21. Filter math accelerator (FMAC)

The FMAC section only applies to STM32H56x/573 MCUs.

21.1 FMAC introduction

The filter math accelerator unit performs arithmetic operations on vectors. It comprises a multiplier/accumulator (MAC) unit, together with address generation logic which allows it to index vector elements held in local memory.

The unit includes support for circular buffers on input and output, which allows digital filters to be implemented. Both finite and infinite impulse response filters can be realized.

The unit allows frequent or lengthy filtering operations to be offloaded from the CPU, freeing up the processor for other tasks. In many cases it can accelerate such calculations compared to a software implementation, resulting in a speed-up of time critical tasks.

21.2 FMAC main features

• • 16 x 16-bit multiplier
• • 24 + 2-bit accumulator with addition and subtraction
• • 16-bit input and output data
• • 256 x 16-bit local memory
• • Up to three areas can be defined in memory for data buffers (two input, one output), defined by programmable base address pointers and associated size registers
• • Input and output buffers can be circular
• • Filter functions: FIR, IIR (direct form 1)
• • Vector functions: Dot product, convolution, correlation
• • AHB slave interface
• • DMA read and write data channels

21.3 FMAC functional description

21.3.1 General description

The FMAC is shown in Figure 99 .

Figure 99. Block diagram

The unit is built around a fixed point multiplier and accumulator (MAC). The MAC can take two 16-bit input signed values from memory, multiply them together and add them to the contents of the accumulator. The address of the input values in memory is determined using a set of pointers. These pointers can be loaded, incremented, decremented or reset by the internal hardware. The pointer and MAC operations are controlled by a built-in sequencer in order to execute the requested operation.

To calculate a dot product, the two input vectors are loaded into the local memory by the processor or DMA controller, and the requested operation is selected and started. Each pair of input vector elements is fetched from memory, multiplied together and accumulated. When all the vector elements have been processed, the contents of the accumulator are stored in the local memory, from where they can be read out by the processor or DMA.

The finite impulse response (FIR) filter operation (also known as convolution) consists in repeatedly calculating the dot product of the coefficient vector and a vector of input samples, the latter being shifted by one sample delay, with the least recent sample being discarded and a new sample added, at each repetition.

The infinite impulse response (IIR) filter operation is the convolution of the feedback coefficients with the previous output samples, added to the result of the FIR convolution.

A more detailed description of the filter operations is given in Section 21.3.6: Filter functions .

21.3.2 Local memory and buffers

The unit contains a 256 x 16-bit read/write memory which is used for local storage:

• • Input values (the elements of the input vectors) are stored in two buffers, X1 and X2.
• • Output values (the results of the operations) are stored in another buffer, Y.
• • The locations and sizes of the buffers are designated as follows:
  • – x1_base: the base address of the X1 buffer
  • – x2_base: the base address of the X2 buffer
  • – y_base: the base address of the Y buffer
  • – x1_buf_size: the number of 16-bit addresses allocated to the X1 buffer
  • – x2_buf_size: the number of 16-bit addresses allocated to the X2 buffer
  • – y_buf_size: the number of 16-bit addresses allocated to the Y buffer.

These parameters are programmed in the corresponding registers when configuring the unit.

The CPU (or DMA controller) can initialize the contents of each buffer using the Initialization functions ( Section 21.3.5: Initialization functions ) and writing to the write data register. The data is transferred to the location within the target buffer indicated by a write pointer. After each new write, the write pointer is incremented. When the write pointer reaches the end of the allocated buffer space, it wraps back to the base address. This feature is used to load the elements of a vector prior to an operation, or to initialize a filter and load filter coefficients.

Buffer configuration

The buffer sizes and base address offsets must be configured in the X1, X2 and Y buffer configuration registers. For each function, the required buffer size is specified in the function description in Section 21.3.6: Filter functions . The base addresses can be chosen anywhere in internal memory, provided that all buffers fit within the internal memory address range (0x00 to 0xFF), that is, base address + buffer size must be less than 256.

There is no constraint on the size and location of the buffers (they can overlap or even coincide exactly). For filter functions it is recommended not to overlap buffers as this can lead to erroneous behavior.

When circular buffer operation is required, an optional “headroom”, d, can be added to the buffer size. Furthermore, a watermark level can be set, to regulate the CPU or DMA activity. The value of d and the watermark level must be chosen according to the application performance requirements. For maximum throughput, the input buffer must never go empty, so d must be somewhat greater than the watermark level, allowing for any interrupt or DMA latency. On the other hand, if the input data can not be provided as fast as the unit can process them, the buffer can be allowed to empty waiting for the next data to be written, so d can be equal to the watermark level (to ensure that no overflow occurs on the input).

21.3.3 Input buffers

The X1 and X2 buffers are used to store data for input to the MAC. Each multiplication takes a value from the X1 buffer and a value from the X2 buffer and multiplies them together. A pointer in the control unit generates the read address offset (relative to the buffer base address) for each value. The pointers are managed by hardware according to the current function.

Figure 100. Input buffer areas

The X1 buffer can be used as a circular buffer, in which case new data are continually transferred into the input buffer whenever space is available. Pre-loading this buffer is optional for digital filters, since if no input samples have been written in the buffer when the operation is started, it is flagged as empty, which triggers the CPU or DMA to load new samples until there are enough to begin operation. Pre-loading is nevertheless useful in the case of a vector operation, that is, the input data is already available in system memory and circular operation is not required.

