Port CLCol2ImKernel to ClCol2ImKernel

Resolves: COMPMID-4517

Change-Id: I50cb02116a1ab86fc29200371944c4774e830746
Signed-off-by: Manuel Bottini <manuel.bottini@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/5949
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>

12 files changed, 220 insertions, 181 deletions

diff --git a/Android.bp b/Android.bp
index 0778b636ce..83e1e0a338 100644
--- a/Android.bp
+++ b/Android.bp
@@ -86,7 +86,6 @@ cc_library_static {
         "src/core/CL/kernels/CLBitwiseKernel.cpp",
         "src/core/CL/kernels/CLBoundingBoxTransformKernel.cpp",
         "src/core/CL/kernels/CLChannelShuffleLayerKernel.cpp",
-        "src/core/CL/kernels/CLCol2ImKernel.cpp",
         "src/core/CL/kernels/CLComparisonKernel.cpp",
         "src/core/CL/kernels/CLDeconvolutionLayerUpsampleKernel.cpp",
         "src/core/CL/kernels/CLDeconvolutionReshapeOutputKernel.cpp",
@@ -330,6 +329,7 @@ cc_library_static {
         "src/core/gpu/cl/kernels/ClActivationKernel.cpp",
         "src/core/gpu/cl/kernels/ClBatchConcatenateKernel.cpp",
         "src/core/gpu/cl/kernels/ClCastKernel.cpp",
+        "src/core/gpu/cl/kernels/ClCol2ImKernel.cpp",
         "src/core/gpu/cl/kernels/ClConvertFullyConnectedWeightsKernel.cpp",
         "src/core/gpu/cl/kernels/ClCopyKernel.cpp",
         "src/core/gpu/cl/kernels/ClCropKernel.cpp",
diff --git a/arm_compute/runtime/CL/functions/CLGEMMConvolutionLayer.h b/arm_compute/runtime/CL/functions/CLGEMMConvolutionLayer.h
index 082b481047..564fb1ecde 100644
--- a/arm_compute/runtime/CL/functions/CLGEMMConvolutionLayer.h
+++ b/arm_compute/runtime/CL/functions/CLGEMMConvolutionLayer.h
@@ -41,7 +41,13 @@
 
 namespace arm_compute
 {
-class CLCol2ImKernel;
+namespace opencl
+{
+namespace kernels
+{
+class ClCol2ImKernel;
+} // namespace kernels
+} // namespace opencl
 class CLIm2ColKernel;
 class CLWeightsReshapeKernel;
 class ICLTensor;
@@ -171,7 +177,7 @@ private:
  * -# @ref CLGEMM (if the data type is FP32 or FP16)
  * -# @ref CLGEMMLowpMatrixMultiplyCore (if the data type is QASYMM8/QASYMM8_SIGNED)
  * -# @ref CLGEMMLowpOutputStage with QUANTIZE_DOWN_FIXEDPOINT type of quantization (if the data type is QASYMM8/QASYMM8_SIGNED)
- * -# @ref CLCol2ImKernel (if NCHW data layout)
+ * -# @ref opencl::kernels::ClCol2ImKernel (if NCHW data layout)
  */
 class CLGEMMConvolutionLayer : public IFunction
 {
@@ -318,10 +324,12 @@ private:
     std::unique_ptr<CLIm2ColKernel>                                    _im2col_kernel;
     CLGEMM                                                             _mm_gemm;
     CLGEMMLowpMatrixMultiplyCore                                       _mm_gemmlowp;
-    std::unique_ptr<CLCol2ImKernel>                                    _col2im_kernel;
+    std::unique_ptr<opencl::kernels::ClCol2ImKernel>                   _col2im_kernel;
     CLActivationLayer                                                  _activationlayer_function;
 
     const ICLTensor *_original_weights;
+    const ICLTensor *_gemm_output_to_use;
+    ICLTensor       *_output;
 
