I've been looking hard and can't find an example of using fragment/tile imageblocks WITHOUT it writing to the framebuffer color attachments. The apple demo (pixie lights) seems to write to the color attachments for some reason, introducing the performance penalty that imageblocks seem to have been designed to overcome (pixel stores have been verified visually and through debugger). I've found no documentation that explicitly shows how to pass through the pipe using imageblock memory only.
Does anyone have a link/sample that explicitly demonstrates going from vertex->fragment->compute without ever writing to an attachment? This is for a very specific processing pipeline that is purpose designed for speed and should not write pixels.