brainstorm.md

Scope

A very specific narrow topic within GPGPU computing (implemented using OpenCL / CUDA)

Potential topics

• OpenCL: Introduction

  • Command Queue model
  • Host memory model
    • opencl native memory types
    • shared virtual memory
  • Device memory model
    • global memory
    • local memory
    • constant memory
    • private memory
  • Kernel execution domain
    • work item
    • work group
    • events
    • out-of-order command queue
    • use event to implement kernel dependency
  • Synchronization techniques
    • barriers
    • fences
    • atomics
    • memory ordering

• OpenCL: Implementation on different hardware

  • Multi-core CPU
    • multiplexing work group on a single physical GPU
    • vectorization
  • GPU
    • GPU threading
    • Overall Sea Island ISA
    • SIMD unit mapping to work items & work groups

• OpenCL: live demo - 95% efficiency improvement

  • Image clustering example from the book
    • naive CPU implementation
    • naive GPU implementation
    • GPU implementation with coalesced memory
    • GPU implementation with vectorization
    • GPU implementation with local memory (programmable scratch pad memory)
    • GPU implementation with constant memory
  • FFT implementation