Experimental workgraph - Phase 4

[HansKristian-Work]

This is the first one with some real meat to it.

• Adds a missing meta .md document I forgot to include in the first rebase.
• Adds meta operations for workgraphs. The high-level details are explained in the .md, but the TLDR is that we have four shaders:

Setup GPU input

Only used when processing a GPU_INPUT entry point. Just generates indirect dispatches.

distribute_workgroups

• Counts number of records produced by each node, prefix sums it and allocates space for the expanded offsets.
• Sets up indirect command for each node. TODO: Potentially allocate a EXT_dgc entry for non-empty nodes.
• Sets up one indirect command for payload expander.

distribute_payload_offsets

Loops over { node index, payload offset, count } entries and expands this into a simple array of u32 payload offsets which will be consumed by the next stage's nodes. The space for these offsets were allocated by distribute_workgroups.

complete_compaction

This was recently added. Adjusts the broadcast amplification rate based on the data we analyzed in payload_offsets, once per node.

[doitsujin]

only thing that would come to mind is rewriting the loop as something like

bool thread_mask = count != 0;

while (subgroupAny(thread_mask)) {
  uint lane = subgroupBallotFindLSB(subgroupBallot(thread_mask));
  thread_mask = thread_mask && gl_SubgroupInvocationID > lane;
  ...
}

[HansKristian-Work]

I think that comparison would require vector ops.

[doitsujin]

Should be good now.