CN101645758B - Symbol detector and sphere decoding method - Google Patents

Abstract

The invention provides a symbol detector and a sphere decoding method. The symbol detector can receive a wireless signal and search a maximum similarity solution. In the symbol detector, a QR decomposer can perform a QR decomposing operation on a passage response matrix to generate a Q matrix and an R matrix. A matrix converter generates an inner product matrix according to the Q matrix and the wireless signal. A scheduler plans a searching task into a plurality of independent sub-searching tasks. A plurality of Euclidean distance calculators operate in parallel under the control of the scheduler. A plurality of calculating units in the Euclidean distance calculators read the R matrix and the inner product matrix through a pipelined construction to search the maximum similarity solution.

Description

Symbol detector and hybrid sphere decoding method

Technical field

The invention relates to hybrid sphere decoding, the symbol detector hardware structure in particular to improvement search pattern and account form can significantly reduce the search complexity.

Background technology

Along with the growth of wireless telecommunications demand, the utilization benefit of frequency spectrum also becomes the main emphasis of research and development.Multiple-input and multiple-output (Multiple Input Multiple Output; MIMO) system disposes a plurality of antennas respectively in transmission end and receiving terminal, makes it can transmit many piece of data simultaneously in frequency range, improves efficiency of transmission thus.Yet the signal that transmits can produce interference at receiving terminal, so must carry out separation and the detecting of signal at receiving terminal.

Maximum similarity detection method (Maximum Likelihood Detection; MLD) be a kind of common algorithm, can all possible symbol combination be all searched once at receiving terminal, to find out the solution the most similar to transmitting signal.Fig. 1 is the search path schematic diagram that constitutes for a known search tree.Suppose that there are four antennas the transmission end, then search process is divided into four level L ₄To L ₁If each symbol adopts the 64QAM modulation, then each level node to be searched has 64, P ₁To P ₆₄As shown in Figure 1, the possible path 110 between per two levels has 64 * 64, thus four levels may make up 64 altogether ⁴Kind.For receiving terminal, calculate whole 64 ⁴It is quite consuming time that kind may make up.

In order to address this problem hybrid sphere decoding (sphere decoding; SD) utilize a kind of method for searching that limits the cost function higher limit, the scope of restriction search excludes most unnecessary combination of paths effectively.Higher limit r in the hybrid sphere decoding ²Can be considered as the radius of a ball equivalently.In search process, path 110 whenever is connected to down certain node of one deck, just can produce the Euclidean distance increment (EDI) of a correspondence, and the Euclidean distance increment of each layer that add up can obtain the part Euclidean distance (PED) of this search path.PED can be used to and higher limit r ²Relatively, judge thus whether present local search result is effective.When all levels are all visited whole Euclidean distance (the Euclidean Di stance that can obtain a complete search path when finishing; ED).

In known Schnorr-Euchner (SE) enumeration method, defined one with reference to the centre of sphere.When searching to the m layer, the SE enumeration method can be with reference to centered by the centre of sphere, to contiguous near and far search of the rigid result of decision, can guarantee that so it is to serve as preferential with the EDI smaller that each layer is visited node sequence, this can make and find the chance of optimum solution to occur in advance, and then has reduced unnecessary search combination.On the other hand, once the someone proposed the fixedly hybrid sphere decoding method (Fixed-ComplexitySD) of complexity, utilized corrected V-BLAST sort method, selected N _FullBe placed on and search the treetop layer and carry out the universe search method, and remainder only the fill order to search method.Yet utilize the V-BLAST sort method to need very big amount of calculation, be only applicable to the indoor wireless networks environment, passage changes at any time in the environment in action, and each symbol is wanted once order of extra computation, and the V-BLAST method is obviously inappropriate.Therefore generally, the technology of hybrid sphere decoding is reducing on the complexity of searching tree, still has progressive space.

Summary of the invention

This case proposes a kind of sort method of Improvement type to this, only need many programs of ordering together in traditional Q R decomposes, just can with known corrections after the V-BLAST sort method reach almost identical usefulness.In addition, the present invention proposes a kind of different hardware structure, with multi-line and multitask parallel processing ability.The unidirectional search scheduling method that this framework collocation proposes can search the result more efficiently.At last, the present invention utilizes the characteristic of star-plot, proposes to simplify Calculation Method, can avoid the use of multiplier in a large number.

In one embodiment, the present invention proposes a kind of symbol detector, in order to receive a wireless signal and to search a maximum similarity solution with the hybrid sphere decoding algorithm.Comprise a QR decomposer in this symbol detector, in order to a channel response matrix is carried out a QR decomposition operation, to produce a Q matrix and a R matrix.One matrix converter produces product matrix in according to this Q matrix and this wireless signal.One internal memory can be kept in this R matrix and is somebody's turn to do interior product matrix.One scheduler is searched mission planning with one and is become a plurality of independently sub-search tasks.A plurality of Euclidean distance calculators are subjected to this scheduler control and parallel running.A plurality of computing units in the Euclidean distance calculator read this R matrix with pipeline architecture and are somebody's turn to do interior product matrix to seek this maximum similarity solution.Comprise a plurality of row vectors in this channel response matrix, and this QR decomposer rearranges described row vector according to an ordering rule in the process of carrying out the QR decomposition operation, make in this R matrix that produces the N in the lower right corner _FullIndividual diagonal element has minimum value, and all the other diagonal elements increase progressively toward the lower right corner in regular turn from the upper left corner as far as possible, and rearranges the hierarchic sequence of searching tree according to ranking results.Thus, the interior product matrix and the Q matrix that produce can meet the condition of using fast searching.

