML-DSA-87 keypair`	`268465` cycles	`268259` cycles	`1.00`
`ML-DSA-87 sign`	`724063` cycles	`724173` cycles	`1.00`
`ML-DSA-87 verify`	`272280` cycles	`272251` 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: `37d7c09`	Previous: `314ad8e`	Ratio
`ML-DSA-44 keypair`	`226343` cycles	`235855` cycles	`0.96`
`ML-DSA-44 sign`	`669578` cycles	`697418` cycles	`0.96`
`ML-DSA-44 verify`	`227619` cycles	`235679` cycles	`0.97`
`ML-DSA-65 keypair`	`413279` cycles	`412522` cycles	`1.00`
`ML-DSA-65 sign`	`1121384` cycles	`1115820` cycles	`1.00`
`ML-DSA-65 verify`	`395696` cycles	`392924` cycles	`1.01`
`ML-DSA-87 keypair`	`675837` cycles	`654605` cycles	`1.03`
`ML-DSA-87 sign`	`1486942` cycles	`1432265` cycles	`1.04`
`ML-DSA-87 verify`	`653743` cycles	`637053` 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 (opt)'.
Benchmark result of this commit is worse than the previous benchmark result exceeding threshold 1.03.

Benchmark suite	Current: `37d7c09`	Previous: `314ad8e`	Ratio
`ML-DSA-87 keypair`	`675837` cycles	`654605` cycles	`1.03`
`ML-DSA-87 sign`	`1486942` cycles	`1432265` cycles	`1.04`

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

oqs-bot

AMD EPYC 3rd gen (c6a)

Benchmark suite	Current: `37d7c09`	Previous: `314ad8e`	Ratio
`ML-DSA-44 keypair`	`69634` cycles	`69408` cycles	`1.00`
`ML-DSA-44 sign`	`185866` cycles	`185588` cycles	`1.00`
`ML-DSA-44 verify`	`69632` cycles	`69022` cycles	`1.01`
`ML-DSA-65 keypair`	`119519` cycles	`119527` cycles	`1.00`
`ML-DSA-65 sign`	`296085` cycles	`296151` cycles	`1.00`
`ML-DSA-65 verify`	`115496` cycles	`115310` cycles	`1.00`
`ML-DSA-87 keypair`	`201515` cycles	`201740` cycles	`1.00`
`ML-DSA-87 sign`	`385521` cycles	`385807` cycles	`1.00`
`ML-DSA-87 verify`	`193326` cycles	`193642` cycles	`1.00`

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

oqs-bot

Graviton2

Benchmark suite	Current: `37d7c09`	Previous: `314ad8e`	Ratio
`ML-DSA-44 keypair`	`117109` cycles	`115867` cycles	`1.01`
`ML-DSA-44 sign`	`380933` cycles	`377525` cycles	`1.01`
`ML-DSA-44 verify`	`121685` cycles	`120368` cycles	`1.01`
`ML-DSA-65 keypair`	`200537` cycles	`200218` cycles	`1.00`
`ML-DSA-65 sign`	`623530` cycles	`623046` cycles	`1.00`
`ML-DSA-65 verify`	`198644` cycles	`198371` cycles	`1.00`
`ML-DSA-87 keypair`	`328184` cycles	`327430` cycles	`1.00`
`ML-DSA-87 sign`	`792111` cycles	`790806` cycles	`1.00`
`ML-DSA-87 verify`	`325565` cycles	`324857` cycles	`1.00`

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

oqs-bot

AMD EPYC 4th gen (c7a)

Benchmark suite	Current: `37d7c09`	Previous: `314ad8e`	Ratio
`ML-DSA-44 keypair`	`42714` cycles	`42050` cycles	`1.02`
`ML-DSA-44 sign`	`130840` cycles	`130623` cycles	`1.00`
`ML-DSA-44 verify`	`44222` cycles	`44275` cycles	`1.00`
`ML-DSA-65 keypair`	`72977` cycles	`72347` cycles	`1.01`
`ML-DSA-65 sign`	`210931` cycles	`211393` cycles	`1.00`
`ML-DSA-65 verify`	`73284` cycles	`72769` cycles	`1.01`
`ML-DSA-87 keypair`	`110228` cycles	`109988` cycles	`1.00`
`ML-DSA-87 sign`	`248784` cycles	`248764` cycles	`1.00`
`ML-DSA-87 verify`	`111640` cycles	`109522` cycles	`1.02`

