WO2008042210A2

WO2008042210A2 - Architecture for joint detection hardware accelerator

Info

Publication number: WO2008042210A2
Application number: PCT/US2007/020832
Authority: WO
Inventors: John Zijun Shen; Paul D. Krivacek; Thomas J. Barber, Jr.; Lidwine Martinot; Aiguo Yan; Marko Kocic
Original assignee: Mediatek Inc.
Priority date: 2006-09-29
Filing date: 2007-09-27
Publication date: 2008-04-10
Also published as: US7953958B2; EP2067268B1; US20080080468A1; EP2067268A2; WO2008042210A3

Abstract

A joint detection system is configured to perform joint detection of received signals and includes a joint detection accelerator and a host processor. The joint detection accelerator may include a memory unit to store input data values, intermediate results and output data values; one or more computation units to process the input data values and the intermediate results, and to provide output data values to the memory unit; a controller to control the memory and the one or more computation units to perform joint detection processing; and an external interface to receive the input data values from the host processor and to provide output data values to the host processor. The computation units may include a complex multiply accumulate unit, a simplified complex multiply accumulate unit and a normalized floating point divider. The memory unit may include an input memory, a matrix memory, a main memory and an output memory.

Description

ARCHITECTURE FOR JOINT DETECTION HARDWARE ACCELERATOR

Field of the Invention

The present application relates to joint detection methods and circuits for wireless communication.

Background of the Invention

TD-SCDMA (Time Division Synchronized Code Division Multiple Access) is a wireless communication standard that combines TDD/TDMA (Time Division Duplexing / Time Division Multiple Access) operation with synchronous CDMA (Code Division Multiple Access). TD-SCDMA can assign users to different time slots and spreading codes, whereby each time slot can contain data associated with different users distinguished by a distinct spreading code. FIG. 1 illustrates a TD-SCDMA scheme 100 wherein a frequency band 110 may be used for communication associated with multiple users via the assignment of users to different time slots 121, 122, 123, 124, etc. and different spreading codes 1, 2, 3, etc. (e.g., up to 16 spreading codes may be utilized by a TD-SCDMA slot). Current TD-SCDMA utilizes up to 16 spreading codes per time slot, such that a maximum of 16 users may be simultaneously allocated different spreading codes in a given time slot. In some instances, a user may be allocated to multiple spreading codes.

TD-SCDMA supports asymmetric traffic and services, whereby uplink and downlink traffic allocations are modified using a flexible frame structure that enables uplink and downlink allocations to be dynamically modified during a call. TD-SCDMA also allows for multiple access interference (MAI) to be reduced using joint detection and smart antenna systems. In a joint detection scheme, data from multiple users that may be associated with a time slot are estimated in parallel and a specific user's data is extracted from a received signal. In this way, interference due to signals associated with other users is accounted for and data provided to a user can exhibit reduced interference. Summary of the Invention

A joint detection system is configured to perform joint detection of received signals. The joint detection system comprises a joint detection accelerator to perform joint detection of input data values and a programmable DSP or other host processor coupled to the joint detection accelerator.

According to a first aspect of the invention, a joint detection accelerator is provided to perform joint detection of received signals. The joint detection accelerator comprises a memory unit to store input data values, intermediate results and output data values; an address generation unit to generate addresses for accessing the memory unit; one or more computation units, each comprising a data path, to process the input data values and the intermediate results, and to provide output data values to the memory unit; a data selector to provide the input data values and the intermediate results to the one or more computation units; a controller to control the memory, the address generation unit, the one or more computation units and the data selector to perform joint detection processing; and an external interface to receive the input data values from a host processor and to provide output data values to the host processor.

According to a second aspect of the invention, a joint detection accelerator is provided to perform joint detection of received signals. The joint detection accelerator comprises an input memory to store input data values; at least one intermediate memory to store intermediate results that are calculated during joint detection processing; an output memory to store output data values that are calculated during joint detection processing; a computation block to perform joint detection processing using the input data values from the input memory and the intermediate results from the at least one intermediate memory and to provide the output data values to the output memory; an external interface to supply the input data values from a host processor to the input memory and to supply the output data values from the output memory to the host processor; and a controller to control the input memory, the at least one intermediate memory, the output memory and the computation block to perform joint detection processing. According to a third aspect of the invention, a joint detection accelerator is provided to perform joint detection of received signals. The joint detection accelerator comprises an external interface to receive input data values from a host processor and to provide output data values to the host processor; a memory unit to store the input data values and the output data values; a primary data path comprising a complex multiply accumulate unit; a secondary data path comprising a simplified complex multiply accumulate unit; a divider data path comprising a normalized floating point divider; and a controller to control the memory unit, the primary data path, the secondary data path and the divider data path to perform joint detection processing.

According to a fourth aspect of the invention, a joint detection accelerator is provided to perform joint detection of received signals. The joint detection accelerator comprises an external interface to receive input data values from a host processor and to provide output data values to the host processor; a memory unit to store the input data values and the output data values; an address generation unit to generate addresses for accessing the memory unit; a computation block to perform joint detection processing using the input data values from the memory unit and to provide the output data values to the memory unit; a controller to control the memory unit, the address generation unit and the computation block to perform joint detection processing, wherein the controller is configured to produce a memory access operation on each clock cycle of an iterative computation after an initial latency.

Other aspects, embodiments and features of the invention will become apparent from the following detailed description of the invention when considered in conjunction with the accompanying figures. The accompanying figures are schematic and are not intended to be drawn to scale. In the figures, each identical or substantially similar component that is illustrated in various figures is represented by a single numeral or notation.

For purposes of clarity, not every component is labeled in every figure. Nor is every component of each embodiment of the invention shown where illustration is not necessary to allow those of ordinary skill in the art to understand the invention. All patent applications and patents incorporated herein by reference are incorporated by reference in their entirety. In case of conflict, the present specification, including definitions, will control.

Brief Description of the Drawings

In the drawings, in which like reference numerals represent like elements: - A -

FIG. 1 is a schematic of TD-SCDMA time slots and spreading codes;

FIG. 2 is a schematic of a TD-SCDMA downlink channel model;

FIG. 3 is a schematic representation of various matrices involved in a joint detection process; FIG. 4 is a block diagram of a receiver system implementing a joint detector system;

FIG. 5 is a block diagram of an inner receiver chain;

FIG. 6 is a block diagram of a joint detection process;

FIG. 7 is a schematic of a slot format of a received signal; FIG. 8 is a schematic of the contents of an accumulator before and after a shifting process;

FIG. 9 is a flowchart of a process by which the contents of an accumulator are shifted and a subset of bits of the accumulator are stored to memory;

FIG. 10 is a block diagram of hardware architecture that can implement a shifting process;

FIG. 1 1 is a flowchart of a pre-scaling procedure for propagation channel estimates;

FIG. 12 is a schematic representation of channel estimate pre-scaling;

FIG. 13 is a simplified block diagram of a joint detection accelerator architecture; FIG. 14 is a flow chart of a process that may be performed by a joint detection accelerator;

FIG. 15 is a simplified block diagram of the memory unit and related components of the joint detection accelerator of FIG. 13;

FIG. 16 is a block diagram of the primary data path of the joint detection accelerator of FIG. 13;

FIG. 17 is a block diagram of the secondary data path of the joint detection accelerator of FIG. 13;

FIG. 18 is a block diagram of the divider data path of the joint detection accelerator of FIG. 13; and FIGs. 19A and 19B are pipeline diagrams that illustrate non-optimized and optimized pipeline operation, respectively. Detailed Description

A joint detection system may be implemented as a software solution, for example, using a digital signal processor (DSP), or via a hardware solution, in the form of a circuit referred to as a joint detection accelerator (JDA). In contrast to joint detection systems implemented in software, a JDA allows for reduced power consumption and chip area, as well as improved performance

Applicants have appreciated that joint detection systems can include some processing operations that may benefit from the flexibility of a programmable software implementation and other processing operations may benefit from the reduced power consumption and speed of a JDA. Operations implemented in a JDA may include mature algorithms that are not likely to change and are not prone to customization by different handset manufacturers. Conversely, operations implemented in a programmable DSP may include algorithms that are likely to change and are prone to customization by different handset manufacturers.

Applicants have also appreciated that a JDA may benefit from a fixed point implementation that can reduce chip area and power consumption, while meeting performance specifications, such as block error rate performance. Reducing a data bit- width of a fixed point implementation may in turn enable both reduced chip area and reduced power consumption. Furthermore, a reduced bit-width implies that processing can execute in less time, and therefore longer periods of chip sleep or idle modes are possible. Applicants have realized that a JDA having a reduced data bit-width can still retain high precision during multiply and accumulate operations by performing operations in an accumulator that has a large number of bits, and only saving a reduced number of bits into memory. Applicants have further appreciated that one or more shift values may be determined internally by the JDA and/or configured by an external source, such as a programmable DSP.

Applicants have further appreciated that current initial channel estimation in joint detection systems may demand a large bit- width to accommodate differences in the amplitude of each channel. Such a situation may arise as a result of the manner in which channel estimation is performed. For example, in TD-SCDMA systems, one or more midambles are provided in each burst, and the receiver utilizes midambles to estimate the propagation channel between the transmitter and receiver. However, the receiver performs an initial channel estimation, in which differences between the power levels of midambles and the number of midambles may not be accounted for. Although such discrepancies are ultimately accounted for using scaling factors generated by an active code detection algorithm, the channel estimates generated by the initial channel estimate may demand a larger bit- width, in a fixed point implementation of a JDA, than would otherwise be necessary if the aforementioned effects were accounted for initially. Applicants have realized that a joint detection system may benefit from a pre-scaling of one or more propagation channels prior to being sent to a JDA, thereby allowing for a reduced bit- width in a fixed point implementation of a JDA.

