This application claims priority under 35 USC §119(e)(1) of European Patent Application Number 02293152.1 filed Dec. 18, 2002.
Not Applicable
Not Applicable
1. Technical Field
This invention relates in general to communications and, more particularly, to a method and apparatus for optimizing RSSE calculations.
2. Description of the Related Art
Compared to binary modulation commonly used in 2G standards, advanced wireless communication systems may use high-level modulation schemes. From a system point of view, this results in spectral efficiency at the expense of an increase of Inter-Symbol Interference (ISI).
In order to mitigate ISI, efficient equalizers need to be implemented on the receiver side. The optimal solution to this problem is known as Maximum Likelihood Sequence Estimation (MLSE). The goal of this method is to subtract every possible sequence of L constellation points weighted by the channel estimation factors from the received symbol vector in order to determine the most probable symbols. L corresponds to the channel estimation length.
The concept behind MLSE is to choose the minimum error square between the received symbol yk and every possible linear combination of the last L symbols âk-i weighted by the channel response samples hi, all values being complex. In a mathematical form, the approximated symbol frame of length Lframe is given by:
This inner term of the equation, the branch metric, can be rewritten as:
The branch metric (bm) is composed by three terms. The first one is the soft value of the received symbol, the second one is the estimated contribution of the ISI coming from the L-1 preceding symbols and the third one is the supposed received symbol multiplied by a channel coefficient.
The âk-i(s) for i=1, . . . ,L−1 fully determine a state of the MLSE in a trellis representation. Thus, the number of possible combinations or states, taking into account that âk-i(s) belongs to a set of M symbols and that K=L−1 with the channel response being of length L, is equal to MK. Furthermore, M transitions output each state. This makes MLSE too complicated for implementation in terminals receivers.
A more adapted method, the Reduced State Sequence Equalizer (RSSE) is derived from MLSE. The simplification lies in the reduction of the number of states. Like for the MLSE, the RSSE method consists in minimizing a sum of square values (branch metrics) among all the symbols of the frame under consideration. Therefore, a trellis structure is well adapted to carry out this task. Compared to convolutional decoding, there are, for the general case, no redundancy properties among the branch metrics.
For each symbol received, i.e. on trellis transition, the following units are executed: (1) Branch Metric Computation (BMC) Unit: Compute individually one branch metric per transition, (2) Add Compare and Select (ACS) Unit. With more than two transitions per state and (3) Post-Processing. Current implementations do not use generic structure to perform the BMC and ACS functions. Each branch metric is computed, added to the previous state value and the minimum path is selected for all next states.
While RSSE requires only a fraction of calculations required by MLSE, it still is computation intensive, using significant processing resources. Therefore, a need has arisen for a method and apparatus for performing RSSE computations efficiently.
In the present invention, a sequence of probable symbols is generated from a sequence of received symbols using Reduced State Sequence Estimation. Butterfly circuitry computes terms sm1′=min{sm1+m,sm2−m} and sm2′=min{sm1−m, sm2+m}. Path metrics are computed between a first state and a second state responsive to the received symbols and reference constellation symbols and a best scenario path is determined at the second state using said butterfly circuitry.
The present invention allows branch metric and the accumulate, compare and select (ACS) functions to be aided by use of a butterfly hardware accelerator available to support Viterbi decoding.
For a more complete understanding of the present invention, and the advantages thereof, reference is now made to the following descriptions taken in conjunction with the accompanying drawings, in which:
a-d illustrate four different possible cosets for 8PSK modulation;
The present invention is best understood in relation to
The Reduced State Sequence Equalizer (RSSE) is derived from the MLSE. In its most complicated form, it is equivalent to the MLSE and in its most simplified form to the DFE (Decision Feedback Equalization). Thus, its complexity is comprised between the MLSE complexity and the DFE complexity.
