Method and system for performing calculation operations and a device
A method for performing calculation operations uses a pipelined calculation device comprising a group of at least two pipeline stages, at least one data interface for input of data, and at least one data interface for output of data. The pipeline stages include at least one data interface for input of data and at least one data interface for output of data. Data for performing a first and a second calculation operation is input to the device. In the first calculation operation, output data of at least one pipeline stage is stored into a memory. In the second calculation operation the stored data is used as input data to a pipeline stage. The invention further relates to a system and a device, in which the method is utilized.
Latest Nokia Corporation Patents:
Priority is claimed under 35 U.S.C. § 119 from Finnish application FI 20021984 filed Nov. 6, 2002.
BACKGROUND OF THE INVENTION1. Technical Field
The present invention relates to a method for performing calculation operations using a pipelined calculation device comprising a group of at least two pipeline stages, at least one data interface for input of data, and at least one data interface for output of data, said pipeline stages having at least one data interface for input of data and at least one data interface for output of data, in which method data for performing a first and a second calculation operation is input to the device. The invention also relates to a device and a system for performing the method.
2. Discussion of Related Art
There are many applications in which multiplication, multiplyaccumulate (MAC) and other calculation operations are needed. As a nonrestrictive example many signal processing applications, such as digital signal filtering applications, video/image processing applications etc., are applications in which real time multiplication operations are implemented. Also other applications in which vector and/or matrix operations are needed use multiplication and MAC operations. Multiplication operations are normally implemented as summation and bitwise shifting operations. Such multiplication operations are resource demanding tasks because one multiplication operation between two operands needs many summation and shifting operations to calculate the multiplication result.
Specific to video/image processing algorithms is the vast amount of computations that have to be implemented in real time. Therefore, the highspeed performance has been the driving force in developing parallel specialized structures (accelerators) for different video/image processing algorithms or subtasks. A prior art video/image processing system involves several such accelerators (e.g. Sum of Absolute Differences (SAD), cosine transform, etc.) each being composed of large amount of hardware elements. However, with developing mobile communication systems, the hardware area, which affects to the costs of the system, and the power/energy consumption are as important properties as the highspeed performance. One way towards satisfying all these criteria is further modernization of Digital Signal Processors (DSPs) and reduction of the number of specialized accelerators. Although some improvements exist on this area the developed systems still does not always meet the highspeed and power consumption requirements.
Table 1 summarizes some core arithmetic patterns along with examples of video/image processing algorithms where these patterns are frequently implemented. Operations involved in these patterns are very basic ones and are very familiar. A vast amount of the literature is devoted to their implementations. Here two specifics are emphasized concerning these operations in the video/image processing context. First, operands of operations are most often medium (8 to 16bit) precision integers. Secondly, most of the algorithms use massively parallel operations. In some cases these parallel operations share the same operands. For example, in a scalar quantization, the same number is multiplied to many pixels of the image, in a matrixvector product different rows of the matrix are multiplied to the same input vector, in Finite Impulse Response (FIR) filtering the same coefficients are involved in a number of MAC operations, etc.
The prior art architectures for video and image processing and other signal processing tasks are usually based on conventional multiplier structures. Many multiplication methods and a very large variety of devices for implementing multiplication and/or multiplyaccumulate operations have been developed. In the following, only multiplication methods and general multiplier structures for the case where both the operands (the multiplicand and the multiplier) are unknown, multiplication of two fixedpoint signed integers presented in two's complement arithmetic, and the so called radixT methods will be considered.
The two's complement representation of the nbit (including the sign) multiplier a will be denoted as ã=a_{n−1}a_{n−2 }. . . a_{1}a_{0}, and the two's complement representation of the mbit (including the sign) multiplicand x will be denoted as {tilde over (x)}=x_{m−1}x_{m−2 }. . . x_{1}x_{0}, respectively. The relation between a and ã (and a similar relation between x and {tilde over (x)}) is as follows:
In a radixT parallel multiplication method the two's complement {tilde over (y)}=y_{m+n−1}y_{m+n−2 }. . . y_{1}y_{0 }of the product y=a·x is obtained according to the formula
in the following two main steps:
 Step 1. Generation of partial products (PP) A_{r}·x, r=0, . . . , n_{radixT}−1 such that Equation (2) is valid.
 Step 2. Summing up in parallel all the partial products preliminary shifting the rth partial product A_{r}·x, r=0, . . . , n_{radixT}−1 for rt_{radixT }positions to the left.
A radixT MAC unit operates in a similar way with the difference that another number (accumulating term) is added along with partial products at the Step 2.
The Step 1 will now be considered in more detail. Depending on how the numbers A_{r}·x, r=0, . . . , n_{radixT}−1 are defined and obtained, different multiplication methods can be derived. In turn, the choice of the numbers A_{r}·x, r=0, . . . , n_{radixT}−1 is, in fact, dictated by representation of the multiplier a. The simplest multiplication method is a radix2 method, which uses the basic two's complement representation of a given in the left equation of (1). In this case, the two's complement of the product will be obtained as:
that is, n_{radix2}=n, and the partial products A_{r}·x, r=0, . . . , n−1 are defined by A_{r}=a_{r }for r=0, . . . , n−2, and A_{n−1}=−a_{n−1 }for r=n−1. These partial products may simply be (and are usually) formed using an array of 2input AND gates between every two's complement bit of the multiplier ã and the multiplicand {tilde over (x)}. The value of A_{r}·x, r=0, . . . , n−1 is multiplied to 2^{r }(i.e. is shifted to the left for r positions) before being accumulated at the second step. It should be noted that in this method the partial product A_{n−1}·x, which sometimes is also called a correction factor, is treated differently from the other partial products.
The nonuniform nature of the partial products is avoided in another radix2 multiplication method based on Booth recoding of the two's complement bits ã of the multiplier into redundant signed digit numbers. The product can now be presented as:
That is, n_{radix2}=n, as before but the partial products A_{r}·x, r=0, . . . , n−1 are all now defined by A_{r}=−a_{r}+a_{r−1}. Similarly to the previous method, the value of A_{r}·x, r=0, . . . , n−1, is multiplied to 2^{r }before being added at the second step. In this scheme the partial products are selected among 0,±x. The two of these values (0 and x) are readily available while finding −x requires inverting the bits of {tilde over (x)} and adding unity. Normally, addition of unity is performed in the Step 2 where the partial products are summed.
There are totally n_{radix2}=n partial products to be summed in a radix2 multiplication method irrespective of if the Booth recoded or nonrecoded method is used. In order to reduce the number of partial products and, hence, the delay of the second stage (summing up partial products), the radix4 Modified Booth Algorithm (MBA) based method has been developed. The MBA is one of the most popular multiplication methods and is being extensively studied and optimized.
In order to simplify the formulae below, in every case where a term like n/k occurs, it is assumed that n is an integer multiple of k. This is a valid assumption since a two's complement number may be complemented with an arbitrary number of bits (repeating the most significant bit).
In MBA, the two's complement of the product is obtained as the sum
of n_{radix4}=n/2 partial products, where the value of A_{r}∈{−2,−1,0,1,2}, r=0, 1, . . . , n/2−1, is chosen according to three consecutive bits a_{2r+1},a_{2r},a_{2r−1 }(a_{−1}=0) of the two's complement representation of the multiplier ã. The partial product A_{r}x, r=0, 1, . . . , n/2−1, is multiplied by 2^{2r }(i.e. hardware shifted to the left for 2r positions) before being added at the Step 2.
It is also possible to use radices higher than 2 with nonrecoded multiplication methods for reducing the number of partial products. For example, in radix4 nonrecoded multiplication method the partial products A_{r}x, A_{r}∈{0,1,2,3}, r=0, 1, . . . , n/2−1 are chosen according to two consecutive bits a_{2r+1},a_{2r }of the multiplier. There are n_{radix4}=n/2 partial products in this method. The potential partial product 2x can be generated by shifting potential partial product x once to the left. The odd partial product 3x needs an additional summation of x. If multiplications of negative numbers are also used, the sign extension must be used, in which the most significant bit (i.e. the sign bit) of each partial product is copied as many times as is necessary to achieve the required bitlength.
In radix8 nonrecoded multiplication method the partial products A_{r}x A_{r}∈{0,1,2,3,4,5,6,7}, r=0, 1, . . . , n/3−1, are chosen according to three consecutive bits of the multiplier. The list of potential partial products is 0, x, 2x, 3x, . . . , 7x all of which become available by implementing three independent additions/subtractions in order to obtain 3x=x+2x, 5x=x+4x, 7x=8x−x. The potential partial product 6x can be formed by shifting the potential partial product 3x one position to the left. For the cases of higher radices (>=16), however, there are some potential partial products (e.g. 11x and 13x) that cannot be obtained in one addition/subtraction.
In order to further reduce the number of partial products, the Booth encoding has further been extended to multibit (arbitrary radixT) recoding. The general equation for the product is now given as:
That is, there are n_{radixT}=n/t partial products (T=2^{t}) and every partial product is selected according to t+1 consecutive bits of the multiplier ã from the list of potential partial products Ax with A ranging between −2^{t−1 }and 2^{t−1}. Every potential partial product may be relatively easily formed by addition of two (for T=8,16) or more (for T>16) powerof2 multiples of x, and, possibly, inversion of the bits followed by addition of 1 (at Step 2). For example, in the case of radix8 recoding, the list of potential partial products is 0, ±x, ±2x, ±3x, ±4x. All the nonnegative multiples from this list are readily available except for 3x which may be obtained in one addition: 3x=x+2x. Negative multiples may be found by invertadd1 method, as before. In the case of radix16 recoding, the list of potential partial products is 0, ±x, ±2x, ±3x, . . . , ±8x all of which become available by implementing three independent additions/subtractions in order to obtain 3x=x+2x, 5x=x+4x, 7x=−x+8x. The potential partial product 6x can be formed by shifting the potential partial product 3x one position to the left. For the cases of higher radices, however, there are some potential partial products (e.g. 11x and 13x) that cannot be obtained in one addition/subtraction.
The array of adders of a typical prior art radixhigherthanfour multiplier comprises s adders/subtractors, where s is the number of odd positive multiples of x involved in the list of potential partial products (s=1 in the cases of T=8 Booth recoded and T=4 nonrecoded multipliers, and s=3 in the cases of T=16 Booth recoded or T=8 nonrecoded multipliers, etc.). Usually, fast carrylookahead (CLA) adders are used since forming the list of potential partial products is rather time consuming part of such multipliers. In a patent U.S. Pat. No. 5,875,125 a special x+2x adder has been proposed which may be used in radix8 multipliers. It should be noted that mixed radix4/8 multipliers have also been proposed, for example in U.S. Pat. No. 4,965,762, which, however, are mainly useful for iterative (not parallel) multipliers where the partial products are generated and accumulated serially. U.S. Pat. No. 5,646,877 describes a multiplier structure where all the potential partial products for an arbitrary radix are obtained as sums or differences of shifted versions of 3x and of x within the array of adders comprises an x+2x adder for generating 3x, two shifters and an adder/subtracter.
The selection block of a typical prior art radixhigherthanfour multiplier comprises n/t radixT Booth encoders and equal number of decoder rows. Each encoder analyzes the corresponding (t+1)tuple of the multiplier and outputs a plurality of control signals according to which the corresponding partial products are formed by the decoder rows. Remarks on how to extend the radix4 Booth encoders and decoders to higher radices are given, for example, in a patent U.S. Pat. No. 6,240,438.
In the following, the summing up the partial products, i.e. the Step 2, will be considered in more detail. Most of the parallel multiplier/MAC unit structures use summation blocks composed of a compression array followed by a fast adder (final adder) for summing up the partial products formed at Step 1 (see
It should be noted that, if a Booth recoding scheme is utilized, then, as a result of performing the unity addition in Step 2 instead of in Step 1, every partial product row is accompanied with a onebit value, which is zero if the partial product is a nonnegative multiple of the multiplicand and is unity otherwise. Thus, actually, the number of rows is 2n_{radixT}. Even though these onebit values may be merged into partial product rows in such a way that the number of rows is again n_{radixT }or maybe n_{radixT}+1 but with the price of increasing the length of partial product rows (by one bit) and making them irregular. In a nonrecoded scheme, there is at most one extra onebit value so simpler compression arrays can be designed.
Another problem associated with the summation block in a Booth recoded multiplier is how to handle sign extensions since the partial product rows are shifted with respect to each other before adding them. In a straightforward implementation every partial product (after shifting) should have been extended to a (n+m)bit number, which is very wasteful approach. Special sign extension methods and circuits have been developed to reduce the number of sign extended bits to two in every row. In the case of nonrecoded multipliers, sign extensions may be handled easier, with no extra sign bits, since all but possibly one partial products are of the same sign.
