On Saturday, 2 September 2017 at 15:41:54 UTC, Joakim wrote:
On Saturday, 2 September 2017 at 14:49:30 UTC, thinwybk wrote:
On Friday, 14 July 2017 at 13:29:30 UTC, Joakim wrote:
Yes, D's compile-time regex are still the fastest in the
world.
I've been benching it recently for a marketing-oriented blog
post I'm preparing for the official D blog, std.regex beats
out the top C and Rust entries from the benchmarks game on
linux/x64 with a single core:
http://benchmarksgame.alioth.debian.org/u64q/regexredux.html
https://github.com/joakim-noah/regex-bench
D comes in third on Android/ARM, but not far behind,
suggesting it would still be third on that list if run with a
bunch of other languages on mobile. Dmitry thinks it might
be alignment issues, the bane of cross-platform,
high-performance code on ARM, as he hasn't optimized his
regex code for ARM.
Do you plan to implement a version for the fastest benchmark
"n-body"
(http://benchmarksgame.alioth.debian.org/u64q/nbody.html) as
well?
No, the goal is to demonstrate the nice, super-speedy regex
engine in the D standard library by using that same benchmark
that Dmitry exhibited years ago, to show D still does really
well at regex. It's not to try and compete across all those
benchmarks, which D used to dominate at one time.
I did wonder how D does on that n-body benchmark now, so I
built and ran the fastest C++ version, #8, on my single-core
linux/x64 VPS. Here's the source link for each benchmark, the
compiler version, and the command I used to build it:
C++:
http://benchmarksgame.alioth.debian.org/u64q/program.php?test=nbody&lang=gpp&id=8
clang/llvm 4.0.1
clang -O3 -std=c++11 nbody.cpp -lm -onbody-cpp
D:
https://bitbucket.org/qznc/d-shootout/raw/898f7f3b3c5d55680229113e973ef95ece6f711a/progs/nbody/nbody.d
ldc 1.4 beta1, llvm 4.0.1
ldc2 -O3 nbody.d
The D version averages 2.5 seconds, the C++ version 6 seconds,
which means D would likely still be at the top of that n-body
ranking today.
Sorry, I assumed the D version worked fine and didn't bother to
check the output, turns out it needs two foreach loops changed in
advance(dt). Specifically, "Body i" should be changed to "ref
Body i" in both foreach statements, so the data is actually
updated. ;)
After that change, the C++ version wins by a little, 6 seconds
vs. 6.5 seconds for the D translation. I see that the C++
version directly invokes SIMD intrinsics, so perhaps that's to be
expected.
