Parallel data processing in a single processor

<u>Background</u>

The present invention concerns parallel data processing in a single processor system.

In general, single processor systems sequentially perform operations on two operands. For example, in a 32-bit computer, each integer operand is 32 bits. In a 64-bit computer, each integer operand is 64 bits. Thus an integer "add" instruction, in a 64-bit computer, adds two 64-bit integer operands to produce a 64-bit integer result. In most pipelined 64-bit processors, a 64-bit add instruction takes one cycle of execution time.

In many instances the pertinent range of operands is 16 bits or less. In current 32-bit and 64-bit computers, however, it still takes a full instruction to perform an operation on a pair of 16-bit operands. Thus the number of execution cycles required to perform an operation on two 16-bit operands is the same as the number of execution cycles required to perform the operation on two 32-bit operands in a 32-bit computer or two 64-bit operands in a 64-bit computer.

In the prior art, parallel data processing required replicating of functional units, each functional unit able to handle full word length data. See for example, Michael Flynn, <u>Very High-Speed Computing Systems,</u><i>Proceedings of IEEE</i>, Vol. 54, Number 12, December 1966, pp. 1901-1909. However, such implementations of parallel processing is significantly costly both in terms of hardware required and complexity in design.

<u>Summary of the Invention</u>

In accordance with the preferred embodiment of the present invention as defined in the appended claims, a system is presented which allows parallel data processing within a single processor. In order to allow for parallel processing of data, an arithmetic logic unit or other operation executing entity within the processing system such as a shifter is partitioned. Within each partition operations are performed. When the operation is to be performed on full word length operands, there is no parallel processing. Thus data is allowed to freely propagate across boundaries between the partitions. When performing the operation in parallel using a plurality of less than one full word length operands, data is prevented from being propagated across at least one boundary between the partitions.

For example, when the operation is an addition operation (e.g., a two's complement addition), each of the plurality of partitions performs an addition operation. When the addition is to be performed on full word length operands, carries are allowed to propagate between the partitions. When performing the addition operation in parallel on a plurality of less than one full word length operand sets, a carry is prevented from propagating across at least one boundary between the partitions.

Likewise, when the operation is a shift, each of the plurality of partitions performs a shift operation. When the shift is to be performed on full word length operands, shifts are allowed between the partitions. When performing the operation in parallel using a plurality of less than one full word length operands, a shift is prevented from crossing at least one boundary between the partitions.

Also in accordance with a preferred embodiment of the present invention, a multiplier implements both multiplication of whole word multiplicands and parallel multiplication of sub-word multiplicands. Circuitry, for example an array of logic AND gates (or their equivalent), generates partial products. Partial product sum circuitry, sums the partial products to produce a result. Partial product gating means, in response to the selection of parallel multiplication of sub-word multiplicands, forces selected partial products to have a value of 0, thereby implementing parallel multiplication of sub-word multiplicands. When the multiplier is implementing whole word multiplication, none of the partial products are forced to have a value of 0. The partial product gating means may be implemented, for example, using third inputs to at least a portion of the logic AND gates.

The present invention allows for a single processor system to significantly increase performance by facilitating parallel processing operations when operands are less than the full word length. This inexpensive use of parallelism results in a huge increase in performance for computations that can utilize this type of data parallelism without significant additional cost in silicon space on a processor chip or complexity in design. The present invention also allows for parallel processing operations performed by a processor in response to a single instruction.

<u>Brief Description of the Drawings</u>

Figure 1 shows a simplified block diagram of an operation execution data path within a processor in accordance with preferred embodiments of the present invention.

Figure 2 shows a simplified block diagram of an arithmetic logic unit (ALU) shown in Figure 1 in accordance with a preferred embodiment of the present invention.

Figure 3 shows an implementation of a two's complement adder within the ALU shown in Figure 2 in accordance with a preferred embodiment of the present invention.

Figure 4 shows an alternate simplified block diagram of the arithmetic logic unit (ALU) shown in Figure 1 in accordance with an . alternate preferred embodiment of the present invention.

Figure 5 shows another alternate simplified block diagram of the arithmetic logic unit (ALU) shown in Figure 1 in accordance with another alternate preferred embodiment of the present invention.

Figure 6 shows an implementation of a shifter shown in Figure 1 in accordance with a preferred embodiment of the present invention.

Figure 7 shows a multiplier in accordance with the prior art.

Figure 8 and Figure 9 show a multiplier implemented in accordance with preferred embodiments of the present invention.

Figure 10 shows an implementation of a carry look-ahead adder within the ALU shown in Figure 1 in accordance with an alternate preferred embodiment of the present invention.

Figure 11 shows an example of an instruction layout in accordance with an alternate preferred embodiment of the present invention.

<u>Description of the Preferred Embodiments</u>

Figure 1 shows a simplified block diagram of an operation execution data path within a processor in accordance with preferred embodiments of the present invention. Operands for upcoming operations and results from accomplished operations are stored within general registers 25. When operations are performed, a first operand stored in a first register within general registers 25 is placed on a first source bus 21. If the operation requires another operand, a second operand stored in a second register within general registers 25 is placed on a second source bus 22.

