COMPMID-2062 Rework NEON ConcatenateLayer

Change-Id: I2703c99f651e4f06c4e44e39a85a8a8d201c5362
Signed-off-by: Michalis Spyrou <michalis.spyrou@arm.com>
Reviewed-on: https://review.mlplatform.org/c/944
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>

6 files changed, 126 insertions, 139 deletions

diff --git a/arm_compute/runtime/NEON/functions/NEConcatenateLayer.h b/arm_compute/runtime/NEON/functions/NEConcatenateLayer.h
index b70d6ebc7c..7dfbcf9199 100644
--- a/arm_compute/runtime/NEON/functions/NEConcatenateLayer.h
+++ b/arm_compute/runtime/NEON/functions/NEConcatenateLayer.h
@@ -77,13 +77,9 @@ public:
     void run() override;
 
 private:
-    void configure_h_concatenate(std::vector<ITensor *> inputs_vector, ITensor *output);
-    static Status validate_h_concatenate(const std::vector<ITensorInfo *> &inputs_vector, const ITensorInfo *output);
-
-    std::unique_ptr<IFunction>                        _concat_function;
-    std::unique_ptr<NEHeightConcatenateLayerKernel[]> _hconcat_kernels;
-    unsigned int                                      _num_inputs;
-    unsigned int                                      _axis;
+    std::vector<std::unique_ptr<INEKernel>> _concat_kernels;
+    unsigned int                            _num_inputs;
+    unsigned int                            _axis;
 };
 } // namespace arm_compute
 #endif /* __ARM_COMPUTE_NECONCATENATELAYER_H__ */
diff --git a/arm_compute/runtime/NEON/functions/NEDepthConcatenateLayer.h b/arm_compute/runtime/NEON/functions/NEDepthConcatenateLayer.h
index e2162ef042..da38151e73 100644
--- a/arm_compute/runtime/NEON/functions/NEDepthConcatenateLayer.h
+++ b/arm_compute/runtime/NEON/functions/NEDepthConcatenateLayer.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2018 ARM Limited.
+ * Copyright (c) 2017-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -41,6 +41,8 @@ class ITensor;
  * -# @ref NEFillBorderKernel (executed if input's lowest two dimensions are smaller than respective output's dimensions)
  * -# @ref NEDepthConcatenateLayerKernel
  *
+ * @deprecated This function is deprecated and will be removed in release 19.08
+ *
  */
 class NEDepthConcatenateLayer : public IFunction
 {
diff --git a/arm_compute/runtime/NEON/functions/NEWidthConcatenateLayer.h b/arm_compute/runtime/NEON/functions/NEWidthConcatenateLayer.h
index 70a81b2788..0aaba7987f 100644
--- a/arm_compute/runtime/NEON/functions/NEWidthConcatenateLayer.h
+++ b/arm_compute/runtime/NEON/functions/NEWidthConcatenateLayer.h
@@ -44,6 +44,8 @@ class ITensor;
 /** Basic function to execute concatenate tensors along x axis. This function calls the following kernel:
  *
  * -# @ref NEWidthConcatenateLayerKernel
+ *
+ * @deprecated This function is deprecated and will be removed in release 19.08
  */
 class NEWidthConcatenateLayer : public IFunction
 {
diff --git a/src/runtime/NEON/functions/NEConcatenateLayer.cpp b/src/runtime/NEON/functions/NEConcatenateLayer.cpp
index 1897915d33..fa7b91c3ca 100644
--- a/src/runtime/NEON/functions/NEConcatenateLayer.cpp
+++ b/src/runtime/NEON/functions/NEConcatenateLayer.cpp
@@ -38,36 +38,16 @@
 namespace arm_compute
 {
 NEConcatenateLayer::NEConcatenateLayer()
-    : _concat_function(nullptr),
-      _hconcat_kernels(),
+    : _concat_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)
+void NEConcatenateLayer::configure(const std::vector<ITensor *> &inputs_vector, ITensor *output, DataLayoutDimension axis)
 {
+    ARM_COMPUTE_ERROR_ON(output == nullptr);
+    _axis       = get_data_layout_dimension_index(output->info()->data_layout(), axis);
     _num_inputs = inputs_vector.size();
 
     std::vector<ITensorInfo *> inputs_vector_info;
@@ -76,98 +56,108 @@ void NEConcatenateLayer::configure_h_concatenate(std::vector<ITensor *> inputs_v
         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);
+    TensorShape output_shape{};
+    if(_axis == Window::DimZ)
+    {
+        output_shape = arm_compute::misc::shape_calculator::calculate_depth_concatenate_shape(inputs_vector);
+    }
+    else
+    {
+        output_shape = arm_compute::misc::shape_calculator::calculate_concatenate_shape(inputs_vector, _axis);
+    }
 
