GlobalParameters: MinimumRequiredVersion: 4.4.0 PrintLevel: 1 ForceRedoBenchmarkProblems: True ForceRedoLibraryLogic: True ForceRedoLibraryClient: True CMakeBuildType: Release EnqueuesPerSync: 1 SyncsPerBenchmark: 1 NumElementsToValidate: 0 ValidationMaxToPrint: 4 ValidationPrintValids: False Platform: 0 Device: 0 KernelTime: True PinClocks: True SleepPercent: 200 DataInitTypeBeta : 0 BenchmarkProblems: ######################################## ######################################## ### ### NN ### ######################################## ######################################## - - # ProblemType OperationType: GEMM DataType: s TransposeA: False TransposeB: False UseBeta: True Batched: True ######################################## # NN - Super Skinny ######################################## - # Benchmark Group InitialSolutionParameters: BenchmarkCommonParameters: - EdgeType: ["ShiftPtr"] - LoopTail: [True] - KernelLanguage: ["Assembly"] - PrefetchLocalRead: [True] - WorkGroupMapping: [1] ForkParameters: - PrefetchGlobalRead: [False, True] - ThreadTile: - [ 2, 2 ] - [ 4, 2 ] - [ 8, 2 ] - [ 4, 4 ] - WorkGroup: - [ 16, 8, 2 ] - [ 32, 4, 2 ] - [ 16, 4, 4 ] - [ 32, 2, 4 ] - [ 8, 4, 8 ] - [ 16, 2, 8 ] - GlobalSplitU: [16, 32] - DepthU: [ 16, 32, 64 ] - VectorWidth: [-1] BenchmarkForkParameters: JoinParameters: BenchmarkJoinParameters: BenchmarkFinalParameters: - ProblemSizes: - Exact: [ 512, 1, 1, 500000 ] - Exact: [ 512, 2, 1, 500000 ] - Exact: [ 512, 4, 1, 500000 ] - Exact: [ 512, 8, 1, 500000 ] - Exact: [ 512, 16, 1, 500000 ] - Exact: [ 1024, 1, 1, 500000 ] - Exact: [ 1024, 2, 1, 500000 ] - Exact: [ 1024, 4, 1, 500000 ] - Exact: [ 1024, 8, 1, 500000 ] - Exact: [ 1024, 16, 1, 500000 ] - Exact: [ 64, 1, 1, 1216 ] - Exact: [ 128, 1, 1, 1024 ] - Exact: [ 128, 1, 1, 1408 ] - Exact: [ 512, 1, 1, 512 ] - Exact: [ 3072, 1, 1, 128 ] - Exact: [ 512, 2, 1, 512 ] - Exact: [ 1024, 1, 1, 512 ] - Exact: [ 4224, 1, 1, 128 ] - Exact: [ 512, 4, 1, 512 ] - Exact: [ 1024, 2, 1, 512 ] - Exact: [ 1024, 4, 1, 512 ] - Exact: [ 3072, 1, 1, 1024 ] - Exact: [ 3072, 1, 1, 1024 ] - Exact: [ 512, 16, 1, 512 ] - Exact: [ 3072, 2, 1, 1024 ] - Exact: [ 4608, 1, 1, 1536 ] - Exact: [ 1024, 16, 1, 512 ] - Exact: [ 3072, 4, 1, 1024 ] - Exact: [ 4608, 2, 1, 1536 ] - Exact: [ 6144, 1, 1, 2560 ] - Exact: [ 7680, 1, 1, 2560 ] - Exact: [ 8448, 1, 1, 2816 ] - Exact: [ 4608, 4, 1, 1536 ] - Exact: [ 6144, 2, 1, 2560 ] - Exact: [ 7680, 2, 1, 2560 ] - Exact: [ 8448, 2, 1, 2816 ] - Exact: [ 1760, 16, 1, 1760 ] - Exact: [ 3072, 16, 1, 1024 ] - Exact: [ 6144, 4, 1, 2560 ] - Exact: [ 2048, 16, 1, 2048 ] - Exact: [ 7680, 4, 1, 2560 ] - Exact: [ 8448, 4, 1, 2816 ] - Exact: [ 2560, 16, 1, 2560 ] - Exact: [ 4608, 16, 1, 1536 ] - Exact: [ 6144, 16, 1, 2560 ] - Exact: [ 4096, 16, 1, 4096 ] - Exact: [ 7680, 16, 1, 2560 ] - Exact: [ 8448, 16, 1, 2816 ] ######################################## # NN - 3 Waves/WG ######################################## - # Benchmark Group InitialSolutionParameters: BenchmarkCommonParameters: - EdgeType: ["ShiftPtr"] - LoopTail: [True] - KernelLanguage: ["Assembly"] - PrefetchGlobalRead: [True] - PrefetchLocalRead: [True] - WorkGroupMapping: [1] ForkParameters: - ThreadTile: - [ 3, 3 ] - [ 3, 4 ] - [ 3, 6 ] - [ 4, 3 ] - [ 4, 4 ] - [ 4, 6 ] - [ 6, 3 ] - [ 6, 4 ] - [ 6, 6 ] - WorkGroup: - [ 24, 8, 1 ] - [ 12, 16, 1 ] - [ 6, 32, 1 ] - [ 8, 12, 2 ] - [ 8, 6, 4 ] - [ 16, 16, 1 ] - GlobalSplitU: [2, 4, 8] - DepthU: [ 8, 16, 24, 32 ] - VectorWidth: [-1] BenchmarkForkParameters: JoinParameters: BenchmarkJoinParameters: BenchmarkFinalParameters: - ProblemSizes: - Exact: [ 35, 700, 1, 2048 ] - Exact: [ 35, 700, 1, 2048 ] - Exact: [ 35, 700, 1, 2560 ] - Exact: [ 35, 1500, 1, 2048 ] - Exact: [ 35, 1500, 1, 2560 ] - Exact: [ 35, 8457, 1, 1760 ] - Exact: [ 35, 8457, 1, 2048 ] - Exact: [ 35, 8457, 1, 2560 ] - Exact: [ 35, 8457, 1, 4096 ] ######################################## # NN - Small or Skinny ######################################## - # Benchmark Group InitialSolutionParameters: BenchmarkCommonParameters: - EdgeType: ["ShiftPtr"] - LoopTail: [True] - KernelLanguage: ["Source"] - PrefetchGlobalRead: [True] - PrefetchLocalRead: [True] ForkParameters: - ThreadTile: - [ 2, 2 ] - [ 4, 2 ] - [ 4, 4 ] - [ 8, 4 ] - [ 8, 8 ] - WorkGroup: - [ 32, 4, 1 ] - [ 16, 4, 1 ] - [ 8, 8, 1 ] - [ 16, 8, 1 ] - WorkGroupMapping: [1, 8] - GlobalSplitU: [1] - DepthU: [ 8, 16, 24, 32 ] - VectorWidth: [-1] BenchmarkForkParameters: JoinParameters: BenchmarkJoinParameters: BenchmarkFinalParameters: - ProblemSizes: - Range: [ [64, 128], [64, 64, 64, 7000], [1], [128] ] # skinny-0, source code for k<=128 - Range: [ [64, 64, 64, 7000], [64, 