Perlin noise benchmark speed

[David Nadlinger via Digitalmars-d]

On Friday, 20 June 2014 at 13:46:26 UTC, Mattcoder wrote:
> On Friday, 20 June 2014 at 13:14:04 UTC, dennis luehring wrote:
>> write, printf etc. performance is benchmarked also - so not 
>> clear
>> if pnoise is super-fast but write is super-slow etc...
>
> Indeed and using Windows (At least 8), the size of 
> command-window (CMD) interferes in the result drastically... 
> for example: running this test with console maximized will 
> take: 2.58s while the same test but in small window: 2.11s!

Before I wrote the above, I briefly ran the benchmark on my local 
(OS X) machine, and verified that the bulk of the time is indeed 
spent in the noise calculation loop (with stdout piped into 
/dev/null). Still, the LDC-compiled code is only about half as 
fast as the Clang-compiled version, and there is no good reason 
why it should be.

My new guess is a difference in inlining heuristics (note also 
that the Rust version uses inlining hints). The big difference 
between GCC and Clang might be a hint that the performance drop 
is caused by a rather minute difference in optimizer tuning.

Thus, we really need somebody to sit down with a 
profiler/disassembler and figure out what is going on.

David