Signal Processing Components

compander = compressor + expander
compressor =
discriminator = demodulates an EM signal
expander =

Operations

Fourier Transforms are used for signal analysis, whereas Laplace Transforms are use for systems analysis. An FT is not exactly a filter per se. It is used for traveling between the time and frequency domains.

Bode Plots

These plots use a logarithmic scale compression, which is very useful for plotting a wide range of frequency and amplitude.

What is aliasing?

Results if you sample below the Nyquist rate. You have a reconstructed sinusoid that appears to be a lower frequency than the original.

Anti-aliasing is a band limit to the input signal. You apply a band-stop/band-reject filter, setting a max past which rapid attenuation occurs. In graphics, this results in an image that is "smoothed" on the high-frequency jagged edges.

What is convolution?

An integral at the core. An operation on two functions to produce a third function that expresses how the shape of one is modified by the other. The term convolution refers to both the result function and to the process of computing it. It is defined as the integral of the product of the two functions after one is reversed and shifted. Another way to describe this is to pass two plots through each other.

When you convolve two square pulses, you end up with a triangular form.

What is oversampling?

Sampling at much higher than the Nyquist rate (> than 2x highest signal freq). This offers advantages such as separation in frequency between spectral images. Oversampling allows quantizing with fewer bits.

DSP Chips

TI is the biggest name traditionally, started in audio processing. These chips are more efficient and focus on loop processing. Computation-intensive operations include analytics, FFTs, matrices. They can have unique memory structures as the designers navigate performance vs. memory tradeoffs. The applications for the chips are in biometrics, machine vision, audio and video processing.

In general, the DSP process is to take a real world analog signal and measure->filter->compress. DSP code is often written in assembly.

What makes DSPs awesome, what they do best is multiply-accumulate operations (a = a + (b*c))
Architecture types include SIMD (single instruction, multiple data) and VLIW (very long instruction word for doing more concurrently).

A Harvard architecture allows fetching of data and instructions simultaneously because they have separate storage.

Loop processing acceleration techniques:

simultaneous execution
overlapping loop iterations
SW pipeline
loop unrolling (copy the loop for iterator overlap)