On Monday, 22 February 2016 at 07:10:23 UTC, Kapps wrote:
If you do want to test the differences between the range
approach and the loop approach, something like:
auto sumtest4(Range)(Range range) @safe pure {
return range.reduce!((a, b) => a + b);
}
is a more fair comparison. I get results within 15% of sumtest2
with this using dmd. I think with ldc this would be identical,
but the version in homebrew is too old to compile this.
Using LDC with the mir version of ndslice so it compiles, and the
following code:
sw.reset();
sw.start();
foreach (unused; 0..times) {
for (int i=0; i<N; ++i) {
res4[i] = sumtest4(f[i]);
}
}
t3 = sw.peek().msecs;
and
auto sumtest4(Range)(Range range) {
return range.reduce!((a, b) => a + b);
}
I get:
145 ms
19 ms
19 ms
19 ms
So, with LDC, there is no performance hit doing this. The only
performance hit is when .sum uses a different algorithm for a
more accurate result. Also, the LDC version appears to be roughly
5x faster than the DMD version.
