IVGCVSW-3729 Added neon slice workload and supporting neon layer tests

* Support added for ACL neon slice workload
* Utility function created to translate ArmNN slice layer params to ACL neon slice layer equivalent
* Neon slice layer tests added as per SliceTestImpl.hpp

Signed-off-by: josh minor <josh.minor@arm.com>
Change-Id: Id583465311879af139e8e977f16ed2280c937ac7

11 files changed, 157 insertions, 0 deletions

diff --git a/src/backends/neon/NeonLayerSupport.cpp b/src/backends/neon/NeonLayerSupport.cpp
index 0eeb948d3f..5410ef42df 100644
--- a/src/backends/neon/NeonLayerSupport.cpp
+++ b/src/backends/neon/NeonLayerSupport.cpp
@@ -46,6 +46,7 @@
 #include "workloads/NeonQuantizedLstmWorkload.hpp"
 #include "workloads/NeonResizeWorkload.hpp"
 #include "workloads/NeonRsqrtWorkload.hpp"
+#include "workloads/NeonSliceWorkload.hpp"
 #include "workloads/NeonSoftmaxBaseWorkload.hpp"
 #include "workloads/NeonSpaceToDepthWorkload.hpp"
 #include "workloads/NeonSplitterWorkload.hpp"
@@ -619,6 +620,18 @@ bool NeonLayerSupport::IsRsqrtSupported(const TensorInfo& input,
     FORWARD_WORKLOAD_VALIDATE_FUNC(NeonRsqrtWorkloadValidate, reasonIfUnsupported, input, output);
 }
 
+bool NeonLayerSupport::IsSliceSupported(const TensorInfo& input,
+                                        const TensorInfo& output,
+                                        const SliceDescriptor& descriptor,
+                                        Optional<std::string&> reasonIfUnsupported) const
+{
+    FORWARD_WORKLOAD_VALIDATE_FUNC(NeonSliceWorkloadValidate,
+                                   reasonIfUnsupported,
+                                   input,
+                                   output,
+                                   descriptor);
+}
+
 bool NeonLayerSupport::IsSoftmaxSupported(const TensorInfo& input,
                                           const TensorInfo& output,
                                           const SoftmaxDescriptor& descriptor,
diff --git a/src/backends/neon/NeonLayerSupport.hpp b/src/backends/neon/NeonLayerSupport.hpp
index 9c1ca2a729..5d4fbad97f 100644
--- a/src/backends/neon/NeonLayerSupport.hpp
+++ b/src/backends/neon/NeonLayerSupport.hpp
@@ -218,6 +218,11 @@ public:
                           const TensorInfo& output,
                           Optional<std::string&> reasonIfUnsupported = EmptyOptional()) const override;
 
+    bool IsSliceSupported(const TensorInfo& input,
+                          const TensorInfo& output,
+                          const SliceDescriptor& descriptor,
+                          Optional<std::string&> reasonIfUnsupported = EmptyOptional()) const override;
+
     bool IsSoftmaxSupported(const TensorInfo& input,
                             const TensorInfo& output,
                             const SoftmaxDescriptor& descriptor,
diff --git a/src/backends/neon/NeonWorkloadFactory.cpp b/src/backends/neon/NeonWorkloadFactory.cpp
index dda1d7a132..dd11af4484 100644
--- a/src/backends/neon/NeonWorkloadFactory.cpp
+++ b/src/backends/neon/NeonWorkloadFactory.cpp
@@ -428,6 +428,12 @@ std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateRsqrt(const RsqrtQueueDesc
     return std::make_unique<NeonRsqrtWorkload>(descriptor, info);
 }
 
+std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateSlice(const SliceQueueDescriptor& descriptor,
+                                                            const WorkloadInfo& info) const
+{
+    return std::make_unique<NeonSliceWorkload>(descriptor, info);
+}
+
 std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateSoftmax(const SoftmaxQueueDescriptor& descriptor,
                                                               const WorkloadInfo& info) const
 {
diff --git a/src/backends/neon/NeonWorkloadFactory.hpp b/src/backends/neon/NeonWorkloadFactory.hpp
index f876092c45..89a495985e 100644
--- a/src/backends/neon/NeonWorkloadFactory.hpp
+++ b/src/backends/neon/NeonWorkloadFactory.hpp
@@ -188,6 +188,9 @@ public:
     std::unique_ptr<IWorkload> CreateRsqrt(const RsqrtQueueDescriptor& descriptor,
                                            const WorkloadInfo& info) const override;
 
