IVGCVSW-6170 Add CpuAcc Conv3d Workload

Signed-off-by: Teresa Charlin <teresa.charlinreyes@arm.com>
Change-Id: I3a5bfef5a0085d172fd3689e67f25af909ace2ee

4 files changed, 162 insertions, 0 deletions

diff --git a/src/backends/neon/workloads/CMakeLists.txt b/src/backends/neon/workloads/CMakeLists.txt
index 6451f4c5d7..0c64a19bf9 100644
--- a/src/backends/neon/workloads/CMakeLists.txt
+++ b/src/backends/neon/workloads/CMakeLists.txt
@@ -36,6 +36,8 @@ list(APPEND armnnNeonBackendWorkloads_sources
     NeonConvertFp32ToFp16Workload.hpp
     NeonConvolution2dWorkload.cpp
     NeonConvolution2dWorkload.hpp
+    NeonConvolution3dWorkload.cpp
+    NeonConvolution3dWorkload.hpp
     NeonDepthToSpaceWorkload.cpp
     NeonDepthToSpaceWorkload.hpp
     NeonDepthwiseConvolutionWorkload.cpp
diff --git a/src/backends/neon/workloads/NeonConvolution3dWorkload.cpp b/src/backends/neon/workloads/NeonConvolution3dWorkload.cpp
new file mode 100644
index 0000000000..1891981162
--- /dev/null
+++ b/src/backends/neon/workloads/NeonConvolution3dWorkload.cpp
@@ -0,0 +1,116 @@
+//
+// Copyright © 2021 Arm Ltd and Contributors. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#include "NeonConvolution3dWorkload.hpp"
+
+#include <aclCommon/ArmComputeTensorUtils.hpp>
+#include <aclCommon/ArmComputeUtils.hpp>
+#include <armnn/utility/PolymorphicDowncast.hpp>
+#include <backendsCommon/TensorHandle.hpp>
+#include <neon/workloads/NeonWorkloadUtils.hpp>
+
+#include <arm_compute/runtime/NEON/functions/NEConv3D.h>
+
+#include <armnn/Types.hpp>
+#include <Half.hpp>
+
+namespace armnn
+{
+
+using namespace armcomputetensorutils;
+
+arm_compute::Status NeonConvolution3dWorkloadValidate(const TensorInfo& input,
+                                                      const TensorInfo& output,
+                                                      const Convolution3dDescriptor& descriptor,
+                                                      const TensorInfo& weights,
+                                                      const Optional<TensorInfo>& biases,
+                                                      bool isFastMathEnabled,
+                                                      const ActivationDescriptor* activationDescriptor)
+{
+    const arm_compute::TensorInfo aclInputInfo = BuildArmComputeTensorInfo(input, descriptor.m_DataLayout);
+    const arm_compute::TensorInfo aclWeightsInfo = BuildArmComputeTensorInfo(weights, descriptor.m_DataLayout);
+    arm_compute::TensorInfo aclBiasesInfo;
+    arm_compute::TensorInfo *optionalAclBiasesInfo = nullptr;
+    if (descriptor.m_BiasEnabled)
+    {
+        ARMNN_ASSERT(biases.has_value());
+
+        aclBiasesInfo = BuildArmComputeTensorInfo(biases.value(), descriptor.m_DataLayout);
+        optionalAclBiasesInfo = &aclBiasesInfo;
+    }
+    const arm_compute::TensorInfo aclOutputInfo = BuildArmComputeTensorInfo(output, descriptor.m_DataLayout);
+
+    const arm_compute::Conv3dInfo aclConv3DInfo = ComputeConv3DInfo(descriptor,
+                                                                    isFastMathEnabled,
+                                                                    activationDescriptor);
+
+    return arm_compute::NEConv3D::validate(&aclInputInfo,
+                                           &aclWeightsInfo,
+                                           optionalAclBiasesInfo,
+                                           &aclOutputInfo,
+                                           aclConv3DInfo);
+}
+
+NeonConvolution3dWorkload::NeonConvolution3dWorkload(const Convolution3dQueueDescriptor& descriptor,
+                                                     const WorkloadInfo& info,
+                                                     std::shared_ptr<arm_compute::MemoryManagerOnDemand>& memoryManager,
+                                                     const bool isFastMathEnabled)
+    : BaseWorkload<Convolution3dQueueDescriptor>(descriptor, info)
+{
+    IgnoreUnused(memoryManager);
+
+    using arm_compute::NEConv3D;
+    uint32_t numInputs = m_Data.m_Parameters.m_BiasEnabled ? 3: 2;
+    m_Data.ValidateInputsOutputs("NeonConvolution3dWorkload", numInputs, 1);
+
+    arm_compute::ITensor& input = PolymorphicDowncast<IAclTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
+    arm_compute::ITensor& weights = PolymorphicDowncast<IAclTensorHandle*>(m_Data.m_Inputs[1])->GetTensor();
+    arm_compute::ITensor* biasesPtr = nullptr;
+    if (m_Data.m_Parameters.m_BiasEnabled)
+    {
+        biasesPtr = &PolymorphicDowncast<IAclTensorHandle *>(m_Data.m_Inputs[2])->GetTensor();
+    }
+    arm_compute::ITensor& output = PolymorphicDowncast<IAclTensorHandle*>(m_Data.m_Outputs[0])->GetTensor();
+
+    arm_compute::DataLayout aclDataLayout = ConvertDataLayout(m_Data.m_Parameters.m_DataLayout);
+    input.info()->set_data_layout(aclDataLayout);
+    weights.info()->set_data_layout(aclDataLayout);
+    output.info()->set_data_layout(aclDataLayout);
+
+    const arm_compute::Conv3dInfo aclConv3DInfo = ComputeConv3DInfo(descriptor, isFastMathEnabled);
+
+    auto convolutionLayer = std::make_unique<arm_compute::NEConv3D>();
+    convolutionLayer->configure(&input,
+                                &weights,
+                                biasesPtr,
+                                &output,
+                                aclConv3DInfo);
+
+    // Add details for profiling output
+    WorkloadInfo detailsInfo;
+
+    detailsInfo.m_InputTensorInfos = info.m_InputTensorInfos;
+    detailsInfo.m_OutputTensorInfos = info.m_OutputTensorInfos;
+
+    // Report Profiling Details
+    ARMNN_REPORT_PROFILING_WORKLOAD_DESC("NeonConvolution3dWorkload_Construct",
+                                         descriptor.m_Parameters,
+                                         detailsInfo,
+                                         this->GetGuid());
+
+    m_ConvolutionLayer.reset(convolutionLayer.release());
+
+    ARMNN_ASSERT(m_ConvolutionLayer);
+
+    m_ConvolutionLayer->prepare();
+}
+
+void NeonConvolution3dWorkload::Execute() const
+{
+    ARMNN_SCOPED_PROFILING_EVENT_NEON_GUID("NeonConvolution3dWorkload_Execute", this->GetGuid());
+    m_ConvolutionLayer->run();
+}
+
+} //namespace armnn
diff --git a/src/backends/neon/workloads/NeonConvolution3dWorkload.hpp b/src/backends/neon/workloads/NeonConvolution3dWorkload.hpp
new file mode 100644
index 0000000000..b5175e963f
--- /dev/null
+++ b/src/backends/neon/workloads/NeonConvolution3dWorkload.hpp
@@ -0,0 +1,43 @@
+//
+// Copyright © 2021 Arm Ltd and Contributors. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#pragma once
+
+#include <backendsCommon/Workload.hpp>
+
+#include <arm_compute/runtime/IFunction.h>
+#include <arm_compute/runtime/Tensor.h>
+#include <arm_compute/runtime/MemoryManagerOnDemand.h>
+
+#include <memory>
+
+namespace armnn
+{
+
+arm_compute::Status NeonConvolution3dWorkloadValidate(const TensorInfo& input,
+                                                      const TensorInfo& output,
+                                                      const Convolution3dDescriptor& descriptor,
+                                                      const TensorInfo& weights,
+                                                      const Optional<TensorInfo>& biases,
+                                                      bool isFastMathEnabled = false,
+                                                      const ActivationDescriptor* activationDescriptor = nullptr);
+
+class NeonConvolution3dWorkload : public BaseWorkload<Convolution3dQueueDescriptor>
+{
+public:
+    using BaseWorkload<Convolution3dQueueDescriptor>::m_Data;
+
+    NeonConvolution3dWorkload(const Convolution3dQueueDescriptor& descriptor,
+                              const WorkloadInfo& info,
+                              std::shared_ptr<arm_compute::MemoryManagerOnDemand>& memoryManager,
+                              const bool isFastMathENabled = false);
+
+    void Execute() const override;
+
+private:
+    std::unique_ptr<arm_compute::IFunction> m_ConvolutionLayer;
+};
+
+} //namespace armnn
diff --git a/src/backends/neon/workloads/NeonWorkloads.hpp b/src/backends/neon/workloads/NeonWorkloads.hpp
index 4d51d186e9..a8134a130b 100644
--- a/src/backends/neon/workloads/NeonWorkloads.hpp
+++ b/src/backends/neon/workloads/NeonWorkloads.hpp
@@ -20,6 +20,7 @@
 #include "NeonConvertFp32ToBf16Workload.hpp"
 #include "NeonConvertFp32ToFp16Workload.hpp"
 #include "NeonConvolution2dWorkload.hpp"
+#include "NeonConvolution3dWorkload.hpp"
 #include "NeonDepthToSpaceWorkload.hpp"
 #include "NeonDepthwiseConvolutionWorkload.hpp"
 #include "NeonDequantizeWorkload.hpp"