2 subpass use same depth_stencil attachment as READ_ONLY

[liufachen]

My vulkan version is 1.1.106.0

The problem is : subpass “1” use color attachment A and subpass “2” use color attachment B, both use the same depth_stencil attachment as VK_IMAGE_LAYOUT_DEPTH_STENCIL_READ_ONLY_OPTIMAL.
The validation say:“A dependency between subpasses 1 and 2 must exist but one is not specified”, but Doc say:“If two subpasses use the same attachment in different layouts, and both layouts are read-only, no subpass dependency needs to be specified between those subpasses.”.

My problem is what mistake i made so i need provide a subpass-dependendcy.

I appreciate your time and help, i fought this problem almost 2 days, and hope someone could give me a help and thanks again.

and more information please refer: https://community.khronos.org/t/2-subpass-use-same-depth-stencil-attachment-as-read-only/104058/5

[tobine]

I believe there's a bug in the layers related to this. Specifically in subpass dependency checking we definitely don't account for the case quoted from the spec:

If two subpasses use the same attachment in different layouts, and both layouts are read-only, no subpass dependency needs to be specified between those subpasses.

For this specific bug the attachment is not "in different layouts" it's in the same layout. However, I think this may be a spec issue as I don't believe there should be a subpass dependency if both subpasses access the attachment in read-only layouts regardless of whether the layouts are the same or not.

[liufachen]

I appreciate for your help and hope this issue will be closed in the feature.

Thank you tobine.

[jzulauf-lunarg]

@tobine -- did you open spec issue (or propose an MR) for the suspected spec bug? It always makes me nervous to get Validation ahead of the spec in this sense, without at least some grunts of affirmation out of the WG. (not that I disagree, if different read-only layouts don't require dependencies**, then read-only attachments with the same layout certainly don't seem like they should).

@mikes-lunarg

** It strikes me that the implementation has got to be potentially inserting some additional dependencies for the cases were the logical image layouts require actual image layout transitions... but it looks like the spec authors were willing to sign up to that implementation complexity.

[jzulauf-lunarg]

@tobine -- does the following give enough cover to say that read-only attachments (input or read only depth stencil) don't require subpass dependency w.r.t. other read-only references to the same attachment? Then the "different layouts" section of the spec is then just further clarification that even image layout transitions between read only formats (and image layout transitions can write) don't require them either and not a restriction for only "different layouts".

Care should be taken to avoid a data race here; if any subpasses access attachments with overlapping memory locations, and one of those accesses is a write, a subpass dependency needs to be included between them.

[tobine]

@jzulauf-lunarg That helps but doesn't seem conclusive. I submitted a spec edit and tagged you on it to clarify that no subpass dependency is required between subpasses using the same attachment in any read-only layouts (regardless of if the layouts are the same or not)

[oddhack]

The spec part of this should be addressed in the 1.1.117 spec update. Please close if that's all that needs to be done for this issue.