128], [1], [128] ] # skinny-1, source code for k<=128 - Range: [ [64, 64, 64, 700], [64, 64, 64, 700], [1], [128] ] # small, source code for k<=128 - # Benchmark Group InitialSolutionParameters: BenchmarkCommonParameters: - EdgeType: ["ShiftPtr"] - LoopTail: [True] - KernelLanguage: ["Assembly"] - PrefetchGlobalRead: [True] - PrefetchLocalRead: [True] ForkParameters: - ThreadTile: - [ 2, 2 ] - [ 4, 2 ] - [ 4, 4 ] - [ 8, 4 ] - [ 8, 8 ] - WorkGroup: - [ 32, 4, 2 ] - [ 16, 4, 4 ] - [ 8, 8, 2 ] - [ 8, 8, 4 ] - [ 16, 8, 2 ] - WorkGroupMapping: [1, 8] - GlobalSplitU: [2, 4, 8] - DepthU: [ 8, 16, 24, 32 ] - VectorWidth: [-1] BenchmarkForkParameters: JoinParameters: BenchmarkJoinParameters: BenchmarkFinalParameters: - ProblemSizes: - Range: [ [64, 128], [64, 64, 64, 7000], [1], [256, 1024, 1024, 4096] ] # skinny-0 - Range: [ [64, 64, 64, 7000], [64, 128], [1], [256, 1024, 1024, 4096] ] # skinny-1 - Range: [ [64, 64, 64, 700], [64, 64, 64, 700], [1], [256, 1024, 1024, 4096] ] # small - Exact: [ 64, 1, 1, 1216 ] - Exact: [ 128, 1, 1, 1024 ] - Exact: [ 128, 1, 1, 1408 ] - Exact: [ 512, 1, 1, 512 ] - Exact: [ 3072, 1, 1, 128 ] - Exact: [ 512, 2, 1, 512 ] - Exact: [ 1024, 1, 1, 512 ] - Exact: [ 4224, 1, 1, 128 ] - Exact: [ 512, 4, 1, 512 ] - Exact: [ 1024, 2, 1, 512 ] - Exact: [ 1024, 4, 1, 512 ] - Exact: [ 3072, 1, 1, 1024 ] - Exact: [ 3072, 1, 1, 1024 ] - Exact: [ 512, 16, 1, 512 ] - Exact: [ 3072, 2, 1, 1024 ] - Exact: [ 4608, 1, 1, 1536 ] - Exact: [ 512, 32, 1, 512 ] - Exact: [ 1024, 16, 1, 512 ] - Exact: [ 3072, 4, 1, 1024 ] - Exact: [ 4608, 2, 1, 1536 ] - Exact: [ 6144, 1, 1, 2560 ] - Exact: [ 1024, 32, 1, 512 ] - Exact: [ 7680, 1, 1, 2560 ] - Exact: [ 8448, 1, 1, 2816 ] - Exact: [ 4608, 4, 1, 1536 ] - Exact: [ 6144, 2, 1, 2560 ] - Exact: [ 7680, 2, 1, 2560 ] - Exact: [ 8448, 2, 1, 2816 ] - Exact: [ 1760, 16, 1, 1760 ] - Exact: [ 35, 700, 1, 2048 ] - Exact: [ 35, 700, 1, 2048 ] - Exact: [ 3072, 16, 1, 1024 ] - Exact: [ 35, 700, 1, 2560 ] - Exact: [ 6144, 4, 1, 2560 ] - Exact: [ 2048, 16, 1, 2048 ] - Exact: [ 7680, 4, 1, 2560 ] - Exact: [ 8448, 4, 1, 2816 ] - Exact: [ 1760, 32, 1, 1760 ] - Exact: [ 3072, 32, 1, 1024 ] - Exact: [ 2560, 16, 1, 2560 ] - Exact: [ 35, 1500, 1, 2048 ] - Exact: [ 4608, 16, 1, 1536 ] - Exact: [ 2048, 32, 1, 2048 ] - Exact: [ 35, 1500, 1, 2560 ] - Exact: [ 1760, 64, 1, 1760 ] - Exact: [ 3072, 64, 1, 1024 ] - Exact: [ 2560, 32, 1, 2560 ] - Exact: [ 4608, 32, 1, 1536 ] - Exact: [ 128, 1500, 1, 1280 ] - Exact: [ 6144, 16, 1, 2560 ] - Exact: [ 2048, 64, 1, 2048 ] - Exact: [ 4096, 16, 1, 4096 ] - Exact: [ 7680, 16, 1, 2560 ] - Exact: [ 176, 1500, 1, 1408 ] - Exact: [ 8448, 16, 1, 2816 ] - Exact: [ 1760, 128, 1, 1760 ] - Exact: [ 3072, 128, 1, 1024 ] - Exact: [ 2560, 64, 1, 2560 ] - Exact: [ 6144, 32, 1, 2560 ] - Exact: [ 35, 8457, 1, 1760 ] - Exact: [ 2048, 128, 1, 2048 ] - Exact: [ 4096, 32, 1, 4096 ] - Exact: [ 35, 8457, 1, 2048 ] - Exact: [ 7680, 32, 1, 2560 ] - Exact: [ 35, 8457, 1, 2560 ] - Exact: [ 8448, 32, 1, 2816 ] - Exact: [ 2560, 128, 1, 2560 ] - Exact: [ 4096, 64, 1, 4096 ] - Exact: [ 35, 8457, 1, 4096 ] - Exact: [ 7680, 64, 1, 2560 ] - Exact: [ 4096, 128, 1, 4096 ] - Exact: [ 7680, 128, 1, 2560 ] ######################################## # NN - Large ######################################## - # Benchmark Group InitialSolutionParameters: BenchmarkCommonParameters: - EdgeType: ["ShiftPtr"] - LoopTail: [True] - KernelLanguage: ["Source"] - PrefetchLocalRead: [True] - GlobalSplitU: [1] ForkParameters: - PrefetchGlobalRead: [False, True] - ThreadTile: - [ 4, 4 ] - [ 4, 8 ] - [ 6, 8 ] - [ 8, 4 ] - [ 8, 6 ] - [ 8, 8 ] - WorkGroup: - [ 16, 16, 1 ] - WorkGroupMapping: [1, 8] - DepthU: [ 8, 16, 24, 32 ] - VectorWidth: [-1] BenchmarkForkParameters: JoinParameters: BenchmarkJoinParameters: BenchmarkFinalParameters: - ProblemSizes: - Range: [ [64, 64, 64, 7000], [64, 64, 64, 7000], [1], [128] ] # large, source code for k<=128 - # Benchmark Group InitialSolutionParameters: BenchmarkCommonParameters: - EdgeType: ["ShiftPtr"] - LoopTail: [True] - KernelLanguage: ["Assembly"] - PrefetchLocalRead: [True] - GlobalSplitU: [1] ForkParameters: - PrefetchGlobalRead: [False, True] - ThreadTile: - [ 4, 4 ] - [ 4, 8 ] - [ 6, 8 ] - [ 8, 4 ] - [ 8, 6 ] - [ 8, 8 ] - WorkGroup: - [ 16, 16, 1 ] - WorkGroupMapping: [1, 8] - DepthU: [ 8, 16, 24, 32 ] - VectorWidth: [-1] BenchmarkForkParameters: JoinParameters: BenchmarkJoinParameters: BenchmarkFinalParameters: - ProblemSizes: - Range: [ [64, 64, 64, 7000], [64, 64, 64, 7000], [1], [256, 1024, 1024, 4096] ] # large - Exact: [ 1024, 1024, 1, 1024 ] - Exact: [ 1024, 700, 1, 512 ] - Exact: [ 1024, 700, 1, 512 ] - Exact: [ 3072, 128, 1, 1024 ] - Exact: [ 3072, 