Name

    NV_cuda_event

Name Strings

    EGL_NV_cuda_event

Contributors

    Debalina Bhattacharjee
    Michael Chock
    James Jones

Contact

    Michael Chock (mchock 'at' nvidia.com)

Status

    Complete

Version

    Version 1, June 20, 2014

Number

    EGL Extension #75

Extension Type

    EGL display extension

Dependencies

    This extension is written against the language of EGL 1.5.

    Either EGL_KHR_fence_sync and the EGLAttrib type or EGL 1.5 are
    required.

    This extension interacts with EGL_NV_device_cuda.

Overview

    This extension allows creating an EGL sync object linked to a CUDA
    event object, potentially improving efficiency of sharing images and
    compute results between the two APIs.

IP Status

    No known claims.

New Types

    A pointer to type cudaEvent_t, defined in the CUDA header files, may
    be included in the attribute list passed to eglCreateSync.

New Procedures and Functions

    None.

New Tokens

    Accepted as attribute names in the <attrib_list> argument
    of eglCreateSync:

        EGL_CUDA_EVENT_HANDLE_NV        0x323B

    Returned in <values> for eglGetSyncAttrib <attribute>
    EGL_SYNC_TYPE:

        EGL_SYNC_CUDA_EVENT_NV          0x323C

    Returned in <values> for eglGetSyncAttrib <attribute>
    EGL_SYNC_CONDITION:

        EGL_SYNC_CUDA_EVENT_COMPLETE_NV 0x323D

Changes to Chapter 3 of the EGL 1.5 Specification (EGL Functions and
Errors)

    In section 3.8.1 (Sync Objects) Replace the sixth paragraph:

   "A <OpenCL event sync object> reflects the status of a corresponding
    OpenCL object to which the sync object is linked. Likewise, a <CUDA
    event sync object> reflects the status of a corresponding CUDA
    object. These provide another method of sharing images or compute
    results between EGL and the CUDA and OpenCL client APIs (see
    Chapter 9 of the OpenCL Specification and the cl_khr_egl_image
    extension for a second method of synchronization with OpenCL).
    Waiting on either type of sync object is equivalent to waiting for
    completion of the corresponding linked CUDA or OpenCL event object."

    Add a new section following section 3.8.1.2 (Creating OpenCL Event
    Sync Objects):

   "Section 3.8.1.X Creating CUDA Event Sync Objects

    If <type> is EGL_SYNC_CUDA_EVENT_NV, a CUDA event sync object is
    created. In this case <attrib_list> must contain the attribute
    EGL_CUDA_EVENT_HANDLE_NV, set to a pointer to a cudaEvent_t object.
    The object must be properly initialized and recorded by the CUDA API
    (using cudaCreateEvent and cudaEventRecord), and the CUDA device
    used to create the event must correspond to <dpy>[fn1]. Note that
    EGL_CUDA_EVENT_HANDLE_NV is not a queryable property of a sync
    object.

    [fn1] If EGL_NV_device_cuda is supported, it is sufficient that the
          CUDA device used to create the CUDA event matches the
          EGL_CUDA_DEVICE_NV attribute of <dpy>'s underlying EGL
          device.

    Attributes of the CUDA event sync object are set as follows:

        Attribute Name          Initial Attribute Value(s)
        -------------           --------------------------
        EGL_SYNC_TYPE           EGL_SYNC_CUDA_EVENT_NV
        EGL_SYNC_STATUS         Depends on status of <event>
        EGL_SYNC_CONDITION      EGL_SYNC_CUDA_EVENT_COMPLETE_NV

    The status of such a sync object depends on the state of <event> at
    the time eglCreateSync was called. If all device work preceding the
    most recent call to cudaEventRecord on the event has not yet
    completed, the status of the linked sync object will be
    EGL_UNSIGNALED. If all such work has completed, the status of the
    linked sync object will be EGL_SIGNALED. Calling cudaEventRecord has
    no effect on a previously created sync object.

    The only condition supported for CUDA event sync objects is
    EGL_SYNC_CUDA_EVENT_COMPLETE_NV. It is satisfied when all device
    work prior to the most recent call to cudaEventRecord at sync
    creation time has completed."

    Add to the list of errors following 3.8.1.X:

   "If <type> is EGL_SYNC_CUDA_EVENT_NV and EGL_CUDA_EVENT_HANDLE_NV is
    not specified in <attrib_list>, then an EGL_BAD_ATTRIBUTE error is
    generated. If its attribute value is not a valid CUDA event pointer
    or has not been initialized as described above, then
    EGL_BAD_ATTRIBUTE may be generated, but the results are undefined
    and may include program termination."

    Modify the third paragraph of section 3.8.1.4 (Querying Sync Object
    Attributes):

   "If any eglClientWaitSync or eglWaitSync commands are blocking on
    <sync> when eglDestroySync is called, <sync> is flagged for deletion
    and will be deleted when the associated fence command, OpenCL event
    object, or CUDA event object has completed, and <sync> is no longer
    blocking any such egl*WaitSync command. Otherwise, the sync object
    is destroyed immediately.

    Replace the EGL_SYNC_CONDITION row of table 3.9 with:

   "Attribute              Description                Supported Sync Objects
    -----------------      -----------------------    ----------------------
    EGL_SYNC_CONDITION     Signaling condition        EGL_SYNC_FENCE,
                                                      EGL_SYNC_CL_EVENT, or
                                                      EGL_SYNC_CUDA_EVENT_NV

    Table 3.9  Attributes Accepted by eglGetSyncAttrib"


Interactions with EGL versions prior to 1.5

    This extension may be used with earlier versions of EGL, provided
    that the EGL_KHR_fence_sync extension is supported. In this case,
    replace all references to sync functions and tokens with
    corresponding KHR-suffixed versions (e.g., replace eglCreateSync
    with eglCreateSyncKHR).

    Additionally, this extension may be used with the 64-bit types and
    functions added to EGL_KHR_fence_sync introduced by
    EGL_KHR_cl_event2 (EGLAttribKHR and eglCreateSync64KHR). Support
    for OpenCL events is not required.

Issues

    None

Revision History

    Version 1, 2014/06/20 (Michael Chock)
        - initial version.