Differences between revisions 19 and 20
Revision 19 as of 2009-05-17 23:07:11
Size: 2716
Editor: WilliamHart
Comment:
Revision 20 as of 2009-05-17 23:08:42
Size: 2819
Editor: WilliamHart
Comment:
Deletions are marked like this. Additions are marked like this.
Line 50: Line 50:
 * Up to 48 8.5Gb/s tranceivers or 24 11.3 gbps tranceivers
 * Deal with 533MHz DDR3 memory directly

MPIR - Parallel Algorithms and CUDA

Present : Carl Witty, Bill Hart, Michael Abshoff, Glenn Tarbox Virtually Present : Jeff Gilchrist, Gonzalo Tornaria

You can chat in a Linux text console by installing "irssi" and running: "irssi -c irc.freenode.net" and then type "/join #sage-devel"

Parallel algorithms:

  • Multimodular algorithms
  • Scalar algorithms
  • Peter Montgomery's remainder algorithm a mod b, precompute b1 = B mod b, b2 = B2 mod b, b3 = B3 mod b, then write a = a0 + a1*B + a2*B^2 +..., then compute a0 + a1*b1 + a2*b2 +.... and do final reduction mod b. Multiplications can be done in parallel.

  • Addition and subtraction can be parallelised using nails - non-unique representation of numbers
  • Classical algorithm is embarrassingly parallel - bad if you have an n log n algorithm in that range

Glenn Tarbox (Owner of cuda1, AMD K10 with NVIDA CUDA card - expert on large scale parallelisation)

  • What are the top level integration issues, e.g. by libraries using MPIR

Michael Abshoff (Sage release manager)

  • Link into Sage via cython and link in CUDA

CUDA documentation:

CUDA issues:

  • Memory bandwidth limits algorithms - matrices n**2 entries to get in and out, matrix multiplication O(n**2.7), but for integers n limbs to get in and out O(n log n log log n) operations to multiply

Other Options:

  • AMD Math library AML provides BLAS interface uses GPU - but that's for linear algebra
  • PTX NVIDIA GPU assembler code for inner loops

Gonzalo Tornaria (theta functions expert)

  • Is there a way to encode integer multiplication in linear algebra? (A. Perhaps vectors - multimodular, but not matrices)
  • Kernel
  • Launch threads - issues based on hierarchy of memory - CPU registers-> memory per processor block-> main graphics memory-> system memory

  • Can launch all the threads on all cpus in a couple of cycles
  • How GPU would compare to carefuly programmed FPGA?
  • E.g a Stratix IV can have around 1000 18x18 multipliers, but maybe that's not too much, and this is probably very expensive hardware
  • http://www.altera.com/products/devices/stratix-fpgas/stratix-iv/stxiv-index.jsp

  • Carl Witty does FPGA programming - says it is probably very expensive
  • Accoding to the spec the stratix can have parallel high-bandwith communication
  • Up to 48 8.5Gb/s tranceivers or 24 11.3 gbps tranceivers
  • Deal with 533MHz DDR3 memory directly

Jeff Gilchrist

  • What about ATI hardware - why not support OpenCL?
  • Carl Witty says -

Bill Hart

  • Cell port will happen as it is funded by EPSRC Grant - will be proof of principle code to apply for a port to Cell2Xi

Glenn Tarbox

  • Flame - for overall integration of libraries

CUDA (last edited 2009-05-17 23:53:03 by WilliamHart)