TestParameters:
  marks: [skip-gfx900, skip-gfx906, skip-gfx908, skip-gfx90a, skip-gfx1010, skip-gfx1011, skip-gfx1012, skip-gfx1030, skip-gfx1100, skip-gfx1101, skip-gfx1102, skip-gfx1200, skip-gfx1201, skip-gfx1250] # not supported by arch

GlobalParameters:
  SyncsPerBenchmark: 0
  NumElementsToValidate: -1
  MinimumRequiredVersion: 5.0.0
  DataInitTypeAlpha: 1
  DataInitTypeBeta: 0
  CSVExportWinner: true
  CSVMergeSameProblemID: true
  Device: 0

BenchmarkProblems:
  ########################################
  # TN - standard
  ########################################
  -
    - # ProblemType
      OperationType: GEMM
      DataType: h
      DestDataType: h
      ComputeDataType: s
      HighPrecisionAccumulate: True
      TransposeA: True
      TransposeB: False
      UseBeta: True
      UseBias: 1
      Batched: True
      Activation: True
      ActivationType: all
    - # BenchmarkProblemSizeGroup - Standard - All problem
      InitialSolutionParameters:
      BenchmarkCommonParameters:
        - KernelLanguage: ["Assembly"]
      ForkParameters:
        - MatrixInstruction:
          - [16, 16, 16, 1,  1,  2, 2,  2, 2] #64x64
          - [32, 32, 8, 1,  1,  1, 1,  2, 2] #64x64
        - PrefetchGlobalRead: [2]
        - PrefetchLocalRead: [1]
        - ClusterLocalRead: [1]
        - DepthU: [64]
        - LocalReadVectorWidth: [8]
        - ScheduleIterAlg: [3]
        - ExpandPointerSwap: [false]
        - TransposeLDS: [-1]
        - LdsBlockSizePerPadA: [-1]
        - LdsBlockSizePerPadB: [-1]
        - LdsPadA: [-1]
        - LdsPadB: [-1]
        - 1LDSBuffer: [-1]
        - GlobalSplitU: [1,2,4]
        - PreloadKernArgs: [true]
        - GlobalSplitUAlgorithm: ["MultipleBuffer"]
        - GlobalSplitUCoalesced: [False,True]
        - GlobalSplitUWorkGroupMappingRoundRobin: [False,True]
        - SourceSwap: [true]
      BenchmarkJoinParameters:
      BenchmarkFinalParameters:
        - ProblemSizes:
          - Exact: [255, 255, 1, 255]
          - Exact: [255, 255, 1, 256]
          - Exact: [255, 255, 1, 257]
          - Exact: [255, 256, 1, 255]
          - Exact: [255, 256, 1, 256]
          - Exact: [255, 256, 1, 257]
          - Exact: [255, 257, 1, 255]
          - Exact: [255, 257, 1, 256]
          - Exact: [255, 257, 1, 257]

          - Exact: [256, 255, 1, 255]
          - Exact: [256, 255, 1, 256]
          - Exact: [256, 255, 1, 257]
          - Exact: [256, 256, 1, 255]
          - Exact: [256, 256, 1, 256]
          - Exact: [256, 256, 1, 257]
          - Exact: [256, 257, 1, 255]
          - Exact: [256, 257, 1, 256]
          - Exact: [256, 257, 1, 257]

          - Exact: [257, 255, 1, 255]
          - Exact: [257, 255, 1, 256]
          - Exact: [257, 255, 1, 257]
          - Exact: [257, 256, 1, 255]
          - Exact: [257, 256, 1, 256]
          - Exact: [257, 256, 1, 257]
          - Exact: [257, 257, 1, 255]
          - Exact: [257, 257, 1, 256]
          - Exact: [257, 257, 1, 257]
        - BiasTypeArgs: ['h']
        - ActivationArgs:
          - [Enum: relu]