[Mesa-dev] [PATCH v5 25/70] glsl: Add std430 related member functions to glsl_type class

[Samuel Iglesias Gonsálvez]

On 16/09/15 09:11, Ilia Mirkin wrote:
> On Wed, Sep 16, 2015 at 1:14 AM, Samuel Iglesias Gonsálvez
> <siglesias at igalia.com> wrote:
>>
>>
>> On 15/09/15 21:03, Jordan Justen wrote:
>>> On 2015-09-10 22:48:55, Samuel Iglesias Gonsálvez wrote:
>>>> On 10/09/15 20:13, Jordan Justen wrote:
>>>>> On 2015-09-10 06:35:41, Iago Toral Quiroga wrote:
>>>>>> From: Samuel Iglesias Gonsalvez <siglesias at igalia.com>
>>>>>>
>>>>>> They are used to calculate size, base alignment and array stride values
>>>>>> for a glsl_type following std430 rules.
>>>>>>
>>>>>> Signed-off-by: Samuel Iglesias Gonsalvez <siglesias at igalia.com>
>>>>>> ---
>>>>>>  src/glsl/glsl_types.cpp | 209 ++++++++++++++++++++++++++++++++++++++++++++++++
>>>>>>  src/glsl/glsl_types.h   |  19 +++++
>>>>>>  2 files changed, 228 insertions(+)
>>>>>>
>>>>>> diff --git a/src/glsl/glsl_types.cpp b/src/glsl/glsl_types.cpp
>>>>>> index 755618a..d97991a 100644
>>>>>> --- a/src/glsl/glsl_types.cpp
>>>>>> +++ b/src/glsl/glsl_types.cpp
>>>>>> @@ -1357,6 +1357,215 @@ glsl_type::std140_size(bool row_major) const
>>>>>>     return -1;
>>>>>>  }
>>>>>>
>>>>>> +unsigned
>>>>>> +glsl_type::std430_base_alignment(bool row_major) const
>>>>>> +{
>>>>>> +
>>>>>> +   unsigned N = is_double() ? 8 : 4;
>>>>>> +
>>>>>> +   /* (1) If the member is a scalar consuming <N> basic machine units, the
>>>>>> +    *     base alignment is <N>.
>>>>>> +    *
>>>>>> +    * (2) If the member is a two- or four-component vector with components
>>>>>> +    *     consuming <N> basic machine units, the base alignment is 2<N> or
>>>>>> +    *     4<N>, respectively.
>>>>>> +    *
>>>>>> +    * (3) If the member is a three-component vector with components consuming
>>>>>> +    *     <N> basic machine units, the base alignment is 4<N>.
>>>>>> +    */
>>>>>> +   if (this->is_scalar() || this->is_vector()) {
>>>>>> +      switch (this->vector_elements) {
>>>>>> +      case 1:
>>>>>> +         return N;
>>>>>> +      case 2:
>>>>>> +         return 2 * N;
>>>>>> +      case 3:
>>>>>> +      case 4:
>>>>>> +         return 4 * N;
>>>>>> +      }
>>>>>> +   }
>>>>>> +
>>>>>> +   /* OpenGL 4.30 spec, section 7.6.2.2 "Standard Uniform Block Layout":
>>>>>> +    *
>>>>>> +    * "When using the "std430" storage layout, shader storage
>>>>>> +    * blocks will be laid out in buffer storage identically to uniform and
>>>>>> +    * shader storage blocks using the "std140" layout, except that the base
>>>>>> +    * alignment of arrays of scalars and vectors in rule (4) and of structures
>>>>>
>>>>> Looking at the 4.3 spec (and 4.5), it actually adds "and stride"
>>>>> following "base alignment". The extension spec *does not* have the
>>>>> "and stride" text.
>>>>>
>>>>
>>>> OK. If you agree, I will keep OpenGL 4.3 (and later) spec wording in all
>>>> the places where this snippet is pasted.
>>>>
>>>>> This seems to be an inconsistency between the extension spec and the
>>>>> actual spec, but the OpenGL spec form would produce more tightly
>>>>> packed arrays.
>>>>>
>>>>> Maybe we want to confirm what another implementation does?
>>>>
>>>> Both NVIDIA and ATI proprietary drivers don't round up the stride of
>>>> arrays of vectors to a multiple of a vec4 size, i.e., they are following
>>>> the OpenGL spec. For example: for an array of vec2, they are returning
>>>> an stride value of 8, not 16 as in std140.
>>>
>>> Well, my concern was that the 'and stride' part might mean that vec3
>>> array stride should be 12 rather than 16. But, I tested NVidia, and
>>> they seem to use a stride of 16 for a vec3 array. So, I think your
>>> interpretation is correct.
>>>
>>> I still say we could still use an update to idr's ubo-lolz branch to
>>> handle ssbo and std430, but this would also involve extending shader
>>> runner to better support ssbo.
>>>
>>
>> I have already done that work. I have a ubo-lolz modified branch [0]
>> with an initial support of SSBOs and std430.
>>
>> About ssbo support for shader_runner, I have sent a couple of patches to
>> piglit [1] and I plan to send a new version of them today with a generic
>> approach (so it is not only for SSBOs but for other interface types
>> defined in ARB_program_interface_query extension).
>>
>> FWIW, I executed [0] with no errors during 15 minutes.
> 
> As way of validation, have you tried running your modified script
> against any other drivers? They may well have bugs in them as well,
> but it should be possible to determine if the bug is in the script or
> the other impl, should they not match up.
> 
>   -ilia
> 

I tested it on NVIDIA proprietary driver version 352.21. It has an issue
when we query shader storage block members when they are arrays of
structs and the index is different than zero -> it doesn't find them as
active. For example:

struct B {
	vec4 a[2];
}

layout(std430) buffer Block {
	B[2] s;
        vec3 v;
};

NVIDIA marked v, s[0].a[0] and s[0].a[1] as active but s[1].a[0] and
s[0].a[1] as inactive even when they are referenced in main().

I have not checked yet what ARB_program_interface_query spec says
regarding to that to see if it is a bug in our driver or in NVIDIA's.

However, I modified by hand the failed autogenerated tests removing the
queries to those "inactive" variables. As we are interested in offsets,
arrays strides and so on, polling index zero of all the top level array
of struct members is enough. With that modifications, those tests pass.

I plan to try to adapt my script to NVIDIA behavior so I can validate it
properly. However, these preliminary results indicate that I don't have
made any big mistake.

Sam