Patent Grant
7418467
The present application may be related to the following commonly owned United States patent applications, which were filed on even date herewith and are hereby incorporated herein by reference in their entireties:
U.S. patent application Ser. No. 10/871,794 entitled DIGITAL FILTER USING MEMORY TO EMULATE A VARIABLE SHIFT REGISTER; and
U.S. patent application Ser. No. 10/871,509 entitled MICRO-PROGRAMMABLE FILTER ENGINE.
The present invention relates generally to communication systems, and more particularly to a micro-programmable digital filter having multiple programmable filter elements that can be configured, controlled, and combined in different ways to implement different types of filters.
Certain communication devices transmit and receive signals over a communication medium, such as a wireline, wireless, or optical communication medium. These communication devices typically include digital filters for performing various filtering operations, such as Finite Impulse Response (FIR) filtering, Infinite Impulse Response (IIR) filtering, decimation, interpolation, and echo cancellation. The types and configurations of digital filters (e.g., the number of taps or bi-quads) for a particular implementation are typically selected based on the type of communication system and the expected characteristics of the communication medium. Often, the types and configurations of digital filters selected for a particular implementation are tradeoffs to obtain acceptable performance over a range of possible conditions.
In one aspect of the invention, there is provided a micro-programmable digital filter including a plurality of programmable filter elements, an instruction memory for storing a control program, at least one instruction decoder for programming the plurality of programmable filter elements based on the control program, and arithmetic logic for selectively scaling and accumulating output values received from the plurality of programmable filter elements and providing accumulated values as inputs to the plurality of programmable filter elements. Each filter element may include at least one memory for storing data is samples and coefficients, a multiplier for multiplying data samples read from the at least one memory with corresponding coefficients read from the at least one memory, an accumulator for summing multiplier outputs, and control logic for, among other things, logically shifting the data samples read from the at least one memory and writing the logically shifted data samples back into the at least one memory so as to emulate a shift register. The filter elements are typically coupled in series so that a coefficient output from one filter element can be passed to an adjacent filter element for implementing longer filters. Preferred embodiments of the present invention include a separate instruction decoder for each filter element. Execution of the separate instruction decoders is typically staggered so that the instruction decoders can share the instruction memory and an accumulator without collision. Each filter element can be programmed to implement a different filtering function, or multiple filter elements can be programmed to operate on a single filtering function.
One or more of the filter elements may include at least one memory for storing data samples and coefficients, a multiplier for multiplying data samples read from the at least one memory with corresponding coefficients read from the at least one memory, an accumulator for summing multiplier outputs, and control logic for controlling the at least one memory, the multiplier, and the accumulator. The control logic logically shifts the data samples read from the at least one memory and writes the logically shifted data samples back into the at least one memory so as to emulate a shift register.
The filter elements may be coupled in series. Thus, a first filter element may include a coefficient output, and a second filter element may include a coefficient input coupled to the coefficient output.
A separate instruction decoder may be included for each of the plurality of programmable filter elements. Execution of the separate instruction decoders may be staggered so as to allow the instruction decoders to share the instruction memory and/or share a common accumulator connected to the filters' outputs. For example, the at least one instruction decoder may include a first decoder and a second decoder, where execution of the second decoder is delayed by one or more clock cycles from execution of the first decoder. A single program counter may be used to control execution of the control program by the instruction decoders.
The programmable interconnection logic may include a multiplexer coupled to the plurality of programmable filter elements for selectively driving the output values received from the plurality of programmable filter elements, a barrel shifter coupled to the multiplexer for receiving the values and selectively scaling the values according to a scaling factor programmed by the filter controller, an accumulator coupled to the barrel shifter for selectively accumulating scaled values from the barrel shifter, and feedback logic coupled to the accumulator and to the plurality of filter elements for selectively providing the accumulated values as an inputs to the plurality of programmable filter units.
The control program may allow each programmable filter element to operate independently to perform different filtering functions. The control program may allow operation of multiple programmable filter elements to be combined to perform a single filtering function.
In another aspect of the invention, there is provided apparatus for digital filtering including a plurality of programmable filter elements, means for programming the plurality of programmable filter elements based on a control program, and means for selectively combining, scaling, and accumulating output values received from the plurality of programmable filter elements and for selectively providing accumulated values as inputs to the plurality of programmable filter elements.
One or more of the filter elements may include at least one memory for storing data samples and coefficients, means for combining data samples read from the at least one memory with corresponding coefficients read from the at least one memory, and means for logically shifting the data samples read from the at least one memory and writing the logically shifted data samples back into the at least one memory so as to emulate a shift register.
