Cache sizes

[mborland]

In the discussion on the prime sieve PR (#400) we have run into the issue several times of not having a cross-platform way of querying the size of the L1 cache. This repo should solve that issue. Is there an existing library in boost that this should go into as a PR? I think it would be useful outside of math, but it is not clear to me where it should go.

[NAThompson]

I suspect, but do not know, that your syscalls will compile down to a cpuid asm instruction. This is one of the slowest instruction on a modern CPU, so it might not pay off to determine the cache size.

Could you godbolt + benchmark your cachesize code?

[mborland]

sysconf definitely compiles down to cpuid. I added the benchmark (generally copied from constants_performance), and it takes 143ns to run on my machine. My thought was having a default L1D value of 32768, and exposing an interface to change that value.

[jzmaddock]

Presumably this would only be called once during program initialization anyway - in which case the cost would be negligible?

As for where it belongs.... I have no idea!

[NAThompson]

What about boost::predef?

I've always thought that the functionality of cpuid should be exposed somewhere in boost.

[mborland]

@jzmaddock With the user being responsible for memory allocation with the OI approach this could just be a extra step in initialization. Querying the L1D size should be trivial in comparison to a cache miss/under-utilization.

@NAThompson I will open an issue with them to see what they say.

[ckormanyos]

...a cpuid asm instruction. This is one of the slowest...

On the architecture being discussed, the cpuid instruction is one of several instructions with the special characteristic of being a barrier instruction, also known as a serializing instruction. In addition to returning the CPU-ID information in the specified registers, this instruction will block and clear the instruction pipeline. Another such instruction is rdtsc on that core.

These instructions are great if you are writing a real-time operating system or similar stuff on the metal. They will, however, hit the performance with the branch and pipeline clear.