IVGCVSW-1860 Support NHWC for Pooling2D

Change-Id: I9059a292e47867df82a5efbba5808fd264949ec9

8 files changed, 148 insertions, 13 deletions

diff --git a/include/armnn/Descriptors.hpp b/include/armnn/Descriptors.hpp
index 2de031e94a..c2efa211a6 100644
--- a/include/armnn/Descriptors.hpp
+++ b/include/armnn/Descriptors.hpp
@@ -180,6 +180,7 @@ struct Pooling2dDescriptor
     , m_StrideY(0)
     , m_OutputShapeRounding(OutputShapeRounding::Floor)
     , m_PaddingMethod(PaddingMethod::Exclude)
+    , m_DataLayout(DataLayout::NCHW)
     {};
 
     PoolingAlgorithm    m_PoolType;
@@ -193,6 +194,7 @@ struct Pooling2dDescriptor
     uint32_t            m_StrideY;
     OutputShapeRounding m_OutputShapeRounding;
     PaddingMethod       m_PaddingMethod;
+    DataLayout          m_DataLayout;
 };
 
 struct FullyConnectedDescriptor
diff --git a/src/backends/cl/workloads/ClPooling2dBaseWorkload.cpp b/src/backends/cl/workloads/ClPooling2dBaseWorkload.cpp
index 98911856fe..e61ad4f28e 100644
--- a/src/backends/cl/workloads/ClPooling2dBaseWorkload.cpp
+++ b/src/backends/cl/workloads/ClPooling2dBaseWorkload.cpp
@@ -17,8 +17,8 @@ arm_compute::Status ClPooling2dWorkloadValidate(const TensorInfo& input,
     const TensorInfo& output,
     const Pooling2dDescriptor& descriptor)
 {
-    const arm_compute::TensorInfo aclInputInfo = BuildArmComputeTensorInfo(input);
-    const arm_compute::TensorInfo aclOutputInfo = BuildArmComputeTensorInfo(output);
+    const arm_compute::TensorInfo aclInputInfo = BuildArmComputeTensorInfo(input, descriptor.m_DataLayout);
+    const arm_compute::TensorInfo aclOutputInfo = BuildArmComputeTensorInfo(output, descriptor.m_DataLayout);
 
     arm_compute::PoolingLayerInfo layerInfo = BuildArmComputePoolingLayerInfo(descriptor);
 
@@ -35,6 +35,10 @@ ClPooling2dBaseWorkload<dataTypes...>::ClPooling2dBaseWorkload(
     arm_compute::ICLTensor& input = static_cast<IClTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
     arm_compute::ICLTensor& output = static_cast<IClTensorHandle*>(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);
+
     arm_compute::PoolingLayerInfo layerInfo = BuildArmComputePoolingLayerInfo(m_Data.m_Parameters);
 
     // Run the layer.
diff --git a/src/backends/neon/workloads/NeonPooling2dBaseWorkload.cpp b/src/backends/neon/workloads/NeonPooling2dBaseWorkload.cpp
index 109e856506..054f9c5f75 100644
--- a/src/backends/neon/workloads/NeonPooling2dBaseWorkload.cpp
+++ b/src/backends/neon/workloads/NeonPooling2dBaseWorkload.cpp
@@ -17,8 +17,8 @@ arm_compute::Status NeonPooling2dWorkloadValidate(const TensorInfo& input,
     const TensorInfo& output,
     const Pooling2dDescriptor& descriptor)
 {
-    const arm_compute::TensorInfo aclInputInfo = BuildArmComputeTensorInfo(input);
-    const arm_compute::TensorInfo aclOutputInfo = BuildArmComputeTensorInfo(output);
+    const arm_compute::TensorInfo aclInputInfo = BuildArmComputeTensorInfo(input, descriptor.m_DataLayout);
+    const arm_compute::TensorInfo aclOutputInfo = BuildArmComputeTensorInfo(output, descriptor.m_DataLayout);
 
     arm_compute::PoolingLayerInfo layerInfo = BuildArmComputePoolingLayerInfo(descriptor);
 
