[Mesa-dev] [PATCH] meta: Use result of texture coordinate clamping operation

[Markus Wick]

Am 2015-09-10 04:00, schrieb Anuj Phogat:
> On Wed, Sep 9, 2015 at 6:19 PM, Ian Romanick <idr at freedesktop.org> 
> wrote:
>> On 09/09/2015 04:55 PM, Anuj Phogat wrote:
>>> On Wed, Sep 9, 2015 at 11:30 AM, Ian Romanick <idr at freedesktop.org>
> wrote:
>>>> I'm pretty sure our implementation of this extension is complete
>>>> rubbish.  I have attached an image from the blit-scaled test.  This
>>>> result cannot be useful to anyone.
>>>> 
>>> Paul Berry and I agreed to follow the "NVIDIA Implementation Details"
>>> given in the extension spec. i965 driver currently use bilinear 
>>> filtering
>>> for both SCALED_RESOLVE_FASTEST and SCALED_RESOLVE_NICEST.
>>> Do you think the implementation doesn't match what is suggested in
>>> the spec? I'll be happy to work on improving the implementation if 
>>> you
>>> have any specific comments.
>>> 
>>> Extension spec mentions the potential loss of quality when using a
>>> single pass scaled resolve blit. Also the quality of image generated
>>> depends on sample count and scaling factor. I vaguely remember
>>> Paul mentioned the scaling factor of ~1.5, beyond which single pass
>>> scaled resolve blit quickly becomes useless. I think the attached
>>> image is using a scaling factor of 2.4.
>> 
>> Yes, I think so.  It was the last image produced by the test.
>> 
>>> I'm attaching an old image generated by me doing the quality
>>> comparison with Nvidia's proprietary driver implementation. I posted
>>> the image on mesa-dev before enabling the extension on i965. Please
>>> see the images under "New implementation" for the comparison.
>> 
>> Comparing our images with NVIDIA's image, I don't see very much
>> difference.  That's both reassuring and depressing.  Looking at all of
>> those images, I can't imagine why anyone would bother using this
>> functionality.  I doubt there's any scenario where the little "spikes"
>> that grow of the sides of the triangles is acceptable.  Oh well.
> 
> Markus Wick (IRC name: degasus) shared these dolphin emulator
> images (rendered using i965 drivers) with me on dri-devel. They show
> the effect of using the extension over resolve and liner scaling. I 
> think
> most of the images look crisper with little loss of smooth edges.
> https://www.dropbox.com/sh/vmt7tolzd6qlnpo/AABlGkXZ4W6_ucFk0Gk3tfNpa?dl=0

But my conclusion was similiar. Bilinear filtering is pretty useless for 
downscaling. But I see a big usecase in upscaling on mobile devices with 
very high DPI. So the shading can be done at a lower resolution, but the 
framebuffer resolution matches the display.

For dolphin, we don't use glBlitFramebuffer for scaling any more, so 
testing this stuff is harder now.