GlobalParameters:
  MinimumRequiredVersion: 4.1.0
  CMakeBuildType: Release
  NumElementsToValidate: 0
  ValidationMaxToPrint: 16
  PinClocks: False
  SleepPercent: 0
  DataInitTypeA: 0
  DataInitTypeB: 0
  EnqueuesPerSync: 100
  KernelTime: True
  PrintConvolutionUsage: 0
  PrintSolutionRejectionReason: 1

BenchmarkProblems:
  -
    - # ProblemType
      OperationType: ConvolutionForward
      DataType: s
      ConvolutionConfig:
      - Dilation: 1x1
      - Filter: 1x1
      - Stride: 1x1
      - TensorAFormat: NCHW
      SetConstStrideB: [[2, 0]]

    - # BenchmarkProblemSizeGroup - Standard
      InitialSolutionParameters:
      BenchmarkCommonParameters:
        - KernelLanguage: ["Assembly"]
        - EdgeType: ["ShiftPtr"]
      ForkParameters:
        - PrefetchGlobalRead: [1]
        - PrefetchLocalRead: [0,1]
        - ThreadTile:
          - [ 8, 4 ]
          - [ 4, 8 ]
          - [ 8, 8 ]
        - WorkGroup:
          - [ 8, 16, 1 ]
          - [ 32, 8, 1 ]
          - [ 16, 16, 1 ]
          - [ 32, 16, 1 ]
          - [ 16, 32, 1 ]
        - GlobalSplitU: [1]
        - DepthU: [8,16]
        - GlobalReadVectorWidth: [1,-1]
        - StaggerU: [0,32]
        - StaggerUMapping: [0,1]
        - WorkGroupMapping: [1,8,64]
        - VectorWidth: [1]
        - FractionalLoad: [0,1]
        - PackBatchDims: [1]
        - VectorStore: [0]
      BenchmarkForkParameters:
      BenchmarkJoinParameters:
      BenchmarkFinalParameters:
        - ProblemSizes:
          - Exact: [ 289, 192, 64, 768 ]