IVGCVSW-3598 Add Neon backend support for Strided_Slice

Signed-off-by: FinnWilliamsArm <Finn.Williams@arm.com>
Change-Id: Ia9bf55ca93e36ad7fa052b70e477cd46078e0a30

10 files changed, 190 insertions, 1 deletions

diff --git a/src/backends/neon/NeonLayerSupport.cpp b/src/backends/neon/NeonLayerSupport.cpp
index bddee11f50..26abc16c46 100644
--- a/src/backends/neon/NeonLayerSupport.cpp
+++ b/src/backends/neon/NeonLayerSupport.cpp
@@ -45,6 +45,7 @@
 #include "workloads/NeonSpaceToDepthWorkload.hpp"
 #include "workloads/NeonSplitterWorkload.hpp"
 #include "workloads/NeonStackWorkload.hpp"
+#include "workloads/NeonStridedSliceWorkload.hpp"
 #include "workloads/NeonSubtractionWorkload.hpp"
 #endif
 
@@ -609,6 +610,18 @@ bool NeonLayerSupport::IsStackSupported(const std::vector<const TensorInfo*>& in
                                    descriptor);
 }
 
+bool NeonLayerSupport::IsStridedSliceSupported(const TensorInfo& input,
+                                               const TensorInfo& output,
+                                               const StridedSliceDescriptor& descriptor,
+                                               Optional<std::string&> reasonIfUnsupported) const
+{
+    FORWARD_WORKLOAD_VALIDATE_FUNC(NeonStridedSliceWorkloadValidate,
+                                   reasonIfUnsupported,
+                                   input,
+                                   output,
+                                   descriptor);
+}
+
 bool NeonLayerSupport::IsSubtractionSupported(const TensorInfo& input0,
                                               const TensorInfo& input1,
                                               const TensorInfo& output,
diff --git a/src/backends/neon/NeonLayerSupport.hpp b/src/backends/neon/NeonLayerSupport.hpp
index acaebc4c58..744683f77a 100644
--- a/src/backends/neon/NeonLayerSupport.hpp
+++ b/src/backends/neon/NeonLayerSupport.hpp
@@ -212,6 +212,11 @@ public:
                           const StackDescriptor& descriptor,
                           Optional<std::string&> reasonIfUnsupported = EmptyOptional()) const override;
 
+    bool IsStridedSliceSupported(const TensorInfo& input,
+                                 const TensorInfo& output,
+                                 const StridedSliceDescriptor& descriptor,
+                                 Optional<std::string&> reasonIfUnsupported = EmptyOptional()) const override;
+
     bool IsSubtractionSupported(const TensorInfo& input0,
                                 const TensorInfo& input1,
                                 const TensorInfo& output,
diff --git a/src/backends/neon/NeonWorkloadFactory.cpp b/src/backends/neon/NeonWorkloadFactory.cpp
index 77660c3b0a..36b176626d 100644
--- a/src/backends/neon/NeonWorkloadFactory.cpp
+++ b/src/backends/neon/NeonWorkloadFactory.cpp
@@ -371,7 +371,7 @@ std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateBatchToSpaceNd(const Batch
 std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateStridedSlice(const StridedSliceQueueDescriptor& descriptor,
                                                                    const WorkloadInfo& info) const
 {
-    return MakeWorkloadHelper<NullWorkload, NullWorkload>(descriptor, info);
+    return std::make_unique<NeonStridedSliceWorkload>(descriptor, info);
 }
 
 std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateMinimum(const MinimumQueueDescriptor& descriptor,
diff --git a/src/backends/neon/backend.mk b/src/backends/neon/backend.mk
index d5483b0c7d..7fc0c1c57f 100644
--- a/src/backends/neon/backend.mk
+++ b/src/backends/neon/backend.mk
@@ -53,6 +53,7 @@ BACKEND_SOURCES := \
         workloads/NeonSpaceToDepthWorkload.cpp \
         workloads/NeonSplitterWorkload.cpp \
         workloads/NeonStackWorkload.cpp \
+        workloads/NeonStridedSliceWorkload.cpp \
         workloads/NeonSubtractionWorkload.cpp
 