The RSSE principle restricts the number of possible combination of past symbols to be tested by regrouping them inside so-called cosets. A coset is a collection of one or more symbols of the signal set. All symbols of the constellation need to be in one and only one coset. The ith coset at time n-k is represented by Ai(k). Jk-i is the number of cosets at time n-k and is naturally comprised between 1 and the number of symbols of the signal set: 1≦Jk≦M for ∀k. Furthermore, the analysis here below is restricted to values of Jk equal to power of 2, e.g. for a 8-PSK constellation we have Jk ∈[1,2,4,8], and decreasing or equal values of Jk for increasing values of k: Jk≦Jk-1. This last restriction means that the older the symbol of the ISI sum, the less cosets at that time.
a-d illustrates four different possible cosets for 8-PSK modulation. As stated above, the idea behind the RSST is to reduce the number of states of the MLSE. Thus, while a MLSE implementation would have MK states, the RSSE implementation would have
states, where
while both the MLSE and RSSE implementation have M branch metrics outputting every state. In the RSSE, the number of states is given by
In the MLSE, Jk is always equal to the number M of symbols in the constellation.
The RSSE selects the best symbol sequence by minimizing the sum of each Square Euclidean Distance of the estimated ISI term subtracted from the received samples. At each transition, each previous state issues branch metrics related to every symbol ai belonging to the reference constellation that reach the next states where the symbol is contained in the coset Aj.
The branch metric between states is
where f(s) may include any corrective terms to compensate for the reduced symbol set.
Many digital signal processors include one or more hardware accelerators (HWAs). A possible HWA is the butterfly-HWA for trellis calculations in Viterbi decoding. The butterfly structure is shown functionally in
sm1′=min{sm1+m, sm2−m} and
sm2′=min{sm1−m, sm2+m}
Hardware accelerators can be of a number of different varieties. In some cases, the HWA can be implemented as a specific instruction that is an extension of the ISA (Instruction Set Architecture) which executes a complex function. The hardware that supports the instruction may be part of the processing core or coupled to the processing core. There may also be variations on the computations shown above.
For purposes of discussion, it will be assumed that the butterfly structure has the following restrictions: (1) the branch metrics of a butterfly are symmetric, (2) the branch metrics entering a state are of opposite sign and (3) the branch metrics are real values. Thus, the branch metric has to be a linear function for at least the part containing the parameter corresponding to the bm (the symbol for the equalizer). If not, the structure behind the HWA may not be fully used.
The following describes modification of the general RSSE methods to allow use of the butterfly structure. As described above:
The equation can be rewritten as:
Thus, bm(s, âk)=∥S−h0âk∥2=∥S∥2+∥h0âk ∥2−2Re{h0âkS*}. Without loss of generality, we can set the coefficient h0=1 and for 8PSK constellation ∥ak∥2 is equal to 1. Consequently,
bm(s,âk)=∥S∥2−2Re[h0âkS*].
For a given state, the ∥S∥2 is constant. Differences come from the targeted symbol âk. The branch metric can be decomposed further by inserting the real and imaginary parts of the right terms: âk=ax+j·ay, h0=hx+j·hy and S=Sx+j·SyThus:
bm(s,âk)=∥S∥2−2[ax(hxSx+hySy) +ay(hxSy−hySx)], or more concisely:
bm(s,âk)=∥S∥2−2[axRe{h0*S}+ay Im{h0*S}]
Taking into account that h and Sare related to the initiating state and not to the symbol, the equation can be rewritten as:
bm(s, âk)=∥S∥2−2[ax(s)+ay(s) ]=∥S∥2−2ax(s), with
ax(s)=axRe{h0*S} and ay(s)=ayIm{h0* S}.
Examining this expression, it can be seen that the branch metric is composed by three terms, the first one is equal for every branch coming out of one state, the second and third are linear combinations of h0, the S term (both determined by the state) and respectively the real and the imaginary parts of the symbol under consideration.
All the terms depend on the initiating state s. This excludes any use of a butterfly structure with two different states (since the butterfly structure has a m, -m,-m, m symmetry relation). Thus, the only solution is to use one butterfly to compute four transitions issued from one state. This would represent two butterflies per RSSE initiating state for an 8-PSK Constellation.
The RSSE applied to the 8PSK constellation, as defined in
As shown graphically in
sm1=smold+∥S∥2−2ax(s)
sm2=smold+∥S∥2+2ax(s)
m=2ay(s)
The butterfly structure computes the transition corresponding to four diagonal symbols of the 8PSK constellation by adding and subtracting their real and imaginary components. The minimum of the two branch metrics determines the best symbol for each coset. It should be noted however, that in this form the butterfly structure is applicable only to cosets that have the symbols with coordinates (±ax, ±ay), ax and ay being two real numbers.
Also as shown in
A way to use the butterfly structure for the computation of all minimums of two value cosets in 8-PSK constellations is to operate a change of referential. The role of the equalizer is to produce hard decisions. A change of referential will therefore not affect the other parts of the DBB (digital baseband). A π/8 counterclockwise rotation of the constellation shows interesting properties concerning the 8-PSK constellation symbols. As shown in
After the rotation, there are two, four-symbol groups ({A0(k), A3(k)} and {A1(k), A2(k)}) that fulfill butterfly conditions exposed above. Thus, the butterfly structure can be used for all transitions and minimum computations in such a configuration as shown in
Using a π/8 counterclockwise rotation (any rotation of (2k+1)*π/8, where k is an integer can be used), the butterfly HWA can be applied with the following values:
For the first butterfly of symbol groups {A0(k),A3(k)}:
ax(s)=βRe{h0*S} and ay(s)=αIm{h0* S}
sm1=smold+∥S∥2−2βRe{h*S}
sm2=smold+∥S∥2+2βRe{h*S}
m=2αIm{h*S}
For the second butterfly of symbol groups {A1(k),A2(k)}
ax(s)=αRe{h0*S} and ay(s)=βIm{h0* S}
sm1=smold+∥S∥2−2αRe{h*S}
sm2=smold+∥S∥2+2αRe{h*S}
m=2βIm{h*S}
with α=cos(π/8), β=sin(π/8) and
Re{h*S}=hxSx+hySy and Im{h*S}=hxSy−hy
A pseudo-code implementation for a modified 8PSK 2-state RSSE method is provided below:
The π/8 rotation method has some implications: (1) π/8 rotation of received samples yi, (2) π/8 rotation of channel estimation hk and (3) only applicable when Jk≦M/2 and when the constellation coset have symbols of opposite signs.
Using a butterfly structure in the RSSE method can significantly reduce the computing resources needed for RSSE calculations, while having no effect on the performance of the RSSE function.
It should also be noted that the RSSE method described above could be used in either a loss-less or a loss variation. In the loss-less variation, both y (received symbol sample vector) and reference constellation symbols are rotated by angle equal to (2k+1)*π/8 with k being a whole number. The angle may vary depending upon the number of symbols in the constellation.
In the loss variation, both y (received symbol sample vector) and reference constellation symbols are rotated by angle equal to (2k+1)*π/8. Then, either the reference constellation or y is rotated by an angle a bit higher or a bit smaller than the first one. The error is proportional to the absolute difference of the rotation angle between y and the reference constellation symbols.
An alternative way to the constellation rotation solution discussed above in connection with the computation of the four axis symbols (cosets A0 and A2 in
Although the Detailed Description of the invention has been directed to certain exemplary embodiments, various modifications of these embodiments, as well as alternative embodiments, will be suggested to those skilled in the art. The invention encompasses any modifications or alternative embodiments that fall within the scope of the Claims.
Number | Date | Country | Kind |
---|---|---|---|
02293152 | Dec 2002 | EP | regional |
Number | Name | Date | Kind |
---|---|---|---|
6163581 | Kang | Dec 2000 | A |
6397240 | Fernando et al. | May 2002 | B1 |
6707849 | Zangi et al. | Mar 2004 | B1 |
6970520 | Lindbom et al. | Nov 2005 | B1 |
7116732 | Worm et al. | Oct 2006 | B2 |
20020056068 | Aymar et al. | May 2002 | A1 |
20020122507 | Eidson | Sep 2002 | A1 |
20020162074 | Bickerstaff | Oct 2002 | A1 |
Number | Date | Country |
---|---|---|
0 786 872 | Jul 1997 | EP |
1 058 392 | Dec 2000 | EP |
WO 0103308 | Jan 2001 | WO |
WO 0160002 | Jan 2001 | WO |
WO 0239689 | May 2001 | WO |
WO 0178239 | Oct 2001 | WO |
Number | Date | Country | |
---|---|---|---|
20040128334 A1 | Jul 2004 | US |