Operator List

The current version of AOE supports only the tuning of AI Core operators whose compute logic is implemented using DSL APIs. This section lists the supported operators.

• Abs
• AccumulateNV2
• Add
• AddN
• Adds
• BatchNorm3D
• BiasAdd
• BiasAddGrad
• BNInfer
• BNTrainingReduce
• BNTrainingReduceGrad
• BNTrainingUpdate
• BNTrainingUpdateGrad
• BNTrainingUpdateV2
• ClipByNormNoDivSum
• ClipByValue
• DiagPartD
• Div
• DivNoNan
• DynamicRNN
• Elu
• EluGrad
• Exp
• ExpandD
• Floor
• FloorDiv
• Gelu
• InstanceNorm
• L2Loss
• LambUpdateWithLrV2
• LeakyRelu
• Log
• Log1p
• Maximum
• MaxPool
• MaxPool3D
• MaxPoolV3
• Minimum
• Mul
• Muls
• Neg
• Pow
• PRelu
• RealDiv
• ReduceAllD
• ReduceMaxD
• ReduceMeanD
• ReduceMinD
• ReduceSumD
• Relu
• ReluV2
• Round
• Rsqrt
• RsqrtGrad
• Sigmoid
• Sign
• SoftmaxCrossEntropyWithLogits
• Softplus
• SqrtGrad
• Square
• SquaredDifference
• SquareSumV1
• SquareSumV2
• StridedSliceD
• Sub
• Tanh
• TileD
• Unpack
• ReluGradV2

• AvgPool
• AvgPoolV2
• AvgPool3D
• BatchMatMulV3
• Deconvolution
• Conv1D
• Conv2D
• Conv2DCompress
• Conv2DBackpropInput
• Conv2DBackpropFilter
• DeformableConv2D
• GEMM
• MatMul
• MatMulV2
• MatMulV3
• MatMulV2Compress
• BatchMatMul
• BatchMatMulV2
• FullyConnection
• FullyConnectionCompress
• DepthwiseConv2D
• DepthwiseConv2DBackpropInput
• DepthwiseConv2DBackpropFilter
• Conv3D
• Conv3DBackpropInput
• Conv3DBackpropFilter
• AvgPool3DGrad
• Pooling
• Conv2DTranspose
• Conv3DTranspose
• GatherV2 (only supported by the Atlas inference products and Atlas training products)

The preceding list summarizes all supported operators, but might vary according to the Ascend AI Processor model. For details, see the Operator Library API Reference.