After performance of the operation, the result is placed on a result bus 23 and loaded into a register within general registers 25. The operation is performed by arithmetic logic unit (ALU) 26 or by a shifter 29. A pre-shifter 27 and complement circuitry 28 may each be used to modify operands before they are received by ALU 26. For general background about the architecture of single processor systems constructed similarly to the present invention see, for example, Ruby B. Lee, <u>Precision Architecture</u>, <i>IEEE Computer,</i> Volume 22, No. 1, January 1989, pp. 78-91.

In accordance with the preferred embodiments of the present invention, the ALU may be partitioned to allow parallel data processing. For example, Figure 2 shows ALU 26 divided into two partitions. A first partition 41 performs operations on low order bits 42 of a first operand and low order bits 43 of a second operand to produce low order bit results 44. A second partition 51 performs operations on high order bits 52 of the first operand and high order bits 53 of the second operand to produce high order bit results 54.

In response to a control input 49, a selector 50 is used to allow information on data path 45 to propagate from first partition 41 to second partition 51 or to intercept information on data path 45 before it is propagated from first partition 41 to second partition 51. Particularly, for arithmetic operations performed on full-word operands, information is allowed to propagate from first partition 41 through selector 50 to second partition 51. For the performance of parallel arithmetic operations on half-word operands, selector 50 prevents information from propagating from first partition 41 to second partition 51. Generally, in logic operations, there is no propagation of information from first partition 41 to second partition 51.

For example, in a computer which has a thirty-two bit wide data path, each full-word operand is 32 bits. Therefore, when performing operations using 32-bit full word operands, selector 50 allows information to propagate from first partition 41 through selector 50 to second partition 51. When performing two parallel operations using 16-bit half word operands, selector 50 prevents information from propagating from first partition 41 through selector 50 to second partition 51. Instead the value on a line 59 is forwarded to partition 51. When an "add" is being performed, a logic 0 is placed on input line 59. When a "subtract" is being performed, a logic 1 is placed on input line 59.

In the preferred embodiment of the present invention, a common arithmetic operation performed by ALU 26, shown in Figure 1, is two's complement addition. As is understood by those skilled in the art, the use of two's complement circuitry 28 to perform a two's complement on an operand before performing a two's complement addition operation in the ALU implements a two's complement subtraction. Also, the use of pre-shifter 27 to pre-shift an operand before performing a two's complement addition operation in the ALU implements a shift and add operation.

Figure 3 shows implementation of a two's complement adder with carry propagate addition within ALU 26 in accordance with a preferred embodiment of the present invention. Alternately, ALU 26 includes a two's complement adder with carry look-ahead. A half adder 60 receives a single bit X₀ of a first operand and a single bit Y₀ of a second operand. Half adder 60 produces a sum bit Z₀ and a carry bit C₀. A full adder 61 receives a single bit X₁ of the first operand, a single bit Y₁ of the second operand and carry bit C₀. Full adder 61 produces a sum bit Z₁ and a carry bit C₁. A full adder 65 receives a single bit X_i-1 of the first operand, a single bit Y_i-1 of the second operand and a carry bit from a previous adder (i.e., C_i-2, not shown). Full adder 65 produces a sum bit Z_i-1 and a carry bit C_i-1. A full adder 66 receives a single bit X_i of the first operand and a single bit Y_i of the second operand. Depending on a value of enable bit 49, full adder 66 also receives, through selector 50 (or equivalent logic circuitry as will be understood by persons of ordinary skill in the art), carry bit C_i-1. Full adder 66 produces a sum bit Z_i and a carry bit C_i. A full adder 69 receives a single bit X_j-1 of the first operand, a single bit Y_j-1 of the second operand and a carry bit from a previous adder (not shown). Full adder 69 produces a sum bit Z_j-1 and a carry bit C_j-1.

In the embodiment of the adder shown in Figure 3, "j" is the size of the data path and the bit length of full word operations. Also, "i" is equal to "j" divided by 2. For example, "j" is equal to 32 and "i" is equal to 16.

Selector 50 is also shown in Figure 3. When performing operations using "j"-bit full word operands, enable bit 49 is equal to logic one and allows a carry to propagate through selector 50 to full adder 66. When performing two parallel operations using "i"-bit half word operands, enable bit 49 is equal to logic zero and prevents the carry from propagating through selector 50 to full adder 66. Instead the value on line 59 is forwarded to full adder 66. When an "add" is being performed, a logic 0 is placed on input line 59. When a "subtract" is being performed, a logic 1 is placed on input line 59.

While Figures 2 and 3 discuss implementations of ALU 26 with two partitions, an ALU designed in accordance with other preferred embodiments of the present invention may variously partition an ALU. For example, Figure 4 shows an alternate simplified block diagram of ALU 26 in accordance with an alternate preferred embodiment of the present invention. In Figure 4, ALU 26 is divided into four partitions. A first partition 71 performs operations on low order bits 72 of a first operand and low order bits 73 of a second operand to produce low order bit results 74. A second partition 81 performs operations on bits 82 of the first operand and bits 83 of the second operand to produce result bits 84. A third partition 91 performs operations on bits 92 of the first operand and bits 93 of the second operand to produce result bits 94. A fourth partition 101 performs operations on high order bits 102 of the first operand and high order bits 103 of the second operand to produce high order bit results 104.