@@ -35,6 +35,10 @@ NeonPooling2dBaseWorkload<dataTypes...>::NeonPooling2dBaseWorkload(
     arm_compute::ITensor& input = boost::polymorphic_downcast<INeonTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
     arm_compute::ITensor& output = boost::polymorphic_downcast<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);
+
     arm_compute::PoolingLayerInfo layerInfo = BuildArmComputePoolingLayerInfo(m_Data.m_Parameters);
 
     m_PoolingLayer.configure(&input, &output, layerInfo);
diff --git a/src/backends/test/ArmComputeCl.cpp b/src/backends/test/ArmComputeCl.cpp
index faa3d4f37b..f4ead34c58 100644
--- a/src/backends/test/ArmComputeCl.cpp
+++ b/src/backends/test/ArmComputeCl.cpp
@@ -122,6 +122,7 @@ ARMNN_AUTO_TEST_CASE(IgnorePaddingL2Pooling2dSize3, IgnorePaddingL2Pooling2dSize
 ARMNN_AUTO_TEST_CASE(UNSUPPORTED_IgnorePaddingL2Pooling2dSize3Uint8, IgnorePaddingL2Pooling2dSize3Uint8Test)
 
 ARMNN_AUTO_TEST_CASE(SimpleAveragePooling2d, SimpleAveragePooling2dTest)
+ARMNN_AUTO_TEST_CASE(SimpleAveragePooling2dNhwc, SimpleAveragePooling2dNhwcTest)
 ARMNN_AUTO_TEST_CASE(SimpleAveragePooling2dUint8, SimpleAveragePooling2dUint8Test)
 ARMNN_AUTO_TEST_CASE(IgnorePaddingAveragePooling2dSize3x2Stride2x2,
                      IgnorePaddingAveragePooling2dSize3x2Stride2x2Test,
diff --git a/src/backends/test/ArmComputeNeon.cpp b/src/backends/test/ArmComputeNeon.cpp
index 7a60c31f73..045aa30889 100644
--- a/src/backends/test/ArmComputeNeon.cpp
+++ b/src/backends/test/ArmComputeNeon.cpp
@@ -232,6 +232,7 @@ BOOST_AUTO_TEST_CASE(DepthwiseConv2dUtils)
 ARMNN_AUTO_TEST_CASE(SimpleMaxPooling2dSize3x3Stride2x4, SimpleMaxPooling2dSize3x3Stride2x4Test, true)
 ARMNN_AUTO_TEST_CASE(SimpleMaxPooling2dSize3x3Stride2x4Uint8, SimpleMaxPooling2dSize3x3Stride2x4Uint8Test, true)
 ARMNN_AUTO_TEST_CASE(SimpleAveragePooling2d, SimpleAveragePooling2dTest)
+ARMNN_AUTO_TEST_CASE(SimpleAveragePooling2dNhwc, SimpleAveragePooling2dNhwcTest)
 ARMNN_AUTO_TEST_CASE(SimpleAveragePooling2dUint8, SimpleAveragePooling2dUint8Test)
 
 ARMNN_AUTO_TEST_CASE(LargeTensorsAveragePooling2d, LargeTensorsAveragePooling2dTest)
diff --git a/src/backends/test/LayerTests.cpp b/src/backends/test/LayerTests.cpp
index a7fb6a824e..17f3ae12e1 100644
--- a/src/backends/test/LayerTests.cpp
+++ b/src/backends/test/LayerTests.cpp
@@ -4929,6 +4929,11 @@ LayerTestResult<float, 4> SimpleAveragePooling2dTest(armnn::IWorkloadFactory& wo
     return SimpleAveragePooling2dTestCommon<float>(workloadFactory);
 }
 
+LayerTestResult<float, 4> SimpleAveragePooling2dNhwcTest(armnn::IWorkloadFactory& workloadFactory)
+{
+    return SimpleAveragePooling2dNhwcTestCommon<float>(workloadFactory);
+}
+
 LayerTestResult<uint8_t, 4> SimpleAveragePooling2dUint8Test(armnn::IWorkloadFactory& workloadFactory)
 {
     return SimpleAveragePooling2dTestCommon<uint8_t>(workloadFactory, 0.5, -1);
diff --git a/src/backends/test/LayerTests.hpp b/src/backends/test/LayerTests.hpp
index 71a468b36b..f5abd985c8 100644
--- a/src/backends/test/LayerTests.hpp
+++ b/src/backends/test/LayerTests.hpp
@@ -88,6 +88,7 @@ LayerTestResult<float,   4> IgnorePaddingMaxPooling2dSize3Test(armnn::IWorkloadF
 LayerTestResult<uint8_t, 4> IgnorePaddingMaxPooling2dSize3Uint8Test(armnn::IWorkloadFactory& workloadFactory);
 
 LayerTestResult<float,   4> SimpleAveragePooling2dTest(armnn::IWorkloadFactory& workloadFactory);
+LayerTestResult<float,   4> SimpleAveragePooling2dNhwcTest(armnn::IWorkloadFactory& workloadFactory);
 LayerTestResult<uint8_t, 4> SimpleAveragePooling2dUint8Test(armnn::IWorkloadFactory& workloadFactory);
 LayerTestResult<float,   4> IgnorePaddingAveragePooling2dSize3x2Stride2x2Test(armnn::IWorkloadFactory& workloadFactory,
                                                                               bool forceNoPadding);
diff --git a/src/backends/test/Pooling2dTestImpl.hpp b/src/backends/test/Pooling2dTestImpl.hpp
index c87548cd5b..7366bcaf44 100644
--- a/src/backends/test/Pooling2dTestImpl.hpp
+++ b/src/backends/test/Pooling2dTestImpl.hpp
@@ -5,25 +5,23 @@
 #pragma once
 
 #include <armnn/ArmNN.hpp>
-#include <armnn/Tensor.hpp>
-#include <armnn/TypesUtils.hpp>
 
 #include <test/TensorHelpers.hpp>
 #include "QuantizeHelper.hpp"
 
 #include <backends/CpuTensorHandle.hpp>
 #include <backends/WorkloadFactory.hpp>
-
+#include <backends/WorkloadInfo.hpp>
 #include <algorithm>
 
 template<typename T>
 LayerTestResult<T, 4> SimplePooling2dTestImpl(
-    armnn::IWorkloadFactory& workloadFactory,
-    armnn::Pooling2dDescriptor descriptor,
-    float qScale,
-    int32_t qOffset,
-    const boost::multi_array<T, 4>& input,
-    const boost::multi_array<T, 4>& outputExpected)
+                                              armnn::IWorkloadFactory& workloadFactory,
+                                              armnn::Pooling2dDescriptor descriptor,
+                                              float qScale,
+                                              int32_t qOffset,
+                                              const boost::multi_array<T, 4>& input,
+                                              const boost::multi_array<T, 4>& outputExpected)
 {
     unsigned int inputHeight = boost::numeric_cast<unsigned int>(input.shape()[2]);
     unsigned int inputWidth = boost::numeric_cast<unsigned int>(input.shape()[3]);
@@ -88,6 +86,80 @@ LayerTestResult<T, 4> SimplePooling2dTestImpl(
     return result;
 }
 
+template<typename T>
+LayerTestResult<T, 4> SimplePooling2dNhwcTestImpl(
+                                                  armnn::IWorkloadFactory& workloadFactory,
+                                                  armnn::Pooling2dDescriptor descriptor,
+                                                  float qScale,
+                                                  int32_t qOffset,
+                                                  const boost::multi_array<T, 4>& input,
+                                                  const boost::multi_array<T, 4>& outputExpected)
+{
+    unsigned int inputHeight = boost::numeric_cast<unsigned int>(input.shape()[1]);
+    unsigned int inputWidth = boost::numeric_cast<unsigned int>(input.shape()[2]);
+    unsigned int inputChannels = boost::numeric_cast<unsigned int>(input.shape()[3]);
+    unsigned int inputBatchSize = boost::numeric_cast<unsigned int>(input.shape()[0]);
+
+    unsigned int outputHeight = boost::numeric_cast<unsigned int>(outputExpected.shape()[1]);
+    unsigned int outputWidth = boost::numeric_cast<unsigned int>(outputExpected.shape()[2]);
+    unsigned int outputChannels = boost::numeric_cast<unsigned int>(outputExpected.shape()[3]);
+    unsigned int outputBatchSize = boost::numeric_cast<unsigned int>(outputExpected.shape()[0]);
+
+    armnn::TensorInfo inputTensorInfo({ inputBatchSize, inputHeight, inputWidth, inputChannels },
+                                      armnn::GetDataType<T>());
+    armnn::TensorInfo outputTensorInfo({ outputBatchSize, outputHeight, outputWidth, outputChannels },
+                                       armnn::GetDataType<T>());
+
+    // Set quantization parameters if the requested type is a quantized type.
+    if(armnn::IsQuantizedType<T>())
+    {
+        inputTensorInfo.SetQuantizationScale(qScale);
+        inputTensorInfo.SetQuantizationOffset(qOffset);
+        outputTensorInfo.SetQuantizationScale(qScale);
+        outputTensorInfo.SetQuantizationOffset(qOffset);
+    }
+
+    LayerTestResult<T, 4> result(outputTensorInfo);
+
+    std::unique_ptr<armnn::ITensorHandle> inputHandle = workloadFactory.CreateTensorHandle(inputTensorInfo);
+    std::unique_ptr<armnn::ITensorHandle> outputHandle = workloadFactory.CreateTensorHandle(outputTensorInfo);
+
+    armnn::Pooling2dQueueDescriptor queueDescriptor;
+    queueDescriptor.m_Parameters = descriptor;
+    queueDescriptor.m_Parameters.m_DataLayout = armnn::DataLayout::NHWC;
+
+    armnn::WorkloadInfo workloadInfo;
+    AddInputToWorkload(queueDescriptor, workloadInfo, inputTensorInfo, inputHandle.get());
+    AddOutputToWorkload(queueDescriptor, workloadInfo, outputTensorInfo, outputHandle.get());
+
+    // Don't execute if Pooling is not supported, as an exception will be raised.
+    armnn::Compute compute = workloadFactory.GetCompute();
+    const size_t reasonIfUnsupportedMaxLen = 255;
+    char reasonIfUnsupported[reasonIfUnsupportedMaxLen+1];
+    result.supported = armnn::IsPooling2dSupported(compute, inputTensorInfo, outputTensorInfo,
+                                                   queueDescriptor.m_Parameters,
+                                                   reasonIfUnsupported, reasonIfUnsupportedMaxLen);
+    if (!result.supported)
+    {
+        return result;
+    }
+
+    std::unique_ptr<armnn::IWorkload> workload = workloadFactory.CreatePooling2d(queueDescriptor, workloadInfo);
+
+    inputHandle->Allocate();
+    outputHandle->Allocate();
+
+    CopyDataToITensorHandle(inputHandle.get(), &input[0][0][0][0]);
+
+    workload->Execute();
+
+    CopyDataFromITensorHandle(&result.output[0][0][0][0], outputHandle.get());
+
+    result.outputExpected = outputExpected;
+
+    return result;
+}
+
 //
 // Tests max pooling with the following parameters:
 //
@@ -263,6 +335,51 @@ LayerTestResult<T, 4> SimpleAveragePooling2dTestCommon(armnn::IWorkloadFactory&
 }
 
