Method and apparatus for parallel loads equalizing utilizing instruction sorting by columns based on predicted instruction execution time

Claim 2

A parallel computer, comprising: input means for accepting an input group of instructions having rows and columns; a group of n processing elements, including means for simultaneously processing a group of instructions in the form of a matrix of m rows and n columns, in units of columns; data feed means for fetching simultaneously a plurality of columns of instructions from said input means; sort means of sorting, including a first sort means for sorting a first column of instructions in descending order according to instruction execution predict count information associated with a given instruction, and a second sort means for simultaneously sorting a second column of instructions in ascending order; and means for feeding said first column and said second column consecutively to said processing elements.

Claim 3

A parallel computer according to claim 1, wherein said data feed means includes means for outputting a block of columns of instructions simultaneously; and wherein said sort means further includes a plurality of said first sort means, each first sort means operating independently; a plurality of said second sort means, each second sort means operating independently; and means for switching successive blocks of columns of instructions alternatively to a first sort means and to a second sort means.

Claim 4

A parallel computer according to claim 2, further comprising an instruction queue circuit that is connected between said sort means and said group of n processing elements.

Claim 5

A partial computer according to claim 3, wherein said instruction execution predict count information is data, in a data field of a given instruction, indicative of a number of clock cycles associated with execution of the given instruction.

Claim 6

A parallel computer according to claim 2, wherein said instruction execution predict count information is data, in a data field of a given instruction, indicative of a number of clock cycles associated with execution of the given instruction.

Claim 7

A parallel computer according to claim 2, wherein said sort means further includes a sufficient quantity of said first sort means and said second sort means such that a sort time associated with a block of instructions does not exceed a mean execution time of a block of instructions.

Claim 8

A parallel computer according to claim 1, wherein said instruction execution predict count information is data, in a data field of a given instruction, indicative of a number of clock cycles associated with execution of the given instruction.

Claim 9

A parallel computer comprising: fetch means for fetching a plurality of columns of instructions simultaneously from an input group of instructions in the form of a matrix of m rows and n columns; sort means for sorting, including first sort means for sorting a first column of instructions in descending order, and second sort means for sorting a second column of instructions in ascending order, said first sort means and said second sort means sorting simultaneously and independently according to instruction execution predict count information associated with a given instruction; switch means for alternately switching a first sort means and a second sort means to receive a plurality of columns of instructions from said fetch means; instruction queue means for receiving said first column and said second column from said sort means, and for outputting said first column and said second column consecutively; and processing elements having means for receiving said first column and said second column from said instruction queue means, and having means for processing instructions simultaneously in units of columns.

Claim 10

A parallel computer comprising: fetch means for fetching a plurality of columns of instructions simultaneously from an input group of instructions in the form of a matrix of m rows and n columns; sort means for sorting, including first sort means for sorting a first column of instructions in descending order, and second sort means for sorting a second column of instructions in ascending order, said first sort means and said second sort means sorting simultaneously and independently according to instruction execution predict count information associated with a given instruction; switch means for alternately switching a first sort means and a second sort means to receive a plurality of columns of instructions from said fetch means; instruction queue means for receiving said first column and said second column from said sort means, and for outputting said first column and said second column consecutively; and processing elements having means for receiving said first column and said second column from said instruction queue means, and having means for processing instructions simultaneously in units of columns.

Claim 11

A parallel computer according to claim 8, wherein said sort means further includes a sufficient quantity of said first sort means and said second sort means such that a sort time associated with a block of instructions does not exceed a mean execution time of a block of instruction.

Claim 12

A data feed method, for use in a parallel computer having n processing elements for simultaneously processing a group of instructions in the form of a matrix of m rows and n columns in units of columns, comprising the steps of: (a) accepting a plurality of columns of instructions; (b) dividing the plurality of columns of instructions into a plurality of blocks, each block including a plurality of first alternate columns and a plurality of second alternate columns; (c) sorting said first plurality of alternate columns simultaneously in descending order according to instruction execution predict count information associated with a given instruction to produce a first output; (d) sorting said second plurality of alternate columns simultaneously in ascending order according to instruction execution predict count information associated with a given instruction to produce a second output; (e) feeding said first output and said second output alternately as a third output in units of columns to said processing elements; and (f) executing steps (a) to (e) simultaneously.

