572651 - ThebesLayer shader program (GetBGRXLayerProgram) is slow on mobile...

Status: RESOLVED FIXED

(Core :: Graphics, defect)

Description

[Oleg Romashin (:romaxa)]

Attachment: Patch which is making layer scrolling 1.3x faster

I was testing ThebesLayer scrolling performance on EGL (maemo). and simple layer texture scrolling is slow there.

I have found (with Bas help) that sBGRXTextureLayerFS  using uLayerOpacity calculation.

I've tested on device with and without uLayerOpacity, and without opacity we have ~1.3x faster layer scrolling.

Comment 1

[Oleg Romashin (:romaxa)]

Comment on attachment 451909 [details] [diff] [review]
Patch which is making layer scrolling 1.3x faster

Need feedback and some explanation how to proceed with this...
Should I create separate shader program without opacity for each existing program, or something else should be done here?

Comment 2

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

mmm, annoying that the extra multiply costs that much, but I guess I can believe that it would dominate.  I like this here (though I don't like the "2" name :-), though it would be nice if we could autogenerate this.. let me think about it for a bit.

Comment 3

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

Oleg, can you try using the original shader program, but before the draw call for those layers, if opacity == 1.0, can you try disabling GL_BLEND (and reenabling it afterward)?  See what that does to performance.

Comment 4

[Oleg Romashin (:romaxa)]

> for those layers, if opacity == 1.0, can you try disabling GL_BLEND (and

checked, gl()->fDisable(LOCAL_GL_BLEND) does not help at all.

Comment 5

[Oleg Romashin (:romaxa)]

Attachment: GetBGRX/[A]OpaqueLayerProgram's added

Comment 6

[Oleg Romashin (:romaxa)]

Attachment: Shader version with no opacity

Comment 7

[Oleg Romashin (:romaxa)]

Attachment: Updated patch to upstream

Comment 8

[Oleg Romashin (:romaxa)]

Here is the results of Simple scrolling with Shadow layers GL on Maemo

With patch        ~ 55 FPS
Without patch m-c ~ 39 FPS

Vlad asked me to try some other option
>Oleg, for the immediate problem, since you're able to get good performance numbers -- can you try changing all the precision to 'lowp' and see if that makes a difference? 
> Also try changing the expression to be vec4(texture2D(..) * uLayerOpacity, uLayerOpacity), and see what change that makes (and the two together).

mediump - lowp change test - ~ 55 FPS

-  gl_FragColor = vec4(texture2D(uTexture, vTexCoord).rgb, 1.0) * uLayerOpacity;
+  gl_FragColor = vec4(texture2D(uTexture, vTexCoord).rgb * uLayerOpacity, uLayerOpacity);
vec4(texture2D(..) * uLayerOpacity, uLayerOpacity) ~ 38 FPS


lowp + vec4(texture2D(..) * uLayerOpacity, uLayerOpacity)  ~ 54.9 FPS

Comment 9

[Oleg Romashin (:romaxa)]

Attachment: mediump->lowp precision... not sure

No sure if it does not break anything...

Comment 10

[Oleg Romashin (:romaxa)]

Comment on attachment 487467 [details] [diff] [review]
mediump->lowp precision... not sure

mediump->lowp, tested with CSS prop "opacity: 0.5;" with webgl layer... layer is transparent and works fine

Comment 11

[Oleg Romashin (:romaxa)]

Attachment: Use right program in canvas and image

Comment 12

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

Comment on attachment 488285 [details] [diff] [review]
Use right program in canvas and image

This really should have gone into a separate bug with some more detail :-)

Also I'm pretty sure the patch commit message isn't correct here.

Comment 13

[Oleg Romashin (:romaxa)]

This is valid part from previous patch.
and yep, that is message from different patch...

Comment 14

[Robert O'Callahan (:roc) (email my personal email if necessary)]

+                                      useGLContext != 0 ? PR_TRUE : PR_FALSE)

You can just write "useGLContext != 0"

Would similar optimizations be at all useful on D3D, I wonder?

Comment 15

[Tuukka Tolvanen (sp3000)]

http://hg.mozilla.org/mozilla-central/pushloghtml?changeset=6511fac93b8b

Comment 16

[Oleg Romashin (:romaxa)]

> 
> Would similar optimizations be at all useful on D3D, I wonder?

probably yes, but I don't have D3D on mobile ;(

Comment 17

[Robert O'Callahan (:roc) (email my personal email if necessary)]

Hmm, so with low precision enabled, removing the multiply by uLayerOpacity doesn't give any improvement? Weird.

Comment 18

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

I think at that point it's the same number of vector multiplies, so it ends up being a wash -- the hardware can do (iirc) 4 low precision multiplies in a cycle, so we can either do (RGB,1.0)*opacity or (RGB*opacity,opacity) in the same amount of time.