IVGCVSW-6174 Add Cl Pooling3d Workload

 * Add IsSupported for Pooling3d
 * Add CreateWorkload case for Pooling3d
 * Create new ClPooling3dWorkload header and source files
 * Add Pooling3d workload to ClWorkloads.hpp
 * Add tests for Pooling3d workload
 * Add Pooling3d build function to ArmComputeTensorUtils

Change-Id: Ia270b0fe809a171ed73af14376de8708b346d500
Signed-off-by: Ryan OShea <ryan.oshea3@arm.com>

9 files changed, 224 insertions, 2 deletions

diff --git a/src/backends/cl/ClLayerSupport.cpp b/src/backends/cl/ClLayerSupport.cpp
index e52f578bc0..4dcaca9576 100644
--- a/src/backends/cl/ClLayerSupport.cpp
+++ b/src/backends/cl/ClLayerSupport.cpp
@@ -59,6 +59,7 @@
 #include "workloads/ClPadWorkload.hpp"
 #include "workloads/ClPermuteWorkload.hpp"
 #include "workloads/ClPooling2dWorkload.hpp"
+#include "workloads/ClPooling3dWorkload.hpp"
 #include "workloads/ClPreluWorkload.hpp"
 #include "workloads/ClQLstmWorkload.hpp"
 #include "workloads/ClQuantizedLstmWorkload.hpp"
@@ -449,6 +450,11 @@ bool ClLayerSupport::IsLayerSupported(const LayerType& type,
                                         infos[1],
                                         *(PolymorphicDowncast<const Pooling2dDescriptor*>(&descriptor)),
                                         reasonIfUnsupported);
+        case LayerType::Pooling3d:
+            return IsPooling3dSupported(infos[0],
+                                        infos[1],
+                                        *(PolymorphicDowncast<const Pooling3dDescriptor*>(&descriptor)),
+                                        reasonIfUnsupported);
         case LayerType::Prelu:
             return IsPreluSupported(infos[0], infos[1], infos[2], reasonIfUnsupported);
         case LayerType::QLstm:
@@ -1194,6 +1200,14 @@ bool ClLayerSupport::IsPooling2dSupported(const TensorInfo& input,
     FORWARD_WORKLOAD_VALIDATE_FUNC(ClPooling2dWorkloadValidate, reasonIfUnsupported, input, output, descriptor);
 }
 
+bool ClLayerSupport::IsPooling3dSupported(const TensorInfo& input,
+                                          const TensorInfo& output,
+                                          const Pooling3dDescriptor& descriptor,
+                                          Optional<std::string&> reasonIfUnsupported) const
+{
+    FORWARD_WORKLOAD_VALIDATE_FUNC(ClPooling3dWorkloadValidate, reasonIfUnsupported, input, output, descriptor);
+}
+
 bool ClLayerSupport::IsPreluSupported(const armnn::TensorInfo &input,
                                       const armnn::TensorInfo &alpha,
                                       const armnn::TensorInfo &output,
diff --git a/src/backends/cl/ClLayerSupport.hpp b/src/backends/cl/ClLayerSupport.hpp
index 103944e05d..b4d0e8298a 100644
--- a/src/backends/cl/ClLayerSupport.hpp
+++ b/src/backends/cl/ClLayerSupport.hpp
@@ -232,6 +232,11 @@ public:
                               const Pooling2dDescriptor& descriptor,
                               Optional<std::string&> reasonIfUnsupported = EmptyOptional()) const override;
 
