Laboratory Exercise 3: A DSP-Based Interference Filter and
Contents
1. for play back The Analog Devices Corporation has graciously donated six SHARC evaluation boards with ADSP21061 processors as the basis for the lab Each board also has a two channel codec chip that contains both analog to digital and digital to analog converters These codecs are the basis for SoundBlaster cards and similar motherboard capabilities in PCs The two channels are the left and right stereo signals You only have to deal with 34 Engineering 164 Lab Manual Spring 2001 one of them The gift package included a program for taking data from the ADC s and putting it back out to the DAC s I have modified that program to store blocks of data on the way through This allows you to manipulate the data with an assembly language pro gram To do the entire job requires about 80 lines of code This is not easy code to write you can make very subtle mistakes I have tried to tell you how to avoid problems but you should build and test the code section by section and you may need to test your work with the simulator I hope to offer a shell program for exercising your code in the simu lator without your having to worry about setting up input data or simulating the DMA input and output transfers Requirements Write an assembly language program for an Analog Devices SHARC processor to do notch filtering and or AGC on a digital signal stream There are two switches on the board that your program will read to determine whether data2. or CD system the output of which has a sine wave at a little above 4 Ki lohertz added to it This sinusoid is the interference that your filter must remove The combined signal goes to the codec on the evaluation board There the signal is sampled that is measured repeatedly at a constant rate of 44 1 Kilosamples per second The pro gram you use to download the code actually lets you choose the sampling rate from a Windows menu You will find that the frequency of the filter notch is proportional to this 38 Engineering 164 Lab Manual Spring 2001 rate Use the 44 1 KHz sampling rate for testing I designed around that frequency be cause it is the standard for audio data You should play around to see the effects of changes FLAG3 Output for Timing Audio Source Tape or CD Powered Speaker Analog Devices Evaluation Board Test Setup EZ KIT LITE Figure 12 Block diagram of the SHARC system for Lab 3 The codec data stream goes to the SHARC processor where my program separates one channel of data stores 512 words in memory for you and calls your program You proc ess the data and write it back to an output buffer that I send on to the DAC You can lis ten to the results on a speaker There are several other inputs and outputs to the proces sor I do not yet know how useful it is in this lab but we have an In Circuit Emulator or ICE available that lets you control running the code and lets you examine memory You can
3. that all operations on the line are done simultaneously This means the product being accumulated in a step is not the one being calculated in that step but is the one calculated in the previous instruction To start the system up takes a little thought so that the first sum will add the first product to an empty register Because the filtering loop is the most computationally intensive part of the whole code using this instruction can make the code much more efficient In fact at the moment the code will only work if you use this instruction successfully We are working on relieving the constraint but don t count on our success Finally a few words about how the DSP instruction set makes signal processing code more efficient My example is the use of circular buffer for your filtering operation No NCOEFFS 1 tice that in the expression for the output of the filter code x yc f the input j 0 data f is indexed backwards that is the current output x depends on not only the cur rent input value f but also on many earlier samples because the subscript is i j Since memory is a finite resource you cannot keep putting data into new locations Instead one keeps only two frames of data the one you are working on and the immediately preced ing one Figure 13 shows a circular buffer for this purpose In the figure the data for the current frame has been put in the left half and the program is processing an output word corr
4. the job However if you are curious about why the whole thing works I recommend the book Digital Signal Processing by Richard Hamming This is a skinny well written book on the basic ideas of digital filters that I found to be a nice in troduction I will make sure a copy is on reserve in the Sciences Library There is an enormous amount of documentation on the SHARC processor and the EZ Kit evaluation board We have put this on the class website and on the Engineering server You are welcome to read and copy the material including the programs as much as you like The software for the assembler and simulator is a product of the Free Software Foundation which encourages its distribution This software is not Analog Devices latest and greatest Their newer software concentrates on automating DSP design within Microsoft s Studio environment and it hides the processor properties from the user The older software is quite serviceable for my pedagogical purposes but you should not hold its clunkiness against the Company However please do not print out big chunks of stuff That wastes paper and does you little good To make it a little easier to write a program I have prepared a handout that contains selected material taken from the SHARC reference manual Appendices A and B in that manual have the complete de scription of the instruction set The handout contains the chapter on the use of the DAG s the Data Address Generators It also has t