It should be appreciated that the techniques presented herein may be implemented in any way, as they are not limited to any particular implementation. Examples of implementations are discussed below, but it should be understood that these implementations are presented simply as illustrative examples, and the embodiments may be implemented in other ways. The examples presented below are described for use in a joint detection system that may be used with a TD-SCDMA scheme. However it should be appreciated that the techniques described herein may be used with other suitable communication schemes, and/or with other joint detection systems that may be implemented in numerous ways, and are not limited to use with any particular type of joint detection system.

As discussed below, one application for the techniques described herein is for use with a joint detection system in a TD-SCDMA receiver. However, this example is merely illustrative, as the techniques described herein can be used with any suitable types of system wherein joint detection of received signals may be performed. As illustrated in FIG. 2, a TD-SCDMA downlink channel model 200 can include channelization and scrambling codes c_/, C₂, ..., Cκ_a, channel impulse responses hi, fi₂, ..., hκ_a, random noise z added to the channels, and a joint detection data receiver 210. Data d], d₂, ..., dκ_a is respectively multiplied by the channelization and scrambling codes C_/, c ₂, •■•_• C_Ka in the base station and transmitted into the channel. Each code channel can be modeled as an impulse response hi, h₂, ..., hκ_a followed by noise z. Due to the use of smart antennas, the channel impulse response for each code channel can be independent. Received data r is sampled by an analog portion of the receiver 210 and inputted to a joint detection system of the receiver 210. Output x of the joint detection system includes user data which may be further decoded by a downlink bit rate processor.

The cumulative effect of the channelization/scrambling codes and the channel impulse response is the convolution of the channelization/scrambling codes C_x and channel impulse response h_x. The cumulative effect of all of the channels over a single data symbol may be represented by a matrix V wherein the columns of the matrix Fare the convolution of a channelization/scrambling code and the channel impulse response for that code channel. The number of columns of matrix V is the number of active code channels K₈. A combined response matrix T over the entire data field can be constructed by arranging the V matrix along the diagonal of the T matrix.

FIG. 3 is a schematic representation of a channel impulse response matrix H, a channelization/scrambling code matrix C, a V matrix, and a T matrix. As shown in the figure, the //matrix has K_a columns and W rows, and the C matrix has K_a columns and Q rows, where W is the length of the channel impulse response, Q is the spreading factor, and K₃ is the number of active channels. The V matrix has K₃ columns and W+Q- 1 rows, and the T matrix has N*Ka columns and N*Q+W-1 rows, where N is the number of data symbols per block. The number of rows of the V matrix depends on the length of the data field, which is N*Q chips for TD-SCDMA, plus the length W of the channel impulse response, which is between 1 and 17 chip periods, minus one (N*Q + W - 1). Using the above-mentioned matrix definitions, the received data r can be expressed in terms of the transmitted data d and the cumulative effect of the channelization/scrambling codes and the channel impulse response, via the matrix T, plus the noise z: r = Td + z . Joint detection algorithms can be used to recover the transmitted data d from the received data r. A first algorithm that can be used to solve for the transmitted data d uses a least squares (LS) criteria:

2 min Td - r d = (T^HT) -\ r Ip Hr , wherein 7** is the transpose conjugate of the T matrix. The least squares algorithm may not perform well for low received signal to noise ratio (SNR) so another algorithm based on a minimum mean squared error (MMSE) criteria can be used: minE[d -

d = (T^HT + σ²iy^ιT^Hr ,

where σ² is the noise power given by z². Both the LS and MMSE algorithms can be reduced to the same equation:

Ad = y , where y = T^Hr , and is referred to as a matched filter output, and A = (T^HT) for the LS algorithm or A = (T "T + σ²/) for the MMSE algorithm.

Typically solving such equations would involve inverting the A matrix. Due to the properties of the A matrix, Cholesky decomposition can be used to express the A matrix in terms of an upper triangular matrix L and a diagonal matrix D, using the equation A=L^HDL, wherein L^H is the transpose conjugate of the L matrix, which can in turn be used to solve the equation Ad=y recursively. For illustration, FIG. 3 also includes a schematic representation of an A matrix and L matrix. The L matrix is composed of N blocks arranged along the diagonal of the matrix, and as shall be discussed further, the L matrix can be approximated by only computing a limited number of the blocks (e.g., block Bi and B₂) and setting the remaining blocks equal to the values of the last computed block (e.g., setting blocks B₃, B₄, ... B_N to the value of B₂).

A solution procedure to solve the equation Ad=y can include forward substitution, dot division, and backward substitution to arrive at a value for the transmitted data: (1) forward substitution: L^Hf=y

(2) dot division: g=f/D

(3) backward substitution: Ld=g, where/is an intermediate vector solved for in the forward substitution equation and g is an intermediate vector solved for in the dot division equation. Output of the joint detection algorithm can include data for the given user equipment (UE). Data for UEs other than the given UE may be removed so that a final output can include only data for the given UE.

In one embodiment, a joint detector system implementing a joint detection algorithm may include a JDA and a programmable DSP, wherein the programmable DSP performs one or more of the processing operations involved in the joint detection algorithm. A programmable DSP allows for the customization, via software, of the one or more joint detection processing operations that the DSP may perform. The programmable DSP may perform processing operations before the JDA receives the data, may perform intermediate processing operations after the JDA has performed some processing, and/or may perform post-processing after the JDA has completed processing the data. In some embodiments where the DSP performs intermediate processing operations, the JDA may include JDA front-end processing performed before the intermediate processing by the DSP and JDA back-end processing performed after the intermediate processing by the DSP. In one embodiment, an intermediate processing operation performed by the DSP is an active code detection process, as discussed further below. In some embodiments where the DSP performs processing on data prior to sending the data to the JDA, the DSP may perform a channel estimation procedure which can generate H and C matrices. The JDA can be used to solve the linear equation Ad=y, and the DSP may provide received data r, Hand C matrices, and noise power c? to the JDA.

FIG. 4 is a block diagram of an illustrative receiver system 400 implementing a joint detection system including a JDA 415 and a programmable DSP 425, in accordance with one embodiment. Programmable DSP 425 can perform one or more processing operations involved in a joint detection algorithm. System 400 can include a radio and analog base band (combined unit 450), where the radio component can receive a signal that may be transmitted by a base station and the analog base band component can process the received signal provided by radio component. Digital domain component 440 can in turn process the signal provided by analog base band component.

Digital domain component 440 may include a digital base band component and a co-processor that can facilitate digital domain processing. The digital base band component may include programmable DSP 425 that can perform digital processing of the received signal. The digital base band component may communicate with the coprocessor which can facilitate processing of the received signal in the digital domain.

The co-processor can include JDA 415 and bit rate processor (BRP) 416. In one embodiment, JDA 415 may perform one or more processing operations of a joint detection algorithm, and DSP 425 may also perform one or more processing operations of the joint detection algorithm. JDA 415 and DSP 425 may communicate, thereby allowing one or more joint detection processing operations to be performed by DSP 425, with the results of those processing operations being sent to JDA 415 for further processing. Additionally, or alternatively, JDA 415 can also perform one or more processing operations of the joint detection algorithm, with the results of those operations being sent to DSP 425. In this manner, any number of operations in the joint detection algorithm may be performed by DSP 425, and any number of operations in the joint detection algorithm may be performed by JDA 415. The JDA 415 can generate soft decision outputs that may then be converted to the most likely hard decisions by the bit rate processor 416. Bit rate processor 416 can perform channel decoding and error detection on transport channels, de-interleaving to increase channel coding performance, de-rate matching to adjust data rates, de-multiplexing of transport channels, and de- mapping of coded composite transport channels on physical channels.

FIG. 5 is a block diagram of a TD-SCDMA inner receiver chain 500, in accordance with one embodiment. Inner receiver chain 500 may include a receive-side raised root cosine filter 520 which may be implemented in the analog base band (e.g., component 450 of system 400). Raised root cosine filter 520 may provide received signals to one or more pre-processing components, such as DC removal component 530 and I/Q compensation component 540. In one embodiment, DC removal component 530 and I/Q compensation component 540 are implemented by a programmable DSP, such as DSP 425 of system 400. I/Q samples collected from one time slot may be pre-processed by DC removal component 530, which can perform DC offset correction, and by I/Q compensation component 540, which can perform I/Q phase imbalance correction, before the received signal is sent to joint detection system 550. In some embodiments, joint detection system 550 includes a JDA and a programmable DSP that enables customization via software of one or more joint detection processing operations. In one embodiment, the DSP may perform pre-processing operations before sending data to the JDA. Pre-processing performed by the DSP may include channel estimation and/or midamble interference cancellation, as discussed further below.

FIG. 6 is a block diagram 600 of a joint detection process in accordance with one embodiment. The joint detection process illustrated by block diagram 600 may be performed by a joint detection system, such as the joint detection system 550 of the inner receiver chain of FIG. 5. The operations of the joint detection process may be performed by a JDA, in conjunction with a programmable DSP.