+    bool IsPooling3dSupported(const TensorInfo& input,
+                              const TensorInfo& output,
+                              const Pooling3dDescriptor& descriptor,
+                              Optional<std::string&> reasonIfUnsupported = EmptyOptional()) const override;
+
     bool IsPreluSupported(const TensorInfo& input,
                           const TensorInfo& alpha,
                           const TensorInfo& output,
diff --git a/src/backends/cl/ClWorkloadFactory.cpp b/src/backends/cl/ClWorkloadFactory.cpp
index c561bf2157..213f474a97 100644
--- a/src/backends/cl/ClWorkloadFactory.cpp
+++ b/src/backends/cl/ClWorkloadFactory.cpp
@@ -581,6 +581,11 @@ std::unique_ptr<IWorkload> ClWorkloadFactory::CreateWorkload(LayerType type,
             auto pooling2dQueueDescriptor = PolymorphicDowncast<const Pooling2dQueueDescriptor*>(&descriptor);
             return MakeWorkload<ClPooling2dWorkload>(*pooling2dQueueDescriptor, info, m_CLCompileContext);
         }
+        case LayerType::Pooling3d :
+        {
+            auto pooling3dQueueDescriptor = PolymorphicDowncast<const Pooling3dQueueDescriptor*>(&descriptor);
+            return MakeWorkload<ClPooling3dWorkload>(*pooling3dQueueDescriptor, info, m_CLCompileContext);
+        }
         case LayerType::PreCompiled :
         {
             auto preCompiledQueueDescriptor = PolymorphicDowncast<const PreCompiledQueueDescriptor*>(&descriptor);
diff --git a/src/backends/cl/backend.mk b/src/backends/cl/backend.mk
index fde6b7eeb2..ed0d9252ad 100644
--- a/src/backends/cl/backend.mk
+++ b/src/backends/cl/backend.mk
@@ -67,6 +67,7 @@ BACKEND_SOURCES := \
         workloads/ClPadWorkload.cpp \
         workloads/ClPermuteWorkload.cpp \
         workloads/ClPooling2dWorkload.cpp \
+        workloads/ClPooling3dWorkload.cpp \
         workloads/ClPreluWorkload.cpp \
         workloads/ClQLstmWorkload.cpp \
         workloads/ClQuantizedLstmWorkload.cpp \
diff --git a/src/backends/cl/test/ClLayerTests.cpp b/src/backends/cl/test/ClLayerTests.cpp
index 1c342149db..fd24043405 100644
--- a/src/backends/cl/test/ClLayerTests.cpp
+++ b/src/backends/cl/test/ClLayerTests.cpp
@@ -469,7 +469,7 @@ ARMNN_AUTO_TEST_FIXTURE_WITH_THF(AcrossChannelNormalization,
                                  ClContextControlFixture,
                                  AcrossChannelNormalizationTest)
 
-// Pooling
+// Pooling2d
 ARMNN_AUTO_TEST_FIXTURE_WITH_THF(SimpleMaxPooling2dSize3x3Stride2x4,
                                  ClContextControlFixture,
                                  SimpleMaxPooling2dSize3x3Stride2x4Test,
@@ -611,6 +611,85 @@ ARMNN_AUTO_TEST_FIXTURE_WITH_THF(UNSUPPORTED_L2Pooling2dSize7Uint8,
 ARMNN_AUTO_TEST_FIXTURE_WITH_THF(L2Pooling2dSize9, ClContextControlFixture, L2Pooling2dSize9Test)
 ARMNN_AUTO_TEST_FIXTURE_WITH_THF(UNSUPPORTED_L2Pooling2dSize9Uint8, ClContextControlFixture, L2Pooling2dSize9Uint8Test)
 
