Conversation

Notices

1. Adrian Cochrane (alcinnz@floss.social)'s status on Wednesday, 15-Mar-2023 03:32:02 JST Adrian Cochrane

   GNU MultiPrecision includes a LibMPN subsystem implemented in Assembly for a wide variety of CPUs. Normally I'd discuss the x86_64 implementation since that's what I'm running, but instead I'll study the ARM64 implementation since that's simpler!

   This includes routines for:
   * Inverting a digit using bitwise-logic, adds, multiplies, & subtracts.
   * Hamming distance *mostly* consists of bitwise logic with extensive controlflow, & an accumulator.
   * Digit squaring with low, high, & mid branches.

   1/

   • Adrian Cochrane (alcinnz@floss.social)'s status on Wednesday, 15-Mar-2023 03:53:32 JST Adrian Cochrane

     in reply to

     * Tightloops supporting datatable lookups.
     * Greatest-Common-Divisor tightloop.
     * Data-copy tightloop routines.
     * Lef-shift routine wrapped controlflow over initial bits.
     * Various bitwise routines branching over initial bits to select a "top" or "mid" codepaths.
     * Another Greatest-Common-Divisor tightloop, with fastpath for smaller numbers.
     * Examine initial bits to choose "top" or "mid" codepaths for multiplication.
     * Accumulator routines for computing remainders.

     2/?

   • Adrian Cochrane (alcinnz@floss.social)'s status on Wednesday, 15-Mar-2023 04:11:36 JST Adrian Cochrane

     in reply to

     * Examine initial bits to choose "top" or "mid" codepaths for addition or subtraction or comparison routines.
     * Various other routines are implemented similarly, though they have a couple "lo" codepaths necessitating altering the examination of the bits. I won't continue going over them.
     * This includes popcount, with a divide&conquer coldpath.
     * There's division on digits implemented using multiplication.
     * Tightloop specifically for pi (judging by name) with even tighter trailing loop.

     3/?

   • Adrian Cochrane (alcinnz@floss.social)'s status on Wednesday, 15-Mar-2023 04:17:33 JST Adrian Cochrane

     in reply to

     There's dataheaders determining which opcodes to use on Cora53, Cora57, Cora72, Cora73, XGene1, & other ARM CPUs.

     On Cora53 it LibMPN implements special variations of it's comparison routines.

     3.1/3.1 Fin for today! Tomorrow: MPQ!

     As mentioned yesterday, MPF calls down to MPN. Also: This may be implemented differently for different CPU architectures.