TestParameters: marks: [skip-gfx900, skip-gfx906, skip-gfx908, skip-gfx90a, skip-gfx950, skip-gfx1010, skip-gfx1011, skip-gfx1012, skip-gfx1030, skip-gfx1100, skip-gfx1101, skip-gfx1102, skip-gfx1200, skip-gfx1201, skip-gfx1250] # not supported by arch GlobalParameters: SyncsPerBenchmark: 0 NumElementsToValidate: -1 MinimumRequiredVersion: 5.0.0 PrintLevel: 1 #PrintSolutionRejectionReason: True Device: 0 CMakeBuildType: Release KernelTime: True MaxWorkspaceSize: 13421772800 DataInitTypeA: 13 DataInitTypeB: 12 DataInitTypeAlpha: 1 DataInitTypeBeta: 1 DataInitTypeBias: 13 DataInitTypeScaleAlphaVec: 12 BoundsCheck: 2 BenchmarkProblems: ######################################## # HHS TN SWIZZLED_A + BIAS + Activation + SAV ######################################## - - # ProblemType OperationType: GEMM DataType: h DestDataType: h ComputeDataType: s HighPrecisionAccumulate: True TransposeA: True TransposeB: False SwizzleTensorA: True UseBeta: True Batched: True UseBias: 1 Activation: True BiasDataTypeList: ['h'] UseScaleAlphaVec: 1 - # BenchmarkProblemSizeGroup - Standard - All problem InitialSolutionParameters: BenchmarkCommonParameters: - KernelLanguage: ["Assembly"] ForkParameters: - MatrixInstruction: - [16, 16, 16, 1, 1, 2, 8, 4,1 ] # MT = 128x128 - [16, 16, 16, 1, 1, 2, 16, 4,1 ] # MT = 128x256 - [16, 16, 16, 1, 1, 4, 8, 4,1 ] # MT = 256x128 - [16, 16, 16, 1, 1, 4, 16, 4,1 ] # MT = 256x256 - [16, 16, 16, 1, 1, 8, 8, 4,1 ] # MT = 512x128 - [16, 16, 16, 1, 1, 8, 16, 4,1 ] # MT = 512x256 - [16, 16, 16, 1, 1, 5, 8, 2,1 ] # MT = 160x128 - [16, 16, 16, 1, 1, 2, 4, 2,2 ] # MT = 64x128 - [16, 16, 16, 1, 1, 2, 8, 2,2 ] # MT = 64x256 - [16, 16, 16, 1, 1, 4, 4, 2,2 ] # MT = 128x128 - [16, 16, 16, 1, 1, 4, 8, 2,2 ] # MT = 128x256 - [16, 16, 16, 1, 1, 8, 4, 2,2 ] # MT = 256x128 - [16, 16, 16, 1, 1, 8, 8, 2,2 ] # MT = 256x256 - [16, 16, 16, 1, 1, 16, 4, 2,2 ] # MT = 512x128 - [16, 16, 16, 1, 1, 16, 8, 2,2 ] # MT = 512x256 - [16, 16, 16, 1, 1, 5, 4, 2,2 ] # MT = 160x128 - WorkGroup: - [16, 16, 1] # LSU=1, minDU=32 - [16, 16, 2] # LSU=2, minDU=64 - [16, 16, 4] # LSU=4, minDU=128 - GlobalReadVectorWidthA: [8] - GlobalReadVectorWidthB: [-1] - PrefetchGlobalRead: [1,2] - PrefetchLocalRead: [1,4] - ClusterLocalRead: [1] - NumElementsPerBatchStore: [0] - DepthU: [32,64,128] - VectorWidthA: [1] - VectorWidthB: [-1] - LocalWritePerMfma: [-1] - StaggerU: [4] - StaggerUStride: [256] - StaggerUMapping: [0] - WorkGroupMappingXCC: [8] - ScheduleIterAlg: [3] - LdsBlockSizePerPadA: [-1] - LdsBlockSizePerPadB: [-1] - StorePriorityOpt: [false] - VectorStore: [-1] - StoreSyncOpt: [0] - LdsPadA: [-1] - LdsPadB: [-1] - 1LDSBuffer: [1] - GlobalSplitU: [1,3] - GlobalSplitUAlgorithm: ["MultipleBuffer", "MultipleBufferSingleKernel"] - LocalReadVectorWidth: [2,4,8] - DirectToVgprA: [true] - UseSgprForGRO: [0,1] BenchmarkJoinParameters: BenchmarkFinalParameters: - ProblemSizes: - Exact: [160, 256, 1, 224] - Exact: [160, 256, 1, 256] - Exact: [160, 256, 1, 288] - Exact: [1600, 512, 1, 992] - Exact: [1600, 512, 1, 1024] - Exact: [1600, 512, 1, 1056] - Exact: [512, 256, 1, 224] - Exact: [512, 256, 1, 256] - Exact: [512, 256, 1, 288] - Exact: [127, 127, 1, 127] - Exact: [128, 127, 1, 127] - Exact: [127, 128, 1, 127] - Exact: [1, 127, 1, 127] - Exact: [127, 1, 1, 127] - BiasTypeArgs: ['h'] - ActivationArgs: - [Enum: relu] ######################################## # HSS TN SWIZZLED_A ######################################## - - # ProblemType OperationType: GEMM DataType: h DestDataType: s ComputeDataType: s HighPrecisionAccumulate: True TransposeA: True TransposeB: False SwizzleTensorA: True UseBeta: True Batched: True - # BenchmarkProblemSizeGroup - Standard - All problem InitialSolutionParameters: BenchmarkCommonParameters: - KernelLanguage: ["Assembly"] ForkParameters: - MatrixInstruction: - [16, 16, 16, 1, 1, 1, 8, 4,1 ] # MT = 64x128 - [16, 16, 16, 1, 1, 1, 16, 4,1 ] # MT = 64x256 - [16, 16, 16, 1, 1, 2, 8, 4,1 ] # MT = 128x128 - [16, 16, 16, 1, 1, 2, 16, 4,1 ] # MT = 128x256 - [16, 16, 16, 1, 1, 2, 4, 2,2 ] # MT = 64x128 - [16, 16, 16, 1, 1, 2, 8, 2,2 ] # MT = 64x256 - [16, 16, 16, 1, 1, 4, 4, 2,2 ] # MT = 128x128 - WorkGroup: - [16, 16, 1] # LSU=1, minDU=32 - [16, 16, 2] # LSU=2, minDU=64 - GlobalReadVectorWidthA: [8] - GlobalReadVectorWidthB: [-1] - PrefetchGlobalRead: [2] - PrefetchLocalRead: [1] - ClusterLocalRead: [1] - NumElementsPerBatchStore: [0] - DepthU: [32,64] - VectorWidthA: [1,2] - VectorWidthB: [-1] - LocalWritePerMfma: [-1] - StaggerU: [4] - StaggerUStride: [256] - StaggerUMapping: [0] - WorkGroupMappingXCC: [8] - ScheduleIterAlg: [3] - LdsBlockSizePerPadA: [-1] - LdsBlockSizePerPadB: [-1] - StorePriorityOpt: [false] - VectorStore: [-1] - StoreSyncOpt: [0] - LdsPadA: [-1] - LdsPadB: [-1] - 1LDSBuffer: [1] - GlobalSplitU: [1,3] - GlobalSplitUAlgorithm: ["MultipleBuffer"] - LocalReadVectorWidth: [2,4,8] - DirectToVgprA: [true] - UseSgprForGRO: [0,1] BenchmarkJoinParameters: BenchmarkFinalParameters: - ProblemSizes: - Exact: [160, 256, 1, 224] - Exact: [160, 256, 1, 256] - Exact: [160, 256, 1, 288] - Exact: [127, 127, 1, 127] - Exact: [128, 127, 1, 127] - Exact: [1, 127, 1, 127] - Exact: [127, 1, 1, 127] ######################################## # BBS TN SWIZZLED_A ######################################## - - # ProblemType OperationType: GEMM DataType: b DestDataType: b ComputeDataType: s HighPrecisionAccumulate: True TransposeA: True TransposeB: False SwizzleTensorA: True UseBeta: True Batched: True - # BenchmarkProblemSizeGroup - Standard - All problem InitialSolutionParameters: BenchmarkCommonParameters: - KernelLanguage: ["Assembly"] ForkParameters: - MatrixInstruction: - [16, 16, 16, 1, 1, 1, 8, 4,1 ] # MT = 64x128 - [16, 16, 16, 1, 1, 1, 16, 4,1 ] # MT = 64x256 - [16, 16, 16, 1, 1, 2, 8, 4,1 ] # MT = 128x128 - [16, 16, 16, 1, 1, 2, 16, 4,1 ] # MT = 128x256 - [16, 16, 16, 1, 1, 2, 4, 2,2 ] # MT = 64x128 - [16, 16, 16, 1, 1, 2, 8, 2,2 ] # MT = 64x256 - [16, 16, 16, 1, 1, 4, 4, 2,2 ] # MT = 128x128 - WorkGroup: - [16, 16, 1] # LSU=1, minDU=32 - [16, 16, 2] # LSU=2, minDU=64 - GlobalReadVectorWidthA: [8] - GlobalReadVectorWidthB: [-1] - PrefetchGlobalRead: [2] - PrefetchLocalRead: [1] - ClusterLocalRead: [1] - NumElementsPerBatchStore: [0] - DepthU: [32,64] - VectorWidthA: [1,2] - VectorWidthB: [-1] - LocalWritePerMfma: [-1] - StaggerU: [4] - StaggerUStride: [256] - StaggerUMapping: [0] - WorkGroupMappingXCC: [8] - ScheduleIterAlg: [3] - LdsBlockSizePerPadA: [-1] - LdsBlockSizePerPadB: [-1] - StorePriorityOpt: [false] - VectorStore: [-1] - StoreSyncOpt: [0] - LdsPadA: [-1] - LdsPadB: [-1] - 1LDSBuffer: [1] - GlobalSplitU: [1,3] - GlobalSplitUAlgorithm: ["MultipleBuffer"] - LocalReadVectorWidth: [2,4,8] - DirectToVgprA: [true] - UseSgprForGRO: [0,1] BenchmarkJoinParameters: BenchmarkFinalParameters: - ProblemSizes: - Exact: [160, 256, 1, 224] - Exact: [160, 256, 1, 256] - Exact: [160, 256, 1, 288] - Exact: [127, 127, 1, 127] - Exact: [128, 127, 1, 127] - Exact: [1, 127, 1, 127] - Exact: [127, 1, 1, 127] ######################################## # F8B8NHS TN SWIZZLED_A ######################################## - - # ProblemType OperationType: GEMM DataType: F8B8N DestDataType: h ComputeDataType: s HighPrecisionAccumulate: True TransposeA: True TransposeB: False SwizzleTensorA: True UseBeta: True Batched: True - # BenchmarkProblemSizeGroup - Standard - All problem InitialSolutionParameters: BenchmarkCommonParameters: - KernelLanguage: ["Assembly"] ForkParameters: - MatrixInstruction: - [16, 16, 32, 1, 1, 1, 8, 4,1 ] # MT = 64x128 - [16, 16, 32, 1, 1, 1, 16, 4,1 ] # MT = 64x256 - [16, 16, 32, 1, 1, 2, 8, 4,1 ] # MT = 128x128 - [16, 16, 32, 1, 1, 2, 16, 4,1 ] # MT = 128x256 - [16, 16, 32, 1, 1, 16, 4, 2,2 ] # MT = 512x128 - [16, 16, 32, 1, 1, 16, 8, 2,2 ] # MT = 512x256 - [16, 16, 32, 1, 1, 5, 4, 2,2 ] # MT = 160x128 - WorkGroup: - [16, 16, 1] # LSU=1, minDU=64 - [16, 16, 2] # LSU=2, minDU=128 - GlobalReadVectorWidthA: [16] - GlobalReadVectorWidthB: [-1] - PrefetchGlobalRead: [2] - PrefetchLocalRead: [1] - ClusterLocalRead: [1] - NumElementsPerBatchStore: [0] - DepthU: [64,128] - VectorWidthA: [1,4] - VectorWidthB: [-1] - LocalWritePerMfma: [-1] - StaggerU: [4] - StaggerUStride: [256] - StaggerUMapping: [0] - WorkGroupMappingXCC: [8] - ScheduleIterAlg: [3] - LdsBlockSizePerPadA: [-1] - LdsBlockSizePerPadB: [-1] - StorePriorityOpt: [false] - VectorStore: [-1] - StoreSyncOpt: [0] - LdsPadA: [-1] - LdsPadB: [-1] - 1LDSBuffer: [1] - GlobalSplitU: [1,3] - GlobalSplitUAlgorithm: ["MultipleBuffer"] - LocalReadVectorWidth: [16] - DirectToVgprA: [true] - UseSgprForGRO: [0,1] BenchmarkJoinParameters: BenchmarkFinalParameters: - ProblemSizes: - Exact: [160, 256, 1, 224] - Exact: [160, 256, 1, 256] - Exact: [160, 256, 1, 288] - Exact: [127, 127, 1, 127] - Exact: [128, 127, 1, 127] - Exact: [1, 127, 1, 127] - Exact: [127, 1, 1, 127] ######################################## # XFloat TN SWIZZLED_A ######################################## - - # ProblemType OperationType: GEMM DataType: s DestDataType: s ComputeDataType: s F32XdlMathOp: X HighPrecisionAccumulate: False TransposeA: True TransposeB: False SwizzleTensorA: True UseBeta: True Batched: True - # BenchmarkProblemSizeGroup - Standard - All problem InitialSolutionParameters: BenchmarkCommonParameters: - KernelLanguage: ["Assembly"] ForkParameters: - MatrixInstruction: - [16, 16, 8, 1, 1, 4, 16, 4,1 ] # MT = 256x256 - [16, 16, 8, 1, 1, 8, 8, 4,1 ] # MT = 512x128 - [16, 