IVGCVSW-1922 Unit test for DepthwiseConvolution with NHWC

Change-Id: I3e6e5b9a62f30d03c05bd7178adea8f4c8275da8

4 files changed, 52 insertions, 4 deletions

diff --git a/src/backends/neon/test/NeonCreateWorkloadTests.cpp b/src/backends/neon/test/NeonCreateWorkloadTests.cpp
index 244002f132..ac0451f11b 100644
--- a/src/backends/neon/test/NeonCreateWorkloadTests.cpp
+++ b/src/backends/neon/test/NeonCreateWorkloadTests.cpp
@@ -219,6 +219,45 @@ BOOST_AUTO_TEST_CASE(CreateConvolution2dFloatNhwcWorkload)
     NeonCreateConvolution2dWorkloadTest<NeonConvolution2dFloatWorkload, DataType::Float32>(DataLayout::NHWC);
 }
 
+template <typename DepthwiseConvolution2dFloat32WorkloadType, typename armnn::DataType DataType>
+static void NeonCreateDepthWiseConvolutionWorkloadTest(DataLayout dataLayout)
+{
+    Graph graph;
+    NeonWorkloadFactory factory;
+
+    auto workload = CreateDepthwiseConvolution2dWorkloadTest<DepthwiseConvolution2dFloat32WorkloadType,
+                                                             DataType>(factory, graph, dataLayout);
+
+    // Checks that inputs/outputs are as we expect them (see definition of CreateNormalizationWorkloadTest).
+    DepthwiseConvolution2dQueueDescriptor queueDescriptor = workload->GetData();
+    auto inputHandle  = boost::polymorphic_downcast<INeonTensorHandle*>(queueDescriptor.m_Inputs[0]);
+    auto outputHandle = boost::polymorphic_downcast<INeonTensorHandle*>(queueDescriptor.m_Outputs[0]);
+
+    std::initializer_list<unsigned int> inputShape  = (dataLayout == DataLayout::NCHW)
+            ? std::initializer_list<unsigned int>({ 2, 2, 5, 5 })
+            : std::initializer_list<unsigned int>({ 2, 5, 5, 2 });
+    std::initializer_list<unsigned int> outputShape = (dataLayout == DataLayout::NCHW)
+            ? std::initializer_list<unsigned int>({ 2, 2, 5, 5 })
+            : std::initializer_list<unsigned int>({ 2, 5, 5, 2 });
+
+    BOOST_TEST(TestNeonTensorHandleInfo(inputHandle, TensorInfo(inputShape, DataType)));
+    BOOST_TEST(TestNeonTensorHandleInfo(outputHandle, TensorInfo(outputShape, DataType)));
+}
+
+BOOST_AUTO_TEST_CASE(CreateDepthWiseConvolution2dFloat32NhwcWorkload)
+{
+    NeonCreateDepthWiseConvolutionWorkloadTest<NeonDepthwiseConvolutionFloatWorkload,
+                                               DataType::Float32>(DataLayout::NHWC);
+}
+
+#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+BOOST_AUTO_TEST_CASE(CreateDepthWiseConvolution2dFloat16NhwcWorkload)
+{
+    NeonCreateDepthWiseConvolutionWorkloadTest<NeonDepthwiseConvolutionFloatWorkload,
+                                               DataType::Float16>(DataLayout::NHWC);
+}
+#endif
+
 template <typename FullyConnectedWorkloadType, typename armnn::DataType DataType>
 static void NeonCreateFullyConnectedWorkloadTest()
 {
diff --git a/src/backends/neon/test/NeonLayerTests.cpp b/src/backends/neon/test/NeonLayerTests.cpp
index 2d4ee996a4..36138b3c3f 100644
--- a/src/backends/neon/test/NeonLayerTests.cpp
+++ b/src/backends/neon/test/NeonLayerTests.cpp
@@ -84,6 +84,7 @@ BOOST_AUTO_TEST_CASE(Conv2dUtils)
 
 // Depthwise Convolution
 ARMNN_AUTO_TEST_CASE(DepthwiseConvolution2dDepthMul1, DepthwiseConvolution2dDepthMul1Test, true)
+ARMNN_AUTO_TEST_CASE(DepthwiseConvolution2dDepthNhwc, DepthwiseConvolution2dDepthNhwcTest, false)
 ARMNN_AUTO_TEST_CASE(UnbiasedDepthwiseConvolution2dDepthMul1, DepthwiseConvolution2dDepthMul1Test, false)
 ARMNN_AUTO_TEST_CASE(DepthwiseConvolution2dDepthMul1Uint8, DepthwiseConvolution2dDepthMul1Uint8Test, true)
 ARMNN_AUTO_TEST_CASE(UnbiasedDepthwiseConvolution2dDepthMul1Uint8, DepthwiseConvolution2dDepthMul1Uint8Test, false)
diff --git a/src/backends/neon/workloads/NeonDepthwiseConvolutionFloatWorkload.cpp b/src/backends/neon/workloads/NeonDepthwiseConvolutionFloatWorkload.cpp
index 742a768b94..4b266f3288 100644
--- a/src/backends/neon/workloads/NeonDepthwiseConvolutionFloatWorkload.cpp
+++ b/src/backends/neon/workloads/NeonDepthwiseConvolutionFloatWorkload.cpp
@@ -20,12 +20,12 @@ NeonDepthwiseConvolutionFloatWorkload::NeonDepthwiseConvolutionFloatWorkload(
     const TensorInfo& weightInfo = m_Data.m_Weight->GetTensorInfo();
 
