[Chart] Fwd: [diagrams] Re: [GSoC 2013] Progress - Porting Charts to Diagrams

Brent Yorgey byorgey at seas.upenn.edu
Tue Jul 2 12:56:53 BST 2013


On Tue, Jul 02, 2013 at 11:17:57AM +0200, Jan Bracker wrote:
> 
> It seems like cabal does not support the separate compilation of
> executable. So we have to split the test package to manage dependencies. 

That's not true.  (Assuming I understand what you are saying.) 		   You
can use some (standard) tricks with cabal flags	  to make		   this work.
For example, see the diagrams-builder package, which has several
executables (each with different dependencies) but you control which
are installed via flags, and you only incur the	   dependencies of	the
ones you request to be installed.  The idea is to make a flag like

flag cairo
  description: install cairo-specific builder tool
  default:     False
  manual:      True

("default" means how the flag is set if	the user does not explicitly
specify; "manual" means that the cabal s

and then in the
executable section you make both 'buildable' and the dependencies
conditional on the flag, like this:

executable diagrams-builder-cairo
  main-is:             diagrams-builder-cairo.hs
  hs-source-dirs:      src/tools
  default-language:    Haskell2010
  other-extensions:    DeriveDataTypeable
                       RecordWildCards

  if !flag(cairo)
    buildable: False

  if flag(cairo)
    build-depends:     base >= 4 && < 5,
                       filepath,
                       directory,
                       diagrams-builder,
                       diagrams-lib >= 0.6 && < 0.7,
                       diagrams-cairo >= 0.6 && < 0.7,
                       cmdargs >= 0.6 && < 0.11

