COMPMID-1761: NEON: Implement Pack

Change-Id: Icc3392494b1e3361e8fd925da200827c494351b3
Reviewed-on: https://review.mlplatform.org/430
Reviewed-by: Manuel Bottini <manuel.bottini@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Giuseppe Rossini <giuseppe.rossini@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>

4 files changed, 263 insertions, 16 deletions

diff --git a/src/core/CL/kernels/CLStackLayerKernel.cpp b/src/core/CL/kernels/CLStackLayerKernel.cpp
index bac8992f7b..ac179ba5f6 100644
--- a/src/core/CL/kernels/CLStackLayerKernel.cpp
+++ b/src/core/CL/kernels/CLStackLayerKernel.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018 ARM Limited.
+ * Copyright (c) 2018-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -69,19 +69,9 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, unsi
     auto_init_if_empty(*output, input->clone()->set_tensor_shape(compute_stack_shape(*input, axis, num_tensors)));
 
     // Configure kernel window
-    constexpr unsigned int num_elems_processed_per_iteration = 1;
+    Window win = calculate_max_window(*input);
 
-    // The window needs to be based on input as we copy all the depths of input
-    Window win = calculate_max_window(*input, Steps(num_elems_processed_per_iteration));
-
-    AccessWindowHorizontal input_access(input, 0, num_elems_processed_per_iteration);
-    AccessWindowHorizontal output_access(output, 0, num_elems_processed_per_iteration);
-
-    bool window_changed = update_window_and_padding(win, input_access, output_access);
-    output_access.set_valid_region(win, ValidRegion(Coordinates(), output->tensor_shape()));
-
-    Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
-    return std::make_pair(err, win);
+    return std::make_pair(Status{}, win);
 }
 } // namespace
 
diff --git a/src/core/NEON/kernels/NEStackLayerKernel.cpp b/src/core/NEON/kernels/NEStackLayerKernel.cpp
new file mode 100644
index 0000000000..cc60609f9b
--- /dev/null
+++ b/src/core/NEON/kernels/NEStackLayerKernel.cpp
@@ -0,0 +1,166 @@
+/*
+ * 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/NEStackLayerKernel.h"
+
+#include "arm_compute/core/Error.h"
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/IAccessWindow.h"
+#include "arm_compute/core/ITensor.h"
+#include "arm_compute/core/TensorInfo.h"
+#include "arm_compute/core/Utils.h"
+#include "arm_compute/core/Validate.h"
+#include "arm_compute/core/Window.h"
+#include "arm_compute/core/utils/misc/ShapeCalculator.h"
+
+#include "support/ToolchainSupport.h"
+
+using namespace arm_compute;
+using namespace arm_compute::misc::shape_calculator;
+
+namespace
+{
+Status validate_arguments(const ITensorInfo *input, unsigned int axis, unsigned int idx_input, unsigned int num_tensors, const ITensorInfo *output)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::U8, DataType::S8,
+                                                         DataType::U16, DataType::S16, DataType::U32, DataType::S32,
+                                                         DataType::F16, DataType::F32);
+    ARM_COMPUTE_RETURN_ERROR_ON(idx_input >= num_tensors);
+    ARM_COMPUTE_RETURN_ERROR_ON(axis > input->num_dimensions());
+    ARM_COMPUTE_RETURN_ERROR_ON(input->num_dimensions() > 4);
+
+    if(output->total_size() != 0)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), compute_stack_shape(*input, axis, num_tensors));
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
+    }
+
+    return Status{};
+}
+
+std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, unsigned int axis, unsigned int num_tensors, ITensorInfo *output)
+{
+    // Output auto inizialitation if not yet initialized
+    auto_init_if_empty(*output, input->clone()->set_tensor_shape(compute_stack_shape(*input, axis, num_tensors)));
+
+    // Configure kernel window
+    Window win = calculate_max_window(*input);
+
+    return std::make_pair(Status{}, win);
+}
+
+inline Coordinates shift_from_axis_and_replace_coordinate(const Coordinates &id, unsigned int axis, unsigned int idx_input)
+{
+    constexpr int   max_out_coord = 5; // Input shape is max a 4D shape, output is max 5D
+    Coordinates id_out        = id;
+    for(unsigned int i = max_out_coord - 1; i > axis; --i)
+    {
+        id_out.set(i, id[i - 1]);
+    }
+    id_out.set(axis, idx_input);
+    return id_out;
+}
+} // namespace
+
+NEStackLayerKernel::NEStackLayerKernel()
+    : _input(nullptr), _output(nullptr), _axis(), _idx_input(), _func(nullptr)
+{
+}
+
+void NEStackLayerKernel::configure(const ITensor *input, unsigned int axis, unsigned int idx_input, unsigned int num_tensors, ITensor *output)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), axis, idx_input, num_tensors, output->info()));
+
+    _input     = input;
+    _output    = output;
+    _axis      = axis;
+    _idx_input = idx_input;
+
+    switch(input->info()->element_size())
+    {
+        case 1:
+            _func = &NEStackLayerKernel::run_stack<uint8_t>;
+            break;
+        case 2:
+            _func = &NEStackLayerKernel::run_stack<uint16_t>;
+            break;
+        case 4:
+            _func = &NEStackLayerKernel::run_stack<uint32_t>;
+            break;
+        default:
+            ARM_COMPUTE_ERROR("Element size not supported");
+            break;
+    }
+
+    // Configure kernel window
+    auto win_config = validate_and_configure_window(input->info(), axis, num_tensors, output->info());
+
+    ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
+    INEKernel::configure(win_config.second);
+}
+
+Status NEStackLayerKernel::validate(const ITensorInfo *input, unsigned int axis, unsigned int idx_input, unsigned int num_tensors, const ITensorInfo *output)
+{
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, axis, idx_input, num_tensors, output));
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), axis, num_tensors, output->clone().get()).first);
+    return Status{};
+}
+
+void NEStackLayerKernel::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);
+
+    if(_func != nullptr)
+    {
+        (this->*_func)(window);
+    }
+}
+
+template <typename T>
+void NEStackLayerKernel::run_stack(const Window &window)
+{
+    Window window_out;
+    window_out.use_tensor_dimensions(_output->info()->tensor_shape());
+
+    Iterator input(_input, window);
+    Iterator output(_output, window_out);
+
+    const int stride_x = _output->info()->strides_in_bytes()[0];
+    const int stride_y = _output->info()->num_dimensions() >= 1 ? _output->info()->strides_in_bytes()[1] : 0;
+    const int stride_z = _output->info()->num_dimensions() >= 2 ? _output->info()->strides_in_bytes()[2] : 0;
+    const int stride_w = _output->info()->num_dimensions() >= 3 ? _output->info()->strides_in_bytes()[3] : 0;
+    const int stride_k = _output->info()->num_dimensions() >= 4 ? _output->info()->strides_in_bytes()[4] : 0;
+
+    execute_window_loop(window, [&](const Coordinates & id)
+    {
+        Coordinates id_out                           = shift_from_axis_and_replace_coordinate(id, _axis, _idx_input);
+        const int   idx                              = id_out[0] * stride_x + id_out[1] * stride_y + id_out[2] * stride_z + id_out[3] * stride_w + id_out[4] * stride_k;
+        *(reinterpret_cast<T *>(output.ptr() + idx)) = *(reinterpret_cast<const T *>(input.ptr()));
+    },
+    input);
+}
diff --git a/src/runtime/CL/functions/CLStackLayer.cpp b/src/runtime/CL/functions/CLStackLayer.cpp
index 85adcad90c..71327fead4 100644
--- a/src/runtime/CL/functions/CLStackLayer.cpp
+++ b/src/runtime/CL/functions/CLStackLayer.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018 ARM Limited.
+ * Copyright (c) 2018-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -61,15 +61,19 @@ void CLStackLayer::configure(const std::vector<ICLTensor *> &input, int axis, IC
 Status CLStackLayer::validate(const std::vector<ITensorInfo *> &input, int axis, const ITensorInfo *output)
 {
     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(output);
-    ARM_COMPUTE_RETURN_ERROR_ON(input.size() < 2);
+    ARM_COMPUTE_RETURN_ERROR_ON(input.empty());
 
