MLCE-77 Depthwise Convolution with depth multiplier > 1 doesn't work

 * Unified ArmNN's weight format to [ M, I, H, W ] for the depthwise convolution
 * Added conversion utilities to permute/reshape the weights as appropriate
   when using CL and Neon backends
 * Updated the reference implementation of the convolution
 * Updated the relevant unit tests accordingly

!android-nn-driver:459

Change-Id: I07d0818efa9d1ca1e5dad82983aac1fe78eadb18

1 files changed, 30 insertions, 11 deletions

diff --git a/src/backends/backendsCommon/WorkloadUtils.hpp b/src/backends/backendsCommon/WorkloadUtils.hpp
index 2b07b2b0d2..a1a8d2a475 100644
--- a/src/backends/backendsCommon/WorkloadUtils.hpp
+++ b/src/backends/backendsCommon/WorkloadUtils.hpp
@@ -6,35 +6,42 @@
 #pragma once
 
 #include "ITensorHandle.hpp"
+#include "CpuTensorHandle.hpp"
 
 #include <armnn/Tensor.hpp>
 
+#include <Permute.hpp>
+#include <Profiling.hpp>
+#include <Half.hpp>
+
 #include <boost/cast.hpp>
 
 namespace armnn
 {
 namespace
 {
+
 template<typename ArrayType, typename Arg>
 void AssignValues(unsigned int num, unsigned int& idx, const ArrayType& array, Arg& arg)
 {
- if (idx >= num)
- {
-     return;
- }
+    if (idx >= num)
+    {
+        return;
+    }
 
- arg = array[(num - 1) - idx];
- idx++;
-};
+    arg = array[(num - 1) - idx];
+    idx++;
+}
 
 template<typename T, typename ArrayType, typename ...Args>
 void AssignValues(unsigned int num, unsigned int idx, const ArrayType& array, T& assignee, Args& ... args)
 {
- AssignValues(num, idx, array, assignee);
+    AssignValues(num, idx, array, assignee);
 
- AssignValues(num, idx, array, args...);
+    AssignValues(num, idx, array, args...);
 }
-} // namespace
+
+} // anonymous namespace
 
 template<typename CopyFunc>
 void CopyTensorContentsGeneric(const ITensorHandle* srcTensor, ITensorHandle* dstTensor, CopyFunc copy)
@@ -142,4 +149,16 @@ void GatherTensorHandlePairs(const DescriptorType& descriptor,
     }
 }
 
-} //namespace armnn
\ No newline at end of file
+armnn::ConstTensor PermuteTensor(const ConstCpuTensorHandle* tensor,
+                                 const PermutationVector& permutationVector,
+                                 void* permuteBuffer);
+
+void ReshapeWeightsForAcl(TensorInfo& weightInfo, DataLayout dataLayout);
+
+TensorInfo ConvertWeightTensorInfoFromArmnnToAcl(const TensorInfo& weightInfo, DataLayout dataLayout);
+
+armnn::ConstTensor ConvertWeightTensorFromArmnnToAcl(const ConstCpuTensorHandle* weightTensor,
+                                                     DataLayout dataLayout,
+                                                     void* permuteBuffer);
+
+} //namespace armnn