COMPMID-1769: Add support for NEON StridedSlice,Split,Slice,Unstack

Change-Id: I7d3c5e6858fed090410720f76947327e39bc72f8
Reviewed-on: https://review.mlplatform.org/450
Reviewed-by: Giuseppe Rossini <giuseppe.rossini@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michalis Spyrou <michalis.spyrou@arm.com>
Reviewed-by: VidhyaSudhan Loganathan <vidhyasudhan.loganathan@arm.com>

5 files changed, 554 insertions, 0 deletions

diff --git a/src/core/NEON/kernels/NEStridedSliceKernel.cpp b/src/core/NEON/kernels/NEStridedSliceKernel.cpp
new file mode 100644
index 0000000000..bb5756de93
--- /dev/null
+++ b/src/core/NEON/kernels/NEStridedSliceKernel.cpp
@@ -0,0 +1,187 @@
+/*
+ * Copyright (c) 2018-2019 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "arm_compute/core/NEON/kernels/NEStridedSliceKernel.h"
+
+#include "arm_compute/core/CPP/Validate.h"
+#include "arm_compute/core/IAccessWindow.h"
+#include "arm_compute/core/ITensor.h"
+#include "arm_compute/core/TensorInfo.h"
+#include "arm_compute/core/Window.h"
+
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/utils/helpers/bit_ops.h"
+#include "arm_compute/core/utils/helpers/tensor_transform.h"
+#include "arm_compute/core/utils/misc/ShapeCalculator.h"
+
+namespace arm_compute
+{
+namespace
+{
+Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output,
+                          const Coordinates &starts, const Coordinates &ends, const BiStrides &strides,
+                          int32_t begin_mask, int32_t end_mask, int32_t shrink_axis_mask)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
+    ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(input);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1,
+                                                         DataType::U8, DataType::S8, DataType::QASYMM8,
+                                                         DataType::U16, DataType::S16,
+                                                         DataType::U32, DataType::S32,
+                                                         DataType::F16, DataType::F32);
+
+    ARM_COMPUTE_RETURN_ERROR_ON(input->tensor_shape().num_dimensions() > 4);
+    ARM_COMPUTE_RETURN_ERROR_ON(starts.num_dimensions() > input->num_dimensions());
+    ARM_COMPUTE_RETURN_ERROR_ON(ends.num_dimensions() > input->num_dimensions());
+    ARM_COMPUTE_RETURN_ERROR_ON(strides.num_dimensions() > input->num_dimensions());
+    ARM_COMPUTE_RETURN_ERROR_ON(std::any_of(strides.cbegin(), strides.cbegin() + strides.num_dimensions(), [](int i)
+    {
+        return i == 0;
+    }));
+
+    // Get expected output shape
+    const TensorShape exp_output_shape = arm_compute::misc::shape_calculator::compute_strided_slice_shape(*input,
+                                                                                                          starts, ends, strides,
+                                                                                                          begin_mask, end_mask, shrink_axis_mask);
+    ARM_COMPUTE_RETURN_ERROR_ON(exp_output_shape.total_size() == 0);
+
+    // Checks output if configured
+    if(output->total_size() != 0)
+    {
+        const TensorInfo exp_output_info = output->clone()->set_tensor_shape(exp_output_shape);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(output, &exp_output_info);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
+    }
+
+    return Status{};
+}
+
+std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *output,
+                                                        const Coordinates &starts, const Coordinates &ends, const BiStrides &strides,
+                                                        int32_t begin_mask, int32_t end_mask, int32_t shrink_axis_mask)
+{
+    // Output tensor auto initialization if not yet initialized
+    const TensorShape output_shape = arm_compute::misc::shape_calculator::compute_strided_slice_shape(*input,
+                                                                                                      starts, ends, strides,
+                                                                                                      begin_mask, end_mask, shrink_axis_mask);
+    auto_init_if_empty(*output, input->clone()->set_tensor_shape(output_shape));
+
+    // Create window
+    const unsigned int num_elems_processed_per_iteration = 1;
