GlobalParameters:
  MinimumRequiredVersion: 4.4.0
  PrintLevel: 1
  ForceRedoBenchmarkProblems: True
  ForceRedoLibraryLogic: True
  ForceRedoLibraryClient: True
  CMakeBuildType: Release
  EnqueuesPerSync: 5
  SyncsPerBenchmark: 3
  LibraryPrintDebug: False
  NumElementsToValidate: 128
  ValidationMaxToPrint: 4
  ValidationPrintValids: False
  ShortNames: False
  MergeFiles: False
  Platform: 0
  Device: 0
  KernelTime: True
  PinClocks: True
  SleepPercent: 0
  DataInitTypeBeta : 0
  PrintWinnersOnly: 0

BenchmarkProblems:
  ########################################
  # NN - standard
  ########################################
  -
    - # ProblemType
      OperationType: GEMM
      DataType: s
      TransposeA: False
      TransposeB: False
      UseBeta: True
      Batched: True

    - # BenchmarkProblemSizeGroup - Standard
      InitialSolutionParameters:
      BenchmarkCommonParameters:
        - ProblemSizes:
          - Exact: [ 5504, 5504, 1, 3104 ]
        - KernelLanguage: ["Assembly"]
        - EdgeType: ["ShiftPtr"]
        - LoopTail: [True]
        - PrefetchLocalRead: [True]
      ForkParameters:
        - PrefetchGlobalRead: [False, True]
        - ThreadTile:
          - [ 4, 4 ]
          - [ 8, 4 ]
          - [ 4, 8 ]
          - [ 8, 8 ]
        - WorkGroup:
          - [ 16, 16, 1 ]
          - [ 8, 8, 4 ]
        - WorkGroupMapping: [1, 8]
        #- GlobalSplitU: [1, 2, 4, 8]
        - GlobalSplitU: [1]
        - DepthU: [ 8, 16, 32 ]
        - VectorWidth: [-1, 1]
      BenchmarkForkParameters:
      JoinParameters:
        - MacroTile
        - GlobalSplitU
      BenchmarkJoinParameters:
      BenchmarkFinalParameters:
        - ProblemSizes:
          - Range: [ [64, 64, 64, 7000], [64, 64, 64, 7000], [1], [256, 512, 512, 4096] ]

  ########################################
  # NN - VectorWidth Correctness
  ########################################
    - # Benchmark Group
      InitialSolutionParameters:
      BenchmarkCommonParameters:
        - EdgeType: ["ShiftPtr"]
        - LoopTail: [True]
        - KernelLanguage: ["Assembly"]
        - PrefetchGlobalRead: [True]
        - PrefetchLocalRead: [True]
        - WorkGroupMapping: [1]
      ForkParameters:
        - ThreadTile:
          - [ 2, 2 ]
        - WorkGroup:
          - [ 4, 4, 4 ]
        #- GlobalSplitU: [2, 4, 8]
        - GlobalSplitU: [1]
        - DepthU: [ -1 ]
        - VectorWidth: [-1, 1]
      BenchmarkForkParameters:
      JoinParameters:
      BenchmarkJoinParameters:
      BenchmarkFinalParameters:
        - ProblemSizes:
          - Range: [ [4], [4], [1], [256, 512, 512, 4096] ] # corner
          - Range: [ [4], [64, 64, 64, 7000], [1], [256, 512, 512, 4096] ] # skinny-0
          - Range: [ [64, 64, 64, 7000], [4], [1], [256, 512, 512, 4096] ] # skinny-1

  ########################################
  # NT - standard
  ########################################
  -
    - # ProblemType
      OperationType: GEMM
      DataType: s
      TransposeA: False
      TransposeB: True
      UseBeta: True
      Batched: True

