> Message: 1
> To: email@hidden
> From: Kyros Yakinthos <email@hidden>
> Subject: FORTRAN and AltiVec
> Date: Fri, 9 Apr 2004 10:53:38 +0300
>
> Hello All,
>
> I have noticed a strange result.
> I have a fortran code with the following command line
>
> res= a*b-c*d
[snip]
>
> When I compile using xlf 8.1 (evaluation) with the -g option, AltiVec
> is faster about 50%
> When I compile using xlf 8.1 with the -O5 option for full optimization
> AltiVec is slower about 40%
>
> Is xlf doing such a good optimization or is something wrong with the C
> subroutine for
> AltiVec instructions? Further, is this related to something about
> alignment or... what else?
After communicating with Kyros, I took a look at this and found something
interesting -- at higher levels of optimization, XLF can do vector-adds
(really scalar array adds in a loop) nearly as fast or faster than passing
the data to AltiVec for a true vector computation. This is on my old 400MHz
PowerBook G4 by the way, so I would expect a G5 to be spectacular.
xlf90 add.f -o addX -O0 vecadd.o -qextname=dtime
n scalar-time vector-time AltiVec speedup
4 2.970000029 0.5500000119 5.400000095
16 2.680000067 0.3400000036 7.882352829
64 2.640000105 0.2800000012 9.428571701
256 2.629999876 0.2099999934 12.52380943
1024 2.690000057 0.2500000000 10.76000023
4096 2.750000000 0.4600000083 5.978260994
16384 2.740000010 0.5099999905 5.372549057
65536 3.279999971 1.409999967 2.326241255
262144 3.359999895 1.289999962 2.604651213
xlf90 add.f -o addX -O3 vecadd.o -qextname=dtime
n scalar-time vector-time AltiVec speedup
4 0.4199999869 0.4499999881 0.9333333373
16 0.3000000119 0.3199999928 0.9375000596
64 0.2700000107 0.2399999946 1.125000119
256 0.2500000000 0.2500000000 1.000000000
1024 0.2399999946 0.2500000000 0.9599999785
4096 0.4000000060 0.4799999893 0.8333333731
16384 0.4399999976 0.4899999797 0.8979592323
65536 0.8199999928 1.419999957 0.5774648190
262144 1.179999948 1.309999943 0.9007633328
As you can see, the vector times are about the same from case to case, it's
the scalar time that really comes down with -O3.
I also compiled with Absoft for comparison:
f90 -s add.f -o addA vecadd.o -lU77
n scalar-time vector-time AltiVec speedup
4 3.80000 0.710000 5.35211
16 3.18000 0.370000 8.59459
64 3.13000 0.270000 11.5926
256 3.19000 0.230000 13.8696
1024 3.25000 0.230000 14.1304
4096 3.46000 0.480000 7.20833
16384 3.46000 0.460000 7.52174
65536 3.79000 0.810000 4.67901
262144 4.23000 1.21000 3.49587
f90 -s add.f -o addA vecadd.o -lU77 -O3
n scalar-time vector-time AltiVec speedup
4 1.66000 0.480000 3.45833
16 1.55000 0.300000 5.16667
64 1.44000 0.280000 5.14286
256 1.49000 0.250000 5.96000
1024 1.47000 0.230000 6.39130
4096 1.67000 0.490000 3.40816
16384 1.69000 0.500000 3.38000
65536 2.15000 0.840000 2.55952
262144 2.45000 1.25000 1.96000
Here, the vector times are pretty consistent case to case (and roughly
consistent with the XLF compilation) but you can see that the scalar times
don't improve quite so much with optimization. So, AltiVec looks pretty
good here.
So, for anybody using XLF and considering AltiVec, be sure to run some tests
before doing your coding/porting! At least for vector-add, XLF is doing
quite well. If you're using Absoft, you'll still see a nice gain from
implementing AltiVec for vector-add.
Kyros noted that AltiVec/vecLib still wins for exp(x) compared to XLF, so I
suspect general computations may still benefit from AltiVec overall. It
would be interesting to test sqrt(x) on a G5, as it can do this in hardware.
On a final note, I was wondering if XLF perhaps used AltiVec for vector-add,
so I repeated the scalar benchmark with double precision data. It still
looked good. For most vector lengths, XLF could do a double-precision
vector-add faster than using AltiVec on single precision data!
Craig
--
Dr. Craig A. Hunter
NASA Langley Research Center
AAAC / Configuration Aerodynamics Branch
(757) 864-3020
email@hidden (NEW!!)
_______________________________________________
scitech mailing list | email@hidden
Help/Unsubscribe/Archives: http://www.lists.apple.com/mailman/listinfo/scitech
Do not post admin requests to the list. They will be ignored.
