C++Talk.NET C++ language newsgroups   Archives   FAQ   Search   Memberlist   Usergroups   Register   Profile   Log in to check your private messages   Log in  Compared valarray and blitz. Surprise.          C++Talk.NET Forum Index -> C++ Language (Moderated) View previous topic :: View next topic   Author Message Piotr L. Guest Posted: Tue Sep 21, 2004 11:39 am    Post subject: Compared valarray and blitz. Surprise. Hallo I'd like to write a small program, which reads and manipulates not so small data (scanned pictures aprox. 100MB each). So in the beginning I wanted to choose some kind of container to store the image data in memory. I have written simple 'benchmark test' for native C array, STL::vector, STL::valarray, boost::array, blitz::Array. This test does nothing more but assigns value to each element of container (sometimes in different ways (loops, algorithms, = operators)). Here is the source of my "benchmarks" http://main.ams.edu.pl/~piotrlg/vector-test.cc Compiled on linux, using g++ 3.3.3 Both boost and blitz libs were rather new ones. I wanted to find the fastest one. The planned operations on the image data are rather simple ones, so sophisticated functions of the container aren't the top priority. I was very surprised with the results. On my P4 2GHz the winner is valarry/vector! I have seen rather bad opinions about valarray class. Not only about its design but also about performance. The advice was to use blitz libraries instead of valarray. So could someone explain me what should I think about the following table? Sorted results ----------------------- fill vect time: 0.108265 for_each vect time: 0.108998 valarray time: 0.111239 vect C loop time: 0.115843 boost array for_each: 0.123644 boost array assign: 0.124048 C array time: 0.205496 boost array C loop: 0.211093 generate vect time: 0.256162 blitz fill array time: 0.329419 blitz for_each array : 0.33393 blitz array time: 1.13735 Please note, that blitz Array is almost 3 times slower than vector or valarray. Please also note the poor C array is not the fastest ;-) So valarray is not that bad and blitz is hm slow. I hoped that the library mentioned and suggested as the valarray killer could be at least that good. Could someone explain me why I see such poor blitz performance? Is it because the nature of the very primitive "test suite"? So I'm confused now. What to choose? BTW. I'm not a programmer. I just want to write a program for my own use. Please also note that this is one of my first C++ programs. One more thing. Please note, that when I run this test second time (just right after the first is finished) I can see that 'blitz array time: ' accelerates a lot. Strange (?). Sorted results ----------------------- fill vect time: 0.10828 vect C loop time: 0.10846 valarray time: 0.10907 for_each vect time: 0.109135 boost array for_each: 0.115181 boost array assign: 0.117081 C array time: 0.196561 boost array C loop: 0.206079 blitz array time: 0.209327 <-- previously over 1 sec. generate vect time: 0.25631 blitz for_each array : 0.322798 blitz fill array time: 0.323558 Regards Piotr Legiecki [ See http://www.gotw.ca/resources/clcm.htm for info about ] [ comp.lang.c++.moderated. First time posters: Do this! ] Back to top Thomas Richter Guest Posted: Tue Sep 21, 2004 6:24 pm    Post subject: Re: Compared valarray and blitz. Surprise. Hi, Quote: I have written simple 'benchmark test' for native C array, STL::vector, STL::valarray, boost::array, blitz::Array. This test does nothing more but assigns value to each element of container (sometimes in different ways (loops, algorithms, = operators)). On my P4 2GHz the winner is valarry/vector! I have seen rather bad opinions about valarray class. Not only about its design but also about performance. The advice was to use blitz libraries instead of valarray. So could someone explain me what should I think about the following table? IMHO, your tests are not really very realistic. You typically have more to do on the vectors than just to fill them with constant values. Thus, you're currently only testing one function that could even end up as a built- in. My personal test case - because it was realistic for my applications - was the implementation of a FIR filter on arrays, or rather, an elementary filter step. Take the sum of two vectors, multiply by a