In response to a control input 79, a selector 80 is used to allow information on data path 75 to propagate from first partition 71 to second partition 81 or to intercept information on data path 75 before it is propagated from first partition 71 to second partition 81. Particularly, for arithmetic operations performed on full-word operands or half-word operands, information is allowed to propagate from first partition 71 through selector 80 to second partition 81. For the performance of parallel arithmetic operations on quarter-word operands, selector 80 prevents information from propagating from first partition 71 to second partition 81. Instead the value on a line 88 is forwarded to partition 81. When an "add" is being performed, a logic 0 is placed on line 88. When a "subtract" is being performed, a logic 1 is placed on line 88. Generally, in logic operations, there is no propagation of information between partitions.

In response to a control input 89, a selector 90 is used to allow information on data path 85 to propagate from second partition 81 to third partition 91 or to intercept information on data path 85 before it is propagated from second partition 81 to third partition 91. Particularly, for arithmetic operations performed on full-word operands, information is allowed to propagate from second partition 81 through selector 90 to third partition 91. For the performance of parallel arithmetic operations on quarter-word operands or half-word operands, selector 90 prevents information from propagating from second partition 81 to third partition 91. Instead the value on a line 98 is forwarded to partition 91. When an "add" is being performed, a logic 0 is placed on line 98. When a "subtract" is being performed, a logic 1 is placed on line 98.

In response to a control input 99, a selector 100 is used to allow information on data path 95 to propagate from third partition 91 to fourth partition 101 or to intercept information on data path 95 before it is propagated from third partition 91 to fourth partition 101. Particularly, for arithmetic operations performed on full-word operands and half-word operands, information is allowed to propagate from third partition 91 through selector 100 to fourth partition 101. For the performance of parallel arithmetic operations on quarter-word operands, selector 100 prevents information from propagating from third partition 91 to fourth partition 101. Instead the value on a line 108 is forwarded to partition 101. When an "add" is being performed, a logic 0 is placed on line 108. When a "subtract" is being performed, a logic 1 is placed on line 108.

For example, in a computer which has a sixty-four bit wide data path, each full-word operand is 64 bits. Therefore, when performing operations using 64-bit full word operands, selector 80 allows information to propagate from first partition 71 through selector 80 to second partition 81, selector 90 allows information to propagate from second partition 81 through selector 90 to third partition 91, and selector 100 allows information to propagate from third partition 91 through selector 100 to fourth partition 101. When performing two parallel operations using 32-bit half word operands, selector 80 allows information to propagate from first partition 71 through selector 80 to second partition 81, selector 90 prevents information from propagating from second partition 81 through selector 90 to third partition 91, and selector 100 allows information to propagate from third partition 91 through selector 100 to fourth partition 101. When performing four parallel operations using 16-bit quarter word operands, selector 80 prevents information from propagating from first partition 71 through selector 80 to second partition 81, selector 90 prevents information from propagating from second partition 81 through selector 90 to third partition 91, and selector 100 prevents information from propagating from third partition 91 through selector 100 to fourth partition 101.

Figure 5 shows an another alternate simplified block diagram of ALU 26 in accordance with another alternate preferred embodiment of the present invention. In Figure 5, ALU 26 is divided into partitions which are each one bit wide. A first partition 111 performs operations on a low order bit 112 of a first operand and on a low order bit 113 of a second operand to produce a low order result bit 114. A second partition 121 performs operations on a bit 122 of the first operand and a bit 123 of the second operand to produce a result bit 124. A partition 131 performs operations on a bit 132 of the first operand and a bit 133 of the second operand to produce a result bit 134. A partition 141 performs operations on a bit 142 of the first operand and a bit 143 of the second operand to produce a result bit 144. A partition 151 performs operations on a high order bit 152 of the first operand and a high order bit 153 of the second operand to produce a high order result bit 154.

In response to a control input 119, a selector 120 is used to allow information on data path 115 to propagate from first partition 111 to second partition 121 or to intercept information on data path 115 before it is propagated from first partition 111 to second partition 121. When data is intercepted the value on a line 128 is forwarded to partition 121. When an "add" is being performed, a logic 0 is placed on line 128. When a "subtract" is being performed, a logic 1 is placed on line 128.

In response to a control input 129, a selector 130 is used to allow information on a data path from an immediately prior partition (not shown) to propagate from the immediately prior partition to partition 131 or to intercept information on the data path from the immediately prior partition before it is propagated to partition 131. When data is intercepted the value on a line 138 is forwarded to partition 131. When an "add" is being performed, a logic 0 is placed on line 138. When a "subtract" is being performed, a logic 1 is placed on line 138.