When building, if you want to turn on a specific flag you do something like

  cabal install -fcairo diagrams-builder

-Brent

> 
> The most current stable version of the port can be found here [0] in
> commit 2e2277b4a65dc426d2da2a92b4c27eaa6bc11ecd .
> 
> Greets
> Jan
> 
> [0]:
> https://github.com/jbracker/haskell-chart/tree/2e2277b4a65dc426d2da2a92b4c27eaa6bc11ecd
> 
> 
> 2013/6/25 Jan Bracker <jan.bracker at googlemail.com>
> 
> > Hello everybody,
> >
> > the last week brought some major changes. The current stable version is
> > here [0]. All changes are in the master branch of my fork. What has been
> > done / changed:
> >
> > - Implemented a new type for 'Path' that supports to be closed. Like this
> > 'fillPath' and 'strokePath' do not need a boolean flag anymore. This should
> > also make future changes easier, since the 'Path' type is abstract.
> > - Cleaned up the code and added documentation for the 'Backend',
> > 'Geometry' and 'Drawing' functions.
> > - 'withClipRegion' now works in a hierarchical manner. You can only make
> > the clipping region smaller and you can only clip on areas in your current
> > clip. The clip region is represented by a 'Limit Rect'. This properly
> > reflects the possibilities that the clip region can be empty 'LMin' or an
> > infinite plane 'LMax'.
> > - Removed the associated 'ChartOutput' type from the 'ChartBackend' class.
> > Instead backend modules now deliver functions specific to their backend.
> > - Removed all function with legacy names and switched the complete library
> > to use the new API and functions.
> > - Marked all 'default...' values as deprecated to push the transition to
> > Data.Default.
> > - Removed the 'Types' module again and distributed all contents into
> > 'Backend', 'Geometry' and 'Drawing'.
> > - Moved the cairo backend and the tests to their own packages (
> > chart-cairo and chart-tests ).
> > - Updated/Fixed the 'chart-gtk' package to work together with all changes.
> > Its tests are now also in the chart-tests package.
> >
> > Some questions that I have:
> >
> > - When implementing the new version of 'withClipRegion' I stumbled upon
> > the question how the empty clip region looks in cairo. Right now I use the
> > empty 'Rect' at location (0,0) to represent this. This seems to work fine.
> > - Right now the 'Renderable' class is commented out. Should I try to
> > reintroduce it? Is this really a good Idea or should we leave it away all
> > together? None of Charts uses it anymore.
> >
> > My plan for the next week would be:
> >
> > - Implement a test that only uses the backend API so we have something
> > more basic to check if the backends work as expected. This will also show
> > if the current documentation is specific enough.
> > - Try to bring 'Renderable' back into scope (If that is a good idea)
> >
> > The next big step after this would be to merge all the changes back into
> > the main Charts repository and start developing the Diagrams backend.
> >
> > Greets
> > Jan
> >
> > [0]:
> > https://github.com/jbracker/haskell-chart/tree/dcd35c767c69db40e1494608029736ae60148378
> >
> >
> > 2013/6/17 Jan Bracker <jan.bracker at googlemail.com>
> >
> >> Hello everybody,
> >>
> >> I made good progress over the last week. I implemented the minimal chart
> >> drawing API (  http://hpaste.org/89758 ) that Tim suggested. It worked
> >> with minor problems. I tried to stay pixel equal to the PNGs that the
> >> current release produces when running the test suite. I did not manage to
> >> stay 100% pixel equal, but all differences are not visible with the naked
> >> eye.
> >>
> >> Here the changes I had to apply to the API:
> >>
> >> - TextSize also contains the Y-bearing since it is used in some of the
> >> text adjustment calculations.
> >> - withSourceColor does not exists any more. The line style provides the
> >> color for stroking, fill style the color for filling and font style the
> >> color for fonts. There are several utilities to aid implementors of
> >> backends with keeping the local environment up to date.
> >> - The strokePath and fillPath functions are now backendStrokePath and
> >> backendFillPath. They take an extra parameter that tells them if the given
> >> path shall be close before rendering. This was necessary to produce equal
> >> output. Convenience functions that default to not closed paths are provided.
> >> - The matrix implementation is basically a copy of the matrix
> >> implementation from cairo.
> >> - The withClipRegion function now takes a Maybe Rect. I thought this
