On Thursday, 19 May 2016 at 18:22:48 UTC, Timon Gehr wrote:
On 19.05.2016 08:04, Joakim wrote:
On Wednesday, 18 May 2016 at 17:10:25 UTC, Timon Gehr wrote:
It's not just slightly worse, it can cut the number of useful
bits in
half or more! It is not unusual, I have actually run into
those
problems in the past, and it can break an algorithm that is
in Phobos
today!
I wouldn't call that broken. Looking at the hex output by
replacing %f
with %A in writefln, it appears the only differences in all
those
results is the last byte in the significand.
Argh...
// ...
void main(){
//double[] data=[1e16,1,-9e15];
import std.range;
double[] data=1e16~repeat(1.0,100000000).array~(-9e15);
import std.stdio;
writefln("%f",sum(data)); // baseline
writefln("%f",kahan(data)); // kahan
writefln("%f",kahanBroken(data)); // broken kahan
}
dmd -run kahanDemo.d
1000000000000000.000000
1000000100000000.000000
1000000000000000.000000
dmd -m32 -O -run kahanDemo.d
1000000000000000.000000
1000000000000000.000000
1000000000000000.000000
Better?
Obviously there is more structure in the data that I invent
manually than in a real test case where it would go wrong. The
problems carry over though.
I looked over your code a bit. If I define sum and c as reals in
"kahanBroken" at runtime, this problem goes away. Since that's
what the CTFE rule is actually doing, ie extending all
floating-point to reals at compile-time, I don't see what you're
complaining about. Try it, run even your original naive
summation algorithm through CTFE and it will produce the result
you want:
enum double[] ctData=[1e16,1,-9e15];
enum ctSum = sum(ctData);
writefln("%f", ctSum);
As Don's talk pointed out,
all floating-point calculations will see loss of precision
starting there.
...
This is implicitly assuming a development model where the
programmer first writes down the computation as it would be
correct in the real number system and then naively replaces
every operation by the rounding equivalent and hopes for the
best.
No, it is intrinsic to any floating-point calculation.
It is a useful rule if that is what you're doing. One might be
doing something else. Consider the following paper for an
example where the last bit in the significant actually carries
useful information for many of the values used in the program.
http://www.jaist.ac.jp/~s1410018/papers/qd.pdf
Did you link to the wrong paper? ;) I skimmed it and that paper
explicitly talks about error bounds all over the place. The only
mention of "the last bit" is when they say they calculated their
constants in arbitrary precision before rounding them for runtime
use, which is ironically similar to what Walter suggested doing
for D's CTFE also.
In this case, not increasing precision gets the more accurate
result,
but other examples could be constructed that _heavily_ favor
increasing
precision.
Sure. In such cases, you should use higher precision. What is
the problem? This is already supported (the compiler is not
allowed to use lower precision than requested).
I'm not the one with the problem, you're the one complaining.
In fact, almost any real-world, non-toy calculation would
favor it.
In any case, nobody should depend on the precision out that
far being
accurate or "reliable."
IEEE floating point has well-defined behaviour and there is
absolutely nothing wrong with code that delivers more accurate
results just because it is actually aware of the actual
semantics of the operations being carried out.
You just made the case for Walter doing what he did. :)