The filter elements may be coupled in series. Thus, the apparatus may include means for coupling a coefficient output of a first filter element to a coefficient input of a second filter element.
The means for programming the plurality of programmable filter elements based on a control program may include an instruction memory and means for sharing the instruction memory among a plurality of separate instruction decoders so as to allow the instruction decoders to share the instruction memory. The means for sharing the instruction memory among a plurality of instruction decoders may include means for staggering execution of the separate instruction decoders. The means for staggering execution of the separate instruction decoders may include means for delaying execution of a first instruction decoder by one or more clock cycles from execution of a second instruction decoder. The apparatus may include means for controlling execution of the control program by a plurality of instruction decoders.
The means for selectively scaling and accumulating output values received from the plurality of programmable filter elements and providing accumulated values as inputs to the plurality of programmable filter elements may include means for selecting an output from one of the plurality of programmable filter elements, means for scaling the selected output according to a programmable scaling factor, and means for accumulating the scaled output and selectively providing the accumulated value as an input to at least one of the plurality of programmable filter units.
The control program may include means for allowing each programmable filter element to operate independently to perform different filtering functions. The control program may include means for allowing operation of multiple programmable filter elements to be combined to perform a single filtering function.
In another aspect of the invention, there is provided a communication device including a transceiver for transmitting and receiving communication signals and a digital filter in communication with the transceiver for processing digitized data samples corresponding to the communication signals. The digital filter includes a plurality of programmable filter elements, programmable interconnection logic coupled to the plurality of programmable filter elements for selectively combining, scaling, and accumulating output values received from the plurality of programmable filter elements and providing accumulated values as inputs to the plurality of programmable filter elements, and a filter controller coupled to the plurality of programmable filter elements and the programmable interconnection logic for controlling the plurality of programmable filter elements and the programmable interconnection logic. The filter controller includes an instruction memory for storing a control program and at least one instruction decoder coupled to the instruction memory for programming the plurality of programmable filter elements and the programmable interconnection logic based on the control program.
One or more of the filter elements may include at least one memory for storing data samples and coefficients, a multiplier for multiplying data samples read from the at least one memory with corresponding coefficients read from the at least one memory, an accumulator for summing multiplier outputs, and control logic for controlling the at least one memory, the multiplier, and the accumulator. The control logic logically shifts the data samples read from the at least one memory and writes the logically shifted data samples back into the at least one memory so as to emulate a shift register. A separate instruction decoder may be included for each of the plurality of programmable filter elements.
In the accompanying drawings:
In accordance with embodiments of the present invention, a digital filter includes a plurality of programmable filter elements, an instruction memory for storing a control program, at least one instruction decoder for programming the plurality of programmable filter elements based on the control program, and arithmetic logic for selectively scaling and accumulating output values received from the plurality of programmable filter elements and providing accumulated values as inputs to the plurality of programmable filter elements. Each filter element may include at least one memory for storing data samples and coefficients, a multiplier for multiplying data samples read from the at least one memory with corresponding coefficients read from the at least one memory, an accumulator for summing multiplier outputs, and control logic for, among other things, logically shifting the data samples read from the at least one memory and writing the logically shifted data samples back into the at least one memory so as to emulate a shift register. The filter elements are typically coupled in series so that a coefficient output from one filter element can be passed to an adjacent filter element for implementing longer filters. Preferred embodiments of the present invention include a separate instruction decoder for each filter element. Execution of the separate instruction decoders is typically staggered so that the instruction decoders can share the instruction memory without collision. Each filter element can be programmed to implement a different filtering function, or multiple filter elements can be programmed to operate on a single filtering function.
The described digital filter can be used alone, or multiple such digital filters can be used in combinations to implement different length filters and/or different types of filters. An exemplary embodiment of the digital filter is described below, and is referred to as an FE2 filter element. An embodiment employing multiple FE2 filter elements is described below, and is referred to as a micro-programmable filter engine (MFE).
In accordance with certain embodiments of the present invention, a micro-programmable filter engine (MFE) provides a flexible and programmable digital filter architecture for implementing various digital filters in hardware in a communication device. The MFE includes multiple programmable filter elements and a microcode controller. The filter elements can be configured, controlled, and combined in different ways to implement different types of filters. The MFE preferably supports multiple-execution instructions that allow a single instruction to perform multiple moves into accumulators for efficient data movement inside MFE. Various aspects of the present invention are described herein with reference to embodiments for a digital subscriber line (DSL) modem application, although the present invention is in no way limited to such an application, and various embodiments of the present invention can be used in other types of communication devices and applications.
