TheRock/rocm-libraries/projects/hipblaslt/tensilelite/Tensile/Tests/common/gemm/gsuasb.yaml

TestParameters:
  marks: [skip-gfx1100, skip-gfx1101, skip-gfx1102, skip-gfx1200, skip-gfx1201, skip-gfx1250] # not supported by arch

GlobalParameters:
  SyncsPerBenchmark: 0
  MinimumRequiredVersion: 5.0.0
  NumWarmups: 0
  SleepPercent: 50
  EnqueuesPerSync: 1
  NumElementsToValidate: -1
  DataInitTypeE: 16
  DataInitTypeBeta: 0
  DataInitTypeAlpha: 1
  DataInitTypeScaleAlphaVec: 2
  Device: 0
  CSVExportWinner: true
  CSVMergeSameProblemID: true
  MaxWorkspaceSize: 3355443200
  PrintSolutionRejectionReason: True

BenchmarkProblems:
  ########################################
  # NN - standard
  ########################################
  -
    - # ProblemType
      OperationType: GEMM
      # DataTypeA: f8
      DataType: h
      DestDataType: h
      ComputeDataType: s
      HighPrecisionAccumulate: True
      TransposeA: False
      TransposeB: False
      UseBeta: True
      Batched: True
      BiasDataTypeList: [s]
      UseBias:       0
      UseE:          False
      Activation:    False
      UseScaleAB: ""
      # SupportUserArgs: True
    - # BenchmarkProblemSizeGroup - Standard
      InitialSolutionParameters:
      BenchmarkCommonParameters:
        - KernelLanguage: ["Assembly"]
      ForkParameters:
        - MatrixInstruction:
          - [16, 16, 16, 1, 1, 2, 2, 2, 2]
        - AssertFree0ElementMultiple: [1]
        - DepthU: [32]
        - ScheduleIterAlg: [3]
        - InnerUnroll: [1]
        - ExpandPointerSwap: [true]
        - TransposeLDS: [1]
        - LdsPadA: [-1]
        - LdsPadB: [-1]
        - StaggerU: [4]
        - WaveSeparateGlobalReadB: [1]
        - 1LDSBuffer: [-1]
        - GlobalReadPerMfma: [1.0]
        - LocalWritePerMfma: [-1]
        - StoreVectorWidth: [-1]
        - SourceSwap: [true]
        - GlobalSplitU: [1, 2, 9, 17]
        - GlobalSplitUAlgorithm: ["MultipleBuffer", "SingleBuffer", "MultipleBufferSingleKernel"]
      BenchmarkJoinParameters:
      BenchmarkFinalParameters:
        - ProblemSizes:
          - Exact: [ 2048, 128, 1, 1024 ]
        - BiasTypeArgs: ['s']
        - ActivationArgs:
          - [Enum: none]

  ########################################
  # NN - Activation
  ########################################
  -
    - # ProblemType
      OperationType: GEMM
      # DataTypeA: f8
      DataType: h
      DestDataType: h
      ComputeDataType: s
      HighPrecisionAccumulate: True
      TransposeA: False
      TransposeB: False
      UseBeta: True
      Batched: True
      BiasDataTypeList: [s]
      UseBias:       1
      UseE:          False
      Activation:    True
      ActivationHPA: True
      UseScaleAB: "Scalar"
      # SupportUserArgs: True
    - # BenchmarkProblemSizeGroup - Standard
      InitialSolutionParameters:
      BenchmarkCommonParameters:
        - KernelLanguage: ["Assembly"]
      ForkParameters:
        - MatrixInstruction:
          - [16, 16, 16, 1, 1, 2, 2, 2, 2]
        - AssertFree0ElementMultiple: [1]
        - DepthU: [32]
        - ScheduleIterAlg: [3]
        - InnerUnroll: [1]
        - ExpandPointerSwap: [true]
        - TransposeLDS: [1]
        - LdsPadA: [-1]
        - LdsPadB: [-1]
        - StaggerU: [4]
        - WaveSeparateGlobalReadB: [1]
        - 1LDSBuffer: [-1]
        - GlobalReadPerMfma: [1.0]
        - LocalWritePerMfma: [-1]
        - StoreVectorWidth: [-1]
        - SourceSwap: [true]
        - GlobalSplitU: [1, 2, 9, 17]
        - GlobalSplitUAlgorithm: ["MultipleBuffer", "MultipleBufferSingleKernel"]
      BenchmarkJoinParameters:
      BenchmarkFinalParameters:
        - ProblemSizes:
          - Exact: [ 2048, 128, 1, 1024 ]
        - BiasTypeArgs: ['s']
        - ActivationArgs:
          - [Enum: none]

  ########################################
  # NN - AdaptiveGemmGSUA
  ########################################
  -
    - # ProblemType
      OperationType: GEMM
      # DataTypeA: f8
      DataType: h
      DestDataType: h
      ComputeDataType: s
      HighPrecisionAccumulate: True
      TransposeA: False
      TransposeB: False
      UseBeta: True
      Batched: True
      BiasDataTypeList: [s]
      UseBias:       1
      UseE:          False
      Activation:    True
      ActivationHPA: True
      UseScaleAB: "Scalar"
      # SupportUserArgs: True
    - # BenchmarkProblemSizeGroup - Standard
      InitialSolutionParameters:
      BenchmarkCommonParameters:
        - KernelLanguage: ["Assembly"]
      ForkParameters:
        - MatrixInstruction:
          - [16, 16, 16, 1, 1, 2, 2, 2, 2]
        - AssertFree0ElementMultiple: [1]
        - DepthU: [32]
        - ScheduleIterAlg: [3]
        - InnerUnroll: [1]
        - ExpandPointerSwap: [true]
        - TransposeLDS: [1]
        - LdsPadA: [-1]
        - LdsPadB: [-1]
        - StaggerU: [4]
        - WaveSeparateGlobalReadB: [1]
        - 1LDSBuffer: [-1]
        - GlobalReadPerMfma: [1.0]
        - LocalWritePerMfma: [-1]
        - StoreVectorWidth: [-1]
        - SourceSwap: [true]
        - AdaptiveGemmGSUA: [1]
        - GlobalSplitU: [1, 2, 4, 8, 16]
        - GlobalSplitUAlgorithm: ["MultipleBuffer", "MultipleBufferSingleKernel"]
      BenchmarkJoinParameters:
      BenchmarkFinalParameters:
        - ProblemSizes:
          - Exact: [ 127, 127, 1, 2047 ]
          - Exact: [ 128, 128, 1, 2048 ]
          - Exact: [ 255, 255, 1, 4095 ]
          - Exact: [ 256, 256, 1, 4096 ]
        - BiasTypeArgs: ['s']
        - ActivationArgs:
          - [Enum: none]