opencl_cuda.md

OpenCL & CUDA

Sources for learning

please add university courses and informative videos

• Parallel Reduction [Slides]
• GPU Memory bootcamp - Tony Scudiero [git repo]
  • Best Practices [Slides] [Video]
  • Beyond the Best Practices [Slides] [Video]
  • Collaborative Access Patterns [Slides] [Video]
• CUB: CUDA Collective primitives library [Git] [Slides] [Video]
• Best Practices Guide by PRACE [HTML] [PDF]

Documentation

• OpenCL specification [1.2] [2.0]
• CUDA Toolkit [latest]
  • CUDA Programming Guide
  • CUDA Runtime API

Source-to-source translation between CUDA and OpenCL

• vtsynergy (https://github.com/vtsynergy)
  • This was shown to work on DAS5 after copying /usr/include/limits.h to $PWD and commenting out the lines around # include_next (122-125) :
    "cu2cl-tool host_code.cc device_code.cu -- -DGPU_ON -I$PWD:/usr/include -I/usr/lib/gcc/x86_64-redhat-linux/4.8.2/include".
• cutocl (https://github.com/benvanwerkhoven/cutocl)

Overview of libraries

• OpenCL-based libraries
  • CLBlast
  • clFFT
• CUDA-based libraries
  • cuBLAS
  • NVBLAS
  • cuFFT
  • nvGRAPH
  • cuRAND
  • cuSPARSE

Foreign Function Interfaces for CUDA and OpenCL

• C++: [Cuda], [OpenCL]
• Python: [PyCuda], [PyOpenCL]
• Java: [JCuda], [JOCL]

Testing

• Unit Testing
  • Example of a unit test for CUDA kernel using the Kernel Tuner
• comparing floating-point results

Debugging and Profiling Tools

• Nvidia Visual Profiler [User Guide]
• CUDA-GDB
• CUDA-MEMCHECK

Performance Optimization

• Resources:

  • Better Performance at Lower Occupancy [Slides] [Video]
  • Maxwell Tuning Guide
  • Pascal Tuning Guide

• Generic Auto Tuners:

  • Kernel Tuner (Python)
  • CLTune (C++)