An MDCT Hardware Accelerator for MP3 Audio

Xingdong Dai Meghanad D. Wagh

LSI Corporation Dept. of ECE
1110 American Parkway NE Lehigh University
Allentown, PA 18109 Bethlehem, PA 18015
xingdong.dai@lsi.com mdw0@lehigh.edu

Abstract— With the increasing popularity of MP3 audio, there subbands to be processed by the MDCT/IMDCT, resulting a
is a need to develop cost and power efficient architectures for long block of 36 samples and a short block of 12 samples. The
the MP3 encoder and decoder. This paper describes dedicated switch from a long block to a short block is called window
architectures for computing the modified discrete cosine trans-
form (MDCT) and its inverse (IMDCT). Recent profiling studies switching and it is used to supress distortions frequently
have shown that these operations represent about 30% of the associated with frequency doamin coding of an audio signal.
total MP3 computations. MP3 format defines two frame sizes The ability to process both long and short blocks is crucial to
that can occur in the same data stream. We have developed the efficient implementations of MP3 audio encoder and decoder.
most efficient algorithms for MDCT and IMDCT suitable for Because the MP3 block lengths are not power of two, only
both sizes. Unlike previous algorithms, our computations can be
unified in a single ASIC architecture. This unified architecture a few fast algorithms are published and the focus has been
implemented in 90 nm TSMC library is still 25% smaller and on software implementations [5]–[7]. However, over the years
25% faster than any previous single frame size architectures the cost of application specific architectures has decreased
designed in the same technology. In addition, at 128 Kbits/sec substantially. In addition, such dedicated architectures can be
data rates, our algorithms save nearly 1800 multiplications per embedded in field programmable gate arrays making them just
second (18%) which can help reduce the power consumption.
as flexible as software routines. Such an approach permits one
I. I NTRODUCTION to utilize a low-cost and low-power processor to work on data
The MPEG-1/2 layer-III (MP3) standard is widely employed management and other computationally simple tasks while
in music industry because of its efficient audio compression. A the dedicated hardware can take care of the computationally
key enabler in MP3 coding is the prefect reconstruction (PR) challenging tasks.
cosine modulated filter bank based on the concept of time This paper is devoted to the development of the MDCT/
domain aliasing cancellation (TDAC) [1]. For the encoder, IMDCT hardware accelerator for MP3 audio applications.
this analysis filter bank is realized by applying the modified Based on a group theoretic partitioning of the transform kernel,
discrete cosine transform (MDCT) to sliding blocks of data, our solutions employ carefully crafted bilinear algorithms
while the synthesis filter bank of the decoder uses the inverse which can be directly mapped to hardware architectures. We
modified discrete cosine transform (IMDCT). Since the MDCT show for the first time that both the long and short MP3 audio
and IMDCT require intensive computations, fast algorithms blocks can be seamlessly processed with a single hardware
and efficient implementations for theses transforms is im- architecture. Our MDCT/IMDCT algorithms require only 9
portant to the realization of high quality audio compression, multiplications for the short block and 36 multiplications for
especially when most MP3 audio players are battery operated the long block. These may be compared with the current
[2]–[4]. best algorithms for the same tasks which use at least 11
The N point MDCT of a sequence {x(i)} is defined as and 43 multiplications respectively [5]–[7]. However, the main
advantage of the proposed algorithms is its bilinearity which
N
 −1
π(2i + 1 + N2 )(2k + 1) implies that all the multiplications are independent and can be
X(k) = x(i) cos( ),
i=0
2N carried out concurrently. Thus when implemented in hardware,
our algorithms have only one multiplication on the critical path
0 ≤ k < N/2.
for both the long and short block MDCT/IMDCT. Previous
The inverse MDCT (IMDCT) is defined as reported algorithms have at least 2 multiplications along the
N critical path [5]–[7]. As a result, our VLSI implementations re-
2 −1
2  π(2i + 1 + N2 )(2k + 1) duce the critial path delay by about 25% while simultaneously
x(i) = X(k) cos( ),
N 2N saving about 25% of the chip area.
k=0
0 ≤ i < N. II. A LGORITHM AND ARCHITECTURE
Note that MDCT converts N signal samples into only N/2 A bilinear algorithm is made up of an addition stage
transform samples. MP3 standard defines two data frames of followed by a stage of independent multiplications and a final
1152 and 384 samples. These frames are further divided in 32 addition stage. Our procedure consists of (a) converting the

