# A config to generate a fast kernel(s) for  DLRM model GEMM problem sizes
# problem sizes comes from Kaggle and Medium_4 data sets
# update parameters to produce best kernel(S)
# to keep number of kernel produced minimum
#   1) add new sections if certain paramters are more conducive to certain problem sizes
#   2) create separate sections for replacement kernel(s)

GlobalParameters:
  MinimumRequiredVersion: 4.14.0
  PrintLevel: 1
  ForceRedoBenchmarkProblems: True
  ForceRedoLibraryLogic: True
  ForceRedoLibraryClient: True
  CMakeBuildType: Release
  EnqueuesPerSync: 1
  SyncsPerBenchmark: 1
  LibraryPrintDebug: False
  NumElementsToValidate: 1
  ValidationMaxToPrint: 0
  ValidationPrintValids: False
  ShortNames: False
  MergeFiles: True
  Platform: 0
  Device: 0
  KernelTime: True
  PinClocks: False
  SleepPercent: 0
  PrintSolutionRejectionReason: True
  DataInitTypeA: 3
  DataInitTypeB: 3
  DataInitTypeBeta: 0
  DataInitTypeAlpha: 1
  PrintTensorA: 0
  PrintTensorB: 0
  PrintTensorD: 0


BenchmarkProblems:

  ########################################
  # TN -
  ########################################
  -
    - # ProblemType
      OperationType: GEMM
      DataType: s
      TransposeA: True
      TransposeB: False
      UseBeta: True
      Batched: True

  ########################################
  # TN - standard 2Blocks/Inst
  #    - medium_4 dataset
  ########################################

    - # BenchmarkProblemSizeGroup - Standard
      InitialSolutionParameters:
      BenchmarkCommonParameters:
        - KernelLanguage: ["Assembly"]
        - EdgeType: ["ShiftPtr"]
        - PrefetchLocalRead: [True]
        - PrefetchGlobalRead: [True]
      ForkParameters:
        - MatrixInstruction:
          - [16, 16, 4, 1]
        - ThreadTile:
          - [ 8, 16 ]
          - [ 8, 32 ]
        - WorkGroup:
          - [ 16, 16, 1 ]
        - TransposeLDS: [1]
        - LdsBlockSizePerPad: [128,256]
        - LdsPadA: [4]
        - LdsPadB: [4]
        - DepthU: [ 16 ]
        - VectorWidth: [4,2]
        - DisableVgprOverlapping: [True,False]
        - WaveSeparateGlobalReadA: [True,False]
        - WaveSeparateGlobalReadB: [True,False]
        - GlobalReadVectorWidth: [4]
        - InnerUnroll: [1,2]
        - ScheduleIterAlg: [1,3]
        - OptNoLoadLoop: [1,2]
        - StoreVectorWidth: [-1,2,4]
      BenchmarkForkParameters:
      JoinParameters:
      BenchmarkJoinParameters:
      BenchmarkFinalParameters:
        - ProblemSizes:
          - Exact: [ 4096,9520,1,1024,1056,1056,1088,1088]
          - Exact: [ 4096,8160,1,1024,1056,1056,1088,1088]

    - # BenchmarkProblemSizeGroup - Standard
      InitialSolutionParameters:
      BenchmarkCommonParameters:
        - KernelLanguage: ["Assembly"]
        - EdgeType: ["ShiftPtr"]
        - PrefetchLocalRead: [True]
        - PrefetchGlobalRead: [True]
      ForkParameters:
        - MatrixInstruction:
          - [16, 16, 1, 4]
        - ThreadTile:
          - [ 2, 16 ]
          - [ 1, 32 ]
          - [ 2, 32 ]
        - WorkGroup:
          - [ 64, 4, 1 ]
        - TransposeLDS: [1]
        - WaveSeparateGlobalReadA: [True,False]
        - WaveSeparateGlobalReadB: [True,False]
        - LdsBlockSizePerPad: [128,256]
        - LdsPadA: [2,4]
        - LdsPadB: [2,4]
        - DepthU: [ 16 ]
        - VectorWidth: [4,2]
        - DisableVgprOverlapping: [False]
        - GlobalReadVectorWidth: [4]
        - InnerUnroll: [1,2]
        - ScheduleIterAlg: [1,3]
        - OptNoLoadLoop: [1,2]
        - StoreVectorWidth: [-1,2,4]
      BenchmarkForkParameters:
      JoinParameters:
      BenchmarkJoinParameters:
      BenchmarkFinalParameters:
        - ProblemSizes:
          - Exact: [ 4096,9520,1,1024,1056,1056,1088,1088]
          - Exact: [ 4096,8160,1,1024,1056,1056,1088,1088]

    - # BenchmarkProblemSizeGroup - Standard
      InitialSolutionParameters:
      BenchmarkCommonParameters:
        - KernelLanguage: ["Assembly"]
        - EdgeType: ["ShiftPtr"]
        - PrefetchLocalRead: [True]
        - PrefetchGlobalRead: [True]
      ForkParameters:
        - MatrixInstruction:
          - [16, 16, 4, 1]
        - ThreadTile:
          - [ 8, 16 ]
          - [ 8, 32 ]
        - WorkGroup:
          - [ 16, 16, 1 ]
        - TransposeLDS: [0]
        - DepthU: [ 16 ]
        - VectorWidth: [4,2]
        - DisableVgprOverlapping: [False]
        - GlobalReadVectorWidth: [4]
        - InnerUnroll: [1,2]
        - ScheduleIterAlg: [1,3]
        - OptNoLoadLoop: [1,2]
        - StoreVectorWidth: [-1,2,4]
      BenchmarkForkParameters:
      JoinParameters:
      BenchmarkJoinParameters:
      BenchmarkFinalParameters:
        - ProblemSizes:
          - Exact: [ 4096,9520,1,1024,1056,1056,1088,1088]
          - Exact: [ 4096,8160,1,1024,1056,1056,1088,1088]

    - # BenchmarkProblemSizeGroup - Standard
      InitialSolutionParameters:
      BenchmarkCommonParameters:
        - KernelLanguage: ["Assembly"]
        - EdgeType: ["ShiftPtr"]
        - PrefetchLocalRead: [True]
        - PrefetchGlobalRead: [True]
      ForkParameters:
        - MatrixInstruction:
          - [16, 16, 1, 4]
        - ThreadTile:
          - [ 2, 16 ]
          - [ 1, 32 ]
          - [ 2, 32 ]
        - WorkGroup:
          - [ 64, 4, 1 ]
        - TransposeLDS: [0]
        - DepthU: [ 16 ]
        - VectorWidth: [4,2]
        - DisableVgprOverlapping: [False]
        - GlobalReadVectorWidth: [4]
        - InnerUnroll: [1,2]
        - ScheduleIterAlg: [1,3]
        - OptNoLoadLoop: [1,2]
        - StoreVectorWidth: [-1,2,4]
      BenchmarkForkParameters:
      JoinParameters:
      BenchmarkJoinParameters:
      BenchmarkFinalParameters:
        - ProblemSizes:
          - Exact: [ 4096,9520,1,1024,1056,1056,1088,1088]
          - Exact: [ 4096,8160,1,1024,1056,1056,1088,1088]

    - # BenchmarkProblemSizeGroup - Standard
      InitialSolutionParameters:
      BenchmarkCommonParameters:
        - KernelLanguage: ["Assembly"]
        - EdgeType: ["ShiftPtr"]
        - PrefetchLocalRead: [True]
        - PrefetchGlobalRead: [True]
      ForkParameters:
        - MatrixInstruction:
          - [32, 32, 2, 1]
        - ThreadTile:
          - [ 2, 32 ]
          - [ 4, 32 ]
        - WorkGroup:
          - [ 16, 16, 1 ]
        - TransposeLDS: [1]
        - LdsBlockSizePerPad: [128,256]
        - LdsPadA: [4]
        - LdsPadB: [4]
        - DepthU: [ 16 ]
        - VectorWidth: [4,2]
        - DisableVgprOverlapping: [False]
        - GlobalReadVectorWidth: [4]
        - WaveSeparateGlobalReadA: [True,False]
        - WaveSeparateGlobalReadB: [True,False]
        - InnerUnroll: [1,2]
        - ScheduleIterAlg: [1,3]
        - OptNoLoadLoop: [1,2]
        - StoreVectorWidth: [-1,2,4]
      BenchmarkForkParameters:
      JoinParameters:
      BenchmarkJoinParameters:
      BenchmarkFinalParameters:
        - ProblemSizes:
          - Exact: [ 4096,9520,1,1024,1056,1056,1088,1088]
          - Exact: [ 4096,8160,1,1024,1056,1056,1088,1088]

    - # BenchmarkProblemSizeGroup - Standard
      InitialSolutionParameters:
      BenchmarkCommonParameters:
        - KernelLanguage: ["Assembly"]
        - EdgeType: ["ShiftPtr"]
        - PrefetchLocalRead: [True]
        - PrefetchGlobalRead: [True]
      ForkParameters:
        - MatrixInstruction:
          - [32, 32, 1, 2]
        - ThreadTile:
          - [ 1, 32 ]
          - [ 2, 32 ]
        - WorkGroup:
          - [ 64, 4, 1 ]
        - TransposeLDS: [1]
        - LdsBlockSizePerPad: [128,256]
        - LdsPadA: [2,4]
        - LdsPadB: [2,4]
        - DepthU: [ 16 ]
        - VectorWidth: [4,2]
        - DisableVgprOverlapping: [True,False]
        - GlobalReadVectorWidth: [4]
        - InnerUnroll: [1,2]
        - WaveSeparateGlobalReadA: [True,False]
        - WaveSeparateGlobalReadB: [True,False]
        - ScheduleIterAlg: [1,3]
        - OptNoLoadLoop: [1,2]
        - StoreVectorWidth: [-1,2,4]
      BenchmarkForkParameters:
      JoinParameters:
      BenchmarkJoinParameters:
      BenchmarkFinalParameters:
        - ProblemSizes:
          - Exact: [ 4096,9520,1,1024,1056,1056,1088,1088]
          - Exact: [ 4096,8160,1,1024,1056,1056,1088,1088]

    - # BenchmarkProblemSizeGroup - Standard
      InitialSolutionParameters:
      BenchmarkCommonParameters:
        - KernelLanguage: ["Assembly"]
        - EdgeType: ["ShiftPtr"]
        - PrefetchLocalRead: [True]
        - PrefetchGlobalRead: [True]
      ForkParameters:
        - MatrixInstruction:
          - [32, 32, 2, 1]
        - ThreadTile:
          - [ 2, 32 ]
          - [ 4, 32 ]
        - WorkGroup:
          - [ 16, 16, 1 ]
        - TransposeLDS: [0]
        - DepthU: [ 16 ]
        - VectorWidth: [4,2]
        - DisableVgprOverlapping: [True,False]
        - GlobalReadVectorWidth: [4]
        - InnerUnroll: [1,2]
        - ScheduleIterAlg: [1,3]
        - OptNoLoadLoop: [1,2]
        - StoreVectorWidth: [-1,2,4]
      BenchmarkForkParameters:
      JoinParameters:
      BenchmarkJoinParameters:
      BenchmarkFinalParameters:
        - ProblemSizes:
          - Exact: [ 4096,9520,1,1024,1056,1056,1088,1088]
          - Exact: [ 4096,8160,1,1024,1056,1056,1088,1088]

    - # BenchmarkProblemSizeGroup - Standard
      InitialSolutionParameters:
      BenchmarkCommonParameters:
        - KernelLanguage: ["Assembly"]
        - EdgeType: ["ShiftPtr"]
        - PrefetchLocalRead: [True]
        - PrefetchGlobalRead: [True]
      ForkParameters:
        - MatrixInstruction:
          - [32, 32, 1, 2]
        - ThreadTile:
          - [ 1, 32 ]
          - [ 2, 32 ]
        - WorkGroup:
          - [ 64, 4, 1 ]
        - TransposeLDS: [0]
        - DepthU: [ 16 ]
        - VectorWidth: [4,2]
        - DisableVgprOverlapping: [True,False]
        - GlobalReadVectorWidth: [4]
        - InnerUnroll: [1,2]
        - ScheduleIterAlg: [1,3]
        - OptNoLoadLoop: [1,2]
        - StoreVectorWidth: [-1,2,4]
      BenchmarkForkParameters:
      JoinParameters:
      BenchmarkJoinParameters:
      BenchmarkFinalParameters:
        - ProblemSizes:
          - Exact: [ 4096,9520,1,1024,1056,1056,1088,1088]
          - Exact: [ 4096,8160,1,1024,1056,1056,1088,1088]

    - # BenchmarkProblemSizeGroup - Standard
      InitialSolutionParameters:
      BenchmarkCommonParameters:
        - KernelLanguage: ["Assembly"]
        - EdgeType: ["ShiftPtr"]
        - PrefetchLocalRead: [True]
        - PrefetchGlobalRead: [True]
      ForkParameters:
        - MatrixInstruction:
          - [16, 16, 4, 1]
        - ThreadTile:
          - [ 1, 32 ]
          - [ 2, 32 ]
        - WorkGroup:
          - [ 64, 4, 1 ]
        - TransposeLDS: [1]
        - WaveSeparateGlobalReadA: [True,False]
        - WaveSeparateGlobalReadB: [True,False]
        - LdsBlockSizePerPad: [128,256]
        - LdsPadA: [4]
        - LdsPadB: [4]
        - DepthU: [ 16 ]
        - VectorWidth: [4,2]
        - DisableVgprOverlapping: [False]
        - GlobalReadVectorWidth: [4]
        - InnerUnroll: [1,2]
        - ScheduleIterAlg: [1,3]
        - OptNoLoadLoop: [1,2]
        - StoreVectorWidth: [-1,2,4]
      BenchmarkForkParameters:
      JoinParameters:
      BenchmarkJoinParameters:
      BenchmarkFinalParameters:
        - ProblemSizes:
          - Exact: [ 4096,9520,1,1024,1056,1056,1088,1088]
          - Exact: [ 4096,8160,1,1024,1056,1056,1088,1088]

########################################
LibraryLogic:
    ScheduleName: "arcturus"
    DeviceNames: ["Device 7380", "Device 7388", "Device 738c", "Device 7390"]
    ArchitectureName: "gfx908"