Created attachment 167306 [details] Add options for LTO, OpenMP and BLAS/LAPACK of choice There are a couple of enhancements that can be made to math/R regarding optimization: one would be to enable LTO and OpenMP whenever the compiler supports it, another one would be to allow linking to OpenBLAS besides ATLAS or R's internal BLAS implementation. The patch was inspired by math/octave's handling of BLAS. Please fin attached a patch to add these options to the user.
Created attachment 170792 [details] Add new options to math/R Add some new options to the port: build using LTO, OpenMP and link with different BLAS libraries. Additionally fix some warnings reported by poudriere testport (x11 missing modules mainly).
testbuilds@work
Building with LTO selected fails on 10.3amd: http://people.freebsd.org/~pi/logs/math__R-103-1465089577.txt Building fails on 9.3a, can you investigate ? http://people.freebsd.org/~pi/logs/math__R-93a-1465090000.txt
Sure, I'll have a look at it. The port is compiling fine for me on 10.3/amd64. First thing that I notice in the logs is that gcc is version 5 while I am compiling with 4.8. I'll check that.
Thanks for the pointer to gcc. I have DEFAULT_VERSIONS=gcc=5 in my make.conf for the 10.3a poudriere, that's why. I'm surprised it turns out to cause trouble.
Ok, that solves one of my questions. Somehow gcc5 seems to not find/install the lto plugin. I'll do some research on that. As to 9.3, this seems an issue with older FreeBSD versions having limited support for long double math [1], thus the configure script finds support for long double type, but the compiler does not find expl nor longl. There is a a configure option to disable long double support, I'll add that for older versions. What I wonder now is how the original port is compiling on 9.3. I don't see the long-double option disabled anywhere. [1] https://lists.freebsd.org/pipermail/freebsd-hackers/2009-March/028030.html
Looking at svn log /usr/src/lib/msun/src/math.h, I have found the following commit which brings an actual implementation for expl (logl should have been added in a previous commit): ---------- r271779 | tijl | 2014-09-18 17:10:22 +0200 (jue, 18 sep 2014) | 10 lines MFC r257770 r257818 r257823 r260066 r260067 r260089 r260145 r268587 r268588 r268589 r268590 r268593 r268597 r269758 r270845 r270847 r270893 r270932 r270947 r271147 Merge libm work by kargl, bde and das from the past few months. Besides optimisations and small bug fixes this includes new implementations for C99 functions expl, coshl, sinhl, tanhl, erfl and erfcl. ---------- Now checking the OS version for that commit with svn cat -r r271779 sys/sys/param.h: #define __FreeBSD_version 1000716 /* Master, propagated to newvers */ How should I handle this in the port so that long double support is disabled in older FreeBSD? Should I check for OS version 1000716 or rather a different one? Thanks!
Created attachment 171764 [details] Adds performance related options and fix compilation on 9.3 Fixes build issues on 9.3 with Long Double data type, plus adds the options from previous patches.
Testbuilds@work
Committed, thanks. I was busy the last few days, sorry for the delay.
A commit references this bug: Author: pi Date: Sat Jun 25 19:10:20 UTC 2016 New revision: 417477 URL: https://svnweb.freebsd.org/changeset/ports/417477 Log: math/R: Add optimization related options to Makefile There are a couple of enhancements that can be made to math/R regarding optimization: one would be to enable LTO and OpenMP whenever the compiler supports it, another one would be to allow linking to OpenBLAS besides ATLAS or R's internal BLAS implementation. The patch was inspired by math/octave's handling of BLAS. PR: 207425 Submitted by: elferdo@gmail.com Approved by: bf (maintainer timeout) Changes: head/math/R/Makefile head/math/R/pkg-plist
There is a problem with these changes: the default package does not include any more the shared lib (only if the option RBLAS is selecte"d, which is not the default). That breaks the depending ports which need this lib, e.g. math/rkward-kde4. See http://beefy6.nyi.freebsd.org/data/101amd64-default/417508/logs/rkward-kde4-0.6.5.log
Thanks for the pointer, I'll check if RBLAS can be made the default.
A quick glance: looks like the LIBR option should be set per default, which would fix the build problem of rkward-kde4 ?
My bad, sorry! I can take the fix if you will. As it stands right now, handling those two slave ports in math/R is somewhat painful. I'll try to rewrite the Makefile, but I wonder if they are useful at all...
(In reply to Thierry Thomas from comment #12) test-build rkward-kde4 with LIBR set in math/R. So probably this is enough ?
(In reply to Kurt Jaeger from comment #16) Yes, it is. Could it be set as DEFAULT?
A commit references this bug: Author: pi Date: Sun Jun 26 14:03:10 UTC 2016 New revision: 417525 URL: https://svnweb.freebsd.org/changeset/ports/417525 Log: math/R: add LIBR to OPTIONS_DEFAULT to allow math/rkward-kde4 to build PR: 207425 Reported by: thierry Changes: head/math/R/Makefile
Fixed, thanks!