Granted Patent B1

US 12,474,903 · App. 18/386,178 · Granted Nov 18, 2025

Inventors: Ivan Garcia Alsina (Framingham, MA); Partha Biswas (Wayland, MA); Jianyi Li (Framingham, MA)

Assignee: The Math Works, Inc.

G06F8/41G06F8/10G06F8/31G06F8/315G06F8/35G06F8/36G06F8/42G06F8/443G06F8/447G06F8/451G06F8/51G06F8/52G06F8/75G06F9/3005G06F9/30174G06F9/38873G06F9/45558G06F9/4881

Abstract

Claims (53)

1 . A computer-implemented method comprising:

accessing a computer program written in a first programming language, the computer program including tensor-based operations that are implemented as programming idioms or vectorized statements of the first programming language;

receiving information on resources of a target hardware system;

receiving an identification of a scheduler of a second programming language different from the first programming language;

generating, by one or more processors, a first Intermediate Representation (IR) of the computer program;

partitioning the first IR of the computer program into a plurality of partitions wherein the plurality of partitions include the tensor-based operations, the partitioning including searching the first IR for the programming idioms or the vectorized statements;

generating, by the one or more processors, an IR for the plurality of partitions, the Intermediate Representation for the plurality of partitions compatible with the second programming language;

generating, by the scheduler of the second programming language, a schedule for the plurality of partitions, wherein the generating the schedule includes analyzing the IR for the plurality of partitions, further wherein the scheduler considers the information on the resources of the target hardware system;

generating, by the one or more processors, a second IR for the computer program, wherein the second IR for the computer program includes the schedule for the plurality of partitions as generated by the scheduler of the second programming language; and

generating code based on the second IR for the computer program, wherein the code implements the schedule for the plurality of partitions as generated by the scheduler of the second programming language.

2 . The computer-implemented method of claim 1 further comprising:

receiving identification of a scheduler of a third programming language; and

repeating

the generating the IR for the plurality of partitions,

the analyzing,

the generating the second IR for the computer program, and

the generating code,

with the scheduler of the third programming language.

3 . The computer-implemented method of claim 1 wherein the first programming language is a functional language or an imperative language, the second programming language is a Domain Specific Language (DSL), and the DSL is Halide, Tiramisu, Tensor Algebra Compiler (TACO), Seq, Tensor Comprehensions (TC), Fireiron, or TVM.

4 . The computer-implemented method of claim 1 wherein the first programming language is graphical programming language, Modelica, Octave, Julia, or Python.

5 . The computer-implemented method of claim 1 wherein the second IR for the computer program is a tensor directed acyclic graph (DAG).

6 . The computer-implemented method of claim 1 wherein the resources of the target hardware system include one or more of:

number of cores;

vectorization option;

instruction set extensions;

application-specific instruction set (ASIP);

single instruction multiple data (SIMD) register size;

memory hierarchy;

cache attributes; or

graphics processor unit (GPU).

7 . The computer-implemented method of claim 6 wherein the scheduler considers the information on the resources of the target hardware system to improve execution performance of the plurality of partitions at the target hardware system.

8 . The computer-implemented method of claim 1 wherein the tensor-based operations include one or more of:

neighborhood operations;

stencil operations;

element-wise mathematical operations;

convolution operations;

batch normalization operations;

operations applying a specified function to each element of an array operations applying a given function to each page of an array; or

matrix multiplication operations.

9 . The computer-implemented method of claim 1 wherein the code generated from the second IR for the computer program is:

C code;

C++ code;

CUDA code;

OpenCL code,

Mojo code,

LLVM assembly code, or

MLIR code.

10 . The computer-implemented method of claim 1 wherein the second programming language supports separating an algorithm from an execution schedule for the algorithm.

11 . The computer-implemented method of claim 1 further comprising:

annotating, by the one or more processors, the computer program with scheduler directives or configuration settings that represent the schedule for the tensor-based operations as determined by the scheduler.

12 . The computer-implemented method of claim 1 further comprising:

running the code on the target hardware system; and

evaluating the code as run on the target hardware system.

Assignments (1)

Continuity (1)

References Cited (21)