My understanding is that register based prefetching should increase the number of registers compared to the baseline and thus start limiting occupancy (when the prefetch distances increases greatly).
I'm finding that the number of registers/thread for RPF is 66 for a prefetch distance of 4 compared to 74 in the baseline. This is counter-intuitive to me. At a prefetch distance of 8, RPF gives 74 registers/thread.
I compiled this on a Hopper GPU.
Let me know if these numbers are to be expected.
My understanding is that register based prefetching should increase the number of registers compared to the baseline and thus start limiting occupancy (when the prefetch distances increases greatly).
I'm finding that the number of registers/thread for RPF is 66 for a prefetch distance of 4 compared to 74 in the baseline. This is counter-intuitive to me. At a prefetch distance of 8, RPF gives 74 registers/thread.
I compiled this on a Hopper GPU.
Let me know if these numbers are to be expected.