FOR CALCULATING MULTIPLE TERMS OF A POLYNOMIAL

IN A SINGLE COMPUTING ELEMENT

Technical Field

This invention relates generally to computer ci rcuits and, more particularly, to computational arrays.

Background of the Invention

A computational array includes a number of computing elements that can be employed for various computer and electronics applications requiring parallel calculations. Each o f the computing elements performs a different calculation i n parallel as part of this computation. For some computations, a number of parallel calculations may be required which exceeds the number of parallel calculations typically required by most computations. As a result, not all computations can be performed by a computational array having what is normally a sufficient number of computing elements for the applications to which it i s typically applied. To accomplish the additional parallel calculations necessary in such circumstances, additional computing elements can be provided. However, some of the computing elements then go unused at other times when only a more typical number of calculations is required, which results i n a waste of valuable resources. Thus, it is desirable to provide a computational array which makes more efficient use of the computing elements therein.

Brief Description of the Drawings

FIG. 1 is a block circuit and data flow diagram of a computer system having a computational array provided i n accordance with a preferred embodiment of the invention.

FIG. 2 is a block circuit and data flow diagram of a computing element within the computational array provided i n accordance with the preferred embodiment of the invention.

FIG. 3 is a flowchart of the method performed by the computing element in accordance with the preferred embodiment of the invention.

FIG. 4 is a block diagram of a random access memory provided in accordance with the preferred embodiment of the invention and an illustration of the contents thereof.

FIG. 5 is a block diagram of a result register set provided in accordance with the preferred embodiment of the invention.

FIG. 6 is a block diagram of a weight register set provided in accordance with the preferred embodiment of the invention.

Description of the Preferred Embodiment

In a preferred embodiment of the invention, a computational array is provided which calculates polynomials efficiently. One method of calculating polynomials is to calculate each term of the polynomial in parallel, assigning a different computing element in the computational array to each different term, such that the computing elements employed each calculate exactly one term. In such a calculation, each term in the polynomial has a coefficient, or weight, multiplied by one or more variables, which may each be raised to a power. Each computing element calculates a different term in the polynomial by exponentiating each variable in the term and weighting each term , as appropriate. For some polynomial calculations, a number o f terms must be calculated which exceeds the number typically present in a polynomial. In such a case, where each computing element of a computational array calculates only one term of the polynomial, as described above, a number of computing elements are required which exceeds a more typical number of computing elements otherwise sufficient to calculate the polynomial. As a result, a computational array that has only such a typical number of computing elements cannot calculate the polynomial unless more computing elements are included in the computational array. However, some of the computing elements then go unused at times when a more typical number of calculations is required, which results in a waste of valuable resources.

The preferred embodiment of the invention provides the advantage of overcoming this problem by providing a computational array which includes at least one computing element that calculates multiple terms of a polynomial. The computing element obtains an input value of a variable in one o f the multiple terms and obtains a subscript uniquely identifying the variable. The computing element reads a term identifier and an exponent from a memory at a memory location based on the

subscript. The term identifier uniquely identifies a term th at includes the variable identified by the subscript. Because a variable is used to identify the term to be calculated, the computing element only calculates multiple terms that do not share any of the same variables. Thus, the term can be identified based on the variable. The exponent indicates a power to wh ich the variable is to be raised. The computing element m u ltipl ies the input value by the input value itself a number of times based on the exponent to produce an exponentiated value. This process is performed for each of one or more variables in each different term calculated by the computing element. For each different term, if the term includes multiple variables, the computing element multiplies the corresponding multiple exponentiated values together, to produce a term value, and multiplies the term value by a weight value corresponding to the term identifier to produce a weighted term value. Thus, where the computing element calculates multiple terms, the computing element produces corresponding multiple weighted term values. The weighted term value for each different term is stored in the memory at a different memory location based on each different term identifier corresponding to each different term.

