(19)
(11)EP 1 049 025 B1

(12)EUROPEAN PATENT SPECIFICATION

(45)Mention of the grant of the patent:
19.08.2009 Bulletin 2009/34

(21)Application number: 99900659.6

(22)Date of filing:  21.01.1999
(51)Int. Cl.: 
G06F 7/57  (2006.01)
(86)International application number:
PCT/JP1999/000237
(87)International publication number:
WO 1999/038088 (29.07.1999 Gazette  1999/30)

(54)

METHOD AND APPARATUS FOR ARITHMETIC OPERATIONS

VERFAHREN UND APPARAT FÜR ARITHMETISCHE OPERATIONEN

PROCEDE ET DISPOSITIF POUR DES OPERATIONS ARITHMETIQUES


(84)Designated Contracting States:
DE FR GB NL

(30)Priority: 21.01.1998 JP 937198

(43)Date of publication of application:
02.11.2000 Bulletin 2000/44

(73)Proprietor: Panasonic Corporation
Kadoma-shi Osaka 571-8501 (JP)

(72)Inventors:
  • KANAKOGI, Tomochika
    Osaka 569-0081 (JP)
  • NAKAJIMA, Masaitsu
    Osaka 536-0007 (JP)

(74)Representative: Grünecker, Kinkeldey, Stockmair & Schwanhäusser Anwaltssozietät 
Leopoldstrasse 4
80802 München
80802 München (DE)


(56)References cited: : 
WO-A-97/08610
JP-A- 7 044 533
JP-A- 10 171 778
US-A- 5 657 262
WO-A-97/14090
JP-A- 8 022 451
US-A- 5 636 155
  
      
    Note: Within nine months from the publication of the mention of the grant of the European patent, any person may give notice to the European Patent Office of opposition to the European patent granted. Notice of opposition shall be filed in a written reasoned statement. It shall not be deemed to have been filed until the opposition fee has been paid. (Art. 99(1) European Patent Convention).


    Description

    TECHNICAL FIELD



    [0001] The present invention relates to an arithmetic logic unit and arithmetic/logical operation method for realizing high-speed multiply, multiply-and-accumulate and other operations that are frequently used in signal processing.

    BACKGROUND ART



    [0002] In the past, multi-media data was processed using a microprocessor and a dedicated LSI in combination. However, thanks to recent amazing performance enhancement of microprocessors, it is now possible for a microprocessor to execute some types of multi-media data processing by itself. The development of a register-divided operation method was one of the factors contributing to this performance enhancement achieved. Also, in the fields of image processing, audio processing and so on, an operation method of deriving a single accumulated data word from input array elements is often used.

    [0003] Figure 12 illustrates a configuration for a known arithmetic logic unit performing a multiply-and-accumulate operation using a divided register.

    [0004] In Figure 12, a register 105 stores 32-bit accumulated data words ZU and ZL as its high- and low-order 32 bits, respectively. A multiplier 101 receives and multiplies together the high-order 16 bits of an input 32-bit data word X (hereinafter, referred to as "XU") and the high-order 16 bits of another input 32-bit data word Y (hereinafter, referred to as "YU") and outputs a 32-bit product. A multiplier 102 receives and multiplies together the low-order 16 bits of the input data word X (hereinafter, referred to as "XL") and the low-order 16 bits of the input data word Y (hereinafter, referred to as "YL") and outputs a 32-bit product. An adder 103 adds up the output data of the multiplier 101 and the data word ZU retained as high-order 32 bits in the register 105 and outputs a 32-bit sum. An adder 104 adds up the output data of the multiplier 102 and the data word ZL retained as low-order 32 bits in the register 105 and outputs a 32-bit sum. The output data of the adder 103 is stored as the high-order 32 bits in the register 105, while the output data of the adder 104 is stored as the low-order 32 bits in the register 105.

    [0005] In the arithmetic logic unit with such a configuration, the multiplier 101 performs the multiplication XU·YU, the adder 103 adds up the product obtained by the multiplier 101 and ZU that has been stored in the high-order 32 bits in the register 105, and the register 105 stores the result of the multiply-and-accumulate operation XU · YU+ZU, which is the output of the adder 103, as its high-order 32 bits.

    [0006] In the same way, the multiplier 102 performs the multiplication XL · YL, the adder 104 adds up the product obtained by the multiplier 102 and ZL that has been stored as low-order 32 bits in the register 105, and the register 105 stores the result of the multiply-and-accumulate operation YL+ZL, which is the output of the adder 104, as its low-order 32 bits.

    [0007] Suppose the multiply-and-accumulate operation is performed N times with the array elements shown in Figure 13 provided as the input data words X and Y to the arithmetic logic unit and with i, or the number of times the data words are input, changed from 0 through N-1. In that case, (x0 · y0 +x2 · y2+...+x2n-2 · y2n-2) will be stored as the result of the operation in the high-order 32 bits of the register 105, while (x1 · y1+x3 · y3+...+x2n-1 · y2n-1) will be stored as the result of the operation in the low-order 32 bits of the register 105.

    [0008] However, the conventional arithmetic logic unit must perform the multiply-and-accumulate operation N times and then add together (x0 · y0+x2 · y2+...+x2n-2 · y2n-2) stored in the high-order 32 bits in the register 105 and (x1 · y1+x3 · y3+... +x2n-1·y2n-1) stored in the low-order 32 bits in the register 105 to obtain (x0·y0+x1·y1+x2·y2+...+x2n-2·y2n-2+x2n-1·y2n-1).

    [0009] To carry out this addition, only the high-order 32 bits of the data stored in the register 105 should be transferred to another register and only the low-order bits of the data stored in the register 105 should be transferred to still another register (or the same register as that receiving the high-order 32 bits). Then, these data bits transferred must be added together.

    [0010] As can be seen, to obtain a single accumulation result from multiple input data words divided, the conventional arithmetic logic unit needs to perform not only the multiply-and-accumulate operation but also data transfer and addition, thus adversely increasing its processing cycle.

    [0011] WO 97/14090 A relates to a signed multiplier circuit that is capable of multiplying input data words of 9 bit or 18 bit. In view of prior-art solutions using multiplier circuits specifically designed for a given size of data words to be multiplied, WO 97/14090 criticizes the area occupied by such multiplier designs on an integrated circuit (IC) and aims to reducing the area required by a multiplier circuit capable of performing multiplications of data words of different sizes. In view of this object, WO 97/14090 suggests a multiplier circuit that is using four multipliers that can realize either one unsigned multiplication of two 18-bit words, i.e. A[17:0].B[17:0] or two unsigned multiplications two pairs of 9-bit words, i.e. A[17:9]-B[17:9] and A[8:0].B[8:0]. The multiplier circuit can be extended to allow for signed multiplications.

    [0012] US 5,636,155 relates to an arithmetic processor capable of performing two types of operations: a multiplication of two input operands X and Y, as well as a product-sum operation. The arithmetic processor performs the multiplication operation in a non-pipelined mode, while the successive product-sum operations can be performed in a pipelined mode. In all three embodiments of the arithmetic processor of US 5,636,155, the arithmetic processor comprises four partial product generators that are supplied by a record circuit of Booth - accordingly the arithmetic processor performs binary multiplications according to Booth. The four partial product generators are linked to two redundant binary adder arrays that sum the results of two of the partial product generators respectively. In the arithmetic processor of Fig. 1, the output of redundant binary adder arrays 20 and 21 is applied to another redundant binary adder array 22 and the result of the multiplication is latched in latch 17. In Fig. 3 and Fig. 8, the output of redundant binary adder arrays 20 and 21 is latched in latches 12 and 13. In the arithmetic processor shown in Fig. 8, the partial products of the input source operands X and Y are aligned to be properly summed in redundant binary adder arrays 20 and 21 prior to the output of the being redundant binary adder arrays 20 and 21 latched.

    [0013] US 5,657,262 relates to arithmetic and logic computation device. The arithmetic and logic computation device is capable of performing two types of operations: a double precision multiplications and multiplication-accumulation. The arithmetic and logic computation device includes a multiplier MULT, register P to accumulate the multiplication result and registers A0 and A1 that store the output of the ALU, and a shifter BD. The input to the multiplier MULT is provided by the two source registers R and L. The source registers R and L have a size of 2xn = 32 bit. The multiplication-accumulation operation is taking several cycles. In each cycle, the multiplier MULT is multiplying the content of one of the sub-registers L0, L1 of register L and the content of one of sub-registers R0, R1 and outputs the result to register P. Partly in parallel, the ALU performs additions of the multiplication results in register P and the content of either register A0 or A1 and outputs the result to the respective register A0 or A1. As the result of the multiplication of the 32 bit values in registers is resulting in a 64 bit value, but the register size of A0 and A1 is only 32 bits, the cross-products Pr2 and Pr3 need to be shifted twice by a shifter BD prior to their addition so that they can be properly added in part to the relevant portions of registers A0 and A1. The double precision multiplication can be obtained replacing the last instruction MAC by an accumulation on register A1 and initializing registers A0 and A1 at zero.

