BenchmarkProblems:
- - AllowNoFreeDims: true
    DataType: s
    IndexAssignmentsA: [0, 2]
    IndexAssignmentsB: [1, 2, 0]
    NumIndicesC: 2
    OperationType: TensorContraction
    SetConstStrideA: []
    SetConstStrideB:
    - [0, 0]
    UseBeta: false
    UseInitialStridesAB: true
  - BenchmarkCommonParameters:
    - EdgeType: [ShiftPtr]
    BenchmarkFinalParameters:
    - ProblemSizes:
      - Exact: [16, 16, 8]
    BenchmarkForkParameters: null
    BenchmarkJoinParameters: null
    ForkParameters:
    - PrefetchGlobalRead: [1]
    - KernelLanguage: [Source,Assembly]
    - ThreadTile:
      - [2, 2]
    - WorkGroup:
      - [8, 8, 1]
    - DepthU: [8]
    - GlobalReadVectorWidth: [-1]
    - VectorWidth: [1]
    - FractionalLoad: [0]
    - PackBatchDims: [1]
    - StaggerU: [0]
    InitialSolutionParameters: null
    JoinParameters: null
GlobalParameters: {CMakeBuildType: Release, EnqueuesPerSync: 1, ForceRedoBenchmarkProblems: true,
  ForceRedoLibraryClient: true, ForceRedoLibraryLogic: true, KernelTime: true, LibraryPrintDebug: false,
  MergeFiles: true, MinimumRequiredVersion: 4.2.0, NumElementsToValidate: 1000,
  ShortNames: false, SolutionSelectionAlg: 1, SyncsPerBenchmark: 1,
  ValidationMaxToPrint: 4, ValidationPrintValids: false,
  DataInitTypeAB: 1,
  DataInitTypeBeta: 0,
  DataInitTypeD: 2,
  PrintIndexAssignments: 1,
  PrintCodeCommands: 1,
  OldClientSourceTmp: 0,
  #PrintTensorD: 2,
  # PrintTensorRef: 2,


  # OldClientSourceTmp=0 causes missing KernelHeader.h, probably something with the pushdir/pop sourceTmp stuff.