AVR Freaks

As long as the corner frequency of the filter is below about 1/10 of the PWM repetition frequency, it makes little difference as far as voltage "accuracy". Of course, the lower you make the filter corner frequency, the slower it is to step changes in PWM duty cycle.

I would not spend much time on getting the filter "just so" unless you are trying to do something like recover audio from a PWM signal. THEN, it is important to get the filter corner as close as you can to the PWM repetition frequency.

By the way, what is a "valid" filter depends, in huge measure, on the allowable voltage ripple in the output. This is NOT the filter passband ripple, but is the amplitude ripple you can see with an oscilloscope. It is synchronized with the PWM period. A side note: the voltage ripple is worst when the duty cycle is 50%.

Jim

I think it will work. You get a mouse report every 5ms, so you are 'sampling' mouse position at 200 hz and updating the pwm every 5ms, so if you have a 100hz 2 section lo pass rc filter, the x box will think you have your thumb on the lever.

The good old fashioned way would go along the following line (I may have omitted some parts, soldering 0603's for four hours instead of sleeping).

* Get impulse response of your RC filer.

* Construct a piecewise input signal with duty cycle DTa for t < tx and with duty cycle DTb for t >= tx. Setting tx = 0 might be a good idea.

* Laplace transform.

* Multiply signal and impulse response in the frequency domain. (multiplication of the Laplace transforms is equivalent to convolution in the time domain).

* Inverse Laplace transformation. This then gives the transfer function.

* Define a criteria for what constitutes settling (e.g. when the output for t >= ts is always within 99,9% of desired value -> settled). When doing so, take LSB of DAC into account. Assuming you want to get 1 LSB resolution then your criteria would be lower than the equivalent of the voltage the LSB of the DAC represents.

* Construct inequation from criteria and transfer function.

* Reorder inequation to get a function of time t (or maybe of delta_t = ts - tx).

* Find maximum of that function of time t (differentiate).

At that point you might consider that the lazy way - simulation - might have been faster to get a good enough answer.

When you yank the joystick over, the pwm goes to 100% and you want the volts to get there real fast. It takes five RC time constants for the cap to get charged up to 99%. I think we need a 100hz filter to track the 200 hz updates, so the time const is 10ms. 5 of those is 50ms (20hz). The volts will get to full scale 50ms after you yank the joystick. Good enough?

Me thinks you are making this much harder than it really is.

Spice, such as the free LTSpice, does wonders for visualizing what happens.

Jim

You want the visual feedback to be indistinguishable when comparing human digit on xbox joystick to avr sensing some voltage, adjusting the pwm, and feeding the voltage to the joystick input. I claim 50ms or less is visually undetectable. Thats 20 somethings a sec... fields, frames, whatever. I'd be surprised if the xbox responded to a step function on the joystick within one video frame... 30ms or so. If the mouse is reporting at 200hz, and the filter is half of that, 100hz or 10ms time const, it should settle to 99% of its avg value in 50ms, as I claimed in an earlier message. If the filter is 10hz, it will be a lot smoother. I'd also try it with a 1khz filter. It will have little 200hz wiggles in it, but there might be a window in the xbox program that smooths this out (defend against elcheapo noisy joysticks).

One additional rule of thumb to consider:

If a PWM signal is modulated (to create a lower frequency analog signal), then the lowpass smoothing filter affects the analog signal in EXACTLY the same way as if that signal were the direct input to the filter.

What this tells you: the PWM repetition frequency and the low frequency signal must be well separated for accurate reconstruction. Take the example of a simple RC lowpass filter. AT the corner, the filter has an attenuation of 3db. At half the corner frequency, the attenuation is around 1db (this may be a LITTLE off). At twice the corner frequency, the attenuation is about 6db (-6db gain). Thus, if you were to use a 2KHz PWM to create a 500Hz signal and used a 1KHz lowpass, the 500Hz reconstruction would NOT be full amplitude; it would low by about 1db. The PWM would not be smoothed out "much" because the fundamental is reduced by only 6db. Remember that -10db means a gain of 0.1 so -6db is a gain of 0.25, more or less.

This will give you some guidance when trying to choose a filter to reconstruct a low frequency signal from a PWM.

Be aware that the more poles the filter has, the closer you can get the low frequency signal to the PWM frequency for a given amount of ripple.

Here, "ripple" refers to the "leakage" of the PWM signal repetition frequency into the filter output.

Hope this does not confuse too much...

Jim

I developed many of these rules of thumb about a year ago when I was working on a tutorial on PWM filtering for programmers. I got abandoned. Sounds like I need to resurrect it and complete it.

There were several others that were useful in estimating the ripple amplitude. One was the fact that ripple is largest at 50% duty cycle. This means that you don't have to check other duty cycles, that is where it is largest.

Jim

If your graphics card is showing 60 frames a sec (16.67ms) and your video captures fields at that rate, and they are somehow synchronized, you might be able to detect if something moves on the next field after moving the eyepoint. I think you need an a/d converter on a joystick. Then you can grab positions faster than the 200hz the mouse gives. You might be able to speed up the mouse baudrate... microsoft mouse send 3 bytes per report... logitech/mouse systems sends 5 bytes per report. The mouse sends 2 xy locs per report, so you can calc the delta between postions on ea report to get vel, and diff the vel between reports to get accel. This lets you 'toss' the cursor across the screen (ballistically?) by reading the accel of the mouse position. Silver Bullet. Adversaries will be using mouse to give position. They will be eliminated.

OK, so if you can extract the velocity component from the mouse movement, you can 'pre distort' the pwm to charge up the cap faster... you will be adding a little overshoot to the step function. Like boosting the bass to overcome an lf rolloff.