Name

    EXT_shader_image_load_store

Name Strings

    EXT_shader_image_load_store

Contact

    Jeff Bolz, NVIDIA Corporation (jbolz 'at' nvidia.com)
    Pat Brown, NVIDIA Corporation (pbrown 'at' nvidia.com)

Contributors

    Barthold Lichtenbelt, NVIDIA
    Bill Licea-Kane, AMD
    Eric Werness, NVIDIA
    Graham Sellers, AMD
    Greg Roth, NVIDIA
    Nick Haemel, AMD
    Pierre Boudier, AMD
    Piers Daniell, NVIDIA

Status

    Shipping.

Version

    Last Modified Date:         10/16/2013
    NVIDIA Revision:            7

Number

    386

Dependencies

    This extension is written against the OpenGL 3.2 specification
    (Compatibility Profile).

    This extension is written against version 1.50 (revision 09) of the OpenGL
    Shading Language Specification.

    OpenGL 3.0 and GLSL 1.30 are required.

    This extension interacts trivially with OpenGL 3.2 (Core Profile).

    This extension interacts trivially with OpenGL 3.1,
    ARB_uniform_buffer_object, and EXT_bindable_uniform.

    This extension interacts trivially with ARB_draw_indirect.

    This extension interacts trivially with NV_vertex_buffer_unified_memory.

    This extension interacts trivially with OpenGL 3.2 and
    ARB_texture_multisample.

    This extension interacts trivially with OpenGL 4.0 and ARB_sample_shading.

    This extension interacts trivially with OpenGL 4.0 and
    ARB_texture_cube_map_array.

    This extension interacts trivially with OpenGL 3.3 and
    ARB_texture_rgb10_a2ui.

    This extension interacts trivially with NV_shader_buffer_load.

    This extension interacts trivially with OpenGL 4.0, ARB_gpu_shader5, and
    NV_gpu_shader5.

    This extension interacts trivially with OpenGL 4.0 and
    ARB_tessellation_shader.

    This extension interacts trivially with EXT_depth_bounds_test.

    This extension interacts with EXT_separate_shader_objects.

    This extension interacts with NV_gpu_program5.

Overview

    This extension provides GLSL built-in functions allowing shaders to load
    from, store to, and perform atomic read-modify-write operations to a
    single level of a texture object from any shader stage.  These built-in
    functions are named imageLoad(), imageStore(), and imageAtomic*(),
    respectively, and accept integer texel coordinates to identify the texel
    accessed.  The extension adds the notion of "image units" to the OpenGL
    API, to which texture levels are bound for access by the GLSL built-in
    functions.  To allow shaders to specify the image unit to access, GLSL
    provides a new set of data types ("image*") similar to samplers.  Each
    image variable is assigned an integer value to identify an image unit to
    access, which is specified using Uniform*() APIs in a manner similar to
    samplers.  For implementations supporting the NV_gpu_program5 extensions,
    assembly language instructions to perform image loads, stores, and atomics
    are also provided.

    This extension also provides the capability to explicitly enable "early"
    per-fragment tests, where operations like depth and stencil testing are
    performed prior to fragment shader execution.  In unextended OpenGL,
    fragment shaders never have any side effects and implementations can
    sometimes perform per-fragment tests and discard some fragments prior to
    executing the fragment shader.  Since this extension allows fragment
    shaders to write to texture and buffer object memory using the built-in
    image functions, such optimizations could lead to non-deterministic
    results.  To avoid this, implementations supporting this extension may not
    perform such optimizations on shaders having such side effects.  However,
    enabling early per-fragment tests guarantees that such tests will be
    performed prior to fragment shader execution, and ensures that image
    stores and atomics will not be performed by fragment shader invocations
    where these per-fragment tests fail.

    Finally, this extension provides both a GLSL built-in function and an
    OpenGL API function allowing applications some control over the ordering
    of image loads, stores, and atomics relative to other OpenGL pipeline
    operations accessing the same memory.  Because the extension provides the
    ability to perform random accesses to texture or buffer object memory,
    such accesses are not easily tracked by the OpenGL driver.  To avoid the
    need for heavy-handed synchronization at the driver level, this extension
    requires manual synchronization.  The MemoryBarrierEXT() OpenGL API
    function allows applications to specify a bitfield indicating the set of
    OpenGL API operations to synchronize relative to shader memory access.
    The memoryBarrier() GLSL built-in function provides a synchronization
    point within a given shader invocation to ensure that all memory accesses
    performed prior to the synchronization point complete prior to any started
    after the synchronization point.

New Procedures and Functions

    void BindImageTextureEXT(uint index, uint texture, int level,
                             boolean layered, int layer, enum access,
                             int format);

    void MemoryBarrierEXT(bitfield barriers);

New Tokens

    Accepted by the <pname> parameter of GetBooleanv, GetIntegerv,
    GetFloatv, and GetDoublev: 

        MAX_IMAGE_UNITS_EXT                             0x8F38
        MAX_COMBINED_IMAGE_UNITS_AND_FRAGMENT_OUTPUTS_EXT 0x8F39
        MAX_IMAGE_SAMPLES_EXT                           0x906D

    Accepted by the <target> parameter of GetIntegeri_v and GetBooleani_v:

        IMAGE_BINDING_NAME_EXT                          0x8F3A
        IMAGE_BINDING_LEVEL_EXT                         0x8F3B
        IMAGE_BINDING_LAYERED_EXT                       0x8F3C
        IMAGE_BINDING_LAYER_EXT                         0x8F3D
        IMAGE_BINDING_ACCESS_EXT                        0x8F3E
        IMAGE_BINDING_FORMAT_EXT                        0x906E

    Accepted by the <barriers> parameter of MemoryBarrierEXT:

        VERTEX_ATTRIB_ARRAY_BARRIER_BIT_EXT             0x00000001
        ELEMENT_ARRAY_BARRIER_BIT_EXT                   0x00000002
        UNIFORM_BARRIER_BIT_EXT                         0x00000004
        TEXTURE_FETCH_BARRIER_BIT_EXT                   0x00000008
        SHADER_IMAGE_ACCESS_BARRIER_BIT_EXT             0x00000020
        COMMAND_BARRIER_BIT_EXT                         0x00000040
        PIXEL_BUFFER_BARRIER_BIT_EXT                    0x00000080
        TEXTURE_UPDATE_BARRIER_BIT_EXT                  0x00000100
        BUFFER_UPDATE_BARRIER_BIT_EXT                   0x00000200
        FRAMEBUFFER_BARRIER_BIT_EXT                     0x00000400
        TRANSFORM_FEEDBACK_BARRIER_BIT_EXT              0x00000800
        ATOMIC_COUNTER_BARRIER_BIT_EXT                  0x00001000
        ALL_BARRIER_BITS_EXT                            0xFFFFFFFF

    Returned by the <type> parameter of GetActiveUniform:

        IMAGE_1D_EXT                                    0x904C
        IMAGE_2D_EXT                                    0x904D
        IMAGE_3D_EXT                                    0x904E
        IMAGE_2D_RECT_EXT                               0x904F
        IMAGE_CUBE_EXT                                  0x9050
        IMAGE_BUFFER_EXT                                0x9051
        IMAGE_1D_ARRAY_EXT                              0x9052
        IMAGE_2D_ARRAY_EXT                              0x9053
        IMAGE_CUBE_MAP_ARRAY_EXT                        0x9054
        IMAGE_2D_MULTISAMPLE_EXT                        0x9055
        IMAGE_2D_MULTISAMPLE_ARRAY_EXT                  0x9056
        INT_IMAGE_1D_EXT                                0x9057
        INT_IMAGE_2D_EXT                                0x9058
        INT_IMAGE_3D_EXT                                0x9059
        INT_IMAGE_2D_RECT_EXT                           0x905A
        INT_IMAGE_CUBE_EXT                              0x905B
        INT_IMAGE_BUFFER_EXT                            0x905C
        INT_IMAGE_1D_ARRAY_EXT                          0x905D
        INT_IMAGE_2D_ARRAY_EXT                          0x905E
        INT_IMAGE_CUBE_MAP_ARRAY_EXT                    0x905F
        INT_IMAGE_2D_MULTISAMPLE_EXT                    0x9060
        INT_IMAGE_2D_MULTISAMPLE_ARRAY_EXT              0x9061
        UNSIGNED_INT_IMAGE_1D_EXT                       0x9062
        UNSIGNED_INT_IMAGE_2D_EXT                       0x9063
        UNSIGNED_INT_IMAGE_3D_EXT                       0x9064
        UNSIGNED_INT_IMAGE_2D_RECT_EXT                  0x9065
        UNSIGNED_INT_IMAGE_CUBE_EXT                     0x9066
        UNSIGNED_INT_IMAGE_BUFFER_EXT                   0x9067
        UNSIGNED_INT_IMAGE_1D_ARRAY_EXT                 0x9068
        UNSIGNED_INT_IMAGE_2D_ARRAY_EXT                 0x9069
        UNSIGNED_INT_IMAGE_CUBE_MAP_ARRAY_EXT           0x906A
        UNSIGNED_INT_IMAGE_2D_MULTISAMPLE_EXT           0x906B
        UNSIGNED_INT_IMAGE_2D_MULTISAMPLE_ARRAY_EXT     0x906C


Additions to Chapter 2 of the OpenGL 3.2 (Compatibility Profile) Specification
(Rasterization)

    (Add new types to table 2.13, pp. 96-98)

      Type Name                                    Keyword
      ------------------------------               -------------------------
      IMAGE_1D_EXT                                 image1D
      IMAGE_2D_EXT                                 image2D
      IMAGE_3D_EXT                                 image3D
      IMAGE_2D_RECT_EXT                            image2DRect
      IMAGE_CUBE_EXT                               imageCube
      IMAGE_BUFFER_EXT                             imageBuffer
      IMAGE_1D_ARRAY_EXT                           image1DArray
      IMAGE_2D_ARRAY_EXT                           image2DArray
      IMAGE_CUBE_MAP_ARRAY_EXT                     imageCubeArray
      IMAGE_2D_MULTISAMPLE_EXT                     image2DMS
      IMAGE_2D_MULTISAMPLE_ARRAY_EXT               image2DMSArray
      INT_IMAGE_1D_EXT                             iimage1D
      INT_IMAGE_2D_EXT                             iimage2D
      INT_IMAGE_3D_EXT                             iimage3D
      INT_IMAGE_2D_RECT_EXT                        iimage2DRect
      INT_IMAGE_CUBE_EXT                           iimageCube
      INT_IMAGE_BUFFER_EXT                         iimageBuffer
      INT_IMAGE_1D_ARRAY_EXT                       iimage1DArray
      INT_IMAGE_2D_ARRAY_EXT                       iimage2DArray
      INT_IMAGE_CUBE_MAP_ARRAY_EXT                 iimageCubeArray
      INT_IMAGE_2D_MULTISAMPLE_EXT                 iimage2DMS
      INT_IMAGE_2D_MULTISAMPLE_ARRAY_EXT           iimage2DMSArray
      UNSIGNED_INT_IMAGE_1D_EXT                    uimage1D
      UNSIGNED_INT_IMAGE_2D_EXT                    uimage2D
      UNSIGNED_INT_IMAGE_3D_EXT                    uimage3D
      UNSIGNED_INT_IMAGE_2D_RECT_EXT               uimage2DRect
      UNSIGNED_INT_IMAGE_CUBE_EXT                  uimageCube
      UNSIGNED_INT_IMAGE_BUFFER_EXT                uimageBuffer
      UNSIGNED_INT_IMAGE_1D_ARRAY_EXT              uimage1DArray
      UNSIGNED_INT_IMAGE_2D_ARRAY_EXT              uimage2DArray
      UNSIGNED_INT_IMAGE_CUBE_MAP_ARRAY_EXT        uimageCubeArray
      UNSIGNED_INT_IMAGE_2D_MULTISAMPLE_EXT        uimage2DMS
      UNSIGNED_INT_IMAGE_2D_MULTISAMPLE_ARRAY_EXT  uimage2DMSArray


    (Add a new subsection after Section 2.14.5, Samplers, p. 106)

    Section 2.14.X, Images

    Images are special uniforms used in the OpenGL Shading Language to
    identify a level of a texture to be read or written using image load,
    store, and atomic built-in functions in the manner described in Section
    3.9.X.  The value of an image uniform is an integer specifying the image
    unit accessed.  Image units are numbered beginning at zero, and there is
    an implementation-dependent number of available image units
    (MAX_IMAGE_UNITS_EXT).  The error INVALID_VALUE is generated if a
    Uniform1i{v} call is used to set an image uniform to a value less than
    zero or greater than or equal to MAX_IMAGE_UNITS_EXT.  Note that image
    units used for image variables are independent of the texture image
    units used for sampler variables; the number of units provided by the
    implementation may differ.  Textures are bound independently and
    separately to image and texture image units.

