Specifically, this is currently about fixed foveated rendering or lens-matched shading, though similar techniques would also be interesting for any potential future eye-tracked system with dynamic foveation.

For reference, the suggestion was to investigate the Qualcomm extensions:

https://www.khronos.org/registry/OpenGL/extensions/QCOM/QCOM_framebuffer_foveated.txt

https://www.khronos.org/registry/OpenGL/extensions/QCOM/QCOM_texture_foveated.txt


The framebuffer_foveated extension is available for Pixel phones using Android O, but very inflexible. It's not possible to modify attachments on an FBO once it was set as foveated. The texture_foveated extension is more flexible, and available on Pixel phones starting with the Android P preview, and also available on the Mirage standalone device. These newer devices no longer support framebuffer_foveated, so we'd want to go with texture_foveated.

The extension reduces the render cost by adjusting the internal tile resolution while rendering, then interpolating the rendered pixel when resolving into the destination texture.

This mainly saves resources when the fragment shader cost is comparatively high. If that's not the case, it may potentially even slow things down due to extra overhead for maintaining the foveated texture internally.

I agree that it's worth trying, especially since it can be applied just to the background texture while keeping the separately composited content quad at full quality.

Handwavy suggestion for an experiment:

- acquire GVR frame
- get hardware buffer:

https://developers.google.com/vr/reference/ios-ndk/group/swap-chain#group__swap__chain_1gaf878e168841607c080e8f41bc39b4e82

  acquired_frame_.BindBuffer(kFramePrimaryBuffer);
  AHardwareBuffer* ahb = static_cast<AHardwareBuffer*>(
      acquired_frame_.GetHardwareBuffer(kFramePrimaryBuffer));

  // or use C API:  gvr_frame_get_hardware_buffer(acquired_frame_.cobj(), kFramePrimaryBuffer));

- bind AHardwarebuffer as GLImage
- bind GLImage as texture
- set foveation on the texture:

  glBindTexture(GL_TEXTURE_2D, shared_buffer_texture_id_);
  glTexParameteri(
      GL_TEXTURE_2D, GL_TEXTURE_FOVEATED_FEATURE_BITS_QCOM,
      GL_FOVEATION_ENABLE_BIT_QCOM | GL_FOVEATION_SCALED_BIN_METHOD_BIT_QCOM);

  // Desired parameters if each eye were its own buffer. Adjusted below
  // for left/right split framebuffer.
  float focal_x = 0.f;
  float focal_y = 0.f;
  float gain_x = 5.f;
  float gain_y = 5.f;
  float fovea_area = 1.4f;

  float focal_x_left = (focal_x - 1.f) / 2.f;
  float focal_x_right = (focal_x + 1.f) / 2.f;
  float gain_x_split = gain_x * 2.f;
  glTextureFoveationParametersQCOM(shared_buffer_texture_id_, 0, 0, focal_x_left,
                                     focal_y, gain_x_split, gain_y, fovea_area);
  glTextureFoveationParametersQCOM(shared_buffer_texture_id_, 0, 1, focal_x_right,
                                     focal_y, gain_x_split, gain_y, fovea_area);


- set up FBO using the texture as color buffer
- draw to FBO
- unbind FBO 
- submit to GVR

See vr_shell_gl's WebVR SharedBuffer code and/or https://cs.chromium.org/chromium/src/chrome/browser/android/vr/arcore_device/ar_image_transport.cc for code examples related to setting up the GLImage, texture, and FBO.

Marking this as available for now. Does somebody have plans on working on this? Maybe it could also go on our feature spreadsheet.