1 files changed, 44 insertions, 205 deletions

diff --git a/src/runtime/CL/functions/CLConcatenateLayer.cpp b/src/runtime/CL/functions/CLConcatenateLayer.cpp
index e97256713f..9df1c34593 100644
--- a/src/runtime/CL/functions/CLConcatenateLayer.cpp
+++ b/src/runtime/CL/functions/CLConcatenateLayer.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018-2020 ARM Limited.
+ * Copyright (c) 2018-2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -23,238 +23,77 @@
  */
 #include "arm_compute/runtime/CL/functions/CLConcatenateLayer.h"
 
-#include "arm_compute/core/CL/kernels/CLBatchConcatenateLayerKernel.h"
-#include "arm_compute/core/CL/kernels/CLDepthConcatenateLayerKernel.h"
-#include "arm_compute/core/CL/kernels/CLHeightConcatenateLayerKernel.h"
-#include "arm_compute/core/CL/kernels/CLWidthConcatenate2TensorsKernel.h"
-#include "arm_compute/core/CL/kernels/CLWidthConcatenate4TensorsKernel.h"
-#include "arm_compute/core/CL/kernels/CLWidthConcatenateLayerKernel.h"
-#include "arm_compute/core/utils/misc/ShapeCalculator.h"
-#include "arm_compute/runtime/CL/CLScheduler.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/MemorySupport.h"
+
+#include "src/common/utils/Log.h"
+#include "src/core/CL/ICLKernel.h"
+#include "src/gpu/cl/operators/ClConcatenate.h"
 
 namespace arm_compute
 {
-CLConcatenateLayer::CLConcatenateLayer()
-    : _concat_kernels(),
-      _num_inputs(0),
-      _axis(Window::DimX)
+struct CLConcatenateLayer::Impl
 {
-}
-
-void CLConcatenateLayer::configure(std::vector<ICLTensor *> &inputs_vector, ICLTensor *output, size_t axis)
+    std::vector<const ICLTensor *>         srcs{};
+    ICLTensor                             *dst{nullptr};
+    unsigned int                           num_inputs{0};
+    unsigned int                           axis{0};
+    std::unique_ptr<opencl::ClConcatenate> op{nullptr};
+};
+
+CLConcatenateLayer::CLConcatenateLayer() : _impl(std::make_unique<Impl>())
 {
-    configure(CLKernelLibrary::get().get_compile_context(), inputs_vector, output, axis);
 }
 
-void CLConcatenateLayer::configure(const CLCompileContext &compile_context, std::vector<ICLTensor *> &inputs_vector, ICLTensor *output, size_t axis)
-{
-    configure_internal(compile_context, std::move(inputs_vector), output, axis);
-}
+CLConcatenateLayer::CLConcatenateLayer(CLConcatenateLayer &&) = default;
 
-void CLConcatenateLayer::configure(std::vector<const ICLTensor *> &inputs_vector, ICLTensor *output, size_t axis)
-{
-    configure(CLKernelLibrary::get().get_compile_context(), inputs_vector, output, axis);
-}
+CLConcatenateLayer &CLConcatenateLayer::operator=(CLConcatenateLayer &&) = default;
 
-void CLConcatenateLayer::configure(const CLCompileContext &compile_context, std::vector<const ICLTensor *> &inputs_vector, ICLTensor *output, size_t axis)
-{
-    configure_internal(compile_context, std::move(inputs_vector), output, axis);
-}
+CLConcatenateLayer::~CLConcatenateLayer() = default;
 
-Status CLConcatenateLayer::validate(const std::vector<ITensorInfo *> &inputs_vector, const ITensorInfo *output, size_t axis)
-{
-    return validate_internal(inputs_vector, output, axis);
-}
-
-Status CLConcatenateLayer::validate(const std::vector<const ITensorInfo *> &inputs_vector, const ITensorInfo *output, size_t axis)
+void CLConcatenateLayer::configure(std::vector<const ICLTensor *> &inputs_vector, ICLTensor *output, size_t axis)
 {
-    return validate_internal(inputs_vector, output, axis);
+    configure(CLKernelLibrary::get().get_compile_context(), inputs_vector, output, axis);
 }
 