    The type of an image variable must match the texture target of the image
    currently bound to the image unit, otherwise the result of the load/
    store/atomic operation is undefined (see Section 4.1.X of the OpenGL 
    Shading Language specification for more detail).

    The location of an image variable needs to be queried with
    GetUniformLocation, just like any uniform variable.  Image values need to
    be set by calling Uniform1i{v}.  Loading image variables with any of the
    other Uniform entry point is not allowed and will result in an
    INVALID_OPERATION error.

    Unlike samplers, there is no limit on the number of active image variables
    that may be used by a program or by any particular shader.  However, given
    that there is an implementation-dependent limit on the number of unique
    image units, the actual number of images that may be used by all shaders
    in a program is limited.


    (Add a new subsection after Section 2.14.7, Shader Execution, p. 109)

    Section 2.14.X, Shader Memory Access

    Shaders may perform random-access reads and writes to texture or buffer
    object memory using built-in image load, store, and atomic functions, as
    described in the OpenGL Shading Language Specification.  The ability to
    perform such random-access reads and writes in system that may be highly
    pipelined results in ordering and synchronization issues discussed in the
    sections below.


    Shader Memory Access Ordering

    The order in which texture or buffer object memory is read or written by
    shaders is largely undefined.  For some shader types (vertex, tessellation
    evaluation, and in some cases, fragment), the number of shader invocations
    that might perform loads and stores is even undefined.  In particular, the
    following rules apply:

      * While a vertex or tessellation evaluation shader will be executed at
        least once for each unique vertex specified by the application (vertex
        shaders) or generated by the tessellation primitive generator
        (tessellation evaluation shaders), it may be executed more than once
        for implementation-dependent reasons.  Additionally, if the same
        vertex is specified multiple times in a collection of primitives
        (e.g., repeating an index in DrawElements), the vertex shader might be
        run only once.

      * For each fragment generated by the GL, the number of fragment shader
        invocations depends on a number of factors.  If the fragment fails the
        pixel ownership test (Section 4.1.1), the fragment shader may not be
        executed.  Otherwise, if the framebuffer has no multisample buffer
        (SAMPLE_BUFFERS is zero), the fragment shader will be invoked exactly
        once.  If the fragment shader specifies per-sample shading, the
        fragment shader will be run once per covered sample.  Otherwise, the
        number of fragment shader invocations is undefined, but must be in the
        range [1,<N>], where <N> is the number of samples covered by the
        fragment.

      * If a fragment shader is invoked to process fragments or samples not
        covered by a primitive being rasterized to facilitate the
        approximation of derivatives for texture lookups, stores and atomics
        have no effect.

      * The relative order of invocations of the same shader type are
        undefined.  A store issued by a shader when working on primitive B
        might complete prior to a store for primitive A, even if primitive A
        is specified prior to primitive B.  This applies even to fragment
        shaders; while fragment shader outputs are written to the framebuffer
        in primitive order, stores executed by fragment shader invocations are
        not.

      * The relative order of invocations of different shader types is largely
        undefined.  However, when executing a shader whose inputs are
        generated from a previous programmable stage, the shader invocations
        from the previous stage are guaranteed to have executed far enough to
        generate final values for all next-stage inputs.  That implies shader
        completion for all stages except geometry; geometry shaders are
        guaranteed only to have executed far enough to emit all needed
        vertices.

    The above limitations on shader invocation order also make some forms of
    synchronization between shader invocations within a single set of
    primitives unimplementable.  For example, having one invocation poll
    memory written by another invocation assumes that the other invocation has
    been launched and can complete its writes.  The only case where such a
    guarantee is made is when the inputs of one shader invocation are
    generated from the outputs of a shader invocation in a previous stage.

    Stores issued to different memory locations within a single shader
    invocation may not be visible to other invocations in the order they were
    performed.  The built-in function memoryBarrier() may be used to provide
    stronger ordering of reads and writes performed by a single invocation.
    Calling memoryBarrier() guarantees that any memory transactions issued by
    the shader invocation prior to the call complete prior to the memory
    transactions issued after the call.  Memory barriers may be needed for
    algorithms that require multiple invocations to access the same memory and
    require the operations need to be performed in a partially-defined
    relative order.  For example, if one shader invocation does a series of
    writes, followed by a memoryBarrier() call, followed by another write,
    then another invocation that sees the results of the final write will also
    see the previous writes.  Without the memory barrier, the final write may
    be visible before the previous writes.

    The atomic memory transaction built-in functions may be used to read and
    write a given memory address atomically.  While atomic built-in functions
    issued by multiple shader invocations are executed in undefined order
    relative to each other, these functions perform both a read and a write of
    a memory address and guarantee that no other memory transaction will write
    to the underlying memory between the read and write.  Atomics allow
    shaders to use shared global addresses for mutual exclusion or as
    counters, among other uses.


    Shader Memory Access Synchronization

    Data written to textures or buffer objects by a shader invocation may
    eventually be read by other shader invocations, sourced by other fixed
    pipeline stages, or read back by the application.  When applications write
    to buffer objects or textures using API commands such as TexSubImage* or
    BufferSubData, the GL implementation knows when and where writes occur and
    can perform implicit synchronization to ensure that operations requested
    before the update see the original data and that subsequent operations see
    the modified data.  Without logic to track the target address of each
    shader instruction performing a store, automatic synchronization of stores
    performed by a shader invocation would require the GL implementation to
    make worst-case assumptions at significant performance cost.  To permit
    cases where textures or buffers may be read or written in different
    pipeline stages without the overhead of automatic synchronization, buffer
    object and texture stores performed by shaders are not automatically
    synchronized with other GL operations using the same memory.

    Explicit synchronization is required to ensure that the effects of buffer
    and texture data stores performed by shaders will be visible to subsequent
    operations using the same objects and will not overwrite data still to be
    read by previously requested operations.  Without manual synchronization,
    shader stores for a "new" primitive may complete before processing of an
    "old" primitive completes.  Additionally, stores for an "old" primitive
    might not be completed before processing of a "new" primitive starts.  The
    command

        void MemoryBarrierEXT(bitfield barriers)

    defines a barrier ordering the memory transactions issued prior to the
    command relative to those issued after the barrier.  For the purposes of
    this ordering, memory transactions performed by shaders are considered to
    be issued by the rendering command that triggered the execution of the
    shader.  <barriers> is a bitfield indicating the set of operations that
    are synchronized with shader stores; the bits used in <barriers> are as
    follows:
    
    - VERTEX_ATTRIB_ARRAY_BARRIER_BIT_EXT:  If set, vertex data sourced from
        buffer objects after the barrier will reflect data written by shaders
        prior to the barrier.  The set of buffer objects affected by this bit
        is derived from the buffer object bindings or GPU addresses used for
        generic vertex attributes (VERTEX_ATTRIB_ARRAY_BUFFER bindings,
        VERTEX_ATTRIB_ARRAY_ADDRESS from NV_vertex_buffer_unified_memory), as
        well as those for arrays of named vertex attributes (e.g., vertex,
        color, normal).

    - ELEMENT_ARRAY_BARRIER_BIT_EXT:  If set, vertex array indices sourced from
        buffer objects after the barrier will reflect data written by shaders
        prior to the barrier.  The buffer objects affected by this bit are
        derived from the ELEMENT_ARRAY_BUFFER binding and the
        NV_vertex_buffer_unified_memory ELEMENT_ARRAY_ADDRESS address.

    - UNIFORM_BARRIER_BIT_EXT:  Shader uniforms and assembly program parameters
        sourced from buffer objects after the barrier will reflect data
        written by shaders prior to the barrier.

    - TEXTURE_FETCH_BARRIER_BIT_EXT:  Texture fetches from shaders, including
        fetches from buffer object memory via buffer textures, after the
        barrier will reflect data written by shaders prior to the barrier.

    - SHADER_IMAGE_ACCESS_BARRIER_BIT_EXT:  Memory accesses using shader image
        load, store, and atomic built-in functions issued after the barrier
        will reflect data written by shaders prior to the barrier.
        Additionally, image stores and atomics issued after the barrier will
        not execute until all memory accesses (e.g., loads, stores, texture
        fetches, vertex fetches) initiated prior to the barrier complete.

    - COMMAND_BARRIER_BIT_EXT:  Command data sourced from buffer objects by
        Draw*Indirect commands after the barrier will reflect data written by
        shaders prior to the barrier.  The buffer objects affected by this bit
        are derived from the DRAW_INDIRECT_BUFFER_EXT binding and the GPU
        address DRAW_INDIRECT_ADDRESS_EXT.

    - PIXEL_BUFFER_BARRIER_BIT_EXT:  Reads/writes of buffer objects via the
        PACK/UNPACK_BUFFER bindings (ReadPixels, TexSubImage, etc.) after the
        barrier will reflect data written by shaders prior to the barrier.
        Additionally, buffer object writes issued after the barrier will wait
        on the completion of all shader writes initiated prior to the barrier.

    - TEXTURE_UPDATE_BARRIER_BIT_EXT:  Writes to a texture via Tex(Sub)Image*,
        CopyTex(Sub)Image*, CompressedTex(Sub)Image*, and reads via
        GetTexImage after the barrier will reflect data written by shaders
        prior to the barrier.  Additionally, texture writes from these
        commands issued after the barrier will not execute until all shader
        writes initiated prior to the barrier complete.

    - BUFFER_UPDATE_BARRIER_BIT_EXT:  Reads/writes via Buffer(Sub)Data,
        MapBuffer(Range), CopyBufferSubData, ProgramBufferParameters, and
        GetBufferSubData after the barrier will reflect data written by
        shaders prior to the barrier.  Additionally, writes via these commands
        issued after the barrier will wait on the completion of all shader
        writes initiated prior to the barrier.

    - FRAMEBUFFER_BARRIER_BIT_EXT:  Reads and writes via framebuffer object
        attachments after the barrier will reflect data written by shaders
        prior to the barrier.  Additionally, framebuffer writes issued after
        the barrier will wait on the completion of all shader writes issued
        prior to the barrier.

    - TRANSFORM_FEEDBACK_BARRIER_BIT_EXT:  Writes via transform feedback
        bindings after the barrier will reflect data written by shaders prior
        to the barrier.  Additionally, transform feedback writes issued after
        the barrier will wait on the completion of all shader writes issued
        prior to the barrier.

    - ATOMIC_COUNTER_BARRIER_BIT_EXT: Accesses to atomic counters after the
        barrier will reflect writes prior to the barrier.

    If <barriers> is ALL_BARRIER_BITS_EXT, shader memory accesses will be
    synchronized relative to all the operations described above.

    Implementations may cache buffer object and texture image memory that
    could be written by shaders in multiple caches; for example, there may be
    separate caches for texture, vertex fetching, and one or more caches for
    shader memory accesses.  Implementations are not required to keep these
    caches coherent with shader memory writes.  Stores issued by one
    invocation may not be immediately observable by other pipeline stages or
    other shader invocations because the value stored may remain in a cache
    local to the processor executing the store, or because data overwritten by
    the store is still in a cache elsewhere in the system.  When MemoryBarrier
    is called, the GL flushes and/or invalidates any caches relevant to the
    operations specified by the <barriers> parameter to ensure consistent
    ordering of operations across the barrier.

    To allow for independent shader invocations to communicate by reads and
    writes to a common memory address, image variables in the OpenGL Shading
    Language may be declared as "coherent".  Buffer object or texture image
    memory accessed through such variables may be cached only if caches are
    automatically updated due to stores issued by any other shader invocation.
    If the same address is accessed using both coherent and non-coherent
    variables, the accesses using variables declared as coherent will observe
    the results stored using coherent variables in other invocations.  Using
    variables declared as "coherent" guarantees only that the results of
    stores will be immediately visible to shader invocations using
    similarly-declared variables; calling MemoryBarrier is required to ensure
    that the stores are visible to other operations.

    The following guidelines may be helpful in choosing when to use coherent
    memory accesses and when to use barriers.

    - Data that are read-only or constant may be accessed without using
      coherent variables or calling MemoryBarrierEXT().  Updates to the
      read-only data via API calls such as BufferSubData will invalidate
      shader caches implicitly as required.

