We can use __builtin_ipu_get_scount_u, together with __builtin_ipu_get_scount_l, to get the full 64-bit cycle counter. The challenge is that when calling both builtins in a row (ideally first the upper and the lower) may overflow the lower and flip the upper, so we need to manually deal with the case where the lower counter is less than 12 (or 6?).
We can use
__builtin_ipu_get_scount_u, together with__builtin_ipu_get_scount_l, to get the full 64-bit cycle counter. The challenge is that when calling both builtins in a row (ideally first the upper and the lower) may overflow the lower and flip the upper, so we need to manually deal with the case where the lower counter is less than 12 (or 6?).