Bug 24648 – Optimize TransformationMatrix

• Status: RESOLVED
• Resolution: FIXED
• Component: Layout and Rendering

[Simon Fraser (smfr)]

We should make a pass over TransformationMatrix and optimize it to reduce copying. For example, translate() and translate3d() could easily be optimized. multiply() could be factored with multLeft() to avoid matrix copies, and more.

[Simon Fraser (smfr)]

Created attachment 29183 [details]
Patch, changelog

[Simon Fraser (smfr)]

Comment on attachment 29183 [details]
Patch, changelog

Landed in http://trac.webkit.org/changeset/42172

Keeping bug open for further optimization.

[Simon Fraser (smfr)]

Created attachment 29214 [details]
Patch to optimize map functions for translations

Comment on attachment 29214 [details]
Patch to optimize map functions for translations

> +        return FloatPoint(static_cast<float>(p.x() + m_matrix[3][0]), static_cast<float>(p.y() + m_matrix[3][1]));

I slightly prefer casting the matrix element to float rather than casting the coordinate to double, adding in doubles, and casting the result down to float. r=me either way.

[Simon Fraser (smfr)]

Last patch checked in as http://trac.webkit.org/changeset/42207

I think this is enough for now.

[Xan Lopez]

(In reply to comment #5)
> Last patch checked in as http://trac.webkit.org/changeset/42207
> 
> I think this is enough for now.
> 

     // Like the version above, except that it rounds the mapped point to the nearest integer value.
-    IntPoint mapPoint(const IntPoint&) const;
+    IntPoint mapPoint(const IntPoint& p) const
+    {
+        return roundedIntPoint(mapPoint(p));
+    }

Won't this call itself until the stack blows up? Because I'm getting exactly this with your patch in reddit.com for example.

[Simon Fraser (smfr)]

Ah, I guess it will. I need to call the FloatPoint version.

[Simon Fraser (smfr)]

Fixed mapPoint in http://trac.webkit.org/changeset/42228