Learning CUDA

Learning GPU programming with CUDA

Some resources:

• Tutorials:
  • TutorialsPoint introduction to CUDA
  • PluralSight introduction to CUDA
  • NVIDIA blog post
  • Presentation on CUDA C/C++ basics
• CUDA By Example
  • apparently a good book; NOT free
• Accelerated libraries for common operations
  • linear algebra, FFT, RNG, algebraic equations, sparse matrices, tensor ops, linear solvers, graph algorithms
  • check this out before rolling your own algos!
• Thrust
  • high-level library of parallel algorithms and data structures

Tips:

• vscode intellisense w/ CUDA:
  • in vscode, there is an extension called vscode-cudacpp that gets rid of the syntax error highlighting you would get from __global__, kernel<<<a,b>>>, etc. in base C/C++
  • vscode does not autocomplete from the cuda libraries by default, so you need to
    1. add association of .cu or .cuh files with C++ in your settings.json (i.e. do this, but replace "cuda" --> "cpp")
    2. in your C++ language properties c_cpp_properties.json, add /usr/local/cuda/include/** (or wherever your CUDA headers are) to your include path
• CUDA-X libraries
  • add -lcurand etc. to your nvcc commands to tell linker where to look
• device specs
  • can find simple list of device specifications on wikipedia
  • a more complete specification can be found on NVIDIA's documentation, but is difficult to find
  • or you can query your device itself by navigating to ~/cuda/samples/1_Utilities/deviceQuery, build using make, and run ./deviceQuery