16, 8, 1, 1, 8, 16, 4,1 ] # MT = 512x256 - [16, 16, 8, 1, 1, 5, 8, 2, 1 ] # MT = 160x128 - [16, 16, 8, 1, 1, 4, 8, 2,2 ] # MT = 128x256 - [16, 16, 8, 1, 1, 8, 4, 2,2 ] # MT = 256x128 - [16, 16, 8, 1, 1, 8, 8, 2,2 ] # MT = 256x256 - WorkGroup: - [16, 16, 1] # LSU=1, minDU=16 - [16, 16, 2] # LSU=2, minDU=32 - [16, 16, 4] # LSU=4, minDU=64 - GlobalReadVectorWidthA: [4] - GlobalReadVectorWidthB: [-1] - PrefetchGlobalRead: [2] - PrefetchLocalRead: [1] - ClusterLocalRead: [1] - NumElementsPerBatchStore: [0] - DepthU: [32,64] - VectorWidthA: [1,8] - VectorWidthB: [-1] - LocalWritePerMfma: [-1] - StaggerU: [4] - StaggerUStride: [256] - StaggerUMapping: [0] - WorkGroupMappingXCC: [8] - ScheduleIterAlg: [3] - LdsBlockSizePerPadA: [-1] - LdsBlockSizePerPadB: [-1] - StorePriorityOpt: [false] - VectorStore: [-1] - StoreSyncOpt: [0] - LdsPadA: [-1] - LdsPadB: [-1] - 1LDSBuffer: [1] - GlobalSplitU: [1,3] - GlobalSplitUAlgorithm: ["MultipleBuffer"] - LocalReadVectorWidth: [2,4,8] - DirectToVgprA: [true] - UseSgprForGRO: [0,1] BenchmarkJoinParameters: BenchmarkFinalParameters: - ProblemSizes: - Exact: [160, 256, 1, 224] - Exact: [160, 256, 1, 256] - Exact: [160, 256, 1, 288] - Exact: [127, 127, 1, 127] - Exact: [128, 127, 1, 127] - Exact: [1, 127, 1, 127] - Exact: [127, 1, 1, 127] ######################################## # SSS TN SWIZZLED_A ######################################## - - # ProblemType OperationType: GEMM DataType: s DestDataType: s ComputeDataType: s HighPrecisionAccumulate: False TransposeA: True TransposeB: False SwizzleTensorA: True UseBeta: True Batched: True - # BenchmarkProblemSizeGroup - Standard - All problem InitialSolutionParameters: BenchmarkCommonParameters: - KernelLanguage: ["Assembly"] ForkParameters: - MatrixInstruction: - [16, 16, 4, 1, 1, 2, 8, 4,1 ] # MT = 128x128 - [16, 16, 4, 1, 1, 2, 16, 4,1 ] # MT = 128x256 - [16, 16, 4, 1, 1, 4, 8, 4,1 ] # MT = 256x128 - [16, 16, 4, 1, 1, 4, 16, 4,1 ] # MT = 256x256 - [16, 16, 4, 1, 1, 16, 4, 2,2 ] # MT = 512x128 - [16, 16, 4, 1, 1, 16, 8, 2,2 ] # MT = 512x256 - [16, 16, 4, 1, 1, 5, 4, 2,2 ] # MT = 160x128 - WorkGroup: - [16, 16, 1] # LSU=1, minDU=16 - [16, 16, 2] # LSU=2, minDU=32 - [16, 16, 4] # LSU=4, minDU=64 - GlobalReadVectorWidthA: [4] - GlobalReadVectorWidthB: [-1] - PrefetchGlobalRead: [2] - PrefetchLocalRead: [1] - ClusterLocalRead: [1] - NumElementsPerBatchStore: [0] - DepthU: [32,64] - VectorWidthA: [1,2] - VectorWidthB: [-1] - LocalWritePerMfma: [-1] - StaggerU: [4] - StaggerUStride: [256] - StaggerUMapping: [0] - WorkGroupMappingXCC: [8] - ScheduleIterAlg: [3] - LdsBlockSizePerPadA: [-1] - LdsBlockSizePerPadB: [-1] - StorePriorityOpt: [false] - VectorStore: [-1] - StoreSyncOpt: [0] - LdsPadA: [-1] - LdsPadB: [-1] - 1LDSBuffer: [1] - GlobalSplitU: [1,3] - GlobalSplitUAlgorithm: ["MultipleBuffer"] - LocalReadVectorWidth: [2,4,8] - DirectToVgprA: [true] - UseSgprForGRO: [0,1] BenchmarkJoinParameters: BenchmarkFinalParameters: - ProblemSizes: - Exact: [160, 256, 1, 224] - Exact: [160, 256, 1, 256] - Exact: [160, 256, 1, 288] - Exact: [127, 127, 1, 127] - Exact: [128, 127, 1, 127] - Exact: [1, 127, 1, 127] - Exact: [127, 1, 1, 127]