Wait until all futexes are available?

I endorse the above. For nodes with a large number of inputs, each bit can represent another bitmap, recursively, until there are enough bits for all the inputs. When any bitmap is filled, the producer sets the bit representing it at the next level up. When the topmost bitmap is filled, the producer that filled it wakes the thread.

Bits are set by compare-and-swap. atomically, with no system call needed.

To do it with a counter, the producer atomically decrements a counter, and wakes the thread when it sees the transition to zero.

Waking the thread might mean writing a byte to a pipe it is sleeping on.

Or, the producer's thread might just add the consumer to a scoreboard of runnable tasks, and some already-running thread will get to it. That way, the kernel is never involved at all. You keep exactly as many threads as you have cores on the job, and they spin when there is no work. Isolate those cores (isolcpus, nohz_full, rcu_nocbs) so the kernel does not steal them.