added a lot of comments

This commit is contained in:
Yann Esposito (Yogsototh) 2013-01-11 15:42:28 +01:00
parent 74aad30fb6
commit b461b33eba

View file

@ -1,17 +1,53 @@
# Flame fractal in Haskell
First computing the flame fractal is difficult to do in parallel.
You can take a look at [electricsheep.org](http://electricsheep.org)
to view some examples.
Here is literate Haskell.
To start your program in Haskell you have to first declare all you imports.
I personally find the number of import annoying.
Not really a language issue but more a usage issue.
For example, instead of the actual
> module Main where
> import Data.Hashable
> import Data.HashMap as Dict -- cabal install HashMap
> import Data.Hashable
> import Data.Maybe as Maybe
> import Data.Word (Word8)
>
> import Codec.Picture -- cabal install juicyPixels FTW
> -- I need to write picture files
> -- I also prefer to declare my own Pixel data type
> import Codec.Picture hiding (Pixel) -- cabal install juicyPixels FTW
> import Control.Monad
> import Control.Monad.State
> import System.Environment (getArgs)
>
I would have preferred a more concise alternative such as:
~~~
import Maybe, Word, Picture, System
import HashMap as Dict
~~~
Not a big deal thought.
Now, instead of using common types like `(Int,Int)` for point,
I prefer to use the power of Haskell types to help me discover errors.
Therefore I create a type for each of the element I need.
I will need a type for 2D points, Colors, extended colors (added color with the number)
Furthermore for efficiency reason I'll use "low level" Haskell.
I will replace `data Point = P Int Int` by an unboxed strict variant.
As it is a GHC optimization it is far more verbose to declare.
We tell GHC to unbox our type using the `{-# UNPACK #-}` comment before each field.
Furthermore to make each field _strict_ we use a `!` before the type.
> -- Data types
> -- Global argument passed to most functions
> data Global = Global { imgWidth :: Int
> -- Global state
> data Global = Global {
> filename :: String
> , imgWidth :: Int
> , imgHeight :: Int
> , nbPoints :: Int }
>
@ -19,45 +55,64 @@
> data Point = P {-# UNPACK #-} !Float
> {-# UNPACK #-} !Float
>
> data Color = Color {-# UNPACK #-} !Word8
> {-# UNPACK #-} !Word8
> {-# UNPACK #-} !Word8
>
> data ExtColor = ExtColor {-# UNPACK #-} !Int
> {-# UNPACK #-} !Int
> {-# UNPACK #-} !Int
> {-# UNPACK #-} !Int
>
> data YPixel = YPixel {-# UNPACK #-} !Int
A Pixel, just a position not the color.
It is mostly like a Point but with integer.
This way I won't mess up between the two representation (on screen) and
into the mathematical plane.
> data Pixel = Pixel {-# UNPACK #-} !Int
> {-# UNPACK #-} !Int
> deriving (Eq,Ord)
I need to talk a bit about the color data structure I use.
Instead of simple RGB each color coded on 8bit.
I need something a bit more complex.
Each time a point will be lighten I'll give it a color.
But some point can be lighten multiple times and in different ways.
Then each time I light the same point I add the new light color to this point.
But in order to retrieve the right brightness,
I remember how many times the point was lighten up.
Thus this strange structure for color.
Note that for drawing the image I'll use a more standard `PixelRGB8` which is
simply three `Word8` values.
> data Color = Color {-# UNPACK #-} !Int
> {-# UNPACK #-} !Int
> {-# UNPACK #-} !Int
> {-# UNPACK #-} !Int
My hash key will be pixel, I then need to make my Pixel data type
an instance of Hashable.
> instance Hashable Pixel where
> hashWithSalt n (Pixel x y) = hashWithSalt n (x,y)
Now I can use `Map` with `Pixel`s as key:
> type YMap = Map Pixel Color
I want to be able to add two `Color` and
to translate to `PixelRGB8` colors which are used to make the image.
