[LLVMdev] RFC - Improvements to PGO profile support
bob.wilson at apple.com
Tue Mar 24 10:54:57 PDT 2015
> On Mar 24, 2015, at 10:53 AM, Diego Novillo <dnovillo at google.com> wrote:
> On Tue, Mar 24, 2015 at 1:48 PM, Bob Wilson <bob.wilson at apple.com> wrote:
>>> On Mar 24, 2015, at 10:27 AM, Xinliang David Li <davidxl at google.com> wrote:
>>> 2.3) remove the 'laplace rule of succession' which can be very harmful
>> I have not seen any consensus on this. I’m not at all convinced this would be a good idea. From what I’ve seen, using LaPlace’s rule avoids really bad behavior in cases where the counts are very small and has almost no impact when the counts are large.
> Duncan and I chatted briefly about this on IRC. I'm going to
> experiment with only applying laplace smoothing if any scaled weights
> are 0. AFAIU, usage of this rule is really just trying to avoid having
> downstream users deal with 0 weights. Duncan, please whack me with a
> clue stick if I totally misrepresented our conclusions.
Zero is the extreme case, but it’s also important for other small counts. I’d like to see some specific examples of why you think the current behavior is harmful.
More information about the llvm-dev