2 files changed, 25 insertions, 26 deletions
diff --git a/src/backends/neon/test/NeonLayerTests.cpp b/src/backends/neon/test/NeonLayerTests.cpp
index edc8cb995c..62864f82dc 100644
--- a/src/backends/neon/test/NeonLayerTests.cpp
+++ b/src/backends/neon/test/NeonLayerTests.cpp
@@ -216,6 +216,11 @@ ARMNN_AUTO_TEST_CASE(DepthToSpaceNhwcInt16_3, DepthToSpaceTest3<DataType::QSymmS
 ARMNN_AUTO_TEST_CASE(DepthToSpaceNhwcInt16_4, DepthToSpaceTest4<DataType::QSymmS16>, DataLayout::NHWC);
 
 // Depthwise Convolution
+ARMNN_AUTO_TEST_CASE_WITH_THF(DepthwiseConvolution2d, DepthwiseConvolution2dTest, true, DataLayout::NCHW)
+ARMNN_AUTO_TEST_CASE_WITH_THF(DepthwiseConvolution2dUint8, DepthwiseConvolution2dUint8Test, true, DataLayout::NCHW)
+
+ARMNN_AUTO_TEST_CASE_WITH_THF(UnbiasedDepthwiseConvolution2d, DepthwiseConvolution2dTest, false, DataLayout::NCHW)
+
 ARMNN_AUTO_TEST_CASE_WITH_THF(DepthwiseConvolution2dDepthMul1,
                      DepthwiseConvolution2dDepthMul1Test, true, DataLayout::NCHW)
 ARMNN_AUTO_TEST_CASE_WITH_THF(UnbiasedDepthwiseConvolution2dDepthMul1,
@@ -291,16 +296,15 @@ TensorInfo CreateOutputTensorInfo(const TensorInfo& inputInfo,
     unsigned int inHeight = inputShape[2];
     unsigned int inBatchSize = inputShape[0];
 
-    unsigned int filterWidth = filterShape[3];
+    unsigned int filterWidth = filterShape[2];
     unsigned int readWidth = (inWidth + descriptor.m_PadLeft + descriptor.m_PadRight) - (filterWidth);
     unsigned int outWidth =  1u + (readWidth / descriptor.m_StrideX);
 
-    unsigned int filterHeight = filterShape[2];
+    unsigned int filterHeight = filterShape[1];
     unsigned int readHeight = (inHeight + descriptor.m_PadTop + descriptor.m_PadBottom) - (filterHeight);
     unsigned int outHeight = 1u + (readHeight / descriptor.m_StrideY);
-    unsigned int depthMultiplier = filterShape[0];
 
-    unsigned int outChannels = filterShape[1] * depthMultiplier;
+    unsigned int outChannels = filterShape[3];
     unsigned int outBatchSize = inBatchSize;
 
     TensorShape outputShape({outBatchSize, outChannels, outHeight, outWidth});
@@ -314,7 +318,7 @@ TEST_CASE("DepthwiseConv2dUtils")
 
     TensorInfo inputInfo({1, 1, 10, 10 }, dataType);
     TensorInfo outputInfo;
-    TensorInfo weightsInfo3x3({ 1, 1, 3, 3 }, dataType);
+    TensorInfo weightsInfo3x3({ 1, 3, 3, 1 }, dataType); // [1,H,W,I*M]
     TensorInfo biasesInfo;
 
     DepthwiseConvolution2dDescriptor descriptor;
@@ -380,7 +384,7 @@ TEST_CASE("DepthwiseConv2dUtils")
                                                             weightsInfo1x1, biasesInfo));
 
     // Supported shape 2x2
-    TensorInfo weightsInfo2x2({ 1, 1, 2, 2 }, DataType::Float32);
+    TensorInfo weightsInfo2x2({ 1, 2, 2, 1 }, DataType::Float32);
     descriptor = MakeDepthwiseConv2dDesc(1, 1);
     outputInfo = CreateOutputTensorInfo(inputInfo, weightsInfo2x2, descriptor, dataType);
     CHECK(layerSupport.IsDepthwiseConvolutionSupported(inputInfo, outputInfo, descriptor,
diff --git a/src/backends/neon/workloads/NeonDepthwiseConvolutionWorkload.cpp b/src/backends/neon/workloads/NeonDepthwiseConvolutionWorkload.cpp
index ad509076b4..589a951825 100644
--- a/src/backends/neon/workloads/NeonDepthwiseConvolutionWorkload.cpp
+++ b/src/backends/neon/workloads/NeonDepthwiseConvolutionWorkload.cpp
@@ -36,12 +36,11 @@ arm_compute::Status NeonDepthwiseConvolutionWorkloadValidate(const TensorInfo& i
     const arm_compute::TensorInfo aclInputInfo  = BuildArmComputeTensorInfo(input,  descriptor.m_DataLayout);
     const arm_compute::TensorInfo aclOutputInfo = BuildArmComputeTensorInfo(output, descriptor.m_DataLayout);
 