    - Data that are shared between shader invocations at a fine granularity
      (e.g., written by one invocation, consumed by another invocation) should
      use coherent variables to read and write the shared data.

    - Data written by one shader invocation and consumed by other shader
      invocations launched as a result of its execution ("dependent
      invocations") should use coherent variables in the producing shader
      invocation and call memoryBarrier() after the last write.  The consuming
      shader invocation should also use coherent variables.

    - Data written to image variables in one rendering pass and read by the
      shader in a later pass need not use coherent variables or
      memoryBarrier().  Calling MemoryBarrierEXT() with the
      SHADER_IMAGE_ACCESS_BARRIER_BIT_EXT set in <barriers> between passes is
      necessary.

    - Data written by the shader in one rendering pass and read by another
      mechanism (e.g., vertex or index buffer pulling) in a later pass need
      not use coherent variables or memoryBarrier().  Calling
      MemoryBarrierEXT() with the appropriate bits set in <barriers> between
      passes is necessary.


Additions to Chapter 3 of the OpenGL 3.2 (Compatibility Profile) Specification
(Rasterization)

    (insert new section immediately before Section 3.8, Texturing, p. 210)

    Section 3.X, Early Per-Fragment Tests

    Once fragments are produced by rasterization (sections 3.4 through 3.8), a
    number of per-fragment operations may be performed prior to fragment
    shader execution.  If a fragment is discarded during any of these
    operations, it will not be processed by any subsequent stage, including
    fragment shader execution.

    Up to six operations are performed on each fragment, in the following
    order:

      * the pixel ownership test, described in section 4.1.1;

      * the scissor test, described in section 4.1.2;

      * the depth bounds test, described in section 4.1.X (of the
        EXT_depth_bounds_test specification);

      * the stencil test, described in section 4.1.5;

      * the depth buffer test, described in section 4.1.6; and

      * occlusion query sample counting, described in section 4.1.7.

    The pixel ownership and scissor tests are always performed.

    The other operations are performed if and only if early fragment tests are
    enabled in the active fragment shader (section 3.12.2).  When early
    per-fragment operations are enabled, the depth bounds test, stencil test,
    depth buffer test, and occlusion query sample counting operations are
    performed prior to fragment shader execution, and the stencil buffer,
    depth buffer, and occlusion query sample counts will be updated
    accordingly.  When early per-fragment operations are enabled, these
    operations will not be performed again after fragment shader execution.
    When there is no active program, the active program has no fragment
    shader, or the active program was linked with early fragment tests
    disabled, these operations are performed only after fragment program
    execution, in the order described in chapter 4.

    If early fragment tests are enabled, any depth value computed by the
    fragment shader has no effect.  Additionally, the depth buffer, stencil
    buffer, and occlusion query sample counts may be updated even for
    fragments or samples that would be discarded after fragment shader
    execution due to per-fragment operations such as alpha-to-coverage or
    alpha tests.


    (Add new section after Section 3.9.19, Texture Application, p. 268)

    Section 3.9.X, Texture Image Loads and Stores

    The contents of a texture may be made available for shaders to read and
    write by binding the texture to one of a collection of image units.  The
    GL implementation provides an array of image units numbered beginning with
    zero, with the total number of image units provided given by the
    implementation-dependent constant MAX_IMAGE_UNITS_EXT.  Unlike texture
    image units, image units do not have a separate attachment for each
    texture target texture; each image unit may have only one texture bound at
    a time.

    A texture may be bound to an image unit for use by image loads and stores
    by calling:

        void BindImageTextureEXT(uint index, uint texture, int level,
                                 boolean layered, int layer, enum access,
                                 int format);

    where <index> identifies the image unit, <texture> is the name of the
    texture, and <level> selects a single level of the texture.  If <texture>
    is zero, <level> is ignored and the currently bound texture to image unit
    <index> is unbound.  If <index> is less than zero or greater than or equal
    to MAX_IMAGE_UNITS_EXT, or if <texture> is not the name of an existing
    texture object, the error INVALID_VALUE is generated.

    If the texture identified by <texture> is a one-dimensional array,
    two-dimensional array, three-dimensional, cube map, cube map array, or
    two-dimensional multisample array texture, it is possible to bind either
    the entire texture level or a single layer or face of the texture level.
    If <layered> is TRUE, the entire level is bound.  If <layered> is FALSE,
    only the single layer identified by <layer> will be bound.  When <layered>
    is FALSE, the single bound layer is treated as a different texture target
    for image accesses:

      * one-dimensional array texture layers are treated as one-dimensional
        textures;

      * two-dimensional array, three-dimensional, cube map, cube map array
        texture layers are treated as two-dimensional textures; and

      * two-dimensional multisample array textures are treated as
        two-dimensional multisample textures.

    For cube map textures where <layered> is FALSE, the face is taken by
    mapping the layer number to a face according to table 4.13.  For cube map
    array textures where <layered> is FALSE, the selected layer number is
    mapped to a texture layer and cube face using the following equations and
    mapping <face> to a face according to table 4.13.

      layer  = floor(layer_orig / 6)
      face   = layer_orig - (layer * 6)

    <format> specifies the format that the elements of the image will be 
    treated as when doing formatted stores, as described later in this 
    section. This is referred to as the "image unit format". This must be one
    of the formats listed in Table X.2, otherwise the error INVALID_VALUE is
    generated.

    <access> specifies whether the texture bound to the image will be treated
    as READ_ONLY, WRITE_ONLY, or READ_WRITE.  If a shader reads from an image
    unit with a texture bound as WRITE_ONLY, or writes to an image unit with a
    texture bound as READ_ONLY, the results of that shader operation are
    undefined and may lead to application termination.

    If a texture object bound to one or more image units is deleted by
    DeleteTextures, it is detached from each such image unit, as though
    BindImageTextureEXT were called with <index> identifying the image unit and
    <texture> set to zero.

    When a shader accesses the texture bound to an image unit using a built-in
    image load, store, or atomic function, it identifies a single texel by
    providing a one-, two-, or three-dimensional coordinate.  Multisample
    texture accesses also specify a sample number.  A coordinate vector is
    mapped to an individual texel tau_i, tau_i_j, or tau_i_j_k according to
    the target of the texture bound to the image unit using Table X.1.  As
    noted above, single-layer bindings of array or cube map textures are
    considered to use a texture target corresponding to the bound layer,
    rather than that of the full texture.

                                                   face/
                                          i  j  k  layer
                                          -- -- -- -----
        TEXTURE_1D                        x  -  -    -
        TEXTURE_2D                        x  y  -    -
        TEXTURE_3D                        x  y  z    -
        TEXTURE_RECTANGLE                 x  y  -    -
        TEXTURE_CUBE_MAP                  x  y  -    z
        TEXTURE_BUFFER                    x  -  -    -
        TEXTURE_1D_ARRAY                  x  -  -    y
        TEXTURE_2D_ARRAY                  x  y  -    z
        TEXTURE_CUBE_MAP_ARRAY            x  y  -    z
        TEXTURE_2D_MULTISAMPLE            x  y  -    -
        TEXTURE_2D_MULTISAMPLE_ARRAY      x  y  -    z

        Table X.1, Mapping of image load, store, and atomic texel coordinate
        components to texel numbers.

    If the texture target has layers or cube map faces, the layer or face
    number is taken from the <layer> argument of BindImageTextureEXT if the
    texture is bound with <layered> set to FALSE, or from the coordinate
    identified by Table X.1 otherwise.  For cube map and cube map array
    textures with <layered> set to TRUE, the coordinate is mapped to a layer
    and face in the same manner as the <layer> argument of
    BindImageTextureEXT.

    If the individual texel identified for an image load, store, or atomic
    operation doesn't exist, the access is treated as invalid.  Invalid image
    loads will return zero.  Invalid image stores will have no effect.
    Invalid image atomics will not update any texture bound to the image unit
    and will return zero.  An access is considered invalid if:

      * no texture is bound to the selected image unit;

      * the texture bound to the selected image unit is incomplete;

      * the texture level bound to the image unit is less than the base
        level or greater than the maximum level of the texture;

      * the texture bound to the image unit is bordered;

      * the internal format of the texture bound to the image unit is not
        found in Table X.2;

      * the internal format of the texture is incompatible with the specified
        <format> according to Table X.2.

      * the texture bound to the image unit has layers, is bound with
        <layered> set to TRUE, and the selected layer or cube map face doesn't
        exist;

      * the selected texel tau_i, tau_i_j, or tau_i_j_k doesn't exist;

      * the <x>, <y>, or <z> coordinate is not listed in the selected row of
        Table X.1 and is non-zero; or

      * the texture bound to the image unit has layers, is bound with
        <layered> set to FALSE, and the corresponding coordinate in the
        face/layer column of Table X.1 is non-zero.

      * the image has more samples than the implementation-dependent value of
        MAX_IMAGE_SAMPLES_EXT.

      * the access is a load and the format is not compatible with the 
        "size" layout qualifier of the image uniform.

     For textures with multiple samples per texel, the sample selected for an
     image load, store, or atomic is undefined if the <sample> coordinate is
     negative or greater than or equal to the number of samples in the
     texture.

     If a shader performs an image load, store, or atomic operation using an
     image variable declared as an array, and if the index used to select an
     individual out of bounds is negative or greater than or equal to the size
     of the array, the results of the operation are undefined but may not lead
     to termination.

     Accesses to textures bound to image units do format conversions based on 
     the <format> argument specified when the image is bound. Loads always
     return a value as a vec4, ivec4, or uvec4, and stores always take the
     source data as a vec4, ivec4, or uvec4. Data is converted to/from the 
     specified format as if it were passed through a TexImage2D or GetTexImage
     command with <format> and <type> as RGBA and FLOAT for vec4 data, with 
     <format> and <type> as RGBA_INTEGER and INT for ivec4 data, or with 
     <format> and <type> as RGBA_INTEGER and UNSIGNED_INT for uvec4 data.
     Unused components are filled in with (0,0,0,1) (where "1" is either a
     float or integer depending on the format). 
     
     The formats that are supported for image loads are dependent on the 
     layout(size*) qualifier of the image uniform. The following formats 
     are supported for image loads:

     - size1x8: R8I, R8UI
     - size1x16: R16I, R16UI
     - size1x32: R32F, R32I, R32UI
     - size2x32: RG32F, RG32I, RG32UI
     - size4x32: RGBA32F, RGBA32I, RGBA32UI

     Image stores support all formats in Table X.2.

     Table X.2 specifies how each format is stored in memory, which must be 
     made explicit because a single image can be viewed with multiple formats 
     according to the <format> argument. The "R", "G", "B", and "A" columns
     indicate which bits of which 32-bit word correspond to that component.
     For example, an entry of "1[15:0]" indicates that the selected component
     uses sixteen bits with its most significant bit in bit 15 of the second 
     word of memory and its least significant bit in bit 0. Floating-point 
     textures with 32-bit components are stored using the IEEE standard 
     representation; textures with 10-, 11-, or 16-bit floating-point 
     components are stored according to Sections 2.1.2 and 2.1.3.  

     The "equivalence" column of Table X.2 defines a set of equivalence 
     classes for formats, such that if the internal format of a texture level
     is in the same equivalence class as the <format> argument to 
     BindImageTextureEXT then the image may be viewed with that format. 
     Otherwise, the access is considered invalid as described above.

