CN101390317A - Improved method and device of multicarrier modulation systems - Google Patents

Abstract

The invention provides a new approach which is better suited to FFT design as applied to multicarrier modulation systems such as OFDM. The signals are scaled so that overflow, rather than being completely avoided, occurs with low probability throughout the IFFT and FFT structures. The size of the error that results from an overflow depends on how overflow is handled in the DSP. To minimize the degradation, overflow should result in saturation of the value at the maximum positive or negative value option. This is equivalent to clipping the signal. Using the new technique, signals within the FFT structure are scaled to balance the effect of clipping and round-off. Clipping may result in comparatively large errors in a few signal values but because of the spreading effect of the FFT and because OFDM systems typically include error coding/correction, system performance depends on the total error or, in other words the total noise power, across all of the FFT outputs rather than on any individual value.

Description

The improving one's methods and installing of multicarrier modulation systems

Technical field

The present invention relates to a kind of method of improving multicarrier modulation systems, especially refer to use transfer process to handle a kind of method of improving multicarrier modulation systems by digital representative signal.The invention further relates to data transmission and reception in multicarrier modulation systems, it uses quadrature conversion that (pair) communicated by letter on a plurality of carriers with permission, but the present invention is not limited to this purposes.

The present invention especially refers to carry out reverse fast fourier transform (IFFT) and fast fourier transform (FFT), and it is used for the multichip carrier communication system, comprising: orthogonal frequency division multitask (OFDM) reflector and receiver.It is relevant for the conversion and the quantification of the fixing point numeral in the FFT/IFFT structure.

Background technology

This multicarrier modulation systems is operated by the sequence data flow point being slit into several constituent data streams arranged side by side, and each these parallel data stream is transmitted on indivedual carriers.Receive each these parallel datas stream at each receiving terminal, and be configured to sequence data stream and corresponding to the sequence data stream that offers reflector.Therefore, in this kind formal system.It is to carry on each time carrier that whole data have only sub-fraction.

When the power spectrum of each parallel data stream overlaps, then can communicate by letter when orthogonal when this grade carrier is common during symbol.This is the direct result that uses the quadrature conversion in reflector and receiver.Use N point conversion (and therefore N time carrier be provided) will increase N during the symbol doubly.

There are various design problems and limit the actual execution of multicarrier modulation systems.

This design problem explanation is because the restriction that counts and handle.Orthogonal frequency division multitask (OFDM) is to use the modulation tech in many new and emerging wide-band communication systems, and these systems comprise: in WLAN (LAN), hd-tv (HDTV) and the 4G system.

This key component in the OFDM reflector is contrary discrete Fourier conversion (IDFT), and is discrete Fourier conversion (DFT) in receiver.Reverse DFT implemented by common reflector and receiver is implemented forward DFT.Usually use certain form of fast fourier transform (FFT), to carry out these conversions.

Computing function that this DSP increased and expressive ability make it be used for the actual execution of OFDM function ideally.Consumer products is usually to cost and power consumption sensitivity, and reason for this reason, and this fixing point DSP method is subjected to preference.Yet the fixing point system has limited dynamic range, causes the relevant issues of noiseization whole (round-off) and count super excessive (overflow).

In the most applications of OFDM, it is important that the complexity of signal processing is minimized.Depend on application can: with cost minimization, reach high data rate or reduce power consumption.When this IFFT in reflector and receiver and FFT are operating as significantly computation intensive operations, then with these function optimization particular importances.Should notice that in this whole specification this DSP (digital signal processor) only uses as an example, it can be replaced with lower device: but any processor, microprocessor, microcontroller, microcomputer, FPGA, CPLD, PLD, ASIC, sea of gates timer, discrete logic chip, this numeral of discrete class or passive component or use any other execution technique that uses with OFDM.

About FFT documents has widely been arranged.Yet the demand in OFDM is with very different in most FFT uses.This main key is not all: the error minimize in OFDM, all should being exported, but not will minimize in the worst error in each output.Design has significant impact for FFT for this.

This FFT is a kind of algorithm that is used to calculate DFT.Have several different algorithms at present, for example: the time is method (decimation in time), frequency ten minutes method (frequency in time), binary system (radix2), the quaternary (radix4) etc. very.Yet it all uses butterfly structure.For binary system, at each butterfly place, two outputs are multiplied each other with two conversion factors, and calculate this equivalance and with poor.

This numeral in FFT is can fixing point or unsteady form representative of putting.The point form of floating has big accuracy usually, but implements in the OFDM computing at fixing point, because its lower cost and lower power consumption are subjected to preference usually.This position presents mark part and sign bit usually in fixing point is implemented, so all digital translation become scope ± 1.In following hypothesis (but not losing its generality) is that this kind situation is because its simplified illustration.

Have two kinds of mechanisms can be reduced in the correctness that each mathematical operation presents: truncation (truncation), for example useization is whole, and saturated (saturation).

Saturated is to take place when this result calculated is outside scope ± 1.This can be only since add operation but not multiplying take place because two its products of the numeral in scope ± 1 are also in scope ± 1.This saturated accurate effect depends on details how to implement addition.This importantly designs this FFT so that will be set at 1 greater than+1 value, and will be set at-1 less than-1 value.In other words, this signal by " brachymemma " (clipped).Some arithmetic operators are for example: 2 complementation can cause " circulation " (wrap-around), so that be transformed into negative greater than 1 numeral, and be transformed into positive number less than 1 number.A kind of mode that saturated probability is minimized or removes is: with this digital translation so that it is in ± 1 scope.

Relative, truncation for example round-off error only since multiplication but not addition take place.When two b-bit digital multiplied each other, this result was that the 2b position is long, and least important b position is cast out.This causes little error at each butterfly shape place.This error size with respect to wanting signal depend on want the size of signal, therefore, for relative error is minimized, this signal should should be converted in scope ± 1 only may be big.Therefore, saturated probability is minimized with will be owing to the error minimize of for example changing whole truncation between exist and conflict.

The exact value of this signal in FFT is unknown, but can calculate the statistical property of this signal.So this design problem is the design conversion, so that under the Distribution Statistics situation of given this signal, for example between round-off error the suitableeest substituting for can be arranged in brachymemma and truncation.

This FFT design conventional method is with conversion of signals, so that can avoid super excessive.For the ofdm system that uses big FFT and fixing point computing,, then need big character length if want this to change the performance performance that whole error reduces this system not significantly.Conversion is distributed in the whole FFT structure usually, because this can reduce the group effect of this round-off error.In order to remove super excessive probability in binary arithmetic operation fully, after each butterfly form class, digital translation must be become 0.5 times.Another kind avoids super excessive method to be to use module floating-point (BFP) conversion.It is adjusted at FFT conversion at different levels according to data, so that can not take place super excessive for given input data.Yet single large-signal sample can cause conversion, and it causes big rounding error in all other signal values.For avoiding this problem, some researcher's suggestions are called the compromising method of restraining BFP, and use different conversion factors for the different sections of data.

Purpose of the present invention is at least one part that alleviates one or more problem of prior art, or a kind of alternative method is provided at least.Refer to that more specifically embodiments of the invention attempt providing a kind of multicarrier modulation communication system, it provides to be chosen among the DSP and uses than the low number position, and is used for the IFFT and the FFT operation of ofdm system.

Summary of the invention

The invention provides a kind of method, it is launched in multicarrier modulation systems and receives data, wherein changes allowing to use the communication of a plurality of carriers, and the method comprises: allow the signal of predetermined number saturated in this reflector and receiver architecture.

This launches the method with received signal in this multicarrier modulation systems, this system has reflector and receiver, and it respectively has input, computing module and output; This reflector and receiver computing module one of at least has: a plurality of calculation stage that are connected, be used for calculating at the enterprising line number word of input received signal, and at least one conversion partly is used for substantially active window and keeps numerical calculation, and provides according to the output of this numerical calculation from deriving at the input received signal; The method comprises: by with the conversion in both or the arbitrary computing module of reflector and receiver, and allow the saturated default remote probability of this signal to surpass active window.Can in a plurality of calculation stage of preestablishing of this computing module, allow saturated.

The method is preferable to be used for orthogonal frequency division multitask (OFDM) and to use together, its can be included in the receiver error detecting with revise step, with essence compensation saturation signal.

This computing module is preferable to be implemented according to digital signal processor (DSP).

This computing module can be at least partly be made of following transformational structure, for example: and fast fourier transform (FFT) and reverse fast fourier transform (IFFT), and the method comprises permission predetermined number saturated in this FFT and IFFT structure or one.

The method can comprise the signal by plural number (complex) representative, and has calculation stage, it passes through the statistical property control transformation of basis signal in computing module with predetermined manner, and implements the arithmetic operator of this representative plural number, saturated being restricted that wherein this allowed.

The real number of this signal and imaginary part can have Gauss or approximate Gaussian distribution.

Select real number and the imaginary part and the headroom (headroom) of this signal with Gauss or approximate Gaussian distribution, it is: maximum quantization standard is to the ratio of gaussian signal standard deviation, so that the suitableeest signal distortion to be provided.

The method can have the real number and the imaginary part of signal, and it has Gauss or approximate Gaussian distribution and headroom, is the ratio of maximum quantization standard to the gaussian signal standard deviation, and it is selected so that the suitableeest signal distortion to be provided.

In addition, the present invention provides reflector and receiver accordingly, and it respectively comprises device and as above-mentioned and according to digital signal processor (DSP), can with method emission of the present invention with receive data.