     // 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()));
+    ARM_COMPUTE_ERROR_THROW_ON(NEConcatenateLayer::validate(inputs_vector_info, output->info(), axis));
 
     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);
-    _axis = get_data_layout_dimension_index(output->info()->data_layout(), axis);
-    switch(_axis)
-    {
-        case 0:
+        switch(_axis)
         {
-            auto func = support::cpp14::make_unique<NEWidthConcatenateLayer>();
-            func->configure(inputs_vector, output);
-            _concat_function = std::move(func);
-            break;
-        }
-        case 1:
-        {
-            configure_h_concatenate(inputs_vector, output);
-            break;
-        }
-        case 2:
-        {
-            auto func = support::cpp14::make_unique<NEDepthConcatenateLayer>();
-            func->configure(inputs_vector, output);
-            _concat_function = std::move(func);
-            break;
+            case Window::DimX:
+            {
+                auto kernel = support::cpp14::make_unique<NEWidthConcatenateLayerKernel>();
+                kernel->configure(inputs_vector.at(i), offset, output);
+                _concat_kernels.emplace_back(std::move(kernel));
+                break;
+            }
+            case Window::DimY:
+            {
+                auto kernel = support::cpp14::make_unique<NEHeightConcatenateLayerKernel>();
+                kernel->configure(inputs_vector.at(i), offset, output);
+                _concat_kernels.emplace_back(std::move(kernel));
+                break;
+            }
+            case Window::DimZ:
+            {
+                auto kernel = support::cpp14::make_unique<NEDepthConcatenateLayerKernel>();
+                kernel->configure(inputs_vector.at(i), offset, output);
+                _concat_kernels.emplace_back(std::move(kernel));
+                break;
+            }
+            default:
+                ARM_COMPUTE_ERROR("Axis not supported");
         }
-        default:
-            ARM_COMPUTE_ERROR("Concatenation is supported across width, height and depth only!");
+        offset += inputs_vector.at(i)->info()->dimension(_axis);
     }
 }
 
 Status NEConcatenateLayer::validate(const std::vector<ITensorInfo *> &inputs_vector, const ITensorInfo *output, DataLayoutDimension axis)
 {
-    ARM_COMPUTE_RETURN_ERROR_ON(output == nullptr);
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(output);
+    ARM_COMPUTE_RETURN_ERROR_ON(inputs_vector.size() < 2);
+    const unsigned int _axis = get_data_layout_dimension_index(inputs_vector[0]->data_layout(), axis);
 
-    switch(get_data_layout_dimension_index(output->data_layout(), axis))
+    // Output auto inizialitation if not yet initialized
+    TensorInfo  tmp_output_info = *output->clone();
+    TensorShape output_shape{};
+    if(_axis == Window::DimZ)
     {
-        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;
-        default:
-            ARM_COMPUTE_RETURN_ERROR_MSG("Concatenation is supported across width and depth only!");
+        output_shape = arm_compute::misc::shape_calculator::calculate_depth_concatenate_shape(inputs_vector);
     }
-    return Status{};
-}
+    else
+    {
+        output_shape = arm_compute::misc::shape_calculator::calculate_concatenate_shape(inputs_vector, _axis);
+    }
+    auto_init_if_empty(tmp_output_info, output_shape, 1, inputs_vector[0]->data_type());
 
