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:
  NumElementsToValidate: -1
  SyncsPerBenchmark: 0
  PrintSolutionRejectionReason: True
  DataInitTypeA: 12
  DataInitTypeB: 13
  DataInitTypeC: 0
  DataInitTypeD: 0
  DataInitTypeBeta: 0
  DataInitTypeAlpha: 1
  DataInitTypeBias: 0
  DataInitTypeScaleAlphaVec: 1
  DataInitTypeScaleA: 1
  DataInitTypeScaleB: 1
  CSVExportWinner: true
  CSVMergeSameProblemID: true
  Device: 0
  MaxWorkspaceSize: 33554432
  KernelTime: True
  RotatingBufferSize: 512

BenchmarkProblems:
  ########################################
  # NN - standard
  ########################################
  -
    - # ProblemType
      OperationType: GEMM
      DataType: b
      DestDataType: b
      ComputeDataType: s
      HighPrecisionAccumulate: True
      TransposeA: True
      TransposeB: False
      UseBeta: True
      Batched: True
      UseBias: 1
      Activation: True
      ActivationType: all
      UseScaleAlphaVec: 1
    - # BenchmarkProblemSizeGroup - Standard
      InitialSolutionParameters:
      BenchmarkCommonParameters:
        - KernelLanguage: ["Assembly"]
      ForkParameters:
        - MatrixInstruction:
          - [16, 16,16, 1,  1,   1, 2,  1,1 ]
        - PrefetchGlobalRead: [2]
        - PrefetchLocalRead: [1]
        - DepthU: [32,64]
        - VectorWidthA: [-1]
        - VectorWidthB: [-1]
        - GlobalReadVectorWidthA: [-1]
        - GlobalReadVectorWidthB: [-1]
        - LocalReadVectorWidth: [-1]
        - TransposeLDS: [1]
        - LdsBlockSizePerPadA: [-1]
        - LdsBlockSizePerPadB: [-1]
        - LdsPadA: [-1]
        - LdsPadB: [-1]
        - WorkGroupMapping: [-1, -8, -300]
        - WorkGroupMappingXCC: [1, 8]
        - WorkGroupMappingXCCGroup: [-1, 0]
        - StaggerUMapping: [1]
        - 1LDSBuffer: [-1]
        - GlobalSplitU: [1]
        - GlobalSplitUAlgorithm: ["MultipleBuffer"]
        - GlobalReadPerMfma: [1.0]
        - LocalWritePerMfma: [-1]
        - StoreRemapVectorWidth: [0]
        - StoreVectorWidth: [-1]
        - SourceSwap: [true]
        - NumElementsPerBatchStore: [16]
        - ClusterLocalRead: [1]
      BenchmarkJoinParameters:
      BenchmarkFinalParameters:
        - ProblemSizes:
          - Exact: [2160,10240,1,16384]
          - Exact: [10240,2160,1,16384]
        - BiasTypeArgs: ['s']
        - ActivationArgs:
          - [Enum: none]