Figure 101. Circular input buffer

The X2 buffer can only be used in vector mode (that is not circular), and needs to be pre-loaded, except if the contents of the buffer do not change from one operation to the next. For filter functions, the X2 buffer is used to store the filter coefficients.

When operating as a circular buffer, the space allocated to the buffer ( \( x1\_buf\_size \) ) must generally be bigger than the number of elements in use for the current calculation, so that there are always new values available in the buffer. Figure 101 illustrates the layout of the buffer for a filter operation. While calculating an output sample \( y[n] \) , the unit uses a set of \( N+1 \) input samples, \( x[n-N] \) to \( x[n] \) . When this is finished, the unit starts the calculation of \( y[n+1] \) , using the set of input samples \( x[n-N+1] \) to \( x[n+1] \) . The least-recent input sample, \( x[n-N] \) , drops out of the input set, and a new sample, \( x[n+1] \) , is added to it.

The processor, or DMA controller, must ensure that the new sample \( x[n+1] \) is available in the buffer space when required. If not, the buffer is flagged as empty, which stalls the execution of the unit until a new sample is added. No underflow condition is signaled on the X1 buffer.

Note: If the flow of samples is controlled by a timer or other peripheral such as an ADC, the buffer regularly goes empty, since the filter processes each new sample faster than the source can provide it. This is an essential feature of filter operation.

If the number of free spaces in the buffer is less than the watermark threshold programmed in the FULL_WM bitfield of the FMAC_X1BUFCFG register, the buffer is flagged as full. As long as the full flag is not set, interrupts are generated, if enabled, to request more data for the buffer. The watermark allows several data to be transferred under one interrupt, without danger of overflow. Nevertheless, if an overflow does occur, the OVFL error flag is set and the write data is ignored. The write pointer is not incremented in the event of an overflow.

The operation of the X1 buffer during a filtering operation is illustrated in Figure 102 . This example shows an 8-tap FIR filter with a watermark set to four.

Figure 102. Circular input buffer operation

21.3.4 Output buffer

The Y (output) buffer is used to store the output of an accumulation. Each new output value is stored in the buffer until it is read by the processor or DMA controller. Each time a read access is made to the read data register, the read data is fetched from the address indicated by the read pointer. This pointer is incremented after each read, and wraps back to the base address when it reaches the end of the allocated Y buffer space.

Figure 103. Circular output buffer

The Y buffer can also operate as a circular buffer. If the address for the next output value is the same as that indicated by the read pointer (an unread sample), then the buffer is flagged as full and execution stalled until the sample is read.

In the case of IIR filters, the Y buffer is used to store the set of \( M \) previous output samples, \( y[n-M] \) to \( y[n-1] \) , used for calculating the next output sample \( y[n] \) . Each time a new sample is added to the set, the least recent sample \( y[n-M] \) drops out.

If the number of unread data in the buffer is less than the watermark threshold programmed in the EMPTY_WM bitfield of the FMAC_YBUFCFG register, the buffer is flagged as empty. As long as the empty flag is not set, interrupts or DMA requests are generated, if enabled, to request reads from the buffer. The watermark allows several data to be transferred under one interrupt, without danger of underflow. Nevertheless, if an underflow does occur, the UNFL error flag is set. In this case, the read pointer is not incremented and the read operation returns the content of the memory at the read pointer address.

The operation of the Y buffer in circular mode is illustrated in Figure 104 . This example shows a 7-tap IIR filter with a watermark set to four.

Figure 104. Circular output buffer operation

21.3.5 Initialization functions

The following functions initialize the FMAC unit. They are triggered by writing the appropriate value in the FUNC bitfield of the FMAC_PARAM register, with the START bit set. The P and Q bitfields must also contain the appropriate parameter values for each function as detailed below. The R bitfield is not used. When the function completes, the START bit is automatically reset by hardware.

During initialization, it is recommended that the DMA requests and interrupts be disabled. The transfer of data into the FMAC memory can be done by software or by memory-to-memory DMA transfers, since no flow control is required.

Load X1 buffer

This function pre-loads the X1 buffer with N values, starting from the address in X1_BASE. Successive writes to the FMAC_WDATA register load the write data into the X1 buffer and increment the write address. The write pointer points to the address X1_BASE + N when the function completes.

The function can be used to pre-load the buffer with the elements of a vector, or to initialize the input storage elements of a filter.

Parameters

• • The parameter P contains the number of values, N, to be loaded into the X1 buffer.
• • The parameters Q and R are not used.

The function completes when N writes have been performed to the FMAC_WDATA register.

Load X2 buffer

This function pre-loads the X2 buffer with \( N + M \) values, starting from the address in X2_BASE. Successive writes to the FMAC_WDATA register load the write data into the X2 buffer and increment the write address.

The function can be used to pre-load the buffer with the elements of a vector, or the coefficients of a filter. In the case of an IIR, the \( N \) feed-forward and \( M \) feed-back coefficients are concatenated and loaded together into the X2 buffer. The total number of coefficients is equal to \( N + M \) . For an FIR, there are no feedback coefficients, so \( M = 0 \) .

Parameters

• • The parameter P contains the number of values, \( N \) , to be loaded into the X2 buffer starting from address X2_BASE.
• • The parameter Q contains the number of values, \( M \) , to be loaded into the X2 buffer starting from address X2_BASE + \( N \) .
• • The parameter R is not used.

The function completes when \( N + M \) writes have been performed to the FMAC_WDATA register.

Load Y buffer

This function pre-loads the Y buffer with \( N \) values, starting from the address in Y_BASE. Successive writes to the FMAC_WDATA register load the write data into the Y buffer and increment the write address. The read pointer points to the address Y_BASE + \( N \) when the function completes.

The function can be used to pre-load the feedback storage elements of an IIR filter.

Parameters

• • The parameter P contains the number of values to be loaded into the Y buffer.
• • The parameters Q and R are not used.

The function completes when \( N \) writes have been performed to the FMAC_WDATA register.

21.3.6 Filter functions

The following filter functions are supported by the FMAC unit. These functions are triggered by writing the corresponding value in the FUNC bitfield of the FMAC_PARAM register with the START bit set. The P, Q and R bitfields must also contain the appropriate parameter values for each function as detailed below. The filter functions continue to run until the START bit is reset by software.

Convolution (FIR filter)

This function performs a convolution of a vector B of length \( N+1 \) and a vector X of indefinite length. The elements of Y for incrementing values of \( n \) are calculated as the dot product,

\( y_n = \underline{\mathbf{B}} \cdot \underline{\mathbf{X}}_n \) , where \( \underline{\mathbf{X}}_n = [x_{n-N}, \dots, x_n] \) is composed of the \( N+1 \) elements of \( \underline{\mathbf{X}} \) at indexes \( n - N \) to \( n \) .

This function corresponds to a finite impulse response (FIR) filter, where vector \( \underline{\mathbf{B}} \) contains the filter coefficients and vector \( \underline{\mathbf{X}} \) the sampled data.

The structure of the filter (direct form) is shown in Figure 105 .

Figure 105. FIR filter structure

Note that the cross correlation vector can be calculated by reversing the order of the coefficient vector \( \underline{\mathbf{B}} \) .

Input:

• • X1 buffer contains the elements of vector \( \underline{\mathbf{X}} \) . It is a circular buffer of length \( N + 1 + d \) .
• • X2 buffer contains the elements of vector \( \underline{\mathbf{B}} \) . It is a fixed buffer of length \( N + 1 \) .

Output:

• • Y buffer contains the output values, \( y_n \) . It is a circular buffer of length \( d \) .

• • The parameter P contains the length, N+1, of the coefficient vector B in the range [2:127].
• • The parameter R contains the gain to be applied to the accumulator output. The value output to the Y buffer is multiplied by \( 2^R \) , where R is in the range [0:7]
• • The parameter Q is not used.

The function completes when the START bit in the FMAC_PARAM register is reset by software.

This function implements an infinite impulse response (IIR) filter. The filter output vector Y is the convolution of a coefficient vector B of length N+1 and a vector X of indefinite length, plus the convolution of the delayed output vector Y' with a second coefficient vector A , of length M. The elements of Y for incrementing values of n are calculated as \( y_n = \mathbf{B} \cdot \mathbf{X}_n + \mathbf{A} \cdot \mathbf{Y}_{n-1} \) , where \( \mathbf{X}_n = [x_{n-N}, \dots, x_n] \) comprises the N+1 elements of X at indexes n - N to n, while \( \mathbf{Y}_{n-1} = [y_{n-M}, \dots, y_{n-1}] \) comprises the M elements of Y at indexes n - M to n - 1. The structure of the filter (direct form 1) is shown in Figure 106 .

Figure 106. IIR filter structure (direct form 1)

MSV47127V1

• • X1 buffer contains the elements of vector X . It is a circular buffer of length \( N + 1 + d \) .
• • X2 buffer contains the elements of coefficient vectors B and A concatenated ( \( b_0, b_1, b_2, \dots, b_N, a_1, a_2, \dots, a_M \) ). It is a fixed buffer of length \( M+N+1 \) .

• • Y buffer contains the output values, \( y_n \) . It is a circular buffer of length \( M + d \) .

• • The parameter P contains the length, \( N + 1 \) , of the coefficient vector B in the range [2:64].
• • The parameter Q contains the length, \( M \) , of the coefficient vector A in the range [1:63].
• • The parameter R contains the gain to be applied to the accumulator output. The value output to the Y buffer is multiplied by \( 2^R \) , where R is in the range [0:7].

The function completes when the START bit in the FMAC_PARAM register is reset by software.

21.3.7 Fixed point representation

The FMAC operates in fixed point signed integer format. Input and output values are q1.15.

In q1.15 format, numbers are represented by one sign bit and 15 fractional bits (binary decimal places). The numeric range is therefore \( -1 \) (0x8000) to \( 1 - 2^{-15} \) (0x7FFF).

The accumulator has 26 bits, of which 22 are fractional and 4 are integer/sign (q4.22). This allows it to support partial accumulation sums in the range \( -8 \) (0x2000000) to \( +7.99999976 \) (0x1FFFFFF). A programmable gain from 0dB to 42dB in steps of 6dB can be applied at the output of the accumulator.