-void NEConcatenateLayer::run()
-{
-    switch(_axis)
+    unsigned int offset = 0;
+    for(const auto &input : inputs_vector)
     {
-        case 0:
-        case 2:
-        {
-            ARM_COMPUTE_ERROR_ON(_concat_function == nullptr);
-            _concat_function->run();
-            break;
-        }
-        case 1:
+        ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input);
+        switch(_axis)
         {
-            for(unsigned i = 0; i < _num_inputs; ++i)
+            case Window::DimX:
             {
-                NEScheduler::get().schedule(_hconcat_kernels.get() + i, Window::DimY);
+                ARM_COMPUTE_RETURN_ON_ERROR(NEWidthConcatenateLayerKernel::validate(input, offset, &tmp_output_info));
+                break;
             }
-            break;
-        }
-        default:
-        {
-            ARM_COMPUTE_ERROR("Axis not supported.");
-            break;
+            case Window::DimY:
+            {
+                ARM_COMPUTE_RETURN_ON_ERROR(NEHeightConcatenateLayerKernel::validate(input, offset, &tmp_output_info));
+                break;
+            }
+            case Window::DimZ:
+            {
+                ARM_COMPUTE_RETURN_ON_ERROR(NEDepthConcatenateLayerKernel::validate(input, offset, &tmp_output_info));
+                break;
+            }
+            default:
+                ARM_COMPUTE_ERROR("Axis not supported");
         }
+        offset += input->dimension(_axis);
+    }
+
+    return Status{};
+}
+
+void NEConcatenateLayer::run()
+{
+    for(auto &kernel : _concat_kernels)
+    {
+        NEScheduler::get().schedule(kernel.get(), _axis);
     }
 }
 } // namespace arm_compute
diff --git a/tests/validation/NEON/DepthConcatenateLayer.cpp b/tests/validation/NEON/DepthConcatenateLayer.cpp
index 24e7649e7d..1b355ae17d 100644
--- a/tests/validation/NEON/DepthConcatenateLayer.cpp
+++ b/tests/validation/NEON/DepthConcatenateLayer.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2018 ARM Limited.
+ * Copyright (c) 2017-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -22,7 +22,7 @@
  * SOFTWARE.
  */
 #include "arm_compute/core/Types.h"
-#include "arm_compute/runtime/NEON/functions/NEDepthConcatenateLayer.h"
+#include "arm_compute/runtime/NEON/functions/NEConcatenateLayer.h"
 #include "arm_compute/runtime/Tensor.h"
 #include "arm_compute/runtime/TensorAllocator.h"
 #include "tests/NEON/Accessor.h"
@@ -31,7 +31,7 @@
 #include "tests/framework/Macros.h"
 #include "tests/framework/datasets/Datasets.h"
 #include "tests/validation/Validation.h"
-#include "tests/validation/fixtures/DepthConcatenateLayerFixture.h"
+#include "tests/validation/fixtures/ConcatenateLayerFixture.h"
 
 namespace arm_compute
 {
@@ -73,48 +73,29 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(
         inputs_vector_info_raw.emplace_back(&input);
     }
 
-    bool is_valid = bool(NEDepthConcatenateLayer::validate(inputs_vector_info_raw,
-                                                           &output_info.clone()->set_is_resizable(false)));
+    bool is_valid = bool(NEConcatenateLayer::validate(inputs_vector_info_raw, &output_info.clone()->set_is_resizable(false), DataLayoutDimension::CHANNEL));
     ARM_COMPUTE_EXPECT(is_valid == expected, framework::LogLevel::ERRORS);
 }
 // clang-format on
 // *INDENT-ON*
 
-TEST_CASE(Configuration, framework::DatasetMode::ALL)
-{
-    // Create tensors
-    Tensor src1 = create_tensor<Tensor>(TensorShape(32U, 32U, 128U), DataType::F32, 1);
-    Tensor src2 = create_tensor<Tensor>(TensorShape(32U, 32U, 32U), DataType::F32, 1);
-    Tensor dst;
-
-    ARM_COMPUTE_EXPECT(src1.info()->is_resizable(), framework::LogLevel::ERRORS);
-    ARM_COMPUTE_EXPECT(src2.info()->is_resizable(), framework::LogLevel::ERRORS);
-    ARM_COMPUTE_EXPECT(dst.info()->is_resizable(), framework::LogLevel::ERRORS);
-
-    // Create and configure function
-    NEDepthConcatenateLayer concat_layer;
-
-    concat_layer.configure({ &src1, &src2 }, &dst);
-
-    // Validate valid region
-    const ValidRegion valid_region = shape_to_valid_region(TensorShape(32U, 32U, 160U));
-    validate(dst.info()->valid_region(), valid_region);
-}
-
 template <typename T>
-using NEDepthConcatenateLayerFixture = DepthConcatenateLayerValidationFixture<Tensor, ITensor, Accessor, NEDepthConcatenateLayer, T>;
+using NEDepthConcatenateLayerFixture = ConcatenateLayerValidationFixture<Tensor, ITensor, Accessor, NEConcatenateLayer, T>;
 
 TEST_SUITE(Float)
 #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
 TEST_SUITE(FP16)
-FIXTURE_DATA_TEST_CASE(RunSmall, NEDepthConcatenateLayerFixture<half>, framework::DatasetMode::PRECOMMIT, combine(datasets::Small2DShapes(), framework::dataset::make("DataType",
-                                                                                                                  DataType::F16)))
+FIXTURE_DATA_TEST_CASE(RunSmall, NEDepthConcatenateLayerFixture<half>, framework::DatasetMode::PRECOMMIT, combine(combine(concat(datasets::Small2DShapes(), datasets::Tiny4DShapes()),
+                                                                                                                  framework::dataset::make("DataType",
+                                                                                                                          DataType::F16)),
+                                                                                                                  framework::dataset::make("Axis", 2)))
 {
     // Validate output
     validate(Accessor(_target), _reference);
 }
-FIXTURE_DATA_TEST_CASE(RunLarge, NEDepthConcatenateLayerFixture<half>, framework::DatasetMode::NIGHTLY, combine(datasets::DepthConcatenateLayerShapes(), framework::dataset::make("DataType",
-                                                                                                                DataType::F16)))
+FIXTURE_DATA_TEST_CASE(RunLarge, NEDepthConcatenateLayerFixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::ConcatenateLayerShapes(), framework::dataset::make("DataType",
+                                                                                                                        DataType::F16)),
+                                                                                                                framework::dataset::make("Axis", 2)))
 {
     // Validate output
     validate(Accessor(_target), _reference);
@@ -123,14 +104,17 @@ TEST_SUITE_END()
 #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
 
