[Numpy-svn] [numpy/numpy] aa3668: ENH: optimize count_boolean_trues further via unro...

[GitHub]

  Branch: refs/heads/master
  Home:   https://github.com/numpy/numpy
  Commit: aa36680aac25fb5560f28b6ed94d4e4f654d1d81
      https://github.com/numpy/numpy/commit/aa36680aac25fb5560f28b6ed94d4e4f654d1d81
  Author: Julian Taylor <jtaylor.debian at googlemail.com>
  Date:   2014-03-25 (Tue, 25 Mar 2014)

  Changed paths:
    M numpy/core/src/multiarray/item_selection.c

  Log Message:
  -----------
  ENH: optimize count_boolean_trues further via unrolling

redoing the computation if the booleans do not not contain 0 or 1 allows
sharing registers for the addition and or operations so we can more
aggressively unroll the loop to save 3 multiplies and shifts. Also it
allows better pipelining of integer operations.
Improvements range from 10%-50% on a range of amd64 cpus.


  Commit: 9d9c9527b23b24ba236df92d90548d2c47d19be5
      https://github.com/numpy/numpy/commit/9d9c9527b23b24ba236df92d90548d2c47d19be5
  Author: Charles Harris <charlesr.harris at gmail.com>
  Date:   2014-03-25 (Tue, 25 Mar 2014)

  Changed paths:
    M numpy/core/src/multiarray/item_selection.c

  Log Message:
  -----------
  Merge pull request #4548 from juliantaylor/boolcount-improve

ENH: optimize count_boolean_trues further via unrolling


Compare: https://github.com/numpy/numpy/compare/2f0494c2601b...9d9c9527b23b