There are principally two ways of extending multiplier structures to MAC units as depicted in
As a summary of highradix multipliers, it should be noted that the higher the radix the higher the complexity of the Step 1 (generating partial products) but lower the complexity of the Step 2 (summing up partial products). The “radixhigherthanfour” multiplication methods have not gained popularity, perhaps due to the necessity of having rather time and area consuming partial product generators, including both the array of adders and the selection block. Commonly, the radix4 MBA is considered the best prior art parallel multiplication method and is used in many industrial multipliers.
A method called pipelining can be used in connection with calculation operations. Hence, a device utilizing pipelining comprises two or more pipeline stages. Each pipeline stage is intended to perform a certain part or parts of calculation operations (i.e. suboperations). In the prior art the calculation operations of the pipeline stages relate to each other so that each pipeline stage performs one or more suboperations of the calculation operation to be performed, and the output of the last pipeline stage provides the result of the calculation operation. In such a device different pipeline stages operate in succession, where the next pipeline stage begins the calculation of the suboperation after the previous pipeline stage has finished the calculation of its suboperation. If pipeline stages are poorly balanced (i.e. some stages are significantly faster than the others), this means that all but one pipeline stage is waiting or is in an idle state most of the time. Furthermore, all the pipeline stages are reserved for a certain task (calculation of a certain suboperation) and they cannot be used for performing other calculation operations.
In the following, some multiplier/MAC unit features, which are desired from a video and image processing point of view but are absent or weak in prior art solutions, will be presented. First, the most popular radix4 Booth recoded multiplier/MAC method will be considered. As a general drawback of this method it is more power consuming than higher radix methods. Another general drawback is that, even though the number of partial products is reduced to half compared to radix2 multiplication, it still could be reduced using higher radices. That is, the complexity of this method is mainly concentrated in the Step 2 (summing up partial products). When pipelining a radix4 Boothrecoded multiplier/MAC structure, usually the partial product generation block is considered as the first pipeline stage which, however, is poorly balanced with (i.e. faster than) the other pipeline stages.
Considering the “radixhigherthanfour” Booth recoded multipliers, it has been shown that different realizations of these multipliers, when considered for implementation of only the multiplication operation, perform competitively to the radix4 multipliers with respect to the area and time criteria while outperforming those with respect to the power consumption. The main drawback of the “radixhigherthanfour” methods is the necessity of having an array of adders at the partial product generation block.
The Booth recoded “radixhigherthanfour” multipliers also have a drawback related to the necessity of handling the negative multiples of the multiplicand as well as the sign extensions.
A radixT nonrecoded multiplier involves the same number of adders in the potential partial product generation block as the radix(2T) Booth recoded one. When “radixhigherthaneight” nonrecoded or “radix−higherthansixteen” Booth recoded multipliers are used, more than onelevel addition is needed to generate potential partial products.
Table 2a presents various characteristics of different blocks used in prior art Booth recoded radixT multipliers/MAC units for multiplying an nbit multiplier and an mbit multiplicand, and Table 2b presents various characteristics of different blocks used in new Nonrecoded radixT multipliers/MAC units for multiplying an nbit multiplier and an mbit multiplicand, and Table 2b presents various characteristics of different blocks used in prior art Nonrecoded radixT multipliers for multiplying an nbit multiplier and an mbit multiplicand, respectively. Table 2c presents the acronyms utilized in Tables 2a and 2b. Analyzing Tables 2a and 2b, one can see that there are essential differences in the delays of different blocks of every multiplier in every multiplier/MAC unit type for most of the values of n and m. That is, a direct pipelined realization of these multipliers would suffer from poor balancing between pipeline stages. Trying to achieve better balancing between these pipeline stages, one could flexibly increase the throughputs of the first and the last pipeline stages by designing carrylookahead (CLA) adders with different numbers of FAs within one carryripple block. This is why the delays of these blocks are not indicated in Tables 2a and 2b. In some cases, e.g. small n, higher radix T, this may mean very small carryripple blocks, and therefore large area. Anyhow, using CLAs may be a solution for speeding up these two stages, though this solution is not the most efficient. The situation is different for the middle two pipeline stages (the SB and the CA) since the stateoftheart circuits for these blocks are designed by optimizing critical paths in such a way that internal pipelining of these circuits would not be reasonable. Moreover, relative differences between the delays of these blocks are very different for every type (BR (T=8,16), NR1 (T=4,8)) of prior art multiplier/MAC unit structures for different values of n. For the cases of smaller n and higher radix T (e.g., T=16 and arbitrary n) the Selection Block is slower than the Compression Array while in other cases the situation is the opposite. This means that designing faster circuits for one or another block will not give a general solution to the problem of balancing these stages. All of these make it difficult to derive a systematic method for pipelining prior art multiplier/MAC structures with wellbalanced pipeline stages.
On the other hand, if a parallel array of independent prior art multipliers/MAC units were to be employed to implement a plurality of corresponding operations, a large silicon area would be required even though the faster blocks could be shared without affecting overall throughput. In addition, as illustrated above, the pipeline stages would be poorly balanced if pipelined multipliers/MAC units were to be employed in such an array.
Prior art parallel architectures for matrixvector arithmetic are designed so that they use a plurality of independent multipliers combined with adders or a plurality of independent MAC units. It is also usual to have a plurality of specific circuits for different kind of arithmetic operations. However, high radix multipliers include blocks (pipeline stages) that may be reused for other operations such as additions/subtractions, accumulations, etc. It should also be mentioned that, in matrix vector arithmetics, especially those used in video/image processing, there are many situations where one multiplicand is to be multiplied with several multipliers meaning that potential partial products of the multiplicand could have been reused in all of these multiplications if highradix multiplication methods would be used. For example, in vector and matrix calculation operations, in which the same multiplier will be used for many times, the same partial products will be used more than once during the calculation. When such calculation operations are implemented in prior art architectures, they calculate the partial products each time they are needed, leading to inefficient utilization of hardware resources and additional power consumption. In addition, as was mentioned above, there are many situations in which pipeline stages of prior art multipliers are poorly balanced thus reducing the efficiency of the device.
OBJECTIVES AND SUMMARY OF THE INVENTIONAn aim of the present invention is to implement an improved calculation structure and method for performing calculation operations with the capability to store intermediate results of the calculation operations, especially the capability to store potential partial products of a multiplication operation. In the method according to the present invention output data of at least one calculation stage in an operation, e.g., a pipeline stage is stored into a memory, wherein the stored data is used as input data with another calculation operation. In addition to storing the potential partial products, the memory of the calculation device can be used with many calculation operations. An advantageous embodiment of the calculation device according to the present invention is multifunctional/configurable to several different structures by a set of control signals. The structures of the present invention can advantageously be implemented as a VLSI architecture.
There are different embodiments of the present invention some of which will be illustrated in more detail later in this description.
