Re: [Haskell-cafe] benchmarking c/c++ and haskell
-BEGIN PGP SIGNED MESSAGE- Hash: SHA1 On 9/15/10 06:41 , Daniel Fischer wrote: > play a role. Since ghc fares pretty well on 64-bit linux (David Terei and > Don), both, via the NCG and via C, it seems its 64-bit code generator is > better than its 32-bit code generator :( amd64 has more registers; IIRC ghc needs a *lot* of registers to play with to perform well (TBH *everything* does), so x86 is always going to be behind the curve. - -- brandon s. allbery [linux,solaris,freebsd,perl] allb...@kf8nh.com system administrator [openafs,heimdal,too many hats] allb...@ece.cmu.edu electrical and computer engineering, carnegie mellon university KF8NH -BEGIN PGP SIGNATURE- Version: GnuPG v2.0.10 (Darwin) Comment: Using GnuPG with Mozilla - http://enigmail.mozdev.org/ iEYEARECAAYFAkySRUMACgkQIn7hlCsL25XG5QCfWOqBnm/jwFXM9L5cDU33yJiO JYQAoLnytUMIZ2XHuWKcnin7fEWJk66u =0FBK -END PGP SIGNATURE- ___ Haskell-Cafe mailing list Haskell-Cafe@haskell.org http://www.haskell.org/mailman/listinfo/haskell-cafe
Re: [Haskell-cafe] benchmarking c/c++ and haskell
On Wednesday 15 September 2010 12:22:24, Vo Minh Thu wrote: > Thanks all for your answsers. I still wonder why some people get very > different results between gcc and ghc, and some others don't. A > difference in processor? Architecture (32/64-bit, x86/..., ...), processor, gcc version, OS, all play a role. Since ghc fares pretty well on 64-bit linux (David Terei and Don), both, via the NCG and via C, it seems its 64-bit code generator is better than its 32-bit code generator :( And judging from David's 32-bit Windows timings, GHC's 32-bit code generator is better adapted to Windows than to linux, but its generated C code is horrible for Windows. > > I guess I will crank up a little package using criterion and producing > two executables to make sure anyone who run the benchmark use the same > options, and ease the whole process. > > Thanks, > Thu ___ Haskell-Cafe mailing list Haskell-Cafe@haskell.org http://www.haskell.org/mailman/listinfo/haskell-cafe
Re: [Haskell-cafe] benchmarking c/c++ and haskell
Thanks all for your answsers. I still wonder why some people get very different results between gcc and ghc, and some others don't. A difference in processor? I guess I will crank up a little package using criterion and producing two executables to make sure anyone who run the benchmark use the same options, and ease the whole process. Thanks, Thu 2010/9/15 Daniel Fischer : > On Wednesday 15 September 2010 02:50:15, David Terei wrote: >> On 13 September 2010 20:41, Vo Minh Thu wrote: >> > ... the post is from 2008. No LLVM goodness. So I thought GHC 6.12.1 >> > (not the latest and greatest HEAD) would be enough. >> >> I compiled the two programs myself out of curiosity and got the >> following times. >> >> Linux, 64bit, Ubuntu 10.10: >> >> 1e8 >> clang: 0.180s >> gcc: 0.179s >> ghc 6.12.1 (viac): 0.187s >> ghc 6.12.1 (fasm): 0.218s >> ghc HEAD (viac): 0.186s >> ghc HEAD (fasm): 0.179s >> ghc HEAD (llvm): 0.174s >> >> 1e9 >> clang: 1.657s >> gcc: 1.647s >> ghc 6.12.1 (viac): 1.653s >> ghc 6.12.1 (fasm): 1.975s >> ghc HEAD (viac): 1.648s >> ghc HEAD (fasm): 1.658s >> ghc HEAD (llvm): 1.646s >> >> So basically all have the same time except ghc 6.12.1 where fasm is a >> little slow. >> >> On windows xp 32bit I get quite different results which I trust less >> as the times are jumping around much more then they were on linux: >> >> 1e8 >> gcc: 0.365s >> ghc 6.12.1 (viac): 5.287s >> ghc 6.12.1 (fasm): 1.332s >> ghc HEAD (viac): 5.292s >> ghc HEAD (fasm): 0.875s >> ghc HEAD (llvm): 0.359s >> >> Not sure why the results on windows are so different. > > I have no idea why, but I remember that on several occasions timings for > via-C compiled programmes on Windows have been abysmal. > Whether it's a general Windows/gcc mismatch or something GHC-specific, I've > no idea (and can't find out, not having Windows). > > For what it's worth, the approximate timings on my 32-bit linux (openSUSE > 11.1) box have been > > 1e8 > gcc: ~0.4s > 6.12.3 (viac): ~1.2s > 6.12.3 (fasm): ~2.7s > HEAD the same as 6.12.3 > > (I have deleted the programmes and I don't remember the exact timings, if > you're interested, I could reconstruct them). > >> If anyone else >> wants to run the two programs on Windows and check that would be >> great. >> >> Cheers, >> David > > ___ Haskell-Cafe mailing list Haskell-Cafe@haskell.org http://www.haskell.org/mailman/listinfo/haskell-cafe