Can produce two kinds of pattern errors owing to implement the fixing point computing: brachymemma, its numeral that causes being produced when addition the time is greater than or less than possible range; And truncation round-off error for example, when casting out in this product minimum value position, it takes place.

The present invention also provides a kind of method to launch in multicarrier modulation systems and receives data, this system has reflector and receiver is used for launching severally and received signal, this reflector and receiver have: a plurality of inputs, respectively be used to receive input digit, it includes about the real number of signal some or imaginary number; Computing module is used for that institute is received a plurality of input digits and implements digital translation; And a plurality of outputs, respectively being used to provide the output that comprises real number and imaginary number numeral, it is derived by input digit according to digital translation; Computing module, it is implemented digital translation and comprises a plurality of relevant calculation stage, be used for by implementing digital operation on the numeral that a part of relevant input digit of signal derives therewith, and this computing module comprises that more at least one conversion partly, be used for one or more numeral being implemented conversion, and in all digital translation, numeral remained in fact in the predetermined scope of active window at computing module; The method comprises conversion, and it is partly changed by at least one conversion, and allows the saturated scope that surpasses active window of the default remote probability of numeral in this computing module, and this saturated numeral is truncated to the scope of active window.

The invention provides a kind of new method, its preferable FFT of being applicable to design and be applied to for example be the multicarrier modulation systems of OFDM.With this conversion of signals is not excessively to avoid fully super, but it is taken place with remote probability in IFFT and FFT structure.This by the super excessive size that derives error depend in DSP, how to handle super excessive.For this degeneration is minimized, this super excessive should when maximum plus or minus is selected, cause being worth saturated, this equals the signal brachymemma.Use new technology, will be in the conversion of signals in this FFT structure with this brachymemma of balance effect whole with change.This brachymemma meeting causes error bigger in some signal values, but because the diffusion effect of FFT, and because ofdm system typically comprises error coding/correction, whole error is depended in the performance of the performance of this system, promptly depends on to stride all FFT outputs but not whole noise power on any individual values.

Allow owing to saturated brachymemma in the present invention, but must be to take place with the control of changing according to the statistical property of signal.The present invention relates in the FFT/IFFT structure the fixing point digital translation, its mode is to consider the statistical property of signal and use conversion factor, its truncation error and truncation for example between round-off error to the suitableeest balance to OFDM.The real number of many these signals of point and imaginary part have Gauss or approximate Gaussian distribution in FFT/IFFT.Put that this number is converted so that be fit in this scope in it at these, this scope shows Gaussian Profile for the suitableeest.

This normal design process that is used for FFT/IFFT is with digital translation so that super excessive generation never.This is usually by fixed conversion or module converts and reach.Yet, must be quite big for fear of super excessive this quantization step, and this causes for example round-off error of inessential big truncation.Though do not offer some clarification in the literature, this implicit design standard FFT for this reason should be designed to: the worst error quantity in any output is minimized.

This is not the enforcement of the suitableeest FFT/IFFT for OFDM.In OFDM the final measurement of its performance performance be generally the bit error rate in received signal under the given computation complexity.Because FFT handles in receiver and use error coding, this depends on the statistical property of error in all outputs of FFT/IFFT again, but not in arbitrary output the maximum of error.

The present invention allows: handle the minimal error rate of reaching for given the counting of institute, or the processing that will count minimizes for given maximum error rate.By reducing required bits number, it also allows new structure, for example presents two 8 bit digital in 16 words (word) length, therefore reduces and carries out the required operation times of this FFT/IFFT.

The present invention also can be used in: spike to average power reduce technology, be used among the OFDM in the FFT/IFFT computing in the pulse mitigation technique, extra.The present invention also can comprise the conversion of using oversized dimensions and filter so that be easier to.It also can comprise using to have some inputs and/or to be output as zero conversion, for example " orthogonal frequency division multiple access " (OFDMA) in, or be used in the ofdm system, a plurality of antennas that its use comprises a plurality of inputs and a plurality of output systems.

Below by describing embodiment in detail and, the present invention further being illustrated with reference to appended graphic.

Description of drawings

Fig. 1 has the diagram that variance is 1 gaussian variable 100 samples;

Fig. 2 has the diagram that variance is the probability density of 1 gaussian variable 100,000 samples;

Fig. 3 has the diagram that standard deviation is 1 and 17 accurate gaussian random parameter of quantization;

Fig. 4 is used for the diagram of gaussian random parameter brachymemma to the probability of headroom;

Fig. 5 is used to have the diagram of each brachymemma of Gaussian Profile signal to the headroom average energy;

Fig. 6 shows and to be used for the accurate number in variable quantization position, the gaussian random parameter is quantized error to headroom;

Fig. 7 is the histogram in the sample error power that is used for 16 ADC and 16dB headroom;

Fig. 8 is for using 16 to be used for the complimentary cumulative distribution table of error in the gaussian signal quantification with the 16dB headroom;

Fig. 9 is the histogram in the sample error power that is used for 16 ADC and 12.5dB headroom;

Figure 10 is for using 16 to be used for the complementary cumulative distribution table of error power in the gaussian signal quantification with the 12.5dB headroom;

Figure 11 is the calcspar that the typical case simplifies the OFDM receiver;

Figure 12 be used for Gaussian Profile, be used to have the ofdm system of 64 carriers and be used to have 2048 carriers ofdm system be used for 16 and complementary cumulative distribution table with 12.5dB headroom;

Figure 13 (a) is the calcspar that the typical case simplifies the OFDM reflector;

Figure 13 (b) is the calcspar that the typical case simplifies the OFDM receiver;

Figure 14 (a) is the butterfly shape structure of 8 binary system FFT;

It shows because the noise that round-off error caused Figure 14 (b) for butterfly;

Figure 14 (c) is for being used in the model of each the FFT level in the Matlab simulation;

It illustrates that this is used for the complementary cumulative distribution of 4QAM and the real number part of the IFFT structure signal of 64 IFFT and Binary Conversion to Figure 15 (a) for synoptic diagram;

It illustrates that this is used for the complementary cumulative distribution of 64QAM and the real number part of the IFFT structure signal of 64 IFFT and Binary Conversion to Figure 15 (b) for synoptic diagram;

Figure 16 is the diagram of headroom, and it gives the minimum average B configuration square error for the accurate number of gaussian signal vs quantization;

It is used in one embodiment of the invention receiver FFT of evaluate application in digital signal processor to Figure 17 for a kind of model;

What Figure 18 illustrated one embodiment of the invention that this is applied to digital signal processor has the butterfly shape structure of conversion factor from binary system FFT;

Figure 19 illustrates that this is used to change FFT design, 16 accuracy and N=64 in the every sample standardization MSE vs headroom to the ADC input;

Figure 20 illustrates that this is used for

Conversion, 8 accuracy and change N in every sample standardization MSE vs headroom to the ADC input;

Figure 21 illustrates that this is used for 1/2 conversion, 8 accuracy and changes N in the every sample standardization MSE vs headroom to the ADC input;

Figure 22 illustrates that this is used for module floating-point, 8 accuracy and changes N in the every sample standardization MSE vs headroom to the ADC input;

Figure 23 illustrates that this is used for 0.5 flexible strategy, 16 fixing points, N=64 and hr=14,16,18 and the complementary cumulative distribution of every symbol M SE of 20dB;

Figure 24 illustrates that this is used for alternately flexible strategy 0.5,1,16 fixing points, N=64 and hr=14,16,18 and the complementary cumulative distribution of every symbol M SE of 20dB;

Figure 25 explanation is in the comparison of two ADSP-BF533 computings of 256-point binary system plural number time ten fens FFT;

Figure 26 illustrates as the function of the input signal power that is used for IFFT that at the signal of IFFT output to noise ratio (SNR), this IFFT is designed to have in the fixed conversion factors at different levels

With 8 fixing point accuracy;

Figure 27 illustrates as the function of the input signal power that is used for IFFT that at the signal of IFFT output to noise ratio (SNR), this IFFT is designed to have in the fixed conversion factors at different levels

64 IFFT and 64QAM modulation;

Figure 28 illustrates as the function of the input signal power that is used for IFFT that at the signal of IFFT output to noise ratio (SNR), this IFFT is designed to have in the fixed conversion factors 0.5 that replace at different levels and 1 and 8 fixing point accuracy;

Figure 29 illustrates as the function of the input signal power that is used for IFFT that at the signal of IFFT output to noise ratio (SNR), this IFFT is designed to have in the fixed conversion factors 0.5 at different levels and 8 fixing point accuracy; And

Figure 30 illustrates as the function of the input signal power that is used for IFFT that at the signal of IFFT output to noise ratio (SNR), this IFFT is designed in use module float points at different levels and 8 fixing point accuracy.

Embodiment

Embodiments of the invention are with reference to orthogonal frequency division multitask (OFDM) system specialization, and it is a widely used multichip carrier system, wherein uses the discrete Fourier conversion to implement modulation and to separate modulation.

In order to understand and to analyze new technology, need the background information of several different themes, it is:

Signal characteristic:

OFDM, and especially at the OFDM noise effect;

The effect of FFT structure and fixing point computing and the limited accuracy that counts;

Signal statistics in OFDM among the IFFT/FFT; And

The quantification of gaussian signal.

Background is described in the following description, and the analog result that is provided for new technology, and with traditional technology relatively.This also illustrates the actual enforcement of new technology on fixing point DSP.

