Refactor `@round` to transform `a+b` to `+(a, b, RoundDown)` #188

dpsanders · 2017-01-09T16:02:26Z

Transforms e.g.

@round(a.lo+b.lo, a.hi+b.hi)

to

Interval(+(a.lo, b.lo, RoundDown), +(a.hi, b.hi, RoundUp))

lbenet · 2017-01-09T17:27:20Z

Tests are not passing...

dpsanders · 2017-01-09T17:33:47Z

Yes, unfortunately that will require quite a lot of work. But I think this is a good direction to go in.

coveralls · 2017-01-24T16:03:47Z

Coverage decreased (-27.7%) to 64.851% when pulling b724ebe on refactor_rounding into 50eb702 on master.

codecov-io · 2017-01-24T16:03:50Z

Current coverage is 90.70% (diff: 85.90%)

Merging #188 into master will decrease coverage by 1.80%

@@             master       #188   diff @@
==========================================
  Files            21         23     +2   
  Lines           988        990     +2   
  Methods           0          0          
  Messages          0          0          
  Branches          0          0          
==========================================
- Hits            914        898    -16   
- Misses           74         92    +18   
  Partials          0          0

Powered by Codecov. Last update 50eb702...7e67cbd

coveralls · 2017-01-24T16:12:06Z

Coverage decreased (-27.8%) to 64.683% when pulling a7fc08e on refactor_rounding into 50eb702 on master.

coveralls · 2017-01-24T16:18:17Z

Coverage decreased (-27.8%) to 64.683% when pulling e6b1f6f on refactor_rounding into 50eb702 on master.

coveralls · 2017-01-24T16:20:04Z

Coverage decreased (-27.8%) to 64.683% when pulling e6b1f6f on refactor_rounding into 50eb702 on master.

coveralls · 2017-01-24T16:38:47Z

Coverage decreased (-16.0%) to 76.511% when pulling e02bf46 on refactor_rounding into 50eb702 on master.

coveralls · 2017-01-24T16:38:47Z

Coverage decreased (-16.0%) to 76.511% when pulling e02bf46 on refactor_rounding into 50eb702 on master.

coveralls · 2017-01-24T16:38:47Z

Coverage decreased (-16.0%) to 76.511% when pulling e02bf46 on refactor_rounding into 50eb702 on master.

dpsanders · 2017-01-26T13:03:46Z

Tests are passing on 0.4 and 0.5, and almost on 0.6 with known issues (one in FixedSizeArrays.jl, and the other in Base). cc @lbenet

dpsanders · 2017-01-28T18:12:03Z

I split out the rounding macros into a new file, rounding_macros.jl.
(Actually there is now only one rounding macro.)

dpsanders · 2017-01-28T18:14:20Z

I should say that part of the motivation for this whole change, which was in my head but not necessarily mentioned anywhere, was that this will make it much easier to experiment with different types of rounding: the only place that needs to be changed are now the definitions in rounding.jl!

EDIT: See the rounding_modes branch for an example. The ROUNDING variable can be taken e.g. from an environment variable. In Julia 0.6 it will be possible to change this interactively by just redefining all these functions!

coveralls · 2017-01-28T18:38:50Z

Coverage decreased (-1.9%) to 90.597% when pulling bafad26 on refactor_rounding into 50eb702 on master.

coveralls · 2017-01-28T18:48:34Z

Coverage decreased (-1.8%) to 90.698% when pulling bafad26 on refactor_rounding into 50eb702 on master.

coveralls · 2017-01-28T18:48:34Z

Coverage decreased (-1.8%) to 90.698% when pulling bafad26 on refactor_rounding into 50eb702 on master.

dpsanders · 2017-01-28T19:03:19Z

Tests are passing on 0.4 and 0.5. I think this is ready to merge!

coveralls · 2017-01-28T19:19:45Z

Coverage decreased (-1.8%) to 90.698% when pulling 7e67cbd on refactor_rounding into 50eb702 on master.

coveralls · 2017-01-28T19:19:45Z

Coverage decreased (-1.8%) to 90.698% when pulling 7e67cbd on refactor_rounding into 50eb702 on master.

coveralls · 2017-01-28T19:36:52Z

Coverage decreased (-1.8%) to 90.707% when pulling 7e67cbd on refactor_rounding into 50eb702 on master.

coveralls · 2017-01-28T19:36:52Z

Coverage decreased (-1.8%) to 90.707% when pulling 7e67cbd on refactor_rounding into 50eb702 on master.

lbenet

Have you included the documentation for the macros? I think it would be useful in this PR.

I must admit that I feel unconfortable with the current implementation: For x = @biginterval(27), x^(1/3) does not include the true answer! Simply, try it with displaymode(format=:full).

In any case, this is almost ready!

lbenet · 2017-01-28T18:07:22Z

src/intervals/functions.jl

 end

+^(a::Interval{BigFloat}, x::AbstractFloat) = a^big(x)


What about setting it back as it was before?

dpsanders · 2017-01-29T12:35:33Z

The macro documentation is in rounding_macros.jl -- see https://github.com/dpsanders/ValidatedNumerics.jl/blob/7e67cbd3cbf3b895a29c8839221f3f9e0bb7ff3e/src/intervals/rounding_macros.jl

Is that what you were looking for?

Since 1/3 is not the true one third, x^(1/3) doesn't need to contain the "true" answer for the true one-third power.

I see this as analogous to the fact that you can do Interval(1/3) if you want to use the floating-point number closest to the true one third, or use @interval(1/3) if you want to include the true one third.

Similarly you can do x^(1/3) to use the floating-point value 1/3, or e.g. x^(1//3) to use the true real number one third.

dpsanders · 2017-01-29T12:53:35Z

I think we should merge this and can come back and change things later if necessary.

lbenet · 2017-01-30T03:47:43Z

Ok. I'm merging.

dpsanders · 2017-01-30T16:10:15Z

Thanks!

tkelman · 2017-02-01T05:28:46Z

src/intervals/arithmetic.jl

@@ -155,6 +155,25 @@ end

 //(a::Interval, b::Interval) = a / b    # to deal with rationals

+if VERSION >= v"0.6.0-dev"


this won't work for all 0.6-dev versions, should be more specific about version-depedent cutoffs or use isdefined

You mean isdefined(:Iterators)? I don't know how to find the exact cutoff where this change was made. In any case, if it's a dev version isn't the expectation that people will keep updating it to the latest version?

There should be a PR linked to in NEWS or otherwise not too hard to find, and there's the contrib/commit-name.sh script you can use

tkelman · 2017-02-01T05:29:45Z

src/intervals/arithmetic.jl

+
+if VERSION < v"0.5"
+    min(x) = x
+    max(x) = x


since you've imported these, this is type piracy and could change the behavior of other unrelated code

These are just not defined on 0.4. Do you mean that I should change this to VERSION <= v"0.4"?

No, I mean do you need to be extending a base method on base types? If these are backporting something that worked in later versions of Julia, is it available in Compat?

These are indeed backporting something that is in 0.5 onwards. It is not available in Compat, but probably should be. JuliaLang/Compat.jl#313

Updated in master to use the correct precise versions, thanks.

tkelman · 2017-02-01T05:32:07Z

src/intervals/rounding.jl

@@ -0,0 +1,75 @@
+# Define rounded versions of elementary functions


didn't you have PR's to add these to base?

No, just for BigFloat constructors.

dpsanders changed the title ~~Refactor so transforms to~~ Refactor @round to transform a+b to +(a, b, RoundDown) Jan 9, 2017

