IVGCVSW-2074: Updated DepthwiseConvolution2dAsymmetricTest for NHWC

Change-Id: I1bc8d2f270d4abfaa62e4843b8d96bcf462c6b09

6 files changed, 71 insertions, 21 deletions

diff --git a/src/backends/cl/test/ClLayerTests.cpp b/src/backends/cl/test/ClLayerTests.cpp
index 057b5d84b2..beb523f84e 100755
--- a/src/backends/cl/test/ClLayerTests.cpp
+++ b/src/backends/cl/test/ClLayerTests.cpp
@@ -75,8 +75,14 @@ ARMNN_AUTO_TEST_CASE(UnbiasedDepthwiseConvolution2dDepthMul1, DepthwiseConvoluti
 ARMNN_AUTO_TEST_CASE(DepthwiseConvolution2dDepthMul1Uint8, DepthwiseConvolution2dDepthMul1Uint8Test, true)
 ARMNN_AUTO_TEST_CASE(UnbiasedDepthwiseConvolution2dDepthMul1Uint8, DepthwiseConvolution2dDepthMul1Uint8Test, false)
 
-ARMNN_AUTO_TEST_CASE(DepthwiseConvolution2dAsymmetric, DepthwiseConvolution2dAsymmetricTest, true)
-ARMNN_AUTO_TEST_CASE(UnbiasedDepthwiseConvolution2dAsymmetric, DepthwiseConvolution2dAsymmetricTest, false)
+ARMNN_AUTO_TEST_CASE(DepthwiseConvolution2dAsymmetric,
+                     DepthwiseConvolution2dAsymmetricTest, true, armnn::DataLayout::NCHW)
+ARMNN_AUTO_TEST_CASE(UnbiasedDepthwiseConvolution2dAsymmetric,
+                     DepthwiseConvolution2dAsymmetricTest, false, armnn::DataLayout::NCHW)
+ARMNN_AUTO_TEST_CASE(DepthwiseConvolution2dAsymmetricNhwc,
+                     DepthwiseConvolution2dAsymmetricTest, true, armnn::DataLayout::NHWC)
+ARMNN_AUTO_TEST_CASE(UnbiasedDepthwiseConvolution2dAsymmetricNhwc,
+                     DepthwiseConvolution2dAsymmetricTest, false, armnn::DataLayout::NHWC)
 
 // Softmax
 BOOST_AUTO_TEST_CASE(Softmax4dSupport)
diff --git a/src/backends/neon/test/NeonLayerTests.cpp b/src/backends/neon/test/NeonLayerTests.cpp
index 349dbf8fb3..c7b0050311 100644
--- a/src/backends/neon/test/NeonLayerTests.cpp
+++ b/src/backends/neon/test/NeonLayerTests.cpp
@@ -98,8 +98,14 @@ ARMNN_AUTO_TEST_CASE(UnbiasedDepthwiseConvolution2dDepthMul1, DepthwiseConvoluti
 ARMNN_AUTO_TEST_CASE(DepthwiseConvolution2dDepthMul1Uint8, DepthwiseConvolution2dDepthMul1Uint8Test, true)
 ARMNN_AUTO_TEST_CASE(UnbiasedDepthwiseConvolution2dDepthMul1Uint8, DepthwiseConvolution2dDepthMul1Uint8Test, false)
 
-ARMNN_AUTO_TEST_CASE(DepthwiseConvolution2dAsymmetric, DepthwiseConvolution2dAsymmetricTest, true)
-ARMNN_AUTO_TEST_CASE(UnbiasedDepthwiseConvolution2dAsymmetric, DepthwiseConvolution2dAsymmetricTest, false)
+ARMNN_AUTO_TEST_CASE(DepthwiseConvolution2dAsymmetric,
+                     DepthwiseConvolution2dAsymmetricTest, true, armnn::DataLayout::NCHW)
+ARMNN_AUTO_TEST_CASE(UnbiasedDepthwiseConvolution2dAsymmetric,
+                     DepthwiseConvolution2dAsymmetricTest, false, armnn::DataLayout::NCHW)
+ARMNN_AUTO_TEST_CASE(DepthwiseConvolution2dAsymmetricNhwc,
+                     DepthwiseConvolution2dAsymmetricTest, true, armnn::DataLayout::NHWC)
+ARMNN_AUTO_TEST_CASE(UnbiasedDepthwiseConvolution2dAsymmetricNhwc,
+                     DepthwiseConvolution2dAsymmetricTest, false, armnn::DataLayout::NHWC)
 
 namespace
 {
diff --git a/src/backends/reference/test/RefLayerTests.cpp b/src/backends/reference/test/RefLayerTests.cpp
index cb5a1c4158..6a92b67e57 100644
--- a/src/backends/reference/test/RefLayerTests.cpp
+++ b/src/backends/reference/test/RefLayerTests.cpp
@@ -67,8 +67,14 @@ ARMNN_AUTO_TEST_CASE(DepthwiseConvolution2dDepthMul1Uint8, DepthwiseConvolution2
 ARMNN_AUTO_TEST_CASE(UnbiasedDepthwiseConvolution2dDepthMul1, DepthwiseConvolution2dDepthMul1Test, false)
 ARMNN_AUTO_TEST_CASE(UnbiasedDepthwiseConvolution2dDepthMul1Uint8, DepthwiseConvolution2dDepthMul1Uint8Test, false)
 
-ARMNN_AUTO_TEST_CASE(DepthwiseConvolution2dAsymmetric, DepthwiseConvolution2dAsymmetricTest, true)
-ARMNN_AUTO_TEST_CASE(UnbiasedDepthwiseConvolution2dAsymmetric, DepthwiseConvolution2dAsymmetricTest, false)
+ARMNN_AUTO_TEST_CASE(DepthwiseConvolution2dAsymmetric,
+                     DepthwiseConvolution2dAsymmetricTest, true, armnn::DataLayout::NCHW)
+ARMNN_AUTO_TEST_CASE(UnbiasedDepthwiseConvolution2dAsymmetric,
+                     DepthwiseConvolution2dAsymmetricTest, false, armnn::DataLayout::NCHW)
+ARMNN_AUTO_TEST_CASE(DepthwiseConvolution2dAsymmetricNhwc,
+                     DepthwiseConvolution2dAsymmetricTest, true, armnn::DataLayout::NHWC)
+ARMNN_AUTO_TEST_CASE(UnbiasedDepthwiseConvolution2dAsymmetricNhwc,
+                     DepthwiseConvolution2dAsymmetricTest, false, armnn::DataLayout::NHWC)
 
 // Pooling
 ARMNN_AUTO_TEST_CASE(SimpleMaxPooling2dSize2x2Stride2x2, SimpleMaxPooling2dSize2x2Stride2x2Test, false)
diff --git a/src/backends/test/Conv2dTestImpl.hpp b/src/backends/test/Conv2dTestImpl.hpp
index 41a0d1b095..993794e325 100755
--- a/src/backends/test/Conv2dTestImpl.hpp
+++ b/src/backends/test/Conv2dTestImpl.hpp
@@ -337,11 +337,12 @@ LayerTestResult<T, 4> SimpleConvolution2dNhwcTestImpl(armnn::IWorkloadFactory& w
 template<typename T, typename B>
 LayerTestResult<T, 4> DepthwiseConvolution2dAsymmetricTestImpl(armnn::IWorkloadFactory& workloadFactory,
                                                                const boost::multi_array<T, 4>& input,
-                                                               const boost::multi_array<T, 4>& kernel,
+                                                               const boost::multi_array<T, 4>& originalKernel,
                                                                const boost::multi_array<B, 1>& bias,
                                                                const boost::multi_array<T, 4>& outputExpected,
                                                                float qScale,
                                                                int32_t qOffset,
+                                                               const armnn::DataLayoutIndexed& layout,
                                                                uint32_t padLeft = 0,
                                                                uint32_t padTop = 0,
                                                                uint32_t padRight = 0,
@@ -353,10 +354,10 @@ LayerTestResult<T, 4> DepthwiseConvolution2dAsymmetricTestImpl(armnn::IWorkloadF
     unsigned int inputChannels  = boost::numeric_cast<unsigned int>(input.shape()[1]);
     unsigned int inputHeight    = boost::numeric_cast<unsigned int>(input.shape()[2]);
     unsigned int inputWidth     = boost::numeric_cast<unsigned int>(input.shape()[3]);
-    unsigned int kernelChanMul  = boost::numeric_cast<unsigned int>(kernel.shape()[0]);
-    unsigned int kernelChannels = boost::numeric_cast<unsigned int>(kernel.shape()[1]);
-    unsigned int kernelHeight   = boost::numeric_cast<unsigned int>(kernel.shape()[2]);
-    unsigned int kernelWidth    = boost::numeric_cast<unsigned int>(kernel.shape()[3]);
+    unsigned int kernelChanMul  = boost::numeric_cast<unsigned int>(originalKernel.shape()[0]);
+    unsigned int kernelChannels = boost::numeric_cast<unsigned int>(originalKernel.shape()[1]);
+    unsigned int kernelHeight   = boost::numeric_cast<unsigned int>(originalKernel.shape()[2]);
+    unsigned int kernelWidth    = boost::numeric_cast<unsigned int>(originalKernel.shape()[3]);
     unsigned int outputNum      = boost::numeric_cast<unsigned int>(outputExpected.shape()[0]);
     unsigned int outputChannels = boost::numeric_cast<unsigned int>(outputExpected.shape()[1]);
     unsigned int outputHeight   = boost::numeric_cast<unsigned int>(outputExpected.shape()[2]);
@@ -367,10 +368,9 @@ LayerTestResult<T, 4> DepthwiseConvolution2dAsymmetricTestImpl(armnn::IWorkloadF
     BOOST_ASSERT(!biasEnabled || bias.size() == outputChannels);
 
     // Creates the tensors.
-    armnn::TensorInfo inputTensorInfo({inputNum, inputChannels, inputHeight, inputWidth}, armnn::GetDataType<T>());
-    armnn::TensorInfo outputTensorInfo({outputNum, outputChannels, outputHeight, outputWidth},
-                                       armnn::GetDataType<T>());
-    armnn::TensorInfo kernelDesc({kernelChanMul, kernelChannels, kernelHeight, kernelWidth}, armnn::GetDataType<T>());
+    armnn::TensorInfo inputTensorInfo = GetTensorInfo<T>(inputNum, inputChannels, inputHeight, inputWidth, layout);
+    armnn::TensorInfo outputTensorInfo = GetTensorInfo<T>(outputNum, outputChannels, outputHeight, outputWidth, layout);
+    armnn::TensorInfo kernelDesc = GetTensorInfo<T>(kernelChanMul, kernelChannels, kernelHeight, kernelWidth, layout);
     armnn::TensorInfo biasDesc({static_cast<unsigned int>(bias.size())}, armnn::GetDataType<B>());
 
     // Set quantization parameters if the requested type is a quantized type.
@@ -389,6 +389,16 @@ LayerTestResult<T, 4> DepthwiseConvolution2dAsymmetricTestImpl(armnn::IWorkloadF
     // Construct the input data.
     std::vector<T> inputData;
     inputData.assign(input.data(), input.data() + inputChannels*inputHeight*inputWidth);
+
+    // At this point if we require it permute the input data
+    const armnn::PermutationVector NCHWToNHWC = { 0, 3, 1, 2 };
+    if (layout.GetDataLayout() == armnn::DataLayout::NHWC)
+    {
+        std::vector<T> tmp(inputData.size());
+        armnnUtils::Permute(inputTensorInfo.GetShape(), NCHWToNHWC, inputData.data(), tmp.data());
+        inputData = tmp;
+    }
+
     auto batchedInput = MakeTensor<T, 4>(inputTensorInfo, inputData);
 
     // Construct the output data, with bias applied, as appropriate.
@@ -404,12 +414,29 @@ LayerTestResult<T, 4> DepthwiseConvolution2dAsymmetricTestImpl(armnn::IWorkloadF
     }
 
     LayerTestResult<T, 4> ret(outputTensorInfo);
+
+    // At this point if we require it permute the expected output
+    if (layout.GetDataLayout() == armnn::DataLayout::NHWC)
+    {
+        std::vector<T> tmp(outputData.size());
+        armnnUtils::Permute(outputTensorInfo.GetShape(), NCHWToNHWC, outputData.data(), tmp.data());
+        outputData = tmp;
+    }
+
     ret.outputExpected = MakeTensor<T, 4>(outputTensorInfo, outputData);
 
     std::unique_ptr<armnn::ITensorHandle> inputHandle = workloadFactory.CreateTensorHandle(inputTensorInfo);
     std::unique_ptr<armnn::ITensorHandle> outputHandle = workloadFactory.CreateTensorHandle(outputTensorInfo);
 
     armnn::ScopedCpuTensorHandle weightsTensor(kernelDesc);
+
+    // Permute the kernel if necessary
+    boost::multi_array<T, 4> kernel = boost::multi_array<T, 4>(originalKernel);
+    if (layout.GetDataLayout() == armnn::DataLayout::NHWC)
+    {
+        armnnUtils::Permute(kernelDesc.GetShape(), NCHWToNHWC, originalKernel.data(), kernel.data());
+    }
+
     AllocateAndCopyDataToITensorHandle(&weightsTensor, &kernel[0][0][0][0]);
 