 template<typename T>
+LayerTestResult<T, 4> SimpleAveragePooling2dNhwcTestCommon(armnn::IWorkloadFactory& workloadFactory,
+                                                           float qScale = 1.0f,
+                                                           int32_t qOffset = 0)
+{
+    armnn::Pooling2dDescriptor descriptor;
+    descriptor.m_PoolType = armnn::PoolingAlgorithm::Average;
+    descriptor.m_PoolWidth = descriptor.m_PoolHeight = 2;
+    descriptor.m_StrideX = descriptor.m_StrideY = 2;
+    descriptor.m_PadLeft = 1;
+    descriptor.m_PadRight = 1;
+    descriptor.m_PadTop = 1;
+    descriptor.m_PadBottom = 1;
+    descriptor.m_PaddingMethod = armnn::PaddingMethod::Exclude;
+
+    armnn::TensorInfo inputTensorInfo({ 1, 4, 4, 1 }, armnn::GetDataType<T>());
+    armnn::TensorInfo outputTensorInfo({ 1, 3, 3, 1 }, armnn::GetDataType<T>());
+
+    // Set quantization parameters if the requested type is a quantized type.
+    if(armnn::IsQuantizedType<T>())
+    {
+        inputTensorInfo.SetQuantizationScale(qScale);
+        inputTensorInfo.SetQuantizationOffset(qOffset);
+        outputTensorInfo.SetQuantizationScale(qScale);
+        outputTensorInfo.SetQuantizationOffset(qOffset);
+    }
+
+    auto input = MakeTensor<T, 4>(inputTensorInfo,
+                                  QuantizedVector<T>(qScale, qOffset, {
+                                          1.0f, 2.0f, 3.0f, 4.0f,
+                                          1.0f, 2.0f, 3.0f, 4.0f,
+                                          1.0f, 2.0f, 3.0f, 4.0f,
+                                          1.0f, 2.0f, 3.0f, 4.0f,
+                                  }));
+
+    auto outputExpected = MakeTensor<T, 4>(outputTensorInfo,
+                                           QuantizedVector<T>(qScale, qOffset, {
+                                                   1.0f, 2.5f, 4.0f,
+                                                   1.0f, 2.5f, 4.0f,
+                                                   1.0f, 2.5f, 4.0f,
+                                           }));
+
+    return SimplePooling2dNhwcTestImpl<T>(workloadFactory, descriptor, qScale, qOffset, input, outputExpected);
+}
+
+template<typename T>
 LayerTestResult<T, 4> LargeTensorsAveragePooling2dTestCommon(armnn::IWorkloadFactory& workloadFactory,
                                                        float qScale = 1.0f,
                                                        int32_t qOffset = 0)