607540 - consider using SSE to speed up fetch_scanline_a8/store_scanline_a8

Status: NEW

(Core :: Graphics, defect)

Description

[Timothy Nikkel (:tnikkel)]

While profiling for bug 579488 I found that fetch_scanline_a8 was the top function, taking over 10% of the time spent in Firefox. store_scanline_a8 took another 3.6%. 1.9% was spent in fetch_scanline_x8r8g8b8. The rest was under 1% or had sse2 in the name.

If the READ macro is the simple one in gfx/cairo/libpixman/src/pixman-accessor.h (is it? I'm not sure) then I would think these functions should be fairly easy to speed up with SSE.

Comment 1

[Vladimir Vukicevic [:vlad] [:vladv] (needinfo me, slow to respond)]

would be intersting to know the full sequence of operations that takes place here, from fetch to store.  I think if you hit the generic fetch/store paths you're already going down a multi-step path, and there's probably a single-step function that could be written.

Comment 2

[Timothy Nikkel (:tnikkel)]

Attachment: simple testcase

Sorry for the delay.

Basically all you need is to set up a rounded rectangle clip and do some painting, even a simple color fill, aka drawing a div with border-radius and a background color.

I tried clipping to the regular rect, pushing a group, and then popping it, applying the rounded rect clip and then painting. That has a similar profile.

Comment 3

[Timothy Nikkel (:tnikkel)]

I observe a 3x slowdown in the testcase on Linux when going from no-border-radius to border-radius.

Comment 4

[Siarhei Siamashka]

fetch_scanline_a8 is now SSE2 optimized: http://cgit.freedesktop.org/pixman/commit/?id=08e855f15cba24aac83145b994069d0bb50be5a1

Regarding the missing single-step function. I have a pixman branch (*) with the additional code which can help to identify such cases. The information is reported to syslog and then can be decoded into a human readable form using a script.

* http://cgit.freedesktop.org/~siamashka/pixman/log/?h=playground/slow-path-reporter