Address Processor for a Signal Processor

The present invention relates to an address processor for a signal processor for calculating addresses in a read/write memory containing state variables for digital filters.

In connection with implementation of digital filters in signal processors and where input data samples to said filters are stored in a read/write memory some form of address processor is required for writing and reading the state variables of said filters. These filters can be both recursive (IIR) and non-recursive (FIR). A filter is calculated by reading the state variables from the memory and calculating output data. These output data then become input data or state variables to the next filter and are therefore written back to said memory. Generally a filter system can comprise several filters, which also can have different sampling frequencies.

After a calculation of a filter value all the state variables in a filter are shifted one position in the memory. The oldest state variable is written over by the next oldest, and a new sample is stored in the memory position that has been freed through the shifting operation.

A drawback of this previously known method is that the shifting in said memory is a time consuming operation, since all the state variables have to be read and written into new memory positions.

Another drawback is that this shifting operation also is power consuming. This is important for instance in connection with signal processors in battery driven telephones, where the available energy is limited.

A common method for addressing and shifting state variables for a digital filter implemented by a signal processor and a read/write memory is to use a circular buffer of size 2.sup.n, where n is an integer greater than 0. A pointer indicates the memory position of the oldest state variable, that is the folding address of the circular buffer. After reading the oldest state variable a new sample is written into the indicated memory position, said oldest state variable being written over. If the pointer now is changed to point to the next oldest state variable, said pointer will indicate a new folding address. The previously next oldest state variable is now the oldest, and the result of this procedure is equivalent to shifting all state variables in a circular buffer one memory position without physically moving them. Only the pointer has been updated to indicate the new folding address. The method is therefore called virtual shifting. When said pointer has reached the first memory position of the circular buffer it will, after the next shift, point to the last memory position. Thus, the addressing of the buffer can be denoted "addressing modulo 2.sup.n ".

However, an important drawback of this previously known method is that the buffer only can be of size 2.sup.n, so that the filters have to be of size 2.sup.n. This makes the method suitable only for certain special filters.

Another previously known method uses a buffer of any length and makes use of a pointer that is shifted down from the end to the beginning of the buffer during the filter calculation. When the pointer has reached the beginning of the buffer, said pointer is once again loaded with the address of the end of the buffer, so that the procedure can be repeated.

A drawback of this previously known method is that it is not possible at an arbitrary time to address an arbitrary state variable in an arbitrary filter. The method will only work if the filters are calculated in sequence in accordance with the above. This makes the method unusable as a starting point for a general address processor.

An object of the present invention is to provide an address processor of the type mentioned above, which processor is capable of addressing and virtually shifting state variables of at least one circular buffer for at least two digital filters stored one after the other in said buffer, said filters having an arbitrary number of delay elements or an arbitrary length, but the same sampling frequency.

In accordance with the invention the above object is achieved with an address processor with the feature of claim 1.

The invention will described in detail below with reference to the accompanying drawing, in which:

FIG. 1 shows the division of the data memory in connection with the present invention;

FIG. 2 shows the division of a buffer in connection with the present invention;

FIG. 3 shows a filter diagram for the filters in the buffer of FIG. 2;

FIGS. 4a-4h illustrate the location of the filters and the folding address of the buffer in FIG. 2 during consecutive calculation steps;

FIG. 5 shows a preferred embodiment of an address processor in accordance with the present invention;

FIG. 6 shows a calculation time diagram for the address processor in accordance with FIG. 5; and

FIG. 7 shows a signal time diagram for the address processor of FIG. 5 for three filters A, B and C with 2, 1 and N taps, respectively, where N is an arbitrary positive integer.

FIG. 1 shows the memory organisation for that part of the RAM that contains state variables for different filters. In this case it is assumed that the memory comprises filters for eight channels CH 0-CH 7. Each channel contains buffers BUFF 0 - BUFF 2. The number of buffers is arbitrary, but each channel has the same number of buffers with corresponding lengths. The buffers in turn contain filters FILTER 0 - FILTER 2. The buffers normally have different lengths, but the filters in each buffer correspond to the same sampling frequency. In the present invention at least one buffer contains two or more filters. The buffer lengths depend of the sum of the filter lengths; in fact the buffer length is formed by the sum of the number of delay elements of the included filters plus 1.