By using each variable to determine the term identifier, and by using the term identifier to distinguish each different term calculated by the computing element, apply the appropriate weight value to each term, and store the weighted term value calculated for each different term in a memory location based on the term identifier, a single computing element is able to calculate multiple terms. As a result, the computational array of the preferred embodiment provides the advantage of calculating a greater number of terms than a computational array with a same number of computing elements which only calculates one term per computing element. Similarly, the computational array of the preferred embodiment provides the advantage of calculating a same number of terms as a computational array having even more computing elements but

which only calculates one term per computing element. This provides the advantages of reducing or eliminating unused computing elements, and thus conserving resources.

In a particularly preferred embodiment of the invention, each computing element obtains the input value from a host computer and obtains a subscript address as the subscript from the host computer. The host computer initially assigns the terms to be calculated by different computing elements such th at no computing element calculates terms sharing any of the same variables. Each computing element reads the term identifier and the exponent from a random access memory at the subscript address. The computing element selects a corresponding coefficient, or weight value, to apply to each term calculated by the computing element by initially storing the weight value as a result value in a selected result register corresponding to the term identifier. The computing element multiplies the result value, which is initially the weight value, by the input value a number of times based on the value of the exponent, storing each resulting product back into the selected result register each ti m e as the result value. For each additional variable having the same term identifier stored at the subscript address, and thus belonging to the same term, the computing element again multiplies the result value by the input value a number of times based on the value of the exponent, storing each resulting product back into the selected result register each time as the resu lt value. The result value stored in the result register is thus a partially calculated value of the term until completion o f calculation of the term, at which time the result value becomes the finally calculated value of the term.

By maintaining the result values in different result registers for different terms, the computing element correctly calculates the terms regardless of the order in which it receives the variables. That is, the computing element may receive first a variable in a first term, then a variable in a second term, then

another variable in the first term, and so forth. Since the resu lt value is maintained in a different result register for each different term, each term will be calculated correctly. Thus, the host computer does not have to sort the input values and subscript addresses for the variables, thereby providing the advantage o f faster and more efficient calculation.

FIG. 1 is a block circuit and data flow diagram of a computer system having a computational array provided i n accordance with a preferred embodiment of the invention. In FIG. 1 , a host computer 110 is connected to a computational array 120. The host computer 1 10 is, for example, any conventional computer which runs software that provides inputs to the computational array 120 and obtains outputs from the computational array 1 20 to calculate the terms of a polynomial. In the preferred embodiment, the computational array 120 is implemented i n digital circuitry such as an integrated circuit, which can be provided in a separate computer, in a separate device, as a separate hardware element or as part of the host computer 1 1 0. Where the computational array 120 is implemented as an integrated circuit, the invention accordingly provides the advantage of conserving chip area. One of ordinary skill in the a rt will recognize that the computational array can be implemented in a variety of ways in accordance with the invention described herein.

The host computer 1 10 utilizes the computational array 120 to perform applications requiring parallel computations, such as the polynomial calculation described herein. The computational array 120 includes a number o f computing elements 130 which perform computations in parallel. The computing elements 130 receive various inputs from the host computer 110 through the computational array 120, perform the appropriate computations and provide outputs to the host computer 1 10 based on the results of those computations. An example of a computational array which calculates a polynomial

is described in fMNEQ0315l . entitled "COMPUTATIONAL ARRAY CIRCUIT FOR PROVIDING PARALLEL MULTIPLICATION, filed March 3. 1995 . which is hereby incorporated by reference. Additional ly, a description of an existing computing element can be found i n U.S. Patent 5,390,136, entitled "ARTIFICIAL NEURON AND METHOD OF USING SAME", issued February 14, 1995, which is also hereby incorporated by reference.