 else
diff --git a/src/backends/neon/test/NeonLayerTests.cpp b/src/backends/neon/test/NeonLayerTests.cpp
index ed99461b31..8ef146280f 100644
--- a/src/backends/neon/test/NeonLayerTests.cpp
+++ b/src/backends/neon/test/NeonLayerTests.cpp
@@ -688,6 +688,27 @@ ARMNN_AUTO_TEST_CASE(ResizeNearestNeighborMagUint8Nhwc,
                      ResizeNearestNeighborMagTest<armnn::DataType::QuantisedAsymm8>,
                      armnn::DataLayout::NHWC, 0.1f, 50, 0.1f, 50)
 
+// Strided Slice
+ARMNN_AUTO_TEST_CASE(StridedSlice4DFloat32, StridedSlice4DFloat32Test)
+ARMNN_AUTO_TEST_CASE(StridedSlice4DReverseFloat32, StridedSlice4DReverseFloat32Test)
+ARMNN_AUTO_TEST_CASE(StridedSliceSimpleStrideFloat32, StridedSliceSimpleStrideFloat32Test)
+ARMNN_AUTO_TEST_CASE(StridedSliceSimpleRangeMaskFloat32, StridedSliceSimpleRangeMaskFloat32Test)
+ARMNN_AUTO_TEST_CASE(StridedSliceShrinkAxisMaskFloat32, StridedSliceShrinkAxisMaskFloat32Test)
+ARMNN_AUTO_TEST_CASE(StridedSlice3DFloat32, StridedSlice3DFloat32Test)
+ARMNN_AUTO_TEST_CASE(StridedSlice3DReverseFloat32, StridedSlice3DReverseFloat32Test)
+ARMNN_AUTO_TEST_CASE(StridedSlice2DFloat32, StridedSlice2DFloat32Test)
+ARMNN_AUTO_TEST_CASE(StridedSlice2DReverseFloat32, StridedSlice2DReverseFloat32Test)
+
+ARMNN_AUTO_TEST_CASE(StridedSlice4DUint8, StridedSlice4DUint8Test)
+ARMNN_AUTO_TEST_CASE(StridedSlice4DReverseUint8, StridedSlice4DReverseUint8Test)
+ARMNN_AUTO_TEST_CASE(StridedSliceSimpleStrideUint8, StridedSliceSimpleStrideUint8Test)
+ARMNN_AUTO_TEST_CASE(StridedSliceSimpleRangeMaskUint8, StridedSliceSimpleRangeMaskUint8Test)
+ARMNN_AUTO_TEST_CASE(StridedSliceShrinkAxisMaskUint8, StridedSliceShrinkAxisMaskUint8Test)
+ARMNN_AUTO_TEST_CASE(StridedSlice3DUint8, StridedSlice3DUint8Test)
+ARMNN_AUTO_TEST_CASE(StridedSlice3DReverseUint8, StridedSlice3DReverseUint8Test)
+ARMNN_AUTO_TEST_CASE(StridedSlice2DUint8, StridedSlice2DUint8Test)
+ARMNN_AUTO_TEST_CASE(StridedSlice2DReverseUint8, StridedSlice2DReverseUint8Test)
+
 // Quantize
 ARMNN_AUTO_TEST_CASE(QuantizeSimpleUint8, QuantizeSimpleUint8Test)
 ARMNN_AUTO_TEST_CASE(QuantizeClampUint8, QuantizeClampUint8Test)
diff --git a/src/backends/neon/workloads/CMakeLists.txt b/src/backends/neon/workloads/CMakeLists.txt
index 34fe0723af..1115df1c20 100644
--- a/src/backends/neon/workloads/CMakeLists.txt
+++ b/src/backends/neon/workloads/CMakeLists.txt
@@ -72,6 +72,8 @@ list(APPEND armnnNeonBackendWorkloads_sources
     NeonSplitterWorkload.hpp
     NeonStackWorkload.cpp
     NeonStackWorkload.hpp
+    NeonStridedSliceWorkload.cpp
+    NeonStridedSliceWorkload.hpp
     NeonSubtractionWorkload.cpp
     NeonSubtractionWorkload.hpp
     NeonWorkloads.hpp
diff --git a/src/backends/neon/workloads/NeonStridedSliceWorkload.cpp b/src/backends/neon/workloads/NeonStridedSliceWorkload.cpp
new file mode 100644
index 0000000000..9c24728004
--- /dev/null
+++ b/src/backends/neon/workloads/NeonStridedSliceWorkload.cpp
@@ -0,0 +1,90 @@
+//
+// Copyright © 2017 Arm Ltd. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#include "NeonStridedSliceWorkload.hpp"
+
+#include "NeonWorkloadUtils.hpp"
+#include <neon/NeonTensorHandle.hpp>
+#include <aclCommon/ArmComputeUtils.hpp>
+#include <aclCommon/ArmComputeTensorUtils.hpp>
+
+
+namespace armnn
+{
+
+arm_compute::Status NeonStridedSliceWorkloadValidate(const TensorInfo& input,
+                                                     const TensorInfo& output,
+                                                     const StridedSliceDescriptor& descriptor)
+{
+    const arm_compute::TensorInfo aclInput = armcomputetensorutils::BuildArmComputeTensorInfo(input);
+    const arm_compute::TensorInfo aclOutput = armcomputetensorutils::BuildArmComputeTensorInfo(output);
+
+    arm_compute::Coordinates starts;
+    arm_compute::Coordinates ends;
+    arm_compute::Coordinates strides;
+
+    std::tie(starts, ends, strides) = SetNeonStridedSliceData(descriptor.m_Begin,
+                                                              descriptor.m_End,
+                                                              descriptor.m_Stride);
+
+    int32_t begin_mask = descriptor.m_BeginMask;