Within the DSL modem 110, digital filtering is used to perform such functions as finite impulse response (FIR) filtering, infinite impulse response (IIR) filtering, echo cancellation, decimation, and interpolation. It should be noted that the type(s) of filtering, the topology of the filters (i.e., where in the transmitter and receiver paths the filtering gets done), and the filter parameters can be selected for a particular implementation, DSL version, or line condition.
In order to support various types of digital filtering functions, embodiments of the present invention preferably include a micro-programmable filter engine (MFE) that can be programmed to implement multiple types of filter and perform multiple filtering operations essentially in parallel.
Within the MFE 700, the FEs can be configured individually and can be configured so as to perform multiple filtering functions simultaneously. A single FE 710 can be used to perform a particular filtering function, or multiple FEs 710 can be “cascaded” to form longer filters as discussed below. The microcode controller 750 can control the multiplexers 730 to direct any FE 710 to any accumulator 740 and can control the multiplexers 730 so that multiple FEs 710 are directed to one accumulator 740.
Thus, a collection of filter elements can be used to implement a single, large filter or a number of small filters, by simply configuring the operation of each filter element. The filter element has the property of being cascadable, or connected in series, to allow the implementation of large filters. A large N-tap filter is implemented across multiple filter elements by computing partial sum of products in each filter element and then summing the filter element outputs. Preferred embodiments are flexible enough to allow execution of FIR filters, IIR filters, and different FIR variations such as decimation and interpolation.
In an exemplary embodiment of the present invention, the MFE includes two types of FEs. The first type of FE (referred to hereinafter as the FE0 filter element) can be used to implement FIR filters as well as other functions such as decimation and interpolation. The second type of FE (referred to hereinafter as the FE2 filter element) includes two FE0 filter elements and additional logic, and can be used to implement both IIR filters and FIR filters as well as other functions such as decimation and interpolation. The FE0 and FE2 filter elements are described more fully below.
The FE0 filter element is a basic filter element designed specifically for implementing FIR filters, but can also be used to implement other FIR-like functions such as decimation and interpolation. The preferred FE0 can support up to a 64-tap filter, although multiple FE0s can be cascaded to form longer FIR filters.
In traditional N-tap FIR filter implementations, data passes through an N-stage shift register where the output of each stage is multiplied with a corresponding coefficient, and the sum of the products constitutes the filter output.
In preferred embodiments of the present invention, the FE0 filter element uses memory instead of registers to store both data and filter coefficients in order to reduce the amount of hardware necessary. Specifically, the shift registers and the coefficient registers are replaced with small random access memories and associated control logic. For convenience, the memory that replaces the shift register is referred to as the Data Memory (DM) and the memory that replaces the coefficient registers is referred to as the Coefficient Memory (CM). Among other things, the control logic manipulates the data in the DM so as to emulate a shift register, specifically by performing appropriate “read-modify-write” operations (e.g., read a word from the memory, shift it one bit, and write it back to the memory). The FE0 architecture takes advantage of the speed of modern integrated circuitry and iterates in time the multiply-accumulate function of the filter in such a way that allows on-the-fly configuration of the filter properties. A much faster clock than the frequency of the incoming data clock is used to operate the filter. The size of the DM and the CM is determined by the ratio of the system clock to that of the incoming data, such that, for each data input, a filter output is generated.
The FE2 filter element is a more complex filter element designed specifically for implementing both FIR and IIR filters, but can also be used to implement other FIR-like functions such as decimation and interpolation. In preferred embodiments of the invention, each FE2 contains two FE0 filter elements and additional hardware, including a filter controller (FC), a barrel shifter, and an output accumulator. The FE0 blocks can operate independently or can be cascaded to form various filters. The FE2 can be used to implement some number of bi-quads for IIR filtering (where the number of bi-quads depends on, among other things, the number of clock cycles per symbol) or up to a 128-tap FIR filter. In the FE2, the computation resolution of the IIR filter is twice that of the FIR filter. The IIR filter execution is based on microcode stored in a small instruction memory within the FC. Each instruction is capable of executing multiple data moves.
In certain embodiments of the present invention, the FC includes an instruction memory, a program counter, and two identical instruction decoders, one for each FE0 filter element. A delay is introduced between the two instruction decoders so that the two instruction decoders operate in a ping-pong fashion in order to share the single instruction memory. A program to execute a bi-quad starts at time n, is decoded with the first instruction decoder, and runs on the first FE0 filter element. The same program, delayed by one clock, is decoded with the second instruction decoder, and runs on the second FE0 filter element. The delay facilitates resource sharing when partial results need to be scaled in the barrel shifter or partially summed at the output accumulator.
