TestParameters:
  marks: [skip-gfx900, skip-gfx906, skip-gfx1010, skip-gfx1011, skip-gfx1012, skip-gfx1030, skip-gfx1031, skip-gfx1032, skip-gfx1033, skip-gfx1034, skip-gfx1035, skip-gfx1036, skip-gfx1100, skip-gfx1101, skip-gfx1102, skip-gfx1103, skip-gfx1150, skip-gfx1151, skip-gfx1152, skip-gfx1153, skip-gfx1200, skip-gfx1201, skip-gfx1250] # not supported by arch

GlobalParameters:
  NumBenchmarks: 3 # we need to run it a few more times to repro the waitcnt bug
  NumElementsToValidate: 16384
  BoundsCheck: True
  KernelTime: True

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

    - # BenchmarkProblemSizeGroup - Source
      InitialSolutionParameters:
      BenchmarkCommonParameters:
        - KernelLanguage: ["Assembly"]
        - EdgeType: ["ShiftPtr"]
        - LoopTail: [True]
      ForkParameters:
        - PrefetchLocalRead: [1]
        - PrefetchGlobalRead: [1]
        - AssertSummationElementMultiple: [2]
        - AssertFree0ElementMultiple: [2]
        - MatrixInstruction:
          - [32, 32, 2, 2]
          - [32, 32, 4, 2]
        - ThreadTile:
          - [ 1, 32 ]
        - WorkGroup:
          - [ 64, 4, 1 ]
        - WorkGroupMapping: [1]
        - GlobalSplitU: [1]
        - DepthU: [32]
        - VectorWidth: [2]
        - SuppressNoLoadLoop: [1] # this triggers the waitcnt bug
      BenchmarkFinalParameters:
        - ProblemSizes:
          - Exact: [256, 8976, 1, 96]
#########################################
LibraryLogic:
  ScheduleName: "arcturus"
  DeviceNames: ["Device 7380", "Device 7388", "Device 738c", "Device 7390"]
  ArchitectureName: "gfx908"