FIG. 1 shows a data memory divided into eight channels CH 0 - CH 7. The invention is not limited to eight channels, rather an arbitrary number of channels is possible. Furthermore, FIG. 1 indicates a data memory containing 512 memory cells. However, this number can also vary. FIG. 2 shows the contents of a buffer in detail. In the example a buffer comprises three filters with filter start addresses filter.sub.-- addr(0), filter.sub.-- addr(1), filter.sub.-- addr(2). The absolute start address is denoted mod.sub.-- start. The state variables in the first filter are denoted X(n), X(n-1), X(n-2), X(-3). Here X(n) indicates the latest sample, X(n-1) the next latest sample, etc.

In order to facilitate the following description it is initially assumed that the three filters in the buffer in FIG. 2 are non-recursive.

In this case the state variables are fetched one after the other from the memory during the calculation of FILTER 0, starting from the oldest state variable X(n-3), and multiplied by the corresponding filter coefficients, which are fetched from RAM or ROM. The filter is calculated by summing these multiplied state variables. When the filter has been calculated the obtained value Y(n) is written into the memory cell for the oldest state variable X(n-3). In FIG. 2 this has been indicated by letting the state variable or sample X(n-3) share the memory cell with the filter output signal Y(n). The calculated filter values Y(n), Y(n-1) . . . in this way form input signals to FILTER 1. FILTER 1 is calculated in a corresponding way, and after calculation its value is written into the memory cell for the oldest state variable Y(n-2) as input signal to FILTER 2. This procedure, where the output signal from one filter forms the input signal to the next filter is possible due to the fact that the filters in the buffer correspond to the same sampling frequency.

FIGS. 3 and 4 further illustrate the calculation process, in particular also the complications that arise in connection with recursive filters.

FIG. 3 shows a more general filter diagram for the filters in the buffer of FIG. 2. As previously the first filter, FILTER 0, is a non-recursive filter that only contains a transversal part. However, the second filter, FILTER 1, contains a recursive left part and a transversal right part. The third filter, FILTER 2, contains only a recursive part and no transversal part.

In FIG. 3 the delay elements have been indicated by T. As can be seen from FIG. 2 FILTER 0 contains three delay elements T and four coefficients a.sub.0. . . a.sub.3. FILTER 1 contains two delay elements T and two coefficients b.sub.1, b.sub.2 in the recursive part and three coefficients c.sub.0. . . c.sub.2 in the transversal part. FILTER 2 contains two delay elements T and two coefficients d.sub.1, d.sub.2. According to the calculation rule above the buffer has a length of 3+2+2+1, that is 8.

FIG. 4 shows a sequence of calculation for the three filters in FIG. 3.

It is essential to note two things in FIG. 4.

Firstly, the values of the state variables X(n) . . . , Y(n) . . . , W(n) . . . always lie still in the memory cell they have once been allocated. For each new figure (a)-(h) n is increased by 1, that is X(n) will be denoted X(n-1), etc. The values of the state variables, however, still remain in the original memory cell. The oldest state variables are then overwritten with new state variable values when they are no longer needed. Thus, it is not necessary to move the state variables.

Secondly, the folding address, mod.sub.-- pointer, of the buffer is continuously updated. This pointer always points to the memory address where the circular buffer folds, in this case the border between FILTER 2 and FILTER 0. The pointer mod.sub.-- pointer forms a "zero-reference", with which it is possible to obtain a correct absolute address for a current state variable, which will be illustrated in detail below.

For the calculation of the address to a certain state variable the following variables are utilized:

During the calculation of the address to a state variable the sum

data.sub.-- addr+mod.sub.-- pointer

is first calculated. Thereafter it is tested whether this sum is less than mod.sub.-- numb. If this is the case rel.sub.-- addr is set equal to the obtained sum, that is

rel.sub.-- addr=data.sub.-- addr+mod.sub.-- pointer

However, if the condition is not fulfilled, that is the sum is greater that or equal to mod.sub.-- numb, rel.sub.-- addr is set to

rel.sub.-- addr=data.sub.-- addr+mod.sub.-- pointer-mod.sub.-- numb

In both cases the displacement of the sought memory position relative to the buffer start address, mod.sub.-- start, is obtained. The final, absolute data memory address, dmaddr, is obtained by adding mod.sub.-- start to the obtained result, that is

dmaddr=rel.sub.-- addr+mod.sub.-- start

Initially mod.sub.-- pointer is set to

mod.sub.-- pointer=mod.sub.-- numb-1

so that

rel.sub.-- addr=data.sub.-- addr-1

A write address is calculated in the same way, with the difference that data.sub.-- addr is replaced by write.sub.-- addr.

The calculation of FILTER 0, FILTER 1 and FILTER 2 is performed in the following way.

1) The input signal is written into position X(n).

2) FILTER 0 is calculated in the following way. First data.sub.-- addr is set to filter.sub.-- addr(0)+index, where index=3. With this value for data.sub.-- addr the address to X(n-3) is calculated in accordance with the above. X(n-3) is fetched from the memory cell that is pointed to and multiplied by coefficient a.sub.3. Thereafter index and therefore also data.sub.-- addr is decremented by 1. X(n-2) is fetched and multiplied by a.sub.2. The product is added to the previously obtained product. In a corresponding way index is stepped down all the way to 0, so that all the variables in FILTER 0 are fetched in turn one after the other, multiplied by the corresponding coefficient and added. Thereafter FILTER 0 has been calculated.

3) FILTER 1 is calculated in the following way. First the recursive left part is calculated by setting data.sub.-- addr to filter.sub.-- addr(1)+index, where index=2. Hereby Y(n-2) is pointed to and multiplied by b.sub.2, whereafter index and therefore also data.sub.-- addr is reduced by 1. The obtained product is added to the result obtained from FILTER 0. In the corresponding way Y(n-1) multiplied by coefficient b.sub.1 is also added to the previously obtained result. At the next decrementation of index the address of position X(n-3)/Y(n) is obtained. The result Y(n) obtained is now written over the previous state variable X(n-3) in this memory position. Now all necessary state variables Y(n), Y(n-1) and Y(n-2) for calculating the transversal write part of FILTER 1 are available. This calculation is done in the same way as the calculation of FILTER 0, the difference being that filter.sub.-- addr(1) and index=2 are used for calculating data.sub.-- addr in the first step.

4) FILTER 2 is calculated in the same way as the recursive part of FILTER 1, the difference being that filter.sub.-- addr(2) is used for the calculation of data.sub.-- addr and d.sub.1, d.sub.2 are used as coefficients. In the initial step index is set equal to 2. When index has reached the value 0 the result W(n) is written over the value Y(n-2).

All these calculation are performed at the stage that is shown in FIG. 4(a). In the transition to FIG. 4(b) mod.sub.-- pointer is decremented by 1 and n is incremented by 1, so that X(n) now is called X(n-1), etc. The next input signal value is written into the position that is denoted X(n), whereafter the calculation cycle above is repeated.

In this way mod.sub.-- pointer is counted down one unit for each of the FIG. 4(a)-(h). When the filters have been calculated in FIG. 4(h) and mod.sub.-- pointer is to be updated a folding back to the position shown in FIG. 4(a) is performed, whereafter the whole procedure is repeated.

From the above it is clear that the memory area of the buffer is utilized very efficiently. In particular the buffer can be addressed modulo n, where n is an arbitrary positive integer. This method has proven to be especially advantageous if many short filters are to be implemented.

A preferred embodiment of the address processor will be described below with reference to FIGS. 5-7.

In accordance with FIG. 5 the address processor can logically be divided into three different sections. In the first, top section buffer constants are stored. In the second, middle section the address calculation is performed. Finally, in the third, bottom section the addresses are stored for accessing the memories at suitable points in time.

For each sampling frequency, sample.sub.-- rate, a pointer, mod.sub.-- pointer, is stored. In the preferred embodiment the address processor comprises a register set S containing four registers for storing four different values of mod.sub.-- pointer, one for each sampling frequency and thereby for each buffer. In the preferred embodiment these registers have a length of five bits. A decrementing circuit DEC, for instance comprising a half subtracter, decrements a chosen pointer and stores the same in the corresponding register after use in accordance with the above. When the pointer reaches the value 0, that is when a folding of the buffer is appropriate, a constant is loaded into the register. This constant can be hard wired to the inputs of the register. The value of the constant is mod.sub.-- numb-1, where mod.sub.-- numb is the length of the buffer. A decrementation of mod.sub.-- pointer is performed when the signal dec.sub.-- mod.sub.-- count is high. A chosen pointer is loaded into a register mod.sub.-- reg before it is decremented or loaded into the calculation section. The constants for buffer size and buffer location in the memory can also be hard wired on the register inputs. The signal sample.sub.-- rate chooses the constants corresponding to a certain sampling frequency for storing in the different registers. Each register in the preferred embodiment comprises four inputs D1-D4, one for each sampling frequency. The constant for buffer size, mod.sub.-- numb, is loaded as a two-complement value, that is as -mod.sub.-- numb. The constant is loaded in a register mod.sub.-- num.sub.-- reg with a length of five bits. The constant for the buffer location in the memory is loaded as mod.sub.-- start in a register mod.sub.-- sta.sub.-- reg with a word length of 6 bits.

The calculation section comprises two registers A.sub.-- reg, that has a length of 6 bits, B-reg, that has a length of 5 bits, and also a 6 bit adder, for instance a so called carry boost adder, where the most significant bit of the adder is a half adder cell.

The address calculation starts by loading either data.sub.-- addr or write.sub.-- addr in B.sub.-- reg, depending on whether a read or write operation is to performed. A.sub.-- reg is loaded with either mod.sub.-- pointer from mod.sub.-- reg on input D2 or 0 on input D1. The loading of register A.sub.-- reg is controlled by a modulo deactivating signal mod.sub.-- dis depending on the chosen addressing mode, either modulo or absolute addressing.

In absolute addressing mode (mod.sub.-- dis is high) 0 is loaded into register A.sub.-- reg, so that the input address, data.sub.-- addr or write.sub.-- addr, is added to 0. This makes the calculation unit transparent for absolute addressing.

In modulo mode the mode.sub.-- pointer loaded into A.sub.-- reg is added to the address loaded in B.sub.-- reg, either data.sub.-- addr or write.sub.-- addr. The result is fed back to input D3 of register B.sub.-- reg.

The two-complement value of mod.sub.-- numb, that is -mod.sub.-- numb, stored in register mod.sub.-- num.sub.-- reg is transferred to register A.sub.-- reg and added to the contents of register B.sub.-- reg. The result is stored in B.sub.-- reg if said result is non-negative. If the result is negative the old value is stored in register B.sub.-- reg for the next operation. Since the register mod.sub.-- num.sub.-- reg has been loaded with the two-complement of mod.sub.-- numb, that is -mod.sub.-- numb, this test procedure is facilitated since the sign of the result from the edition directly can be read out from the most significant bit in said result. This bit over line +/- together with the signals CHSST and IXZ in FIG. 7 controls the loading of register B.sub.-- reg.

The constant mod.sub.-- start transferred from register mod.sub.-- sta.sub.-- reg to register A.sub.-- reg in the next step is added to the contents of register B.sub.-- reg, and the result is stored in a register RREG for a read operation or in register WREG for a write operation. This result is the final memory address for reading or writing.

For reading of coefficients, where absolute addressing is used, the coefficient address coe.sub.-- addr is fed directly to a register CAREG.

The bottom section of the address processor, which is provided for storing of addresses and accessing of the memory, comprises five six bit registers. Registers RREG and WREG have already been mentioned and comprise only one input D1. Registers DAREG and CAREG are connected to the data memory bus DMADDRSST and coefficient memory data bus CMADDRSST, respectively. The fifth register, FBREG, stores read and write addresses in a cyclic way. The addresses are fed back from the output busses CMADDRSST and DMADDRSST, respectively, to the corresponding inputs D1 and D2. This feedback operation is controlled by signals PDMWSST and PCMWSST in FIG. 7.

The feedback addressing scheme is used only when the address processor alternatively gives write and read addresses. This happens when a filter calculation is complete and the result is written into a memory cell simultaneously with the start of a new filter calculation for the next filter and data are read from the memory for this calculation. Writing is performed when CHSST is even and reading is performed when CHSST is odd, see FIG. 6 (in FIGS. 6 and 7 R stands for reading and W for writing).

In the preferred embodiment of the invention eight channels are processed in a cyclic read/write process. Two channels are processed simultaneously so that the last read/write operation will be performed in cycles 6 and 7.

For a more detailed description of the operation of the preferred embodiment of the address processor in accordance with the present invention it is referred to the following signal definitions for the address processor and to the following simulation of the address processor in the simulation language RTL from Texas Instruments.

Sets the address processor in absolute address calculation mode. The calculation of the write address for the coefficient memory starts when this signal goes to high level. "apcsst" forms a part of the instruction word that comes from the controller unit.

Data and write addresses are decremented by one each time the index counter is decremented. The index counter is loaded with the filter length "number.sub.-- of.sub.-- taps". This is all done in the controller unit

It is appreciated that the invention is not limited to the specifically described embodiments and that many modifications can be performed by a man skilled in the art. For instance, the buffer length can be reduced by one so that it becomes the sum of the number of delay elements in the filters that are contained in the buffer, if the latest sample X(n) is stored in an extra register instead. Furthermore, the number of channels, the number of buffers in each channels, the word length for the different registers and the method of loading constants in the different registers can be varied without departure from the inventive idea that is defined in the following patents claims.