The computing element 130 of the preferred embodiment will now be more specifically described. An example of a polynomial that might be calculated by the computational array of the preferred embodiment is

y = W-| Xι ³ X2 ² X4 ⁺ W2X1 X3 ^{2 +} W3X3X5

wherein w-| X-| ³ X2 ² X ₄ , W ₂ X ₁ X3 , and W3X3X5 are terms, x-i , X2, X ₃ , X ₄ , and x ₅ are variables, and w-i , W2, and w ₃ are weights. The computing element 130 has the capability of calculating m ulti pl e terms in the polynomial, and thus can calculate, for example, both the w -| Xι ³ X X4 term and the W ₃ X3X ₅ term in the above example, which do not have the same variables. FIG. 2 is a block circuit and data flow diagram of a computing element 130 within the computational array 120. As shown in FIG. 2, the computing element 130 includes a control unit 210 which receives an in put value and a subscript address from the host computer 1 10 of FIG. 1 . The input value is the value of a variable in a term in the polynomial to be calculated by the computing element 130. The subscript address uniquely corresponds to the variable having the input value.

The control unit 210 reads a term identifier and an exponent from a random access memory 220 at the subscript address obtained from the host computer 1 10. Alternatively, as one of ordinary skill in the art will recognize, the random access memory 220 could be any type of memory or storage device. Term identifiers and exponents are previously stored by the host

computer based on the terms and variables therein to be calculated by the computing element, as will later be explained i n more detail. The term identifier uniquely identifies the term th at includes the variable having the currently received input value. The exponent is the exponent of the variable having the input value, and is thus the power to which the input value is to be raised. The control unit 210 provides the input value, the term identifier and the exponent to a computational unit 230.

The computational unit 230 calculates each term to be calculated based on the input value and exponent of all of the variables in all of the terms to be calculated and based on a selected weight value for each term. Specifically, the weight register set 240 contains a number of weight registers each storing a weight value to be applied to a corresponding term among the multiple terms which the computing element 1 30 calculates. The weight register set 240 obtains the weight values as weight inputs from the host computer 1 10. One o f ordinary skill in the art will understand the process of selecting and modifying weight values in calculating a polynomial. The weight inputs are each stored as a weight value for a different term in a different weight register in the weight register set 240. Prior to each calculation of the polynomial, the weight register set provides the weight values in the weight registers t o corresponding result registers in a result register set 250. Thus, initially, before the polynomial is calculated, each result register in the result register set 250 stores the weight value to be applied to each different term. As will be explained in more detail below, during calculation of the polynomial, each resu lt register stores for a corresponding term a changing result value while calculating the term.

One of ordinary skill will in the art will recognize that the computational unit 230 can be implemented in a variety of ways. For example, the computational unit can include a multiplier or can include an adder where the computational array

is provided in a logarithmic calculation environment. For example, the computational array could by provided between a logarithmic conversion circuit and an antilogarithmic conversion circuit, similar to the system which is described in fMNE00315) . entitled "COMPUTATIONAL ARRAY CIRCUIT FOR PROVIDING PARALLEL MULTIPLICATION, filed March 3. 1995. as mentioned above.

The method performed by the computing element 1 30 of FIG. 2 will now be explained in detail. Initially, the host computer 1 10 stores a term identifier and an exponent in the random access memory 220 of each computing element 130 a t different memory locations, each corresponding to one of the variables in all of the terms to be calculated by the computing element 130. As noted above, the host computer 1 10 assigns t o each computing element 130 only terms that do not have the same variables. Thus, each variable in the terms calculated by a computing element 130 is represented once in the random access memory 220. Prior to each polynomial calculation, the host computer 1 10 provides weight inputs to the computing element 130 which the computing element 130 stores as weight values i n different weight registers corresponding to each of the different terms among the multiple terms to be calculated by the computing element. This initializes the computing element 130 to calculate multiple terms. FIG. 3 is a flowchart of the method performed by the computing element 130 to calculate a polynomial term i n accordance with the preferred embodiment of the invention. I n step 310 of FIG. 3, for each term "t" of "n" terms to be calculated by the computing element 130, the computing element 1 30 initializes the result value, called herein result[t], stored in a corresponding one of the result registers in a corresponding one of the result registers in the result register set 250 to the weight value for the same term t, called herein weight[t], in the weight register set 240.

