You are on page 1of 41

Digital Signal Processing

Discrete Fourier Transform


Linear Convolution: Direct computation

Direct computation for “1” output sample: M complex multiplications


& (M-1) complex additions
Linear Convolution: Direct computation

Direct computation for M output samples: M^2 complex multiplications

& M(M-1) complex additions

total (approximate) ≈ 2*(M^2) computations (for large M)

 Convolution processing is notoriously computationally intensive.

 For typical audio signals, a 3 second impulse response sampled at 44.1 kHz requires
around 35 billion computations to convolve with the same length input.
Linear Convolution: Using DFT

 This method is much more efficient than direct time-domain convolution.

 The number of operations ≈ 2(L+M-1)(L+M-1) for the DFTs, (L+M-1) operations for the
complex multiplication, (L+M-1)(L+M-1) for the IDFT.

 For a 3 second input this results in about 17 million calculations ≈ (1/1000)th


computations required for direct convolution.
Linear Convolution ≡ Circular convolution ??
To summarize, we conclude that

You might also like