TENSORIZE: Fast synthesis of Tensor programs from legacy code using symbolic tracing, sketching and solving

Alexander Brauckmann; Luc Jaulmes; José W. de Souza Magalhães; Elizabeth Polgreen; Michael F. P. O'Boyle

doi:10.1145/3696443.3708956

TENSORIZE: Fast synthesis of Tensor programs from legacy code using symbolic tracing, sketching and solving

Alexander Brauckmann, Luc Jaulmes, José W. de Souza Magalhães, Elizabeth Polgreen, Michael F. P. O'Boyle

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution

Abstract

Tensor domain specific languages (DSLs) achieve substantial performance due to high-level compiler optimization and hardware acceleration. However, to achieve such performance for existing applications requires the programmer to manual rewrite their legacy code in evolving Tensor DSLs. Prior efforts to automate this translation face significant scalability issues which greatly reduces their applicability to real-world code.

This paper presents TENSORIZE, a novel MLIR-based compiler approach to automatically lift legacy code to high level Tensor DSLs using program synthesis. TENSORIZE uses a symbolic trace of the legacy program as a specification and automatically selects sketches from the target Tensor DSLs to drive the program synthesis. It uses an algebraic solver to rapidly simplify the specification, resulting in a fast, automatic approach that is correct by design. We evaluate TENSORIZE on several legacy code benchmarks and compare against state-of-the-art techniques. Tensorize is able to lift more code than prior schemes, is an order of magnitude faster in synthesis time, and guarantees correctness by construction.

Original language	English
Title of host publication	CGO '25
Subtitle of host publication	Proceedings of the 23rd ACM/IEEE International Symposium on Code Generation and Optimization
Place of Publication	New York, NY, USA
Publisher	Association for Computing Machinery (ACM)
Pages	15-30
Number of pages	16
ISBN (Electronic)	9798400712753
DOIs	https://doi.org/10.1145/3696443.3708956
Publication status	Published - 1 Mar 2025
Event	The International Symposium on Code Generation and Optimization (CGO) 2025 - The Westin Las Vegas Hotel & Spa, Las Vegas, United States Duration: 1 Mar 2025 → 5 Mar 2025 Conference number: 23 https://2025.cgo.org/

Publication series

Name	Proceedings of the CGO
Publisher	Association for Computing Machinery (ACM)
ISSN (Print)	2164-2397
ISSN (Electronic)	2643-2854

Symposium

Symposium	The International Symposium on Code Generation and Optimization (CGO) 2025
Abbreviated title	CGO 2025
Country/Territory	United States
City	Las Vegas
Period	1/03/25 → 5/03/25
Internet address	https://2025.cgo.org/

Keywords / Materials (for Non-textual outputs)

program synthesis
tensor compilers
lifting

Access to Document

10.1145/3696443.3708956Licence: Creative Commons: Attribution (CC-BY)

BrauckmannEtalCGO2025TENSORIZEFastSynthesisofTensorProgramsAccepted author manuscript, 953 KBLicence: Creative Commons: Attribution (CC-BY)

Cite this

Brauckmann, A., Jaulmes, L., de Souza Magalhães, J. W., Polgreen, E., & O'Boyle, M. F. P. (2025). TENSORIZE: Fast synthesis of Tensor programs from legacy code using symbolic tracing, sketching and solving. In CGO '25: Proceedings of the 23rd ACM/IEEE International Symposium on Code Generation and Optimization (pp. 15-30). (Proceedings of the CGO). Association for Computing Machinery (ACM). https://doi.org/10.1145/3696443.3708956

Brauckmann, Alexander ; Jaulmes, Luc ; de Souza Magalhães, José W. et al. / TENSORIZE : Fast synthesis of Tensor programs from legacy code using symbolic tracing, sketching and solving. CGO '25: Proceedings of the 23rd ACM/IEEE International Symposium on Code Generation and Optimization. New York, NY, USA : Association for Computing Machinery (ACM), 2025. pp. 15-30 (Proceedings of the CGO).

@inproceedings{57c9f20655284a0698559c8ebfc66a5d,

title = "TENSORIZE: Fast synthesis of Tensor programs from legacy code using symbolic tracing, sketching and solving",

abstract = "Tensor domain specific languages (DSLs) achieve substantial performance due to high-level compiler optimization and hardware acceleration. However, to achieve such performance for existing applications requires the programmer to manual rewrite their legacy code in evolving Tensor DSLs. Prior efforts to automate this translation face significant scalability issues which greatly reduces their applicability to real-world code. This paper presents TENSORIZE, a novel MLIR-based compiler approach to automatically lift legacy code to high level Tensor DSLs using program synthesis. TENSORIZE uses a symbolic trace of the legacy program as a specification and automatically selects sketches from the target Tensor DSLs to drive the program synthesis. It uses an algebraic solver to rapidly simplify the specification, resulting in a fast, automatic approach that is correct by design. We evaluate TENSORIZE on several legacy code benchmarks and compare against state-of-the-art techniques. Tensorize is able to lift more code than prior schemes, is an order of magnitude faster in synthesis time, and guarantees correctness by construction.",

