COMPMID-1933: Implement NEHeightConcatenateLayer.

Added support to concactenate tensors along the Y axis in NEConcatenateLayer.

Change-Id: Ib714bfcf9954cc35918efa7d52fc9164bb08bdf6
Signed-off-by: Pablo Tello <pablo.tello@arm.com>
Reviewed-on: https://review.mlplatform.org/c/841
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

5 files changed, 236 insertions, 15 deletions

diff --git a/src/core/NEON/kernels/NEHeightConcatenateLayerKernel.cpp b/src/core/NEON/kernels/NEHeightConcatenateLayerKernel.cpp
new file mode 100644
index 0000000000..b8e204cfd8
--- /dev/null
+++ b/src/core/NEON/kernels/NEHeightConcatenateLayerKernel.cpp
@@ -0,0 +1,139 @@
+/*
+ * Copyright (c) 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/NEHeightConcatenateLayerKernel.h"
+
+#include "arm_compute/core/Error.h"
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/IAccessWindow.h"
+#include "arm_compute/core/ITensor.h"
+#include "arm_compute/core/NEON/NEAsymm.h"
+#include "arm_compute/core/NEON/wrapper/wrapper.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 <cstdint>
+
+using namespace arm_compute;
+
+namespace
+{
+std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *output)
+{
+    const unsigned int num_elems_processed_per_iteration = 16 / output->element_size();
+
+    // The window needs to be based on input as we copy all the widths 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);
+
+    Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
+    return std::make_pair(err, win);
+}
+
+Status validate_arguments(const ITensorInfo *input, unsigned int height_offset, const ITensorInfo *output)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
+    // Note: ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(input) is not needed here as this kernel doesn't use NEON FP16 instructions.
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1,
+                                                         DataType::U8, DataType::S8, DataType::QASYMM8,
+                                                         DataType::U16, DataType::S16, DataType::F16,
+                                                         DataType::U32, DataType::F32);
+    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
+    ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(Window::DimX) != output->dimension(Window::DimX));
+    ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(Window::DimY) + height_offset > output->dimension(Window::DimY));
+    for(size_t i = 2; i < Coordinates::num_max_dimensions; ++i)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(i) != output->dimension(i));
+    }
+
+    return Status{};
+}
+} // namespace
+
+NEHeightConcatenateLayerKernel::NEHeightConcatenateLayerKernel()
+    : _input(nullptr), _output(nullptr), _height_offset(0)
+{
+}
+
+void NEHeightConcatenateLayerKernel::configure(const ITensor *input, unsigned int height_offset, ITensor *output)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), height_offset, output->info()));
+
+    _input         = input;
+    _output        = output;
+    _height_offset = height_offset;
+
+    // Configure kernel window
+    auto win_config = validate_and_configure_window(input->info(), output->info());
+    ARM_COMPUTE_ERROR_THROW_ON(std::get<0>(win_config));
+
+    INEKernel::configure(std::get<1>(win_config));
+}
+
+Status NEHeightConcatenateLayerKernel::validate(const ITensorInfo *input, unsigned int height_offset, const ITensorInfo *output)
+{
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, height_offset, output));
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), output->clone().get()).first);
+    return Status{};
+}
+
+void NEHeightConcatenateLayerKernel::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);
+
+    // Offset output pointer to the correct position
+    uint8_t *output_ptr = _output->buffer() + _output->info()->offset_first_element_in_bytes() + _height_offset * _output->info()->strides_in_bytes()[Window::DimY];
+
+    // Create iterators
+    Iterator                input(_input, window);
+    Iterator                output(_output, window);
+    const DataType          dt           = _input->info()->data_type();
+    const QuantizationInfo &input_qinfo  = _input->info()->quantization_info();
+    const QuantizationInfo &output_qinfo = _output->info()->quantization_info();
+    if(dt == DataType::QASYMM8 && input_qinfo != output_qinfo)
+    {
+        execute_window_loop(window, [&](const Coordinates &)
+        {
+            vst1q_u8(output_ptr + output.offset(), vquantize(vdequantize(vld1q_u8(input.ptr()), input_qinfo), output_qinfo));
+        },
+        input, output);
+    }
+    else
+    {
+        execute_window_loop(window, [&](const Coordinates &)
+        {
+            const auto in_ptr  = input.ptr();
+            const auto out_ptr = output_ptr + output.offset();
+
+            wrapper::vstore(out_ptr, wrapper::vloadq(in_ptr));
+        },
+        input, output);
+    }
+}
diff --git a/src/graph/nodes/ConcatenateLayerNode.cpp b/src/graph/nodes/ConcatenateLayerNode.cpp
index ade3f6e1a9..3ce09d0073 100644
--- a/src/graph/nodes/ConcatenateLayerNode.cpp
+++ b/src/graph/nodes/ConcatenateLayerNode.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018 ARM Limited.
+ * Copyright (c) 2018-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -71,10 +71,9 @@ TensorDescriptor ConcatenateLayerNode::compute_output_descriptor(const std::vect
         shapes.emplace_back(&input_descriptor.shape);
     }
 