Claim 13

A data feed method, for use in a parallel computer having n processing elements for simultaneously processing a group of instructions in the form of a matrix of m rows and n columns in units of columns, comprising the steps of: (a) accepting a plurality of columns of instructions; (b) dividing the plurality of columns of instructions into a plurality of blocks, each block including a plurality of first alternate columns and a plurality of second alternate columns; (c) sorting said first plurality of alternate columns simultaneously in descending order according to instruction execution predict count information associated with a given instruction to produce a first output; (d) sorting said second plurality of alternate columns simultaneously in ascending order according to instruction execution predict count information associated with a given instruction to produce a second output; (e) feeding said first output and said second output alternately as a third output in units of columns to said processing elements; and (f) executing steps (a) to (e) simultaneously.

Claim 14

A data feed method according to claim 10, wherein said instruction execution predict count information is data, in a data field of a given instruction, indicative of a number of clock cycles associated with execution of the given instruction.

Claim 15

A parallel computing method, for use in a parallel computer having n processing elements for simultaneously processing a group of instructions in the form of a matrix of m rows and n columns in units of columns, comprising the steps of: (a) accepting a plurality of columns of instructions; (b) dividing the plurality of columns of instructions into a plurality of blocks, each block including a plurality of first alternate columns and a plurality of second alternate columns; (c) sorting said first plurality of alternate columns simultaneously in descending order according to instruction execution predict count information associated with a given instruction to produce a first output; (d) sorting said second plurality of alternate columns simultaneously in ascending order according to instruction execution predict count information associated with a given instruction to produce a second output; (e) feeding said first output and said second output alternately as a third output in units of columns to said processing elements; (f) executing simultaneously a plurality of instructions from said third output in units of columns; and (g) executing steps (a) to (f) simultaneously.

Claim 16

A parallel computing method, for use in a parallel computer having n processing elements for simultaneously processing a group of instructions in the form of a matrix of m rows and n columns in units of columns, comprising the steps of: (a) accepting a plurality of columns of instructions; (b) dividing the plurality of columns of instructions into a plurality of blocks, each block including a plurality of first alternate columns and a plurality of second alternate columns; (c) sorting said first plurality of alternate columns simultaneously in descending order according to instruction execution predict count information associated with a given instruction to produce a first output; (d) sorting said second plurality of alternate columns simultaneously in ascending order according to instruction execution predict count information associated with a given instruction to produce a second output; (e) feeding said first output and said second output alternately as a third output in units of columns to said processing elements; (f) executing simultaneously a plurality of instructions from said third output in units of columns; and (g) executing steps (a) to (f) simultaneously.

Claim 17

A parallel computing method according to claim 12, wherein said instruction predict count information is data, in a data field of a given instruction, indicative of a number of clock cycles associated with execution of the given instruction.

Claim 18

A parallel computing method, for use in a parallel computer having n processing elements for simultaneously processing a group of instructions in the form of a matrix of m rows and n columns, each instruction containing an execution predict count information data field, in units of columns, comprising the steps of: (a) inserting in the data field of a given instruction a value representing the number of clock cycles required to execute the given instruction; (b) accepting a plurality of columns of instructions; (c) dividing the plurality of columns of instructions into a plurality of blocks, each block including a plurality of first alternate columns and a plurality of second alternate columns; (d) sorting said first plurality of alternate columns simultaneously in descending order according to instruction execution predict count information associated with a given instruction to produce a first output; (e) sorting said second plurality of alternate columns simultaneously in ascended order according to instruction execution predict count information associated with a given instruction to produce a second input; (f) feeding said first output and said second output alternately as a third output in units of columns to said processing elements; and (g) executing simultaneously a plurality of instructions from said third output in units of columns.

Claim 19

A parallel computing method, for use in a parallel computer having n processing elements for simultaneously processing a group of instructions in the form of a matrix of m rows and n columns, each instruction containing an execution predict count information data field, in units of columns, comprising the steps of: (a) inserting in the data field of a given instruction a value representing the number of clock cycles required to execute the given instruction; (b) accepting a plurality of columns of instructions; (c) dividing the plurality of columns of instructions into a plurality of blocks, each block including a plurality of first alternate columns and a plurality of second alternate columns; (d) sorting said first plurality of alternate columns simultaneously in descending order according to instruction execution predict count information associated with a given instruction to produce a first output; (e) sorting said second plurality of alternate columns simultaneously in ascended order according to instruction execution predict count information associated with a given instruction to produce a second input; (f) feeding said first output and said second output alternately as a third output in units of columns to said processing elements; and (g) executing simultaneously a plurality of instructions from said third output in units of columns.