-    // ArmNN's weight format is [ M, I, H, W ]
-    const unsigned int aclDepthMultiplier = weights.GetShape()[0];
-
-    // Convert the weight format from ArmNN's [ M, I, H, W ] (does NOT depend on the data layout) to either
-    // [ 1, H, W, I * M ] (if NHWC) or [ 1, I * M, H, W ] (if NCHW), as required by the compute library
-    TensorInfo weightsPermuted = ConvertWeightTensorInfoFromArmnnToAcl(weights, descriptor.m_DataLayout);
+    // ArmNN's weight format is usually [ M, I, H, W ] but for depthwise its [ 1, H, W, I*M]
+    // Permute to [ 1, I * M, H, W ] (if NCHW), as required by the compute library
+    unsigned int aclDepthMultiplier;
+    TensorInfo weightsPermuted;
+    std::tie(weightsPermuted, aclDepthMultiplier) = Convert1HWOTensorInfoToAcl(weights, input,descriptor.m_DataLayout);
 
     // Convert the weights into the compute library format
     const arm_compute::TensorInfo aclWeightsInfo = BuildArmComputeTensorInfo(weightsPermuted, descriptor.m_DataLayout);
@@ -79,21 +78,20 @@ NeonDepthwiseConvolutionWorkload::NeonDepthwiseConvolutionWorkload(
     const WorkloadInfo& info)
     : BaseWorkload<DepthwiseConvolution2dQueueDescriptor>(descriptor, info)
 {
-    // ArmNN's weight format is [ M, I, H, W ]
+    // ArmNN's weight format for depthwise is [ 1, H, W, I*M ]
     auto& weightInfo = m_Data.m_Weight->GetTensorInfo();
 
-    // Allocate a buffer for the swizzling of the weight tensor
-    std::unique_ptr<unsigned char[]> permuteBuffer(new unsigned char[m_Data.m_Weight->GetTensorInfo().GetNumBytes()]);
-
-    // Convert the weight format from ArmNN's [ M, I, H, W ] (does NOT depend on the data layout) to either
-    // [ 1, H, W, I * M ] (if NHWC) or [ 1, I * M, H, W ] (if NCHW), as required by the compute library
-    ConstTensor weightPermuted = ConvertWeightTensorFromArmnnToAcl(m_Data.m_Weight,
-                                                                   m_Data.m_Parameters.m_DataLayout,
-                                                                   permuteBuffer.get());
+    ConstTensor weightsPermuted;
+    unsigned int depthMultiplier;
+    std::unique_ptr<unsigned char[]> permuteBuffer(new unsigned char[weightInfo.GetNumBytes()]);
+    std::tie(weightsPermuted, depthMultiplier) = Convert1HWOTensorToAcl(m_Data.m_Weight,
+                                                                              info.m_InputTensorInfos[0],
+                                                                              m_Data.m_Parameters.m_DataLayout,
+                                                                              permuteBuffer.get());
 
     // Convert the weights into the compute library format
     m_KernelTensor = std::make_unique<arm_compute::Tensor>();
-    BuildArmComputeTensor(*m_KernelTensor, weightPermuted.GetInfo(), m_Data.m_Parameters.m_DataLayout);
+    BuildArmComputeTensor(*m_KernelTensor, weightsPermuted.GetInfo(), m_Data.m_Parameters.m_DataLayout);
 
     if (m_Data.m_Parameters.m_BiasEnabled)
     {
@@ -116,9 +114,6 @@ NeonDepthwiseConvolutionWorkload::NeonDepthwiseConvolutionWorkload(
     input.info()->set_data_layout(aclDataLayout);
     output.info()->set_data_layout(aclDataLayout);
 
-    // Get the depth multiplier
-    const unsigned int depthMultiplier = weightInfo.GetShape()[0];
-
     arm_compute::PadStrideInfo padStrideInfo = BuildArmComputePadStrideInfo(m_Data.m_Parameters);
 
     const arm_compute::ActivationLayerInfo activationInfo = ConvertAdditionalInfoToAclActivationLayerInfo(descriptor);
@@ -136,7 +131,7 @@ NeonDepthwiseConvolutionWorkload::NeonDepthwiseConvolutionWorkload(
 
     ARMNN_ASSERT(m_pDepthwiseConvolutionLayer);
 
-    ScopedTensorHandle weightsPermutedHandle(weightPermuted);
+    ScopedTensorHandle weightsPermutedHandle(weightsPermuted);
     InitializeArmComputeTensorData(*m_KernelTensor, &weightsPermutedHandle);
 
     if (m_Data.m_Parameters.m_BiasEnabled)