> addColor (Color r g b n) (Color r' g' b' n') =
> Color (r+r') (g+g') (b+b') (n+n')
>
> instance Hashable YPixel where hashWithSalt n (YPixel x y) = hashWithSalt n (x,y)
>
> type YMap = Map YPixel ExtColor
>
> addExtColor (ExtColor r g b n) (ExtColor r' g' b' n') =
> ExtColor (r+r') (g+g') (b+b') (n+n')
>
> cmap f (Color r g b) = Color (f r) (f g) (f b)
> ecmap f (ExtColor r g b n) = ExtColor (f r) (f g) (f b) n
>
> gammaCorrection :: Float -> Color -> Color
> gammaCorrection gamma = cmap (round . (**(1/gamma)) . fromIntegral)
>
> colorToPixelRGB8 :: Color -> PixelRGB8
> colorToPixelRGB8 (Color r g b) = PixelRGB8 r g b
>
> colorFromExt :: ExtColor -> Color
> colorFromExt (ExtColor r g b n) = Color (fromIntegral $ div r n)
> colorFromExt :: Color -> PixelRGB8
> colorFromExt (Color r g b n) = PixelRGB8 (fromIntegral $ div r n)
> (fromIntegral $ div g n)
> (fromIntegral $ div b n)
> -- Basic functions
> neg x = 0-x
>
> rgb :: Word8 -> Word8 -> Word8 -> Color
> rgb r g b = Color r g b
> -- Colors (theme is solarized)
In most of my project I use the solarized theme.
I generally use only a small part of these colors.
But it is simply generally better to use a basic scheme than to use hard colors.
> -- Colors from the theme solarized
> rgb :: Int -> Int -> Int -> Color
> rgb r g b = Color r g b 1
> black = rgb 0 0 0
> base03 = rgb 0 43 54
> base02 = rgb 7 54 66
@ -75,12 +130,13 @@
> blue = rgb 38 139 210
> cyan = rgb 42 161 152
> green = rgb 133 153 0
>
> extend :: Color -> ExtColor
> extend (Color r g b) = ExtColor (fromIntegral r) (fromIntegral g) (fromIntegral b) 1
>
> pixelFromPoint (P x y) = YPixel (round x) (round y)
>
> -- very basic change of representation between point and pixel
> pixelFromPoint (P x y) = Pixel (round x) (round y)
Next I needed some pseudo random number generation.
I don't need real good random number generation inside the Random for now.
> -- PSEUDO RANDOM NUMBER GENERATION
> -- !!!!!!!! DONT WORK ON 32 BITS Architecture !!!!!!!
> nextint n = (a*n + c) `rem` m
@ -89,28 +145,86 @@
> c = 1
> m = 2^32
> -- generate a random sequence of length k starting with some seed
> randlist seed n = take n $ iterate nextint seed
> randlist seed = iterate nextint seed
> -- END OF PSEUDO RANDOM NUMBER GENERATION
>
>
> {-
> - Flame Set
> -
> - S = U_{i} F_i(S)
> -
> - F_i being transformations
> - General form:
> - F = affine . linearcomp [variation] . affine
> - affine is a linear function (x,y) -> (ax+by+c,dx+ey+f)
> - variation is some kind of function with some contraction properties
> ex: (x,y) -> (x,y), (sin x, sin y), etc...
> - linearcomp [f] is a linear composition of functions: (x,y) -> Sum vi*f(x,y)
> -}
>
## The Flame Set
Let $F_i$ be a finite family of functions.
And let consider the set $S$ which is the minimal non trivial set which is
the fixed point of the union of these function.
More precisely, let F_i be transformations.
Then consider the set S such that for all i F_i(S)=S
Or equivalently :
$$ S = U_{i} F_i(S) $$
Consider the set is non trivial (understand non empty), then There is at least
one point in it.
Finding the exact set is a difficult task.
But finding an approximation can be done this way:
Let S_0 = {x} where x is a random point in the unit square.
Let S_1 = x_1 = F_i(x) for a random i
Let S_2 = x_2 = F_j(x_1) for a random j
...
Let S_n = x_n = F_k(x_{n-1}) for a random k
...
Each S_n will be closer to S.
At each step you add another point to S_i.
Also to remove bad initialization we generally don't consider the 20th firsts
steps. And we return only {x_21,....,x_n}.