In the first embodiment, a memory unit is utilized in which the potential partial products can be stored and reused in those cases where the same multiplicand is to be multiplied to a number of multipliers. Nonlimiting examples of such operations are matrixvector multiplication, FIR filtering, etc. This way, only the first of the number of multiplications having the same multiplicand is completely implemented while the others are only partially implemented avoiding the most resource demanding part of computing the potential partial products. For example, when multiplying a matrix to a vector, complete multiplications are implemented only for computing the first component of the output vector while the rest of the multiplications for computing the other components of the output vector are implemented only partially. When implementing FIR filtering, complete multiplications are implemented only once, when computing the first output sample, but the rest of the multiplications for computing the other output samples are implemented incompletely.
The second embodiment utilizes a parallel multiplier structure in matrixvector arithmetic and at the same time is flexible, giving an opportunity to configure it to smaller tasks.
The structure of the second embodiment of the invention provides a structure that either may be realized as a device that may be configured to implement a list of operations including: (a) a plurality of multiplications; (b) a plurality of MAC operations; (c) a plurality of additions/subtractions of different precision (bitwidth); and/or (d) parallel accumulation.
The third embodiment is a combination of the parallel multiplier structure and a sum of absolute differences (SAD) accelerator structure for video and image processing (MAVIP). The structure is a multifunctional architecture in such a way that it can be configured to different architectures by a set of control signals.
To be exact, the method according to the present invention is primarily characterized in that in said first calculation operation, output data of at least one pipeline stage is stored into a memory, and in said second calculation operation the stored data is used as input data to a pipeline stage. The system according to the present invention is primarily characterized in that the system further includes a memory for storing in said first calculation operation output data of at least one pipeline stage, and a data retriever for using the stored data in said second calculation operation as input data to a pipeline stage. The device according to the present invention is primarily characterized in that the device further includes a memory for storing in said first calculation operation output data of at least one pipeline stage, and a data retriever for using the stored data in said second calculation operation as input data to a pipeline stage.
The present invention shows remarkable advantages compared to solutions of prior art. A VLSI architecture suitable for efficient implementation of a wide range of arithmetic patterns frequently implemented in signal processing applications such as video/image processing algorithms can be developed. By using a method according to the invention the multiplication operations in the context of video/image processing can be performed faster and with less power than with prior art methods and devices. The invention is especially efficient for implementing a plurality of multiplications/MACs, and for implementing a plurality of multiplications/MACs with one shared operand (shared multiplicand). Significant reduction in the area and power consumption can be achieved due to the ability to store the potential partial products and use them with successive calculation operations. Also the ability of another advantageous embodiment of the present invention to set the different pipeline stages to power saving mode when they are not needed and/or to use different pipeline stages to perform different calculation operations substantially simultaneously results in reduced power consumption. High speed performance (high throughput) is due to utilization of pipelining with a flexible possibility of balancing the delays of different pipeline stages and reducing the period of the total structure to the delay of the fastest block. Increasing of the throughput is achieved with a minimal increase in the area and without increasing the input/output bus width as compared to a single pipelined prior art (radix4 or higher) multiplier.
The memory unit also allows utilization of the first pipeline stage for another task in parallel with the other stages, that is, simultaneously with implementing a multiplication based operation, parallel addition/subtraction operations (of varying precision) may be implemented for another task.
Still another advantage of one advantageous embodiment of the present invention is that it is not necessary to balance the first pipeline stage of the architecture, i.e. the arrays of adders for computing potential partial products, with other stages since it is out of the critical path during most of the cycles.
The configurable structure according to an advantageous embodiment of the present invention makes it possible to reuse the same device for different operations. Moreover, operations may be implemented with different bitprecision of input arguments, e.g. eight parallel 8bit additions or four parallel 16bit additions.
Significant time and power/energy consumption savings are achieved since the most costly part, the computation of the potential partial products is discarded from most of the multiplication operations.
Being multifunctional, an advantageous embodiment of the present invention can replace several hardware accelerators specialized for a certain signal processing operations, for example, video/image processing operations.
In the following the present invention will be described in more detail with reference to the appended drawings, in which
In the following, the first advantageous embodiment of the present invention will be described in more detail. The general structure of the first embodiment of the device according to the present invention is illustrated on
Pipelining is a common way to increase the throughput of the system for implementing a plurality of similar operations, e.g. multiplication and/or MAC. Considering a typical pipelined realization of a prior art radixT Boothrecoded (T=8,16) or nonrecoded (T=4,8) multiplier structures, the first pipeline stage P1 would be the Array of s Adders with s given in Table 2a, the second stage P2 would be the Selection Block, the third stage P3 would be the Compression Array, and last several stages P4 (their number being different in different realizations) would constitute the Final Adder.
In one realization of the device according to the first advantageous embodiment of the present invention incorporates a plurality (bank) of multipliers/MAC units. The general structure of the variety of banks of radixT Booth recoded (T=8,16) or nonrecoded (T=4,8) multipliers is schematically shown on
The principle of timeinterlaced operation of a plurality of functional blocks is briefly explained below. Consider a functional block FB to be used in a pipeline stage of a bigger system, and suppose that it has an estimated delay of D_{FB}≈pT_{des }(but D_{FB}≦pT_{des}) where T_{des }is the desired operating step duration of the system. Then, in order to make the pipeline stage P1, . . . , P4 to operate with the throughput matching the duration of one step, the serial Functional Block may be replaced with a plurality FB_{1}, FB_{2}, . . . , FB_{w }of w Functional Blocks having pipeline registers (latches) R_{1}, . . . , R_{w }at their inputs and a w:1 multiplexer 10, 11, 12, 13 at their outputs. The parameter w depends on the pipeline stage in question. For the first pipeline stage P1 the parameter w corresponds to the parameter p i.e. there are p pipeline registers R_{1}, . . . , R_{p }and the multiplexer 10 of the first pipeline stage selects one of the p inputs for its output. For the second pipeline stage P2 the parameter w corresponds to the parameter q i.e. there are q pipeline registers R_{1}, . . . , R_{q }and the multiplexer 11 of the second pipeline stage P2 selects one of the q inputs for its output, respectively. For the third pipeline stage P3 the parameter w corresponds to the parameter u i.e. there are u pipeline registers R_{1}, . . . , R_{u }and the multiplexer 12 of the third pipeline stage P3 comprises two multiplexers (not shown) each of which select one of the u inputs for its output. For the fourth pipeline stage P4 the parameter w corresponds to the parameter v i.e. there are v pipeline registers R_{1}, . . . , R_{v }and the multiplexer 13 of the fourth pipeline stage P4 selects one of the v inputs for its output 414. The shared input to the pipeline stage P1, P2, P3, P4 is connected to everyone of w pipeline registers R_{1}, . . . , R_{w}. However, everyone of these registers is open only one out of every w operating steps with a time offset of one operating step. This way, at every operating step, the input I is actually connected to the input of only one Functional Block while the other connections are inactive since the input registers of all the other Functional Blocks within their plurality forming the considered pipeline stage are closed for writing. The register R_{1 }at the input of the first functional block FB_{1 }is open during the first of every w operating steps, the register R_{2 }at the input of the second functional block FB_{2 }is open during the second of every w operating steps, etc. In general, at the operating step t=1, 2, . . . , K, the input I is actually connected to the input of only the FB_{ƒ}, where ƒ=(t−1)mod w+1. Thus, the input I is spatially shared but temporally distributed among the functional blocks FB_{1}, . . . , FB_{w }of the considered pipeline stage. During the operation of the system, at the operating step t=1, 2, . . . , K, one sample X_{t }of the input data stream to that stage enters to the functional block FB_{ƒ}, ƒ=(t−1)mod w+1 which is initiated at that stage. Once the FB_{ƒ}completes its operation over X_{t }at the end of the operating step t+w−1, t=1, 2, . . . , K, the w:1 multiplexer 10, 11, 12, 13 at the outputs of the functional blocks passes the results obtained in FB_{ƒ}to the output of the stage at the next operating step t+w, t=1, 2, . . . , K. For this reason the multiplexer 10, 11, 12, 13 operates according to circular mod w rule.