 TEST_SUITE(FP32)
-FIXTURE_DATA_TEST_CASE(RunSmall, NEDepthConcatenateLayerFixture<float>, framework::DatasetMode::PRECOMMIT, combine(datasets::Small2DShapes(), framework::dataset::make("DataType",
-                                                                                                                   DataType::F32)))
+FIXTURE_DATA_TEST_CASE(RunSmall, NEDepthConcatenateLayerFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(concat(datasets::Small3DShapes(), datasets::Tiny4DShapes()),
+                                                                                                                   framework::dataset::make("DataType",
+                                                                                                                           DataType::F32)),
+                                                                                                                   framework::dataset::make("Axis", 2)))
 {
     // Validate output
     validate(Accessor(_target), _reference);
 }
-FIXTURE_DATA_TEST_CASE(RunLarge, NEDepthConcatenateLayerFixture<float>, framework::DatasetMode::NIGHTLY, combine(datasets::DepthConcatenateLayerShapes(), framework::dataset::make("DataType",
-                                                                                                                 DataType::F32)))
+FIXTURE_DATA_TEST_CASE(RunLarge, NEDepthConcatenateLayerFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::ConcatenateLayerShapes(), framework::dataset::make("DataType",
+                                                                                                                 DataType::F32)),
+                                                                                                                 framework::dataset::make("Axis", 2)))
 {
     // Validate output
     validate(Accessor(_target), _reference);
@@ -140,14 +124,18 @@ TEST_SUITE_END()
 
 TEST_SUITE(Quantized)
 TEST_SUITE(QASYMM8)
-FIXTURE_DATA_TEST_CASE(RunSmall, NEDepthConcatenateLayerFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(datasets::Small2DShapes(), framework::dataset::make("DataType",
-                                                                                                                     DataType::QASYMM8)))
+FIXTURE_DATA_TEST_CASE(RunSmall, NEDepthConcatenateLayerFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(concat(datasets::Small3DShapes(), datasets::Tiny4DShapes()),
+                                                                                                                     framework::dataset::make("DataType",
+                                                                                                                             DataType::QASYMM8)),
+                                                                                                                     framework::dataset::make("Axis", 2)))
 {
     // Validate output
     validate(Accessor(_target), _reference);
 }
-FIXTURE_DATA_TEST_CASE(RunLarge, NEDepthConcatenateLayerFixture<uint8_t>, framework::DatasetMode::NIGHTLY, combine(datasets::DepthConcatenateLayerShapes(), framework::dataset::make("DataType",
-                                                                                                                   DataType::QASYMM8)))
+FIXTURE_DATA_TEST_CASE(RunLarge, NEDepthConcatenateLayerFixture<uint8_t>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::ConcatenateLayerShapes(),
+                                                                                                                   framework::dataset::make("DataType",
+                                                                                                                           DataType::QASYMM8)),
+                                                                                                                   framework::dataset::make("Axis", 2)))
 {
     // Validate output
     validate(Accessor(_target), _reference);
diff --git a/tests/validation/reference/ConcatenateLayer.cpp b/tests/validation/reference/ConcatenateLayer.cpp
index 1440878829..93b1768ebd 100644
--- a/tests/validation/reference/ConcatenateLayer.cpp
+++ b/tests/validation/reference/ConcatenateLayer.cpp
@@ -114,6 +114,15 @@ SimpleTensor<T> concatenate_layer(std::vector<SimpleTensor<T>> &srcs, SimpleTens
             dst = reference::permute<T>(dst, PermutationVector(1U, 0U));
             return reference::permute<T>(widthconcatenate_layer(srcs, dst), PermutationVector(1U, 0U));
         }
+        case Window::DimZ:
+        {
+            for(auto &t : srcs)
+            {
+                t = reference::permute<T>(t, PermutationVector(2U, 1U, 0U));
+            }
+            dst = reference::permute<T>(dst, PermutationVector(2U, 1U, 0U));
+            return reference::permute<T>(widthconcatenate_layer(srcs, dst), PermutationVector(2U, 1U, 0U));
+        }
         default:
         {
             ARM_COMPUTE_ERROR("Not supported");