[DUG] Int64 or floating point faster?

Mon Aug 18 09:00:53 NZST 2014

Pieter, better to stick to what tests *show *to be best, not what
"intuition" suggests *should* be.  ;)

A Pentium CPU doesn't have opcodes for floating point... but it does have
an FPU which does.  The 32-bit Delphi compiler may not make the absolute
best use of the available silicon (e.g. MMX), but it will at least make use
of the FPU.

But the 32-bit Delphi compiler doesn't use CPU opcodes for Int64 operations
- even if they are present in the hardware) which I think is why - in
*tests* - Int64 arithmetic Is significantly less efficient than Double
arithmetic.

Something was bugging me that I had forgotten something... and I had...

There is also the "Comp" type.  This is listed as a "real" type in most
documentation but is in fact an Int64.  The difference is that operations
involving this type are performed in the FPU, rather than through software
emulation of a 64-bit integer.  So it's possible that Comp might provide a
performance advantage over Double.

Again, testing will be the key to answering the question of whether or not
it does in practice.

On 18 August 2014 08:37, Pieter De Wit <pieter at insync.za.net> wrote:

>  Hi,
>
>
>
> No, it appears I was wrong.
>
>
>
> Delphi (at least pascal and by default, I assume then delphi) stores
> "real"'s as per the IEEE std.
>
>
>
> See :
> https://www.informatik.uni-hamburg.de/RZ/software/SUNWspro/pascal/lang_ref/ref_data.doc.html
>
>
>
> It's a weird layout which results in (if I understand this correctly) a
> multiply instruction to even set the value:
>
> A real number is represented by this form:
>
> (-1)sign * 2 exponent-bias *1.f
> f is the bits in the fraction
>
>
>
> I would hate to work out the CPU cycles needed to multiply real's...
>
>
>
> Either way - you going to end up with a LOT more CPU usage multiplying
> real's vs int - *unless* you need a real, stick to int :)
>
>
>
> A tip from me - I would load the source ints into a few arrays and then
> use threading to make it faster/use more cores.
>
>
>
> (This did spark off a talk about if Intel/AMD include a math co-pro with
> each core - if not, you really have to stick to threaded int64 - even more
> so if your CPUs are hyperthreaded, then you need int64 as a single core
> serves 2 HT cores - research the early days of HT for more info on this :) )
>
>
>
> Cheers,
>
>
>
> Pieter
>
>
>
> On 17/08/2014 20:09, Cameron Hart wrote:
>
>  I'm confused now as I'm pretty sure Delphi uses a standard format to
> represent float (the same format used anywhere else for that matter).   In
> which case a float is essentially two int32 (or other int's depending on
> the scale of the float).  Ie a single used two int16.
>
>
>
> One int represented the mantissa the other the exponent (in essence the
> decimal portion).  Together they resulted in the floating point value.
>
>
>
> How would you describe this otherwise?
>
>
>
>
>
> *From:* delphi-bounces at listserver.123.net.nz [mailto:
> delphi-bounces at listserver.123.net.nz] *On Behalf Of *Jolyon Smith
> *Sent:* Sunday, 17 August 2014 12:54 p.m.
> *To:* NZ Borland Developers Group - Delphi List
> *Subject:* Re: [DUG] Int64 or floating point faster?
>
>
>
> That's curious.  Who are "they" ?  It doesn't sound like any floating
> point implementation I ever came across in Delphi (or anywhere else, for
> that matter).  O.o
>
>
>
> On 17 August 2014 12:28, Pieter De Wit <pieter at insync.za.net> wrote:
>
> Hi Jolyon,
>
> From memory, they used 2 int32's to make a float - this could have been
> int16's - memory is very vague on this :) The one was used to represent the
> whole numbers and the other was to show the decimal numbers
>
> Cheers,
>
> Pieter
>
>
>
> On 17/08/2014 12:05, Jolyon Smith wrote:
>
>  @Pieter - I don't understand what you mean when you say that "float was
> int32.int32".  For starters, "float" is an imprecise term.  If you mean
> "single" then the entire value was always 32 bit in it's entirety.  If you
> mean double then it was always 64 bit.  What is this "in32.int32" type of
> which you speak ?  O.o
>
>
>
> On 17 August 2014 11:52, Jolyon Smith <jsmith at deltics.co.nz> wrote:
>
> I think there are too many variables involved to give an answer to this
> question without some of those variables being reduced to known values.
>
> e.g.  what hardware ?  what version of Delphi ?  x64 target or x86 ?  what
> precision of floating point ?
>
> Having said that, in a quick test knocked up in my Smoketest framework I
> found that Double comfortably outperforms Int64 when compiling for Win32
> but that both Double and Int64 demonstrated improved performance when
> compiling for Win64 and that whilst Double still showed some advantage it
> was not as significant (and in some test runs the difference was
> negligible).
>
> If you are targeting FireMonkey you will have to bear in mind that the
> back-end compiler is different to the x86/x64 backend, so results obtained
> using the WinXX compilers will not necessarily be indicative of performance
> on the ARM or LLVM platforms.
>
>
>
> Conditions:
>
>
>  - Delphi XE4
>
>  - Running in a 64-bit Win 7 VM
>
>  - No testing was done for correctness of the results.
>
>
>
>
>
>
>
> On 16 August 2014 15:30, Ross Levis <ross at stationplaylist.com> wrote:
>
>   Would I be correct that int64 multiplications would be faster than
> floating point in Delphi?  My app needs to do several million.
>
>
>
>
>
> _______________________________________________
> NZ Borland Developers Group - Delphi mailing list
> Post: delphi at listserver.123.net.nz
> Admin: http://delphi.org.nz/mailman/listinfo/delphi
> Unsubscribe: send an email to delphi-request at listserver.123.net.nz with
> Subject: unsubscribe
>
>
>
>
>
>
>
> _______________________________________________
>
> NZ Borland Developers Group - Delphi mailing list
>
> Post: delphi at listserver.123.net.nz
>
> Admin: http://delphi.org.nz/mailman/listinfo/delphi
>
> Unsubscribe: send an email to delphi-request at listserver.123.net.nz with Subject: unsubscribe
>
>
>
>
> _______________________________________________
> NZ Borland Developers Group - Delphi mailing list
> Post: delphi at listserver.123.net.nz
> Admin: http://delphi.org.nz/mailman/listinfo/delphi
> Unsubscribe: send an email to delphi-request at listserver.123.net.nz with
> Subject: unsubscribe
>
>
>
> _______________________________________________
> NZ Borland Developers Group - Delphi mailing list
> Post: delphi at listserver.123.net.nz
> Admin: http://delphi.org.nz/mailman/listinfo/delphi
> Unsubscribe: send an email to delphi-request at listserver.123.net.nz with Subject: unsubscribe
>
>
>
>
> _______________________________________________
> NZ Borland Developers Group - Delphi mailing list
> Post: delphi at listserver.123.net.nz
> Admin: http://delphi.org.nz/mailman/listinfo/delphi
> Unsubscribe: send an email to delphi-request at listserver.123.net.nz with
> Subject: unsubscribe
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://listserver.123.net.nz/pipermail/delphi/attachments/20140818/96f27ad6/attachment.html 