COMPMID-1365: Add support for NHWC in CLDepthConcatenateLayer

Change-Id: I3ed55bdb95d888aff0b0b76fb841bf1669659308
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/139963
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>

4 files changed, 195 insertions, 43 deletions

diff --git a/src/core/CL/kernels/CLDepthConcatenateLayerKernel.cpp b/src/core/CL/kernels/CLDepthConcatenateLayerKernel.cpp
index 72dc21197d..4055d1c7ab 100644
--- a/src/core/CL/kernels/CLDepthConcatenateLayerKernel.cpp
+++ b/src/core/CL/kernels/CLDepthConcatenateLayerKernel.cpp
@@ -41,6 +41,53 @@
 
 using namespace arm_compute;
 
+namespace
+{
+std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, unsigned int depth_offset, ITensorInfo *output)
+{
+    ARM_COMPUTE_UNUSED(depth_offset);
+
+    // Configure kernel window
+    const int left_right = (output->dimension(0) - input->dimension(0)) / 2;
+    const int top_bottom = (output->dimension(1) - input->dimension(1)) / 2;
+
+    const unsigned int num_elems_processed_per_iteration = 16 / input->element_size();
+    const unsigned int num_elems_read_per_iteration      = 16 / input->element_size();
+    const unsigned int num_rows_read_per_iteration       = 1;
+
+    // The window needs to be based on input as we copy all the depths of input
+    Window win = calculate_max_window(*output, Steps(num_elems_processed_per_iteration));
+    win.set(Window::DimZ, Window::Dimension(0, input->tensor_shape().z(), 1));
+
+    AccessWindowRectangle  input_access(input, -left_right, -top_bottom, num_elems_read_per_iteration, num_rows_read_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);
+}
+Status validate_arguments(const ITensorInfo *input, unsigned int depth_offset, const ITensorInfo *output)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
+    ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F16, DataType::F32);
+    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
+
+    ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(2) + depth_offset > output->dimension(2));
+    ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(0) > output->dimension(0));
+    ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(1) > output->dimension(1));
+    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(3, input, output);
+
+    // The gaps between the two lowest dimensions of input and output need to be divisible by 2
+    // Otherwise it is not clear how the padding should be added onto the input tensor
+    ARM_COMPUTE_RETURN_ERROR_ON((output->dimension(0) - input->dimension(0)) % 2);
+    ARM_COMPUTE_RETURN_ERROR_ON((output->dimension(1) - input->dimension(1)) % 2);
+
+    return Status{};
+}
+} // namespace
+
 CLDepthConcatenateLayerKernel::CLDepthConcatenateLayerKernel()
     : _input(nullptr), _output(nullptr), _top_bottom(0), _left_right(0), _depth_offset(0)
 {
@@ -53,59 +100,41 @@ BorderSize CLDepthConcatenateLayerKernel::border_size() const
 
 void CLDepthConcatenateLayerKernel::configure(const ICLTensor *input, unsigned int depth_offset, ICLTensor *output)
 {
-    static std::map<int, std::pair<std::string, int>> configs_map =
-    {
-        { 1, { "uchar", 16 } },
-        { 2, { "ushort", 8 } },
-        { 4, { "uint", 4 } },
-        { 8, { "ulong", 2 } },
-    };
-
-    ARM_COMPUTE_ERROR_ON_F16_UNSUPPORTED(input);
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32);
-    ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
-    ARM_COMPUTE_ERROR_ON(input->info()->dimension(2) + depth_offset > output->info()->dimension(2));
-    ARM_COMPUTE_ERROR_ON(input->info()->dimension(0) > output->info()->dimension(0));
-    ARM_COMPUTE_ERROR_ON(input->info()->dimension(1) > output->info()->dimension(1));
-    ARM_COMPUTE_ERROR_ON_MISMATCHING_SHAPES(3, input, output);
-    ARM_COMPUTE_ERROR_ON(configs_map.find(input->info()->element_size()) == configs_map.end());
-
-    // The gaps between the two lowest dimensions of input and output need to be divisible by 2
-    // Otherwise it is not clear how the padding should be added onto the input tensor
-    ARM_COMPUTE_ERROR_ON((output->info()->dimension(0) - input->info()->dimension(0)) % 2);
-    ARM_COMPUTE_ERROR_ON((output->info()->dimension(1) - input->info()->dimension(1)) % 2);
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), depth_offset, output->info()));
 
     _input        = input;
     _output       = output;
     _depth_offset = depth_offset;
 
+    const unsigned int num_elems_processed_per_iteration = 16 / input->info()->element_size();
+
     // Add build options
-    auto                  config = configs_map.find(static_cast<int>(input->info()->element_size()));
-    std::set<std::string> build_opts;
-    build_opts.emplace(("-DDATA_TYPE=" + config->second.first));
-    build_opts.emplace(("-DVEC_SIZE=" + support::cpp11::to_string(config->second.second)));
+    CLBuildOptions build_opts;
+    build_opts.add_option("-DDATA_TYPE=" + get_underlying_cl_type_from_data_type(input->info()->data_type()));
+    build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration));
 
