Re: OGR-NG on Xeon Phi?
"Dead J. Dona" <[email protected]> Thu, 13 Dec 2012 17:37:04 +0200
| Newsgroups | gmane.comp.distributed-net.rc5 |
|---|---|
| Message-ID | <CA+YvA3oHdzsLAHN907DqWCM_aybCkfAP9u+WZqDcHw0h1t42Pw@mail.gmail.com> |
found this on intel forums. Intel finally released the documentation of the instruction set of the announced Xeon Phi which is the new brand name for Knights Corner. The document is downloadable at http://software.intel.com/file/44500. Very interesting to see is that this marvelous CPU features most of the instructions present up to Pentium (incl. x87) but misses MMX, SSE, AVX, and other things like CMOV. On the other hand there is a whopping set of 32 (!) registers of 512 (!) bits each and a bunch of pepped up commands (incl. FMA and scatter/gather) for single/double floats and dword/qword integers IMHO better than AVX2. The omnipresent vector masks ease odd loops and can do other magics as well (similar to PDEP/PEXT). Swizzling/converting is built into most commands. The only thing I'm missing is a direct support for byte and word integer arithmetic but this is only a minor speed penalty. The new "coprocessor" is announced to have at least 50 cores (probably 62) with 4-fold hyperthreading. 248 threads! === wbr, Me. Dead J. Dona On Thu, Dec 13, 2012 at 2:17 PM, Roman Trunov <[email protected]> wrote: > [email protected] wrote: > > > Additional, the Phi-Cores are equipped with 512Bit SIMD-Units. I dont > know any detail of their architecture or commandset, but maybe > this could be a great benefit for OGR? > > The answer really depends on architecture and command set of these > units, and latencies of operations. > > OGR-NG operates on bitmaps of 256 bits long, 512 bits units looks > excessive. More important is which kinds of rotations are supported. > High-bitness SIMD architectures usually cannot rotate a register by > arbitrary number of bits, which is required for OGR. Instead, you have > to use two operations - rotate by a whole number of bytes first, then by > remaining 0-7 bits. If new command set could overcome this limitation, > it could give us some gain. Otherwise, it's quite useless. > > Roman > _______________________________________________ > rc5 mailing list > rc5-Ra3b/QYEcJ3d140v2zMXi0fjHoOT/h/[email protected] > http://lists.distributed.net/mailman/listinfo/rc5 > _______________________________________________ rc5 mailing list rc5-Ra3b/QYEcJ3d140v2zMXi0fjHoOT/h/[email protected] http://lists.distributed.net/mailman/listinfo/rc5