Analog Devices ADSP-SC58 Series Hardware Reference Manual page 3346

Figure 51-5: Single Channel Filtering Flow
Floating-Point Data Format
The FIR accelerator treats data and coefficients in 32-bit floating-point format as the default functional mode.
Fixed-Point Data Format
In fixed-point mode, the 32-bit input data or coefficient is treated as fixed point. A 32-bit fixed-point MAC opera-
tion generates an 80-bit result. Fixed-point data or coefficients can be unsigned integer, unsigned fractional and
signed integer.
NOTE: In fixed-point mode, the entire 80-bit result register is always written back in bursts of 3 × 32 bits. The
first word is the LSW, the second word is the MSW, and the third word is a 16-bit overflow. The remain-
ing 16 bits are padded with zeros. Therefore, for fixed-point mode: WINDOWSIZE = WINDOWSIZE ×
3.
If the signed fractional format is used, the output must be scaled by 2. The MAC does not right shift to
remove the redundant sign bit. A final routine must decimate the output buffer to the desired samples.
Multi-iteration mode is not supported in this format. Therefore, the maximum TAP length is 1024.
ADSP-SC58x/ADSP-2158x SHARC+ Processor Hardware Reference
LOAD I, M, B, L REGISTERS
FOR COEFFICIENT AND DATA I/0
LOAD COEFFICIENTS INTO
COEFFICIENT MEMORY
PRELOAD DELAY LINE
FROM INTERNAL MEMORY
COMPUTE RESULT AND
STORE IN INTERNAL MEMORY
NO
WINDOW
OVER?
YES
NO LAST
ITERATION?
YES
UPDATE INDEX REGISTER VALUES
OF TCB IN INTERNAL MEMORY
TO NEXT CHANNEL PROCESSING
FETCH NEXT DATA
Operating Modes
51–11