    [0014] WO 97/08610 A relates to a processor specifically adapted to perform so-called "packet data instructions", in particular an apparatus in a processor that allows for performing two multiply-add operations using one single multiply-add instruction. WO 97/08610 describes a "circuit" for performing multiply-add operations on packet data. As shown in Fig. 8, the source operands 831 and 833 consist of 64 bit that provide four data values, each of them comprising 16 bit. Each of the 16 bit data values of the two sources is applied to a respective 16x16 multiplier 810-813 which produces 32 bit multiplication results respectively. The multiplication results of multipliers 810 and 811 are added in adder 850, while the multiplication results of multipliers 812 and 813 are added in adder 851. The outputs of the two adders 850 and 851 are subsequently combined to the 64 bit result communicated to register 871. The "circuit" shown in Fig. 8 is performing the multiply-add operation.

    DISCLOSURE OF INVENTION



    [0015] An object of the present invention is providing an arithmetic logic unit that can obtain a single accumulation result from multiple input data words divided without performing the data transfer and addition.

    [0016] The object is solved by the subject-metter of the independent claims. Advantageous embodiments are object to the dependent claims.

    [0017] An arithmetic logic unit receives (n × M)-bit data words X and Y and outputs a single independent data word Z, where X and Y are each composed of a number n of M-bit data units that are independent of each other. The arithmetic logic unit includes: 1St through nth multipliers, each multiplying together associated data units with the same digit position of the data words X and Y; 1st through nth shifters, each being able to perform bit shifting on an output of associated one of the 1st through nth multipliers; and an adder for adding up outputs of the 1st through nth shifters. If a sum of the outputs of the 1st through nth multipliers is obtained as the data word Z, the 1st through nth shifters perform no bit shifting. But if the outputs of the 1st through nth multipliers are obtained separately for the data word z, the 1st through nth shifters perform a bit-shifting control in such a manner that the outputs of the 1st through nth multipliers are shifted to respective digit positions not overlapping each other.

    [0018] In such a configuration, a multiply-and-accumulate operation can be performed with the number of steps reduced. Also, by switching the modes of control performed by the shifter, multiple lines of multiplication can be performed in parallel.

    [0019] Another arithmetic logic unit receives (n × M)-bit data words X and Y and outputs a single independent data word Z, where X and Y are each composed of a number n of M-bit data units that are independent of each other. The arithmetic logic unit includes: a register for storing the data word z; 1st through nth multipliers, each multiplying together associated data units with the same digit position of the data words X and Y; and an adder for adding up outputs of the 1st through nth multipliers and an output of the register and inputting the sum to the register. The arithmetic logic unit performs a sum-of-products operation with the data words X and Y input for multiple cycles.

    [0020] In such a configuration, even though an increased number of inputs should be provided to a multi-input adder, the increase in circuit size of the adder can be relatively small. Thus, a multiply-and-accumulate operation is realizable with the increase in circuit size minimized.

    [0021] A further arithmetic logic unit receives (n × M)-bit data words X and Y and outputs a single independent data word Z, where X and Y are each composed of a number n of M-bit data units that are independent of each other. The arithmetic logic unit includes: a register for storing the data word Z; 1st through nth multipliers, each multiplying together associated data units with the same digit position of the data words X and Y; 1st through nth shifters, each being able to perform bit shifting on an output of associated one of the 1st through nth multipliers; and an adder for adding up outputs of the 1st through nth shifters and an output of the register and inputting the sum to the register. In performing a sum-of-products operation with the data words X and Y input for multiple cycles, if a cumulative sum of products of the 1st through nth multipliers is obtained as the data word Z, the 1st through nth shifters perform no bit shifting. But if the sums of products of the 1st through nth multipliers are obtained separately for the data word Z, the 1st through nth shifters perform such a control that the outputs of the 1st through nth multipliers are shifted to respective digit positions not overlapping each other.

    [0022] In such a configuration, a sum-of-products operation can be performed with the number of steps reduced. Also, by switching the modes of control performed by the shifter, multiple lines of sum-of-products operations can be performed in parallel.

    [0023] Another arithmetic logic unit receives (n × M)-bit data words X and Y and outputs a data word Z, where X and Y are each composed of a number n of M-bit data units. The arithmetic logic unit is characterized by including: a register for storing the data word Z; a selector for selecting one of the number n of data units of which the data word Y is made up; 1st through nth multipliers, each selecting one of the number n of data units, of which the data word X is made up, and multiplying together the data unit selected and an output of the selector, the data units selected by the multipliers not overlapping each other; 1st through nth shifters, each being able to perform bit shifting on an output of associated one of the 1st through nth multipliers; and an adder for adding up outputs of the 1st through nth shifters and an output of the register and inputting the sum to the register. The arithmetic logic unit is also characterized in that in a pth cycle, the selector selects a pth least significant one of the data units and a qth least significant one of the shifters performs a bit shifting control by (p+q-2)M bits.

    [0024] In this configuration, even a multiplicand with a bit number equal to or greater than the number of bits input to the multipliers included in the arithmetic logic unit can be multiplied.

    [0025] Another arithmetic logic unit receives 2M-bit data words X and Y and outputs a 4M-bit data word Z. The arithmetic logic unit includes: a first register for storing bit-by-bit carries C resulting from additions; a second register for storing bit-by-bit sums S resulting from the additions; a third register for storing the data word Z; a first decoder for receiving and decoding high-order M bits of the data word X; a second decoder for receiving and decoding low-order M bits of the data word X; first and second selectors, each selecting either high- or low-order M bits of the data word Y; a first partial product generator for receiving output data of the first decoder and the first selector and generating partial products for a multiply-and-accumulate operation; a second partial product generator for receiving output data of the second decoder and the second selector and generating partial products for the multiply-and-accumulate operation; a first full adder for adding up the partial products generated by the first partial product generator; a second full adder for adding up the partial products generated by the second partial product generator; a data extender/shifter that receives output data of the first and second full adders and can perform data extension and data shifting on the data; a carry-propagation adder for receiving, and performing a carry-propagation addition on, the bit-by-bit carries C and the bit-by-bit sums S that have been stored in the first and second registers and outputting the result to the third register; a third selector for selectively outputting either the data stored in the third register or zero data; a fourth selector for selectively outputting either the output data of the carry-propagation adder or zero data; and a third full adder for receiving, and performing a full addition on, the output data of the data extender/shifter and the output data of the third and fourth selectors and for inputting the bit-by-bit carries C and the bit-by-bit sums S to the first and second registers, respectively.

    [0026] In such a configuration, a sum-of-products operation can be performed with the number of steps and the circuit size both reduced. Also, by switching the modes of control performed by the shifter, multiple lines of sum-of-products operations can be performed in parallel or a multiplicand with a bit number equal to or greater than the number of bits input to the multipliers can be multiplied with the circuit size reduced.

    [0027] A solution is an arithmetic/logical operation method for calculating a single independent data word Z from input (nXM)-bit data words X and Y, where X and Y are each composed of a number n of M-bit data units that are independent of each other. The method includes the steps of: multiplying together associated data units with the same digit position of the data words X and Y, thereby obtaining respective products; shifting bits of the products obtained in the multiplying step; and adding up values obtained in the shifting step, thereby obtaining a sum. If a sum of the number n of products obtained in the multiplying step is calculated as the data word Z, no bit shifting is performed in the shifting step. But if the number n of products are obtained separately in the multiplying step for the data word Z, bit-shifting is performed in the shifting step in such a manner that the number n of products are shifted to respective digit positions not overlapping each other.

    [0028] According to such configuration, a multiply-and-accumulate operation can be performed with the number of steps reduced.

    [0029] Another solution is an arithmetic/logical operation method for calculating a single independent data word Z from input (n × M)-bit data words X and Y, where X and Y are each composed of a number n of M-bit data units that are independent of each other. The method includes the steps of: multiplying together associated data units with the same digit position of the data words X and Y, thereby obtaining respective products; adding up the number n of products obtained in the multiplying step, thereby obtaining a sum; and performing a sum-of-products operation on the sums obtained in the adding step with the data words X and Y input for multiple cycles.

    [0030] According to such configuration, a multiply-and-accumulate operation can be performed with the increase in circuit size minimized.

    [0031] A further solution is an arithmetic/logical operation method for calculating a single independent data word Z from input (n × M)-bit data words X and Y, where X and Y are each composed of a number n of M-bit data units that are independent of each other. The method includes the steps of: multiplying together associated data units with the same digit position of the data words X and Y, thereby obtaining respective products; shifting bits of the products obtained in the multiplying step; adding up values obtained in the shifting step, thereby obtaining a sum; and performing a sum-of-products operation on the sums obtained in the adding step with the data words X and Y input for multiple cycles. If a cumulative sum of the number n of products, which have been generated in the multiplying step, is obtained as the data word Z, no bit shifting is performed in the shifting step. But if the sums of the number n of products, which have been generated in the multiplying step, are obtained separately for the data word Z, bits of the number n of products are shifted to respective digit positions not overlapping each other in the shifting step.