121
1-4244-2334-7/08/$20.00 
c 2008 IEEE
MDCT/IMDCT computation to a DCT of type IV (using are in the same set. Further, the summation in (1) is carried out
additions and subtractions only), (b) decomposing the DCT over the two signal index sets separately. We will denote the
kernel into cyclic convolutions and Hankel matrix products, computation of X(k) with signal component indices restricted
and finally, (c) employing bilinear algorithms for each of the to sets S1 and S2 by X1 (k) and X2 (k) respectively. Clearly,
convolutions and Hankel matrix products to obtain the required X(k) = X1 (k) + X2 (k).
bilinear algorithm for the MDCT. We illustrate each of these Let p denote the DCT kernel element cos(πp/4N ) and
steps in this section. p, the element − cos(πp/4N ). Then from (1), the transform
components X1 (k), k ∈ S1 are given by:
A. Conversion of MDCT into DCT
    
To obtain MDCT using a DCT, introduce a new data X1 (1) 9 3 x(1)
= . (2)
sequence X1 (4) 3 9 x(4)

−x(i + 3N/4), if 0 ≤ i < N/4,
y(i) = Since the constant matrix in (2) is a Hankel matrix, one can
x(i − N/4), if N/4 ≤ i < N.
use a bilinear algorithm to obtain this product.
Then defining Similarly, using the same shorthand notation, transform
components X2 (k), k ∈ S1 are given by:
z(i) = y(i) − y(N − 1 − i), 0 ≤ i < N/2,
    
an N point MDCT can be expressed as an N/2 point DCT-IV X2 (1) 3 9 x(0) − x(3)
= . (3)
as X2 (4) 9 3 −x(5) − x(2)
N/2−1
  
π(2i + 1)(2k + 1) Note that (3) exploits the fact that some kernel matrix elements
X(k) = z(i) cos ,
i=0
2N are equal in magnitude and therefore the signal sequence may
be folded to reduce the number of multiplications. Since the
0 ≤ k < N/2.
constant matrix in (3) is a Hankel matrix, one can once again
Thus the MDCT computation of short and long blocks is use a bilinear algorithm to obtain this product.
transformed into DCTs of 6 and 18 points respectively. When both the signal and transform indices are restricted to
set S2 , the computation can be transformed into a multidimen-
B. Conversion of IMDCT into DCT
sional convolution. To do this, observe that indices in S2 are
To obtain the IMDCT, we first compute the N/2 point type- elements of A(8N ), a group formed by non-negative integers
IV DCT of X as less than and relatively prime to 8N under the operation of
N/2−1   multiplication modulo 8N . The structure of this group can
2  π(2i + 1)(2k + 1)
z(i) = X(k) cos , be used to order the indices in S2 as well as to define a
N i=0 2N sign function. By using the order {0, 5, 3, 2} suggested by
0 ≤ i < N/2. A(8N ) and the corresponding sign function {1, −1, 1, 1}, one
can express X2 (k), k ∈ S2 as:
Then defining a new data sequence
 ⎡ ⎤ ⎡ ⎤⎡ ⎤
z(i), if 0 ≤ i ≤ N/2 − 1, X2 (0) 1 11 7 5 x(0)
y(i) = ⎢ −X2 (5) ⎥ ⎢ 11 1 5 ⎥
7 ⎥ ⎢ −x(5) ⎥
⎢
−z(N − 1 − i), if N/2 ≤ i < N, ⎢ ⎥ ⎢ ⎥ . (4)
⎣ X2 (3) ⎦ = ⎣ 7 5 1 11 ⎦ ⎣ x(3) ⎦
One can recover the signal sequence as
 X2 (2) 5 7 11 1 x(2)
y(i + N/4), if 0 ≤ i < 3N/4,
x(i) =
−y(i − 3N/4), if 3N/4 ≤ i < N. The structure of the matrix in (4) is predicted by the structure
of the group A(8N ). By partitioning the matrix in (4) as
Thus the IMDCT computation of short and long blocks is also shown, one can see that it is a block cyclic matrix with
transformed into DCTs of 6 and 18 points respectively. each block being a Hankel matrix. Thus this computation
C. DCT algorithm for MP3 short block corresponds to a two dimensional convolution with a 2 point
cyclic convolution along one dimension and a 2 point Hankel
Recall that a DCT-IV of an N point sequence {x(i)} is
product along the other. Again, appropriate bilinear algorithms
given by
for these small lengths can be combined to obtain the bilinear
N
 −1
π(2i + 1)(2k + 1) algorithm for (4).
X(k) = x(i) cos( ), 0 ≤ k < N. (1) Finally, when the signal indices are restricted to the set S1 ,
i=0
4N
one the computation of X2 (k), k ∈ S2 gives:
For MP3 short block, the DCT length N = 6. To compute (1) ⎡ ⎤ ⎡ ⎤
for this N ,we partition the signal and transform indices in two X1 (0) 3 9  
sets: set S1 = {1, 4} is made up of those i’s for which (2i+1) ⎢ X1 (5) ⎥ ⎢ 9 3 ⎥ x(1)
⎢ ⎥=⎢ ⎥
is a multiple of 3 and set S2 = {0, 2, 3, 5}, of the rest. We ⎣ X1 (3) ⎦ ⎣ 3 9 ⎦ x(4) . (5)
compute those transform components together whose indices X1 (2) 9 3

