Granted Patent B2

US 12,737,186 · App. 18/770,560 · Granted Sep 15, 2026

Inventors: Allan Kennedy Porterfield (Durham, NC); Skyler Arron Windh (McKinney, TX); Bashar Romanous (Allen, TX)

Assignee: Micron Technology, Inc.

G06F9/30181G06F9/30043

Abstract

Claims (68)

1 . A device, comprising:

a memory; and

at least one processor configured via instructions stored in the memory to:

receive a program specifying a plurality of memory variables, a plurality of data flows through the memory variables, and first instructions to transform data along the data flows;

receive a profile specifying a coarse grained reconfigurable array having a plurality of tiles operable in parallel, wherein the coarse grained reconfigurable array having latency constraints in propagating data through elements in the tiles;

map the data flows to operations of data through the elements in the tiles;

generate a memory map identifying, for each respective memory variable among the plurality of memory variables, one of the tiles that contains a memory location represented by the respective memory variable;

wherein the data flows are mapped to the operations based on the memory map.

2 . The device of claim 1 , wherein the at least one processor is further configured to assign, based on the memory map, the first instructions to the tiles.

3 . The device of claim 2 , wherein the schedule is configured to identify timing of execution of the first instructions in the tiles.

4 . The device of claim 3 , wherein the at least one processor is further configured to partition the memory variables and the first instructions into a plurality of groups, each of the groups configured to be implemented on one of the tiles.

5 . The device of claim 4 , wherein the at least one processor is further configured to partition the memory variables and the first instructions to balance a number of instructions implemented per tile, to balance a number of memory variables implemented per tile, or to balance an amount of memory usage implemented per tile.

6 . The device of claim 4 , wherein the at least one processor is further configured to assign each respective instruction among the first instructions to a tile based on the tile containing memory variables having data to be operated upon by the respective instruction.

7 . The device of claim 6 , wherein to determine the timing of execution of the first instructions the at least one processor is configured to:

select a current instruction for scheduling;

identify a slot in a first tile containing memory variables used by the current instruction;

search for a clock cycle for execution of the current instruction in the slot; and

select a next instruction for scheduling, in response to a determination that a valid spoke RAM slot is found for execution of the current instruction;

wherein the at least one processor is configured to, if no valid spoke RAM slot is found for execution of the current instruction in the slot:

search for an available slot in the first tile; and

search, in response to the available slot being found, for a clock cycle for execution of the current instruction in the available slot; and

wherein the at least one processor is configured to, if no available slot is found:

determine a prior instruction scheduled before the current instruction;

identify a first schedule previously determined for the prior instruction as invalid; and

start determination of a second valid schedule for the prior instruction.

8 . A method, comprising:

receiving a program specifying a plurality of memory variables, a plurality of data flows through the memory variables, and first instructions to transform data along the data flows;

receiving a profile specifying a coarse grained reconfigurable array having a plurality of tiles operable in parallel;

mapping the data flows to operations of data through the elements in the tiles;

determining a schedule to execute the first instructions in the tiles;

receiving the profile specifying a coarse grained reconfigurable array having the plurality of tiles operable in parallel, wherein the coarse grained reconfigurable array includes latency constraints in propagating data through elements in the tiles;

mapping the data flows to operations of propagating data through the elements in the tiles;

determining the schedule to execute the first instructions in the tiles without violating the latency constraints;

generating a memory map identifying, for each respective memory variable among the plurality of memory variables, one of the tiles that contains a memory location represented by the respective memory variable; and

wherein the data flows are mapped to the operations based on the memory map.

9 . The method of claim 8 , further comprising:

assigning, based on the memory map, the first instructions to the tiles.

10 . The method of claim 9 , wherein the schedule is configured to identify timing of execution of the first instructions in the tiles.

11 . The method of claim 10 , further comprising:

partitioning the memory variables and the first instructions into a plurality of groups to implement each of the groups on one of the tiles.

12 . The method of claim 11 , wherein the partitioning of the memory variables and the first instructions is based on balancing a number of instructions implemented per tile, balancing a number of memory variables implemented per tile, or balancing an amount of memory usage implemented per tile.

13 . The method of claim 11 , wherein the assigning includes assigning each respective instruction among the first instructions to a tile based on the tile containing memory variables having data to be operated upon by the respective instruction.

14 . The method of claim 13 , wherein the timing of execution of the first instructions is determined via:

selecting a current instruction for scheduling;

identifying a slot in a first tile containing memory variables used by the current instruction;

searching for a clock cycle for execution of the current instruction in the slot; and

selecting a next instruction for scheduling, in response to a determination that a valid spoke RAM slot is found for execution of the current instruction;

wherein if no valid spoke RAM slot is found for execution of the current instruction in the slot, the timing of execution of the first instructions is further determined via:

searching for an available slot in the first tile; and

searching, in response to the available slot being found, for a clock cycle for execution of the current instruction in the available slot; and

wherein if no available slot is found, the timing of execution of the first instructions is further determined via:

determining a prior instruction scheduled before the current instruction;

identifying a first schedule previously determined for the prior instruction as invalid; and

starting determination of a second valid schedule for the prior instruction.

15 . A non-transitory computer storage medium storing instructions which, when executed by a computing device, cause the computing device to perform a method, comprising:

receiving a program specifying a plurality of memory variables, a plurality of data flows through the memory variables, and first instructions to transform data along the data flows;

receiving a profile specifying a coarse grained reconfigurable array having a plurality of tiles operable in parallel;

mapping the data flows to operations of data through the elements in the tiles;

determining a schedule to execute the first instructions in the tiles without violating the latency constraints;

receiving the profile specifying a coarse grained reconfigurable array having the plurality of tiles operable in parallel, wherein the coarse grained reconfigurable array includes latency constraints in propagating data through elements in the tiles;

mapping the data flows to operations of propagating data through the elements in the tiles;

determining the schedule to execute the first instructions in the tiles without violating the latency constraints, wherein the schedule is configured to identify timing of execution of the first instructions in the tiles;

generating a memory map identifying, for each respective memory variable among the plurality of memory variables, one of the tiles that contains a memory location represented by the respective memory variable;

wherein the data flows are mapped to the operations based on the memory map.

16 . The non-transitory computer storage medium of claim 15 , wherein the method further comprises:

assigning, based on the memory map, the first instructions to the tiles.

17 . The non-transitory computer storage medium of claim 16 , wherein the method further comprises:

partitioning the memory variables and the first instructions into a plurality of groups to implement each of the groups on one of the tiles.

Assignments (1)

Continuity (2)

References Cited (49)