IVGCVSW-5826 Change weights layout for depthwise to [1,H,W,I*M]

 * This change is necessary because tflite uses a [1,H,W,I*M] format
   and uses the I*M dimension for per axis quantization. Our previous
   layout [M,I,H,W] can't handle the correlating quantization scales.
 * Updates Onnx-, TfLiteParser and TfliteDelegate
 * Updates the CpuRef, CpuAcc and GpuAcc backends
 * Adjusts unit tests
 * Adds test to ensure models with old layout can still be read and
   executed
 * Adds conversion function to previous layout [1,H,W,I*M] --> [M,I,H,W]
   which can be used by backend developers

!android-nn-driver:5553

Signed-off-by: Jan Eilers <jan.eilers@arm.com>
Change-Id: Ifef23368b8c3702cf315a5838d214f7dc13c0152

2 files changed, 5 insertions, 17 deletions

diff --git a/delegate/src/Convolution.hpp b/delegate/src/Convolution.hpp
index 6566ffff44..96612e0214 100644
--- a/delegate/src/Convolution.hpp
+++ b/delegate/src/Convolution.hpp
@@ -289,8 +289,6 @@ TfLiteStatus VisitDepthwiseConv2dOperator(DelegateData& delegateData,
     const armnn::TensorInfo& inputTensorInfo  = GetTensorInfoForTfLiteTensor(tfLiteInputTensor);
     const armnn::TensorInfo& outputTensorInfo = GetTensorInfoForTfLiteTensor(tfLiteOutputTensor);
 
-    // Mappings from TensorflowLite filter tensors to the ArmNN filter tensors (ArmNN weights have to be [M, I, H, W])
-    armnn::PermutationVector permutationVector{ 2, 3, 1, 0 }; // [H, W, I, M] -> [M, I, H, W]
     armnn::TensorInfo filterTensorInfo = GetTensorInfoForTfLiteTensor(tfLiteFilterTensor);
 
     // Assuming input is NHWC
@@ -301,12 +299,6 @@ TfLiteStatus VisitDepthwiseConv2dOperator(DelegateData& delegateData,
     unsigned int filterHeight = filterTensorInfo.GetShape()[1];
     unsigned int filterWidth  = filterTensorInfo.GetShape()[2];
 
-    // Reshape weights as [ H, W, I, M ]
-    filterTensorInfo.SetShape({ filterHeight,
-                                filterWidth,
-                                inputTensorInfo.GetShape()[3],
-                                filterTensorInfo.GetShape()[3] / inputTensorInfo.GetShape()[3] });
-
     // Calculate padding
     CalcPadding(inputHeight, filterHeight, descriptor.m_StrideY, descriptor.m_DilationY,
                 descriptor.m_PadTop, descriptor.m_PadBottom, params->padding);
@@ -340,12 +332,8 @@ TfLiteStatus VisitDepthwiseConv2dOperator(DelegateData& delegateData,
         biasTensorInfo = armnn::TensorInfo(armnn::TensorShape({1}), GetDataType(tfLiteInputTensor));
     }
 
-    std::vector<uint8_t> swizzledData(filterTensorInfo.GetNumBytes());
-    auto filter =
-        CreateConstTensor(&tfLiteFilterTensor,
-                          filterTensorInfo,
-                          armnn::Optional<armnn::PermutationVector&>(permutationVector),
-                          swizzledData.data());
+    // For depthwise the weights layout is the same as for tflite [1, H, W, I*M]. No permutation required.
+    auto filter = CreateConstTensor(&tfLiteFilterTensor, filterTensorInfo);
 
     if (!delegateData.m_Network)
     {
@@ -369,8 +357,7 @@ TfLiteStatus VisitDepthwiseConv2dOperator(DelegateData& delegateData,
     {
         auto biases =
             CreateConstTensor(&tfLiteContext->tensors[tfLiteNode->inputs->data[2]],
-                              biasTensorInfo,
-                              armnn::Optional<armnn::PermutationVector&>());
+                              biasTensorInfo);
         layer = delegateData.m_Network->AddDepthwiseConvolution2dLayer(descriptor,
                                                                        filter,
                                                                        armnn::Optional<armnn::ConstTensor>(biases));
diff --git a/delegate/src/DelegateUtils.hpp b/delegate/src/DelegateUtils.hpp
index 5dea567761..b04baac36e 100644
--- a/delegate/src/DelegateUtils.hpp
+++ b/delegate/src/DelegateUtils.hpp
@@ -472,7 +472,8 @@ armnn::TensorInfo GetTensorInfoForTfLiteTensor(const TfLiteTensor& tfLiteTensor)
 
 armnn::ConstTensor CreateConstTensor(const TfLiteTensor* tfLiteTensor,
                                      armnn::TensorInfo& tensorInfo,
-                                     armnn::Optional<armnn::PermutationVector&> permutationVector,
+                                     armnn::Optional<armnn::PermutationVector&>
+                                             permutationVector = armnn::EmptyOptional(),
                                      void* permutationData = nullptr)
 {
     if (tfLiteTensor->allocation_type != kTfLiteMmapRo)