# A config to generate a fast kernel for BERT model GEMMs
# 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)
#   3) create separate sections for problem sizes belongs to specific model parameter sizes
GlobalParameters:
  MinimumRequiredVersion: 4.14.0
  PrintLevel: 1
  ForceRedoBenchmarkProblems: True
  ForceRedoLibraryLogic: True
  ForceRedoLibraryClient: True
  CMakeBuildType: Release
  EnqueuesPerSync: 1
  SyncsPerBenchmark: 1
  LibraryPrintDebug: False
  NumElementsToValidate: 1
  ValidationMaxToPrint: 4
  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 -
  # max_sequenth_lenth = 128
  ########################################
  -
    - # ProblemType
      OperationType: GEMM
      DataType: s
      TransposeA: True
      TransposeB: False
      UseBeta: True
      Batched: True

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

    - # BenchmarkProblemSizeGroup - Standard
      InitialSolutionParameters:
      BenchmarkCommonParameters:
        - KernelLanguage: ["Assembly"]
        - EdgeType: ["ShiftPtr"]
        - PrefetchLocalRead: [True]
        - PrefetchGlobalRead: [True]
      ForkParameters:
        - MatrixInstruction:
          - [32, 32, 1, 2]
          #- [32, 32, 2, 1]
        - ThreadTile:
          - [ 2, 32 ]
          - [ 1, 64 ]
        - WorkGroup:
          - [ 64, 4, 1 ]
          - [ 32, 8, 1 ]
        - WorkGroupMapping: [1]
        #- DirectToLds: [0]
        - TransposeLDS: [1]
        - LdsBlockSizePerPad: [128]
        - LdsPadA: [4]
        - LdsPadB: [4]
        - GlobalSplitU: [1]
        - DepthU: [ 16 ]
        - VectorWidth: [1]
        - GlobalReadVectorWidth: [4]
        - InnerUnroll: [1]
        - ScheduleIterAlg: [1]
        - OptNoLoadLoop: [0]
      BenchmarkForkParameters:
      JoinParameters:
      BenchmarkJoinParameters:
      BenchmarkFinalParameters:
        - ProblemSizes:
          - Exact: [ 1024,3072,1,1024,1056,1056,1088,1056 ]
          - Exact: [ 1024,3072,1,1024,1056,1056,1088,1056 ]
          - Exact: [ 1024,4096,1,4096,1056,1056,4128,4160 ]
          - Exact: [ 4096,4096,1,1024,4128,4128,1088,1056 ]
          - Exact: [ 1024,2048,1,1024,1056,1056,1088,1056 ]
          - Exact: [ 1024,3072,1,3072,1056,1056,3136,3104 ]
          - Exact: [ 3072,3072,1,1024,3104,3104,1088,1056 ]
          - Exact: [ 1024,2048,1,3072,1056,1056,3136,3104 ]
          - Exact: [ 3072,2048,1,1024,3104,3104,1088,1056 ]
          - Exact: [ 1024,512,1,1024,1056,1056,1088,1056 ]
          - Exact: [ 1024,512,1,3072,1056,1056,3136,3104 ]
          - Exact: [ 512,512,1,64,544,544,128,96 ]
          - Exact: [ 64,512,1,512,92,96,576,544 ]
          - Exact: [ 3072,512,1,1024,3104,3104,1088,1056 ]
          - Exact: [ 30528,4096,1,1024,30554,30554,1088,1056 ]
          - Exact: [ 30528,2048,1,1024,30554,30554,1088,1056 ]

  ########################################
  # TN - standard 1Blocks/Inst
  ########################################