+    int32_t end_mask = descriptor.m_EndMask;
+    int32_t shrink_axis_mask = descriptor.m_ShrinkAxisMask;
+
+    return arm_compute::NEStridedSlice::validate(&aclInput,
+                                                 &aclOutput,
+                                                 starts,
+                                                 ends,
+                                                 strides,
+                                                 begin_mask,
+                                                 end_mask,
+                                                 shrink_axis_mask);
+}
+
+NeonStridedSliceWorkload::NeonStridedSliceWorkload(const StridedSliceQueueDescriptor& descriptor,
+                                                   const WorkloadInfo& info)
+        : BaseWorkload<StridedSliceQueueDescriptor>(descriptor, info)
+{
+    m_Data.ValidateInputsOutputs("NeonStridedSliceWorkload", 1, 1);
+
+    arm_compute::ITensor& input = boost::polymorphic_downcast<IAclTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
+    arm_compute::ITensor& output = boost::polymorphic_downcast<IAclTensorHandle*>(m_Data.m_Outputs[0])->GetTensor();
+
+    arm_compute::Coordinates starts;
+    arm_compute::Coordinates ends;
+    arm_compute::Coordinates strides;
+
+    std::tie(starts, ends, strides) = SetNeonStridedSliceData(m_Data.m_Parameters.m_Begin,
+                                                              m_Data.m_Parameters.m_End,
+                                                              m_Data.m_Parameters.m_Stride);
+
+    int32_t begin_mask = m_Data.m_Parameters.m_BeginMask;
+    int32_t end_mask = m_Data.m_Parameters.m_EndMask;
+    int32_t shrink_axis_mask = m_Data.m_Parameters.m_ShrinkAxisMask;
+
+    arm_compute::DataLayout aclDataLayout = ConvertDataLayout(m_Data.m_Parameters.m_DataLayout);
+    input.info()->set_data_layout(aclDataLayout);
+    output.info()->set_data_layout(aclDataLayout);
+
+    auto layer = std::make_unique<arm_compute::NEStridedSlice>();
+
+    layer->configure(&input,
+                     &output,
+                     starts,
+                     ends,
+                     strides,
+                     begin_mask,
+                     end_mask,
+                     shrink_axis_mask);
+    m_Layer.reset(layer.release());
+}
+
+void NeonStridedSliceWorkload::Execute() const
+{
+    ARMNN_SCOPED_PROFILING_EVENT_NEON("NeonStridedSliceWorkload_Execute");
+    m_Layer->run();
+}
+
+} //namespace armnn
\ No newline at end of file
diff --git a/src/backends/neon/workloads/NeonStridedSliceWorkload.hpp b/src/backends/neon/workloads/NeonStridedSliceWorkload.hpp
new file mode 100644
index 0000000000..458db3f7b4
--- /dev/null
+++ b/src/backends/neon/workloads/NeonStridedSliceWorkload.hpp
@@ -0,0 +1,34 @@
+//
+// Copyright © 2017 Arm Ltd. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#pragma once
+
+#include <backendsCommon/Workload.hpp>
+
+#include <arm_compute/core/Error.h>
+#include <arm_compute/runtime/IFunction.h>
+#include <arm_compute/runtime/NEON/functions/NEStridedSlice.h>
+
+#include <memory>
+
+
+namespace armnn
+{
+
+arm_compute::Status NeonStridedSliceWorkloadValidate(const TensorInfo& input,
+                                                     const TensorInfo& output,
+                                                     const StridedSliceDescriptor& descriptor);
+
+class NeonStridedSliceWorkload : public BaseWorkload<StridedSliceQueueDescriptor>
+{
+public:
+    NeonStridedSliceWorkload(const StridedSliceQueueDescriptor& descriptor, const WorkloadInfo& info);
+    virtual void Execute() const override;
+
+private:
+    mutable std::unique_ptr<arm_compute::NEStridedSlice> m_Layer;
+};
+
+} //namespace armnn
\ No newline at end of file
diff --git a/src/backends/neon/workloads/NeonWorkloadUtils.hpp b/src/backends/neon/workloads/NeonWorkloadUtils.hpp
index 22ffece6a2..8cf97d3909 100644
--- a/src/backends/neon/workloads/NeonWorkloadUtils.hpp
+++ b/src/backends/neon/workloads/NeonWorkloadUtils.hpp
@@ -54,4 +54,26 @@ inline void InitializeArmComputeTensorData(arm_compute::Tensor& tensor,
     }
 };
 