The real number and the imaginary part of this ofdm signal have approximate Gaussian distribution.This cause high spike to mean value than (PARA, peak to average ratio).This with the big the most common relevant problem of PAPR is: the design with reflector output amplifier of great dynamic range.Yet it also means the quantizing range that must carefully be chosen in the receiver input signal.

Use the finite population position to represent in force: the signal in reflector and receiver numeral part.This restriction can be set in the following manner: the simulation in receiver front end-to the accuracy of-digital translation, or the number by uses position, and to represent the numeral in fixing point Digital Signal Processing (DSP) enforcement.

Fig. 1 shows the sample waveform of Gaussian random variable.For this example, the variance of this random variable is 1.Fig. 2 shows that the probability of same signal value distributes.

Though most value is close to 0 and concentrates, yet it has long-tail, and for for example shown in Fig. 2, some is worth near 5.Take place with remote probability though have the value of high amplitude, it still can take place with little and limited probability.In the content of high data rate OFDM receiver, its per second can receive millions of samples, so the remote probability incident is quite taking place during the short time.For example, every several minutes faults TV recipient once can't make us accepting, therefore must this receiver of design, even when considerably less high levels sample takes place and can suitably carry out.

Two kinds of pattern errors when quantizing, ofdm signal can take place: truncation error and quantization error.Truncation error is to be when this value outside this " analog-to-digital converter " active window scope (ADC) (it also is called saturated or overload region) and must take place when quantizing with minimum bit is accurate maximum.This quantization error be because: the analog signal in this scope at quantizer (being commonly referred to the granulomere) takes place when approximate with one of this quantization almost finite number.

When ofdm signal is quantized, exist between for the brachymemma of given number position and quantizing noise and substitute for.Because Gaussian Profile has very long afterbody, therefore indivedual brachymemma incidents can cause very large error, though the considerably less generation of this equal error.Fig. 3 be presented at-2 and+2 have 17 gaussian signals in the quantization Barebone, and the sample outside the granulomere can be by brachymemma." headroom " is the ratio of maximum quantization standard to the gaussian signal standard deviation in this manual.Usually convenient with dB measurement headroom.Therefore headroom is 20log10 (2/1) dB or 6dB in Fig. 3.Distance between this quantization standard is represented with d.Therefore, for system with k quantization standard:

h _r＝(k-1)d/2 (1)

Because because the error that quantizes is shown as noise in communication system, it is called " brachymemma noise " and " quantizing noise ".This quantification has very different statistical properties with the brachymemma noise.In this ofdm system, this brachymemma is unusual rare event, and most input sample can be subjected to quantizing noise.This quantizing noise is a stochastic variable.The characteristic of the analog signal that receives is depended in the distribution of this quantizing noise, but for ofdm signal, its will-d/2 and+roughly evenly distribute between d/2.This result is: when these signals evenly distributed between these quantization standards, this was owing to the Mean Square Error that quantizes is:

E〔n _q ²〕＝d ²/12 (2)

N wherein _qBe the number of quantization standard, and the distance between d=quantization standard.

For given headroom, increase the accurate number in position and can reduce d and therefore reduce Mean Square Error.If therefore the accurate number in position doubles, then can reduce 6dB according to this quantizing noise of formula (2).Similarly, if the accurate number in position is kept constant and headroom h _rDouble (that is, increasing 6dB), then quantizing noise can increase 6dB.Therefore increase for each dB in headroom, this quantizing noise also increases 1dB.Therefore, the power and the distribution of quantizing noise can be described simply.

The brachymemma noise has more complex features.The shape of this brachymemma noise is a pulsed.The pulse of this brachymemma noise does not often take place, and quantizes big many amplitudes but can have.When this headroom increased, the probability of this brachymemma reduced.

Fig. 4 shows probability: this sample can be punctured into the function of headroom.This is converted to numerical expression.The increase that headroom is quite little can reduce brachymemma probability several magnitude.For example: headroom from 10dB increase to 14dB then the probability of brachymemma from 10 ^-3Be reduced to 10 ^-6Fig. 5 shows how the average energy of this each brachymemma changes with headroom.The increase of headroom only reduces the average energy of each brachymemma lentamente.

The suitableeest headroom of decision is not a simple thing, and it depends on several aspects of whole system design.This quantizing noise and brachymemma noise sum change with headroom.When headroom increased, this quantizing noise increased, and the brachymemma noise reduces.

Fig. 6 shows the Mean Square Error that quantizes for the Gaussian random variable analog result for (vs) headroom, and has the resolution that changes bits number.There is a headroom in its demonstration for bits number, and it gives minimum average B configuration square error (MMSE).This can be envisioned for the suitableeest headroom, yet in most situation, this suitableeest headroom in ofdm system is greater than being used for several dB of MMSE headroom.

, because to the right side of minimum value and the error profile of left quantized result, considering the situation of 16 bit representations, gaussian signal and 16dB headroom and 12.5dB headroom are simulated in order to seek, to watch error profile in the minimum value both sides.

Fig. 7 shows the histogram that is used for 16 ADC and 16dB result.This error power is normalized to averaged signal level, so that this error power is to measure noise to signal power.But the brachymemma of negligible number is arranged with this headroom quantity, and the distribution of sample between the quantization standard is represented in the distribution of this error.Be not easy to find out the frequency of " bad sample " in this histogram: promptly it has the mean error of being higher than power person.The figure of this complementary cumulative distribution (CCD) makes that the generation of remote probability incident of this higher-wattage is more obvious.

Fig. 8 shows the CCD be used for identical data, this figure-there is rapid decline at the 82dB place.This probability of error power that surpasses this standard in sample is very low.Fig. 9 shows that with Figure 10 working as headroom is reduced to the identical result of 12.5dB.

Difference between Fig. 7 and Fig. 9 is very unobvious.Yet the effect that this cumulative distribution in Figure 10 makes headroom reduce is more obvious.Sample is arranged now, and it has error power and approaches 0dB.Equal the average power of this signal in 0dB error power originally.

Figure among Figure 10 has two zoness of different.This on the left side zone is the result of quantizing noise.This zone, figure the right is the result of brachymemma noise.This probability value of reduction place fast depends on interval d between the quantization standard.Therefore, this can the be descended dB of nidus of the number that increases or reduce the quantization standard moves.The accurate probability that shows this brachymemma in the position on this " plateau " therefore increases headroom and can make this plateau take place than remote probability.

In order to understand due substituting between this mean square noise and brachymemma probability, this OFDM receiver need be done whole consideration.Figure 11 shows the calcspar of typical OFDM receiver.This gaussian signal sampled and that quantize comes across the input of series connection-extremely-transducer in parallel.The critical nature of this ofdm signal is in fact also uncorrelated at this equal samples of this point.Therefore, brachymemma incident is also in fact also uncorrelated.This sample is by the true of brachymemma and change this sample not significantly by the probability of brachymemma.This to series connection-to the input of-transducer in parallel with export from the data of error decoder between be several modules, it has average effect.These are FFT and error decoder.

The analysis of effect is closely related with the analysis of impulsive noise in OFDM between this brachymemma quantizing noise.This document on impulsive noise shows that this effect can predict by using " noise bucket " idea.Through find in theory with this error performance performance in fact can prediction exactly by considering following factor: the noise sum is gone up in all inputs to FFT that are used for conversion operations, and Energy distribution has little effect between this FFT input.The average effect of this error coding depends on this system.In digital video is play (DVB), there is no at intersymbol staggered, so this error decoder round-off error in FFT operation only.In some other systems, exist staggered at intersymbol.

Figure 12 is presented to FFT is arranged side by side and imports upward the noise averaged result.Herein the x axle by noise average on the picture of 64 samples and 2048 samples to the signal ratio.This shows the effect respectively use noise in 64 and 2048 carrier systems.When using 64 carriers, the probability of this brachymemma in one or more input increases about 64 times, but average noise power reduces by 64 times (18dB).This carrier number is increased to 2048 strengthen this average effect.For this example, if headroom increases, then this whole system performance can improve, so that the probability of this brachymemma reduces.This with adequate remedy in the medium and small increase of quantizing noise.

Therefore, though this can advise that 12.5dB is near the suitableeest headroom being used for the Mean Square Error result shown in Fig. 6, when considering whole OFDM system, this higher headroom gives preferable general performance.

The whole and saturated effect of change during this fixing point digital signal processor (DSP) at IFFT and FFT is implemented, be to be shown as disclosed embodiment with reference to orthogonal frequency division multitask (OFDM), this OFDM is that a kind of modulation tech is used in many emerging wide-band communication systems, and it comprises: WLAN, HDTV and 4G system.

This key component in the OFDM reflector is contrary fast fourier transform (IFFT), and is FFT in receiver.The size of this conversion depends on application, and its scope is: in some WLAN (LAN) standard 64 to some play 8192 points in digital television systems.Rated output that this DSP increased and expressive ability make it become the ideal candidates person who is used for the actual execution of OFDM function.Real consumption person's product is usually for cost and power consumption sensitivity, and reason for this reason, and this fixing point DSP method is subjected to preference.Yet this fixing point system has limited dynamic range, the whole noise of its causing and the super excessive relevant issues that count.

This to the method for FFT design for conversion of signals being avoided super excessive.For the ofdm system that uses big FFT with the fixing point algorithm, if this round-off error can not make significantly degeneration of system's performance, character length greatly then.In situation worst, for N-point FFT, the N that this maximum in the FFT output valve is a maximum input level doubly so that its by depend between input and output the N of palpus be that conversion factor is changed.This conversion under normal circumstances is to be distributed in the FFT structure, because this reduces the group effect of round-off error.In order to remove super excessive possibility in binary system is implemented fully, these numerals must be converted 1/2 times after each butterfly level.Avoid super excessive another kind of method for using the conversion of module float point.This is adjusted at FFT conversion at different levels according to data so that for given input data can not take place super excessive.