The joint detection process may begin by receiving a signal that may have undergone DC offset correction, I/Q phase imbalance compensation, and/or any other pre-processing. The received signal may include two data fields, separated by a midamble. FIG. 7 illustrates a time slot format 700 of a received signal including a first data field 710 followed by a midamble field 720, followed by a second data field 730, and then followed by a guard period 740. Due to air propagation channel spread, data in the tail of data field 710 interferes with the midamble 720 and data in the tail of the midamble interferes with the second data field 730, resulting in data field 712 (rl) and data field 722 (r2).

Data split operation 610 of the joint detection process illustrated in FIG. 6 can process the received signal so as to split the signal into two separate signals, the data field (rl or r2) and the midamble. The joint detection process can be such that data fields rl and r2 can be processed sequentially, in any desired order, wherein processing of one data field (e.g., rl) may reuse some processing results obtained from the processing of the other data field (e.g., rl), as discussed further below. As such, the description that follows can refer to the processing of data field rl and/or r2. In some embodiments, the processing of data field r2 is performed before the processing of data field rl. Such an embodiment may be preferable when the second data field includes command instructions, such as synchronism and power control instructions which may be processed during the processing of the data stream of data field r2 and then rl.

Channel estimation operation 615 can process the midamble provided by data split operation 610 and generate channel estimation matrix Hand code matrix C, such as the Hand C matrices illustrated in FIG. 3. As is known, channel estimation processes can estimate air propagation channels from a base station to the receiver by using a known signal pattern, for example the midamble signal. If a smart antenna is used, each code channel of a TD-SCDMA scheme may be associated with a different propagation channel. A crude estimation of the number of active codes can be made using the result of the channel estimation, but over-estimating the number of active channels is prudent so to not designate an active code as inactive. Ultimately, the active channel detection process of the joint detection can provide a better determination of the active codes. Output of channel estimation operation 615 may include H and C matrices.

Midamble interference cancellation operation 620 can process the data fields outputted by operation 610 by performing a midamble interference cancellation operation that removes the effect of midamble interference on the data fields. The midamble interference cancellation process may involve the use of channel estimates from the channel estimation operation 615. Output of operation 620 may be a data field that has undergone midamble interference cancellation. In one embodiment, data splitting, midamble interference cancellation, and/or channel estimation are performed by the DSP. This allows one or more of these operations to be customized without changing the receiver chip-set.

In one embodiment, the results of pre-processing performed by the DSP are then sent to a JDA for front-end processing. The JDA can receive H and C matrices sent by the DSP (e.g., via an external co-processor interface, as illustrated in the system of FIG. 4) and construct a V matrix in operation 625. The construction of the matrix V may involve using the channel estimation matrix Hand code matrix C. The i^Λ column of the V matrix is the convolution of the i^th column of the H matrix with the i^Λ column of the C matrix. As described below, in some embodiments, a JDA can implement a shift of the result prior to saving the result into memory. Additionally, operation 630 of the JDA can perform a matched filter computation and thereby construct the matched filter output ^T^r, where r is rl and/or r2. The matched filter operation can construct j> using the matrix V and the vector r, and the entire T*¹ matrix does not have to be constructed since many entries of the T^ matrix are zero, as shown in the illustration of FIG. 3. The matched filter operation 630 can receive the V matrix constructed by operation 625. Furthermore, the matched filter operation 630 can also receive the rl and r2 data fields from the midamble interference cancellation operation 620. As described below, in some embodiments, the JDA can implement a shift of the result prior to saving the result into memory.

The JDA can also calculate the power of the y and V matrix columns in operation 635, which can in turn facilitate active code detection. The calculation of the power of each column of the V matrix may involve an operation wherein the squared magnitudes of the V matrix entries are summed for a given column. The calculation of the power of y can be performed on the vector yl and the calculation need not necessarily be performed for_y2, since yl may be sufficient for the purposes of active code detection. In some embodiments, a JDA can implement a shift of the resulting power values prior to saving the results into memory. In one embodiment, active code detection is performed by the DSP. The DSP can receive the y and V matrix (optional) power computation results from the JDA and use the power values to determine the active codes and scaling factors for each channelization code, in operation 640. When the active code detection process is performed in a DSP, customization of the active code detection algorithm is possible. DSP software can be modified as the active code detection process evolves, and the same chip-set can be used to implement the modified active code detection process.

Any suitable algorithm can be used to perform the active code detection. For example, the active code detection procedure may involve determining the codes for which the power of the matched filter output (y) is greater than a threshold level. It should be appreciated that this is just one example of a simple active code detection procedure and any active code detection process may be used. The active code detection process may also determine scaling factors that should be applied to each channelization code. Scaling factors for each channelization code may be represented by mantissa values and exponent values. It should be appreciated that active code detection need not necessarily be performed and the operation may be by-passed in certain circumstances, for example, when using a spreading factor of one, or when a user's equipment already possesses an indication of which codes are active in a given time slot.

Results of the active code detection operation can be used by a SIR estimation operation 655, which may also be performed by the DSP. SIR estimation operation 655 may use both the results of the channel estimation operation 615 and the active code detection operation 640. SIR estimation operation may output the noise power </. It should be appreciated that in some embodiments, the SIR estimation may be performed without using the results of the active code detection. In such cases, the SIR estimation may be performed by the DSP after the channel estimation and can be sent to the JDA prior to when the JDA performs front-end processing. Alternatively, the SIR estimation may be performed by the DSP, at least in part, while the JDA is performing front-end processing. In some embodiments, an indication of the active codes and scaling factors as determined by the active code detection process performed by the DSP, and/or noise computed by the DSP are sent to a JDA for back-end processing. The JDA back-end processing may include ay rescaling and reordering operation 645 and a Frescaling and reordering operation 650. These operations may reorder and rescale the y and V matrix columns on the basis of results sent by active code detection operation 640, where the reordering eliminates any columns corresponding to non-active codes. As a result of such a reordering, the JDA back-end processing may use the same matrix indexing irrespective of which codes are active. The back-end processing performed by the JDA may also include an A matrix computation operation 660 that receives the rescaled and reordered V matrix generated by operation 650 and the noise generated by operation 655 to construct the A matrix by evaluating the matrix operation ^T+ σ I. The construction of the A matrix does not necessarily involve the construction of the T matrix since the elements of the A matrix may be computed using the V matrix directly, and since many elements of the T matrix are zero. Thus, the computation of only the non-zero elements of the A matrix may be performed, and these non-zero elements of the A matrix can be stored (e.g., known zero elements need not be stored). In some embodiments, a JDA can implement a shift of the resulting A matrix values prior to saving the results into memory. The JDA back-end processing may also include a Cholesky decomposition operation 655 that can decompose the A matrix into L and D matrices. The Cholesky decomposition may be performed without the computation of all the L matrix elements. The L matrix can be divided into blocks which converge in value, and the number of blocks computed may depend on the degree of accuracy desired. In one implementation, the number of blocks computed for the L matrix is 2. Using a reduced number of blocks for the L matrix results in a reduced number of divisions in the dot division computation, thereby facilitating the implementation of the joint detection algorithm. Thus, the computation of only a subset of the non-zero elements of the L matrix may be performed, and these non-zero elements of the L matrix can be stored (e.g., known zero elements need not be stored).

The JDA back-end processing can further include a linear equation solver operation 670 that solves linear equation Ax=y (e.g., using forward substitution, dot di vision, and backward substitution, as described previously). Linear equation solver operation 670 can receive data fields from y rescaling and reordering operation 645 and L and D matrices from Cholesky decomposition operation 665. Linear equation solver operation 670 can generate data fields (xl and x2). In some embodiments, a JDA can implement a shift of the results of the forward substitution, dot division, and/or backward substitution processes prior to saving the results into memory.

Data fields xl an x2 can be processed by user extraction operation 675 which can extract the particular UE data using the code(s) which that UE is utilizing. The two data fields xl and x2 can be generated sequentially by the linear equation solver 670 and the user data extraction block 670 can also combine the two data fields so as to create one unified data field x, which the data extraction operation 670 then can output for processing by other components. For example, post joint detection processing may be performed by the DSP. If codes in addition to the UE codes are desired, for example, for power measurements, other codes may be included in the output. In some embodiments, a JDA may be implemented in a fixed point implementation, wherein the result of an operation (e.g., in an accumulator) are shifted prior to saving a reduced number of bits, at fixed bit locations of the accumulator, into memory. Such an operation is equivalent to selecting which bits of the accumulator to save to memory without having to shift the contents of the accumulator. The shift value and which fixed bit locations of the accumulator are to be saved to memory may be chosen so as to ensure that the value in the accumulator is adequately represented in memory (e.g., an accurate value without any significant bit clipping).

In one embodiment, the JDA includes a memory component wherein variables are stored as signed N-bit fractions. As such, the numerical values of the stored variables are between -1 and +1, including -1 and excluding +1. Alternatively, variables in the memory of a JDA may be signed N-bit integers, as the techniques presented herein are not limited to only use with fractions. When operations are performed on two or more stored variables in a JDA, the result of the operation (e.g., stored in an accumulator) may not fit the aforementioned variable range for storage in the JDA memory. The shifting techniques presented herein allow for the storage of values using a desired bit- width.