In order to find only interesting elements we much choose our F_i family wisely.
Here are standard form of the F_i:
F_i = affine . linear_combination [variations] . affine
affine are function of the following form:
affine (x,y) = (ax+by+c , dx+ey+f)
It correspond to a composition of translation, rotation and scaling.
linear_combination are function of the following form:
linear_combination [v_i] = sum p_iv_i
affine have 6 parameters,
linear_combination [x] have length [x] parameters.
And we can use many different fonctions for the variations.
Example of variations:
- (x,y) → (x,y)
- (x,y) → (sin x,sin y)
- (x,y) → (x/r^2,y/r^2)
- (x,y) → (x sin r^2 - y cos r^2, x cos r^2 + y sin r^2)
- (x,y) → ((x-y)(x+y)/r,2xy/r)
Wich are coded here:
> -- Some variations
> vs :: [Point -> Point]
> vs = [ \ (P x y) -> P x y
> , \ (P x y) -> P (sin x) (sin y)
> , \ (P x y) -> let r2 = x*x+y*y in P (x/r2) (y/r2)
> , \ (P x y) -> let r2 = x*x+y*y in P (x*(sin r2) - y*(cos r2)) (x*(cos r2) + y * (sin r2))
> , \ (P x y) -> let r = sqrt (x^2+y^2) in P ((x - y)*(x + y)/r) (2*x*y/r)
> ]
To define affine function a standard usage is to use matrices.
> data Matrice = M Float Float Float Float Float Float
> aff :: Matrice -> Point -> Point
> aff (M a b c d e f) (P x y) = P (a*x + b*y + c) (d*x + e*y +f)
>
If you use the identity variation,
the following functions generate the sierpinsky set.
> -- Some affine functions to generate the sierpinsky set
> -- Equivalent to
> -- sierp = [ \(x,y)->(x/2,y/2)
@ -127,56 +241,68 @@
> 0.5 0.0 0.0
> 0.0 0.5 0.5
> ]
Here are the functions for the fern functions.
> fern :: [ Point -> Point ]
> fern = [ aff $ M
> 0.0 0.0 0.0
> 0.0 0.16 0.0
> , aff $ M
> 0.85 0.04 0.0
> (neg 0.04) 0.85 1.6
> (- 0.04) 0.85 1.6
> , aff $ M
> 0.2 (neg 0.26) 0.0
> 0.2 (- 0.26) 0.0
> 0.23 0.22 1.6
> , aff $ M
> (neg 0.15) 0.28 0.0
> (- 0.15) 0.28 0.0
> 0.26 0.24 0.44
> ]
>
> -- Some variations
> vs :: [Point -> Point]
> vs = [ \ (P x y) -> P x y
> , \ (P x y) -> P (sin x) (sin y)
> , \ (P x y) -> let r2 = x*x+y*y in P (x/r2) (y/r2)
> , \ (P x y) -> let r2 = x*x+y*y in P (x*(sin r2) - y*(cos r2)) (x*(cos r2) + y * (sin r2))
> , \ (P x y) -> let r = sqrt (x^2+y^2) in P ((x - y)*(x + y)/r) (2*x*y/r)
> ]
>
> -- Some final functions
> fs :: [((Int,ExtColor),Point -> Point)]
> fs = [ (( 1,extend red),(vs !! 0) . (fern !! 0))
> , (( 86,extend green),(vs !! 0) . (fern !! 1))
> , (( 95,extend blue),(vs !! 0) . (fern !! 2))
> , ((100,extend yellow),(vs !! 0) . (fern !! 3))
> ]
>
Also in order to zoom on all points we generally add a final transformation
which is applied to all points. It helps zoom on the fractal for example.
> -- Transformation functions
> -- translate
> trans :: (Float,Float) -> Point -> Point
> trans (tx,ty) = aff $ M 1 0 tx 0 1 ty
> -- rotate
> rot :: Float -> Point -> Point
> rot phi = aff $ M (cos phi) (sin phi) 0.0 (neg (sin phi)) (cos phi) 0.0
> rot phi = aff $ M (cos phi) (sin phi) 0.0 (- (sin phi)) (cos phi) 0.0
> -- zoom
> zoom :: Float -> Point -> Point
> zoom z = aff $ M z 0 0 0 z 0
>
As the final function goal is to help the final rendering position,
it seems natural to add the size of the view as parameter.
