OVERVIEW
CONVOLUTION
FAST CONVOLUTION
COOK-TOOM ALGORITHM
EXAMPLE
ALGORITHM
MODIFIED COOK-TOOM ALGORITHM
IMPLEMENTATION OF FILTERBANK
ADVANTAGES
REFERENCES
CONVOLUTION
Mathematical operation on two functions to produce a third
variable, typically viewed as a modified version of the
original function.
The convolution theorem allows one to mathematically
convolve in the time domain by simply multiplying in the
frequency domain.
Two types: linear and cyclic
Linear-(M-1)(N-1) additions and MN multiplications
Cyclic-N(N-1) additions and N*N multiplications
For speeding up the calculation-FFT
Commonly used for fast computation of convolutions
Disadvantage-complex arithmetic
Another method is to convert 1D convolution into
multidimensional convolution
By using efficient short length algorithms
FAST CONVOLUTION
Convolution using fewer number of operations
Algorithmic strength reduction: Number of strong
operations is reduced at the expense of an increase in
the number of weak operations.
Best suited for implementation using programmable
or dedicated hardware.
Assume (a+jb)(c+jd) = e+jf ,
where (a+jb) is the signal sample
(c+jd) is a coefficient
implemented using 4 multiplications and 2 additions
Using fast convolution algorithm arithmetic complexity
is reduced to 3 multiplications and 3 additions
COOK-TOOM ALGORITHM
Minimum number of multiplications
For non cyclic convolution
( 1)
This requires 2N-1 multiplications
Define the generating polynomial of a sequence xi, by
(2)
Then W(z)=X(z)H(z), (3)
where H(z) and W(z) are generating polynomials of hi
and wi.
W(z) is a 2N-2 degree polynomial
To determine the 2N- 1 wi's, select 2N- 1 distinct
numbers αj, j = 0, 1, . . , 2N - 2, and substitute them
for z in (3) to obtain the 2N - 1 products
mj=W(α j)=H(α j)X(αj), j=0,1,…..2N-2 (4)
Using Lagrange interpolation formula
(5)
Cost is 2N-1 multiplications
(4) can be written in matrix form as, m = ( Ah)x(Ax)
where
From (5) the coefficients of W(z) will be the linear
combinations of the mj’s and can be written as
w=C*m
where C* is a 2N-1 by 2N-1 matrix
To calculate cyclic convolution, compute
Y(z)=W(z)mod(zN-1)
which leads to y=Cm, where C is an N by 2N-1 matrix
obtained from C* by performing row operations.
General form is m=(Ah)x(Bx)
Value of (Ah) is precomputed.
B and C will have no multiplications
Only multiplications are the element by element
multiplication of Ah by Bx.
Cook-Toom algorithm yields large integer coefficients
in A,B,C matrices which is costly as multiplication.
Example
(Qn). Calculate the non cyclic 2 point convolution
Using (1)
wo=hoxo
w1=h0x1+h1x0
w2=h1x1
In terms of z transforms, this is equivalent to
w0+w1z+w2z2=hoxo +(h0x1+h1x0)z+h1x1z2
= ho(x0+x1z)+h1z(x0+x1z)
= (h0+h1z)(x0+x1z)
Let αj=-1,0,1 for j=0,1,2 in (4)
m0=(h0-h1)(x0-x1)
m1=h0x0
m2=(ho+h1)(x0+x1)
[put z=-1,0,1 in the above eqn. to obtain m0,m1,m2)
From (5)
So that
w0=m1
w1=(m2-m0)/2
w2=[(m0+m2)/2]-m1
Transferring denominators from C to A matrix,
combine the factor ½ with hj’s and store the
precomputed constants
a0=(h0-h1)/2
a1=h0
a2=(h0+h1)/2
Hence
m0=a0(x0-x1)
m1=a1x0
m2=a2(x0-x1)
w0 = m 1
w1=m2-m0
w2=m0+m2-m1
Only 3 multiplications and 5 additions are required
instead of 4 multiplications and 3 addition
Algorithm
1. Choose L+N-1 different real numbers β0, β1,…. βL+N-2
