Sunday, March 9, 2014

Two At Twice The Price

There's a line in "Contact", spoken by Hadden, that I have taken as a personal mantra:

"First rule in government spending; why build one when you can have two, at twice the price."

That's just before his other immortal line. (And just after some really bad zero-G special effects)

A good friend of mine who studies governments once told me that while it sounds stupidly obvious, governments often alter the price of a service merely by trying to use it, and it doesn't always get cheaper in bulk. If you're in government, "two at twice the price" is a bargain.

The same thing happens in computer programming, for much the same reason - overheads. It's not just the transactions, the framework containing them has to be accounted for.

Here's a little hint that has made vast wodges of my programming more efficient: don't pass one, when you can pass an array.

A case in point would be my new Raster classes: The 'primitive' method for modifying the screen contents is a function called fragment() which takes a linear array of pixels, and paints the strip to the surface. My equivalent of drawPixel() stuffs the colour value into a one-pixel array, and then calls fragment().

Therefore, drawing a single pixel is actually a special case of drawing n pixels. Not the other way around.

So what? The difference is between calling a single function that internally loops around, (keeping all its state intact) or having a loop that calls the function n times. Those function calls aren't free, you know. And you'll rarely update only one pixel total. That's why my font routines are twice as fast.

Write the array case first. Optimize it. Then build the easy-use single-case version on top of that, because if the user is yanking single entries they don't care about performance. If they care, they'll be calling the array version to fast-fill buffers - and if that is merely a loop that calls the single case over and over, you've fooled them good and lost an opportunity for real performance.

Don't do one when you can do n at n times the price. Because the price gets worse when you lose scope or context, so build a 'vector machine' which amortizes the scope costs over as much data as possible. That's how Seymour Cray rolled.

No comments:

Post a Comment