     // Create kernel
-    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("concatenate_depth", build_opts));
+    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("concatenate_depth", build_opts.options()));
 
     // Configure kernel window
     _left_right = (output->info()->dimension(0) - input->info()->dimension(0)) / 2;
     _top_bottom = (output->info()->dimension(1) - input->info()->dimension(1)) / 2;
 
-    const unsigned int num_elems_processed_per_iteration = 16 / input->info()->element_size();
-    const unsigned int num_elems_read_per_iteration      = 16 / input->info()->element_size();
-    const unsigned int num_rows_read_per_iteration       = 1;
-
-    // The window needs to be based on input as we copy all the depths of input
-    Window win = calculate_max_window(*output->info(), Steps(num_elems_processed_per_iteration));
-    win.set(Window::DimZ, Window::Dimension(0, input->info()->tensor_shape().z(), 1));
+    // Configure kernel window
+    auto win_config = validate_and_configure_window(input->info(), depth_offset, output->info());
+    ARM_COMPUTE_ERROR_THROW_ON(std::get<0>(win_config));
 
-    AccessWindowRectangle  input_access(input->info(), -_left_right, -_top_bottom, num_elems_read_per_iteration, num_rows_read_per_iteration);
-    AccessWindowHorizontal output_access(output->info(), 0, num_elems_processed_per_iteration);
-    update_window_and_padding(win, input_access, output_access);
-    output_access.set_valid_region(win, ValidRegion(Coordinates(), output->info()->tensor_shape()));
+    ICLKernel::configure(std::get<1>(win_config));
+}
 
-    ICLKernel::configure(win);
+Status CLDepthConcatenateLayerKernel::validate(const arm_compute::ITensorInfo *input,
+                                               unsigned int                    depth_offset,
+                                               const arm_compute::ITensorInfo *output)
+{
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, depth_offset, output));
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), depth_offset, output->clone().get()).first);
+    return Status{};
 }
 
 void CLDepthConcatenateLayerKernel::run(const Window &window, cl::CommandQueue &queue)
diff --git a/src/runtime/CL/functions/CLConcatenateLayer.cpp b/src/runtime/CL/functions/CLConcatenateLayer.cpp
new file mode 100644
index 0000000000..f4bc1ff4ac
--- /dev/null
+++ b/src/runtime/CL/functions/CLConcatenateLayer.cpp
@@ -0,0 +1,90 @@
+/*
+ * Copyright (c) 2018 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/CL/functions/CLConcatenateLayer.h"
+
+#include "arm_compute/runtime/CL/functions/CLDepthConcatenateLayer.h"
+#include "arm_compute/runtime/CL/functions/CLWidthConcatenateLayer.h"
+
+#include "arm_compute/core/CL/ICLTensor.h"
+#include "arm_compute/core/Error.h"
+#include "arm_compute/core/TensorInfo.h"
+#include "arm_compute/core/Types.h"
+#include "support/ToolchainSupport.h"
+
+namespace arm_compute
+{
+CLConcatenateLayer::CLConcatenateLayer()
+    : _concat_function(nullptr)
+{
+}
+
+void CLConcatenateLayer::configure(std::vector<ICLTensor *> inputs_vector, ICLTensor *output, DataLayoutDimension axis)
+{
+    ARM_COMPUTE_ERROR_ON(output == nullptr);
+
+    switch(get_data_layout_dimension_index(output->info()->data_layout(), axis))
+    {
+        case 0:
+        {
+            auto func = support::cpp14::make_unique<CLWidthConcatenateLayer>();
+            func->configure(std::move(inputs_vector), output);
+            _concat_function = std::move(func);
+            break;
+        }
+        case 2:
+        {
+            auto func = support::cpp14::make_unique<CLDepthConcatenateLayer>();
+            func->configure(std::move(inputs_vector), output);
+            _concat_function = std::move(func);
+            break;
+        }
+        default:
+            ARM_COMPUTE_ERROR("Concatenation is supported across width and depth only!");
+    }
+}
+
+Status CLConcatenateLayer::validate(const std::vector<ITensorInfo *> &inputs_vector, const ITensorInfo *output, DataLayoutDimension axis)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON(output == nullptr);
+
+    switch(get_data_layout_dimension_index(output->data_layout(), axis))
+    {
+        case 0:
+            ARM_COMPUTE_RETURN_ON_ERROR(CLWidthConcatenateLayer::validate(inputs_vector, output));
+            break;
+        case 2:
+            ARM_COMPUTE_RETURN_ON_ERROR(CLDepthConcatenateLayer::validate(inputs_vector, output));
+            break;
+        default:
+            ARM_COMPUTE_RETURN_ERROR_MSG("Concatenation is supported across width and depth only!");
+    }
+    return Status{};
+}
+
+void CLConcatenateLayer::run()
+{
+    ARM_COMPUTE_ERROR_ON(_concat_function == nullptr);
+    _concat_function->run();
+}
+} // namespace arm_compute
diff --git a/src/runtime/CL/functions/CLDepthConcatenateLayer.cpp b/src/runtime/CL/functions/CLDepthConcatenateLayer.cpp
index 0b26f55a29..174be94410 100644
--- a/src/runtime/CL/functions/CLDepthConcatenateLayer.cpp
+++ b/src/runtime/CL/functions/CLDepthConcatenateLayer.cpp
@@ -27,7 +27,9 @@
 #include "arm_compute/core/Error.h"
 #include "arm_compute/core/Helpers.h"
 #include "arm_compute/core/PixelValue.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/CL/CLScheduler.h"
 #include "support/ToolchainSupport.h"
 