This QR decomposer carries out QR and decomposes, and makes this Q matrix of generation and the product of this R matrix equal this channel response matrix.This QR decomposer can produce an objective matrix in decomposable process, by the capable square value (column norm) that continues to calculate each row vector in the objective matrix, and according to this ordering rule this objective matrix is carried out delegation's exchange, and the square value of will going is ranked (N _Full+ 1) little capable vector drains into the leftmost side of pending part.Decomposable process lasts till that always the pending part of objective matrix is left N _FullTill the individual capable vector, last N _FullIndividual capable vector then can not limit its order.

In another embodiment, the present invention proposes a kind of hybrid sphere decoding method, is implemented in the above-mentioned symbol detector.At first, a channel response matrix is carried out a QR decomposition operation, to produce a Q matrix and a R matrix.Then produce product matrix in according to this Q matrix and this wireless signal.By this, can organize one to search tree, comprise a plurality of levels, wherein N _FullIndividual level is used the universe search method, and other level is used unidirectional search method.At last, according to this R matrix and should in product matrix, search from this and to seek this maximum similarity solution tree.Comprise a plurality of row vectors in this channel response matrix, and this step of carrying out the QR decomposition operation comprises, and rearranges described row vector according to an ordering rule, makes in this R matrix that produces the N in the lower right corner _FullIndividual diagonal element has minimum value, and all the other diagonal elements increase progressively toward the lower right corner in regular turn from the upper left corner as far as possible.

Description of drawings

Be further to disclose concrete technology contents of the present invention, below in conjunction with embodiment and accompanying drawing describes in detail as after, wherein:

Fig. 1 is the search path schematic diagram that constitutes for a known search tree;

Fig. 2 is one search path schematic diagram for the embodiment of the invention;

Fig. 3 is one symbol detector for the embodiment of the invention;

Fig. 4 is the embodiment for computing unit among Fig. 3 of the present invention;

Fig. 5 is the embodiment for decision-making device among Fig. 4 of the present invention;

Fig. 6 is QPSK quantizer among Fig. 5 of the present invention, the transformation curve of 16QAM quantizer and 64QAM quantizer;

Fig. 7 is the flow chart for hybrid sphere decoding method of the present invention;

Fig. 8 is the flow chart that decomposes for QR among Fig. 7 of the present invention; And

Fig. 9 a is the process schematic diagram that decomposes for QR of the present invention to Fig. 9 d.

Embodiment

The following example specific description realizes the present invention as how preferable mode.The mode that embodiment only generally uses for explanation, but not in order to limit scope of the present invention.Actual range is as the criterion so that claim is listed.

Find in our simulation process that in the search process of the depth-first hybrid sphere decoding algorithm of collocation SE enumeration method, just find the probability of maximum similarity solution quite high when visiting to the bottom (m=1) for the first time, especially when SNR was more high, this probit value was more high.Change speech, if the quality that original search is set according to SNR reorganizes the level framework, make the bad level of SNR adopt the universe search method, and the good level of SNR is only carried out the unidirectional search method, still can find the maximum similarity solution, can more effectively reduce and search the complexity of setting.

Fig. 2 is one search path schematic diagram for the embodiment of the invention.In the present embodiment, a search tree is that the real number model that receives signal is set up, and comprises a plurality of levels.For instance, if a receiver has four antennas, the signal that every antenna receives comprises real part and imaginary part, then can set up eight level L ₈To L ₁, real part and the imaginary part of respectively corresponding every antenna.On the other hand, if signal is with the 64QAM modulation, then the value of symbol of each real part or imaginary part representative has-7, and-5 ,-3 ,-1,1,3,5,7 eight kinds may.So eight nodes are arranged, each corresponding value of symbol on each level.Present embodiment uses a kind of sort method in the process that QR decomposes, set up the hierarchic sequence of this search tree, makes this search tree meet the condition of using this fast searching.More particularly, this sort method makes former levels such as L ₈And L ₇Be applicable to the universe search method, and other level L ₆To L ₁Between then be suitable for unidirectional search method.For instance, the universe search method makes the 8th layer of L ₈With the 7th layer of L ₇Each node all searched (having 64 paths 110), and (comprising the 6th layer) after the 6th layer, each paths 110 uses unidirectional search method, each level only has according to the SE enumeration method that node is selected to come out to search (as path 120), and the complete search path number of therefore whole search tree is exactly [8 * 8 * 1 * 1 * 1 * 1 * 1 * 1].The default node of each layer searched number predefined on demand in addition, such as the 8th layer of L ₈With the 7th layer of L ₇The use universe is searched, and the 6th layer and the 5th layer can have respectively that two nodes are selected to come out to carry out unidirectional search, and has only a node to be selected after the 4th layer, and then the complexity set of whole search can be expressed as [8 * 8 * 2 * 2 * 1 * 1 * 1 * 1].At last by searching the maximum similarity solution of coming out, can infer the value of symbol of real part and imaginary part on four corresponding antennas, further solve the value of symbol of 64QAM thus by star-plot (constellation).About above-mentioned sort method, and the node of each level chooses method, will describe in detail among the embodiment below.