dpsanders mentioned this pull request Jan 24, 2017

Type instabilities all over the place #189

Closed

dpsanders added 17 commits January 26, 2017 08:04

Proof of concept for generic rounding macros in rounding.jl: + and sin

b677125

More rounding

462e7a1

Define rounding for sqrt

786338f

Up-arrow and down-arrow macros with code re-use

2aae7c2

Add fallback for rounding of non-expressions

7d63258

@Rounding in terms of up-arrow and down-arrow

ae96e46

Remove old definitions

fe65e4c

Start replacing with new version that does not have explicit type T

9cc37f7

More @round conversions

caa20cd

Fix escaping for 0.6

91b8882

Make multiline comment

d32bead

0.6 compatibility for round_string

f4c834e

Add \times for constructing IntervalBox

8e69dec

0.6 deprecation of takebuf_string

4eec8ef

round for max and min; multiplication now works

1b46ed5

Refactor ^

309c41f

0.6 deprecation of invoke

fe40635

dpsanders added 2 commits January 28, 2017 13:19

Include rounding_macros.jl

d76083b

Re-add at-round_up and at-round_down macros

bafad26

dpsanders added 2 commits January 28, 2017 13:53

Allow Travis failures on Julia nightly

27fd40b

Add fallback for hull:

7e67cbd

lbenet reviewed Jan 29, 2017

View reviewed changes

src/intervals/functions.jl

end

^(a::Interval{BigFloat}, x::AbstractFloat) = a^big(x)

Copy link

Member

lbenet Jan 28, 2017

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

What about setting it back as it was before?

lbenet merged commit d059a3b into master Jan 30, 2017

dpsanders deleted the refactor_rounding branch January 30, 2017 17:32

This was referenced Jan 30, 2017

Update macros for 0.6 #184

Closed

Add × for making IntervalBoxes #181

Closed

Incorrect definition of float #174

Closed

@interval should always return an interval #169

Closed

lbenet mentioned this pull request Jan 30, 2017

Eliminate extra convert when creating an interval #170

Closed

tkelman reviewed Feb 1, 2017

View reviewed changes

tkelman mentioned this pull request Feb 1, 2017

Tag ValidatedNumerics.jl v0.6.0 JuliaLang/METADATA.jl#7770

Merged

		@@ -155,6 +155,25 @@ end

		//(a::Interval, b::Interval) = a / b # to deal with rationals

		if VERSION >= v"0.6.0-dev"

		@@ -0,0 +1,75 @@
		# Define rounded versions of elementary functions

Refactor @round to transform a+b to +(a, b, RoundDown) #188

Refactor @round to transform a+b to +(a, b, RoundDown) #188

Conversation

dpsanders commented Jan 9, 2017 • edited Loading

lbenet commented Jan 9, 2017

dpsanders commented Jan 9, 2017

coveralls commented Jan 24, 2017 • edited Loading

codecov-io commented Jan 24, 2017 • edited Loading

Current coverage is 90.70% (diff: 85.90%)

coveralls commented Jan 24, 2017 • edited Loading

coveralls commented Jan 24, 2017 • edited Loading

coveralls commented Jan 24, 2017 • edited Loading

coveralls commented Jan 24, 2017

coveralls commented Jan 24, 2017

coveralls commented Jan 24, 2017

dpsanders commented Jan 26, 2017

dpsanders commented Jan 28, 2017

dpsanders commented Jan 28, 2017 • edited Loading

coveralls commented Jan 28, 2017

coveralls commented Jan 28, 2017

coveralls commented Jan 28, 2017

dpsanders commented Jan 28, 2017

coveralls commented Jan 28, 2017

coveralls commented Jan 28, 2017

coveralls commented Jan 28, 2017

coveralls commented Jan 28, 2017

lbenet left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

dpsanders commented Jan 29, 2017

dpsanders commented Jan 29, 2017

lbenet commented Jan 30, 2017

dpsanders commented Jan 30, 2017

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Refactor `@round` to transform `a+b` to `+(a, b, RoundDown)` #188

Refactor `@round` to transform `a+b` to `+(a, b, RoundDown)` #188

dpsanders commented Jan 9, 2017 •

edited

Loading

coveralls commented Jan 24, 2017 •

edited

Loading

codecov-io commented Jan 24, 2017 •

edited

Loading

coveralls commented Jan 24, 2017 •

edited

Loading

coveralls commented Jan 24, 2017 •

edited

Loading

coveralls commented Jan 24, 2017 •

edited

Loading

dpsanders commented Jan 28, 2017 •

edited

Loading