        Internal format  Equivalence      R        G         B         A
        ---------------  -----------   -------  -------   -------   -------
        RGBA32F             4x32       0[31:0]  1[31:0]   2[31:0]   3[31:0]
        RGBA16F             2x32       0[15:0]  0[31:16]  1[15:0]   1[31:16]
        RG32F               2x32       0[31:0]  1[31:0] 
        RG16F               1x32       0[15:0]  0[31:16] 
        R11F_G11F_B10F      1x32       0[10:0]  0[21:11]  0[31:22] 
        R32F                1x32       0[31:0]  
        R16F                1x16       0[15:0] 

        RGBA32UI            4x32       0[31:0]  1[31:0]   2[31:0]   3[31:0]
        RGBA16UI            2x32       0[15:0]  0[31:16]  1[15:0]   1[31:16]
        RGB10_A2UI          1x32       0[9:0]   0[19:10]  0[29:20]  0[31:30]
        RGBA8UI             1x32       0[7:0]   0[15:8]   0[23:16]  0[31:24]
        RG32UI              2x32       0[31:0]  1[31:0]
        RG16UI              1x32       0[15:0]  0[31:16] 
        RG8UI               1x16       0[7:0]   0[15:8] 
        R32UI               1x32       0[31:0]  
        R16UI               1x16       0[15:0] 
        R8UI                1x8        0[7:0] 

        RGBA32I             4x32       0[31:0]  1[31:0]   2[31:0]   3[31:0]
        RGBA16I             2x32       0[15:0]  0[31:16]  1[15:0]   1[31:16]
        RGBA8I              1x32       0[7:0]   0[15:8]   0[23:16]  0[31:24]
        RG32I               2x32       0[31:0]  1[31:0]
        RG16I               1x32       0[15:0]  0[31:16] 
        RG8I                1x16       0[7:0]   0[15:8] 
        R32I                1x32       0[31:0]  
        R16I                1x16       0[15:0] 
        R8I                 1x8        0[7:0] 

        RGBA16              2x32       0[15:0]  0[31:16]  1[15:0]   1[31:16]
        RGB10_A2            1x32       0[9:0]   0[19:10]  0[29:20]  0[31:30]
        RGBA8               1x32       0[7:0]   0[15:8]   0[23:16]  0[31:24]
        RG16                1x32       0[15:0]  0[31:16] 
        RG8                 1x16       0[7:0]   0[15:8] 
        R16                 1x16       0[15:0] 
        R8                  1x8        0[7:0] 

        RGBA16_SNORM        2x32       0[15:0]  0[31:16]  1[15:0]   1[31:16]
        RGBA8_SNORM         1x32       0[7:0]   0[15:8]   0[23:16]  0[31:24]
        RG16_SNORM          1x32       0[15:0]  0[31:16] 
        RG8_SNORM           1x16       0[7:0]   0[15:8] 
        R16_SNORM           1x16       0[15:0] 
        R8_SNORM            1x8        0[7:0] 

        Table X.2, Supported texture formats, component packing, and 
        equivalence classes for formatted image accesses.

     Implementations may support a limited combined number of image units and
     active fragment shader outputs (section 4.2.1).  A link error will be 
     generated if the number of active image uniforms used in all shaders and 
     the number of active fragment shader outputs exceeds the implementation-
     dependent value (MAX_COMBINED_IMAGE_UNITS_AND_FRAGMENT_OUTPUTS_EXT).


   Modify Section 3.12.2, Shader Execution, p. 274

   (add new unnumbered subsection section at the end of the section, p. 279)

   Early Fragment Tests

   An explicit control is provided to allow fragment shaders to enable early
   fragment tests.  If the fragment shader specifies the
   "early_fragment_tests" layout qualifier, the per-fragment tests described
   in Section 3.X will be performed prior to fragment shader execution.
   Otherwise, they will be performed after fragment shader execution.


Additions to Chapter 4 of the OpenGL 3.2 (Compatibility Profile) Specification
(Per-Fragment Operations and the Framebuffer)

    None.


Additions to Chapter 5 of the OpenGL 3.2 (Compatibility Profile) Specification
(Special Functions)

    Modify Section 5.4.1, Commands Not Usable In Display Lists (p. 358)

    (add "MemoryBarrierEXT" to the list of commands not allowed in a display
     list, in the "Buffer objects" paragraph)


Additions to Chapter 6 of the OpenGL 3.2 (Compatibility Profile) Specification
(State and State Requests)

    None.


New Implementation Dependent State

                                                        Minimum
    Get Value                    Type  Get Command      Value      Description             Sec.       Attrib
    ---------                    ----  -----------      -------    -----------             ----       ------
    MAX_IMAGE_UNITS_EXT          Z+    GetIntegerv      8          number of units for     3.9.X        -
                                                                   image load/store/atom
    MAX_COMBINED_IMAGE_UNITS_    Z+    GetIntegerv      8          limit on active image   3.9.X        -
      AND_FRAGMENT_OUTPUTS_EXT                                     units + fragment outputs
    MAX_IMAGE_SAMPLES_EXT        Z     GetIntegerv      0          max allowed samples     3.9.X        -
                                                                   for a texture level
                                                                   bound to an image unit  

New State

    Add a new Table 6.X, Image Stage (state per image unit)

    Get Value                            Type    Get Command     Initial Value   Sec     Attribute
    ---------                            ----    -----------     -------------   ---     ---------
    IMAGE_BINDING_NAME_EXT               8*xZ+   GetIntegeri_v    0             3.9.X      none
    IMAGE_BINDING_LEVEL_EXT              8*xZ+   GetIntegeri_v    0             3.9.X      none
    IMAGE_BINDING_LAYERED_EXT            8*xB    GetBooleani_v    FALSE         3.9.X      none
    IMAGE_BINDING_LAYER_EXT              8*xZ+   GetIntegeri_v    0             3.9.X      none
    IMAGE_BINDING_ACCESS_EXT             8*xZ3   GetIntegeri_v    READ_ONLY     3.9.X      none
    IMAGE_BINDING_FORMAT_EXT             8*xZ+   GetIntegeri_v    R8            3.9.X      none


Additions to Appendix A of the OpenGL 3.2 (Compatibility Profile)
Specification (Invariance)

    None.


Additions to the AGL/GLX/WGL Specifications

    None.


GLX Protocol

    !!! TBD !!!

    
Modifications to the OpenGL Shading Language Specification, Version 1.50

    Including the following line in a shader can be used to control the
    language features described in this extension:

      #extension GL_EXT_shader_image_load_store : <behavior>

    where <behavior> is as specified in section 3.3.

    New preprocessor #defines are added to the OpenGL Shading Language:

      #define GL_EXT_shader_image_load_store    1


    Modify Section 3.6, Keywords, p. 14

    (add the following to the list of keywords, p. 14)

    coherent
    volatile
    restrict

    image1D             iimage1D                uimage1D
    image2D             iimage2D                uimage2D
    image3D             iimage3D                uimage3D
    image2DRect         iimage2DRect            uimage2DRect
    imageCube           iimageCube              uimageCube
    imageBuffer         iimageBuffer            uimageBuffer
    image1DArray        iimage1DArray           uimage1DArray
    image2DArray        iimage2DArray           uimage2DArray
    imageCubeArray      iimageCubeArray         uimageCubeArray
    image2DMS           iimage2DMS              uimage2DMS
    image2DMSArray      iimage2DMSArray         uimage2DMSArray

    (remove from the list of reserved keywords, p. 15)

    volatile


    (Insert a new section immediately after Section 4.1.7, Samplers, p. 23)

    Section 4.1.X, Images

    Like samplers, images are opaque handles to one-, two-, or
    three-dimensional images corresponding to all or a portion of a single
    level of a texture image bound to an image unit.  There are distinct 
    image variable types for each texture target, and for each of float, 
    integer, and unsigned integer data types.  Image accesses should use
    an image type that matches the target of the texture whose level is
    bound to the image unit, or for non-layered bindings of 3D or array 
    images should use the image type that matches the dimensionality of
    the layer of the image (i.e. a layer of 3D, 2DArray, Cube, or 
    CubeArray should use image2D, a layer of 1DArray should use image1D, 
    and a layer of 2DMSArray should use image2DMS). If the image target type 
    does not match the bound image in this manner, if the data type does not
    match the bound image, or if the "size" layout qualifier does not match 
    the image unit format as described in Section 3.9.X of the OpenGL 
    Specification, the results of image accesses are undefined but may not 
    include program termination.

    Image variables are used in the image load, store, and atomic functions
    described in Section 8.X, "Image Functions" to specify an image to access.
    They can only be declared as function parameters or uniform variables (see
    Section 4.3.5 "Uniform").  Except for array indexing, structure field
    selection, and parentheses, images are not allowed to be operands in
    expressions.  Images may be aggregated into arrays within a shader (using
    square brackets [ ]) and can be indexed with general integer expressions.
    The results of accessing an image array with an out-of-bounds index are
    undefined.  Images cannot be treated as l-values; hence, they cannot be
    used as out or inout function parameters, nor can they be assigned into.
    As uniforms, they are initialized only with the OpenGL API; they cannot be
    declared with an initializer in a shader.  As function parameters, images
    may only be passed to samplers of matching type.


    Modify Section 4.3, Storage Qualifiers, p. 29

    (add new qualifiers to the first table, p. 29)

        Qualifier       Meaning
        ------------    -------------------------------------------------
        coherent        memory variable where reads and writes are coherent
                        with reads and writes from other shader invocations

        volatile        memory variable whose underlying value may be
                        changed at any point during shader execution by
                        some source other than the current shader invocation

        restrict        memory variable where use of that variable is the
                        only way to read and write the underlying memory
                        in the relevant shader stage


    Modify Section 4.3.2, Constant Qualifier (p. 30)

    (add after last paragraph of section)  

    Because image variables can not be built from constant expressions, the
    "const" qualifier may not be used to create a compile-time constant image
    variable.  However, the "const" qualifier may be used to declare image
    variables whose image data are treated as constant, as described in
    Section 4.3.X.


    Modify Section 4.3.8.1 (Input Layout Qualifiers), p. 39

    Remove "only" from the sentence:

    Fragment shaders can have an input layout only for redeclaring the 
    built-in variable gl_FragCoord...

    Add to the end of the section:

    Fragment shaders also allow an input layout qualifier on the qualifier
    "in". The only valid layout qualifier is:

      layout-qualifier-id
        early_fragment_tests

    to indicate that fragment tests will be performed before fragment shader
    execution, as described in Section 3.12.2 of the OpenGL Specification. 
    For example,

      layout(early_fragment_tests) in;


    (Insert immediately after Section 4.3.8.3, Uniform Block Layout
     Qualifiers, p. 40)

    Section 4.3.8.X, Image Qualifiers

    Layout qualifiers can be used for image variable declarations.  The layout
    qualifier identifiers for image variable declarations are

      layout-qualifier-id
        size1x8
        size1x16
        size1x32
        size2x32
        size4x32

    The "size" identifiers indicate the set of image formats that the image
    variable can be used to access.  Only one "size" identifier may be
    specified for any variable declaration.  A layout of "size1x8" is illegal
    for image variables associated with floating-point data types.

    All image variable declarations, including function parameter
    declarations, must specify a "size" layout qualifier.  It is an error to
    declare an image uniform variable or function parameter without a size
    qualifier.


    (Insert immediately after Section 4.3.9, Interpolation, p. 42)

    Section 4.3.X, Memory Access Qualifiers

    The "coherent", "volatile", "restrict", and "const" storage qualifiers can
    be specified in image variable declarations to control memory accesses
    using the declared variables.

    Memory accesses to image variables declared using the "coherent" storage
    qualifier are performed coherently with similar accesses from other shader
    invocations.  In particular, when reading a variable declared as
    "coherent", the values returned will reflect the results of previously
    completed writes performed by other shader invocations.  When writing a
    variable declared as "coherent", the values written will be reflected in
    subsequent coherent reads performed by other shader invocations.  As
    described in the Section 2.20.X of the OpenGL Specification, shader memory
    reads and writes complete in a largely undefined order.  The built-in
    function memoryBarrier() can be used if needed to guarantee the completion
    and relative ordering of memory accesses performed by a single shader
    invocation.

    When accessing memory using variables not declared as "coherent", the
    memory accessed by a shader may be cached by the implementation to service
    future accesses to the same address.  Memory stores may be cached in such
    a way that the values written may not be visible to other shader
    invocations accessing the same memory.  The implementation may cache the
    values fetched by memory reads and return the same values to any shader
    invocation accessing the same memory, even if the underlying memory has
    been modified since the first memory read.  While variables not declared
    as "coherent" may not be useful for communicating between shader
    invocations, using non-coherent accesses may result in higher performance.

    Memory accesses to image variables declared using the "volatile" storage
    qualifier must treat the underlying memory as though it could be read or
    written at any point during shader execution by some source other than the
    executing shader invocation.  When a volatile variable is read, its value
    must be re-fetched from the underlying memory, even if the shader
    invocation performing the read had already fetched its value from the same
    memory once.  When a volatile variable is written, its value must be
    written to the underlying memory, even if the compiler can conclusively
    determine that its value will be overwritten by a subsequent write.  Since
    the external source reading or writing a "volatile" variable may be
    another shader invocation, variables declared as "volatile" are
    automatically treated as coherent.

    Memory accesses to image variables declared using the "restrict" storage
    qualifier may be compiled assuming that the variable used to perform the
    memory access is the only way to access the underlying memory using the
    shader stage in question.  This allows the compiler to coalesce or reorder
    loads and stores using "restrict"-qualified image variables in ways that
    wouldn't be permitted for image variables not so qualified, because the
    compiler can assume that the underlying image won't be read or written by
    other code.  Applications are responsible for ensuring that image memory
    referenced by variables qualified with "restrict" will not be referenced
    using other variables in the same scope; otherwise, accesses to
    "restrict"-qualified variables will have undefined results.

