xref: /third_party/openGLES/extensions/AMD/AMD_query_buffer_object.txt

Name

    AMD_query_buffer_object

Name Strings

    GL_AMD_query_buffer_object

Contact

    Daniel Rakos (daniel.rakos 'at' amd.com)

Contributors

    Daniel Rakos, AMD
    Graham Sellers, AMD
    Christophe Riccio, AMD

Status

    Shipping in Catalyst 12.6

Version

    Last Modified Date:         03/07/2016
    Author Revision:            6

Number

    420

Dependencies

    OpenGL 1.5 is required.

    This extension is written against the OpenGL 4.2 (core) specification.

Overview

    Statistics about the operation of the OpenGL pipeline, such as the number
    of samples that passed the depth test, the elapsed time between two events
    or the number of vertices written to a transform feedback buffer may
    be retrieved from the GL through query objects. The current value of a
    query object may be retrieved by the application through the OpenGL API.
    Should the result returned by the API be required for use in a shader,
    it must be passed back to the GL via a program uniform or some other
    mechanism. This requires a round-trip from the GPU to the CPU and back.

    This extension introduces a mechanism whereby the current value of a query
    result may be retrieved into a buffer object instead of client memory.
    This allows the query result to be made available to a shader without a
    round-trip to the CPU for example by subsequently using the buffer object
    as a uniform buffer, texture buffer or other data store visible to the
    shader. This functionality may also be used to place the results of
    many query objects into a single, large buffer and then map or otherwise
    read back the entire buffer at a later point in time, avoiding a per-query
    CPU-GPU synchronization event.

    The result of any query object type supported by the GL implementation
    may be retrieved into a buffer object. The implementation will determine
    the most efficient method of copying the query result to the buffer.

New Procedures and Functions

    None.

New Tokens

    Accepted by the <pname> parameter of GetQueryObjectiv, GetQueryObjectuiv,
    GetQueryObjecti64v and GetQueryObjectui64v:

        QUERY_RESULT_NO_WAIT_AMD                        0x9194

    Accepted by the <target> parameters of BindBuffer, BufferData,
    BufferSubData, MapBuffer, UnmapBuffer, MapBufferRange, GetBufferSubData,
    GetBufferParameteriv
    and GetBufferPointerv:

        QUERY_BUFFER_AMD                                0x9192

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

        QUERY_BUFFER_BINDING_AMD                        0x9193

Additions to Chapter 2 of the OpenGL 4.2 (core) Specification (OpenGL Operation)

    Modify Section 2.9, Buffer Objects

    Add to Table 2.8: Buffer object binding targets (p. 43)

    +---------------------+--------------------------+-------------------------+
    | Target name         | Purpose                  | Described in section(s) |
    +---------------------+--------------------------+-------------------------+
    | QUERY_BUFFER_AMD    | Query result buffer      | 6.1.7                   |
    +---------------------+--------------------------+-------------------------+

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

    None.

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

    None.

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

    None.

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

  Modify Section 6.1.7, Asynchronous Queries

  Add new paragraph before the third paragraph on p. 358

        Initially, zero is bound for the QUERY_BUFFER_AMD, indicating that
    <params> is a pointer into client memory. However, if a non-zero buffer
    object is bound as the current query result buffer, then <params> is
    treated as an offset into the designated buffer object.

  Replace last sentence of the third paragraph on p. 358
  beginning with "The state of a query object can be queried with ..."

    pname must be QUERY_RESULT, QUERY_RESULT_NO_WAIT_AMD
    or QUERY_RESULT_AVAILABLE.

  Add new paragraph before the last paragraph on p. 358

        If <pname> is QUERY_RESULT_NO_WAIT_AMD, then the query object's
    result value is returned as a single integer in <params> if the result
    is available at the time of the state query. If the result is not
    available then the destination memory location is not overwritten.

Additions to the AGL/GLX/WGL Specifications

    None.

GLX Protocol

    None.

Errors

    INVALID_OPERATION is generated by GetQueryObjectiv, GetQueryObjectuiv,
    GetQueryObjecti64v, or GetQueryObjectui64v if the command would cause data
    to be written beyond the bounds of the buffer currently currently bound
    to the QUERY_BUFFER_AMD target.

New State

    Append to Table 6.41, "Query Object State"

    +----------------------------+-------+-------------------+---------------+-------------------------------------------+-------+
    | Get Value                  | Type  | Get Command       | Initial Value | Description                               | Sec.  |
    +----------------------------+-------+-------------------+---------------+-------------------------------------------+-------+
    | QUERY_BUFFER_BINDING_AMD   | Z+    | GetIntegeriv      | 0             | Query result buffer binding.              | 6.1.7 |
    +----------------------------+-------+-------------------+---------------+-------------------------------------------+-------+

New Implementation Dependent State

    None.

Usage Examples

    Convenient macro definition for specifying buffer offsets:

        #define BUFFER_OFFSET(i)    ((void*)NULL + (i))

    Example 1: Using occlusion query result in shader

        // Create a buffer object for the query result
        glGenBuffers(1, &queryBuffer);
        glBindBuffer(GL_QUERY_BUFFER_AMD, queryBuffer);
        glBufferData(GL_QUERY_BUFFER_AMD, sizeof(GLuint),
                     NULL, GL_DYNAMIC_COPY);

        // Perform occlusion query
        glBeginQuery(GL_SAMPLES_PASSED, queryId)
        ...
        glEndQuery(GL_SAMPLES_PASSED);

        // Get query results to buffer object
        glBindBuffer(GL_QUERY_BUFFER_AMD, queryBuffer);
        glGetQueryObjectuiv(queryId, GL_QUERY_RESULT, BUFFER_OFFSET(0));

        // Bind query result buffer as uniform buffer
        glBindBufferBase(GL_UNIFORM_BUFFER, 0, queryBuffer);
        ...

        --- Shader ---

        ...
        uniform queryResult {
            uint samplesPassed;
        };
        ...
        void main() {
            ...
            if (samplesPassed > threshold) {
                // complex processing
                ...
            } else {
                // simplified processing
                ...
            }
            ...
        }

    Example 2: Using occlusion query result in shader only if result is available

        // Create a buffer object for the query result
        glGenBuffers(1, &queryBuffer);
        glBindBuffer(GL_QUERY_BUFFER_AMD, queryBuffer);
        glBufferData(GL_QUERY_BUFFER_AMD, 2 * sizeof(GLuint),
                     NULL, GL_DYNAMIC_COPY);

        // Perform occlusion query
        glBeginQuery(GL_SAMPLES_PASSED, queryId)
        ...
        glEndQuery(GL_SAMPLES_PASSED);

        // Get query availability and result (if available) to buffer object
        glBindBuffer(GL_QUERY_BUFFER_AMD, queryBuffer);
        glGetQueryObjectuiv(queryId, GL_QUERY_RESULT_AVAILABLE, BUFFER_OFFSET(0));
        glGetQueryObjectuiv(queryId, GL_QUERY_RESULT_NO_WAIT_AMD, BUFFER_OFFSET(4));

        // Bind query result buffer as uniform buffer
        glBindBufferBase(GL_UNIFORM_BUFFER, 0, queryBuffer);
        ...

        --- Shader ---
        
        ...
        uniform queryResult {
            uint resultAvailable;
            uint samplesPassed;
        };
        ...
        void main() {
            ...
            if (resultAvailable) {
                if (samplesPassed > threshold) {
                    // complex processing
                    ...
                } else {
                    // simplified processing
                    ...
                }
            } else {
                // default processing if no query result is available
                ...
            }
            ...
        }
        
    Example 3: Using a default value and QUERY_RESULT_NO_WAIT_AMD

        // Create a buffer object for the query result        
        // Store a default value in the buffer that will be used
        // if the query results are not available
        glGenBuffers(1, &queryBuffer);
        glBindBuffer(GL_QUERY_BUFFER_AMD, queryBuffer);
        GLuint defaultValue = 42;
        glBufferData(GL_QUERY_BUFFER_AMD, sizeof(GLuint),
                     &defaultValue, GL_DYNAMIC_COPY);
        
        // Perform occlusion query
        glBeginQuery(GL_SAMPLES_PASSED, queryId)
        ...
        glEndQuery(GL_SAMPLES_PASSED);

        // Get query results to buffer object with no wait
        // Default value remains untouched if results are not available
        glBindBuffer(GL_QUERY_BUFFER_AMD, queryBuffer);
        glGetQueryObjectuiv(queryId, GL_QUERY_RESULT_NO_WAIT_AMD, BUFFER_OFFSET(0));
        ...

    Example 4: Using transform feedback query result to fill indirect draw buffer

        // Create a buffer object for the indirect draw command        
        glGenBuffers(1, &drawIndirectBuffer);

        // Initialize draw command
        DrawArraysIndirectCommand cmd = { ... };
        glBindBuffer(GL_DRAW_INDIRECT_BUFFER, drawIndirectBuffer);
        glBufferData(GL_DRAW_INDIRECT_BUFFER, sizeof(DrawArraysIndirectCommand),
                     &cmd, GL_DYNAMIC_COPY);
        
        // Perform transform feedback query
        glBeginQuery(GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN, queryId)
        ...
        glEndQuery(GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN);

        // Write query result to the primCount field of the indirect draw command
        glBindBuffer(GL_QUERY_BUFFER_AMD, drawIndirectBuffer);
        glGetQueryObjectuiv(queryId, GL_QUERY_RESULT,
                            BUFFER_OFFSET(offsetof(DrawArraysIndirectCommand, primCount)));

        // Execute the indirect draw command
        glDrawArraysIndirect(GL_TRIANGLES, BUFFER_OFFSET(0));
        ...

Issues

    1) What is QUERY_RESULT_NO_WAIT_AMD useful for?

    RESOLVED: The application may decide that it does not want to wait for the
    result of the query object if it's not available at the time. This is not
    a problem without this extension as the application can always query the
    availability using QUERY_RESULT_AVAILABLE and decide to actually get the
    results using QUERY_RESULT only if the result is available. However, when
    using query buffers, QUERY_RESULT_AVAILABLE and QUERY_RESULT alone cannot
    provide the same flexibility for shader based decision making as the
    results are either always available (if QUERY_RESULT was used) or never.
    QUERY_RESULT_NO_WAIT_AMD provides a way to query the result only if it's
    available. Combined with QUERY_RESULT_AVAILABLE, the shader can decide
    to use the result or not based on the availability (see usage example 2).

    2) Should QUERY_RESULT_NO_WAIT_AMD be accepted by GetQueryObjectiv,
    GetQueryObjectuiv, GetQueryObjecti64v and GetQueryObjectui64v in case
    there is no buffer bound to QUERY_BUFFER_AMD?

    RESOLVED: YES, for completeness.

    3) Is there any guarantee that GetQueryObject* will not wait for the
    query results to become available if <pname> is QUERY_RESULT_NO_WAIT_AMD?

    RESOLVED: There is no need to have such guarantee. An implementation may
    choose to wait for the results even in case of QUERY_RESULT_NO_WAIT_AMD,
    however, that may incur a potential performance hit in case the application
    expects it to not wait.

    4) Should the INVALID_OPERATION error be generated if a GetQueryObject*
    command would access data outside the rage of the bound query buffer?

    RESOLVED: YES. This requires considering the value of <params> and the
    size of the type of the written value to determine the maximum addressed
    byte for the state query command.

    Note: This follows the precedence of the language introduced by
    ARB_pixel_buffer_object.

    5) Should we support 64 bit versions of the state query commands
    (GetQueryObjecti64v and GetQueryObjectui64v)?

    RESOLVED: YES, both for completeness and to support timer queries. In this
    case a 64 bit integer value is written to the buffer.

    6) Should the extension support all query types supported by the current
    GL implementation?

    RESOLVED: YES. There is not much value in providing additional capability
    queries that would allow the application to find out which query object
    types are supported. Also, the GL implementation can always choose to
    implement the functionality in software for query types that cannot be
    supported in hardware.

    7) Is there any precedence that the pointer parameter of a glGet* command
    is treated as an offset into a bound buffer object?

    RESOLVED: YES, glGetTexImage accepts an offset into the pixel pack buffer
    in case a pixel pack buffer is bound. As pixel buffer objects are part of
    the core specification since version 2.1, no precedence is introduced by
    this extension.
    
    8) What should be written to the query buffer when QUERY_RESULT_NO_WAIT_AMD
    is used but the results are not available?
    
    DISCUSSION: Leaving the written value undefined is an option, however
    in many cases the application can benefit from having a default value in
    the query buffer if the results are not available. The following options
    were considered to support such use cases:

        a. Write a predefined fixed value to the buffer.
        b. Write a user specified value to the buffer.
        c. Don't write anything to the buffer, leave the current value
           untouched.

    The problem with option a. is that it may be difficult to select such
    a predefined value that would not potentially conflict with a valid
    value. Option b. could be fine, however it requires new API to specify
    this default value. Thus, option c. is considered to most preferable.

    RESOLVED: Nothing, the current value in the specified offset of the
    query buffer is leaved untouched.


Revision History

    Rev.    Date      Author    Changes
    ----  --------    --------  ---------------------------------------------

      6   03/07/2016  drakos    Fixed typo in example #4
      5   06/01/2012  gsellers  Minor cleanup. Shipped and ready for posting.
      4   03/21/2012  drakos    Removed undefined behavior when using
                                QUERY_RESULT_NO_WAIT and added issue #8
                                based on Graham's comments
      3   03/12/2012  drakos    Assigned enums
      2   03/07/2012  drakos    Clarified issues #4 and #5
                                Added issue #7 based on Christophe's comments
      1   02/28/2012  drakos    Initial revision