    [0032] According to such configuration, a sum-of-products operation can be performed with the number of steps reduced.

    BRIEF DESCRIPTION OF DRAWINGS



    [0033] 

    Figure 1 illustrates a configuration for an arithmetic logic unit according to an embodiment of the present invention.

    Figure 2 illustrates a detailed configuration for a data extender/shifter according to the embodiment of the present invention.

    Figure 3 illustrates a detailed configuration for a third full adder according to the embodiment of the present invention.

    Figure 4 illustrates a detailed configuration for a carry-propagation adder according to the embodiment of the present invention.

    Figure 5 illustrates how the arithmetic logic unit shown in Figure 1 performs a multiply-and-accumulate operation.

    Figure 6 is a timing diagram illustrating the multiply-and-accumulate operation performed by the arithmetic logic unit shown in Figure 1.

    Figure 7 illustrates how the arithmetic logic unit shown in Figure 1 performs a dual multiply-and-accumulate operation.

    Figure 8 is a timing diagram illustrating the dual multiply-and-accumulate operation performed by the arithmetic logic unit shown in Figure 1.

    Figure 9 illustrates how the arithmetic logic unit shown in Figure 1 performs a multiplication.

    Figure 10 is a timing diagram illustrating the multiplication performed by the arithmetic logic unit shown in Figure 1.

    Figure 11 illustrates various types of operations performed by the arithmetic logic unit shown in Figure 1.

    Figure 12 illustrates a configuration for a conventional arithmetic logic unit with a divided register.

    Figure 13 illustrates array elements for a multiply-and-accumulate operation.


    BEST MODE FOR CARRYING OUT THE INVENTION



    [0034] Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings.

    [0035] Figure 1 illustrates a schematic configuration for an arithmetic logic unit according to an embodiment of the present invention. In Figure 1, X and Y are 32-bit input data words, each composed of two independent 16-bit data half-words, and Z is a 64-bit accumulated data word. S and C are 64-bit intermediate data words, which are collections of bit-by-bit sums and carries, respectively. These intermediate data words S and C are herein used because in performing additions repeatedly, obtaining sums and carries on a bit-by-bit basis and then a sum of all the bits at last would be more efficient than obtaining the sum of all the bits every time. Booth decoders 511 and 512 decode XU and XL, respectively, and output the results. Each of selectors 521 and 522 selects and outputs YU or YL. A partial product generator 531 receives the output data of the first decoder 511 and the data selected by the first selector 521 and generates and outputs partial products for multiplication. A partial product generator 532 receives the output data of the second decoder 512 and the data selected by the second selector 522 and generates and outputs partial products for multiplication. Full adders 541 and 542 perform full addition on the partial products generated by the first and second partial product generators 531 and 532, respectively, and then output the results. A data extender/shifter 55 receives, and performs data extension and shifting on, the output data of the first and second full adders 541 and 542, and outputs the results. A third selector 523 selectively outputs "0" or the data word Z stored in a third register 59. A fourth selector 524 selectively outputs "0" or the output S + C of a carry-propagation adder 58. A full adder 56 receives, and performs full addition on, the data that has been extended and shifted by the data extender/shifter 55 and the data that has been selected by the third and fourth selectors 523 and 524, thereby outputting the 64-bit intermediate data words C and S. Registers 571 and 572 store the data words C and S, which are the outputs of the third full adder 56. The carry-propagation adder 58 receives, and performs carry-propagation addition on, the data words C and S that have been stored in the first and second registers 571 and 572, respectively, and then outputs the result S+C. The register 59 stores the output of the carry-propagation adder 58.

    [0036] Figure 2 illustrates a detailed configuration for the data extender/shifter 55. In Figure 2, a data extension section 61, consisting of data extenders 611 through 614, performs sign or zero extension on the input data and outputs the results. A data shifting section 62, consisting of data shifters 621 through 624, receives, and performs data shifting on, the output data of the data extension section and outputs the results. In Figure 2, SFT1.O through SFT4.O are output data words of the data shifters 621 through 624, respectively.

    [0037] Next, Figure 3 illustrates a detailed configuration for the third full adder 56. In Figure 3, a full adder 72 receives, and performs full addition on, the respective low-order 32 bits of the output data words SFT1.O through SFT4.O of the data shifters 621 through 624 and those of the output data words SEL1.O and SEL2 O of the third and fourth selectors 523 and 524. A selector 73 selects either an overflow signal of the fourth full adder 72 or "0" and outputs the result. Specifically, while a control signal CTRL 50 is "H", the fifth selector 73 selects the overflow signal of the fourth full adder 72 and outputs it to a fifth full adder 71. On the other hand, while the control signal CTRL 50 is "L", the fifth selector 73 selects "0" and outputs it to the fifth full adder 71. The full adder 71 receives, and performs full addition on, the respective high-order 32 bits of the output data words SFT1.O through SFT4.O of the data shifters 621 through 624, those of the output data words SEL1.O and SEL2.O of the third and fourth selectors 523 and 524 and the output data of the fifth selector 73. In Figure 3, CU and SU are output data words of the fifth full adder 71, while CL and SL are output data words of the fourth full adder 72.

    [0038] Next, Figure 4 illustrates a detailed configuration for the carry-propagation adder 58.

    [0039] In Figure 4, a carry-propagation adder 82 receives, and performs carry-propagation addition on, CL and SL, which have been stored as low-order 32 bits in the first and second registers 571 and 572, respectively, and outputs the result. A selector 83 selects either an overflow signal of the carry-propagation adder 82 or "0" and outputs the result. Specifically, while the control signal CTRL 50 is "H", the sixth selector 83 selects the overflow signal of the carry-propagation adder 82 and outputs it to a carry-propagation adder 81. On the other hand, while the control signal CTRL 50 is "L", the sixth selector 83 selects "0" and outputs it to the carry-propagation adder 81. The carry-propagation adder 81 receives, and performs carry-propagation addition on, CU and SU, which have been stored as high-order 32 bits in the first and second registers 571 and 572, respectively, and the output data of the sixth selector 83 and then outputs the result.