+
+    Window win = calculate_max_window(*output, Steps(num_elems_processed_per_iteration));
+    output->set_valid_region(ValidRegion(Coordinates(), output->tensor_shape()));
+
+    return std::make_pair(Status{}, win);
+}
+
+void strided_slice_generic(const ITensor *input, ITensor *output,
+                           const Coordinates &starts, const BiStrides &strides, int32_t shrink_axis_mask,
+                           const Window &window)
+{
+    Iterator     output_it(output, window);
+    const size_t width_size = input->info()->element_size();
+
+    const bool is_shrink_w = arm_compute::helpers::bit_ops::is_bit_set(shrink_axis_mask, 0);
+    const bool is_shrink_h = arm_compute::helpers::bit_ops::is_bit_set(shrink_axis_mask, 1);
+    const bool is_shrink_c = arm_compute::helpers::bit_ops::is_bit_set(shrink_axis_mask, 2);
+    const bool is_shrink_n = arm_compute::helpers::bit_ops::is_bit_set(shrink_axis_mask, 3);
+
+    unsigned int index = 0;
+    const int    idx_w = is_shrink_w ? 0 : index++;
+    const int    idx_h = is_shrink_h ? 0 : index++;
+    const int    idx_c = is_shrink_c ? 0 : index++;
+    const int    idx_n = is_shrink_n ? 0 : index;
+
+    BiStrides shrinked_strides;
+    shrinked_strides.set(0, is_shrink_w ? 0 : strides[0]);
+    shrinked_strides.set(1, is_shrink_h ? 0 : strides[1]);
+    shrinked_strides.set(2, is_shrink_c ? 0 : strides[2]);
+    shrinked_strides.set(3, is_shrink_n ? 0 : strides[3]);
+
+    execute_window_loop(window, [&](const Coordinates & id)
+    {
+        const int w_coord = starts[0] + (id[idx_w] * shrinked_strides[0]);
+        const int h_coord = starts[1] + (id[idx_h] * shrinked_strides[1]);
+        const int c_coord = starts[2] + (id[idx_c] * shrinked_strides[2]);
+        const int n_coord = starts[3] + (id[idx_n] * shrinked_strides[3]);
+
+        Coordinates in_coords(w_coord, h_coord, c_coord, n_coord);
+        std::copy_n(input->ptr_to_element(in_coords), width_size, output_it.ptr());
+    },
+    output_it);
+}
+} // namespace
+
+NEStridedSliceKernel::NEStridedSliceKernel()
+    : _input(nullptr), _output(nullptr), _starts_abs(), _final_strides(), _shrink_mask()
+{
+}
+
+void NEStridedSliceKernel::configure(const ITensor *input, ITensor *output,
+                                     const Coordinates &starts, const Coordinates &ends, const BiStrides &strides,
+                                     int32_t begin_mask, int32_t end_mask, int32_t shrink_axis_mask)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), starts, ends, strides, begin_mask, end_mask, shrink_axis_mask));
+
+    _input       = input;
+    _output      = output;
+    _shrink_mask = shrink_axis_mask;
+
+    const TensorShape &input_shape = input->info()->tensor_shape();
+
+    Coordinates ends_abs;
+    std::tie(_starts_abs, ends_abs, _final_strides) = arm_compute::helpers::tensor_transform::calculate_strided_slice_coords(
+                                                          input_shape,
+                                                          starts, ends, strides,
+                                                          begin_mask, end_mask, shrink_axis_mask);
+
+    // Configure kernel window
+    auto win_config = validate_and_configure_window(input->info(), output->info(), starts, ends, strides, begin_mask, end_mask, shrink_axis_mask);
+    ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
+    INEKernel::configure(win_config.second);
+}
+
+Status NEStridedSliceKernel::validate(const ITensorInfo *input, const ITensorInfo *output,
+                                      const Coordinates &starts, const Coordinates &ends, const BiStrides &strides,
+                                      int32_t begin_mask, int32_t end_mask, int32_t shrink_axis_mask)
+{
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, starts, ends, strides, begin_mask, end_mask, shrink_axis_mask));
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), output->clone().get(),
+                                                              starts, ends, strides, begin_mask, end_mask, shrink_axis_mask)
+                                .first);
+
+    return Status{};
+}
+
+void NEStridedSliceKernel::run(const Window &window, const ThreadInfo &info)
+{
+    ARM_COMPUTE_UNUSED(info);
