]> Cypherpunks repositories - gostls13.git/commit
hash/crc32: use LDP instead of LDR in crc32 computation
authorruinan <ruinan.sun@arm.com>
Mon, 16 May 2022 01:53:39 +0000 (01:53 +0000)
committerEric Fang <eric.fang@arm.com>
Wed, 10 Aug 2022 08:57:33 +0000 (08:57 +0000)
commitc81dfdd47aa44bb1da0f60e36742cc0103da4837
treead83d9676b9f911ef1877ec2ad5a0b07d3fc046e
parent5531838c046eb1a94b142deeb89d7d8d9be3fc14
hash/crc32: use LDP instead of LDR in crc32 computation

Since recent ARM CPUs support CRC late forwarding of results from a
producer to a consumer, the second CRC instruction can be executed
before the first CRC instruction complete. By loading 16 bytes once with
LDP and ordering two CRC instructions here we can make the data
forwarding happen to get better performance.

Benchmarks                             old         ThisRun    delta
CRC32/poly=IEEE/size=15/align=0-160    1.24GB/s    1.26GB/s   +1.92%
CRC32/poly=IEEE/size=15/align=1-160    1.24GB/s    1.26GB/s   +1.92%
CRC32/poly=IEEE/size=40/align=0-160    2.89GB/s    2.87GB/s   -0.72%
CRC32/poly=IEEE/size=40/align=1-160    2.86GB/s    2.87GB/s     ~
CRC32/poly=IEEE/size=512/align=0-160   9.29GB/s   14.47GB/s  +55.69%
CRC32/poly=IEEE/size=512/align=1-160   9.26GB/s   13.88GB/s  +49.89%
CRC32/poly=IEEE/size=1kB/align=0-160   10.2GB/s    17.6GB/s  +72.97%
CRC32/poly=IEEE/size=1kB/align=1-160   10.1GB/s    17.2GB/s  +69.29%
CRC32/poly=IEEE/size=4kB/align=0-160   10.5GB/s    20.9GB/s  +99.01%
CRC32/poly=IEEE/size=4kB/align=1-160   10.5GB/s    20.5GB/s  +95.02%
CRC32/poly=IEEE/size=32kB/align=0-160  11.1GB/s    22.0GB/s  +98.40%
CRC32/poly=IEEE/size=32kB/align=1-160  11.1GB/s    21.6GB/s  +94.80%

CRC32/poly=Castagnoli/size=15/align=0-160   1.26GB/s   1.26GB/s     ~
CRC32/poly=Castagnoli/size=15/align=1-160   1.25GB/s   1.26GB/s     ~
CRC32/poly=Castagnoli/size=40/align=0-160   2.88GB/s   3.02GB/s   +5.18%
CRC32/poly=Castagnoli/size=40/align=1-160   2.85GB/s   3.01GB/s   +5.57%
CRC32/poly=Castagnoli/size=512/align=0-160  9.24GB/s  14.71GB/s  +59.29%
CRC32/poly=Castagnoli/size=512/align=1-160  9.21GB/s  13.45GB/s  +45.92%
CRC32/poly=Castagnoli/size=1kB/align=0-160  10.1GB/s   17.8GB/s  +75.81%
CRC32/poly=Castagnoli/size=1kB/align=1-160  10.1GB/s   17.0GB/s  +67.80%
CRC32/poly=Castagnoli/size=4kB/align=0-160  10.5GB/s   21.0GB/s  +99.67%
CRC32/poly=Castagnoli/size=4kB/align=1-160  10.5GB/s   20.5GB/s  +94.26%
CRC32/poly=Castagnoli/size=32kB/align=0-160 11.1GB/s   22.0GB/s  +98.39%
CRC32/poly=Castagnoli/size=32kB/align=1-160 11.1GB/s   21.7GB/s  +95.63%

Change-Id: Ifc7be5048cafac242e7b75f652e4aafa9aeae844
Reviewed-on: https://go-review.googlesource.com/c/go/+/407854
Reviewed-by: Keith Randall <khr@google.com>
Reviewed-by: Cherry Mui <cherryyz@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Run-TryBot: Eric Fang <eric.fang@arm.com>
Reviewed-by: Keith Randall <khr@golang.org>
src/hash/crc32/crc32_arm64.s