In response to a control input 139, a selector 140 is used to allow information on data path 135 to propagate from partition 131 to partition 141 or to intercept information on data path 135 before it is propagated from partition 131 to partition 141. When data is intercepted the value on a line 148 is forwarded to partition 141. When an "add" is being performed, a logic 0 is placed on line 148. When a "subtract" is being performed, a logic 1 is placed on line 148.

In response to a control input 149, a selector 150 is used to allow information on a data path from an immediately prior partition (not shown) to propagate from the immediately prior partition to partition 151 or to intercept information on the data path from the immediately prior partition before it is propagated to partition 151. When data is intercepted the value on a line 158 is forwarded to partition 151. When an "add" is being performed, a logic 0 is placed on line 158. When a "subtract" is being performed, a logic 1 is placed on line 158.

The control inputs to the selectors may be used to allow parallel processing of operands of varying length. For example, in a processing system with a sixty-four bit wide data path, the control inputs could be selected so that parallel processing of two sixteen bit and four eight-bit arithmetic operations are all performed simultaneously. Additionally any bit combination which add up to no more than the word size could be used. For example, parallel processing of seventeen bit, three bit, sixteen bit, twelve bit, five bit, and eleven bit arithmetic operations can also be performed simultaneously.

The principles discussed above also apply to a carry look-ahead adder. For example, Figure 10 shows implementation of a two's complement adder with carry look-ahead within ALU 26 in accordance with another preferred embodiment of the present invention. A carry look-ahead circuit 470 produces carries for the adder. A half adder 460 receives a single bit X₀ of a first operand and a single bit Y₀ of a second operand. Half adder 460 produces a sum bit Z₀. A full adder 461 receives a single bit X₁ of the first operand, a single bit Y₁ of the second operand and a carry bit carry bit C₀. Full adder 461 produces a sum bit Z₁. A full adder 465 receives a single bit X_i-1 of the first operand, a single bit Y_i-1 of the second operand and a carry bit C_i-2. Full adder 465 produces a sum bit Z_i-1. A full adder 466 receives a single bit Xi of the first operand, a single bit Yi of the second operand and a carry bit C_i-1. Full adder 466 produces a sum bit Z_i. A full adder 469 receives a single bit X_j-1 of the first operand, a single bit Y_j-1 of the second operand and a carry bit C_j-2. Full adder 469 produces a sum bit Z_j-1.

In the embodiment of the adder shown in Figure 10, "j" is the size of the data path and the bit length of full word operations. Also, "i" is equal to "j" divided by 2. For example, "j" is equal to 32 and "i" is equal to 16. Alternately, when j is equal to 32, i may be equal to any integer less than 32.

When performing operations using "j"-bit full word operands, an enable bit 452 is equal to logic one and allows all carries to propagate. When performing two parallel operations using "i"-bit sub-word operands partitioned between bits i and i+1, enable bit 452 is equal to logic zero and prevents the carry propagating across the partition boundary. Instead the value on line 451 is used as the value forwarded to full adder 466. When an "add" is being performed, a logic 0 is placed on input line 451. When a "subtract" is being performed, a logic 1 is placed on input line 451.

Operation of carry look-ahead adders are well understood in the art. For example, suppose A[i] is one bit of an input, B[i] is one bit of the other input and S[i] is one bit of the sum from the adder. Then, the sum from one bit of the adder is given by Equation 1 below: <maths id="math0001"><math display="block"><mrow><mtext>Equation 1</mtext><mspace linebreak="newline" /><mtext>S [i] = A [i] XOR B[i] XOR C[i-1]</mtext></mrow></math><img file="EP0924601B1_D0001.tif" /></maths>In equation 1, C[i-1] is the carry out of the previous bits of the carry look-ahead adder. The carry look-ahead adder works on generating these carry bits quickly.

Let G[i] be a signal which signifies that a carry is to be generated by this bit and P[i] be a signal that a carry may propagate from the previous bits to the output of this bit. These are determined in accordance with Equation 2 below: <maths id="math0002"><math display="block"><mrow><mtext>Equation 2</mtext><mspace linebreak="newline" /><mtext>G[i] = A[i] AND B[i];</mtext><mspace linebreak="newline" /><mtext>P[i] = A[i] OR B[i];</mtext></mrow></math><img file="EP0924601B1_D0002.tif" /></maths>

Therefore, for four bits within a carry look-ahead adder, the carry bits may be generated as in Equation 3 below: <maths id="math0003"><math display="block"><mrow><mtext>Equation 3</mtext><mspace linebreak="newline" /><mtext>C[i] = G[i] + P[i] * (G[i-1] + P[i-1] *(G[i-2] + P[i-2] * (G[i-3] + P[i-3] * C[i-4])))</mtext><mspace linebreak="newline" /><mtext>C[i-1] = G[i-1] + P[i-1] * (G[I-2] + P[i-2] * (G[i-3] + P[i-3] * C[i-4]))</mtext><mspace linebreak="newline" /><mtext>C[i-2] = G[i-2] + P[i-2] * (G[i-3] + P[i-3] * C[i-4J)</mtext><mspace linebreak="newline" /><mtext>C[i-3] = G[i-3] + P[i-3] * C[i-4]</mtext></mrow></math><img file="EP0924601B1_D0003.tif" /></maths>

In equation 3 above, "*" is equivalent to a logic AND operation and "+" is equivalent to a logic OR operation.

When implementing a preferred embodiment of the present invention, a carry is stopped at a particular bit if the Generate G[i] and Propagate P[i] are forced to be false. For instance, in equation 3 above, if G[i-3] and P[i-3] are false, C[i-3] will be false and C[i-4] can never effect the value of C[i-2], C[i-1], and C[i]. Likewise, if G[i-2] and P[i-2] are false, C[i-2] will be false and G[i-3] and P[i-3] and C[i-4] can never effect the value of C[i-1] and C[i].

If we let M [i] be a mask bit that breaks the carry-chain between bit [i] and bit [i+1] when M [i] is 1, then a new Equation 4 can be generated as follows: <maths id="math0004"><math display="block"><mrow><mtext>Equation 4</mtext><mspace linebreak="newline" /><mtext>Gm[i] = !M [i] * (A[i] * B[i])</mtext><mspace linebreak="newline" /><mtext>Pm[i] = !M [i] * (A[i] + B[i])</mtext></mrow></math><img file="EP0924601B1_D0004.tif" /></maths>

Now if M [i] is 1, a carry will not be allowed to be generated from bit [i] or to propagate through bit [i].

For subtraction by creating the one's complement of one of the operands and adding it to the other operand with a carry in (two's complement arithmetic), a carry must be forced to be generated in a bit when M[i] is 1.

Let F be a signal that when true will force a carry to be generated in a bit when M [i] is 1. The equation for Gs[i] and Ps[i] becomes as set out in Equation 5 below: <maths id="math0005"><math display="block"><mrow><mtext>Equation 5</mtext><mspace linebreak="newline" /><mtext>Gs[i] = (M [i] * F) + (!M [i] * (A[i] * B[i]))</mtext><mspace linebreak="newline" /><mtext>= (M[i]*F) + (!M[i] * G[i])</mtext><mspace linebreak="newline" /><mtext>Ps[i] = Pm[i]</mtext></mrow></math><img file="EP0924601B1_D0005.tif" /></maths>Now if M[i] is 1, the value of Gs[i] is determined by F. If M [i] is 0 the value of Gs[i] is determined by A[i] and B[i] as it was previously. The propagate does not have to be forced by the F signal.

The equation of the carry out is given by Equation 6 below:         Equation 6 C[i] = Gs [i] + Ps [i] * C[i-1]

As will be understood by persons of skill in the art, principles of the present invention are not confined to arithmetic operations within computer system ALUs. For example, partitioning as shown in the ALU may also be extended to other entities within the computer system which operate on data. For example, Figure 6 shows the present invention embodied in pre-shifter 27. The same embodiment of the present invention may also be used to implement shifter 29. Partitioning of pre-shifter 27 and shifter 29 allows, for example, for the implementation of parallel shift-and-add operations and parallel shift operations

Pre-shifter 27 is shown to include a shift register one-bit slice 160, a shift register one-bit slice 161, a shift register one-bit slice 165, a shift register one-bit slice 166 and a shift register one-bit slice 169.

When data is shifted to the left, a datum on input 171, typically a logic 0 value, is used as input to shift register one-bit slice 160. When data is shifted to the right, a selector 175 in response to a control input 182 selects either a datum on input 181 (a logic 0 value or a logic 1 value) or selects the value currently stored by shift register one-bit slice 169 to be input to shift register one-bit slice 169.

Wherever the shifter is to be partitioned, additional selectors are added to the shifter. For example, Figure 6 shows the shifter partitioned between shift register one-bit slice 165 and shift register one-bit slice 166. There a selector 174 and a selector 173 have been added. For shift operations on partitioned operands, when data is shifted to the left, selector 173, in response to a control input 185, selects a datum on input 172, typically a logic 0 value, to be used as input to shift register one-bit slice 166. For shift operations on full word operands, when data is shifted to the left, selector 173 selects output from shift register one-bit slice 165 to be used as input to shift register one-bit slice 166.

For shift operations on partitioned operands, when data is shifted to the right, selector 174 in response to a control input 184 selects either a datum on input 183 (a logic 0 value or a logic 1 value) or selects the value currently stored by shift register one-bit slice 166 to be input to shift register one-bit slice 165. For shift operations on full word operands, when data is shifted to the right, selector 174 selects output from shift register one-bit slice 166 to be used as input to shift register one-bit slice 165.

Figure 6 shows a shifter with only two partitions. As will be understood from the foregoing discussion of partitions in an ALU, the shifter can be partitioned in a variety of ways. For example, a 64-bit shifter may be partitioned into two, four, eight, sixteen, thirty-two or sixty-four bit equal size partitions. Additionally, it is not a requirement of the present invention that partitions each operate on equal number of bits.

While the above embodiment describes the pre-shifter 27 and shifter 29 implemented as a shift register consisting of a series of one bit slices, alternative preferred embodiments are pre-shifters and shifters implemented with multiplexors. Typically, pre-shifter 27 is implemented by a one level of multiplexors, since it can usually shift by at most a small number of bits, for example, 0, 1, 2, 3 or 4 bits. Shifter 29 is typically implemented by three levels of multiplexors, where each level of multiplexor is a four-to-one multiplexor. For example, in a 64-bit shifter 29, the first level of multiplexors will shift either 0, 16, 32 or 48 bits. The second level of multiplexors can shift either 0, 4, 8 or 12 bits. The third level of multiplexors can shift 0, 1, 2 or 3 bits. This gives a shift of any number of bits from 0 to 63. In such a shifter built up of 3 stages of multiplexors, one-bit slices can still be identified. However the blocking of the shifts between any two bits may need to be done in one or more of the three multiplexer stages, as will be understood by those of ordinary skill in the art.

Principles of the present invention may also be extended to other elements in a computer system. For example, a multiplier may be implemented in accordance with a preferred embodiment of the present invention to allow for partial word parallel multiplications in addition to whole word multiplications.

For example, Figure 7 shows a four-bit multiplier in accordance with the prior art. The multiplier multiplies a four-bit first multiplicand X₃X₂X₁X₀ (base 2) with a four-bit second multiplicand Y₃Y₂Y₁Y₀ (base 2) to produce an eight-bit result Z₇Z₆Z₅Z₄Z₃Z₂Z₁Z₀ (base 2). As is understood by those skilled in the art, logic AND gates 201, 202, 203, 204, 205, 206, 207, 208, 209, 210, 211, 212, 213, 214, 215 and 216 may be used to generate partial products for the multiplication. A partial product sum circuit 220 sums the partial products generated by logic AND gates 201 through 216 to produce the result.

The two multiplicands, X₃X₂X₁X₀ and Y₃Y₂Y₁Y₀, the partial products generated by logic AND gates 201 through 216, and the result produced by partial product sum circuit 220 may be placed in a table in such a way as to summarize operation of the multiplier. For example, such a table is shown as Table 1 below: <tables id="tabl0001" num="0001"><img file="EP0924601B1_D0006.tif" /></tables>

In the notation used in Table 1 above, the bit position of each bit of both multiplicands and the result is specifically identified. Additionally, the bits of the multiplicand which are used to form each partial product are specifically set out. As is understood by those skilled in the art, the information shown in Table 1 above may be set out using abbreviated or simplified notation, as in Table 2 below: <tables id="tabl0002" num="0002"><img file="EP0924601B1_D0007.tif" /></tables>

In Table 2 above, each bit of the first multiplicand is represented by an "X", each bit of the second multiplicand is represented by a "T", each bit of a partial product is represented by a "z", and each bit of the result is represented by a "Z". Using the simpler notation of Table 2, an eight-bit multiplier may be described as shown in Table 3 below: <tables id="tabl0003" num="0003"><img file="EP0924601B1_D0008.tif" /></tables>

The multiplier shown in Table 3 multiplies an eight-bit first multiplicand XXXXXXXX_{(base 2)} with an eight-bit second multiplicand YYYYYYYY_{(base 2)} to produce an sixteen-bit result ZZZZZZZZZZZZZZZZ_{(base 2)}.

Similarly, using the simpler notation of Table 2 and Table 3 (but eliminating spaces between bits positions) a sixteen-bit multiplier may be described as shown in Table 4 belows: <tables id="tabl0004" num="0004"><img file="EP0924601B1_D0009.tif" /></tables>

The multiplier shown in Table 4 multiplies a sixteen bit first multiplicand XXXXXXXXXXXXXXXX_{(base 2)} with a sixteen-bit second multiplicand YYYYYYYYYYYYYYYY_{(base 2)} to produce a thirty-two-bit result ZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZ_{(base 2).}

In accordance with preferred embodiments of the present invention, a standard multiplier may be modified to implement a multiplier which provides parallel multiplication of partial words in addition to multiplication of whole words. For example, Figure 8 shows a four-bit multiplier in accordance with the preferred embodiment of the present invention. Logic AND gates 301,302,303,304,305,306,307,308,309,310,311, 312, 313, 314, 315 and 316 generate partial products for the multiplication. A partial product sum circuit 320 sums the partial products generated by logic AND gates 301 through 316 to produce the result.

In the multiplier shown in Figure 8, partial product sum circuit 320 may be implemented exactly the same as partial product sum circuit 220 shown in Figure 7. The difference between the multiplier shown in Figure 8 and the multiplier shown in Figure 7, is the addition of a control line 321, which is connected to an additional input included in each of logic AND gates 303, 304, 307, 308, 309, 310, 313 and 314.

