powturbo / turbo-histogram Goto Github PK

I got a substantial boost of speed on your hist_8_64 by changing the size of your bin[] arrays to prevent 4K aliasing. The issue is that when the CPU is trying to disambiguate memory accesses, it looks only at the low 12 bits (4KB). If two addresses have the same low bits (and maybe some other stuff?), it falsely thinks that they collide. The net effect is that you are only getting the benefit of 4 bins rather than 8. If you change the size of your buffers to be (256+16) bytes, the same character will no longer alias in any of the bins.

(I'm about to post a comment on Yann's blog, and will include a link to my updated source)

I'm able to test the speeds on Haswell, but wanted to use the same test files that you did so the numbers are directly comparable. I'm not familiar with the standard benchmarks in this area, and worry I might grab the wrong file if I guess. Could you provide a direct download link to the the files in the README, or exact directions for creating them?