By replacing every functional block of a prior art multiplier/MAC unit structure with a plurality of similar functional blocks and by making use of the principle of timeinterlaced operation, a better balancing of the pipeline stages may be achieved with respect to a predetermined operating step duration by appropriate selection of the numbers of blocks within every pipeline stage (that is, parameters p, q, u, and v). Suppose that the delays of the four pipeline stages within a prior art multiplier/MAC unit structure are D_{AA}, D_{SB}, D_{CA}, D_{CLA}, respectively, and the desired throughput of the system is N_{des }multiplications MAC/operations per second, then the desired operating step duration will be chosen to be T_{des }seconds, so that T_{des}≦1 N_{des}, and the mentioned design parameters will be selected as p=┌D_{AA}/T_{des}┐, q=┌D_{SB}/T_{des}┐, u=┌D_{CA}/T_{des}┐, and v=┌D_{CLA}/T_{des}┐, in which the notation ┌x┐ means rounding up to the next integer value.
When all the control signals c_{1}, c_{2}, c_{3 }are set to logical 1state, i.e. c_{1}=c_{2}=c_{3}=1, the structure behaves as a multiplication/MAC operation device. In a situation in which the first control signal c_{1 }is in logical 0state and the other two control signals c_{2}, c_{3 }are in logical 1state, i.e. c_{1}=0; c_{2}=c_{3}=1, the structure behaves as an array of sp adder/subtracters. Only the first pipeline stage P1 (i.e. the array of adders 2) is activated while the others are idle. The third alternative is that the first control signal c_{1}, is set to logical 1state while the second c_{2 }and the third control signals c_{3 }are set to logical 0state (c_{1}=1; c_{2}=c_{3}=0). This means that the structure behaves as an accumulator. Only compression arrays 4 and the final adders 5 operate while the arrays of adders and the selection blocks are idle. In the fourth alternative the first c_{1 }and second control signals c_{2 }are both set to logical 1state and the third control signal c_{3 }is set to logical 0state (c_{1}=c_{2}=1: c_{3}=0). In that situation the structure behaves as a fast adder of a wide precision. Yet in another alternative, when c_{1}=c_{2}=c_{3}=0, the structure simultaneously behaves as an array of sp adders/subtractors and an accumulator. It should be evident that the above mentioned selection means 16, 17, 20 as well as the dependence of the structure/operation modes on the states of the control signals c_{1}, c_{2}, c_{3 }are just one possible alternative but also other alternatives can be applied within the scope of the present invention. This also applies to other control signals which will be described later in this description.
In the structure of
Architectures for MatrixVector Arithmetic
In the following, the advantageous embodiments of the multiplier/MAC unit structures according to the present invention will be used for performing matrixvector operations such as a scalar to a vector or to a matrix multiplication, a matrix to vector multiplication (and thus matrixto matrix multiplication, an inverse matrix computation, color conversions, fast orthogonal transforms, e.g., FFT, fast DCT, etc.), and convolution or FIR filtering, and thus filter bank and, in particular, discrete wavelet transform implementations. Common to these operations is that one multiplicand x is multiplied with a number of multipliers a_{i}, i=1, . . . , K. On the other hand, an essential feature of the high radix multiplication methods is that a significant complexity is moved to the first step (forming partial products) from the second step (summing up partial products). More complexity is moved with higher radices.
The main idea of this advantageous application is to compute potential partial products of the common multiplicand x only once when multiplying it by the first multiplier a_{1 }and then reuse these potential partial products when multiplying x with the rest of multipliers a_{i}, i=2, . . . , K. This way, only one out of K multiplications will be completely executed while K−1 multiplications will be executed incompletely, without executing the most complicated part of a highradix multiplication method (that is the computation of the potential partial products). Clearly, significant savings in time and power/energy consumption will then be achieved.
According to this idea, the general structure of the proposed architecture for matrixvector arithmetic can be derived from the multiplier/MAC unit structure of the first as well as from the reconfigurable device of the second advantageous embodiments of the present invention. This is achieved by utilizing the memory 21 as depicted on
Actually, any read/write memory can be used in the proposed structure for matrixvector arithmetic. Two recommended memory types are depicted on
The second memory type depicted in
In the following, execution of three operations frequently implemented in video and image processing will be considered in more detail. It should be noted that these are just example implementations for only particular operations but the present invention is not restricted to only these operations nor exactly to the presented methods. All the other operations that involve multiplication of a common multiplicand by a number of multipliers may similarly be executed by the architecture according to the invention for matrixvector arithmetic. Also other similar algorithms for the considered operations may be developed.
