I didn't spend much time looking at the code. These kinds of tests do
very little to show what real-world differences may or may not exist
between languages. The real question isn't whether C++ is or is not
significantly faster than C#. It's whether the C# program is able to
meet your specific performance requirements. I've yet to run into a
real-world situation where C# genuinely could not be made to run as well
as C++ code for the intended purpose.
That said, one thing stands out, in that the C# code is using a much
less accurate timer mechanism than the C++ code. It's entirely possible
that a significant portion of the apparent difference is simply due to
measurement error.
The fact that the C++ code shows times evenly distributed between 20 and
23 ms, while the C# code shows basically just two different timing
results (140/141 and 156 ms) strongly suggests at least _some_ sort of
measurement error is going on. Real programs don't quantize timings
like that, but imprecise timers do.
The other thing that stands out with respect to the post itself is that
the C# test program does no "warm-up" execution at all (to ensure the
code has been JITted and the various CPU caches have stabilized before
timing), and yet the first iteration is within the tests own margin of
error relative to the subsequent eight iterations.
I've also never seen a C++ compiler that would inline a function as
large as the one in that test, nor would inlining have any noticeable
effect on the overall performance of the program (none of real time
spent in the code involves any function calls at all). The idea that
the poster needed to "downgrade the optimizations" is suspicious to me.
Finally, without a complete description of the test conditions,
including whether a debugger was attached, and what compiler switches
were used to compile the code, it's impossible to confirm that the
person was comparing apples to apples.
I won't go so far to say that the person is just making his test results
up. But if I actually cared about whether the test was valid or not
(and frankly, I don't…as I said, even if the test is completely
accurate, it doesn't tell us anything useful), I would make damn sure I
verified the test myself, using a proper timing object (i.e. Stopwatch),
and doing a warm-up run (or just ignoring the first iteration). I
definitely wouldn't take the word of a post you found on the Internet.
Pete
