Defragmentation relies on undefined memory aliasing behavior about vulkanmemoryallocator

[adam-sawicki-a]

Thank you for reporting this issue and for the detailed description. I think you are right, it is a major issue. We will analyze it and go back to it when we find some proposed solution.

[adam-sawicki-a]

I've made a clarification in the documentation stating that only buffers can be defragmented, not images.

I also made a fix that limits GPU defragmentation only to those memory types that support creating a buffer used to do GPU transfers.

Do you think that defragmentation with these limitations is now compliant to Vulkan specification?

[jeffbolznv]

Yes, that should be compliant. The pre-edited comment where you mentioned linear images would be as well, since linear images are allowed to alias with buffers.

[Tobski]

Hey @jeffbolznv, I suggested @adam-sawicki-amd remove the linear image part, as linear images still have an expected image "header" with undefined contents - buffers only have defined aliasing over the texel data, which occurs after an implementation defined offset. So that falls under the same conditions as the rest of a regular images AFAICT.

BTW I definitely think we need some spec clarifications here...

[jeffbolznv]

OK, fine with me.

I think the spec is fairly clear about what can alias with what in the resources-memory-aliasing section.

[Tobski]

Ah I forgot about that language - you're right that's fairly clear.

[adam-sawicki-a]

Thank you for your help with sorting out this issue.

[adam-sawicki-a]

Please note a new defragmentation is being developed as #90 !