• Trass3r (23/23) Aug 07 2011 I ported smallpt to D some time ago and now that I've working versions o...
• bearophile (10/12) Aug 07 2011 I have converted to D many of the commonly found benchmarks, this is my ...
• %u (3/3) Aug 08 2011 This is off topic but if you're interested you can get it to run using
• Christian Kamm (5/23) Aug 09 2011 bearophile, just out of interest, what's the performance like if you ran...
• bearophile (5/8) Aug 09 2011 It's the only way I have found to perform link-time-optimization with LD...
• Jacob Carlborg (5/9) Aug 09 2011 ldmd is a wrapper that converts dmd arguments to ldc arguments. So you
• Trass3r (1/6) Aug 10 2011 Is there a more convenient way to get LTO with LDC?
• bearophile (4/5) Aug 10 2011 This is the best/only one I have found for LDC1. Feel free to search for...
• Trass3r (4/8) Aug 10 2011 Well, obviously you need to install the gold linker and that llvm plugin...
• Trass3r (2/3) Aug 10 2011 Had a quick look at the source, no indications that -O4 and O5 do anythi...
• Robert Clipsham (7/10) Aug 10 2011 They don't, there's no way to do LTO built into LDC. I seem to recall
• Trass3r (4/10) Aug 10 2011 Why is there no way?
• Robert Clipsham (7/18) Aug 10 2011 Of course it can - I said "built into", not "in". The only reason it's

Trass3r <un known.com> writes:

I ported smallpt to D some time ago and now that I've working versions of  
LDC2 and GDC2 I did a quick comparison:

Original code:
http://kevinbeason.com/smallpt/explicit.cpp

D version:
https://bitbucket.org/trass3r/smallptd/src


C++ -- g++ -O3 explicit.cpp:

Rendering (4 spp) 100.00%
real	0m28.477s
user	0m26.470s
sys	0m0.160s

D -- dmd -release -O -inline -noboundscheck smallpt.d -ofsmallptDMD:
real	1m32.687s
user	1m30.820s
sys	0m0.300s

D -- gdc -frelease -fno-bounds-check -O3 smallpt.d -o smallptGDC:
real	0m52.598s
user	0m48.900s
sys	0m0.150s

D -- ldc2 -O3 -release -enable-inlining smallpt.d -ofsmallptLDC:
real	0m39.622s
user	0m36.100s
sys	0m0.240s

bearophile <bearophileHUGS lycos.com> writes:

Trass3r:

 I ported smallpt to D some time ago and now that I've working versions of  
 LDC2 and GDC2 I did a quick comparison:

I have converted to D many of the commonly found benchmarks, this is my version:
http://codepad.org/ZbmSfseY

If I compile it with with LDC1 with Link Time Optimization + Interning:
ldc -O3 -release -inline -output-bc smallpt2_d.d

opt -std-compile-opts smallpt2_d.bc > smallpt2_do.bc

llvm-ld -native -ltango-base-ldc -ltango-user-ldc -ldl -lm -lpthread
-internalize-public-api-list=_Dmain -o=smallpt2_do smallpt2_do.bc

I produce a binary that's faster than the C++ version (22.7 seconds instead of
29 seconds).

Bye,
bearophile

%u <me emails.com> writes:

This is off topic but if you're interested you can get it to run using
float if you change the Sphere code to:

FP eps = 1e-2; instead of 1e-4;

Christian Kamm <kamm-incasoftware removethis.de> writes:

bearophile wrote:

 Trass3r:
 
 I ported smallpt to D some time ago and now that I've working versions of
 LDC2 and GDC2 I did a quick comparison:

 
 I have converted to D many of the commonly found benchmarks, this is my
 version: http://codepad.org/ZbmSfseY
 
 If I compile it with with LDC1 with Link Time Optimization + Interning:
 ldc -O3 -release -inline -output-bc smallpt2_d.d
 
 opt -std-compile-opts smallpt2_d.bc > smallpt2_do.bc
 
 llvm-ld -native -ltango-base-ldc -ltango-user-ldc -ldl -lm -lpthread
 -internalize-public-api-list=_Dmain -o=smallpt2_do smallpt2_do.bc
 
 I produce a binary that's faster than the C++ version (22.7 seconds
 instead of 29 seconds).

bearophile, just out of interest, what's the performance like if you ran 
ldmd -O3 -release -inline smallpt2_d.d ? If everything worked right -std-
compile-opts should do the same that LDC does at -O3. And why would 
internalizing _Dmain have a noticable performance impact?

bearophile <bearophileHUGS lycos.com> writes:

Christian Kamm:

 bearophile, just out of interest, what's the performance like if you ran 
 ldmd -O3 -release -inline smallpt2_d.d ?

