TestParameters:
  marks: [skip-gfx900, skip-gfx906, skip-gfx908, skip-gfx90a, skip-gfx950, 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: 1
  # SyncsPerBenchmark is not set to 0: this test uses LibraryLogic which
  # requires benchmark data to build the library table.
  NumElementsToValidate: -1
  MinimumRequiredVersion: 5.0.0
  PrintLevel: 1
  PrintSolutionRejectionReason: True
  Device: 0
  CMakeBuildType: Release
  KernelTime: True
  MaxWorkspaceSize: 13421772800
  DataInitTypeAlpha: 1
  DataInitTypeBeta: 0
  BoundsCheck: 2

BenchmarkProblems:
  ########################################
  # F32X NT
  ########################################
  -
    - # ProblemType
      OperationType: GEMM
      DataType: S
      DestDataType: S
      ComputeDataType: S
      F32XdlMathOp: X
      TransposeA: False
      TransposeB: True
      UseBeta: True
      Batched: True
    - # BenchmarkProblemSizeGroup - Standard
      InitialSolutionParameters:
      BenchmarkCommonParameters:
        - KernelLanguage: ["Assembly"]
      ForkParameters:
        - MatrixInstruction:
          - [16, 16,8, 1,  1,   4, 4,  2,2 ]  # 128x128
        - DepthU: [ 16 ]
        - PrefetchGlobalRead: [2]
        - PrefetchLocalRead: [1]
        - VectorWidthA: [1,2,4]
        - VectorWidthB: [1,2,4]
        - ScheduleIterAlg: [3]
        - SourceSwap: [false, true]
      BenchmarkJoinParameters:
      BenchmarkFinalParameters:
        - ProblemSizes:
          - Exact: [128, 128, 1, 128]
  ########################################
  # F32X TN
  ########################################
  -
    - # ProblemType
      OperationType: GEMM
      DataType: S
      DestDataType: S
      ComputeDataType: S
      F32XdlMathOp: X
      TransposeA: True
      TransposeB: False
      UseBeta: True
      Batched: True
    - # BenchmarkProblemSizeGroup - Standard
      InitialSolutionParameters:
      BenchmarkCommonParameters:
        - KernelLanguage: ["Assembly"]
      ForkParameters:
        - MatrixInstruction:
          - [16, 16,8, 1,  1,   4, 4,  2,2 ]  # 128x128
        - DepthU: [ 16 ]
        - PrefetchGlobalRead: [2]
        - PrefetchLocalRead: [1]
        - VectorWidthA: [1,2,4]
        - VectorWidthB: [1,2,4]
        - ScheduleIterAlg: [3]
        - SourceSwap: [false, true]
        - LocalReadVectorWidth: [1,2,4]
      BenchmarkJoinParameters:
      BenchmarkFinalParameters:
        - ProblemSizes:
          - Exact: [128, 128, 1, 128]

LibraryLogic:
    ScheduleName: "aquavanjaram"
    DeviceNames: ["Device 0049", "Device 0050"]
    ArchitectureName: "gfx942"
    LibraryType: "GridBased"

LibraryClient: