Comparison of performance/complexity of stencil code implementations for different architectures. Metrics: Performance (measured in GLUPS) and code complexity.

1. Rationale ============

• Naming Scheme: BENCHMARK_ARCHITECTURE_VARIANT.SUFFIX e.g. jacobi3d_opensse_vanilla.cpp.

• Architectures:

  • serial = standard, sequential code
  • openmp = parallelized with OpenMP
  • cuda = manually parallelized with CUDA

• Variants:

  • vanilla: straightforward code
  • avx: manually vectorized code with AVX intrinsics

• Benchmarks:

  • jacobi3d: simple 3D Jacobi iteration

2. Compiling/Running ====================

stencilProbe comes with a standard CMake script, configuration of the compiler is hence done with "-D CMAKE_CXX_COMPILER=..." and "-D CMAKE_CXX_FLAGS=...".

cd stencilProbe mkdir build cd build cmake ../ make -j10 ./jacobi3d_serial_vanilla 128 128 128 10 >/dev/null

3. (not so) FAQ ===============

Q: All those codes are horribly redundant. Why? A: The purpose of the codes is to track which (minimal) modifications to a sequential, vanilla code are required to implement certain features.