This course introduces GPU programming through CUDA with a focus on understanding performance at the kernel level. Students explore how modern GPUs execute parallel workloads and how memory hierarchy and execution structure affect efficiency. The course centers around the design and iterative improvement of matrix multiplication kernels as a concrete case study for learning core GPU concepts. Emphasis is placed on reasoning about design choices, understanding performance tradeoffs, and interpreting how low-level implementation details map to hardware behavior.

- Basic linear algebra (matrix operations) - Programming experience (C++ and Python) - Familiarity with basic computer architecture concepts (CPU execution, memory hierarchy, instruction-level parallelism)

- GPU architecture fundamentals (SIMT, warps, occupancy) - Memory hierarchy (global, shared, registers, caches) - CUDA programming model (threads, blocks, grids) - Kernel optimization strategies (tiling, unrolling, memory reuse) - Performance concepts (bandwidth, latency, coalescing)

In presenza

Presentazione report scritto

P.D.1-1 - Gennaio