Consider addmul on the k8 , running speed we get
mpn_addmul_1.333
1 8.06
2 17.12
3 18.14
4 21.15
5 23.17
6 27.20
7 31.23
8 32.24
9 36.27
10 37.27
11 41.30
12 41.31
13 46.34
14 47.35
15 51.38
16 51.38
17 56.42
18 57.43
19 61.46
20 61.46
1000 2543.00
Calculating the overhead for 1000 limbs we get 2543-2.5*1000=43 cycles
and the overhead for 18 limbs is 57-2.5*18=12 cycles
13 cycles is a branch mis-predict you get on the final loop of a loop with a
count >8 , so we have 30 cycles overhead on a 1000 limbs .
Why?
Surely the overheads should be the same ?
This is not restricted to just addmul , all the functions are the same , we
seem have that the overhead is proportional to the total runtime. Could the
time stamp counter be at fault , or do we have a pipeline bubble every so
often?
Jason
--~--~---------~--~----~------------~-------~--~----~
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
mpir-devel+unsubscr...@googlegroups.com
For more options, visit this group at
http://groups.google.com/group/mpir-devel?hl=en
-~----------~----~----~----~------~----~------~--~---
