[Piglit] [PATCH] arb_texture_multisample: stress test of very large textures (v3)

[Brian Paul]

On 12/26/2017 11:57 PM, Arthur Huillet wrote:
> On Sun, 24 Dec 2017 15:41:38 -0700
> Brian Paul <brianp at vmware.com> wrote:
>
>> + * Some drivers/GPUs may fail this test.  NVIDIA, for example, appears to
>> + * only store the MSAA coverage info, not the sample colors, for samples
>> + * beyond the 8th sample.  We may tune the way this test operates over time
>> + * to be more useful.  Maybe we should test all MSAA samples/pixels (2x,
>> + * 4x, 8x, etc).
>
> If you believe the NVIDIA driver is wrong, please let us know what needs to be fixed and we'll look into it.

Just to be clear, I'm referring to the NVIDIA proprietary driver, not 
the gallium driver.


> If on the other hand you can't prove that the NVIDIA driver is wrong,
> then please don't write a test that "fails" when you have no evidence
> that it isn't the test that is incorrect. Otherwise this tends to
> claim "this is required OpenGL driver behavior" and confuse everyone.

I haven't done an exhaustive search of the specifications, but so far I 
haven't found any language that permits "tricks" like NVIDIAs CSAA 
technique for "standard" MSAA textures.  That is, the spec seems to 
expect that if a texture stores N samples per texel, then there are N 
distinct colors stored per texel.  It's a different story with the 
GL_NV_framebuffer_multisample_coverage extension, though.

If you have more insight into this, I'm all ears.

FWIW, from Googling it sounds like more recent NVIDIA GPUs (Kepler and 
later) no longer advertise 16x and 32x MSAA so those devices would 
probably pass this test.

-Brian