+inline auto SetNeonStridedSliceData(const std::vector<int>& m_begin,
+                                    const std::vector<int>& m_end,
+                                    const std::vector<int>& m_stride)
+{
+    arm_compute::Coordinates starts;
+    arm_compute::Coordinates ends;
+    arm_compute::Coordinates strides;
+
+    unsigned int num_dims = static_cast<unsigned int>(m_begin.size());
+
+    for (unsigned int i = 0; i < num_dims; i++)
+    {
+        unsigned int revertedIndex = num_dims - i - 1;
+
+        starts.set(i, static_cast<int>(m_begin[revertedIndex]));
+        ends.set(i, static_cast<int>(m_end[revertedIndex]));
+        strides.set(i, static_cast<int>(m_stride[revertedIndex]));
+    }
+
+    return std::make_tuple(starts, ends, strides);
+}
+
 } //namespace armnn
diff --git a/src/backends/neon/workloads/NeonWorkloads.hpp b/src/backends/neon/workloads/NeonWorkloads.hpp
index 8fc684e3e9..c201f6c3cd 100644
--- a/src/backends/neon/workloads/NeonWorkloads.hpp
+++ b/src/backends/neon/workloads/NeonWorkloads.hpp
@@ -37,4 +37,5 @@
 #include "NeonSpaceToDepthWorkload.hpp"
 #include "NeonSplitterWorkload.hpp"
 #include "NeonStackWorkload.hpp"
+#include "NeonStridedSliceWorkload.hpp"
 #include "NeonSubtractionWorkload.hpp"