Note that the content of the accumulator is not saturated if the numeric range is exceeded. Partial sums whose value is greater than \( +7.99999976 \) or less than \( -8 \) , wrap but this is harmless provided subsequent accumulations undo the wrapping. Nevertheless, the SAT flag in the FMAC_SR register is set if wrapping occurs, and generates an interrupt if the SATIEN bit is set in the FMAC_CR register. This helps in debugging the filter.

The data output by the accumulator can optionally be saturated, after application of the programmable gain, by setting the CLIPEN bit in the FMAC_CR register. If this bit is set, then any value which exceeds the numeric range of the q1.15 output, is set to \( 1 - 2^{-15} \) or \( -1 \) , according to the sign. If clipping is not enabled, the unused accumulator bits after applying the gain is simply truncated.

21.3.8 Implementing FIR filters with the FMAC

The FMAC supports FIR filters of length N, where N is the number of taps or coefficients. The minimum local memory requirement for a FIR filter of length N is \( 2N + 1 \) :

• – N coefficients
• – N input samples
• – 1 output sample

Since the local memory size is 256, the maximum value for N is 127.

If maximum throughput is required, it may be necessary to allocate a small amount of extra space, d1 and d2, to the input and output sample buffers respectively, to ensure that the filter never stalls waiting for a new input sample, or waiting for the output sample to be read. In this case, the local memory requirement is \( 2N + d1 + d2 \) .

The buffers must be configured as follows:

• • \( X1\_BUF\_SIZE = N + d1 \) ;
• • \( X2\_BUF\_SIZE = N \) ;
• • \( Y\_BUF\_SIZE = d2 \) (or 1 if no extra space is required)

The buffer base addresses can be allocated anywhere, but the X2 buffer must not overlap with the others, or else the coefficients are overwritten. An example configuration is:

• • \( X2\_BASE = 0 \) ;
• • \( X1\_BASE = N \) ;
• • \( Y\_BASE = 2N + d1 \)

However, if the memory space is limited, the X1 and Y buffer areas can be overlapped, such that each output sample takes the place of the oldest input sample, which is no longer required:

• • X2_BASE = 0;
• • X1_BASE = N;
• • Y_BASE = N

In this case, Y_BUF_SIZE = X1_BUF_SIZE = N + d1, so that the buffers remain in sync.

Note: The FULL_WM bitfield of X1 buffer configuration register must be programmed with a value less than or equal to log ₂ (d1), otherwise the buffer is flagged full before N input samples have been written, and no more samples are requested. Similarly, the EMPTY_WM bitfield of the Y buffer configuration register must be less than or equal to log ₂ (d2).

The filter coefficients must be pre-loaded into the X2 buffer, using the Load X2 Buffer function. The X1 buffer can optionally be pre-loaded with any number of samples up to a maximum of N. There is no point in pre-loading the Y buffer, since for the FIR filter there is no feedback path.

After configuring and initializing the buffers, the FMAC_CR register must be programmed according to the method used for writing and reading data to and from the FMAC memory.

Three methods are supported:

• • Polling: No DMA request or Interrupt request is generated. Software must check that the X1_FULL flag is low before writing to WDATA, or that the Y_EMPTY flag is low before reading from RDATA.
• • Interrupt: The interrupt request is asserted while the X1_FULL flag is low, for writes, or when the Y_EMPTY flag is low, for reads.
• • DMA: DMA requests are asserted on the DMA write channel while the X1_FULL flag is low, and on the read channel while the Y_EMPTY flag is low.

Different methods can be used for read and for write. However it is not recommended to use both interrupts and DMA requests for the same operation ^(a) . The valid combinations are listed in Table 172 .

Table 172. Valid combinations for read and write methods

a. If both interrupts and DMA requests are enabled then only DMA must perform the transfer.

The filter is started by writing to the FMAC_PARAM register with the following bitfield values:

• • FUNC = 8 (FIR filter);
• • P = N (number of coefficients);
• • Q = “Don’t care”;
• • R = Gain;
• • START = 1;

If less than \( N + d - 2^{\text{FULL\_WM}} \) values have been pre-loaded in the X1 buffer, the X1FULL flag remains low. If the WIEN bit is set in the FMAC_CR register, then the interrupt request is asserted immediately to request the processor to write \( 2^{\text{FULL\_WM}} \) additional samples into the buffer, via the FMAC_WDATA register. It remains asserted until the X1FULL flag goes high in the FMAC_SR register. The interrupt service routine must check the X1FULL flag after every \( 2^{\text{FULL\_WM}} \) writes to the FMAC_WDATA register, and repeat the transfer until the flag goes high. Similarly, if the DMAWEN bit is set in the FMAC_CR register, DMA write channel requests are generated until the X1FULL flag goes high.

The filter calculates the first output sample when at least N samples have been written into the X1 buffer (including any pre-loaded samples).

When \( 2^{\text{EMPTY\_WM}} \) output samples have been written into the Y buffer, the YEMPTY flag in the FMAC_SR register goes low. If the RIEN bit is set in the FMAC_CR register, the interrupt request is asserted to request the processor to read \( 2^{\text{EMPTY\_WM}} \) samples from the buffer, via the FMAC_RDATA register. It remains asserted until the YEMPTY flag goes high. The interrupt service routine must check the YEMPTY flag after every \( 2^{\text{EMPTY\_WM}} \) reads from the FMAC_RDATA register, and repeat the transfer until the flag goes high. If the DMAREN bit is set in the FMAC_CR, DMA read channel requests are generated until the YEMPTY flag goes high.