Yet, the invention provides a kind of new method, its preferable FFT design that is suitable for OFDM.In OFDM, this should be striden the statistics optimization of all Mean Square Errors of all outputs of receiver FFT, but not with spike error minimize in arbitrary value.This is because of the diffusion effect of FFT and because OFDM under normal circumstances uses the error correction or the detecting of certain form, therefore, as long as the error in other output is little, can revise big error in some FFT output.

These the suitableeest statistics can be super excessive with certain low probability generation but not by avoiding super excessive reaching by allowing.It is super excessive how this handles in DSP owing to the super excessive size that produces error depends on.For degeneration is minimized, this super excessive should causing is selected the saturated of duration in maximum plus or minus value.This equals this signal brachymemma.

Use this new technology conversion of signals in the FFT structure can be minimized Mean Square Error.If this signal is too big, then the probability of this brachymemma increases.If this signal is too little, then this round-off error becomes bigger with respect to signal level.The statistical property of signals at different levels in this IFFT/FFT structure is depended in this suitableeest conversion.

The result shows can use new technology design IFFT/FFT structure, and it is used for given fixing point accuracy, and its performance performance significantly surmounts the conventional art that this is avoided brachymemma.This attainable improvement can increase with the increase of FFT size.This system complexity is similar to the complexity of conventional fixed conversion, and is lower than the complexity of the unsteady point of square.Can also determine the result of the actual execution of above-mentioned design on fixing point DSP.

This super excessive problem that counts is even more important for the ofdm system that uses big FFT.If this FFT design has conversion, so that there do not have input condition to cause to be super excessive, then this round-off error becomes extremely important, and can cause being used for signal unacceptable " signal is to noise ratio ".

Another kind of mode be allow to count super excessive with certain low probability generation.This will cause some noise in signal.This noise level depends on how to handle super excessively in DSP, and take place super excessive in the FFT butterfly structure.In order to reduce degeneration, this super excessive saturated in the value of whole plus or minus value of should causing.Promptly be equivalent to this signal brachymemma.The brachymemma of this output stage only influences the output at single butterfly place.Cause these all outputs of depending on butterfly to be affected in the brachymemma of level early.

Figure 13 (a) and Figure 13 (b) respectively show the simplification calcspar of OFDM reflector and receiver.Some unessential squares are omitted in this FFT design context.The key component of this reflector is the IFFT module, and the key component of receiver is the FFT module.

These data that are launched are at first done series connection-to-conversion in parallel, and mapping is to plural number then, and its representative is from the value of signal constellation (in digital modulation), and this constellation is used as for example in 16 signal constellation (in digital modulation)s shown in Figure 13 (a).Each input is controlled at the signal of a frequency.This IFFT carries out the multitask of modulation and this grade carrier of each time carrier in a computing.Then this signal is done in parallel-become analog signal to-serial conversion, and on the modulation tremendously high frequency carrier.The details how this reaches and these modules configured order can be looked actual needs and be changed.

Changed downwards at this signal of receiver, and before inputing to FFT, convert digital signal to and do series connection-to-conversion in parallel.This FFT carries out separating modulation and separating multitask of each time carrier.If this channel does not cause noise or distortion, then the output of this receiver FFT equals the input of reflector IFFT.Actually, this channel can make distorted signals and add noise.Distortion in this passage has the effect with the phase place of each time carrier and amplitude change.This can be by the correction of single tap equalizer, its change of compensation of phase and amplitude with each output of a CM FFT.Then these data are inputed to error decoder.Under normal circumstances OFDM is used in error correction with the error code translator, because can have very large error rate at declining carrier significantly.The existence of this error code translator means, so that this global error rate does not depend on the noise in arbitrary FFT output, but depends on the statistics of striding all FFT output noises.This is closely related with " noise bucket " effect, and observes the impulsive noise that is used for OFDM.

FFT is the algorithm that is used to calculate DFT.Have some algorithms of different at present, for example: the time is sampling method, frequency ten minutes sampling method very, binary law and quaternary method etc.Yet it all uses butterfly structure.Figure 14 (a) and Figure 14 (b) show: be used for the butterfly structure of 8 binary system FFT, it shows because the truncation noise of round-off error for example with indivedual butterflies.This fft algorithm is according to butterfly structure.

Each butterfly relates to multiplication and addition.When using fixing point to count, this changes whole and super excessive may the generation.This contingent accurate point depends on the details of execution in butterfly.Between this grade, also can use conversion factor.This Mathworks's

In the simulation, the structure among use Figure 14 (c) is with modellings at different levels.In outputs at different levels with quantitative modelization.The unlimited accuracy of hypothesis at different levels.Each butterfly two inputs are multiplied each other with two conversion factors, and calculate this product and with poor.

This butterfly structure of 8 binary system FFT with conversion factor of Figure 14 (a) is applied to digital signal processor.This processor that embodiment had with digital signal processor (DSP) of fixed number processing position has: be used for the input of received signal, the output that is used for the computing module of processing signals and is used to export this treated signal.This computing module comprises: fast fourier transform (FFT) or contrary fast fourier transform (IFFT) structure, it is designed to allow predetermined amount saturated of the input received signal in FFT or IFFT structure.

Design the computing module of this digital signal processor, with the real number of processing signals partly and imaginary part, it has Gauss or approximate Gaussian distribution, and this headroom for this reason the maximum quantization standard of gaussian signal to the ratio of standard deviation, its selected to provide because E (n _q ²)=d ²/ 12 given quantifications preestablish Mean Square Error (E), and n _qBe the number of quantization standard, d is the distance between the quantization standard.

Digital signal processor has a plurality of levels in computing module, its computing on the cumulative distribution of real number with Gauss or approximate Gaussian distribution signal and imaginary part.Shown in Figure 18, this conversion in each calculation stage is that what had as shown is converted to 1/2 on each input partly, and the not conversion (that is, being converted to 1) on the right side.Design this computing module handling real number and the imaginary number composition of signal in the active window, and this signal is quantized to one of finite population quantization standard, is used for according to handling by predefined a plurality of butterfly levels conversion of the calculation stage of computing module.

In order to understand the performance of whole system performance, must consider the statistics of signal distortion.Because error correction in many systems and detecting be with symbol to symbol serve as the basis on symbol, operate, this signal distortion in each symbol is for important.High distortion in a symbol is can't be with the low distortion compensation in another symbol.

Statistics optimization will be used in order to design FFT or IFFT, the statistical nature of signal in whole reflector FFT and receiver IFFT need be determined in the distortion of fixing point computing given number bit sign.

This depends on signal constellation (in digital modulation) to reflector IFFT input signal in reflector.If use 4QAM, then each plural number input has one of four probable value ± 1 ± j (may multiply by a scale factor).For 16QAM 16 possibility input values are arranged, and 64 probable values are arranged for 64QAM.

Signal complex random variable in ofdm system in IFFT and FFT.This changes the probability distribution that whole effect with brachymemma depends on signal in FFT and IFFT structure.Figure 15 (a) and Figure 15 (b) are various level real numbers complementary cumulative distribution partly in reflector IFFT structure.It does not use conversion, therefore in mean square value growth of working as its signal of signal plus of first prime at different levels.Use the expression of this form and make its afterbody that relatively distributes easily, this is important in the design of using brachymemma, because must guarantee that the probability of being clipped can be too not high.Figure 15 (a) shows the distribution that is used for 4QAM.For three levels at first, its output does not have Gaussian Profile significantly, yet the 4th level, the afterbody of its distribution is very near Gaussian Profile, and it distributes and the difference of Gaussian Profile can be ignored after the fourth stage.Figure 15 (b) shows the distribution that is used for 64QAM.Because the possible input value of greater number, this signal is in early the level approximate Gaussian distribution that becomes.

Another vital point of Figure 15 (a) and Figure 15 (b) is: when arbitrary grade input signal was Gaussian Profile, it was output as the signal with Gaussian Profile that 3dB power increases.This promptly is that average power increases twice and amplitude increases

Doubly.Though please note that peak signal increases twice and mean amplitude of tide increases

Doubly.This means in use 0.5 conversion factor FFT design " tradition " methods at different levels, cause reducing about 3dB in average powers at different levels.

Most ofdm system use 64 or more times carrier, and through finding that this is in close proximity to Gaussian Profile at the real number of the signal of reflector IFFT output and the distribution of imaginary part.

Analysis at receiver FFT signal is more complicated, because it depends on communication port.In the situation of " ideal " passage, can not cause distortion and can not increase noise.This signal distributions in receiver IFFT but its reversed in order identical with person in the reflector.In other words, this is a Gaussian Profile at the signal of level early, and obtains centrifugal pump in output from signal constellation (in digital modulation).Even in actual applications good wireless electric channel can cause the phase change to receive time carrier, it can distribute to the scope of+π equably at-π in fact, and therefore in the signal distributions of FFT output be continuously but not disperse.Though its signal distributions depends on passage and is not Gaussian Profile usually that its character often takes place for the big rare and little value of value.