In step 320, the computing element 130 obtains the input value for a variable in a term to be calculated from the host computer 110. In step 330, the computing element 130 obtains the subscript address for the variable from the host computer 1 10. In step 340, the computing element 130 reads the term identifier from the random access memory 220 shown in FIG. 2, at the subscript address obtained in step 330. For example, if the variable of the polynomial

y = W ₁ X1 ³ X2 ² X4 ⁺ W2X1 X3 ^{2 +} W3X3X5

is X ₅ from the term W ₃ X ₃ Xs, the corresponding subscript is 5 and the corresponding subscript address is 0005, then the computing element 130 reads the term identifier from the random access memory 220 at address 0005. In the preferred embodiment, the random access memory 220 is as illustrated in FIG. 4. Each byte of the random access memory 220 contains the term identifier and the exponent. In one possible embodiment, the random access memory 220, or a portion of the random access memory 220 devoted to the functions described herein, is 128 bytes, as shown in FIG. 4. One of ordinary skill in the art would recognize, however, that the memory storing the term identifier and exponent could be of many potential forms and sizes, including differently sized or identified storage locations for the term identifier and the exponent stored at a location based on the subscript, and will implement a memory having the attributes best suited for the application at hand.

The random access memory 220 in FIG. 4 contains a one byte variable field 410 corresponding to each variable. The subscript address of the 128 byte random access memory 220 shown in FIG. 4 thus ranges from address 0000 to 0127. Each variable field contains a term field 420 and an exponent field 430. The term field 420 stores the term identifier and the exponent field 430 stores the exponent. Thus, returning to the example above, the computing element 130 reads the term identifier from

address 0005 in the random access memory 220. One of ordinary skill will recognize that the amount of memory devoted to each field can be selected based on the application at hand. For example, four bits might be designated for each variable, where 2 bits represent the term identifier, which would thus have fou r possible values 0-3, and 2 bits represent the exponent, which would thus have four possible values 0-3.

Returning to FIG. 3, in step 350, the computing element 130 reads the exponent from the random access memory 220 shown in FIG. 2, at the subscript address obtained in step 330. For example, if the current variable is X5 and the subscript is "5", then the computing element 130 reads the exponent from the random access memory 220 at address 0005. Because X ₅ has an exponent of "1 ", the computing element 130 would read the number "1 " from the exponent field 430 at address 0005. Alternatively, if the variable is, for example, x^, then the exponent is "3", and thus the computing element 130 would read the number "3" from the exponent field 430 at address 0001.

In step 360 - 380 of FIG. 3, the computing element 130 computes the value of the variable based on the input value and the exponent. In step 360, the computing element 130 sets a count value to the value of the exponent. If the variable has an exponent of "3", the count would be set to a value of 3. In step 365, the computing element 130 determines whether the count remains above zero. In the case of a constant, that is, a coefficient or weight value that is not multiplied by a variable, then the exponent is equal to 0. In such a case, the value of the count is 0 so that control branches to step 385 the first time step 365 is executed, and steps 370 and 380 are never performed. Thus, the result value result[t] remains equal to the weight value weight[t] stored in the weight register 600 for the same term. Continuing the previous example, however, where the value of the count is 3, control proceeds to step 370. In step 370, the computing element 130 reads result[t] from a memory location

based on the term identifier, multiplies result[t] by the input value and stores the resulting product as the new result value resultft] back into the same memory location. For example, the computing element 130 reads resultft] from a selected resu lt register from the result register set 250 corresponding to the term identifier. Initially, the value in the selected resu lt register is the same as the value in the weight register in the weight register set 240 corresponding to the same term identifier.

