GlobalParameters:
  MinimumRequiredVersion: 4.1.0
  CMakeBuildType: Release
  LibraryPrintDebug: False
  NumElementsToValidate: 1000
  ValidationMaxToPrint: 16
  PinClocks: False
  SleepPercent: 0
  DataInitTypeA: 3
  DataInitTypeB: 3
  ClientArgs: "--stride_b 0"
  EnqueuesPerSync: 100 # avoid weird timing issues
  KernelTime: True

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

    - # BenchmarkProblemSizeGroup - Standard
      InitialSolutionParameters:
      BenchmarkCommonParameters:
        - KernelLanguage: ["Assembly"]
        - EdgeType: ["ShiftPtr"]
      ForkParameters:
        - ThreadTile:
          - [ 8, 8 ]
        - WorkGroup:
          - [ 8, 16, 1 ]
          - [ 16, 16, 1 ]
          - [ 32, 16, 1 ]
          - [ 16, 32, 1 ]
        - GlobalSplitU: [1]
        - DepthU: [16]
        - GlobalReadVectorWidth: [4]
        - StaggerU: [0,32]
        - StaggerUMapping: [0,1]
        - WorkGroupMapping: [1,8,64]
        - VectorWidth: [4]
        - FractionalLoad: [1]
      BenchmarkForkParameters:
      BenchmarkJoinParameters:
      BenchmarkFinalParameters:
        - ProblemSizes:
          #- Exact: [ 784, 128, 64, 512 ] # original
          #- Exact: [ 768, 128, 64, 512 ]
          - Exact: [ 768, 128, 64, 2048 ] # modified to user more challenging 8K stride
          #- Exact: [ 384, 128, 64, 2048 ]

  ########################################
  #
  #Cijk_Ailk_Bljk_S_MT128x128x16_GRVW04_SUM01_TT08_08_USFGRO0_VW04_WG16_16_01_WGM0

    # ./rocblas-bench -f gemm_strided_batched_ex --transposeA N --transposeB T -m 3136 -n 256 -k 64 --alpha 1 --a_type s --lda 3136 --stride_a 200704 --b_type s --ldb 256 --stride_b 0 --beta 0 --c_type s --ldc 3136 --stride_c 802816 --d_type s --ldd 3136 --stride_d 802816 --batch 64 --compute_type s --algo 0 --solution_index 0 --flags 0 --workspace_size 0x7ffc72d7d248
    #
    # ./rocblas-bench -f gemm_strided_batched_ex --transposeA N --transposeB N -m 784 -n 128 -k 512 --alpha 1 --a_type s --lda 784 --stride_a 401408 --b_type s --ldb 512 --stride_b 0 --beta 0 --c_type s --ldc 784 --stride_c 100352 --d_type s --ldd 784 --stride_d 100352 --batch 64 --compute_type s --algo 0 --solution_index 0 --flags 0 --workspace_size 0x7ffeb19c1c08
    # Cijk_Ailk_Bljk_SB_MT128x064x16_APM0_AF0EM01_AF1EM01_ASEM01_BL1_DTL0_EPS1_FL01_GRVW04_GSU01_ISA906_K1_KLA_LPA00_LPB00_MGWVW01_NLCA01_NLCB01_PGR1_PLR1_SNLL1_TT08_04_USFGRO0_VAW01_VW04_WG16_16_01_WGM01;