Fig. 3 is one symbol detector for the embodiment of the invention.The input value of this symbol detector 300 is to receive signal matrix y and channel response matrix H.One QR decomposer 304 at first carries out a QR decomposition operation to this channel response matrix H, to produce a Q matrix and a R matrix.One matrix converter 302 then receives product matrix in the signal matrix y generation one according to this Q matrix and this.In QR decomposition operation process of the present invention, used a kind of special sort method, make the interior product matrix of generation and R matrix be applicable to that the universe search method adds unidirectional search method, detailed embodiment will illustrate in Fig. 8.In the process of hybrid sphere decoding, desired maximum similarity solution can be expressed as:

${\hat{s}}_{\mathrm{ML}}=\underset{s\∈\mathrm{\Λ}}{\min }{\left|\right|Q^{H}y-R\·s\left|\right|}^2---\left(1\right)$

Wherein s is one group of variable vector, represents one group of possible symbolic solution, and s can correspond to certain eight node in these eight levels in the present embodiment.The output result of this matrix converter 302 and this QR decomposer 304 can be temporarily stored in the internal memory 306.The present invention proposes a scheduler 320, in order to assign can parallel running a plurality of Euclidean distance calculators 310 of search task to the symbol detector 300 in handle.Each Euclidean distance calculator 310 reads this R matrix and is somebody's turn to do interior product matrix to carry out the computing as (1) formula from this internal memory 306.More particularly, all path plannings that this scheduler 320 may be visited universe search method among Fig. 2 and unidirectional search method become a plurality of independently unidirectional branches to search task, and assign described Euclidean distance calculator 310 multiplex (MUX)s ground to carry out described unidirectional branch by control signal #task and search task.Each unidirectional Branch Tasks only can be searched one group of feasible solution basically.When one Euclidean distance calculator 310 is carried out a unidirectional Branch Tasks, can visit each unidirectional wherein node of searching layer according to the SE enumeration method, and EDI is added up.At last all EDI add up the ED that forms again with buffer 308 in higher limit r ²Relatively, to judge whether to find minimum ED.By the control of scheduler 320, can carry out parallel processing for a plurality of Euclidean distance calculators 310 all unidirectional Branch Tasks mean allocation.In each Euclidean distance calculator 310, be connected in series a plurality of computing units 312, formed a kind of pipeline architecture, the different levels of a plurality of unidirectional Branch Tasks are carried out simultaneously, with the maximum effect of performance hardware.

Fig. 4 is the embodiment for computing unit 312 among Fig. 3 of the present invention.Because each computing unit 312 is repeatedly to carry out with a kind of expression formula, so its hardware structure can be represented by the computing unit 312 of Fig. 4, and each computing unit 312 is to carry out the partly computing of Euclidean distance (PED) behind the decision-making of the newly-increased node of this layer and the newly-increased node.Product matrix in wherein an estimation unit 410 receives and is somebody's turn to do, the nodal value of this R matrix and former levels is to calculate one with reference to the centre of sphere

This carries out the PED computing except being used for reference to the centre of sphere, can also be for scheduler 320 reference during as the unidirectional Branch Tasks of planning.Similar with known Schnorr-Euchner (SE) enumeration method, if scheduler 320 is when planning tasks, when a certain level must branch out more than one unidirectional Branch Tasks, then determine the priority of described unidirectional Branch Tasks according to this value with reference to the centre of sphere.More particularly, more near this node with reference to the centre of sphere, more preferential the remittance carried out its unidirectional branch search task in the computing unit.

Wherein when carrying out the PED computing, this estimation unit 410 at first calculates this with reference to the centre of sphere according to following formula:

${\tilde{s}}_{m|m+1}=\frac{-1}{r_{m,m}}\{\underset{j=m+1}{\overset{M}{\mathrm{\Σ}}}{r}_{m,j}(s_j-1)+\underset{j=m+1}{\overset{M}{\mathrm{\Σ}}}{r}_{m,j}-{\tilde{y}}_m\}---\left(2\right)$

Wherein m represents present number of levels, and number of levels is to be begun to successively decrease with level by M.Should

Be to be this value with reference to the centre of sphere, and r _{M, m}Represent m diagonal element in this R matrix, r _{M, j}Represent (m, j) individual element in this R matrix.s _jBe the nodal value for the j level node of visiting in the unidirectional Branch Tasks instantly, in 64QAM only can be ± 1, ± 3, ± 5 and ± 7 eight kinds of probable values.

Be product matrix in this.In the process that a unidirectional Branch Tasks carries out, in (2) formula

$\underset{j=m+1}{\overset{M}{\mathrm{\Σ}}}{r}_{m,j}-{\tilde{y}}_m$ Can be a fixed value, so only need set of circuits on calculating.Because (s _j-1) might be 0, ± 2 only, ± 4, ± 6 and 8 these eight kinds of values be so only need simple translation and add operation just can obtain in (2) formula on circuit design

$\underset{j=m+1}{\overset{M}{\mathrm{\Σ}}}{r}_{m,j}(s_j-1),$ The use that design like this can keep away multiplier, design and the cost of simplification circuit.

