Digital Signal Processing/DSP Programming

Digital Signal Processing units, on first glance, are very similar to other programmable microprocessors, microcontrollers, and FPGAs. However, DSP chips frequently have certain features and limitations involved that other categories of chips don't have. This chapter will attempt to explain, in broad language-neutral terms, some of the issues involved with DSP programming.

Saturation Arithmetic
Intel-compatible microprocessors, something that most programmers are familiar with, have several features that people have come to expect from programmable chips. In particular, Intel-compatible chips have an arithmetic mode known as "Roll-Over Arithmetic", also called modular arithmetic. Let's say that we have the integer number 250 (decimal) stored in a byte-sized register. If we add 10 to that number, the processor will overflow, and will roll-over the answer, giving us the answer 4. However, DSP chips frequently do not roll-over, and instead saturate to the maximum value.

Let's say we are processing a grayscale image. This image has byte-sized integer values of white/black value as such: Pure white is 255, and pure black is 0. Now, let's say we want to "brighten" every pixel by 10 points, to improve visibility. So we create a loop, and we go through each pixel in the image, adding 10 to the value at any particular point. But what happens if we add 10 to a pixel that is already pure white? In a regular Intel-compatible chip, the value 255 + 10 = 9. In effect, if we make the white pixels brighter, we are turning them black! now, to avoid this, DSP chips saturate, and will go up to the highest value (255), and won't roll over. So, on a DSP chip, white can't get any whiter. Here are some examples (using unsigned bytes):


 * $$255 + 10 = 255$$
 * $$0 - 10 = 0$$

Multiply and Accumulate (MAC)
In difference equations, we have seen that multiply and addition operations are the most common. Let's look at a general example:


 * $$A x_n + B x_{n-1} + C x_{n-2} + D x_{n-3} + ...$$

Now, we notice that each element in this equation is multiplied by a coefficient, and added to the total. DSP engineers have noticed this pattern, and have optimized DSP chips to be able multiply and add very quickly (at the expense of other operations missing or slow). In fact, many DSP chips will have instructions known as multiply and accumulate, which will perform both operations simultaneously, much quicker than a normal processor could do.