Partly have essence Gaussian Profile at least owing to understand the real number of many these signals of point in IFFT and FFT, therefore can determine the optimal doseization of gaussian signal with imaginary part.Therefore, how in this way this key issue is: representation signal and with the statistics optimization of signal distortion.This problem in OFDM in the FFT design content before not analyzed as yet mistake, yet if allow brachymemma and change wholely, the optimal dose problem of this problem and gaussian signal is very closely related.For the position of institute's given number, then at big quantization step, it causes the brachymemma and big round-off error of little probability, and the small quantization step, it has to exist between the brachymemma of big probability and the little round-off error and substitutes for.This problem is in simulation-to the document of-digital quantizer content discussion was arranged once.

The problem that this Gaussian random variable quantizes general in the past be used for the simulation of OFDM-broad research is all arranged to-digital quantizer (ADC) design content.These results can directly be applied to fixing point in counting saturated with change whole problem.In order to describe the character of this uniform quantizer, must define three relevant parameters.These three parameters are: d, and it is the distance between the quantization standard; K, it is the number of quantization standard; And h _r, it is the headroom of maximum quantization standard.H in this manual _rBe standardized as root mean square (RMS) value of signal.When this quantized for 0 symmetry, the pass of these three parameters was

h _r＝(k-1)d/2 (3)

Quantize to have very different statistical properties with the brachymemma noise.In ofdm system, this is punctured into considerably less event, but most sample can be subjected to quantizing noise.Usually this quantizing noise is evenly distribution between-d/2 and+d/2.For given h _rIf, k is doubled, then quantizing noise will reduce about 6dB.Similarly, if keep k constant and with h _rIncrease 6dB, then quantizing noise can increase 6dB.The brachymemma noise has more complex characteristics.Its form is a pulsed.The pulse of this brachymemma noise seldom takes place, but that its amplitude can quantize noise is a lot of greatly.Work as h _rDuring increase, the probability of this brachymemma reduces.

Figure 16 shows the function of minimum average B configuration square error (MMSE) as headroom, is used to change the resolution of number position.This " headroom " is defined as: maximum quantization standard is for the ratio of signal real number part root mean square (rms) value.Figure 10 shows as headroom h _rBe 12.5dB and k=216, that is, use 16 to represent the complementary cumulative distribution that is used for when respectively being worth the gaussian variable quantization error.This figure has two zoness of different.The zone of this on the left side is the result of quantizing noise.This zone on the right is the result of brachymemma noise.This noise in interregional conversion depends on d to signal level.The accurate probability that shows brachymemma in the position on this " plateau " therefore increases h _rCan make this plateau to take place than remote probability.Yet, because Gaussian Profile endless afterbody, always it has the brachymemma of non-zero probability, no matter and parameter value why this figure has identical basic configuration.

Get back to Fig. 6 now, its demonstration: be used for the value of the Mean Square Error of gaussian signal to the accurate number of the used quantization of value of (vs) 8,10,12,14 and 16 bit resolutions.Its demonstration has a headroom for each bits number, and it gives minimum average B configuration square error (MMSE).Yet this MMSE headroom also need not give optimizer system performance in OFDM.Also need consider the effect of this remote probability, strong noise and brachymemma incident.Its demonstration as can be seen from Figure 6: when use 8 or more multidigit when representing gaussian variable, this with the minimized headroom of every sample mean square error be between 10 and 15dB between.

Also usefully know that how responsive this Mean Square Error be for actual clearance.In other words, whether can significantly increase Mean Square Error from smallest point is mobile a little.

Fig. 6 shows that when bits number increased, then the minimum average B configuration square error reduced.This minimum value is the highstrung function of headroom, and especially this headroom is when following.The brachymemma of too many sample of signal palpus is arranged in this case.Relative, if this headroom is too big, then this Mean Square Error can't increase very soon.The effect of this too big headroom is to increase a little in round-off error.

In order to design preferable OFDM reflector and receiver, then need to consider to be used for the fixed conversion factor of transformational structure, and with the signal distortion optimization.This performance be and the conversion of using the module float point processing relatively, and conversion is that to depend on whether this particular data can cause super excessive and in enforcements at different levels.

Now above information is combined to show: how can design this and have, use and be used for specific OFDM through controlling the FFT of brachymemma.

Below sum up the key results of first previous paragraphs:

This performance performance that is used for the OFDM symbol is depended on: the statistics that all noises added to symbol;

The signal of many points has Gaussian Profile in this IFFT and FFT;

When gaussian signal was quantized, the headroom that then depends on quantizer exists between truncation error and round-off error substituted for; And

At different levels, for Gauss or approximate Gaussian distribution signal, this signal power doubles in FFT at least, so mean amplitude of tide increases

Doubly.

Signal power among this whole FFT/IFFT is adjusted in the suitableeest design of this conversion and selection, so that put in order the suitableeest compromise of intercropping in the headroom of each point in brachymemma and change.This just when the many aspects that depend on the whole system design, comprising: the power and the statistics in the number of inferior carrier, other source of noise and be used in the bits number of fixing point in presenting.For this suitableeest headroom of gaussian signal usually less times greater than the headroom that is used for MMSE.In order in whole conversion, to keep the suitableeest headroom, should keep constant signal power or approximately constant.

For this binary system is main structure, therefore in the suitableeest conversion factor at different levels is

A kind of time the suitableeest structure a little of easier execution is to change with 1 or 1/2 (simple left dislocation) in the level that replaces on DSP.

Use

Emulation is with the performance performance of check OFDM receiver, and this receiver comprises the FFT that uses new technology to design, and compares with using the conventional fixed conversion and the performance performance of module float point processing.Figure 17 shows the model that is used in the receiver emulation.

Be the effect of design in the whole system performance understand this receiver FFT, must know at this this signal of input of FFT is presented as fixing point form how.When this signal in input has been subjected to the simulation-to-narrow restriction of digital translation (ADC) inevitably, as and if meaningless design can avoid the FFT of brachymemma fully.

This represents with " ADC " module in Figure 17.Yet in reality was implemented, this related to some functions by the high frequency analog-converted for the fundamental frequency numeral.The headroom that in FFT, the brachymemma quantity that takes place is directly depended on ADC.This simulation comprises the ADC effect.Each these binary system bowknot has in the structure shown in Figure 14 (c) in Figure 17.

In this input generation conversion to module, and in the living quantification of its output output, except being used for the 17th figure first order, it does not have conversion, because this equals to change the form of input power and input ADC.This input signal is molded as complex signal with Gauss's real number and imaginary part.In reality was implemented, the form of this ofdm signal that receives depended on passage.This performance performance that is used for Gauss's input is for may the showing of: most passage, give good expression with the relative performance of different designs FFT.

FFT for different designs implements simulation, calculates the MSE that is exported by new technology in two ways: every sample and every symbol.Calculate this " every sample error " and be used for the complex values of each conversion operations in FFT output.Average by every sample MSE that will be used for each conversion operations, and calculate " every error in label ".This second method gives the demonstration of better performance performance in ofdm system, because in most system, this error coding is only operated on symbol with symbolic base, and not staggered at intersymbol.Consider that this is used for the result of four kinds of different FFT designs: in uses at different levels

The new technology of conversion; Use the new technology of 1 and 0.5 conversion at alternate level; Has 0.5 a conversion traditional FFT design at different levels; And the FFT that uses BFT.Also can calculate the independent infringement that causes by ADC.Implement simulation for some values: the change number position b of FFT size, N and fixing point accuracy.

Figure 19 shows the every sample MSE that is used for different FFT designs and is used to use 16 independent ADC of accuracy.This for the independent error of ADC along with headroom increases to about 14dB and descends from 0dB.This result who is used for FFT has analogous shape usually.This minimum value increases 3dB, but this be because to the input of ADC with quantize first butterfly between the power increase of 3dB is arranged.In other words, this minimum value also occurs in the about 14dB headroom that is used for quantized level.This is consistent with the result who is used for the gaussian variable quantification in Fig. 6.On the right side of this minimum value, most figure tilts to rise with 1 gradient.This traditional design in use 0.5 conversions at different levels has the poorest performance, and this is in declines at different levels because of signal power.Two versions of this new technology produce preferable performance, have only little expending for being used alternatingly 0.5 and 1 conversion.

This is used for the BFP performance and shows very great change when headroom changes a little.This result who is used for BFP also changes between simulated operation sharp.This is by very " bad symbol " institute master control of remote probability because of its performance.It has very high " spike is to average specific " to exist symbol, and causes BFP to change to avoid brachymemma with 0.5 in each level.This causes the whole noise of change of high value.For in the medium and small change of headroom, then the performance for this bad symbol can significantly change.This be because if at ADC with the high peak brachymemma, then the module floating-point need lessly be changed.

Figure 20, Figure 21 and Figure 22 show how changing along with inferior carrier number for different designs of MSE.Compare Figure 20 and Figure 21 as can be seen, less degeneration is many along with the N that is increased for its performance of new design.When this variable conversion compensation when avoiding brachymemma, more insensitive for its performance of BFP for headroom.

In OFDM, the statistics of this each symbol also are important.Figure 23 and Figure 24 show that CCD is used for: traditional 0.5 flexible strategy; And new method, it is for h _rValue be used alternatingly 0.5 and 1 flexible strategy, this value just in time be under the value that is used for every sample MMSE with on.These parameters are identical with person shown in Figure 19.In Figure 25 as can be seen, though h _r=14dB causes for the low every symbol M SE of most symbol individual plateau being arranged near 0.05.In other words, work as h _rDuring=14dB, about 5% symbol has every symbol M SE of high value.When its critical value increased, the probability on this plateau descended fast, and for 18 with 20dB its on figure no longer as seen.Yet because the unlimited essence that prolongs of this Gaussian Profile, it always has the plateau.Headroom should be selected so that but this probability relevant with the plateau is the recipient.It is usually less times greater than the headroom that is used for MMSE.Figure 24 shows the result who is used for new method.Please note that this noise power significantly reduces.

