Rewrite dynamic fusion

The new version introduces the following major changes:

* Change public interface to simplify and standardize the user experience
    - Use the term "Workload" uniformly
    - Simplify operator interface to be a set of static methods:
      validate_op(), create_op()

* Separate the kernel writing into its own component (template_writer).
  This is to allow the co-development of GpuKernelWriter, and to allow
  easy replacement once GpuKernelWriter is mature.

* Optimize the core fusion algorithm used by the component graph. The
  details can be found in GpuKernelComponentGraph::fuse()

* Use Gpu instead of Cl prefixes for most of the Workload interfaces
  (except for runtime and kernel components, which have to be language specific)
  This allows the potential extension to other Gpu langauges in the
  future.

* Refactor runtime memory interface so that auxiliary tensor handling
  is separate from the user tensor passing. This is because the former
  is less stable and may require extension in the future.

* Hide source code object from the user as it is not required at the
  moment

* Deprecate the old prototype entirely by disabling it in SCons build

Resolves COMPMID-5510, COMPMID-5512, COMPMID-5513

Change-Id: If69d2362856f2de4503546b7b6cf48a525cf3079
Signed-off-by: SiCong Li <sicong.li@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8406
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Reviewed-by: Jakub Sujak <jakub.sujak@arm.com>
Reviewed-by: Viet-Hoa Do <viet-hoa.do@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>

3 files changed, 627 insertions, 0 deletions

diff --git a/src/dynamic_fusion/runtime/gpu/cl/ClKernelRuntime.cpp b/src/dynamic_fusion/runtime/gpu/cl/ClKernelRuntime.cpp
new file mode 100644
index 0000000000..93fbdfed63
--- /dev/null
+++ b/src/dynamic_fusion/runtime/gpu/cl/ClKernelRuntime.cpp
@@ -0,0 +1,200 @@
+/*
+ * Copyright (c) 2022 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 "ClKernelRuntime.h"
+#include "arm_compute/core/CL/ICLTensor.h"
+#include "src/core/CL/CLUtils.h"
+#include "src/dynamic_fusion/sketch/gpu/GpuKernelSourceCode.h"
+#include "src/gpu/cl/ClKernelLibrary.h"
+
+#include "support/Cast.h"
+namespace arm_compute
+{
+namespace experimental
+{
+namespace dynamic_fusion
+{
+using namespace arm_compute::opencl;
+
+void ClKernelRuntime::configure(const ClCompileContext &compile_ctx, const GpuKernelSourceCode &code)
+{
+    // Create kernel from kernel source string
+    opencl::ClKernelLibrary &klib = opencl::ClKernelLibrary::get();
+    _kernel                       = static_cast<cl::Kernel>(compile_ctx.create_kernel(code.name(),
+                                                                                      "" /* Program name: Used to as part of a unique string for built kernel cache. Not needed */,
+                                                                                      code.code(),
+                                                                                      klib.kernel_path() /* Kernel path: Used in cases of embedded kernels */,
+                                                                                      code.build_options().options(),
+                                                                                      false /* Is source binary */));
+
+    // Configure execution window
+    IClKernel::configure_internal(code.window());
+
+    // Set config id for lws tuning
+    _config_id = code.config_id();
+
+    // Set kernel arguments
+    _arguments = code.arguments();
+}
+
+inline void ClKernelRuntime::add_tensor_argument(unsigned int &idx, const GpuKernelArgumentInfo &arg, const ICLTensor *tensor, const Window &arg_slice, std::vector<cl::Image2D> &cl_images)
+{
+    switch(arg.type)
+    {
+        case GpuKernelArgumentInfo::Type::Scalar:
+        {
+            ARM_COMPUTE_ERROR("Unsupported yet");
+            break;
+        }
+
+        case GpuKernelArgumentInfo::Type::Vector:
+        {
+            add_1D_tensor_argument(idx, tensor, arg_slice);
+            break;
+        }
+
+        case GpuKernelArgumentInfo::Type::Image:
+        {
+            add_2D_tensor_argument(idx, tensor, arg_slice);
+            break;
+        }
+        case GpuKernelArgumentInfo::Type::Image_Reinterpret_As_3D:
+        {
+            add_2D_tensor_argument(idx, tensor, arg_slice);
+            const unsigned int total_cross_plane_pad = tensor->info()->padding().top + tensor->info()->padding().bottom;
+            _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(total_cross_plane_pad));
+            break;
+        }
+        case GpuKernelArgumentInfo::Type::Image_Export_To_ClImage2D:
+        {
+            const TensorShape shape2d(tensor->info()->dimension(0) / 4, tensor->info()->dimension(1) * tensor->info()->dimension(2) * tensor->info()->dimension(3));
+            const size_t      image_row_pitch = tensor->info()->strides_in_bytes()[1];
+            cl::Image2D       tensor_image2d  = create_image2d_from_buffer(CLKernelLibrary::get().context(), tensor->cl_buffer(), shape2d, tensor->info()->data_type(), image_row_pitch);
+            cl_images.push_back(tensor_image2d);
+            _kernel.setArg(idx++, tensor_image2d);
+            break;
+        }
+
+        case GpuKernelArgumentInfo::Type::Image_3D:
+        {
+            add_2D_tensor_argument(idx, tensor, arg_slice);
+            _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(tensor->info()->strides_in_bytes()[2]));
+            break;
+        }
+        case GpuKernelArgumentInfo::Type::Image_3D_Export_To_ClImage2D:
+        {
+            const TensorShape shape2d(tensor->info()->dimension(0) / 4, tensor->info()->dimension(1) * tensor->info()->dimension(2) * tensor->info()->dimension(3));
+            const size_t      image_row_pitch = tensor->info()->strides_in_bytes()[1];
+            cl::Image2D       tensor_image2d  = create_image2d_from_buffer(CLKernelLibrary::get().context(), tensor->cl_buffer(), shape2d, tensor->info()->data_type(), image_row_pitch);
+            cl_images.push_back(tensor_image2d);
+            _kernel.setArg(idx++, tensor_image2d);
+            _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(tensor->info()->strides_in_bytes()[2]));
+            break;
+        }
+
+        case GpuKernelArgumentInfo::Type::Tensor_3D:
+        {
+            add_3D_tensor_argument(idx, tensor, arg_slice);
+            break;
+        }
+
+        case GpuKernelArgumentInfo::Type::Tensor_4D:
+        {
+            add_4D_tensor_argument(idx, tensor, arg_slice);
+            break;
+        }
+        case GpuKernelArgumentInfo::Type::Tensor_4D_t_Buffer:
+        {
+            add_4d_tensor_nhwc_argument(idx, tensor);
+            break;
+        }
+        case GpuKernelArgumentInfo::Type::Tensor_4D_t_Image:
+        {
+            const size_t image_w        = tensor->info()->dimension(0) / 4;
+            const size_t image_h        = tensor->info()->tensor_shape().total_size_upper(1);
+            const size_t image_stride_y = tensor->info()->strides_in_bytes()[1];
+
+            cl::Image2D tensor_image2d = create_image2d_from_buffer(CLKernelLibrary::get().context(), tensor->cl_buffer(),
+                                                                    TensorShape(image_w, image_h), tensor->info()->data_type(), image_stride_y);
+            cl_images.push_back(tensor_image2d);
+
+            _kernel.setArg(idx++, tensor_image2d);
+            add_4d_tensor_nhwc_argument(idx, tensor);
+            break;
+        }
+        default:
+        {
+            ARM_COMPUTE_ERROR("Unsupported");
+        }
+    }
+}
+
+void ClKernelRuntime::run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue)
+{
+    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
+
+    Window slice = window.first_slice_window_3D();
+    // Don't slice matrix along the z dimension if matrix has just 2 dimensions and matrix A more than 2
+    // This scenario can happen when the matrix multiplication is used to perform a convolution operation
+    Window slice_fixed_z = slice;
+    slice_fixed_z.set(Window::DimX, Window::Dimension(0, 1, 1));
+    slice_fixed_z.set(Window::DimY, Window::Dimension(0, 1, 1));
+
+    /// NOTE: Parameters extracted from old kernels. So far they seem to be constant
+    /// but we may need to make them into another configuration passed from GpuWorkloadSourceCode if needed in the future
+    constexpr bool slide_along_dimz     = true;
+    constexpr bool skip_sliding_window  = false;
+    constexpr bool use_dummy_work_items = false;
+
+    unsigned int idx = 0;
+    do
+    {
+        // Set kernel arguments
+        Window arg_slice = slice;
+        // CLImages created from tensor arguments. Need to be retained until enqueue
+        std::vector<cl::Image2D> cl_images;
+        for(auto id_arg : _arguments)
+        {
+            const auto arg    = id_arg.second;
+            auto       tensor = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(id_arg.first));
+            ARM_COMPUTE_ERROR_ON_NULLPTR(tensor);
+            ARM_COMPUTE_ERROR_ON_NULLPTR(tensor->info());
+            if(!slide_along_dimz)
+            {
+                // The stride_z for matrix must be zero if we do not slice
+                ARM_COMPUTE_ERROR_ON(tensor->info()->strides_in_bytes()[3] != 0);
+                arg_slice = slice_fixed_z;
+            }
+            add_tensor_argument(idx, *arg.kernel_argument_info(), tensor, arg_slice, cl_images);
+        }
+
+        // Dispatch kernel
+        enqueue(queue, *this, slice, lws_hint(), use_dummy_work_items);
+    }
+    while(skip_sliding_window && window.slide_window_slice_3D(slice));
+}
+
+} // namespace dynamic_fusion
+} // namespace experimental
+} // namespace arm_compute
diff --git a/src/dynamic_fusion/runtime/gpu/cl/ClKernelRuntime.h b/src/dynamic_fusion/runtime/gpu/cl/ClKernelRuntime.h
new file mode 100644
index 0000000000..acc2380031
--- /dev/null
+++ b/src/dynamic_fusion/runtime/gpu/cl/ClKernelRuntime.h
@@ -0,0 +1,76 @@
+/*
+ * Copyright (c) 2022 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.
+ */
+#ifndef SRC_DYNAMIC_FUSION_RUNTIME_GPU_CL_CLKERNELRUNTIME
+#define SRC_DYNAMIC_FUSION_RUNTIME_GPU_CL_CLKERNELRUNTIME
+
+#include "src/dynamic_fusion/sketch/gpu/GpuKernelArgument.h"
+#include "src/dynamic_fusion/sketch/gpu/GpuKernelSourceCode.h"
+#include "src/gpu/cl/ClCompileContext.h"
+#include "src/gpu/cl/IClKernel.h"
+
+namespace arm_compute
+{
+namespace experimental
+{
+namespace dynamic_fusion
+{
+struct GpuKernelSourceCode;
+
+/** OpenCL runtime to run a single kernel */
+class ClKernelRuntime final : public opencl::IClKernel
+{
+public:
+    /** Configure the kernel runtime
+     *
+     * @param[in] compile_ctx OpenCL compile context
+     * @param[in] code        Kernel source code
+     */
+    void configure(const opencl::ClCompileContext &compile_ctx, const GpuKernelSourceCode &code);
+    /** Run the kernel
