From 0843a0bb177109b70ffd4ee2cacc14b8d014e706 Mon Sep 17 00:00:00 2001
From: Abhishek Gola <abhishek.gola@bigvisionllc.com>
Date: Mon, 6 Jul 2026 13:36:53 +0530
Subject: [PATCH] fusion fix

---
 modules/dnn/src/net_impl.cpp  |  8 +++++++
 modules/dnn/src/net_impl.hpp  | 15 ++++++++++++
 modules/dnn/src/net_impl2.cpp | 44 +++++++++++++++++++++++++++++++++++
 3 files changed, 67 insertions(+)

diff --git a/modules/dnn/src/net_impl.cpp b/modules/dnn/src/net_impl.cpp
index b2c07c6584..796d9a1a2d 100644
--- a/modules/dnn/src/net_impl.cpp
+++ b/modules/dnn/src/net_impl.cpp
@@ -154,6 +154,9 @@ void Net::Impl::clear()
 
     prepared = false;
     finalizeLayers = true;
+    finalized = false;
+    fusedSnapshotValid = false;
+    fusedSnapshot.clear();
 }
 
 
@@ -1419,6 +1422,9 @@ void Net::Impl::getLayerShapes(const ShapesVec& netInputShapes,
         LayerShapes& shapes)
 {
     if (mainGraph) {
+        // Fusion emits block-layout layers; their TransformLayout conversions are
+        // only inserted by finalize(), so shape inference must run post-finalize.
+        finalize();
         std::vector<MatShape> shapeCache;
         std::vector<int> typeCache;
         CV_Assert(layerId == 0);
@@ -2551,6 +2557,7 @@ int64 Net::Impl::getFLOPS(const std::vector<MatShape>& netInputShapes,
                           const std::vector<MatType>& netInputTypes) /*const*/
 {
     if (mainGraph) {
+        finalize();
         // The new graph engine executes in FP32 on CPU regardless of the requested
         // target, so FP16 input types (e.g. coming from an OpenCL FP16 target) would be
         // rejected by Layer::getTypes(). Normalize them to FP32 for shape/FLOPS inference.
@@ -2584,6 +2591,7 @@ int64 Net::Impl::getFLOPS(
         const std::vector<MatType>& netInputTypes) /*const*/
 {
     if (mainGraph) {
+        finalize();
         std::vector<MatType> inputTypes = filterFP16InputTypes(netInputTypes);
         LayerShapes shapes;
         std::vector<MatShape> shapeCache;
diff --git a/modules/dnn/src/net_impl.hpp b/modules/dnn/src/net_impl.hpp
index 2deee06458..6afdaadb71 100644
--- a/modules/dnn/src/net_impl.hpp
+++ b/modules/dnn/src/net_impl.hpp
@@ -144,6 +144,18 @@ struct Net::Impl : public detail::NetImplBase
     bool prepared; // need to rerun graph transformations/optimizations
     bool finalizeLayers; // need to initialize each layer
     bool finalized = false; // executors have been selected for the current backend/target
+
+    // Post-fusion (pre block-layout) snapshot so finalize() can re-run from a clean
+    // state on a backend/target change; useBlockLayout() is destructive and must
+    // run after backend assignment (see deviceOp handling in graph_block_layout.cpp).
+    struct FusedGraphSnapshot {
+        Ptr<Graph> graph;
+        std::vector<Ptr<LayerInfo> > prog;
+        std::vector<std::vector<Arg> > inputs;
+        std::vector<std::vector<Arg> > outputs;
+    };
+    bool fusedSnapshotValid = false;
+    std::vector<FusedGraphSnapshot> fusedSnapshot;
     std::vector<Ptr<BackendWrapper> > argWrappers;
     std::vector<const void*> argWrapperData;
     TracingMode tracingMode;
@@ -426,6 +438,9 @@ struct Net::Impl : public detail::NetImplBase
     void finalize();
     // Selects executors for a single graph (recursing into subgraphs).
     void finalizeGraph(const Ptr<Graph>& graph, bool useCUDA);
+    // Save/restore the fused graph so finalize() is re-entrant across backend changes.
+    void saveFusedSnapshot();
+    void restoreFusedSnapshot();
 #ifdef HAVE_CUDA
     Ptr<BackendWrapper> getCudaArgWrapper(Arg arg, Mat& hostMat);
 #endif
diff --git a/modules/dnn/src/net_impl2.cpp b/modules/dnn/src/net_impl2.cpp
index b0b66017b9..b9c526a915 100644
--- a/modules/dnn/src/net_impl2.cpp
+++ b/modules/dnn/src/net_impl2.cpp
@@ -624,6 +624,41 @@ void Net::Impl::finalizeGraph(const Ptr<Graph>& graph, bool useCUDA)
     }
 }
 
+void Net::Impl::saveFusedSnapshot()
+{
+    fusedSnapshot.clear();
+    for (const Ptr<Graph>& g : allgraphs) {
+        FusedGraphSnapshot snap;
+        snap.graph = g;
+        const std::vector<Ptr<LayerInfo> >& prog = g->prog();
+        snap.prog = prog;
+        snap.inputs.reserve(prog.size());
+        snap.outputs.reserve(prog.size());
+        for (const Ptr<LayerInfo>& op : prog) {
+            snap.inputs.push_back(op ? op->inputs : std::vector<Arg>());
+            snap.outputs.push_back(op ? op->outputs : std::vector<Arg>());
+        }
+        fusedSnapshot.push_back(std::move(snap));
+    }
+}
+
+void Net::Impl::restoreFusedSnapshot()
+{
+    // Roll the graph back to its post-fusion state: undo the layer-input rewiring
+    // and remove the TransformLayout ops inserted by a previous useBlockLayout().
+    for (const FusedGraphSnapshot& snap : fusedSnapshot) {
+        for (size_t i = 0; i < snap.prog.size(); i++) {
+            const Ptr<LayerInfo>& op = snap.prog[i];
+            if (!op)
+                continue;
+            op->inputs = snap.inputs[i];
+            op->outputs = snap.outputs[i];
+        }
+        snap.graph->setProg(snap.prog);
+    }
+    totalLayers = updateGraphOfs(mainGraph, 0, true);
+}
+
 void Net::Impl::finalize()
 {
 #ifdef HAVE_ONNXRUNTIME
@@ -637,6 +672,15 @@ void Net::Impl::finalize()
     if (finalized)
         return;
 
+    // Snapshot the fused graph once so finalize() can re-run cleanly on a
+    // backend/target change (block layout + buffer assignment are destructive).
+    if (!fusedSnapshotValid) {
+        saveFusedSnapshot();
+        fusedSnapshotValid = true;
+    } else {
+        restoreFusedSnapshot();
+    }
+
     bool useCUDA = false;
 #ifdef HAVE_CUDA
     argWrappers.clear();