998 replies to this topic

[WanWizard]

Hi WanWizard,

So the flashing issue is still there with the modified code?

Yes. Less obvious, but still visible.
 

Has OpenPLi looked at the Enigma2 build options that OpenViX use to get around the hardware / driver issues on HiSi boxes?

 

We don't do workarounds.

 

If there are manufacturer related issues, they have to be passed on to the manufacturer so the root cause can be fixed.

[IanSav]

Hi WanWizard,
 

 

So the flashing issue is still there with the modified code?

Yes. Less obvious, but still visible.

 

 

Would you like me to commit this code or wait for a driver fix?
 

 

Has OpenPLi looked at the Enigma2 build options that OpenViX use to get around the hardware / driver issues on HiSi boxes?

We don't do workarounds.

If there are manufacturer related issues, they have to be passed on to the manufacturer so the root cause can be fixed.

 

Does this mean that you will now take this issue up with HiSi to get a fix. I am sure that all the images would appreciate a proper fix.

Regards,
Ian.

[WanWizard]

Would you like me to commit this code or wait for a driver fix?

 

As it clearly fixes the response issues, I'd say commit.

 

Does this mean that you will now take this issue up with HiSi to get a fix. I am sure that all the images would appreciate a proper fix.

 

If someone can formulate needs to be fixed, yes.

 

As you say OpenVIX has worked around it, can you point out the commits that do?

[IanSav]

Hi All,

 

To answer my own question:
 

Transparency Mapping and Trimming with Alpha

 

The alpha component of textures can be used to solve a number of interesting problems. Intricate shapes such as an image of a tree can be stored in texture memory with the alpha component acting as a matte (1 where the image is opaque, 0 where it is transparent, and a fractional value along the edges). When the texture is applied to geometry, blending can be used to composite the image into the color buffer or the alpha test can be used to discard pixels with a 0 alpha component using GL_ EQUALS test. To maximize performance, set the alpha test to GL_ LESS and discard pixels with a small alpha value, for example less than .05. This way some more pixels are discarded that do not contribute significantly to the image.

 

The advantage of using the alpha test instead of alpha blending is that blending typically degrades the performance of fragment processing. With alpha testing fragments with zero alpha are rejected before they get to the color buffer. A disadvantage of alpha testing is that the edges will not be blended into the scene so the edges will not be properly antialiased.

 

The alpha component of a texture can be used in other ways, for example, to cut holes in polygons or to trim surfaces. An image of the trim region is stored in a texture map and when it is applied to the surface, alpha testing or blending can be used to reject the trimmed region. This method can be useful for trimming complex surfaces in scientific visualization applications.

 

I will assume that this description also applies to Enigma2.

 

This would seem to explain why AlphaTest is so much faster than AlphaBlend and why we are having issues seeing a difference.  Given that a poorer quality of antialiasing is likely to be hard to see in most aspects of the UI it would seem that using AlphaTest will be a worthwhile change even if the AlphaBlend code is fixed.

 

Anyone have any thoughts or comments?

 

Regards,

Ian.

[IanSav]

Hi WanWizard,

 

As you say OpenVIX has worked around it, can you point out the commits that do?

 

This is not an area of Enigma2 with which I work but I was told that this is simply a configuration option in "configure.ac".  Something to do with an option "FORCE_NO_BLENDING_ACCELERATION".  Apparently a number of boxes need the option set.

 

Regards,

Ian.

Edited by IanSav, 1 June 2020 - 16:06.

[WanWizard]

There seems to be a lot more going on that that: https://github.com/O...mpare/hisilicon

 

This is above my paygrade I'm afraid.

[IanSav]

Hi WanWizard,

 

The "Blend" to "Test" change that improves performance for some hardware has been committed to all repositories.

 

Here is the OpenPLi pull request: https://github.com/O...igma2/pull/2593

 

Regards,

Ian.

[IanSav]

Hi WanWizard,

 

This is above my paygrade I'm afraid.

 

Are you getting paid?  How can I get a cut?   

 

Regards,

Ian.

[Frenske]

He definitely forgot the quotationmarks left and right to the paygrade.

[betacentauri]

https://github.com/O...re.ac#L457-L461

 

I guess it has something to do with CONFIG_HISILICON_FB.

[Taapat]

@IanSav It seems to me that you are looking for answers to your questions in the wrong places.

What is true for a computer with unlimited resources for software decoding is not true for a satellite receiver that requires hardware decoding for fast operation.

I now can't quickly find the relevant commit, but as far as I can remember then AlphaBlend in enigma was created because it worked many times faster if it was supported by hadware.

 

However, if you check the configure.ac file, then you will see that this option is not supported by many receivers.

Unfortunately, receiver support is specified manually and therefore I assume that there are many receivers that do not support this option, but it is not specified in configure.ac or in the manufacturer's BSP and therefore the enigma is compiled with support that does not work and causes this behavior.

 

I agree that it is better not to use this option at this time because there is a bug.

Once the problem with AlphaBlend is resolved, then can try to use it.

[Taapat]

Sorry, I'm not right.

It seems that the use of blend has performance degradation: https://github.com/O...7982de7671734a7

And I assume it's ten times bigger if the hadware doesn't support it but the enigma doesn't know it and tries to use.

Edited by Taapat, 1 June 2020 - 17:48.

[littlesat]

Why not keep it blend an arrange that test is executed....???

[IanSav]

Hi Littlesat,

 

Why not keep it blend an arrange that test is executed....???

 

That may be what OpenViX has done.  I don't know the details.

 

In any case I am happy to use AlphaTest as that will always perform better with very little visual compromise.

 

Regards,

Ian.

[Huevos]

 

Hi WanWizard,

So the flashing issue is still there with the modified code?

Yes. Less obvious, but still visible.
 

Has OpenPLi looked at the Enigma2 build options that OpenViX use to get around the hardware / driver issues on HiSi boxes?

 

We don't do workarounds.

 

If there are manufacturer related issues, they have to be passed on to the manufacturer so the root cause can be fixed.

 

I don't understand this. If the manufacturer needs to fix, why are you supporting the hardware at all?

 

For us it is really simple, if a box doesn't work properly and the manufacturer doesn't fix it we don't introduce it to our line up.

 

Otherwise you are encouraging users to by broken hardware because that hardware is endorsed by PLi.

Edited by Huevos, 2 June 2020 - 10:00.

[WanWizard]

Afaik the manufacturers haven't been asked to fix anything yet, so we are not at that point. This issue was only raised yesterday.

 

Besides that, I've looked at TWOL's code, and (most of) it doesn't look like a workaround, Enigma has specific "ifdef"'s for other hardware platforms in its low-level code too.

 

As I wrote, I don't have the skillset to evaluate this code, so someone else needs to pick this up.

[ims]

Hi Ian,

 

I think, there could be changed MultiContentEntryPixmapAlphaBlend to MultiContentEntryPixmap for self.bg_l, self.bg_m, self.bg_r only. It would be good compromise for Hisil boxes.

On H7, vs1500 and e4hd is there VK almost without color rectangles - for my eyes it is not visibled well

 

M.

[IanSav]

Hi Ims,

 

The latest commit changed all the AlphaBlend definitions into AlphaTest definitions.  From what I understand of this change we get acceptable image overlay with a significantly reduced performance cost.  Are you advocating for further change?  (I didn't think it would be neccessary.)

 

Regards,

Ian.

[foxbob]

I have these lines in the log. Maybe somewhere it needs to be enabled.

 

[fb] double buffering available!

[gFBDC] resolution: 1280x720x32 stride=5120, 17100kB available for acceleration surfaces.