    - # BenchmarkProblemSizeGroup - Standard
      InitialSolutionParameters:
      BenchmarkCommonParameters:
        - ProblemSizes:
          - Exact: [ 5504, 5504, 1, 3104 ]
        - KernelLanguage: ["Assembly"]
        - EdgeType: ["ShiftPtr"]
        - LoopTail: [True]
        - PrefetchLocalRead: [True]
      ForkParameters:
        - PrefetchGlobalRead: [False, True]
        - ThreadTile:
          - [ 4, 4 ]
          - [ 8, 4 ]
          - [ 4, 8 ]
          - [ 8, 8 ]
        - WorkGroup:
          - [ 16, 16, 1 ]
          - [ 8, 8, 4 ]
        - WorkGroupMapping: [1, 8]
        #- GlobalSplitU: [1, 2, 4, 8]
        - GlobalSplitU: [1]
        - DepthU: [ 8, 16, 32 ]
        - VectorWidth: [-1, 1]
      BenchmarkForkParameters:
      JoinParameters:
        - MacroTile
        - GlobalSplitU
      BenchmarkJoinParameters:
      BenchmarkFinalParameters:
        - ProblemSizes:
          - Range: [ [64, 64, 64, 7000], [64, 64, 64, 7000], [1], [256, 512, 512, 4096] ]

  ########################################
  # NT - VectorWidth Correctness
  ########################################
    - # Benchmark Group
      InitialSolutionParameters:
      BenchmarkCommonParameters:
        - EdgeType: ["ShiftPtr"]
        - LoopTail: [True]
        - KernelLanguage: ["Assembly"]
        - PrefetchGlobalRead: [True]
        - PrefetchLocalRead: [True]
        - WorkGroupMapping: [1]
      ForkParameters:
        - ThreadTile:
          - [ 2, 2 ]
        - WorkGroup:
          - [ 4, 4, 4 ]
        #- GlobalSplitU: [2, 4, 8]
        - GlobalSplitU: [1]
        - DepthU: [ -1 ]
        - VectorWidth: [-1, 1]
      BenchmarkForkParameters:
      JoinParameters:
      BenchmarkJoinParameters:
      BenchmarkFinalParameters:
        - ProblemSizes:
          - Range: [ [4], [4], [1], [256, 512, 512, 4096] ] # corner
          - Range: [ [4], [64, 64, 64, 7000], [1], [256, 512, 512, 4096] ] # skinny-0
          - Range: [ [64, 64, 64, 7000], [4], [1], [256, 512, 512, 4096] ] # skinny-1

  ########################################
  # TN - standard
  ########################################
  -
    - # ProblemType
      OperationType: GEMM
      DataType: s
      TransposeA: True
      TransposeB: False
      UseBeta: True
      Batched: True

    - # BenchmarkProblemSizeGroup - Standard
      InitialSolutionParameters:
      BenchmarkCommonParameters:
        - ProblemSizes:
          - Exact: [ 5504, 5504, 1, 3104 ]
        - KernelLanguage: ["Assembly"]
        - EdgeType: ["ShiftPtr"]
        - LoopTail: [True]
        - PrefetchLocalRead: [True]
      ForkParameters:
        - PrefetchGlobalRead: [False, True]
        - ThreadTile:
          - [ 4, 4 ]
          - [ 8, 4 ]
          - [ 4, 8 ]
          - [ 8, 8 ]
        - WorkGroup:
          - [ 16, 16, 1 ]
          - [ 8, 8, 4 ]
        - WorkGroupMapping: [1, 8]
        #- GlobalSplitU: [1, 2, 4, 8]
        - GlobalSplitU: [1]
        - DepthU: [ 8, 16, 32 ]
        - VectorWidth: [-1, 1]
      BenchmarkForkParameters:
      JoinParameters:
        - MacroTile
        - GlobalSplitU
      BenchmarkJoinParameters:
      BenchmarkFinalParameters:
        - ProblemSizes:
          - Range: [ [64, 64, 64, 7000], [64, 64, 64, 7000], [1], [256, 512, 512, 4096] ]

  ########################################
  # TN - VectorWidth Correctness
  ########################################
    - # Benchmark Group
      InitialSolutionParameters:
      BenchmarkCommonParameters:
        - EdgeType: ["ShiftPtr"]
        - LoopTail: [True]
        - KernelLanguage: ["Assembly"]
        - PrefetchGlobalRead: [True]
        - PrefetchLocalRead: [True]
        - WorkGroupMapping: [1]
      ForkParameters:
        - ThreadTile:
          - [ 2, 2 ]
        - WorkGroup:
          - [ 4, 4, 4 ]
        #- GlobalSplitU: [2, 4, 8]
        - GlobalSplitU: [1]
        - DepthU: [ -1 ]
        - VectorWidth: [-1, 1]
      BenchmarkForkParameters:
      JoinParameters:
      BenchmarkJoinParameters:
      BenchmarkFinalParameters:
        - ProblemSizes:
          - Range: [ [4], [4], [1], [256, 512, 512, 4096] ] # corner
          - Range: [ [4], [64, 64, 64, 7000], [1], [256, 512, 512, 4096] ] # skinny-0
          - Range: [ [64, 64, 64, 7000], [4], [1], [256, 512, 512, 4096] ] # skinny-1

