United States Patent m
Casavant et al.
 COMPUTER-AIDED DESIGN METHOD FOR RESTRUCTURING COMPUTATIONAL NETWORKS TO MINIMIZE SHIMMING DELAYS
 Inventors: Albert E. Casavant; Richard I.
Hartley, both of Schenectady, N.Y.
 Assignee: General Electric Company, Schenectady, N.Y.
 Appl. No.: 428,808
 Filed: Oct. 30, 1989
 Int. CI.' G06F9/00
 U.S. CI 395/500; 364/DIG. 1;
364/271.5: 364/259; 364/259.1  Field of Search ... 364/200 MS File. 900 MS File.
 References Cited
U.S. PATENT DOCUMENTS
3.978.452 8/197o Barton el al 364/200
4.447.875 5/1984 Bolton el al 364/200
4.654.780 3/1987 Logsdon et al 364/200
4.734.848 3/1988 Yamano et al 364/200
4.827.428 5/1989 Dunlop et al 364/491
[ii] Patent Number: 5,175,843  Date of Patent: Dec. 29, 1992
4,942.396 1/1990 Hartley et al 341/101
4.956.773 9/1990 Saito et al 364/200
Primary Examiner—Lawrence E. Anderson
Assistant Examiner—Mehmet Geckil
Attorney, Agent, or Firm—Marvin Snyder
A computer-aided design method for restructuring computational networks to minimize latency and shim delay, suitable for use by a silicon compiler. Data-flow graphs for computational networks which use trees of operators, each performing associative and commutative combining of its respective imput operands to generate a respective output operand, are converted to data-flow graphs with multiple-input operators. Dataflow graphs with multiple-input operators, after being optimally scheduled, are converted to data-flow graphs which use trees of dual-input operators or of dual-input and three-input operators, those trees having minimum latency and shim delay associated with them. These data-flow graphs then have shim delay minimized in them, e.g. by being subjected to linear programming.
15 Claims, 13 Drawing Sheets
MODIFY DATA-FLOW GRAPH BY REPLACING EACH TREE OF PLURAL-INPUT OPERATORS OF A TYPE WITH A SINGLE MULTIPLE-INPUT OPERATOR OF THE SAME TYPE
DETERMINE EARLIEST SCHEDULE
FOR THE COMPUTATIONAL NETWORK
REPLACE EACH MULTIPLE-INPUT OPERATOR WITH A MINIMUM-DELAY TREE OF DUAL-INPUT OPERATORS OF THE SAME TYPE
END WITH CONSTRUCTION OF THE REVISED^ ^COMPUTATIONAL NETWORK AS DESCRIBED BYj THE FINAL DATA FLOW GRAPH