    Memory accesses to image variables declared using the "const" storage
    qualifier may only read the underlying memory, which is treated as
    read-only.  It is an error to pass an image variable qualified with
    "const" to imageStore() or imageAtomic*().

    In image variable declarations, the "coherent", "volatile", "restrict",
    and "const" qualifiers can be positioned anywhere in the declaration,
    either before or after the data type of the variable being qualified.
    Qualifiers before the type name apply to the image data referenced by the
    image variable; qualifiers after the type name apply to the image variable
    itself.  It is an error to specify "restrict" prior to the type name, as
    "restrict" can only qualify the image variable itself.

    The "coherent", "volatile", and "restrict" storage qualifiers may only be
    used on image variables, and may not be used on variables of any other
    type.  "const" may be used in declarations with non-image variable types,
    as described in Section 4.3.2.

    The values of variables qualified with "coherent", "volatile", "restrict",
    or "const" may not be assigned to function parameters lacking such
    qualifiers.  It is legal to add qualifiers in a function call, but not to
    remove them.

      vec4 funcA(layout(size4x32) image2D restrict a)   { ... }
      vec4 funcB(layout(size4x32) image2D a)            { ... }
      layout(size4x32) uniform image2D img1;
      layout(size4x32) coherent uniform image2D img2;

      funcA(img1);              // OK, adding "restrict" is allowed
      funcB(img2);              // illegal, stripping "coherent" is not

    
    (Insert a new numbered section at the end of Chapter 8, Built-in
    Functions, p. 69)

    Section 8.X, Image Functions

    Variables using one of the image data types may be used in the built-in
    shader image memory functions defined in this section to read and write
    individual texels of a texture.  Each image variable is an integer scalar
    that references an image unit, which has a texture image attached.

    When image memory functions access memory, an individual texel in the
    image is identified using an i, (i,j), or (i,j,k) coordinate corresponding
    to the values of <coord>.  For image2DMS and image2DMSArray variables (and 
    the corresponding int/unsigned int types) corresponding to multisample 
    textures, each texel may have multiple samples and an individual sample is 
    identified using the integer <sample> parameter.  The coordinates and 
    sample number are used to select an individual texel in the manner 
    described in Section 3.9.X of the OpenGL specification.

    Loads and stores support float, integer, and unsigned integer types. The
    data types "gimage*" serve as placeholders meaning either "image*", 
    "iimage*", or "uimage*" in the same way as "gvec" or "gsampler".

    The "IMAGE_INFO" in the prototypes below is a placeholder representing
    33 separate functions, each for a different type of image variable.  The
    "IMAGE_INFO" placeholder is replaced by one of the following argument
    lists:

        gimage1D image, int coord
        gimage2D image, ivec2 coord
        gimage3D image, ivec3 coord
        gimage2DRect image, ivec2 coord
        gimageCube image, ivec3 coord
        gimageBuffer image, int coord
        gimage1DArray image, ivec2 coord
        gimage2DArray image, ivec3 coord
        gimageCubeArray image, ivec3 coord
        gimage2DMS image, ivec2 coord, int sample
        gimage2DMSArray image, ivec3 coord, int sample

    (Note that each of the "gimage*" lines represents one of three different
    image variable types.)

    Syntax:

      gvec4 imageLoad(const IMAGE_INFO);

    Description:

    Loads the texel at the coordinate <coord> from the image unit specified 
    by <image>.  For multisample loads, the sample number is given by 
    <sample>.  When <image>, <coord>, and <sample> identify a valid texel, 
    the bits used to represent the selected texel in memory are converted to 
    a vec4, ivec4, or uvec4 in the manner described in Section 3.9.X of the 
    OpenGL Specification and returned.


    Syntax:

      void imageStore(IMAGE_INFO, gvec4 data);

    Description:

    Stores the value of <data> into the texel at the coordinate <coord> from
    the image specified by <image>.  For multisample stores, the sample number
    is given by <sample>.  When <image>, <coord>, and <sample> identify a
    valid texel, the bits used to represent <data> are converted to the format
    of the image unit in the manner described in Section 3.9.X of the OpenGL
    Specification and stored to the specified texel.
    

    Syntax:

      uint      imageAtomicAdd(IMAGE_INFO, uint data);
      int       imageAtomicAdd(IMAGE_INFO, int data);

      uint      imageAtomicMin(IMAGE_INFO, uint data);
      int       imageAtomicMin(IMAGE_INFO, int data);

      uint      imageAtomicMax(IMAGE_INFO, uint data);
      int       imageAtomicMax(IMAGE_INFO, int data);

      uint      imageAtomicIncWrap(IMAGE_INFO, uint wrap);

      uint      imageAtomicDecWrap(IMAGE_INFO, uint wrap);

      uint      imageAtomicAnd(IMAGE_INFO, uint data);
      int       imageAtomicAnd(IMAGE_INFO, int data);

      uint      imageAtomicOr(IMAGE_INFO, uint data);
      int       imageAtomicOr(IMAGE_INFO, int data);

      uint      imageAtomicXor(IMAGE_INFO, uint data);
      int       imageAtomicXor(IMAGE_INFO, int data);

      uint      imageAtomicExchange(IMAGE_INFO, uint data);
      int       imageAtomicExchange(IMAGE_INFO, int data);

      uint      imageAtomicCompSwap(IMAGE_INFO, uint compare, uint data);
      int       imageAtomicCompSwap(IMAGE_INFO, int compare, int data);

    Description:

    These functions perform atomic operations on individual texels or samples
    of an image variable.  Atomic memory operations read a value from the
    selected texel, compute a new value using one of the operations described
    below, writes the new value to the selected texel, and returns the
    original value read.  The contents of the texel being updated by the
    atomic operation are guaranteed not to be updated by any other image store
    or atomic function between the time the original value is read and the
    time the new value is written.

    As with image load and store functions, <image>, <coord>, and <sample>
    specify the the individual texel to operate on.  The method for
    identifying the individual texel operated on from <image>, <coord>, and
    <sample>, and the method for reading and writing the texel are specified
    in Section 3.9.X of the OpenGL specification. The format of the image 
    unit must be in the "1x32" equivalence class in Table X.2 in Section 3.9.X
    of the OpenGL specification, otherwise the atomic operation is invalid.

    imageAtomicAdd() computes a new value by adding the value of <data> to the
    contents of the selected texel.  These functions support 32-bit unsigned
    integer operands and 32-bit signed integer operands.

    imageAtomicMin() computes a new value by taking the minimum of the value
    of <data> and the contents of the selected texel.  These functions support
    32-bit signed and unsigned integer operands.

    imageAtomicMax() computes a new value by taking the maximum of the value
    of <data> and the contents of the selected texel.  These functions support
    32-bit signed and unsigned integer operands.

    imageAtomicIncWrap() computes a new value by adding one to the contents of
    the selected texel, and then forcing the result to zero if and only if the
    incremented value is greater than or equal to <wrap>.  These functions
    support only 32-bit unsigned integer operands.

    imageAtomicDecWrap() computes a new value by subtracting one from the
    contents of the selected texel, and then forcing the result to <wrap>-1 if
    the original value read from the selected texel was either zero or greater
    than <wrap>.  These functions support only 32-bit unsigned integer
    operands.

    imageAtomicAnd() computes a new value by performing a bitwise and of the
    value of <data> and the contents of the selected texel.  These functions
    support 32-bit signed and unsigned integer operands.

    imageAtomicOr() computes a new value by performing a bitwise or of the
    value of <data> and the contents of the selected texel.  These functions
    support 32-bit signed and unsigned integer operands.

    imageAtomicXor() computes a new value by performing a bitwise exclusive or
    of the value of <data> and the contents of the selected texel.  These
    functions support 32-bit signed and unsigned integer operands.

    imageAtomicExchange() computes a new value by simply copying the value of
    <data>.  These functions support 32-bit signed and unsigned integer
    operands.

    imageAtomicCompSwap() compares the value of <compare> and the contents of
    the selected texel.  If the values are equal, the new value is given by
    <data>; otherwise, it is taken from the original value loaded from the
    texel.  These functions support 32-bit signed and unsigned integer
    operands.


    (Insert another new numbered section at the end of Chapter 8, Built-in
    Functions, p. 69)

    Section 8.Y, Shader Memory Functions

    Shaders of all types may read and write the contents of textures and
    buffer objects using image variables.  While the order or reads and writes
    within a single shader invocation is well-defined, the relative order of
    reads and writes to a single shared memory address from multiple separate
    invocations is largely undefined.

    Syntax:

      void      memoryBarrier(void);

    Description:

    memoryBarrier() can be used to control the ordering of memory transactions
    issued by a shader invocation.  When called, it will wait on the
    completion of all memory accesses resulting from the use of image
    variables prior to calling the function.  When all memory operations have
    been flushed, memoryBarrier() returns to the caller with no other effect.
    When this function returns, the results of any memory stores performed
    using coherent variables performed prior to the call will be visible to
    any future coherent memory access to the same addresses from other shader
    invocations.  In particular, the values written and flushed this way in
    one shader stage are guaranteed to be visible to coherent memory accesses
    performed by shader invocations in subsequent stages when those
    invocations were triggered by the execution of the original shader
    invocation (e.g., fragment shader invocations for a primitive resulting
    from a particular geometry shader invocation).


    Modify Section 9, Shading Language Grammar (p. 105)

    !!! TBD:  Add grammar constructs for memory access qualifiers, allowing
        memory access qualifiers before or after the type in a variable
        declaration.


Errors

    INVALID_VALUE is generated by Uniform1i{v} if the location refers to an
    image variable and the value specified is less than zero or greater than
    or equal to MAX_IMAGE_UNITS_EXT.

    INVALID_OPERATION is generated by Uniform* functions other than
    Uniform1i{v} if the location refers to an image variable.

    INVALID_VALUE is generated by BindImageTextureEXT if <index> is less than
    zero or greater than or equal to MAX_IMAGE_UNITS_EXT.

    INVALID_VALUE is generated by BindImageTextureEXT if <texture> is not the
    name of an existing texture object.

    INVALID_VALUE is generated by BindImageTextureEXT if <format> is not a 
    legal format.


Dependencies on OpenGL 3.2 (Core Profile)

    If only the core profile of OpenGL 3.2 is supported, references to buffer
    objects for conventional vertex attributes and to the Begin and RasterPos
    commands should be removed.

Dependencies on OpenGL 3.1, ARB_uniform_buffer_object, and
EXT_bindable_uniform

    If OpenGL 3.1, ARB_uniform_buffer_object, and EXT_bindable_uniform are not
    supported, references to UNIFORM_BARRIER_BIT should be removed.

Dependencies on ARB_draw_indirect

    If ARB_draw_indirect is not supported, references to COMMAND_BARRIER_BIT_EXT
    should be removed.

Dependencies on NV_vertex_buffer_unified_memory

    If NV_vertex_buffer_unified_memory is not supported, references to that
    extension and GPU addresses in the discussion of
    VERTEX_ATTRIB_ARRAY_BARRIER_BIT_EXT and ELEMENT_ARRAY_BARRIER_BIT_EXT should
    be removed.

Dependencies on OpenGL 3.2 and ARB_texture_multisample

    If OpenGL 3.2 and ARB_texture_multisample are not supported, references to
    multisample textures should be removed.

Dependencies on OpenGL 4.0 and ARB_sample_shading

    If OpenGL 4.0 or ARB_sample_shading is supported, the discussion of the
    number of shader invocations for a given fragment in the "Shader Memory
    Access" section of the specification should be updated to discuss the
    sample shading enable and the minimum sample shading factor provided in
    that extension.

Dependencies on OpenGL 4.0 and ARB_texture_cube_map_array

    If OpenGL 4.0 or ARB_texture_cube_map_array are not supported, references
    to cube map array textures should be removed.

Dependencies on OpenGL 3.3 and ARB_texture_rgb10_a2ui

    If OpenGL 3.3 or ARB_texture_rgb10_a2ui are not supported, references to
    the RGB10_A2UI texture format should be removed.

Dependencies on NV_shader_buffer_load

    If NV_shader_buffer_load is supported, the new section 2.14.X (Shader
    Memory Access) should be combined with "Section 2.20.X, Shader Memory
    Access" from NV_shader_buffer_load.