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

oqs-bot

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

Benchmark suite	Current: `37d7c09`	Previous: `314ad8e`	Ratio
`ML-DSA-44 keypair`	`135099` cycles	`135380` cycles	`1.00`
`ML-DSA-44 sign`	`540292` cycles	`539702` cycles	`1.00`
`ML-DSA-44 verify`	`148192` cycles	`148584` cycles	`1.00`
`ML-DSA-65 keypair`	`228667` cycles	`228833` cycles	`1.00`
`ML-DSA-65 sign`	`890151` cycles	`892086` cycles	`1.00`
`ML-DSA-65 verify`	`238302` cycles	`238446` cycles	`1.00`
`ML-DSA-87 keypair`	`373734` cycles	`372875` cycles	`1.00`
`ML-DSA-87 sign`	`1105910` cycles	`1106065` cycles	`1.00`
`ML-DSA-87 verify`	`387102` cycles	`387563` cycles	`1.00`

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

oqs-bot

Graviton4

Benchmark suite	Current: `37d7c09`	Previous: `314ad8e`	Ratio
`ML-DSA-44 keypair`	`69576` cycles	`69610` cycles	`1.00`
`ML-DSA-44 sign`	`213709` cycles	`213787` cycles	`1.00`
`ML-DSA-44 verify`	`72626` cycles	`72475` cycles	`1.00`
`ML-DSA-65 keypair`	`123669` cycles	`123352` cycles	`1.00`
`ML-DSA-65 sign`	`350942` cycles	`350469` cycles	`1.00`
`ML-DSA-65 verify`	`120890` cycles	`120415` cycles	`1.00`
`ML-DSA-87 keypair`	`202348` cycles	`201192` cycles	`1.01`
`ML-DSA-87 sign`	`450126` cycles	`449091` cycles	`1.00`
`ML-DSA-87 verify`	`198811` cycles	`197923` 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: `37d7c09`	Previous: `314ad8e`	Ratio
`ML-DSA-44 keypair`	`120412` cycles	`120466` cycles	`1.00`
`ML-DSA-44 sign`	`453079` cycles	`455604` cycles	`0.99`
`ML-DSA-44 verify`	`129899` cycles	`129917` cycles	`1.00`
`ML-DSA-65 keypair`	`205074` cycles	`205155` cycles	`1.00`
`ML-DSA-65 sign`	`734147` cycles	`735520` cycles	`1.00`
`ML-DSA-65 verify`	`210029` cycles	`209895` cycles	`1.00`
`ML-DSA-87 keypair`	`339889` cycles	`337601` cycles	`1.01`
`ML-DSA-87 sign`	`931618` cycles	`924871` cycles	`1.01`
`ML-DSA-87 verify`	`347721` cycles	`345776` cycles	`1.01`

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

oqs-bot

Graviton3

Benchmark suite	Current: `37d7c09`	Previous: `314ad8e`	Ratio
`ML-DSA-44 keypair`	`74307` cycles	`74268` cycles	`1.00`
`ML-DSA-44 sign`	`228607` cycles	`228732` cycles	`1.00`
`ML-DSA-44 verify`	`78259` cycles	`78126` cycles	`1.00`
`ML-DSA-65 keypair`	`130499` cycles	`130397` cycles	`1.00`
`ML-DSA-65 sign`	`378327` cycles	`378266` cycles	`1.00`
`ML-DSA-65 verify`	`129295` cycles	`129145` cycles	`1.00`
`ML-DSA-87 keypair`	`209579` cycles	`211710` cycles	`0.99`
`ML-DSA-87 sign`	`479323` cycles	`479642` cycles	`1.00`
`ML-DSA-87 verify`	`208629` cycles	`210191` cycles	`0.99`

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

oqs-bot

Graviton2 (no-opt)