    [0040] Hereinafter, it will be described how the arithmetic logic unit with such a configuration operates.
    1. (1) First, an arithmetic/logical operation for obtaining a single independent accumulated data word from two divided input data words will be described. In the following example, an operation of obtaining (x0 · y0+x1 · y1+x2 · y2+...+x2n-· y2n-2+x2n-1 · y2n-1) from the input array elements shown in Figure 13 will be explained. Also, in the following example, a signed multiply-and-accumulate operation will be performed as the arithmetic/logical operation for illustrative purposes. In this case, the operation will be executed on the supposition that the control signal CTRL 50 is "H".
      In Figure 1, the first and second selectors 521 and 522 select YU and YL, respectively, for the first cycle. Thus, the first and second partial product generators 531 and 532 generate respective partial products for multiplications XU˙ YU and XL·YL. The first full adder 541 performs full addition on the partial products generated by the first partial product generator 531, thereby outputting two 32-bit intermediate data words for the multiplication XU·YU. In the same way, the second full adder 542 outputs two 32-bit intermediate data words for the multiplication XL·YL.
      The two 32-bit intermediate data words, output from the first full adder 541, are input to the data extenders 611 and 612 of the data extender/shifter 55 so as to be sign-extended to 64 bits. These data words, which have been sign-extended by the data extenders 611 and 612, are input to the data shifters 621 and 622, respectively. The data shifters 621 and 622 do not perform any data shifting but just output the two 64-bit intermediate data words for the multiplication XU·YU.
      In the same way, the two 32-bit intermediate data words, output from the second full adder 542, are input to the data extenders 613 and 614 of the data extender/shifter 55 so as to be sign-extended to 64 bits. These data words, which have been sign-extended by the data extenders 613 and 614, are input to the data shifters 623 and 624, respectively. The data shifters 623 and 624 do not perform any data shifting either, but just output the 64-bit intermediate data words for the multiplication XL · YL.
      The fourth selector 524 selects "0" and then outputs it to the third full adder 56. On the other hand, the third selector 523 selects the 64-bit data word Z, which has been stored in the third register 59, and outputs it to the third full adder 56.
      The third full adder 56 receives, and performs full addition on, the outputs SFT1.O through SFT4.O of the data extender/shifter and the outputs SEL1.O and SEL2.O of the third and fourth selectors 523 and 524. In this case, since the control signal CTRL 50 is "H", the overflow signal of the fourth full adder 72 is input to the fifth full adder 71. As a result, the overflow propagates from the 31st to the 32nd bit in the 64-bit full addition operation.
      Accordingly, the third full adder 56 outputs two 64-bit intermediate data words C and S for the multiply-and-accumulate operation (XU · YU+XL · YL+Z) as shown in Figure 5.
      These 64-bit intermediate data words C and S are stored in the first and second registers 571 and 572, respectively.
      In the second cycle, the carry-propagation adder 58 receives, and performs carry-propagation addition on, the two 64-bit intermediate data words C and S that have been stored in the first and second registers 571 and 572. In this case, since the control signal CTRL 50 is "H", the overflow signal of the carry-propagation adder 82 is input to the carry-propagation adder 81. As a result, the overflow signal propagates from the 31st to the 32nd bit in the 64-bit carry-propagation addition operation.
      Accordingly, the carry-propagation adder 58 outputs the result of the multiply-and-accumulate operation (XU · YU+XL · YL+Z) as shown in Figure 5. The third register 59 stores the output data of the carry-propagation adder 58.
      In this manner, the arithmetic logic unit can obtain a single 64-bit accumulated data word in two cycles as shown in Figure 6 by performing a multiply-and-accumulate operation on the input 32-bit data words X and Y, each composed of two 16-bit data half-words.
      Accordingly, the arithmetic logic unit can obtain (x0 · y0 +x1 · y1+x2 · y2+...+x2n-2·y2n-2+x2n-1- y2n-1) just by performing the multiply-and-accumulate operation N times on the input array elements shown in Figure 13. That is to say, the arithmetic logic unit needs no data transfer and addition indispensable for known arithmetic logic units.
    2. (2) Next, an arithmetic/logical operation of obtaining two independent accumulated data words from two divided input data words will be described. In the following example, it will be described how to obtain two 32-bit accumulated data words in two cycles as shown in Figure 8 by performing an arithmetic/logical operation on two input 32-bit data words X and Y, each composed of two 16-bit data half-words. Also, in the following example, a signed dual multiply-and-accumulate operation will be performed as the arithmetic/logical operation for illustrative purposes. In this case, the operation will be executed on the supposition that the control signal CTRL 50 is "L".
      In Figure 1, the first and second selectors 521 and 522 select YU and YL, respectively, for the first cycle. Thus, the first and second partial product generators 531 and 532 generate respective partial products for multiplications XU· YU and XL·YL. The first full adder 541 performs full addition on the partial products generated by the first partial product generator 531, thereby outputting two 32-bit intermediate data words for the multiplication XU·YU. In the same way, the second full adder 542 outputs two 32-bit intermediate data words for the multiplication XL·YL.
      The two 32-bit intermediate data words, output from the first full adder 541, are input to the data extenders 611 and 612 of the data extender/shifter 55 so as to be zero-extended to 64 bits. These data words, which have been zero-extended by the data extenders 611 and 612, are input to the data shifters 621 and 622, respectively. The data shifters 621 and 622 shift these data words to the left by 32 bits, thereby outputting two 64-bit intermediate data words for an operation (XU·YU)<<32.
      In the same way, the two 32-bit intermediate data words, output from the second full adder 542, are input to the data extenders 613 and 614 of the data extender/shifter 55 so as to be zero-extended to 64 bits. These data words, which have been zero-extended by the data extenders 613 and 614, are input to the data shifters 623 and 624, respectively. The data shifters 623 and 624 do not perform any data shifting but just output these two 64-bit intermediate data words for the multiplication XL·YL.
      The fourth selector 524 selects "0" and then outputs it to the third full adder 56. On the other hand, the third selector 523 selects the 64-bit data word Z, which has been stored in the third register 59, and outputs it to the third full adder 56.
      The third full adder 56 receives, and performs full addition on, the outputs SFT1.O through SFT4.O of the data extender/shifter 55 and the outputs SEL1.O and SEL2.O of the third and fourth selectors 523 and 524. In this case, since the control signal CTRL 50 is "L", "0" is input to the fifth full adder 71. Accordingly, no overflow propagates from the 31st to the 32nd bit in the 64-bit full addition operation.
      As a result, the third full adder 56 outputs intermediate data words C and S, of which the high-order 32 bits represent the operation (XU·YU+ZU) and the low-order 32 bits represent the operation (XL·YL + ZL), as shown in Figure 7. These 64-bit intermediate data words C and S are stored in the first and second registers 571 and 572, respectively.
      In the second cycle, the carry-propagation adder 58 receives, and performs carry-propagation addition on, the two 64-bit intermediate data words C and S that have been stored in the first and second registers 571 and 572. In this case, since the control signal CTRL 50 is "L", "0" is input to the carry-propagation adder 81. As a result, no overflow signal propagates from the 31st to the 32nd bit in the 64-bit carry-propagation addition operation.
      Accordingly, the carry-propagation adder 58 outputs the high-order 32 bits as a result of the operation (XU·YU+ZU) and the low-order 32 bits as a result of the operation (XL· YL+ZL) as shown in Figure 7. The third register 59 stores the output data of the carry-propagation adder 58.
      In this manner, the arithmetic logic unit can obtain two 32-bit accumulated data words in two cycles as shown in Figure 8 by performing an operation on the input 32-bit data words X and Y, each composed of two 16-bit data half-words.
    3. (3) Next, it will be described how to perform an arithmetic/logical operation on non-divided input data. In the following example, an operation of obtaining a 64-bit product by performing multiplication on input 32-bit data words X and Y will be explained. Also, in the following example, a signed multiplication operation will be performed as the arithmetic/logical operation for illustrative purposes. In this case, the operation will be executed on the supposition that the control signal CTRL 50 is "H".


    [0041] In Figure 1, the first and second selectors 521 and 522 select YU and YL, respectively, for the first cycle. Thus, the first and second partial product generators 531 and 532 generate respective partial products for multiplications XU· YU and XL·YL. The first full adder 541 performs full addition on the partial products generated by the first partial product generator 531, thereby outputting two 32-bit intermediate data words for the multiplication XU YU. In the same way, the second full adder 542 outputs two 32-bit intermediate data words for the multiplication XL · YL.

    [0042] The two 32-bit intermediate data words, output from the first full adder 541, are input to the data extenders 611 and 612 of the data extender/shifter 55 so as to be sign-extended to 64 bits. These data words, which have been sign-extended by the data extenders 611 and 612, are input to the data shifters 621 and 622, respectively. The data shifters 621 and 622 perform arithmetic shift on these data words to the left by 32 bits, thereby outputting two 64-bit intermediate data words for a multiplication (XU·YU)<<32.

    [0043] In the same way, the two 32-bit intermediate data words, output from the second full adder 542, are input to the data extenders 613 and 614 of the data extender/shifter 55 so as to be sign-extended to 64 bits. These data words, which have been sign-extended by the data extenders 613 and 614, are input to the data shifters 623 and 624, respectively. The data shifters 623 and 624 do not perform any data shifting but just output the two 64-bit intermediate data words for the operation XL·YL.

    [0044] The third and fourth selectors 523 and 524 both select "0" and then outputs it to the third full adder 56.

    [0045] The third full adder 56 receives, and performs full addition on, the outputs SFT1.O through SFT4.O of the data extender/shifter 55 and the outputs SEL1.O and SEL2.O of the third and fourth selectors 523 and 524. In this case, since the control signal CTRL 50 is "H", the overflow signal of the fourth full adder 72 is input to the fifth full adder 71. As a result, the overflow propagates from the 31st to the 32nd bit in the 64-bit full addition operation.

    [0046] Accordingly, the third full adder 56 outputs two 64-bit intermediate data words C and S for the multiply-and-accumulate operation (((XU·YU)<<32) +XL·YL) as shown in Figure 9.

    [0047] These 64-bit intermediate data words C and S are stored in the first and second registers 571 and 572, respectively.

    [0048] In the second cycle, the carry-propagation adder 58 receives, and performs carry-propagation addition on, the two 64-bit intermediate data words C and S that have been stored in the first and second registers 571 and 572. In this case, since the control signal CTRL 50 is "H", the overflow signal of the carry-propagation adder 82 is input to the carry-propagation adder 81. As a result, the overflow signal propagates from the 31st to the 32nd bit in the 64-bit carry-propagation addition operation.

    [0049] Accordingly, the carry-propagation adder 58 outputs the result of the multiply-and-accumulate operation (((XU·YU) << 32)+XL·YL) as shown in Figure 9.

    [0050] The first and second selectors 521 and 522 also select YL and YU, respectively. Thus, the first and second partial product generators 531 and 532 generate respective partial products for multiplications XU YL and XL·YU. The first full adder 541 performs full addition on the partial products generated by the first partial product generator 531, thereby outputting two 32-bit intermediate data words for the multiplication XU· YL. In the same way, the second full adder 542 outputs two 32-bit intermediate data words for the multiplication XL·YU.

    [0051] The two 32-bit intermediate data words, output from the first full adder 541, are input to the data extenders 611 and 612 of the data extender/shifter 55 so as to be sign-extended to 64 bits. These data words, which have been sign-extended by the data extenders 611 and 612, are input to the data shifters 621 and 622, respectively. The data shifters 621 and 622 shift these data words to the left by 16 bits, thereby outputting two 64-bit intermediate data words for a multiplication (XU·YL)<<16.

