Re: [CFT] faster athlon/duron memory copy implementation

["Dieter Nützel" <Dieter.Nuetzel@hamburg.de>]

[-- Attachment #1: Type: text/plain, Size: 7077 bytes --]

Rober Love wrote:
> The majority of the program is inline assembly so I do not think
> compiler is playing a huge role here.

I think they are...

> Regardless, the numbers are all pretty uniform in saying the new no
> prefetch method is superior so its a mute point.

But all "your" numbers are slow.
Look at mine with the "right" (TM) flags ;-)

processor       : 0
vendor_id       : AuthenticAMD
cpu family      : 6
model           : 6
model name      : AMD Athlon(tm) MP 1900+
stepping        : 2
cpu MHz         : 1600.377
cache size      : 256 KB
fdiv_bug        : no
hlt_bug         : no
f00f_bug        : no
coma_bug        : no
fpu             : yes
fpu_exception   : yes
cpuid level     : 1
wp              : yes
flags           : fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca 
cmov pat pse36 mmx fxsr sse syscall mp mmxext 3dnowext 3dnow
bogomips        : 3145.72

processor       : 1
vendor_id       : AuthenticAMD
cpu family      : 6
model           : 6
model name      : AMD Athlon(tm) MP
stepping        : 2
cpu MHz         : 1600.377
cache size      : 256 KB
fdiv_bug        : no
hlt_bug         : no
f00f_bug        : no
coma_bug        : no
fpu             : yes
fpu_exception   : yes
cpuid level     : 1
wp              : yes
flags           : fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca 
cmov pat pse36 mmx fxsr sse syscall mp mmxext 3dnowext 3dnow
bogomips        : 3194.88


SuSE Linux 7.3

glibc-2.2.4
Addons: db db2 linuxthreads noversion
Build CFLAGS: -O -mcpu=k6 -mpreferred-stack-boundary=2 -malign-functions=4 
-fschedule-insns2 -fexpensive-optimizations -g
Build CC: gcc
Compiler version: 2.95.3 20010315 (SuSE)

Linux 2.5.43-mm2
Kernel compiler FLAGS
HOSTCC          = gcc
HOSTCFLAGS      = -Wall -Wstrict-prototypes -O -fomit-frame-pointer -mcpu=k6 
-mpreferred-stack-boundary=2 -malign-functions=4 -fschedule-insns2 
-fexpensive-optimizations

YES, I only use "-mcpu=k6" and "-O" for ages (since 26. August 1999 ;-) on my 
Athlons.

nuetzel/Entwicklung> ./athlon ; ./athlon ; ./athlon
Athlon test program $Id: fast.c,v 1.6 2000/09/23 09:05:45 arjan Exp $
clear_page() tests
clear_page function 'warm up run'        took 17409 cycles per page
clear_page function '2.4 non MMX'        took 12340 cycles per page
clear_page function '2.4 MMX fallback'   took 12429 cycles per page
clear_page function '2.4 MMX version'    took 9794 cycles per page
clear_page function 'faster_clear_page'  took 4639 cycles per page
clear_page function 'even_faster_clear'  took 4914 cycles per page

copy_page() tests
copy_page function 'warm up run'         took 16506 cycles per page
copy_page function '2.4 non MMX'         took 18412 cycles per page
copy_page function '2.4 MMX fallback'    took 18468 cycles per page
copy_page function '2.4 MMX version'     took 16550 cycles per page
copy_page function 'faster_copy'         took 10239 cycles per page
copy_page function 'even_faster'         took 10816 cycles per page


Athlon test program $Id: fast.c,v 1.6 2000/09/23 09:05:45 arjan Exp $
clear_page() tests
clear_page function 'warm up run'        took 17148 cycles per page
clear_page function '2.4 non MMX'        took 12426 cycles per page
clear_page function '2.4 MMX fallback'   took 12330 cycles per page
clear_page function '2.4 MMX version'    took 9776 cycles per page
clear_page function 'faster_clear_page'  took 4619 cycles per page
clear_page function 'even_faster_clear'  took 4938 cycles per page

copy_page() tests
copy_page function 'warm up run'         took 16640 cycles per page
copy_page function '2.4 non MMX'         took 18434 cycles per page
copy_page function '2.4 MMX fallback'    took 18454 cycles per page
copy_page function '2.4 MMX version'     took 16533 cycles per page
copy_page function 'faster_copy'         took 10418 cycles per page
copy_page function 'even_faster'         took 10707 cycles per page


Athlon test program $Id: fast.c,v 1.6 2000/09/23 09:05:45 arjan Exp $
clear_page() tests
clear_page function 'warm up run'        took 17475 cycles per page
clear_page function '2.4 non MMX'        took 12435 cycles per page
clear_page function '2.4 MMX fallback'   took 12379 cycles per page
clear_page function '2.4 MMX version'    took 9902 cycles per page
clear_page function 'faster_clear_page'  took 4665 cycles per page
clear_page function 'even_faster_clear'  took 4947 cycles per page

copy_page() tests
copy_page function 'warm up run'         took 16606 cycles per page
copy_page function '2.4 non MMX'         took 18439 cycles per page
copy_page function '2.4 MMX fallback'    took 18676 cycles per page
copy_page function '2.4 MMX version'     took 16560 cycles per page
copy_page function 'faster_copy'         took 10239 cycles per page
copy_page function 'even_faster'         took 10728 cycles per page

nuetzel/Entwicklung> ./athlon2 ; ./athlon2 ; ./athlon2
1600.061 MHz
clear_page by 'normal_clear_page'        took 12463 cycles (501.5 MB/s)
clear_page by 'slow_zero_page'           took 12461 cycles (501.6 MB/s)
clear_page by 'fast_clear_page'          took 9555 cycles (654.1 MB/s)
clear_page by 'faster_clear_page'        took 4436 cycles (1408.7 MB/s)

copy_page by 'normal_copy_page'  took 8992 cycles (695.0 MB/s)
copy_page by 'slow_copy_page'    took 9010 cycles (693.7 MB/s)
copy_page by 'fast_copy_page'    took 8134 cycles (768.3 MB/s)
copy_page by 'faster_copy'       took 5546 cycles (1126.8 MB/s)
copy_page by 'even_faster'       took 5616 cycles (1112.9 MB/s)


1600.057 MHz
clear_page by 'normal_clear_page'        took 12555 cycles (497.8 MB/s)
clear_page by 'slow_zero_page'           took 12740 cycles (490.6 MB/s)
clear_page by 'fast_clear_page'          took 9783 cycles (638.8 MB/s)
clear_page by 'faster_clear_page'        took 4459 cycles (1401.4 MB/s)

copy_page by 'normal_copy_page'  took 9123 cycles (685.0 MB/s)
copy_page by 'slow_copy_page'    took 9080 cycles (688.3 MB/s)
copy_page by 'fast_copy_page'    took 8232 cycles (759.3 MB/s)
copy_page by 'faster_copy'       took 5535 cycles (1129.1 MB/s)
copy_page by 'even_faster'       took 5565 cycles (1123.1 MB/s)


1600.060 MHz
clear_page by 'normal_clear_page'        took 12625 cycles (495.1 MB/s)
clear_page by 'slow_zero_page'           took 12541 cycles (498.3 MB/s)
clear_page by 'fast_clear_page'          took 9648 cycles (647.8 MB/s)
clear_page by 'faster_clear_page'        took 4463 cycles (1400.2 MB/s)

copy_page by 'normal_copy_page'  took 9178 cycles (680.9 MB/s)
copy_page by 'slow_copy_page'    took 9011 cycles (693.6 MB/s)
copy_page by 'fast_copy_page'    took 8138 cycles (768.0 MB/s)
copy_page by 'faster_copy'       took 5508 cycles (1134.7 MB/s)
copy_page by 'even_faster'       took 5552 cycles (1125.6 MB/s)

Regards,
	Dieter
-- 
Dieter Nützel
Graduate Student, Computer Science

University of Hamburg
Department of Computer Science
@home: Dieter.Nuetzel at hamburg.de (replace at with @)

[-- Attachment #2: athlon2.c.bz2 --]
[-- Type: application/x-bzip2, Size: 2912 bytes --]