The result register set 250 of the preferred embodiment is shown in FIG. 5. The result register set 250 contains a result register 500 for each term. Thus, for n possible terms, the result register set 250 contains result registers #0 , #1 , #2 ... #n. The weight register set 240 of the preferred embodiment is shown in FIG. 6. Similarly, the weight register set 240 of the preferred embodiment contains weight registers #0, #1 , #2 ... #n. Continuing the above example where the variable i s X ₅ , as explained above, the computing element 130 reads the term identifier from the random access memory 220 at the subscript address. Assuming the term identifier stored in the term f i el d 420 of the variable field 410 at subscript address 0005 is "2", then the selected result register 500 is result register #2. The value of result[2] initially stored in result register #2 is the value of weight[2] stored in weight register #2. Assuming a weight input for term #2 equal to 0.5, then 0.5 is stored in resu lt register #2. Thus, as described above, the value of result[term identifier] r[t] stored in result register #2 is also equal to 0.5. Assuming an input value for the variable xi is equal to 10.0, then the first time step 370 is executed, the input value 10.0 i s multiplied by the weight[t] 0.5 and the result value 5.0 is stored in result register #2 as a weighted input value.

In step 380, the count is decremented and control loops back to step 365. This ensures that steps 370 and 380 w i l l be executed a number of times equal to the value of the exponent,

which in turn assures that the value of the current variable i s exponentiated to the correct degree. Continuing with the above example, the count is decremented to 2 and control proceeds t o step 370. In step 370, the current value of result[2] stored i n result register #2 is again multiplied by the input value. Thus, 5.0 would be multiplied by 10.0 and the result value 50.0 would be stored back in result register #2. In step 380, the count i s decremented to 1 and control again proceeds to 370, where the value 50.0 stored in result register #2 is again multiplied by the input value 10.0 and the result value 500.0 is stored back i n result register #2 as a weighted exponentiated value. The count is decremented to 0, and control loops to step 365. This time, i n step 365, control branches to step 385. In step 385, the computing element 130 determines whether more input i s received from the host computer 1 10. If so, control loops to step 320 and the next input value and subscript address are obtained.

If, in the above example, the entire term to be calculated were w^i ³ , then 500.0 would be the final result stored in result register #2 when, in step 385, the computing element 130 finally determines that no more input is received. I f the next input value and subscript address received correspond to a variable in a different term, a term identifier other than "2" will be stored in the term field 420 of the appropriate variable field 410, at the subscript address received, in the random access memory 220. Thus, the new variable will be calculated and stored in a different result register 500 in the result register set 250. A different weight[t] will be applied to each different term, since a different weight[t] was initially stored in the result register 500 for each term in step 310. On the other hand, if the next input value and subscript address received correspond to another variable in the same term as before, the term identifier stored i the term field 420 of the variable field 410, at the subscript address received, in the random access memory 220 will also be "2". As a result, the next input value will be multiplied by the current result[t] in result register #2, which is initially 500.0,

the number of times indicated by the exponent stored at the new subscript address and the result value will be stored in resu lt register #2 each time.

Upon completion of calculation of all of the terms the computing element 130 is to calculate, the computing element 130 then outputs the final output values in the result registers 500 to the host computer 1 10. Alternatively, the host computer 1 10 reads the output value from result registers 500. The host computer 1 10 may contain a sum circuit, implemented in either hardware or software, which calculates the sum of every output value from all the computing elements 130 to calculate the polynomial. Alternatively, the computer system of FIG. 1 may contain an additional sum circuit (not shown) provided, fo r example, between the computational array and the host computer, which calculates a sum of every output value from all the computing elements 130 to calculate the polynomial.

The computational array 120 described above provides the advantages of calculating a greater number of terms than a computational array with a same number of computing elements that only calculates one term per computing element. Simi larly, the computational array 120 of the preferred embodiment provides the advantages of calculating a same number of terms as a computational array having even more computing elements which only calculates one term per computing element. Th is provides the advantage of conserving resources. For example, where the computational array 120 is implemented as an integrated circuit, chip area is conserved. While specific embodiments of the invention have been shown and described, further modifications and improvements will occur to those skilled in the art. It is understood that this invention is not limited to the particular forms shown and it is intended for the appended claims to cover all modifications of the invention which fall within the true spirit and scope of the invention. What is claimed is: