# A config to generate a fast kernel for large DGEMMs
GlobalParameters:
  EnqueuesPerSync: 1
  LibraryPrintDebug: False
  NumElementsToValidate: 0
  ValidationMaxToPrint: 4
  ValidationPrintValids: False
  KernelTime: True
  #PinClocks: True
  SleepPercent: 0
  DataInitTypeBeta : 0
  PrintSolutionRejectionReason: 1
  ClientArgs: "--stride_b 0"

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

    - # BenchmarkProblemSizeGroup - Standard
      InitialSolutionParameters:
      BenchmarkCommonParameters:
        - KernelLanguage: ["Assembly"]
        - EdgeType: ["ShiftPtr"]
      ForkParameters:
        - FractionalLoad: [1]
        - PrefetchGlobalRead: [ False, True ]
        - PrefetchLocalRead: [ True]
        - ThreadTile:
          - [ 4, 4 ]
          - [ 7, 8 ]
          - [ 9, 8 ]
          - [ 5, 8 ]
          - [ 8, 4 ]
          - [ 8, 8 ]
          - [ 10, 6 ]
          - [ 6, 10 ]
        - WorkGroup:
          #- [ 16, 8, 2 ]
          #- [ 16, 4, 4 ]
          - [ 16,  8, 1 ]
          - [ 8, 32, 1 ]
          - [ 16, 16, 1 ]
          - [ 32,  8, 1 ]
        - WorkGroupMapping: [1,8,64]
        - DepthU: [ 8,16,32 ]
        - VectorWidth: [-1] # use biggest available
        - GlobalReadVectorWidth: [1,-1] # 1 allows D2LDS
        - LdsPadA: [0, -1 ]
        - LdsPadB: [0, -1 ]
        - AssertSummationElementMultiple: [4]
        - AssertFree0ElementMultiple: [4]
        - AssertFree1ElementMultiple: [4]
      BenchmarkForkParameters:
      BenchmarkJoinParameters:
      BenchmarkFinalParameters:
        - ProblemSizes:
          # This example is from ResNet50 1x1 conv layer batch64, suffers from granularity loss:
          - Exact: [192, 1024, 64, 256 ]  # much higher perf with smaller size
          - Exact: [196, 1024, 64, 256 ]  # granularity loss


  ########################################