> >> might be useful, because otherwise there would be no possibility to remove
> >> the clip region. Though this is not needed anywhere in charts yet.
> >> - I had to add the fillClip function, because otherwise there would have
> >> been no possibility to fill the clip region if it is not set (infinite
> >> plane clip).
> >> - showText is now drawText and also takes a Point so it know where to
> >> draw the text. The point refers to the left end of the text baseline.
> >> - The drawing module now contains all kinds of higher level drawing
> >> functions and helpers. There are still a lot of legacy functions with
> >> temporary names. I will remove these soon.
> >>
> >> This commit is "stable" [0]: bdb628591132964e94f5b317026605790250359a
> >>
> >> Plans I would like to hear thoughts to:
> >>
> >> - Should I prefix the backend functions and provide wrapper functions
> >> with the nice name? I already did this for backendFillPath since I noticed
> >> that it also needed to know if the path has to be closed, but in 95% of all
> >> cases the version that does not close by default is more practical. Like
> >> this we would also be able to change the backend interface without having
> >> to change several other modules.
> >> - Should I keep the Maybe in withClipRegion? As said nothing in charts
> >> actually ever removes the clip region.
> >>
> >> The next steps would be:
> >>
> >> - Remove legacy functions.
> >> - Document all the drawing functions.
> >> - Create a separate package for the Cairo backend. I don't know if we
> >> should merge with the main repo before this or afterwards.
> >> - There are still some direct connections to the Cairo backend in
> >> Simple.Internal we have to think about how these can be removed.
> >> - I would also suggest to create a separate package for the tests since
> >> that would make dependency management easier.
> >>
> >> Greets
> >> Jan
> >>
> >> [0]:
> >> https://github.com/jbracker/haskell-chart/tree/bdb628591132964e94f5b317026605790250359a
> >>
> >>
> >> 2013/6/13 Jan Bracker <jan.bracker at googlemail.com>
> >>
> >>> I will proceed with 3 then.
> >>>
> >>>
> >>> 2013/6/12 Tim Docker <tim at dockerz.net>
> >>>
> >>>>  Hi,
> >>>>
> >>>> I think your analysis is good. In the context of the backend, it makes
> >>>> sense to separate out the color from the various styles.
> >>>>
> >>>> On the other hand, in the context of the chart use API, it makes sense
> >>>> to have all of the properties of a line in a single data type - it's
> >>>> simpler for the caller to say style this plot line using properties as per
> >>>> this value.
> >>>>
> >>>> Possible solutions are:
> >>>>
> >>>>     1) Leave it and put up with the inconsistency...
> >>>>     2) Have separate types in the backend API / user API for LineStyle,
> >>>> FillStyle and FontStyle
> >>>>     3) Remove withSourceColor from the backend API, and require
> >>>> backends to give the logical behaviour (ie setting the FillStyle doesn't
> >>>> affect the color of lines). This would mean the cairo backend would have to
> >>>> do a bit more work to be correct.
> >>>>
> >>>> Either 2 or 3 is ok with me.
> >>>>
> >>>> Tim
> >>>>
> >>>>
> >>>>
> >>>> On 13/06/13 00:45, Jan Bracker wrote:
> >>>>
> >>>> Hi everybody,
> >>>>
> >>>>  a small follow up: While thinking about and implementing the smaller
> >>>> API I noticed that due to the internals of Cairo the source color is used
> >>>> to render everything. When setting a style (font, fill, line) the source
> >>>> color is changed according to that style. This means each of the
> >>>> "with...Style" or "withSourceColor" functions constantly overrides the
> >>>> source color. Is this really desired behaviour? Shouldn't the fill color
> >>>> apply to fills, the line color to strokes and the font color to text
> >>>> rendering? If so the source color ("withSourceColor") is a useless concept.
> >>>> Of course keeping these color seperate will expand the cairo rendering
> >>>> backend, but I think it is useful to avoid confusion to which color is used
> >>>> where.
> >>>>
> >>>>  Example:
> >>>>
> >>>>  1: withLineStyle ls $ withFillStyle fs $ fillPath p >> strokePath p
> >>>>
> >>>>  should be the same as (according to my thoughts):
> >>>>
> >>>> 2: withFillStyle fs $ withLineStyle ls $ fillPath p >> strokePath p
> >>>>
> >>>>  but with the current implementation 1 would render both with the fill
> >>>> color and 2 would render them with the line color.
> >>>>
> >>>>  Jan
> >>>>
> >>>>
> >>>>
