TestParameters:
  marks: [skip-gfx900, skip-gfx906, skip-gfx908, skip-gfx90a, 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, emulation5] # not supported by arch

GlobalParameters:
  #MinimumRequiredVersion: 4.14.0
  #PrintLevel: 2
  ClientLogLevel: 3
  #ForceRedoBenchmarkProblems: True
  #ForceRedoLibraryLogic: True
  #ForceRedoLibraryClient: True
  #CMakeBuildType: Release
  #EnqueuesPerSync: 1
  #SyncsPerBenchmark: 1
  #LibraryPrintDebug: False
  #NumElementsToValidate: 16384

  NumElementsToValidate: -1   # everything
  ValidationMaxToPrint: 16384

  #ValidationPrintValids: False
  #ShortNames: False
  #MergeFiles: True
  #Platform: 0
  #Device: 0
  KernelTime: True
  #PinClocks: False
  #SleepPercent: 200
  PrintSolutionRejectionReason: True
  #DataInitTypeA: 3
  #DataInitTypeB: 3

  DataInitTypeA: 3
  DataInitTypeB: 3
  DataInitTypeC: 3  # SerialIdx%4

  #DataInitTypeD: 0
  DataInitTypeBeta: 1
  DataInitTypeAlpha: 2

  #BoundsCheck: 1

  #PrintTensorA: 0x1
  #PrintTensorB: 0x1
  #PrintTensorD: 0x3  ## both init and copy back
  NewClient: 2

BenchmarkProblems:
  ########################################
  # NT - standard
  ########################################
  -
    - # ProblemType - NT
      OperationType: GEMM
      DataType: F8
      DestDataType: F8
      ComputeDataType: S
      HighPrecisionAccumulate: True
      TransposeA: False
      TransposeB: True
      UseBeta: True
      Batched: True
      #Fp8NoPackUpConversion: True    # for non-zero beta

    - # BenchmarkProblemSizeGroup - Standard
      InitialSolutionParameters:
      BenchmarkCommonParameters:
        - KernelLanguage: ["Assembly"]
        - EdgeType: ["ShiftPtr"]
        - LoopTail: [True]
        - PrefetchLocalRead: [True]
      ForkParameters:
        - MatrixInstruction:
          - [16, 16, 32, 1, 1, 1, 1, 1, 1]
          - [16, 16, 32, 1, 1, 1, 1, 2, 2]
          - [32, 32, 16, 1, 1, 1, 1, 1, 1]
          - [32, 32, 16, 1, 1, 1, 1, 2, 2]
        - PrefetchGlobalRead: [True]
        #- WorkGroupMapping: [8]
        #- GlobalSplitU: [1]
        - DepthU: [64]
        #- VectorWidth: [-1]
        #- AssertSummationElementMultiple: [4]
        #- AssertFree0ElementMultiple: [4]
        - StoreRemapVectorWidth: [0, 4]
          #- 1LDSBuffer: [0,1]
      JoinParameters:
      BenchmarkJoinParameters:
      BenchmarkFinalParameters:
        - ProblemSizes:
          #- Exact: [ 16, 16, 1, 32 ]
          #- Exact: [ 64, 64, 1, 64 ]
            #- Exact: [ 128, 128, 1, 128 ]
            #- Exact: [ 111, 111, 1, 111 ]
            #- Exact: [ 111, 111, 1, 64 ]
          - Range : [ [99, 1, 117], [111], [1], [64] ]
          #- Exact: [ 128, 128, 1, 256 ]    # larger K to test clipping