It should be appreciated that many operations within a JDA are multiplications and/or summations, such as, the operation C_j=∑j afo. A JDA may perform such an operation so as to preserve a high precision during the multiply and accumulate operations (MAC) via the use of an accumulator having a bit-width that is significantly larger than the memory data bit- width to which the final result of the MAC will be saved to. Upon completion of an operation, such as a MAC, a subset of the accumulator bits may be saved to memory. A selection of which accumulator bits shall be saved to memory may involve shifting the contents of the accumulator by a shift value and saving bit values from fixed bit locations of the accumulator to memory.

FIG. 8 is a diagram of an accumulator and a subset of bits of the accumulator which are saved to memory after a shift operation in accordance with one embodiment. The bit values illustrated in diagram 800 are merely for illustrative purposes, and the techniques presented are not limited in this respect. An accumulator 810 may include a larger number of bits than will be stored to a data memory storage location 820 after an operation (e.g., a MAC) is complete. The accumulator 810 may include any number of bits, such as 28 bits, which may include a sign bit, and the number of bits N of the result stored to memory may be any number less than the number of bits of the accumulator, for example 11 bits, which may also include a sign bit. It should be appreciated that the aforementioned accumulator and data bit- width values are just examples, and the techniques presented herein are not limited in this respect. Also, the accumulator size may be selected to be large enough, based on the size of the data being operated on, so as to ensure that there is no substantial loss of precision.

The number of the bits of the accumulator 810 that should be saved to memory may be selected based on the desired memory data bit-width N. Furthermore, the specific location of the fixed bits in the accumulator 810 that should be saved to memory may be selected arbitrarily, since the shift that will be applied to the accumulator contents prior to saving the result to memory may be adjusted accordingly based on the selected fixed bit locations. In the diagram of FIG. 8, the accumulator data bit values that will be saved to memory are the bits values within the rectangle 840.

The data operated on, and the operation performed, may be such that an accumulator radix point 830 is located between two specific bit locations 831 and 832 of the accumulator 810, as may be the case when the data operated on are signed fractions. The accumulator 810 data bits that will be saved to memory, namely bits within the rectangle 840, have been chosen so that the bit location 831 contains the leftmost bit that will be saved to memory.

After an operation (e.g., a MAC) is complete and the result of the operation is in the accumulator 810, the bit values which should be saved to memory are selected based on a shift applied to the bit contents of the accumulator 810. The diagram of FIG. 8 shows the application of a shift to the contents of the accumulator 810 so as to shift the bit values of the accumulator, as illustrated in the resultant accumulator 810'. Accumulator 810' is the same accumulator as 810, after the bit values in accumulator 810 have been shifted by a shift value S to the right. The shift value is a signed integer and may be determined or set in any suitable manner, as discussed below. A positive shift value S, where S is a positive integer, may be associated with a shift of the accumulator bits to the left. A negative shift value -S, where S is a positive integer, may be associated with a shift of the accumulator contents to the right. It should be appreciated that the sign of the shift value is arbitrary and depends on convention, and, the techniques presented herein are not limited in this respect.

In the illustration of FIG. 8, the shift value is such that the shift operation moves the first sign bit value at bit location 833 into the accumulator bit location 831. Upon completing the shift of the contents of the accumulator, the bit values at the fixed bit locations of the accumulator, as indicated by rectangle 840, are saved to memory. In the illustrative example of FIG. 8, the shift value was selected so that the repeated sign bits of the binary number within the accumulator are not stored within the memory, which is referred to herein as normalization. It should be appreciated that values may be shifted by any amount, as the example presented in FIG. 8 is merely for purposes of illustration. In some embodiments, values stored to memory are signed N-bit fractions, and the shift value applied to the contents of the accumulator is such that the shifted contents to the left of the accumulator radix point only include repeated sign bits (e.g., bits to the left of bit location 833 in accumulator 810).

FIG. 9 is a flowchart of a process 900 by which the result of an operation stored in an accumulator may be stored to memory. Process 900 may be performed by hardware in a JDA, which in the case of a MAC operation may include a multiply accumulate unit and one or more shifters. In act 902, an operation (e.g., a MAC) is performed with the result stored in an accumulator. Upon completion of the operation, the accumulator contents are shifted by a shift value, in act 904. The shift operation is equivalent to multiplying or dividing the contents of the accumulator by 2^SHIFT, where SHIFT is the shift value. The shifting operation may be performed by an output shifter, and the shift value used may be determined internally by the JDA or may be provided by a system external to the JDA (e.g., a programmable DSP, and may be specified by a user).

In act 908, the accumulator contents may be rounded in anticipation of storing a subset of bits into memory. The rounding may be performed by rounding-up (or down) the last bit of the subset of bits of the accumulator that will be stored to memory. However, it should be appreciated that rounding may be performed in any other suitable way, as the techniques presented herein are not limited in this respect. In act 910, overflow verification may be performed to determine whether rounding results in an overflow, as is well known to those skilled in the art. If a determination is made that overflow occurred, the contents of the accumulator that will be saved to memory are saturated (act 912). Saturation involves setting the value to be stored to memory to the largest positive number or the smallest negative number.

If a determination was made that no overflow occurred (in act 910), or overflow occurred and the value was then saturated accordingly, the process proceeds with act 914 where N contiguous bits at a specified fixed location in the accumulator are saved to memory, where N is less than the total number of accumulator bits. The number N of bits saved to memory and the fixed location of the accumulator bits that should be saved to memory may be specified by a hardware designer upon the design of the JDA. Process 900 may then terminate.

In some embodiments, different variables may have different associated shift values. In some embodiments, each element of a vector or matrix stored in memory is assigned the same shift value. In other embodiments, different columns or rows of a matrix are assigned different shift values. Allowing different shift values to be used for different columns or rows of a matrix can allow accuracy to be improved by selecting the shift value for each column or row based on a tailoring of the shifts to the values in each column or row of the matrix.

It should be appreciated that an addition operation performed by a JDA may involve two or more stored variables having been potentially stored using different shifts. It should be understood that stored variables associated with different shift values may be viewed as the storage of mantissa values having different exponents. In such operations, the JDA may ensure that the shift is the same for all of the variables being added prior to performing the addition. For example, when performing an operation such as Cpd_j+Σ, a,b,, a JDA may determine whether one or more of the vector elements have been shifted prior to storage in memory. If one or more of these vectors have been shifted, then the JDA may ensure that all the vectors have the same shift value prior to performing the addition. For example, if vector a has undergone a shift by a shift value A SHIFT prior to being saved into memory, vector b has undergone a shift by a shift value B_SHIFT prior to being saved into memory, and vector d has undergone a shift by a shift value D_SHIFT prior to being saved into memory, then the JDA may shift the vector d retrieved from memory by A_SHIFT+B_SHIFT-D_SHIFT prior to adding the d vector elements to the summation result Σ, a,b,. Such an operation can be represented mathematically as (d_j« A_SHIFT+B_SHIFT-D_SHIFT)+ Σ, a,b,, where the operator "«" represents a shifting operating performed on d_j. The result of the operation may then also be shifted prior to storage in a memory location.

FIG. 10 is a block diagram of hardware architecture 1000 in a JDA that can implement a shifting process for an operation involving the addition of a variable d to the result of a summation ∑_! a,b_b as described above. In one embodiment, the number of bits N used to store variables a, b, and d is 11, and the number of accumulator bits is 28, although other data bit-widths are possible, and the techniques described herein are not limited in this respect. Hardware 1000 can include an input shifter 1008 for shifting the d value prior to loading into an accumulator 1006. The shift value used by input shifter 1008 may be A_SHIFT+B_SHIFT-D_SHIFT, if vector a has undergone a shift by a shift value A_SHIFT prior to being saved into memory, vector b has undergone a shift by a shift value B-SHIFT prior to being saved into memory, and vector d has undergone a shift by a shift value D-SHIFT prior to being saved into memory.

Hardware 1000 can also include a multiplier 1002 for multiplying the a, and b, input values, and an adder 1004 for adding the contents of the accumulator 1006 with the results of the multiplication of a, and b, provided by multiplier 1002. Accumulator 1006 can comprise a number of bits A, which may be greater than the number of bits N of the input data. Upon completion of the multiply and accumulate operation, the value in the accumulator 1006 may be shifted by a shift value C-SHIFT by output shifter 1010, and a subset of the accumulator bits may be saved to memory. The subset of bits of the accumulator may include N bits at fixed locations in the accumulator, as described previously. The shift value C_SHIFT may be selected so that the significant bits of the resultant value in the accumulator are saved to memory. By reducing the number of bits used for storage of intermediate values computed during the joint detection process, a JDA may possess a desirable speed, memory area, and/or power consumption.

In some embodiments, one or more shift values used in a JDA are set by a system external to a JDA (e.g., by a programmable DSP). An external system can include a programmable DSP that allows a designer to program software that determines shift values for one or more variables stored by the JDA. Alternatively, or additionally, a designer sets a fixed shift value (e.g., via the programmable DSP), which is then provided to the JDA. The determination of shift values by the DSP may involve the use of results from processing performed in the front-end processing of a JDA. For example, shift values for one or more variables stored by the JDA may be determined by the DSP based on results of an active code detection procedure that may be performed by the DSP. Such a procedure may be desirable since the selection of shift values for back-end operations of a JDA may depend on the number of additions in summation operations, which may in turn be related to the number of active codes. Shift values therefore may be determined by the DSP based at least partially on the results of the active code detection procedure and the shifts may then be sent to the JDA.