+    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window);
+
+    // Dispatch kernel
+    strided_slice_generic(_input, _output, _starts_abs, _final_strides, _shrink_mask, window);
+}
+} // namespace arm_compute
diff --git a/src/runtime/NEON/functions/NESlice.cpp b/src/runtime/NEON/functions/NESlice.cpp
new file mode 100644
index 0000000000..03c2053b9b
--- /dev/null
+++ b/src/runtime/NEON/functions/NESlice.cpp
@@ -0,0 +1,63 @@
+/*
+ * Copyright (c) 2018-2019 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "arm_compute/runtime/NEON/functions/NESlice.h"
+
+#include "arm_compute/core/ITensor.h"
+#include "arm_compute/core/NEON/kernels/NEStridedSliceKernel.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/Validate.h"
+#include "arm_compute/core/utils/helpers/tensor_transform.h"
+
+#include "support/ToolchainSupport.h"
+
+namespace arm_compute
+{
+void NESlice::configure(const ITensor *input, ITensor *output, const Coordinates &starts, const Coordinates &ends)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input);
+
+    // Get absolute end coordinates
+    const int32_t slice_end_mask = arm_compute::helpers::tensor_transform::construct_slice_end_mask(ends);
+
+    auto k = arm_compute::support::cpp14::make_unique<NEStridedSliceKernel>();
+    k->configure(input, output, starts, ends, BiStrides(), 0, slice_end_mask, 0);
+    _kernel = std::move(k);
+}
+
+Status NESlice::validate(const ITensorInfo *input, const ITensorInfo *output, const Coordinates &starts, const Coordinates &ends)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input);
+
+    // Check start dimensions for being non-negative
+    ARM_COMPUTE_RETURN_ERROR_ON(std::any_of(starts.cbegin(), starts.cbegin() + starts.num_dimensions(), [](int i)
+    {
+        return i < 0;
+    }));
+
+    // Get absolute end coordinates
+    const int32_t slice_end_mask = arm_compute::helpers::tensor_transform::construct_slice_end_mask(ends);
+
+    return NEStridedSliceKernel::validate(input, output, starts, ends, BiStrides(), 0, slice_end_mask, 0);
+}
+} // namespace arm_compute
diff --git a/src/runtime/NEON/functions/NESplit.cpp b/src/runtime/NEON/functions/NESplit.cpp
new file mode 100644
index 0000000000..e947657934
--- /dev/null
+++ b/src/runtime/NEON/functions/NESplit.cpp
@@ -0,0 +1,138 @@
+/*
+ * Copyright (c) 2018-2019 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "arm_compute/runtime/NEON/functions/NESplit.h"
+
+#include "arm_compute/core/Error.h"
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/ITensor.h"
+#include "arm_compute/core/TensorInfo.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/utils/misc/ShapeCalculator.h"
+
+#include "support/ToolchainSupport.h"
+
+namespace arm_compute
+{
+NESplit::NESplit()
+    : _outputs_vector(), _slice_functions(), _num_outputs(0)
+{
+}
+
+void NESplit::configure(const ITensor *input, const std::vector<ITensor *> &outputs, unsigned int axis)
+{
+    // Create Slice functions
+    _num_outputs     = outputs.size();
+    _slice_functions = arm_compute::support::cpp14::make_unique<NESlice[]>(_num_outputs);
+
+    // Get output shape
+    const TensorShape output_shape = arm_compute::misc::shape_calculator::compute_split_shape(input->info(), axis, _num_outputs);
+
+    // Extract output tensor info
+    std::vector<ITensorInfo *> outputs_info;
+    for(auto &output : outputs)
+    {
+        ARM_COMPUTE_ERROR_ON_NULLPTR(output);
+        outputs_info.emplace_back(output->info());
+    }
+
+    // Validate
+    ARM_COMPUTE_ERROR_THROW_ON(NESplit::validate(input->info(), outputs_info, axis));
+
+    const size_t axis_split_step = output_shape[axis];
+    unsigned int axis_offset     = 0;
+
+    // Start/End coordinates
+    Coordinates start_coords;
+    Coordinates end_coords;
+    for(unsigned int d = 0; d < output_shape.num_dimensions(); ++d)
+    {
+        end_coords.set(d, -1);
+    }
+
+    for(unsigned int i = 0; i < _num_outputs; i++)
+    {
+        // Update coordinate on axis
+        start_coords.set(axis, axis_offset);
+        end_coords.set(axis, axis_offset + axis_split_step);
+
+        // Configure slice function
+        _slice_functions[i].configure(input, outputs[i], start_coords, end_coords);
+
+        // Set valid region from shape
+        outputs[i]->info()->set_valid_region(ValidRegion(Coordinates(), output_shape));
