TestParameters:
  marks: [skip-gfx900, skip-gfx906, skip-gfx908, skip-gfx1010, skip-gfx1011, skip-gfx1012, skip-gfx1030, skip-gfx1031, skip-gfx1032, skip-gfx1033, skip-gfx1034, skip-gfx1035, skip-gfx1036, emulation1] # not supported by arch

GlobalParameters:
  NumElementsToValidate: -1
  BoundsCheck: True
  KernelTime: True
  PrintSolutionRejectionReason: True

BenchmarkProblems:
  ########################################
  # NT - standard
  ########################################
  - # sgemm TN
    - # ProblemType
      OperationType: GEMM
      DataType: d
      TransposeA: True
      TransposeB: False
      UseBeta: True
      Batched: True

    - # BenchmarkProblemSizeGroup - Standard
      InitialSolutionParameters:
      BenchmarkCommonParameters:
        - KernelLanguage: ["Assembly"]
        - EdgeType: ["ShiftPtr"]
        - PrefetchLocalRead: [True]
      ForkParameters:
        - MatrixInstruction:
          - [16, 16, 4, 1,  1,  8,2,  1,4]  # 128x128
          # - [16, 16, 4, 1,  1,  4,4,  2,2]  # 128x128
          # - [16, 16, 4, 1,  1,  2,8,  4,1]  # 128x128
        - SourceSwap: [1]
        - PrefetchGlobalRead: [1]
        - ThreadTile:
          - [ 1, 32 ]
        - WorkGroup:
          - [  16, 16, 1 ]
        - WorkGroupMapping: [8]
        - GlobalSplitU: [1]
        - DepthU: [ 8 ]
        - VectorWidth: [1]
      BenchmarkForkParameters:
      JoinParameters:
      BenchmarkJoinParameters:
      BenchmarkFinalParameters:
        - ProblemSizes:
          # - Range: [ [1024], [1024], [1], [1024] ]
          - Range: [ [1031], [1031], [8], [1031] ]

  # - # sgemm NN
  #   - # ProblemType
  #     OperationType: GEMM
  #     DataType: d
  #     TransposeA: False
  #     TransposeB: False
  #     UseBeta: True
  #     Batched: True

  #   - # BenchmarkProblemSizeGroup - Standard
  #     InitialSolutionParameters:
  #     BenchmarkCommonParameters:
  #       - KernelLanguage: ["Assembly"]
  #       - EdgeType: ["ShiftPtr"]
  #       - PrefetchLocalRead: [True]
  #     ForkParameters:
  #       - MatrixInstruction:
  #         - [16, 16, 4, 1,  1,  8,2,  1,4]  # 128x128
  #         - [16, 16, 4, 1,  1,  4,4,  2,2]  # 128x128
  #         - [16, 16, 4, 1,  1,  2,8,  4,1]  # 128x128
  #       - SourceSwap: [1]
  #       - PrefetchGlobalRead: [1]
  #       - ThreadTile:
  #         - [ 1, 32 ]
  #       - WorkGroup:
  #         - [  16, 16, 1 ]
  #       - WorkGroupMapping: [8]
  #       - GlobalSplitU: [1]
  #       - DepthU: [ 8 ]
  #       - VectorWidth: [1]
  #     BenchmarkForkParameters:
  #     JoinParameters:
  #     BenchmarkJoinParameters:
  #     BenchmarkFinalParameters:
  #       - ProblemSizes:
  #         - Range: [ [1024], [1024], [1], [1024] ]
  #         - Range: [ [1031], [1031], [8], [1031] ]

  # - # sgemm NT
  #   - # ProblemType
  #     OperationType: GEMM
  #     DataType: d
  #     TransposeA: False
  #     TransposeB: True
  #     UseBeta: True
  #     Batched: True

  #   - # BenchmarkProblemSizeGroup - Standard
  #     InitialSolutionParameters:
  #     BenchmarkCommonParameters:
  #       - KernelLanguage: ["Assembly"]
  #       - EdgeType: ["ShiftPtr"]
  #       - PrefetchLocalRead: [True]
  #     ForkParameters:
  #       - MatrixInstruction:
  #         - [16, 16, 4, 1,  1,  8,2,  1,4]  # 128x128
  #         - [16, 16, 4, 1,  1,  4,4,  2,2]  # 128x128
  #         - [16, 16, 4, 1,  1,  2,8,  4,1]  # 128x128
  #       - SourceSwap: [1]
  #       - PrefetchGlobalRead: [1]
  #       - ThreadTile:
  #         - [ 1, 32 ]
  #       - WorkGroup:
  #         - [  16, 16, 1 ]
  #       - WorkGroupMapping: [8]
  #       - GlobalSplitU: [1]
  #       - DepthU: [ 8 ]
  #       - VectorWidth: [1]
  #     BenchmarkForkParameters:
  #     JoinParameters:
  #     BenchmarkJoinParameters:
  #     BenchmarkFinalParameters:
  #       - ProblemSizes:
  #         - Range: [ [1024], [1024], [1], [1024] ]
  #         - Range: [ [1031], [1031], [8], [1031] ]

  # - # sgemm TT
  #   - # ProblemType
  #     OperationType: GEMM
  #     DataType: d
  #     TransposeA: True
  #     TransposeB: True
  #     UseBeta: True
  #     Batched: True

  #   - # BenchmarkProblemSizeGroup - Standard
  #     InitialSolutionParameters:
  #     BenchmarkCommonParameters:
  #       - KernelLanguage: ["Assembly"]
  #       - EdgeType: ["ShiftPtr"]
  #       - PrefetchLocalRead: [True]
  #     ForkParameters:
  #       - MatrixInstruction:
  #         - [16, 16, 4, 1,  1,  8,2,  1,4]  # 128x128
  #         - [16, 16, 4, 1,  1,  4,4,  2,2]  # 128x128
  #         - [16, 16, 4, 1,  1,  2,8,  4,1]  # 128x128
  #       - SourceSwap: [1]
  #       - PrefetchGlobalRead: [1]
  #       - ThreadTile:
  #         - [ 1, 32 ]
  #       - WorkGroup:
  #         - [  16, 16, 1 ]
  #       - WorkGroupMapping: [8]
  #       - GlobalSplitU: [1]
  #       - DepthU: [ 8 ]
  #       - VectorWidth: [1]
  #     BenchmarkForkParameters:
  #     JoinParameters:
  #     BenchmarkJoinParameters:
  #     BenchmarkFinalParameters:
  #       - ProblemSizes:
  #         - Range: [ [1024], [1024], [1], [1024] ]
  #         - Range: [ [1031], [1031], [8], [1031] ]