For fun (!) I used my new ldc2 build on Fedora 20 to run my "π by quadrature" examples. On my 7 year old 8 core workstation, the various C ++ versions tend to take about 1.1s using all 8 cores (as do C, Java, Scala, etc.). Indeed generally so do all the D codes using LDC or GDC – DMD takes generally twice as long.

Now something potentially remarkable: using ldc the std.parallelism.map version of the code takes 0.6s. dmd takes 2.6s, I can't test gdc on Fedora as Fedora do not package it and I haven't built it.

There could be many reasons for this result, but it appears at present a genuine "D runs 1.8 times faster than C++".

Clearly this needs std.benchmark to be rolled out, along with something
equivalent for C++ (but I am not sure what just at present). More news
as it happens.

-- 
Russel. ============================================================================= Dr Russel Winder      t: +44 20 7585 2200   voip: sip:russel.winder@ekiga.net 41 Buckmaster Road    m: +44 7770 465 077   xmpp: russel@winder.org.uk London SW11 1EN, UK   w: www.russel.org.uk  skype: russel_winder

Hi Russel!

On Friday, 6 June 2014 at 16:04:44 UTC, Russel Winder via digitalmars-d-ldc wrote:
> There could be many reasons for this result, but it appears at present a
> genuine "D runs 1.8 times faster than C++".

Which branch and which LLVM version did you use?

> Clearly this needs std.benchmark to be rolled out, along with something
> equivalent for C++ (but I am not sure what just at present). More news
> as it happens.
> 

Yes, this has to be investigated. A "magical" increase in performance might also be a problem with measurement or a compiler bug. But hopefully not! :-)

Regards,
Kai

On Fri, 2014-06-06 at 16:11 +0000, Kai Nacke via digitalmars-d-ldc
wrote:
[…]
> Which branch and which LLVM version did you use?

I am compiling LDC merge-2.065 HEAD, using GCC 4.8.2 with LLVM 3.4 on Fedora 20.

I will do something similar on Debian Sid, but there I try to use GCC 4.9 since I invariably try to compile my own stuff with -std=C++1y.

[…]
> Yes, this has to be investigated. A "magical" increase in performance might also be a problem with measurement or a compiler bug. But hopefully not! :-)

First thought has to be bad user code, but in this case I think not, but it could still be a factor.

I am going to have to create new versions of the code to be working with benchmark frameworks rather than just the "one off timing" stuff I currently have.

-- 
Russel. ============================================================================= Dr Russel Winder      t: +44 20 7585 2200   voip: sip:russel.winder@ekiga.net 41 Buckmaster Road    m: +44 7770 465 077   xmpp: russel@winder.org.uk London SW11 1EN, UK   w: www.russel.org.uk  skype: russel_winder

> Now something potentially remarkable: using ldc the std.parallelism.map version of the code takes 0.6s. dmd takes 2.6s, I can't test gdc on Fedora as Fedora do not package it and I haven't built it.

Well, I have good news for you - I am nearly done with GDC RPM SPEC (installs GDC in /opt/gdc directory). The moment it is ready for testing by D community I will inform people on the NG.

(I also have very unstable 4.8 SPEC which replaces system GCC with one that has D enabled...)

-- 
http://dejan.lekic.org

On Mon, 2014-06-09 at 22:32 +0100, Dejan Lekic via digitalmars-d-ldc
wrote:
[…]
> Well, I have good news for you - I am nearly done with GDC RPM SPEC (installs GDC in /opt/gdc directory). The moment it is ready for testing by D community I will inform people on the NG.
> 
> (I also have very unstable 4.8 SPEC which replaces system GCC with one that has D enabled...)

Is /opt/gdc the right place to install to: given this is an RPM can't it
just use the standard places. In particular gdc really needs to be
in /usr/bin.

Is the intention to get this RPM into RPMFusion?

-- 
Russel. ============================================================================= Dr Russel Winder      t: +44 20 7585 2200   voip: sip:russel.winder@ekiga.net 41 Buckmaster Road    m: +44 7770 465 077   xmpp: russel@winder.org.uk London SW11 1EN, UK   w: www.russel.org.uk  skype: russel_winder