Convolution algorithms on DSP processors

For state-of-the-art digital signal processors (DSPs) alternative ways to implement FIR (finite impulse response) filters are compared with respect to the necessary number of instruction cycles per tap. Also, the accuracy attainable both with fixed-point and floating-point arithmetic is discussed. Besides direct form filtering and fast convolution algorithms using FFTs (fast Fourier transforms), radix FIR filter structures are also considered. The latter are very interesting for most DSP architectures, offering speed improvements of up to 25% compared to the direct form