Use FastRounding for speed #25
Conversation
Codecov Report
@@ Coverage Diff @@
## master #25 +/- ##
==========================================
- Coverage 92.29% 91.64% -0.66%
==========================================
Files 21 21
Lines 935 945 +10
==========================================
+ Hits 863 866 +3
- Misses 72 79 +7
Continue to review full report at Codecov.
|
src/intervals/arithmetic.jl
Outdated
|
|
||
| mid{T}(a::Interval{Rational{T}}) = (1//2) * (a.lo + a.hi) | ||
|
|
||
| function simple_mid(a::Interval) |
There was a problem hiding this comment.
Do we really need simple_mid as a new function? I think its sole use is inside that method for mid.
In any case, it is now good timing to add docstrings and tests.
There was a problem hiding this comment.
I think you're right that we no longer need simple_mid.
src/intervals/rounding.jl
Outdated
| @@ -9,6 +9,7 @@ interval rounding types, e.g. | |||
| +(IntervalRounding{:none}, a, b, RoundDown) | |||
|
|
|||
| The current allowed rounding types are | |||
| - :errorfree # use errorfree arithmetic via FastRounding.jl | |||
There was a problem hiding this comment.
I suggest the following changes for naming the rounding modes:
:errorfreeshould be named:fast
-:fastshould be named:wider(or:wide?)
There was a problem hiding this comment.
It should also be mentioned somewhere (the docstrings and perhaps a warning too) that :errorfree works only with Float64 and Float32.
There was a problem hiding this comment.
I shall still refer to :errorfree instead of (the new) :fast to avoid confusions.
There was a problem hiding this comment.
Agree with :fast and :wide, thanks for the suggestions.
There was a problem hiding this comment.
Following the standard, how about :errorfree -> :fast; :fast -> :accurate; :correct-> :tight?
There was a problem hiding this comment.
The problem is that this implies that :fast is not :tight, which is incorrect. (Errorfree arithmetic does give tight bounds.)
src/intervals/rounding.jl
Outdated
| @@ -157,23 +199,31 @@ doc""" | |||
| setrounding(Interval, rounding_type::Symbol) | |||
|
|
|||
| Set the rounding type used for all interval calculations on Julia v0.6 and above. | |||
| Valid `rounding_type`s are `:correct`, `:fast` and `:none`. | |||
| Valid `rounding_type`s are `:correct`, `:fast` and `:none`, `:errorfree`. | |||
There was a problem hiding this comment.
I think this should read as "Valid rounding_types are :correct, :fast, :none, and :errorfree".
|
I forgot to say that this should probably be rebased against current master; the tests related to |
|
I just read in the standard (cf page 9, definition of tightness) about what we are discussing as rounding modes. They use the nomenclature "tightest", "accurate" and "valid", which correspond respectively to |
dpsanders
force-pushed
the
fast_rounding
branch
2 times, most recently
from
05b5e71
to
0228710
Compare
|
What is faster, I actually thought that the bounds provided by I suggest to have the docstring in decreasing order of tightness. |
|
What was |
|
So maybe we can replace |
|
Maybe you are right; give me a bit of time, since I thought i had some counterexample. |
I just run all ITF1788 suite tests and they pass with :errorfree / :fast. So, I agree: it should be renamed as Sorry for being so skeptic... |
|
I was quite sceptical actually that this would ever be possible. But thanks to @JeffreySarnoff and some hard work fixing difficult corner cases, it seems that it is finally finished! |
|
So far, I have |
|
! |
|
That's ok with me! Once travis shows the green lights, I'll merge this. |
2 similar comments
|
Everything is passing! |
|
This is really nice! Merging... |
|
Thanks! |
Use
FastRounding.jlas a drop-in replacement for changing the rounding mode. This gives a 3-4x speed-up for basic arithmetic operations.