Analysis for vector size of 1,000 elements -Assume that block size is 256 threads -8 warps in each block All threads in Blocks 0,1,and 2 are within valid range i values from 0 to 767 There are 24 warps in these three blocks,none will have control divergence Most warps in Block 3 will not control divergence Threads in the warps 0-6 are all within valid range,thus no control divergence One warp in Block 3 will have control divergence Threads with i values 992-999 will all be within valid range Threads with i values of 1000-1023 will be outside valid range Effect of serialization on control divergence will be small 1 out of 32 warps has control divergence The impact on performance will likely be less than 3% 电子神越女学 O13 Analysis for vector size of 1,000 elements – Assume that block size is 256 threads – 8 warps in each block – All threads in Blocks 0, 1, and 2 are within valid range – i values from 0 to 767 – There are 24 warps in these three blocks, none will have control divergence – Most warps in Block 3 will not control divergence – Threads in the warps 0-6 are all within valid range, thus no control divergence – One warp in Block 3 will have control divergence – Threads with i values 992-999 will all be within valid range – Threads with i values of 1000-1023 will be outside valid range – Effect of serialization on control divergence will be small – 1 out of 32 warps has control divergence – The impact on performance will likely be less than 3%