     armnn::ScopedCpuTensorHandle biasTensor(biasDesc);
@@ -428,6 +455,7 @@ LayerTestResult<T, 4> DepthwiseConvolution2dAsymmetricTestImpl(armnn::IWorkloadF
     data.m_Parameters.m_PadTop = padTop;
     data.m_Parameters.m_PadBottom = padBottom;
     data.m_Parameters.m_BiasEnabled = biasEnabled;
+    data.m_Parameters.m_DataLayout = layout.GetDataLayout();
 
     armnn::WorkloadInfo info;
     AddInputToWorkload(data, info, inputTensorInfo, inputHandle.get());
diff --git a/src/backends/test/LayerTests.cpp b/src/backends/test/LayerTests.cpp
index bc9e116f92..43a42f305c 100755
--- a/src/backends/test/LayerTests.cpp
+++ b/src/backends/test/LayerTests.cpp
@@ -432,7 +432,8 @@ template<typename T>
 LayerTestResult<T, 4> DepthwiseConvolution2dAsymmetricTestCommon(armnn::IWorkloadFactory& workloadFactory,
                                                                  float qScale,
                                                                  int32_t qOffset,
-                                                                 bool biasEnabled)
+                                                                 bool biasEnabled,
+                                                                 const armnn::DataLayoutIndexed& layout)
 {
     // Use a single-batch 2-channel 5x5 image as input.
     armnn::TensorInfo inputTensorInfo({ 1, 2, 5, 5 }, armnn::GetDataType<T>());
@@ -490,6 +491,7 @@ LayerTestResult<T, 4> DepthwiseConvolution2dAsymmetricTestCommon(armnn::IWorkloa
         expectedOutput,
         qScale,
         qOffset,
+        layout,
         1,  // Padding left.
         1,  // Padding top.
         2,  // Padding right.
@@ -643,13 +645,14 @@ LayerTestResult<float, 4> DepthwiseConvolution2dDepthMul1Test(armnn::IWorkloadFa
 }
 
 LayerTestResult<float, 4> DepthwiseConvolution2dAsymmetricTest(armnn::IWorkloadFactory& workloadFactory,
-                                                               bool                     biasEnabled)
+                                                               bool biasEnabled,
+                                                               const armnn::DataLayoutIndexed& layout)
 {
-    return DepthwiseConvolution2dAsymmetricTestCommon<float>(workloadFactory, 0.0f, 0, biasEnabled);
+    return DepthwiseConvolution2dAsymmetricTestCommon<float>(workloadFactory, 0.0f, 0, biasEnabled, layout);
 }
 
 LayerTestResult<uint8_t, 4> DepthwiseConvolution2dUint8Test(armnn::IWorkloadFactory& workloadFactory,
-                                                            bool                     biasEnabled)
+                                                            bool biasEnabled)
 {
     return DepthwiseConvolution2dTestImpl<uint8_t, int32_t>(workloadFactory, 0.5f, 50, biasEnabled);
 }
@@ -671,7 +674,7 @@ LayerTestResult<uint8_t, 4> Convolution1dUint8Test(armnn::IWorkloadFactory& work
 }
 
 LayerTestResult<float,4> CompareConvolution2dTest(armnn::IWorkloadFactory& workloadFactory,
-                                                armnn::IWorkloadFactory& refWorkloadFactory)
+                                                  armnn::IWorkloadFactory& refWorkloadFactory)
 {
     return CompareConvolution2dTestImpl<float>(workloadFactory, refWorkloadFactory);
 }
diff --git a/src/backends/test/LayerTests.hpp b/src/backends/test/LayerTests.hpp
index 8846297ff1..fec73d08f4 100644
--- a/src/backends/test/LayerTests.hpp
+++ b/src/backends/test/LayerTests.hpp
@@ -81,7 +81,8 @@ LayerTestResult<float, 4> DepthwiseConvolution2dDepthMul1Test(armnn::IWorkloadFa
                                                               bool biasEnabled);
 
 LayerTestResult<float, 4> DepthwiseConvolution2dAsymmetricTest(armnn::IWorkloadFactory& workloadFactory,
-                                                               bool biasEnabled);
+                                                               bool biasEnabled,
+                                                               const armnn::DataLayoutIndexed& layout);
 
 LayerTestResult<float,   4> SimpleMaxPooling2dSize2x2Stride2x2Test(armnn::IWorkloadFactory& workloadFactory,
                                                                    bool forceNoPadding);