| 377 | } |
| 378 | |
| 379 | void NodeFusionMutator::mutate(Graph &g) |
| 380 | { |
| 381 | // Supported activations when fusing |
| 382 | const std::set<Activation> supported_fused_activations = { |
| 383 | Activation::ABS, Activation::BOUNDED_RELU, Activation::ELU, |
| 384 | Activation::HARD_SWISH, Activation::IDENTITY, Activation::LEAKY_RELU, |
| 385 | Activation::LINEAR, Activation::LOGISTIC, Activation::LU_BOUNDED_RELU, |
| 386 | Activation::RELU, Activation::SOFT_RELU, Activation::SQRT, |
| 387 | Activation::SQUARE, Activation::TANH}; |
| 388 | |
| 389 | // Preconditions |
| 390 | auto empty_prec = [](INode &) { return true; }; |
| 391 | auto cl_target_prec = [](INode &n) { return n.assigned_target() == Target::CL; }; |
| 392 | auto qs8_prec = [&g](INode &n) |
| 393 | { |
| 394 | ARM_COMPUTE_ERROR_ON(n.output(0) == nullptr); |
| 395 | |
| 396 | const auto output_edge_id = *n.output_edges().begin(); |
| 397 | const auto output_edge = g.edge(output_edge_id); |
| 398 | // To perform fusion the two nodes must have same output quantization information |
| 399 | const bool same_qinfo = n.output(0)->desc().quant_info == output_edge->producer()->output(0)->desc().quant_info; |
| 400 | const bool output_qasymm8 = n.output(0)->desc().data_type == DataType::QASYMM8; |
| 401 | |
| 402 | return (output_qasymm8 && same_qinfo) || !output_qasymm8; |
| 403 | }; |
| 404 | |
| 405 | // Fusion mutations |
| 406 | |
| 407 | detail::fuse_layer<PadLayerNode, ConvolutionLayerNode>(g, empty_prec, |
| 408 | detail::fuse_pad_with_convolution<ConvolutionLayerNode>); |
| 409 | detail::fuse_layer<PadLayerNode, DepthwiseConvolutionLayerNode>( |
| 410 | g, empty_prec, detail::fuse_pad_with_convolution<DepthwiseConvolutionLayerNode>); |
| 411 | detail::fuse_layer<BatchNormalizationLayerNode, ActivationLayerNode>( |
| 412 | g, empty_prec, detail::fuse_node_with_activation<BatchNormalizationLayerNode>, supported_fused_activations); |
| 413 | detail::fuse_layer<ConvolutionLayerNode, ActivationLayerNode>( |
| 414 | g, empty_prec, detail::fuse_node_with_activation<ConvolutionLayerNode>, supported_fused_activations); |
| 415 | detail::fuse_layer<DepthwiseConvolutionLayerNode, ActivationLayerNode>( |
| 416 | g, qs8_prec, detail::fuse_node_with_activation<DepthwiseConvolutionLayerNode>, supported_fused_activations); |
| 417 | detail::fuse_layer<FullyConnectedLayerNode, ActivationLayerNode>( |
| 418 | g, empty_prec, detail::fuse_node_with_activation<FullyConnectedLayerNode>, supported_fused_activations); |
| 419 | detail::fuse_layer<EltwiseLayerNode, ActivationLayerNode>( |
| 420 | g, cl_target_prec, detail::fuse_node_with_activation<EltwiseLayerNode>, supported_fused_activations); |
| 421 | // The fusion of BatchNormalizationLayer must occur after the fusion of ActivationLayer. Because FusedConvolutionBatchNormalizationNode assumes the BatchNormalization is already fused with activation, if any |
| 422 | detail::fuse_layer<ConvolutionLayerNode, BatchNormalizationLayerNode>( |
| 423 | g, empty_prec, detail::fuse_convolution_with_batch_normalization); |
| 424 | detail::fuse_layer<DepthwiseConvolutionLayerNode, BatchNormalizationLayerNode>( |
| 425 | g, empty_prec, detail::fuse_depthwise_convolution_with_batch_normalization); |
| 426 | } |
| 427 | } // namespace graph |
| 428 | } // namespace arm_compute |