+
+        // Update axis offset
+        axis_offset += axis_split_step;
+    }
+}
+
+Status NESplit::validate(const ITensorInfo *input, const std::vector<ITensorInfo *> &outputs, unsigned int axis)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input);
+    ARM_COMPUTE_RETURN_ERROR_ON(axis >= input->num_dimensions());
+    ARM_COMPUTE_RETURN_ERROR_ON(outputs.size() < 2);
+
+    // Get output shape
+    const TensorShape output_shape = arm_compute::misc::shape_calculator::compute_split_shape(input, axis, outputs.size());
+    ARM_COMPUTE_RETURN_ERROR_ON(output_shape.total_size() == 0);
+
+    const size_t axis_split_step = output_shape[axis];
+    unsigned int axis_offset     = 0;
+
+    // Start/End coordinates
+    Coordinates start_coords;
+    Coordinates end_coords;
+    for(unsigned int d = 0; d < output_shape.num_dimensions(); ++d)
+    {
+        end_coords.set(d, -1);
+    }
+
+    // Validate output tensors
+    for(const auto &output : outputs)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(output);
+
+        // Output auto inizialitation if not yet initialized
+        TensorInfo tmp_output_info = *output->clone();
+        auto_init_if_empty(tmp_output_info, input->clone()->set_is_resizable(true).set_tensor_shape(output_shape));
+
+        // Update coordinate on axis
+        start_coords.set(axis, axis_offset);
+        end_coords.set(axis, axis_offset + axis_split_step);
+
+        ARM_COMPUTE_RETURN_ON_ERROR(NESlice::validate(input, output, start_coords, end_coords));
+        axis_offset += axis_split_step;
+    }
+
+    return Status{};
+}
+
+void NESplit::run()
+{
+    for(unsigned i = 0; i < _num_outputs; ++i)
+    {
+        _slice_functions[i].run();
+    }
+}
+} // namespace arm_compute
diff --git a/src/runtime/NEON/functions/NEStridedSlice.cpp b/src/runtime/NEON/functions/NEStridedSlice.cpp
new file mode 100644
index 0000000000..53eb2b03b0
--- /dev/null
+++ b/src/runtime/NEON/functions/NEStridedSlice.cpp
@@ -0,0 +1,47 @@
+/*
+ * Copyright (c) 2018-2019 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "arm_compute/runtime/NEON/functions/NEStridedSlice.h"
+
+#include "arm_compute/core/NEON/kernels/NEStridedSliceKernel.h"
+#include "arm_compute/core/Types.h"
+#include "support/ToolchainSupport.h"
+
+namespace arm_compute
+{
+void NEStridedSlice::configure(const ITensor *input, ITensor *output,
+                               const Coordinates &starts, const Coordinates &ends, const BiStrides &strides,
+                               int32_t begin_mask, int32_t end_mask, int32_t shrink_axis_mask)
+{
+    auto k = arm_compute::support::cpp14::make_unique<NEStridedSliceKernel>();
+    k->configure(input, output, starts, ends, strides, begin_mask, end_mask, shrink_axis_mask);
+    _kernel = std::move(k);
+}
+
+Status NEStridedSlice::validate(const ITensorInfo *input, const ITensorInfo *output,
+                                const Coordinates &starts, const Coordinates &ends, const BiStrides &strides,
+                                int32_t begin_mask, int32_t end_mask, int32_t shrink_axis_mask)
+{
+    return NEStridedSliceKernel::validate(input, output, starts, ends, strides, begin_mask, end_mask, shrink_axis_mask);
+}
+} // namespace arm_compute
diff --git a/src/runtime/NEON/functions/NEUnstack.cpp b/src/runtime/NEON/functions/NEUnstack.cpp
new file mode 100644
index 0000000000..7532020973
--- /dev/null
+++ b/src/runtime/NEON/functions/NEUnstack.cpp
@@ -0,0 +1,119 @@
+/*
+ * Copyright (c) 2018-2019 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "arm_compute/runtime/NEON/functions/NEUnstack.h"
+
+#include "arm_compute/core/Error.h"
+#include "arm_compute/core/ITensor.h"
+#include "arm_compute/core/TensorInfo.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/utils/misc/ShapeCalculator.h"
+
+namespace arm_compute
+{
+namespace
+{
+inline unsigned int wrap_axis(int axis, const ITensorInfo *const tensor)
+{
+    return wrap_around(axis, static_cast<int>(tensor->num_dimensions()));
+}
+
+inline void setup_slice_coordinates_and_mask(Coordinates &slice_start, int32_t &slice_end_mask, const unsigned int input_num_dimensions)
+{
+    // Setups up coordinates to slice the input tensor: start coordinates to all 0s and the unstacking axis of both Start/End to slice just one 2d tensor at a time.