In some embodiments, one or more shift values used by the JDA are determined internally by the JDA. Internal determination by the JDA of one or more shift values to be used by the JDA may involve analysis of results of the data to be stored in memory. In the case of an A matrix, it may be difficult to compute a maximum possible output shift in software because the V matrix is stored internally in the JDA, and therefore it may be desirable to have the JDA determine shift values for the A matrix. In some embodiments, a maximum possible output shift for the A matrix can be computed by the JDA internally. If the output shift for the A matrix is computed by the JDA internally, the maximum possible output shift can be determined based on the largest element in the A matrix. The largest element in the A matrix is along the diagonal since the elements on the diagonal represent the autocorrelation for each code channel (plus noise) and the channelization codes, and the autocorrelation for each code channel is greater than the cross-correlation to any other code channel. To determine the maximum possible output shift for the A matrix, each element of the diagonal of the A matrix may each be computed and the maximum possible shift value of the largest element may be used as the maximum possible output shift for all A matrix elements. It should be appreciated that the automatic internal determination of shift values may be performed using only a small number of cycles as compared to the total number of cycles used to perform the entire joint detection process.

The storage of other variables in the JDA may also benefit from the internal determination of shift values by the JDA. For example, the results of the dot division process may also benefit from the internal determination of shift values that should be applied to the result prior to storing the result in memory. In the case of a dot division, such as the //D operation of the linear equation solver operation, as described previously, the operation includes a multiplication of a fraction with the inverse of a diagonal D matrix. Since the elements of D are positive fractions, there exists a possibility that the dot division result is not a fraction. An internally derived shift for the inverse diagonal elements of the D matrix (i.e., J/du) can be used to ensure that the result of the dot division is also a fraction. In some embodiments, a single shift value for all elements of the matrix is used, which can minimize computational complexity and memory area. In such embodiments, the single shift value may be determined by determining a maximum possible shift value of the largest element of the matrix, and then using that maximum possible shift value for all elements of the matrix.

Division for the D matrix may be executed for each element of the D matrix in multiple parts. First, each diagonal element of the D matrix may be normalized via a shifting process that shifts each element by a shift value, thereby eliminating repeated sign bits, as is known in the art. The shift values applied to all elements may be the same, to reduce computational complexity, or may be different, as the techniques presented herein are not limited in this respect. The normalized values of the D matrix elements da, referred to as normalized(du), are less than 1 and greater than or equal to 0.5. Therefore, each value 0.5/normalized(da) is greater than 0.5 and less than or equal to 1. The values 0.5/normalized(da) may be computed in an intermediate divider having a larger number of bits (e.g., 21 bits) and then a reduced number of those bits may be saved in memory in the form of a mantissa (e.g., as an 11 bit value) and an exponent (e.g., a 5 bit value). Furthermore, the maximum exponent for the values 0.5/normalized(dji) may be determined and used as a shift value prior to storing the result of the dot division operation, g=f/.D, where it should be understood that the maximum exponent can be used a shift value for all of the elements of the g vector.

In some embodiments, a designer may select whether the JDA uses a shift value for a given variable such that the shift is determined internally by the JDA or set by a system external to the JDA (e.g., by the programmable DSP). A designer may be able to set a bit variable, for example via the programmable DSP that communicates with the JDA, wherein the bit variable indicates whether the JDA should use a shift value for a given variable that is determined internally (e.g., as described above for the A and IJD matrices) by a JDA or set by a system external to a JDA (e.g., as programmed into a DSP by a designer). This allows a designer to select which variables should be stored using shifts determined internally and which variables should be stored using shifts as determined or set by an external source (e.g., a programmable DSP). This approach offers flexibility by enabling a designer to select which bits are significant via externally determined shift values for some variables (e.g., shifts values programmed by a designer), while at the same time allowing the JDA to internally determine shift values for other variables using the results of intermediate processing. The shift values set by a system external to the JDA may involve a calculation of the shift values using intermediate results provided to the external system, or may be fixed shift values that may be provided by a designer.

It should be appreciated that one or more techniques for using, determining, and/or setting shifts for a fixed point implementation of a JDA may be used alone or in combination with other techniques presented herein. Shifts may be used in a JDA that communicates with a programmable DSP to perform one or more processing operations, such as one or more intermediate processing operations, but the technique of shifting variables in a JDA may also be used by JDAs that do not possess all the features presented herein (e.g., JDAs that do not necessarily use a DSP for intermediate processing operations).

In some embodiments, a joint detection algorithm may include an operation of pre-scaling one or more propagation channel estimates (e.g., one or more columns of the H matrix of FIG. 3) prior to sending the channel estimates to a JDA. A pre-scaling operation may be included in a channel estimation operation, and performed once initial channel estimation is complete and before outputting of the channel estimates. Such pre- scaling may be performed in a programmable DSP that may also perform the initial channel estimate process. Pre-scaling of one or more propagation channel estimates prior to those channel estimates being sent to a JDA may allow for improved accuracy in a fixed-point implementation of a JDA.

FIG. 11 illustrates a flowchart 1100 of a pre-scaling procedure for the propagation channel estimates. The process can begin with act 1102, wherein one or more pre-scaling factors are determined. The pre-scaling factors will be applied to propagation channel estimates that may have been determined by an initial channel estimation procedure. The pre-scaling factors for each propagation channel may be different, but it should be appreciated that the technique is not limited in this respect. The determination of the one or more pre-scaling factors may be performed in any suitable manner.

Pre-scaling factors for the propagation channel estimates may be selected based at least partially on the maximum element of the propagation channel estimate and/or the power of the propagation channel estimate. Pre-scaling factors may be selected to achieve a variety of objectives, including, but not limited to: (1) scaling the propagation channel estimates so that after pre-scaling the propagation channel estimates have the same exponent for their maximum absolute elements, (2) scaling the propagation channel estimates so that after pre-scaling the propagation channel estimates have substantially the same maximum absolute elements (e.g., same exponent and absolute value of mantissa for their maximum absolute elements), or (3) scaling the propagation channel estimates so that after pre-scaling the propagation channel estimates have substantially the same power.

In one embodiment, the pre-scaling factors may be selected so as to ensure that, upon pre-scaling, the exponents of the maximum absolute element of each channel estimate are the same. When channel estimates comprise complex numbers entries, the maximum absolute element can be selected to be the maximum value of the set including the absolute value of the real components and absolute value of the imaginary components of the entries. In this context, a complex number entry is formed of two real number elements, namely the real and imaginary components of the complex number. If the initial propagation channel estimates are given by

/z, = [Λ(θ)Λ(i),-Λ(_w - i)]

Pt₂ = [H₂(OIh₂ (I),- -Λ(_W - 1)]

where hi, h_∑, ..., h_ka are the initial propagation channel estimates (columns of the H matrix) and each initial propagation channel estimate is a vector having w-1 complex entries. Each complex entry h,(j) is formed a real element real(h,(j)) and imaginary element imag(h,(j)). Thus, as referred to herein, the maximum absolute element (also referred to as the maximum absolute value of the plurality of values that form an initial propagation channel estimate) of a given propagation channel estimate hi may be represented as the maximum of the set given by {abs(real(h,(j))), abs(imag(h,(j))), j=0, ..., w-1}.

In another embodiment, the maximum absolute element of a given propagation channel estimate (e.g., a given column of the H matrix) is determined and the pre-scaling factor for the given channel may be set to the inverse of the maximum absolute element, thereby ensuring that after pre-scaling, the elements of the given propagation channel are less than or equal to unity (e.g., fractions). The pre-scaling factors for each propagation channel estimate may be represented using a separate mantissa and exponent. In another embodiment, the power of each propagation channel estimate (e.g., the norm squared of each column of the H matrix) is determined and each channel may be scaled with a pre-scaling factor so as to have the substantially the same power after scaling. The pre-scaling factors for each propagation channel estimate may thus be selected to be the inverse of the power of each propagation channel estimate. In act 1104, each propagation channel estimate is pre-scaled using the pre-scaling factors determined in act 1102. Pre-scaled propagation channel estimates and corresponding pre-scaling factors may then be provided to a JDA.

FIG. 12 shows an illustrative embodiment wherein the initial propagation channel estimates are pre-scaled using a pre-scaling factor for each propagation channel in accordance with one embodiment. The initial channel estimates may be represented by columns in the //matrix, as illustrated in the H matrix 1210 which includes one column for each channel, and wherein the length of each column of the H matrix is W. Therefore, each column of the //matrix includes W entries (e.g., complex numbers). FIG. 12 illustrates the pre-scaling of each initial propagation channel estimate (each column of the H matrix) by a pre-scaling factor Pj, so as to construct a pre-scaled H matrix 1250. Each pre-scaling factor may be determined in any suitable manner, for example, as described using processes such as those described for act 1102 of the method illustrated in the flowchart of FIG. 11. It should be appreciated that the technique of pre-scaling may be used alone or in combination with other techniques presented herein. Pre-scaling may be used in a JDA that communicates with a programmable DSP to perform one or more processing operations, such as one or more intermediate processing operations, but the technique of pre-scaling may also be used in other types of JDAs. The effect of pre-scaling factors may be accounted for during the joint detection process, or at the end of the joint detection process, once the user data has been extracted. In some embodiments, all factors used to scale or shift intermediate results produced during the joint detection process may be accounted for at the end of the joint detection process. Such factors may include pre-scaling factors, scaling factors as dictated by active code detection, and shift values used for JDA memory storage. For example, if the T matrix has been scaled, the^ influence of those scaling factors (e.g., from active code detection and/or from pre-scaling) may be removed from the final output. Additionally, or alternatively, any shifts performed on intermediate results produced during the joint detection process can be accounted for by shifting the final output by the negative of the net shift value.