And in this computing unit 312, one apart from summation circuit 420 then according to this with reference to the PED of the centre of sphere and last level to calculate the PED of present level.This calculates the PED of present level according to following formula apart from summation circuit 420:

${\mathrm{PED}}_m={\mathrm{PED}}_{m+1}+r_{m.m}^2{(s_m-{\tilde{s}}_{m|m+1})}^2---\left(3\right)$

PED wherein _mBe the PED for present level, PED _M+1Be the part Euclidean distance for last level, and s _mBe to be this rigid decision value with reference to the centre of sphere.(3) formula can be realized by simple circuit.For instance, comprise a subtracter 404 at this in 420, in order to subtracting each other with reference to centre of sphere decision value rigid with it.Then a multiplier 406 is with the diagonal values r in the R matrix _{M, m}Multiply each other with this output valve of 404, again by a squarer 408 with in (3) formula

$r_{m.m}^2{(s_m-{\tilde{s}}_{m|m+1})}^2$ Partly calculate.Last adder 412 namely gets the PED value of level up till now with the output valve that the Euclidean distance PEDm+1 of last level adds this squarer 408.

In Fig. 3, this symbol detector 300 further comprises a buffer 308, in order to store a higher limit r ²And among Fig. 4 each is apart from comprising a comparator 414 in the summation circuit 420, can be with this 412 PED and this higher limit r that calculates gained ²Compare.A level add up and PED all can be examined once.If this PED is less than this higher limit r ², then this unidirectional Branch Tasks can be proceeded.On the contrary, if this PED greater than this higher limit r ², represent that then the search of this branch should be ended, seemingly be not maximum because it is spent mutually.Aspect in addition, when an Euclidean distance calculator 310 was finished once unidirectional Branch Tasks, the PED of last level output namely equaled the whole Euclidean distance (ED) of this unidirectional Branch Tasks.This ED also with this higher limit r ²Compare.If this ED is less than this higher limit r ², then new maximum similarity solution is found in expression, so the higher limit r in this buffer 308 ²Just be updated to the value of this ED.Because a plurality of Euclidean distance calculators 310 are parallel processing, find new higher limit r so work as any one Euclidean distance calculator 310 ²During value, any PED surpasses higher limit r ²Task all can end at once, and the complexity of searching also can be along with quick convergence.

Fig. 5 is the embodiment for decision-making device 402 among Fig. 4 of the present invention.Possess 510, two 16QAM quantizers 520 of a QPSK quantizer and three 64QAM quantizers 530 in this decision-making device 402, carry out quantization operations in order to real part or imaginary values to starlike signal.Before carrying out quantization operations, one first subtracter 502 is at first with an input value V _IN Subtract 1, then by first shifter 504 operation result of this first subtracter 502 is moved to right one.By one group of first switch 512 and second switch 522, visual demand selects signal #SE selectivity to connect this QPSK quantizer 510 according to one, 16QAM quantizer 520 and 64QAM quantizer 530 one of them, the operation result of first shifter 504 is sent to wherein carries out quantization operations.On the other hand, one second shifter 506 receives these QPSK quantizers 510 by this second switch 522, one of them quantized result of 16QAM quantizer 520 or 64QAM quantizer 530, and it is moved to left one.At last by a first adder 508 operation result of this second shifter 506 is added 1 to produce an output valve V _OUTAnd when in fact being applied in the computing unit 312, this input value V _INBe for this with reference to the centre of sphere, this output valve V _OUTBe to be this rigid decision value.Fig. 6 is the transformation curve for 510, the 1 6QAM quantizers 520 of QPSK quantizer among Fig. 5 of the present invention and 64QAM quantizer 530.By selecting the selection of signal #SE, can make the decision-making device 402 of Fig. 5 be applicable to QPSK, 16QAM or three kinds of different modulation systems of 64QAM.

Fig. 7 is the flow chart for hybrid sphere decoding method of the present invention.Based on above-mentioned hardware structure, reality of the present invention is made method can be summarized as the following step.At first in step 701, receive signal matrix y by many antennas, and measure the channel response matrix H.In step 703, determine to use the number of plies (N of universe search method according to factors such as antenna amount, modulation type _Full), and other level is used unidirectional search method.Then in step 705, this channel response matrix H is carried out a QR decomposition operation, to produce a Q matrix and a R matrix.Traditionally the QR method of decomposing have a variety of, Householder for example, Modified Gram-Schmidt and Givens rotations etc., present embodiment is in the Householder mode as an illustration.This channel response matrix H is made up of a plurality of row vectors, and when carrying out QR decomposition operation of the present invention, rearranges described row vector according to an ordering rule, makes in this R matrix that produces the N in the lower right corner _FullIndividual diagonal element has minimum value, and all the other diagonal elements increase progressively toward the lower right corner in regular turn from the upper left corner as far as possible, and rearranges the hierarchic sequence of searching tree according to ranking results.Thus, the interior product matrix and the Q matrix that produce can meet the condition of using fast searching.Be example with Fig. 2, work as N _Full=2 o'clock, the 8th layer of L ₈With the 7th layer of L ₇Adopted the universe search method, and other level adopts unidirectional search method.It is described that yet the level that adopts the universe search method is not limited to present embodiment.In step 707, after having determined to search the structure of tree, further can be planned to a plurality of unidirectional Branch Tasks.As discussed previously, a complexity of searching tree may be different because of search strategy, for example [8 * 8 * 1 * 1 * 1 * 1 * 1 * 1] or [8 * 8 * 2 * 2 * 1 * 1 * 1 * 1].With the example of complexity for [8 * 8 * 1 * 1 * 1 * 1 * 1 * 1], can be assigned as 64 unidirectional Branch Tasks in the equivalence, each unidirectional Branch Tasks is responsible for seeking one group of possible solution.In step 709, utilize a plurality of Euclidean distance calculators 310 among Fig. 3, can handle a plurality of unidirectional Branch Tasks simultaneously.Say further, owing to comprise a plurality of computing units 312 in each Euclidean distance calculator 310, form a kind of line construction, so under the arrangement and appointment of scheduler 320, each Euclidean distance calculator 310 can be handled the different levels in the unidirectional Branch Tasks of a plurality of differences further simultaneously.