For IIR filtering, the number of bi-quads that can be executed by the FE2 filter element 1100 is dependent on the rate of the incoming data and the execution length for each bi-quad. Assuming that an input sample arrives every M clock cycles and that a bi-quad is computed in W clock cycles, the maximum number of bi-quads that can be calculated is the integer result of the ratio M/W. During execution of K bi-quads (K>1), the microcode repeats the same instruction sequence K times in a zero-overhead loop. When in FIR mode, only the filter elements 1120 and 1130, multiplexer 1140, barrel shifter 1150, and output accumulator 1160 portions of the block are active. The number of taps N in the FIR filter is dictated by the ratio of the system clock to the data clock, where the data clock is defined as the slower rate of either rate at which data is produced by the filter or applied to the filter. Each FE01120, 1130 is capable of computing up to N taps and the two FE0s 1120, 1130 can be cascaded to compute up to 2N taps. Running concurrently, each FE0 produces a partial result that gets summed to the output accumulator 1160 in the final tally.
In preferred embodiments of the invention, the double-precision data flow graph of the FE2 is split into two merged single precision data flow graphs. This allows double-precision operations of the FE2 to be run on two FE0 functional blocks. Specifically, as described above, each FE0 functional block is a 16-bit block, and the FE2 functional block is a 32-bit block. Implementing an IIR filter using the FE2 functional block involves multiplying a 16-bit coefficient by a 32-bit value to produce a 48-bit value (i.e., C[15:0]*D[31:0]). The multiplication problem is preferably split into two parts, specifically C[15:0]*D[31:16]+C[15:0]*D[15:0]. The barrel shifter and accumulator allow the upper product and the lower product to be added. The 48-bit values are fed back into the two FE0 blocks (see
A particular feedback technique combines inter-bi-quad scaling with the summation of intermediate values. With reference to
Within the MFE, each FE0 is typically connected to two adjacent filter elements. Specifically, the CREG_OUT output 1228 of one FE0 is connected to the CREG_IN input 1226 of the adjacent FE0. A similar interconnection between the two FE0 filter elements of the FE2 filter element exists.
The FE0 is event driven. Therefore, a single pulse on XCLK 1230 triggers the element to complete a single sweep of execution, as programmed in CFR 1210. Single loops or nested loops can be run, as described below.
MAC(n)={DS(n)*C0+DS(n−1)*C1+ . . . +DS(n−7) *C7}
or,
MAC(n)=Σ7i=0DS(n−i)*Ci
where DS(x) represents data sample x and Cy represents coefficient y. These conventions will be continued in the discussion below.
When programmed as an FIR filter, the FE0 essentially works as follows. First, it is assumed that DM[7:0] contains {DS(n−8) . . . DS(n−1)} and CM[7:0] contains {C7 . . . C0}. The MAC accumulates [Ry*Rc], where Ry is register Y 1208 and Rc is register C 1214.
At Mclk(1), the current content of DM(0), which is DS(n−1), is loaded into Ry 1208 and the current content of CM(0), which is C0, is loaded into Rc 1214. Rx(n) is selected through multiplexer DMnxt 1236 and is written into DM(0), or, DS(n) which was loaded into register Rx 1202 on the last Sclk 1234 event is DM location 0. The MAC is loaded with DS(n−8)*C7. The MAC content at this point is transferred to a holding register 1224 because it contains a sample output.
At Mclk(2), DM(1), which is DS(n−2), is loaded into Ry 1208 and CM(1) is loaded into Rc 1214. DMnxt 1236 points to Ry 1208 and the current content of Ry 1208, which is DS(n−1), is written into DM(1). The MAC is overwritten with DS(n−1)*C0. The overwrite operation is important because it sets up the MAC for computing the next sample output.
At Mclk(3)-Mclk(8), the process described above for Mclk(2) is repeated. DM(k) is DS(n−k−1) and is loaded into Ry 1208 while CM(K) is loaded into Rc 1214. The current content of Ry 1208 is written into DM(k). The MAC is accumulated with Ry*Rc.
