Integration and evaluation of Triton with OmpSs-2 and RISC-V CPUs

The increasing complexity of AI workloads has amplified the demand for efficient, high-performance computing solutions. While GPUs remain the primary platform for accelerating such workloads, programming these devices effectively often requires specialized knowledge of vendor-specific APIs like CUDA...

Descripción completa

Detalles Bibliográficos
Autor: Aguirre López, Alejandro
Tipo de recurso: tesis de maestría
Fecha de publicación:2025
País:España
Institución:Universitat Politècnica de Catalunya (UPC)
Repositorio:UPCommons. Portal del coneixement obert de la UPC
Idioma:inglés
OAI Identifier:oai:upcommons.upc.edu:2117/449427
Acceso en línea:https://hdl.handle.net/2117/449427
Access Level:acceso abierto
Palabra clave:RISC microprocessors
Graphics processing units
Artificial intelligence
Paral·lelisme basat en tasques
Triton
RISC-V
OmpSs-2
Avaluació comparativa
Vectorització
GPU
Task-based parallelism
Microprocessadors RISC
Processadors gràfics
Intel·ligència artificial
Àrees temàtiques de la UPC::Informàtica::Arquitectura de computadors
Descripción
Sumario:The increasing complexity of AI workloads has amplified the demand for efficient, high-performance computing solutions. While GPUs remain the primary platform for accelerating such workloads, programming these devices effectively often requires specialized knowledge of vendor-specific APIs like CUDA and ROCm. Triton is an emerging open-source language and compiler designed to simplify GPU programming by providing a higher-level abstraction while retaining competitive performance. This work evaluates the applicability of Triton across different hardware platforms, including GPUs and CPUs, with a particular focus on task-based programming models like OmpSs-2 and emerging architectures such as RISC-V. We port key kernels from HPCCG and GPT-2 applications to Triton and assess their performance on multiple systems. Our results indicate that while Triton enables easier development and achieves respectable performance on GPUs, its CPU backend currently falls short of optimized vendor libraries and compilers. Finally, we identify specific bottlenecks related to vectorization and register management on RISC-V and suggest avenues for future improvement.