Benchmark suite	Current: `37d7c09`	Previous: `314ad8e`	Ratio
`ML-DSA-44 keypair`	`214694` cycles	`214295` cycles	`1.00`
`ML-DSA-44 sign`	`782368` cycles	`794948` cycles	`0.98`
`ML-DSA-44 verify`	`230498` cycles	`230029` cycles	`1.00`
`ML-DSA-65 keypair`	`385625` cycles	`385839` cycles	`1.00`
`ML-DSA-65 sign`	`1309399` cycles	`1307192` cycles	`1.00`
`ML-DSA-65 verify`	`375860` cycles	`376224` cycles	`1.00`
`ML-DSA-87 keypair`	`607551` cycles	`606942` cycles	`1.00`
`ML-DSA-87 sign`	`1624842` cycles	`1625604` cycles	`1.00`
`ML-DSA-87 verify`	`618005` cycles	`617304` cycles	`1.00`

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

oqs-bot

Intel Xeon 3rd gen (c6i)

Benchmark suite	Current: `37d7c09`	Previous: `314ad8e`	Ratio
`ML-DSA-44 keypair`	`57069` cycles	`57296` cycles	`1.00`
`ML-DSA-44 sign`	`180312` cycles	`180557` cycles	`1.00`
`ML-DSA-44 verify`	`61275` cycles	`61279` cycles	`1.00`
`ML-DSA-65 keypair`	`99888` cycles	`99806` cycles	`1.00`
`ML-DSA-65 sign`	`296356` cycles	`296199` cycles	`1.00`
`ML-DSA-65 verify`	`100376` cycles	`100207` cycles	`1.00`
`ML-DSA-87 keypair`	`154114` cycles	`154547` cycles	`1.00`
`ML-DSA-87 sign`	`353395` cycles	`352845` cycles	`1.00`
`ML-DSA-87 verify`	`152549` cycles	`152958` cycles	`1.00`

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

github-actions

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

Benchmark suite	Current: `37d7c09`	Previous: `314ad8e`	Ratio
`ML-DSA-44 keypair`	`826071` cycles	`824630` cycles	`1.00`
`ML-DSA-44 sign`	`3328025` cycles	`3328785` cycles	`1.00`
`ML-DSA-44 verify`	`919467` cycles	`918716` cycles	`1.00`
`ML-DSA-65 keypair`	`1403179` cycles	`1404566` cycles	`1.00`
`ML-DSA-65 sign`	`5468362` cycles	`5454585` cycles	`1.00`
`ML-DSA-65 verify`	`1466427` cycles	`1465675` cycles	`1.00`
`ML-DSA-87 keypair`	`2303521` cycles	`2303418` cycles	`1.00`
`ML-DSA-87 sign`	`6816366` cycles	`6808316` cycles	`1.00`
`ML-DSA-87 verify`	`2403866` cycles	`2402810` cycles	`1.00`

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

oqs-bot

Graviton4 (no-opt)

Benchmark suite	Current: `37d7c09`	Previous: `314ad8e`	Ratio
`ML-DSA-44 keypair`	`128332` cycles	`128280` cycles	`1.00`
`ML-DSA-44 sign`	`456357` cycles	`457208` cycles	`1.00`
`ML-DSA-44 verify`	`136286` cycles	`136322` cycles	`1.00`
`ML-DSA-65 keypair`	`220788` cycles	`220689` cycles	`1.00`
`ML-DSA-65 sign`	`746911` cycles	`746178` cycles	`1.00`
`ML-DSA-65 verify`	`220358` cycles	`220384` cycles	`1.00`
`ML-DSA-87 keypair`	`365198` cycles	`365074` cycles	`1.00`
`ML-DSA-87 sign`	`944750` cycles	`944382` cycles	`1.00`
`ML-DSA-87 verify`	`368963` cycles	`368862` cycles	`1.00`

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

oqs-bot

Graviton3 (no-opt)

