GlobalParameters: MinimumRequiredVersion: 4.1.0 NumElementsToValidate: 0 ValidationMaxToPrint: 16 PinClocks: False SleepPercent: 0 DataInitTypeA: 0 DataInitTypeB: 0 EnqueuesPerSync: 10 KernelTime: True PrintConvolutionUsage: 0 PrintSolutionRejectionReason: 0 BenchmarkProblems: - - # ProblemType OperationType: ConvolutionForward DataType: s ConvolutionConfig: - Dilation: 1x1 - Filter: NxN - Stride: 1x1 - TensorAFormat: NCHW - PadStart: NxN - PadEnd: NxN - # BenchmarkProblemSizeGroup - Standard InitialSolutionParameters: BenchmarkCommonParameters: - KernelLanguage: ["Assembly"] - EdgeType: ["ShiftPtr"] ForkParameters: - PrefetchGlobalRead: [1] - PrefetchLocalRead: [0,1] - PackSummationDims: [1] - ThreadTile: - [ 8, 4 ] - [ 4, 8 ] - [ 8, 8 ] - WorkGroup: - [ 8, 16, 1 ] - [ 32, 8, 1 ] - [ 16, 16, 1 ] - [ 32, 16, 1 ] #- [ 16, 32, 1 ] #- GlobalSplitU: [1,2,4,8] - GlobalSplitU: [1,2,8] - DepthU: [8,16] - GlobalReadVectorWidth: [1, -1] #- StaggerU: [0,32] #- StaggerUMapping: [0,1] #- WorkGroupMapping: [1,8,64] - VectorWidth: [1,-1] - FractionalLoad: [1] - PackBatchDims: [1] BenchmarkForkParameters: BenchmarkJoinParameters: BenchmarkFinalParameters: - ProblemSizes: - Conv: { count: 2 ,n: 32 ,c: 128 ,h: 17 ,w: 17 ,k: 128 ,y: 1 ,x: 7 ,p: 0 ,q: 3 ,u: 1 ,v: 1 ,l: 1 ,j: 1 } - Conv: { count: 1 ,n: 32 ,c: 128 ,h: 17 ,w: 17 ,k: 192 ,y: 1 ,x: 7 ,p: 0 ,q: 3 ,u: 1 ,v: 1 ,l: 1 ,j: 1 } - Conv: { count: 4 ,n: 32 ,c: 160 ,h: 17 ,w: 17 ,k: 160 ,y: 1 ,x: 7 ,p: 0 ,q: 3 ,u: 1 ,v: 1 ,l: 1 ,j: 1 } - Conv: { count: 2 ,n: 32 ,c: 160 ,h: 17 ,w: 17 ,k: 192 ,y: 1 ,x: 7 ,p: 0 ,q: 3 ,u: 1 ,v: 1 ,l: 1 ,j: 1 } - Conv: { count: 4 ,n: 32 ,c: 192 ,h: 17 ,w: 17 ,k: 192 ,y: 1 ,x: 7 ,p: 0 ,q: 3 ,u: 1 ,v: 1 ,l: 1 ,j: 1 } - Conv: { count: 0 ,n: 32 ,c: 2048 ,h: 17 ,w: 18 ,k: 192 ,y: 1 ,x: 7 ,p: 0 ,q: 3 ,u: 1 ,v: 1 ,l: 1 ,j: 1 }