Curso de programación en GPU

Post-PASI training: Week 1 » Panamerican Advanced Studies Institute » Boston University

Class 1 (August 8th):

Lecture 1 – slides

• Understand the need for multi-core in applications
• Manycore architecture:
  • GPU vs CPU chip design
  • Data parallelism
  • Concepts behind a CUDA-friendly algorithm
• Basic CUDA:
  • C-like language
  • Threads and thread hierarchy
  • Launching a CUDA kernel

Lab 1 (August 9th):
Lab1 – slides
Lab1 – Codes
matrixMultiplication.cu

• Familiarize with CUDA and nvcc compiler
• Device query
• Launch a simple vector add
• Implement a matrix matrix multiplication

• References:
  • Kirk, D. and Hwu, W. Programming Massively Parallel Processors. (Ch. 1, Ch. 2, Ch. 3)
  • CUDA C Programming Guide. Version 4. Ch. 1-2

Class 2 (August 10th):
Lecture 2 – slides

• Programming model: mapping the discretized model to the GPU threads
• Warps and warp scheduling
• Multilevel memory hierarchy
  • Shared, global, registers, textures, constant, texture memories
  • Sizes and latency
• Fundamentals of the finite difference method

Lab 2 (August 11th):
Lab2 – Slides
Lab2 – Codes
FD_2D_global.cu
FD_2D_texture_pad.cu

• Implementation 2D explicit heat transfer with global memory
• Implementation 2D explicit heat transfer with texture memory

• References:
  • Kirk, D. and Hwu, W. Programming Massively Parallel Processors (Ch. 4)
  • Micikevicius P. 3D Finite Difference Computations on GPUs using CUDA
  • Sanders, J. And Kandrot E. CUDA by Example. (Ch. 7)

Class 3 (August 12th):
Lecture 3 – Slides

• Shared memory in detail
• Tiling
• Bank conflicts
• Race conditions and atomic operations

• References:
  • Kirk, D. and Hwu, W. Programming Massively Parallel Processors. (Ch. 4, Ch. 5)
  • Micikevicius P. 3D Finite Difference Computations on GPUs using CUDA