Re: [PATCH] include/linux/etherdevice.h, kernel 2.6.14

["J.A. Magallon" <jamagallon@able.es>]

On 2005.10.29, at 18:00, Russell King wrote:

> Please do not drop CC's from discussions on mailing lists.
>
> On Sat, Oct 29, 2005 at 05:40:27PM +0200, Michal Srajer wrote:
>
> > On Sat, Oct 29, 2005 at 03:17:57PM +0100, Russell King wrote:
> >
> > > On Sat, Oct 29, 2005 at 04:10:46PM +0200, Michal Srajer wrote:
> > >
> > > > Description: Very small optimization patch for include/linux/ 
> > > > etherdevice.h in 2.6.14 kernel.
> > >
> > > How is this an optimisation?
> >
> > I wrote C proggram which is about two times faster
> > when using is_zero_ether_addr2 than is_zero_ether_addr1.
> >
> > --------cut--------
> > typedef unsigned char u8;
> >
> > static inline int is_zero_ether_addr1(const u8 *addr)
> > {
> >        return !(addr[0] | addr[1] | addr[2] | addr[3] | addr[4] |  
> > addr[5]);
> > }
> >
> > static inline int is_zero_ether_addr2(const u8 *addr)
> > {
> >        return !(addr[0] || addr[1] || addr[2] || addr[3] || addr 
> > [4] || addr[5]);
> > }
> >
> > main () {
> >         long i;
> >         u8 test_data[6] = {0x00,0x12,0xF0,0x0E,0xC9,0xDE};
> >         u8 test_data0[6] = {0x00,0x00,0x00,0x00,0x00,0x00};
> >         for (i=0; i<50000000; i++) {
> >                 is_zero_ether_addr1(test_data);
> >                 is_zero_ether_addr1(test_data0);
> >         }
> >         return 0;
> > }
> > --------cut--------
> > $ time ./is_zero_ether_addr1_test
> > real    0m5.986s
> > user    0m5.976s
> > sys     0m0.004s
> > $ time ./is_zero_ether_addr2_test
> > real    0m3.092s
> > user    0m3.076s
> > sys     0m0.004s
> >
> > I use gcc 4.0.3.
> > $ gcc is_zero_ether_addr1_test.c -o is_zero_ether_addr1_test
> > Should I use some special gcc options?
>
> The test is data dependent.  is_zero_ether_addr1() provides a  
> determinstic
> execution time irrespective of the supplied data.
>
> is_zero_ether_addr2() depends on the data supplied, and whether the
> architecture is able to optimise it sufficiently well (x86 may be able
> to, RISC architectures less so.)
>
> Therefore, the existing code is far more preferable, at least to me.
> This is what I get on ARM:
>
> $ /usr/bin/time ./t1
> 0.66user 0.02system 0:00.68elapsed 98%CPU (0avgtext+0avgdata  
> 0maxresident)k
> 0inputs+0outputs (0major+106minor)pagefaults 0swaps
> $ /usr/bin/time ./t2
> 1.10user 0.02system 0:01.13elapsed 99%CPU (0avgtext+0avgdata  
> 0maxresident)k
> 0inputs+0outputs (0major+106minor)pagefaults 0swaps
> $ /usr/bin/time ./t1
> 0.67user 0.01system 0:00.68elapsed 98%CPU (0avgtext+0avgdata  
> 0maxresident)k
> 0inputs+0outputs (0major+106mino^[[Ar)pagefaults 0swaps
> $ /usr/bin/time ./t2
> 1.11user 0.02system 0:01.12elapsed 100%CPU (0avgtext+0avgdata  
> 0maxresident)k
> 0inputs+0outputs (0major+106minor)pagefaults 0swaps
> $ /usr/bin/time ./t1
> 0.67user 0.02system 0:00.69elapsed 99%CPU (0avgtext+0avgdata  
> 0maxresident)k
> 0inputs+0outputs (0major+106minor)pagefaults 0swaps
> $ /usr/bin/time ./t2
> 1.11user 0.01system 0:01.12elapsed 99%CPU (0avgtext+0avgdata  
> 0maxresident)k
> 0inputs+0outputs (0major+106minor)pagefaults 0swaps
>
> where t1 is using is_zero_ether_addr1 and t2 is using
> is_zero_ether_addr2.  That's almost twice as long for your "optimised"
> version than for the present version.

Just for curiosity, could you both benchmark this also:

int is_zero_ether_addr0(const unsigned char *addr)
{
    return !(((unsigned long *)addr)[0] | ((unsigned short*)addr)[2]);
}

Assembler in x86 is

is_zero_ether_addr0:
    pushl   %ebp
    movl    %esp, %ebp
    movl    8(%ebp), %edx
    movzwl  4(%edx), %eax
    orl (%edx), %eax
    sete    %al
    movzbl  %al, %eax
    popl    %ebp
    ret


--
J.A. Magallon <jamagallon()able!es>   \          Software is like sex:
wolverine                              \    It's better when it's free
MacOS X 10.4.2, Darwin Kernel Version 8.2.0


-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/