keywords = "program synthesis, tensor compilers, lifting",

author = "Alexander Brauckmann and Luc Jaulmes and {de Souza Magalh{\~a}es}, {Jos{\'e} W.} and Elizabeth Polgreen and O'Boyle, {Michael F. P.}",

year = "2025",

month = mar,

day = "1",

doi = "10.1145/3696443.3708956",

language = "English",

series = "Proceedings of the CGO",

publisher = "Association for Computing Machinery (ACM)",

pages = "15--30",

booktitle = "CGO '25",

address = "United States",

note = "The International Symposium on Code Generation and Optimization (CGO) 2025, CGO 2025 ; Conference date: 01-03-2025 Through 05-03-2025",

url = "https://2025.cgo.org/",

}

Brauckmann, A, Jaulmes, L, de Souza Magalhães, JW, Polgreen, E & O'Boyle, MFP 2025, TENSORIZE: Fast synthesis of Tensor programs from legacy code using symbolic tracing, sketching and solving. in CGO '25: Proceedings of the 23rd ACM/IEEE International Symposium on Code Generation and Optimization. Proceedings of the CGO, Association for Computing Machinery (ACM), New York, NY, USA, pp. 15-30, The International Symposium on Code Generation and Optimization (CGO) 2025, Las Vegas, Nevada, United States, 1/03/25. https://doi.org/10.1145/3696443.3708956

TENSORIZE: Fast synthesis of Tensor programs from legacy code using symbolic tracing, sketching and solving. / Brauckmann, Alexander; Jaulmes, Luc; de Souza Magalhães, José W. et al.
CGO '25: Proceedings of the 23rd ACM/IEEE International Symposium on Code Generation and Optimization. New York, NY, USA: Association for Computing Machinery (ACM), 2025. p. 15-30 (Proceedings of the CGO).

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution

TY - GEN

T1 - TENSORIZE

T2 - The International Symposium on Code Generation and Optimization (CGO) 2025

AU - Brauckmann, Alexander

AU - Jaulmes, Luc

AU - de Souza Magalhães, José W.

AU - Polgreen, Elizabeth

AU - O'Boyle, Michael F. P.

N1 - Conference code: 23

PY - 2025/3/1

Y1 - 2025/3/1

N2 - Tensor domain specific languages (DSLs) achieve substantial performance due to high-level compiler optimization and hardware acceleration. However, to achieve such performance for existing applications requires the programmer to manual rewrite their legacy code in evolving Tensor DSLs. Prior efforts to automate this translation face significant scalability issues which greatly reduces their applicability to real-world code. This paper presents TENSORIZE, a novel MLIR-based compiler approach to automatically lift legacy code to high level Tensor DSLs using program synthesis. TENSORIZE uses a symbolic trace of the legacy program as a specification and automatically selects sketches from the target Tensor DSLs to drive the program synthesis. It uses an algebraic solver to rapidly simplify the specification, resulting in a fast, automatic approach that is correct by design. We evaluate TENSORIZE on several legacy code benchmarks and compare against state-of-the-art techniques. Tensorize is able to lift more code than prior schemes, is an order of magnitude faster in synthesis time, and guarantees correctness by construction.

AB - Tensor domain specific languages (DSLs) achieve substantial performance due to high-level compiler optimization and hardware acceleration. However, to achieve such performance for existing applications requires the programmer to manual rewrite their legacy code in evolving Tensor DSLs. Prior efforts to automate this translation face significant scalability issues which greatly reduces their applicability to real-world code. This paper presents TENSORIZE, a novel MLIR-based compiler approach to automatically lift legacy code to high level Tensor DSLs using program synthesis. TENSORIZE uses a symbolic trace of the legacy program as a specification and automatically selects sketches from the target Tensor DSLs to drive the program synthesis. It uses an algebraic solver to rapidly simplify the specification, resulting in a fast, automatic approach that is correct by design. We evaluate TENSORIZE on several legacy code benchmarks and compare against state-of-the-art techniques. Tensorize is able to lift more code than prior schemes, is an order of magnitude faster in synthesis time, and guarantees correctness by construction.

KW - program synthesis

KW - tensor compilers

KW - lifting

U2 - 10.1145/3696443.3708956

DO - 10.1145/3696443.3708956

M3 - Conference contribution

T3 - Proceedings of the CGO

SP - 15

EP - 30

BT - CGO '25

PB - Association for Computing Machinery (ACM)

CY - New York, NY, USA

Y2 - 1 March 2025 through 5 March 2025

ER -

Brauckmann A, Jaulmes L, de Souza Magalhães JW, Polgreen E , O'Boyle MFP. TENSORIZE: Fast synthesis of Tensor programs from legacy code using symbolic tracing, sketching and solving. In CGO '25: Proceedings of the 23rd ACM/IEEE International Symposium on Code Generation and Optimization. New York, NY, USA: Association for Computing Machinery (ACM). 2025. p. 15-30. (Proceedings of the CGO). doi: 10.1145/3696443.3708956

TENSORIZE: Fast synthesis of Tensor programs from legacy code using symbolic tracing, sketching and solving

Abstract

Publication series

Symposium

Keywords / Materials (for Non-textual outputs)

Access to Document

Fingerprint

Cite this