The filter continues to operate in this fashion until it is stopped by the software resetting the START bit.

21.3.9 Implementing IIR filters with the FMAC

The FMAC supports IIR filters of length N, where N is the number of feed-forward taps or coefficients. The number of feedback coefficients, M, can be any value from 1 to N-1. Only direct form 1 implementations can be realized, so filters designed for other forms need to be converted.

The minimum memory requirement for an IIR filter with N feed-forward coefficients and M feed-back coefficients is \( 2N + 2M \) :

• • N + M coefficients
• • N input samples
• • M output samples

If \( M = N-1 \) , then the maximum filter length that can be implemented is \( N = 64 \) .

As for the FIR, for maximum throughput, a small amount of additional space, d1 and d2, is allowed in the input and output buffer size respectively, making the total memory requirement \( 2M + 2N + d1 + d2 \) .

The buffers must be configured as follows:

• • X1_BUF_SIZE = N + d1;
• • X2_BUF_SIZE = N + M;
• • Y_BUF_SIZE = M + d2;

The buffer base addresses can be allocated anywhere, but must not overlap. An example configuration is given below:

• • X2_BASE = 0;
• • X1_BASE = N + M;
• • Y_BASE = 2N + M + d1;

Note: The FULL_WM bitfield of X1 buffer configuration register must be programmed with a value less than or equal to \( \log_2(d1) \) , otherwise the buffer is flagged full before N input samples have been written, and no more samples are requested. Similarly, the EMPTY_WM bitfield of the Y buffer configuration register must be less than or equal to \( \log_2(d2) \) .

The filter coefficients (N feed-forward followed by M feedback) must be pre-loaded into the X2 buffer, using the Load X2 Buffer function. The X1 buffer can optionally be pre-loaded with any number of samples up to a maximum of N. The Y buffer can optionally be pre-loaded with any number of values up to a maximum of M. This has the effect of initializing the feedback delay line.

After configuring the buffers, the FMAC_CR register must be programmed in the same way as for the FIR filter (see Section 21.3.8: Implementing FIR filters with the FMAC ).

The filter is started by writing to the FMAC_PARAM register with the following bitfield values:

• • FUNC = 9 (IIR filter);
• • P = N (number of feed-forward coefficients);
• • Q = M (number of feed-back coefficients);
• • R = Gain;
• • START = 1;

If less than \( N + d - 2^{\text{FULL\_WM}} \) values have been pre-loaded in the X1 buffer, the X1FULL flag remains low. If the WIEN bit is set in the FMAC_CR register, then the interrupt request is asserted immediately to request the processor to write \( 2^{\text{FULL\_WM}} \) additional samples into the buffer, via the FMAC_WDATA register. It remains asserted until the X1FULL flag goes high in the FMAC_SR register. The interrupt service routine must check the X1FULL flag after every \( 2^{\text{FULL\_WM}} \) writes to the FMAC_WDATA register, and repeat the transfer until the flag goes high. Similarly, if the DMAWEN bit is set in the FMAC_CR register, DMA write channel requests are generated until the X1FULL flag goes high.

The filter calculates the first output sample when at least N samples have been written into the X1 buffer (including any pre-loaded samples). The first sample is calculated using the first N samples in the X1 buffer, and the first M samples in the Y buffer (whether or not they are preloaded). The first output sample is written into the Y buffer at Y_BASE + M.

When \( 2^{\text{EMPTY\_WM}} \) new output samples have been written into the Y buffer, the YEMPTY flag in the FMAC_SR register goes low. If the RIEN bit is set in the FMAC_CR register, the interrupt request is asserted to request the processor to read \( 2^{\text{EMPTY\_WM}} \) samples from the buffer, via the FMAC_RDATA register. It remains asserted until the YEMPTY flag goes high. The interrupt service routine must check the YEMPTY flag after every \( 2^{\text{EMPTY\_WM}} \) reads from the FMAC_RDATA register, and repeat the transfer until the flag goes high. If the DMAREN bit is set in the FMAC_CR, DMA read channel requests are generated until the YEMPTY flag goes high

The filter continues to operate in this fashion until it is stopped by the software resetting the START bit.

21.3.10 Examples of filter initialization

Figure 107. X1 buffer initialization

MSv47128V1

The example in Figure 107 illustrates an X1 buffer pre-load with four samples (P = 4). The buffer size is six (X1_BUF_SIZE = 6). The initialization is launched by programming the FMAC_PARAM register with the START bit set. The four samples are then written to FMAC_WDATA, and transferred into local memory from X1_BASE onwards. The START bit resets after the fourth sample has been written. At this point, the X1 buffer contains the four samples, in order of writing, and the write pointer (next empty space) is at X1_BASE + 0x4.

21.3.11 Examples of filter operation

Figure 108. Filtering example 1

The example in Figure 108 illustrates the beginning of a filter operation. The filter has four taps (P=4). The X1 buffer size is six and the Y buffer size is two. The FULL_WM and EMPTY_WM bitfields are both set to 0. Prior to starting the filter, the X1 buffer has been pre-loaded with four samples, x[0:3] as in Figure 107. So the filter starts calculating the first output sample, y[0], immediately after the START bit is set. Since the X1FULL flag is not set (due to two uninitialized spaces in the X1 buffer), the interrupt is asserted straight away, to request new data. The processor writes two new samples, x[4] and x[5], to the FMAC_WDATA register, which are transferred to the empty locations in the X1 buffer.

In the mean time, the FMAC finishes calculating the first output sample, y[0], and writes it into the Y buffer, causing the Y_EMPTY flag to go low. At the same time, the x[0] sample is discarded, as it is no longer required, freeing up its location in memory (at X1_BASE). The FMAC can immediately start work on the second output sample, y[1], since all the required input samples x[1:5] are present in the X1 buffer.

Since the Y_EMPTY flag is low, the interrupt remains active after the processor finishes writing x[5]. The processor reads y[0] from the FMAC_RDATA register, freeing up its location in the Y buffer. There are now no samples in the output buffer since y[1] is still being calculated, so the Y_EMPTY flag goes high. Nevertheless, the interrupt remains active, because there is still free space in the X1 buffer, which the processor next fills with x[6], and so on.

Note: In this example, the processor can fill the input buffer more quickly than the FMAC can process them, so the X1_full flag regularly goes active. However, it struggles to read the Y buffer fast enough, so the FMAC stalls regularly waiting for space to be freed up in the Y buffer. This means the filter is not executing at maximum throughput. The reason is that the

filter length is small and the processor relatively slow, in this example. So increasing the Y buffer size does not help.

Figure 109. Filtering example 2

The example in Figure 109 illustrates the beginning of the same filter operation, but this time the filter has six taps ( \( P=6 \) ). The X1 buffer size is six and the Y buffer size is two. The FULL_WM and EMPTY_WM bitfields are both set to 0. Prior to starting the filter, the X1 buffer has been pre-loaded with four samples, \( x[0:3] \) as in Figure 107 . Because there are not enough samples in the input buffer, the X1FULL flag is not set, so the interrupt is asserted straight away, to request new data. The FMAC is stalled.

The processor writes two new samples, \( x[4] \) and \( x[5] \) , to the FMAC_WDATA register, which are transferred to the empty locations in the X1 buffer. As soon as there are six unused samples in the X1 buffer, the X1_FULL flag goes active (since the buffer size is six), causing the interrupt to go inactive. The FMAC starts calculating the first output sample, \( y[0] \) . Since this requires all six input samples, there are no free spaces in the X1 buffer and so the X1_FULL flag remains active. Only when the FMAC finishes calculating \( y[0] \) and writes it into the Y buffer, can \( x[0] \) be discarded, freeing up a space in the X1 buffer, and deasserting X1_FULL. At the same time, the Y_EMPTY flag goes inactive. Both these flag states cause the interrupt to be asserted, requesting the processor to write a new input sample, first of all, and then read the output sample just calculated. The FMAC remains stalled until a new input sample is written.

In this example, the processor has to wait for the FMAC to finish calculating the current output sample, before it can write a new input sample, and therefore the X1 buffer regularly goes empty, stalling the FMAC. This can be avoided by allowing some extra space in the input buffer.

21.3.12 Filter design tips

The FMAC architecture imposes some constraints detailed below, on the design of digital filters.

1. 1. Implementation of direct form 2, or transposed forms, is not efficient. Filters which have been designed for such forms must be converted to direct form 1.
2. 2. Cascaded filters must either be combined into a single stage, or implemented as separate filters. In the latter case, multiple sets of filter coefficients can be pre-loaded into the memory, one set per stage, and only the X2_BASE address changed to select which set is used. The most efficient method of implementing a multi-stage filter is to pre-load a large X1 buffer with input samples, run the IIR filter function on it using the first stage coefficients, and store the output samples back in memory. Then change the X2_BASE pointer to point to the 2nd stage coefficients, and reload the input buffer with the output of the first stage (with a gain if required), before running the IIR function again. The procedure is repeated for all stages. Once the final stage samples have been transferred back into system memory, the input buffer can be loaded with the next set of input samples, and a new round of calculations started. Note that the N sample input buffer of each stage must be pre-loaded first of all with the N-1 last inputs from the previous round, plus one new sample, in order to keep continuity between each round. Similarly, the output buffer of each stage must be loaded with the last M samples from the previous round, for the same reason.
3. 3. The use of direct form 1 for IIR designs can lead to large positive or negative partial sums in the accumulator, if for example a large step occurs on the input, or some of the filter coefficients' absolute values are \( >1 \) . Since the accumulator is limited to 26 bits, the biggest value that it can handle without wrapping (changing sign) is \( 0x1FFFFFF \) positive or \( 0x2000000 \) negative. This corresponds to 3.99999988 and -4 respectively in q3.23 fixed point format. Wrapping does not represent a problem provided the wrapping is “undone” before the end of the accumulation. However this is not always the case when a filter is starting up and can lead to unexpected results. Consider pre-loading the output buffer with suitable values to avoid this.
4. 4. The IIR filter has feed-forward (numerator) coefficients \( [b_0, b_1, \dots, b_{N-1}] \) , and feed-back (denominator) coefficients \( [1, a_1, \dots, a_M] \) . Many IIR filters require some of the denominator coefficients to have an absolute value greater than 1 to achieve a steep roll-off in the frequency response. Given that the coefficients are coded in fixed point q1.15 format, this is not possible. Nevertheless, by scaling the denominator coefficients by a factor \( 2^{-R} \) , such that \( 2^{-R} \cdot [1, a_1, \dots, a_M] \) are all less than 1, such filters can be implemented. However an inverse gain of \( 2^R \) must be applied at the output of the accumulator to compensate the scaling. This has an adverse effect on the signal-to-noise ratio.

