mirror
/
go
mirror of https://github.com/golang/go.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

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>
This commit is contained in:
ruinan 2022-05-16 01:53:39 +00:00 committed by Eric Fang
parent 5531838c04
commit c81dfdd47a
1 changed files with 22 additions and 16 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

38
src/hash/crc32/crc32_arm64.s
View File

@ -12,19 +12,22 @@ TEXT ·castagnoliUpdate(SB),NOSPLIT,$0-36
MOVD p+8(FP), R13 // data pointer
MOVD p_len+16(FP), R11 // len(p)
CMP $8, R11
BLT less_than_8
update:
MOVD.P 8(R13), R10
CMP $16, R11
BLT less_than_16
LDP.P 16(R13), (R8, R10)
CRC32CX R8, R9
CRC32CX R10, R9
SUB $8, R11
CMP $8, R11
BLT less_than_8
SUB $16, R11
JMP update
less_than_16:
TBZ $3, R11, less_than_8
MOVD.P 8(R13), R10
CRC32CX R10, R9
less_than_8:
TBZ $2, R11, less_than_4
@ -55,19 +58,22 @@ TEXT ·ieeeUpdate(SB),NOSPLIT,$0-36
MOVD p+8(FP), R13 // data pointer
MOVD p_len+16(FP), R11 // len(p)
CMP $8, R11
BLT less_than_8
update:
MOVD.P 8(R13), R10
CMP $16, R11
BLT less_than_16
LDP.P 16(R13), (R8, R10)
CRC32X R8, R9
CRC32X R10, R9
SUB $8, R11
CMP $8, R11
BLT less_than_8
SUB $16, R11
JMP update
less_than_16:
TBZ $3, R11, less_than_8
MOVD.P 8(R13), R10
CRC32X R10, R9
less_than_8:
TBZ $2, R11, less_than_4