Be used for analogue means in order to be presented at the feasibility that DSP upward carries out these new conversion methods, therefore to write The program code of processor.This processor be characterized as DSP and MCU function.This Blackfin processor that is used in this situation is ADSP-BF533, and its characteristic is: a high-speed SRAM and a low-down power consumption until 1512MMAC (per second 1,000,000 phase multiply accumulatings), 1.2Mbits.

These yards are adapted by the present scope of fft algorithm to write, and can be obtained easily by the Blackfin website.The execution of these yards is by the height optimization, and makes full use of the existing Blackfin instruction set support that is used for FFT.Use this " analogue means video signal DSP++IDE (integrated development environment) " with execution and with its debug.

This example with principle of " time is method very " butterfly is: each butterfly has two output results: top and bottom.This top is called a, and this bottom is called b.This conversion factor is W _nThis is used for comprising following formula at the sign indicating number of the embodiment of simulation DIT butterfly:

Top output: O _a=I _a+ W _n.I _b

Bottom output: O _b=I _a-W _n.I _b

All numerals are plural number.This DSP need use discrete multiplication to carry out these calculating.Therefore, it need be broken down into real number part and imaginary part.In with following formula for the sake of clarity, with " O _b" it " a " and " b " partly omission.

W＝W _r+W _i.i

I＝I _r+I _i.i

I.W＝(W _r+W _i.i).(I _r+I _i.i)

＝W _r.I _r-W _i.I _i+(W _i.I _r+W _r.I _i).i

When calculating O _aThe time, calculate two indivedual results: real number and imaginary number

Real number part: O (b)=I (a) _r-+[W _r.I (b) _r-W _i.I (b) _i]

Imaginary part: O (b)=I (a) _i-+[W _i.I (b) _r+ W _r.I (b) _i]

O _bCalculating identical in fact:

Real number part: O (b)=I (a) _r-[W _r.I (b) _r-W _i.I (b) _i]

Imaginary part: O (b)=I (a) _i-[W _i.I (b) _r+ W _r.I (b) _i]

Therefore real number partly has two multiplication and three addition/subtraction.In following shown sign indicating number, on each butterfly level, implement conversion, and show an example.

At first, set the loop that is used for level-3 numbers.The general execution that it calculates for butterfly.Set first loop and be used for half at butterfly numbers at different levels.Set this second multiple circuit (nested loop) and be used for number at each butterfly center line.On the output array, implement to calculate.After will exporting, its storage is used to change purpose divided by 2.In with the loop, read and handle two butterfly data.

I0＝B0；

I2＝B2；

I3=B2; The address of // output array

P0＝P3<<2；

M2=P0; //M2 preserves the compensation of relative line

P0+＝-4；

M0＝P0；

P5＝P5>>1；

R7＝R7>>>1‖I3+＝M2；

M1＝R7；

P3+＝-1；

lsetup(Loop1_strt，Loop1_end)LC0＝P5；

// setting is used for the loop of butterfly number

Loop1_strt：

I1=B3; The address of // conversion factor

R2＝[I2++]；

R3＝[I1++M1]‖R4＝[I3++]；

lsetup(Loop2_strt，Loop2_end)LC1＝P3；

// setting is used for the loop of line number

Loop2_strt：

R5＝R2+|+R4，

R6＝R2-|-R4(ASR)‖R3＝[I1++M1]

‖R4＝[I3++]；

A1＝R3.L*R4.H，

A0＝R3.L*R4.L‖[I0++M2]＝R5‖R2＝[I2++]；

Loop2_end： R4.H＝(A1+＝R3.H*R4.L)，R4.L＝(A0-＝R3.H

*R4.H)

‖I0-＝M0‖[I0]＝R6；

R5＝R2+|+R4，R6＝R2-|-R4(ASR)‖I2+＝M2；

I3+＝M2?‖[I0++M2]＝R5；

Loop1_end：

[I0++]＝R6；

P3+＝1；

P3＝P3<<1；

R0+＝-1；

B1＝B0；

B0＝B2；

B2＝B1；

CC＝R0＝＝0；

If！CC?Jump?Loopfor_m(BP)；

//Loopfor?m.

jump?Esc_mid；

This is the example of execution butterfly, and it implements to move (shift) on each butterfly level.In this example, only move and on a butterfly level, implementing.

cc＝bittst(R0，0)；

if?cc?jump?NO_SHIFT；

I0＝B0；

I2＝B2；

I3=B2; The address of // output array

P0＝P3<<2；

K-M2=P0; //M2 preserves the compensation of relative line

P0+＝-4；

M0＝P0；

P5＝P5>>1；

R7＝R7>>>1‖I3+＝M2；

M1＝R7；

P3+＝-1；

lsetup(Loop1a_strt，Loop1a_end)LC0＝P5；

// setting is used for the loop of butterfly number

Loop1a_strt：

I1=B3; The address of // conversion factor

R2＝[I2++]；

R3＝[I1++M1]‖R4＝[I3++]；

lsetup(Loop2a_strt，Loop2a_end)LCl＝P3；

// setting is used for the loop of line number

Loop2a_strt：

R5＝R2+|+R4，R6＝R2-|-R4(ASR)‖R3＝[I1++M1]

‖R4＝[I3++]；

A1＝R3.L*R4.H，

A0＝R3.L*R4.L‖[I0++M2]＝R5‖R2＝[I2++]；

Loop2a_end：R4.H＝(A1+＝R3.H*R4.L)，

R4.L＝(A0-＝R3.H*R4.H)

‖I0-＝M0‖[I0]＝R6；

R5＝R2+|+R4，R6＝R2-|-R4(ASR) ‖I2+＝M2；

I3+＝M2?‖[I0++M2]＝R5；

Loop1a_end：

[I0++]＝R6；

P3+＝1；

P3＝P3<<1；

R0+＝-1；

B1＝B0；

B0＝B2；

B2＝B1；

CC＝R0＝＝0；

If！CC?Jump?Loopfor_m(BP)；

// be used for the loop of m

NO_SHIFT：

I0＝B0；

I2＝B2；

I3=B2; The address of // output array

P0＝P3<<2；

K-M2=P0; //M2 preserves the compensation of relative line

P0+＝-4；

M0＝P0；

P5＝P5>>1；

R7＝R7>>>1‖I3+＝M2；

M1＝R7；

P3+＝-1；

lsetup(Loop1b_strt，Loop1b_end) LC0＝P5；

// setting is used for the loop of butterfly number

Loop1b_strt：

I1=B3; The address of // conversion factor

R2＝[I2++]；

R3＝[I1++M1]?‖R4＝[I3++]；

lsetup(Loop2b_strt，Loop2b_end) LC1＝P3；

// setting is used for the loop of line number

Loop2b_strt：

R5＝R2+|+R4，R6＝R2-|-R4?‖R3＝[I1++M1]

‖R4＝[I3++]；

A1＝R3.L*R4.H，

A0＝R3.L*R4.L ‖[I0++M2]＝R5‖R2＝[I2++]；

Loop2b_end：?R4.H＝?(A1+＝R3.H*R4.L)，

R4.L＝?(?A0-＝R3.H*R4.H)

‖I0-＝M0‖[I0]＝R6；

R5＝R2+|+R4，R6＝R2-|-R4 ‖I2+＝M2；

I3+＝M2?‖[I0++M2]＝R5；

Loop1b_end：

[I0++]＝R6；

P3+＝1；

P3＝P3<<1；

R0+＝-1；

B1＝B0；

B0＝B2；

B2＝B1；

CC＝R0＝＝0；

If！CC?Jump?Loopfor_m(BP)；

Use Produce and analysis tool as data.Use this

Middle FFT function is to provide with reference to the FFT computing and also to be used to produce the input noise sample.Also write

Script (script) is to analyze this result.

In this special case, select to be used for the binary system 256-point time ten minutes algorithm of computing.This algorithm is implemented 16 plural FFT.Compare two conversion methods: this first method is, in the standard handovers method of each butterfly level with 1/2 conversion; This second method be every a butterfly level with 1/2 the conversion conversion method.

Produce testing tool, it produces Gaussian Profile input data, and is sent to the DSP computing via the input archives.This DSP testing tool reads the input data, and via DSP sign indicating number deal with data producing the output archives, its by Script reads and analyzes.

The very useful instrument of editing simulator that is called that provides in Visual DSP++ is provided in this DSP computing.It is that the master can carry out x86 with personal computer (PC) that this editor's simulator produces one, and its emulation DSP program is accurately accurate to the position, position, and this PC can carry out x86 and allow complete debugger capacity, and carries out fast thousands of times than standard analog.It is particularly useful to edit simulator in this case because its can by

Directly call and carry out.This can carry out DSP through editor and read archives from the PC hard disk, allows simple produce batch test and analysis.

In the analysis, with this from the dateout of DSP with

In the reference FFT computing that provided relatively.Calculate MSE by comparing these two results.This standard is striden complete accurately 16 scopes of this input and is changed, and MSE is drawn to this standard.

Show this result among Figure 25.All curves are the results that handle the Gaussian Profile input among Figure 25.How the enforcement that please notes FFT improves with the signal level increase.This is owing to this round-off error when signal level increases becomes more inessential.With standard handovers, this sees that best MSE is-59.3dB.With alternately conversion, this sees that best MSE is-73.3dB.This improvement can be owing to the conversion of FFT M signal, and it reduces the effect that noiseization is whole, and allows saturated through what control.

This need carry out on ADSP-BF533TM that this FFT must circulate is 3171.The size of this yard is 500 bytes.The size of these data is made of conversion table, and this input and about altogether 3000 bytes of dateout array, and is used for the plural FFT that 256-is ordered.Replace conversion designs and only need for carrying out this: 80 extra circulations, extra 100 bytecode spaces and do not need the excessive data space.

Very large purpose FFT possibility form, size and accurate probability are arranged at present.This special case is presented near the improvement 15dB among result's the MSE and has very little circulation consume.Bigger for big this improvement of FFT size.This be improved as enough big so that its can: improve the performance in the OFDM computing, allow less on DSP/comparatively fast computing; Or permission use less bits is the computing of main design with silicon.This is that the improvement of main design is from lower memory requirements and required less bits in the unit that counts with silicon.This converts to: less chip area, lower power consumption and the cost that reduces.

This is depended in importing errors at different levels by fixing point global error that computing causes.Suppose in the following description error in one-level be with what its level in office in error irrelevant.This is entirely true for round-off error, and to because the error of brachymemma is roughly correct.If take place in the saturated value in one-level, then it can increase a little and uses this value as the probability of importing brachymemma generation in the butterfly subsequently, but this is the second grade effect and can ignoring in force.Therefore, this problem can be by simplifying independent design at different levels in the FFT structure.

In reflector IFFT, for reflector IFFT initial what, this signal value has discrete distribution.This number with level of discrete distribution depends on the size of constellation.In this grade, should design this conversion, so that this maximum possible value can be by brachymemma.This is for taking place very frequently because of this equivalence, so that the brachymemma noise that this caused is many greatly than the minimizing of any noise in round-off error.This conversion should be in level after a while

Emulation widely (its result does not show) shows:: the meticulous adjustment that gains in initial several discrete stages only can produce very trickle improvement, especially for the situation of 64QAM.

Figure 26 to Figure 30 is presented at IFFT output and is used for the different designs standard " signal is to noise ratio " diagram, these standards are: fixing

0.5 conversion and the conversion of module floating-point are alternately changed, are fixed in conversion, 0.5 and 1.Figure 26 is presented at IFFT output and is used for 8 SNR that count and change the FFT size, and it is the function in IFFT input power position standard.It shows that can only use the 8-position to count with careful design reaches the very noise of low level.This minimum average B configuration square error is-15dB place generation approximately accurate in the input position.This is to be used for the above 3dB of gaussian signal lowest order standard, is to take place two signal plus are produced average 3dB rising in signal level after because this changes whole with brachymemma.The general shape of this figure is similar to Fig. 6 but has the x-axle of reversing.This is corresponding to the headroom that reduces because of the input power that is increased.This SNR descends fast on this suitableeest power input because this brachymemma takes place more frequently, under this suitableeest power input this SNR descend slower because this round-off error only slowly increases.This SNR reduces along with the increase of size conversion, because double to increase another level for the IFFT size in conversion at every turn.Yet noise only increases with progression is linear, therefore along with the increase of N has only a spot of degeneration.

Figure 27 shows the effects that are used for 64QAM and 64 IFFT accuracy increases.Increase the reduction that accuracy causes the 12dB noise with two positions.

Figure 26 is for being used for the function of IFFT input signal power in the SNR of IFFT output conduct, this IFFT design has: in the fixed conversion factor at different levels

And 8 fixing point accuracy, these results are used for 64,128,256,512,1024 IFFT and 64QAM modulation.

Figure 27 is for being used for the function of IFFT input signal power in the SNR of IFFT output conduct, this IFFT design has: in the fixed conversion factor of 64 IFFT at different levels and 64QAM modulation

These results are used for 6,8,10,12,14,16 fixing point accuracy.

This conversion factor

And be not easy in Digital Signal Processing, to carry out, therefore implement simulation and whether be used alternatingly 1 and 0.5 with research and can significantly reduce performance as conversion factor and show.Can be by the conversion of only these gts being carried out 0.5 factor in Digital Signal Processing.In Figure 28, show its result.This maximum S R reduces by 2～3dB.

Also implement emulation for other IFFT structure.Its common technique is in use 0.5 conversion factors at different levels, because this guarantees super excessively can not take place.Figure 29 shows the simulation result of this technology.It is many that its performance is used for new technology person's difference, and more serious for its degeneration of more a little bigger IFFT.This is because reduce in signal level at different levels.The whole reduction is that noise increases because signal reduces among this SNR.

The performance performance of this module float dot system is used in research at last.At different levels, check possible super excessive of this system for each conversion operations.If super overflowing taken place, then all values be multiply by 0.5 in any output of this grade.In Figure 30, show its result.Use this module float point to cause the result of plane graph.The accurate selection in this input position is not so good as the important in to adjust compensation of conversion factor selection.This minimum average B configuration square error performance is similar to new technology for N=64, yet its performance can be degenerated quickly when N increases.This is that the probability of this large-signal sample is bigger because for bigger N.

Figure 28 and Figure 30 be as can be seen: use new method can reach significantly better performance than module float point method for bigger N.This preferable performance is to handle complexity than low signal and to reach.

Figure 28 is for being used for the function of IFFT input signal power in the SNR of IFFT output conduct, this IFFT design has: in the alternately fixed conversion factor 0.5 and 1 at different levels, and 8 fixing point accuracy, these results are used for 64,128,256,512,1024 IFFT and 64QAM modulation.

Figure 29 is for being used for the function of IFFT input signal power in the SNR of IFFT output conduct, this IFFT design has: in the fixed conversion factor 0.5 at different levels, and 8 fixing point accuracy, these results are used for 64,128,256 IFFT and 64QAM modulation.

Figure 30 is for being used for the function of IFFT input signal power in the SNR of IFFT output conduct, this IFFT design is used: at module float point at different levels, and 8 fixing point accuracy, these results are used for 64,128,256,512,1024 IFFT and 64QAM modulation.

Only be used for more than should noting the explanation but be not for restriction of the present invention.Have the knack of this technical staff's understanding significantly and need not increase any novelty: these variation and improvement to the foregoing description are to be included in the scope of the present invention.

For example, said system can be applied to not coordination FFT and the conversion that is applied to different carries.Different carries has different performances, and higher carry FFT can show preferable.And, can use different saturated and index detecting designs.

In addition, can complete or partly DFT to allow maximum likelihood.Early use DFT on the level at FFT, and use butterfly then.Can utilize in this way in fourier transform complete accuracy in the level early.

Embodiments of the invention can be applied to multicarrier modulation systems, and it makes any suitable design with modulation time carrier.For example, can use phase shift keying or orthogonal amplitude modulating and changing.In principle, these embodiment can use modulation scheme, and it is applicable to multicarrier modulation systems usually.

Embodiments of the invention can be applied to input data, and it has any Distribution Statistics.For example, these input data can by original coding or any other correlative coding design example as " sign indicating number is cut apart multiple access " (CDMA), make error correcting code and previous coding.

Embodiments of the invention are applicable to ground and wireless system.The various application of this specific embodiments comprise: (a) Digital Transmission, the broadcast of (b) digit news, (c) Groupe Speciale Mobile, the broadcast of (d) Digital Television, (e) satellite communication, (f) WLAN on public telephone network (for example: Asymmetrical Digital Subscriber Line (ADSL) and high-bit-rate digital subscriber line (HDSL)).

Other application examples is as also being possible relevant for big wideband data storing technology.Though embodiments of the invention are to illustrate that about carrier system principle of the present invention also can be applied to the baseband system.

More than disclosed only be several specific embodiment of the present invention, still, the present invention is not limited thereto, any those skilled in the art can think variation all should fall into protection scope of the present invention.

Claims (49)

1. improving one's methods of a multicarrier modulation systems is used for emission and receives data, and this system has reflector and receiver, and this reflector and receiver respectively have input, computing module and output; And

At least one computing module of this reflector or receiver has: the enterprising line number word of signal that the calculation stage of a plurality of connections is used for being received in input calculates, and at least one conversion partly is used for numerical calculation is maintained in the active window in fact, and provides according to this numerical calculation and from then on import the output that received signal is derived;

The method comprises:

By the conversion in both or arbitrary computing module of reflector and receiver architecture, and allow to pre-determine the saturated active window that surpasses of signal of remote probability.

2. the method for claim 1 is characterized in that,

In a plurality of calculation stage of pre-determining of this computing module, allow to pre-determine the saturated of remote probability.

3. method as claimed in claim 2 is characterized in that,

Be included in error detecting and correction step in the receiver.

4. method as claimed in claim 3 is characterized in that,

Select a plurality of calculation stage of pre-determining of this computing module, so that the signal of Gaussian Profile in fact to be provided.

5. as each described method of claim 1 to 4, it is characterized in that,

Use this orthogonal frequency division multitask OFDM.

6. method as claimed in claim 5 is characterized in that, also comprises:

Error detecting in receiver and correction step.

7. as each described method of claim 1 to 6, it is characterized in that,

This computing module is to implement on digital signal processor DSP.

8. as each described method of claim 1 to 6, it is characterized in that,

This computing module is according to the arbitrary or a plurality of middle enforcement of array apparatus down: but processor, microprocessor, microcontroller, microcomputer, on-site programmable gate array FPGA, complex programmable logic device (CPLD), programmable logic device PLD, application-specific integrated circuit ASIC, sea of gates timer, discrete logic chip or discrete analog(ue) numeral or passive component.

9. as each described method of claim 1 to 8, it is characterized in that,

This computing module at least partly is made of following institute: transformational structure; And method, it is saturated that it comprises that permission pre-determines the signal of remote probability in this transformational structure.

10. method as claimed in claim 9 is characterized in that,

This computing module at least partly is made of following institute: fast fourier transform FFT; Contrary fast fourier transform IFFT structure; And method its be included in the saturated quantity that allows to pre-determine number in FFT and IFFT structure or the one.

11. as claim 9 or 10 described methods, it is characterized in that,

This signal is to be represented by plural number, and this calculation stage implements the numerical calculation that counts of this representative plural number, and wherein this is because the saturated brachymemma that allows is to be limited by following mode: with pre-determine mode according to the statistical property of accumulating signal in computing module by control transformation.

12. method as claimed in claim 11 is characterized in that,

The real number of this signal and imaginary part have Gauss or approximate Gaussian distribution at least one calculation stage.

13. as claim 11 or 12 described methods, it is characterized in that,

This pre-determines remote probability saturated is to determine by balance truncation error and quantification, it comprises the transformation result that numerical calculation is provided and has headroom by computing module, it is the ratio of maximum quantization standard to this signal real number part root mean square rms value, it is greater than the headroom that is used for the minimum average B configuration square error owing to quantification, and wherein this is because the minimum average B configuration square error E that quantizes is E (n _q ²)=d ²/ 12, and n _qBe the number of quantization standard, and d is the distance between the quantization standard.

14. method as claimed in claim 13 is characterized in that, also comprises:

Error detecting in receiver and correction step.

15. method as claimed in claim 14 is characterized in that,

Use this orthogonal frequency division multitask OFDM.

16. as claim 13,14 or 15 described methods, it is characterized in that,

This computing module is to implement on digital signal processor DSP.

17. as claim 13,14 or 15 described methods, it is characterized in that,

This computing module is according to the arbitrary or a plurality of middle enforcement of array apparatus down: but processor, microprocessor, microcontroller, microcomputer, FPGA, CPLD, PLD, ASIC, sea of gates timer, discrete logic chip or discrete analog(ue) numeral or passive component.

18. as each described method of claim 11 to 17, it is characterized in that,

This conversion of each calculation stage in according to Binary Conversion at different levels with

Implement.

19. as each described method of claim 11 to 17, it is characterized in that,

This conversion of each calculation stage in according to Binary Conversion in level subsequently with 1 and 1/2 alternately enforcement.

20. as each described method of claim 11 to 17, it is characterized in that,

This conversion in each calculation stage is to implement with 1/2 in quaternary conversion at different levels.

21. as each described method of claim 1 to 20, it is characterized in that,

The real number of signal and imaginary part in active window are quantized into one of finite population quantization standard, are used for the processing that pre-determines a plurality of calculation stage at computing module.

22. method as claimed in claim 21 is characterized in that,

The real number of this signal and imaginary part have Gauss or approximate Gaussian distribution, and this headroom, it is the ratio of maximum quantization standard to this signal real number part root mean square rms value, and be the ratio of maximum quantization standard to this gaussian signal standard deviation, it is selected to pre-determine the signal distortion statistics to provide.

23. method as claimed in claim 22 is characterized in that,

This headroom be with respect to the average power of signal between 8 and 30dB between.

24. method as claimed in claim 22 is characterized in that,

This headroom substitute for be with respect to the average power of signal between 9 and 25dB between.

25. method as claimed in claim 22 is characterized in that,

This headroom substitute for be with respect to the average power of signal between 10 and 20dB between.

26. method as claimed in claim 22 is characterized in that,

This headroom substitute for be with respect to the average power of signal between 10 and 15dB between.

27. method as claimed in claim 22 is characterized in that,

This headroom substitute for be with respect to the average power of signal between 10 and 13dB between.

28. method as claimed in claim 21 is characterized in that,

The real number of this signal and imaginary part have Gauss or approximate Gaussian distribution, and this headroom, it is the ratio of maximum quantization standard to this gaussian signal standard deviation, it is selected to pre-determine the signal distortion statistics to provide, and this error in label is to stride this Mean Square Error sum from all outputs of conversion operations.

29. the device with fixed number digital processing position is characterized in that, comprising:

Input is used for received signal; Computing module is used for processing signals; And

Output is used for this treated signal of output; This computing module comprises fast fourier transform FFT or contrary fast fourier transform IFFT structure, and by the pre-determine saturated quantity of sequencing with the input received signal of permission in FFT or IFFT structure.

30. device as claimed in claim 29 is characterized in that, also comprises:

A plurality of level in computing module, with the cumulative distribution of real number that this signal is provided and imaginary part to have Gauss or approximate Gaussian distribution.

31. as claim 29 or 30 described devices, it is characterized in that,

It is saturated that this pre-determines remote probability is by providing numerical calculation to determine, and changes by a plurality of butterfly levels of computing module, and its result is a headroom greater than owing to quantize headroom in the minimum average B configuration square error.

32. as claim 29,30 or 31 described devices, it is characterized in that,

This conversion of each calculation stage in according to Binary Conversion with

Implement.

33. as claim 29,30 or 31 described devices, it is characterized in that,

This conversion of each calculation stage in according to Binary Conversion in level subsequently with 1 and 1/2 alternately enforcement.

34. as claim 29,30 or 31 described devices, it is characterized in that,

This conversion in each calculation stage is to implement with 1/2 in quaternary conversion.

35. as each described device of claim 29 to 34, it is characterized in that,

Design the imaginary part and real number part of this computing module, and it is quantized to one of finite population quantization standard, be used for preestablishing in a plurality of calculation stage and handle at computing module to handle signal in active window.

36. device as claimed in claim 35 is characterized in that,

Design this computing module has Gauss or approximate Gaussian distribution signal with processing real number and imaginary part, and select headroom, it is the ratio of maximum quantization standard to the gaussian signal standard deviation, preestablish the signal distortion statistics to provide, and signal distortion is for striding this from all output Mean Square Error sums of conversion operations.

37. device as claimed in claim 36 is characterized in that,

This headroom be with respect to average power signal between 8 and 30dB between.

38. device as claimed in claim 36 is characterized in that,

This headroom be with respect to average power signal between 9 and 25dB between.

39. device as claimed in claim 36 is characterized in that,

This headroom be with respect to average power signal between 10 and 20dB between.

40. device as claimed in claim 36 is characterized in that,

This headroom be with respect to average power signal between 10 and 15dB between.

41. device as claimed in claim 36 is characterized in that,

This headroom be with respect to average power signal between 10 and 13dB between.

42. device as claimed in claim 35 is characterized in that,

Design this computing module has Gauss or approximate Gaussian distribution signal with processing real number and imaginary part, and select headroom, it is the ratio of maximum quantization standard to the gaussian signal standard deviation, preestablish the signal distortion statistics to provide, and the statistics distortion is for striding this from all output Mean Square Error sums of conversion operations.

43. as each described device of claim 29 to 42, it is characterized in that,

This computing module is to implement according to digital signal processor DSP.

44. as each described device of claim 29 to 42, it is characterized in that,

This computing module is that array apparatus is arbitrary or a plurality of to be implemented according to descending: but processor, microprocessor, microcontroller, microcomputer, FPGA, CPLD, PLD, ASIC, sea of gates timer, discrete logic chip or discrete analog(ue) numeral or passive component.

45. emission and the method that receives data in the multicarrier modulation systems with reflector and receiver, it is used for individually launching and received signal, it is characterized in that,

This reflector has with receiver: a plurality of inputs, respectively be used for input digit, this numeral and comprise relevant signal real number and imaginary number partly, computing module, be used for implementing digital translation in a plurality of input digits of receive, and a plurality of output, respectively be used to provide the output numeral, it comprises according to digital translation from real number and imaginary number that input digit derived;

Computing module, implement digital translation and comprise a plurality of relevant calculation stage, be used for going up the enforcement digit manipulation to deriving numeral with the relevant input digit of signal part, and this computing module also comprises at least one converter section, be used for one or more digital translation of computing module, during digital translation, to maintain these numerals in the preset range of active window in fact;

The method comprises: by the conversion of at least one converter section, and allowing the digital saturated scope of default remote probability in computing module above active window, and the scope that will this saturated numeral be truncated to active window.

46. method as claimed in claim 45 is characterized in that,

This digital translation is a fourier transform.

47. method as claimed in claim 46 is characterized in that,

This digital translation is fast fourier transform FFT, and use with receiver with orthogonal frequency division multitask OFDM reflector, wherein saturated by allowing by the default remote probability of computing module decision, and with the saturated and quantification with balance of conversion of signals in the FFT structure, and numerical calculation is provided and has the transformation result of headroom, this headroom is the ratio of maximum quantization standard to signal real number part root mean square rms value, this is greater than the headroom in the minimum average B configuration square error owing to quantification, and wherein this is owing to the Mean Square Error E that quantizes is: E (n _q ²)=d ²/ 12, and n _qBe the number of quantization standard, and d is the distance between the quantization standard.

48. emission and the method that receives data in a multimedia modulation system,

It is essentially the method for above-mentioned reference accompanying drawing illustration.

49. emission and the device that receives data in a multimedia modulation system,

It is essentially the device of the graphic illustration of above-mentioned reference.

Images