     CLTensor _im2col_output;
     CLTensor _weights_reshaped;
diff --git a/docs/user_guide/release_version_and_change_log.dox b/docs/user_guide/release_version_and_change_log.dox
index 93fbfb01af..d8258a10cd 100644
--- a/docs/user_guide/release_version_and_change_log.dox
+++ b/docs/user_guide/release_version_and_change_log.dox
@@ -1384,7 +1384,7 @@ v17.02.1 Sources preview
  - New OpenCL kernels / functions:
    - CLLogits1DMaxKernel, CLLogits1DShiftExpSumKernel, CLLogits1DNormKernel / @ref CLSoftmaxLayer
    - CLPoolingLayerKernel / @ref CLPoolingLayer
-   - @ref CLIm2ColKernel, @ref CLCol2ImKernel, CLConvolutionLayerWeightsReshapeKernel / CLConvolutionLayer
+   - @ref CLIm2ColKernel, CLCol2ImKernel, CLConvolutionLayerWeightsReshapeKernel / CLConvolutionLayer
    - @ref CLRemapKernel / @ref CLRemap
    - CLGaussianPyramidHorKernel, CLGaussianPyramidVertKernel / CLGaussianPyramid, CLGaussianPyramidHalf, CLGaussianPyramidOrb
    - CLMinMaxKernel, CLMinMaxLocationKernel / CLMinMaxLocation
diff --git a/filelist.json b/filelist.json
index 3b549b2483..6babf70eab 100644
--- a/filelist.json
+++ b/filelist.json
@@ -357,7 +357,7 @@
       "GEMMConv2d": {
         "files": {
           "kernel": [
-            "src/core/CL/kernels/CLCol2ImKernel.cpp",
+            "src/core/gpu/cl/kernels/ClCol2ImKernel.cpp",
             "src/core/CL/kernels/CLIm2ColKernel.cpp"
           ]
         }
diff --git a/src/core/CL/CLKernels.h b/src/core/CL/CLKernels.h
index 09f8109445..d2d5b928bc 100644
--- a/src/core/CL/CLKernels.h
+++ b/src/core/CL/CLKernels.h
@@ -31,7 +31,6 @@
 #include "src/core/CL/kernels/CLBitwiseKernel.h"
 #include "src/core/CL/kernels/CLBoundingBoxTransformKernel.h"
 #include "src/core/CL/kernels/CLChannelShuffleLayerKernel.h"
-#include "src/core/CL/kernels/CLCol2ImKernel.h"
 #include "src/core/CL/kernels/CLComparisonKernel.h"
 #include "src/core/CL/kernels/CLDeconvolutionLayerUpsampleKernel.h"
 #include "src/core/CL/kernels/CLDeconvolutionReshapeOutputKernel.h"
diff --git a/src/core/CL/kernels/CLCol2ImKernel.h b/src/core/CL/kernels/CLCol2ImKernel.h
deleted file mode 100644
index 710e048bca..0000000000
--- a/src/core/CL/kernels/CLCol2ImKernel.h
+++ /dev/null
@@ -1,106 +0,0 @@
-/*
- * Copyright (c) 2017-2020 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 ARM_COMPUTE_CLCOL2IMKERNEL_H
-#define ARM_COMPUTE_CLCOL2IMKERNEL_H
-
-#include "src/core/CL/ICLKernel.h"
-
-namespace arm_compute
-{
-class ICLTensor;
-
-/** Interface for the col2im reshaping kernel.
- *
- * Rearranges each matrix column into image blocks. It's the inverse operation of @ref CLIm2ColKernel.
- *
- * For example, a vector of 9 elements can be reshaped to a block(image) of 3x3:
- *
- * @f[
- * \left( \begin{array}{ccccccccc}
- * a0 & a1 & a2 & a3 & a4 & a5 & a6 & a7 & a8 \\
- * \end{array} \right)
- * \rightarrow
- * \left( \begin{array}{ccc}
- * a0 & a1 & a2 \\
- * a3 & a4 & a5 \\
- * a6 & a7 & a8 \\
- * \end{array} \right)
- * @f]
- */
-class CLCol2ImKernel : public ICLKernel
-{
-public:
-    /** Default constructor */
-    CLCol2ImKernel();
-    /** Prevent instances of this class from being copied (As this class contains pointers) */
-    CLCol2ImKernel(const CLCol2ImKernel &) = delete;
-    /** Prevent instances of this class from being copied (As this class contains pointers) */
-    CLCol2ImKernel &operator=(const CLCol2ImKernel &) = delete;
-    /** Allow instances of this class to be moved */
-    CLCol2ImKernel(CLCol2ImKernel &&) = default;
-    /** Allow instances of this class to be moved */
-    CLCol2ImKernel &operator=(CLCol2ImKernel &&) = default;
-    /** Default destructor */
-    ~CLCol2ImKernel() = default;
-    /** Set the input and output of the kernel.
-     *
-     * @param[in]  input          The input tensor to convert. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32
-     * @param[out] output         The output tensor. 3 lower dimensions represent a single output [width, height, OFM],
-     *                            while the rest represent batch of outputs. Data types supported: Same as @p input. Data layout: NCHW
-     * @param[in]  convolved_dims Output convolved dimensions.
-     * @param[in]  num_groups     (Optional) Number of groups when performing a grouped convolution
-     */
-    void configure(const ICLTensor *input, ICLTensor *output, const Size2D &convolved_dims, unsigned int num_groups = 1);
-    /** Set the input and output of the kernel.
-     *
-     * @param[in]  compile_context The compile context to be used.
-     * @param[in]  input           The input tensor to convert. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32
-     * @param[out] output          The output tensor. 3 lower dimensions represent a single output [width, height, OFM],
-     *                             while the rest represent batch of outputs. Data types supported: Same as @p input. Data layout: NCHW
-     * @param[in]  convolved_dims  Output convolved dimensions.
-     * @param[in]  num_groups      (Optional) Number of groups when performing a grouped convolution
-     */
-    void configure(const CLCompileContext &compile_context, const ICLTensor *input, ICLTensor *output, const Size2D &convolved_dims, unsigned int num_groups = 1);
-    /** Static function to check if given info will lead to a valid configuration of @ref CLCol2ImKernel
-     *
-     * @param[in] input          The input tensor to convert. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32
-     * @param[in] output         The output tensor. 3 lower dimensions represent a single output [width, height, OFM],
-     *                           while the rest represent batch of outputs. Data types supported: Same as @p input. Data layout: NCHW
-     * @param[in] convolved_dims Output convolved dimensions.
-     * @param[in] num_groups     (Optional) Number of groups when performing a grouped convolution
-     *
-     * @return a status
-     */
-    static Status validate(const ITensorInfo *input, const ITensorInfo *output, const Size2D &convolved_dims, unsigned int num_groups = 1);
-
-    // Inherited methods overridden:
-    void run(const Window &window, cl::CommandQueue &queue) override;
-
-public:
-    const ICLTensor *_input;
-    ICLTensor       *_output;
-    Size2D           _convolved_dims;
-};
-} // namespace arm_compute
-#endif /*ARM_COMPUTE_CLCOL2IMKERNEL_H */
diff --git a/src/core/CL/kernels/CLCol2ImKernel.cpp b/src/core/gpu/cl/kernels/ClCol2ImKernel.cpp
index 6d7b83471f..a3d57115f9 100644
--- a/src/core/CL/kernels/CLCol2ImKernel.cpp
+++ b/src/core/gpu/cl/kernels/ClCol2ImKernel.cpp
@@ -21,7 +21,7 @@
  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  * SOFTWARE.
  */
-#include "src/core/CL/kernels/CLCol2ImKernel.h"
+#include "src/core/gpu/cl/kernels/ClCol2ImKernel.h"
 
