IVGCVSW-3734 Add NEON workload for DepthToSpace

Signed-off-by: Aron Virginas-Tar <Aron.Virginas-Tar@arm.com>
Change-Id: I4fa8cba57f3a2277112c02062f4d2790089d1eb5

10 files changed, 155 insertions, 0 deletions

diff --git a/src/backends/neon/NeonLayerSupport.cpp b/src/backends/neon/NeonLayerSupport.cpp
index 0d4d4c3459..b713abaef3 100644
--- a/src/backends/neon/NeonLayerSupport.cpp
+++ b/src/backends/neon/NeonLayerSupport.cpp
@@ -23,6 +23,7 @@
 #include "workloads/NeonActivationWorkload.hpp"
 #include "workloads/NeonBatchNormalizationWorkload.hpp"
 #include "workloads/NeonConvolution2dWorkload.hpp"
+#include "workloads/NeonDepthToSpaceWorkload.hpp"
 #include "workloads/NeonDepthwiseConvolutionWorkload.hpp"
 #include "workloads/NeonDequantizeWorkload.hpp"
 #include "workloads/NeonGreaterWorkload.hpp"
@@ -249,6 +250,18 @@ bool NeonLayerSupport::IsConvolution2dSupported(const TensorInfo& input,
                                    biases);
 }
 
+bool NeonLayerSupport::IsDepthToSpaceSupported(const TensorInfo& input,
+                                               const TensorInfo& output,
+                                               const DepthToSpaceDescriptor& descriptor,
+                                               Optional<std::string&> reasonIfUnsupported) const
+{
+    FORWARD_WORKLOAD_VALIDATE_FUNC(NeonDepthToSpaceWorkloadValidate,
+                                   reasonIfUnsupported,
+                                   input,
+                                   output,
+                                   descriptor);
+}
+
 bool NeonLayerSupport::IsDepthwiseConvolutionSupported(const TensorInfo& input,
                                                        const TensorInfo& output,
                                                        const DepthwiseConvolution2dDescriptor& descriptor,
diff --git a/src/backends/neon/NeonLayerSupport.hpp b/src/backends/neon/NeonLayerSupport.hpp
index cc4798eea5..31f9e576d7 100644
--- a/src/backends/neon/NeonLayerSupport.hpp
+++ b/src/backends/neon/NeonLayerSupport.hpp
@@ -58,6 +58,12 @@ public:
                                   const Optional<TensorInfo>& biases,
                                   Optional<std::string&> reasonIfUnsupported = EmptyOptional()) const override;
 
+    bool IsDepthToSpaceSupported(const TensorInfo& input,
+                                 const TensorInfo& output,
+                                 const DepthToSpaceDescriptor& descriptor,
+                                 Optional<std::string&> reasonIfUnsupported = EmptyOptional()) const override;
+
+
     bool IsDepthwiseConvolutionSupported(const TensorInfo& input,
                                          const TensorInfo& output,
                                          const DepthwiseConvolution2dDescriptor& descriptor,
diff --git a/src/backends/neon/NeonWorkloadFactory.cpp b/src/backends/neon/NeonWorkloadFactory.cpp
index fb81008644..f0b738c350 100644
--- a/src/backends/neon/NeonWorkloadFactory.cpp
+++ b/src/backends/neon/NeonWorkloadFactory.cpp
@@ -170,6 +170,12 @@ std::unique_ptr<armnn::IWorkload> NeonWorkloadFactory::CreateConvolution2d(
                                                        m_MemoryManager->GetIntraLayerManager());
 }
 
+std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateDepthToSpace(const DepthToSpaceQueueDescriptor& descriptor,
+                                                                   const WorkloadInfo& info) const
+{
+    return std::make_unique<NeonDepthToSpaceWorkload>(descriptor, info);
+}
+
 std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateDepthwiseConvolution2d(
     const DepthwiseConvolution2dQueueDescriptor& descriptor, const WorkloadInfo& info) const
 {
diff --git a/src/backends/neon/NeonWorkloadFactory.hpp b/src/backends/neon/NeonWorkloadFactory.hpp
index 0bee365d8a..4bdbc8e859 100644
--- a/src/backends/neon/NeonWorkloadFactory.hpp
+++ b/src/backends/neon/NeonWorkloadFactory.hpp
@@ -76,6 +76,9 @@ public:
     std::unique_ptr<IWorkload> CreateConvolution2d(const Convolution2dQueueDescriptor& descriptor,
                                                    const WorkloadInfo& info) const override;
 
+    std::unique_ptr<IWorkload> CreateDepthToSpace(const DepthToSpaceQueueDescriptor& descriptor,
+                                                  const WorkloadInfo& info) const override;
+
     std::unique_ptr<IWorkload> CreateDepthwiseConvolution2d(const DepthwiseConvolution2dQueueDescriptor& descriptor,
                                                             const WorkloadInfo& info) const override;
 
diff --git a/src/backends/neon/backend.mk b/src/backends/neon/backend.mk
index 5eae9e5137..3e46387ae4 100644
--- a/src/backends/neon/backend.mk
+++ b/src/backends/neon/backend.mk
@@ -30,6 +30,7 @@ BACKEND_SOURCES := \
         workloads/NeonConvertFp16ToFp32Workload.cpp \
         workloads/NeonConvertFp32ToFp16Workload.cpp \
         workloads/NeonConvolution2dWorkload.cpp \
+        workloads/NeonDepthToSpaceWorkload.cpp \
         workloads/NeonDepthwiseConvolutionWorkload.cpp \
         workloads/NeonDequantizeWorkload.cpp \
         workloads/NeonFloorFloatWorkload.cpp \
diff --git a/src/backends/neon/test/NeonLayerTests.cpp b/src/backends/neon/test/NeonLayerTests.cpp
index dbd2d329fc..745d2ceece 100644
--- a/src/backends/neon/test/NeonLayerTests.cpp
+++ b/src/backends/neon/test/NeonLayerTests.cpp
@@ -110,6 +110,37 @@ ARMNN_AUTO_TEST_CASE(DepthwiseConvolution2dMult2,
                      false,
                      armnn::DataLayout::NCHW)
 
+// DepthToSpace
+ARMNN_AUTO_TEST_CASE(DepthToSpaceNchwFloat32_1, DepthToSpaceTest1<DataType::Float32>, DataLayout::NCHW);
+ARMNN_AUTO_TEST_CASE(DepthToSpaceNchwFloat32_2, DepthToSpaceTest2<DataType::Float32>, DataLayout::NCHW);
+ARMNN_AUTO_TEST_CASE(DepthToSpaceNchwFloat32_3, DepthToSpaceTest3<DataType::Float32>, DataLayout::NCHW);
+ARMNN_AUTO_TEST_CASE(DepthToSpaceNchwFloat32_4, DepthToSpaceTest4<DataType::Float32>, DataLayout::NCHW);
+
+ARMNN_AUTO_TEST_CASE(DepthToSpaceNchwUint8_1, DepthToSpaceTest1<DataType::QuantisedAsymm8>, DataLayout::NCHW);
+ARMNN_AUTO_TEST_CASE(DepthToSpaceNchwUint8_2, DepthToSpaceTest2<DataType::QuantisedAsymm8>, DataLayout::NCHW);
+ARMNN_AUTO_TEST_CASE(DepthToSpaceNchwUint8_3, DepthToSpaceTest3<DataType::QuantisedAsymm8>, DataLayout::NCHW);
+ARMNN_AUTO_TEST_CASE(DepthToSpaceNchwUint8_4, DepthToSpaceTest4<DataType::QuantisedAsymm8>, DataLayout::NCHW);
+
+ARMNN_AUTO_TEST_CASE(DepthToSpaceNchwInt16_1, DepthToSpaceTest1<DataType::QuantisedSymm16>, DataLayout::NCHW);
+ARMNN_AUTO_TEST_CASE(DepthToSpaceNchwInt16_2, DepthToSpaceTest2<DataType::QuantisedSymm16>, DataLayout::NCHW);
+ARMNN_AUTO_TEST_CASE(DepthToSpaceNchwInt16_3, DepthToSpaceTest3<DataType::QuantisedSymm16>, DataLayout::NCHW);
+ARMNN_AUTO_TEST_CASE(DepthToSpaceNchwInt16_4, DepthToSpaceTest4<DataType::QuantisedSymm16>, DataLayout::NCHW);
+
+ARMNN_AUTO_TEST_CASE(DepthToSpaceNhwcFloat32_1, DepthToSpaceTest1<DataType::Float32>, DataLayout::NHWC);
+ARMNN_AUTO_TEST_CASE(DepthToSpaceNhwcFloat32_2, DepthToSpaceTest2<DataType::Float32>, DataLayout::NHWC);
+ARMNN_AUTO_TEST_CASE(DepthToSpaceNhwcFloat32_3, DepthToSpaceTest3<DataType::Float32>, DataLayout::NHWC);
+ARMNN_AUTO_TEST_CASE(DepthToSpaceNhwcFloat32_4, DepthToSpaceTest4<DataType::Float32>, DataLayout::NHWC);
+
+ARMNN_AUTO_TEST_CASE(DepthToSpaceNhwcUint8_1, DepthToSpaceTest1<DataType::QuantisedAsymm8>, DataLayout::NHWC);
+ARMNN_AUTO_TEST_CASE(DepthToSpaceNhwcUint8_2, DepthToSpaceTest2<DataType::QuantisedAsymm8>, DataLayout::NHWC);
+ARMNN_AUTO_TEST_CASE(DepthToSpaceNhwcUint8_3, DepthToSpaceTest3<DataType::QuantisedAsymm8>, DataLayout::NHWC);
+ARMNN_AUTO_TEST_CASE(DepthToSpaceNhwcUint8_4, DepthToSpaceTest4<DataType::QuantisedAsymm8>, DataLayout::NHWC);
+
+ARMNN_AUTO_TEST_CASE(DepthToSpaceNhwcInt16_1, DepthToSpaceTest1<DataType::QuantisedSymm16>, DataLayout::NHWC);
+ARMNN_AUTO_TEST_CASE(DepthToSpaceNhwcInt16_2, DepthToSpaceTest2<DataType::QuantisedSymm16>, DataLayout::NHWC);
+ARMNN_AUTO_TEST_CASE(DepthToSpaceNhwcInt16_3, DepthToSpaceTest3<DataType::QuantisedSymm16>, DataLayout::NHWC);
+ARMNN_AUTO_TEST_CASE(DepthToSpaceNhwcInt16_4, DepthToSpaceTest4<DataType::QuantisedSymm16>, DataLayout::NHWC);
+
 // Depthwise Convolution
 ARMNN_AUTO_TEST_CASE(DepthwiseConvolution2dDepthMul1,
                      DepthwiseConvolution2dDepthMul1Test, true, DataLayout::NCHW)
diff --git a/src/backends/neon/workloads/CMakeLists.txt b/src/backends/neon/workloads/CMakeLists.txt
index 90043911cf..42ac641369 100644
--- a/src/backends/neon/workloads/CMakeLists.txt
+++ b/src/backends/neon/workloads/CMakeLists.txt
@@ -22,6 +22,8 @@ list(APPEND armnnNeonBackendWorkloads_sources
     NeonConvertFp32ToFp16Workload.hpp
     NeonConvolution2dWorkload.cpp
     NeonConvolution2dWorkload.hpp
+    NeonDepthToSpaceWorkload.cpp
+    NeonDepthToSpaceWorkload.hpp
     NeonDepthwiseConvolutionWorkload.cpp
     NeonDepthwiseConvolutionWorkload.hpp
     NeonDequantizeWorkload.cpp
diff --git a/src/backends/neon/workloads/NeonDepthToSpaceWorkload.cpp b/src/backends/neon/workloads/NeonDepthToSpaceWorkload.cpp
new file mode 100644
index 0000000000..b30dfcd80e
--- /dev/null
+++ b/src/backends/neon/workloads/NeonDepthToSpaceWorkload.cpp
@@ -0,0 +1,61 @@
+//
+// Copyright © 2019 Arm Ltd. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#include "NeonDepthToSpaceWorkload.hpp"
