Hi Kenneth, Thanks for the extensive response! From your proposed changes, it seems enhancing the existing EasyBlock for GCC sounds like the best solution and sounds 'doable'. I hope to find some time to experiment with the EasyBlock soon. Indeed, I had the same idea that the hands-on sessions during the EasyBuild user meeting might be a good to work out any issues I might encounter along the way...
Regards, Caspar From: "Kenneth Hoste" <[email protected]> To: "easybuild" <[email protected]> Sent: Saturday, 23 December, 2017 20:18:04 Subject: Re: [easybuild] GCC with OpenMP 4.5 / OpenACC offloading capabilities Hi Caspar, On 22/12/2017 10:52, Caspar van Leeuwen wrote: Dear Easybuilders, We would like to install a GCC with OpenMP 4.5 / OpenACC with offloading capabilities to nvptx targets as described here [ https://gcc.gnu.org/wiki/Offloading#How_to_try_offloading_enabled_GCC | https://gcc.gnu.org/wiki/Offloading#How_to_try_offloading_enabled_GCC ] We would of course prefer to do this installation with EasyBuild. As far as I can determine, no EasyConfigs are available for this yet, and I even doubt whether this can be build with the conventional GCC EasyBlock: the installation procedure for a GCC with offloading support is quite different, as nvptx-tool, nvptx-newlib, an accelerator GCC compiler and a host GCC compiler all need to be built (and it is suggested to install all in the same directory), through repeated configure/make/make installs (see e.g. this build script: [ https://gist.github.com/kristerw/4e9a735f2d755ffa73f9bf27edbf3c29 | https://gist.github.com/kristerw/4e9a735f2d755ffa73f9bf27edbf3c29 ] ). My questions to you: 1) Has anyone already tried building a GCC with offloading support though EasyBuild? 2) Specifically to the Gent people who wrote the GCC EasyBlock: would you think this can somehow be done with the current EasyBlock (my guess would be: no)? If not, would you think the existing EasyBlock should be adapted to support this, or would you feel a separate EasyBlock would be better? I personally wouldn't like 2 GCC EasyBlocks; on the other hand, incorporating it in the existing one results in a more complicated EasyBlock. I must immediately say that I am not yet sure if I can invest the time to write / adapt an EasyBlock for this, but I just want to explore my options and your thoughts on this. After taking a quick look at this, it looks like you could enhance the existing easyblock for GCC with fairly limited effort to build a GCC with offloading capabilities. A lot of the logic in the current easyblock for GCC is there to first build a GCC that can be used to build the CLooG & PPL libraries, to finally obtain a GCC that has support for using these libraries. See the code paths guarded with "if self.stagedbuild". All this stuff is probably irrelevant for you (unless you want to build a GCC that supports both offloading and CLooG/PPL...). >From the looks of it, these are a couple of changes you would need to do: * add an extra custom easyconfig parameter for GCC to enable building with offloading support, let's call it "with_offloading" * disable building with "make bootstrap" in the build step if "with_offloading" is enabled (or maybe not, you may still want to do this for the host compiler actually...) * build the accel compiler (either for Intel MIC or Nvidia PTX) if "with_offloading" is enabled, this should probably be done in the build step before the current last line there ("super(EB_GCC, self).build_step()") * build the host compiler with the right configure options when "with_offloading" is enabled (this should probably be done at the end of build step) You could probably just ignore the GCC easyblock all together and use a (very) complicated easyconfig file that leverages the generic ConfigureMake easyblock (since we already support running multiple cycles of configure-make-install in a generic way by using a list of string values for configopts rather than a single string value, see for example the older FFTW easyconfigs), but you'll probably find yourself making nasty hacks to get it to work the way you need to (just a hunch, it may actually work just fine). Forking off another GCC easyblock specifically for this doesn't seem like a good idea either, I agree with you. It may feel cleaner initially, but it would probably bite you later (for example when the existing EB_GCC is tweaked, which somehow causes the EB_GCC_offloading easyblock that would sit on top of EB_GCC to break..). Note that the support for "eb -x" would be of great help here, since it gives you a pretty good idea of the effect of changes you make to an easyblock, see [ http://easybuild.readthedocs.io/en/latest/Extended_dry_run.html | http://easybuild.readthedocs.io/en/latest/Extended_dry_run.html ] for more info. This may be worth looking at during the hands-on sessions during the EasyBuild User Meeting... regards, Kenneth BQ_BEGIN Regards, Caspar van Leeuwen SURFsara BQ_END