-    // Calculate output shape
-    if(axis_idx == 0)
+    if(axis_idx < 2)
     {
-        output_descriptor.shape = arm_compute::misc::shape_calculator::calculate_width_concatenate_shape(shapes);
+        output_descriptor.shape = arm_compute::misc::shape_calculator::calculate_concatenate_shape(shapes, axis_idx);
     }
     else if(axis_idx == 2)
     {
@@ -138,4 +137,4 @@ void ConcatenateLayerNode::accept(INodeVisitor &v)
     v.visit(*this);
 }
 } // namespace graph
-} // namespace arm_compute
\ No newline at end of file
+} // namespace arm_compute
diff --git a/src/runtime/CL/functions/CLWidthConcatenateLayer.cpp b/src/runtime/CL/functions/CLWidthConcatenateLayer.cpp
index d0801a6768..6e42377a07 100644
--- a/src/runtime/CL/functions/CLWidthConcatenateLayer.cpp
+++ b/src/runtime/CL/functions/CLWidthConcatenateLayer.cpp
@@ -51,7 +51,7 @@ Status CLWidthConcatenateLayer::validate(const std::vector<ITensorInfo *> &input
 
     // Output auto inizialitation if not yet initialized
     TensorInfo        tmp_output_info = *output->clone();
-    const TensorShape output_shape    = arm_compute::misc::shape_calculator::calculate_width_concatenate_shape(inputs_vector);
+    const TensorShape output_shape    = arm_compute::misc::shape_calculator::calculate_concatenate_shape(inputs_vector, Window::DimX);
     auto_init_if_empty(tmp_output_info, output_shape, 1, inputs_vector[0]->data_type());
 
     switch(num_inputs)
@@ -90,7 +90,7 @@ void CLWidthConcatenateLayer::configure(std::vector<ICLTensor *> inputs_vector,
     {
         inputs_vector_info.emplace_back(inputs_vector.at(i)->info());
     }
-    const TensorShape output_shape = arm_compute::misc::shape_calculator::calculate_width_concatenate_shape(inputs_vector);
+    const TensorShape output_shape = arm_compute::misc::shape_calculator::calculate_concatenate_shape(inputs_vector, Window::DimX);
 
     // Output auto inizialitation if not yet initialized
     auto_init_if_empty(*output->info(), output_shape, 1, inputs_vector[0]->info()->data_type());
diff --git a/src/runtime/NEON/functions/NEConcatenateLayer.cpp b/src/runtime/NEON/functions/NEConcatenateLayer.cpp
index 21ab47d3fe..f764a126a0 100644
--- a/src/runtime/NEON/functions/NEConcatenateLayer.cpp
+++ b/src/runtime/NEON/functions/NEConcatenateLayer.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018 ARM Limited.
+ * Copyright (c) 2018-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -26,6 +26,9 @@
 #include "arm_compute/runtime/NEON/functions/NEDepthConcatenateLayer.h"
 #include "arm_compute/runtime/NEON/functions/NEWidthConcatenateLayer.h"
 
+#include "arm_compute/core/utils/misc/ShapeCalculator.h"
+#include "arm_compute/runtime/NEON/NEScheduler.h"
+
 #include "arm_compute/core/Error.h"
 #include "arm_compute/core/ITensor.h"
 #include "arm_compute/core/TensorInfo.h"
@@ -35,15 +38,66 @@
 namespace arm_compute
 {
 NEConcatenateLayer::NEConcatenateLayer()
-    : _concat_function(nullptr)
+    : _concat_function(nullptr),
+      _hconcat_kernels(),
+      _num_inputs(0),
+      _axis(Window::DimX)
 {
 }
 