21.4 FMAC registers

21.4.1 FMAC X1 buffer configuration register (FMAC_X1BUFCFG)

Address offset: 0x00

Reset value: 0x0000 0000

Access: word access

This register can only be modified if START = 0 in the FMAC_PARAM register.

Bits 31:26 Reserved, must be kept at reset value.

Bits 25:24 FULL_WM[1:0] : Watermark for buffer full flag

Defines the threshold for setting the X1 buffer full flag when operating in circular mode. The flag is set if the number of free spaces in the buffer is less than \( 2^{\text{FULL\_WM}} \) .

0: Threshold = 1

1: Threshold = 2

2: Threshold = 4

3: Threshold = 8

Setting a threshold greater than 1 allows several data to be transferred into the buffer under one interrupt.

Threshold must be set to 1 if DMA write requests are enabled (DMAWEN = 1 in FMAC_CR register).

Bits 23:16 Reserved, must be kept at reset value.

Bits 15:8 X1_BUF_SIZE[7:0] : Allocated size of X1 buffer in 16-bit words

The minimum buffer size is the number of feed-forward taps in the filter (+ the watermark threshold - 1).

Bits 7:0 X1_BASE[7:0] : Base address of X1 buffer

21.4.2 FMAC X2 buffer configuration register (FMAC_X2BUFCFG)

Address offset: 0x04

Reset value: 0x0000 0000

Access: word access

Bits 31:16 Reserved, must be kept at reset value.

Bits 15:8 X2_BUF_SIZE[7:0] : Size of X2 buffer in 16-bit words

This bitfield can not be modified when a function is ongoing (START = 1).

Bits 7:0 X2_BASE[7:0] : Base address of X2 buffer

The X2 buffer base address can be modified while START=1, for example to change coefficient values. The filter must be stalled when doing this, since changing the coefficients while a calculation is ongoing affects the result.

21.4.3 FMAC Y buffer configuration register (FMAC_YBUFCFG)

Address offset: 0x08

Reset value: 0x0000 0000

Access: word access

This register can only be modified if START = 0 in the FMAC_PARAM register.

Bits 31:26 Reserved, must be kept at reset value.

Bits 25:24 EMPTY_WM[1:0] : Watermark for buffer empty flag

Defines the threshold for setting the Y buffer empty flag when operating in circular mode. The flag is set if the number of unread values in the buffer is less than \( 2^{\text{EMPTY\_WM}} \) .

0: Threshold = 1

1: Threshold = 2

2: Threshold = 4

3: Threshold = 8

Setting a threshold greater than 1 allows several data to be transferred from the buffer under one interrupt.

Threshold must be set to 1 if DMA read requests are enabled (DMAREN = 1 in FMAC_CR register).

Bits 23:16 Reserved, must be kept at reset value.

Bits 15:8 Y_BUF_SIZE[7:0] : Size of Y buffer in 16-bit words

For FIR filters, the minimum buffer size is 1 (+ the watermark threshold). For IIR filters the minimum buffer size is the number of feedback taps (+ the watermark threshold).

Bits 7:0 Y_BASE[7:0] : Base address of Y buffer

21.4.4 FMAC parameter register (FMAC_PARAM)

Address offset: 0x0C

Reset value: 0x0000 0000

Access: word access

Bit 31 START : Enable execution

0: Stop execution

1: Start execution

Setting this bit triggers the execution of the function selected in the FUNC bitfield. Resetting it by software stops any ongoing function. For initialization functions, this bit is reset by hardware.

Bits 30:24 FUNC[6:0] : Function

0: Reserved

1: Load X1 buffer

2: Load X2 buffer

3: Load Y buffer

4 to 7: Reserved

8: Convolution (FIR filter)

9: IIR filter (direct form 1)

10 to 127: Reserved

This bitfield can not be modified when a function is ongoing (START = 1)

Bits 23:16 R[7:0] : Input parameter R.

The value of this parameter is dependent on the function.

This bitfield can not be modified when a function is ongoing (START = 1)

Bits 15:8 Q[7:0] : Input parameter Q.

The value of this parameter is dependent on the function.

This bitfield can not be modified when a function is ongoing (START = 1)

Bits 7:0 P[7:0] : Input parameter P.

The value of this parameter is dependent on the function

This bitfield can not be modified when a function is ongoing (START = 1)

21.4.5 FMAC control register (FMAC_CR)

Address offset: 0x10

Reset value: 0x0000 0000

Access: word access

Bits 31:17 Reserved, must be kept at reset value.

Bit 16 RESET : Reset FMAC unit

This resets the write and read pointers, the internal control logic, the FMAC_SR register and the FMAC_PARAM register, including the START bit if active. Other register settings are not affected. This bit is reset by hardware.

0: Reset inactive

1: Reset active

Bit 15 CLIPEN : Enable clipping

0: Clipping disabled. Values at the output of the accumulator which exceed the q1.15 range, wrap.

1: Clipping enabled. Values at the output of the accumulator which exceed the q1.15 range are saturated to the maximum positive or negative value (+1 or -1) according to the sign.