First, execution of the scalar to a vector multiplication operation xa is considered, which is the most evident case where a common multiplicand (scalar) x is to be multiplied with a big number of multipliers a_{i}, i=1, . . . , K, the components of the vector a. At the first step, when implementing multiplication of the common multiplicand with the first multiplier a_{1}, the sixth control signal c_{6 }is set to low so that potential partial products of x directly pass from the arrays of adders 2 of the first pipeline stage P1 to the plurality of selection blocks 3 of the second pipeline stage P2, and, at the same time, are written into the memory 21. The memory 21 is set to operate according to mod 1 mode and all its registers except for the 0th one are halted. All but the first array of adders 2a of the first pipeline stage are also halted if multiplication of only one scalar to one or to several vectors is to be implemented. Starting from the second operating step, the sixth control signal c_{6 }is set to high so that the memory 21 is closed for writing and potential partial products of x, that have been written in the memory after the first step, are fetched from the memory 21 to the selection blocks 3 through the second input of the corresponding 2:1 multiplexer 22. At the same time, all the arrays of adders 2 (including the first one) are either halted or switched to another task. Thus, time savings are achieved because of the reduction of the critical path of the multiplication operation by discarding the most costly part, namely, the computation of potential partial products. Significant savings in power and energy consumption are achieved because the first pipeline stage is halted or is switched to another task during K1 out of K operating steps. Another advantage is that, in this computational scheme, it is not necessary to balance the first pipeline stage P1 with other pipeline stages P2 to P4 since it is out of the critical path during K1 out of K operating steps. Yet another advantage is that the multiplicand x is fetched only once from the external memory (not shown) while in a conventional implementation it would have to be fetched K times.
The second operation that will be considered here is the matrixvector multiplication

 s=Ax, or equivalently,
This operation is a part of many linear algebra and video/image processing algorithms such as matrixtomatrix multiplication, inverse matrix computation, image color space conversions, fast orthogonal transforms, (e.g., FFT, fast DCT, etc.), geometric manipulations, edge detection, and others. The matrixtovector multiplication could be considered as multiple vectortovector multiplications and so it could be implemented with a MAC unit or with a bank of them. However, it should be noted that every component (multiplicand) x_{j}, j=1, . . . , P, of the input vector x is multiplied by K multipliers, components of the jth column of the matrix A. Utilization of the memory 21 within the structure for matrixvector arithmetic according to this embodiment of the present invention allows for computation of potential partial products only once for every component x_{j}, j=1, . . . , P, of the input vector x. The potential partial products are then saved into the memory 21 provided that it contains enough registers (i.e., P≦R).
As is known from graph theory the critical path mentioned above expresses the longest route from a starting node to an end node. When applied with the present invention the critical path expresses the longest route from the beginning of the calculation process to the end of the calculation process in the pipelined system. The length of the critical path describes the time taken for performance of the calculation operation. By using critical path analysis it is possible to find such parts of the system which affect to the length of the critical path. When such parts are found they can be analyzed to find if there exist any possibilities to shorten the length of such parts. Respectively, it is normally not necessary or useful to try to shorten such parts which do not belong to the critical path.
In order to initialize multiplication of a (K×P)matrix to a vector the structure according to an advantageous embodiment of the present invention is configured by setting appropriate values for the control signals as follows. The memory 21 is set to operate according to mod P mode and the registers from P to R of the memory are halted. The compression array 4 of the architecture is set to operate as an accumulator by setting the fifth control signal c_{5 }to high. The sixth control signal c_{6 }is set to low in the beginning while potential partial products for every component x_{j}, j=1, . . . , P, are formed during P operating steps t=p+1, . . . , p+P, where p steps is the delay of the first pipeline stage which is equal to the number of arrays of adders within it. During these operating steps potential partial products for component x_{j}, j=1, . . . , P, are directly passed from the arrays of adders 2 of the first pipeline stage P1 via the multiplexer 22 to the plurality of selection blocks 3 of the second pipeline stage P2, and, at the same time, are written into the memory 21. Thus, multiplication of the first row of the matrix A by the input vector x is implemented, and substantially simultaneously, potential partial products for every component x_{j}, j=1, . . . , P, are stored in consecutive registers. After the operating step p+P, the control signal c_{6 }is set to high so that the memory 21 is closed for writing and potential partial products for x_{j}, j=1, . . . , P, are circularly fetched to the input of the plurality of selection blocks 3 from the memory 21 through the second input of the 2:1 multiplexer. Also, the fifth control signal c_{5 }is set to low for one step in order to restart new accumulation. This is repeated K−1 times until all the rows of the matrix A will be multiplied with the input vector x. Starting from the operating step p+P, all the arrays of adders 2 (including the first one) are either halted or switched to another task. Thus, only P out of PK multiplications are completely implemented while P(K−1) of them are implemented partially, without the first, most complicated step. Similarly to the case of the scalartovector multiplication operation, this leads to significant savings in time and power/energy consumption.
In the following, convolution or FIR filtering implementation on the structure according to the present invention will be considered in more detail. The operation to be performed is
where usually P<<K. Two strategies are possible to implement FIR filtering on the structure according to the present invention for matrixvector arithmetic.
In the first strategy the FIR filtering will be considered as multiplication of the Pdiagonal (or Ptaped) matrix A with the input vector, which is appended with P components x_{i}, i=−P+1, . . . , 0 corresponding to negative indices and is shifted for P positions down. Taking into account that the matrix A is sparse (it has nonzero entries only on P diagonals) the above general strategy for matrixvector multiplication will be slightly modified to the case of FIR filtering. The ith row, i=1, . . . , K, of the matrix A has only P nonzero entries, the filter coefficients a_{j}, j=1, . . . , P, at the columns i, i+1, . . . , i+P−1. In this embodiment the potential partial products are not simultaneously stored in the memory for all the components of the input vector, but potential partial products for only P components are simultaneously stored. Every P operating steps a new set of potential partial products for the current input sample is computed within the plurality of the arrays of adders and it replaces the set of potential partial products stored in the memory 21 P operating steps earlier. This is achieved by setting the sixth control signal c_{6 }to low every P operating steps. Therefore, the plurality of the arrays of adders 2 cannot be fully halted but they can operate P times slower than in an implementation without storing potential partial products. Only P+K out of PK multiplications are completely implemented while the others are only partially implemented.
In the second strategy for FIR filtering implementation potential partial products for the P filter coefficients are stored in the memory 21 while computing the first output sample. These potential partial products are then circularly fetched from the memory 21 when computing the other K−1 output samples. This means that only P out of PK multiplications will be completely implemented while the others are only partially implemented. All the arrays of adders 2 may be halted or switched to another task as soon as potential partial products for all the filter coefficients will be stored in the memory 21. Thus, the second strategy is even more advantageous than the first one. However, this strategy is possible only if the adder/subtracters 2 of the first pipeline stage P1 have enough precision to compute potential partial products for the filter coefficients.
Multifunctional Architectures for Video/Image Processing
The structure of the present invention can advantageously be implemented as a multifunctional architecture for video/image processing (MAVIP).
