00

Code Generation for Custom Architectures using Constraint Programming

Paper ☆ WATCH ⎘ SHARE

First seen 6/9/2026

Last seen 6/9/2026

Evidence 19 chunks

01

NEIGHBORHOOD

40 nodes · 64 edges

graph · Code Generation for Custom Architectures using Constraint Programming · depth=1

02

RELATIONSHIPS

39 connections

Code Generation uses → 100% 2e evidence

The thesis addresses automatic code generation for custom architectures.

Data Assignment uses → 100% 2e evidence

Data assignment is addressed in the thesis as a key subproblem of code generation.

Integer Linear Programming mentions → 85% 2e evidence

Integer linear programming is mentioned as an alternative approach to constraint programming for code generation.

Register Allocation uses → 95% 2e evidence

Register allocation is discussed and used within the code generation framework.

Modulo Scheduling uses → 95% 2e evidence

Modulo scheduling is used as an additional set of constraints for loop scheduling in the thesis.

intermediate representation uses → 95% 2e evidence

The code generation framework uses intermediate representations as input to the back end.

Constraint Programming uses → 100% 2e evidence

The thesis uses constraint programming as the core paradigm for modeling code generation problems.

Very Large Instruction Word mentions → 85% 2e evidence

VLIW architectures are mentioned in the context of related code generation work.

ePUMA Architecture evaluates → 95% 2e evidence

The thesis targets and evaluates code generation for the ePUMA architecture.

EIT Architecture evaluates → 95% 2e evidence

The thesis targets and evaluates code generation for the EIT architecture.

Global Constraints uses → 95% 2e evidence

Global constraints are used extensively throughout the thesis for modeling code generation.

Cumulative Global Constraint uses → 95% 2e evidence

The cumulative global constraint is used for VLIW and SIMD groupings in scheduling.

Non-overlapping Rectangles uses → 95% 2e evidence

Data assignment is modeled as non-overlapping rectangles using the diff2 global constraint.

Macro-SIMDization mentions → 85% 1e evidence

Macro-SIMDization is mentioned as a related approach for streaming applications.

Micro-SIMDization mentions → 90% 1e evidence

Micro-SIMDization is the approach taken in the thesis for kernel-level SIMD code generation.

Polyhedral Compilation mentions → 80% 1e evidence

Polyhedral compilation is mentioned as a related technique used by other approaches.

Superblock Instruction Scheduling mentions → 85% 1e evidence

Superblock instruction scheduling is discussed in the related work section.

Dynamic Programming mentions → 80% 1e evidence

Dynamic programming is mentioned as used by related work for combined instruction selection and scheduling.

Subgraph Isomorphism mentions → 80% 1e evidence

Subgraph isomorphism is mentioned as used by Unison for instruction selection.

Directed Acyclic Graph mentions → 85% 1e evidence

Directed acyclic graphs are used to represent application basic blocks in code generation.

Coarse-Grained Reconfigurable Architecture mentions → 80% 1e evidence

Coarse-grained reconfigurable architectures are mentioned in the context of modulo scheduling.

Stream Graph Scheduling mentions → 80% 1e evidence

Stream graph scheduling is mentioned in the context of modulo scheduling for embedded systems.

Unison mentions → 90% 1e evidence

Unison is mentioned as a related integrated code generation project.

Application-Set Driven Architecture Exploration mentions → 90% 1e evidence

Application-set driven architecture exploration is a topic covered in one of the included papers.

Inverse Discrete Cosine Transformation mentions → 90% 1e evidence

IDCT is mentioned as an example of a complex instruction implemented in the ePUMA architecture.

Mehmet Ali Arslan authored by → 100% 1e evidence

The doctoral thesis is authored by Mehmet Ali Arslan.

Krzysztof Kuchcinski authored by → 70% 1e evidence

Krzysztof Kuchcinski supervised the thesis and is acknowledged.

Lund University published by → 100% 1e evidence

The thesis was published by the Department of Computer Science, Lund University.

Instruction Selection uses → 100% 1e evidence

Instruction selection is one of the code generation subproblems addressed in the thesis.

Instruction Scheduling uses → 100% 1e evidence

Instruction scheduling is one of the code generation subproblems addressed in the thesis.

JaCoP uses → 100% 1e evidence

JaCoP is used as the constraint solver framework throughout the thesis.

Diff2 Global Constraint uses → 95% 1e evidence

The diff2 global constraint is used to model data assignment as non-overlapping rectangles.

Search Space Heuristics uses → 90% 1e evidence

Search space heuristics are employed to guide the constraint solver.

Branch-and-Bound uses → 90% 1e evidence

Branch-and-bound is used as a technique to search for optimal solutions in the constraint programming framework.

Pareto Points uses → 90% 1e evidence

Pareto points are used in the architecture exploration paper included in the thesis.

Lifetime Analysis uses → 90% 1e evidence

Lifetime analysis is used to determine when data is live for data assignment.

Permutation Vector Optimization uses → 90% 1e evidence

Permutation vector optimization is addressed as a custom memory problem for SIMD architectures.

Loop Unrolling uses → 85% 1e evidence

Loop unrolling is used in combination with modulo scheduling in the thesis.

Domain Specific Language uses → 85% 1e evidence

A domain specific language is used as a higher-level programming interface in one of the included papers.