+Status NEConcatenateLayer::validate_h_concatenate(const std::vector<ITensorInfo *> &inputs_vector, const ITensorInfo *output)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(output);
+    ARM_COMPUTE_RETURN_ERROR_ON(inputs_vector.size() < 2);
+
+    // Output auto inizialitation if not yet initialized
+    TensorInfo  tmp_output_info = *output->clone();
+    TensorShape output_shape    = arm_compute::misc::shape_calculator::calculate_concatenate_shape(inputs_vector, Window::DimY);
+    auto_init_if_empty(tmp_output_info, output_shape, 1, inputs_vector[0]->data_type());
+
+    unsigned int offset = 0;
+    for(const auto &input : inputs_vector)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input);
+        ARM_COMPUTE_RETURN_ON_ERROR(NEHeightConcatenateLayerKernel::validate(input, offset, &tmp_output_info));
+        offset += input->dimension(Window::DimY);
+    }
+
+    return Status{};
+}
+
+void NEConcatenateLayer::configure_h_concatenate(std::vector<ITensor *> inputs_vector, ITensor *output)
+{
+    _num_inputs = inputs_vector.size();
+
+    std::vector<ITensorInfo *> inputs_vector_info;
+    for(unsigned int i = 0; i < _num_inputs; ++i)
+    {
+        ARM_COMPUTE_ERROR_ON_NULLPTR(inputs_vector.at(i));
+        inputs_vector_info.emplace_back(inputs_vector.at(i)->info());
+    }
+    TensorShape output_shape = arm_compute::misc::shape_calculator::calculate_concatenate_shape(inputs_vector, Window::DimY);
+
+    // Output auto inizialitation if not yet initialized
+    auto_init_if_empty(*output->info(), output_shape, 1, inputs_vector[0]->info()->data_type());
+    ARM_COMPUTE_ERROR_THROW_ON(validate_h_concatenate(inputs_vector_info, output->info()));
+
+    unsigned int offset = 0;
+
+    _hconcat_kernels = arm_compute::support::cpp14::make_unique<NEHeightConcatenateLayerKernel[]>(_num_inputs);
+
+    for(unsigned int i = 0; i < _num_inputs; ++i)
+    {
+        _hconcat_kernels[i].configure(inputs_vector.at(i), offset, output);
+        offset += inputs_vector.at(i)->info()->dimension(Window::DimY);
+    }
+}
+
 void NEConcatenateLayer::configure(const std::vector<ITensor *> &inputs_vector, ITensor *output, DataLayoutDimension axis)
 {
     ARM_COMPUTE_ERROR_ON(output == nullptr);
-
-    switch(get_data_layout_dimension_index(output->info()->data_layout(), axis))
+    _axis = get_data_layout_dimension_index(output->info()->data_layout(), axis);
+    switch(_axis)
     {
         case 0:
         {
@@ -52,6 +106,11 @@ void NEConcatenateLayer::configure(const std::vector<ITensor *> &inputs_vector,
             _concat_function = std::move(func);
             break;
         }
+        case 1:
+        {
+            configure_h_concatenate(inputs_vector, output);
+            break;
+        }
         case 2:
         {
             auto func = support::cpp14::make_unique<NEDepthConcatenateLayer>();
@@ -73,6 +132,9 @@ Status NEConcatenateLayer::validate(const std::vector<ITensorInfo *> &inputs_vec
         case 0:
             ARM_COMPUTE_RETURN_ON_ERROR(NEWidthConcatenateLayer::validate(inputs_vector, output));
             break;
+        case 1:
+            ARM_COMPUTE_RETURN_ON_ERROR(NEConcatenateLayer::validate_h_concatenate(inputs_vector, output));
+            break;
         case 2:
             ARM_COMPUTE_RETURN_ON_ERROR(NEDepthConcatenateLayer::validate(inputs_vector, output));
             break;
@@ -84,7 +146,28 @@ Status NEConcatenateLayer::validate(const std::vector<ITensorInfo *> &inputs_vec
 
 void NEConcatenateLayer::run()
 {
-    ARM_COMPUTE_ERROR_ON(_concat_function == nullptr);
-    _concat_function->run();
+    switch(_axis)
+    {
+        case 0:
+        case 2:
+        {
+            ARM_COMPUTE_ERROR_ON(_concat_function == nullptr);
+            _concat_function->run();
+            break;
+        }
+        case 1:
+        {
+            for(unsigned i = 0; i < _num_inputs; ++i)
+            {
+                NEScheduler::get().schedule(_hconcat_kernels.get() + i, Window::DimY);
+            }
+            break;
+        }
+        default:
+        {
+            ARM_COMPUTE_ERROR("Axis not supported.");
+            break;
+        }
+    }
 }
 } // namespace arm_compute
diff --git a/src/runtime/NEON/functions/NEWidthConcatenateLayer.cpp b/src/runtime/NEON/functions/NEWidthConcatenateLayer.cpp
index 17c352b8f3..9fce13cbd7 100644
--- a/src/runtime/NEON/functions/NEWidthConcatenateLayer.cpp
+++ b/src/runtime/NEON/functions/NEWidthConcatenateLayer.cpp
@@ -48,7 +48,7 @@ inline Status NEWidthConcatenateLayer::validate_internal(const std::vector<Tenso
 
     // Output auto inizialitation if not yet initialized
     TensorInfo  tmp_output_info = *output->clone();
-    TensorShape output_shape    = arm_compute::misc::shape_calculator::calculate_width_concatenate_shape(inputs_vector);
+    TensorShape output_shape    = arm_compute::misc::shape_calculator::calculate_concatenate_shape(inputs_vector, Window::DimX);
     auto_init_if_empty(tmp_output_info, output_shape, 1, inputs_vector[0]->data_type());
 
     unsigned int width_offset = 0;
@@ -71,7 +71,7 @@ inline void NEWidthConcatenateLayer::configure_internal(std::vector<TensorType *
     {
         inputs_vector_info.emplace_back(inputs_vector.at(i)->info());
     }
-    TensorShape output_shape = arm_compute::misc::shape_calculator::calculate_width_concatenate_shape(inputs_vector);
+    TensorShape output_shape = arm_compute::misc::shape_calculator::calculate_concatenate_shape(inputs_vector, Window::DimX);
 
     // Output auto inizialitation if not yet initialized
     auto_init_if_empty(*output->info(), output_shape, 1, inputs_vector[0]->info()->data_type());