    [0052] In the same way, the two 32-bit intermediate data words, output from the second full adder 542, are input to the data extenders 613 and 614 of the data extender/shifter 55 so as to be sign-extended to 64 bits. These data words, which have been sign-extended by the data extenders 613 and 614, are input to the data shifters 623 and 624, respectively. The data shifters 623 and 624 shift these data words to the left by 16 bits, thereby outputting two 64-bit intermediate data words for the operation (XL·YU) < <16.

    [0053] The third and fourth selectors 523 and 524 select "0" and the output of the carry-propagation adder 58, respectively, and then output them to the third full adder 56.

    [0054] The third full adder 56 receives, and performs full addition on, the outputs SFT1.O through SFT4.O of the data extender/shifter and the outputs SEL1.O and SEL2.O of the third and fourth selectors 523 and 524. In this case, since the control signal CTRL 50 is "H", the overflow signal of the fourth full adder 72 is input to the fifth full adder 71. As a result, the overflow propagates from the 31st to the 32nd bit in the 64-bit full addition operation.

    [0055] Accordingly, the third full adder 56 outputs two 64-bit intermediate data words C and S for the operation (((XU·YU)< <32)+XL·YL)+(((XU·YL)<<16)+((XL·YU)<<16)) as shown in Figure 9. These 64-bit intermediate data words C and S are stored in the first and second registers 571 and 572, respectively.

    [0056] In the third cycle, the carry-propagation adder 58 receives, and performs carry-propagation addition on, the two 64-bit intermediate data words C and S that have been stored in the first and second registers 571 and 572. In this case, since the control signal CTRL 50 is "H", the overflow signal of the carry-propagation adder 82 is input to the carry-propagation adder 81. As a result, the overflow signal propagates from the 31st to the 32nd bit in the 64-bit carry-propagation addition operation.

    [0057] Accordingly, the carry-propagation adder 58 outputs the result of the operation (((XUYU)<<32)+XL·YL)+(((XU·YL)< <16)+((XL·YU)<<16)), i.e., the result of the multiplication X·Y, as shown in Figure 9.

    [0058] In this manner, the arithmetic logic unit can obtain a 64-bit product in three cycles as shown in Figure 10 by performing multiplication on the input 32-bit data words X and Y.

    [0059] The arithmetic logic unit shown in Figure 1 can execute various arithmetic/logical operations other than those exemplified above.

    [0060] Figure 11 illustrates those operations and their controls. In Figure 11, the first through fourth selectors 521 through 524 represent the data words selected and the shifters 621 through 624 represent the numbers of bits shifted and the directions thereof. m < < represents shifting to the left by m bits, while m > > represents shifting to the right by m bits. S+C is the output of the carry-propagation adder 58.

    [0061] The selectors and shifters belong to the first stage of the two-pipeline-stage arithmetic logic unit. Thus, controls for the first cycle are described for the two-cycle operations, and controls for the first and second cycles are described for the three-cycle operations.

    [0062] In Figure 1, the outputs of the third and fourth selectors 523 and 524 are input to the third full adder 56. However, the same effects are attainable even if those outputs are input to the first and second full adders 541 and 542, respectively.

    [0063] Also, the arithmetic logic unit shown in Figure 1 has a two-stage pipeline structure. Alternatively, the same effects are attainable even when the arithmetic logic unit has a multi-stage pipeline structure.

    [0064] Furthermore, in the foregoing embodiment, data words are input and subjected to arithmetic/logical operations with the register divided into two. Similar effects are attainable, however, if more than two divided data words are input.

    [0065] As is apparent from the foregoing description, the present invention is advantageous in that no data transfer and addition processes are needed even in obtaining a single independent accumulated data word from multiple divided input data words. In addition, the present invention is also advantageous in that the various types of operations shown in Figure 11 are executable.

    [0066] As can be understood from the foregoing description, the inventive arithmetic logic unit and arithmetic/logical operation method are effectively applicable to image or audio processing and is suitably applicable to multi-media data processing including these types of processing.


    Claims

    1. An arithmetic logic unit for performing plural arithmetic/logical operations on data words X and Y input to the arithmetic logic unit capable of performing the plural arithmetic/logical operations, wherein the data words X and Y are each composed of a number n>1 of data units and each of the data units consists of M bits, the arithmetic logic unit comprising:

    n multipliers (531, 541; 532, 542), respectively corresponding to one of the n data units of the data word X, wherein each of the multipliers (531, 541; 532, 542) is adapted to multiply a respective data unit of the data word X and a respective data unit of the data word Y and to output a multiplication result,

    n shifters (55), respectively corresponding to one of the n multipliers, wherein the shifters are adapted to, depending on the arithmefic/ogical operation to be performed by the arithmetic logic unit, either shift the multiplication results in such a manner that the outputs of the n shifters are shifted to respective different bit positions, or not to shift the multiplication results, and

    an adder (56, 58) adapted to add the outputs of the n shifters.


     
    2. The arithmetic logic unit according to claim 1, wherein the plural arithmetic/logical operations include a multiply-and-accumulate operation of the n data units of the data words X and Y and n independent multiply-and-accumulate operations for respective pairs of data units of the data words X and Y,
    wherein the shifters are adapted to shift the multiplication results input by the multipliers in such a manner that the outputs of the n shifters are shifted to respective bit positions not overlapping each other, if the arithmetic logic unit is to perform n independent multiply-and-accumulate operations for respective pairs of data units of the data words X and Y, and
    wherein the shifters are adapted not to shift the multiplication results, if the arithmetic logic unit is to perform a multiply-and-accumulate operation of the n data units of the data words X and Y,
    wherein the arithmetic logic unit further comprises a register (59) adapted to store the output of the adder (56, 58) as a data word Z, which can be fed back to the adder (56, 58),
    wherein the adder (56, 58) is adapted to add the outputs of the n shifters and the stored data word Z, if the arithmetic logic unit is to perform a multiply-and-accumulate operation of the n data units of the data words X and Y, or n independent multiply-and-accumulate operations for respective pairs of data units of the data words X and Y.
     
    3. The arithmetic logic unit according to claim 1 or 2, further comprising n selectors (521, 522) wherein each of the n selectors (521, 522) is adapted to selectively input to a respective one of the n multipliers (531, 541; 532, 542) one data unit from the n data units of the data word Y.
     
    4. The arithmetic logic unit according to claim 3, wherein, if the arithmetic logic unit is to perform n independent multiply-and-accumulate operations for respective pairs of data units of the data words X and Y, the pth selector of the n selectors is adapted to select the pth least significant data unit of data word Y, so as to cause the pth multiplier of the n multipliers to multiply the pth least significant data units of the data words X and Y, and the shifters receiving the multiplication results from the multipliers are adapted to shift the multiplication results to non-overlapping bit positions and
    wherein, if the arithmetic logic unit is to perform a multiply-and-accumulate operation of the n data units of the data words X and Y, the pth selector of the n selectors is adapted to select the pth least significant data unit of data word Y, so as to cause the pth multiplier to multiply the pth least significant data units of the data words X and Y and the shifters are adapted not to shift the respective received multiplication results.
     
    5. The arithmetic logic unit according to claim 3, wherein the plural arithmetlc/logical operations includes a multiplication operation for multiplying the data words X and Y,
    wherein, if the arithmetic logic unit is to perform a multiplication operation for multiplying the data words X and Y, the Pth selector of the n selectors is adapted to select the qth least significant data unit of data word Y, so as to cause the pth multiplier of the n multipliers to multiply the pth data unit of the data word X and the qth data unit of the data word Y, and the pth shifter receiving the multiplication result of the pth data unit of data word X and the qth data unit of data word Y from the pth multiplier is adapted to shift the multiplication result by (p+q-2)M bits.
     
    6. A method for performing plural arithmetic/logical operations on data words X and Y input to an arithmetic logic unit capable of performing the plural arithmetic logical operations, wherein the data words X and Y are each composed of a number n>1 of data units that are independent of each other and each of the data units insists of M bits, the method comprising the following steps performed by the arithmetic logic unit:

    a multiplication step performed by n multipliers (531, 541; 532, 542), respectively corresponding to one of the n data units of the data word X, in which each of the multipliers (531, 541; 532, 542) is multiplying a respective data unit of the data word X and a respective data unit of the data word Y and to output a multiplication result,

    a shifting step performed by n shifters (55), respectively corresponding to one of the n multipliers, in which the shifters are shifting the multiplication results in such a manner that the outputs of the n shifters are shifted to respective different bit positions, or are performing no shifting of the multiplication results, wherein the operation performed by the shifters is depending on the arithmetic/logical operation to be performed by the arithmetic logic unit, and

    an adding step performed by an adder (56, 58), in which the adder (56, 58) is adding the outputs of the n shifters.


