US 12,307,217 B1
Dynamic adjustment of floating point exponent bias for exponent compression
Eric F. Dellinger, Longmont, CO (US); and Philip B. James-Roxby, Longmont, CO (US)
Assigned to XILINX, INC., San Jose, CA (US)
Filed by Xilinx, Inc., San Jose, CA (US)
Filed on May 6, 2021, as Appl. No. 17/313,224.
Int. Cl. G06F 7/483 (2006.01); G06F 7/499 (2006.01); H03M 7/24 (2006.01); H03M 7/30 (2006.01)
CPC G06F 7/483 (2013.01) [G06F 7/49942 (2013.01); H03M 7/24 (2013.01); H03M 7/30 (2013.01)] 20 Claims
OG exemplary drawing
 
1. A method comprising:
inputting a first set and a second set of floating point values to a computer-implemented application over a first time period and a subsequent second time period, respectively;
accumulating a distribution of values of exponents of the first set of floating point values by a statistics circuit;
compressing the exponents of the first set of floating point values by a compression circuit into a compressed exponent bit-width as a function of a compressed exponent bias;
processing the first set of floating point values having compressed exponents by the application;
adjusting the compressed exponent bit-width and the compressed exponent bias by a bias adjustment circuit based on the distribution of values of exponents of the first set of floating point values;
accumulating the distribution of values of exponents of the first set of floating point values with values of exponents of the second set of floating point values by the statistics circuit;
compressing, after the adjusting of the compressed exponent bit-width and the compressed exponent bias, the exponents of second set of floating point values into the compressed exponent bit-width by the compression circuit as a function of the compressed exponent bias; and
processing the second set of floating point values having compressed exponents by the application,
wherein processing floating point values having compressed exponents by the application comprises at least one of performing arithmetic operations by a processor and storing the floating point values having compressed exponents in a memory.