Hi Stuart,
On 2018-04-28 02:37, Stuart Marks wrote:
Hi Claes,
I finally got a chance to look at this. Overall looks fine. Although
the randomness is "weaker" iterating in both directions will, I think,
still have the desired effect of flushing out order dependencies.
thanks, glad you agree with the direction here...
I have a test (forthcoming) that verifies that iteration order will
vary from run to run, and it passes with both the old and new code.
Great!
Not an issue, but I'm curious as to why the initial index is chosen
this way:
idx = Math.floorMod(SALT ^ remaining, elements.length);
The floorMod with a divisor of elements.length makes perfect sense.
I'm wondering why the dividend is (SALT ^ remaining) and not just SALT
for example.
Right, adding " ^ remaning" was a play to see if I could increase the
randomness of initial idx distribution for "free" when comparing a set
of N elements with a set of N+1 elements. While it has near zero
performance cost (no added loads, just a quick xor on values we load
anyway), I never got around to measuring if it improved or even worsened
the starting point "randomness". In the end it's not that important, so
I'll just remove it.
/Claes