> >>>> 2013/6/12 Jan Bracker <jan.bracker at googlemail.com>
> >>>>
> >>>>> Hi everybody,
> >>>>>
> >>>>>  Brent: Thanks you for the offer and thanks for your great help on
> >>>>> those nasty type errors yesterday.
> >>>>>
> >>>>>  On IO in Diagrams: We looked into the SVGFont rendering stuff and
> >>>>> noticed that it has a call to unsafePerformIO (which let us scream in pain
> >>>>> and agony :-P ). We though it would be nice to remove this hidden call and
> >>>>> actually offer a function that makes it explicit. I am not sure why it is
> >>>>> there and assume its purpose is to keep everything pure. Your suggestion to
> >>>>> produce a IO Diagram is what we thought of too. We are not sure which
> >>>>> changes would be required up to now, but we will come back to you when we
> >>>>> do know.
> >>>>>
> >>>>>  On general note: I have worked on abstracting the backend the last
> >>>>> two days. Here what I accomplished by now:
> >>>>>
> >>>>>  - I have fully converted my branch to use the abstract rendering
> >>>>> backend ChartBackend [0] instead of the CRender monad directly.
> >>>>>
> >>>>>  - There is a implementation of the backend with Cairo in the
> >>>>> Backend.Cairo module [1].
> >>>>>
> >>>>>  - The current HEAD of the backend-experiment branch [2] is stable
> >>>>> (commit ad715c4d216e8fcb494f9919d4fb5459ce7d85c2).
> >>>>>
> >>>>>  - I added reference images [3] of all tests to compare with the
> >>>>> results that the original code delivered.
> >>>>>
> >>>>>  - All tests are running using the new backend code and pixel-diffs
> >>>>> [4] of the produced PNGs show that they render exactly the same.
> >>>>>
> >>>>>  - I ran into the issue that the ToRenderable class was complicated
> >>>>> to abstract. My abstraction involved a associated type with two parameters
> >>>>> that are in relation to each other [5,6]. After talking with Tim I removed
> >>>>> the ToRenderable code completely and instead just exported specific
> >>>>> functions for all types directly. This also made the examples easier to
> >>>>> adjust, since the new ToRenderable class made type annotations necessary
> >>>>> everywhere, which, in my point of view, defeats the purpose of the
> >>>>> toRenderable function. It was easier just calling the specific function
> >>>>> directly. I think the change of API is acceptable, since the generalised
> >>>>> backend lets type parameters popup everywhere anyway.
> >>>>>
> >>>>>  The next steps:
> >>>>>
> >>>>>  - Tim suggested this simplified API instead of the rather extensive
> >>>>> [0] version used right now: http://hpaste.org/89758 . I will work on
> >>>>> trimming down the current API and providing all complex functions in terms
> >>>>> of the new API.
> >>>>> - The only downside I see is the Matrix type which (I assume) comes
> >>>>> from Cairo.Matrix [7]. I would just give a naive implementation that
> >>>>> contains 6 double values for the affine transformations and provide a
> >>>>> similar custom API.
> >>>>> - I will also try to provide a small EDSL for writing paths to make
> >>>>> that easier (I guess some short combinators and a monoid instance will do
> >>>>> fine).
> >>>>>
> >>>>>  Jan
> >>>>>
> >>>>>  [0]:
> >>>>> https://github.com/jbracker/haskell-chart/blob/backend-experiment/chart/Graphics/Rendering/Chart/Types.hs#L85
> >>>>> [1]:
> >>>>> https://github.com/jbracker/haskell-chart/blob/backend-experiment/chart/Graphics/Rendering/Chart/Backend/Cairo.hs#L80
> >>>>> [2]:
> >>>>> https://github.com/jbracker/haskell-chart/tree/ad715c4d216e8fcb494f9919d4fb5459ce7d85c2
> >>>>> [3]:
> >>>>> https://github.com/jbracker/haskell-chart/tree/backend-experiment/chart/tests-reference
> >>>>>  [4]:
> >>>>> https://github.com/jbracker/haskell-chart/blob/backend-experiment/chart/gen-diffs-png.sh
> >>>>>  [5]:
> >>>>> https://github.com/jbracker/haskell-chart/blob/backend-experiment/chart/Graphics/Rendering/Chart/Renderable.hs#L77
> >>>>> [6]:
> >>>>> https://github.com/jbracker/haskell-chart/blob/backend-experiment/chart/Graphics/Rendering/Chart/Layout.hs#L163
> >>>>> [7]:
> >>>>> http://hackage.haskell.org/packages/archive/cairo/0.12.4/doc/html/Graphics-Rendering-Cairo-Matrix.html
> >>>>>
> >>>>>
> >>>>> 2013/6/11 Brent Yorgey <byorgey at seas.upenn.edu>
> >>>>>
> >>>>>> Hi Jan and all,
> >>>>>>
> >>>>>> Looks like some great progress!  I'm excited. =) Just a couple
> >>>>>> thoughts: one, let me know if you want help staring at the funny type
> >>>>>> errors with Legend and whatnot.  I'm not familiar with the Chart code
