TestParameters: marks: [skip-gfx1100, skip-gfx1101, skip-gfx1102, skip-gfx1200, skip-gfx1201, skip-gfx1250] # not supported yet GlobalParameters: SyncsPerBenchmark: 0 MinimumRequiredVersion: 5.0.0 SleepPercent: 50 NumElementsToValidate: 128 DataInitTypeBeta: 0 DataInitTypeAlpha: 1 CSVExportWinner: true CSVMergeSameProblemID: true Device: 0 BenchmarkProblems: ######################################## # NN - standard ######################################## - - # ProblemType OperationType: GEMM DataType: h DestDataType: h ComputeDataType: s HighPrecisionAccumulate: True TransposeA: False TransposeB: False UseBeta: True UseBias: 1 Batched: True Activation: True ActivationType: all - # BenchmarkProblemSizeGroup - Standard - All problem InitialSolutionParameters: BenchmarkCommonParameters: - KernelLanguage: ["Assembly"] ForkParameters: - MatrixInstruction: - [16, 16, 16, 1, 1, 2, 1, 4,1 ] - PrefetchGlobalRead: [2] - PrefetchLocalRead: [1] - ClusterLocalRead: [1] - DepthU: [64] - ScheduleIterAlg: [3] - ExpandPointerSwap: [false] - TransposeLDS: [0] - GlobalSplitU: [1,2,4] - GlobalSplitUAlgorithm: ["MultipleBuffer", "MultipleBufferSingleKernel"] - SourceSwap: [false] - StoreVectorWidth: [8] - AdaptiveGemm: [0,1] BenchmarkJoinParameters: BenchmarkFinalParameters: - ProblemSizes: # edge n - Exact: [248, 127, 1, 128] # ag 0 svw 1 <-> ag 1 svw 8 - Exact: [249, 127, 1, 128] # ag 0 svw 1 <-> ag 1 svw 1 - Exact: [250, 127, 1, 128] # ag 0 svw 1 <-> ag 1 svw 2 - Exact: [252, 127, 1, 128] # ag 0 svw 1 <-> ag 1 svw 4 - Exact: [256, 127, 1, 128] # ag 0 svw 8 <-> ag 1 svw 8 # non edge n - Exact: [248, 128, 1, 128] # ag 0 svw 1 <-> ag 1 svw 8 - Exact: [249, 128, 1, 128] # ag 0 svw 1 <-> ag 1 svw 1 - Exact: [250, 128, 1, 128] # ag 0 svw 1 <-> ag 1 svw 2 - Exact: [252, 128, 1, 128] # ag 0 svw 1 <-> ag 1 svw 4 - Exact: [256, 128, 1, 128] # ag 0 svw 8 <-> ag 1 svw 8 # edge n - Exact: [248, 129, 1, 128] # ag 0 svw 1 <-> ag 1 svw 8 - Exact: [249, 129, 1, 128] # ag 0 svw 1 <-> ag 1 svw 1 - Exact: [250, 129, 1, 128] # ag 0 svw 1 <-> ag 1 svw 2 - Exact: [252, 129, 1, 128] # ag 0 svw 1 <-> ag 1 svw 4 - Exact: [256, 129, 1, 128] # ag 0 svw 8 <-> ag 1 svw 8 - BiasTypeArgs: ['h'] - ActivationArgs: - [Enum: relu] - [Enum: silu] - [Enum: swish] ######################################## # TN - standard ######################################## - - # ProblemType OperationType: GEMM DataType: h DestDataType: h ComputeDataType: s HighPrecisionAccumulate: True TransposeA: True TransposeB: False UseBeta: True UseBias: 1 Batched: True Activation: True ActivationType: all - # BenchmarkProblemSizeGroup - Standard - All problem InitialSolutionParameters: BenchmarkCommonParameters: - KernelLanguage: ["Assembly"] ForkParameters: - MatrixInstruction: - [16, 16, 16, 1, 1, 2, 1, 4,1 ] - PrefetchGlobalRead: [2] - PrefetchLocalRead: [1] - ClusterLocalRead: [1] - DepthU: [64] - ScheduleIterAlg: [3] - ExpandPointerSwap: [false] - TransposeLDS: [0] - GlobalSplitU: [1,2,4] - GlobalSplitUAlgorithm: ["MultipleBuffer", "MultipleBufferSingleKernel"] - SourceSwap: [false] - StoreVectorWidth: [8] - AdaptiveGemm: [0,1] BenchmarkJoinParameters: BenchmarkFinalParameters: - ProblemSizes: # edge n - Exact: [248, 127, 1, 128] # ag 0 svw 1 <-> ag 1 svw 8 - Exact: [249, 127, 1, 128] # ag 0 svw 1 <-> ag 1 svw 1 - Exact: [250, 127, 1, 128] # ag 0 svw 1 <-> ag 1 svw 2 - Exact: [252, 127, 1, 128] # ag 0 svw 1 <-> ag 1 svw 4 - Exact: [256, 127, 1, 128] # ag 0 svw 8 <-> ag 1 svw 8 # non edge n - Exact: [248, 128, 1, 128] # ag 0 svw 1 <-> ag 1 svw 8 - Exact: [249, 128, 1, 128] # ag 0 svw 1 <-> ag 1 svw 1 - Exact: [250, 128, 1, 128] # ag 0 svw 1 <-> ag 1 svw 2 - Exact: [252, 128, 1, 128] # ag 0 svw 1 <-> ag 1 svw 4 - Exact: [256, 128, 1, 128] # ag 0 svw 8 <-> ag 1 svw 8 # edge n - Exact: [248, 129, 1, 128] # ag 0 svw 1 <-> ag 1 svw 8 - Exact: [249, 129, 1, 128] # ag 0 svw 1 <-> ag 1 svw 1 - Exact: [250, 129, 1, 128] # ag 0 svw 1 <-> ag 1 svw 2 - Exact: [252, 129, 1, 128] # ag 0 svw 1 <-> ag 1 svw 4 - Exact: [256, 129, 1, 128] # ag 0 svw 8 <-> ag 1 svw 8 - BiasTypeArgs: ['h'] - ActivationArgs: - [Enum: relu] - [Enum: silu] - [Enum: swish] ######################################## # NT - standard ######################################## - - # ProblemType OperationType: GEMM DataType: h DestDataType: h ComputeDataType: s HighPrecisionAccumulate: True TransposeA: False TransposeB: True UseBeta: True UseBias: 1 Batched: True Activation: True ActivationType: all - # BenchmarkProblemSizeGroup - Standard - All problem InitialSolutionParameters: BenchmarkCommonParameters: - KernelLanguage: ["Assembly"] ForkParameters: - MatrixInstruction: - [16, 16, 16, 1, 1, 2, 1, 4,1 ] - PrefetchGlobalRead: [2] - PrefetchLocalRead: [1] - ClusterLocalRead: [1] - DepthU: [64] - ScheduleIterAlg: [3] - ExpandPointerSwap: [false] - TransposeLDS: [0] - GlobalSplitU: [1,2,4] - GlobalSplitUAlgorithm: ["MultipleBuffer", "MultipleBufferSingleKernel"] - SourceSwap: [false] - StoreVectorWidth: [8] - AdaptiveGemm: [0,1] BenchmarkJoinParameters: BenchmarkFinalParameters: - ProblemSizes: # edge n - Exact: [248, 127, 1, 128] # ag 0 svw 1 <-> ag 1 svw 8 - Exact: [249, 127, 1, 128] # ag 0 svw 1 <-> ag 1 svw 1 - Exact: [250, 127, 1, 128] # ag 0 svw 1 <-> ag 1 svw 2 - Exact: [252, 127, 1, 128] # ag 