One or more of the techniques described herein may be implemented using any suitable hardware architecture, as the techniques are not limited in this respect. FIG. 13 is a block diagram 1300 of a JDA architecture in accordance with embodiments of the invention. The JDA architecture of FIG. 13 includes an external bus interface 1320 that allows the JDA to communicate with external components, such as a programmable DSP or other host computer, via a communication bus 1321. The JDA may include a controller 1350 that controls a plurality of hardware blocks so as to execute a joint detection algorithm, such as the JDA algorithm illustrated in the block diagram of FIG. 6. The JDA hardware blocks may include a data address generator 1303, a register file 1302, a JDA memory unit 1304, and a computation block 1305. The computation block 1305 may include one or more computation units, or compute data paths, to perform joint detection processing. In the embodiment of FIG. 13, computation block 1305 includes a primary data path 1306 (e.g., a complex number multiply accumulate unit), a secondary data path 1307 (e.g., a complex number multiply accumulate unit), and a divider data path 1308. Register file 1302 can store configuration and status information, and JDA memory unit 1304 can store input data values, intermediate results and output data values associated with the joint detection processing algorithm. An input multiplexer, or data selector, 1314 may be included in the architecture and configured to direct input data to the primary data path 1306, the secondary data path 1307, and the divider data path 1308. An output multiplexer 1316 may be configured to direct the results of executed operations back to the JDA memory unit 1304. A buffer 1330 may be used to hold data values temporarily as described below.

As illustrated in FIG. 13, the JDA may include compute data paths that can perform different types of operations. The JDA architecture illustrated in the block diagram of FIG. 13 includes primary data path 1306, secondary data path 1307, and divider data path 1308. As discussed below, primary data path 1306 includes a complex number multiply accumulate unit and secondary data path 1307 includes a simplified complex number multiply accumulate unit.

The primary data path 1306, which includes a complex multiply and accumulate unit, can perform operations such as ∑j ajbj + d_j. The illustrated primary data path includes a shifted input register 1310 and a quantization unit 1312 which can multiply or divide values by powers of two (i.e., multiply by 2 ). Shifted input register 1310 may be used to supply a shifted input to the accumulator, and quantization unit 1312 may be used to shift the bits of a resultant value in the accumulator before a subset of those bits (e.g., at fixed bit locations in the accumulator) are stored in joint detection memory unit 1304. In some embodiments, the primary data path may be used to execute multiplication and accumulation operations other than the computation of the V matrix. The primary data path 1306 is shown in greater detail in FIG. 16. The secondary data path 1307, which includes a simplified complex multiply and accumulate unit, can perform operations such as Σ, aib, where b, is either +1, -1, +j, or -j. Such an operation may be performed during the computation of the V matrix, which may involve the convolution of the //matrix and C matrix columns. Since the code matrix C may be restricted to include elements belonging to the set {+l,-l,+j,-j}, the secondary data path may be used to compute the V matrix elements. The secondary data path 1307 may be configured to more optimally perform the multiplication of a,b, via the use of one or more multiplexers that can select the real or imaginary component of input value a, on the basis of whether b, is +1, -1, +j, or -j. The accumulation operation performed by the secondary data path does not necessarily include the addition of a d_j value to the summation of a,b,, nor an accompanying input shift of d_j. It should be appreciated that an output shifter that operates on the result of the accumulation may be included in the secondary data path, so as to shift the output. In some embodiments, the secondary data path 1307 may be used to execute the computation of the V matrix. The divider data path 1308 may be used to compute Vd₁₁ operations during the dot division process. The division data path may be used to perform the normalized division process wherein each diagonal element of the D matrix may be normalized via a shifting process that shifts each element by a shift value, thereby eliminating repeated sign bits, as is known in the art. The shift values applied to all elements may be the same, to reduce computational complexity, or may be different, as the techniques presented herein are not limited in this respect. Values Q.5/normalized(d_l() may be computed in an intermediate divider having a larger number of bits (e.g., 21 bits) and then a reduced number of those bits may be saved in the JDA memory unit 1304 in the form of a mantissa (e.g., as an 11 bit value) and an exponent (e.g., a 5 bit value). It should be understood that the storage of a mantissa and exponent for the inverse elements of the D matrix does not necessarily require a specific memory, as the mantissa can be stored in place of a real part of a complex number and the exponent can be stored in place of an imaginary part of a complex number.

Controller 1350 controls the operation of the hardware blocks, which may be configured as pipelined stages including an address generation stage 1360, a data fetch stage 1370, an execution stage 1380, and a data write stage 1390. A joint detection operation is divided into several sub-operations, and the sub-operations are performed in respective pipeline stages on consecutive clock cycles of the JDA. The address generation stage 1360 may include the data address generator 1303. The data fetch stage 1370 may include the JDA memory unit 1304 and the input multiplexer 1314. The execution stage 1380 may include the primary data path 1306 and the secondary data path 1307. The divider data path 1308 requires several clock cycles to complete an operation and thus is not considered part of execution stage 1380. The data write stage 1390 may be associated with writeback of the result from the execution stage 1380 to the JDA memory unit 1304. Controller 1350 generates control and timing signals for performing joint detection processing as described herein. In one embodiment, except for an initial latency at the start of each joint detection operation, a memory access, either a read access or a write access, occurs on each clock cycle.

JDA processing operations, such as those illustrated in the block diagram of FIG. 6, may be performed by the architecture of FIG. 13 under the control of the controller 1350. In some embodiments, controller 1350 is implemented as a finite state machine for control of the hardware blocks of the joint detection accelerator. The finite state machine sequences through a series of states to perform joint detection processing. In further embodiments, controller 1350 is implemented as a programmable controller, including a program sequencer and an instruction memory, such as a RAM or ROM, to hold instructions for performing joint detection processing. In these embodiments, the controller executes the instructions in the instruction memory. The instructions control the hardware blocks of the joint detection accelerator to perform joint detection processing.

FIG. 14 is a flow chart 1400 of a process that may be implemented by the controller 1350 to control the hardware blocks of the JDA architecture and perform one or more tasks of the joint detection process. In each operation of the joint detection process, controller 1350 controls the blocks of the joint detection accelerator to perform the indicated operation.

The controller 1350 may begin with the construction of the V matrix (act 1402), by computing the V matrix elements. The secondary data path 1307 may be used to construct the V matrix. The controller may proceed with computing a first matched filter output yl (act 1404). The primary data path 1306 may be used to compute the first matched filter output. The controller may proceed with the computation of the power of the V matrix and/or the previously computed matched filter output, where the power of only one matched filter output need be computed in some embodiments (act 1406). The primary data path 1306 may be used to compute the power of the first matched filter output and the V matrix. The controller 1350 may then proceed with computing a second matched filter output >>2 (act 1408). The primary data path 1306 may be used to compute the second matched filter output. The controller may also proceed to control the execution of acts related to active code detection, which may be performed simultaneously with other acts, such as the construction of the matched filter output^. In some embodiments, the acts associated with active detection may include a determination of whether active code detection is desired. An indication of whether active code detection is desired may be provided by a designer. The indication may include setting a parameter to specify whether active code detection is to be performed or skipped. The parameter setting may be provided to the JDA via any suitable means, as the embodiments are not limited in this respect. The parameter setting that indicates whether active code detection is to be skipped may be provided from the DSP to the JDA. The controller 1350 may determine whether to skip the active code detection based on the value of the aforementioned parameter (act 1409). If a determination is made that active code detection is to be skipped, the controller may skip the active code detection and continue with the remaining acts in the process. If a determination is made that active code detection is not to be skipped, and hence should be performed, the controller may direct the initiation of the active code detection, which may be performed by a programmable DSP coupled to the JDA. In this manner, while the DSP is performing the active code detection, the JDA may be simultaneously performing other operations that do not use the results of the active code detection, such as constructing the matched filter output y2 (act 1408).

The controller 1350 may then wait for the active code detection to be completed (act 1410), wherein, as previously described, in some embodiments, the active code detection may be performed by a component other than the JDA₅ for example a programmable DSP. When active code detection is complete (act 1410) and constructing the matched filter output y2 is complete (act 1408), the controller may control the rescaling and reordering of the matched filter outputs yl, y2, and the V matrix (act 1412). The primary data path 1306 may be used for rescaling and reordering of the matched filter outputs and the V matrix. The controller may then control computation of the A matrix (act 1413). The primary data path 1306 may be used for computation of the A matrix. The controller may then perform a Cholesky decomposition (act 1414). The primary data path 1306 and the divider data path 1308 may be used to perform Cholesky decomposition. The controller may then proceed with the computations to solve the linear equation A *xl =yl for the value of xl (act 1416). The controller may proceed with the extraction of user data from the solution xl (act 1418). The controller may then proceed with the computations to solve the linear equation A *x2=y2 for the value of x2 (act 1420). The controller may proceed with the extraction of user data from the solution x2 (act 1422). As discussed above, solving the linear equations for the values of xl and x2 may involve forward substitution, dot division and backward substitution. The primary data path 1306 and the divider data path 1308 may be used to solve the linear equations in acts 1416 and 1420. The primary data path 1306 may be used for extraction of user data in acts 1418 and 1422. The controller may then wait for the next time slot (act 1424) and repeat the process beginning again at act 1402 when a next time slot is received.