Dependencies on OpenGL 4.0, ARB_gpu_shader5, and NV_gpu_shader5

    If OpenGL 4.0, ARB_gpu_shader5, and NV_gpu_shader5 are not supported, the
    modifications to the OpenGL Shading Language Specification should be
    removed.

Dependencies on OpenGL 4.0 and ARB_tessellation_shader

    If OpenGL 4.0 and ARB_tessellation_shader are not supported, references to
    tessellation control and evaluation shaders should be removed.

Dependencies on EXT_shader_atomic_counters

    If EXT_shader_atomic_counters is not supported, remove references to
    ATOMIC_COUNTER_BARRIER_BIT_EXT.

Dependencies on EXT_depth_bounds_test

    If EXT_depth_bounds_test is not supported, references to the depth bounds
    test should be removed.

Dependencies on EXT_separate_shader_objects

    If EXT_separate_shader_objects is supported, early depth tests are enabled
    if and only if (a) there is an active program for the fragment shader
    stage and (b) the fragment shader in that program enables early depth
    tests using a layout qualifier.

Dependencies on NV_gpu_program5

    If NV_gpu_program5 is supported, the following edits are made to extend
    the assembly programming model documented in the NV_gpu_program4 extension
    and extended by NV_gpu_program5.  No "OPTION" line is required; the
    following capability is implied by NV_gpu_program5 program headers such as
    "!!NVfp5.0".

    If NV_gpu_program5 is not supported, the contents of this dependencies
    section should be ignored.

    Section 2.X.2, Program Grammar

    (add the following rules to the grammar)

      <namingStatement>       ::= IMAGE_statement

      <IMAGE_statement>       ::= "IMAGE" <establishName> <imageSingleInit>
                                | "IMAGE" <establishName> <optArraySize>
                                    <imageMultipleInit>

      <imageSingleInit>       ::= "=" <imageUseDS>

      <imageMultipleInit>     ::= "=" "{" <imageItemList> "}"

      <imageItemList>         ::= <imageUseDM>
                                | <imageUseDM> "," <imageItemList>

      <imageUseDS>            ::= "image" <arrayMemAbs>

      <imageUseDM>            ::= <imageUseDS>
                                | "image" <arrayRange>


      <instruction>           ::= <ImageInstruction>

      <ImageInstruction>:     ::= <LOADIMop_instruction>
                                | <STOREIMop_instruction>
                                | <ATOMIMop_instruction>

      <LOADIMop_instruction>  ::= <LOADIMop> <opModifiers> <instResult> "," 
                                       <instOperandV> "," <imageAccess>

      <STOREIMop_instruction> ::= <STOREIMop> <opModifiers> <imageUnit> "," 
                                       <instOperandV> "," <instOperandV> "," 
                                       <imageTarget>

      <ATOMIMop_instruction>  ::= <ATOMIMop> <opModifiers> <instResult> "," 
                                       <instOperandV> "," <instOperandV> "," 
                                       <imageAccess>

      <LOADIMop>              ::= "LOADIM"
      <STOREIMop>             ::= "STOREIM"
      <ATOMIMop>              ::= "ATOMIM"

      <imageAccess>           ::= <imageUnit> "," <imageTarget>

      <imageUnit>             ::= "image" <arrayMemAbs>
                                | <imageVarName> <optArrayMem>

      <imageTarget>           ::= "1D"
                                | "2D"
                                | "3D"
                                | "RECT"
                                | "CUBE"
                                | "BUFFER"
                                | "ARRAY1D"
                                | "ARRAY2D"
                                | "ARRAYCUBE"
                                | "2DMS"
                                | "ARRAY2DMS"

    Section 2.X.3.X, Program Image Variables

    Program image variables are used as constants during program execution
    and refer the image objects bound to one or more image units. All 
    image variables have associated bindings and are read-only during
    program execution.  Image variables retain their values across program
    invocations, and the set of image units to which they refer is
    constant.  The texture object a variable refers to may be changed by
    binding a new texture object to the corresponding image unit.  Image 
    variables may only be used to identify a texture object in image 
    instructions, and may not be used as operands in any other instruction.
    Image variables may be declared explicitly via the <IMAGE_statement> 
    grammar rule, or implicitly by using an image unit binding in an 
    instruction.

    Image array variables may be declared as arrays, but the list of image 
    units assigned to the array must increase consecutively.

      Binding          Components  Underlying State
      ---------------  ----------  ------------------------------------------
      image[a]             x       image object bound to image unit a
      image[a..b]          x       image objects bound to image units a
                                     through b

      Table X.12.2:  Image Unit Bindings.  <a> and <b> indicate image unit
      numbers.

    If an image binding matches "image[a]", the image variable is filled
    with a single integer referring to image unit <a>.

    If an image binding matches "image[a..b]", the image variable is
    filled with an array of integers referring to image units <a> through 
    <b>, inclusive.  A program will fail to compile if <a> or <b> is
    negative or greater than or equal to the number of image units
    supported, or if <a> is greater than <b>.


    Modify Section 2.X.4, Program Execution Environment

      Instr-      Modifiers 
      uction  V  F I C S H D  Out Inputs    Description
      ------- -- - - - - - -  --- --------  --------------------------------
      ATOMIM  50 - - X - - -  s   v,vs,i    atomic image operation
      LOADIM  50 - - X X - F  v   vs,i      image load
      MEMBAR  50 - - - - - -  -   -         memory barrier
      STOREIM 50 X X - - - F  -   i,v,vs    image store

      ...

      The input and output columns describe the formats of the operands and
      results of the instruction.

        i:  IMAGE variable, read-only


    Modify Section 2.X.4.1, Program Instruction Modifiers

    (add to Table X.14 of the NV_gpu_program4 specification.)

      Modifier  Description
      --------  ---------------------------------------------------
      COH       Mark LOADIM and STOREIM operations as coherent
      VOL       Make LOADIM and STOREIM operations as volatile

    For image load and store operations, the "COH" modifier controls whether
    the operation is performed in a manner guaranteed to be coherent with
    loads and stores performed by other shader invocations.

    For image load and store operations, the "VOL" modifier controls whether
    the operation should treat the contents of the image accessed as volatile,
    where the underlying image contents may be changed at any point during
    shader execution by some source other than the current shader thread.


    Section 2.X.8.Z, LOADIM:  Image Load

    The LOADIM instruction takes the components of a single signed integer
    vector operand and uses them as coordinates to perform an unformatted 
    image load from the texture bound to the image unit specified by 
    <imageUnit>. Unformatted loads read the data from memory without 
    converting from the image unit format, by copying raw bits from memory 
    to the destination variable according to the bit layouts described in 
    Table X.2, where word 0 is written to the .x component, word 1 to .y, 
    etc..

    Eleven image targets are supported:  1D, 2D, 3D, RECT, CUBE, BUFFER, 
    ARRAY1D, ARRAY2D, ARRAYCUBE, 2DMS, and ARRAY2DMS.  The texel coordinate 
    is a one-, two- or three-dimensional vector, taken from the <x>, <y>, and
    <z> components of the operand.  For the 2DMS and ARRAY2DMS, the texel
    coordinate is a two- or three-dimensional vector, taken from the <x>, 
    <y>, and <z> components of the operand, and a sample number is taken from
    the <w> component of the operand.

        coords = VectorLoad(op0);
        if (target == 1D || target == BUFFER) {
          coords.y = 0;
        }
        if (target == 1D || target == 2D ||
            target == BUFFER || target == RECT || 
            target == 2DMS) {
          coords.z = 0;
        }
        if (target != 2DMS && target != ARRAY2DMS) {
          coords.w = 0;
        }
        result = ImageLoad(image, coords);

    When an image load uses the "S8", "U8", "S16", "U16", "F32", "S32", or
    "U32" storage modifiers, the <x> component of the result contains the
    loaded value and the <y>, <z>, and <w> components of the result are zero,
    zero, and one (int or float, depending on the type of the opModifier),
    respectively. For "S8" and "S16" modifiers, the loaded value is sign-
    extended; for "U8" and "U16", the loaded value is zero-extended.  When 
    an image load uses the "F32X2", "S32X2", or "U32X2" storage modifiers, 
    the <x> and <y> components of the result contain the loaded values and 
    the <z>, and <w> components of the result are zero and one, respectively.
    When an image load uses the "F32X4", "S32X4", or "U32X4" storage 
    modifiers, all four components of the result contain the loaded values.
    If the image load is invalid for any of the reasons described in Section
    3.9.X, the result vector will be undefined.

    LOADIM supports no base data type modifiers, but requires exactly one
    storage modifier.  An image load is treated as invalid unless the storage
    modifier matches the image unit format, as described in Table X.3.  The 
    base data type of the result vector is derived from the storage modifier.
    The single operand is always interpreted as a signed integer vector.

        Data Type    Supported Modifers
        ---------    -------------------
          4x32       F32X4, S32X4, U32X4
          2x32       F32X2, S32X2, U32X2
          1x32       F32,   S32,   U32
          1x16              S16,   U16
          1x8               S8,    U8

      Table X.3, Supported Storage Modifiers.  Unformatted image operations
      are considered invalid unless the storage modifier is compatible with
      the "Data Type" entry for the image unit format, as described in Table
      X.2.


    Section 2.X.8.Z, STOREIM:  Image Store

    The STOREIM instruction takes the components of the second signed integer
    vector operand, uses them as coordinates to perform a formatted or
    unformatted image store to the texture bound to the image unit specified
    by <imageUnit> using the data specified in the first vector operand.  The
    store is performed in the manner described in Section 3.9.X.

    Eleven image targets are supported:  1D, 2D, 3D, RECT, CUBE, BUFFER, 
    ARRAY1D, ARRAY2D, ARRAYCUBE, 2DMS, and ARRAY2DMS.  The texel coordinate 
    is a one-, two- or three-dimensional vector, taken from the <x>, <y>, and
    <z> components of the operand.  For the 2DMS and ARRAY2DMS, the texel
    coordinate is a two- or three-dimensional vector, taken from the <x>, 
    <y>, and <z> components of the operand, and a sample number is taken from
    the <w> component of the operand.

        data = VectorLoad(op0);
        coords = VectorLoad(op1);
        if (target == 1D || target == BUFFER) {
          coords.y = 0;
        }
        if (target == 1D || target == 2D ||
            target == BUFFER || target == RECT || 
            target == 2DMS) {
          coords.z = 0;
        }
        if (target != 2DMS && target != ARRAY2DMS) {
          coords.w = 0;
        }
        ImageStore(image, coords, data);

    STOREIM supports an optional base data type or storage modifier.  If a
    storage modifier is specified, the store is unformatted; otherwise, it is
    formatted.  Formatted stores operate as described in Section 3.9.X. 
    Unformatted stores write the data to memory without converting to the 
    image unit format, by copying raw bits from the source variable to 
    memory according to the bit layouts described in Table X.2, where word 
    0 is taken from the <x> component, word 1 from <y>, etc..
    
    An unformatted image store is treated as invalid unless the
    storage modifier matches image unit format, as described in Table X.3.  
    When performing an unformatted store using the "S8", "U8", "S16", or 
    "U16" modifiers, all bits but the least significant eight or sixteen 
    are dropped as part of the store.  When performing a formatted store, 
    the first operand will be converted to the image unit format as part 
    of the store.

    The base data type of the first vector operand is derived from the data
    type or storage modifier.  The second operand is always interpreted as a
    signed integer vector.


    Section 2.X.8.Z, ATOMIM:  Image Atomic Memory Operation

    The ATOMIM instruction takes the components of the second signed integer
    vector operand, uses them as coordinates to perform an unformatted image
    load from the texture bound to the image unit specified by <imageUnit>,
    performs a computation using the loaded value and the first vector
    operand, performs an unformatted store of the result of the computation to
    the same texel, and then returns the loaded value in the vector result.
    The atomic operation is performed in the manner described in Section
    3.9.X.

    The ATOMIM instruction has two required instruction modifiers.  The atomic
    modifier specifies the type of computation to be performed.  The storage
    modifier specifies the size and data type of the operand read from the
    image unit and the base data type of the operation used to compute the
    value to be written back.

      atomic     storage
      modifier   modifiers   operation
      --------   ---------   --------------------------------------
       ADD       U32, S32    compute a sum
       MIN       U32, S32    compute minimum
       MAX       U32, S32    compute maximum
       IWRAP     U32         increment memory, wrapping at operand
       DWRAP     U32         decrement memory, wrapping at operand
       AND       U32, S32    compute bit-wise AND
       OR        U32, S32    compute bit-wise OR
       XOR       U32, S32    compute bit-wise XOR
       EXCH      U32, S32    exchange memory with operand
       CSWAP     U32, S32    compare-and-swap

     Table X.4, Supported atomic and storage modifiers for the ATOMIM
     instruction.

