Re: [Devel] Question about optimizing FT_MulFix.

[Nathan Hurst]

On Mon, 30 Apr 2001, David Turner wrote:

> "mult_fix2" corresponds to the original code in FT_MulFix, and it's
> interesting to see that it's the fastest on a "modern" processor
> like a PIII. I'd like to see the results for a PowerPC or an Alpha
> though, just to get sure of what's hapenning..
> 
> Apparently, it seems that "mult_fix3" is a winner, because it
> povides a 12-15% speedup on "old" processors, whild loses less
> than 1% of performance compared to the original FT_MulFix on
> a Pentium III..
> 
> I'd like to have more data however to be certain about that..
> Testers are then welcome :-)

On a 400MHz G4:
address@hidden:~/timecpu$ gcc fixmul64.c test_fixmul.c 
address@hidden:~/timecpu$ ./a.out 
mult_fix1 : time = 14.07000
mult_fix2 : time = 13.46000
mult_fix3 : time = 13.06000
mult_fix4 : time = 14.87000
mult_fix5 : time = 19.28000
address@hidden:~/timecpu$ gcc -O6 fixmul64.c test_fixmul.c 
address@hidden:~/timecpu$ ./a.out 
mult_fix1 : time = 6.63000
mult_fix2 : time = 5.43000
mult_fix3 : time = 6.22000
mult_fix4 : time = 7.44000
mult_fix5 : time = 6.42000
address@hidden:~/timecpu$ cat /proc/cpuinfo 
cpu             : 7400 (G4)
clock           : 400MHz
bogomips        : 796.76
L2 cache        : 1024K unified
memory          : 128MB

On a 400MHz G3:
bash-2.05$ gcc fixmul64.c test_fixmul.c
bash-2.05$ ./a.out 
mult_fix1 : time = 13.97000
mult_fix2 : time = 12.70000
mult_fix3 : time = 12.38000
mult_fix4 : time = 14.52000
mult_fix5 : time = 17.99000
bash-2.05$ gcc -O6 fixmul64.c test_fixmul.c
bash-2.05$ ./a.out 
mult_fix1 : time = 6.65000
mult_fix2 : time = 5.44000
mult_fix3 : time = 6.25000
mult_fix4 : time = 7.45000
mult_fix5 : time = 6.45000
bash-2.05$ cat /proc/cpuinfo 
cpu             : 740/750
clock           : 400MHz
bogomips        : 798.40
L2 cache        : 1024K unified
memory          : 256MB

Dual G4 400MHz:
address@hidden:~/timecpu$ gcc fixmul64.c test_fixmul.c
address@hidden:~/timecpu$ ./a.out 
mult_fix1 : time = 12.32000
mult_fix2 : time = 11.42000
mult_fix3 : time = 11.10000
mult_fix4 : time = 12.84000
mult_fix5 : time = 17.12000
address@hidden:~/timecpu$ gcc -O6 fixmul64.c test_fixmul.c
address@hidden:~/timecpu$ ./a.out 
mult_fix1 : time = 5.93000
mult_fix2 : time = 4.83000
mult_fix3 : time = 5.55000
mult_fix4 : time = 6.65000
mult_fix5 : time = 5.83000

Which just proves that benchmarks are meaningless.

njh