GlobalParameters:
  MinimumRequiredVersion: 4.2.0
  ForceRedoBenchmarkProblems: True
  ForceRedoLibraryLogic: True
  ForceRedoLibraryClient: True
  CMakeBuildType: Debug
  EnqueuesPerSync: 1
  SyncsPerBenchmark: 1
  LibraryPrintDebug: False
  NumElementsToValidate: 1000
  ValidationMaxToPrint: 4
  ValidationPrintValids: False
  ShortNames: False
  MergeFiles: True
  KernelTime: True
  SolutionSelectionAlg: 1
  DataInitTypeA: 1
  DataInitTypeB: 3
  DataInitTypeAlpha: 1
  #PrintTensorD: 2
BenchmarkProblems:
  -
    -
      OperationType: GEMM
      DataType: s
      TransposeA: False
      TransposeB: True
      SetConstStrideA: [[0, 1]]
      SetConstStrideB: [[2, 0]]
      UseBeta: False
      Batched: True
    -
      InitialSolutionParameters:
      BenchmarkCommonParameters:
        - EdgeType: ["ShiftPtr"]
      ForkParameters:
        - PrefetchGlobalRead: [1]
        - PrefetchLocalRead: [0]
        - KernelLanguage: ["Assembly"]
        - DirectToLdsA: [False, True]
        - DirectToLdsB: [False, True]
        - ThreadTile:
          - [ 8, 8 ]
        - WorkGroup:
          - [  8, 8, 1 ]
        - DepthU: [4]
          #- GlobalReadVectorWidth: [1]
        - VectorWidth: [1]

          #- FractionalLoad: [0]
      BenchmarkForkParameters:
      JoinParameters:
      BenchmarkJoinParameters:
      BenchmarkFinalParameters:
         - ProblemSizes:
           #- Exact: [64, 16, 1, 8]
           - Exact: [64, 64, 1, 8]
             #- Range: [[64],[64],[1],[16,1,30]]

# Bug with PFR=1 and D2L=1.
# Conservative sync does not help.
# Seems to be issue on the load path.
# Could be issue with waitcnts ?