WebGL Extension Proposal

Byungseon Shin, LG

Motivations

• YouTube on TV HTML5
  • YouTube 360 does not play at smooth 60FPS on embedded devices
• WebGL for 360-degree video.
• Must support VideoElement source
• Should render smoothly (~60 fps)
  • Most of the GPU resources are consumed by Color Space Conversion (CSC) processing

VR Technology

• 360 VR Tech: spherical mapping of equirectangular textures
• Maping a flat surface onto a 3D model (in this case, a sphere).
• Equirectanular Format: -- A bit like a map projection.
• Native OpenGL ES implementation was 60 fps. Video -> libvt - { texture } -> Sphere Mapping (OpenGL) -> Frame Buffer.
• Initially tried WebGL, but was SLOW. (~30 fps). Video -> libvt - { texture } -> WebGL (YUV -> RGBA) -> Sphere Mappig (OpenGL) -> Frame Buffer.
  • There's an extra step of converting from libvt to RGBA
  • In the case of 360VR, final screen size will be far smaller than input screen, but we have to convert full screen to use WebGL
  • Three bottlenecks:
    • Color Space Conversion (YUV -> RGB): 90%
    • Texture Copy inside Video Decoder: 8-9%
    • Compositing of WebGL canvas: 1-2%

• Currently WebGL only supports TEXTURE_2D (RGBA) as an import. Not YUV.
• YUV -> RGB conversion is very slow.

Proposal

OES_EGL_image_external

• Add an OES_EGL_image_external extension on WebGL. -- Allows YUV textures to be handled without color space conversion.
• Solution: currently, WebGL only supports TEXTURE_2D (RGBA). We should make it support TEXTURE_EXTERNAL_OES as well
• Expose OES_EGL_image_external functionality to WebGL (define a new texture target TEXTURE_EXTERNAL_OES)
• Advantage over existing proposals: focus on extending texture format of WebGL. Make it easy to port and use
• 2 issues pointed out in the WG:
  • EGL image sync issue (syncing EGL image between video decoder)
  • Audio sync (need exact timestamp of the frame currently being rendered)
• Also need other per-frame metadata like dimensions of the texture
• This approach allows avoiding converting the entire video image from YUV -> RGBA. Instead, only need to decode the region presented in the frame buffer.
• This extension already exists in OpenGL. This is just a matter of exposing through WebGL.
• Some competing proposals work on a similar problem, but are not as good a fit

Challenges

• Audio sync is a key challenge to overcome.

Questions

• Why not just decode the portion of the image that will be displayed?
  • Might work if the viewport is fixed, but this won't have optimal performance when changing the viewport