+    std::unique_ptr<IWorkload> CreateSlice(const SliceQueueDescriptor& descriptor,
+                                           const WorkloadInfo& info) const override;
+
     std::unique_ptr<IWorkload> CreateSoftmax(const SoftmaxQueueDescriptor& descriptor,
                                              const WorkloadInfo& info) const override;
 
diff --git a/src/backends/neon/backend.mk b/src/backends/neon/backend.mk
index c66f5da6a5..23289345ea 100644
--- a/src/backends/neon/backend.mk
+++ b/src/backends/neon/backend.mk
@@ -54,6 +54,7 @@ BACKEND_SOURCES := \
         workloads/NeonReshapeWorkload.cpp \
         workloads/NeonResizeWorkload.cpp \
         workloads/NeonRsqrtWorkload.cpp \
+        workloads/NeonSliceWorkload.cpp \
         workloads/NeonSoftmaxBaseWorkload.cpp \
         workloads/NeonSoftmaxFloatWorkload.cpp \
         workloads/NeonSoftmaxUint8Workload.cpp \
diff --git a/src/backends/neon/test/NeonLayerTests.cpp b/src/backends/neon/test/NeonLayerTests.cpp
index ea9c813d34..cd9a55d3ee 100644
--- a/src/backends/neon/test/NeonLayerTests.cpp
+++ b/src/backends/neon/test/NeonLayerTests.cpp
@@ -756,6 +756,20 @@ ARMNN_AUTO_TEST_CASE(ResizeNearestNeighborMagUint8Nhwc,
                      ResizeNearestNeighborMagTest<DataType::QuantisedAsymm8>,
                      DataLayout::NHWC, 0.1f, 50, 0.1f, 50)
 