    - # BenchmarkProblemSizeGroup - Standard
      InitialSolutionParameters:
      BenchmarkCommonParameters:
        - KernelLanguage: ["Assembly"]
        - EdgeType: ["ShiftPtr"]
        - PrefetchLocalRead: [True]
        - PrefetchGlobalRead: [True]
      ForkParameters:
        - MatrixInstruction:
          - [32, 32, 2, 1]
        - ThreadTile:
          - [ 2, 32 ]
          - [ 1, 64 ]
        - WorkGroup:
          - [ 64, 4, 1 ]
          - [ 32, 8, 1 ]
        - WorkGroupMapping: [1]
        #- DirectToLds: [0]
        - TransposeLDS: [1]
        - LdsBlockSizePerPad: [128]
        - LdsPadA: [4]
        - LdsPadB: [4]
        - GlobalSplitU: [1]
        - DepthU: [ 16 ]
        - VectorWidth: [1]
        - GlobalReadVectorWidth: [4]
        - InnerUnroll: [1]
        - ScheduleIterAlg: [1]
        - OptNoLoadLoop: [0]
      BenchmarkForkParameters:
      JoinParameters:
      BenchmarkJoinParameters:
      BenchmarkFinalParameters:
        - ProblemSizes:
          - Exact: [ 1024,3072,1,1024,1056,1056,1088,1056 ]
          - Exact: [ 1024,3072,1,1024,1056,1056,1088,1056 ]
          - Exact: [ 1024,4096,1,4096,1056,1056,4128,4160 ]
          - Exact: [ 4096,4096,1,1024,4128,4128,1088,1056 ]
          - Exact: [ 1024,2048,1,1024,1056,1056,1088,1056 ]
          - Exact: [ 1024,3072,1,3072,1056,1056,3136,3104 ]
          - Exact: [ 3072,3072,1,1024,3104,3104,1088,1056 ]
          - Exact: [ 1024,2048,1,3072,1056,1056,3136,3104 ]
          - Exact: [ 3072,2048,1,1024,3104,3104,1088,1056 ]
          - Exact: [ 1024,512,1,1024,1056,1056,1088,1056 ]
          - Exact: [ 1024,512,1,3072,1056,1056,3136,3104 ]
          - Exact: [ 512,512,1,64,544,544,128,96 ]
          - Exact: [ 64,512,1,512,92,96,576,544 ]
          - Exact: [ 3072,512,1,1024,3104,3104,1088,1056 ]
          - Exact: [ 30528,4096,1,1024,30554,30554,1088,1056 ]
          - Exact: [ 30528,2048,1,1024,30554,30554,1088,1056 ]

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

  ########################################
  # NN - standard 2Blocks/Inst
  ########################################

    - # BenchmarkProblemSizeGroup - Standard
      InitialSolutionParameters:
      BenchmarkCommonParameters:
        - KernelLanguage: ["Assembly"]
        - EdgeType: ["ShiftPtr"]
        - PrefetchLocalRead: [True]
        - PrefetchGlobalRead: [True]
      ForkParameters:
        - MatrixInstruction:
          - [32, 32, 1, 2]
        - ThreadTile:
          - [ 2, 32 ]
          - [ 1, 64 ]
        - WorkGroup:
          - [ 64, 4, 1 ]
          - [ 32, 8, 1 ]
        - WorkGroupMapping: [1]
        #- DirectToLds: [0]
        - TransposeLDS: [0]
        - GlobalSplitU: [1]
        - DepthU: [ 16 ]
        - VectorWidth: [1]
        - GlobalReadVectorWidth: [4]
        - InnerUnroll: [2]
        - ScheduleIterAlg: [1]
        - OptNoLoadLoop: [0]
      BenchmarkForkParameters:
      JoinParameters:
      BenchmarkJoinParameters:
      BenchmarkFinalParameters:
        - ProblemSizes:
          - Exact: [ 1024,3072,1,1024,1056,1056,1088,1056 ]
          - Exact: [ 1024,3072,1,1024,1056,1056,1088,1056 ]
          - Exact: [ 1024,4096,1,4096,1056,1056,4128,4160 ]
          - Exact: [ 4096,4096,1,1024,4128,4128,1088,1056 ]
          - Exact: [ 1024,2048,1,1024,1056,1056,1088,1056 ]
          - Exact: [ 1024,3072,1,3072,1056,1056,3136,3104 ]
          - Exact: [ 3072,3072,1,1024,3104,3104,1088,1056 ]
          - Exact: [ 1024,2048,1,3072,1056,1056,3136,3104 ]
          - Exact: [ 3072,2048,1,1024,3104,3104,1088,1056 ]
          - Exact: [ 1024,512,1,1024,1056,1056,1088,1056 ]
          - Exact: [ 1024,512,1,3072,1056,1056,3136,3104 ]
          - Exact: [ 512,512,1,64,544,544,128,96 ]
          - Exact: [ 64,512,1,512,92,96,576,544 ]
          - Exact: [ 3072,512,1,1024,3104,3104,1088,1056 ]
          - Exact: [ 30528,4096,1,1024,30554,30554,1088,1056 ]
          - Exact: [ 30528,2048,1,1024,30554,30554,1088,1056 ]

  ########################################
  # NN - standard 1Blocks/Inst
  ########################################

    - # BenchmarkProblemSizeGroup - Standard
      InitialSolutionParameters:
      BenchmarkCommonParameters:
        - KernelLanguage: ["Assembly"]
        - EdgeType: ["ShiftPtr"]
        - PrefetchLocalRead: [True]
        - PrefetchGlobalRead: [True]
      ForkParameters:
        - MatrixInstruction:
          - [32, 32, 2, 1]
        - ThreadTile:
          - [ 2, 32 ]
          - [ 1, 64 ]
        - WorkGroup:
          - [ 64, 4, 1 ]
          - [ 32, 8, 1 ]
        - WorkGroupMapping: [1]
        #- DirectToLds: [0]
        - TransposeLDS: [1]
        - LdsBlockSizePerPad: [128]
        - LdsPadA: [4]
        - LdsPadB: [4]
        - GlobalSplitU: [1]
        - DepthU: [ 16 ]
        - VectorWidth: [1]
        - GlobalReadVectorWidth: [4]
        - InnerUnroll: [2]
        - ScheduleIterAlg: [1]
        - OptNoLoadLoop: [0]
      BenchmarkForkParameters:
      JoinParameters:
      BenchmarkJoinParameters:
      BenchmarkFinalParameters:
        - ProblemSizes:
          - Exact: [ 1024,3072,1,1024,1056,1056,1088,1056 ]
          - Exact: [ 1024,3072,1,1024,1056,1056,1088,1056 ]
          - Exact: [ 1024,4096,1,4096,1056,1056,4128,4160 ]
          - Exact: [ 4096,4096,1,1024,4128,4128,1088,1056 ]
          - Exact: [ 1024,2048,1,1024,1056,1056,1088,1056 ]
          - Exact: [ 1024,3072,1,3072,1056,1056,3136,3104 ]
          - Exact: [ 3072,3072,1,1024,3104,3104,1088,1056 ]
          - Exact: [ 1024,2048,1,3072,1056,1056,3136,3104 ]
          - Exact: [ 3072,2048,1,1024,3104,3104,1088,1056 ]
          - Exact: [ 1024,512,1,1024,1056,1056,1088,1056 ]
          - Exact: [ 1024,512,1,3072,1056,1056,3136,3104 ]
          - Exact: [ 512,512,1,64,544,544,128,96 ]
          - Exact: [ 64,512,1,512,92,96,576,544 ]
          - Exact: [ 3072,512,1,1024,3104,3104,1088,1056 ]
          - Exact: [ 30528,4096,1,1024,30554,30554,1088,1056 ]
          - Exact: [ 30528,2048,1,1024,30554,30554,1088,1056 ]

  ########################################
  # TN - standard 2Blocks/Inst
  # BERT BIG model
  # batch size 1, 4, 6
  # seq_size : 512
  ########################################

    - # BenchmarkProblemSizeGroup - Standard
      InitialSolutionParameters:
      BenchmarkCommonParameters:
        - KernelLanguage: ["Assembly"]
        - EdgeType: ["ShiftPtr"]
        - PrefetchLocalRead: [True]
        - PrefetchGlobalRead: [True]
      ForkParameters:
        - MatrixInstruction:
          - [32, 32, 1, 2]
          - [32, 32, 2, 1]
          - [16, 16, 4, 1]
          - [16, 16, 1, 4]
        - ThreadTile:
          - [ 2, 32 ]
          - [ 1, 64 ]
          - [ 2, 16 ]
          - [ 4, 32 ]
        - WorkGroup:
          - [ 64, 4, 1 ]
          - [ 32, 8, 1 ]
          - [ 16, 16, 1 ]
        - WorkGroupMapping: [1]
        #- DirectToLds: [0]
        - TransposeLDS: [1]
        - LdsBlockSizePerPad: [128]
        - LdsPadA: [4]
        - LdsPadB: [4]
        - GlobalSplitU: [1]
        - DepthU: [ 16 ]
        - VectorWidth: [1]
        - GlobalReadVectorWidth: [4]
        - InnerUnroll: [1]
        - ScheduleIterAlg: [1]
        - OptNoLoadLoop: [0]
      BenchmarkForkParameters:
      JoinParameters:
      BenchmarkJoinParameters:
      BenchmarkFinalParameters:
        - ProblemSizes:
          - Exact: [ 1024,2048,1,3072,1056,1056,3136,3104 ]
          - Exact: [ 1024,3072,1,1024,1056,1056,1088,1056 ]
          - Exact: [ 1024,512,1,1024,1056,1056,1088,1056 ]
          - Exact: [ 3072,3072,1,1024,3104,3104,1088,1056 ]
          - Exact: [ 3072,512,1,1024,3104,3104,1088,1056 ]
          - Exact: [ 512,512,1,64,544,544,128,96 ]
          - Exact: [ 1024,80,1,1024,1056,1056,1088,1056 ]
          - Exact: [ 30522,120,1,1024,30554,30554,1088,1056 ]
          - Exact: [ 30522,20,1,1024,30554,30554,1088,1056 ]
          - Exact: [ 30522,80,1,1024,30554,30554,1088,1056 ]
          - Exact: [ 30522,160,1,1024,30554,30554,1088,1056 ]
          - Exact: [ 30522,120,1,1024,30554,30554,1088,1056 ]
          - Exact: [ 1024,120,1,1024,30554,30554,1088,1056 ]
          - Exact: [ 1024,20,1,1024,30554,30554,1088,1056 ]

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