constant, add result to aonther vector, measure time. The result was that the standard C array was the fastest possibility (without choosing some vector formats, i.e. SIMD instructions, the platform had to offer). The compiler generated code for the valarray looked pretty bad, worse than the C code, and no compiler I tried was able to generate the SIMD form "out of the box". My personal opinion about the valarrays is that I really don't see a use for them; if it really matters, I've to do it by hand anyhow. Quote: So I'm confused now. What to choose? Write a test case that minimics your typical operations on the data you'll perform later, then test again. Unless, of course, your typical operation is filling of vectors with constants, your benchmark doesn't really help. Quote: One more thing. Please note, that when I run this test second time (just right after the first is finished) I can see that 'blitz array time: ' accelerates a lot. Strange (?). As a general hint: Always run your main program loop several times, dispose the timing of the first loops, then take the average of the last twenty loops or so. The first loops might not only allocate memory, they may also move data into the cache. Depending on the size of your vectors, memory might then be partially in the cache and the code will perform much better. This is especially true for the P4 architecture and its rather "microscopic" first level cache of 8K. If you can arrange your data to fit in there, you win. Everything else is rather secondary as branch prediction and code cache do a good job on the P4. So long, Thomas [ See http://www.gotw.ca/resources/clcm.htm for info about ] [ comp.lang.c++.moderated. First time posters: Do this! ] Back to top Brooks Moses Guest Posted: Tue Sep 21, 2004 10:51 pm    Post subject: Re: Compared valarray and blitz. Surprise. Thomas Richter wrote: [attribution lost] Quote: One more thing. Please note, that when I run this test second time (just right after the first is finished) I can see that 'blitz array time: ' accelerates a lot. Strange (?). As a general hint: Always run your main program loop several times, dispose the timing of the first loops, then take the average of the last twenty loops or so. The first loops might not only allocate memory, they may also move data into the cache. Another thing potentially worth doing is trying a second test with the order of which array index corresponds to the outer and inner loop variables reversed. Some array implementations (most notably those that are attempting to provide behavior similar to Fortran arrays) use a different order than the standard C array, and it's also worth making sure you've got the order right. Which reminds me: while the original poster is testing implementations, the Template Numerical Toolkit that NIST distributes (http://math.nist.gov/tnt/) might be worth checking out. It looks to me like it should be a fairly straightforward wrapper around the standard C array, but may be a little easier to work with. Quote: Depending on the size of your vectors, memory might then be partially in the cache and the code will perform much better. This is especially true for the P4 architecture and its rather "microscopic" first level cache of 8K. If you can arrange your data to fit in there, you win. Everything else is rather secondary as branch prediction and code cache do a good job on the P4. Indeed -- but this is rather difficult to do with the original poster's 100MB image files! - Brooks -- The "bmoses-nospam" address is valid; no unmunging needed. [ See http://www.gotw.ca/resources/clcm.htm for info about ] [ comp.lang.c++.moderated. First time posters: Do this! ] Back to top Glen Low Guest Posted: Wed Sep 22, 2004 10:06 am    Post subject: Re: Compared valarray and blitz. Surprise. Quote: My personal test case - because it was realistic for my applications - was the implementation of a FIR filter on arrays, or rather, an elementary filter step. Take the sum of two vectors, multiply by a constant, add result to aonther vector, measure time. The result was that the standard C array was the fastest possibility (without choosing some vector formats, i.e. SIMD instructions, the platform had to offer). The compiler generated code for the valarray looked pretty bad, worse than the C code, and no compiler I tried was