The dual loop mode is used to generate the proper addressing for executing the decimator and interpolator functions as described previously. The following is a description of how this mode operates. It is assumed that LP1 counts down LP1_N times and its initial value is LP1_BASE (loaded in LP1_CNTR) and that LP0 counts LP0_N times (loaded in LP0_CNTR) with a displacement of LP0_INCR. The first value of LP1 (LP1_BASE) is loaded into the output REG to coincide with the assertion of Xclk, LP1_CNTR decrements by 1 so that LP1_CNTR contains (LP1_BASE−1), and LP0_CNTR pre-loads with the value LP0_N. On the next LP0_N Mclk, the content of output REG is LP1_BASE+n*LP0_INCR, where n∈1, 2 . . . LP0_N. The corresponding value in LP0_CNTR is LP0_N−n. When LP0_CNTR=0, then the next clock coincides with Xclk and the new value of LP1_BASE is loaded into the output REG, and the second pass of LP1 begins. The entire process described above is repeated until both loop counters LP1 and LP0 expire, or equal zero. The output REG is loaded with LP1_BASE in anticipation of the next Xclk. The number of Mclks for completing a sweep is (LP1_BASE+1)*(LP0_N+1).
The single loop mode provides the addressing generation function for simple N-tap FIR filters. It is assumed that LP0 counts LP0_N times (loaded in LP0_CNTR) with a displacement of LP0_INCR. During each sweep, LP0_CNTR counts up LP0_N times and the output REG=n*LP0_INCR. At the end of the sweep, the output REG is reset to zero.
The memory by pass mode is optimized for a fast, symmetric or non-symmetric FIR filter operation. CFG1 and CFG2 must have a value of one, CFG0 a value of one or two depending on its position in the chain of cascaded filters, and the end filter position requires a value of two. Every Xclk, samples are shifted serially through the filter element and the sum of Ry+Rc is multiplied by the CM output. The result is stored in the accumulator (Acc). In non-bypass operation, Rc is multiplied by the default CFG3 value of zero. CFG3 controls the symmetry operation of the filter (zero for non-symmetric, one for symmetric).
The ADG 1700 operates on the principle that Xclk triggers a single sweep and a sweep is composed of one or two nested loops, namely LP0 and LP1. The INIT CNTR 1720 counts up or down by one. The rest of the structure counts by the displacement of INCR REG 1710 content. LP0 and LP1 counters 1760 and 1770, respectively, control the operation of the nested loops.
The INCR REG 1710 content is a two's complement value ranging from −32 to +31. The adder (ADDR) 1730 output is always positive and ranges between 0 and 63. The INIT CNTR 1720 is always positive. The adder 1730 receives as inputs the contents of INCR REG 1710 and the contents of REG 1750 and sign extends both inputs to seven bits in order to behave correctly and yield a result in the proper range. The multiplexer (mux) 1740 allows either the contents of the INIT CNTR 1720 or the output of the adder 1730 to be directed to REG 1750. The CTL REG 1780 controls operation of the ADG.
An instruction based controller, or engine, can take over the filter element control through the filter controller (FC) port. The FC port provides the flexibility to customize the filter operation being implemented in the FE0 data path. The FE0 is thus capable of implementing FIR and IIR filtering functions. In FE2 configuration, the FC port is used to pass control to FE0.
The following is a description of various FE0 signals including the signal name, direction (I=input, 0=output), and width:
For the purpose of testing, registers, adders, multiplexers and MAC will be tested using scan. Memory testing uses special provisioning. Specifically, DM and CM are enclosed in a special wrapper that bypasses the physical memory during scan test. The bypass permits all inputs to connect to the 16-bit output bus of the wrapper in order to provide visibility to the memory control and data inputs. The MPU port is used to write data directly to either DM or CM of any FE0. The read back path to the MPU is not direct but goes through Ry and Rc for DM, or, Rc for CM, of the FE0 under test, plus Rc of each FE0 between the first FE0 and the FE0 under test. The first FE0 is the unit where CREG_OUT drives the MPU data bus.
In an exemplary embodiment of the present invention, the configuration registers (CFG) are defined as follows:
Name=CFG CTL Addr=0x00
Name=LP0 CTL Addr=0x01
Name=LP1CTL Addr=0x02
Name=INIT Addr=0x03
Name=INCR Addr=0x04
Name=intctl Addr=0x05
The following table describes the FE2 internal registers:
The FE2 instruction memory is 32 words long with a width of 16 bits. Instruction bits [15:14] indicate the type of instruction, where [0:0] is used for setup instructions, [0:1] is used for start/stop instructions, [1:0] is used for external math instructions, and [1:1] is used for internal math instructions. Setup instructions control the configuration registers. Start instructions operate on external data and perform internal data pre-fetch. The stop instruction controls operations that occur at the end of a loop. The internal math instruction controls internal data storage and math operators. The external math instruction controls the accumulator and shifter at the FE2 top level.
Setup instructions control all of the configuration registers. Setup instructions have the following format:
The wt iir [13] and iir [12] bits are used to control IIR mode. When IIR mode is disabled (0=default), the FE0s operate using their own control registers; when IIR mode is enabled (1), the FE0s are controlled by the FC program instructions. When the value in wt iir [13] is high (1), then IIR mode is either enabled or disabled according to the value of iir [12]; when the value in wt iir [13] is low (0), then the IIR mode is unchanged.