I don't remember what ldmd is. Without LTO the performance of the LDC compile
was a bit lower than the G++ compile.


 And why would internalizing _Dmain have a noticable performance impact?

It's the only way I have found to perform link-time-optimization with LDC1. It
assumes the whole program is compiled at once, so it's able to perform more
cross optimizations.

Bye,
bearophile

Jacob Carlborg <doob me.com> writes:

On 2011-08-09 22:04, bearophile wrote:
 Christian Kamm:

 bearophile, just out of interest, what's the performance like if you ran
 ldmd -O3 -release -inline smallpt2_d.d ?

 I don't remember what ldmd is. Without LTO the performance of the LDC compile
was a bit lower than the G++ compile.

ldmd is a wrapper that converts dmd arguments to ldc arguments. So you 
can call ldc with the same arguments you would call dmd.

-- 
/Jacob Carlborg

Trass3r <un known.com> writes:

 If I compile it with with LDC1 with Link Time Optimization + Interning:
 ldc -O3 -release -inline -output-bc smallpt2_d.d

 opt -std-compile-opts smallpt2_d.bc > smallpt2_do.bc

 llvm-ld -native -ltango-base-ldc -ltango-user-ldc -ldl -lm -lpthread  
 -internalize-public-api-list=_Dmain -o=smallpt2_do smallpt2_do.bc

Is there a more convenient way to get LTO with LDC?

bearophile <bearophileHUGS lycos.com> writes:

Trass3r:

 Is there a more convenient way to get LTO with LDC?

This is the best/only one I have found for LDC1. Feel free to search for
something better/shorter (and tell me if you find it, please).

Bye,
bearophile

Trass3r <un known.com> writes:

Am 10.08.2011, 14:54 Uhr, schrieb bearophile <bearophileHUGS lycos.com>:

 Trass3r:

 Is there a more convenient way to get LTO with LDC?

 This is the best/only one I have found for LDC1. Feel free to search for  
 something better/shorter (and tell me if you find it, please).

Well, obviously you need to install the gold linker and that llvm plugin.
But I don't know how to go on after that.
LDC does have -O4 and -O5 switches but they probably don't work?!

Trass3r <un known.com> writes:

 LDC does have -O4 and -O5 switches but they probably don't work?!

Had a quick look at the source, no indications that -O4 and O5 do anything  
more than -O3.

Robert Clipsham <robert octarineparrot.com> writes:

On 10/08/2011 14:49, Trass3r wrote:
 LDC does have -O4 and -O5 switches but they probably don't work?!

 Had a quick look at the source, no indications that -O4 and O5 do
 anything more than -O3.

They don't, there's no way to do LTO built into LDC. I seem to recall 
some conversation about adding it, I may just be remembering this thread 
though *g*

-- 
Robert
http://octarineparrot.com/

Trass3r <un known.com> writes:

Am 10.08.2011, 18:16 Uhr, schrieb Robert Clipsham  
<robert octarineparrot.com>:
 On 10/08/2011 14:49, Trass3r wrote:
 Had a quick look at the source, no indications that -O4 and O5 do
 anything more than -O3.

 They don't, there's no way to do LTO built into LDC. I seem to recall  
 some conversation about adding it, I may just be remembering this thread  
 though *g*

Why is there no way?
Can't LDC be modified to support gold+plugin just like clang?

Robert Clipsham <robert octarineparrot.com> writes:

On 10/08/2011 17:43, Trass3r wrote:
 Am 10.08.2011, 18:16 Uhr, schrieb Robert Clipsham
 <robert octarineparrot.com>:
 On 10/08/2011 14:49, Trass3r wrote:
 Had a quick look at the source, no indications that -O4 and O5 do
 anything more than -O3.

 They don't, there's no way to do LTO built into LDC. I seem to recall
 some conversation about adding it, I may just be remembering this
 thread though *g*

 Why is there no way?

Patches welcome ;)

 Can't LDC be modified to support gold+plugin just like clang?

Of course it can - I said "built into", not "in". The only reason it's 
not there is because no one has written support for it.

-- 
Robert
http://octarineparrot.com/