@@ -43,20 +45,24 @@ CLDepthConcatenateLayer::CLDepthConcatenateLayer() // NOLINT
 
 void CLDepthConcatenateLayer::configure(std::vector<ICLTensor *> inputs_vector, ICLTensor *output) // NOLINT
 {
-    ARM_COMPUTE_ERROR_ON(inputs_vector.size() < 2);
-
     _num_inputs = inputs_vector.size();
 
-    unsigned int depth_offset = 0;
+    std::vector<ITensorInfo *> inputs_vector_info;
+    for(unsigned int i = 0; i < _num_inputs; i++)
+    {
+        inputs_vector_info.emplace_back(inputs_vector.at(i)->info());
+    }
 
     _concat_kernels_vector  = arm_compute::support::cpp14::make_unique<CLDepthConcatenateLayerKernel[]>(_num_inputs);
     _border_handlers_vector = arm_compute::support::cpp14::make_unique<CLFillBorderKernel[]>(_num_inputs);
 
-    TensorShape output_shape = calculate_depth_concatenate_shape(inputs_vector);
+    TensorShape output_shape = arm_compute::misc::shape_calculator::calculate_depth_concatenate_shape(inputs_vector_info);
 
     // 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(CLDepthConcatenateLayer::validate(inputs_vector_info, output->info()));
 
+    unsigned int depth_offset = 0;
     for(unsigned int i = 0; i < _num_inputs; i++)
     {
         _concat_kernels_vector[i].configure(inputs_vector.at(i), depth_offset, output);
@@ -69,6 +75,27 @@ void CLDepthConcatenateLayer::configure(std::vector<ICLTensor *> inputs_vector,
     output->info()->set_valid_region(ValidRegion(Coordinates(), output_shape));
 }
 
+Status CLDepthConcatenateLayer::validate(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_depth_concatenate_shape(inputs_vector);
+    auto_init_if_empty(tmp_output_info, output_shape, 1, inputs_vector[0]->data_type());
+
+    unsigned int depth_offset = 0;
+    for(const auto &input : inputs_vector)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input);
+        ARM_COMPUTE_RETURN_ON_ERROR(CLDepthConcatenateLayerKernel::validate(input, depth_offset, &tmp_output_info));
+        depth_offset += input->dimension(2);
+    }
+
+    return Status{};
+}
+
 void CLDepthConcatenateLayer::run()
 {
     cl::CommandQueue q = CLScheduler::get().queue();
diff --git a/src/runtime/NEON/functions/NEDepthConcatenateLayer.cpp b/src/runtime/NEON/functions/NEDepthConcatenateLayer.cpp
index 3d47ec2ac2..cb0157664b 100644
--- a/src/runtime/NEON/functions/NEDepthConcatenateLayer.cpp
+++ b/src/runtime/NEON/functions/NEDepthConcatenateLayer.cpp
@@ -28,6 +28,7 @@
 #include "arm_compute/core/ITensor.h"
 #include "arm_compute/core/PixelValue.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"
 
@@ -49,7 +50,12 @@ void NEDepthConcatenateLayer::configure(std::vector<ITensor *> inputs_vector, IT
     _concat_kernels_vector  = arm_compute::support::cpp14::make_unique<NEDepthConcatenateLayerKernel[]>(_num_inputs);
     _border_handlers_vector = arm_compute::support::cpp14::make_unique<NEFillBorderKernel[]>(_num_inputs);
 
-    TensorShape output_shape = calculate_depth_concatenate_shape(inputs_vector);
+    std::vector<ITensorInfo *> inputs_vector_info;
+    for(unsigned int i = 0; i < _num_inputs; i++)
+    {
+        inputs_vector_info.emplace_back(inputs_vector.at(i)->info());
+    }
+    TensorShape output_shape = arm_compute::misc::shape_calculator::calculate_depth_concatenate_shape(inputs_vector_info);
 
     // Output auto inizialitation if not yet initialized
     auto_init_if_empty(*output->info(), output_shape, 1, inputs_vector[0]->info()->data_type());