US 12,135,990 B2
Modeling and compiling tensor processing applications for a computing platform using multi-layer adaptive data flow graphs
Chia-Jui Hsu, Santa Clara, CA (US); Mukund Sivaraman, Palo Alto, CA (US); and Vinod Kathail, Palo Alto, CA (US)
Assigned to XILINX, INC., San Jose, CA (US)
Filed by XILINX, INC., San Jose, CA (US)
Filed on Dec. 30, 2022, as Appl. No. 18/091,907.
Prior Publication US 2024/0220316 A1, Jul. 4, 2024
Int. Cl. G06F 9/48 (2006.01); G06F 8/41 (2018.01); G06F 9/54 (2006.01)
CPC G06F 9/4881 (2013.01) [G06F 8/451 (2013.01); G06F 9/544 (2013.01)] 17 Claims
OG exemplary drawing
 
1. A machine-implemented method, comprising:
receiving a multi-layer adaptive data flow (ML-ADF) graph specifying an application for execution on a data processing array (DPA) that comprises data processing elements (DPEs) and local, shared, and external memories;
folding the ML-ADF graph onto the DPA to provide an overlay graph, wherein the overlay graph represents the DPEs and the local and shared memories, and wherein resources of the DPA are temporally shared amongst multiple layers of the ML-ADF graph;
constructing DPE schedules for compute nodes of the ML-ADF graph and a data transfer schedule corresponding to shared-data and external-data nodes of the ML-ADF graph in order to coordinate runtime execution of the layers of the ML-ADF graph by the DPEs and sharing of data amongst the layers of the ML-ADF graph through the shared and external memories;
generating DPE code for the DPEs based on the respective DPE schedules; and
generating controller code for controllers and data movers of the local, shared and external memories based on the data transfer schedule.