Filippo Valsorda
2dbfe9001f
poly1305: rewrite the Go implementation with 64-bit limbs
...
The new code is meant to be readable without external references for
Poly1305, and explains the field logic. The generic code is now 30-50%
faster on a Intel(R) Xeon(R) CPU E5-2690 v3 @ 2.60GHz, and even better
on a 3.1 GHz i7 MacBook.
name old time/op new time/op delta
64-48 126ns ± 0% 80ns ± 1% -36.24% (p=0.000 n=16+20)
1K-48 1.07µs ± 0% 0.81µs ± 2% -23.63% (p=0.000 n=19+20)
2M-48 2.07ms ± 0% 1.61ms ± 1% -22.31% (p=0.000 n=20+20)
Write64-48 79.3ns ± 0% 58.0ns ± 1% -26.89% (p=0.000 n=20+19)
Write1K-48 1.02µs ± 0% 0.79µs ± 1% -22.91% (p=0.000 n=19+19)
Write2M-48 2.07ms ± 0% 1.61ms ± 2% -22.33% (p=0.000 n=17+20)
name old speed new speed delta
64-48 508MB/s ± 0% 797MB/s ± 1% +56.95% (p=0.000 n=16+20)
1K-48 960MB/s ± 0% 1257MB/s ± 2% +30.94% (p=0.000 n=18+20)
2M-48 1.01GB/s ± 0% 1.30GB/s ± 1% +28.73% (p=0.000 n=20+20)
Write64-48 807MB/s ± 0% 1104MB/s ± 1% +36.78% (p=0.000 n=18+19)
Write1K-48 1.00GB/s ± 0% 1.30GB/s ± 1% +29.71% (p=0.000 n=18+19)
Write2M-48 1.01GB/s ± 0% 1.31GB/s ± 2% +28.77% (p=0.000 n=17+20)
The assembly is still 50-90% faster on the Xeon, 30-60% on the MacBook.
The Go code does not use all the arithmetic tricks the assembly does,
and it does not have access to the three operand wide shift instruction.
name old time/op new time/op delta
64-48 80.3ns ± 1% 54.2ns ± 0% -32.50% (p=0.000 n=20+17)
1K-48 815ns ± 2% 446ns ± 1% -45.27% (p=0.000 n=20+20)
2M-48 1.61ms ± 1% 0.86ms ± 0% -46.54% (p=0.000 n=20+17)
Write64-48 58.0ns ± 1% 34.0ns ± 0% -41.34% (p=0.000 n=19+20)
Write1K-48 790ns ± 1% 427ns ± 0% -45.92% (p=0.000 n=19+17)
Write2M-48 1.61ms ± 2% 0.86ms ± 0% -46.51% (p=0.000 n=20+20)
name old speed new speed delta
64-48 797MB/s ± 1% 1180MB/s ± 0% +48.09% (p=0.000 n=20+19)
1K-48 1.26GB/s ± 2% 2.30GB/s ± 1% +82.71% (p=0.000 n=20+20)
2M-48 1.30GB/s ± 1% 2.44GB/s ± 0% +87.04% (p=0.000 n=20+17)
Write64-48 1.10GB/s ± 1% 1.88GB/s ± 0% +70.52% (p=0.000 n=19+18)
Write1K-48 1.30GB/s ± 1% 2.40GB/s ± 0% +84.84% (p=0.000 n=19+18)
Write2M-48 1.31GB/s ± 2% 2.44GB/s ± 0% +86.93% (p=0.000 n=20+20)
Hopefully this will also avoid the need for an arm64 implementation.
Since now the Go and the amd64/ppc64le assembly use the same limb
schedule, drop the assembly initialize and finalize implementations,
and make the wrapper code match. It comes with a minor slowdown.
name old time/op new time/op delta
64-48 50.3ns ± 0% 54.2ns ± 0% +7.73% (p=0.000 n=20+17)
1K-48 441ns ± 0% 446ns ± 1% +1.10% (p=0.000 n=19+20)
2M-48 860µs ± 0% 859µs ± 0% ~ (p=0.178 n=19+17)
Write64-48 34.0ns ± 0% 34.0ns ± 0% ~ (all equal)
Write1K-48 424ns ± 0% 427ns ± 0% +0.71% (p=0.000 n=17+17)
Write2M-48 860µs ± 0% 859µs ± 0% -0.04% (p=0.000 n=19+20)
name old speed new speed delta
64-48 1.27GB/s ± 0% 1.18GB/s ± 0% -7.20% (p=0.000 n=20+19)
1K-48 2.32GB/s ± 0% 2.30GB/s ± 1% -1.07% (p=0.000 n=18+20)
2M-48 2.44GB/s ± 0% 2.44GB/s ± 0% ~ (p=0.173 n=19+17)
Write64-48 1.88GB/s ± 0% 1.88GB/s ± 0% +0.04% (p=0.000 n=19+18)
Write1K-48 2.41GB/s ± 0% 2.40GB/s ± 0% -0.67% (p=0.000 n=19+18)
Write2M-48 2.44GB/s ± 0% 2.44GB/s ± 0% +0.04% (p=0.000 n=19+20)
Since poly1305/sum_generic.go was almost entirely rewritten, it's
probably best reviewed on gitiles.
This is the implementation published at
https://blog.filippo.io/a-literate-go-implementation-of-poly1305/
Updates #31470
Change-Id: I74f9011d3ee317a43b05ae7f05d96081d08bffd3
Reviewed-on: https://go-review.googlesource.com/c/crypto/+/169037
Reviewed-by: Katie Hockman <katie@golang.org >
2019-11-11 21:33:42 +00:00
bill_ofarrell
4eb8c2c8d8
poly1305: add optimized s390x SIMD implementation with VMSL
...
SIMD implementation based the on the algorithm outlined in:
NEON crypto, Daniel J. Bernstein and Peter Schwabe
https://cryptojedi.org/papers/neoncrypto-20120320.pdf
and as modified for VMSL as described in
Accelerating Poly1305 Cryptographic Message Authentication on the z14
O'Farrell, Gadriwala, et al, CASCON 2017, p48-55
https://ibm.ent.box.com/s/jf9gedj0e9d2vjctfyh186shaztavnht
name old new delta
64 485MB/s 1315 MB/s +171.58%
1K 607MB/s 4352 MB/s +616.97%
64Unaligned 485MB/s 1373 MB/s +183.09%
1KUnaligned 606MB/s 4286 MB/s +607.26%
2M 607MB/s 5529 MB/s +810.87%
Change-Id: I31ccc25ced09180d99ea5c9233f0dcdc8666fc98
Reviewed-on: https://go-review.googlesource.com/110297
Run-TryBot: Michael Munday <mike.munday@ibm.com >
TryBot-Result: Gobot Gobot <gobot@golang.org >
Reviewed-by: Michael Munday <mike.munday@ibm.com >
2018-05-14 22:55:51 +00:00
Filippo Valsorda
Michael Munday
bill_ofarrell