Comparing the SAD structure on
There are several alternatives to combine the SAD structure with the multiplierbased structures.
In one realization, the multifunctional architecture for video/image processing involves arrays of adders 2 which, depending on the first control signal c_{1 }and the fifth control signal c_{5}, can be configured either as adder/subtracters or as DS units and, hence are denoted as AA/DSA. A possible realization of an AA/DSA unit in the general case is schematically shown in
In the considered embodiment of the MAVIP on
Modifications to the summation blocks 4 within the architecture on
Table 3 lists some operations that may be implemented on the example realization of the MAVIP of
There are many alternatives of implementing the above described embodiments of the invention since a number of different extensions of different types of multiplier/MAC unit structures can be applied. In addition, different choices of design parameters in every structure lead to different alternatives.
The advantageous embodiments of the present invention can be applied as, for example, standalone (tightly or loosely coupled) devices. An alternative implementation of the invention can be a functional unit within a processor of a system. This kind of arrangement can replace a multiplication unit within the processor. There are many other alternatives how this can be done. One of them can be an implementation where only a part of one of the different embodiments without including the plurality of arrays of adders (the first pipeline stage P1) and without including a memory 21 is used as a Functional Unit within the processor. The adders/subtracters of the processor will then play the role of the plurality of arrays of adders while the registers will play the role of the memory used in the structure for matrixvector operations.
For example, Texas Instruments TMS320C64x DSP includes eight functional units where six of them can operate as 64bit adders/subtracters, each of them being able to operate as either two 32bit or four 16bit, or eight 8bit ones. It also includes 128 64bit registers each of which may, similarly, be used as either two 32bit or four 16bit or eight 8bit ones. Supposing that another functional unit (FU) similar to the one on
Yet another alternative is the possibility of implementing the advantageous structures of the invention for various bit precisions. For instance, it is possible to implement multipliers/MAC units that can perform either one n×n multiplication, two (n/2)×m or n×(m/2) multiplications, or four (n/2)×(m/2) multiplications in one operating step.
The present invention is not limited solely to the above described embodiments, but it can be modified within the scope of the appended claims.
Claims
1. Method comprising
 inputting data to a pipelined calculation device comprising a group of at least two pipeline stages for performing a first calculation operation and other calculation operations,
 outputting data of at least one pipeline stage,
 said first calculation operation comprising
 storing the calculation results of the first calculation operation into a memory,
 controlling the memory by a control signal to set the memory open for writing when storing the calculation results of the first calculation operation into the memory, and
 said other calculation operations comprising
 using the stored data as input data to a pipeline stage with a multiple of other calculation operations, and
 controlling the memory by the control signal to set the memory closed for writing when the multiple of said other calculation operations are being performed,
 the method further comprising selecting an input data interface to a pipeline stage from at least a first output data interface and a second output data interface, the first output data interface being an output of a previous pipeline stage.
2. A method according to claim 1 comprising using a highradix nonBooth recoded multiplication method in said first and other calculation operations.
3. A method according to claim 1, wherein the first calculation operation further comprises
 multiplying a multiplicand by a multiplier, and
 performing a calculation of potential partial products.
4. A method according to claim 1, wherein the first calculation operation further comprises multiplying one multiplicand x with a number of multipliers to compute potential partial products of the common multiplicand when multiplying it by a first multiplier a1.
5. A method according to claim 4, wherein the calculation operations are performed at least partly in a timeinterlaced manner by dividing at least one calculation operation into partial calculation operations, and initiating each partial calculation operation at a different time.
6. A method according to claim 1, wherein it is used in connection with processing of video information.
7. A method according to claim 1 comprising
 using one multiplication operation as said first calculation operation,
 using another multiplication operation as said other calculation operation,
 performing said one multiplication operation and said another multiplication operation in a pipelined manner, wherein the method further comprises starting said another multiplication operation before said one multiplication operation is finished,
 dividing said one multiplication operation and said another multiplication operation into at least a first and a second sub operation,
 carrying out each of said first and second sub operations in one pipeline stage, and
 performing at least one of said sub operations in parallel in one pipeline, comprising starting a sub operation for the other multiplication operation before the corresponding sub operation for the first multiplication operation is finished.
8. A method according to claim 1 comprising
 selecting an input data interface of said pipeline stages;
 connecting at least one output data interface of said pipeline stages to said selected input data interface for routing data between said at least one output data interface and said selected input data interface.
9. System in a pipelined calculation device comprising:
 at least one input data interface for input of data for a first calculation operation and
 other calculation operations, and
 at least one output data interface for output of data,
 at least two pipeline stages,
 a memory for storing output data of said first calculation operation,
 a control signal to set the memory open or closed for writing, and
 a data retriever for using the output data stored in the memory in said other calculation operation as input data to a pipeline stage, and
 at least one multiplexer for selecting an input data interface to a pipeline stage from at least a first output data interface and a second output data interface, the first output data interface being an output of a previous pipeline stage,
 wherein the system is configured to use the output data stored in the memory with a multiple of said other calculation operations, and to control the memory by the control signal to set the memory open for writing when storing the calculation results of the first calculation operation into the memory, and to set the memory closed for writing when performing the multiple of said other calculation operations.
10. A system according to claim 9, wherein the calculation device is a Highradix nonBooth recoded multiplier.
11. A system according to claim 9, wherein the data retriever comprises a multiplexer to connect the output of the memory to an input data interface of at least one of said other pipeline stages for using the stored calculation results with said multiple of other calculation operations.
12. A system according to claim 9, a first pipeline stage including arrays of adders, a second pipeline stage including selection blocks, a third pipeline stage including compression arrays, and a fourth pipeline stage including carrylookahead adders, that an output data interface of the first pipeline stage is connected to the input of the memory and to one input of the data retriever, the output of the memory is connected to another input of the data retriever, and the output of the data retriever is connected to an input data interface of the second pipeline stage.
13. A system according to claim 9, the memory comprising a multiple of registers each register being configured to store one value, and an addressing unit for addressing one of the registers at a time for reading/writing data from/to the register.
14. A system according to claim 9, the memory comprising a multiple of registers each register being configured to store one value, a multiplexer for selecting either the input of the memory or an output of one of the registers for storing the value of a selected source into the first register, and that the output of the first register is connected to the output of the memory.
