Skip to content
Snippets Groups Projects
user avatar
Joel Sing authored
This provides an assembly implementation of subVW for riscv64,
processing up to four words per loop, resulting in a significant
performance gain.

On a StarFive VisionFive 2:

                  │   subvw.1    │               subvw.2               │
                  │    sec/op    │   sec/op     vs base                │
SubVW/1-4            57.43n ± 0%   41.45n ± 0%  -27.82% (p=0.000 n=10)
SubVW/2-4            69.31n ± 0%   48.15n ± 0%  -30.53% (p=0.000 n=10)
SubVW/3-4            76.12n ± 0%   54.87n ± 0%  -27.92% (p=0.000 n=10)
SubVW/4-4            85.47n ± 0%   56.14n ± 0%  -34.32% (p=0.000 n=10)
SubVW/5-4            96.15n ± 0%   62.83n ± 0%  -34.65% (p=0.000 n=10)
SubVW/10-4          149.60n ± 0%   89.55n ± 0%  -40.14% (p=0.000 n=10)
SubVW/100-4         1115.0n ± 0%   549.3n ± 0%  -50.74% (p=0.000 n=10)
SubVW/1000-4        10.732µ ± 0%   5.071µ ± 0%  -52.75% (p=0.000 n=10)
SubVW/10000-4        153.0µ ± 0%   103.7µ ± 0%  -32.21% (p=0.000 n=10)
SubVW/100000-4       1.542m ± 0%   1.046m ± 0%  -32.13% (p=0.000 n=10)
SubVWext/1-4         57.42n ± 0%   41.45n ± 0%  -27.81% (p=0.000 n=10)
SubVWext/2-4         69.33n ± 0%   48.15n ± 0%  -30.55% (p=0.000 n=10)
SubVWext/3-4         76.12n ± 0%   54.93n ± 0%  -27.84% (p=0.000 n=10)
SubVWext/4-4         85.47n ± 0%   56.14n ± 0%  -34.32% (p=0.000 n=10)
SubVWext/5-4         96.15n ± 0%   62.83n ± 0%  -34.65% (p=0.000 n=10)
SubVWext/10-4       149.60n ± 0%   89.56n ± 0%  -40.14% (p=0.000 n=10)
SubVWext/100-4      1115.0n ± 0%   549.3n ± 0%  -50.74% (p=0.000 n=10)
SubVWext/1000-4     10.732µ ± 0%   5.061µ ± 0%  -52.84% (p=0.000 n=10)
SubVWext/10000-4     152.5µ ± 0%   103.7µ ± 0%  -32.02% (p=0.000 n=10)
SubVWext/100000-4    1.533m ± 0%   1.046m ± 0%  -31.75% (p=0.000 n=10)
geomean              1.005µ        633.7n       -36.92%

                  │   subvw.1    │                subvw.2                 │
                  │     B/s      │      B/s       vs base                 │
SubVW/1-4           132.9Mi ± 0%    184.1Mi ± 0%   +38.54% (p=0.000 n=10)
SubVW/2-4           220.1Mi ± 0%    316.9Mi ± 0%   +43.95% (p=0.000 n=10)
SubVW/3-4           300.7Mi ± 0%    417.1Mi ± 0%   +38.72% (p=0.000 n=10)
SubVW/4-4           357.1Mi ± 0%    543.6Mi ± 0%   +52.24% (p=0.000 n=10)
SubVW/5-4           396.7Mi ± 0%    607.2Mi ± 0%   +53.03% (p=0.000 n=10)
SubVW/10-4          510.1Mi ± 0%    851.9Mi ± 0%   +67.01% (p=0.000 n=10)
SubVW/100-4         684.2Mi ± 0%   1388.9Mi ± 0%  +102.99% (p=0.000 n=10)
SubVW/1000-4        710.9Mi ± 0%   1504.5Mi ± 0%  +111.63% (p=0.000 n=10)
SubVW/10000-4       498.7Mi ± 0%    735.7Mi ± 0%   +47.52% (p=0.000 n=10)
SubVW/100000-4      494.8Mi ± 0%    729.1Mi ± 0%   +47.34% (p=0.000 n=10)
SubVWext/1-4        132.9Mi ± 0%    184.1Mi ± 0%   +38.53% (p=0.000 n=10)
SubVWext/2-4        220.1Mi ± 0%    316.9Mi ± 0%   +44.00% (p=0.000 n=10)
SubVWext/3-4        300.7Mi ± 0%    416.7Mi ± 0%   +38.57% (p=0.000 n=10)
SubVWext/4-4        357.1Mi ± 0%    543.6Mi ± 0%   +52.24% (p=0.000 n=10)
SubVWext/5-4        396.7Mi ± 0%    607.2Mi ± 0%   +53.04% (p=0.000 n=10)
SubVWext/10-4       510.1Mi ± 0%    851.9Mi ± 0%   +67.01% (p=0.000 n=10)
SubVWext/100-4      684.2Mi ± 0%   1388.9Mi ± 0%  +102.99% (p=0.000 n=10)
SubVWext/1000-4     710.9Mi ± 0%   1507.6Mi ± 0%  +112.07% (p=0.000 n=10)
SubVWext/10000-4    500.1Mi ± 0%    735.7Mi ± 0%   +47.10% (p=0.000 n=10)
SubVWext/100000-4   497.8Mi ± 0%    729.4Mi ± 0%   +46.52% (p=0.000 n=10)
geomean             387.6Mi         614.5Mi        +58.51%

Change-Id: I9d7fac719e977710ad9db9121fa298db6df605de
Reviewed-on: https://go-review.googlesource.com/c/go/+/595398


Reviewed-by: default avatarMark Ryan <markdryan@rivosinc.com>
Reviewed-by: default avatarDmitri Shuralyov <dmitshur@google.com>
Reviewed-by: default avatarMeng Zhuo <mengzhuo1203@gmail.com>
Reviewed-by: default avatarCherry Mui <cherryyz@google.com>
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
4f18477d
History
Name Last commit Last update