5. will rename and edit You should copy the contents of this directory to your ac count and edit code there Open Windows Explorer go to your root directory and drag the source folder to your directory In the same directory there is a batch file called Mk bat Open this with Notepad and edit the end of the first line changing Lab3_wrp asm to the name of your file for this lab When you have entered the code 40 Engineering 164 Lab Manual Spring 2001 you can assemble and link the program by double clicking on this file The executable is given the name tt 21k You download this file with the EZ Kit Lite Host program We have also written a C program that can be used with the simulator to exercise your code The same Mk bat file links your code to that program as well giving the file Lab3_sim 21k You can run the EZ KIT Lite Simulator on that executable The simulator program will let you pick a set of input data look at the output buffer and ex amine how your code works line by line The simulator is similar in operation to the one for the PIC16CS505 It is a little slow but should be fine for your purposes In addition there are shortcuts in the EN164 Source Code program group to folders containing example files and source code including information about the Analog De vices Talk Thru program referenced in this handout The TalkThru program relies heavily on assembly languag
6. Engineering 164 Lab Manual Spring 2001 Laboratory Exercise 3 A DSP Based Interference Filter and Automatic Gain Control System It is now about seventeen years since the compact disk developed by Phillips Laborato ries in Briarcliff Manor New York began to change the way we listen to music and to improve the quality of recorded music tremendously Digital technology will eventually pervade the entire home music system All sources will be digital and the digital data will be manipulated to do tone compensation and room equalization before the data is fed to a DAC and power amplifier Even the radio receiver will come to depend on digital filtering techniques to lower cost and improve performance The basic method for this filtering is one of the concepts behind this lab While I am not trying to teach the mathematics of digital filtering I do hope you pay attention to the mechanics of the proc ess and come to understand some of what one can do in this field I also have a continuing interest in another area of new product development even though I do not work in it I have a nephew who is profoundly hearing impaired for whom hearing aids have been a great help and a continuous hassle Such aids are still largely analog designs but Siemens and other companies have begun offering digital ones The basic function of a hearing aid is to filter the microphone signal to match the limited range of frequency response typically a band only a kilohe
7. also stop the program and read memory back from the board with the PC See the EZ Kit Reference Manual chapter 3 We will have at least one station set up with the ICE but I may not have had time to figure out exactly how best to use it Again feel free to experiment the instructions and software will be available Finally there are four FLAGn pins on the processor that allow you to put in bits or take them out by hard ware FLAGO is dedicated to the codec We have set up FLAG3 as an output pin with a jack to connect it to a scope or logic analyzer This pin goes HIGH when your code is called and goes LOW when it returns The HIGH time measures how efficient your code is If your code isn t fast enough this test will show it the effect on the audio stream is pretty clear as well but this test allows you to separate causes more easily Flags 1 and 2 are input pins connected to switches mounted to the EZ Kit We read those switches before each call to your code and pass the result a number between 0 and 3 u in the memory location _InFlags You use this for controlling whether filtering and or AGC is done This requirement makes it easier to test I suggest you apply the filter only if _InFlags contains or 3 You already apply AGC only if the number is 2 or 3 39 Engineering 164 Lab Manual Spring 2001 This is not a course on digital signal processing so I give you the algorithms and the con stants you need to do
8. e library routines They are stored in the Assembly Rou tines folder within the EN164 program group The EZSharc folder in the same group has source code for several applications mentioned in the EZ Kit reference manual Each application is in its own folder and the TalkThru program is in the tt folder You will not find these programs of much use directly but you may want to look at them as exam ples of correct syntax In the interests of making your work more efficient and less painful here are four obser vations about SHARC programming that represent mistakes I have made at one time or another Consider them warnings about the subtle errors one can make on this processor First of all the general register set has only 16 actual registers These may be referred to as rO to r16 or fO to f16 In the first case the assembler assumes you mean integer data and in the second you mean floating point data There are NO DEDICATED FLOATING POINT REGISTERS Be careful that you don t overwrite data you need by using a given register under two different names e g f1 and rl Second most compute instructions allow you to fetch or store two pieces of data at the same time by accessing both program and data memory By a peculiarity of the assem bler the order of program and data access is important For example the line rO rl r3 dm il m6 r4 r3 pm i9 m15 is perfectly legal It adds the contents of rl and r3