+     *
+     * @param[in,out] tensors @ref ITensorPack object containing run-time tensor memories
+     * @param[in]     window  Execution window
+     * @param[in]     queue   OpenCL command queue
+     */
+    virtual void run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue) override;
+
+private:
+    /** Set a kernel tensor argument
+     *
+     * @param[in,out] idx       Index at which to start adding the tensor's arguments. Will be incremented by the number of kernel arguments set.
+     * @param[in]     arg       Kernel argument descriptor accompanying @p tensor
+     * @param[in]     tensor    Tensor to set as an argument of the object's kernel
+     * @param[in]     arg_slice Window the kernel will be run on
+     * @param[out]    cl_images Extra cl images created from the tensor (will need to be retained until the kernel is enqueued)
+     */
+    inline void add_tensor_argument(unsigned int &idx, const GpuKernelArgumentInfo &arg, const ICLTensor *tensor, const Window &arg_slice, std::vector<cl::Image2D> &cl_images);
+
+private:
+    GpuKernelArgumentList _arguments{}; /** All kernel arguments required by the runtime */
+};
+
+} // namespace dynamic_fusion
+} // namespace experimental
+} // namespace arm_compute
+#endif /* SRC_DYNAMIC_FUSION_RUNTIME_GPU_CL_CLKERNELRUNTIME */
diff --git a/src/dynamic_fusion/runtime/gpu/cl/ClWorkloadRuntime.cpp b/src/dynamic_fusion/runtime/gpu/cl/ClWorkloadRuntime.cpp
new file mode 100644
index 0000000000..549c6d4abb
--- /dev/null
+++ b/src/dynamic_fusion/runtime/gpu/cl/ClWorkloadRuntime.cpp
@@ -0,0 +1,351 @@
+/*
+ * Copyright (c) 2022 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/dynamic_fusion/runtime/gpu/cl/ClWorkloadRuntime.h"
+
+#include "arm_compute/core/experimental/Types.h"
+#include "arm_compute/runtime/CL/CLTensor.h"
+#include "src/dynamic_fusion/runtime/gpu/cl/ClKernelRuntime.h"
+#include "src/dynamic_fusion/sketch/gpu/GpuWorkloadSketchImpl.h"
+#include "src/dynamic_fusion/sketch/gpu/GpuWorkloadSourceCode.h"
+#include "support/Cast.h"
+
+#include <algorithm>
+
+namespace arm_compute
+{
+namespace experimental
+{
+namespace dynamic_fusion
+{
+namespace
+{
+/** Holder of any auxiliary @ref CLTensor required by a @ref GpuWorkloadSourceCode.
+ *
+ * @note The tensors are not allocated by default, and require the user to explicitly allocate them using the associated @ref TensorInfo and @ref AuxMemoryInfo
+ *
+ * @note This data holder must remain valid until the @ref ClWorkloadRuntime that uses it, is out of scope
+ */
+class ClAuxTensors
+{
+public:
+    /** A view of a single auxiliary data and the associated @ref TensorInfo and @ref AuxMemoryInfo
+     */
+    struct DataView
+    {
+        DataView() = default;
+        DataView(CLTensor *tensor, const TensorInfo &tensor_info, const AuxMemoryInfo &memory_info)
+            : tensor{ tensor }, tensor_info{ tensor_info }, memory_info{ memory_info }
+        {
+        }
+        ~DataView()                     = default;
+        DataView(const DataView &other) = default;
+        DataView &operator=(const DataView &other) = default;
+        DataView(DataView &&other)                 = default;
+        DataView &operator=(DataView &&other) = default;
+        CLTensor     *tensor{};      /**< Pointer to the auxiliary tensor */
+        TensorInfo    tensor_info{}; /**< Associated tensor info */
+        AuxMemoryInfo memory_info{}; /**< Memory requirement */
+    };
+
+    /** Get views of all auxiliary tensors. This is mainly used for allocating the auxiliary tensors. */
+    std::vector<DataView> get_tensors()
+    {
+        return _tensors;
+    }
+    std::vector<DataView> get_tensors() const
+    {
+        return _tensors;
+    }
+
+    friend Status create_aux_tensors(ClAuxTensors *aux_tensors, const GpuWorkloadSourceCode &code);
+
+private:
+    /** Add auxiliary tensor.
+     *
+     * @param[in] tensor_info @ref ITensorInfo of the auxiliary tensor
+     * @param[in] memory_info Memory requirements of the auxiliary tensor
+     *
+     * @return CLTensor*  Corresponding tensor memory if successfully added, otherwise nullptr
+     */
+    CLTensor *add_aux_tensor(const ITensorInfo &tensor_info, const AuxMemoryInfo &aux_memory_info)
+    {
+        const auto t_id             = tensor_info.id();
+        auto       find_tensor_pair = _owned_tensors.find(t_id);
+        if(find_tensor_pair == _owned_tensors.end())
+        {
+            return find_tensor_pair->second.get();
+        }
+        else
+        {
+            auto tensor        = std::make_unique<CLTensor>();
+            auto inserted_pair = _owned_tensors.emplace(t_id, std::move(tensor)).first;
+            auto new_tensor    = inserted_pair->second.get();
+            _tensors.emplace_back(new_tensor, tensor_info, aux_memory_info);
+            return new_tensor;
+        }
+    }
+
+    std::map<ITensorInfo::Id, std::unique_ptr<CLTensor>> _owned_tensors{};
+    std::vector<DataView> _tensors{};
+};
+/** Construct auxiliary tensors required by @ref GpuWorkloadSourceCode
+ *
+ * @note This is the only recommended method for user to create @ref ClAuxTensors
+ *
+ * @param[out] aux_tensors Auxiliary tensors required by the workload code
+ * @param[in]  code        @ref GpuWorkloadSourceCode which all tensors bind to
+ *
+ * @return Status
+ */
+Status create_aux_tensors(ClAuxTensors *aux_tensors, const GpuWorkloadSourceCode &code)
+{
+    for(auto t_id : code.tensors())
+    {
+        // Get tensor object
+        const auto workload_arg  = code.query_tensor(t_id);
+        ICLTensor *tensor_object = nullptr;
+        if(workload_arg->memory_descriptor()->memory_type == MemoryType::Auxiliary)
+        {
+            // Create aux tensor CLTensor object
+            const TensorInfo tensor_info = *workload_arg->tensor_info();
+            ARM_COMPUTE_ERROR_ON(tensor_info.id() != t_id);
+            const auto aux_memory_info = workload_arg->memory_descriptor()->aux_memory_info;
+            tensor_object              = aux_tensors->add_aux_tensor(tensor_info, aux_memory_info);
+        }
+        if(tensor_object == nullptr)
+        {
+            return ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Failed to construct an auxiliary tensor");
+        }
+    }
+    return Status{};
+}
+
+/** A fast tensor lookup table for runtime tensor objects retrieval
+ */
+class ClTensorLUT
+{
+public:
+    /** Find a tensor pack associated with the @ref UnitWorkloadId @p uwk_id
+     *
+     * @param[in] uwk_id @ref UnitWorkloadId associated with the tensor pack
+     *
+     * @return ITensorPack*
+     */
+    ITensorPack *find_tensor_pack(UnitWorkloadId uwk_id)
+    {
+        auto tensor_pack = _tensor_packs.find(uwk_id);
+        if(tensor_pack != _tensor_packs.end())
+        {
+            return &(tensor_pack->second);
+        }
+        return nullptr;
+    }
+    /** Get a tensor pack associated with @p uwk_id. Throws a exception if it cannot be found.
+     *
+     * @param[in] uwk_id @ref UnitWorkloadId associated with the tensor pack
+     *
+     * @return ITensorPack*
+     */
+    ITensorPack &get_tensor_pack(UnitWorkloadId uwk_id)
+    {
+        return _tensor_packs.at(uwk_id);
+    }
+
+    friend Status create_tensor_lut(ClTensorLUT *tensor_lut, const GpuWorkloadSourceCode &code, const std::vector<CLTensor *> &user_tensors, const ClAuxTensors &aux_tensors);
+
+private:
+    /** Add a tensor pack and associate it with @ref UnitWorkloadId @p uwk_id
+     *
+     * @param[in] uwk_id      @ref UnitWorkloadId associated with the tensor pack
+     * @param[in] tensor_pack Tensor pack to be added
+     */
+    void add_tensor_pack(UnitWorkloadId uwk_id, const ITensorPack &tensor_pack)
+    {
+        _tensor_packs[uwk_id] = tensor_pack;
+    }
+    std::map<UnitWorkloadId, ITensorPack> _tensor_packs{};
+};
+
+/** Create a fast tensor lookup table for runtime tensor retrieval
+ *
+ * @param[out] tensor_lut   @ref ClTensorLUT used by the runtime to feed tensor memories to underlying kernels
+ * @param[in]  code         @ref GpuWorkloadSourceCode which all tensors bind to
+ * @param[in]  user_tensors User tensors
+ * @param[in]  aux_tensors  Auxiliary tensors required by the workload code
+ *
+ * @return Status
+ */
+Status create_tensor_lut(ClTensorLUT *tensor_lut, const GpuWorkloadSourceCode &code, const std::vector<CLTensor *> &user_tensors, const ClAuxTensors &aux_tensors)
+{
+    // Combine user tensors and aux tensors
+    std::map<ITensorInfo::Id, CLTensor *> tensor_map;
+    for(auto tensor : user_tensors)
+    {
+        const auto t_id = tensor->info()->id();
+        if(tensor_map.find(t_id) != tensor_map.end())
+        {
+            return ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Clashing tensor ids");
+        }
+        tensor_map[t_id] = tensor;
+    }
+    for(const auto &data : aux_tensors.get_tensors())
+    {
+        const auto t_id   = data.tensor_info.id();
+        const auto tensor = data.tensor;
+        if(tensor_map.find(t_id) != tensor_map.end())
+        {
+            return ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Clashing tensor ids");
+        }
+        tensor_map[t_id] = tensor;
+    }
+
+    // Add tensor objects into corresponding tensor packs
+    for(auto id_tensor : tensor_map)
+    {
+        const auto t_id          = id_tensor.first;
+        const auto tensor_object = id_tensor.second;
+        if(tensor_object == nullptr)
+        {
+            return ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Trying to add a nullptr into the tensor packs");
+        }
+        if(tensor_object->allocator()->info().total_size() == 0U)
+        {
+            return ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "No allocated memory found in tensor");
+        }
+
+        for(auto uwk_id : code.get_unit_workloads_from_tensor(t_id))
+        {
+            ITensorPack *tensor_pack = tensor_lut->find_tensor_pack(uwk_id);
+            if(tensor_pack == nullptr)
+            {
+                tensor_lut->add_tensor_pack(uwk_id, ITensorPack{ { t_id, tensor_object } });
+            }
+            else
+            {
+                tensor_pack->add_tensor(t_id, tensor_object);
+            }
+        }
+    }
+    return Status{};
+}
+
+} // namespace
+
+struct ClWorkloadRuntime::Implementation
+{
+    std::map<UnitWorkloadId, std::unique_ptr<ClKernelRuntime>> _kernels{};
+    std::map<UnitWorkloadId, std::unique_ptr<ClKernelRuntime>> _kernels_prep{};
+    bool                  _is_configured{ false };
+    bool                  _is_prepared{ false };
+    ClTensorLUT           _tensor_lut{};
+    ClAuxTensors          _aux_tensors{};
+    GpuWorkloadSourceCode _source_code{};
+};
+
+ClWorkloadRuntime::ClWorkloadRuntime()
+    : _impl{ std::make_unique<Implementation>() }
+{
+}
+
+ClWorkloadRuntime::~ClWorkloadRuntime() = default;
+
+Status ClWorkloadRuntime::configure(const GpuWorkloadSketch &sketch)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(_impl->_is_configured, "ClWorkloadRuntime cannot be re-configured");
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(sketch.gpu_context()->gpu_language() != GpuLanguage::OpenCL, "ClWorkloadRuntime cannot be configured with non-OpenCL workload sketch");
+    // Generate source code
+    _impl->_source_code = sketch.implementation().generate_source_code();
+    // Configure unit workload from source code
+    for(auto uwk_id : _impl->_source_code.unit_workloads())
+    {
+        const auto work  = _impl->_source_code.query_unit_workload(uwk_id);
+        const auto stage = work.stage().stage;
+        auto       k     = std::make_unique<ClKernelRuntime>();
+        k->configure(*sketch.gpu_context()->cl_compile_context(), work.code());
+
+        switch(stage)
+        {
+            case UnitWorkloadStage::Stage::Run:
+                _impl->_kernels.emplace(work.id(), std::move(k));
+                break;
+            case UnitWorkloadStage::Stage::Prepare:
+                _impl->_kernels_prep.emplace(work.id(), std::move(k));
+                break;
+            default:
+                ARM_COMPUTE_ERROR("Invalid unit workload stage");
+        }
+        break;
+    }
+    // Create auxiliary tensor objects
+    create_aux_tensors(&_impl->_aux_tensors, _impl->_source_code);
+    _impl->_is_configured = true;
+    return Status{};
+}
+
+void ClWorkloadRuntime::prepare()
+{
+    if(!_impl->_is_prepared)
+    {
+        for(auto &id_kernel_pair : _impl->_kernels_prep)
+        {
+            const bool flush_queue = false;
+            const auto uwk_id      = id_kernel_pair.first;
+            auto       kernel      = id_kernel_pair.second.get();
+            CLScheduler::get().enqueue_op(*kernel, _impl->_tensor_lut.get_tensor_pack(uwk_id), flush_queue);
+        }
+
+        _impl->_is_prepared = true;
+    }
+}
+
+Status ClWorkloadRuntime::run(const std::vector<CLTensor *> &tensors)
+{
+    // Need to create the tensor lut in every run, unless the user can guarantee the binding remains fixed,
+    // in which case the lut can be cached during prepare
+    const auto st = create_tensor_lut(&_impl->_tensor_lut, _impl->_source_code, tensors, _impl->_aux_tensors);
+    ARM_COMPUTE_RETURN_ON_ERROR(st);
+    prepare();
+    for(auto &id_kernel_pair : _impl->_kernels)
+    {
+        // Flush the command queue on the last kernel
+        const bool flush_queue = false;
+        const auto uwk_id      = id_kernel_pair.first;
+        auto       kernel      = id_kernel_pair.second.get();
+        CLScheduler::get().enqueue_op(*kernel, _impl->_tensor_lut.get_tensor_pack(uwk_id), flush_queue);
+    }
+    return Status{};
+}
+
+std::vector<std::pair<CLTensor *, AuxMemoryInfo>> ClWorkloadRuntime::get_auxiliary_tensors()
+{
+    std::vector<std::pair<CLTensor *, AuxMemoryInfo>> aux_tensors;
+    for(const auto &data : _impl->_aux_tensors.get_tensors())
+    {
+        aux_tensors.emplace_back(data.tensor, data.memory_info);
+    }
+    return aux_tensors;
+}
+} // namespace dynamic_fusion
+} // namespace experimental
+} // namespace arm_compute