[webkit-dev] Transformations precision: double -> float

[Gabor Rapcsanyi]

On 10/12/2012 11:23 AM, Tom Hudson wrote:
> Do you have particular bottleneck functions that you have reason to
> believe would benefit from ARM SIMD optimizations? Under what
> workloads, and how much benefit?
>
> We've see a few ridiculously expensive transform functions in WebKit
> that my colleagues in Chapel Hill might be able to recall, but usually
> it's the rendering engines that are hot, not WebKit itself.
Yes, you're right that the rendering engine is usually the basic bottleneck,
but why shouldn't we optimize WebKit where it is possible.
For example the function what you asking is 
TransformationMatrix::multiply().
I made a test which could be speed up with optimizing this matrix 
multiplication.
I uploaded here: https://bugs.webkit.org/show_bug.cgi?id=98913
> We have seen places where double math appears to be absolutely
> required for correctness: matrix inversion and path convexity testing;
> there are some other areas we're experimenting with that seem to
> require it as well. You'd probably need to survey WebKit for stability
> (or just hack in doubles and run layout tests) to see how many other
> places will break if you throw away the extra precision.
I tried it in the Qt port and there were just a few (47) layout fails with
0.0 and -0.0 problems and one or two pixel differences. But we need more
investigations here.
> The rest of Chromium's renderer (Skia) uses float, and we've deployed
> SSE/NEON and played with float vs double quite a bit there. The
> float->double conversion at the WebKit/Skia boundary doesn't seem to
> be too much of a drag, but doing it in any sort of inner loop can
> really kill performance.
>
> Meanwhile, there are a lot of places where switching from ARM assembly
> to NEON in the renderer hasn't given us as much win as we'd hoped for,
> which is why the question "how much benefit?" seems important to
> answer before spending too much time polishing a set of typedefs.
>
> Tom
I can say the same as above. If there is a way to optimize WebKit,
but the benefit is just a half or one percent on an avrage page now
or maybe less, why shouldn't we do that, and of course on a specific
page/case we would benefit more.

-Gabor

>
> On Fri, Oct 12, 2012 at 10:08 AM, Gabor Rapcsanyi
> <rgabor at inf.u-szeged.hu> wrote:
>> Yes I agree that we need to do some measures to see the benefits of floats.
>> But it's not just about the performance and memory gain of the floats, I
>> would like to make ARM SIMD optimizations for the transformation functions
>> to fasten up them, but ARM NEON technology only works with floats that's why
>> I want to change the doubles.
>> I would prefer a typedef solution to make this configurable as Ryosuke said,
>> and then the ports could chose whether they need floats, doubles or long
>> doubles.
>>
>> - Gabor
>>
>>> On Oct 10, 2012, at 9:00 AM, Gabor Rapcsanyi <rgabor at inf.u-szeged.hu>
>>> wrote:
>>>
>>>> That was a long time ago and there were no objections
>>>>> Unless there's something in the spec requiring double precision it makes
>>>>> sense to move away from double precision throughout WebKit.
>>> I’m a little concerned about this.
>>>
>>> The programming interface to the graphics system on OS X uses doubles, so
>>> it’s likely we’ll be introducing float to double conversions for at least
>>> that platform if we change things inside WebKit that are currently doubles
>>> to float en masse.
>>>
>>> Similarly, the native data type for numbers in JavaScript is double. So
>>> any time we are passing a number from JavaScript to WebKit it’s likely that
>>> additional double to float conversions will be required as we change
>>> internals to use doubles instead of floats.
>>>
>>> On the other hand, there may also be performance boost or memory gains
>>> from using float instead of double.
>>>
>>> If we do make this change, I think we need to do performance testing as we
>>> do so, perhaps on more than one platform, to see if these changes are
>>> leading to benefits or not.
>>>
>>> I also think that “we should always use float unless a spec mandates
>>> double” is probably not the right rule of thumb. There are many factors
>>> other than “a spec” at play here.
>>>
>>> -- Darin
>>
>> _______________________________________________
>> webkit-dev mailing list
>> webkit-dev at lists.webkit.org
>> http://lists.webkit.org/mailman/listinfo/webkit-dev

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.webkit.org/pipermail/webkit-dev/attachments/20121015/022a3115/attachment.html>