Reduce unrolling in Panama dotProduct float variant

[ChrisHegarty]

Reduce unrolling in Panama dotProduct float variants.

[msokolov]

Q: were you able to confirm what happens on ARM ?

[ChrisHegarty]

yeah, I do have some ARM results. I'll post them shortly.

[rmuir]

You have rocket lake right? with only 2 fma units? so the 2x may work fine for you because of that.

I haven't looked at the assembly, but if the jvm is unrolling 4x, we shouldnt need to unroll at all to keep your CPU busy. Last time i checked, it didnt do this.

I can run the script in the repo against various aws instances so that we are sure.

[rmuir]

did a quick run: looks ok, but it hurts the haswell, zen2, and zen3. I'll do a pass on the instance types, since they are a bit outdated and try to bring them up to speed (e.g. no graviton4 represented).

cascadelake: ['0', 'GenuineIntel', 'Intel(R) Xeon(R) Platinum 8275CL CPU @ 3.00GHz', '1', 'GenuineIntel', 'Intel(R) Xeon(R) Platinum 8275CL CPU @ 3.00GHz']

main

Benchmark                                  (size)   Mode  Cnt   Score   Error   Units
VectorUtilBenchmark.floatDotProductVector    1024  thrpt   75  14.038 ± 0.140  ops/us

patch

Benchmark                                  (size)   Mode  Cnt   Score   Error   Units
VectorUtilBenchmark.floatDotProductVector    1024  thrpt   75  14.229 ± 0.092  ops/us

graviton2: ['0', '1']

main

Benchmark                                  (size)   Mode  Cnt  Score   Error   Units
VectorUtilBenchmark.floatDotProductVector    1024  thrpt   75  7.671 ± 0.040  ops/us

patch

Benchmark                                  (size)   Mode  Cnt  Score   Error   Units
VectorUtilBenchmark.floatDotProductVector    1024  thrpt   75  8.116 ± 0.172  ops/us

graviton3: ['0', '1']

main

Benchmark                                  (size)   Mode  Cnt   Score   Error   Units
VectorUtilBenchmark.floatDotProductVector    1024  thrpt   75  10.894 ± 0.233  ops/us

patch

Benchmark                                  (size)   Mode  Cnt   Score   Error   Units
VectorUtilBenchmark.floatDotProductVector    1024  thrpt   75  11.194 ± 0.241  ops/us

haswell: ['0', 'GenuineIntel', 'Intel(R) Xeon(R) CPU E5-2666 v3 @ 2.90GHz', '1', 'GenuineIntel', 'Intel(R) Xeon(R) CPU E5-2666 v3 @ 2.90GHz']

main

Benchmark                                  (size)   Mode  Cnt   Score   Error   Units
VectorUtilBenchmark.floatDotProductVector    1024  thrpt   75  13.091 ± 0.079  ops/us

patch

Benchmark                                  (size)   Mode  Cnt   Score   Error   Units
VectorUtilBenchmark.floatDotProductVector    1024  thrpt   75  11.177 ± 0.046  ops/us

icelake: ['0', 'GenuineIntel', 'Intel(R) Xeon(R) Platinum 8375C CPU @ 2.90GHz', '1', 'GenuineIntel', 'Intel(R) Xeon(R) Platinum 8375C CPU @ 2.90GHz']

main

Benchmark                                  (size)   Mode  Cnt   Score   Error   Units
VectorUtilBenchmark.floatDotProductVector    1024  thrpt   75  14.924 ± 0.777  ops/us

patch

Benchmark                                  (size)   Mode  Cnt   Score   Error   Units
VectorUtilBenchmark.floatDotProductVector    1024  thrpt   75  15.803 ± 0.974  ops/us

sapphirerapids: ['0', 'GenuineIntel', 'Intel(R) Xeon(R) Platinum 8488C', '1', 'GenuineIntel', 'Intel(R) Xeon(R) Platinum 8488C']

main

Benchmark                                  (size)   Mode  Cnt   Score   Error   Units
VectorUtilBenchmark.floatDotProductVector    1024  thrpt   75  20.011 ± 0.716  ops/us

patch

Benchmark                                  (size)   Mode  Cnt   Score   Error   Units
VectorUtilBenchmark.floatDotProductVector    1024  thrpt   75  21.082 ± 1.011  ops/us

zen2: ['0', 'AuthenticAMD', 'AMD EPYC 7R32', '1', 'AuthenticAMD', 'AMD EPYC 7R32']

main

Benchmark                                  (size)   Mode  Cnt   Score   Error   Units
VectorUtilBenchmark.floatDotProductVector    1024  thrpt   75  16.704 ± 0.239  ops/us

patch

Benchmark                                  (size)   Mode  Cnt   Score   Error   Units
VectorUtilBenchmark.floatDotProductVector    1024  thrpt   75  10.656 ± 0.016  ops/us

zen3: ['0', 'AuthenticAMD', 'AMD EPYC 7R13 Processor', '1', 'AuthenticAMD', 'AMD EPYC 7R13 Processor']

main

Benchmark                                  (size)   Mode  Cnt   Score   Error   Units
VectorUtilBenchmark.floatDotProductVector    1024  thrpt   75  18.279 ± 0.395  ops/us

patch

Benchmark                                  (size)   Mode  Cnt   Score   Error   Units
VectorUtilBenchmark.floatDotProductVector    1024  thrpt   75  17.248 ± 0.092  ops/us

zen4: ['0', 'AuthenticAMD', 'AMD EPYC 9R14', '1', 'AuthenticAMD', 'AMD EPYC 9R14']

main

Benchmark                                  (size)   Mode  Cnt   Score   Error   Units
VectorUtilBenchmark.floatDotProductVector    1024  thrpt   75  18.610 ± 0.328  ops/us

patch

Benchmark                                  (size)   Mode  Cnt   Score   Error   Units
VectorUtilBenchmark.floatDotProductVector    1024  thrpt   75  19.149 ± 0.456  ops/us

[rmuir]

For the graviton4 it also hurts. I'll open a PR adding that one to the instance type list.

graviton4: ['0', '1']

main

Benchmark                                  (size)   Mode  Cnt   Score   Error   Units
VectorUtilBenchmark.floatDotProductVector    1024  thrpt   75  12.062 ± 0.306  ops/us

patch

Benchmark                                  (size)   Mode  Cnt   Score   Error   Units
VectorUtilBenchmark.floatDotProductVector    1024  thrpt   75  10.983 ± 0.112  ops/us

[rmuir]

See #14072 PR with benchie fixes, if you want to reproduce or play with it. We at least need to fix the hardcoded lucene 10 version so that benchie works again :)