forums.silverfrost.com

[LitusSaxonicum]

There's something wrong with the Polly Benchmarks, that's for sure. When I looked at the source codes some years ago I decided that the effort to understand them was a waste of time as they were not what I would hold up as good examples of coding style.

I have a feeling that the Intel compiler makes a better job of using multiple cores than others, like FTN95, do, and a more realistic (for most users) measure would be the performance on a less highly-specified machine (as J-S suggests). Also, the Polly Benchmarks don't use the latest FTN95, nor do they use 64-bit FTN95, which might have a bearing. Frankly, I'd like to see a comparison of 32-bit, 64-bit and .NET flavours of FTN95, and also with and without /OPT and so on, so I know which options to choose rather than comparisons with other compilers.

Mind you, I'm impressed by Polly, because they can get the Absoft compiler to run, which I couldn't.

The execution speed of the Intel compiler on the worst benchmarks where it is around 10x faster than FTN95 is immaterial, as it doesn't have Clearwin+, and is therefore rather useless! Also, it wouldn't make any perceptible difference to my users if the software were 10 or a 100 times faster.

Eddie

[DanRRight]

In 90th this compiler was light years ahead of everyone. Then dark times and recessions came when no one interested in engineering, science and technology. Now healing of the king definitely needed.
https://www.youtube.com/watch?v=Y6wE2W3ag1g

[JohnCampbell]

There are 4 main areas where FTN95 is behind iFort ( I started at 2!)
1) optimisation of inner loops
2) implementation of vector instructions
3) recognition of bad coding
4) using array sections, especially as arguments
5) OpenMP support

It is the 3rd that is interesting, as there are a few examples in the Polyhedron benchmarks where bad coding is common. Some compilers have been tuned to recognising these poor constructs, also called optimisation. ( eg real**2.0 )
( my posts from a few years ago have more details )

There are 2 main areas where FTN95 is ahead of iFort
1) very good diagnostics.
2) clearwin

What this says is that if you can eliminate some bad coding constructs, the relative performance of FTN95 vs iFort (or gFortran) is not as bad as reported.

mecej4 should be thanked for helping to make FTN95 more reliable for /opt which helps FTN95 to address most of the points above.

[LitusSaxonicum]

John,

Was the miscounting a reference to further Monty Python sketches - ('No one expects the Spanish Inquisition if you need reminding)?

The use of vector instructions comes are a risk of increased roundoff, and at a guess, worse performance with transcendental functions
.
I think it's your 5. that gives the huge differences in the run times, and that would be far less apparent with fewer cores. I hesitate to adduce the appeal of filthy lucre, but perhaps Polly makes more from a sale of iFort than FTN95, and hence their choice of machine to run the benchmarks on? It would, I think, be a closer run thing on a single or dual core machine.

Eddie

[JohnCampbell]

Eddie,

My counting 4 and listing 5 was my sense of humour ! There are always more for the list.

64-bit does not support 80-bit reals, so that "increased roundoff" has been with us for a while.

Regarding the choice of processor, the Polyhedron benchmark being quoted is a single thread, 32-bit test, so I am not sure why they are now using 2 x Xeon E5-2643 and 64GB memory. I am definitely "not a fan" of this class of processor. I find the i7 to be much better for computation and my i7-8700K definitely gives good performance for the analysis I am now performing.
Probably most of the referring to slower recent performance could be due to the 3.3 GHz Xeon change. Based on my coding approaches, they really are not suited to a single thread, 32-bit test, or to numerical computation in general. (perhaps I don't know how to use them effectively, but I was forced to use Xeon for a number of years, but no more.)

[LitusSaxonicum]

Then perhaps you don't know the Spanish Inquisition sketch: https://www.dailymotion.com/video/xsefeg

Eddie