[Mesa-dev] [RFC PATCH 03/26] glsl: introduce new base types for bindless samplers/images

[Nicolai Hähnle]

On 19.04.2017 11:09, Nicolai Hähnle wrote:
>>> At least I haven't seen anything in the spec to contradict that.
>>>
>>> Side note: As far as I can tell, you're even allowed to do:
>>>
>>>    layout (bound_sampler) uniform sampler2D tex;
>>>
>>>    ...
>>>
>>>    uvec2 handle = uvec2(tex);
>>
>> I think this is correct as well, at least it compiles with NV. However,
>> it makes just no sense to me. The spec says:
>>
>> "The error INVALID_OPERATION is generated by UniformHandleui64{v}ARB if
>> the sampler or image uniform being updated has the "bound_sampler" or
>> "bound_image" layout qualifier.""
>>
>> So, you can only assign a 32-bit uniform here, converting to a pair of
>> 32-bit will make the .y component 0.
>
> The way I understand it, converting to uvec2 should give a pointer to
> the descriptor (at least for the GCN backend).
>
> So bound_samplers are implemented as simply old-style samplers living in
> the good old descriptor table. When they are converted to uvec2, we need
> to emit code that returns a pointer into that descriptor table. That's a
> fully 64-bit value.

By the way, this is also how Ilia's example would be implemented. When 
inlined, it could become something like:

   layout (bound_sampler) uniform sampler2D u_tex;
   in sampler2D in_tex;

   void main()
   {
     ...
     sampler2D tmp = foo ? u_tex : in_tex;
     vec4 color = texture2D(tmp, ...);
   }

This should compile. At the GLSL level, the type is always sampler2D. At 
the TGSI and LLVM IR level, u_tex is converted to a 64-bit value at the 
point of the ?: operation. To keep the craziness low in TGSI, there 
should probably be some S2U64 instruction for explicitly converting 
old-style samplers/images to 64-bit values, even though most cases of 
explicit conversion between sampler types and uvec2 are actually no-ops 
in TGSI.

Cheers,
Nicolai