Name

    NV_shader_buffer_store

Name Strings

    none (implied by GL_NV_gpu_program5 or GL_NV_gpu_shader5)

Contact

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

Status

    Shipping.

Version

    Last Modified Date:         August 13, 2012
    NVIDIA Revision:            5

Number

    390

Dependencies

    OpenGL 3.0 and GLSL 1.30 are required.

    This extension is written against the OpenGL 3.2 (Compatibility Profile)
    specification, dated July 24, 2009.

    This extension is written against version 1.50.09 of the OpenGL Shading
    Language Specification.

    OpenGL 3.0 and GLSL 1.30 are required.

    NV_shader_buffer_load is required.

    NV_gpu_program5 and/or NV_gpu_shader5 is required.

    This extension interacts with EXT_shader_image_load_store.

    This extension interacts with NV_gpu_shader5.

    This extension interacts with NV_gpu_program5.

    This extension interacts with GLSL 4.30, ARB_shader_storage_buffer_object, 
    and ARB_compute_shader.

Overview

    This extension builds upon the mechanisms added by the
    NV_shader_buffer_load extension to allow shaders to perform random-access
    reads to buffer object memory without using dedicated buffer object
    binding points.  Instead, it allowed an application to make a buffer
    object resident, query a GPU address (pointer) for the buffer object, and
    then use that address as a pointer in shader code.  This approach allows
    shaders to access a large number of buffer objects without needing to
    repeatedly bind buffers to a limited number of fixed-functionality binding
    points.

    This extension lifts the restriction from NV_shader_buffer_load that
    disallows writes.  In particular, the MakeBufferResidentNV function now
    allows READ_WRITE and WRITE_ONLY access modes, and the shading language is
    extended to allow shaders to write through (GPU address) pointers.
    Additionally, the extension provides built-in functions to perform atomic
    memory transactions to buffer object memory.

    As with the shader writes provided by the EXT_shader_image_load_store
    extension, writes to buffer object memory using this extension are weakly
    ordered to allow for parallel or distributed shader execution.  The
    EXT_shader_image_load_store extension provides mechanisms allowing for
    finer control of memory transaction order, and those mechanisms apply
    equally to buffer object stores using this extension.


New Procedures and Functions

    None.

New Tokens

    Accepted by the <barriers> parameter of MemoryBarrierNV:

        SHADER_GLOBAL_ACCESS_BARRIER_BIT_NV             0x00000010

    Accepted by the <access> parameter of MakeBufferResidentNV:

        READ_WRITE
        WRITE_ONLY


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

    Modify Section 2.9, Buffer Objects, p. 46

    (extend the language inserted by NV_shader_buffer_load in its "Append to
     Section 2.9 (p. 45) to allow READ_WRITE and WRITE_ONLY mappings)

    The data store of a buffer object may be made accessible to the GL 
    via shader buffer loads and stores by calling:

        void MakeBufferResidentNV(enum target, enum access);

    <access> may be READ_ONLY, READ_WRITE, and WRITE_ONLY.  If a shader loads
    from a buffer with WRITE_ONLY <access> or stores to a buffer with
    READ_ONLY <access>, the results of that shader operation are undefined and
    may lead to application termination.  <target> may be any of the buffer
    targets accepted by BindBuffer.

    The data store of a buffer object may be made inaccessible to the GL
    via shader buffer loads and stores by calling:
    
        void MakeBufferNonResidentNV(enum target);


    Modify "Section 2.20.X, Shader Memory Access" introduced by the
    NV_shader_buffer_load specification, to reflect that shaders may store to
    buffer object memory.

    (first paragraph) Shaders may load from or store to buffer object memory
    by dereferencing pointer variables.  ...

    (second paragraph) When a shader dereferences a pointer variable, data are
    read from or written to buffer object memory according to the following
    rules:

    (modify the paragraph after the end of the alignment and stride rules,
    allowing for writes, and also providing rules forbidding reads to
    WRITE_ONLY mappings or vice-versa) If a shader reads or writes to a GPU
    memory address that does not correspond to a buffer object made resident
    by MakeBufferResidentNV, the results of the operation are undefined and
    may result in application termination.  If a shader reads from a buffer
    object made resident with an <access> parameter of WRITE_ONLY, or writes
    to a buffer object made resident with an <access> parameter of READ_ONLY,
    the results of the operation are also undefined and may lead to
    application termination.

    Incorporate the contents of "Section 2.14.X, Shader Memory Access" from
    the EXT_shader_image_load_store specification into the same "Shader memory
    Access", with the following edits.

