RegpackTester

Compares performance of LIBXSMM versions (and to GiMMiK).

Process is scripted from code generation all the way to plotting. (Generate/get matrices before-hand)

Made to compare the Register Packing addition to LIBXSMM small-sparse A operator MM routine.

Requirements

• An AVX-512 CPU
• GCC
• Python
• numpy, matplotlib
• Reference LIBXSSM in "./../libxsmm_reference"
• Custom LIBXSSM in "./../libxsmm_custom"
• OpenBlas build in "./../OpenBlas-build"
• If also testing GiMMiK:
  • ICC
  • GiMMiK in "./../GiMMiK" and installed

Flow

Checkout to the commits you want to test in reference and custom LIBXSMMs.

Run ./scripts/benchmark.sh with the following options

• If also testing GiMMiK:
  • -g : 1 (Default is 0)
• -o : log output dir
• -p : plot output dir
• -m : source matrix suite dir
• -t : type of matrix suite (pyfr or synth)
• -n : number of benchmark runs (Default is 3)
• -d : if reference XSMM is dense routine (1) (Default is 0)
• -s : skip benchmark process, plot a previous run
  • Other options above still required
  • Default is 0 (don't skip)
  • Provide the timestamp of the previous run you want to plot

Hardcoded values

• Block alignment = 48

  • src/benchmark/common.h
  • src/plots/tools.py

• XEON_8175M Stats

  • src/plots/cpu_stats.py