     
    7. The method according to claim 6, wherein the plural arithmetic/logical operations include a multiply-and-accumulate operation of the n data units of the data words X and Y and n independent multiply-and-accumulate operations for respective pairs of data units of the data words X and Y,
    wherein in the shifting step, the shifters shift the multiplication results output by multipliers in such a manner that the outputs of the n shifters are shifted to respective bit positions not overlapping each other, if the arithmetic logic unit is to perform n independent multiply-and-accumulate operations for respective pairs of data units of the data words X and Y, and
    wherein the shifters do not shift the multiplication results, if the arithmetic logic unit is to perform a multiply-and-accumulate operation of the n data units of the data words X and Y, and
    wherein the method further comprises the steps of storing the output of the adder (56, 58) in a register (59) of the arithmetic logic unit as a data word Z, and feeding the data word Z back to the adder (56, 58),
    wherein in the step of adding, the adders (58, 58) adds the outputs of the n shifters and the stored data word Z, if the arithmetic logic unit is to perform a multiply-and-accumulate operation of the n data units of the data words X and Y, or n independent multiply-and-accumulate operations for respective pairs of data units of the data words X and Y.
     
    8. The method according to claim 6 or 7, further comprising an input step in which each of n selectors (521, 522) selectively inputs one data unit from the n data units of the data word Y to a respective one of the n multipliers (531, 541; 532, 542).
     
    9. The method according to claim 8, wherein, if the arithmetic logic unit is to perform n independent multiply-and-accumulate operations for respective pairs of data units of the data words X and Y, in the input step,the pth selector of the n selectors selects the pth least significant data unit of data word Y, so as to cause the pth multiplier of the n multipliers to multiply the pth least significant data units of the data words X and Y in the multiplication step, and in the shifting step, the shifters receiving the multiplication results from the multipliers shift the multiplication results to non-overlapping bit positions, and
    wherein, if the arithmetic logic unit is to perform a multiply-and-accumulate operation of the n data units of the data words X and Y, in the input step, the pth selector of the n selectors selects the pth least significant data unit of data word Y, so as to cause the pth multiplier of the n multipliers to multiply the pth least significant data units of the data words X and Y, and in the shifting step the shifters do not shift the respective received multiplication result.
     
    10. The method according to claim 8, wherein the plural anthmetic/logical operation includes a multiplications operation for multiplying the data words X and Y and
    wherein, if the arithmetic logic unit is to perform a multiplication operation for multiplying the data words X and Y, in the input step, the pth selector of the n selectors selects the qth least significant data unit of data word Y, so as to cause the pth multiplier of the n multipliers to multiply the pth data unit of the data word X and the qth data unit of the data word Y,in the multiplication step, and in the shifting step, the pth shifter receiving the multiplication result of the p"' data unit of data word X and the qth data unit of data word Y from the pth multiplier shifts the multiplication result by (p+q-2)M bits.
     


    Ansprüche

    1. Arithmetik-Logikeinheit zum Durchführen mehrerer arithmetischer/logischer Operationen an Datenwörtern X und Y, die in die Arithmetik-Logikeinheit eingegeben werden, die in der Lage ist, die mehreren arithmetischen/logischen Operationen durchzuführen, wobei die Datenwörter X und Y jeweils aus einer Anzahl n>1 von Dateneinheiten zusammengesetzt sind und jede der Dateneinheiten aus M Bits besteht, wobei die Arithmetik-Logikeinheit umfasst:

    n Multipliziereinrichtungen (531, 541; 532, 542), die jeweils einer der n Dateneinheiten des Datenworts X entsprechen, wobei jede der Multipliziereinrichtungen (531, 541; 532, 542) so eingerichtet ist, dass sie eine jeweilige Dateneinheit des Datenworts X und eine jeweilige Dateneinheit des Datenworts Y multipliziert und ein Multiplikationsergebnis ausgibt,

    n Verschiebeeinrichtungen (55), die jeweils einer der n Multipliziereinrichtungen entsprechen, wobei die Verschiebeeinrichtungen so eingerichtet sind, dass sie in Abhängigkeit von der durch die Arithmetik-Logikeinheit durchzuführenden arithmetischen/logischen Operation entweder die Multiplikationsergebnisse so verschieben, dass die Ausgänge der n Verschiebeeinrichtungen an jeweilige andere Bitpositionen verschoben werden, oder die Multiplikationsergebnisse nicht verschieben, und

    eine Addiereinrichtung (56, 58), die so eingerichtet ist, dass sie die Ausgänge der n Verschiebeeinrichtungen addiert.


     
    2. Arithmetik-Logikeinheit nach Anspruch 1, wobei die mehreren arithmetischen/logischen Operationen eine Multiplizier-und-Summier-Operation der n Dateneinheiten der Datenwörter X und Y und n unabhängige Multiplizier-und-Summier-Operationen für jeweilige Paare von Dateneinheiten der Datenwörter X und Y einschließen,
    wobei die Verschiebeeinrichtungen so eingerichtet sind, dass sie die Multiplikationsergebnisse, die durch die Multipliziereinrichtungen eingegeben werden, so verschieben, dass die Ausgänge der n Verschiebeeinrichtungen an jeweilige Bitpositionen verschoben werden, die einander nicht überlappen, wenn die Arithmetik-Logikeinheit n unabhängige Multiplizier-und-Summier-Operationen für jeweilige Paare von Dateneinheiten der Datenwörter X und Y durchführen soll, und
    wobei die Verschiebeeinrichtungen so eingerichtet sind, dass sie die Multiplikationsergebnisse nicht verschieben, wenn die Arithmetik-Logikeinheit eine Multiplizier-und-Summier-Operation der n Dateneinheiten der Datenwörter X und Y durchführen soll,
    wobei die Arithmetik-Logikeinheiten des Weiteren ein Register (59) umfasst, das so eingerichtet ist, dass es den Ausgang der Addiereinrichtung (56, 58) als ein Datenwort Z speichert, das zu der Addiereinrichtung (56, 58) zurückgeführt werden kann,
    wobei die Addiereinrichtung (56, 58) so eingerichtet ist, dass sie die Ausgänge der n Verschiebeeinrichtungen und das gespeicherte Datenwort Z addiert, wenn die Arithmetik-Logikeinheit eine Multiplizier-und-Summier-Operation der n Dateneinheiten der Datenwörter X und Y, oder n unabhängige Multiplizier-und-Summier-Operationen für jeweilige Paare von Dateneinheiten der Datenwörter X und Y durchführen soll.
     
    3. Arithmetik-Logikeinheit nach Anspruch 1 oder 2, die des Weiteren n Wähleinrichtungen (521, 522) umfasst, wobei jede der n Wähleinrichtungen (521, 522) so eingerichtet ist, dass sie selektiv eine Dateneinheit aus den n Dateneinheiten des Datenwortes Y in eine entsprechende der n Multipliziereinrichtungen (531, 541; 532, 542) eingibt.
     
    4. Arithmetik-Logikeinheit nach Anspruch 3, wobei die p-te Wähleinrichtung der n Wähleinrichtungen so eingerichtet ist, dass sie, wenn die Arithmetik-Logikeinheit n unabhängige Multiplizier-und-Summier-Operationen für jeweilige Paare von Dateneinheiten der Datenwörter X und Y durchführen soll, die p-te niedrigstwertige Dateneinheit von Datenwort Y auswählt, um die p-te Multipliziereinrichtung der n Multipliziereinrichtungen zu veranlassen, die p-ten niedrigstwertigen Dateneinheiten der Datenwörter X und Y zu multiplizieren, und die Verschiebeeinrichtungen, die die Multiplikationsergebnisse von den Multipliziereinrichtungen empfangen, so eingerichtet sind, dass sie die Multiplikationsergebnisse an nicht überlappende Bitpositionen verschieben, und
    wobei die p-te Wähleinrichtung der n Wähleinrichtungen so eingerichtet ist, dass sie, wenn die Arithmetik-Logikeinheit eine Multiplizier-und-Summier-Operation der n Dateneinheiten der Datenwörter X und Y durchführen soll, die p-te niedrigstwertige Dateneinheit von Datenwort Y auswählt, um die p-te Multipliziereinrichtung zu veranlassen, die p-ten niedrigstwertigen Dateneinheiten der Datenwörter X und Y zu multiplizieren, und die Verschiebeeinrichtungen so eingerichtet sind, dass sie die jeweiligen empfangenen Multiplikationsergebnisse nicht verschieben.
     
    5. Arithmetik-Logikeinheit nach Anspruch 3, wobei die mehreren arithmetischen/logischen Operationen eine Multiplikationsoperation zum Multiplizieren der Datenwörter X und Y einschließen,
    wobei die p-te Wähleinrichtung der n Wähleinrichtungen so eingerichtet ist, dass sie, wenn die Arithmetik-Logikeinheit eine Multiplikationsoperation zum Multiplizieren der Datenwörter X und Y durchführen soll, die q-te niedrigstwertige Dateneinheit von Datenwort Y auswählt, um die p-te Multipliziereinrichtung der n Multipliziereinrichtungen zu veranlassen, die p-te Dateneinheit des Datenworts X und die q-te Dateneinheit des Datenworts Y zu multiplizieren, und die p-te Verschiebeeinrichtung, die das Multiplikationsergebnis der p-ten Dateneinheit von Datenwort X und der q-ten Dateneinheit von Datenwort Y von der p-ten Multipliziereinrichtung empfängt, so eingerichtet ist, dass sie das Multiplikationsergebnis um (p+q-2)M Bits verschiebt.
     
