I am looking at the SaschaWillems subpass example for getting some insight into subpass depdendencies but its hard to understand whats going on without any comments. Also there is not a lot of documentation on subpass dependencies overall.

Looking at the code, I can see that user specifies the src subpass, dst subpass and src state, dst state. But there is no mention of which resource the dependency is on. Is a subpass dependency like a pipeline barrier. If yes, how does it issue the barrier? Is the pipeline barrier issued on all attachments in the subpass with the input src and dst access flags? Any explanation will really clear a lot of doubts on subpass dependencies.

Thank you

I recently wrote an abstraction for this mechanism so my graphics API would not be D3D12 specific.  Given that, I can only really describe this from the point of view of writing the code but since things seem to be working, I believe the details I figured out are pretty close to accurate.

First off, you need to look at the three related info structures again since they most certainly do tell you exactly which images are being referenced, it is just a bit indirect. Basically there is an array of all images used in the overall pass found in the render pass info structure, sub passes reference these images via 0 based indexing.

As to the behavior, at the start and end of each subpass the API issues an image transition barrier if needed to put the attachment in the requested format.  So, for instance, if you were doing a post processing blur, you might end up with the following chain of events:

NextSubPass
Transition attachment 0 to writable
.. Draw your scene
NextSubPass
Transition attachment 0 to readable
Transition attachment 1 to writable
.. Draw post processing quad to run vertical blur with input attachment 0 and output attachment 1
NextSubPass
Transition attachment 0 to writable
Transition attachment 1 to readable
.. Draw post processing quad to run horizontal blur with input attachment 1 and output attachment 0

So the attachments involved are ping ponging from readable to writable as required for the post processing to occur.

Hopefully this makes sense and helps you out.  I had to look at those structures quite a few times till I figured out the details.  The structures themselves are pretty simple, it's just the relationships that are hard to see until you try and fail a couple times to get the correct behavior.

Thanks for the explanation.

1 hour ago, Hiwas said:

NextSubPass
Transition attachment 0 to writable
.. Draw your scene
NextSubPass
Transition attachment 0 to readable
Transition attachment 1 to writable
.. Draw post processing quad to run vertical blur with input attachment 0 and output attachment 1

Here are you talking about the attachment in the subpass or the renderpass? (Is attachment0 relative to the pColorAttachments in the subpass or pAttachments in the renderpass)

In the subpass descriptions you have arrays of VkAttachmentReference which is a uint and layout.  The uint is the 0 based index into the VkRenderPassCreateInfo structure's pAttachment array where you listed all of the attachments for the render pass.  So, effectively, what I'm saying with those is:

// assume you have pRenderPass and pSubPass pointers to the respective Vk structures.
theImageWeWantToMessWith = pRenderPass->pAttachments[ pSubPass->pInputAttachments.attachment ]

That is effectively what is going on behind the scenes to figure out which image to call memory barriers on.
So, when I said attachment 0 and 1, I was talking about the index into the VkRenderPassCreateInfo structure's pAttachments array.  Note that render pass info does not separate inputs/outputs etc, it just takes one big list, only subpasses care about usage.

Hope that clarifies things.

So how is the image barrier issued. Is the logic something like this:

for (uint32_t i = 0; i < dependencyCount; ++i)
{
	if (pDependencies[i].srcSubpass == currentSubpass)
    {
    	for (uint32_t att = 0; att < pRenderPass->attachmentCount; ++att)
        {
        	if (pRenderPass->pAttachments[att]->srcAccessFlag == pDepdendencies[i].srcAccessFlag)
            {
            	// transition the attachment to pDependencies.dstAccess?
            }
        }
    }
}

In a general way, that is fairly close to a very simplistic solution.  Unfortunately at this level it is really all about how clever the drivers get when they solve the path through the dag generated by the subpasses.  They could do the very simplistic solution of just issuing a vkCmdPipelineBarrier with top and bottom of pipe flags set between subpasses with dependencies or they could look at the subpass attachments in detail and figure out a more refined approach.  Since this is all just a state transition chain, building a simple DAG allows for a much more optimized approach to issuing a mix of pipeline and memory barriers.

I can't find the article I remember that describes some of this but this one may be of interest: https://gpuopen.com/vulkan-barriers-explained/ as it is related.