If you’re using a typical consumer computer, you’ll get limitations of memory bandwidth, which it seems will limit simple calculations on large amounts of data. It would probably be worth ensuring one’s installation of PDL is compiled with -O3 just in case; -O2 (the usual default) enables vectorisation on clang, but not on GCC which only does so on -O3.
I just did a bit more experimenting with very latest PDL on a MacBook with 6 cores/12 hyperthreads (which apparently defaults to -O3). For comparison, normal Perl takes about 28ms for 1000 iterations, so C will be about 1ms. Best performance was with PDL_AUTOPTHREAD_SIZE=0 PDL_AUTOPTHREAD_TARG=10 (11 was about 1.5x as long), where 1000 iterations took about 0.31ms, or a bit over 3x quicker than C, and comparable with the JavaScript (which I suspect benefits from using GPU or maybe just multicore). This 2019 presentation (https://indico.cern.ch/event/814979/contributions/3401203/attachments/1831468/3115808/VectorParallelismMultiCoreProc.pdf) discusses the various issues in making parallel process go Really Fast. For me, a key takeaway is the problem is generally quite hard, and it’s wise to use e.g. BLAS where all the possible optimisations have been wrung out. PDL could benefit from that by parsing the “Code” etc sections, and inserting BLAS calls. Similarly, we should probably start using LAPACK in core, like GNU Octave etc do. An interesting possibility would be to use the “Matriplex” library for vectorising operations on many smallish matrices (it even generates code using Perl). It also mentions Amdahl’s Law, which gives limits to parallelism speedups (fundamentally, the non-parallelisable bits impose limits, including main-memory access). From: Karl Glazebrook<mailto:karlglazebr...@mac.com> Sent: 26 October 2021 08:57 To: Ed .<mailto:ej...@hotmail.com> Cc: Luis Mochan<mailto:moc...@icf.unam.mx>; pdl-de...@lists.sourceforge.net<mailto:pdl-de...@lists.sourceforge.net>; perldl<mailto:pdl-general@lists.sourceforge.net> Subject: Re: [Pdl-devel] benchmarks This thread is interesting. I was wondering if anyone has ever seen speedups of 2x or better with PDL_AUTOPTHREAD_TARG > 2? I find it tends to max out at around 1.5-1.7x whatever I set. I know about overhead etc. but kind of feel for some of the basic stuff (e.g. A=B*C for large arrays with big chunks) I should see 4x for PDL_AUTOPTHREAD_TARG=4 and never do) The various numbers in the tests reported by Ed show <2x. Nice getting faster than C! Karl On 4 Oct 2021, at 1:05 am, Ed . <ej...@hotmail.com<mailto:ej...@hotmail.com>> wrote: Thank you for the independent measurement! From: Luis Mochan<mailto:moc...@icf.unam.mx> Sent: 03 October 2021 15:03 To: pdl-de...@lists.sourceforge.net<mailto:pdl-de...@lists.sourceforge.net>; perldl<mailto:pdl-general@lists.sourceforge.net> Subject: Re: [Pdl-devel] benchmarks Now I have run the C benchmark and Ed's. My results are: | Program | # iterations | time (s) | speed (K/s) | factor | |--------------+--------------+----------+-------------+--------| | ansi c | 150e6 | 133 | 1127.8195 | 1. | | perl | 1.5e6 | 56 | 26.785714 | 42.1 | | my pdl | 15e6 | 67 | 223.88060 | 5.0 | | Ed's pdl | 15e6 | 16 | 937.5 | 1.2 | | Ed's 4 cores | 15e6 | 11 | 1363.6364 | 0.8 | So, as Ed wrote, just by stting and environment variable, perl+pdl+pp_def can be made faster than c. On Sat, Oct 02, 2021 at 07:03:50PM -0500, Luis Mochan wrote: > I made my own version of the ray-tracing program (as I tried to > understand it). I didn't use pp_def, only Perl and ordinary PDL. I used > ... -- o W. Luis Mochán, | tel:(52)(777)329-1734 /<(*) Instituto de Ciencias Físicas, UNAM | fax:(52)(777)317-5388 `>/ /\ Av. Universidad s/n CP 62210 | (*)/\/ \ Cuernavaca, Morelos, México | moc...@fis.unam.mx<mailto:moc...@fis.unam.mx> /\_/\__/ GPG: 791EB9EB, C949 3F81 6D9B 1191 9A16 C2DF 5F0A C52B 791E B9EB _______________________________________________ pdl-devel mailing list pdl-de...@lists.sourceforge.net<mailto:pdl-de...@lists.sourceforge.net> https://lists.sourceforge.net/lists/listinfo/pdl-devel _______________________________________________ pdl-devel mailing list pdl-de...@lists.sourceforge.net<mailto:pdl-de...@lists.sourceforge.net> https://lists.sourceforge.net/lists/listinfo/pdl-devel
_______________________________________________ pdl-general mailing list pdl-general@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/pdl-general