    Not all storage modifiers are supported by ATOMIM, and the set of
    modifiers allowed for any given instruction depends on the atomic modifier
    specified.  Table X.4 enumerates the set of atomic modifiers supported by
    the ATOMIM instruction, and the storage modifiers allowed for each.

        data = VectorLoad(op0);
        coords = VectorLoad(op1);
        if (target == 1D || target == BUFFER) {
          coords.y = 0;
        }
        if (target == 1D || target == 2D ||
            target == BUFFER || target == RECT || 
            target == 2DMS) {
          coords.z = 0;
        }
        if (target != 2DMS && target != ARRAY2DMS) {
          coords.w = 0;
        }
        result = ImageLoad(coords, data);
        switch (atomicModifier) {
        case ADD:
          writeval = tmp0.x + result;
          break;
        case MIN:
          writeval = min(tmp0.x, result);
          break;
        case MAX:
          writeval = max(tmp0.x, result);
          break;
        case IWRAP:
          writeval = (result >= tmp0.x) ? 0 : result+1; 
          break;
        case DWRAP:
          writeval = (result == 0 || result > tmp0.x) ? tmp0.x : result-1;
          break;
        case AND:
          writeval = tmp0.x & result;
          break;
        case OR:
          writeval = tmp0.x | result;
          break;
        case XOR:
          writeval = tmp0.x ^ result;
          break;
        case EXCH:
          break;
        case CSWAP:
          if (result == tmp0.x) {
            writeval = tmp0.y;
          } else {
            writeval = result;
          }
          break;
        }
        ImageStore(image, writeval);

    ATOMIM performs a scalar atomic operation.  The <y>, <z>, and <w>
    components of the result vector are undefined.
      
    ATOMIM supports no base data type modifiers, but requires exactly one
    storage and one atomic modifier.  An image atomic is treated as invalid
    unless the storage modifier matches the format of the texture bound to the
    image unit, as described in Table X.3.  The base data type of the result
    and the first operand is derived from the storage modifier.  The second
    operand is always interpreted as a signed integer vector.


    Section 2.X.8.Z, MEMBAR:  Memory Barrier

    The MEMBAR instruction synchronizes memory transactions to ensure that
    memory transactions resulting from any instruction executed by the thread
    prior to the MEMBAR instruction complete prior to any memory transactions
    issued after the instruction.

    MEMBAR has no operands and generates no result.

    Modify Section 3.9.X, Texture Image Loads and Stores, as added above.

    (Add a separate paragraph and table describing how the four-component
    coordinate vector used in image load, store, and atomic opcodes are mapped
    to individual texels.)

    When a program accesses the texture bound to an image unit using the
    LOADIM, STOREIM, or ATOMIM opcodes, it provides a four-component
    coordinate vector used to select individual texels or samples.  This
    (x,y,z,w) vector is used to select an individual texel tau_i, tau_i_j, or
    tau_i_j_k according to the target of the texture bound to the image unit
    using Table X.5.  As noted above, single-layer bindings of array or cube
    map textures are considered to use a texture target corresponding to the
    bound layer, rather than that of the full texture.

                                                   face/
                                          i  j  k  layer sample
                                          -- -- -- ----- ------
        TEXTURE_1D                        x  -  -    -     -
        TEXTURE_2D                        x  y  -    -     -
        TEXTURE_3D                        x  y  z    -     -
        TEXTURE_RECTANGLE                 x  y  -    -     -
        TEXTURE_CUBE_MAP                  x  y  -    z     -
        TEXTURE_BUFFER                    x  -  -    -     -
        TEXTURE_1D_ARRAY                  x  -  -    z     -
        TEXTURE_2D_ARRAY                  x  y  -    z     -
        TEXTURE_CUBE_MAP_ARRAY_ARB        x  y  -    z     -
        TEXTURE_2D_MULTISAMPLE            x  y  -    -     w
        TEXTURE_2D_MULTISAMPLE_ARRAY      x  y  -    z     w

        Table X.5, Mapping of image load, store, and atomic texel coordinate
        components to texel numbers.


Issues

    (1) How are the format and type of the load/store determined?

      RESOLVED:  There is a natural desire to load and store using a 
      canonical 4-vector in the shader with hardware converting to/from a 
      format compatible with the bound image, to be consistent with how
      texture loads and fragment shader outputs currently behave. There is
      also good reason to allow some flexibility in the format used for image
      accesses being different from the internal format of the texture level.
      We allow format conversions to and from any format that image units 
      support. We make the format be selected when the image is bound to an 
      image unit, and define which image unit formats can be used for which
      texture level internal formats. For example, it is legal to access an
      image whose internal format is RGBA8 with an image unit format of 
      R32UI.
      
    (2) What set of texture formats should be supported for image loads and
        stores?

      RESOLVED:  We allow textures to be bound to image units if and only if
      the implementation supports formatted stores for the texture format.
      Any texture formats not explicitly enumerated in this extension may not
      be bound to an image unit, although future extensions may add new
      formats to the set of supported formats.

      In particular, this extension supports one-, two-, and four-component
      textures with 8-, 16-, and 32-bit components, including floating-point,
      signed integer, unsigned integer, as well as signed and unsigned
      normalized formats.  Additionally, a small number of other formats are
      supported, including the 11/11/10 RGB format from EXT_packed_float and
      10/10/10/2 unsigned normalized RGBA.

    (3) Should we general support image loads and stores for three-component
        "RGB" formats?

      RESOLVED:  Not in this extension.  If an application needs to perform
      image loads and stores on a three-component texture, it could use an
      equivalent RGBA format and ignore the alpha component.  The
      EXT_texture_swizzle extension could be used to make the values returned
      by texture appear identical to an RGB texture, if required.

    (4) Should textures be unbound from image units when they are deleted?

      RESOLVED:  Yes, this matches behavior of existing bind points.

    (5) Should we support image loads and stores for the deprecated LUMINANCE,
        LUMINANCE_ALPHA, and ALPHA formats?

      RESOLVED:  No, only support the RGBA-style formats. EXT_texture_swizzle
      can be used to mimic luminance and alpha if required.

    (6) Should we support 64-bit atomics on images?  Should we support atomics
        at all on formats with 8-, 16-, 64-, or 128-bit texels?

      RESOLVED:  No, we will only support 32-bit atomic operations on images.

    (7) How do shader image loads and stores interact with texture
        completeness?  What happens if you bind a texture with inconsistent
        mipmaps?

      RESOLVED:  The image unit is treated as if nothing were bound, where 
      all accesses are treated as invalid.

    (8) What happens if the value passed to Uniform1i to specify the image
        unit corresponding to a image variable refers to a non-existent image
        unit (i.e., is negative or greater than or equal to the number of
        image units supported)?

      RESOLVED:  Values referring to invalid image units will be rejected and
      produce an INVALID_VALUE error.

    (9) Should we provide counting rules for image variable use in different
        shaders like we have for samplers?  In particular, there are limits
        on the amount of state, the number of active samplers in each shader
        stage, and the sum of the active sampler counts in each stage.

      RESOLVED:  No.  It was considered sufficient to have just a limit on the
      total number of image units in the implementation (i.e., the number of
      distinct values that the variable can be set to).

    (10) Can this extension be used to load and store values into a buffer
         object?  Into a renderbuffer?

      RESOLVED:  Yes, indirectly.  The BUFFER_TEXTURE target provided by
      OpenGL 3.0 and the EXT_texture_buffer_object extension allows an
      application to create a one-dimensional buffer texture using the data
      store of a buffer object. This buffer texture may be bound to an image
      unit and accessed with an imageBuffer variable in the Shading Language.

      This extension adds support for image accesses to multisample textures, 
      but not renderbuffers. Note that with the ARB_texture_multisample 
      extension, there is no longer a good reason to use renderbuffers. 
      Existing 2D or rectangle targets already provided a superset of single-
      sample renderbuffer functionality; the new ARB extension provides a 
      superset of multisample renderbuffer functionality.

    (11) What amount of automatic synchronization is provided for image loads
         and stores?  In particular, is the use of MemoryBarrierEXT() required
         to ensure consistent ordering relative to other GL operations?  Or is
         some other mechanism (e.g., unbinding a texture from an image unit
         and then binding it to a texture image unit) sufficient?

      RESOLVED:  Use of MemoryBarrierEXT is required, and there is no
      automatic synchronization when images are bound or unbound.

      Implicit synchronization is difficult, as it might require some
      combination of:

        - tracking which images might be written (randomly) in the shader
          itself;

        - assuming that if a shader that performs writes is executed, all
          texels of all bound images could be modified and thus must be
          treated as dirty;

        - idling at the end of each primitive or draw call, so that the
          results of all previous commands are complete.

      Since normal OpenGL operation is pipelined, idling would result in a
      significant performance impact since pipelining would otherwise allow
      fragment shader execution for draw call N while simultaneously
      performing vertex shader execution for draw call N+1.

    (12) Should image loads and stores be allowed for all shader types?

      RESOLVED:  Yes, it seems useful.

      Note that some shader types pose specific implementation complexities
      (e.g., reuse of vertices in vertex shaders, number of fragment shader
      invocations in multisample modes, relative order of execution within and
      between shader groups).  We have explicitly specify several cases where
      the invocation count and execution order are undefined.  While these
      cases may be a problem for some algorithms, we expect that many
      algorithms will not be adversely impacted.

    (13) Should an implementation be required to throw INVALID_OPERATION
         errors if the dimension of the texture coordinates implied by the
         image variable type doesn't match the structure of the texture
         level/layer bound to the corresponding image unit?  If not, what
         happens in such a mismatch?

      RESOLVED:  No.  The results of image accesses are undefined.

    (14) Should shader image variable types include a "format" implying the
         data type accepted/returned by shader image loads and stores?  For
         example, an image variable corresponding to a 2D texture with format
         of RGBA32F might have a type "image2Dvec4", with the "vec4"
         indicating that the image data lines up with a four-component
         floating-point vector.

      RESOLVED:  No.  Separate types are provided for float vs. int vs. 
      unsigned int, but not for each image format.

    (15) If shader image variable types include information on the texel
         components returned or written by shader image accesses, should an
         implementation be required to enforce errors if the variable type is
         incompatible with the format of the referenced texture?  If not, or
         if the image variable type doesn't include format information, what
         happens in case of a mismatch between the texture format and the
         shader access format?

      RESOLVED:  We aren't including types in the variable that correspond
      to the image format, so an error check in the driver is not possible.

      If an individual load, store, or atomic uses a data type incompatible
      with the texture bound to the image unit, loads will return and stores
      will write undefined values.

    (16) Is it possible to bind the "default texture" (numbered zero) for a
         given texture target to an image unit?

      RESOLVED:  No.  Passing zero to BindImageTexture unbinds and texture
      currently bound to the selected image unit.  If this ability were
      provided, it would also be necessary to provide some mechanism to
      specify a texture target because there is a separate default "zero"
      texture for each target.  

      Note that existing framebuffer objects have a similar behavior; default
      textures can't be attached to an FBO.

    (17) May bordered textures be used with image loads and stores?

      RESOLVED:  No.

    (18) Should we have defined behavior if invalid coordinates are passed to
         an image load, store, or atomic operation?  If so, what happens?

      RESOLVED:  Yes. We define the behavior to return zeroes on a load and 
      atomic and to have no effect on any bound texture on stores and 
      atomics.

    (19) Should we have a limit on the total number of combined image units
         and draw buffers, and if so, what should that be?

      RESOLVED:  Yes, some hardware requires this. The program will fail to 
      link.

    (20) What happens if a shader specifies an image store or atomic operation
         for killed/discarded pixels?

      RESOLVED:  For GLSL shaders that execute a "discard" instruction, any
      image stores or atomics performed before executing the discard will
      behave normally.  When the "discard" instruction is executed, the shader
      invocation will be terminated and will perform no further image store or
      atomic operations.

      For assembly shaders (NV_gpu_program5) that execute a "KIL" instruction,
      any image stores or atomics performed before executing the KIL will
      behave normally.  Unlike GLSL's "discard", the "KIL" instruction does
      not terminate program invocations.  However, any image store or atomic
      operations performed after the KIL instruction do not update memory, and
      the value returned by atomic operations is undefined.

    (21) When enabling early depth tests in a program, what happens if a
         fragment fails one of the tests (e.g., depth test)?

      RESOLVED:  The specification indicates that the fragment shader is not
      executed.  Implementations might still end up running fragment shader
      for implementation-dependent reasons.  For example, the fragment shader
      may be run in order to approximate derivatives for neighboring pixels
      that did pass all per-fragment tests.  In these cases, implementations
      must guarantee that image stores have no effect.

    (22) If implementations run fragment shaders for fragments that aren't
         covered by the primitive or fail early depth tests (e.g., "helper
         pixels"), how does that interact with stores and atomics?

      RESOLVED:  The current OpenGL specification has no formal notion of
      "helper" pixels.  In practice, implementations may run fragment shaders
      for pixels near the boundaries of rasterized primitives to allow
      derivatives to be approximated by differencing.  Typically, these shader
      invocations have no effect.  While they may produce outputs, the outputs
      for these pixels will be discarded without affecting the framebuffer.
      The spec basically treats these shader invocations as though they don't
      exist.

      If such a shader invocation performs store or atomic operations, we need
      to define what happens.  In our definition, stores will have no effect,
      atomics will not update memory, and the values returned by atomics will
      be undefined.  The fact that these invocations don't affect memory is
      consistent with the notion of helper pixel shader invocations not
      existing.

      However, it is possible to write a fragment shader where flow control
      depends on the (undefined) values returned by the atomic.  In this case,
      the undefined values returned for helper pixels could result in very
      long execution time (appearing to be hang) or an infinite loop.  To
      avoid hangs in such cases, it is possible to use the fragment shader
      input sample mask to identify helper pixels:

        // If the input sample mask is non-zero, at least one sample is
        // covered and the invocation should be treated as a real invocation.
        // If the sample mask is zero, nothing is covered and this should be
        // treated as a helper pixel.  If more than 32 samples are supported,
        // additional words of gl_SampleMaskIn would need to be checked.
        if (gl_SampleMaskIn[0] != 0)  {
          // "real" pixel, perform atomic operations
        } else {
          // "helper" pixel, skip atomics
        }

      It may be desirable to formalize the notion of helper pixels in a future
      addition to the shading language.

    (23) What API should we use to specify early depth tests?

      RESOLVED:  Use a layout qualifier in a fragment shader rather than
      having a separate program parameter or other piece of GL state.

    (24) For formatted loads where the format doesn't include some component,
         what values are filled in? (0,0,0,1)? (0,0,0,0)?
         
      RESOLVED: Prefer (0,0,0,1) to match other APIs.

    (25) How does the combined-image-and-fragment-output limit interact with
         separate shader objects?  For example, an application may want to
         share a single image unit between two shader stages and not have it
         count twice against the limit.

      RESOLVED:  The known implementations of this extension do not have this
      issue, so we chose not to include any spec language.  Perhaps a
      Begin-time error could be specified in the future if this limit is
      exceeded.

    (26) What sort of qualifiers should we provide relevant to memory
         referenced by image variables?

      RESOLVED:  We will support the qualifiers "coherent", "volatile",
      "restrict", and "const" to be used in image variable declarations.

      "coherent" is used to ensure that memory accesses from different shader
      invocations are cached coherently (i.e., one invocation will be able to
      observe writes from another when the other invocation's writes
      complete).  This coherence may mean the use of "coherent"-qualified
      image variables may perform more slowly than of otherwise equivalent
      unqualified variables.

      "volatile" behaves is as in C, and may be needed if an algorithm
      requires reading image memory that may be written asynchronously by
      other shader invocations.

      "restrict" behaves as in the C99 standard, and can be used to indicate
      that no other image variable points to the same underlying data.  This
      permits optimizations that would otherwise be impossible if the compiler
      has to assume that a pair of images might end up pointing to the same
      data.  For example, in standard C/C++, a loop like:

        int *a, *b;
        a[0] = b[0] + b[0];
        a[1] = b[0] + b[1];
        a[2] = b[0] + b[2];

      would need to reload b[0] for each assignment because a[0] or a[1] might
      point at the same data as b[0].  With restrict, the compiler can assume
      that b[0] is not modified by any of the instructions and load it just
      once.  The same considerations apply to accesses using imageLoad(),
      imageStore(), and imageAtomic*() builtins.

      "const" behaves as in C, and indicates that the image memory should be
      treated as read-only.  Note that the use of "const" in image variable
      declarations is different from the normal "const" qualifier, as it
      treats the image data referenced by the variable as constant.

    (27) How should shaders be able to express qualifiers for image variables?

      RESOLVED:  This extension borrows from C/C++ syntax rules where a
      qualifier may be specified before or after the type.  For example,

        layout(size4x32) const uniform image2D imageVariable;

      declare an image uniform whose image data are treated as read-only.  We
      permit qualifiers to be provided either before or after the type name
      (image2D).  The position of the qualifier is meaningful.  Qualifiers
      before the type name apply to the data referenced by the variable.
      Qualifiers after the type name apply to the variable itself.

      The closest C/C++ equivalent to the declarations above would turn
      declarations like:

        layout(size4x32) const uniform image2D firstImage;
        layout(size4x32) uniform image2D const secondImage;

      into:

        const struct image2D_data * firstImage;
        struct image2D_data * const secondImage;

      where "image2D" is replaced with "struct image2D_data *".  In this
      model, the former declares <firstImage> to be a pointer to constant
      image data.  The latter declares <secondImage> to be a constant pointer
      to non-constant image data.  

      For "coherent", "volatile", and "const", the qualifier should typically
      go before the image type.  For "restrict", the qualifier must go after
      the image type, since "restrict" applies to the pointer, not the data
      being pointed to.

      Note that a qualifier could theoretically be specified before and after
      the type name, such as:

        const image2D const imageVariable;

      which would declare <imageVariable> to be constant and to reference
      constant image data.  In this extension, declaring an image variable to
      be constant isn't meaningful, as such variables can never be used as
      l-values.

    (28) What is the meaning of "restrict" on a system that might run either
         multiple invocations of the same shader simultaneously, or multiple
         invocations of different shaders (vertex and fragment)
         simultaneously?

      RESOLVED:  When an image variable is qualified with "restrict", the only
      guarantee is that no other image variable in the same shader invocation
      references the same underlying image data.  There is no guarantee that
      the same image couldn't be referenced by another invocation of the same
      shader, or by an invocation of a different shader.

      The main function of "restrict" is to allow compilers to generate more
      efficient code for a single shader invocation than it could if it had to
      conservatively assume that accesses to other images could touch the same
      image data.

    (29) What is the purpose of the memoryBarrier() built-in function?

      RESOLVED:  The memoryBarrier() function can be used to ensure that if
      another shader invocation or other portions observe image memory being
      written by a shader, that accesses appear in a predictable order.  For
      example, consider the following code:

        uniform imageBuffer buf1;
        uniform imageBuffer buf2;
        int offset1, offset2;
        vec4 data1, data2;
        imageStore(buf1, offset1, data1);
        imageStore(buf2, offset2, data2);

      This specification doesn't require that writes be committed to memory in
      the order specified in the shader.  It is possible that another shader
      invocation or some other observer would see <data2> before seeing
      <data1>.  If an algorithm involved multiple shader invocations with one
      possibly needing to wait on data written by another, observing <data2>
      in the second shader would not ensure that <data1> has been written.
      However, if memoryBarrier() were used, as in the following code, the
      second shader would have such a guarantee.

        imageStore(buf1, offset1, data1);
        memoryBarrier();
        imageStore(buf2, offset2, data2);

    (30) What happens if the texel identified by the coordinates given to an
         image load, store, or atomic built-in doesn't exist?  (i.e.,
         coordinates are out of bounds)

      RESOLVED:  The results of image loads return zero.  Stores do not update
      image memory.  Atomics do not update image memory and return zero.
      These same considerations apply if no texture is bound to an image unit,
      the texture is incomplete, and various other conditions.  We do not ever
      apply wrap modes on image operations.

    (31) Why do we have a <format> parameter on BindImageTextureEXT?

      RESOLVED:  It allows some amount of bit-casting, to view a texture with
      one format using another format.  This parameter allows applications to
      work around several limitations of the specification:

        * Image loads do not support all formats supported for stores.  In
          particular, the only formats supported are 1x8, 1x16, 1x32, 2x32,
          and 4x32.  Using the <format> parameter allows an application to
          view an RGBA8 texture as "R32UI" and examine the component bits
          itself.

        * Image atomics are single-component 32-bit operations.  The ability
          to view some other formats as "size1x32" allows atomic operations to
          be done on some multi-component formats, such as RGBA8.

    (32) Do we support image atomics on multi-component texture formats?

      RESOLVED:  Only using the formats in the "size1x32" equivalence class,
      and then only as 32-bit scalar integer operations.  Atomics do not
      operate on a component-by-component basis in this extension.

    (33) What happens if early fragment testing is enabled, the early depth
         test passes, and a fragment shader that computes a new depth value is
         executed?

      RESOLVED:  The depth value produced by the fragment shader has no effect
      if early depth and stencil tests are enabled.  The depth value computed
      by a fragment shader is used only by the post-fragment shader stencil
      and depth tests, and those tests always have no effect when early
      fragment tests is enabled.

    (34) How do early fragment tests interact with occlusion queries?

      RESOLVED:  When early fragment tests are enabled, sample counting for
      occlusion queries also happens prior to fragment shader execution.
      Enabling early fragment tests can change the overall sample count,
      because samples killed by alpha test and alpha to coverage will still be
      counted if early fragment tests are enabled.

    (35) If we provide support for multiple active program objects (e.g., one
         containing a vertex shader, another containing a fragment shader, as
         in EXT_separate_shader_object), how will early fragment tests be
         handled?

      RESOLVED:  The early fragment test enable should be taken from the
      active program object corresponding to the fragment shader stage.

    (36) When specifying a coordinate vector to specify a texel for a
         TEXTURE_1D_ARRAY target, what coordinate is used to specify the
         layer?

      RESOLVED:  For GLSL functions, a two-component vector is specified and
      the second (y) component is used to select a layer.  When using the
      LOADIM, STOREIM, and ATOMIM NV_gpu_program5 assembly opcodes, a
      four-component vector is provided and the third (z) component selects
      the layer.

Revision History

    Rev.    Date    Author    Changes
    ----  --------  --------  -----------------------------------------
     7    10/16/13  pbrown    Update issue (20) to clarify that any image
                              stores and atomics issued before a "discard" do
                              have an effect.  Update issue (22) to better
                              define the behavior of stores and atomics on
                              "helper" pixels and to suggest a workaround for
                              shaders that need to use values returned by
                              atomics (undefined for helper pixels) in flow
                              control constructs.

     6    12/12/10  pbrown    Fix minor errata reported by spec reviewers
                              (bugs 6870 and 6991).

     5    09/17/10  pbrown    Clean up the spec language specifying the
                              mapping of coordinates to texels according to
                              the texture target.  For 1D arrays, GLSL wants
                              the layer in the second component of a
                              two-component vector while NV_gpu_program5 wants
                              it in the third component of a four-component
                              vector.  Also clarify that single-layer bindings
                              of an array or cube map texture use a target
                              appropriate to the bound layer.
                              
     4    03/23/10  pbrown    Add interaction with EXT_separate_shader_objects.
                              Update issues section to include some issues
                              left behind in NV_gpu_shader5 when specs were
                              refactored.

     3    03/21/10  pbrown    Update spec overview, interactions, and issues 
                              sections; miscellaneous minor clarifications.

     2    03/16/10  pbrown    Add a separate #extension line for this
                              extension; needed since the became packaged
                              separately from ARB_gpu_shader5.  Added C99-like
                              "restrict" qualifier to indicate that an image
                              variable won't share underlying image contents
                              with any other variable.  Added support for
                              "const" qualifiers on images to allow indicate
                              read-only image data.  Added language describing
                              the significance of the position of image
                              variable qualifiers.  Clarified rules on use of
                              image variables as function parameters; adding
                              qualifiers is OK, stripping them off is not.
                              Updated image layout qualifier section to
                              clarify that "size" layout qualifiers are
                              required on both uniform and function parameter
                              declarations.  Added "const" qualifier on the
                              image argument in imageLoad() prototypes.
                              Updated extension names in dependency sections.
                              Add support for stores to the RGB10_A2 texture
                              format from OpenGL 3.3.  Add several issues.

     1              jbolz     Internal revisions.