+// Slice
+ARMNN_AUTO_TEST_CASE(Slice4dFloat32, Slice4dFloat32Test)
+ARMNN_AUTO_TEST_CASE(Slice3dFloat32, Slice3dFloat32Test)
+ARMNN_AUTO_TEST_CASE(Slice2dFloat32, Slice2dFloat32Test)
+ARMNN_AUTO_TEST_CASE(Slice1dFloat32, Slice1dFloat32Test)
+ARMNN_AUTO_TEST_CASE(Slice4dUint8, Slice4dUint8Test)
+ARMNN_AUTO_TEST_CASE(Slice3dUint8, Slice3dUint8Test)
+ARMNN_AUTO_TEST_CASE(Slice2dUint8, Slice2dUint8Test)
+ARMNN_AUTO_TEST_CASE(Slice1dUint8, Slice1dUint8Test)
+ARMNN_AUTO_TEST_CASE(Slice4dInt16, Slice4dInt16Test)
+ARMNN_AUTO_TEST_CASE(Slice3dInt16, Slice3dInt16Test)
+ARMNN_AUTO_TEST_CASE(Slice2dInt16, Slice2dInt16Test)
+ARMNN_AUTO_TEST_CASE(Slice1dInt16, Slice1dInt16Test)
+
 // Strided Slice
 ARMNN_AUTO_TEST_CASE(StridedSlice4dFloat32, StridedSlice4dFloat32Test)
 ARMNN_AUTO_TEST_CASE(StridedSlice4dReverseFloat32, StridedSlice4dReverseFloat32Test)
diff --git a/src/backends/neon/workloads/CMakeLists.txt b/src/backends/neon/workloads/CMakeLists.txt
index 46e4928882..19ac9c9d00 100644
--- a/src/backends/neon/workloads/CMakeLists.txt
+++ b/src/backends/neon/workloads/CMakeLists.txt
@@ -70,6 +70,8 @@ list(APPEND armnnNeonBackendWorkloads_sources
     NeonResizeWorkload.hpp
     NeonRsqrtWorkload.cpp
     NeonRsqrtWorkload.hpp
+    NeonSliceWorkload.cpp
+    NeonSliceWorkload.hpp
     NeonSoftmaxBaseWorkload.cpp
     NeonSoftmaxBaseWorkload.hpp
     NeonSoftmaxFloatWorkload.cpp
diff --git a/src/backends/neon/workloads/NeonSliceWorkload.cpp b/src/backends/neon/workloads/NeonSliceWorkload.cpp
new file mode 100644
index 0000000000..171edc6c59
--- /dev/null
+++ b/src/backends/neon/workloads/NeonSliceWorkload.cpp
@@ -0,0 +1,57 @@
+//
+// Copyright © 2019 Arm Ltd. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#include "NeonSliceWorkload.hpp"
+
+#include "NeonWorkloadUtils.hpp"
+
+#include <aclCommon/ArmComputeTensorUtils.hpp>
+
+#include <neon/NeonTensorHandle.hpp>
+
+#include <arm_compute/core/Error.h>
+
+namespace armnn
+{
+
+arm_compute::Status NeonSliceWorkloadValidate(const TensorInfo& input,
+                                              const TensorInfo& output,
+                                              const SliceDescriptor& descriptor)
+{
+    const arm_compute::TensorInfo aclInputInfo = armcomputetensorutils::BuildArmComputeTensorInfo(input);
+    const arm_compute::TensorInfo aclOutputInfo = armcomputetensorutils::BuildArmComputeTensorInfo(output);
+
+    arm_compute::Coordinates starts;
+    arm_compute::Coordinates ends;
+
+    std::tie(starts, ends) = SetNeonSliceData(descriptor.m_Begin, descriptor.m_Size);
+
+    return arm_compute::NESlice::validate(&aclInputInfo, &aclOutputInfo, starts, ends);
+}
+
+NeonSliceWorkload::NeonSliceWorkload(const SliceQueueDescriptor& descriptor,
+                                     const WorkloadInfo& info)
+        : BaseWorkload<SliceQueueDescriptor>(descriptor, info)
+{
+    m_Data.ValidateInputsOutputs("NeonSliceWorkload", 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;
+
+    std::tie(starts, ends) = SetNeonSliceData(m_Data.m_Parameters.m_Begin, m_Data.m_Parameters.m_Size);
+
+    m_SliceFunction.configure(&input, &output, starts, ends);
+}
+
+void NeonSliceWorkload::Execute() const
+{
+    ARMNN_SCOPED_PROFILING_EVENT_NEON("NeonSliceWorkload_Execute");
+    m_SliceFunction.run();
+}
+
+} // namespace armnn
diff --git a/src/backends/neon/workloads/NeonSliceWorkload.hpp b/src/backends/neon/workloads/NeonSliceWorkload.hpp
new file mode 100644
index 0000000000..d1f32a9d7e
--- /dev/null
+++ b/src/backends/neon/workloads/NeonSliceWorkload.hpp
@@ -0,0 +1,31 @@
+//
+// Copyright © 2019 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/NESlice.h>
+
+namespace armnn
+{
+
+arm_compute::Status NeonSliceWorkloadValidate(const TensorInfo& input,
+                                              const TensorInfo& output,
+                                              const SliceDescriptor& descriptor);
+
+class NeonSliceWorkload : public BaseWorkload<SliceQueueDescriptor>
+{
+public:
+    NeonSliceWorkload(const SliceQueueDescriptor& descriptor, const WorkloadInfo& info);
+    virtual void Execute() const override;
+
+private:
+    mutable arm_compute::NESlice m_SliceFunction;
+};
+
+} // 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 8cf97d3909..f63946ec07 100644
--- a/src/backends/neon/workloads/NeonWorkloadUtils.hpp
+++ b/src/backends/neon/workloads/NeonWorkloadUtils.hpp
@@ -76,4 +76,28 @@ inline auto SetNeonStridedSliceData(const std::vector<int>& m_begin,
     return std::make_tuple(starts, ends, strides);
 }
 
+inline auto SetNeonSliceData(const std::vector<unsigned int>& m_begin,
+                             const std::vector<unsigned int>& m_size)
+{
+    // This function must translate the size vector given to an end vector
+    // expected by the ACL NESlice workload
+    arm_compute::Coordinates starts;
+    arm_compute::Coordinates ends;
+
+    unsigned int num_dims = static_cast<unsigned int>(m_begin.size());
+
+    // For strided slices, we have the relationship size = (end - begin) / stride
+    // For slice, we assume stride to be a vector of all ones, yielding the formula
+    // size = (end - begin) therefore we know end = size + begin
+    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_begin[revertedIndex] + m_size[revertedIndex]));
+    }
+
+    return std::make_tuple(starts, ends);
+}
+
 } //namespace armnn
diff --git a/src/backends/neon/workloads/NeonWorkloads.hpp b/src/backends/neon/workloads/NeonWorkloads.hpp
index 7d99d260a6..77d819e3ab 100644
--- a/src/backends/neon/workloads/NeonWorkloads.hpp
+++ b/src/backends/neon/workloads/NeonWorkloads.hpp
@@ -37,6 +37,7 @@
 #include "NeonReshapeWorkload.hpp"
 #include "NeonResizeWorkload.hpp"
 #include "NeonRsqrtWorkload.hpp"
+#include "NeonSliceWorkload.hpp"
 #include "NeonSoftmaxFloatWorkload.hpp"
 #include "NeonSoftmaxUint8Workload.hpp"
 #include "NeonSpaceToDepthWorkload.hpp"