-template <typename TensorType>
-void CLConcatenateLayer::configure_internal(const CLCompileContext &compile_context, std::vector<TensorType *> &&inputs_vector, ICLTensor *output, size_t axis)
+void CLConcatenateLayer::configure(const CLCompileContext         &compile_context,
+                                   std::vector<const ICLTensor *> &inputs_vector,
+                                   ICLTensor                      *output,
+                                   size_t                          axis)
 {
     ARM_COMPUTE_ERROR_ON(output == nullptr);
-    _axis       = axis;
-    _num_inputs = inputs_vector.size();
+    ARM_COMPUTE_LOG_PARAMS(inputs_vector, output, axis);
 
-    std::vector<ITensorInfo *> inputs_vector_info(inputs_vector.size());
-    std::transform(inputs_vector.begin(), inputs_vector.end(), inputs_vector_info.begin(), [](TensorType * t)
-    {
-        ARM_COMPUTE_ERROR_ON_NULLPTR(t);
-        return t->info();
-    });
-    TensorShape output_shape = arm_compute::misc::shape_calculator::calculate_concatenate_shape(inputs_vector, _axis);
+    _impl->srcs       = inputs_vector;
+    _impl->dst        = output;
+    _impl->axis       = axis;
+    _impl->num_inputs = inputs_vector.size();
+    _impl->op         = std::make_unique<opencl::ClConcatenate>();
 
-    // 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(CLConcatenateLayer::validate(inputs_vector_info, output->info(), axis));
-
-    unsigned int offset = 0;
-    switch(_axis)
+    std::vector<ITensorInfo *> inputs_vector_info;
+    for (unsigned int i = 0; i < inputs_vector.size(); ++i)
     {
-        case Window::DimX:
-        {
-            switch(_num_inputs)
-            {
-                case 2:
-                {
-                    // Configure WidthConcatenate2Tensors kernel
-                    auto kernel = support::cpp14::make_unique<CLWidthConcatenate2TensorsKernel>();
-                    kernel->configure(compile_context, inputs_vector.at(0), inputs_vector.at(1), output);
-                    _concat_kernels.emplace_back(std::move(kernel));
-                    break;
-                }
-                case 4:
-                {
-                    // Configure WidthConcatenate4Tensors kernel
-                    auto kernel = support::cpp14::make_unique<CLWidthConcatenate4TensorsKernel>();
-                    kernel->configure(compile_context, inputs_vector.at(0), inputs_vector.at(1), inputs_vector.at(2), inputs_vector.at(3), output);
-                    _concat_kernels.emplace_back(std::move(kernel));
-                    break;
-                }
-                default:
-                {
-                    // Configure generic case WidthConcatenate kernels
-                    for(unsigned int i = 0; i < _num_inputs; ++i)
-                    {
-                        auto kernel = support::cpp14::make_unique<CLWidthConcatenateLayerKernel>();
-                        kernel->configure(compile_context, inputs_vector.at(i), offset, output);
-                        offset += inputs_vector.at(i)->info()->dimension(_axis);
-                        _concat_kernels.emplace_back(std::move(kernel));
-                    }
-                    break;
-                }
-            }
-            break;
-        }
-        case Window::DimY:
-        {
-            for(unsigned int i = 0; i < _num_inputs; ++i)
-            {
-                auto kernel = support::cpp14::make_unique<CLHeightConcatenateLayerKernel>();
-                kernel->configure(compile_context, inputs_vector.at(i), offset, output);
-                offset += inputs_vector.at(i)->info()->dimension(_axis);
-                _concat_kernels.emplace_back(std::move(kernel));
-            }
-            break;
-        }
-        case Window::DimZ:
-        {
-            for(unsigned int i = 0; i < _num_inputs; ++i)
-            {
-                auto kernel = support::cpp14::make_unique<CLDepthConcatenateLayerKernel>();
-                kernel->configure(compile_context, inputs_vector.at(i), offset, output);
-                offset += inputs_vector.at(i)->info()->dimension(_axis);
-                _concat_kernels.emplace_back(std::move(kernel));
-            }
-            break;
-        }
-        case 3:
-        {
-            for(unsigned int i = 0; i < _num_inputs; ++i)
-            {
-                auto kernel = support::cpp14::make_unique<CLBatchConcatenateLayerKernel>();
-                kernel->configure(compile_context, inputs_vector.at(i), offset, output);
-                offset += inputs_vector.at(i)->info()->dimension(_axis);
-                _concat_kernels.emplace_back(std::move(kernel));
-            }
-            break;
-        }
-        default:
-            ARM_COMPUTE_ERROR("Axis not supported");
+        ARM_COMPUTE_ERROR_ON_NULLPTR(inputs_vector.at(i));
+        inputs_vector_info.emplace_back(inputs_vector.at(i)->info());
     }
+    _impl->op->configure(compile_context, inputs_vector_info, _impl->dst->info(), axis);
 }
 
-template <typename TensorInfoType>
-Status CLConcatenateLayer::validate_internal(const std::vector<TensorInfoType *> &inputs_vector, const ITensorInfo *output, size_t axis)
+Status CLConcatenateLayer::validate(const std::vector<const ITensorInfo *> &inputs_vector,
+                                    const ITensorInfo                      *output,
+                                    size_t                                  axis)
 {
-    ARM_COMPUTE_RETURN_ERROR_ON(output == nullptr);
-    const unsigned int num_inputs = inputs_vector.size();
-
-    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(output);
-    ARM_COMPUTE_RETURN_ERROR_ON(num_inputs < 2);
-
-    unsigned int offset = 0;
-    switch(axis)
-    {
-        case Window::DimX:
-        {
-            switch(num_inputs)
-            {
-                case 2:
-                    // Validate WidthConcatenate2Tensors kernels if there are 2 inputs
-                    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(inputs_vector[0], inputs_vector[1]);
-                    ARM_COMPUTE_RETURN_ON_ERROR(CLWidthConcatenate2TensorsKernel::validate(inputs_vector[0], inputs_vector[1], output));
-                    break;
-                case 4:
-                    // Validate WidthConcatenate4Tensors kernels if there are 4 inputs
-                    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(inputs_vector[0], inputs_vector[1], inputs_vector[2], inputs_vector[3]);
-                    ARM_COMPUTE_RETURN_ON_ERROR(CLWidthConcatenate4TensorsKernel::validate(inputs_vector[0], inputs_vector[1], inputs_vector[2], inputs_vector[3], output));
-                    break;
-                default:
-                    // Validate generic case of WidthConcatenate kernel
-                    for(const auto &input : inputs_vector)
-                    {
-                        ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input);
-                        ARM_COMPUTE_RETURN_ON_ERROR(CLWidthConcatenateLayerKernel::validate(input, offset, output));
-                        offset += input->dimension(axis);
-                    }
-                    break;
-            }
-            break;
-        }
-        case Window::DimY:
-        {
-            for(const auto &input : inputs_vector)
-            {
-                ARM_COMPUTE_RETURN_ON_ERROR(CLHeightConcatenateLayerKernel::validate(input, offset, output));
-                offset += input->dimension(axis);
-            }
-            break;
-        }
-        case Window::DimZ:
-        {
-            for(const auto &input : inputs_vector)
-            {
-                ARM_COMPUTE_RETURN_ON_ERROR(CLDepthConcatenateLayerKernel::validate(input, offset, output));
-                offset += input->dimension(axis);
-            }
-            break;
-        }
-        case 3:
-        {
-            for(const auto &input : inputs_vector)
-            {
-                ARM_COMPUTE_RETURN_ON_ERROR(CLBatchConcatenateLayerKernel::validate(input, offset, output));
-                offset += input->dimension(axis);
-            }
-            break;
-        }
-        default:
-            ARM_COMPUTE_ERROR("Axis not supported");
-    }
-
-    if(output->total_size() != 0)
-    {
-        TensorShape output_shape = arm_compute::misc::shape_calculator::calculate_concatenate_shape(inputs_vector, axis);
-        ARM_COMPUTE_RETURN_ERROR_ON(output_shape.total_size() != output->tensor_shape().total_size());
-    }
-
-    return Status{};
+    return opencl::ClConcatenate::validate(inputs_vector, output, axis);
 }
 
 void CLConcatenateLayer::run()
 {
-    for(auto &kernel : _concat_kernels)
+    ITensorPack pack;
+    for (unsigned i = 0; i < _impl->num_inputs; ++i)
     {
-        CLScheduler::get().enqueue(*kernel, true);
+        pack.add_tensor(TensorType::ACL_SRC_VEC + i, _impl->srcs.at(i));
     }
+    pack.add_tensor(TensorType::ACL_DST, _impl->dst);
+
+    _impl->op->run(pack);
 }
 } // namespace arm_compute