GlobalParameters:
  EnqueuesPerSync: 30
  LibraryPrintDebug: False
  NumElementsToValidate: 000
  KernelTime: True
  SleepPercent: 0
  DataInitTypeBeta : 0
  PrintSolutionRejectionReason: 0
  PrintWinnersOnly: 1

BenchmarkProblems:
  -
    - # ProblemType
      OperationType: TensorContraction
      IndexAssignmentsA: [ 3, 0, 2]
      IndexAssignmentsB: [ 3, 1, 2]
      DataType: s
      UseBeta: False
      NumIndicesC: 2

    - # BenchmarkProblemSizeGroup - Standard
      InitialSolutionParameters:
      BenchmarkCommonParameters:
        - KernelLanguage: ["Assembly"]
        - EdgeType: ["ShiftPtr"]
      ForkParameters:
        - PrefetchLocalRead: [False,True]
        - PrefetchGlobalRead: [False,True]
        - ThreadTile:
          - [ 4, 4 ]
          - [ 8, 4 ]
          - [ 4, 8 ]
          - [ 8, 8 ]
        - WorkGroup:
          #- [ 4, 16, 1 ]
          - [ 8, 8, 1 ]
          - [ 8, 16, 1 ]
          - [ 8, 32, 1 ]
            #- [ 16, 32, 1 ]
            #- [ 32, 16, 1 ]
            #- [ 16, 16, 1 ]
        - WorkGroupMapping: [1,8,64]
          #- StaggerU: [0,32]  # bozo
        - DepthU: [ 8,16 ]
        - VectorWidth: [1,-1]
        - FractionalLoad: [0]
        - AssertFree0ElementMultiple: [4]
        - AssertFree1ElementMultiple: [4]
        - GlobalSplitU: [1,2,4,6,8,12]
          #- DisableKernelPieces: [1]
          #- DisableAtomicFail: [1]
      BenchmarkForkParameters:
      BenchmarkJoinParameters:
      BenchmarkFinalParameters:
        - ProblemSizes:
          # ./rocblas-bench -f gemm_ex --transposeA T --transposeB N -m 1024 -n 256 -k 196 --alpha 1 --a_type f32_r --lda 196 --b_type f32_r --ldb 196 --beta 1 --c_type f32_r --ldc 1024 --d_type f32_r --ldd 1024 --compute_type f32_r --algo 0 --solution_index 0 --flags 0
          - Exact: [ 1024, 256, 64,  196]
          - Exact: [ 1024, 256, 2,  6272]

          # ./rocblas-bench -f gemm_ex --transposeA T --transposeB N -m 256 -n 1024 -k 196 --alpha 1 --a_type f32_r --lda 196 --b_type f32_r --ldb 196 --beta 1 --c_type f32_r --ldc 256 --d_type f32_r --ldd 256 --compute_type f32_r --algo 0 --solution_index 0 --flags 0
          - Exact: [ 256, 1024, 64,  196]
          - Exact: [ 256, 1024, 2,  6272]

            # ./rocblas-bench -f gemm_ex --transposeA T --transposeB N -m 512 -n 2048 -k 49 --alpha 1 --a_type f32_r --lda 49 --b_type f32_r --ldb 49 --beta 1 --c_type f32_r --ldc 512 --d_type f32_r --ldd 512 --compute_type f32_r --algo 0 --solution_index 0 --flags 0
          - Exact: [ 512, 2048, 64,  49]
          - Exact: [ 512, 2048, 2,  1568]