Benchmark suite	Current: `37d7c09`	Previous: `314ad8e`	Ratio
`ML-DSA-44 keypair`	`138815` cycles	`138766` cycles	`1.00`
`ML-DSA-44 sign`	`493048` cycles	`493633` cycles	`1.00`
`ML-DSA-44 verify`	`148351` cycles	`148356` cycles	`1.00`
`ML-DSA-65 keypair`	`242409` cycles	`242242` cycles	`1.00`
`ML-DSA-65 sign`	`809815` cycles	`809933` cycles	`1.00`
`ML-DSA-65 verify`	`240722` cycles	`240609` cycles	`1.00`
`ML-DSA-87 keypair`	`396739` cycles	`396619` cycles	`1.00`
`ML-DSA-87 sign`	`1027495` cycles	`1027094` cycles	`1.00`
`ML-DSA-87 verify`	`401595` cycles	`401371` 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: `37d7c09`	Previous: `314ad8e`	Ratio
`ML-DSA-44 keypair`	`159949` cycles	`157993` cycles	`1.01`
`ML-DSA-44 sign`	`571755` cycles	`566309` cycles	`1.01`
`ML-DSA-44 verify`	`172156` cycles	`169599` cycles	`1.02`
`ML-DSA-65 keypair`	`271448` cycles	`271036` cycles	`1.00`
`ML-DSA-65 sign`	`926175` cycles	`925911` cycles	`1.00`
`ML-DSA-65 verify`	`276594` cycles	`275788` cycles	`1.00`
`ML-DSA-87 keypair`	`451462` cycles	`451772` cycles	`1.00`
`ML-DSA-87 sign`	`1183639` cycles	`1183487` cycles	`1.00`
`ML-DSA-87 verify`	`461547` cycles	`460706` cycles	`1.00`

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: `37d7c09`	Previous: `314ad8e`	Ratio
`ML-DSA-44 keypair`	`309643` cycles	`304613` cycles	`1.02`
`ML-DSA-44 sign`	`1177956` cycles	`1238978` cycles	`0.95`
`ML-DSA-44 verify`	`327779` cycles	`342557` cycles	`0.96`
`ML-DSA-65 keypair`	`558913` cycles	`574348` cycles	`0.97`
`ML-DSA-65 sign`	`1934782` cycles	`2000543` cycles	`0.97`
`ML-DSA-65 verify`	`521783` cycles	`544249` cycles	`0.96`
`ML-DSA-87 keypair`	`865936` cycles	`896114` cycles	`0.97`
`ML-DSA-87 sign`	`2492678` cycles	`2558078` cycles	`0.97`
`ML-DSA-87 verify`	`895228` cycles	`929425` cycles	`0.96`

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

github-actions

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

Benchmark suite	Current: `37d7c09`	Previous: `314ad8e`	Ratio
`ML-DSA-44 keypair`	`115721` cycles	`115633` cycles	`1.00`
`ML-DSA-44 sign`	`377208` cycles	`377229` cycles	`1.00`
`ML-DSA-44 verify`	`120332` cycles	`120215` cycles	`1.00`
`ML-DSA-65 keypair`	`200117` cycles	`200075` cycles	`1.00`
`ML-DSA-65 sign`	`622821` cycles	`622903` cycles	`1.00`
`ML-DSA-65 verify`	`198196` cycles	`198200` cycles	`1.00`
`ML-DSA-87 keypair`	`327645` cycles	`326771` cycles	`1.00`
`ML-DSA-87 sign`	`791183` cycles	`789996` cycles	`1.00`
`ML-DSA-87 verify`	`325316` cycles	`324410` 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: `37d7c09`	Previous: `314ad8e`	Ratio
`ML-DSA-44 keypair`	`214094` cycles	`213856` cycles	`1.00`
`ML-DSA-44 sign`	`781550` cycles	`782246` cycles	`1.00`
`ML-DSA-44 verify`	`230039` cycles	`230309` cycles	`1.00`
`ML-DSA-65 keypair`	`385262` cycles	`385241` cycles	`1.00`
`ML-DSA-65 sign`	`1327081` cycles	`1314056` cycles	`1.01`
`ML-DSA-65 verify`	`375526` cycles	`375747` cycles	`1.00`
`ML-DSA-87 keypair`	`606484` cycles	`606555` cycles	`1.00`
`ML-DSA-87 sign`	`1621418` cycles	`1622829` cycles	`1.00`
`ML-DSA-87 verify`	`617191` cycles	`617434` 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: `37d7c09`	Previous: `314ad8e`	Ratio
`ML-DSA-44 keypair`	`292600` cycles	`292046` cycles	`1.00`
`ML-DSA-44 sign`	`929575` cycles	`928646` cycles	`1.00`
`ML-DSA-44 verify`	`296239` cycles	`295851` cycles	`1.00`
`ML-DSA-65 keypair`	`492833` cycles	`494958` cycles	`1.00`
`ML-DSA-65 sign`	`1524447` cycles	`1521425` cycles	`1.00`
`ML-DSA-65 verify`	`481202` cycles	`484056` cycles	`0.99`
`ML-DSA-87 keypair`	`838549` cycles	`846764` cycles	`0.99`
`ML-DSA-87 sign`	`2061507` cycles	`2058225` cycles	`1.00`
`ML-DSA-87 verify`	`820738` cycles	`820143` cycles	`1.00`