2. Compute h(βi) and x(βi), for i=0,1,….L+N-2
3. Compute s(βi)=h(βi)x(βi), for i=0,1,….L+N-2
4. Compute s(p) using the equation
Reduction of operation count occurs if the numbers
β0, β1,…. Βn are carefully chosen
A better algorithm using Chinese remainder theorem
to compute the convolution as:
S(x)=[D(x)G(x)mod
Modified Cook-Toom Algorithm
Choose L+N-2 different real numbers β1,… β0, βL+N-2
Compute h(βi) and x(βi),for i=0,1,…L+N-3
Compute s(βi)=h(βi)x(βi), for i=0,1,…L+N-3
Compute s’(βi)= s(βi)-sL+N-2 βiL+N-2, for i=0,1,…L+N-3
Compute s’(p) using the equation
Compute s(p)= s’(p)+sL+N-2pL+N-2
IMPLEMENTATION OF FILTER BANK
Consider two polynomials, g(z-1 )=g0+g1z-1+….+gL-1z-L+1 and
u(z-1)=u0+u1z-1+…uN-1z-N+1
By computing in normal form it’s product y(z-1) requires
NL multiplications
Cook-toom algorithm will reduce it to M>=N+L-1
First, choose a set of interpolation points{ρi}i=0:M-1 that
are the roots of r(z-1) =
Evaluate y(ρi)=g(ρi)u(ρi)
Perform Lagrange interpolation to restore
y(z-1)= with
Li(z-1)=
Eg: Consider g(z-1)=g0+g1z-1(L=2), and u(z-1)=u0+u1z-1 (N=2)
Choose interpolation points{0,1,-1}
Then y(0)=u0g0
y(1)=(u0+u1)(g0+g1)
y(-1)=(u0-u1)(g0-g1)
Lagrange polynomials are calculated as
L0(z-1)=(1-z-2)
L1(z-1)=(z-1+z-2)/2
L-1(z-1)=(-z-1+z-2)/2
ie, y(z-1)=y(0)L0(z-1)+y(1)L1(z-1)+y(-1)L-1(z-1) is
reconstructed
Multiplication can be written as y=Gu, then algorithm can be
represented as the matrix decomposition G=CDA, with
D=diag(Bg)
G is the (N+L-1) x N Toeplitz matrix defining the filter g(z-1)
A is the Vandermonde matrix with Am,n= (n=0:N-1)
B is also the MxL Vandermonde matrix with Bm,l= (l=0:L-1)
C is the (N+L-1) x M matrix whose ith column contains the first
N+L-1 coefficients of Li(z-1)
ADVANTAGES
The number of multiplications have been reduced to
L+N-1 at the expense of an increase in the number of
additions
Adder has much smaller area and computation time
than multiplier
Low hardware complexity
Pre addition and post addition matrices are not simple
REFERENCES
[1]. Zdenka Babic,Danilo P.Mandic, “A Fast Algorithm for
Linear Convolution of Discrete Time signals” , TELSIKS,
pp.no-595-598,September 2001.
[2]. Yuke Wang,Keshab Parhi, “Explicit Cook-Toom Algorithm
for linear convolution” ,IEEE,2000.
[3]. Geert Van Meerbergen, Marc Moonen, Hugo De Man,
“Critically Subsampled filterbanks implementing Reed-
Solomon codes” ,vol 2, pp.no-989-992,IEEE,2004
[4].Keshab K. Parhi, “VLSI digital signal Processing Systems,
Design and Implementation”, pp.no:227-237,New
Delhi,1999.
[5].Ivan W.Selesnick,C. Sidney Burrus, “Fast Convolution and
Filtering” , Digital Signal Processing Handbook, CRC Press
LLC, 1999.
[6]. R. Meyer, R. Reng and K. Schwarz, Convolution
Algorithms On DSP Processors, IEEE,1991.
[7].R.E.Blahut, “Fast Convolution Algorithms for Digital
Signal Processing” , Addison-Wesley,1985.
[8]. H.J.Nussbaumer, “Fast Fourier Transform and
Convolution Algorithms” , Springer-Verlag, Berlin,
Heidelberg, and New York,1981
[9]. Ramesh C Agarwal,James W. Cooley, ”New Algorithms for
Digital Convolution”, IEEE Transactions on Accoustics,
Speech and Signal Processing, Vol. ASSP-25, No.5,October
1997.
[10]. Alberto Zanoni, Toom-”Cook 8-way For Long Integers
Multiplication”, 11th International Symposium on Symbolic
and Numeric Algorithms for Scientific Computing,2009
