On 4/9/12 11:29 AM, Francois Chabot wrote:
Which is great, unless the program wants to measure the cache memory
itself, in which case it would use special assembler instructions or
large memset()s. (We do such at Facebook.)
I disagree. If a regression suddenly causes a function to become
heavily cache-bound, it should show up in benchmarks somehow,
regardless of the previous expected behavior of the function.
But cache binding depends on a variety of cache characteristics, i.e.
the machine. The question is whether we accept a heavy dependence of the
benchmark on the machine.
Andrei