Re: [Haskell-cafe] benchmarking c/c++ and haskell
On Wednesday 15 September 2010 02:50:15, David Terei wrote: > On 13 September 2010 20:41, Vo Minh Thu wrote: > > ... the post is from 2008. No LLVM goodness. So I thought GHC 6.12.1 > > (not the latest and greatest HEAD) would be enough. > > I compiled the two programs myself out of curiosity and got the > following times. > > Linux, 64bit, Ubuntu 10.10: > > 1e8 > clang: 0.180s > gcc: 0.179s > ghc 6.12.1 (viac): 0.187s > ghc 6.12.1 (fasm): 0.218s > ghc HEAD (viac): 0.186s > ghc HEAD (fasm): 0.179s > ghc HEAD (llvm): 0.174s > > 1e9 > clang: 1.657s > gcc: 1.647s > ghc 6.12.1 (viac): 1.653s > ghc 6.12.1 (fasm): 1.975s > ghc HEAD (viac): 1.648s > ghc HEAD (fasm): 1.658s > ghc HEAD (llvm): 1.646s > > So basically all have the same time except ghc 6.12.1 where fasm is a > little slow. > > On windows xp 32bit I get quite different results which I trust less > as the times are jumping around much more then they were on linux: > > 1e8 > gcc: 0.365s > ghc 6.12.1 (viac): 5.287s > ghc 6.12.1 (fasm): 1.332s > ghc HEAD (viac): 5.292s > ghc HEAD (fasm): 0.875s > ghc HEAD (llvm): 0.359s > > Not sure why the results on windows are so different. I have no idea why, but I remember that on several occasions timings for via-C compiled programmes on Windows have been abysmal. Whether it's a general Windows/gcc mismatch or something GHC-specific, I've no idea (and can't find out, not having Windows). For what it's worth, the approximate timings on my 32-bit linux (openSUSE 11.1) box have been 1e8 gcc: ~0.4s 6.12.3 (viac): ~1.2s 6.12.3 (fasm): ~2.7s HEAD the same as 6.12.3 (I have deleted the programmes and I don't remember the exact timings, if you're interested, I could reconstruct them). > If anyone else > wants to run the two programs on Windows and check that would be > great. > > Cheers, > David ___ Haskell-Cafe mailing list Haskell-Cafe@haskell.org http://www.haskell.org/mailman/listinfo/haskell-cafe
Re: [Haskell-cafe] benchmarking c/c++ and haskell
On Tue, Sep 14, 2010 at 5:50 PM, David Terei wrote: > On 13 September 2010 20:41, Vo Minh Thu wrote: >> ... the post is from 2008. No LLVM goodness. So I thought GHC 6.12.1 >> (not the latest and greatest HEAD) would be enough. > > I compiled the two programs myself out of curiosity and got the following > times. > > Linux, 64bit, Ubuntu 10.10: > > 1e8 > clang: 0.180s > gcc: 0.179s > ghc 6.12.1 (viac): 0.187s > ghc 6.12.1 (fasm): 0.218s > ghc HEAD (viac): 0.186s > ghc HEAD (fasm): 0.179s > ghc HEAD (llvm): 0.174s > > 1e9 > clang: 1.657s > gcc: 1.647s > ghc 6.12.1 (viac): 1.653s > ghc 6.12.1 (fasm): 1.975s > ghc HEAD (viac): 1.648s > ghc HEAD (fasm): 1.658s > ghc HEAD (llvm): 1.646s > > So basically all have the same time except ghc 6.12.1 where fasm is a > little slow. > > On windows xp 32bit I get quite different results which I trust less > as the times are jumping around much more then they were on linux: Thanks for your rather extensive effort to pin down the performance numbers. I just wanted to add a suggestion: I would highly recommend using criterion for this. It's easy, simple, and good at giving you statistically robust measures of the time it takes. http://hackage.haskell.org/package/criterion I resisted using it for a while, but once I took the plunge I was quite happy with the results. It's a library that is definitely worth the spin up time (for me at least). Jason ___ Haskell-Cafe mailing list Haskell-Cafe@haskell.org http://www.haskell.org/mailman/listinfo/haskell-cafe
Re: [Haskell-cafe] benchmarking c/c++ and haskell