Bits 14:10 Reserved, must be kept at reset value.

Bit 9 DMAWEN : Enable DMA write channel requests

0: Disable. No DMA requests are generated

1: Enable. DMA requests are generated while the X1 buffer is not full.

This bit can only be modified when START= 0 in the FMAC_PARAM register. A read returns the current state of the bit.

Bit 8 DMAREN : Enable DMA read channel requests

0: Disable. No DMA requests are generated

1: Enable. DMA requests are generated while the Y buffer is not empty.

This bit can only be modified when START= 0 in the FMAC_PARAM register. A read returns the current state of the bit.

Bits 7:5 Reserved, must be kept at reset value.

Bit 4 SATIEN : Enable saturation error interrupts

0: Disabled. No interrupts are generated upon saturation detection.

1: Enabled. An interrupt request is generated if the SAT flag is set

This bit is set and cleared by software. A read returns the current state of the bit.

Bit 3 UNFLIEN : Enable underflow error interrupts

0: Disabled. No interrupts are generated upon underflow detection.

1: Enabled. An interrupt request is generated if the UNFL flag is set

This bit is set and cleared by software. A read returns the current state of the bit.

Bit 2 OVFLIEN : Enable overflow error interrupts

0: Disabled. No interrupts are generated upon overflow detection.

1: Enabled. An interrupt request is generated if the OVFL flag is set

This bit is set and cleared by software. A read returns the current state of the bit.

Bit 1 WIEN : Enable write interrupt

0: Disabled. No write interrupt requests are generated.

1: Enabled. An interrupt request is generated while the X1 buffer FULL flag is not set.

This bit is set and cleared by software. A read returns the current state of the bit.

Bit 0 RIEN : Enable read interrupt

0: Disabled. No read interrupt requests are generated.

1: Enabled. An interrupt request is generated while the Y buffer EMPTY flag is not set.

This bit is set and cleared by software. A read returns the current state of the bit.

21.4.6 FMAC status register (FMAC_SR)

Address offset: 0x14

Reset value: 0x0000 0001

Access: word access

Bits 31:11 Reserved, must be kept at reset value.

Bit 10 SAT : Saturation error flag

Saturation occurs when the result of an accumulation exceeds the numeric range of the accumulator.

0: No saturation detected

1: Saturation detected. If the SATIEN bit is set, an interrupt is generated.

This flag is cleared by a reset of the unit.

Bit 9 UNFL : Underflow error flag

An underflow occurs when a read is made from FMAC_RDATA when no valid data is available in the Y buffer.

0: No underflow detected

1: Underflow detected. If the UNFLIEN bit is set, an interrupt is generated.

This flag is cleared by a reset of the unit.

Bit 8 OVFL : Overflow error flag

An overflow occurs when a write is made to FMAC_WDATA when no free space is available in the X1 buffer.

0: No overflow detected

1: Overflow detected. If the OVFLIEN bit is set, an interrupt is generated.

This flag is cleared by a reset of the unit.

Bits 7:2 Reserved, must be kept at reset value.

The buffer is flagged as full if the number of available spaces is less than the FULL_WM threshold. The number of available spaces is the difference between the write pointer and the least recent sample currently in use.

0: X1 buffer not full. If the WIEN bit is set, the interrupt request is asserted until the flag is set. If DMAWEN is set, DMA write channel requests are generated until the flag is set.

1: X1 buffer full.

This flag is set and cleared by hardware, or by a reset.

Note: after the last available space in the X1 buffer is filled there is a delay of 3 clock cycles before the X1FULL flag goes high. To avoid any risk of overflow it is recommended to insert a software delay after writing to the X1 buffer before reading the FMAC_SR. Alternatively, a FULL_WM threshold of 2 can be used.

The buffer is flagged as empty if the number of unread data is less than the EMPTY_WM threshold. The number of unread data is the difference between the read pointer and the current output destination address.

0: Y buffer not empty. If the RIEN bit is set, the interrupt request is asserted until the flag is set. If DMAREN is set, DMA read channel requests are generated until the flag is set.

1: Y buffer empty.

This flag is set and cleared by hardware, or by a reset.

Note: after the last sample is read from the Y buffer there is a delay of 3 clock cycles before the YEMPTY flag goes high. To avoid any risk of underflow it is recommended to insert a software delay after reading from the Y buffer before reading the FMAC_SR. Alternatively, an EMPTY_WM threshold of 2 can be used.

21.4.7 FMAC write data register (FMAC_WDATA)

Address offset: 0x18

Reset value: 0x0000 0000

Access: word access

Bits 31:16 Reserved, must be kept at reset value.

Bits 15:0 WDATA[15:0] : Write data

When a write access to this register occurs, the write data are transferred to the address offset indicated by the write pointer. The pointer address is automatically incremented after each write access.

21.4.8 FMAC read data register (FMAC_RDATA)

Address offset: 0x1C

Reset value: 0x0000 0000

Access: word access

Bits 31:16 Reserved, must be kept at reset value.

Bits 15:0 RDATA[15:0] : Read data
When a read access to this register occurs, the read data are the contents of the Y output buffer at the address offset indicated by the READ pointer. The pointer address is automatically incremented after each read access.

21.4.9 FMAC register map

Table 173. FMAC register map and reset values

Refer to Section 2.3: Memory organization for the register boundary addresses.