 #include "arm_compute/core/CL/CLHelpers.h"
 #include "arm_compute/core/CL/CLKernelLibrary.h"
@@ -32,47 +32,51 @@
 #include "src/core/CL/CLValidate.h"
 #include "src/core/helpers/AutoConfiguration.h"
 #include "src/core/helpers/WindowHelpers.h"
+#include "support/Cast.h"
 #include "support/StringSupport.h"
 
 #include <cmath>
 
-using namespace arm_compute::misc::shape_calculator;
-
 namespace arm_compute
 {
+using namespace misc::shape_calculator;
+namespace opencl
+{
+namespace kernels
+{
 namespace
 {
-Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const Size2D &convolved_dims, unsigned int num_groups)
+Status validate_arguments(const ITensorInfo *src, const ITensorInfo *dst, const Size2D &convolved_dims, unsigned int num_groups)
 {
-    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::QASYMM8_SIGNED, DataType::F16, DataType::F32);
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, dst);
+    ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(src);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::F16, DataType::F32);
 
     // Checks performed when output is configured
-    if(output->total_size() != 0)
+    if(dst->total_size() != 0)
     {
-        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), compute_col2im_shape(*input, convolved_dims, true, num_groups));
-        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
-        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(input, output);
-        ARM_COMPUTE_RETURN_ERROR_ON_MSG(output->data_layout() != DataLayout::NCHW, "Col2Im output's data layout must always be NCHW");
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(dst->tensor_shape(), compute_col2im_shape(*src, convolved_dims, true, num_groups));
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, dst);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(src, dst);
+        ARM_COMPUTE_RETURN_ERROR_ON_MSG(dst->data_layout() != DataLayout::NCHW, "Col2Im output's data layout must always be NCHW");
     }
 
