TestParameters:
  marks: [emulation2]

# benchmark assembly and source kernels
GlobalParameters:
  MinimumRequiredVersion: 4.2.0
  CMakeBuildType: Release
  PrintLevel: 1
  ForceRedoBenchmarkProblems: True
  ForceRedoLibraryLogic: True
  ForceRedoLibraryClient: True
  EnqueuesPerSync: 1
  SyncsPerBenchmark: 1
  NumElementsToValidate: -1
  BoundsCheck: True
  ValidationMaxToPrint: 4
  ValidationPrintValids: False
  ShortNames: False
  MergeFiles: True
  DataInitTypeAB: 3
  DataInitTypeC: 3
  KernelTime: True

BenchmarkProblems:

  - # sgemm NN
    - # ProblemType
      OperationType: GEMM
      DataType: d
      DestDataType: d
      TransposeA: False
      TransposeB: False
      UseBeta: True
      Batched: True

    - # BenchmarkProblemSizeGroup - Assembly
      InitialSolutionParameters:
      BenchmarkCommonParameters:
        - LoopTail: [True]
        - EdgeType: ["ShiftPtr"]
        - KernelLanguage: ["Assembly"]
      ForkParameters:
        - GlobalSplitU: [1, 3]
        - PrefetchLocalRead: [False]
        - PrefetchGlobalRead: [True]
        - ThreadTile:
          - [ 3, 5 ]
          - [ 4, 8 ]
          #- [ 8, 8 ]
        - WorkGroup:
          - [ 32,  4,  1 ]
          - [  8,  8,  1 ]
        - DepthU: [-3]
        - VectorWidth: [-1]
      BenchmarkForkParameters:
      JoinParameters:
      BenchmarkJoinParameters:
      BenchmarkFinalParameters:
        - ProblemSizes:
          - Range: [ [127,1,129], 0, [2], [63,1,64] ]

    # - # BenchmarkProblemSizeGroup - Assembly
    #   InitialSolutionParameters:
    #   BenchmarkCommonParameters:
    #     - LoopTail: [True]
    #     - EdgeType: ["ShiftPtr"]
    #   ForkParameters:
    #     - KernelLanguage: ["Assembly"]
    #     - GlobalSplitU: [1, 3]
    #     - PrefetchLocalRead: [False]
    #     - PrefetchGlobalRead: [True]
    #     - ThreadTile:
    #       #- [ 3, 3 ]
    #       - [ 4, 4 ]
    #       - [ 5, 5 ]
    #       #- [ 8, 8 ]
    #     - WorkGroup:
    #       - [ 16, 16,  1 ]
    #       - [  8,  8,  1 ]
    #     - DepthU: [-1]
    #     - VectorWidth: [1]
    #   BenchmarkForkParameters:
    #   JoinParameters:
    #   BenchmarkJoinParameters:
    #   BenchmarkFinalParameters:
    #     - ProblemSizes:
    #       - Range: [ [127,1,129], 0, [2], [63,1,64] ]

  - # sgemm NT
    - # ProblemType
      OperationType: GEMM
      DataType: d
      DestDataType: d
      TransposeA: False
      TransposeB: True
      UseBeta: True
      Batched: True

    - # BenchmarkProblemSizeGroup - Assembly
      InitialSolutionParameters:
      BenchmarkCommonParameters:
        - LoopTail: [True]
        - EdgeType: ["ShiftPtr"]
        - KernelLanguage: ["Assembly"]
      ForkParameters:
        - GlobalSplitU: [1, 3]
        - PrefetchLocalRead: [True]
        - PrefetchGlobalRead: [True]
        - ThreadTile:
          - [ 3, 5 ]
          - [ 4, 8 ]
          #- [ 8, 8 ]
        - WorkGroup:
          - [ 16, 16,  1 ]
          - [  8, 16,  1 ]
        - DepthU: [-3]
        - VectorWidth: [-1]
      BenchmarkForkParameters:
      JoinParameters:
      BenchmarkJoinParameters:
      BenchmarkFinalParameters:
        - ProblemSizes:
          - Range: [ [127,1,129], 0, [2], [63,1,64] ]

    # - # BenchmarkProblemSizeGroup - Assembly
    #   InitialSolutionParameters:
    #   BenchmarkCommonParameters:
    #     - LoopTail: [True]
    #     - EdgeType: ["ShiftPtr"]
    #   ForkParameters:
    #     - KernelLanguage: ["Assembly"]
    #     - GlobalSplitU: [1, 3]
    #     - PrefetchLocalRead: [True]
    #     - PrefetchGlobalRead: [True]
    #     - ThreadTile:
    #       - [ 3, 3 ]
    #       - [ 4, 4 ]
    #       #- [ 5, 5 ]
    #       #- [ 8, 8 ]
    #     - WorkGroup:
    #       - [ 16, 16,  1 ]
    #       - [ 16,  8,  1 ]
    #     - DepthU: [-1]
    #     - VectorWidth: [1]
    #   BenchmarkForkParameters:
    #   JoinParameters:
    #   BenchmarkJoinParameters:
    #   BenchmarkFinalParameters:
    #     - ProblemSizes:
    #       - Range: [ [127,1,129], 0, [2], [63,1,64] ]

  - # sgemm TN
    - # ProblemType
      OperationType: GEMM
      DataType: d
      DestDataType: d
      TransposeA: True
      TransposeB: False
      UseBeta: True
      Batched: True

    - # BenchmarkProblemSizeGroup - Assembly
      InitialSolutionParameters:
      BenchmarkCommonParameters:
        - LoopTail: [True]
        - EdgeType: ["ShiftPtr"]
        - KernelLanguage: ["Assembly"]
      ForkParameters:
        - GlobalSplitU: [1, 3]
        - PrefetchLocalRead: [True]
        - PrefetchGlobalRead: [False]
        - ThreadTile:
          - [ 3, 5 ]
          - [ 4, 8 ]
          #- [ 8, 8 ]
        - WorkGroup:
          - [ 16, 16,  1 ]
          - [ 32,  4,  1 ]
        - DepthU: [-4]
        - VectorWidth: [-1]
      BenchmarkForkParameters:
      JoinParameters:
      BenchmarkJoinParameters:
      BenchmarkFinalParameters:
        - ProblemSizes:
          - Range: [ [127,1,129], 0, [2], [63,1,64] ]

    # - # BenchmarkProblemSizeGroup - Assembly
    #   InitialSolutionParameters:
    #   BenchmarkCommonParameters:
    #     - LoopTail: [True]
    #     - EdgeType: ["ShiftPtr"]
    #   ForkParameters:
    #     - KernelLanguage: ["Assembly"]
    #     - GlobalSplitU: [1, 3]
    #     - PrefetchLocalRead: [True]
    #     - PrefetchGlobalRead: [False]
    #     - ThreadTile:
    #       - [ 3, 3 ]
    #       - [ 4, 4 ]
    #       #- [ 5, 5 ]
    #       #- [ 8, 8 ]
    #     - WorkGroup:
    #       - [ 16, 16,  1 ]
    #       - [  8,  8,  1 ]
    #     - DepthU: [-1]
    #     - VectorWidth: [1]
    #   BenchmarkForkParameters:
    #   JoinParameters:
    #   BenchmarkJoinParameters:
    #   BenchmarkFinalParameters:
    #     - ProblemSizes:
    #       - Range: [ [127,1,129], 0, [2], [63,1,64] ]

  - # sgemm TT
    - # ProblemType
      OperationType: GEMM
      DataType: d
      DestDataType: d
      TransposeA: True
      TransposeB: True
      UseBeta: True
      Batched: True

    - # BenchmarkProblemSizeGroup - Assembly
      InitialSolutionParameters:
      BenchmarkCommonParameters:
        - LoopTail: [True]
        - EdgeType: ["ShiftPtr"]
        - KernelLanguage: ["Assembly"]
      ForkParameters:
        - PrefetchLocalRead: [False]
        - PrefetchGlobalRead: [False]
        - ThreadTile:
          - [ 3, 5 ]
          - [ 4, 8 ]
          #- [ 8, 8 ]
        - WorkGroup:
          - [ 16, 16,  1 ]
          - [  8,  8,  1 ]
        - DepthU: [-3]
        - VectorWidth: [-1]
      BenchmarkForkParameters:
      JoinParameters:
      BenchmarkJoinParameters:
      BenchmarkFinalParameters:
        - ProblemSizes:
          - Range: [ [127,1,129], 0, [2], [63,1,64] ]

    # - # BenchmarkProblemSizeGroup - Assembly
    #   InitialSolutionParameters:
    #   BenchmarkCommonParameters:
    #     - LoopTail: [True]
    #     - EdgeType: ["ShiftPtr"]
    #   ForkParameters:
    #     - KernelLanguage: ["Assembly"]
    #     - PrefetchLocalRead: [False]
    #     - PrefetchGlobalRead: [False]
    #     - ThreadTile:
    #       - [ 3, 3 ]
    #       - [ 4, 4 ]
    #       #- [ 5, 5 ]
    #       #- [ 8, 8 ]
    #     - WorkGroup:
    #       - [ 16, 16,  1 ]
    #       - [  8,  8,  1 ]
    #     - DepthU: [-1]
    #     - VectorWidth: [1]
    #   BenchmarkForkParameters:
    #   JoinParameters:
    #   BenchmarkJoinParameters:
    #   BenchmarkFinalParameters:
    #     - ProblemSizes:
    #       - Range: [ [127,1,129], 0, [2], [63,1,64] ]