Bug 92123 – Fix arithmetic composite filter for auto-vectorization

• Component: SVG
• Status: RESOLVED FIXED

[Allan Sandfeld Jensen]

The arithmetic composite filter can currently not be auto-vectorized by GCC, but with a few changes many common cases could take a path that can.

[Allan Sandfeld Jensen]

Created attachment 154078 [details]
Patch

[WebKit Review Bot]

Comment on attachment 154078 [details]
Patch

Attachment 154078 [details] did not pass chromium-ews (chromium-xvfb):
Output: http://queues.webkit.org/results/13323652

New failing tests:
svg/filters/filter-placement-issue.svg

[WebKit Review Bot]

Created attachment 154086 [details]
Archive of layout-test-results from gce-cr-linux-01

The attached test failures were seen while running run-webkit-tests on the chromium-ews.
Bot: gce-cr-linux-01  Port: <class 'webkitpy.common.config.ports.ChromiumXVFBPort'>  Platform: Linux-2.6.39-gcg-201203291735-x86_64-with-Ubuntu-10.04-lucid

[Allan Sandfeld Jensen]

Created attachment 154293 [details]
Patch

Fixing the detection of whether the unclamped version can be used. This has become more specific, but still represents many common use-cases.

[Nikolas Zimmermann]

Comment on attachment 154293 [details]
Patch

View in context: https://bugs.webkit.org/attachment.cgi?id=154293&action=review

Looks good, r=me with minor suggestions:

> Source/WebCore/platform/graphics/filters/FEComposite.cpp:167
> +        scaledK1 = k1 / 255.f;

Nitpick, we use 255.0f in WebCore/SVG code.

> Source/WebCore/platform/graphics/filters/FEComposite.cpp:189
> +    float upperLimit = fmaxf(0.f, k1) + fmaxf(0.f, k2) + fmaxf(0.f, k3) + k4;

Any specific reason for not using std::max here? Both parameters are floats.

> Source/WebCore/platform/graphics/filters/FEComposite.cpp:193
> +    if (k4 >= 0.f && k4 <= 1.f
> +        && upperLimit >= 0.f && upperLimit <= 1.f
> +        && lowerLimit >= 0.f && lowerLimit <= 1.f) {

No need for additional lines here, we don't care for line lengths in SVG code.

> Source/WebCore/platform/graphics/filters/FEComposite.cpp:198
> +        if (!k4) {
> +            if (!k1)
> +                computeArithmeticPixelsUnclamped<0, 0>(source, destination, pixelArrayLength, k1, k2, k3, k4);
> +            else
> +                computeArithmeticPixelsUnclamped<1, 0>(source, destination, pixelArrayLength, k1, k2, k3, k4);

Can you swap all of these? if (k4) { if (k1) { ...} to avoid the negations.

> Source/WebCore/platform/graphics/filters/FEComposite.cpp:208
> +    if (!k4) {

Ditto.

[Nikolas Zimmermann]

(In reply to comment #0)
> The arithmetic composite filter can currently not be auto-vectorized by GCC, but with a few changes many common cases could take a path that can.

Out of curiosity: how do you test that the auto-vectorization is applied? Do you inspect assembly or are there other tricks to let gcc show you what kind of optimizations it applied?

[Allan Sandfeld Jensen]

(In reply to comment #6)
> (In reply to comment #0)
> > The arithmetic composite filter can currently not be auto-vectorized by GCC, but with a few changes many common cases could take a path that can.
> 
> Out of curiosity: how do you test that the auto-vectorization is applied? Do you inspect assembly or are there other tricks to let gcc show you what kind of optimizations it applied?

I use -ftree-vectorizer-verbose, =2 tells you if it is applied, but I normally use =5 or =6, which tells me why it is not applied when it isn't.

[Allan Sandfeld Jensen]

(In reply to comment #7)
> (In reply to comment #6)
> > (In reply to comment #0)
> > > The arithmetic composite filter can currently not be auto-vectorized by GCC, but with a few changes many common cases could take a path that can.
> > 
> > Out of curiosity: how do you test that the auto-vectorization is applied? Do you inspect assembly or are there other tricks to let gcc show you what kind of optimizations it applied?
> 
> I use -ftree-vectorizer-verbose, =2 tells you if it is applied, but I normally use =5 or =6, which tells me why it is not applied when it isn't.

Btw, I was hoping to be able to use GCC pragma, or function attributes to mark functions that should be auto-vectorized, but I can't seem to get neither 

#pragma GCC optimize("tree-vectorize") 

nor

__attribute__((optimize("tree-vectorize)))

to do anything. I will probably open GCC bug on that, but once that works, these and other functions that could benefit from vectorization could be specifically marked, so only they are optimized aggresively and we do not risk library-bloat from aggresively optimizing irrelevant parts of WebCore.

[Allan Sandfeld Jensen]

Committed r123605: <http://trac.webkit.org/changeset/123605>