As discussed in the February VF2F, WGSL should get a built-in function workgroupBroadcast. This would allow users to branch uniformly on reads from storage buffers for example.

Proposal

Workgroup broadcast is a new collective operation (i.e. must be called only in uniform control flow). It would be compute stage only. The function would return a uniform value (i.e. the same value for each invocation in the workgroup). Uniformity analysis would understand the built-in function and consider the result uniform regardless of the uniformity of the input value.

// Returns "value" for linear local invocation index 0 to all workgroup invocations.
fn workgroupBroadcastFirst(value : T) -> T

// Local invocation index 0 loads p and returns the value to all workgroup invocations.
fn workgroupBroadcastFirst(p : ptr<T, S>) -> T

// Possible, generic broadcast. Takes the local invocation index to broadcast (must be const expression).
fn workgroupBroadcast(value : T, linear_id : i32/u32) -> T
fn workgroupBroadcast(p : ptr<T, S>, linear_id : i32/u32) -> T

A possible implementation in pseudo-WGSL would look like:

var<workgroup> wg_storage : T;
fn workgroupBroadcastFirst(p : ptr<T, S>) -> T {
  workgroupBarrier();
  if (local_invocation_index == 0) {
    wg_storage = *p;
  }
  workgroupBarrier();
  return wg_storage;
}

Depending on the underlying implementation the polyfill could be improved to use subgroup broadcasts to reduce workgroup memory reads.