Le 11 mai 2013 à 12:40, Marc Joliet <mar...@gmx.de> a écrit : > Am Sat, 11 May 2013 12:02:45 +0200 > schrieb Stéphane Letz <l...@grame.fr>: > >>> >>>> Then we know that the resulting C/C++ code can be quite large and >>>> sometimes push the C compilers to their limits… It some sense the Faust >>>> compiler does produce completely "inlined" code that can be quite huge. >>>> There is still a lot of work to do in the quality of this code, so that >>>> for instance to better share common code that is applied on different >>>> datas, these kind of things... >>> >>> Yes, I use both GCC and Clang here, and Clang handles the code that FAUST >>> produces *much* better than GCC. In the case of my project, g++ alone could >>> take several minutes (>10 I think?) and have peak memory use of over 2G (!), >>> while clang++ compiles much faster (a few seconds), uses much less RAM *and* >>> produces better code (when the code is vectorised, that is). >> >> Yes, it is well known that Clang beats GCC in this area.. But GGC is still >> better at auto-vectorization right? >> >> (AFAICS clang still has very limited form of auto-vectorization …) > > Almost two years ago I was trying out LV2 and wrote a peaking EQ following the > EQ cookbook. To benchmark it, I let it filter 48000 samples of audio. At > that > point, GCC beat Clang by a few ms (something like 20 ms for gcc vs. 23 ms for > clang). After learning about auto-vectorisation, I split the filter loop into > the recursive and non-recursive parts, with no other modifications to the > code. > Suddenly, Clang *beat* GCC by a few ms (now something like 18ms for gcc vs. > 15ms > for clang). So my conclusion at the time was: clang is better with vectorised > code, while GCC is better with scalar code. > > [ Aside: I know that if you write C/C++ by hand you can give GCC hints about > your code (stuff like memory alignment, etc.), in which case it will > produce very good, compact vectorised code. ] > > Now, in the case of my Regalia-Mitra-Project, I'll just quote my README: > > NOTE: it turns out that on my computer (AMD Athlon64 X2 4200+ from late > 2006) clang++ produces faster binaries than g++. MBStereophony uses about > 15% CPU on average when compiled with g++ (with CFLAGS="-O3 > -march=native"), while it only uses about 9% CPU on average when compiled > with clang++. > > This is, of course, with vectorisation. And for completeness: I use clang-3.1 > (with llvm-3.1). > > So I believe that Clang *is* better at auto-vectorisation than GCC and has > been > for a while now, or are my observations too simplistic?
Without looking precisely at the generated assembly code I would not be sure ((-; Stéphane ------------------------------------------------------------------------------ Learn Graph Databases - Download FREE O'Reilly Book "Graph Databases" is the definitive new guide to graph databases and their applications. This 200-page book is written by three acclaimed leaders in the field. The early access version is available now. Download your free book today! http://p.sf.net/sfu/neotech_d2d_may _______________________________________________ Faudiostream-users mailing list Faudiostream-users@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/faudiostream-users