9. esponding to an input partway through that section The gray areas show the earlier data that contributes to the new output You set up a circular buffer by loading b1 and 11 with the base address and the buffer length If you then use only the dm il m method of addressing the processor makes the data addresses wrap around Thus you can index backwards through the data with dm il m7 and know that you will bring back data from the high end of the buffer when you need it The m7 register always holds the value 1 and you must not change that It is used for stack management in my code After cal culating the output for one point you can set il to point to the next input point using the instruction modify il m1 Of course you load m1 with what you need to add to il before you start the calculation Usually this value is one to three greater than the num ber of coefficients being one more than the number of times you decrement il during the calculation On the next frame you put the new data into the right half of the buffer and the present data becomes the immediate past data Keeping track of which half of the buffer to use in each frame is the main nuisance in the input part of the code Remember that you have to set 11 0 before leaving your code and therefore have to set up the buffer each time you use it 42 Engineering 164 Lab Manual Spring 2001 Base Address in register b1 t Calculating for output at this input da
10. he index of the instruction set and spe cific data on a few difficult to understand but very useful instructions In writing a pro gram the reference manual is probably the most useful single piece of documentation To access all the documentation and source files in The Instructional Computing Lab you first have to set up your account There is a program group called SHARC EZ KIT Lite in the START menu of the The Instructional Computing Lab machines First run the command file called Environment Setup in this group by double clicking on it This edits your PATH environment variables and registry entries to include all the pro grams necessary to compile link and run your code The same program group also contains two applications One is called EZ KIT Lite Host this is the program you run to test your code When you think you are ready to test your code launch this application and select the file created by the assembler Even though you can t run your code in the Instructional Computing Lab because we don t have SHARCs on the machines the NT machines in room 196 will access both the server software and an evaluation board The other application is the EZ KIT Lite Simulator use this program to simulate your code if you are having problems There is also a link to the folder containing the source files you will use for the lab The shortcut is called EN164 Source Code and it contains the Lab3_00 asm template file that you
11. into r0 fetches a word from program memory to r3 and writes one from r4 into data memory all in the same clock cycle The line rO rl r3 r3 pm i9 m15 dm il m6 r4 is illegal Note that in dual access with computation commands the only method of access is through the use of the DAGs There are not enough instruction bits to en code anything more complicated Third lines that do multiple operations are subject to some subtle errors the assembler will not catch partly because the assembler is no too smart and partly because the opera tions are not actually illegal For example the line r3 rl 13 dm il m6 r4 r3 pm i9 m15 is perfectly legal but will malfunction The problem is that there are two writes to r3 in the same line The hardware has rules about precedence and only the ALU write will actually work The other is lost Finally the digital filtering can be done in very little more than one instruction per filter coefficient if you use the multifunction 41 Engineering 164 Lab Manual Spring 2001 instruction that allows you to do a sum and a multiply at the same time as dual memory reads I have included the description of this Parallel Multiplier and ALU Floating Point instruction in the handout on the instruction set There are two problems using this instruction One is that you have to plan the register use very carefully since it has severe restrictions on the source registers The other is
12. ion il dm _InAddress Similarly the memory location OutAddress has the address of where to place your output data Do not modify these pointers Your code must do the following 1 Retrieve 512 integer inputs from the array at _InAddress convert them to floating point numbers and store them contiguously to the last set of data in a buffer of your own Floating point conversion is necessary because both the processing algorithms use floating point numbers Very likely you would use a circular buffer to make building the filtering algorithm easier The discussion section talks about these buffers as does the material on DAG s in the software handout 2 Filter all the data from the new frame I have provided a set of filter coefficients that is a set of constants in program memory The number of such coefficients is given in a define statement at the top of the code as NCOEFFS Suppose that the data in your floating point buffer are called f where i is an index corre sponding to sample number This index simply increases continuously with time Since there is only a finite amount of memory the buffer locations are cy clic and amounts to an index through the buffer The filtering formula is NCOEFFS 1 x DG PIS j 0 where cj is the jth coefficient and x is the ith output of the filter On subroutine entry the values of these coefficients are found in program memory as an array starting at addres