1500, 1, 128 ] - Exact: [ 2560, 128, 1, 2560 ] - Exact: [ 4224, 1500, 1, 176 ] - Exact: [ 512, 1500, 1, 1536 ] - Exact: [ 512, 1500, 1, 2048 ] - Exact: [ 512, 1500, 1, 2560 ] - Exact: [ 4096, 128, 1, 4096 ] - Exact: [ 512, 1500, 1, 2816 ] - Exact: [ 512, 3000, 1, 1536 ] - Exact: [ 1024, 1500, 1, 1536 ] - Exact: [ 7680, 128, 1, 2560 ] - Exact: [ 512, 3000, 1, 2048 ] - Exact: [ 1024, 1500, 1, 2048 ] - Exact: [ 512, 3000, 1, 2560 ] - Exact: [ 1024, 1500, 1, 2560 ] - Exact: [ 512, 3000, 1, 2816 ] - Exact: [ 1024, 1500, 1, 2816 ] - Exact: [ 512, 6000, 1, 1536 ] - Exact: [ 1024, 3000, 1, 1536 ] - Exact: [ 3072, 1500, 1, 1024 ] - Exact: [ 3072, 1500, 1, 1024 ] - Exact: [ 512, 6000, 1, 2048 ] - Exact: [ 1024, 3000, 1, 2048 ] - Exact: [ 5124, 700, 1, 2048 ] - Exact: [ 5124, 700, 1, 2048 ] - Exact: [ 512, 6000, 1, 2560 ] - Exact: [ 1024, 3000, 1, 2560 ] - Exact: [ 512, 6000, 1, 2816 ] - Exact: [ 1024, 3000, 1, 2816 ] - Exact: [ 5124, 700, 1, 2560 ] - Exact: [ 1024, 6000, 1, 1536 ] - Exact: [ 3072, 3000, 1, 1024 ] - Exact: [ 4608, 1500, 1, 1536 ] - Exact: [ 1024, 6000, 1, 2048 ] - Exact: [ 1024, 6000, 1, 2560 ] - Exact: [ 5124, 1500, 1, 2048 ] - Exact: [ 1024, 6000, 1, 2816 ] - Exact: [ 512, 24000, 1, 1536 ] - Exact: [ 3072, 6000, 1, 1024 ] - Exact: [ 5124, 1500, 1, 2560 ] - Exact: [ 4608, 3000, 1, 1536 ] - Exact: [ 1760, 7000, 1, 1760 ] - Exact: [ 6144, 1500, 1, 2560 ] - Exact: [ 512, 24000, 1, 2048 ] - Exact: [ 2048, 7000, 1, 2048 ] - Exact: [ 7680, 1500, 1, 2560 ] - Exact: [ 512, 24000, 1, 2560 ] - Exact: [ 512, 24000, 1, 2816 ] - Exact: [ 8448, 1500, 1, 2816 ] - Exact: [ 512, 48000, 1, 1536 ] - Exact: [ 1024, 24000, 1, 1536 ] - Exact: [ 4608, 6000, 1, 1536 ] - Exact: [ 2560, 7000, 1, 2560 ] - Exact: [ 6144, 3000, 1, 2560 ] - Exact: [ 512, 48000, 1, 2048 ] - Exact: [ 1024, 24000, 1, 2048 ] - Exact: [ 7680, 3000, 1, 2560 ] - Exact: [ 512, 48000, 1, 2560 ] - Exact: [ 1024, 24000, 1, 2560 ] - Exact: [ 512, 48000, 1, 2816 ] - Exact: [ 1024, 24000, 1, 2816 ] - Exact: [ 8448, 3000, 1, 2816 ] - Exact: [ 1024, 48000, 1, 1536 ] - Exact: [ 3072, 24000, 1, 1024 ] - Exact: [ 5124, 9124, 1, 1760 ] - Exact: [ 6144, 6000, 1, 2560 ] - Exact: [ 5124, 9124, 1, 2048 ] - Exact: [ 1024, 48000, 1, 2048 ] - Exact: [ 4096, 7000, 1, 4096 ] - Exact: [ 7680, 6000, 1, 2560 ] - Exact: [ 5124, 9124, 1, 2560 ] - Exact: [ 1024, 48000, 1, 2560 ] - Exact: [ 1024, 48000, 1, 2816 ] - Exact: [ 8448, 6000, 1, 2816 ] - Exact: [ 3072, 48000, 1, 1024 ] - Exact: [ 4608, 24000, 1, 1536 ] - Exact: [ 5124, 9124, 1, 4096 ] - Exact: [ 4608, 48000, 1, 1536 ] - Exact: [ 6144, 24000, 1, 2560 ] - Exact: [ 7680, 24000, 1, 2560 ] - Exact: [ 8448, 24000, 1, 2816 ] - Exact: [ 6144, 48000, 1, 2560 ] - Exact: [ 7680, 48000, 1, 2560 ] - Exact: [ 8448, 48000, 1, 2816 ] ######################################## # NN - VectorWidth Correctness ######################################## - # Benchmark Group InitialSolutionParameters: BenchmarkCommonParameters: - EdgeType: ["ShiftPtr"] - LoopTail: [True] - KernelLanguage: ["Source"] - PrefetchGlobalRead: [True] - PrefetchLocalRead: [True] - WorkGroupMapping: [1] ForkParameters: - ThreadTile: - [ 2, 2 ] - [ 4, 4 ] - WorkGroup: - [ 8, 8, 1 ] - GlobalSplitU: [1] - DepthU: [ -1 ] - VectorWidth: [-1] BenchmarkForkParameters: JoinParameters: BenchmarkJoinParameters: BenchmarkFinalParameters: - ProblemSizes: - Range: [ [4], [4], [1], [128] ] # corner, source code for k<=128 - Range: [ [4], [64, 64, 64, 7000], [1], [128] ] # skinny-0, source code for k<=128 - Range: [ [64, 64, 64, 7000], [4], [1], [128] ] # skinny-1, source code for k<=128 - # Benchmark Group InitialSolutionParameters: BenchmarkCommonParameters: - EdgeType: ["ShiftPtr"] - LoopTail: [True] - KernelLanguage: ["Assembly"] - PrefetchGlobalRead: [True] - PrefetchLocalRead: [True] - WorkGroupMapping: [1] ForkParameters: - ThreadTile: - [ 2, 2 ] - [ 4, 4 ] - WorkGroup: - [ 8, 8, 1 ] - [ 8, 8, 2 ] - [ 4, 4, 4 ] - [ 8, 8, 4 ] - GlobalSplitU: [1, 2, 4, 8, 16] - DepthU: [ -1 ] - VectorWidth: [-1] BenchmarkForkParameters: JoinParameters: BenchmarkJoinParameters: BenchmarkFinalParameters: - ProblemSizes: - Range: [ [4], [4], [1], [256, 1024, 1024, 4096] ] # corner - Range: [ [4], [64, 64, 64, 7000], [1], [256, 1024, 1024, 4096] ] # skinny-0 - Range: [ [64, 64, 64, 7000], [4], [1], [256, 1024, 1024, 4096] ] # skinny-1 ######################################## # NN - VGPR refactor ######################################## - # Benchmark Group InitialSolutionParameters: BenchmarkCommonParameters: - EdgeType: ["ShiftPtr"] - LoopTail: [True] - KernelLanguage: ["Assembly"] - PrefetchGlobalRead: [ True ] - PrefetchLocalRead: [ True ] ForkParameters: - ThreadTile: - [ 2, 2 ] - [ 2, 4 ] - [ 4, 2 ] - [ 4, 4 ] - [ 4, 8 ] - [ 8, 4 ] - [ 8, 8 ] - WorkGroup: - [ 8, 8, 1 ] - [ 16, 8, 1 ] - [ 16, 16, 1 ] - [ 32, 8, 1 ] - [ 32, 16, 1 ] - WorkGroupMapping: [1, 8] - DepthU: [ 8, 16, 24, 32, 64 ] - VectorWidth: [1,2,4] - GlobalReadVectorWidth: [4] - LdsPadB: [0,1,2,4] BenchmarkForkParameters: JoinParameters: BenchmarkJoinParameters: BenchmarkFinalParameters: - ProblemSizes: - Exact: [ 256, 193600, 1, 64 ] - Exact: [ 64, 193600, 1, 64 ] - Exact: [ 64, 193600, 1, 256 ] - Exact: [ 512, 50176, 1, 128 ] - Exact: [ 128, 50176, 1, 512 ] - Exact: [ 256, 12544, 1, 1024 ] - Exact: [ 1024, 12544, 1, 256 ] - Exact: [ 2048, 3136, 1, 512 ] - Exact: [ 512, 3136, 1, 2048 ] ######################################## # NN - Batch ######################################## - # Benchmark Group - ResNet 1x1: InitialSolutionParameters: BenchmarkCommonParameters: - EdgeType: ["ShiftPtr"] - LoopTail: [True] - KernelLanguage: ["Assembly"] ForkParameters: - PrefetchGlobalRead: [ False, True ] - PrefetchLocalRead: [ False] - ThreadTile: - [ 4, 4 ] - [ 4, 8 ] - [ 8, 4 ] - WorkGroup: - [ 16, 8, 1 ] - [ 8, 32, 1 ] - [ 16, 16, 1 ] - [ 32, 8, 1 ] - WorkGroupMapping: [8] - DepthU: [ 16 ] - VectorWidth: [1,2,4] - GlobalReadVectorWidth: [1,4] - LdsPadB: [0,1,2,4] BenchmarkForkParameters: JoinParameters: BenchmarkJoinParameters: BenchmarkFinalParameters: - ProblemSizes: - Exact: [ 196, 256, 64, 1024 ] - Exact: [ 784, 512, 64, 128 ] - Exact: [ 49, 512, 64, 2048 ] - Exact: [ 3136, 512, 1, 2048 ] - Exact: [ 196, 1024, 64, 256 ] - Exact: [ 49, 2048, 64, 512 ] - Exact: [ 3136, 2048, 1, 512 ] - Exact: [ 3025, 256, 64, 64 ] - Exact: [ 3025, 64, 64, 64 ] ######################################## ######################################## ### ### TN ### ######################################## ######################################## - - # ProblemType OperationType: GEMM DataType: s TransposeA: True TransposeB: False UseBeta: True Batched: True ######################################## # TN - Super Skinny ######################################## - # Benchmark Group InitialSolutionParameters: BenchmarkCommonParameters: - EdgeType: ["ShiftPtr"] - LoopTail: [True] - KernelLanguage: ["Assembly"] - PrefetchLocalRead: [True] - WorkGroupMapping: [1] ForkParameters: - PrefetchGlobalRead: [False, True] - ThreadTile: - [ 2, 2 ] - [ 4, 2 ] - [ 8, 2 ] - [ 4, 4 ] - WorkGroup: - [ 32, 4, 2 ] - [ 16, 4, 4 ] - [ 32, 2, 4 ] - [ 8, 4, 8 ] - [ 16, 2, 8 ] - GlobalSplitU: [16, 32] - DepthU: [ 16, 32, 64 ] - VectorWidth: [-1] BenchmarkForkParameters: JoinParameters: BenchmarkJoinParameters: BenchmarkFinalParameters: - ProblemSizes: - Exact: [ 512, 8, 1, 500000 ] - Exact: [ 512, 16, 1, 500000 ] - Exact: [ 1024, 8, 1, 500000 ] - Exact: [ 1024, 16, 1, 500000 ] - Exact: [ 1760, 16, 1, 1760 ] - Exact: [ 3072, 16, 1, 1024 ] - Exact: [ 2048, 16, 1, 2048 ] - Exact: [ 2560, 16, 1, 2560 ] - Exact: [ 4608, 16, 1, 1536 ] - Exact: [ 6144, 16, 1, 2560 ] - Exact: [ 4096, 16, 1, 4096 ] - Exact: [ 7680, 16, 1, 2560 ] - Exact: [ 8448, 16, 1, 2816 ] ######################################## # TN - 3 Waves/WG ######################################## - # Benchmark Group InitialSolutionParameters: BenchmarkCommonParameters: - EdgeType: ["ShiftPtr"] - LoopTail: [True] - KernelLanguage: ["Assembly"] - PrefetchGlobalRead: [True] - PrefetchLocalRead: [True] - WorkGroupMapping: [1] ForkParameters: - ThreadTile: - [ 3, 3 ] - [ 3, 4 ] - [ 3, 6 ] - [ 4, 3 ] - [ 4, 4 ] - [ 4, 6 ] - [ 6, 3 ] - [ 6, 4 ] - [ 6, 6 ] - WorkGroup: - [ 24, 8, 1 ] - [ 12, 16, 1 ] - [ 6, 32, 1 ] - [ 8, 12, 2 ] - [ 8, 6, 4 ] - [ 16, 16, 1 ] - GlobalSplitU: [2, 4, 8] - DepthU: [ 8, 16, 24, 32 ] - VectorWidth: [-1] BenchmarkForkParameters: JoinParameters: BenchmarkJoinParameters: BenchmarkFinalParameters: - ProblemSizes: - Exact: [ 35, 8457, 1, 1760 ] - Exact: [ 35, 8457, 1, 2048 ] - Exact: [ 35, 8457, 1, 2560 ] - Exact: [ 35, 8457, 1, 4096 ] ######################################## # TN - Small or Skinny ######################################## - # Benchmark Group InitialSolutionParameters: BenchmarkCommonParameters: - EdgeType: ["ShiftPtr"] - LoopTail: [True] - KernelLanguage: ["Source"] - PrefetchGlobalRead: [True] - PrefetchLocalRead: [True] ForkParameters: - ThreadTile: - [ 2, 2 ] - [ 4, 2 ] - [ 4, 4 ] - [ 8, 4 ] - [ 8, 8 ] - WorkGroup: - [ 32, 4, 1 ] - [ 16, 4, 1 ] - [ 8, 8, 1 ] - [ 16, 8, 1 ] - WorkGroupMapping: [1, 8] - GlobalSplitU: [1] - DepthU: [ 8, 16, 24, 32 ] - VectorWidth: [-1] BenchmarkForkParameters: JoinParameters: BenchmarkJoinParameters: BenchmarkFinalParameters: - ProblemSizes: - Range: [ [64, 128], [64, 64, 64, 7000], [1], [128] ] # skinny-0, source code for k<128 - Range: [ [64, 64, 64, 7000], [64, 128], [1], [128] ] # skinny-1, source code for k<128 - Range: [ [64, 64, 64, 700], [64, 64, 64, 700], [1], [128] ] # small, source code for k<128 - # Benchmark Group InitialSolutionParameters: BenchmarkCommonParameters: - EdgeType: ["ShiftPtr"] - LoopTail: [True] - KernelLanguage: ["Assembly"] - PrefetchGlobalRead: [True] - PrefetchLocalRead: [True] ForkParameters: - ThreadTile: - [ 2, 2 ] - [ 4, 2 ] - [ 4, 4 ] - [ 8, 4 ] - [ 8, 8 ] - WorkGroup: - [ 32, 4, 2 ] - [ 16, 4, 4 ] - [ 8, 8, 2 ] - [ 8, 8, 4 ] - [ 16, 8, 2 ] - WorkGroupMapping: [1, 8] - GlobalSplitU: [2, 4, 8] - DepthU: [ 8, 16, 24, 32 ] - VectorWidth: [-1] BenchmarkForkParameters: JoinParameters: BenchmarkJoinParameters: BenchmarkFinalParameters: - ProblemSizes: - Range: [ [64, 128], [64, 64, 64, 7000], [1], [256, 1024, 1024, 4096] ] # skinny-0 - Range: [ [64, 64, 64, 7000], [64, 128], [1], [256, 1024, 1024, 4096] ] # skinny-1 - Range: [ [64, 64, 64, 700], [64, 64, 64, 700], [1], [256, 1024, 1024, 4096] ] # small - Exact: [ 1760, 16, 1, 1760 ] - Exact: [ 3072, 16, 1, 1024 ] - Exact: [ 2048, 16, 1, 2048 ] - Exact: [ 1760, 32, 1, 1760 ] - Exact: [ 3072, 32, 1, 1024 ] - Exact: [ 2560, 16, 1, 2560 ] - Exact: [ 4608, 16, 1, 1536 ] - Exact: [ 2048, 32, 1, 2048 ] - Exact: [ 1760, 64, 1, 1760 ] - Exact: [ 3072, 64, 1, 1024 ] - Exact: [ 2560, 32, 1, 2560 ] - Exact: [ 4608, 32, 1, 1536 ] - Exact: [ 6144, 16, 1, 2560 ] - Exact: [ 2048, 64, 1, 2048 ] - Exact: [ 4096, 16, 1, 4096 ] - Exact: [ 7680, 16, 1, 2560 ] - Exact: [ 8448, 16, 1, 2816 ] - Exact: [ 1760, 128, 1, 1760 ] - Exact: [ 3072, 128, 1, 1024 ] - Exact: [ 2560, 64, 1, 2560 ] - Exact: [ 6144, 32, 1, 2560 ] - Exact: [ 35, 8457, 1, 1760 ] - Exact: [ 2048, 128, 1, 2048 ] - Exact: [ 4096, 32, 1, 4096 ] - Exact: [ 35, 8457, 1, 2048 ] - Exact: [ 7680, 32, 1, 2560 ] - Exact: [ 35, 8457, 1, 2560 ] - Exact: [ 8448, 32, 1, 2816 ] - Exact: [ 2560, 128, 1, 2560 ] - Exact: [ 4096, 64, 1, 4096 ] - Exact: [ 35, 8457, 1, 4096 ] - Exact: [ 7680, 64, 1, 2560 ] - Exact: [ 4096, 128, 1, 4096 ] - Exact: [ 7680, 128, 1, 2560 ] ######################################## # TN - Large ######################################## - # Benchmark Group InitialSolutionParameters: BenchmarkCommonParameters: - EdgeType: ["ShiftPtr"] - LoopTail: [True] - KernelLanguage: ["Source"] - PrefetchGlobalRead: [True] - PrefetchLocalRead: [True] - GlobalSplitU: [1] ForkParameters: - ThreadTile: - [ 4, 4 ] - [ 4, 8 ] - [ 6, 8 ] - [ 8, 4 ] - [ 8, 6 ] - [ 8, 8 ] - WorkGroup: - [ 16, 16, 1 ] - WorkGroupMapping: [1, 8] - DepthU: [ 8, 16, 24, 32 ] - VectorWidth: [-1] BenchmarkForkParameters: JoinParameters: BenchmarkJoinParameters: BenchmarkFinalParameters: - ProblemSizes: - Range: [ [64, 64, 64, 7000], [64, 64, 64, 7000], [1], [128] ] # large, source code for k<=128 - # Benchmark Group InitialSolutionParameters: BenchmarkCommonParameters: - EdgeType: ["ShiftPtr"] - LoopTail: [True] - KernelLanguage: ["Assembly"] - PrefetchGlobalRead: [True] - PrefetchLocalRead: [True] - GlobalSplitU: [1] ForkParameters: - ThreadTile: - [ 4, 4 ] - [ 4, 8 ] - [ 6, 8 ] - [ 8, 4 ] - [ 8, 6 ] - [ 8, 8 ] - WorkGroup: - [ 16, 16, 1 ] - WorkGroupMapping: [1, 8] - DepthU: [ 8, 16, 24, 32 ] - VectorWidth: [-1] BenchmarkForkParameters: JoinParameters: BenchmarkJoinParameters: BenchmarkFinalParameters: - ProblemSizes: - Range: [ [64, 64, 64, 7000], [64, 64, 64, 7000], [1], [256, 1024, 1024, 4096] ] # large - Exact: [ 1024, 1024, 1, 1024 ] # MIOpen sizes - Exact: [ 1760, 800, 1, 1760 ] - Exact: [ 1760, 1600, 1, 1760 ] - Exact: [ 1760, 3200, 1, 1760 ] - Exact: [ 1760, 6400, 1, 1760 ] - Exact: [ 2048, 800, 1, 2048 ] - Exact: [ 2048, 1600, 1, 2048 ] - Exact: [ 2048, 3200, 1, 2048 ] - Exact: [ 2048, 6400, 1, 2048 ] - Exact: [ 2560, 800, 1, 2560 ] - Exact: [ 2560, 1600, 1, 2560 ] - Exact: [ 2560, 3200, 1, 2560 ] - Exact: [ 2560, 6400, 1, 2560 ] - Exact: [ 2048, 400, 1, 512 ] - Exact: [ 2048, 800, 1, 512 ] - Exact: [ 2048, 1600, 1, 512 ] - Exact: [ 2048, 3200, 1, 512 ] - Exact: [ 4096, 400, 1, 1024 ] - Exact: [ 4096, 800, 1, 1024 ] - Exact: [ 4096, 1600, 1, 1024 ] - Exact: [ 4096, 3200, 1, 1024 ] - Exact: [ 8192, 400, 1, 2048 ] - Exact: [ 8192, 800, 1, 2048 ] - Exact: [ 8192, 1600, 1, 2048 ] - Exact: [ 8192, 3200, 1, 2048 ] - Exact: [ 16384, 400, 1, 4096 ] - Exact: [ 16384, 800, 1, 4096 ] - Exact: [ 16384, 1600, 1, 4096 ] - Exact: [ 16384, 3200, 1, 4096 ] - Exact: [ 8448, 48000, 1, 2816 ] - Exact: [ 8448, 24000, 1, 2816 ] - Exact: [ 8448, 12000, 1, 2816 ] - Exact: [ 8448, 5984, 1, 2816 ] - Exact: [ 6144, 48000, 1, 2048 ] - Exact: [ 6144, 24000, 1, 2048 ] - Exact: [ 6144, 12000, 1, 2048 ] - Exact: [ 6144, 5984, 1, 2048 ] - Exact: [ 4608, 48000, 1, 1536 ] - Exact: [ 4608, 24000, 1, 1536 ] - Exact: [ 4608, 12000, 1, 1536 ] - Exact: [ 4608, 5984, 1, 1536 ] - Exact: [ 7680, 48000, 1, 2560 ] - Exact: [ 7680, 24000, 1, 2560 ] - Exact: [ 7680, 12000, 1, 2560 ] - Exact: [ 7680, 5984, 1, 2560 ] # Deepbench sizes - Exact: [ 1024, 700, 1, 512 ] - Exact: [ 512, 24000, 1, 1536 ] - Exact: [ 1760, 7000, 1, 1760 ] - Exact: [ 512, 24000, 1, 2048 ] - Exact: [ 2048, 7000, 1, 2048 ] - Exact: [ 512, 24000, 1, 2560 ] - Exact: [ 512, 24000, 1, 2816 ] - Exact: [ 512, 48000, 1, 1536 ] - Exact: [ 1024, 24000, 1, 1536 ] - Exact: [ 2560, 7000, 1, 2560 ] - Exact: [ 512, 48000, 1, 2048 ] - Exact: [ 1024, 24000, 1, 2048 ] - Exact: [ 512, 48000, 1, 2560 ] - Exact: [ 1024, 24000, 1, 2560 ] - Exact: [ 512, 48000, 1, 2816 ] - Exact: [ 1024, 24000, 1, 2816 ] - Exact: [ 1024, 48000, 1, 1536 ] - Exact: [ 3072, 24000, 1, 1024 ] - Exact: [ 5124, 9124, 1, 1760 ] - Exact: [ 5124, 9124, 1, 2048 ] - Exact: [ 1024, 48000, 1, 2048 ] - Exact: [ 4096, 7000, 1, 4096 ] - Exact: [ 5124, 9124, 1, 2560 ] - Exact: [ 1024, 48000, 1, 2560 ] - Exact: [ 1024, 48000, 1, 2816 ] - Exact: [ 3072, 48000, 1, 1024 ] - Exact: [ 4608, 24000, 1, 1536 ] - Exact: [ 5124, 9124, 1, 4096 ] - Exact: [ 4608, 48000, 1, 1536 ] - Exact: [ 6144, 24000, 1, 2560 ] - Exact: [ 7680, 24000, 1, 2560 ] - Exact: [ 8448, 24000, 1, 2816 ] - Exact: [ 6144, 48000, 1, 2560 ] - Exact: [ 7680, 48000, 1, 2560 ] - Exact: [ 8448, 48000, 1, 2816 ] ######################################## # TN - VectorWidth Correctness ######################################## - # Benchmark Group InitialSolutionParameters: BenchmarkCommonParameters: - EdgeType: ["ShiftPtr"] - LoopTail: [True] - KernelLanguage: ["Source"] - PrefetchGlobalRead: [True] - PrefetchLocalRead: [True] - WorkGroupMapping: [1] ForkParameters: - ThreadTile: - [ 2, 2 ] - [ 4, 4 ] - WorkGroup: - [ 8, 8, 1 ] - GlobalSplitU: [1] - DepthU: [ -1 ] - VectorWidth: [-1] BenchmarkForkParameters: JoinParameters: BenchmarkJoinParameters: BenchmarkFinalParameters: - ProblemSizes: - Range: [ [4], [4], [1], [128] ] # corner, source code for k<=128 - Range: [ [4], [64, 64, 64, 7000], [1], [128] ] # skinny-0, source code for k<=128 - Range: [ [64, 64, 64, 7000], [4], [1], [128] ] # skinny-1, source code for k<=128 - # Benchmark Group InitialSolutionParameters: BenchmarkCommonParameters: - EdgeType: ["ShiftPtr"] - LoopTail: [True] - KernelLanguage: ["Assembly"] - PrefetchGlobalRead: [True] - PrefetchLocalRead: [True] - WorkGroupMapping: [1] ForkParameters: - ThreadTile: - [ 2, 2 ] - [ 4, 4 ] - WorkGroup: - [ 8, 8, 1 ] - [ 8, 8, 2 ] - [ 4, 4, 4 ] - [ 8, 8, 4 ] - GlobalSplitU: [1, 2, 4, 8, 16] - DepthU: [ -1 ] - VectorWidth: [-1] BenchmarkForkParameters: JoinParameters: BenchmarkJoinParameters: BenchmarkFinalParameters: - ProblemSizes: - Range: [ [4], [4], [1], [256, 1024, 1024, 4096] ] # corner - Range: [ [4], [64, 64, 64, 7000], [1], [256, 1024, 1024, 4096] ] # skinny-0 - Range: [ [64, 64, 64, 7000], [4], [1], [256, 1024, 1024, 4096] ] # skinny-1 ######################################## ######################################## ### ### NT ### ######################################## ######################################## - - # ProblemType OperationType: GEMM DataType: s TransposeA: False TransposeB: True UseBeta: True Batched: True ######################################## # NT - Small or Skinny ######################################## - # Benchmark Group InitialSolutionParameters: BenchmarkCommonParameters: - EdgeType: ["ShiftPtr"] - LoopTail: [True] - KernelLanguage: ["Source"] - PrefetchGlobalRead: [True] - PrefetchLocalRead: [True] ForkParameters: - ThreadTile: - [ 2, 2 ] - [ 4, 2 ] - [ 4, 4 ] - [ 8, 4 ] - [ 8, 8 ] - WorkGroup: - [ 32, 4, 1 ] - [ 16, 4, 1 ] - [ 8, 8, 1 ] - [ 16, 8, 1 ] - WorkGroupMapping: [1, 8] - GlobalSplitU: [1] - DepthU: [ 8, 16, 24, 32 ] - VectorWidth: [-1] BenchmarkForkParameters: JoinParameters: BenchmarkJoinParameters: BenchmarkFinalParameters: - ProblemSizes: - Range: [ [64, 128], [64, 64, 64, 7000], [1], [128] ] # skinny-0, source code for k<=128 - Range: [ [64, 64, 64, 7000], [64, 128], [1], [128] ] # skinny-1, source code for k<=128 - Range: [ [64, 64, 64, 700], [64, 64, 64, 700], [1], [128] ] # small, source code for k<=128 - # Benchmark Group InitialSolutionParameters: BenchmarkCommonParameters: - EdgeType: ["ShiftPtr"] - LoopTail: [True] - KernelLanguage: ["Assembly"] - PrefetchGlobalRead: [True] - PrefetchLocalRead: [True] ForkParameters: - ThreadTile: - [ 2, 2 ] - [ 4, 2 ] - [ 4, 4 ] - [ 8, 4 ] - [ 8, 8 ] - WorkGroup: - [ 32, 4, 2 ] - [ 16, 4, 4 ] - [ 8, 8, 2 ] - [ 8, 8, 4 ] - [ 16, 8, 2 ] - WorkGroupMapping: [1, 8] - GlobalSplitU: [2, 4, 8] - DepthU: [ 8, 16, 24, 32 ] - VectorWidth: [-1] BenchmarkForkParameters: JoinParameters: BenchmarkJoinParameters: BenchmarkFinalParameters: - ProblemSizes: - Range: [ [64, 128], [64, 64, 64, 7000], [1], [256, 1024, 1024, 4096] ] # skinny-0 - Range: [ [64, 64, 64, 7000], [64, 128], [1], [256, 1024, 1024, 4096] ] # skinny-1 - Range: [ [64, 64, 64, 700], [64, 64, 64, 700], [1], [256, 1024, 1024, 4096] ] # small - Exact: [ 512, 16, 1, 512 ] - Exact: [ 512, 32, 1, 512 ] - Exact: [ 1024, 16, 1, 512 ] - Exact: [ 1024, 32, 1, 512 ] ######################################## # NT - Large ######################################## - # Benchmark Group InitialSolutionParameters: BenchmarkCommonParameters: - EdgeType: ["ShiftPtr"] - LoopTail: [True] - KernelLanguage: ["Source"] - PrefetchGlobalRead: [True] - PrefetchLocalRead: [True] - GlobalSplitU: [1] ForkParameters: - ThreadTile: - [ 4, 4 ] - [ 4, 8 ] - [ 6, 8 ] - [ 8, 4 ] - [ 8, 6 ] - [ 8, 8 ] - WorkGroup: - [ 16, 16, 1 ] - WorkGroupMapping: [1, 8] - DepthU: [ 8, 16 ] - VectorWidth: [-1] BenchmarkForkParameters: JoinParameters: BenchmarkJoinParameters: BenchmarkFinalParameters: - ProblemSizes: - Range: [ [64, 64, 64, 7000], [64, 64, 64, 7000], [1], [128] ] # large, source code for k<=128 - # Benchmark Group InitialSolutionParameters: BenchmarkCommonParameters: - EdgeType: ["ShiftPtr"] - LoopTail: [True] - KernelLanguage: ["Assembly"] - PrefetchGlobalRead: [True] - PrefetchLocalRead: [True] - GlobalSplitU: [1] ForkParameters: - ThreadTile: - [ 4, 4 ] - [ 4, 8 ] - [ 6, 8 ] - [ 8, 4 ] - [ 8, 6 ] - [ 8, 8 ] - WorkGroup: - [ 16, 16, 1 ] - WorkGroupMapping: [1, 8] - DepthU: [ 8, 16 ] - VectorWidth: [-1] BenchmarkForkParameters: JoinParameters: BenchmarkJoinParameters: BenchmarkFinalParameters: - ProblemSizes: - Range: [ [64, 64, 64, 7000], [64, 64, 64, 7000], [1], [256, 1024, 1024, 4096] ] # large - Exact: [ 1024, 1024, 1, 1024 ] - Exact: [ 1760, 7133, 1, 1760 ] - Exact: [ 2048, 7133, 1, 2048 ] - Exact: [ 2560, 7133, 1, 2560 ] - Exact: [ 3072, 7435, 1, 1024 ] - Exact: [ 4096, 7133, 1, 4096 ] - Exact: [ 7680, 5481, 1, 2560 ] ######################################## # NT - VectorWidth Correctness ######################################## - # Benchmark Group InitialSolutionParameters: BenchmarkCommonParameters: - EdgeType: ["ShiftPtr"] - LoopTail: [True] - KernelLanguage: ["Source"] - PrefetchGlobalRead: [True] - PrefetchLocalRead: [True] - WorkGroupMapping: [1] ForkParameters: - ThreadTile: - [ 2, 2 ] - [ 4, 4 ] - WorkGroup: - [ 8, 8, 1 ] - GlobalSplitU: [1] - DepthU: [ -1 ] - VectorWidth: [-1] BenchmarkForkParameters: JoinParameters: BenchmarkJoinParameters: BenchmarkFinalParameters: - ProblemSizes: - Range: [ [4], [4], [1], [128] ] # corner, source code for k<=128 - Range: [ [4], [64, 64, 64, 7000], [1], [128] ] # skinny-0, source code for k<=128 - Range: [ [64, 64, 64, 7000], [4], [1], [128] ] # skinny-1, source code for k<=128 - # Benchmark Group InitialSolutionParameters: BenchmarkCommonParameters: - EdgeType: ["ShiftPtr"] - LoopTail: [True] - KernelLanguage: ["Assembly"] - PrefetchGlobalRead: [True] - PrefetchLocalRead: [True] - WorkGroupMapping: [1] ForkParameters: - ThreadTile: - [ 2, 2 ] - [ 4, 4 ] - WorkGroup: - [ 8, 8, 1 ] - [ 8, 8, 2 ] - [ 4, 4, 4 ] - [ 8, 8, 4 ] - GlobalSplitU: [1, 2, 4, 8, 16] - DepthU: [ -1 ] - VectorWidth: [-1] BenchmarkForkParameters: JoinParameters: BenchmarkJoinParameters: BenchmarkFinalParameters: - ProblemSizes: - Range: [ [4], [4], [1], [256, 1024, 1024, 4096] ] # corner - Range: [ [4], [64, 64, 64, 7000], [1], [256, 1024, 1024, 4096] ] # skinny-0 - Range: [ [64, 64, 64, 7000], [4], [1], [256, 1024, 1024, 4096] ] # skinny-1 ######################################## ######################################## ### ### TT ### ######################################## ######################################## - - # ProblemType OperationType: GEMM DataType: s TransposeA: True TransposeB: True UseBeta: True Batched: True ######################################## # TT - Small or Skinny ######################################## - # Benchmark Group InitialSolutionParameters: BenchmarkCommonParameters: - EdgeType: ["ShiftPtr"] - LoopTail: [True] - KernelLanguage: ["Source"] - PrefetchGlobalRead: [True] - PrefetchLocalRead: [True] ForkParameters: - ThreadTile: - [ 2, 2 ] - [ 2, 4 ] - [ 4, 4 ] - [ 4, 8 ] - [ 8, 8 ] - WorkGroup: - [ 32, 4, 1 ] - [ 8, 8, 1 ] - [ 16, 16, 1 ] - [ 16, 8, 1 ] - [ 8, 16, 1 ] - WorkGroupMapping: [1, 4] - GlobalSplitU: [1] - DepthU: [ 8 ] - VectorWidth: [-1] BenchmarkForkParameters: JoinParameters: BenchmarkJoinParameters: BenchmarkFinalParameters: - ProblemSizes: - Range: [ [64, 128], [64, 64, 64, 7000], [1], [128] ] # skinny-0, source code for k<=128 - Range: [ [64, 64, 64, 7000], [64, 128], [1], [128] ] # skinny-1, source code for k<=128 - Range: [ [64, 64, 64, 700], [64, 64, 64, 700], [1], [128] ] # small, source code for k<=128 - # Benchmark Group InitialSolutionParameters: BenchmarkCommonParameters: - EdgeType: ["ShiftPtr"] - LoopTail: [True] - KernelLanguage: ["Assembly"] - PrefetchGlobalRead: [True] - PrefetchLocalRead: [True] ForkParameters: - ThreadTile: - [ 2, 2 ] - [ 2, 4 ] - [ 4, 4 ] - [ 4, 8 ] - [ 8, 8 ] - WorkGroup: - [ 32, 4, 2 ] - [ 4, 8, 4 ] - [ 8, 8, 2 ] - [ 16, 8, 4 ] - [ 8, 16, 2 ] - WorkGroupMapping: [1, 4] - GlobalSplitU: [2, 4, 8] - DepthU: [ 8, 16, 24, 32 ] - VectorWidth: [-1] BenchmarkForkParameters: JoinParameters: BenchmarkJoinParameters: BenchmarkFinalParameters: - ProblemSizes: - Range: [ [64, 128], [64, 64, 64, 7000], [1], [256, 1024, 1024, 4096] ] # skinny-0 - Range: [ [64, 64, 64, 7000], [64, 128], [1], [256, 1024, 1024, 4096] ] # skinny-1 - Range: [ [64, 64, 64, 700], [64, 64, 64, 700], [1], [256, 1024, 1024, 4096] ] # small ######################################## # TT - Large ######################################## - # Benchmark Group InitialSolutionParameters: BenchmarkCommonParameters: - EdgeType: ["ShiftPtr"] - LoopTail: [True] - KernelLanguage: ["Source"] - PrefetchLocalRead: [True] - GlobalSplitU: [1] ForkParameters: - PrefetchGlobalRead: [False, True] - ThreadTile: - [ 4, 4 ] - [ 4, 8 ] - [ 6, 8 ] - [ 8, 4 ] - [ 8, 6 ] - [ 8, 8 ] - WorkGroup: - [ 16, 16, 1 ] - WorkGroupMapping: [1, 4] - DepthU: [ 8 ] - VectorWidth: [-1] BenchmarkForkParameters: JoinParameters: BenchmarkJoinParameters: BenchmarkFinalParameters: - ProblemSizes: - Range: [ [64, 64, 64, 7000], [64, 64, 64, 7000], [1], [128] ] # large, source code for k<=128 - # Benchmark Group InitialSolutionParameters: BenchmarkCommonParameters: - EdgeType: ["ShiftPtr"] - LoopTail: [True] - KernelLanguage: ["Assembly"] - PrefetchLocalRead: [True] - GlobalSplitU: [1] ForkParameters: - PrefetchGlobalRead: [False, True] - ThreadTile: - [ 4, 4 ] - [ 4, 8 ] - [ 6, 8 ] - [ 8, 4 ] - [ 8, 6 ] - [ 8, 8 ] - WorkGroup: - [ 16, 16, 1 ] - WorkGroupMapping: [1, 4] - DepthU: [ 8, 16, 24, 32 ] - VectorWidth: [-1] BenchmarkForkParameters: JoinParameters: BenchmarkJoinParameters: BenchmarkFinalParameters: - ProblemSizes: - Range: [ [64, 64, 64, 7000], [64, 64, 64, 7000], [1], [256, 1024, 1024, 4096] ] # large - Exact: [ 1024, 1024, 1, 1024 ] ######################################## # TT - VectorWidth Correctness ######################################## - # Benchmark Group InitialSolutionParameters: BenchmarkCommonParameters: - EdgeType: ["ShiftPtr"] - LoopTail: [True] - KernelLanguage: ["Source"] - PrefetchGlobalRead: [True] - PrefetchLocalRead: [True] - WorkGroupMapping: [1] ForkParameters: - ThreadTile: - [ 2, 2 ] - [ 4, 4 ] - WorkGroup: - [ 8, 8, 1 ] - GlobalSplitU: [1] - DepthU: [ -1 ] - VectorWidth: [-1] BenchmarkForkParameters: JoinParameters: BenchmarkJoinParameters: BenchmarkFinalParameters: - ProblemSizes: - Range: [ [4], [4], [1], [128] ] # corner, source code for k<=128 - Range: [ [4], [64, 64, 64, 7000], [1], [128] ] # skinny-0, source code for k<=128 - Range: [ [64, 64, 64, 7000], [4], [1], [128] ] # skinny-1, source code for k<=128 - # Benchmark Group InitialSolutionParameters: BenchmarkCommonParameters: - EdgeType: ["ShiftPtr"] - LoopTail: [True] - KernelLanguage: ["Assembly"] - PrefetchGlobalRead: [True] - PrefetchLocalRead: [True] - WorkGroupMapping: [1] ForkParameters: - ThreadTile: - [ 2, 2 ] - [ 4, 4 ] - WorkGroup: - [ 8, 8, 1 ] - [ 8, 8, 2 ] - [ 4, 4, 4 ] - [ 8, 8, 4 ] - GlobalSplitU: [1, 2, 4, 8, 16] - DepthU: [ -1 ] - VectorWidth: [-1] BenchmarkForkParameters: JoinParameters: BenchmarkJoinParameters: BenchmarkFinalParameters: - ProblemSizes: - Range: [ [4], [4], [1], [256, 1024, 1024, 4096] ] # corner - Range: [ [4], [64, 64, 64, 7000], [1], [256, 1024, 1024, 4096] ] # skinny-0 - Range: [ [64, 64, 64, 7000], [4], [1], [256, 1024, 1024, 4096] ] # skinny-1 LibraryLogic: # ScheduleName: "vega20" # DeviceNames: ["Device 66a0", "Device 66a1", "Device 66a7", "Device 66af", "Vega 20"] # ArchitectureName: "gfx906" ScheduleName: "vega10" DeviceNames: ["Device 6863", "Device 6862", "Device 687f", "Device 6860", "Device 6861", "Vega 10 XTX [Radeon Vega Frontier Edition]", "Vega [Radeon RX Vega]", Vega, Device 6864, Device 686c] ArchitectureName: "gfx900" # ScheduleName: "mi25" # DeviceNames: ["Device 6860"] # ArchitectureName: "gfx900" # ScheduleName: "r9nano" # DeviceNames: ["Device 7300"] # ArchitectureName: "gfx803" # ScheduleName: "hip" # DeviceNames: ["Device 0000"] # ArchitectureName: "fallback" LibraryClient: