Discrepancy in measuring performance?

bodrato at mail.dm.unipi.it bodrato at mail.dm.unipi.it
Tue Apr 3 08:18:54 CEST 2012


Ciao,

Il Lun, 2 Aprile 2012 10:30 pm, Torbjorn Granlund ha scritto:
> Perhaps it's easier to explicitly link with .libs/libgmp.a?

I tested William's code with Torbjorn's suggestion on a test machine:

$ g++ -march=native -O2 -Wall -std=c++0x
-Wl,--sort-common,--as-needed,-z,relro,--hash-style=gnu -static main.cpp
-o main -lrt .libs/libgmp.a && ./main
GMP info:
cflags:		-Wall -g  -O3
compiler:		x86_64-linux-gnu-gcc -std=gnu99
bits/limb		64
gmp ver.		5.0.90

test info:
nTrials:		100000
nBits:			2048

average TSCs: 13465.6

$ tune/speed -c -s$[2048/64] mpn_mul_n
overhead 8.17 cycles, precision 10000 units of 1.00e-09 secs, CPU freq
1000.00 MHz
            mpn_mul_n
32           13820.00


I don't see a big discrepancy...

Regards,
m

-- 
http://bodrato.it/papers/



More information about the gmp-discuss mailing list