By the way, I should probably say again that we don't expect the superoptimiser to work on Windows, even if it were ported.
We currently believe that the only way to get reliable timings for the Windows ABI is on a Linux server (with the same CPU architecture as the target Windows machine). However, we haven't run any experiments of this since we are still writing the initial version of the superoptimiser, and functionality for the Windows ABI would have to be added once it is working reliably. Currently we keep finding pieces of code that it isn't able to handle, and we are currently having trouble getting reliable enough timings on some machines for the superoptimiser to actually work. That is consuming all of Alex's time for now, though it is working reliably on one of our machines with some of the MPIR asm files and with some of Jens' new code. Bill. On 12 May 2016 at 15:14, Bill Hart <goodwillh...@googlemail.com> wrote: > > > On 12 May 2016 at 14:54, Brian Gladman <b...@gladman.plus.com> wrote: > >> On 12/05/2016 12:52, 'Bill Hart' via mpir-devel wrote: >> > Hi Brian, >> > >> > I don't think the superoptimiser can be made to run on Windows. >> > Experience dictates timings are not stable enough to run a >> > superoptimiser, and we don't have the resources to port it. Alex is only >> > with us another month and the person we hire to replace him will be >> > working on actually using the superoptimiser, not further development >> of it. >> >> Hi Bill, >> >> That is disapointing. But, timing issues apart, is it written with >> portability in mind? >> > > No, that's not been possible. > > We had to start with an existing asm Jit library. This would need to be > ported first. > > >> >> I ask because I am suspicious that unless i can run it myself, MPIR on >> Windows x64 will simply not get the attention it needs in maintaining a >> good level of performance. >> > > That's possible of course. > > We will make the code available publicly, and you are welcome to try and > port it if you want. > > >> > We've discussed alternatives though, such as optimising it for the >> > Windows ABI on Linux (i.e. feed the data in through different registers >> > than normal). We haven't checked, but it seems like it might be >> feasible. >> >> At least we should be able to work as I did with Jason, that is taking >> the gcc assembler, adding register exchanges, prologues and epilogues, >> and then converting to Intel syntax for yasm. But I really hoped that >> we might do better than this. >> > > As I said, we might be able to. It depends how long Alex takes to finish > the work he is doing, and how capable his replacement is. > > >> >> Anyway, what is the priority for new assembler code going into the MPIR >> repository? I must admit that my development machine is getting rather >> dated (ivybridge) so I won't be able to do much with the more recent >> additions such as FMA and the like. Which is partly why I was hoping to >> run the superoptimiser myself as I can't see older machines getting much >> priority. >> > > Our priority is on each of the complete microarchitecture redesigns Intel > and AMD have done. However, we are severely limited in what we have access > to. > > We have a Haswell and Jens has a Skylake. On the AMD side we have nothing > recent. I'm considering purchasing a recent AMD machine and maybe later a > recent Intel. > > >> >> But my three year Dell support contract ends in August so I will be >> spending money on my next wokstation (yes, I have to buy my own >> machines!) and it would help to know that MPIR on Windows x64 has a >> future as this could influence what I do. >> >> with my regards, >> >> Brian >> >> -- >> You received this message because you are subscribed to the Google Groups >> "mpir-devel" group. >> To unsubscribe from this group and stop receiving emails from it, send an >> email to mpir-devel+unsubscr...@googlegroups.com. >> To post to this group, send email to mpir-devel@googlegroups.com. >> Visit this group at https://groups.google.com/group/mpir-devel. >> For more options, visit https://groups.google.com/d/optout. >> > > -- You received this message because you are subscribed to the Google Groups "mpir-devel" group. To unsubscribe from this group and stop receiving emails from it, send an email to mpir-devel+unsubscr...@googlegroups.com. To post to this group, send email to mpir-devel@googlegroups.com. Visit this group at https://groups.google.com/group/mpir-devel. For more options, visit https://groups.google.com/d/optout.
