Systolic array processors for reducing under-utilization of original design parallel-bit processors with digit-serial processors by using maximum common divisor of latency around the loop connection

A method for designing a systolic array of processors including therewithin loop connections of processors as well as internal pipeline connections of processors, which method reduces any under-utilization of the processors, each of said loop connections being adapted to introduce a latency in its respective loop, said method comprising the steps of:

• initially developing an original design for said systolic array, said original design using parallel-bit processors;
• determining the latency around each loop connection of parallel-bit processors in said original design for said systolic array;
• determining a maximum common divisor P of the latencies around the loop connections of parallel-bit processors in said original design for said systolic array;
• if said maximum common divisor P is one, indicating that there are no under-utilized processors in said original design for said systolic array using parallel-bit processors, using said original design for said systolic array using parallel-bit processors as a final design for said systolic array;
• and
• if said maximum common divisor P is larger than one, indicating that there are under-utilized processors in said original design for said systolic array using parallel-bit processors, generating a final design for said systolic array by performing the following substeps, cyclically if necessary,
• (a) in a most recent said systolic array design using parallel-bit processors, replacing each parallel-bit processor with a corresponding digit-serial processor in which, for each digital word comprising a number of digits, the number of digits per digital word equals the maximum common divisor of the respective latencies of said loop connections in said most recent systolic array design using parallel-bit processors, thereby generating a most recent modified design for said systolic array;
• (b) using said most recent modified design for said systolic array as the final design for said systolic array only if the latencies in all the loop connections of said most recent modified design for said systolic array are the same as the latencies in the corresponding loop connections of said most recent systolic array design using parallel-bit processors;
• and
• (c) if the latencies in all the loop connections of said most recent design for said systolic array are not the same as the latencies in the corresponding loop connections of said most recent systolic array design using parallel-bit processors, inserting delay elements into the most recent systolic array design using parallel-bit processors for increasing the latencies in all the loop connections by a factor equal to (Q+1)/Q, where Q is the greatest common divisor of all loop latencies in the most recent design using parallel-bit processors, to generate an updated most recent systolic array design using parallel-bit processors.

A method for designing a digit-serial systolic array of functional cells, each of which functional cells has at least one respective input port, at least one respective output port, and at least one operator, which digit-serial systolic array has internal pipeline connections therein and has at least one loop connection therein, said at least one loop connection being adapted to introduce a latency in said loop, said method comprising the steps of:

• developing a prototypical structure for an individual functional cell of each type of functional cell used in said systolic array, which individual functional cell may be repeated in said systolic array;
• generating digit-serial systolic array designs by assigning a scheduled time to each operator in each functional cell and subsequently inserting delay elements to equalize latencies, and then using integer programming to determine a minimum number P of digits per digital word, said integer programming proceeding from the following constraints:
• for each connection between a preceding and succeeding operator in each functional cell
• S.sub.i +t.sub.i .ltoreq.S.sub.j +t.sub.j
• where t.sub.i is the latency of the preceding operator, t.sub.j is the latency of the succeeding functional cell, S.sub.i is a scheduled time of the preceding functional cell, and S.sub.j is the scheduled time of the succeeding functional cell;
• and
• for each loop connection within said digit-serial systolic array of functional cells having no interior loop connections therewithin and threading a number n of input and output ports of the functional cells of said array ##EQU17## where S.sub.i.sbsb.k and S.sub.j.sbsb.k are for each value of k the scheduled times of a different pair of interconnected functional-cell input and output ports threaded by that loop connection, and m is an integer;
• and
• selecting the digit-serial systolic array design in which P is a minimum.

A method for designing a digit-serial systolic array of functional cells, each of which functional cells has at least one respective input port, at least one respective output port, and at least one operator, which digit-serial systolic array has internal pipeline connections therein and has at least one loop connection therein, said at least one loop connection being adapted to introduce a latency in said loop, said method comprising the steps of:

• developing one or more prototypical structures of an individual functional cell for each type of functional cell used in said systolic array, which type of functional cell may be repeated in said systolic array, the prototypical structures for each type of functional cell where there is more than one such prototype structure differing in their respective minimal latencies;
• for each combination prototypical structures as may be utilized in each loop connection within said systolic array, generating digit-serial systolic array designs by assigning a scheduled time to each operator in each functional cell and subsequently inserting delay elements to equalize latencies, and then using integer programming to determine a minimum number P of digits per digital word, said integer programming proceeding from the following constraints:
• for each connection between a preceding and a succeeding operator in each functional cell
• S.sub.i +t.sub.i .ltoreq.S.sub.j +t.sub.j
• where t.sub.i is the latency of the preceding operator, t.sub.j is the latency of the succeeding functional cell, S.sub.i is a scheduled time of the preceding functional cell, and S.sub.j is the scheduled time of the succeeding functional cell;
• and
• for each loop connection within said digit-serial systolic array of functional cells having no interior loop connections therewithin and threading a number n of input and output ports of the functional cells of said array ##EQU18## where S.sub.i.sbsb.k and S.sub.j.sbsb.k are for each value of k the scheduled times of a different pair of interconnected functional-cell input and output ports threaded by that loop connection, and m is an integer;
• and
• selecting the digit-serial systolic array design in which P is a minimum.

A loop connection comprising first, second and third digit-serial processing elements cyclically operable in accordance with cycles of clock signals supplied to each of said digit-serial processing elements, each of said first, second and third digit-serial processing elements having at least a first input port and at least a first output port, said first output port of said first digit-serial processing element coupled to the first input port of said second digit-serial processing element, said first output port of said second digit-serial processing element coupled to the first input port of the third digit-serial processing element, said first input port of said first digit-serial processing element connected for receiving a respective digit-serial input signal from an additional digit-serial processing element not connected in said loop connection, at least one of said second and third digit-serial processing elements having a second output port for supplying a respective digit-serial output signal to be used outside said loop connection, and a delay element of at least one-cycle of clock signal delay coupled between a second input port of the first digit-serial processing element and an additional output port of the third digit-serial processing element so as to complete said loop connection, said loop connection being adapted to provide a latency of at least two cycles of clock signal therein and said digit-serial signals having as many digits as there are cycles of clock signal latency in said loop connection.

A systolic array of digit-serial processing elements cyclically operable in accordance with cycles of clock signals supplied to each of said digital-serial processing elements and adapted to process signals made up of a number of digits, said systolic array having a plurality of loop connections therein adapted to introduce respective latencies, which latencies have at least one common divisor larger than one, each of said loop connections being comprised of a first, second and third processing element wherein a first output port of said first processing element is coupled to a first input port of said second processing element and a first output port of said second processing element is coupled to a first input port of said third processing element, said first digit-serial processing element in said each loop connection having a first input port connected for receiving a respective digit-serial input signal from a digit-serial processing element not connected in said each loop connection, said each loop connection further including a pipeline connection between a second output port of said third digit-serial processing element and a second input port of said first digit-serial processing element so as to complete said each loop connection and introduce one-cycle of clock signal delay in said each loop connection, and at least one of said second and third digit-serial processing elements having an additional output port for supplying a respective digit-serial output signal to be used outside said each loop connection or pipeline connection, the number of digits in signals for processing in said systolic array of digit-serial processing elements equaling a maximum of such common divisors in order to fully utilize the processing elements of said array.