US 12,242,854 B2
Compressing instructions for machine-learning accelerators
Kyong Ho Lee, Los Altos, CA (US); Miguel Angel Guerrero, Palo Alto, CA (US); and Varun Agarwal, Pleasanton, CA (US)
Assigned to Meta Platforms, Inc., Menlo Park, CA (US)
Filed by Meta Platforms, Inc., Menlo Park, CA (US)
Filed on Feb. 21, 2023, as Appl. No. 18/172,016.
Prior Publication US 2024/0281253 A1, Aug. 22, 2024
Int. Cl. G06F 9/30 (2018.01)
CPC G06F 9/30185 (2013.01) [G06F 9/30178 (2013.01)] 20 Claims
OG exemplary drawing
 
1. A computing system for accelerating machine-learning computations, comprising:
one or more control agents, and
an instruction master, which is communicably coupled to the one or more control agents over an instruction bus, that is configured to:
access an instruction that is to be executed by a control agent among the one or more control agents;
decompress the instruction by replacing each of one or more zero-symbol run-length fields in the instruction with as many continuous zero symbols as indicated by a corresponding value of the zero-symbol run-length field and removing one or more non-zero-symbol run-length fields from the instruction;
determine that the instruction is spatial-delta-encoded based on a compression data header associated with the instruction;
perform, in response to the determination, spatial-delta decoding on the instruction by orderly determining a spatial-delta-decoded value of each bit in the instruction, wherein a spatial-delta-decoded value of a bit is determined to be:
identical to a value of an immediately preceding bit when an encoded bit value is zero; or
opposite of the value of the immediately preceding bit when the encoded bit value is one; and
cause the instruction to be sent to the control agent.