算子融合推荐

UB算子融合

输出潜在可合并算子组合列表，如图1。

====UBModel====

UB fusion operators need to be optimized

# 需要进行UB融合的算子

Identifications of UB fusion operators: the following operators can be used for UB fusion:

# UB融合算子识别：以下算子可以进行UB融合：

List of operators that can be fused in subgraph 0

# 可以在子图中融合的算子运算符列表

Fusion Type: DepthwiseConv2D+Mul; Fusion Operator Detail: block2a_dwconv/depthwiseblock2a_activation/Sigmoidblock2a_activation/mul, block2a_se_excite/mul; Fusion Operator Duration: 336.248993;

# Fusion Type：可融合算子类型；Fusion Operator Detail：可融合算子明细；Duration：可融合算子运行总时间。

Fusion Type: DepthwiseConv2D+Mul; Fusion Operator Detail: block2b_dwconv/depthwiseblock2b_activation/Sigmoidblock2b_activation/mul, block2b_se_excite/mul; Fusion Operator Duration: 284.063004;

Fusion Type: DepthwiseConv2D+Mul; Fusion Operator Detail: block3b_dwconv/depthwiseblock3b_activation/Sigmoidblock3b_activation/mul, block3b_se_excite/mul; Fusion Operator Duration: 220.571999;

优化建议：

建议根据输出结果将可融合算子进行融合。

首层算子融合

输出潜在可合并算子组合列表，如图2。

====AippFusionModel====

Fuse Cast/TransData with Conv needs to be optimized

# 需要进行Aipp首层算子融合的算子

Identifications of AIPP fusion operators: the following operators can be used for AIPP fusion:

# AIPP融合算子识别：以下算子可以进行AIPP融合：

List of operators that can be fused with aipp

# 可以在AIPP中融合的算子运算符列表

1. trans_Cast_0+trans_TransData_1+stem_conv/convolutionstem_activation/Sigmoidstem_activation/mul

优化建议：

建议根据输出结果将可融合算子进行融合。

L2融合（动态Batch切分）

====L2Model====

L2 fusion operators need to be optimized

# L2融合算子需要优化

Identifications of L2 fusion operators: the following operators can be used for L2 fusion：

# L2融合算子的识别：L2融合可以使用以下算子：

List of operators that can be fused in subgraph

# 可以在子图中融合的运算符列表

1. MobilenetV3/expanded_conv/project/Conv2DMobilenetV3/expanded_conv/add, MobilenetV3/expanded_conv_1/expand/Conv2DMobilenetV3/expanded_conv_1/expand/Relu, MobilenetV3/expanded_conv_1/depthwise/depthwiseMobilenetV3/expanded_conv_1/depthwise/Relu, MobilenetV3/expanded_conv_1/project/Conv2D

Op Info

# 算子信息

Op Name: MobilenetV3/expanded_conv/project/Conv2DMobilenetV3/expanded_conv/add; OP Type: Conv2D; Input Shapes: 8,1,112,112,16;1,1,16,16;16;8,1,112,112,16; mac_ratio: 0.063851; vec_ratio: 0.635701; mte2_ratio: 0.980165; mte3_ratio: 0.547420; Hit Rate: 0.989344;

Op Name: MobilenetV3/expanded_conv_1/expand/Conv2DMobilenetV3/expanded_conv_1/expand/Relu; OP Type: Conv2D; Input Shapes: 8,1,112,112,16;1,4,16,16;64; mac_ratio: 0.066542; vec_ratio: 0.608959; mte2_ratio: 0.880165; mte3_ratio: 0.973353; Hit Rate: 0.982700;

Op Name: MobilenetV3/expanded_conv_1/depthwise/depthwiseMobilenetV3/expanded_conv_1/depthwise/Relu; OP Type: DepthwiseConv2D; Input Shapes: 8,4,112,112,16;4,3,3,1,16,16;64; mac_ratio: 0.175088; vec_ratio: 0.119661; mte2_ratio: 0.879285; mte3_ratio: 0.185707; Hit Rate: 0.930730;

Op Name: MobilenetV3/expanded_conv_1/project/Conv2D; OP Type: Conv2D; Input Shapes: 8,4,56,56,16;4,2,16,16;24; mac_ratio: 0.276653; vec_ratio: 0.427420; mte2_ratio: 0.877128; mte3_ratio: 0.668421; Hit Rate: 0.989344;

Recommadation

# 优化建议

1. Open AutoTune

# 开启AutoTune功能。

AutoTune功能开启后能在计算时将模型中所有非1batch的算子自动切分为1batch，从而降低每层算子的数据量，解决L2 cache空间不足，产生DDR写回，引发算子性能恶化的问题。有关AutoTune功能的使用请参见《CANN 开发工具指南》中的“Auto Tune工具使用指南”章节。