ROCm / MIOpen

AMD's Machine Intelligence Library
https://rocm.docs.amd.com/projects/MIOpen/en/latest/
Other
1.05k stars 218 forks source link

Redesign hip-iGEMM solver implementation #597

Open asroy opened 3 years ago

asroy commented 3 years ago
  1. Simplify IsApplicable(): only depends on blockwise GEMM size
  2. Refactor functions that calculate tuning parameters to avoid re-calculation if possible
  3. Simplify EurisiticInit(): choose a blockwise GEMM size, then calculate the rest of tuning parameter, which should be guaranteed to be IsReallyValid
  4. For tuning, use default and "spare" set of tuning range. If generic_search fails to find parameter that is in default set, then fall back to "spare" set.
  5. Some reviewers' comments:

    549

    https://github.com/ROCmSoftwarePlatform/MIOpen/pull/595#pullrequestreview-531881874 https://github.com/ROCmSoftwarePlatform/MIOpen/pull/568#discussion_r517806880 https://github.com/ROCmSoftwarePlatform/MIOpen/pull/568#discussion_r518227026 https://github.com/ROCmSoftwarePlatform/MIOpen/pull/553#discussion_r514432002 https://github.com/ROCmSoftwarePlatform/MIOpen/pull/553#discussion_r514253178 https://github.com/ROCmSoftwarePlatform/MIOpen/pull/605#discussion_r528903383

asroy commented 3 years ago

@atamazov Feel free to give suggestions, we will use this chance to improve hip iGEMM solver quality

atamazov commented 3 years ago

Please resolve this soon: https://github.com/ROCmSoftwarePlatform/MIOpen/pull/595#discussion_r525427847

ppanchad-amd commented 5 months ago

@asroy Is there any outstanding task on this ticket? If not, please close ticket. Thanks!