TheRock/rocm-libraries/shared/tensile/Tensile/Configs/rocblas_hgemm_bufferload_limit.yaml

GlobalParameters:
  BoundsCheck: False
  KernelTime: True
  PrintSolutionRejectionReason: True
  # NumElementsToValidate: 65535
  # CMakeBuildType: Debug

BenchmarkProblems:

  ########################################
  # NN / TLUA = 1, TLUB = 0
  ########################################
  -
    - # ProblemType
      OperationType: GEMM
      DataType: h
      TransposeA: False
      TransposeB: False
      UseBeta: True
      Batched: True

    - # BenchmarkProblemSizeGroup
      InitialSolutionParameters:
      BenchmarkCommonParameters:
        - BufferLoad: [True]
        - BufferStore: [True]
        - KernelLanguage: ["Assembly"]
        - EdgeType: ["ShiftPtr"]
        - LoopTail: [True]
        - WorkGroupMapping: [8]
        - AssertFree0ElementMultiple: [2]
        - AssertFree1ElementMultiple: [2]
      ForkParameters:
        - VectorWidth: [2]
        - ThreadTile:
          - [ 2, 8 ]
        - WorkGroup:
          - [ 32, 2, 1 ]  # 64x16
          - [ 16, 4, 1 ]  # 32x32
          - [ 16, 8, 1 ]  # 32x64
        - DepthU: [32,64]
        - StaggerU: [0]
        - StaggerUStride: [256]
        - StaggerUMapping: [0]
      BenchmarkForkParameters:
      JoinParameters:
      BenchmarkJoinParameters:
      BenchmarkFinalParameters:
        # - ProblemSizes:
        #   - Range: [ [67108864], [32], [1], [32] ]  # strideA = 67108864, should fail on DU >= 32 (For tensorA-N)
        #   - Range: [ [33554432], [64], [1], [64] ]  # strideA = 33554432, should fail on DU >= 64 (For tensorA-N)
        #   - Range: [ [32], [32], [1], [67108864] ]  # strideB = 33554432, should fail on MT1 >= 32 (For tensorB-N)  # TODO - check MT32x32x32/MT32x64x32/MT32x32x64/MT32x64x64: FailAssert but PassValidate
        #   - Range: [ [64], [64], [1], [33554432] ]  # strideB = 33554432, should fail on MT1 >= 64 (For tensorB-N)  # TODO - check            MT32x64x32/           MT32x64x64: FailAssert but PassValidate

  # ########################################
  # # NT / TLUA = 1, TLUB = 1 # Pass
  # ########################################
  # -
  #   - # ProblemType
  #     OperationType: GEMM
  #     DataType: h
  #     TransposeA: False
  #     TransposeB: True
  #     UseBeta: True
  #     Batched: True

  #   - # BenchmarkProblemSizeGroup
  #     InitialSolutionParameters:
  #     BenchmarkCommonParameters:
  #       - BufferLoad: [True]
  #       - BufferStore: [True]
  #       - KernelLanguage: ["Assembly"]
  #       - EdgeType: ["ShiftPtr"]
  #       - LoopTail: [True]
  #       - WorkGroupMapping: [8]
  #       - AssertFree0ElementMultiple: [2]
  #       - AssertFree1ElementMultiple: [2]
  #     ForkParameters:
  #       - VectorWidth: [2]
  #       - ThreadTile:
  #         - [ 2, 2 ]
  #       - WorkGroup:
  #         - [ 16, 16, 1 ]  # 32x32
  #         - [ 16, 32, 1 ]  # 32x64
  #         - [ 32, 16, 1 ]  # 64x32
  #       - DepthU: [16,32,64]
  #       - StaggerU: [0]
  #       - StaggerUStride: [256]
  #       - StaggerUMapping: [0]
  #     BenchmarkForkParameters:
  #     JoinParameters:
  #     BenchmarkJoinParameters:
  #     BenchmarkFinalParameters:
  #       - ProblemSizes:
  #         - Range: [ [67108864], [32], [1], [32] ]  # strideA = 67108864, should fail on DU >= 32 (For tensorA-N)
  #         - Range: [ [33554432], [64], [1], [64] ]  # strideA = 33554432, should fail on DU >= 64 (For tensorA-N)
  #         - Range: [ [32], [67108864], [1], [32] ]  # strideB = 67108864, should fail on DU >= 32 (For tensorB-T)
  #         - Range: [ [64], [33554432], [1], [64] ]  # strideB = 33554432, should fail on DU >= 64 (For tensorB-T)

