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

[Ilia Mirkin]

On Tue, Apr 18, 2017 at 5:26 PM, Nicolai Hähnle <nhaehnle at gmail.com> wrote:
> On 18.04.2017 21:49, Dave Airlie wrote:
>>
>> On 19 April 2017 at 05:30, Samuel Pitoiset <samuel.pitoiset at gmail.com>
>> wrote:
>>>
>>>
>>>
>>> On 04/18/2017 08:14 PM, Nicolai Hähnle wrote:
>>>>
>>>>
>>>> On 11.04.2017 18:48, Samuel Pitoiset wrote:
>>>>>
>>>>>
>>>>> Bindless sampler/image types are really different from the existing
>>>>> sampler/image types. They are considered 64-bit unsigned integers,
>>>>> they can be declared as temporary, shader inputs/outputs and are
>>>>> non-opaque types.
>>>>>
>>>>> For these reasons, it looks more convenient to introduce new
>>>>> internal base types to the GLSL compiler, called
>>>>> GLSL_TYPE_BINDLESS_SAMPLER and respectively GLSL_TYPE_BINDLESS_IMAGE.
>>>>
>>>>
>>>>
>>>> Sorry for taking so long to get to this series, but could you explain
>>>> the
>>>> rationale here a bit more?
>>>
>>>
>>>
>>> No worries, all feedbacks are always welcome, even late. :)
>>>
>>> So, the bindless sampler/image types introduced in ARB_bindless_texture
>>> are
>>> really different from the "standard" ones.
>>>
>>> They are considered as 64-bit unsigned integers (not 32-bit) and they are
>>> non-opaque types. The latter means they can be declared as temporary
>>> variables, as shader inputs/outputs, inside an interface block, etc.
>
>
> I don't know that that's the best way to think about it at the GLSL level.
> Mostly they're still opaque types, it's just that they can be explicitly
> converted to/from 64-bit ints (or uvec2... not having an interaction with
> GL_ARB_gpu_shader_int64 seems like a spec oversight).
>
>
>>> That said, the current sampler/image types are opaque (cf,
>>> glsl_type::is_opaque()) and it seemed quite impossible to change the
>>> glsl_type helpers to fit my needs.
>
>
> I see no is_opaque, maybe you mean contains_opaque? I agree that it's
> annoying that the restrictions expressed in ast_to_hir.cpp need to be
> modified. What other helpers are an issue?
>
>
>>> I tried many different solutions before figuring this one which seems
>>> better
>>> for some reasons:
>>>
>>> - easy to make bindless sampler/image types 64-bit unsigned int
>>> - easy to make bindless sampler/image types non-opaque
>>> - should avoid breakage because the base type is different
>>> - reduce the amount of changes in most places in the compiler
>
>
> I still don't see a *positive* justification for having the two different
> type families. Where do you actually need to *distinguish* between bindless
> and bound samplers in the compiler? The spec really doesn't make that
> distinction at all, except in one single place:
>
>    "These modifiers control whether default-block uniforms of
>     the corresponding types may have their values set via both
>     UniformHandle* and Uniform1i (bindless_sampler and
>     bindless_image) or only via Uniform1i (bound_sampler and
>     bound_image)."
>
> It seems like the only place to distinguish between the two is actually
> *outside* the compiler, in the uniforms API. And that's not even really have
> about the *type* of the variable; it's about the variable's memory layout.
> So it really behaves more similarly to the layout qualifiers (offset,
> alignment) you have in SSBO and UBO blocks rather than a property of a type.
>
> So as I see it, this code is messing with the type system for no good
> reason, and doing so in a fundamentally incoherent way that goes against all
> that is good and sane in compiler design. And I think it shows. Consider the
> following example shader:
>
>    layout (bound_sampler) uniform sampler2D tex_bound;
>    layout (bindless_sampler) uniform sampler2D tex_bindless;
>
>    vec4 f(sampler2D s, ...)
>    {
>       return texture2D(s, ...);
>    }
>
>    void main()
>    {
>       vec4 a = f(tex_bound, ...);
>       vec4 b = f(tex_bindless, ...);
>    }
>
> What's the type of the parameter `s' of function `f'? Will this compile, and
> if so, how?
>
> I guess the answer is: as is, the type is sampler2DBindless, which makes no
> sense whatsoever if the function f appears in a shader which *doesn't*
> enable ARB_bindless_texture (which is allowed!). But having the type of f
> depend on whether the extension is enabled makes even *less* sense.
>
> And does it compile? Well, I don't recall seeing code that would
> automatically do a type conversion -- you probably either end up generating
> an incorrect compiler error, or you end up with IR that assigns a value of
> bindless-type to a variable of non-bindless-type, or vice versa.
>
> None of this gives me confidence from a compiler design perspective. What's
> left are some vague implementation concerns that I don't understand yet,
> admittedly in part because I haven't tried to implement the alternative :-)
>
>
>>> At the Gallium level, the changes are really small. Basically, if the
>>> type
>>> is a bindless sampler, the ir_dereference variable is visited and it can
>>> be
>>> considered as PROGRAM_UNIFORM or PROGRAM_TEMPORARY (for shader
>>> inputs/outputs).
>>>
>>> Hopefully, you are convinced now. :)
>>
>>
>> When I did my initial implementation I went with merged types, I don't
>> think
>> it was a good idea either. Apart from making old non-bindless samples take
>> up
>> twice as much space, there's a lot or corner cases.
>
>
> Non-bindless samplers taking up twice as much space for the accounting of
> the default uniform block is something the spec explicitly calls out as
> being okay, though.
>
> Even if we decide that it isn't, it's really something that belongs into
> struct gl_uniform_storage during linking, because it's clearly a property
> of... uniform storage layout.
>
>
>> Unless we hit some problem in the future where they need to be the same
>> type I'd say this is preferable.
>
>
> I hope the code sample above is convincing evidence that using different
> types for the same thing just makes no sense.