     return Status{};
 }
 
-std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *output, const Size2D &convolved_dims, unsigned int num_groups)
+std::pair<Status, Window> validate_and_configure_window(ITensorInfo *src, ITensorInfo *dst, const Size2D &convolved_dims, unsigned int num_groups)
 {
-    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
+    ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst);
     // Output auto inizialitation if not yet initialized
-    auto_init_if_empty(*output, input->clone()->set_tensor_shape(compute_col2im_shape(*input, convolved_dims, true, num_groups)).set_data_layout(DataLayout::NCHW));
+    auto_init_if_empty(*dst, src->clone()->set_tensor_shape(compute_col2im_shape(*src, convolved_dims, true, num_groups)).set_data_layout(DataLayout::NCHW));
 
     constexpr unsigned int num_elems_read_per_iteration = 8;
 
     // Configure window
-    Window win = calculate_max_window(*input, Steps(num_elems_read_per_iteration));
+    Window win = calculate_max_window(*src, Steps(num_elems_read_per_iteration));
 
     // Update window and padding just for the input tensor as we cannot access out-of-bounds elements in the output one
-    AccessWindowHorizontal input_access(input, 0, num_elems_read_per_iteration);
+    AccessWindowHorizontal input_access(src, 0, num_elems_read_per_iteration);
     bool                   window_changed = update_window_and_padding(win, input_access);
 
     Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
@@ -80,78 +84,74 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITen
 }
 } // namespace
 
-CLCol2ImKernel::CLCol2ImKernel()
-    : _input(nullptr), _output(nullptr), _convolved_dims()
+ClCol2ImKernel::ClCol2ImKernel()
+    : _convolved_dims()
 {
     _type = CLKernelType::ELEMENTWISE;
 }
 
-void CLCol2ImKernel::configure(const ICLTensor *input, ICLTensor *output, const Size2D &convolved_dims, unsigned int num_groups)
+void ClCol2ImKernel::configure(const CLCompileContext &compile_context, ITensorInfo *src, ITensorInfo *dst, const Size2D &convolved_dims, unsigned int num_groups)
 {
-    configure(CLKernelLibrary::get().get_compile_context(), input, output, convolved_dims, num_groups);
-}
-
-void CLCol2ImKernel::configure(const CLCompileContext &compile_context, const ICLTensor *input, ICLTensor *output, const Size2D &convolved_dims, unsigned int num_groups)
-{
-    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
+    ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst);
 
     // Perform validation step
-    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), convolved_dims, num_groups));
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src, dst, convolved_dims, num_groups));
 
-    _input          = input;
-    _output         = output;
     _convolved_dims = convolved_dims;
 
-    const DataType data_type = input->info()->data_type();
+    const DataType data_type = src->data_type();
 
     // Create kernel
     CLBuildOptions build_opts;
     build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(data_type));
-    build_opts.add_option("-DELEMENT_SIZE=" + support::cpp11::to_string(input->info()->element_size()));
-    build_opts.add_option("-DWIDTH_INPUT=" + support::cpp11::to_string(input->info()->dimension(0)));
+    build_opts.add_option("-DELEMENT_SIZE=" + support::cpp11::to_string(src->element_size()));
+    build_opts.add_option("-DWIDTH_INPUT=" + support::cpp11::to_string(src->dimension(0)));
     build_opts.add_option("-DWIDTH_OUTPUT=" + support::cpp11::to_string(_convolved_dims.width));
     build_opts.add_option("-DNUM_GROUPS=" + support::cpp11::to_string(num_groups));
 
     _kernel = create_kernel(compile_context, "col2im", build_opts.options());
 
     // Configure kernel window
-    auto win_config = validate_and_configure_window(input->info(), output->info(), _convolved_dims, num_groups);
+    auto win_config = validate_and_configure_window(src, dst, _convolved_dims, num_groups);
     ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
-    ICLKernel::configure_internal(win_config.second);
+    IClKernel::configure_internal(win_config.second);
 
     // Set config_id for enabling LWS tuning
     _config_id = "col2im_";
-    _config_id += lower_string(string_from_data_type(input->info()->data_type()));
+    _config_id += lower_string(string_from_data_type(src->data_type()));
     _config_id += "_";
     _config_id += support::cpp11::to_string(num_groups);
     _config_id += "_";
-    _config_id += support::cpp11::to_string(input->info()->dimension(0));
+    _config_id += support::cpp11::to_string(src->dimension(0));
     _config_id += "_";
-    _config_id += support::cpp11::to_string(input->info()->dimension(1));
+    _config_id += support::cpp11::to_string(src->dimension(1));
     _config_id += "_";
-    _config_id += support::cpp11::to_string(output->info()->dimension(0));
+    _config_id += support::cpp11::to_string(dst->dimension(0));
     _config_id += "_";
-    _config_id += support::cpp11::to_string(output->info()->dimension(1));
+    _config_id += support::cpp11::to_string(dst->dimension(1));
 }
 
