GPU Architecture Simulation

A simulation of a GPU (Graphics Processing Unit) architecture developed in Python. This project emulates the execution flow of parallel kernels across multiple Streaming Multiprocessors (SMs) and cores.

🏗️ Architectural Overview

The simulation replicates the hardware-software interface of modern GPUs:

• Streaming Multiprocessors (SM): Implemented using multiprocessing.Process to simulate independent hardware units.
• CUDA Cores: Emulated via threading.Thread within each SM to handle concurrent execution of threads.
• Memory Hierarchy: * Global Memory (VRAM): Shared across all SMs through the GPUMemory class.
  • Shared/Local Memory: Fast, per-SM memory implemented in SMMemory.
• Synchronization: Uses threading.Barrier to coordinate SIMT (Single Instruction, Multiple Threads) execution and prevent race conditions.

🧬 Implemented Kernels

The simulator includes several parallel algorithms to demonstrate different memory access patterns:

1. Arithmetic Kernels: Vector Increment and Vector Sum.
2. Generalized Blur: A stencil-based operation using a configurable radius and shared memory optimization.
3. Dot Product (Reduction): Uses atomic-style locking (multiprocessing.Value.get_lock()) to aggregate partial sums into a global result.

🛠️ Execution

To run the simulation and test different kernels, execute the main entry point:

python src/gpu.py

Note: You can switch between kernels (INCR, SUMAR, DIFUMINAR, ESCALAR) by modifying the KERNEL_A_PROBAR variable in gpu.py.

✒️ Authorship

Developed by Alberto Peña and Fabio Torres (March 2026).