+// Pooling3d
+ARMNN_AUTO_TEST_FIXTURE_WITH_THF(SimpleMaxPooling3dSize2x2x2Stride1x1x1,
+                                 ClContextControlFixture,
+                                 SimpleMaxPooling3dSize2x2x2Stride1x1x1Test,
+                                 DataLayout::NDHWC)
+
+ARMNN_AUTO_TEST_FIXTURE_WITH_THF(SimpleMaxPooling3d,
+                                 ClContextControlFixture,
+                                 SimpleMaxPooling3dTest,
+                                 DataLayout::NDHWC)
+
+ARMNN_AUTO_TEST_FIXTURE_WITH_THF(IgnorePaddingSimpleMaxPooling3d,
+                                 ClContextControlFixture,
+                                 IgnorePaddingSimpleMaxPooling3dTest,
+                                 DataLayout::NDHWC)
+
+ARMNN_AUTO_TEST_FIXTURE_WITH_THF(SimpleAveragePooling3d,
+                                 ClContextControlFixture,
+                                 SimpleAveragePooling3dTest,
+                                 DataLayout::NDHWC)
+
+ARMNN_AUTO_TEST_FIXTURE_WITH_THF(LargeTensorsAveragePooling3d,
+                                 ClContextControlFixture,
+                                 LargeTensorsAveragePooling3dTest,
+                                 DataLayout::NDHWC)
+
+ARMNN_AUTO_TEST_FIXTURE_WITH_THF(IgnorePaddingSimpleAveragePooling3d,
+                                 ClContextControlFixture,
+                                 IgnorePaddingSimpleAveragePooling3dTest,
+                                 DataLayout::NDHWC)
+
+ARMNN_AUTO_TEST_FIXTURE_WITH_THF(SimpleL2Pooling3d,
+                                 ClContextControlFixture,
+                                 SimpleL2Pooling3dTest,
+                                 DataLayout::NDHWC)
+
+ARMNN_AUTO_TEST_FIXTURE_WITH_THF(IgnorePaddingSimpleL2Pooling3d,
+                                 ClContextControlFixture,
+                                 IgnorePaddingSimpleL2Pooling3dTest,
+                                 DataLayout::NDHWC)
+
+ARMNN_AUTO_TEST_FIXTURE_WITH_THF(AsymmetricNonSquareMaxPooling3d,
+                                 ClContextControlFixture,
+                                 AsymmetricNonSquareMaxPooling3dTest,
+                                 DataLayout::NDHWC)
+
+ARMNN_AUTO_TEST_FIXTURE_WITH_THF(AsymmetricNonSquareAveragePooling3d,
+                                 ClContextControlFixture,
+                                 AsymmetricNonSquareAveragePooling3dTest,
+                                 DataLayout::NDHWC)
+
+ARMNN_AUTO_TEST_FIXTURE_WITH_THF(AsymmetricNonSquareL2Pooling3d,
+                                 ClContextControlFixture,
+                                 AsymmetricNonSquareL2Pooling3dTest,
+                                 DataLayout::NDHWC)
+
+ARMNN_AUTO_TEST_FIXTURE_WITH_THF(UNSUPPORTED_AsymmetricNonSquareMaxPooling3dWithPaddingOnlyPool,
+                                 ClContextControlFixture,
+                                 AsymmetricNonSquareMaxPooling3dWithPaddingOnlyPoolTest,
+                                 DataLayout::NDHWC)
+
+ARMNN_AUTO_TEST_FIXTURE_WITH_THF(UNSUPPORTED_AsymmetricNonSquareAveragePooling3dWithPaddingOnlyPool,
+                                 ClContextControlFixture,
+                                 AsymmetricNonSquareAveragePooling3dWithPaddingOnlyPoolTest,
+                                 DataLayout::NDHWC)
+
+ARMNN_AUTO_TEST_FIXTURE_WITH_THF(UNSUPPORTEDAsymmetricNonSquareL2Pooling3dWithPaddingOnlyPool,
+                                 ClContextControlFixture,
+                                 AsymmetricNonSquareL2Pooling3dWithPaddingOnlyPoolTest,
+                                 DataLayout::NDHWC)
+
+ARMNN_AUTO_TEST_FIXTURE_WITH_THF(SimpleAveragePooling3d,
+                                 ClContextControlFixture,
+                                 SimpleAveragePooling3dTest,
+                                 DataLayout::NDHWC)
+
+
+
+
 // Add
 ARMNN_AUTO_TEST_FIXTURE_WITH_THF(SimpleAdd, ClContextControlFixture, AdditionTest)
 ARMNN_AUTO_TEST_FIXTURE_WITH_THF(Add5d, ClContextControlFixture, Addition5dTest)
@@ -2010,7 +2089,6 @@ TEST_CASE_FIXTURE(ClContextControlFixture, "ClContextControlFixture") {}
 // The ARMNN_COMPARE_REF_AUTO_TEST_CASE and the ARMNN_COMPARE_REF_FIXTURE_TEST_CASE test units are not available
 // if the reference backend is not built
 
