Yeah, I also came to that conclusion (one FIFO per thread). Adding the
FIFOs wouldn't be that difficult. Just have a fixed size array
(maximum number of threads), which is initialized with NULLs. A
GPrivate (private data per thread) could be used to assign a FIFO to
the thread, the pointer to the FIFO could then be added to the array,
by looping on it and using g_atomic_int_compare_and_exchange() on NULL
values until it returns TRUE.
The synthesis thread could then just use g_atomic_int_get() when
iterating over the array to process the FIFOs and stop at the first
NULL (queues would never go away).