[Mesa-dev] [PATCH 1/8] Import current draft of EXT_shader_samples_identical spec

[Nicolai Hähnle]

On 19.11.2015 00:46, Ian Romanick wrote:
> From: Ian Romanick <ian.d.romanick at intel.com>
>
> Signed-off-by: Ian Romanick <ian.d.romanick at intel.com>
> Cc: "Chris Forbes" <chrisf at ijw.co.nz>
> ---
>   docs/specs/EXT_shader_samples_identical.txt | 174 ++++++++++++++++++++++++++++
>   1 file changed, 174 insertions(+)
>   create mode 100644 docs/specs/EXT_shader_samples_identical.txt
>
> diff --git a/docs/specs/EXT_shader_samples_identical.txt b/docs/specs/EXT_shader_samples_identical.txt
> new file mode 100644
> index 0000000..bae6c73
> --- /dev/null
> +++ b/docs/specs/EXT_shader_samples_identical.txt
> @@ -0,0 +1,174 @@
> +Name
> +
> +    EXT_shader_samples_identical
> +
> +Name Strings
> +
> +    GL_EXT_shader_samples_identical
> +
> +Contact
> +
> +    Ian Romanick, Intel (ian.d.romanick 'at' intel.com)
> +
> +Contributors
> +
> +    Chris Forbes, Mesa
> +    Magnus Wendt, Intel
> +    Graham Sellers, AMD
> +
> +Status
> +
> +    XXX - Not complete yet.
> +
> +Version
> +
> +    Last Modified Date: November 18, 2015
> +    Revision: 5
> +
> +Number
> +
> +    TBD
> +
> +Dependencies
> +
> +    OpenGL 3.2, or OpenGL ES 3.1, or ARB_texture_multisample is required.
> +
> +    This extension is written against the OpenGL 4.5 (Core Profile)
> +    Specification
> +
> +Overview
> +
> +    Multisampled antialiasing has become a common method for improving the
> +    quality of rendered images.  Multisampling differs from supersampling in
> +    that the color of a primitive that covers all or part of a pixel is
> +    resolved once, regardless of the number of samples covered.  If a large
> +    polygon is rendered, the colors of all samples in each interior pixel will
> +    be the same.  This suggests a simple compression scheme that can reduce
> +    the necessary memory bandwidth requirements.  In one such scheme, each
> +    sample is stored in a separate slice of the multisample surface.  An
> +    additional multisample control surface (MCS) contains a mapping from pixel
> +    samples to slices.
> +
> +    If all the values stored in the MCS for a particular pixel are the same,
> +    then all the samples have the same value.  Applications can take advantage
> +    of this information to reduce the bandwidth of reading multisample
> +    textures.  A custom multisample resolve filter could optimize resolving
> +    pixels where every sample is identical by reading the color once.
> +
> +    color = texelFetch(sampler, coordinate, 0);
> +    if (!textureSamplesAllIdenticalEXT(sampler, coordinate)) {

textureSamplesIdenticalEXT

> +        for (int i = 1; i < MAX_SAMPLES; i++) {
> +            vec4 c = texelFetch(sampler, coordinate, i);
> +
> +            //... accumulate c into color
> +
> +        }
> +    }
> +
> +New Procedures and Functions
> +
> +    None.
> +
> +New Tokens
> +
> +    None.
> +
> +Additions to the OpenGL 4.5 (Core Profile) Specification
> +
> +    None.
> +
> +Modifications to The OpenGL Shading Language Specification, Version 4.50.5
> +
> +    Including the following line in a shader can be used to control the
> +    language features described in this extension:
> +
> +        #extension GL_EXT_shader_samples_identical
> +
> +    A new preprocessor #define is added to the OpenGL Shading Language:
> +
> +        #define GL_EXT_shader_samples_identical
> +
> +    Add to the table in section 8.7 "Texture Lookup Functions"
> +
> +    Syntax:
> +
> +        bool textureSamplesIdenticalEXT(gsampler2DMS sampler, ivec2 coord)
> +
> +        bool textureSamplesIdenticalEXT(gsampler2DMSArray sampler,
> +                                        ivec3 coord)
> +
> +    Description:
> +
> +        Returns true if it can be determined that all samples within the texel
> +        of the multisample texture bound to <sampler> at <coord> contain the
> +        same values or false if this cannot be determined."
> +
> +Additions to the AGL/EGL/GLX/WGL Specifications
> +
> +    None
> +
> +Errors
> +
> +    None
> +
> +New State
> +
> +    None
> +
> +New Implementation Dependent State
> +
> +    None
> +
> +Issues
> +
> +    1) What should the new functions be called?
> +
> +    RESOLVED: textureSamplesIdenticalEXT.  Initially
> +    textureAllSamplesIdenticalEXT was considered, but
> +    textureSamplesIdenticalEXT is more similar to the existing textureSamples
> +    function.
> +
> +    2) It seems like applications could implement additional optimization if
> +       they were provided with raw MCS data.  Should this extension also
> +       provide that data?
> +
> +    There are a number of challenges in providing raw MCS data.  The biggest
> +    problem being that the amount of MCS data depends on the number of
> +    samples, and that is not known at compile time.  Additionally, without new
> +    texelFetch functions, applications would have difficulty utilizing the
> +    information.
> +
> +    Another option is to have a function that returns an array of tuples of
> +    sample number and count.  This also has difficulties with the maximum
> +    array size not being known at compile time.
> +
> +    RESOLVED: Do not expose raw MCS data in this extension.
> +
> +    3) Should this extension also extend SPIR-V?
> +
> +    RESOLVED: Yes, but this has not yet been written.
> +
> +    4) Is it possible for textureSamplesAllIdenticalEXT to report false negatives?

Same here: textureSamplesIdenticalEXT

Cheers,
Nicolai

> +    RESOLVED: Yes.  It is possible that the underlying hardware may not detect
> +    that separate writes of the same color to different samples of a pixel are
> +    the same.  The shader function is at the whim of the underlying hardware
> +    implementation.  It is also possible that a compressed multisample surface
> +    is not used.  In that case the function will likely always return false.
> +
> +Revision History
> +
> +    Rev  Date        Author    Changes
> +    ---  ----------  --------  ---------------------------------------------
> +      1  2014/08/20  cforbes   Initial version
> +      2  2015/10/23  idr       Change from MESA to EXT.  Rebase on OpenGL 4.5,
> +                               and add dependency on OpenGL ES 3.1.  Initial
> +                               draft of overview section and issues 1 through
> +                               3.
> +      3  2015/10/27  idr       Typo fixes.
> +      4  2015/11/10  idr       Rename extension from EXT_shader_multisample_compression
> +                               to EXT_shader_samples_identical.
> +                               Add issue #4.
> +      5  2015/11/18  idr       Fix some typos spotted by gsellers.  Change the
> +                               name of the name of the function to
> +                               textureSamplesIdenticalEXT.
>