Re: [css3-background] Where we are with Blur value discussion

On Tue, Jul 20, 2010 at 5:47 PM, Tab Atkins Jr. <> wrote:
> On Mon, Jul 19, 2010 at 3:53 PM, Aryeh Gregor <> wrote:
>> Then it will look different between different browsers.  We should try
>> to not write specs that allow this unless there's a specific good
>> reason, like respecting platform conventions or limiting optimization
>> opportunities.  If browsers use entirely different blur algorithms,
>> then they will diverge a lot in edge cases at the very least, and
>> that's not interoperable.  Mandating that browsers should use one
>> particular *pixel-specific* blur algorithm (or close approximations
>> thereto) is the correct thing to do, if implementers are okay with it
>> -- which isn't clear to me at this point.
> I'm digging in deeper with our Skia devs to see what the precise
> reasoning was behind using what appears to be a single box-blur.  As I
> said before, I think it was for performance reasons, but I'm trying to
> get an answer from the horse's mouth before I come down hard on one
> side of the issue or another.
> If it turns out that the performance savings of Skia blurs really
> aren't that important, and that it would be fine to just change our
> impl to do a triple box-blur like I think everyone else does, then we
> can define the conformance criteria as each pixel having to be within
> X% of a gaussian blur with stdev equal to half the blur length.  (X%
> being probably 3% or 5%.)  That'd pass everyone who's doing pretty
> blurs right now, and pass Chrome on all platforms once we change Skia
> to be prettier.

Well, so far nobody I've talked to actually knows why Skia blurs were
made so ugly, and they all want prettier blurs.  We've had a bug to
make it prettier for a while (though nobody's had the bandwidth to
pick it up).

So, we'll consider it a non-problem, and assume that in the future
Chrome will use pretty blurs everywhere.  Everyone else is already
using pretty blurs, where by "pretty" I mean "closely approximates a
gaussian blur".  So, there doesn't seem to be any good reason not to
just require a gaussian.

I suggest requiring blurs to approximate a gaussian blur with a stdev
equal to half the blur length, with a 5% variance from the value that
would be given by a true gaussian blur permitted.

(I don't know whether the reported "within 3% accuracy" I hear about
for a triple box-blur is +-3%, or +-1.5%.  I'm assuming the former.
If the latter, then a smaller permitted variance is probably


Received on Thursday, 22 July 2010 21:01:17 UTC