Uploaded image for project: 'AI Platform Core Components'
  1. AI Platform Core Components
  2. AIPCC-11651

Improvements for the onnxruntime CPU build

XMLWordPrintable

    • Icon: Story Story
    • Resolution: Unresolved
    • Icon: Undefined Undefined
    • None
    • None
    • Accelerator Enablement
    • None
    • False
    • Hide

      None

      Show
      None
    • False

      • store the output of build_cmd run in a separate build log file (grep the plugins for log_filename, use sdist_root_dir.parent), so we have the output in a log file. Right now we only have a build.log with the wheel build step.
      • update to latest version
      • prevent builds on CUDA, ROCm, and other archs. You can either do that in the plugin or by adding a constraint onnxruntime<0.0 to the constraints files for CUDA, ROCm, and so on. The constraint solution is more flexible.
      • verify that the build works on x86_64v3 and does not require AVX512 instructions. I see nnxruntime/core/mlas/lib/q4gemm_avx512.cpp.o in the output. Does onnxruntime detect CPU features at runtime or will it fail on x86_64v3 machines without AVX512?
      • In the build process there are someĀ 
        CMake vars to enable / disable ISA features:
      ./cmake/CMakeLists.txt:option(onnxruntime_USE_AVX "Use AVX instructions" OFF) 
      ./cmake/CMakeLists.txt:option(onnxruntime_USE_AVX2 "Use AVX2 instructions" OFF) 
      ./cmake/CMakeLists.txt:option(onnxruntime_USE_AVX512 "Use AVX512 instructions" OFF)

      We should enable AVX and AVX2 and keep AVX512 disabled.

              lbarto Lance Barto
              rh-ee-alustosa Andre Lustosa Cabral de Paula Motta
              Frank's Team
              Votes:
              0 Vote for this issue
              Watchers:
              1 Start watching this issue

                Created:
                Updated: