TestParameters: marks: [skip-gfx900, skip-gfx906, skip-gfx1010, skip-gfx1011, skip, emulation1] # not supported by arch GlobalParameters: NumElementsToValidate: 65536 BoundsCheck: True KernelTime: True BenchmarkProblems: ######################################## # NT - standard ######################################## - - # ProblemType OperationType: GEMM DataType: c TransposeA: False TransposeB: True UseBeta: True Batched: True - # BenchmarkProblemSizeGroup - Standard InitialSolutionParameters: BenchmarkCommonParameters: - KernelLanguage: ["Assembly"] - EdgeType: ["ShiftPtr"] ForkParameters: - TransposeLDS: [0, 1] - MatrixInstruction: - [32, 32, 1, 2] # - [32, 32, 2, 1] - PrefetchLocalRead: [0, 1, 3] - PrefetchGlobalRead: [True] - ThreadTile: - [ 2, 32 ] - WorkGroup: - [ 16, 8, 1 ] # - [ 32, 4, 1 ] # - [ 64, 2, 1 ] # - [ 128, 1, 1 ] - WorkGroupMapping: [8] - GlobalSplitU: [1] - InnerUnroll: [2] - DepthU: [8] - VectorWidth: [2, 4] - AssertSummationElementMultiple: [1, 2] - StoreRemapVectorWidth: [2] JoinParameters: BenchmarkJoinParameters: BenchmarkFinalParameters: - ProblemSizes: # - Range: [ [1024], [1024], [1], [1024] ] - Range: [ [1031], [1031], [8], [1031] ] # - # BenchmarkProblemSizeGroup - Standard # InitialSolutionParameters: # BenchmarkCommonParameters: # - KernelLanguage: ["Assembly"] # - EdgeType: ["ShiftPtr"] # ForkParameters: # - TransposeLDS: [0, 1] # - MatrixInstruction: # - [16, 16, 1, 4] # - [16, 16, 4, 1] # - PrefetchLocalRead: [0, 1, 3] # - PrefetchGlobalRead: [True] # - ThreadTile: # - [ 2, 32 ] # - WorkGroup: # - [ 16, 8, 1 ] # - [ 32, 4, 1 ] # - [ 64, 2, 1 ] # - [ 128, 1, 1 ] # - WorkGroupMapping: [8] # - GlobalSplitU: [1] # - InnerUnroll: [2] # - DepthU: [8] # - VectorWidth: [2, 4] # - AssertSummationElementMultiple: [1, 2] # - StoreRemapVectorWidth: [2] # JoinParameters: # BenchmarkJoinParameters: # BenchmarkFinalParameters: # - ProblemSizes: # - Range: [ [1024], [1024], [1], [1024] ] # - Range: [ [1031], [1031], [8], [1031] ] # - # BenchmarkProblemSizeGroup - Standard # InitialSolutionParameters: # BenchmarkCommonParameters: # - KernelLanguage: ["Assembly"] # - EdgeType: ["ShiftPtr"] # ForkParameters: # - TransposeLDS: [0, 1] # - MatrixInstruction: # - [ 4, 4, 1, 16] # - PrefetchLocalRead: [0, 1] # - PrefetchGlobalRead: [True] # - ThreadTile: # - [ 2, 8 ] # - WorkGroup: # - [ 8, 32, 1 ] # - [ 16, 16, 1 ] # - [ 32, 8, 1 ] # - [ 64, 4, 1 ] # - WorkGroupMapping: [8] # - GlobalSplitU: [1] # - InnerUnroll: [2] # - DepthU: [8] # - VectorWidth: [2, 4] # - AssertSummationElementMultiple: [1, 2] # - StoreRemapVectorWidth: [2] # JoinParameters: # BenchmarkJoinParameters: # BenchmarkFinalParameters: # - ProblemSizes: # - Range: [ [1024], [1024], [1], [1024] ] # - Range: [ [1031], [1031], [8], [1031] ] ######################################## # TN - standard ######################################## - - # ProblemType OperationType: GEMM DataType: c TransposeA: True TransposeB: False UseBeta: True Batched: True # - # BenchmarkProblemSizeGroup - Standard # InitialSolutionParameters: # BenchmarkCommonParameters: # - KernelLanguage: ["Assembly"] # - EdgeType: ["ShiftPtr"] # ForkParameters: # - TransposeLDS: [0, 1] # - MatrixInstruction: # - [32, 32, 1, 2] # - [32, 32, 2, 1] # - PrefetchLocalRead: [0, 1, 3] # - PrefetchGlobalRead: [True] # - ThreadTile: # - [ 2, 32 ] # - WorkGroup: # - [ 16, 8, 1 ] # - [ 32, 4, 1 ] # - [ 64, 2, 1 ] # - [ 128, 1, 1 ] # - WorkGroupMapping: [8] # - GlobalSplitU: [1] # - InnerUnroll: [2] # - DepthU: [8] # - VectorWidth: [2, 4] # - AssertSummationElementMultiple: [1, 2] # - StoreRemapVectorWidth: [2] # JoinParameters: # BenchmarkJoinParameters: # BenchmarkFinalParameters: # - ProblemSizes: # - Range: [ [1024], [1024], [1], [1024] ] # - Range: [ [1031], [1031], [8], [1031] ] - # BenchmarkProblemSizeGroup - Standard InitialSolutionParameters: BenchmarkCommonParameters: - KernelLanguage: ["Assembly"] - EdgeType: ["ShiftPtr"] ForkParameters: - TransposeLDS: [0, 1] - MatrixInstruction: - [16, 16, 1, 4] # - [16, 16, 4, 1] - PrefetchLocalRead: [0, 1, 3] - PrefetchGlobalRead: [True] - ThreadTile: - [ 2, 32 ] - WorkGroup: - [ 16, 8, 1 ] # - [ 32, 4, 1 ] # - [ 64, 2, 1 ] # - [ 128, 1, 1 ] - WorkGroupMapping: [8] - GlobalSplitU: [1] - InnerUnroll: [2] - DepthU: [8] - VectorWidth: [2, 4] - AssertSummationElementMultiple: [1, 2] - StoreRemapVectorWidth: [2] JoinParameters: BenchmarkJoinParameters: BenchmarkFinalParameters: - ProblemSizes: - Range: [ [1024], [1024], [1], [1024] ] # - Range: [ [1031], [1031], [8], [1031] ] # - # BenchmarkProblemSizeGroup - Standard # InitialSolutionParameters: # BenchmarkCommonParameters: # - KernelLanguage: ["Assembly"] # - EdgeType: ["ShiftPtr"] # ForkParameters: # - TransposeLDS: [0, 1] # - MatrixInstruction: # - [ 4, 4, 1, 16] # - PrefetchLocalRead: [0, 1] # - PrefetchGlobalRead: [True] # - ThreadTile: # - [ 2, 8 ] # - WorkGroup: # - [ 8, 32, 1 ] # - [ 16, 16, 1 ] # - [ 32, 8, 1 ] # - [ 64, 4, 1 ] # - WorkGroupMapping: [8] # - GlobalSplitU: [1] # - InnerUnroll: [2] # - DepthU: [8] # - VectorWidth: [2, 4] # - AssertSummationElementMultiple: [1, 2] # - StoreRemapVectorWidth: [2] # JoinParameters: # BenchmarkJoinParameters: # BenchmarkFinalParameters: # - ProblemSizes: # - Range: [ [1024], [1024], [1], [1024] ] # - Range: [ [1031], [1031], [8], [1031] ] ######################################## # NN - standard ######################################## - - # ProblemType OperationType: GEMM DataType: c TransposeA: False TransposeB: False UseBeta: True Batched: True - # BenchmarkProblemSizeGroup - Standard InitialSolutionParameters: BenchmarkCommonParameters: - KernelLanguage: ["Assembly"] - EdgeType: ["ShiftPtr"] ForkParameters: - TransposeLDS: [0, 1] - MatrixInstruction: - [32, 32, 1, 2] # - [32, 32, 2, 1] - PrefetchLocalRead: [0, 1, 3] - PrefetchGlobalRead: [True] - ThreadTile: - [ 2, 32 ] - WorkGroup: - [ 16, 8, 1 ] # - [ 32, 4, 1 ] # - [ 64, 2, 1 ] # - [ 128, 1, 1 ] - WorkGroupMapping: [8] - GlobalSplitU: [1] - InnerUnroll: [2] - DepthU: [8] - VectorWidth: [2, 4] - AssertSummationElementMultiple: [1, 2] - StoreRemapVectorWidth: [2] JoinParameters: BenchmarkJoinParameters: BenchmarkFinalParameters: - ProblemSizes: # - Range: [ [1024], [1024], [1], [1024] ] - Range: [ [1031], [1031], [8], [1031] ] # - # BenchmarkProblemSizeGroup - Standard # InitialSolutionParameters: # BenchmarkCommonParameters: # - KernelLanguage: ["Assembly"] # - EdgeType: ["ShiftPtr"] # ForkParameters: # - TransposeLDS: [0, 1] # - MatrixInstruction: # - [16, 16, 1, 4] # - [16, 16, 4, 1] # - PrefetchLocalRead: [0, 1, 3] # - PrefetchGlobalRead: [True] # - ThreadTile: # - [ 2, 32 ] # - WorkGroup: # - [ 16, 8, 1 ] # - [ 32, 4, 1 ] # - [ 64, 2, 1 ] # - [ 128, 1, 1 ] # - WorkGroupMapping: [8] # - GlobalSplitU: [1] # - InnerUnroll: [2] # - DepthU: [8] # - VectorWidth: [2, 4] # - AssertSummationElementMultiple: [1, 2] # - StoreRemapVectorWidth: [2] # JoinParameters: # BenchmarkJoinParameters: # BenchmarkFinalParameters: # - ProblemSizes: # - Range: [ [1024], [1024], [1], [1024] ] # - Range: [ [1031], [1031], [8], [1031] ] # - # BenchmarkProblemSizeGroup - Standard # InitialSolutionParameters: # BenchmarkCommonParameters: # - KernelLanguage: ["Assembly"] # - EdgeType: ["ShiftPtr"] # ForkParameters: # - TransposeLDS: [0, 1] # - MatrixInstruction: # - [ 4, 4, 1, 16] # - PrefetchLocalRead: [0, 1] # - PrefetchGlobalRead: [True] # - ThreadTile: # - [ 2, 8 ] # - WorkGroup: # - [ 8, 32, 1 ] # - [ 16, 16, 1 ] # - [ 32, 8, 1 ] # - [ 64, 4, 1 ] # - WorkGroupMapping: [8] # - GlobalSplitU: [1] # - InnerUnroll: [2] # - DepthU: [8] # - VectorWidth: [2, 4] # - AssertSummationElementMultiple: [1, 2] # - StoreRemapVectorWidth: [2] # JoinParameters: # BenchmarkJoinParameters: # BenchmarkFinalParameters: # - ProblemSizes: # - Range: [ [1024], [1024], [1], [1024] ] # - Range: [ [1031], [1031], [8], [1031] ] ######################################## # TT - standard ######################################## - - # ProblemType OperationType: GEMM DataType: c TransposeA: True TransposeB: True UseBeta: True Batched: True # - # BenchmarkProblemSizeGroup - Standard # InitialSolutionParameters: # BenchmarkCommonParameters: # - KernelLanguage: ["Assembly"] # - EdgeType: ["ShiftPtr"] # ForkParameters: # - TransposeLDS: [0, 1] # - MatrixInstruction: # - [32, 32, 1, 2] # - [32, 32, 2, 1] # - PrefetchLocalRead: [0, 1, 3] # - PrefetchGlobalRead: [True] # - ThreadTile: # - [ 2, 32 ] # - WorkGroup: # - [ 16, 8, 1 ] # - [ 32, 4, 1 ] # - [ 64, 2, 1 ] # - [ 128, 1, 1 ] # - WorkGroupMapping: [8] # - GlobalSplitU: [1] # - InnerUnroll: [2] # - DepthU: [8] # - VectorWidth: [2, 4] # - AssertSummationElementMultiple: [1, 2] # - StoreRemapVectorWidth: [2] # JoinParameters: # BenchmarkJoinParameters: # BenchmarkFinalParameters: # - ProblemSizes: # - Range: [ [1024], [1024], [1], [1024] ] # - Range: [ [1031], [1031], [8], [1031] ] - # BenchmarkProblemSizeGroup - Standard InitialSolutionParameters: BenchmarkCommonParameters: - KernelLanguage: ["Assembly"] - EdgeType: ["ShiftPtr"] ForkParameters: - TransposeLDS: [0, 1] - MatrixInstruction: - [16, 16, 1, 4] # - [16, 16, 4, 1] - PrefetchLocalRead: [0, 1, 3] - PrefetchGlobalRead: [True] - ThreadTile: - [ 2, 32 ] - WorkGroup: - [ 16, 8, 1 ] # - [ 32, 4, 1 ] # - [ 64, 2, 1 ] # - [ 128, 1, 1 ] - WorkGroupMapping: [8] - GlobalSplitU: [1] - InnerUnroll: [2] - DepthU: [8] - VectorWidth: [2, 4] - AssertSummationElementMultiple: [1, 2] - StoreRemapVectorWidth: [2] JoinParameters: BenchmarkJoinParameters: BenchmarkFinalParameters: - ProblemSizes: - Range: [ [1024], [1024], [1], [1024] ] # - Range: [ [1031], [1031], [8], [1031] ] # - # BenchmarkProblemSizeGroup - Standard # InitialSolutionParameters: # BenchmarkCommonParameters: # - KernelLanguage: ["Assembly"] # - EdgeType: ["ShiftPtr"] # ForkParameters: # - TransposeLDS: [0, 1] # - MatrixInstruction: # - [ 4, 4, 1, 16] # - PrefetchLocalRead: [0, 1] # - PrefetchGlobalRead: [True] # - ThreadTile: # - [ 2, 8 ] # - WorkGroup: # - [ 8, 32, 1 ] # - [ 16, 16, 1 ] # - [ 32, 8, 1 ] # - [ 64, 4, 1 ] # - WorkGroupMapping: [8] # - GlobalSplitU: [1] # - InnerUnroll: [2] # - DepthU: [8] # - VectorWidth: [2, 4] # - AssertSummationElementMultiple: [1, 2] # - StoreRemapVectorWidth: [2] # JoinParameters: # BenchmarkJoinParameters: # BenchmarkFinalParameters: # - ProblemSizes: # - Range: [ [1024], [1024], [1], [1024] ] # - Range: [ [1031], [1031], [8], [1031] ]