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

github-actions

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

Benchmark suite	Current: `37d7c09`	Previous: `314ad8e`	Ratio
`ML-DSA-44 keypair`	`469687` cycles	`469757` cycles	`1.00`
`ML-DSA-44 sign`	`2212428` cycles	`2214917` cycles	`1.00`
`ML-DSA-44 verify`	`550456` cycles	`550782` cycles	`1.00`
`ML-DSA-65 keypair`	`782601` cycles	`782351` cycles	`1.00`
`ML-DSA-65 sign`	`3630929` cycles	`3630275` cycles	`1.00`
`ML-DSA-65 verify`	`853006` cycles	`848857` cycles	`1.00`
`ML-DSA-87 keypair`	`1263087` cycles	`1268638` cycles	`1.00`
`ML-DSA-87 sign`	`4489471` cycles	`4502481` cycles	`1.00`
`ML-DSA-87 verify`	`1373034` cycles	`1369173` cycles	`1.00`

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

dev/aarch64_clean/src/mld_polyvecl_pointwise_acc_montgomery_l4.S

dev/aarch64_clean/src/intt.S

hanno-becker · 2025-11-11T17:14:10Z

Stopping CI for now to free resources for #668 and, then, main. I believe we need to explicitly trigger a nix cache build on main once #668 is merged.

hanno-becker · 2025-11-11T17:18:02Z

dev/aarch64_clean/src/mld_polyvecl_pointwise_acc_montgomery_l4.S

    pmull v24, v25, v16, v20
    pmull v26, v27, v17, v21
    pmull v28, v29, v18, v22
    pmull v30, v31, v19, v23


General comment, I didn't notice this when the code was introduced: We should really use symbolic register values (using .req), not architectural ones. It's much easier to read.

Good idea. Thanks!

While coming up with names for the 4 registers for coefficients from a and b, I wanted to use a0, ..., a3, b0, ..., b3. However b0, ..., b3 are built-in register names already.

I'll resort to use a_0, ..., a_3, b_0, ..., b_3 for now. Any other naming schemes are appreciated!

hanno-becker · 2025-11-11T17:27:39Z

dev/aarch64_clean/src/mld_polyvecl_pointwise_acc_montgomery_l4.S

-    // load -q^-1 = 4236238847
-    movz wtmp, #57343
-    movk wtmp, #64639, lsl #16
+    // load q^-1 = 58728449


We don't support it yet in assembly files, but can you add

/* check-magic: 58728449 == unsigned_mod(pow(MLDSA_Q, -1, 2^32), 2^32) */

Good idea. Working on it now.

Signed-off-by: jammychiou1 <jammy.chiou1@gmail.com>

The Montgomery reduction is changed to the positive/subtractive variant to match mld_montgomery_reduce(). This avoids the need to clarify that the positive and the negative variant have exact same bounds despite their different appearance. Signed-off-by: jammychiou1 <jammy.chiou1@gmail.com>

Signed-off-by: jammychiou1 <jammy.chiou1@gmail.com>

jammychiou1 · 2025-11-12T03:36:26Z

Thank you @hanno-becker and @mkannwischer for your suggestions. I've implemented them in the new patch.

I also included some annotation for the range of intermediates inside decompose.

jammychiou1 added the benchmark label Nov 11, 2025