FIFO is overflown by GatherPipe, CPU thread is too fast! (4.0-926)

[tueidj]

Volatiles should be fine, although I'm not sure why you'd need dozens; there should only be the CP FIFO pointers (begin, end, read and write position) and interrupt/watermark thresholds to worry about. The read position should only be updated by the GPU thread and the write position should only be updated by the CPU thread so there's no explicit locking needed. Worst case should be the CPU thread seeing a stale value for the read position when bursting (for platforms with weak ordering i.e. android) resulting in it thinking the GPU is further behind than reality and firing the high watermark exception when not necessary, rather than overflowing.