E-Locus - Institutional Repository of the University of Crete

Home Collections Type of Work Post-graduate theses

Post-graduate theses

Search command : Author="Μπριασούλης" And Author="Γεώργιος"

Current Record: 5 of 6633

[Add to Basket]

Identifier

000463830

Title

Checking memory safety of CUDA kernels

Alternative Title

Έλεγχος ασφάλειας μνήμης σε CUDA kernels

Author

Ελευθερίου, Κωνσταντίνος Χ.

Thesis advisor

Πρατικάκης, Πολύβιος

Reviewer

Μπίλας, Άγγελος
Βασιλειάδης, Γιώργος

Abstract

Graphics processing units (GPUs) are heavily used nowadays for tasks like deep learning model construction and big data analysis, due to their high processing power. Nvidia’s CUDA platform enables the use of GPUs for such tasks. CUDA extends the C++ programming language with special functions, called kernels, that run on multiple CUDA threads. Most of the time, kernels perform operations on data that reside in dynamically allocated memory, which can cause runtime errors during the kernel’s execution, if a memory access occurs in unallocated memory. To identify unsafe kernel executions, we perform static analysis on CUDA’s intermediate assembly, called PTX. The reason for this is that the source code of Nvidia libraries, like cuBLAS, is not publicly available. At first, we convert the kernel’s PTX code into LLVM’s intermediate representation in order to detect the loops and the array accesses inside the kernel using LLVM passes. Then, we detect the loop bounds and the expressions used in array accesses and define them in terms of the kernel’s paremeters. The analysis generates constraints that are solved using Microsoft’s Z3 Theorem Prover. The satisfiability of the constraints determines if the kernel is safe to run.

Language

English

Subject

Constraints

GPU

LLVM

Loop bounds

PTX

Static analysis

Όρια βροχών

Ενδιάμεση αναπαράσταση

Μονάδες επεξεργασίας γραφικών