    (modify first paragraph to reference pointers) Shaders may perform
    random-access reads and writes to texture or buffer object memory using
    pointers or with built-in image load, store, and atomic functions, as
    described in the OpenGL Shading Language Specification.  ...

    (add to list of bits in <barriers> in MemoryBarrierNV)

    - SHADER_GLOBAL_ACCESS_BARRIER_BIT_NV:  Memory accesses using pointers and
        assembly program global loads, stores, and atomics issued after the
        barrier will reflect data written by shaders prior to the barrier.
        Additionally, memory writes using pointers issued after the barrier
        will not execute until memory accesses (loads, stores, texture
        fetches, vertex fetches, etc) initiated prior to the barrier complete.

    (modify second paragraph after the list of <barriers> bits) To allow for
    independent shader threads to communicate by reads and writes to a common
    memory address, pointers and image variables in the OpenGL shading
    language may be declared as "coherent".  Buffer object or texture memory
    accessed through such variables may be cached only if...

    (add to the coherency guidelines)

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


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

    None.


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

    None.


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

    None.


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

    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

    None.
    

Additions to the OpenGL Shading Language Specification, Version 1.50 (Revision
09)

    Modify Section 4.3.X, Memory Access Qualifiers, as added by
    EXT_shader_image_load_store

    (modify second paragraph) Memory accesses to image and pointer variables
    declared using the "coherent" storage qualifier are performed coherently
    with similar accesses from other shader threads.  ...

    (modify fourth paragraph) Memory accesses to image and pointer 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 thread.  ...

    (modify fifth paragraph) Memory accesses to image and pointer 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.  ...

    (modify sixth paragraph) Memory accesses to image and pointer variables
    declared using the "const" storage qualifier may only read the underlying
    memory, which is treated as read-only.  ...

    (insert after seventh paragraph) 

    In pointer variable declarations, the "coherent", "volatile", "restrict",
    and "const" qualifiers can be positioned anywhere in the declaration, and
    may apply qualify either a pointer or the underlying data being pointed
    to, depending on its position in the declaration.  Each qualifier to the
    right of the basic data type in a declaration is considered to apply to
    whatever type is found immediately to its left; qualifiers to the left of
    the basic type are considered to apply to that basic type.  To interpret
    the meaning of qualifiers in pointer declarations, it is useful to read
    the declaration from right to left as in the following examples.

      int * * const a;     // a is a constant pointer to a pointer to int
      int * volatile * b;  // b is a pointer to a volatile pointer to int
      int const * * c;     // c is a pointer to a pointer to a constant int
      const int * * d;     // d is like c
      int const * const *  // e is a constant pointer to a constant pointer
       const e;            //   to a constant int

    For pointer types, the "restrict" qualifier can be used to qualify
    pointers, but not non-pointer types being pointed to.

      int * restrict a;    // a is a restricted pointer to int
      int restrict * b;    // b qualifies "int" as restricted - illegal

    (modify eighth paragraph) The "coherent", "volatile", and "restrict"
    storage qualifiers may only be used on image and pointer variables, and
    may not be used on variables of any other type.  ...

    (modify last paragraph) The values of image and pointer variables
    qualified with "coherent," "volatile," "restrict", or "const" may not be
    assigned to function parameters or l-values lacking such qualifiers.

    (add examples for the last paragraph)

      int volatile * var1;
      int * var2;
      int * restrict var3;
      var1 = var2;              // OK, adding "volatile" is allowed
      var2 = var3;              // illegal, stripping "restrict" is not


    Modify Section 5.X, Pointer Operations, as added by NV_shader_buffer_load

    (modify second paragraph, allowing storing through pointers) The pointer
    dereference operator ...  The result of a pointer dereference may be used
    as the left-hand side of an assignment.


    Modify Section 8.Y, Shader Memory Functions, as added by
    EXT_shader_image_load_store

    (modify first paragraph) Shaders of all types may read and write the
    contents of textures and buffer objects using pointers and image
    variables.  ...

    (modify description of memoryBarrier) memoryBarrier() can be used to
    control the ordering of memory transactions issued by a shader thread.
    When called, it will wait on the completion of all memory accesses
    resulting from the use of pointers and image variables prior to calling
    the function.  ...

    (add the following paragraphs to the end of the section)

    If multiple threads need to atomically access shared memory addresses
    using pointers, they may do so using the following built-in functions.
    The following atomic memory access functions allow a shader thread to
    read, modify, and write an address in memory in a manner that guarantees
    that no other shader thread can modify the memory between the read and the
    write.  All of these functions read a single data element from memory,
    compute a new value based on the value read from memory and one or more
    other values passed to the function, and writes the result back to the
    same memory address.  The value returned to the caller is always the data
    element originally read from memory.

    Syntax:

      uint      atomicAdd(uint *address, uint data);
      int       atomicAdd(int *address, int data);
      uint64_t  atomicAdd(uint64_t *address,  uint64_t data);

      uint      atomicMin(uint *address, uint data);
      int       atomicMin(int *address, int data);

      uint      atomicMax(uint *address, uint data);
      int       atomicMax(int *address, int data);

      uint      atomicIncWrap(uint *address, uint wrap);

      uint      atomicDecWrap(uint *address, uint wrap);

      uint      atomicAnd(uint *address, uint data);
      int       atomicAnd(int *address, int data);

      uint      atomicOr(uint *address, uint data);
      int       atomicOr(int *address, int data);

      uint      atomicXor(uint *address, uint data);
      int       atomicXor(int *address, int data);

      uint      atomicExchange(uint *address, uint data);
      int       atomicExchange(int *address, uint data);
      uint64_t  atomicExchange(uint64_t *address, uint64_t data);

      uint      atomicCompSwap(uint *address, uint compare, uint data);
      int       atomicCompSwap(int *address, int compare, int data);
      uint64_t  atomicCompSwap(uint64_t *address, uint64_t compare, 
                               uint64_t data);

    Description:

    atomicAdd() computes the new value written to <address> by adding the
    value of <data> to the contents of <address>.  This function supports 32-
    and 64-bit unsigned integer operands, and 32-bit signed integer operands.

    atomicMin() computes the new value written to <address> by taking the
    minimum of the value of <data> and the contents of <address>.  This
    function supports 32-bit signed and unsigned integer operands.

    atomicMax() computes the new value written to <address> by taking the
    maximum of the value of <data> and the contents of <address>.  This
    function supports 32-bit signed and unsigned integer operands.

    atomicIncWrap() computes the new value written to <address> by adding one
    to the contents of <address>, and then forcing the result to zero if and
    only if the incremented value is greater than or equal to <wrap>.  This
    function supports only 32-bit unsigned integer operands.

    atomicDecWrap() computes the new value written to <address> by subtracting
    one from the contents of <address>, and then forcing the result to
    <wrap>-1 if the original value read from <address> was either zero or
    greater than <wrap>.  This function supports only 32-bit unsigned integer
    operands.

    atomicAnd() computes the new value written to <address> by performing a
    bitwise and of the value of <data> and the contents of <address>.  This
    function supports 32-bit signed and unsigned integer operands.

    atomicOr() computes the new value written to <address> by performing a
    bitwise or of the value of <data> and the contents of <address>.  This
    function supports 32-bit signed and unsigned integer operands.

    atomicXor() computes the new value written to <address> by performing a
    bitwise exclusive or of the value of <data> and the contents of <address>.
    This function supports 32-bit signed and unsigned integer operands.

    atomicExchange() uses the value of <data> as the value written to
    <address>.  This function supports 32- and 64-bit unsigned integer
    operands and 32-bit signed integer operands.

    atomicCompSwap() compares the value of <compare> and the contents of
    <address>.  If the values are equal, <data> is written to <address>;
    otherwise, the original contents of <address> are preserved.  This
    function supports 32- and 64-bit unsigned integer operands and 32-bit
    signed integer operands.


    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 and the "*"
        characters indicating pointers in a variable declaration.
 

Dependencies on EXT_shader_image_load_store

    This specification incorporates the memory access ordering and
    synchronization discussion from EXT_shader_image_load_store verbatim.  

    If EXT_shader_image_load_store is not supported, this spec should be
    construed to introduce:

      * the shader memory access language from that specification, including
        the MemoryBarrierNV() command and the tokens accepted by <barriers>
        from that specification;

      * the memoryBarrier() function to the OpenGL shading language
        specification; and

      * the capability and spec language allowing applications to enable early
        depth tests.

Dependencies on NV_gpu_shader5

    This specification requires either NV_gpu_shader5 or NV_gpu_program5.  

    If NV_gpu_shader5 is supported, use of the new shading language features
    described in this extension requires 

      #extension GL_NV_gpu_shader5 : enable

    If NV_gpu_shader5 is not supported, modifications to the OpenGL Shading
    Language Specification should be removed.

Dependencies on NV_gpu_program5

    If NV_gpu_program5 is supported, the extension provides support for stores
    and atomic memory transactions to buffer object memory.  Stores are
    provided by the STORE opcode; atomics are provided by the ATOM opcode.  No
    "OPTION" line is required for these features, which are implied by
    NV_gpu_program5 program headers such as "!!NVfp5.0".  The operation of
    these opcodes is described in the NV_gpu_program5 extension specification.

    Note also that NV_gpu_program5 also supports the LOAD opcode originally
    added by the NV_shader_buffer_load and the MEMBAR opcode originally
    provided by EXT_shader_image_load_store.


Dependencies on GLSL 4.30, ARB_shader_storage_buffer_object, and
ARB_compute_shader

    If GLSL 4.30 is supported, add the following atomic memory functions to
    section 8.11 (Atomic Memory Functions) of the GLSL 4.30 specification:

      uint atomicIncWrap(inout uint mem, uint wrap);
      uint atomicDecWrap(inout uint mem, uint wrap);

    with the following documentation

      atomicIncWrap() computes the new value written to <mem> by adding one to
      the contents of <mem>, and then forcing the result to zero if and only
      if the incremented value is greater than or equal to <wrap>.  This
      function supports only 32-bit unsigned integer operands.

      atomicDecWrap() computes the new value written to <mem> by subtracting
      one from the contents of <mem>, and then forcing the result to <wrap>-1
      if the original value read from <mem> was either zero or greater than
      <wrap>.  This function supports only 32-bit unsigned integer operands.

    Additionally, add the following functions to the section:

      uint64_t atomicAdd(inout uint64_t mem, uint data);
      uint64_t atomicExchange(inout uint64_t mem, uint data);
      uint64_t atomicCompSwap(inout uint64_t mem, uint64_t compare, 
                              uint64_t data);

    If ARB_shader_storage_buffer_object or ARB_compute_shader are supported,
    make similar edits to the functions documented in the
    ARB_shader_storage_buffer object extension.

    These functions are available if and only if GL_NV_gpu_shader5 is enabled
    via the "#extension" directive.


Errors

    None

New State

    None.

Issues

    (1) Does MAX_SHADER_BUFFER_ADDRESS_NV still apply?

      RESOLVED:  The primary reason for this limitation to exist was the lack
      of 64-bit integer support in shaders (see issue 15 of 
      NV_shader_buffer_load). Given that this extension is being released at 
      the same time as NV_gpu_shader5 which adds 64-bit integer support, it 
      is expected that this maximum address will match the maximum address
      supported by the GPU's address space, or will be equal to "~0ULL" 
      indicating that any GPU address returned by the GL will be usable in a
      shader.

    (2) What qualifiers should be supported on pointer variables, and how can
        they be used in declarations?

      RESOLVED:  We will support the qualifiers "coherent", "volatile",
      "restrict", and "const" to be used in pointer declarations.  "coherent"
      is taken from EXT_shader_image_load_store and is used to ensure that
      memory accesses from different shader threads are cached coherently
      (i.e., will be able to see each other when complete).  "volatile" and
      "const" behave is as in C.

      "restrict" behaves as in the C99 standard, and can be used to indicate
      that no other pointer points to the same underlying data.  This permits
      optimizations that would otherwise be impossible if the compiler has to
      assume that a pair of pointers 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.

    (3) What amount of automatic synchronization is provided for buffer object
        writes through pointers?

      RESOLVED:  Use of MemoryBarrierEXT() is required, and there is no
      automatic synchronization when buffers are bound or unbound.  With
      resident buffers, there are no well-defined binding points in the first
      place -- all resident buffers are effectively "bound".

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

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

        - assuming that if a shader that performs writes is executed, all
          bytes of all resident buffers 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.


Revision History

    Rev.    Date    Author    Changes
    ----  --------  --------  -----------------------------------------
     5    08/13/12  pbrown    Add interaction with OpenGL 4.3 (and related ARB
                              extensions) supporting atomic{Inc,Dec}Wrap and 
                              64-bit unsigned integer atomics to shared and
                              shader storage buffer memory. 

     4    04/13/10  pbrown    Remove the floating-point version of atomicAdd(). 

     3    03/23/10  pbrown    Minor cleanups to the dependency sections.
                              Fixed obsolete extension names.  Add an issue
                              on synchronization.

     2    03/16/10  pbrown    Updated memory access qualifiers section
                              (volatile, coherent, restrict, const) for
                              pointers.  Added language to document how
                              these qualifiers work in possibly complicated
                              expression.

     1              pbrown    Internal revisions.