738865 - Use Small tiles should not be android only.

Status: RESOLVED FIXED

(Core :: Graphics, defect)

Description

[Oleg Romashin (:romaxa)]

Attachment: More generic sub texture upload detection

I was playing with b2g on beagle board, and tried to figure out why it is fast on S Galaxy II and slow in my case...
by checking vendors/renderers tricks I found that using small tiles and avoid subTexture upload makes rendering as on beagleboard as fast as on S Galaxy II... because powervr SGX 530 there has similar problem.

I'm wondering can we make that check more generic?

with this fix, I have FPS increase for http://bubblemark.com/dhtml.htm from 10FPS->45FPS

Comment 1

[Patrick Walton (:pcwalton)]

Comment on attachment 608928 [details] [diff] [review]
More generic sub texture upload detection

Looks fine to me, but Chris should review this.

Comment 2

[Chris Lord [:cwiiis]]

Comment on attachment 608928 [details] [diff] [review]
More generic sub texture upload detection

Review of attachment 608928 [details] [diff] [review]:
-----------------------------------------------------------------

r+ with the function renamed, but I think BenWa should review this really (sorry for the ping-pong).

::: gfx/gl/GLContext.cpp
@@ +618,5 @@
>          return false;
>  
>      // Don't use small tiles otherwise. (If we implement incremental texture upload,
>      // then we will want to revisit this.)
>      return false;

whoops, we have implemented incremental texture upload, we should probably review this...

::: gfx/gl/GLContext.h
@@ +712,5 @@
>  
>      bool CanUploadSubTextures();
>      bool CanUploadNonPowerOfTwo();
>      bool WantsSmallTiles();
> +    virtual bool SupportRenderToTexture() { return true; }

This doesn't return whether render-to-texture is available, but whether fast texture upload is available. It should be renamed to reflect what it does, and it should return false by default.

Perhaps HasFastDirectUpdate() instead?

Comment 3

[Benoit Girard (:BenWa)]

Comment on attachment 608928 [details] [diff] [review]
More generic sub texture upload detection

Review of attachment 608928 [details] [diff] [review]:
-----------------------------------------------------------------

::: gfx/gl/GLContext.cpp
@@ +578,5 @@
>  
>      // On PowerVR glTexSubImage does a readback, so it will be slower
>      // than just doing a glTexImage2D() directly. i.e. 26ms vs 10ms
>      if (Renderer() == RendererSGX540 || Renderer() == RendererSGX530)
> +        return SupportRenderToTexture();

I don't want to make a change like until we some performance number, and ideally a small benchmark hack patch and/or some details on how it measure. We've made several assumptions in the pass about performance and checking them after a change is made is very tedious.

I'd be happy to take this patch without this change until we get these numbers.

Comment 4

[Oleg Romashin (:romaxa)]

perf numbers are in first comment, also there are perf numbers in right above that check:
>// than just doing a glTexImage2D() directly. i.e. 26ms vs 10ms
Direct texture rendering with RendererSGX530 supported only on N9... all other devices having same problem as android

Comment 5

[Oleg Romashin (:romaxa)]

Attachment: Use small tiles not only on android

Comment 6

[Benoit Girard (:BenWa)]

Comment on attachment 609019 [details] [diff] [review]
Use small tiles not only on android

As discussed on IRC I was asking for perf numbers for KHRLock, feel free to submit a patch for that if we have evidence that it is faster.

Comment 7

[Ryan VanderMeulen [:RyanVM]]

https://hg.mozilla.org/integration/mozilla-inbound/rev/2c14032c09e1

Comment 8

[Ed Morley [:emorley]]

https://hg.mozilla.org/mozilla-central/rev/2c14032c09e1