+
+#include "NeonWorkloadUtils.hpp"
+
+#include <aclCommon/ArmComputeTensorUtils.hpp>
+
+#include <boost/numeric/conversion/cast.hpp>
+#include <boost/polymorphic_pointer_cast.hpp>
+
+namespace armnn
+{
+
+using namespace armcomputetensorutils;
+
+arm_compute::Status NeonDepthToSpaceWorkloadValidate(const TensorInfo& input,
+                                                     const TensorInfo& output,
+                                                     const DepthToSpaceDescriptor& descriptor)
+{
+    DataLayout dataLayout = descriptor.m_DataLayout;
+    const arm_compute::TensorInfo aclInput = BuildArmComputeTensorInfo(input, dataLayout);
+    const arm_compute::TensorInfo aclOutput = BuildArmComputeTensorInfo(output, dataLayout);
+
+    int32_t blockSize = boost::numeric_cast<int32_t>(descriptor.m_BlockSize);
+
+    return arm_compute::NEDepthToSpaceLayer::validate(&aclInput, &aclOutput, blockSize);
+}
+
+NeonDepthToSpaceWorkload::NeonDepthToSpaceWorkload(const DepthToSpaceQueueDescriptor& desc,
+                                                   const WorkloadInfo& info)
+    : BaseWorkload<DepthToSpaceQueueDescriptor>(desc, info)
+{
+    m_Data.ValidateInputsOutputs("NeonDepthToSpaceWorkload", 1, 1);
+
+    arm_compute::DataLayout aclDataLayout = ConvertDataLayout(m_Data.m_Parameters.m_DataLayout);
+
+    arm_compute::ITensor& input =
+        boost::polymorphic_pointer_downcast<IAclTensorHandle>(m_Data.m_Inputs[0])->GetTensor();
+    input.info()->set_data_layout(aclDataLayout);
+
+    int32_t blockSize = boost::numeric_cast<int32_t>(desc.m_Parameters.m_BlockSize);
+
+    arm_compute::ITensor& output =
+        boost::polymorphic_pointer_downcast<IAclTensorHandle>(m_Data.m_Outputs[0])->GetTensor();
+    output.info()->set_data_layout(aclDataLayout);
+
+    m_Layer.configure(&input, &output, blockSize);
+    m_Layer.prepare();
+}
+
+void NeonDepthToSpaceWorkload::Execute() const
+{
+    ARMNN_SCOPED_PROFILING_EVENT_NEON("NeonDepthToSpaceWorkload_Execute");
+    m_Layer.run();
+}
+
+} // namespace armnn
diff --git a/src/backends/neon/workloads/NeonDepthToSpaceWorkload.hpp b/src/backends/neon/workloads/NeonDepthToSpaceWorkload.hpp
new file mode 100644
index 0000000000..95b5e44e4d
--- /dev/null
+++ b/src/backends/neon/workloads/NeonDepthToSpaceWorkload.hpp
@@ -0,0 +1,31 @@
+//
+// Copyright © 2019 Arm Ltd. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#pragma once
+
+#include <armnn/Descriptors.hpp>
+
+#include <backendsCommon/Workload.hpp>
+
+#include <arm_compute/runtime/NEON/functions/NEDepthToSpaceLayer.h>
+
+namespace armnn
+{
+
+arm_compute::Status NeonDepthToSpaceWorkloadValidate(const TensorInfo& input,
+                                                     const TensorInfo& output,
+                                                     const DepthToSpaceDescriptor& descriptor);
+
+class NeonDepthToSpaceWorkload : public BaseWorkload<DepthToSpaceQueueDescriptor>
+{
+public:
+    NeonDepthToSpaceWorkload(const DepthToSpaceQueueDescriptor& descriptor, const WorkloadInfo& info);
+    void Execute() const override;
+
+private:
+    mutable arm_compute::NEDepthToSpaceLayer m_Layer;
+};
+
+} // namespace armnn
diff --git a/src/backends/neon/workloads/NeonWorkloads.hpp b/src/backends/neon/workloads/NeonWorkloads.hpp
index 4f30f33d1b..9d35ed42fe 100644
--- a/src/backends/neon/workloads/NeonWorkloads.hpp
+++ b/src/backends/neon/workloads/NeonWorkloads.hpp
@@ -12,6 +12,7 @@
 #include "NeonConvertFp16ToFp32Workload.hpp"
 #include "NeonConvertFp32ToFp16Workload.hpp"
 #include "NeonConvolution2dWorkload.hpp"
+#include "NeonDepthToSpaceWorkload.hpp"
 #include "NeonDepthwiseConvolutionWorkload.hpp"
 #include "NeonDequantizeWorkload.hpp"
 #include "NeonFloorFloatWorkload.hpp"