/GemmExample

GEMM multi-GPU example program

Primary LanguageCuda

Example multi-GPU kernel transformation

This demonstrates how to transform a GPU kernel into multi-GPU kernel with unified memory.

The gemm source comes from PolyBench/GPU 1.0 test suite.

files

  • gemm.cu : original PolyBench gemm program.
  • gemm_multi_gpu.cu : multi-GPU gemm program.
  • compile.sh : script used for compiling
  • util.hpp : a C++ library, simplifies the transformation

usage of the library

see the comments in util.hpp

run the example

  1. compile
    ./compile.sh
        

    which generate binaries in ./bin/

  2. run
    # run the gemm for 4-GPUs, policy (RM, RM, UM)
    bin/gemm-4-20480-RM-RM-UM