122 2008SymposiumonApplication Specific Processors (SASP 2008)

By comparing (5) with (2) one can see that
⎡ ⎤ ⎡ ⎤ TABLE I
X1 (0) −X1 (4) C OMPLEXITIES AND CRITICAL DELAYS OF VARIOUS 12 POINT MDCT
⎢ X1 (5) ⎥ ⎢ −X1 (1) ⎥
⎢ ⎥ ⎢ ⎥ AND IMDCT ALGORITHMS . N OTE THAT M AND A REFER TO
⎣ X1 (3) ⎦ = ⎣ X1 (4) ⎦ . (6)
MULTIPLICATION AND ADDITION RESPECTIVELY.
X1 (2) −X1 (1) Algorithm complexity Critical delay
Equation (5) shows that X1 (k), k ∈ S2 need not be computed MDCT Proposed 9M + 29A† M + 6A
Ref. [5] 13M + 39A 2M + 6A
separately. Ref. [6] 13M + 27A 2M + 5A
As a last simplification, note that because of the following Ref. [7] 11M + 29A 3M + 7A
trigonometric identities for N = 6, IMDCT Proposed 9M + 23A† M + 5A
Ref. [5] 13M + 33A 2M + 5A
cos(9π/4N ) = cos(π/4N ) − cos(7π/4N ), Ref. [6] 13M + 21A 2M + 4A
− cos(3π/4N ) = − cos(11π/4N ) − cos(5π/4N ), Ref. [7] 11M + 23A 3M + 6A
(7) † Two of these additions can be converted to shifts.
one gets
X2 (1) = −2X2 (2),
(8)
X2 (4) = 2X2 (3).
Therefore computation for (3) can be absorbed into the that x 10
4

12
for (4). The operation of multiplication by 2 may be counted Ref.[5]
as one addition. However, frequently in hardware design, this 11
Ref.[6]
Ref.[7]
scaling by 2 can be realized as a trivial left shift with negligible Bilinear
impact on area and speed of the implementation. 10

The complete bilinear algorithm for 6 point DCT is shown 9

in Fig. 1. The multiplication constants used in Fig. 1 have the

Area (μm2)
8

c1
x(4) X + + X(4) 7
c2
+ X 2x
c3 6
x(1) X + − X(1)
c4
x(0) + X + + X(0) 5
c5
+ X
c6 4
x(5) + X + + X(5) 5 10 15 20
c7 Delay (ns)
x(3) − X + − − X(3)
c8
+ X 2x
c9 Fig. 2. Delay and area for various implementations of 12 point MDCT for
x(2) − X + + − X(2) MP3 short block.
CGT6

Fig. 1. Proposed bilinear implementation of the 6 point DCT-IV.

4
x 10
following values: c1 = 0.5412, c2 = −0.9239, c3 = 1.3066,
9
Ref.[6]
c4 = 0.4687, c5 = 0.3314, c6 = −1.1315, c7 = 0.6533, 8.5 Ref.[7]
Bilinear
c8 = −0.4619 and c9 = 0.2706. 8
The arithmetic complexity and hardware delay of vaious
7.5
MDCT/IMDCT algorithms are compared in Table I. One can
Area (μm )

see that the new proposed algorithm has about 20% less mul-
2

tiplications than other algorithms currently available. Because 6.5

of the bilinear structure, we have only one multiplication on
6
the critical path which results into a very fast implementation.
Figures 2 and 3 show a comparison of various fixed point 5.5
implementations in TSMC 90nm technology. It is evident
5
from this figure that the proposed bilinear algorithm provides
the smallest and fastest MDCT/IMDCT implementation for 4.5
5 10 15 20
Delay (ns)
MP3 audio short block.
D. DCT algorithm for MP3 long block Fig. 3. Delay and area for various implementations of 12 point IMDCT for
We now extend the procedure of Subsection II-C to length MP3 short block.
N = 18 which is used in MP3 audio long block. As before, we

2008 Symposium on Application Specific Processors (SASP 2008) 123

partition the transform components into two sets and compute As in the previous case, this block can be converted into
each set of components together. The partition of transform a multi-dimensional convolution with the use of the group
indices is based on the set of integers A(8N ) = {0 ≤ i < A(8N ) where N = 18. By permuting the inputs and outputs
8N | gcd(i, 8N ) = 1}, i.e., a set of positive integers less as per the index sequence {0, 9, 11, 2, 12, 14, 8, 17, 15, 6, 3, 5}
than 8N which are relatively prime to 8N . The transform and using the sign function values {1, 1, −1, −1, −1, 1, 1, 1, 1,
components X(k) with (2k + 1) ∈ / A(8N ), can be directly 1, −1, −1}, one can translate the computation to a three-
obtained from a DCT of a 6 point sequence y(i) obtained by dimensional convolution with a 2 point cyclic convolution,
folding of the input as a 3 point cyclic convolution and a 2 point Hankel matrix
multiplication along the three dimensions. The flow-chart of
y(i) = x(i) − x(11 − i) − x(12 + i), 0 ≤ i < 6. this multi-dimensional convolution is shown in Fig. 2. Note
that both the index permutation and the sign function sequence
On the other hand, the transform components X(k) with
are obtained systematically from the structure of the group
(2k + 1) ∈ A(8N ) require a computation that we call the N
A(8N ). The details are omitted for brevity.
point Cosine Group Transform, CGTN . This overall structure
of the computation for N = 18 is shown in Fig. 4. x(0) + −
c1
X + − + Y(12)
c2
+ X
c3
−x(9) + − X + − + Y(14)
c4
Cosine Group Transform −x(11) + + + + + Y(11)
X(k) c5
X
CGT 18
2k+1 ε A(8N) +
c6
X

and 0=<k<18. −x(2) + + X + + + Y(2)

c7
x(i) x(12) + − X + − + Y(0)
c8
i=0,1,...,17.
+ X
c9
x(14) + − X + − + −Y(9)
Pre− y(i), i=0,1,...,5. DCT−IV of c10
Addition 6 points X(k) x(8) − − X + − − −Y(3)
2k+1 ε A(8N)
c11
+ X
and 0<k<18. c12
x(17) − − X + − − −Y(5)
c13
x(15) − + X + + − Y(15)
Fig. 4. Flow graph for 18 point bilinear DCT-IV. +
c14
X
c15
x(6) − + X + + − Y(6)
c16
−x(3) − − + − − Y(8)
The Cosine Group Transform, CGTN , can itself be com-
X
c17
+ X
c18
puted recursively by dividing the signal components into sets −x(5) − − X + − − Y(17)

using the group A(N ). In particular, the portion of the CGTN

that uses signal components x(i), with i ∈ / A(N ) forms Fig. 6. Proposed bilinear implementation of multidimensional convolution
involved in the 18 point DCT-IV.
CGTN/3 while the portion using x(i), with i ∈ A(N ) can
be structured as a multidimensional cyclic convolution. The The multiplication constants used in Fig. 6 are as follows:
output of this convolution may be added to the output of c1 = −0.9231, c2 = −0.6528, c3 = 2.2287, c4 = −0.5086,
CGTN/3 to obtain the output of CGTN . The decomposition c5 = −0.3596, c6 = 1.2278, c7 = −0.6025, c8 = −0.4261,
of CGTN is summarized in Fig. 5. c9 = 1.4546, c10 = −0.1628, c11 = 0.2779, c12 = −0.3930,
c13 = 0.1851, c14 = −0.3160, c15 = 0.4469, c16 = 0.7181,
c17 = −1.2258 and c18 = 1.7336.
x(i) Multidimensional
The arithmetic complexity and hardware delay of vaious
Y(k)
2i+1 ε A(8N) cyclic convolution 2k+1 ε A(8N) MDCT/IMDCT algorithms are compared in Table II. The
Post−
Addition X(k)
2k+1 ε A(8N) TABLE II
x(i)
i=0,1,...,17. C OMPLEXITIES AND CRITICAL DELAYS OF VARIOUS 36 POINT MDCT
x(i) X’(k) AND IMDCT ALGORITHMS . N OTE THAT M AND A REFER TO
2i+1 ε A(8N) 2k+1 ε A(8N/3)
CGT 6 MULTIPLICATION AND ADDITION RESPECTIVELY.
Algorithm Complexity Critical delay
MDCT Proposed 36M + 150A† M + 10A
Fig. 5. Flow graph for cosine group transform CGT18 . Ref. [5] 47M + 165A 2M + 9A
Ref. [6] 47M + 129A 2M + 8A
Ref. [7] 43M + 133A 3M + 22A
The computation of 18 point DCT (hence 36 point MDCT/
IMDCT Proposed 36M + 132A† M + 9A
IMDCT) for the MP3 long block requires one to obtain CGT18 Ref. [5] 51M + 151A 2M + 8A
and a 6 point DCT. From Fig. 5, CGT18 computation in turn Ref. [6] 51M + 115A 2M + 7A
involves that of a CGT6 and a large computational block Ref. [7] 43M + 115A 3M + 21A
† Two of these additions can be converted to shifts.
(multidimensional convolution). We already have the 6 point
bilinear DCT algorithm from Subsection II-C. Further, the
CGT6 is also a part of the 6 point DCT as indicated in Fig. 1. proposed bilinear MDCT/IMDCT algorithm has the lowest
Thus we only need to obtain the additional algorithm for the multiplication requirement for MP3 audio short block. For
multidimensional convolution in Fig. 5. 128Kbps MP3 audio stream, our algorithms represent a saving

124 2008 Symposium on Application Specific Processors (SASP 2008)

of nearly 1800 multiplications per second. This can instantly IV. C ONCLUSION
translate into significant power savings for MP3 players, now The modified discrete cosine transform (MDCT) and its
mostly battery operated. Moreover, because the proposed algo- inverse (IMDCT) are amongst the two most compute intensive
rithm is bilinear, a directly mapped hardware will have fastest operations in MP3 audio encoders and decoders. This paper
implementation due to the minimization of multiplications on has developed efficient bilinear algorithms and architectures
the critical path. for computing the MDCT and IMDCT of short and long MP3
III. U NIFIED ARCHITECTURE blocks. Our algorithms have less multiplications than any other
algorithm in literature. In mobile MP3 players this may imply
The computation of MDCT/IMDCT of MP3 long block
a sizable power saving.
requires one to design for a multi-dimensional convolution,
The real advantage of our algorithm is seen when it is cast
a CGT6 and a 6 point DCT. Note that both the CGT6 and the
in Silicon. Because our algorithm is bilinear, it has only one
6 point DCT are used in the computation of MDCT/IMDCT
multiplication on the critical path. All other algorithms have
of MP3 short block. Therefore without introducing any extra
at least two multiplications on the critical path. This implies
computations, we can integrate the computation of MP3 short
a much faster MP3 encoder and decoder architecture allowing
block with that of the long block. Three different unified
much more flexibility in trading speed for area and power of
MDCT/IMDCT architectures are proposed, providing the pro-
the hardware accelerator.
cessing power of one long or one short block (architecture A),
The structure of our algorithm allows us to create unified
one long or two short blocks (architecture B), or half long or
architectures to process both the long and short MP3 blocks.
one short block (folding, architecture C). It is worth empha-
This ability to process both block sizes is crucial to MP3
sizing that the ability to process both block sizes is of central
encoders and decoders because both types of blocks generally
importance to a low cost and low power implementation of
abound in the same data stream. As compared with the
MP3 audio player.
implementations (in the same technology) of other algorithms
We implemented our algorithms and those available in
that process only a single long block, our unified architectures
the literature using fixed point arithmetic in TSMC 90nm
processing both size blocks are still 25% smaller and 25%
technology. The area and delay comparisons are shown in
faster.
Fig. 2. We compare our unified architecture performance with
that of other architectures useful for processing long block R EFERENCES
audio only. One can see that using the same technology, our [1] J. Princen and A. Bradley, “Analysis/synthesis filter bank design based on
unified architectures are still 25% smaller and 25% faster than time domain aliasing cancellation,” IEEE Trans. Acoust. Speech Signal
any previous architectures designed for a single frame size. Processing, vol. ASSP-34, pp. 1153–1161, Oct. 1986.
[2] T. Tsai, T. Chen, and L. Chen, “An MPEG audio decoder chip,” IEEE
Trans. Consumer Electronics, vol. 41, pp. 89–96, Feb. 1995.
x 10
5
[3] S. Tai, C. Wang, and C. Lin, “FFT and IMDCT circuit sharing in DAB
4.5
Ref.[5]
receiver,” IEEE Trans. Broadcasting, vol. 49, pp. 124–131, June 2003.
Ref.[6] [4] Y. Yao, Q. Yao, P. Liu, and Z. Xiao, “Embedded software optimization
4
Ref.[7] for MP3 decoder implemented on RISC core,” IEEE Trans. Consumer
Bilinear Electronics, vol. 50, pp. 1244–1249, June 2005.
Arch.A
Arch.B
[5] V. Britanak and K. Rao, “An efficient implementation of the forward and
3.5 Arch.C inverse MDCT in MPEG audio coding,” IEEE Signal Processing Letters,
vol. 8, pp. 48–50, Feb. 2001.
Area (μm2)

[6] V. Nikolajevic and G. Fettweis, “Improved implementation of MDCT in

3 MP3 audio coding,” 10th Asia-Pacific Conf. Comm. and 5th Intern. Symp.
Multi-Dimen. Mobile Comm., vol. 1, pp. 309–312, Aug. 2004.
[7] S. Lee, “Improved algorithm for efficient computation of the forward and
2.5
inverse MDCT in MPEG audio coding,” IEEE Trans. Circuits Syst. II,
vol. 48, pp. 990–994, Oct. 2001.
2

1.5
6 8 10 12 14 16 18 20 22
Delay (ns)

Fig. 7. Delay and area for unified 12 and 36 point MDCT and IMDCT
architectures (A, B and C) for MP3 audio, with comparison to the 36 point
MDCT architectures in literature.

2008 Symposium on Application Specific Processors (SASP 2008) 125