Because hybrid sphere decoding of the present invention is mainly used in multiple-input and multiple-output (MIMO) system, corresponding second quantity receiving terminal receives the matrix that produces so this channel response matrix H is first quantity transmission end basically.Each level among Fig. 2 in this search tree is real part or the imaginary part of each corresponding transmission end in fact, and each level comprises a plurality of nodes, corresponding all possible real part or imaginary part of symbol value.

Fig. 8 is the flow chart that decomposes for QR among Fig. 7 of the present invention.Wherein formed by a plurality of row vectors in this channel response matrix H, and this QR decomposer 304 can recursively produce one group of instantaneous matrix G in the process of carrying out the QR decomposition operation ₁, G ₂, G ₃G _MWith objective matrix R ₁, R ₂R _MMake

R ₁＝G ₁H

R ₂＝G ₂G ₁H

R ₃＝G ₃G ₂G ₁H

R _M-1＝G _M-1…G ₂G ₁H

R＝R _M＝G _MG _M-1…G ₂G ₁H

Wherein R is a upper triangular matrix, and:

G ₁G ₂…G _M-1G _M＝Q

In decomposable process, the present invention according to a kind of ordering rule respectively to objective matrix H, R ₁, R ₂R _M-1Rearrange its row vector, make in the R matrix of last generation the N in the lower right corner _FullIndividual diagonal element has minimum value, and all the other diagonal elements increase progressively to the lower right corner in regular turn from the upper left corner as far as possible.

In step 801, be objective matrix with the channel response matrix H at first, start the QR decomposing program.In step 803, calculate the capable square value (column norm) of row vector in the objective matrix.Then this QR decomposer 304 carries out delegation's exchange step according to this ordering rule to this objective matrix, the square value of will going seniority among brothers and sisters (N _Full+ 1) little capable vector drains into the leftmost side of pending part.Decomposable process lasts till that always the pending part of objective matrix is left N _FullTill the individual capable vector, last N _FullIndividual capable vector then can not limit its order.The process of row exchange can stay a record, follows the trail of the initial and last position relation of each row in the channel response matrix H thus, after hunting out the maximum similarity solution, also needs to rely on this part record to find out real part and the imaginary part of symbol of original correspondence.In step 805, this objective matrix that rearranged is carried out QR decompose, to produce an intermediary matrix G _iAnd R _iWherein i represents number of levels, scope from 1 to maximum level M.In the step 807, judge whether to finish QR and decompose.If do not finish as yet, then carry out step 809, number of levels advances one-level, with this intermediary matrix R _iBe assigned as objective matrix, recursively execution in step 803.The computing dimension of intermediary matrix can be along with the propelling of level depression of order one by one.For instance, when decomposing for the first time, the computing of row square value is at 8 * 8 elements in the channel response matrix H with the row exchange, and when decomposing for the second time, the computing of row square value is at computing R with the row exchange ₁In the lower right corner 7 * 7 elements, and for the third time the time, the computing of row square value then is at R-with the row exchange ₂6 * 6 elements in the matrix lower right corner ... the rest may be inferred.When the recurrence decomposition proceeds to R ₈After namely accuse and finish output Q matrix and R matrix in step 811.

Fig. 9 a is the process schematic diagram that decomposes for QR of the present invention to Fig. 9 d.For convenience of description, Fig. 9 a shows one 5 * 5 channel response matrix H, and sets N _Full=2.At first pass through the ordering of step 803, row square value ranked third a little N ₅Be moved to the leftmost side, but and remaining row randomize.Then with reference to figure 9b.

That Fig. 9 b shows is the R that produces for according to the channel response matrix H after Fig. 9 a ordering ₁Matrix, wherein first row is except diagonal element R ₁₁Outside other element all be converted to zero.This R ₁Matrix is recursively by step 803 ordering, and 4 * 4 of the lower right corner partly is considered as the pending part of this objective matrix, goes exchange after its row square value is obtained again according to this.Similar with the step of last level, row square value ranked third a little N ₄Be arranged at the leftmost side, remaining row is randomize then.Follow this R that was rearranged ₁Matrix further is decomposed into objective matrix R ₂With a G ₂Matrix (not shown).

What Fig. 9 c showed is to be the R after sorting according to Fig. 9 b ₁The R that matrix decomposition produces ₂Matrix.This R ₂Matrix is recursively by step 803 ordering, the lower right corner 3 * 3 partly by as the pending part of this objective matrix, go exchange after its row square value is obtained again according to this.Similar with the step of last level, row square value ranked third the little leftmost side that is arranged at, and remaining row is randomize then.Above-mentioned transfer process is performed until objective matrix R _iOnly surplus 2 behaviors of pending row vector end.The order of last 2 row then can arbitrarily be arranged.At last obtain a upper triangular matrix R who converts at Fig. 9 d, wherein diagonal entry is R ₁₁To R ₅₅On the other hand, corresponding Q matrix (not shown) is also along with generation, and embodiment is as described in Figure 3 implemented according to this thus.