able to generate the SIMD form "out of the box". My personal opinion about the valarrays is that I really don't see a use for them; if it really matters, I've to do it by hand anyhow. My implementation of valarray transparently uses SIMD to accelerate such code. The PowerPC Altivec implementation is fairly complete now but I'm working on an SSE/SSE2 version and cleaning up the code somewhat. I have gotten close to the theoretical limit for SIMD acceleration and the generated code looks sharp. Stay tuned. http://www.pixelglow.com/macstl/ IIRC, only one other valarray implementation besides mine uses expression templates i.e. libstdc++ (gcc). Metrowerks, SGI STL and Visual C++ (Dinkumware) don't use this technique, which means their code slows down signficantly with longer expressions as large temp valarrays get allocated. Cheers, Glen Low, Pixelglow Software www.pixelglow.com [ See http://www.gotw.ca/resources/clcm.htm for info about ] [ comp.lang.c++.moderated. First time posters: Do this! ] Back to top Tom Widmer Guest Posted: Wed Sep 22, 2004 6:37 pm    Post subject: Re: Compared valarray and blitz. Surprise. On 21 Sep 2004 14:24:32 -0400, Thomas Richter <thor (AT) cleopatra (DOT) math.tu-berlin.de> wrote: Quote: The compiler generated code for the valarray looked pretty bad, worse than the C code, and no compiler I tried was able to generate the SIMD form "out of the box". My personal opinion about the valarrays is that I really don't see a use for them; if it really matters, I've to do it by hand anyhow. There are valarray implementations around that generate vector assembler instructions. I don't know a SIMD one, but there's an Altivec one (for PowerPC) here: http://www.pixelglow.com/macstl/ I have a feeling that the GCC and libstdc++ team are working on adding vector instructions to their valarray implementation. I don't know how far they've got. Tom [ See http://www.gotw.ca/resources/clcm.htm for info about ] [ comp.lang.c++.moderated. First time posters: Do this! ] Back to top Gabriel Dos Reis Guest Posted: Tue Nov 23, 2004 10:53 am    Post subject: Re: Compared valarray and blitz. Surprise. Tom Widmer <tom_usenet (AT) hotmail (DOT) com> writes: Quote: I have a feeling that the GCC and libstdc++ team are working on adding vector instructions to their valarray implementation. I don't know how far they've got. When I get more time But, measurements I've made on my own applications suggested quite good performance. Part of the perceived problem is that array processing means different things for different people... -- Gabriel Dos Reis [email]gdr (AT) integrable-solutions (DOT) net[/email] [ See http://www.gotw.ca/resources/clcm.htm for info about ] [ comp.lang.c++.moderated. First time posters: Do this! ] Back to top Gabriel Dos Reis Guest Posted: Tue Nov 23, 2004 10:54 am    Post subject: Re: Compared valarray and blitz. Surprise. [email]glenlow (AT) pixelglow (DOT) com[/email] (Glen Low) writes: Quote: http://www.pixelglow.com/macstl/ IIRC, only one other valarray implementation besides mine uses expression templates i.e. libstdc++ (gcc). At the time I originally implemented and reworked GCC's valarray, I also had a use for it and I was quite sensitive to performance loss. I've read Thomas Richter complain a couple of time, but as far as I can determine he never gave me a set of testcases I could run in order to see what the problems might be... -- Gabriel Dos Reis [email]gdr (AT) integrable-solutions (DOT) net[/email] [ See http://www.gotw.ca/resources/clcm.htm for info about ] [ comp.lang.c++.moderated. First time posters: Do this! ] Back to top Display posts from previous: All Posts1 Day7 Days2 Weeks1 Month3 Months6 Months1 Year Oldest FirstNewest First         C++Talk.NET Forum Index -> C++ Language (Moderated) All times are GMT Page 1 of 1   Jump to: Select a forum C/C++----------------C LanguageC Language (Moderated)C++ Language (Moderated)C++ language (comp.lang.c++)C++ language, library and standards Non-English C/C++ Newsgroups----------------C++ (German)C++ (French)  You cannot post new topics in this forum You cannot reply to topics in this forum You cannot edit your posts in this forum You cannot delete your posts in this forum You cannot vote in polls in this forum Powered by phpBB © 2001, 2006 phpBB Group SEO toolkit © 2004-2006 webmedic.