[Mesa-dev] [PATCH 4/8] glsl: Add new builtins required by GL_ARB_sample_shading

[Kenneth Graunke]

On 10/16/2013 10:40 AM, Ian Romanick wrote:
> On 10/15/2013 07:50 PM, Kenneth Graunke wrote:
[snip]
>> Right, but gl_SampleMask is not an unsized array.  It's statically sized
>> based on your implementation's maximum supported number of color samples
>> (ceil(samples/32)), and this is totally known at compile time...
> 
> That doesn't match what the spec says.  At least the GLSL 4.00 spec:
> 
>     "In the fragment language, built-in variables are intrinsically
>     declared as:
> 
>         in vec4 gl_FragCoord;
>         in bool gl_FrontFacing;
>         in float gl_ClipDistance[];
>         in vec2 gl_PointCoord;
>         in int gl_PrimitiveID;
>         in int gl_SampleID;
>         in vec2 gl_SamplePosition;
>         out vec4 gl_FragColor;
>         out vec4 gl_FragData[gl_MaxDrawBuffers];
>         out float gl_FragDepth;
>         out int gl_SampleMask[];"
> 
> Note that both gl_ClipDistance and gl_SampleMask are unsized, and both
> have a maximum size based on implementation limits.  Later the GLSL 4.00
> spec says:
> 
>     "This array must be sized in the fragment shader either implicitly
>     or explicitly to be the same size as the implementation-dependent
>     maximum sample-mask (as an array of 32bit elements), determined by
>     the maximum number of samples."
> 
> The maximum size of the array is ceil(max_samples/32), but initially it
> is unsized.

You've got to be kidding me.  This is utter and complete garbage.
"We know exactly how big the array is, at compile time, but we're going
to make you, the technical artist, tell us anyway.  By the way, you have
to guess the exact number we're thinking."

So, about that "explicitly the same size" text...what if my shader sets
it to some /other/ size?  What happens?

* Larger than ceil(float(gl_NumSamples)/32.0):

A compile error?  Truncated to the right size?  Probably the same
behavior as ClipDistance.  That doesn't seem well specified either.

* Smaller than ceil(float(gl_NumSamples)/32.0):

>From the GLSL 4.40 specification,
"If the fragment shader statically assigns a value to gl_SampleMask,
 the sample mask will be undefined for any array elements of any
 fragment shader invocations that fail to assign a value."

The hardware is expecting a certain number of samples, so ultimately,
the fact that you don't have them means they're undefined.  Just like
the above case of forgetting to assign them.  Do we pretend the array is
the right size, ignoring their request and treating any unwritten
elements as unsized?  Or do we reject it as a compile error?

...

Also, considering that the Radeon 8970 (arguably AMD's best card today)
only does 24x MSAA, and it looks like nVidia's GTX Titan doesn't do more
than 32x (though it's harder to tell), this array will always be exactly
size 1 for every implementation in existence.

<sarcasm>
I'm sure applications are fully on board to explicitly declare this and
cope with the consequences of it not being a single integer.
</sarcasm>

That said, at this point, it's about 4 years too late to change the
idiocy of the spec, so we're kind of stuck.  We probably ought to
investigate what other vendors do.

--Ken