As shown in Figure 8, when control line 321 is set at logic 1, the multiplier performs a whole word multiplication on a four-bit first multiplicand X₃X₂X₁X₀ (base 2) and a four-bit second multiplicand Y₃Y₂Y₁Y₀ (base 2) to produce an eight-bit result Z₇Z₆Z₅Z₄Z₃Z₂Z₁Z₀ (base 2). The two multiplicands, X₃X₂X₁X₀ and Y₃Y₂Y₁Y₀, the partial products generated by logic AND gates 301 through 316, and the result produced by partial product sum circuit 320 may be represented in table form as shown in Table 5 below: <tables id="tabl0005" num="0005"><img file="EP0924601B1_D0010.tif" /></tables>

A comparison of Table 5 with Table 1 above, confirms that when line 321 is set to logic 1, operation of the multiplier shown in Figure 8 is identical to operation of the multiplier shown in Figure 7. Therefore, similar to Table 2 above, the simplified notation may be used to describe operation of the multiplier shown in Figure 8 as in Table 6 below: <tables id="tabl0006" num="0006"><img file="EP0924601B1_D0011.tif" /></tables>

Figure 9 shows the multiplier shown in Figure 8, except however that control line 321 is set at logic 0. This forces half the partial products to zero allowing the multiplier to perform parallel multiplication of partial (two-bit) words. That is, in a first multiplication, a two-bit multiplicand A₁A₀ (base 2) is multiplied by a two-bit multiplicand C₁C₀ (base 2) to produce a four-bit result E₃E₂E₁E₀ (base 2). In a second multiplication, a two-bit multiplicand B₁B₀ (base 2) is multiplied by a two-bit multiplicand D₁D₀ (base 2) to produce a four-bit result F₃F₂F₁F₀ (base 2). The partial products not used for the parallel multiplications are forced to logic zero. The parallel multiplication may be represented in table form as shown in Table 7 below: <tables id="tabl0007" num="0007"><img file="EP0924601B1_D0012.tif" /></tables>

Using the simplified notation first introduced in Table 2, the multiplier shown in Figure 9 may be represented as in Table 8 below: <tables id="tabl0008" num="0008"><img file="EP0924601B1_D0013.tif" /></tables>

As illustrated by Table 7 and Table 8, parallel multiplication of partial words is implemented in a multiplier by forcing selected partial products in the multiplier to zero. In general, a standard multiplier of any size may be utilized to perform parallel multiplication by forcing unused partial products to zero. The partial products are forced to logic 0, for example, by using one or more control inputs and three input logic AND gates (or their equivalents).

For example, as discussed above, an eight-bit multiplier may be implemented as described by Table 3. This same multiplier may be utilized to perform parallel multiplication of partial word multiplicands by providing circuitry, such as shown in Figure 8 and Figure 9, to force partial products to zero in accordance with the teaching of the present invention. No modification is necessary to the partial product sum circuitry. Thus, modifying the multiplier described by Table 3 in accordance with the teaching of the present invention allows, for example, the performance of two parallel multiplications using four-bit multiplicands as implemented by Table 9 below: <tables id="tabl0009" num="0009"><img file="EP0924601B1_D0014.tif" /></tables>

As can be seen from Table 9 above, in a first parallel multiplication of partial word multiplicands, a four-bit multiplicand AAAA_{(base 2)} is multiplied by a four-bit multiplicand CCCC_{(base 2)} to produce an eight-bit result EEEEEEEE_{(base 2)}. In a second parallel multiplication of partial word multiplicands, a four-bit multiplicand BBBB_{(base 2)} is multiplied by a four-bit multiplicand DDDD_{(base 2)} to produce an eight-bit result FFFFFFFF_{(base 2)}. Multiplication of two whole word (eight-bit) multiplicands is implemented by the multiplier by not forcing any of the partial products to zero.

Likewise, as discussed above, a sixteen-bit multiplier may be implemented as shown by Table 4. This same multiplier may be utilized to perform parallel multiplications of partial word multiplicands by providing circuitry, such as shown in Figure 8 and Figure 9, to force partial products to zero in accordance with the teaching of the present invention. No modification is necessary to the partial product sum circuitry. Thus, modifying the multiplier described by Table 4 in accordance with the teaching of the present invention allows, for example, the performance of two parallel multiplications using eight-bit (partial word) multiplicands as implemented by Table 10 below: <tables id="tabl0010" num="0010"><img file="EP0924601B1_D0015.tif" /></tables>

As can be seen from Table 10 above, in a first parallel multiplication; an eight-bit multiplicand AAAAAAAA_{(base 2)} is multiplied by an eight-bit multiplicand CCCCCCCC_{(base 2)} to produce a sixteen-bit result EEEEEEEEEEEEEEEE_{(base 2)}. In a second parallel multiplication, an eight-bit multiplicand BBBBBBBB_{(base 2)} is multiplied by an eight-bit multiplicand DDDDDDDD_{(base 2)} to produce a sixteen-bit result FFFFFFFFFFFFFFFF_{(base 2)}. Multiplication of two whole word (sixteen-bit) multiplicands is implemented by the multiplier by not forcing any of the partial products to zero.

While the above description has showed parallel multiplication of half words, it will be understood by persons of ordinary skill in the art that both the number of parallel multiplications performed and the size of the partial word may be varied by selecting the appropriate partial products to force to zero.