On 13 September 2010 20:41, Vo Minh Thu wrote: > ... the post is from 2008. No LLVM goodness. So I thought GHC 6.12.1 > (not the latest and greatest HEAD) would be enough. I compiled the two programs myself out of curiosity and got the following times. Linux, 64bit, Ubuntu 10.10: 1e8 clang: 0.180s gcc: 0.179s ghc 6.12.1 (viac): 0.187s ghc 6.12.1 (fasm): 0.218s ghc HEAD (viac): 0.186s ghc HEAD (fasm): 0.179s ghc HEAD (llvm): 0.174s 1e9 clang: 1.657s gcc: 1.647s ghc 6.12.1 (viac): 1.653s ghc 6.12.1 (fasm): 1.975s ghc HEAD (viac): 1.648s ghc HEAD (fasm): 1.658s ghc HEAD (llvm): 1.646s So basically all have the same time except ghc 6.12.1 where fasm is a little slow. On windows xp 32bit I get quite different results which I trust less as the times are jumping around much more then they were on linux: 1e8 gcc: 0.365s ghc 6.12.1 (viac): 5.287s ghc 6.12.1 (fasm): 1.332s ghc HEAD (viac): 5.292s ghc HEAD (fasm): 0.875s ghc HEAD (llvm): 0.359s Not sure why the results on windows are so different. If anyone else wants to run the two programs on Windows and check that would be great. Cheers, David ___ Haskell-Cafe mailing list Haskell-Cafe@haskell.org http://www.haskell.org/mailman/listinfo/haskell-cafe
Re: [Haskell-cafe] benchmarking c/c++ and haskell
2010/9/13 Daniel Fischer : > On Monday 13 September 2010 11:50:14, Vo Minh Thu wrote: >> 2010/9/13 David Virebayre : >> > Does it help to compile with ghc --make -O2 -funbox-strict-fields ?? >> >> No, it doesn't. Can I assume you don't have the problem I described? > > Currently, GHC's native code generator is not too good at optimising loops. > It might help if you compile via C, > > ghc -O2 -fexcess-precision -fvia-C -optc-O3 > > On my box, that gives a > 2× speedup (unfortunately, that means it takes > almost three times as long as the C version instead of > 6×). With your options, the Haskell code is only 1.5x slower. But still... > If you have the llvm backend for GHC, that is supposedly better for such > code. ... the post is from 2008. No LLVM goodness. So I thought GHC 6.12.1 (not the latest and greatest HEAD) would be enough. Thanks, Thu ___ Haskell-Cafe mailing list Haskell-Cafe@haskell.org http://www.haskell.org/mailman/listinfo/haskell-cafe
Re: [Haskell-cafe] benchmarking c/c++ and haskell
On Monday 13 September 2010 11:50:14, Vo Minh Thu wrote: > 2010/9/13 David Virebayre : > > Does it help to compile with ghc --make -O2 -funbox-strict-fields ?? > > No, it doesn't. Can I assume you don't have the problem I described? Currently, GHC's native code generator is not too good at optimising loops. It might help if you compile via C, ghc -O2 -fexcess-precision -fvia-C -optc-O3 On my box, that gives a > 2× speedup (unfortunately, that means it takes almost three times as long as the C version instead of > 6×). If you have the llvm backend for GHC, that is supposedly better for such code. > > Thanks, > Thu ___ Haskell-Cafe mailing list Haskell-Cafe@haskell.org http://www.haskell.org/mailman/listinfo/haskell-cafe
Re: [Haskell-cafe] benchmarking c/c++ and haskell
2010/9/13 David Virebayre : > Does it help to compile with ghc --make -O2 -funbox-strict-fields ?? No, it doesn't. Can I assume you don't have the problem I described? Thanks, Thu ___ Haskell-Cafe mailing list Haskell-Cafe@haskell.org http://www.haskell.org/mailman/listinfo/haskell-cafe
Re: [Haskell-cafe] benchmarking c/c++ and haskell
Does it help to compile with ghc --make -O2 -funbox-strict-fields ?? David. ___ Haskell-Cafe mailing list Haskell-Cafe@haskell.org http://www.haskell.org/mailman/listinfo/haskell-cafe
Re: [Haskell-cafe] benchmarking c/c++ and haskell
2010/9/12 Andrew Coppin : > Vo Minh Thu wrote: >> >> Hi, >> >> I would like to benchmark C/C++ and Haskell code. The goal is to >> improve the Haskell port[0] of smallpt[1]. >> >> To make sure my approach was reliable, I got the code of two programs >> (one in C, the other in Haskell) from a post[2] by Don. The code is >> reproduced below. When timing the execution of both program, I have a >> >>> >>> 4x difference. It is said on the blog the programs should have >>> >> >> similar performance. >> >> I simply don't get the reason of such a difference. I've tried the >> code on my Atom netbook and also on an older centrino machine. The >> timing are similar (i.e. the C and Haskell program show >4x >> difference). Both machines have GHC 6.12.1 on Linux. >> >> Would you have an idea? >> > > The function floor :: Double -> Int is surprisingly slow under GHC. (IIRC, > it's implemented by converting Double -> (Int, Integer) -> Int, or something > equally absurd.) Poking around GHC.Prim directly allows you to do the same > operation much, much faster. I couldn't say exactly how much of a difference > it makes, but I've had programs go from seconds to microseconds just by > switching this. > > http://hackage.haskell.org/trac/ghc/ticket/2271 > http://hackage.haskell.org/trac/ghc/ticket/1434 > > My solution was to do this: > > http://hackage.haskell.org/packages/archive/AC-Colour/1.1.3/doc/html/src/Data-Colour-FastFloor.html > > Uh... good luck! o_O Thanks for the tip. But I would really get the two little programs from Don's blog to have the same performance as advertised so that I can build from there. I have to make sure things are different for a good reason, not because of a flaw in my setup (which I assume it is since I can't even get those two to run in the same time). Thanks, Thu ___ Haskell-Cafe mailing list Haskell-Cafe@haskell.org http://www.haskell.org/mailman/listinfo/haskell-cafe
Re: [Haskell-cafe] benchmarking c/c++ and haskell
Vo Minh Thu wrote: Hi, I would like to benchmark C/C++ and Haskell code. The goal is to improve the Haskell port[0] of smallpt[1]. To make sure my approach was reliable, I got the code of two programs (one in C, the other in Haskell) from a post[2] by Don. The code is reproduced below. When timing the execution of both program, I have a 4x difference. It is said on the blog the programs should have similar performance. I simply don't get the reason of such a difference. I've tried the code on my Atom netbook and also on an older centrino machine. The timing are similar (i.e. the C and Haskell program show >4x difference). Both machines have GHC 6.12.1 on Linux. Would you have an idea? The function floor :: Double -> Int is surprisingly slow under GHC. (IIRC, it's implemented by converting Double -> (Int, Integer) -> Int, or something equally absurd.) Poking around GHC.Prim directly allows you to do the same operation much, much faster. I couldn't say exactly how much of a difference it makes, but I've had programs go from seconds to microseconds just by switching this. http://hackage.haskell.org/trac/ghc/ticket/2271 http://hackage.haskell.org/trac/ghc/ticket/1434 My solution was to do this: http://hackage.haskell.org/packages/archive/AC-Colour/1.1.3/doc/html/src/Data-Colour-FastFloor.html Uh... good luck! o_O (Of course, I could be speaking complete nonesense and this bug has since been fixed, or the cause of the slowness in smallpt is something else...) ___ Haskell-Cafe mailing list Haskell-Cafe@haskell.org http://www.haskell.org/mailman/listinfo/haskell-cafe
[Haskell-cafe] benchmarking c/c++ and haskell
Hi,
I would like to benchmark C/C++ and Haskell code. The goal is to
improve the Haskell port[0] of smallpt[1].
To make sure my approach was reliable, I got the code of two programs
(one in C, the other in Haskell) from a post[2] by Don. The code is
reproduced below. When timing the execution of both program, I have a
>4x difference. It is said on the blog the programs should have
similar performance.
I simply don't get the reason of such a difference. I've tried the
code on my Atom netbook and also on an older centrino machine. The
timing are similar (i.e. the C and Haskell program show >4x
difference). Both machines have GHC 6.12.1 on Linux.
Would you have an idea?
[0] http://hackage.haskell.org/package/smallpt-hs
[1] http://www.kevinbeason.com/smallpt/
[2]
http://donsbot.wordpress.com/2008/06/04/haskell-as-fast-as-c-working-at-a-high-altitude-for-low-level-performance/
$ gcc -O2 mean.c -omean-c
$ ghc --make -O2 mean.hs -o mean-hs
$ time ./mean-c 1e8
5000.50
real0m1.575s
user0m1.513s
sys 0m0.000s
$ time ./mean-hs 1e8
5000.5
real0m6.997s
user0m6.856s
sys 0m0.013s
-- file mean.hs
module Main where
import System.Environment
import Text.Printf
mean :: Double -> Double -> Double
mean n m = go 0 0 n
where
go :: Double -> Int -> Double -> Double
go s l x | x > m = s / fromIntegral l
| otherwise = go (s+x) (l+1) (x+1)
main = do
[d] <- map read `fmap` getArgs
printf "%f\n" (mean 1 d)
/* file mean.c */
#include
#include
int
main (int argc, char **argv)
{
double d = atof(argv[1]);
double n;
long a;
double b;
for (n = 1,
a = 0,
b = 0; n <= d; b+=n,
n++,
a++)
;
printf("%f\n", b / a);
return 0;
}
___
Haskell-Cafe mailing list
Haskell-Cafe@haskell.org
http://www.haskell.org/mailman/listinfo/haskell-cafe
