Each area of the viewing image is independent, so let's divide up the viewing surface into a number of independent buffered images. We have a little helper function which given a width, height and tile size gives us back a list of positions to draw at.
(defn create-work-list [width height unitX unitY] (let [xs (range 0 width unitX) ys (range 0 height unitY)] (mapcat (fn [x] (mapcat (fn [y] (list (list x y))) ys)) xs)))
For example, if we want to draw use a tile size of 150x150 with a viewing surface of 300x300 we get the following offsets:
user> (create-work-list 300 300 150 150) ((0 0) (0 150) (150 0) (150 150))
Now we should make the
ray-tracefunction take notice of these co-ordinates. The previous version of the ray-trace function wasn't very functional as it performed IO (drawing to the canvas). Side-effects are the enemy of referential transparency.
(defn ray-trace [world w h ox oy] (let [buffered-image (BufferedImage. w h BufferedImage/TYPE_BYTE_GRAY)] (doseq [x (range 0 (dec w))] (doseq [y (range 0 (dec h))] (.setRGB buffered-image x y (color-at (+ x ox) (+ y oy))))) buffered-image))
Now the function is pure because it always gives the same output for the same input and no IO occurs.
Finally, we need to distribute this work across all the cores. Clojure has a built in function, pmap that works the same as map but in parallel.
Initially I performed the drawing to the canvas within a
pmapexpression - this is very wrong!
pmapis lazy, it's not evaluated unless it is needed. Clojure doesn't know that I intended that to always been evaluated, so it was only ever drawing the first four tiles (presumably because that's how many cores my desktop has).
If you want to force the evaluation there's a number of functions which do that:
pmapexpression to produce a list of images, and then use
doseqto perform the side effects.
(def canvas (proxy [JPanel]  (paintComponent [g] (proxy-super paintComponent g) (.setColor g Color/RED) (let [width (.getWidth this) height (.getHeight this) unitX (/ width 16) unitY (/ height 16) work-list (create-work-list width height unitX unitY)] (doseq [image (pmap (fn [pos] (list (apply ray-trace (list world unitX unitY (first pos) (second pos))) (first pos) (second pos))) work-list)] (.drawImage g (first image) (second image) (nth image 2) unitX unitY nil))))))The separation of IO and pure functions is something advocated in "A wish list for the next mainstream programming language". Clojure doesn't force this, whereas something like Haskell does. Haskell uses monads to achieve this, which is something I'll visit at some point. See LtU for some explanations. Ok, enough theory - what difference does this actually make? Well for me, about a 4x difference, and I have 4 cores, so that's good! Timing's aren't that exciting, but you can see the difference with the system monitor.
Pretty much full CPU utilization across all cores, with only a few lines of code changed, funky!