The wt slv [11] and slave [10] bits are used to control slave mode. When slave mode is disabled (0=default), the FE2 receives external input is from a filter operating on a separate filtering function; when slave mode is enabled (1), the external input is from a linked filter operating on the same iir function so that input from a linked filter comes in from a different port, uses the xclk as an input ready signal, and is fed into the FE0s using special multiplexers. When the value in wt slv [11] is high (1), then slave mode is either enabled or disabled according to the value of slave [10]; when the value in wt slv is low (0), then slave mode is unchanged.
The loop [9], on [8], and loop n [7-5] bits are used to control looping. When the loop [9] bit is high (1), the value of the on [8] bit controls whether a loop is active (1) or inactive (0), and the value of the loop n [7-5] bits specifies the number of iterations that the loop will run, where a value of n runs the loop (n+1) times. The loop start and end address values are determined by a separate instruction (eval).
The offset [4], set [3], and offset n [2-0] bits are used to control offsets. If there is a running loop, the default offset is the current iteration of the running loop (3′h0 if there is no loop). The value of Offset_n is appended to the end of the data and the coefficient memory addresses. When Offset_n is set, by Offset_on equaling 1, then the value of Offset_n is offset_n. The write enable signal offset is used to write to the register Offset_on the signal set and Offset_n the signal offset_n.
Start instructions control and operate on external input data, and also allow “pre-fetch” from the data and coefficient memories inside the FE0 filter elements. Start instructions have the following format:
The fd in [12] bit is used to allow external input (from the port FD) to bypass the FE0s to the shifting logic to be fed back to the FE0s on the next clock cycle. This is only used if slave mode is disabled.
Stop instructions control operations that occur at the end of a loop or IIR operation. Stop instructions have the following format:
Internal math instructions control the data storage, multiplier, and accumulator internal to the FE0 modules. The following is the format of internal math instructions:
External math instructions control the accumulator and shifter that are external to the FE0 modules, and also allow for “pre-fetch” of memory. The following is the format of external math instructions:
An IIR filter utilizes five coefficient values plus a coefficient with value one for each second order calculation. Each group of six coefficients is stored in addresses zero through seven. The CM address mask specifies up to eight groups of coefficients. The coefficients are addressed as (offset/loop-cntr[2:0], cmrda[2:0]), i.e., the coefficients for the last stage of a sixth order IIR are contained in the address range 010000-010110. The programmer has the freedom of deciding which address within a range contains a particular coefficient. The FE2 typically finishes a basic IIR bi-quad in 15 mclk cycles, and is capable of finishing n bi-quads in 3+12n mclk cycles. For 8.8 MHz sample clock frequency, one bi-quad per FE2 can be completed within that period. For a 4.4 MHz sample clock, two bi-quads, and at a 2.2 MHz frequency, five bi-quad operations can be performed in a single FE2 unit.
The looping capability facilitates implementing IIR filters in stages of second order configuration. Each loop path executes a second order IIR filter as shown in
The MFE operates in a single zero-overhead tight loop either indefinitely or as specified by the user. Multiple code images loaded in program memory afford the engine the ability of context switching on symbol or frame boundaries, as selected by the user. Arithmetic, rounding, and scaling operations performed by the MFE have inherent overflow protection.
MFE instructions are 25 bits long. The instruction set contains dedicated fields for symbol clocks to the individual elements and a wait field to embed execution control within the instruction set.
Instruction bit 24 is the wait (Wn) field. When set, the Wn bit is an immediate execution instruction that the MFE recognizes on the next clock and causes execution of an indefinite wait following completion of executing the current instruction. This wait is terminated by assertion of the input trigger.
Instruction bits 23-20 (C3-C0, respectively) are dedicated bits for providing four soft clocks to all FEs. Each FE requires two clock rates, namely Xclk and Fclk, which are provided by these four bits C3-C0. A soft clock connection is made between one of C3-C0 to an FE's Xclk or Fclk by executing a setup instruction (CLKSET), as described below. The frequency of each clock is determined by the number of system clocks in a loop and the number of 1's and 0's in each field (C3-C0). For example, if the symbol clock rate is 64 system clocks, then the program loop length should be a multiple of 64 clocks. By placing a single 1 in one instruction at the C0 field, while keeping the same field 0 in the rest of the 63 instructions yields the correct symbol clock rate. Double the frequency is generated by setting 2 bit fields at a distance of 32 instructions apart and so on.
