commit 2ab3032ae39d129ef79115cad9084a7074606b59
author Teresa Charlin <teresa.charlinreyes@arm.com> Thu Feb 29 14:59:30 2024 +0000
committer TeresaARM <teresa.charlinreyes@arm.com> Thu Feb 29 15:01:18 2024 +0000
tree 89293008b7a4a7f7331075d918b627367f2207c6
parent 282881877522d3e94752dfc0839de9bfa0aa5a81

IVGCVSW-8212 Bug Fix: AddMulAdd optimization missing check on second ADD

* Failures was: MultiplicationQueueDescriptor: Tensors input_0 & input_1 must have the same number of dimensions in order to be broadcasted
* When trying to apply AddMulAdd, Mul dimensions were being collapsed before checking the second ADD.
* If the second ADD determined that fusion cannot happened MUL should have not been collapse.



Signed-off-by: Teresa Charlin <teresa.charlinreyes@arm.com>
Change-Id: I6dda131c424e1bf7b22fff40c5bef8977f073b24

src/backends/neon/NeonBackendOptimizationUtils.hpp

diff --git a/src/backends/neon/NeonBackendOptimizationUtils.hpp b/src/backends/neon/NeonBackendOptimizationUtils.hpp
index 3a8bf46..34ab41f 100644
--- a/src/backends/neon/NeonBackendOptimizationUtils.hpp
+++ b/src/backends/neon/NeonBackendOptimizationUtils.hpp
@@ -1,5 +1,5 @@
 //
-// Copyright © 2023 Arm Ltd and Contributors. All rights reserved.
+// Copyright © 2023-2024 Arm Ltd and Contributors. All rights reserved.
 // SPDX-License-Identifier: MIT
 //
 
@@ -100,6 +100,37 @@
     ARMNN_THROW_INVALIDARG_IF_FALSE(IsSequenceLayerType(*layerList[1], BinaryOperation::Mul));
     ARMNN_THROW_INVALIDARG_IF_FALSE(IsSequenceLayerType(*layerList[2], BinaryOperation::Add));
 
+    auto is1D = [](const TensorInfo expanded)
+    {
+        TensorInfo collapsed;
+        if (CollapseLeadingUnitDimensions(expanded, collapsed))
+        {
+            return (collapsed.GetNumDimensions() == 1);
+        }
+        else
+        {
+            return (expanded.GetNumDimensions() == 1);
+        }
+    };
+
+    // One of the 2 inputs for MUL and the Second ADD must be 1D
+    // ref: clframework/src/cpu/kernels/CpuAddMulAddKernel.cpp
+    auto& mulLayer  = *(PolymorphicDowncast<ElementwiseBinaryLayer*>(layerList[1]));
+    auto& add2Layer = *(PolymorphicDowncast<ElementwiseBinaryLayer*>(layerList[2]));
+
+    Layer& mulInput0  = mulLayer.GetInputSlot(0).GetConnectedOutputSlot()->GetOwningLayer();
+    Layer& mulInput1  = mulLayer.GetInputSlot(1).GetConnectedOutputSlot()->GetOwningLayer();
+    Layer& add2Input0 = add2Layer.GetInputSlot(0).GetConnectedOutputSlot()->GetOwningLayer();
+    Layer& add2Input1 = add2Layer.GetInputSlot(1).GetConnectedOutputSlot()->GetOwningLayer();
+    if (!is1D(mulInput0.GetOutputSlot(0).GetTensorInfo()) && !is1D(mulInput1.GetOutputSlot(0).GetTensorInfo()))
+    {
+        return false;
+    }
+    if (!is1D(add2Input0.GetOutputSlot(0).GetTensorInfo()) && !is1D(add2Input1.GetOutputSlot(0).GetTensorInfo()))
+    {
+        return false;
+    }
+
     fuseReLu = (layerList[3] != nullptr);
     if (fuseReLu)
     {