0 svw 1 <-> ag 1 svw 4 - Exact: [256, 127, 1, 128] # ag 0 svw 8 <-> ag 1 svw 8 # non edge n - Exact: [248, 128, 1, 128] # ag 0 svw 1 <-> ag 1 svw 8 - Exact: [249, 128, 1, 128] # ag 0 svw 1 <-> ag 1 svw 1 - Exact: [250, 128, 1, 128] # ag 0 svw 1 <-> ag 1 svw 2 - Exact: [252, 128, 1, 128] # ag 0 svw 1 <-> ag 1 svw 4 - Exact: [256, 128, 1, 128] # ag 0 svw 8 <-> ag 1 svw 8 # edge n - Exact: [248, 129, 1, 128] # ag 0 svw 1 <-> ag 1 svw 8 - Exact: [249, 129, 1, 128] # ag 0 svw 1 <-> ag 1 svw 1 - Exact: [250, 129, 1, 128] # ag 0 svw 1 <-> ag 1 svw 2 - Exact: [252, 129, 1, 128] # ag 0 svw 1 <-> ag 1 svw 4 - Exact: [256, 129, 1, 128] # ag 0 svw 8 <-> ag 1 svw 8 - BiasTypeArgs: ['h'] - ActivationArgs: - [Enum: relu] - [Enum: silu] - [Enum: swish] ######################################## # TT - standard ######################################## - - # ProblemType OperationType: GEMM DataType: h DestDataType: h ComputeDataType: s HighPrecisionAccumulate: True TransposeA: True TransposeB: True UseBeta: True UseBias: 1 Batched: True Activation: True ActivationType: all - # BenchmarkProblemSizeGroup - Standard - All problem InitialSolutionParameters: BenchmarkCommonParameters: - KernelLanguage: ["Assembly"] ForkParameters: - MatrixInstruction: - [16, 16, 16, 1, 1, 2, 1, 4,1 ] - PrefetchGlobalRead: [2] - PrefetchLocalRead: [1] - ClusterLocalRead: [1] - DepthU: [64] - ScheduleIterAlg: [3] - ExpandPointerSwap: [false] - TransposeLDS: [0] - GlobalSplitU: [1,2,4] - GlobalSplitUAlgorithm: ["MultipleBuffer", "MultipleBufferSingleKernel"] - SourceSwap: [false] - StoreVectorWidth: [8] - AdaptiveGemm: [0,1] BenchmarkJoinParameters: BenchmarkFinalParameters: - ProblemSizes: # edge n - Exact: [248, 127, 1, 128] # ag 0 svw 1 <-> ag 1 svw 8 - Exact: [249, 127, 1, 128] # ag 0 svw 1 <-> ag 1 svw 1 - Exact: [250, 127, 1, 128] # ag 0 svw 1 <-> ag 1 svw 2 - Exact: [252, 127, 1, 128] # ag 0 svw 1 <-> ag 1 svw 4 - Exact: [256, 127, 1, 128] # ag 0 svw 8 <-> ag 1 svw 8 # non edge n - Exact: [248, 128, 1, 128] # ag 0 svw 1 <-> ag 1 svw 8 - Exact: [249, 128, 1, 128] # ag 0 svw 1 <-> ag 1 svw 1 - Exact: [250, 128, 1, 128] # ag 0 svw 1 <-> ag 1 svw 2 - Exact: [252, 128, 1, 128] # ag 0 svw 1 <-> ag 1 svw 4 - Exact: [256, 128, 1, 128] # ag 0 svw 8 <-> ag 1 svw 8 # edge n - Exact: [248, 129, 1, 128] # ag 0 svw 1 <-> ag 1 svw 8 - Exact: [249, 129, 1, 128] # ag 0 svw 1 <-> ag 1 svw 1 - Exact: [250, 129, 1, 128] # ag 0 svw 1 <-> ag 1 svw 2 - Exact: [252, 129, 1, 128] # ag 0 svw 1 <-> ag 1 svw 4 - Exact: [256, 129, 1, 128] # ag 0 svw 8 <-> ag 1 svw 8 - BiasTypeArgs: ['h'] - ActivationArgs: - [Enum: relu] - [Enum: silu] - [Enum: swish]