MFE instructions support simultaneous moves or a move plus arithmetic operation or two moves plus either a loop evaluation/jump operation. An Amove operation is from ASRC (Source A) to ADST (Destination A) and a Bmove operation is from BSRC (Source B) to BDST (Destination B). ASRC is selected from the set {E14 . . . E0}, and is encoded as shown in the Source/Destination Resources Address table below. ADST is selected from the set {Ureg (1), Vreg (0)}. BSRC is selected from the set {Ureg (10), Vreg (00), Rxtrmreg (01), IFFTreg (11)}. BDST is selected from the set {E14 . . . E0, FFTreg, Txtrmreg, Ureg, Vreg}, and is encoded as shown in the Source/Destination Resources Address table below.
The MFE supports two nested loops. The outer loop is specified by an immediate jump instruction JPMOV while the inner loop is controlled with a LPMOV instruction, where the number of iterations is declared.
The following is the format of the CLR (clear) instruction for an exemplary embodiment of the present invention:
The CLR (clear) instruction clears all resources in the datapath.
The following is the format of the ADSTSET (destination A set) instruction for an exemplary embodiment of the present invention:
The LP [5] bit is used to select between LPMOV and JPMOV registers that hold ADST for each instruction. Only a single ADST is assigned to each of the two instructions. The EN [4] bit is used to enable (1) the implicit ADST field for instructions LPMOV and JPMOV or disable (0) the Amove operation. The RS [3] bit is used to select Vreg (0) or Ureg (1). The OWR [2] bit is used to overwrite the selected register. The SUB [1] bit is used to subtract ASRC from the selected register (i.e., add the two's complement of ASRC to the register). The RND [0] bit is used to round the sum before storing in the selected register.
The following is the format of the FLGSET (flag set) instruction for an exemplary embodiment of the present invention:
The FLGSET instruction is used to modify flag fields F7-F0.
The following is the format of the STBSET (strobe set) instruction for an exemplary embodiment of the present invention:
The STBSET instruction asserts a four-bit output strobe mfe_stb[3:0] for one clock period. The strobes may be used to request data from an external resource or to signal data available. The data valid input from the resource is used to register the dataset.
The following is the format of the MODESET (mode set) instruction for an exemplary embodiment of the present invention:
The MODESET instruction determines the MFE operational mode. Bits [1] and [0] are input trigger enable bits (1=enable).
The following is the format of the ARITH (arithmetic) instruction for an exemplary embodiment of the present invention:
The OWR [11] bit is used to overwrite the selected register with the value in ASRC [15-12]. The RND [10] bit is used to enable rounding. The AOP [9-8] bits are used to select arithmetic operations, where 00=BSRC+ASRC, 01=BSRC−ASRC, 10=Ureg−Vreg, and 11=Ureg+Vreg. The RS [7] register select bit is used to select a destination register. The SCALE [4-0] bit are used to shift the result.
The following is the format of the CLKSET (clock set) instruction for an exemplary embodiment of the present invention:
The CLKSET instruction writes a four-bit mask register that specifies to the MFE the source of XCLK[FCLK] to the destination FE or resource addressed by the Addr field. The decoding of ADDR is per the Source/Destination Resources Address table below. Each FE has the source of its XCLK[FCLK] specified by the two-bit field in this mask. The decoding of FCSEL/XCSEL is as follows: 00=C0, 01=C1, 10=C2, 11=C3.
The following is the format of the LPSET (loop set) instruction for an exemplary embodiment of the present invention:
The LPSET instruction sets up loop control. The SI [15] bit is used to start an indefinite zero-overhead loop whose end is marked with a LPMOV instruction. An indefinite loop returns to the instruction that follows a LPSET. The SC [14] bit is used to start a loop that repeats LPLEN [11-0] times, whose end is marked with a LPMOV instruction.
The following is the format of the JPSET (jump set) instruction for an exemplary embodiment of the present invention:
The JPSET instruction defines the next instruction address when a JPMOV is executed at JPADDR. The SEL [15-14] bits provide the register address to which the Jump Address is written. The encoding of this field is as follows: 00=JPA0 (Default), 01=JPA1, 10=JPA2, 11=JPA3. The JPADDR [9-0] bits are the physical address.
The following is the format of the JPMOV (ump move) instruction for an exemplary embodiment of the present invention:
The JPMOV instruction is typically the next-to-last instruction in the main loop. Two move operations and an immediate jump to address JPADDR are executed. The ADST field is defined by instruction ADSTSET, which generally must be executed prior to this instruction. The Bmove source BSRC is limited to Ureg (when 1) and Vreg (when 0). BDST is defined in the Source/Destination Resources Address table. A Bmove can move data from Ureg to Vreg but not the reverse. The SCALE field allows scaling to be performed by selecting specified output bits. The SCALE field is encoded according to the Barrel Shift table below. The scale operation applies to the Amove only.