     m_KernelTensor = std::make_unique<arm_compute::Tensor>();
-    BuildArmComputeTensor(*m_KernelTensor, weightInfo, descriptor.m_DataLayout);
+    BuildArmComputeTensor(*m_KernelTensor, weightInfo, m_Data.m_Parameters.m_DataLayout);
 
     if (m_Data.m_Parameters.m_BiasEnabled)
     {
         m_BiasTensor = std::make_unique<arm_compute::Tensor>();
-        BuildArmComputeTensor(*m_BiasTensor, m_Data.m_Bias->GetTensorInfo(), descriptor.m_DataLayout);
+        BuildArmComputeTensor(*m_BiasTensor, m_Data.m_Bias->GetTensorInfo(), m_Data.m_Parameters.m_DataLayout);
     }
 
     arm_compute::PadStrideInfo padStrideInfo(m_Data.m_Parameters.m_StrideX,
@@ -41,6 +41,10 @@ NeonDepthwiseConvolutionFloatWorkload::NeonDepthwiseConvolutionFloatWorkload(
     arm_compute::ITensor& input  = static_cast<INeonTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
     arm_compute::ITensor& output = static_cast<INeonTensorHandle*>(m_Data.m_Outputs[0])->GetTensor();
 
+    arm_compute::DataLayout aclDataLayout = ConvertDataLayout(m_Data.m_Parameters.m_DataLayout);
+    input.info()->set_data_layout(aclDataLayout);
+    output.info()->set_data_layout(aclDataLayout);
+
     bool use3x3Optimisation = weightInfo.GetShape()[3] == 3 && weightInfo.GetShape()[2] == 3;
     if (use3x3Optimisation)
     {
diff --git a/src/backends/neon/workloads/NeonDepthwiseConvolutionUint8Workload.cpp b/src/backends/neon/workloads/NeonDepthwiseConvolutionUint8Workload.cpp
index 722b778eba..6c6c2dfb6c 100644
--- a/src/backends/neon/workloads/NeonDepthwiseConvolutionUint8Workload.cpp
+++ b/src/backends/neon/workloads/NeonDepthwiseConvolutionUint8Workload.cpp
@@ -20,12 +20,12 @@ NeonDepthwiseConvolutionUint8Workload::NeonDepthwiseConvolutionUint8Workload(
     const TensorInfo& weightInfo = m_Data.m_Weight->GetTensorInfo();
 
     m_KernelTensor = std::make_unique<arm_compute::Tensor>();
-    BuildArmComputeTensor(*m_KernelTensor, weightInfo, descriptor.m_DataLayout);
+    BuildArmComputeTensor(*m_KernelTensor, weightInfo, m_Data.m_Parameters.m_DataLayout);
 
     if (m_Data.m_Parameters.m_BiasEnabled)
     {
         m_BiasTensor = std::make_unique<arm_compute::Tensor>();
-        BuildArmComputeTensor(*m_BiasTensor, m_Data.m_Bias->GetTensorInfo(), descriptor.m_DataLayout);
+        BuildArmComputeTensor(*m_BiasTensor, m_Data.m_Bias->GetTensorInfo(), m_Data.m_Parameters.m_DataLayout);
     }
 
     arm_compute::PadStrideInfo padStrideInfo(m_Data.m_Parameters.m_StrideX,
@@ -41,6 +41,10 @@ NeonDepthwiseConvolutionUint8Workload::NeonDepthwiseConvolutionUint8Workload(
     arm_compute::ITensor& input  = static_cast<INeonTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
     arm_compute::ITensor& output = static_cast<INeonTensorHandle*>(m_Data.m_Outputs[0])->GetTensor();
 
+    arm_compute::DataLayout aclDataLayout = ConvertDataLayout(m_Data.m_Parameters.m_DataLayout);
+    input.info()->set_data_layout(aclDataLayout);
+    output.info()->set_data_layout(aclDataLayout);
+
     bool use3x3Optimisation = weightInfo.GetShape()[3] == 3 && weightInfo.GetShape()[2] == 3;
     if (use3x3Optimisation)
     {