594352 - Optimize texture conversion when no conversion needed

Status: RESOLVED FIXED

(Core :: Graphics: CanvasWebGL, defect)

Description

[Benoit Jacob [:bjacob] (mostly away)]

Attachment: Optimize texture conversion when no conversion needed

When srcFormat == dstFormat and srcPremultiplied == dstPremultiplied, we can directly use memcpy(). To allow flipping, this patch does one memcpy() per row.

Comment 1

[Benoit Jacob [:bjacob] (mostly away)]

Attachment: Optimize texture conversion when no conversion needed

Oops, forgot the 'return'. Not this is _actually_ an optimization.

Comment 2

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

Comment on attachment 473032 [details] [diff] [review]
Optimize texture conversion when no conversion needed

this we definitely want, regardless of template optimization.  However, why do we need memcpy at all here?  Can't we just.. not convert?  No extra allocation, memcpy, etc.

Comment 3

[Benoit Jacob [:bjacob] (mostly away)]

This is a separate optimization. What you describe is something we should do when no conversion AND no flipping is needed. When flipping, we still need 1 memcpy per row.

Comment 4

[Benoit Jacob [:bjacob] (mostly away)]

Attachment: Skip unneeded texture conversions

Improved patch:

 * we now avoid doing anything at all (use src buffer in place, no ConvertImage call, no memcpy) when possible, i.e. when the source texture has the right format, premultiplication, stride and no flipY is requested.

 * otherwise, when the format and premultiplication are good but there may still be a need to flip / change stride, we do 1 memcpy per row in ConvertImage()

Comment 5

[Benoit Jacob [:bjacob] (mostly away)]

on second thought --- we never have to change the stride. so the only reason why we may have to do 1 memcpy per row, is when flipping the image.

Comment 6

[Benoit Jacob [:bjacob] (mostly away)]

http://hg.mozilla.org/mozilla-central/rev/4dac41e6edd2