Questions about prefetch and cache flush

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi everyone:

  I am now using nvprof to profile the hit ratios for some benchmarks. But I found that the results some how does not match my expectation. So I am worried that prefetch and cache flush may effect the profiled result. 

  So two questions:

  (1) is there a way to disable or enable the prefetch for the L2 cache?

  (2) is there a way to flush L2 cache before launching the kernels?

  Thanks!

Dong

2

Unfortunately the answer is no to both questions. All global memory accesses go through the L2 cache.
Instead of flushing L2, you could however cache a large enough unrelated memory area to ensure L2 holds no relevant data.