    6. Verfahren zum Durchführen mehrerer arithmetischer/logischer Operationen an Datenwörtern X und Y, die in eine Arithmetik-Steuereinheit eingegeben werden, die in der Lage ist, die mehreren arithmetischen/logischen Operationen durchzuführen, wobei die Datenwörter X und Y jeweils aus einer Anzahl n>1 von Dateneinheiten zusammengesetzt sind, die unabhängig voneinander sind, und jede der Dateneinheiten aus M Bits besteht und wobei das Verfahren die folgenden durch die Arithmetik-Steuereinheit durchgeführten Schritte umfasst:

    einen Multiplikationsschritt, der von n Multipliziereinrichtungen (531, 541; 532, 542) durchgeführt wird, die jeweils einer der n Dateneinheiten des Datenworts X entsprechen, wobei jede der Multipliziereinrichtungen (531, 541; 532, 542) eine jeweilige Dateneinheit des Datenworts X und eine jeweilige Dateneinheit des Datenworts Y multipliziert und ein Multiplikationsergebnis ausgibt,

    einen Verschiebeschritt, der von n Verschiebeeinrichtungen (55) durchgeführt wird, die jeweils einer der n Multipliziereinrichtungen entsprechen, wobei die Verschiebeeinrichtungen die Multiplikationsergebnisse so verschieben, dass die Ausgänge der n Verschiebeeinrichtungen an jeweilige verschiedene Bitpositionen verschoben werden, oder sie keine Verschiebung der Multiplikationsergebnisse durchführen, wobei die durch die Verschiebeeinrichtungen durchgeführte Operation von der arithmetischen/logischen Operation abhängt, die von der Arithmetik-Logikeinheit durchgeführt werden soll, und

    einen Addierschritt, der von einer Addiereinrichtung (56, 58) durchgeführt wird, wobei die Addiereinrichtung (56, 58) die Ausgänge der n Verschiebeeinrichtungen addiert.


     
    7. Verfahren nach Anspruch 6, wobei die mehreren arithmetischen/logischen Operationen eine Multiplizier-und-Summier-Operation der n Dateneinheiten der Datenwörter X und Y und n unabhängige Multiplizier-und-Summier-Operationen für jeweilige Paare von Dateneinheiten der Datenwörter X und Y enthalten,
    wobei bei dem Verschiebeschritt die Verschiebeeinrichtungen die durch die Multipliziereinrichtungen ausgegebenen Multiplikationsergebnisse so verschieben, dass die Ausgänge der n Verschiebeeinrichtungen an jeweilige Bitpositionen verschoben werden, die einander nicht überlappen, wenn die Arithmetik-Logikeinheit n unabhängige Multiplizier- und Summier-Operationen für jeweilige Paare von Dateneinheiten der Datenwörter X und Y durchführen soll, und
    wobei die Verschiebeeinrichtungen die Multiplikationsergebnisse nicht verschieben, wenn die Arithmetik-Logikeinheit eine Multiplizier-und-Summier-Operation der n Dateneinheiten der Datenwörter X und Y durchführen soll, und
    wobei das Verfahren des Weiteren die Schritte des Speicherns des Ausgangs der Addiereinrichtung (56, 58) in einem Register (59) der Arithmetik-Logikeinheit als ein Datenwort Z und das Zurückführen des Datenworts Z zu der Addiereinrichtung (56, 58) umfasst,
    wobei beim Schritt des Addierens die Addiereinrichtung (56, 58) die Ausgänge der n Verschiebeeinrichtungen und das gespeicherte Datenwort Z addiert, wenn die Arithmetik-Logikeinheit eine Multiplizier-und-Summier-Operation der n Dateneinheiten der Datenwörter X und Y oder n unabhängige Multiplizier-und-Summier-Operationen für jeweilige Paare von Dateneinheiten der Datenwörter X und Y durchführen soll.
     
    8. Verfahren nach Anspruch 6 oder 7, das des Weiteren einen Eingabeschritt umfasst, bei dem jede von n Wähleinrichtungen (521, 522) selektiv eine Dateneinheit von den n Dateneinheiten des Datenworts Y in eine jeweilige der n Multipliziereinrichtungen (531, 541; 532, 542) eingibt.
     
    9. Verfahren nach Anspruch 8, wobei, wenn die Arithmetik-Logikeinheit n unabhängige Multiplizier-und-Summier-Operationen für jeweilige Paare von Dateneinheiten der Datenwörter X und Y durchführen soll, die p-te Wähleinrichtung der n Wähleinrichtungen in dem Eingabeschritt die p-te niedrigstwertige Dateneinheit von Datenwort Y auswählt, um die p-te Multipliziereinrichtung der n Multipliziereinrichtungen zu veranlassen, die p-ten niedrigstwertigen Dateneinheiten der Datenwörter X und Y in dem Multiplikationsschritt zu multiplizieren, und die Verschiebeeinrichtungen, die die Multiplikationsergebnisse von den Multipliziereinrichtungen empfangen, in dem Verschiebeschritt die Multiplikationsergebnisse an nicht überlappende Bitpositionen verschieben, und
    wobei, wenn die Arithmetik-Logikeinheit eine Multiplizier-und-Summier-Operation der n Dateneinheiten der Datenwörter X und Y durchführen soll, die p-te Wähleinrichtung der n Wähleinrichtungen in dem Eingabeschritt die p-te niedrigstwertige Dateneinheit von Datenwort Y auswählt, um die p-te Multipliziereinrichtung der n Multipliziereinrichtungen zu veranlassen, die p-ten niedrigstwertigen Dateneinheiten der Datenwörter X und Y zu multiplizieren, und die Verschiebeeinrichtungen in dem Verschiebeschritt das jeweilige empfangene Multiplikationsergebnis nicht verschieben.
     
    10. Verfahren nach Anspruch 8, wobei die mehreren arithmetischen/logischen Operationen eine Multiplikationsoperation zum Multiplizieren der Datenwörter X und Y einschließen, und wenn die Arithmetik-Logikeinheit eine Multiplikationsoperation zum Multiplizieren der Datenwörter X und Y durchführen soll, die p-te Wähleinrichtung der n Wähleinrichtungen in dem Eingabeschritt die q-te niedrigstwertige Dateneinheit von Datenwort Y auswählt, um die p-te Multipliziereinrichtung der n Multipliziereinrichtungen zu veranlassen, die p-te Dateneinheit des Datenworts X und die q-te Dateneinheit des Datenworts Y in dem Multiplikationsschritt zu multiplizieren, und die p-te Verschiebeeinrichtung, die das Multiplikationsergebnis der p-ten Dateneinheit von Datenwort X und der q-ten Dateneinheit von Datenwort Y von der p-ten Multipliziereinrichtung empfängt, in dem Verschiebeschritt das Multiplikationsergebnis um (p+q-2)M Bits verschiebt.
     


    Revendications

    1. Unité arithmétique et logique pour effectuer plusieurs opérations arithmétiques/logiques sur des mots de données X et Y introduits à l'unité arithmétique et logique apte à effectuer la pluralité d'opérations arithmétiques/logiques, dans laquelle les mots de données X et Y sont chacun composé d'un nombre n>1 d'unités de données et chacune des unités de données est constituée de M bits, l'unité arithmétique et logique comprenant:

    n multiplicateurs (531, 541; 532, 542), correspondant respectivement à l'une des n unités de données du mot de données X, où chacun des multiplicateurs (531, 541; 532, 542) est adapté à la multiplication d'une unité de données respective du mot de données X et d'une unité de données respective du mot de données Y et pour délivrer en sortie un résultat de multiplication,

    n dispositifs de décalage (55), correspondant respectivement à l'un des n multiplicateurs, où selon l'opération arithmétique/logique à effectuer par l'unité arithmétique et logique, les dispositifs de décalage sont adaptés soit pour décaler les résultats de multiplication de telle manière que les sorties des n dispositifs de décalage soient décalées à différentes positions de bits respectives, soit pour ne pas décaler les résultats de multiplication, et

    un additionneur (56, 58) adapté pour additionner les sorties des n dispositifs de décalage.


