No abstract

No abstract

Registro de acceso restringido Este recurso no está disponible en acceso abierto por política de la editorial. No obstante, se puede acceder al texto completo desde la Universitat Jaume I o si el usuario cuenta con suscripción. Registre d'accés restringit Aquest recurs no està disponible en accés obert per política de l'editorial. No obstant això, es pot accedir al text complet des de la Universitat Jaume I o si l'usuari compta amb subscripció. Restricted access item This item isn't open access because of publisher's policy. The full--text version is only available from Jaume I University or if the user has a running suscription to the publisher's contents.

The considerable impact of Convolutional Neural Networks on many Artificial Intelligence tasks has led to the development of various high performance algorithms for the convolution operator present in this type of networks. One of these approaches leverages the im2col transform followed by a general matrix multiplication (gemm) in order to take advantage of the highly optimized realizations of the gemm kernel in many linear algebra libraries. The main problems of this approach are 1) the large memory workspace required to host the intermediate matrices generated by the im2col transform; and 2) the time to perform the im2col transform, which is not negligible for complex neural networks. This paper presents a portable high performance convolution algorithm based on the BLIS realization of the gemm kernel that avoids the use of the intermediate memory by taking advantage of the BLIS structure. In addition, the proposed algorithm eliminates the cost of the explicit im2col transform, while maintaining the portability and performance of the underlying realization of gemm in BLIS.

Abstract. In recent years, several lightweight thread (LWT) libraries have emerged to tackle exascale challenges. These offer programming models (PMs) based on user-level threads and incorporate their own lightweight mechanisms. However, each library proposes its own PM, exposing different semantics and hindering portability. To address this drawback, we have designed Generic Lightweight Thread (GLT), an application programming interface that frames the functionality of the most popular LWT libraries for high-performance computing under a single PM. We implement GLT on top of Argobots, MassiveThreads, and Qthreads. We provide GLT as a dynamic library, as well as in the form of a static version based on macro preprocessing resolution to reduce overhead. This paper discusses the GLT PM and demonstrates its minimal performance impact.