mirror/opencv
1
0
Fork 0
mirror of https://github.com/opencv/opencv.git synced 2025-06-12 20:42:53 +08:00
Code Issues Actions Packages Projects Releases Wiki Activity

Handle fusion of conv+eltwise in case of multi-output node (i.e. Split)

Browse Source
This commit is contained in:
Dmitry Kurtaev 2025-05-17 11:31:01 +03:00
parent 250b5003ee
commit fd5b33bb00
1 changed files with 12 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

14
modules/dnn/src/net_impl_fuse.cpp
View File

@ -370,6 +370,7 @@ void Net::Impl::fuseLayers(const std::vector<LayerPin>& blobsToKeep_)
Ptr<NaryEltwiseLayer> nextNaryEltwiseLayer = nextData->layerInstance.dynamicCast<NaryEltwiseLayer>();
if (nextEltwiseLayer.empty() && nextNaryEltwiseLayer.empty())
break;
LayerData *naryOrEltwiseData = nextData;
// TODO: fused the Conv+NaryEltwise on OpenCL backend. At present, we can only support it at CUDA backend.
if (IS_DNN_OPENCL_TARGET(preferableTarget) && nextNaryEltwiseLayer)
@ -605,8 +606,17 @@ void Net::Impl::fuseLayers(const std::vector<LayerPin>& blobsToKeep_)
else if (fuse_eltwise) // conv + eltwise/naryEltwise (note: conv could have fused activations before eltwise)
{
CV_Assert(IS_DNN_CUDA_TARGET(preferableTarget));
CV_Assert_N(biasLayerData->outputBlobsWrappers.size() == 1, ld.inputBlobsWrappers.size() == 1);
ld.inputBlobsWrappers.push_back(biasLayerData->outputBlobsWrappers[0]);
CV_Assert_N(biasLayerData->outputBlobsWrappers.size() >= 1, ld.inputBlobsWrappers.size() == 1);
// Iterate over eltwise inputs to find exact output id
for (const auto& pin : naryOrEltwiseData->inputBlobsId)
{
if (pin.lid == biasLayerData->id)
{
ld.inputBlobsWrappers.push_back(biasLayerData->outputBlobsWrappers[pin.oid]);
break;
}
}
CV_Assert(ld.inputBlobsWrappers.size() == 2); // Check input was found
if (nextEltwiseLayer)
printf_(("\tfused with %s\n", nextEltwiseLayer->name.c_str()));