The following is the format of the LPMOV (loop move) instruction for an exemplary embodiment of the present invention:
The LPMOV instruction is typically the next-to-last instruction in an inner loop with a finite number of iterations. Two move operations and an immediate jump to instruction that follows a LPSET are executed. The ADST field is defined by instruction ADSTSET, which generally must be executed prior to this instruction. The Bmove source BSRC is limited to Ureg (when 1) and Vreg (when 0). BDST is defined in the Source/Destination Resources Address table. A Bmove can move data from Ureg to Vreg but not the reverse. The SCALE field allows scaling to be performed by selecting specified output bits. The SCALE field is encoded according to the Barrel Shift table below. The scale operation applies to the Amove only.
The following is the format of the MOV (move) instruction for an exemplary embodiment of the present invention:
The MOV instruction moves data from ASRC to ADST with scaling applied simultaneously with a move from BSRC to BDST. The Amove is not executed if ASRC=0x0F (i.e., hexadecimal 0F). Similarly, the Bmove is not executed if the BDST=0x1F. The ADST field includes three bits, namely the OWR [18] overwrite bit, the RS [17] register select bit, and the Round [16] bit. The SCALE field allows scaling to be performed by selecting specified output bits. The SCALE field is encoded according to the Barrel Shift table below.
The following is the Barrel Shift table used for scaling, as discussed above:
The following is the Source/Destination Resources Address table used for addressing as discussed above:
The following is an exemplary address map for the MFE 2100. With reference to
FE0_0:
DMA Access
CRB Access
FE0—1:
DMA Access
CRB Access
FE0_2:
DMA Access
CRB Access
FE0_3:
DMA Access
CRB Access
FE0_4:
DMA Access
CRB Access
FE0_5:
DMA access
CRB Access
FE0_6:
DMA Access
CRB Access
FE0_7:
DMA Access
CRB Access
FE0_8:
DMA access
CRB Access
FE2_9:
DMA access
CRB Access
FE2_10:
DMA Access
CRB Access
FE2_11:
DMA Access
CRB Access
FE2_12:
DMA Access
CRB Access
FE2_13:
DMA Access
CRB Access
FE2_14:
DMA Access
CRB Access
Because the MFE is programmable, various filter configuration profiles can be pre-defined for use with various respective line conditions, and the MFE can be programmed with an appropriate configuration profile based on actual or expected line conditions. Each configuration profile can define such things as the type(s) of filters, the topology of the filters (i.e., where in the path the filtering gets done), and the filter parameters. Line conditions, such as echoes, noise, and frequency response, can be characterized, for example, using line probing or other active and/or passive characterization techniques. The appropriate configuration profile can be selected based on the results of the characterization.
It should be noted that the following claims may use the term “at least one” to indicate the inclusion of one or more of a particular element, but the omission of that term from a particular claim element is not to be construed as a limitation to just one of that element.
The present invention may be embodied in other specific forms without departing from the true scope of the invention. The described embodiments are to be considered in all respects only as illustrative and not restrictive.
| Number | Name | Date | Kind |
|---|---|---|---|
| 4491910 | Caudel et al. | Jan 1985 | A |
| 4959815 | Wada | Sep 1990 | A |
| 5047972 | Wada | Sep 1991 | A |
| 5243551 | Knowles et al. | Sep 1993 | A |
| 5432723 | Chen et al. | Jul 1995 | A |
| 5500811 | Corry | Mar 1996 | A |
| 6263354 | Gandhi | Jul 2001 | B1 |
| 6279019 | Oh et al. | Aug 2001 | B1 |
| 6314393 | Zheng et al. | Nov 2001 | B1 |
| 6823353 | Fischer et al. | Nov 2004 | B2 |
| 6952709 | Dujardin et al. | Oct 2005 | B1 |
| 20020013798 | Willson et al. | Jan 2002 | A1 |
| 20020161806 | Shaikh | Oct 2002 | A1 |
| 20040101039 | Glendenning | May 2004 | A1 |
| 20040181564 | MacInnis et al. | Sep 2004 | A1 |
| 20050251542 | Hennedy et al. | Nov 2005 | A1 |
| 20050283509 | Hennedy et al. | Dec 2005 | A1 |
| Number | Date | Country | |
|---|---|---|---|
| 20050283508 A1 | Dec 2005 | US |
