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>

2 files changed, 91 insertions, 8 deletions

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());