15. A system according to claim 9, including means for processing video information.
16. Device comprising
 at least one data interface for input of data for a first calculation operation and other calculation operations,
 at least one data interface for output of data,
 a group of at least two pipeline stages,
 a memory for storing output data of said first calculation operation,
 a control signal to set the memory open or closed for writing, and
 a data retriever for using the stored data in said other calculation operation as input data to a pipeline stage, and
 at least one multiplexer for selecting an input data interface to a pipeline stage from at least a first output data interface and a second output data interface, the first output data interface being an output of a previous pipeline stage,
 wherein the device is configured to use the output data stored in the memory with a multiple of said other calculation operations, and to control the memory by the control signal to set the memory open for writing when storing the calculation results of the first calculation operation into the memory, and to set the memory closed for writing when performing the multiple of said other calculation operations.
17. A device according to claim 16, wherein it is a Highradix nonBooth recoded multiplier.
18. A device according to claim 16, wherein the data retriever comprises a multiplexer to connect the output of the memory to an input data interface of at least one of said other pipeline stages for using the stored calculation results with said multiple of other calculation operations.
19. A device according to claim 16, the first pipeline stage including arrays of adders, the second pipeline stage including selection blocks, a third pipeline stage including compression arrays, and a fourth pipeline stage including carrylookahead adders, an output data interface of the first pipeline stage connected to an input of the memory and to one input of the data retriever, an output of the memory connected to another input of the data retriever, and an output of the data retriever connected to an input data interface of the second pipeline stage.
20. A device according to claim 16, the memory comprising a multiple of registers each register configured to store one value, and an addressing unit for addressing one of the registers at a time for reading/writing data from/to the register.
21. A device according to claim 16, the memory comprising a multiple of registers each register configured to store one value, a multiplexer for selecting either an input of the memory or an output of one of the registers for storing a value of a selected source into a first register, an output of the first register connected to an output of the memory.
22. A device according to claim 16, including means for processing video information.
23. A device according to claim 16, wherein the device includes pipeline registers at an input data interface of at least one pipeline stage, at least one multiplexer at an output of said at least one pipeline stage, and control inputs for controlling said pipeline registers such that only one of the pipeline registers are open at a time for writing data to the input data interface of said pipeline stage, said multiplication operations divided into at least a first and a second sub operation, each of said first and second sub operation carried out in one pipeline stage, and at least one of said sub operations performed in parallel in one pipeline, a sub operation for another multiplication operation started before a corresponding sub operation for a first multiplication operation is finished.
24. A device according to claim 16, including selection means for performing a selective data routing, said selection means comprising means for connecting a data interface to another data interface among at least two other data interfaces for routing data between the connected data interfaces.
25. A device according to claim 16, a first pipeline stage of said at least two pipeline stages comprising a first group of input registers for inputting a multiplicand, and a second pipeline stage of said at least two pipeline stages comprising a second group of input registers for inputting a multiplier, and
 said device is configured to use said multiplicand in said first calculation operation,
 and said multiplier in said other calculation operations.
26. A device according to claim 16, comprising a first and a second selection block, said first selection block comprising a first control input to control the operation of the first selection block by a first multiplier, and said second selection block comprising a second control input to control the operation of the second selection block by a second multiplier.
27. A device according to claim 26, comprising a multiplexer for selecting potential partial products of a same multiplicand or potential partial products of two different multiplicands to be input to said first and second selection block.
28. Mobile communication device having a pipelined calculation device comprising
 at least one data interface for input of data for a first calculation operation and other calculation operations,
 at least one data interface for output of data,
 a group of at least two pipeline stages,
 a memory for storing output data of said first calculation operation,
 a control signal to set the memory open or closed for writing, and
 a data retriever for using the stored data in said other calculation operations as input data to a pipeline stage, and
 at least one multiplexer for selecting an input data interface to a pipeline stage from at least a first output data interface and a second output data interface, the first output data interface being an output of a previous pipeline stage,
 wherein the device is configured to use the output data stored in the memory with a multiple of said other calculation operations, and to control the memory by the control signal to set the memory open for writing when storing the calculation results of the first calculation operation into the memory, and to set the memory closed for writing when performing the multiple of said other calculation operations.
29. Integrated circuit comprising
 at least one data interface for input of data, at least one data interface for output of data,
 a group of at least two pipeline stages,
 a memory for storing output data in a first calculation operation,
 a control signal to set the memory open or closed for writing, and
 a data retriever for using the stored data in other calculation operations as input data to a pipeline stage, and
 at least one multiplexer for selecting an input data interface to a pipeline stage from at least a first output data interface and a second output data interface, the first output data interface being an output of a previous pipeline stage,
 wherein the integrated circuit is configured to store calculation results of the first calculation operation into the memory, and to use the stored calculation results with a multiple of said other calculation operations, and to control the memory by the control signal to set the memory open for writing when storing the calculation results of the first calculation operation into the memory, and to set the memory closed for writing when performing the multiple of said other calculation operations.
30. Computing device comprising
 at least one data interface for input of data, at least one data interface for output of data,
 a group of at least two pipeline stages,
 a memory for storing output data of a first calculation operation,
 a control signal to set the memory open or closed for writing, and
 a data retriever for using the stored data in other calculation operations as input data to a pipeline stage and
 at least one multiplexer for selecting an input data interface to a pipeline stage from at least a first output data interface and a second output data interface, the first output data interface being an output of a previous pipeline stage,
 wherein the computing device is configured to use the output data stored in the memory with a multiple of said other calculation operations, and to control the memory by the control signal to set the memory open for writing when storing the calculation results of the first calculation operation into the memory, and to set the memory closed for writing when performing the multiple of said other calculation operations.
3732410  May 1973  Mackechnie 
5220525  June 15, 1993  Anderson et al. 
5594679  January 14, 1997  Iwata 
5935202  August 10, 1999  Frederick, Jr. 
6052706  April 18, 2000  Wheeler et al. 
6085214  July 4, 2000  De Angel 
6141674  October 31, 2000  Unkrich et al. 
6311203  October 30, 2001  Wada et al. 
6353843  March 5, 2002  Chehrazi et al. 
6367003  April 2, 2002  Davis 
7127482  October 24, 2006  Hou et al. 
0644492  March 1995  EP 
1050828  November 2000  EP 
 M. R. Santoro and M. A. Harowitz, “SPIM: A Pipelined 64×64bit Iterative Multiplier,” IEEE Journal of Solidstate Circuits, vol. 24, No. 2, New York, NY, Apr. 1989.
Type: Grant
Filed: Nov 6, 2003
Date of Patent: May 19, 2009
Patent Publication Number: 20040139131
Assignee: Nokia Corporation (Espoo)
Inventors: David Guevokian (Tampere), Aki Launiainen (Tampere), Petri Liuha (Tampere)
Primary Examiner: David H Malzahn
Application Number: 10/703,161