13. is filtered AGC d both or neither There is a block diagram of the system at the beginning of the discussion section The system has been set up and completely wired on Analog Devices boards You provide and integrate the actual processing code You can receive 65 of the grade for the lab by getting the data transfer and digital filter parts to work properly These operations take 90 of the computation time but the AGC is the hardest part to write It is easier to show than to explain the simple proce dures we have for testing the system For this reason the details on testing the code are a bit vague and will have to be explained more during the lab and help sessions We have placed copies of a template for an assembly language subroutine called Lab3_00 asm both on the class website and on the Instructional Computing Facility machines See the discussion section for access details Make a copy of that file using your own unique file name but with the same asm extension There is also source code and object code for a mixed C and assembler program that I wrote by modifying the Analog Devices TalkThru program You are welcome to look at the source code but its details are of little real interest I am not particularly proud of it it is really very in efficient Still it does the job and could be written and debugged quickly When the Lab3_00 asm routine is properly assembled and linked with my code it is cal
14. led peri odically whenever there are 512 new data points available to process Those data are 32 bit fixed point integers whose magnitudes are always less than 32768 They start as 16 bit two s complement numbers from the codec and are sign extended to 32 bits You must write SHARC assembler code to do the filtering and AGC control on this data and write it back to an output buffer for the calling program to send to the codec The template file has clearly marked places for allocating space for variables and buffers in both program and data memory as well a place for your code There is a header sec tion with definitions for SHARC parameters and for some constants used in the algo rithms Finally the code section has a short piece of program that I originally used to check that the template worked It copies the fixed point input data directly to the output buffer This code is now commented out and you should remove it altogether before turning in your code I left it in place so you would have a simple guide to common syn tax and could use it as an example of retrieving and returning the signal data 35 Engineering 164 Lab Manual Spring 2001 When your subroutine is called the address of the start of the input data array effectively a pointer to the array is stored in data memory at the location _InAddress The lead ing underscore is necessary You can retrieve that pointer to a register say to register 11 with the express
15. or this sample Let Y be a run ning measure of the energy in the signal which we recalculate for each frame with the 511 expression Y 4A Y a x where A is a constant with value 0 982 This will give i 0 an average value of the square of the signal amplitude over roughly a second From this we calculate a new gain G and each s G x Step by step the algo new rithm is 4 for i 0 i lt 512 i G gt x if Is gt SMAX then si CLIP s BY r0 where rO contains SMAX Y D x B 37 Engineering 164 Lab Manual Spring 2001 C B Y G Store s in either the working buffer or the output buffer If you write to the output buffer then this is where you do conversion from float back to integer 5 If not already done store results to the output buffer converting to fixed point integers en route 6 Store the value of Y in memory for use in the next frame The constants A B C D and SMAX have values A 0 988 B 9 77 10 C 3 127 10 D 1 2 10 and SMAX 28300 0 These values are already entered in de fine statements at the top of the template file as A_agc B_agc etc The reason to sepa rate steps 1 and 2 is to improve the accuracy of calculating Y the small terms in x will accumulate more accurately in an empty register than in one filled with a very large number beforehand The CLIP operation replaces s with either 28300 o
