368408 - Tiny SVG performance improvement

Status: RESOLVED WORKSFORME

(Core :: SVG, defect)

Description

[Steve Snyder]

Running SeaMonkey v1.1.0 on a 2.4GHz Pentium4-based Windows system the SVG performance is terrible, even with the MMX re-enabled in the build.  Profiling shows that about 40% of the execution time is spent in Microsoft's _lldiv() routine.

The problem is that 32-bit CPUs are just crushed by the repeated 64-bit fixed-point division used in _compute_x().  My significant improvements to this routine suffer from the occasional fatal division overflow, but I noticed this trivial improvement while doing them.

The signed long long division is so burdensome on 32-bit CPUs that just about any attempts to avoid them are cost-effective.

Comment 1

[Steve Snyder]

Attachment: Reduces the frequency of 64-bit fixed-point division

This patch avoids the 64-bit division in _compute_x() where possible.

In the benchmark http://www.hixie.ch/tests/adhoc/svg/perf/002-with-workarounds-for-asv3.svg the 64-bit division is avoided about 80,000 times, or 1% (yes, one percent) of the of the total calls to _compute_x().

For the Tab Preview feature of SeaMonkey _compute_x() is called for every preview but it seems that the division can be avoided 100% of the time.

Yes, a 1% improvment is pretty trivial, but I wanted to get this "on the record" given that SVG perforamnce needs all the help it can get.

Comment 2

[Robert Longson [:longsonr]]

(In reply to comment #1)
> Created an attachment (id=253013) [details]
> Reduces the frequency of 64-bit fixed-point division
> 

Have you reported this upstream in the cairo project http://cairographics.org/? It needs to get fixed there.

Comment 3

[Steve Snyder]

(In reply to comment #2)
> Have you reported this upstream in the cairo project http://cairographics.org/?

I have now.  I suspect, though, the the Cairo code present in the Gecko 1.8.1 code base is so old (v1.0.2, isn't it?) that it won't be relevant to the current (1.3.x) Cairo code.

Comment 4

[Jonathan Watt [:jwatt]]

No, we're at 1.3.12 on trunk. http://lxr.mozilla.org/seamonkey/source/gfx/cairo/README

Comment 5

[Ryan VanderMeulen [:RyanVM]]

Is this still relevant for Cairo 1.5.2, which is the current version on the trunk?

Comment 6

[Steve Snyder]

(In reply to comment #5)
> Is this still relevant for Cairo 1.5.2, which is the current version on the
> trunk?

The _compute_x() function, in which all that divsion was done, no longer exists in the trunk (24 Dec) copy of Cairo.

Displaying many tab preview images in SeaMonkey v2.0a1pre, I don't see _lldiv() consuming much CPU time.

So it seems that this is *not* relevant to trunk builds.

Comment 7

[:Gavin Sharp [email: [e-mail]]]

Given that this is no longer an issue on the trunk, and doesn't match the criteria for landing on the branch, this bug should be closed as INVALID or WORKSFORME, right?