In addition, because the characteristic of hybrid sphere decoding, if placing, the node that SNR is the poorest searches the misgivings that the tree orlop can't wrong diffusion, so when at the beginning objective matrix H being sorted, also can select first row that comes of capable square value minimum earlier, follow-up objective matrix ordering is then kept former regular.

So in embodiments of the present invention, this ordering rule can have two kinds of selections: 1. can come the objective matrix rightmost side to the row of row square value minimum when sorting for the first time, depression of order calculating process afterwards changes arranges (N with row square value _Full+ 1) N is down to up to exponent number in the little leftmost side that is arranged in the pending part of objective matrix _FullThe time, the succeeding target matrix can be arranged.2. the square value of just will going is at the beginning arranged (N _Full+ 1) little capable vector is arranged in the leftmost side of the pending part of objective matrix, is down to N up to exponent number _FullThe time, the succeeding target matrix can be arranged.

In sum, the present invention proposes a kind of Improvement type hybrid sphere decoding method, the level of searching tree can be sorted in addition, and complexity is simplified in arrange in pairs or groups universe search and unidirectional search.In addition, search tree and can be planned to a plurality of unidirectional search tasks by scheduling, make hardware can realize the processing of multitask parallel pipeline easily.The applied circuit framework of the present invention, except have at a high speed with high extendibility, also utilized the skill of translation and adder, avoid complicated multiplication inner product operation.Because searching tree is to be based upon on the real number model, so can adopt a kind of rigid decision-making device simple in structure.Though embodiments of the invention are tuned as example with four antennas and 64QAM, the disclosed method of the present invention all can be applied in the mimo system of other different antennae number and modulation mode.

Though the present invention with the preferred embodiment explanation as above is understandable that the not necessarily so restriction of scope of the present invention.Relative, any is that apparent improvement is all in covering scope of the present invention based on same spirit or to the known technology person.Therefore the claim scope must be understood in the mode of broad sense.

Claims (28)

1. a symbol detector in order to receive a wireless signal and to search a maximum similarity solution with the hybrid sphere decoding algorithm, is characterized in that, comprises:

One QR decomposer is in order to carry out a QR decomposition operation to a channel response matrix, to produce a Q matrix and a R matrix;

One matrix converter couples this QR decomposer, produces product matrix in according to this Q matrix and this wireless signal;

One internal memory couples this matrix converter and this QR decomposer, in order to keep in this R matrix and to be somebody's turn to do interior product matrix;

One scheduler is searched tree in order to plan one, and a search mission planning is become a plurality of independently sub-search tasks, and wherein this search tree has universe search degree of depth N _Full

A plurality of Euclidean distance calculators couple this scheduler, are subjected to the control of this scheduler and parallel running, and each Euclidean distance calculator comprises a plurality of computing units and is concatenated into a pipeline architecture, in order to according to this R matrix and this maximum similarity solution of this inner product matrix computations; Wherein:

Comprise a plurality of row vectors in this channel response matrix; And

This QR decomposer rearranges described row vector according to an ordering rule, makes in this R matrix that produces minimum N _FullIndividual diagonal element is arranged in the lower right corner, and all the other diagonal elements increase progressively toward the lower right corner in regular turn from the upper left corner.

2. symbol detector as claimed in claim 1 is characterized in that, wherein:

This QR decomposer is carried out a plurality of depression of order computings handling an objective matrix line by line, and calculates capable square value of each the row vector that is untreated in this objective matrix partly.

3. symbol detector as claimed in claim 2 is characterized in that, wherein this ordering rule be for:

When for the first time producing objective matrix, the row of the capable square value minimum in this objective matrix is changed the rightmost side to this objective matrix;

For the second time above when producing objective matrix, will go square value (N _Full+ 1) little row changes the leftmost side partly of being untreated to this objective matrix; And

The surplus N of the line number that in the objective matrix that produces, is untreated _FullWhen row, do not go exchange.

4. symbol detector as claimed in claim 2 is characterized in that, wherein this ordering rule be for:

Each when producing objective matrix, will go square value (N _Full+ 1) little row changes to this objective matrix leftmost side; And

The surplus N of the line number that in the objective matrix that produces, is untreated _FullWhen row, do not go exchange.

5. symbol detector as claimed in claim 1 is to be applied to multi-input multi-output system (MIMO), it is characterized in that, wherein:

This channel response matrix is to be sent by one first a quantity transmission end, receives the matrix that produces by one second a quantity receiving terminal;

This searches real part or the imaginary part of each the corresponding transmission end of each level in the tree; And

Each level comprises a plurality of nodes, corresponding all possible real part or imaginary part of symbol value.

6. symbol detector as claimed in claim 5, it is characterized in that, when wherein this scheduler is organized this search tree, all nodes that universe search method and unidirectional search method may be visited are formulated for a plurality of unidirectional Branch Tasks, and assign this Euclidean distance calculator multiplex (MUX) ground to carry out described unidirectional Branch Tasks.