That's one quirk of this approach - you have to create a combinatorial
explosion of user (and built-in, as you've seen) functions for each
sampler they accept. So if the function takes 5 sampler arguments, you
have to generate 32 copies. However I suspect such functions are rare.
Not sure if Samuel did this, but if he didn't, that's a necessary
component of this approach.

Having it all be one type definitely creates a lot of complications
which you're not seeing here as they're all avoided. Neither approach
is perfect, but I think at the cost of the extra types you end up with
a much more understandable compiler. Here are some odd cases to deal
with for the single type approach:

1. local variables: so doing something like

    sampler2D foo = bar;

is not legal without bindless. So imagine you have a function that
uses a sampler local variable,

    vec4 f(sampler2D s, ...)
    {
       sampler2D tex = foo ? s : t;
       return texture2D(tex, ...);
    }

This function can implicitly not be called with a bound sampler.
Expressing this in a way that the compiler can understand seems like
it couldn't be done without the types being different, or some deeper
analysis being done on the uses of the function's arguments. With the
type explosion approach, this magically works out since there's no
ambiguity about whether a particular sampler is bindless or not.

Furthermore, even without the separate function issue, the fact that
local variables are bindless-only creates a host of issues for
figuring out whether an expression is legal. Perhaps this is doable
with a deep scan of any expression assigning to a local variable. But
things with temporary expression like texture2D(foo ? s : t, ...)
could still be annoying. (All solvable, I'm sure.)

2. uniform storage: this is a weaker argument, but I feel like it
needs to be said anyways -- the uniform layout code is subtle and
quick to anger. Even if it is modified appropriately, seems unlikely
for there to be an easily achievable way that non-bindless GPUs are
allowed to not have to waste uniform space. IIRC these uniforms are
allocated real uniform storage (despite never being used), which will
use up precious uniform space for DX9 and earlier GPUs. However if
this were the only thing, it'd be too weak. More of a "pile on top"
kind of thing.

FWIW I'm in favor of the separate type approach, assuming that we all
agree that the user function explosion is acceptable. IMHO it is
though. [And I had been discussing this approach and the alternative
of doing a single type with Samuel for a while before this patchset
was sent out.]

Cheers,

  -ilia