Storing filter coefficients
Rather than defining the filter coefficients in your main
assembly program file, it is usually more convenient to
store them in a separate file. By defining the
coefficients in a separate assembly (for example,
coeff.asm) file, you can load the
coefficients at a desired memory location at the run time,
although it is not essential for the current simple FIR
filtering lab.
The assembly file containing the filter coefficients can be written as follows:
.def _coef
.sect "coeffs"
_coef:
.short 0ff9bh
.short 0ff06h
.short 0feffh
.short 0ff93h
.short 070h
.short 0117h
.short 0120h
.short 07bh
Each coefficient must be converted to the Q-15 format and
defined by each .short assembly
directive. For your convenience, I wrote a short MATLAB
script save_coef.m that converts the filter
coefficients stored as a MATLAB vector to Q-15 format and
then writes to a file exactly in the above format. (You
can download save_coef.m from the course web
page.) The section coeffs should be defined
in the link command file so that the coefficients are to
be loaded at the correct memory location.
You can simply include the coeff.asm using
the .include directive at the beginning of
your main assembly program.
Exercise 1
Make coefficient files for each of the filters you designed in the previous exercise.
Solution
Assembly implementation
Based on the codec input and output program you have written in the previous labs, you can now implement a real-time FIR filtering algorithm.
Exercise 2
Write an assembly routine that implements the FIR
filter by modifying the inner product program you have
written in Lab 3. Combine the FIR filtering routine
with the interrupt-based codec input-output code you
wrote in the previous lab. Your code should perform
FIR filtering on the input samples and output the
filtered result to the codec. Both the left and right
channels should be filtered. To write the designed
MATLAB vector of filter coefficients as a
.asm file, use the provided
save_coef.m matlab function. First
implement the length-40 lowpass filter with 10kHz
cutoff designed using the remez.m.
Solution
Implementation using circular addressing modes
As you might already have noticed, a lot of cycles are wasted in FIR filtering while maintaining the buffer to see if you reached the end of buffer and update the address pointers properly. To avoid this unnecessary buffer maintenance, the TI DSP processors have a special addressing mode, called circular addressing. Using circular addressing, you can define a block of memory as a circular buffer. As you increase (or decrease) the pointer register pointing to the buffer index beyond the buffer limit, it automatically points to the other end of the buffer, implementing a circle of data array. Instead of moving the data samples themselves, you can move the pointer which specifies the beginning of the buffer, as each new sample is processed. You don't need to check if you reached the end of buffer because the address pointer returns to the beginning of the buffer immediately after reaching the end.
Of the 32 registers on the C6x, 8 of them can perform
circular addressing. These registers are A4 through A7
and B4 through B7. Since circular addressing is not
default, each of these registers must be specified as
circular using the AMR (Address Mode
Register) register. The lower 16 bits of the
AMR are used to select the mode for each of
the 8 registers. The upper 10 bits (6 are reserved) are
used to set the length of the circular buffer. Buffer size
is determined by
AMR register. The top address of the buffer
needs to be aligned with proper physical memory block
address using the .align assembler directive.
Exercise 3
First read TMS320C62x/C67x CPU and Instruction Set Reference Guide to learn how to define circular buffers. Modify your FIR filtering assembly code to use circular addressing modes. After optimizing your code as much as you can, count the number of required clock cycles for each FIR filter output computation. Compare the number with the code written without circular addressing.
