Design Ideas – March 18, 1999

-March 18, 1999

Model fixed-point DSP arithmetic in C

Roger Maher, SSL Ltd, Dublin, Ireland

You can run initial high-level simulations of custom numerical algorithms, such as digital filters, using floating-point numbers in an environment such as C or Matlab. Unfortunately, you won't see include fixed-point effects, such as truncation due to limited precision and register overflow, until you use a Hardware Design Language (HDL), such as Verilog or VHDL. However, a technique that models these effects in C—the function "bit_limit" in Listing 1—provides faster execution and better portability than HDLs and allows early exploration of the trade-off between bus width and performance.

Figure 1 shows a flow diagram that calculates the output E=A/4+A•B. Table 1 shows the effect of the chosen bus width. For example, the output of the multiplier truncates to 10 bits, and the output of the adder truncates to 9 bits. In this example, when input A=201.8 and input B=0.19, the output E=87.5; an ideal model without truncation would give 88.792.

Listing 1, the ANSI C code for the bit_limit function, shows how the code truncates. The code first converts floating-point numbers a_fl and b_fl to justified integers by scaling. The macro PREC sets the number of bits that the code uses to store the fractional part of the input, which are the bits to the right of the "binary" point. The bit_limit function truncates a_fl to 8 bits on the left of the binary point and to 0 bits on the right; the respective figures for b_fl are 0 and 4 bits. Because the code stores the numbers as normal integers, the standard operators can perform addition and multiplication. However, the code must rescale the result of a multiplication because both operands are justified. The bus widths are set at points A, B, C, D and E; the code displays the bit-limited value of E and compares this value to the result from full-precision arithmetic.

Table 1—Truncation effects based on bus width
  Binary point
A 1 1 0 0 1 0 0 1 1 1 0 0 201
B 0 0 0 0 0 0 0 0 0 0 1 1 0.1875
C=AB 0 0 1 0 0 1 0 1 1 0 1 1 37.5
D 0 0 1 1 0 0 1 0 0 1 0 0 50
E=C+D 0 1 0 1 0 1 1 1 1 0 0 0 87.5

You can handle signed arithmetic using 2's complement numbers by extending the bit_limit function to sign-extend the MSB after each truncation. You can perform rounding using the following statement:

bit_limit(8, 0, A + 0.5•(1& SMALL>

The number of bits in the "int long" data type on the simulation platform sets the major limitation of this technique. Normally, the number is 32 bits, so all numbers in the simulation, including the unscaled result of a multiplication, must be less than 231 . You can slightly extend this range by using "unsigned int," or you can double the range by using "int long long" when this type maps to 64 bits (DI #2328).

Two ADC channels double sensor precision

Luke J Barker, Reinke Manufacturing Co Inc, Deshler, NE

The accuracy of the on-chip ADCs of numerous small and inexpensive 8-bit µPs is well-suited for many applications. However, some situations benefit from just a little more accuracy from a resistive position sensor, for example. The resistive circuit in Figure 1 uses two ADC pins on a µP to double the precision of a resistive position sensor. In effect, the 8-bit ADC becomes a 9-bit ADC. The resistive part of the circuit shown in Figure 1 costs as little as $3; the cost is higher if you use a precision potentiometer as the position sensor.

ADC1 0 1 1 2 2 3 . . . 253 254 254 255 255
ADC2 0 0 1 2 2 3 . . . 253 253 254 254 255
9-bit result 0 1 2 3 4 5 . . . 506 507 508 509 510

The concept behind using two ADC channels is simple. One ADC input takes direct measurements of the position sensor. The second ADC input measures the voltage out of a second potentiometer that is a ½ bit behind the first input. This scheme creates a "phasing" of the two analog-to-digital results (Table 1). Adding the two results produces a 9-bit answer that is limited to a value of 510.

To set up the circuit for operation, you will need to accurately measure VREF , which is 5.000V in this case, and set ADC1 (VR1 ) to a known voltage, 3.000V in this case. Then, adjust ADC2 (VR2 ) according to the following equation:


Thus, an ADC2 =2.990V sets input-channel ADC2 to a ½-bit lag. (DI #2332)

Transistor pair lowers PWM IC's start-up current

Christophe Basso, Motorola Semiconductor, Toulouse, France

An auxiliary winding usually provides power for the popular UC384X-based offline switch-mode power supplies (SMPSs). This winding feeds the main PWM controller during steady-state operation. However, the controller always needs a small start-up current, which clearly plagues the efficiency in very low-power standby SMPSs with POUT =500 mW, for example. Any wasted source power, such as for the start-up network and controller supply, adds to the power drawn from the mains and significantly degrades the overall efficiency. Obtaining an acceptable efficiency at high line voltage represents a tough design task. Some tricks totally stop the start-up current by using a high-voltage bipolar or MOSFET (Reference 1). However, the addition of a high-voltage component burdens the bill-of-material cost of a small-power SMPS.

An alternative and inexpensive solution allows you to program the start-up current at any value, whatever the PWM IC (Figure 1). Implemented around two low-voltage off-the-shelf bipolar transistors, Q1 and Q2 , the circuit brutally connects the controller to bulk capacitor C1 when C1's voltage level is adequate. This level determines the CXV product that is large enough to feed the PWM IC until the steady-state operation takes over. The designer can then select any small start-up current to charge the bulk capacitor in agreement with time-constant recommendations. During the charge, series pnp transistor Q1 is locked off, which prevents any current consumption from the main controller. When the voltage reaches a defined level, Q2 starts to pull Q1's base to ground. Q1's collector voltage rises and further saturates Q2 through the reaction resistor, R1 . The charged bulk capacitor is now fully connected to the PWM IC, which starts to oscillate. A few cycles later, the auxiliary winding takes over, and Q1 stays on.

The level VTHRESHOLD at which Q2 turns on is simply given by:


If you assume a linear charge from the start-up network, R4+R5, then the time at which Q2 starts to conduct is equal to




With the values in Figure 1, C1 connects to the IC when its voltage reaches 17V, which occurs in less than 350 msec at VDC=350V. The circuit offers better than 50% efficiency at high-line (250VAC), drawing less than 1W for a 500-mW output power. The MMG05N60D insulated-gate bipolar transistor (Motorola Semiconductor, also contributes to the circuit's performance because of its low capacitive parasitic elements, including a 5-nC gate charge. The circuit thus minimizes commutation losses. Figure 2 shows the start-up phase of the low-power SMPS at very low line, 70V ac. (DI #2330)


1. Basso, Christophe, "Low-cost MOSFET quashes power resistor," EDN, June 9, 1994, pg 140.

Autozero a position-sensing detector

James Zannis, Renishaw S A, Champs-sur-Marne, France

Autozeroing schemes can be necessary to minimize the input-offset voltages of position-sensing detectors (PSDs), particularly when you use these detectors with high dc gain. PSDs are useful optical transducers for accurately measuring displacement. In practice, their typical configuration is as a differential current-to-voltage circuit (Figure 1). The ratio of the photocurrents I1-to-I2 linearly divides between the electrodes, proportional to the incident light beam. The magnitude of the photocurrents is a function of the light intensity.

Although PSDs are photodetectors, using them with high dc gain can quickly get you into trouble. Input-offset voltages can easily multiply due to the relatively low sheet resistance between pins 1 and 2, which typically ranges from 5 to 100 kW. Consider the following nodal equations:


An offset multiplication occurs because of the low intrinsic diode impedance, RS , of the PSD, and the high values of RF. For normal photodetectors, RS is very high, and, therefore, no multiplication occurs. FET-input op amps are usually the choice for photodetector circuits because they have a higher input impedance and lower current-noise density at high-impedance levels than their bipolar counterparts.

For a typical pair of FET op amps mismatched by ±3 mV and with an RF of 1 MW and sheet resistance of 20 kW, VOUTB would be –0.303V, and VOUTA would be 0.303V. For a theoretical pair of op amps mismatched by ±10 µV, an RF of 1 MW, and a sheet resistance of 20 kW, VOUTB would be –0.001V and VOUTA would be 0.001V. Ultimately, it would be nice to use a low-noise, low-bias-current amplifier with a very closely matched front end. However, it is a much more difficult process to match the input FET differential pair than that of a bipolar transistor. Nevertheless, this circuit shortcoming due to the sheet resistance is also the key to its success; you can force the offset voltage to be equal.

You can accomplish this goal by sampling and integrating the difference of the two input-offset voltages and forcing the positive input of Amplifier A with the result (Figure 2). The two amplifier outputs converge to the input offset voltage of Amplifier B with the difference in the two output voltages equal to the offset of amplifier C. VOUTA reduces to


and VOUTB reduces to


Note that biasing the wrong op amp causes the two outputs to diverge, and you must choose an op amp for Amplifier C on the merits of its dc specifications.

The second approach to forcing the offset voltage to be equal is to use a potentiometer to adjust one of the op amps at its offset-adjustment pins. This circuit converges or diverges in a manner similar to that depicted in Figure 2. (DI #2331)

OTP µC controls Boomer amplifier

Wallace Ly, National Semiconductor Corp, Santa Clara, CA

Figure 1 shows a circuitthat uses a one-time-programmable (OTP) µC in an unusual way. A COP8SGR7 µC uses digital signals to "bit-bang" an LM4835 (dubbed "Boomer" by National) amplifier. Although the amplifier is designed for potentiometer control, you can modify it to make it a fully digitally controlled part. Figure 2 shows the flow chart for the control process. A PWM signal and a lowpass filter allow you to use digital signals to control the amplifier. The technique sets the µC in processor-independent mode. You load the values affecting the duty cycle into the appropriate timer registers. When a control bit goes high, the µC delivers a PWM signal.

The PWM signal goes to a first-order lowpass filter. The output from the filter is a dc voltage, whose amplitude is proportional to the duty cycle of the PWM signal. The processor-independent mode allows the µC to perform other duties and calculations while it generates the PWM signals. With the COP8SGR7 µC, three PWM outputs are available; therefore, the µC can control three amplifiers, or six channels of audio. The µC debounces the two (volume-up and -down) pushbuttons. The controller also handles the mute, shutdown, beep, volume, and headphone functions.

Download the C code for controlling the µC (DI #2326).

LED driver displays standing-wave ratio

Richard Panosh, Vista, Bolingbrook, IL

The circuit in Figure 1 uses an LM3914 LED driver to directly display standing-wave ratio (SWR) in a low-cost, rugged instrument. The SWR-sensing head is derived from the ARRL Antenna Handbookin an article that describes the tandem match. The forward voltage and reflected voltage (VF+VR ) signal drives Pin 6 (RHI ), and the VF–VR signal drives the normal signal input at Pin 5. You can use this basic arrangement to display the ratio of two voltages in other applications. The internal circuit of the LM3914 comprises 10 voltage comparators that compare the input voltage at Pin 5 to an internal, 10-step, linear-voltage-divider string between Pin 6 (RHI ) and Pin 4 (RLO ). The voltage-divider levels, VN , are VN =VREFX n/10, where n is the voltage-divider step from 1 to 10, and VREF is the voltage between pins 6 and 4.

When the signal voltage on Pin 5 satisfies the equality in the following equation, the nth LED energizes in the dot-mode display (with approximately 1 mV of hysteresis).


If you force VREF to equal VF +VR and the voltage at Pin 5 to equal VF–Vr, you can arrange the terms as in the following equation.


where VF +VRVF–VR is the definition of the SWR. Two sections of the quad op amp, IC1 , buffer the forward and reverse voltages developed in the SWR bridge. The remaining two sections serve as a difference amplifier to produce the voltage VF–VR and as a noninverting summing amplifier to produce the voltage VF +VR . Even though the circuit does not use the internal voltage regulator, you must properly terminate it to establish the LED current. (DI #2324).

Comparator provides stable hysteresis

Fernando Garcia, Lucent Technologies, Brownsville, TX

Unless a voltage-comparator circuitis sampling an extremely clean signal, the comparator always requires some hysteresis. Traditional comparator circuits obtain the required hysteresis by using positive feedback derived from the ratio of two resistors. The voltage at the noninverting input is the superposition of a fraction of the input and output voltages, each divided by its respective resistor ratio. With an inverting comparator, the noninverting input usually connects to a voltage reference with reasonably low impedance, and you can choose the hysteresis-setting resistors without concern for loading. However, for the noninverting comparator, the input voltage comes from the actual source the circuit is sampling. This source has a series impedance, RSOURCE, which can be a significant fraction of the input resistance R6 (Figure 1). This impedance may not be repeatable or may change value with changing circuit conditions; therefore, it may result in hysteresis errors.

Although you can change the feedback and input resistor values to minimize the source-impedance impact, you face a practical limit on the increases, because the value of the feedback resistor that the resistor requires to maintain the proper resistor ratios can become excessively high. Traditionally, you could use an op amp configured for unity gain to buffer the high-impedance source from the comparator. However, in some applications, cost, board-area, or current-consumption constraints may preclude adding an op amp. The circuit variant in Figure 2 eliminates the source-impedance problem. The voltage source under comparison connects directly to the noninverting input. Hysteresis does not come from resistor feedback, but rather from MOSFET Q1. If the voltage you are sampling is less than the threshold, the comparator's output is low, and Q1 turns off.

The comparator's inverting input essentially sees a reference voltage identical to the reference voltage in Figure 1. However, when the voltage exceeds the threshold, the comparator's output goes high, turning on Q1. The MOSFET shorts out the lower portion (R5) of the resistor divider. This action has the net effect of lowering the reference voltage to the comparator's inverting input. The differential voltage thus increases, providing the required hysteresis. The source impedance basically sees only the comparator's input impedance. This impedance is extremely high, so the source-impedance impact on offset is low. Most small-signal MOSFETs can work in this application, provided that the RDS(ON) is at least one order but preferably two orders of magnitude lower than the resistor it must shunt. This application requires a logic-level FET that turns completely on with VGS =5V. (DI #2335).

Loading comments...

Write a Comment

To comment please Log In