Take the following scenario:
Steps 2 and 3 each require a render pass for rendering. The clear in step 1 can potentially be done through
loadOp of the render pass for step 3, assuming step 2 doesn't use the attachments of FBO1. This optimization is achieved in ANGLE by deferring clears.
When a clear is issued, one of the following happens:
Deferring a clear is done by staging a
Clear update in the
vk::ImageHelper corresponding to the attachment being cleared.
There are two possibilities at this point:
vk::ImageHelperis used in any way other than as a framebuffer attachment (for example it's sampled from), or
In scenario 1, the staged updates in the
vk::ImageHelper are flushed. That includes the
Clear updates which will be done with an out-of-render-pass
In scenario 2,
FramebufferVk::syncState is responsible for extracting the staged
Clear updates, assuming there are no subsequent updates to that subresource of the image, and keep them as deferred clears. The
FramebufferVk call that immediately follows must handle these clears one way or another. In most cases, this implies starting a new render pass and using
loadOps to perform the clear before the actual operation in that function is performed. This also implies that the front-end must always follow a
syncState call with another call (and for example cannot decide to no-op the call in between).
If the subsequent call itself is a clear operation, there could be further optimizations. In particular, the previously deferred clears and can be overridden by and/or re-deferred along with the new clears.