Compute shader: support counter resource?

[WangHoi]

Hello, AppendStructuredBuffer/ConsumeStructuredBuffer seems not working now.

d3d12 device always create UAV with null counter resource.

[TheMostDiligent]

AppendStructuredBuffer/ConsumeStructuredBuffer is simply a buffer with a hidden counter in D3D. No other graphics APIs have the notion of counter buffers and emulating them in Diligent will only add complexity without any value as they can be straightforwardly implemented as structured buffer + counter buffer. Moreover, the explicit approach will be much more flexible as the app will have direct access to the counter.

[WangHoi]

Thanks, I'll port the code to use RWByteAddressBuffer.InterlockedAdd.
It should be slower than UAV counter.

[TheMostDiligent]

Why would it be slower?

[WangHoi]

AFAIK, Some gpu provides a very fast on-chip globally coherent cache called GDS.
AMD’s Cayman GPU Architecture

So beyond RWByteAddressBuffer.InterlockedAdd, D3D11 and D3D12 provides the UAV counter api, to expose this hardware ability.

[TheMostDiligent]

Accessing a single counter from the entire GPU should be done carefully no matter how it is implemented internally.
If your compute shader needs to write 10,000 elements, you'd better off incrementing the global counter once by 10,000 to reserve space for all elements, and then writing the elements within the compute group.
Adding 10,000 elements one-by-one is extremely unlikely to work better even if the counter resides in GDS.

[MikhailGorobets]

Tera Scale is a very old architecture, so this article cannot be taken seriously at the moment. Try writing a benchmark that uses AppendBuffer, and another one that uses atomic operations with InterlockedAdd, and we’ll see if this can actually affect performance.