+    Coordinates slice_end;
+    slice_start.set_num_dimensions(input_num_dimensions);
+    slice_end.set_num_dimensions(input_num_dimensions);
+    for(size_t k = 0; k < input_num_dimensions; ++k)
+    {
+        slice_start.set(k, 0);
+        slice_end.set(k, -1);
+    }
+    slice_end_mask = arm_compute::helpers::tensor_transform::construct_slice_end_mask(slice_end);
+}
+} // namespace
+
+NEUnstack::NEUnstack() // NOLINT
+    : _num_slices(0),
+      _strided_slice_vector()
+{
+}
+
+void NEUnstack::configure(const ITensor *input, const std::vector<ITensor *> &output_vector, int axis)
+{
+    std::vector<ITensorInfo *> outputs_vector_info(output_vector.size());
+    std::transform(output_vector.begin(), output_vector.end(), outputs_vector_info.begin(), [](ITensor * t)
+    {
+        ARM_COMPUTE_ERROR_ON_NULLPTR(t);
+        return t->info();
+    });
+
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input);
+    ARM_COMPUTE_ERROR_THROW_ON(NEUnstack::validate(input->info(), outputs_vector_info, axis));
+
+    // Wrap around negative values
+    const unsigned int axis_u = wrap_axis(axis, input->info());
+    _num_slices               = std::min(outputs_vector_info.size(), input->info()->dimension(axis_u));
+    _strided_slice_vector     = arm_compute::support::cpp14::make_unique<NEStridedSlice[]>(_num_slices);
+
+    Coordinates slice_start;
+    int32_t     slice_end_mask;
+    setup_slice_coordinates_and_mask(slice_start, slice_end_mask, input->info()->tensor_shape().num_dimensions());
+    for(unsigned int slice = 0; slice < _num_slices; ++slice)
+    {
+        // Adjusts start and end coordinates to take a 2D slice at a time
+        slice_start.set(axis_u, slice);
+        _strided_slice_vector[slice].configure(input, output_vector[slice], slice_start, Coordinates(), BiStrides(), 0, slice_end_mask, (1 << axis_u));
+    }
+}
+
+Status NEUnstack::validate(const ITensorInfo *input, const std::vector<ITensorInfo *> &output_vector, int axis)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input);
+    ARM_COMPUTE_RETURN_ERROR_ON(output_vector.empty());
+    ARM_COMPUTE_RETURN_ERROR_ON(axis < (-static_cast<int>(input->tensor_shape().num_dimensions())));
+    ARM_COMPUTE_RETURN_ERROR_ON(axis >= static_cast<int>(input->tensor_shape().num_dimensions()));
+
+    const unsigned int num_slices = std::min(output_vector.size(), input->dimension(wrap_axis(axis, input)));
+    ARM_COMPUTE_RETURN_ERROR_ON(num_slices > input->dimension(wrap_axis(axis, input)));
+    ARM_COMPUTE_RETURN_ERROR_ON(num_slices > output_vector.size());
+
+    Coordinates slice_start;
+    int32_t     slice_end_mask;
+    for(size_t k = 0; k < num_slices; ++k)
+    {
+        slice_start.set(wrap_axis(axis, input), k);
+        setup_slice_coordinates_and_mask(slice_start, slice_end_mask, input->tensor_shape().num_dimensions());
+        ARM_COMPUTE_RETURN_ON_ERROR(NEStridedSlice::validate(input, output_vector[k], slice_start, Coordinates(), BiStrides(), 0, slice_end_mask, (1 << wrap_axis(axis, input))));
+    }
+    return Status{};
+}
+
+void NEUnstack::run()
+{
+    for(unsigned i = 0; i < _num_slices; ++i)
+    {
+        _strided_slice_vector[i].run();
+    }
+}
+} // namespace arm_compute