General Q&A

[mjguynn]

A bunch of questions I've asked and received answers to, hopefully this will be useful to others. Answers might not be up-to-date. Please correct any mistakes I made, I probably misremembered some of these answers -- my fault for not writing them down sooner!

Q: Do value functions reference value funclets or scheduling funclets? Or is it like the pipeline entry point, where it can reference either, and value funclets are explicated as necessary?

A: They reference value funclets. The specific implementation is selected in the schedule.

Q: Why do we see "submit->encode fence->sync" in schedules? Shouldn't fence encoding occur before submission?

A: This isn't necessary under WGPU semantics. It automatically attaches a fence after every submission and "encode fence" really just grabs a reference to the most recent fence.

Q: Follow-up: Would "encode fence->submit->sync" work as well?

A: I don't think so. The fence would actually refer to the previous submission.

Q: What's the "return value" of encode-do-* in CAIR?

A: Nothing. If you look at content.ron, you'll see that EncodeDo actually has a return type of None. The same is true for EncodeCopy. Returning nothing isn't really a concept in CAIR yet, so you need to add a "useless" return value, but you can call it %_ to indicate to human readers that it's useless and should be ignored.

Q: What's the "return value" of submit-* in CAIR?

A: Nothing right now -- but it should return a fence eventually (negating the need for EncodeFence.)

Q: What's the "hierarchy" of funclets?

A: The value, spatial, and temporal funclets all define a contract. The scheduling funclet is the "real" implementation and must satisfy the constraints given by the previous three.
Each scheduling funclet maps to at most one value funclet, at most one timeline funclet, and at most one spatial funclet. The converse is not true. For example, a value funclet which calls a value funclet is implemented by at least two scheduling funclets (one for before the call and one for after the call.) The schedule-to-value bindings are explicitly established in the extras section of the RON/CAIR. (Follow-up: are the timeline/spatial bindings established by the timeline/spatial tags in that section?)

Q: How come some examples encode CPU external function execution on the local queue? Couldn't the local queue theoretically be on the GPU?

A: In the future, yes. Currently the local queue is always on the CPU. Encoding external CPU calls on the local queue shouldn't be allowed, but currently it works better than trying to use the CPU queue.

Q: Why can't you submit or sync on the local queue?

A: The local queue is always up-to-date. Local resources are never in the Encoded stage. (Follow-up: does this mean they're also never in the Submitted stage? In other words, they directly go from Bound to Ready upon local command encoding?)

Q: Why does AllocTemporary need to know the corresponding value funclet node? What do you do if you want to "replicate" the same logical value in multiple physical locations; for example, copy a value from the CPU to the GPU?

A: To answer the first part, IIRC it needs to know the node in order to validate that the schedule actually satisfies the value funclet. To answer the second part: just refer to the same node. I thought this didn't work, but was actually a bug with CPU allocation.

Q: Why can I return a slot that was created via AllocTemporary? Temporaries can't leave their scope.

A: Correct, they can't, but Caiman doesn't check this right now.

[mjguynn]

Q: If the output of bar is assigned no value tag, why is the following legal?

value @foo() -> i64 {
  %x = constant 2i64;
  %y = constant 3i64;
  return %x;
}

schedule @bar() -> $slot {
  %y_loc = allocate @foo.%y;
  encode-do %y() -> %y_loc;
  return %y_loc;
}

A: The return values of value and scheduling funclets don't necessarily match one-for-one. A scheduling funclet can return whatever it wants as long as it doesn't tag those outputs. For example, a scheduling funclet could return a supplementary data structure that informs some calculation in its callee scheduling funclet, even if the value function has no such data structure.

On the other hand, suppose you wrote

value @baz() -> i64 {
  %z_t = call @foo_function();
  %z = extract %z_t 0;
  return %z;
}

and implement a schedule for @baz which attempts

  %z_loc = allocate @baz.%z;
  // call @bar and put its return value in %z_loc, I don't actually know how to write the scheduling code to do this yet

This should fail typechecking. The problem is that the value funclet @baz expects %z to be the first return value of @foo_function, but the first return value of @baz does NOT correspond to the first return value of @foo_function because it isn't tagged as such. And if you DO tag it as the first return value of @foo_function, then @bar will be rejected as an invalid schedule because clearly it's returning %y instead of %x.

The confusing part is the coordinator-Caiman boundary since there isn't a distinction between value-funclet-output-tagged return values and untagged scheduling-funclet-only return values.