On Fri, 6 Nov 2015, Bernd Schmidt wrote: > On 11/05/2015 10:51 PM, Martin Jambor wrote: > > Individual changes are described in slightly more detail in their > > respective messages. If you are interested in how the HSAIL > > generation works in general, I encourage you to have a look at my > > Cauldron slides or presentation, only very few things have changed as > > far as the general principles are concerned. Let me just quickly stress > > here that we do acceleration within a single compiler, as opposed to > > LTO-ways of all the other accelerator teams. > > Realistically we're probably not going to reject this work, but I still want > to ask whether the approach was acked by the community before you started. I'm > really not exactly thrilled about having two different classes of backends in > the compiler, and two different ways of handling offloading.
Realistically the other approaches werent acked either (well, implicitely by review). Not doing an RTL backend for NVPTX would have simplified your life as well. Not doing an RTL backend practically means not going the LTO way as you couldn't easily even build a target without RTL pieces (not sure how big a "dummy" RTL target would be). Richard. > > I also acknowledge that we should add HSA-specific tests to the GCC > > testsuite but we are only now looking at how to do that and will > > welcome any guidance in this regard. > > Yeah, I was looking for any kind of new test, because... > > > the class of OpenMP loops we can handle well is small, > > I'd appreciate more information on what this means. Any examples or > performance numbers? > > > Bernd > > -- Richard Biener <rguent...@suse.de> SUSE LINUX GmbH, GF: Felix Imendoerffer, Jane Smithard, Graham Norton, HRB 21284 (AG Nuernberg)