     
    2. Unité arithmétique et logique selon la revendication 1, dans laquelle la pluralité d'opérations arithmétiques/logiques comporte une opération de multiplication et de totalisation des n unités de données des mots de données X et Y et n opérations de multiplication et de totalisation indépendantes pour des paires respectives d'unités de données des mots de données X et Y,
    où les dispositifs de décalage sont adaptés pour décaler les résultats de multiplication introduits par les multiplicateurs de telle manière que les sorties des n dispositifs de décalage soient décalées à des positions de bits respectives qui ne se recoupent pas, si l'unité arithmétique et logique doit effectuer n opérations de multiplication et de totalisation indépendantes pour des paires respectives d'unités de données des mots de données X et Y, et
    où les dispositifs de décalage sont adaptés pour ne pas décaler les résultats de multiplication, si l'unité arithmétique et logique doit effectuer une opération de multiplication et de totalisation des n unités de données des mots de données X et Y,
    où l'unité arithmétique et logique comprend en plus un registre (59) adapté pour stocker la sortie de l'additionneur (56, 58) sous forme de mot de données Z, qui peut être renvoyé à l'additionneur (56, 58),
    où l'additionneur (56, 58) est adapté pour additionner les sorties des n dispositifs de décalage et le mot de données Z stocké, si l'unité arithmétique et logique doit effectuer une opération de multiplication et de totalisation des n unités de données des mots de données X et Y, ou n opérations de multiplication et de totalisation indépendantes pour des paires respectives d'unités de données des mots de données X et Y.
     
    3. Unité arithmétique et logique selon la revendication 1 ou 2, comprenant en plus n sélectionneurs (521, 522) dans laquelle chacun des n sélectionneurs (521, 522) est adapté pour faire entrer, au choix, à l'un respectif parmi les n multiplicateurs (531, 541; 532, 542) une unité de données parmi les n unités de données du mot de données Y.
     
    4. Unité arithmétique et logique selon la revendication 3, dans laquelle, si l'unité arithmétique et logique doit effectuer n opérations de multiplication et de totalisation indépendantes pour des paires respectives d'unités de données des mots de données X et Y, le pième sélectionneur des n sélectionneurs est adapté pour sélectionner la pième unité de données de poids faible du mot de données Y, de sorte à amener le pième multiplicateur des n multiplicateurs à multiplier les pième unités de données de poids faible des mots de données X et Y, et les dispositifs de décalage recevant les résultats de multiplication à partir des multiplicateurs sont adaptés pour décaler les résultats de multiplication à des positions de bits qui ne se recoupent pas et
    où, si l'unité arithmétique et logique doit effectuer une opération de multiplication et de totalisation des n unités de données des mots de données X et Y, le pième sélectionneur des n sélectionneurs est adapté pour sélectionner la pième unité de données de poids faible du mot de données Y, de sorte à amener le pième multiplicateur à multiplier les pième unités de données de poids faible des mots de données X et Y et les dispositifs de décalage sont adaptés pour ne pas décaler les résultats de multiplication reçus respectifs.
     
    5. Unité arithmétique et logique selon la revendication 3, dans laquelle la pluralité d'opérations arithmétiques et logiques comporte une opération de multiplication pour multiplier les mots de données X et Y,
    où, si l'unité arithmétique et logique doit effectuer une opération de multiplication pour multiplier les mots de données X et Y, le pième sélectionneur des n sélectionneurs est adapté pour sélectionner la qième unité de données de poids faible du mot de données Y, de sorte à amener le pième multiplicateur des n multiplicateurs à multiplier la pième unité de données du mot de données X et la qième unité de données du mot de données Y, et le pième dispositif de décalage recevant le résultat de multiplication de la pième unité de données du mot de données X et la qième unité de données du mot de données Y du pième multiplicateur est adapté pour décaler le résultat de multiplication par (p+q-2)M bits.
     
    6. Procédé pour effectuer plusieurs opérations arithmétiques/logiques sur des mots de données X et Y introduits à une unité arithmétique et logique capable d'effectuer la pluralité d'opérations arithmétiques/logiques, où chacun des mots de données X et Y est composé d'un nombre n> 1 d'unités de données qui sont indépendantes entre elles et chacune des unités de données se compose de M bits, le procédé comprenant les étapes suivantes effectuées par l'unité arithmétique et logique ;
    une étape de multiplication effectuée par n multiplicateurs (531, 541; 532, 542), correspondant respectivement à l'une des n unités de données du mot de données X?, où chacun des multiplicateurs (531, 541; 532, 542) multiplie une unité de données respective du mot de données X et une unité de données respective du mot de données Y et délivre en sortie un résultat de multiplication,
    une étape de décalage effectuée par n dispositifs de décalage (55), correspondant respectivement à l'un des n multiplicateurs, où les dispositifs de décalage décalent les résultats de multiplication de telle manière que les sorties des n dispositifs de décalage soient décalées à différentes positions de bits respectives, ou n'effectuent aucun décalage des résultats de multiplication, où l'opération effectuée par les dispositifs de décalage dépend de l'opération arithmétique et logique à effectuer par l'unité arithmétique et logique, et
    une étape d'addition effectuée par un additionneur (56, 58), où l'additionneur (56, 58) additionne les sorties des n dispositifs de décalage.
     
    7. Procédé selon la revendication 6, dans lequel la pluralité d'opérations arithmétiques/logiques comporte une opération de multiplication et de totalisation des n unités de données des mots de données X et Y et n opérations de multiplication et de totalisation indépendantes pour des paires respectives d'unités de données des mots de données X et Y.
    où dans l'étape de décalage, les dispositifs de décalage décalent les résultats de multiplication délivrés en sortie par les multiplicateurs de telle manière que les sorties des n dispositifs de décalage soient décalées à des positions de bits respectives qui ne se recoupent pas, si l'unité arithmétique et logique doit effectuer n opérations de multiplication et de totalisation indépendantes pour des paires respectives d'unités de données des mots de données X et Y, et
    où les dispositifs de décalage ne décalent pas les résultats de multiplication, si l'unité arithmétique et logique doit effectuer une opération de multiplication et de totalisation des n unités de données des mots de données X et Y, et
    où le procédé comprend en plus les étapes de stockage de la sortie de l'additionneur (56, 58) dans un registre (59) de l'unité arithmétique et logique sous forme de mot de données Z, et de renvoi du mot de données Z à l'additionneur (56, 58),
    où dans l'étape d'addition, l'additionneur (56, 58) additionne les sorties des n dispositifs de décalage et le mot de données Z stocké, si l'unité arithmétique et logique doit effectuer une opération de multiplication et de totalisation des n unités de données des mots de données X et Y, ou n opérations de multiplication et de totalisation indépendantes pour des paires respectives d'unités de données des mots de données X et Y.
     
    8. Procédé selon la revendication 6 ou 7, comprenant en plus une étape d'entrée dans laquelle chacun des n sélectionneurs (521, 522) fait entrer, au choix, une unité de données parmi les n unités de données du mot de données Y à l'un un multiplicateur respectif parmi les n multiplicateurs (531, 541; 532, 542).
     
    9. Procédé selon la revendication 8, dans lequel, si l'unité arithmétique et logique doit effectuer n opérations de multiplication et de totalisation indépendantes pour des paires respectives des unités de données des mots de données X et Y, dans l'étape d'entrée, le pième sélectionneur des n sélectionneurs sélectionne la pième unité de données de poids faible du mot de données Y, de sorte à amener le pième multiplicateur des n multiplicateurs à multiplier les pième unités de données de poids faible des mots de données X et Y dans l'étape de multiplication, et dans l'étape de décalage, les dispositifs de décalage recevant les résultats de multiplication des multiplicateurs décalent les résultats de multiplication à des positions de bits qui ne se recoupent pas, et
    où, si l'unité arithmétique et logique doit effectuer une opération de multiplication et de totalisation des n unités de données des mots de données X et Y, dans l'étape d'entrée, le pième sélectionneur des n sélectionneurs sélectionne la pième unité de données de poids faible du mot de données Y, de sorte à amener le pième multiplicateur des n multiplicateurs à multiplier les pième unités de données de poids faible des mots de données X et Y, et dans l'étape de décalage les dispositifs de décalage ne décalent pas le résultat de multiplication reçu respectif.
     
    10. Procédé selon la revendication 8, dans lequel la pluralité d'opérations arithmétiques/logiques comprend une opération de multiplication pour multiplier les mots de données X et Y et
    où, si l'unité arithmétique et logique doit effectuer une opération de multiplication pour multiplier les mots de données X et Y, dans l'étape d'entrée, le pième sélectionneur des n sélectionneurs sélectionne la qième unité de données de poids faible du mot de données Y, de sorte à amener le pième multiplicateur des n multiplicateurs à multiplier la pième unité de données du mot de données X et la qième unité de données du mot de données Y, dans l'étape de multiplication, et dans l'étape de décalage, le piém dispositif de décalage recevant le résultat de multiplication de la pième unité de données du mot de données X et la qième unité de données du mot de données Y du pième multiplicateur décale le résultat de multiplication de (p+q-2)M bits.
     




    Drawing






























    REFERENCES CITED IN THE DESCRIPTION



    This list of references cited by the applicant is for the reader's convenience only. It does not form part of the European patent document. Even though great care has been taken in compiling the references, errors or omissions cannot be excluded and the EPO disclaims all liability in this regard.

    Patent documents cited in the description