It should be appreciated that in some embodiments, loading data and control parameters (herein sometimes referred to collectively as data) for the next time-slot can occur before the end of the processing for the current time-slot. In some embodiments, loading data and control parameters for the next time-slot can be performed as soon as the JDA has completed loading the control parameters and data for the current time-slot. Control parameters can include shift values, a parameter indicating whether active code detection should be skipped, the length of the channel (W), and/or the number of codes. An embodiment of memory-related components of the j oint detection accelerator is shown in the block diagram of FIG. 15. JDA memory unit 1304 may include several independently-accessible memories to enhance performance. In the embodiment of FIG. 15, JDA memory unit 1304 includes an input memory 1510, a first intermediate memory, or matrix memory, 1512, a second intermediate memory, or main memory, 1514 and an output memory 1516. Each memory may be a dual-port memory to further enhance performance. The memories 1510, 1512, 1514 and 1516 may be addressed by data address generator 1303 (FIG. 13) and/or controller 1350. Input data values from the host processor are written to an input register 1520 and then are written to input memory 1510 under control of controller 1350. Output data values are read from output memory 1516 and placed in an output register 1522 under control of controller 1350. The output values are read from output register 1522 by the host processor. The input register 1520 and the output register 1522 permit exchange of data values with the host processor. Through use of input register 1520 and output register 1522, the host processor writes input data values to a single location and reads output data values from a single location, without accessing JDA memory unit 1304. As shown in FIG. 6, the input data values may include front-end processing inputs received from the DSP after initial preprocessing of received signals and back-end processing inputs received from the DSP after intermediate processing, such as active code detection. The output data values may include front-end processing outputs supplied to the DSP after front-end processing and back-end processing outputs supplied to the DSP after back-end processing. The joint detection accelerator may further include buffer 1330 to enhance performance as described below. Buffer 1330 may be used to hold input values and intermediate results so as to reduce accesses to memory unit 1304 in some situations. For example, buffer 1330 may be used to hold temporarily data values read from memory during an idle clock cycle and processed during a later clock cycle. In addition, buffer 1330 may be used to hold temporarily intermediate results or output values until they can be written to the appropriate memory. Also, buffer 1330 can be used to hold temporarily intermediate results that are to be used by the computation block 1305 in a subsequent computation, thereby avoiding the need for writing the intermediate result to one of the memories. Furthermore, the data in buffer 1330 can be reused in two or more computations when appropriate, thereby reducing memory accesses and saving power.

As shown in FIG. 15, buffer 1330 can receive writeback data values from the computation block 1305 and can receive data values from input memory 1510, matrix memory 1512 and main memory 1514 via a multiplexer 1532. Buffer 1330 can output data values to the computation block via multiplexer 1314, to matrix memory 1512 via a multiplexer 1534, to main memory 1514 via a multiplexer 1536 and to output memory 1516 via a multiplexer 1538. In one embodiment, input memory 1510 is a 1264x22 dual -port memory used to store input data values for a time slot. Input memory 1510 receives input data values from the host processor and supplies the data values to the computation block 1305 via multiplexer 1314. Data values may be written to input memory 1510 by the host processor at the same time that data values are being read from input memory 1510 for joint detection processing by the computation block.

Matrix memory 1512 may be a 512x22 dual-port memory used to store the combined channel estimation and spreading/scrambling matrix and the reordered version of the channel estimation matrix. A separate matrix memory permits matrix generation to execute in parallel with other joint detection operations.

Main memory 1514 may be a 2064x22 dual -port memory used to store the intermediate variables that are generated from the matched filter operation, the reordered version of the matched filter, the A matrix, the L matrix, forward substitution, dot division and backward substitution. The memory space may be reused for these values. The main memory is implemented as a separate memory block so that external memory reads and writes do not cause joint detection processing to stall.

Output memory 1516 may be a dual-port 740x16 FIFO (first-in, first-out) buffer to store the output data values that result from joint detection calculations. The output memory 1516 is implemented as a separate memory block so that external memory reads to not cause the joint detection operations to stall.

As shown in FIG. 15, memories 1512 and 1514 receive writeback data values from computation block 1305 and from buffer 1330. The memories 1512 and 1514 supply intermediate data values to the computation block via multiplexer 1314 for further processing. Output memory 1516 receives writeback data from the computation block 1305 and from buffer 1330. Output memory 1516 supplies output data values to the host processor via output register 1522 and external bus interface 1320.

An embodiment of the primary data path 1306 is shown in FIG. 16. Primary data path 1306 may include dual complex multiply accumulate (MAC) circuits with additional functionality as described below. As shown, each complex MAC includes multipliers 1610 and 1612, each of which receives two 11-bit inputs, an adder 1614 which receives the outputs of multipliers 1610 and 1612, an adder 1616, an accumulate register 1618 and a multiplexer 1620. Adder 1616 receives the outputs of adder 1614 and multiplexer 1620, and accumulate register 1618 receives the output of adder 1616. Accumulate register 1618 supplies an output to quantization unit 1312 and to an input of multiplexer 1620. Multiplexer 1620 receives a second input from shifted input register 1310. A shifted data value from shifted input register 1310 can be loaded into accumulator register 1618. Shifted input register 1310 may perform the functions of shifted input 1008 shown in FIG. 10 and described above, and quantization unit 1312 may perform the functions of output shifter 1010 shown in FIG. 10 and described above. An embodiment of the secondary data path 1307 is shown in FIG. 17. Secondary data path 1307 may include a simplified complex MAC, which does not require the use of multipliers. This configuration permits multiplication of complex values where one of the complex values is +1, -1, +j, or -j. As shown, the secondary data path 1307 includes a first four-input multiplexer 1710 and a second four-input multiplexer 1712. Multiplexer 1710 receives inputs A real, A real inverted, A imaginary inverted and A imaginary at inputs 0-3, respectively. Multiplexer 1712 receives inputs A imaginary, A imaginary inverted, A real and A real inverted at inputs 0-3, respectively. Each of the multiplexers 1710 and 1712 is controlled by the complex number B imaginary. The output of each multiplexer 1710, 1712 is supplied to an accumulator which includes an adder 1720 and an accumulator register 1722. The accumulator registers output the real and imaginary components of the result. An embodiment of the divider data path 1308 is shown in FIG. 18. Divider data path 1308 includes a normalized floating point divider. The divider data path performs the 1/D calculation required for the Cholesky decomposition and the dot division in the joint detection processing. The divider data path 1308 includes an integer divider 1810, shift logic 1812, a shift generator 1814, XOR (exclusive OR) logic 1820, multiplexers 1822 and 1826, an adder 1824 and latches 1830 and 1832.

The divider 1810 performs an integer divide that returns an integer quotient and a remainder. If the numerator is decimal 1 and the denominator is larger than decimal 1, the integer divider 1810 returns a quotient of 0 and a remainder of 1. To produce an 11-bit 2' s complement quotient, the numerator and denominator are normalized. Each quotient from the divider is post-processed to generate a floating point number that is stored in the JDA memory unit 1304. The floating point number 1/D is represented by a mantissa and an exponent. The mantissa is an 11-bit 2's complement number and the exponent is a 5-bit 2's complement number. The remainder may be used in rounding.

The denominator is normalized to a fixed range to allow a fixed range at the input. In this embodiment, the denominator is normalized to the range (-1, -0.5) to (0.5, 1). The numerator (decimal 1) is normalized such that the output will have 11 significant digits. In the worst case, the LSB of the integer divider output is equal to 2"^"1, where n is the bit width of the 2's complement input. Therefore, to produce an 1 1-bit 2's complement quotient with an 11-bit 2's complement denominator, the normalized numerator is a 2²⁰-bits number. Denominator normalization is done by shift generator 1814 and shift logic 1812.

The 11-bit denominator input is shifted left such that the resulting number is in the range (-1, -0.5) to (0.5, 1). This shift is performed by shifting the denominator left until the first non-sign bit of the 2's complement number is in the bit 9 position. The value of the shift is stored as the exponent of the result. The numerator is always equal to 0.5 in the Q(1.20) format, which is equal to 2¹ in integer format. The numerator is 0.5 to keep the result in the range (-1, -0.5) to (0.5, 1). To account for the fact that the numerator is 0.5 rather than 1, a 1 is added to the shift generator output to produce the exponent.

The entire range of output values fits within the mantissa' s Q(1.10) format except the +1 value. Therefore, the +1 value requires special handling. If the divider 1810 returns a value of +1, the denominator was 0.5, which only occurs if the denominator is a power of 2, specifically 2^"n, where n is [1,9]. In this case, the result is shifted right by 1 and 2 is added to the result of the shift generation to produce the exponent.

Exponent generation logic includes shift generator 1814, XOR logic 1820, multiplexer 1822 and adder 1824. The shift generator 1814 produces an initial shift. The bottom 10 LSBs of the input are XOR'd against the MSB to produce a shift value. The XOR logic 1820 determines if the value output from the divider 1810 is a +1. The XOR logic 1820 returns a "1" if the 2 MSBs of the 12-bit quotient output are 01. In all other cases, the XOR logic 1820 returns a "0". If the XOR logic 1820 returns a "0", the bottom 1 1 bits of the quotient output are returned by multiplexer 1826 as the mantissa and a 1 is added to the shift generator 1814 output by adder 1824 and returned as the exponent. If the XOR logic 1820 returns a "1", bits [11 :1] of the quotient output are returned by multiplexer 1826 as the mantissa and 2 is added to the shift generator 1814 by adder 1824 output and returned as the exponent. Latches 1830 and 1832 hold the mantissa and the exponent, respectively.