16. quare root operation for which the code may be taken verbatim from the Analog Devices user s manual The algorithm uses the New ton Raphson iterative method with an initial guess supplied by the RECIPS assembler instruction Remember the homework The description of the syntax of that instruction contains code for the full calculation I have made life a little easier by typing the code from the manual into the Lab3_00 asm template file as a subroutine at the end of the file You will need to consult the Users Manual to figure out how to use the code The idea of the AGC algorithm is to estimate the average amplitude of the incoming signal If this is small then multiply each signal point by a large gain Similarly if the amplitude is large multiply by a smaller gain You can do this easily by making the gain inversely proportional to the average amplitude To avoid an infinite gain when the signal disap pears one adds an extra constant into the denominator to keep it non zero In the algo rithm below the first two steps calculate the mean square amplitude The third step cal culates a new gain which is applied to the data in the fourth step After compensation for overload from abrupt increases in the input the data is converted to integer format and stored The exact algorithm is For each new frame of 512 data points let x be the i output of the FIR filter where in this case 0 lt i lt 511 Let s be the DAC output signal f
17. r 28300 de pending on the sign of s if the initial magnitude exceeds 28300 The format of the CLIP statement is the syntax of the actual assembler instruction where s is understood to mean the register in which si is stored When G is recalculated after the CLIP statement the new gain applies only to the later outputs in the current frame With this choice of con stants the gain for low amplitude signals is 20 DB and the maximum signal to the DAC is 28300 out of a possible 32767 The Analog Devices TalkThru program that is the basis of my code allows you to change the analog gain of the codec with a software button on the PC you use to down load the program Downloading instructions are posted in the lab We will test the AGC control by looking at how your output changes when the codec gain changes If the AGC is working then a 20 DB change in input should produce a 3 to 6 DB change in output When the lab TA certifies that your program works to the level you claim he or she may ask some questions about how you designed it Then to finish getting credit for the lab you email the source code to the recording account Please indicate in the cover mes sage whether both parts or neither of the lab work that is both filtering and AGC Discussion Figure 12 shows a block diagram of the finished system The central part of the drawing is the EZ Kit LITE evaluation board from Analog Devices On the left there is a tape player
18. rtz or so wide around 1KHz of a particular deaf person Ideally this response should be tailored to the per son for the greatest effectiveness and the least damage from acoustic overload Hearing aids also need to adjust their gain to match the level of signal in the room because the main problem in deafness is the lack of dynamic range You have to boost any signal weak or strong to a fairly loud but not damaging level You have to do so in a way that does not distort the signal when the room noise level increases This change of gain has to occur on a fractional second basis with no intervention of the user Analog hearing aids use an Automatic Gain Control circuit for this function Again in the context of pre senting the instruction set and capabilities of a DSP processor we have designed an AGC algorithm into the lab that does the same thing digitally The two capabilities of the program filtering and AGC are applicable to both music systems and hearing aids In audio systems the equivalent of AGC is found in the Dolby compression system that allows limited dynamic range media such as analog cassette tapes and movie sound tracks to seem to have more range The trick is to crank up the gain in recording when the signal is weak and crank the gain down again when the signal is strong You do the reverse in playback Your DSP program is equivalent to the re cording end of this process You can probably figure out an inverse algorithm
19. s _coefs Each filter output is affected only by the current in put and by the last NCOEFFS 1 inputs Thus you will save two frames of input data replacing the oldest with a new set on each frame The old data you have left are needed to calculate the first outputs of the new frame You will have to store the new outputs x in a separate working storage area 3 There is an integer in the range 0 lt n lt 3 passed to you in the memory location InFlags If this number is 2 or 3 apply AGC to the xi values that you just calculated The AGC algorithm is discussed more below 4 After you have finished the calculations convert the x values to fixed point numbers and write those out to the output buffer at _OutAddress The digital filter is primarily a notch filter that is it removes a portion of the frequency spectrum between 4 and 4 5 Kilohertz By remove we mean that it reduces the magni tude of a sinusoid in this frequency range by a factor of 1000 or so To test your filtering code we will have signal sources that mix a pure tone in this band with a normal source like a tape or CD player Without filtering the pure tone masks much of the music With filtering the tone disappears but there is little effect on the rest of the signal 36 Engineering 164 Lab Manual Spring 2001 The AGC algorithm itself is slightly complicated but the required code is not long 46 lines You will need to do an inverse s
20. ta point Length 1024 words 11 1024 Figure 13 Memory use for a two frame circular buffer Memory addresses increase to the right Each half holds one 512 word frame and the left side was the most recently filled The gray regions signify data that contributes to calculating an output for the sample point marked on the left side Note the use of older data extending into the top of the buffer 43
Download Pdf Manuals
Related Search