I been trying to account for each and every cycle in the asm routines , and I
came across this oddity
let mpn_fn be unrolled to X-way
then run
./speed -cD -s X-X*20 -t X mpn_fn
and you see the time differences per loop as you expect , except for the size
9*X , where you have an extra 12-14 cycles (a branch miss-prediction)
for example for my mpn_add_n which is 4-way unroll
~/gmpextra-1.0.2/tune# ./speed -cD -s 4-80 -t 4 mpn_add_n
overhead 6.04 cycles, precision 10000 units of 5.53e-10 secs, CPU freq 1808.23
MHz
mpn_add_n
4 (16.11)
8 5.04
12 5.06
16 6.05
20 6.01
24 6.04
28 6.12
32 6.06
36 19.97
40 9.18
44 5.98
48 6.08
52 6.07
56 6.01
60 6.04
64 6.04
68 6.05
72 6.04
76 6.05
80 6.47
This is quite significant , as it amounts to a 21% slowdown at 36 limbs , and
even at 100 limbs we still losing 10% speed. It seems to be a general problem
as it effects every mpn fn I've tested. The problem appears to be , when you
have a loop with a count >=9 , then the cpu will always have a branch
misprediction once. It's possible that this is a problem only with my exact
cpu(steping) below
/gmp-4.2.4/tune# cat /proc/cpuinfo
processor : 0
vendor_id : AuthenticAMD
cpu family : 15
model : 44
model name : AMD Sempron(tm) Processor 3000+
stepping : 2
cpu MHz : 1808.227
cache size : 128 KB
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 clflush mmx fxsr sse sse2 syscall nx mmxext fxsr_opt lm
3dnowext 3dnow up rep_good pni lahf_lm
bogomips : 3620.38
TLB size : 1024 4K pages
clflush size : 64
cache_alignment : 64
address sizes : 40 bits physical, 48 bits virtual
power management: ts fid vid ttp tm stc
The only fix I could think of is to unroll the function further , and so push
the problem to larger sizes , where the slowdown% will be less.
--~--~---------~--~----~------------~-------~--~----~
You received this message because you are subscribed to the Google Groups
"mpir-devel" group.
To post to this group, send email to mpir-devel@googlegroups.com
To unsubscribe from this group, send email to [EMAIL PROTECTED]
For more options, visit this group at
http://groups.google.com/group/mpir-devel?hl=en
-~----------~----~----~----~------~----~------~--~---