  # ########################################
  # # TN / TLUA = 0, TLUB = 0 TODO - seems all incorrect
  # ########################################
  # -
  #   - # ProblemType
  #     OperationType: GEMM
  #     DataType: h
  #     TransposeA: True
  #     TransposeB: False
  #     UseBeta: True
  #     Batched: True

  #   - # BenchmarkProblemSizeGroup
  #     InitialSolutionParameters:
  #     BenchmarkCommonParameters:
  #       - BufferLoad: [True]
  #       - BufferStore: [True]
  #       - KernelLanguage: ["Assembly"]
  #       - EdgeType: ["ShiftPtr"]
  #       - LoopTail: [True]
  #       - WorkGroupMapping: [8]
  #       - AssertFree0ElementMultiple: [2]
  #       - AssertFree1ElementMultiple: [2]
  #     ForkParameters:
  #       - VectorWidth: [2]
  #       - ThreadTile:
  #         - [ 2, 2 ]
  #       - WorkGroup:
  #         - [ 8, 8, 1 ]   # 16x16
  #         - [ 8, 16, 1 ]  # 16x32
  #         - [ 16, 8, 1 ]  # 32x16
  #         - [ 16, 16, 1 ] # 32x32
  #         - [ 16, 32, 1 ] # 32x64
  #         - [ 32, 16, 1 ] # 64x32
  #       - DepthU: [16]      # DU doesn't affect bufferload limit in TN
  #     BenchmarkForkParameters:
  #     JoinParameters:
  #     BenchmarkJoinParameters:
  #     BenchmarkFinalParameters:
  #       - ProblemSizes:
  #         - Range: [ [32], [32], [1], [67108864] ]  # strideA = 67108864, should fail on MT0 >= 32 (For tensorA-T)  # strideB = 67108864, should fail on MT1 >= 32 (For tensorB-N)
  #         - Range: [ [64], [64], [1], [33554432] ]  # strideA = 33554432, should fail on MT0 >= 64 (For tensorA-T)  # strideB = 33554432, should fail on MT1 >= 64 (For tensorB-N)

  # ########################################
  # # TT / TLUA = 0, TLUB = 1
  # ########################################
  # -
  #   - # ProblemType
  #     OperationType: GEMM
  #     DataType: h
  #     TransposeA: True
  #     TransposeB: True
  #     UseBeta: True
  #     Batched: True

  #   - # BenchmarkProblemSizeGroup
  #     InitialSolutionParameters:
  #     BenchmarkCommonParameters:
  #       - BufferLoad: [True]
  #       - BufferStore: [True]
  #       - KernelLanguage: ["Assembly"]
  #       - EdgeType: ["ShiftPtr"]
  #       - LoopTail: [True]
  #       - WorkGroupMapping: [8]
  #       - AssertFree0ElementMultiple: [2]
  #     ForkParameters:
  #       - VectorWidth: [2]
  #       - ThreadTile:
  #         - [ 2, 2 ]
  #       - WorkGroup:
  #         - [ 8, 8, 1 ]   # 16x16
  #         - [ 16, 8, 1 ]  # 32x16
  #         - [ 32, 8, 1 ]  # 64x16
  #       - DepthU: [16,32,64]
  #     BenchmarkForkParameters:
  #     JoinParameters:
  #     BenchmarkJoinParameters:
  #     BenchmarkFinalParameters:
  #       - ProblemSizes:
  #         - Range: [ [32], [32], [1], [67108864] ]  # strideA = 67108864, should fail on MT0 >= 32 (For tensorA-T)
  #         - Range: [ [64], [64], [1], [33554432] ]  # strideA = 33554432, should fail on MT0 >= 64 (For tensorA-T)  # TODO - check MT64x16x32: FailAssert but PassValidate
  #         - Range: [ [32], [67108864], [1], [32] ]  # strideB = 67108864, should fail on DU >= 32 (For tensorB-T)
  #         - Range: [ [64], [33554432], [1], [64] ]  # strideB = 33554432, should fail on DU >= 64 (For tensorB-T)

LibraryLogic:
    ScheduleName: "vega20"
    DeviceNames: ["Device 66a0", "Device 66a1", "Device 66a7", "Device 66af", "Vega 20"]
    ArchitectureName: "gfx906"

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

LibraryClient: