From 04ea4e88af1bdf9cf34436f7302c73e15f7bd4a5 Mon Sep 17 00:00:00 2001 From: giuros01 Date: Mon, 30 Sep 2019 16:36:17 +0100 Subject: COMPMID-2581: Fuse batch normalization with convolution and depthwise convolution at graph level for OpenCL - FP16 Change-Id: I757e27f45b5c25a5f98dacf2a6d519450d366760 Signed-off-by: giuros01 Reviewed-on: https://review.mlplatform.org/c/2007 Tested-by: Arm Jenkins Reviewed-by: Gian Marco Iodice --- src/graph/mutators/NodeFusionMutator.cpp | 3 +-- 1 file changed, 1 insertion(+), 2 deletions(-) diff --git a/src/graph/mutators/NodeFusionMutator.cpp b/src/graph/mutators/NodeFusionMutator.cpp index 4c3a905598..61d9479fca 100644 --- a/src/graph/mutators/NodeFusionMutator.cpp +++ b/src/graph/mutators/NodeFusionMutator.cpp @@ -318,8 +318,7 @@ void NodeFusionMutator::mutate(Graph &g) // 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 - // TODO (COMPMID-2581): Fuse batch normalization with convolution and depthwise convolution at graph level for OpenCL - FP16 - if(target == Target::CL && (g.nodes()[0].get()->output(0)->desc().data_type == DataType::F32)) + if(target == Target::CL) { //Depthwise Convolution and Batch Normalization Fusion active only for CL detail::fuse_layer(g, empty_prec, detail::fuse_convolution_with_batch_normalization); -- cgit v1.2.1