commit 541880fcf4572887e57658a508623fb5f95ac554
author Cathal Corbett <cathal.corbett@arm.com> Mon May 16 15:20:56 2022 +0100
committer Cathal Corbett <cathal.corbett@arm.com> Wed May 18 18:16:42 2022 +0100
tree f78db6d55bc9feaf884d83212c9ec9195f482225
parent 07307f3c40c4efd8615755ed92ce300a3e150732

IVGCVSW-6147 ConstTensorsAsInput: Optimizer - FusePermuteIntoConstLayer

  * No trailing permute layer after a constant layer
  * Unit test for optimization

Signed-off-by: Cathal Corbett <cathal.corbett@arm.com>
Change-Id: I0d098f5af41d2c55df7cef1ccfb848093320ddc1

src/armnn/Network.cpp

diff --git a/src/armnn/Network.cpp b/src/armnn/Network.cpp
index 9da28ce..fecc766 100644
--- a/src/armnn/Network.cpp
+++ b/src/armnn/Network.cpp
@@ -1158,6 +1158,7 @@
         if(selectedBackend == armnn::Compute::GpuAcc || selectedBackend == armnn::Compute::CpuAcc)
         {
             Optimizer::Pass(optGraph, MakeOptimizations(optimizations::PermuteDepthwiseConv2dWeights()));
+            Optimizer::Pass(optGraph, MakeOptimizations(optimizations::FusePermuteIntoConstLayer()));
         }
 
         // Select sub-graphs based on backend
@@ -1719,6 +1720,10 @@
         optGraph.InferTensorInfos();
     }
 
+    // Need to FusePermuteIntoConstantLayer before FoldPadIntoDepthwiseConvolution2d or
+    // FuseBatchNormIntoDepthwiseConvolution2D optimizations are called.
+    Optimizer::Pass(optGraph, MakeOptimizations(FusePermuteIntoConstLayer()));
+
     // Perform optimisation passes
     Optimizer::Pass(optGraph, MakeOptimizations(SquashEqualPermuteSiblings(),
                                                 SquashEqualTransposeSiblings(),
@@ -1739,8 +1744,7 @@
                                                 FuseBatchNormIntoConvolution2DFloat16(),
                                                 FuseBatchNormIntoDepthwiseConvolution2DFloat32(),
                                                 FuseBatchNormIntoDepthwiseConvolution2DFloat16(),
-                                                ConvertConstDequantisationLayersToConstLayers(),
-                                                RedirectMembersToConstantInputs()));
+                                                ConvertConstDequantisationLayersToConstLayers()));
 
     // If Fp32 to Fp16 optimization is set convert Fp32 network to Fp16
     if (options.m_ReduceFp32ToFp16)