The embodiment of the joint detection accelerator described herein utilizes a pipelined architecture including an address generation stage, a data fetch stage, an execution stage and a data write stage. FIG. 19A illustrates the operation of a non- optimized pipeline, and FIG. 19B illustrates the operation of an optimized pipeline. In FIGS. 19A and 19B, row ADD corresponds to the address generation stage, row RD corresponds to the data fetch stage, row EX corresponds to the execution stage and row WR corresponds to the data write stage.

In the optimized pipeline of FIG. 19B, after an initial latency, a data read R together with an execution EX or a writeback WB occur on every cycle. No stall cycles are incurred for the memory accesses after the initial latency.

In the example of FIG. 19B, a stream of data elements is calculated. Each element requires three multiplications EX to finish its calculation. On each cycle, a memory read/execution or memory writeback occurs (see for example cycles 7 through 10). Except for the first and last elements, only 4 cycles are needed to calculate each of the elements, including the memory write cycle. A writeback WB occurs every 4 cycles. No stall cycles are incurred for the memory accesses during the process. Only 14 cycles are required to compute the first three data elements.

In the non-optimized pipeline of FIG. 19A, stall cycles are incurred in memory accesses. As a result, 18 cycles are required to complete processing of the first three data elements. In FIG. 19A, it may be observed that a writeback WB occurs every 6 cycles. The optimized operation shown in FIG. 19B is achieved by appropriate configuration of controller 1350 (FIG. 13). For example, the finite state machine, rather than incurring a stall cycle, is programmed to perform a data access for a data element used later in the process. The data element is stored in buffer 1330 until needed. The use of buffer 1330 together with the JDA controller 1350 effectively removes stall cycles. By removing stall cycles from joint detection processing, the performance of the joint detection accelerator is enhanced. In particular, a given level of performance is achieved at a lower clock frequency, thereby saving power. In some embodiments, the joint detection accelerator may be implemented as an ASIC (application specific integrated circuit). In other embodiments, the joint detection accelerator may be implemented as a FPGA (field programmable gate array). In each case, the joint detection accelerator serves as a coprocessor for the DSP or other host processor.

Various aspects of the present invention may be used alone, in combination, or in a variety of arrangements not specifically discussed in the embodiments described in the foregoing, and the aspects of the present invention described herein are not limited in their application to the details and arrangements of components set forth in the foregoing description or illustrated in the drawings. The aspects of the invention are capable of other embodiments and of being practiced or of being carried out in various ways. Various aspects of the present invention may be implemented in connection with any type of network, cluster or configuration. No limitations are placed on the network implementation. Accordingly, the foregoing description and drawings are by way of example only.

Also, the phraseology and terminology used herein is for the purpose of description and should not be regarded as limiting. The use of "including," "comprising," or "having," "containing," "involving," and variations thereof herein, is meant to encompass the items listed thereafter and equivalent thereof as well as additional items.

What is claimed is:

Claims

1. A joint detection accelerator to perform joint detection of received signals, comprising: a memory unit to store input data values, intermediate results and output data values; an address generation unit to generate addresses for accessing the memory unit; one or more computation units, each comprising a data path, to process the input data values and the intermediate results, and to provide output data values to the memory unit; a data selector to provide the input data values and the intermediate results to the one or more computation units; a controller to control the memory, the address generation unit, the one or more computation units and the data selector to perform joint detection processing; and an external interface to receive the input data values from a host processor and to provide output data values to the host processor.

2. A joint detection accelerator as defined in claim 1 , wherein the one or more computation units include a primary data path comprising a complex multiply accumulate unit.

3. A joint detection accelerator as defined in claim 2, wherein the one or more computation units further comprise a divider data path including a normalized floating point divider.

4. A joint detection accelerator as defined in claim 3, wherein the one or more computation units further comprise a secondary data path including a simplified complex multiply accumulate unit.

5. A joint detection accelerator as defined in claim 1 , wherein a first pipeline stage includes the address generation unit, a second pipeline stage includes the memory unit and the data selector, and a third pipeline stage includes the one or more computation units.

6. A joint detection accelerator as defined in claim 5, further comprising a fourth pipeline stage including a writeback unit for writeback of results produced by the one or more computation units.

7. A joint detection accelerator as defined in claim 1 , wherein the memory unit includes a plurality of independently-accessible memories.

8. A joint detection accelerator as defined in claim 7, wherein the plurality of memories comprises an input memory to store input data values for a time slot, a matrix memory to store a channel estimation matrix and a reordered version of the channel estimation matrix, a main memory to store intermediate results that are calculated during joint detection, and an output memory to store output data values of joint detection.

9. A joint detection accelerator as defined in claim 5, wherein the controller is configured to produce a memory access operation on each clock cycle of an iterative computation, after an initial latency.

10. A joint detection accelerator as defined in claim 1 , further comprising a buffer to hold data read from the memory unit to thereby reduce the number of memory accesses during joint detection.

11. A joint detection accelerator as defined in claim 1 , further comprising a buffer to hold intermediate results to thereby reduce the number of memory accesses during joint detection.

12. A joint detection accelerator as defined in claim 1 , further comprising a buffer to hold data values from the memory unit and intermediate results.

13. A joint detection accelerator as defined in claim 1 , further comprising a register file to hold joint detection parameters used by the controller to perform joint detection.

14. A joint detection accelerator as defined in claim 1 , wherein the controller comprises a finite state machine.

15. A joint detection accelerator as defined in claim 1 , wherein the controller comprises a program sequencer and an instruction memory to hold instructions for performing joint detection processing.

16. A joint detection accelerator as defined in claim 2, wherein the complex multiplier accumulator includes four multipliers, two adders and two accumulators with a shifted input preloadable directly to each of the accumulators.

17. A joint detection accelerator as defined in claim 3, wherein the normalized floating point divider outputs a mantissa and an exponent.

18. A joint detection accelerator as defined in claim 4, wherein the simplified complex multiply accumulate unit includes two data selectors and two accumulators.

19. A joint detection accelerator as defined in claim 4, wherein the controller controls the secondary data path to perform matrix generation of joint detection, controls the divider data path to perform dot division of joint detection and controls the primary data path to perform other operations of joint detection.

20. A joint detection accelerator as defined in claim 1 , wherein the one or more computation units comprise at least two computation units coupled in parallel to different outputs of the data selector.

21. A joint detection accelerator to perform joint detection of received signals, comprising: an input memory to store input data values; at least one intermediate memory to store intermediate results that are calculated during joint detection processing; an output memory to store output data values that are calculated during joint detection processing; a computation block to perform joint detection processing using the input data values from the input memory and the intermediate results from the at least one intermediate memory and to provide the output data values to the output memory; an external interface to supply the input data values from a host processor to the input memory and to supply the output data values from the output memory to the host processor; and a controller to control the input memory, the at least one intermediate memory, the output memory and the computation block to perform joint detection processing.

22. A joint detection accelerator as defined in claim 21 , wherein the at least one intermediate memory comprises a matrix memory to store a channel estimation matrix and a main memory to store intermediate values.

23. A joint detection accelerator as defined in claim 22, wherein the input memory, the matrix memory, the main memory and the output memory each comprise a dual-port memory.

24. A joint detection accelerator as defined in claim 21 , further comprising a buffer to hold data values from the input memory and the at least one intermediate memory, and intermediate results from the computation block.

25. A joint detection accelerator to perform joint detection of received signals, comprising: an external interface to receive input data values from a host processor and to provide output data values to the host processor; a memory unit to store the input data values and the output data values; a primary data path comprising a complex multiply accumulate unit; a secondary data path comprising a simplified complex multiply accumulate unit; a divider data path comprising a normalized floating point divider; and a controller to control the memory unit, the primary data path, the secondary data path and the divider data path to perform joint detection processing.

26. A joint detection accelerator as defined in claim 25, further comprising a data selector to provide data values to the primary data path, secondary data path and the divider data path.

27. A joint detection accelerator as defined in claim 25, wherein the controller controls the secondary data path to perform matrix generation, controls the divider data path to perform dot division and controls the primary data path to perform other operations of joint detection processing.

28. A joint detection accelerator to perform joint detection of received signals, comprising: an external interface to receive input data values from a host processor and to provide output data values to the host processor; a memory unit to store the input data values and the output data values; an address generation unit to generate addresses for accessing the memory unit; a computation block to perform joint detection processing using the input data values from the memory unit and to provide the output data values to the memory unit; a controller to control the memory unit, the address generation unit and the computation block to perform joint detection processing, wherein the controller is configured to produce a memory access operation on each clock cycle of an iterative computation after an initial latency.

29. A joint detection accelerator as defined in claim 28, further comprising a buffer to hold data values from the memory unit and intermediate results from the computation block.

30. A joint detection accelerator as defined in claim 28, wherein the controller comprises a finite state machine.

31. A joint detection accelerator as defined in claim 28, wherein the controller comprises a program sequencer and an instruction memory to hold instructions for performing joint detection processing.