For example, the sixteen-bit multiplier implemented as described by Table 4 (and/or Table 10) may be utilized to perform three simultaneous parallel multiplications by providing circuitry, such as shown in Figure 8 and Figure 9, to force partial products to zero in accordance with the teaching of the present invention. Thus, modifying the multiplier described by Table 4 in accordance with the teaching of the present invention allows, for example, the performance of one parallel multiplication using eight-bit multiplicands and two parallel multiplication using four-bit multiplicands as implemented by Table 11 below: <tables id="tabl0011" num="0011"><img file="EP0924601B1_D0016.tif" /></tables>

As can be seen from Table 11 above, in a first parallel multiplication, an eight-bit multiplicand AAAAAAAA_{(base 2)} is multiplied by an eight-bit multiplicand DDDDDDDD_{(base 2)} to produce a sixteen-bit result GGGGGGGGGGGGGGGG_{(base 2).} In a second parallel multiplication, a four-bit multiplicand BBBB_{(base 2)} is multiplied by a four-bit multiplicand EEEE_{(base 2)} to produce an eight-bit result H HH H H H H H_{(base 2)}. In a third parallel multiplication, a four-bit multiplicand CCCC_{(base 2)} is multiplied by a four-bit multiplicand FFFF_{(base 2)} to produce an eight-bit result IIIIIIII_{(base 2)}. As will be understood by those skilled in the art, for every partial product shown in Table 11 with a value of zero, it is necessary to have a three input logic AND gate or its logic equivalent to allow the partial product to be forced to zero when performing parallel multiplication operations. However, when a mixture of different sized partitions are done, as in Table 11, in some implementations different control inputs may be needed to force different partial product terms to zero, as will be understood by those of ordinary skill in the art.

As may be understood from the above-discussion, by selectively forcing partial products of a multiplier to zero, parallel multiplication of partial words may be fully implemented in a multiplier. The size of the word, the number of parallel multiplications simultaneously performed and the size of the partial words may be freely varied in accordance with the teaching of the present invention.

Figure 11 shows an example of instructions that can be executed in accordance with the preferred embodiment of the present invention. For example, instruction 500 includes a field 501, a subfield 502 of field 501, a field 503, a field 504 and a field 505. Field 501 sets out the op code. Field 501 sets out, for example, an add, a shift and add, a subtract, a shift and subtract, a shift left, a shift right, a multiply, or any number of other operations. Subfield 502 of 501 indicates whether the operation is to be performed as parallel operations, and if so, what is the size of the operands. Field 503 indicates a first source register. Field 504 indicates a second source register. Field 505 indicates a destination register.

As will be understood in the art, instruction 500 illustrates one of many possible ways an instruction can be organized. For example, instruction 510 shows an alternate embodiment in that the parallel operation indication is in a separate field. Specifically, instruction 510 includes a field 511, a field 512 , a field 513, a field 514 and a field 515. Field 511 sets out the op code. Field 511 sets out, for example, an add, a shift and add, a subtract, a shift and subtract, a shift left, a shift right, a multiply, or any number of other operations. Field 512 indicates whether the operation is to be performed as parallel operations, and if so, what is the size of the operands. Field 513 indicates a first source register. Field 514 indicates a second source register. Field 515 indicates a destination register.

As will be understood in the art, the present invention also works for other multipliers where partial products are generated. For example, the present invention also may be utilized in a Booth-encoded multiplier. In a Booth-encoded multiplier, fewer rows of partial product terms are generated by considering more than one bit of the multiplier (y-multiplicand) for each row of the partial product term. See for example, John Hennessy &amp; David Patterson, <u>Computer Architecture. A Quantitative Approach</u>, Morgan Kaufmann, 1990, Appendix, pp. A-39 through A-49. As, in the case of the multiplier above, the value of some partial product terms generated by the Booth-encoded multiplier are changed to take into account the parallel processing, as will be understood by those skilled in the art.

More specifically, for a Booth-encoded multiplier, the AND gates 301 through 316 shown in Figures 8 and 9 are replaced by multiplexors. For example, a Booth-encoded multiplier using the "overlapping triplets" method examines three bits of the multiplier (i.e., y multiplicand) each time, instead of one bit each time, to generate a row of partial products that is one of +x, +2x, -2x, -x or zero, instead of a row of partial products which is always +x or 0 as in the multiplier shown in Figures 8 and 9. This may be implemented as a five-to-one multiplexor. The name "overlapping triplets" is due to the fact that this method looks at three bits of the multiplier (y multiplicand) and retires two bits of the multiplier (y multiplicand) for each row. The overlapping occurs when, for the next row, the least significant bit of the three multiplier (y multiplicand) bits used by this next row was the most significant bit of the three multiplier bits used from the previous row.

To implement parallel sub-word multiplication, the bits of the x multiplicand that do not correspond to the sub-word product whose partial product rows are being formed are set to zero. This can be implemented with multiplexors like in the unmodified Booth-encoded multiplier, modifying the control signals to the multiplexors. The sign of the partial product row may also be used as an additional input to the multiplexors.