From 5dea19e58a5521b05e95375c8618a37072697bc0 Mon Sep 17 00:00:00 2001
From: Gian Marco Iodice <gianmarco.iodice@arm.com>
Date: Fri, 8 Nov 2019 12:13:48 +0000
Subject: COMPMID-2579: Fuse batch normalization with convolution and depthwise
 convolution at graph level on NEON

Change-Id: Ib263a680bbd2dc1a4947102ee8d6da76b95f02bf
Signed-off-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Reviewed-on: https://review.mlplatform.org/c/2252
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Reviewed-by: Giorgio Arena <giorgio.arena@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
---
 src/graph/mutators/NodeFusionMutator.cpp | 13 ++-----------
 1 file changed, 2 insertions(+), 11 deletions(-)

(limited to 'src/graph/mutators/NodeFusionMutator.cpp')

diff --git a/src/graph/mutators/NodeFusionMutator.cpp b/src/graph/mutators/NodeFusionMutator.cpp
index 61d9479fca..abd6436d74 100644
--- a/src/graph/mutators/NodeFusionMutator.cpp
+++ b/src/graph/mutators/NodeFusionMutator.cpp
@@ -309,21 +309,12 @@ void NodeFusionMutator::mutate(Graph &g)
         return (output_qasymm8 && same_qinfo) || !output_qasymm8;
     };
 
-    Target target = g.nodes()[0].get()->output(0)->desc().target;
-
     // Fusion mutations
     detail::fuse_layer<BatchNormalizationLayerNode, ActivationLayerNode>(g, empty_prec, detail::fuse_node_with_activation<BatchNormalizationLayerNode>, supported_fused_activations);
     detail::fuse_layer<ConvolutionLayerNode, ActivationLayerNode>(g, empty_prec, detail::fuse_node_with_activation<ConvolutionLayerNode>, supported_fused_activations);
     detail::fuse_layer<DepthwiseConvolutionLayerNode, ActivationLayerNode>(g, qs8_prec, detail::fuse_node_with_activation<DepthwiseConvolutionLayerNode>, supported_fused_activations);
-
-    // Currently fuse batch normalization brings performance uplift only on OpenCL with FP32 data type
-    // TODO (COMPMID-2524): Fuse batch normalization with convolution and depthwise convolution at graph level for NEON - FP32
-    if(target == Target::CL)
-    {
-        //Depthwise Convolution and Batch Normalization Fusion active only for CL
-        detail::fuse_layer<ConvolutionLayerNode, BatchNormalizationLayerNode>(g, empty_prec, detail::fuse_convolution_with_batch_normalization);
-        detail::fuse_layer<DepthwiseConvolutionLayerNode, BatchNormalizationLayerNode>(g, empty_prec, detail::fuse_depthwise_convolution_with_batch_normalization);
-    }
+    detail::fuse_layer<ConvolutionLayerNode, BatchNormalizationLayerNode>(g, empty_prec, detail::fuse_convolution_with_batch_normalization);
+    detail::fuse_layer<DepthwiseConvolutionLayerNode, BatchNormalizationLayerNode>(g, empty_prec, detail::fuse_depthwise_convolution_with_batch_normalization);
 }
 } // namespace graph
 } // namespace arm_compute
-- 
cgit v1.2.1