7. symbol detector as claimed in claim 6 is characterized in that, wherein:

Each Euclidean distance calculator carries out hybrid sphere decoding according to the unidirectional Branch Tasks of assigning, and by the pipeline architecture of a plurality of computing unit serial connections, visits various level node in a plurality of unidirectional Branch Tasks simultaneously;

Each computing unit carries out part Euclidean distance PED computing to the node of visiting, and comprises:

One estimation unit receives and is somebody's turn to do interior product matrix, and the nodal value of this R matrix and former levels is to calculate one with reference to the centre of sphere; And

One apart from summation circuit, according to this with reference to the PED of the centre of sphere and last level to calculate the PED of present level.

8. symbol detector as claimed in claim 7 is characterized in that, wherein:

This scheduler is in planning during unidirectional Branch Tasks, if when a certain level must branch out more than one unidirectional Branch Tasks, then determines the priority of described unidirectional Branch Tasks according to this value with reference to the centre of sphere; And

More near this node with reference to the centre of sphere, more preferential the remittance carried out its unidirectional branch search task in the computing unit.

9. symbol detector as claimed in claim 7 is characterized in that, wherein this estimation unit calculates this with reference to the centre of sphere according to following formula:

${\tilde{s}}_{m|m+1}=\frac{-1}{r_{m,m}}\{\underset{j=m+1}{\overset{M}{\mathrm{\Σ}}}{r}_{m,j}(s_j-1)+\underset{j=m+1}{\overset{M}{\mathrm{\Σ}}}{r}_{m,j}-{\tilde{y}}_m\}$

Wherein m represents present number of levels, and the superiors' number of levels is M, and orlop is 1;

Should

Be to be this value with reference to the centre of sphere;

r _{M, m}Represent m diagonal element in this R matrix, r _{M, j}Represent (m, j) individual element in this R matrix;

s _jIt is the nodal value for the j level node of visiting in the unidirectional Branch Tasks instantly; And

Be m element value of product matrix in this.

10. symbol detector as claimed in claim 8 is characterized in that, wherein should calculate the PED of present level apart from summation circuit according to following formula:

${\mathrm{PED}}_m={\mathrm{PED}}_{m+1}+r_{m.m}^2{(s_m-{\tilde{s}}_{m|m+1})}^2$

PED wherein _mBe the PED for present level, PED _M+1Be the PED for last level, r _{M, m}Represent m diagonal element in this R matrix,

Be to be this value with reference to the centre of sphere, and s _mBe to be this rigid decision value with reference to the centre of sphere.

11. symbol detector as claimed in claim 10 is characterized in that, wherein should comprise a decision-making device apart from summation circuit, in order to produce this rigid decision value according to this with reference to the centre of sphere, this decision-making device comprises:

One first subtracter subtracts 1 with an input value;

One first shifter moves to right one with the operation result of this first subtracter;

One 1 QPSK quantizers, one 2 16QAM quantizers and one 3 64QAM quantizers;

One first switch and one second switch select signal to select to connect this QPSK quantizer according to one, 16QAM quantizer and 64QAM quantizer one of them, the operation result of first shifter is sent to wherein carries out quantization operations;

One second shifter receives this QPSK quantizer by this second switch, one of them quantized result of 16QAM quantizer or 64QAM quantizer, and it is moved to left one; And

One first adder adds 1 with the operation result of this second shifter, to produce an output valve; Wherein:

This input value be for this with reference to the centre of sphere, this output valve is to be this rigid decision value.

12. symbol detector as claimed in claim 10 is characterized in that, wherein:

This symbol detector further comprises a buffer, in order to store a higher limit;

Each is compared PED and this higher limit of present level apart from comprising a comparator in the summation circuit;

If this PED less than this higher limit, then continues this unidirectional Branch Tasks; And

If this PED greater than this higher limit, then ends this unidirectional Branch Tasks.

13. symbol detector as claimed in claim 11 is characterized in that, wherein:

When each Euclidean distance calculator was finished once unidirectional Branch Tasks, a whole Euclidean distance ED and this higher limit that last level is exported compared; And

If, then upgrading this higher limit less than this higher limit, this ED is this ED.

14. symbol detector as claimed in claim 11 is characterized in that, wherein when all unidirectional Branch Tasks were finished, the node path of this higher limit correspondence was the maximum similarity solution.

15. a hybrid sphere decoding method in order to receive a wireless signal and to search a maximum similarity solution, is characterized in that, comprises:

One channel response matrix is carried out a QR decomposition operation, to produce a Q matrix and a R matrix;

Produce product matrix in according to this Q matrix and this wireless signal;

Search degree of depth N according to a universe _FullTree is searched in planning one, and a search mission planning is become a plurality of independently sub-search tasks;

According to this R matrix and should in product matrix to seek this maximum similarity solution; Wherein

Comprise a plurality of row vectors in this channel response matrix; And

This QR decomposition operation rearranges described row vector according to an ordering rule, makes in this R matrix that produces, and the Nfull in a lower right corner diagonal element has minimum value, and all the other diagonal elements increase progressively toward the lower right corner in regular turn from the upper left corner.

16. hybrid sphere decoding method as claimed in claim 15 is characterized in that, wherein:

This QR decomposition operation comprises a plurality of depression of order computings, and each depression of order computing is handled an objective matrix line by line and calculated the capable square value of each row vector in this objective matrix.

17. hybrid sphere decoding method as claimed in claim 16 is characterized in that, wherein this ordering rule be for:

When for the first time producing objective matrix, the row of the capable square value minimum in this objective matrix is changed the rightmost side to this objective matrix;

For the second time above when producing objective matrix, will go square value (N _Full+ 1) little row changes the leftmost side partly of being untreated to this objective matrix; And

The surplus N of the line number that in the objective matrix that produces, is untreated _FullWhen row, do not go exchange.

18. hybrid sphere decoding method as claimed in claim 16 is characterized in that, wherein this ordering rule be for:

Each when producing objective matrix, will go square value (N _Full+ 1) little row changes to this objective matrix leftmost side; And

When the surplus Nfull of the line number that is untreated in the objective matrix that produces is capable, do not go exchange.

19. hybrid sphere decoding method as claimed in claim 17 is to be applied to multiple-input and multiple-output (MIMO) system, it is characterized in that, wherein:

This channel response matrix is to be sent by one first a quantity transmission end, receives the matrix that produces by one second a quantity receiving terminal;

This searches real part or the imaginary part of each the corresponding transmission end of each level in the tree; And

Each level comprises a plurality of nodes, corresponding all possible real part or imaginary part of symbol value.

20. hybrid sphere decoding method as claimed in claim 19 is characterized in that, wherein organizes the step of this search tree to comprise, all nodes that universe search method and unidirectional search method may be visited are formulated for a plurality of unidirectional Branch Tasks.

21. hybrid sphere decoding method as claimed in claim 20 is characterized in that, further comprises:

According to the unidirectional Branch Tasks of assigning, visit various level node in a plurality of unidirectional Branch Tasks simultaneously respectively, and the node of visiting is carried out part Euclidean distance PED computing; Wherein a PED computing comprises:

According to product matrix in this, the nodal value of this R matrix and former levels is to calculate one with reference to the centre of sphere; And

According to this with reference to the PED of the centre of sphere and last level to calculate the PED of present level.

22. hybrid sphere decoding method as claimed in claim 21 is characterized in that, wherein:

The step of planning unidirectional Branch Tasks comprises, if when a certain level must branch out more than one unidirectional Branch Tasks, then determines the priority of described unidirectional Branch Tasks according to this value with reference to the centre of sphere; And

More near this node with reference to the centre of sphere, task is searched by its unidirectional branch of more preferential execution.

23. hybrid sphere decoding method as claimed in claim 21 is characterized in that, wherein calculating this step with reference to the centre of sphere is according to following formula:

${\tilde{s}}_{m|m+1}=\frac{-1}{r_{m,m}}\{\underset{j=m+1}{\overset{M}{\mathrm{\Σ}}}{r}_{m,j}(s_j-1)+\underset{j=m+1}{\overset{M}{\mathrm{\Σ}}}{r}_{m,j}-{\tilde{y}}_m\}$

Wherein m represents present number of levels, and the superiors' number of levels is M, and orlop is 1;

Should

Be to be this value with reference to the centre of sphere;

r _{M, m}Represent m diagonal element in this R matrix, r _{M, j}Represent (m, j) individual element in this R matrix;

s _jIt is the nodal value for the j level node of visiting in the unidirectional Branch Tasks instantly; And

Be m element value of product matrix in this.

24. hybrid sphere decoding method as claimed in claim 22 is characterized in that, the step of wherein calculating the PED of present level is according to following formula:

${\mathrm{PED}}_m={\mathrm{PED}}_{m+1}+r_{m.m}^2{(s_m-{\tilde{s}}_{m|m+1})}^2$

PED wherein _mBe the PED for present level, PED _M+1Be the PED for last level, r _{M, m}Represent m diagonal element in this R matrix,

Be to be this value with reference to the centre of sphere, and s _mBe to be this rigid decision value with reference to the centre of sphere.

25. hybrid sphere decoding method as claimed in claim 24 is characterized in that, wherein comprises with reference to the step that the centre of sphere produces this rigid decision value according to this:

Move to right one after one input value subtracted 1, produce an operation result;

One 1 QPSK quantizers are provided, and one 2 16QAM quantizers and one 3 64QAM quantizers respectively possess different transformation curves, in order to carry out quantization operations;

Select signal to select to connect this QPSK quantizer according to one, 16QAM quantizer and 64QAM quantizer one of them, this operation result is sent to wherein carries out quantization operations, to produce a quantized result; And

This quantized result moved to left adds 1 after one, to produce an output valve; Wherein this input value be for this with reference to the centre of sphere, this output valve is to be this rigid decision value.

26. hybrid sphere decoding method as claimed in claim 24 is characterized in that, further comprises:

A temporary higher limit;

After the PED of each unidirectional Branch Tasks computing, a part of Euclidean distance PED and this higher limit that calculate are compared;

If this PED less than this higher limit, then continues this unidirectional Branch Tasks;

If this PED greater than this higher limit, then ends this unidirectional Branch Tasks;

27. hybrid sphere decoding method as claimed in claim 26 is characterized in that, further comprises:

When a unidirectional Branch Tasks was finished, a whole Euclidean distance ED and this higher limit that last level is exported compared; And

If, then upgrading the value of this higher limit less than this higher limit, this ED is this ED.

28. hybrid sphere decoding method as claimed in claim 27 further comprises: when all unidirectional Branch Tasks are finished, the node path of this higher limit correspondence is exported this be the maximum similarity solution.

Images