-// ============================================================================
 // COMPARE tests
 
 ARMNN_COMPARE_REF_AUTO_TEST_CASE_WITH_THF(CompareSoftmaxBeta1WithReference, CompareSoftmaxTest, 1.0f)
@@ -2049,6 +2127,15 @@ ARMNN_COMPARE_REF_AUTO_TEST_CASE_WITH_THF(CompareAveragePooling2dWithRefUint8, C
 
 ARMNN_COMPARE_REF_AUTO_TEST_CASE_WITH_THF(CompareL2Pooling2dWithRef, ComparePooling2dTest, PoolingAlgorithm::L2)
 
+ARMNN_COMPARE_REF_AUTO_TEST_CASE_WITH_THF(CompareMaxPooling3dWithRef, ComparePooling3dTest, PoolingAlgorithm::Max,
+                                          DataLayout::NDHWC)
+
+ARMNN_COMPARE_REF_AUTO_TEST_CASE_WITH_THF(CompareAveragePooling3dWithRef, ComparePooling3dTest,
+                                          PoolingAlgorithm::Average, DataLayout::NDHWC)
+
+ARMNN_COMPARE_REF_AUTO_TEST_CASE_WITH_THF(CompareL2Pooling3dWithRef, ComparePooling3dTest, PoolingAlgorithm::L2,
+                                          DataLayout::NDHWC)
+
 ARMNN_COMPARE_REF_AUTO_TEST_CASE_WITH_THF(CompareAddition, CompareAdditionTest)
 
 ARMNN_COMPARE_REF_AUTO_TEST_CASE_WITH_THF(CompareMultiplicationWithRef, CompareMultiplicationTest)
diff --git a/src/backends/cl/workloads/CMakeLists.txt b/src/backends/cl/workloads/CMakeLists.txt
index 6bb8d68d07..52326ae4a9 100644
--- a/src/backends/cl/workloads/CMakeLists.txt
+++ b/src/backends/cl/workloads/CMakeLists.txt
@@ -86,6 +86,8 @@ list(APPEND armnnClBackendWorkloads_sources
     ClPermuteWorkload.hpp
     ClPooling2dWorkload.cpp
     ClPooling2dWorkload.hpp
+    ClPooling3dWorkload.cpp
+    ClPooling3dWorkload.hpp
     ClPreluWorkload.cpp
     ClPreluWorkload.hpp
     ClQLstmWorkload.cpp
diff --git a/src/backends/cl/workloads/ClPooling3dWorkload.cpp b/src/backends/cl/workloads/ClPooling3dWorkload.cpp
new file mode 100644
index 0000000000..a896110a2e
--- /dev/null
+++ b/src/backends/cl/workloads/ClPooling3dWorkload.cpp
@@ -0,0 +1,73 @@
+//
+// Copyright © 2022 Arm Ltd and Contributors. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#include "ClPooling3dWorkload.hpp"
+#include <cl/ClLayerSupport.hpp>
+#include <cl/ClTensorHandle.hpp>
+#include <aclCommon/ArmComputeUtils.hpp>
+#include <aclCommon/ArmComputeTensorUtils.hpp>
+
+#include "ClWorkloadUtils.hpp"
+
+namespace armnn
+{
+    using namespace armcomputetensorutils;
+
+    arm_compute::Status ClPooling3dWorkloadValidate(const TensorInfo& input,
+                                                    const TensorInfo& output,
+                                                    const Pooling3dDescriptor& descriptor)
+    {
+        const arm_compute::TensorInfo aclInputInfo = BuildArmComputeTensorInfo(input, descriptor.m_DataLayout);
+        const arm_compute::TensorInfo aclOutputInfo = BuildArmComputeTensorInfo(output, descriptor.m_DataLayout);
+
+        arm_compute::Pooling3dLayerInfo layerInfo = BuildArmComputePooling3dLayerInfo(descriptor);
+
+        return arm_compute::CLPooling3dLayer::validate(&aclInputInfo, &aclOutputInfo, layerInfo);
+    }
+
+    ClPooling3dWorkload::ClPooling3dWorkload( const Pooling3dQueueDescriptor& descriptor,
+                                              const WorkloadInfo& info,
+                                              const arm_compute::CLCompileContext& clCompileContext)
+                                              : ClBaseWorkload<Pooling3dQueueDescriptor>(descriptor, info)
+    {
+        // Report Profiling Details
+        ARMNN_REPORT_PROFILING_WORKLOAD_DESC("ClPooling3dWorkload_Construct",
+                                             descriptor.m_Parameters,
+                                             info,
+                                             this->GetGuid());
+
+        m_Data.ValidateInputsOutputs("ClPooling3dWorkload", 1, 1);
+
+        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);
+
+        // flag to use wider accumulators (32 bit instead of 16 for FP16) to improve accuracy
+        // enable fp_mixed_precision for the the FP16 cases that
+        // accumulation reaches a limit beyond which there is no more increment of the value
+        bool fpMixedPrecision = false;
+
+        arm_compute::Pooling3dLayerInfo layerInfo = BuildArmComputePooling3dLayerInfo(m_Data.m_Parameters,
+                                                                                      fpMixedPrecision);
+
+        {
+            ARMNN_SCOPED_PROFILING_EVENT(Compute::Undefined, "ClPooling3dWorkload_configure");
+            // Run the layer.
+            m_PoolingLayer.configure(clCompileContext, &input, &output, layerInfo);
+        }
+    }
+
+    void ClPooling3dWorkload::Execute() const
+    {
+        ARMNN_SCOPED_PROFILING_EVENT_CL_GUID("ClPooling3dWorkload_Execute", this->GetGuid());
+        RunClFunction(m_PoolingLayer, CHECK_LOCATION());
+    }
+
+}
+
+
diff --git a/src/backends/cl/workloads/ClPooling3dWorkload.hpp b/src/backends/cl/workloads/ClPooling3dWorkload.hpp
new file mode 100644
index 0000000000..75777db188
--- /dev/null
+++ b/src/backends/cl/workloads/ClPooling3dWorkload.hpp
@@ -0,0 +1,34 @@
+//
+// Copyright © 2022 Arm Ltd and Contributors. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#pragma once
+
+#include "ClBaseWorkload.hpp"
+
+#include <arm_compute/runtime/CL/functions/CLPooling3dLayer.h>
+
+namespace armnn
+{
+
+    arm_compute::Status ClPooling3dWorkloadValidate(const TensorInfo& input,
+                                                    const TensorInfo& output,
+                                                    const Pooling3dDescriptor& descriptor);
+
+    class ClPooling3dWorkload : public ClBaseWorkload<Pooling3dQueueDescriptor>
+    {
+    public:
+        using BaseWorkload<Pooling3dQueueDescriptor>::m_Data;
+
+        ClPooling3dWorkload(const Pooling3dQueueDescriptor& descriptor,
+                            const WorkloadInfo& info,
+                            const arm_compute::CLCompileContext& clCompileContext);
+
+        void Execute() const override;
+
+    private:
+        mutable arm_compute::CLPooling3dLayer m_PoolingLayer;
+    };
+
+} //namespace armnn
diff --git a/src/backends/cl/workloads/ClWorkloads.hpp b/src/backends/cl/workloads/ClWorkloads.hpp
index 3558da341e..27119bb2dc 100644
--- a/src/backends/cl/workloads/ClWorkloads.hpp
+++ b/src/backends/cl/workloads/ClWorkloads.hpp
@@ -43,6 +43,7 @@
 #include "ClPermuteWorkload.hpp"
 #include "ClPadWorkload.hpp"
 #include "ClPooling2dWorkload.hpp"
+#include "ClPooling3dWorkload.hpp"
 #include "ClPreluWorkload.hpp"
 #include "ClQLstmWorkload.hpp"
 #include "ClQuantizeWorkload.hpp"