> >>>>>> but perhaps I can help figure out a good solution for the abstraction.
> >>>>>>
> >>>>>> My other general comment is that I know you and Tim have talked a bit
> >>>>>> about the problems with having font stuff in IO whereas diagrams is
> >>>>>> pure.  However, from my point of view it wouldn't be that big of a
> >>>>>> deal if Chart produced an IO Diagram rather than just a Diagram.
> >>>>>> Maybe I haven't thought about it sufficiently but it seems like this
> >>>>>> would be an easy solution to the problem of needing IO.  I'm happy to
> >>>>>> talk about it in more detail if you want, and if there are ways that
> >>>>>> the diagrams API could change to make any of this easier I'm happy to
> >>>>>> discuss that as well. I'm in the middle of some big refactoring anyway
> >>>>>> (though don't worry, I don't think it will affect your work on porting
> >>>>>> Chart very much).
> >>>>>>
> >>>>>> -Brent
> >>>>>>
> >>>>>> On Mon, Jun 10, 2013 at 05:24:34PM +0200, Jan Bracker wrote:
> >>>>>> > Hi,
> >>>>>> >
> >>>>>> > Andy: The goal is to abstract Charts from Cairo so we can plug in
> >>>>>> another
> >>>>>> > backend (target is Diagrams). And by that we would also be able to
> >>>>>> use
> >>>>>> > Sunroof through diagrams. Of course in the end it should also be
> >>>>>> possible
> >>>>>> > to implement a Sunroof backend on its own.
> >>>>>> >
> >>>>>> > Tim:
> >>>>>> >
> >>>>>> >  > Rather than have the CRender monad take an extra parameter, I
> >>>>>> guess one
> >>>>>> > > could just have a separate data type for each backend (ie
> >>>>>> CRenderCairo and
> >>>>>> > > CRenderDiagrams) and have the drawing code abstracted over a
> >>>>>> typeclass
> >>>>>> > > implemented by both. Not sure if this would be better...
> >>>>>> > >
> >>>>>> >
> >>>>>> >  I tried your approach in a new branch:
> >>>>>> > https://github.com/jbracker/haskell-chart/tree/backend-experiment
> >>>>>> >
> >>>>>> > It seems to be working quite well. So I started generalizing
> >>>>>> everything
> >>>>>> > outside of the cairo module to use a general ChartBackend m instead
> >>>>>> of a
> >>>>>> > CRender to draw everything. There are a few spots that are giving
> >>>>>> me a hard
> >>>>>> > time:
> >>>>>> >
> >>>>>> > Renderable was implemented in terms of CRender. So I added another
> >>>>>> type
> >>>>>> > parameter and that worked fine until I hit Legend. There I also
> >>>>>> added a
> >>>>>> > type parameter. But that blew up the ToRenderable instance. So I
> >>>>>> had to
> >>>>>> > insert a associated type to relate types to each other. After doing
> >>>>>> so I
> >>>>>> > was not able to fix whatever went wrong in Plot.Pie. The
> >>>>>> 'ToRenderable
> >>>>>> > PieLayout' instance gives me this error:
> >>>>>> >
> >>>>>> > Graphics/Rendering/Chart/Plot/Pie.hs:121:62:
> >>>>>> >     Could not deduce (RenderableT m a0 ~ PieChart)
> >>>>>> >     from the context (ChartBackend m)
> >>>>>> >       bound by the type signature for
> >>>>>> >                  toRenderable :: ChartBackend m =>
> >>>>>> >                                  RenderableT m PieLayout ->
> >>>>>> Renderable m ()
> >>>>>> >       at Graphics/Rendering/Chart/Plot/Pie.hs:(117,5)-(125,29)
> >>>>>> >     The type variable `a0' is ambiguous
> >>>>>> >     Possible fix: add a type signature that fixes these type
> >>>>>> variable(s)
> >>>>>> >     In the return type of a call of `pie_plot_'
> >>>>>> >     In the second argument of `($)', namely `pie_plot_ p'
> >>>>>> >     In the second argument of `addMargins', namely
> >>>>>> >       `(toRenderable $ pie_plot_ p)'
> >>>>>> >
> >>>>>> > I think I know what is going wrong, but I have no idea how to
> >>>>>> specify the
> >>>>>> > instance it should use there. Maybe I am just to tired to see the
> >>>>>> solution.
> >>>>>> >
> >>>>>> > Right now I am trying to make things work without extra type
> >>>>>> parameters,
> >>>>>> > though the definitions that are based on CRender are giving me a
> >>>>>> hard time,
> >>>>>> > because then I end up with rank N types.
> >>>>>> >
> >>>>>>   > --
> >>>>>> > You received this message because you are subscribed to the Google
> >>>>>> Groups "diagrams-discuss" group.
> >>>>>> > To unsubscribe from this group and stop receiving emails from it,
> >>>>>> send an email to diagrams-discuss+unsubscribe at googlegroups.com.
> >>>>>> > For more options, visit https://groups.google.com/groups/opt_out.
> >>>>>> >
> >>>>>> >
> >>>>>>
> >>>>>
> >>>>>
> >>>>
> >>>>
> >>>
> >>
> >



More information about the Chart mailing list