> -- The final transformation to transform the final result (zoom,rotate,translate)
> final :: Int -> Point -> Point
> final width = trans (w/2,w/2) . zoom (w/10) . rot (neg pi)
> final width = trans (w/2,w/2) . zoom (w/10) . rot (- pi)
> where w = fromIntegral width
>
> sierpset :: Int -> Point -> [Int] -> YMap -> YMap
> sierpset w startpoint rands tmpres =
And now the F_i functions.
As we can see, it is not only a list of functions.
But we add informations to each function:
- a probability to be used,
- a color.
> -- F_i
> fs :: [((Int, Color), Point -> Point)]
> fs = [ (( 1, red), (vs !! 0) . (fern !! 0))
> , (( 86, green), (vs !! 0) . (fern !! 1))
> , (( 95, blue), (vs !! 0) . (fern !! 2))
> , ((100,yellow), (vs !! 0) . (fern !! 3))
> ]
Up until now it was only a verbose babbling.
Here is the heart of our program.
Where the interesting stuff is going on.
For now, this is a rather naive implementation.
Naive in the sense that it doesn't use Monad as helper.
> flameset :: Int -> Point -> [Int] -> YMap -> YMap
> flameset w startpoint rands tmpres =
> if rands == []
> then tmpres
> else
@ -193,39 +319,48 @@
> -- Search the old color
> oldvalue = Dict.lookup savepoint tmpres
> -- Set the new color.
> newvalue = addExtColor col (Maybe.fromMaybe (extend black) oldvalue)
> newvalue = addColor col (Maybe.fromMaybe black oldvalue)
> -- update the dict
> newtmpres = Dict.insert savepoint newvalue tmpres
> in
> sierpset w newpoint (tail rands) newtmpres
>
> sierpinsky :: Int -> Int -> YMap
> sierpinsky w n = sierpset w (P 0.13 0.47) (randlist 0 n) Dict.empty
>
> flameset w newpoint (tail rands) newtmpres
The flame function is just a call to the flameset function with initial values.
Clearly there is something to be done here.
> flame :: Int -> Int -> YMap
> flame w n = flameset w (P 0.13 0.47) (take n $ randlist 0) Dict.empty
A function to read the command line arguments.
> initGlobalParams args =
> Global { imgWidth = read (args !! 0)
> , imgHeight = read (args !! 1)
> , nbPoints = read (args !! 2) }
>
> Global { filename = args !! 0
> , imgWidth = read (args !! 1)
> , imgHeight = read (args !! 2)
> , nbPoints = read (args !! 3) }
The functions needed to transform the dictionary as a picture file.
> imageFromDict :: YMap -> Int -> Int -> Image PixelRGB8
> imageFromDict dict width height = generateImage colorOfPoint width height
> where
> colorOfPoint :: Int -> Int -> PixelRGB8
> colorOfPoint x y = colorToPixelRGB8 $ colorFromExt $
> fromMaybe (extend base03)
> (Dict.lookup (YPixel x y) dict)
> colorOfPoint x y = colorFromExt $
> fromMaybe base03 -- background color
> (Dict.lookup (Pixel x y) dict)
>
> writeImage :: Int -> Int -> Int -> YMap -> IO ()
> writeImage w h n dict = writePng "flame.png" $ imageFromDict dict w h
> writeImage :: String -> Int -> Int -> Int -> YMap -> IO ()
> writeImage filename w h n dict = writePng filename $ imageFromDict dict w h
>
> main :: IO ()
> main = do
> args <- getArgs
> if (length args<3)
> then print $ "Usage flame w h n"
> if (length args<4)
> then print $ "Usage flame ficname w h n"
> else do
> env <- return (initGlobalParams args)
> fic <- return (filename env)
> w <- return (imgWidth env)
> h <- return (imgHeight env)
> n <- return (nbPoints env)
> writeImage w h n (sierpinsky w n)
> writeImage fic w h n (flame w n)