     // Wrap around negative values
-    const unsigned int axis_u = wrap_around(axis, static_cast<int>(input[0]->num_dimensions() + 1));
+    const size_t       rank   = input[0]->num_dimensions();
+    const unsigned int axis_u = wrap_around(axis, static_cast<int>(rank + 1));
 
     const unsigned int num_inputs = input.size();
 
     for(unsigned int i = 0; i < num_inputs; i++)
     {
+        // All the tensors must have the same rank
+        ARM_COMPUTE_RETURN_ERROR_ON(input[i]->num_dimensions() != rank);
+        // Validate Kernel
         ARM_COMPUTE_RETURN_ON_ERROR(CLStackLayerKernel::validate(input[i], axis_u, i, num_inputs, output));
     }
 
diff --git a/src/runtime/NEON/functions/NEStackLayer.cpp b/src/runtime/NEON/functions/NEStackLayer.cpp
new file mode 100644
index 0000000000..2f49c225a4
--- /dev/null
+++ b/src/runtime/NEON/functions/NEStackLayer.cpp
@@ -0,0 +1,87 @@
+/*
+ * 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/NEStackLayer.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 "arm_compute/runtime/NEON/NEScheduler.h"
+
+#include "support/ToolchainSupport.h"
+namespace arm_compute
+{
+NEStackLayer::NEStackLayer() // NOLINT
+    : _input(),
+      _stack_kernels(),
+      _num_inputs(0)
+{
+}
+
+void NEStackLayer::configure(const std::vector<ITensor *> &input, int axis, ITensor *output)
+{
+    _num_inputs    = input.size();
+    _stack_kernels = arm_compute::support::cpp14::make_unique<NEStackLayerKernel[]>(_num_inputs);
+
+    // Wrap around negative values
+    const unsigned int axis_u = wrap_around(axis, static_cast<int>(input[0]->info()->num_dimensions() + 1));
+
+    for(unsigned int i = 0; i < _num_inputs; i++)
+    {
+        _stack_kernels[i].configure(input[i], axis_u, i, _num_inputs, output);
+    }
+}
+
+Status NEStackLayer::validate(const std::vector<ITensorInfo *> &input, int axis, const ITensorInfo *output)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(output);
+    ARM_COMPUTE_RETURN_ERROR_ON(input.empty());
+
+    // Wrap around negative values
+    const size_t       rank   = input[0]->num_dimensions();
+    const unsigned int axis_u = wrap_around(axis, static_cast<int>(rank + 1));
+
+    const unsigned int num_inputs = input.size();
+
+    for(unsigned int i = 0; i < num_inputs; i++)
+    {
+        // All the tensors must have the same rank
+        ARM_COMPUTE_RETURN_ERROR_ON(input[i]->num_dimensions() != rank);
+        // Validate Kernel
+        ARM_COMPUTE_RETURN_ON_ERROR(NEStackLayerKernel::validate(input[i], axis_u, i, num_inputs, output));
+    }
+
+    return Status{};
+}
+
+void NEStackLayer::run()
+{
+    for(unsigned i = 0; i < _num_inputs; i++)
+    {
+        NEScheduler::get().schedule(&_stack_kernels[i], Window::DimY);
+    }
+}
+} // namespace arm_compute
\ No newline at end of file