  ########################################
  # TT - standard
  ########################################
  -
    - # ProblemType
      OperationType: GEMM
      DataType: s
      TransposeA: True
      TransposeB: True
      UseBeta: True
      Batched: True

    - # BenchmarkProblemSizeGroup - Standard
      InitialSolutionParameters:
      BenchmarkCommonParameters:
        - ProblemSizes:
          - Exact: [ 5504, 5504, 1, 3104 ]
        - KernelLanguage: ["Assembly"]
        - EdgeType: ["ShiftPtr"]
        - LoopTail: [True]
        - PrefetchLocalRead: [True]
      ForkParameters:
        - PrefetchGlobalRead: [False, True]
        - ThreadTile:
          - [ 4, 4 ]
          - [ 8, 4 ]
          - [ 4, 8 ]
          - [ 8, 8 ]
        - WorkGroup:
          - [ 16, 16, 1 ]
          - [ 8, 8, 4 ]
        - WorkGroupMapping: [1, 4]
        #- GlobalSplitU: [1, 2, 4, 8]
        - GlobalSplitU: [1]
        - DepthU: [ 8, 16, 32 ]
        - VectorWidth: [-1, 1]
      BenchmarkForkParameters:
      JoinParameters:
        - MacroTile
        - GlobalSplitU
      BenchmarkJoinParameters:
      BenchmarkFinalParameters:
        - ProblemSizes:
          - Range: [ [64, 64, 64, 7000], [64, 64, 64, 7000], [1], [256, 512, 512, 4096] ]

  ########################################
  # TT - VectorWidth Correctness
  ########################################
    - # Benchmark Group
      InitialSolutionParameters:
      BenchmarkCommonParameters:
        - EdgeType: ["ShiftPtr"]
        - LoopTail: [True]
        - KernelLanguage: ["Assembly"]
        - PrefetchGlobalRead: [True]
        - PrefetchLocalRead: [True]
        - WorkGroupMapping: [1]
      ForkParameters:
        - ThreadTile:
          - [ 2, 2 ]
        - WorkGroup:
          - [ 4, 4, 4 ]
        #- GlobalSplitU: [2, 4, 8]
        - GlobalSplitU: [1]
        - DepthU: [ -1 ]
        - VectorWidth: [-1, 1]
      BenchmarkForkParameters:
      JoinParameters:
      BenchmarkJoinParameters:
      BenchmarkFinalParameters:
        - ProblemSizes:
          - Range: [ [4], [4], [1], [256, 512, 512, 4096] ] # corner
          - Range: [ [4], [64, 64, 64, 7000], [1], [256, 512, 512, 4096] ] # skinny-0
          - Range: [ [64, 64, 64, 7000], [4], [1], [256, 512, 512, 4096] ] # skinny-1


LibraryLogic:
#   ScheduleName: "vega20"
#   DeviceNames: ["Device 66a0", "Device 66a1", "Device 66a7", "Device 66af", "Vega 20"]
#   ArchitectureName: "gfx906"

    ScheduleName: "navi10"
    DeviceNames: ["Device 731f"]
    ArchitectureName: "gfx1010"

    #ScheduleName: "vega10"
    #DeviceNames: ["Device 6863", "Device 6862", "Device 687f", "Device 6860",
    #              "Device 6861", "Vega 10 XTX [Radeon Vega Frontier Edition]",
    #              "Vega [Radeon RX Vega]", "Vega 10 [Radeon Instinct MI25]",
    #              "Vega", "Device 6864", "Device 686c"]
    #ArchitectureName: "gfx900"

#   ScheduleName: "mi25"
#   DeviceNames: ["Device 6860"]
#   ArchitectureName: "gfx900"

#   ScheduleName: "r9nano"
#   DeviceNames: ["Device 7300"]
#   ArchitectureName: "gfx803"

#   ScheduleName: "hip"
#   DeviceNames: ["Device 0000"]
#   ArchitectureName: "fallback"

LibraryClient: