I modified the code so that render passes with multiple subpasses correctly call Image::access(..) for all subpasses, in order, so that the final layout of the image matches the render pass.

Previously the loop was skipping Image::access(..) calls unless it was the first access of that image for the render pass. This change only affects the scenario where an image is, for instance, written as an attachment and then used as input or some other access type within the same render pass.

Overall, I am not sure the behavior we're seeing for AMD on Linux matches the Vulkan spec, because AFAIK the render pass should act as a barrier and so the additional wait on "Nothing" until "Present" is valid given that the render pass waited on "Color Attachment Write". Either way this change isn't wrong or performance-impacting and brings the logic back to where it was before #81, which explicitly barriers on the previous work.

In any case, your diagnosis was correct and greatly helped me review this issue! If it looks good to you I'm ready to merge it and push an updated patch version.