-Status CLCol2ImKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const Size2D &convolved_dims, unsigned int num_groups)
+Status ClCol2ImKernel::validate(const ITensorInfo *src, const ITensorInfo *dst, const Size2D &convolved_dims, unsigned int num_groups)
 {
-    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
-    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, convolved_dims, num_groups));
-    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), output->clone().get(), convolved_dims, num_groups).first);
+    ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst);
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src, dst, convolved_dims, num_groups));
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(src->clone().get(), dst->clone().get(), convolved_dims, num_groups).first);
     return Status{};
 }
 
-void CLCol2ImKernel::run(const Window &window, cl::CommandQueue &queue)
+void ClCol2ImKernel::run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue)
 {
     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
-    ARM_COMPUTE_ERROR_ON_MISMATCHING_WINDOWS(ICLKernel::window(), window);
+    ARM_COMPUTE_ERROR_ON_MISMATCHING_WINDOWS(IClKernel::window(), window);
 
     bool is_collapsed     = false;
     bool is_collapsed_out = false;
 
+    auto src = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC));
+    auto dst = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST));
+
     Window out_window;
-    out_window.use_tensor_dimensions(_output->info()->tensor_shape());
+    out_window.use_tensor_dimensions(dst->info()->tensor_shape());
 
     Window collapsed     = window.collapse_if_possible(ICLKernel::window(), Window::DimZ, &is_collapsed);
     Window collapsed_out = out_window.collapse_if_possible(out_window, 3, &is_collapsed_out);
@@ -164,10 +164,12 @@ void CLCol2ImKernel::run(const Window &window, cl::CommandQueue &queue)
     {
         // Set inputs
         unsigned int idx = 0;
-        add_3D_tensor_argument(idx, _input, slice);
-        add_4D_tensor_argument(idx, _output, slice_out);
+        add_3D_tensor_argument(idx, src, slice);
+        add_4D_tensor_argument(idx, dst, slice_out);
         enqueue(queue, *this, slice, lws_hint());
     }
     while(collapsed.slide_window_slice_3D(slice) && collapsed_out.slide_window_slice_4D(slice_out));
 }
+} // namespace kernels
+} // namespace opencl
 } // namespace arm_compute
diff --git a/src/core/gpu/cl/kernels/ClCol2ImKernel.h b/src/core/gpu/cl/kernels/ClCol2ImKernel.h
new file mode 100644
index 0000000000..42d0a96075
--- /dev/null
+++ b/src/core/gpu/cl/kernels/ClCol2ImKernel.h
@@ -0,0 +1,89 @@
+/*
+ * Copyright (c) 2017-2021 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 ARM_COMPUTE_CL_COL2IM_KERNEL_H
+#define ARM_COMPUTE_CL_COL2IM_KERNEL_H
+
+#include "arm_compute/core/Size2D.h"
+#include "src/core/common/Macros.h"
+#include "src/core/gpu/cl/ClCompileContext.h"
+#include "src/core/gpu/cl/IClKernel.h"
+
+namespace arm_compute
+{
+namespace opencl
+{
+namespace kernels
+{
+/** Interface for the col2im reshaping kernel.
+ *
+ * Rearranges each matrix column into image blocks. It's the inverse operation of @ref CLIm2ColKernel.
+ *
+ * For example, a vector of 9 elements can be reshaped to a block(image) of 3x3:
+ *
+ * @f[
+ * \left( \begin{array}{ccccccccc}
+ * a0 & a1 & a2 & a3 & a4 & a5 & a6 & a7 & a8 \\
+ * \end{array} \right)
+ * \rightarrow
+ * \left( \begin{array}{ccc}
+ * a0 & a1 & a2 \\
+ * a3 & a4 & a5 \\
+ * a6 & a7 & a8 \\
+ * \end{array} \right)
+ * @f]
+ */
+class ClCol2ImKernel : public IClKernel
+{
+public:
+    /** Default constructor */
+    ClCol2ImKernel();
+    ARM_COMPUTE_DISALLOW_COPY_ALLOW_MOVE(ClCol2ImKernel);
+    /** Set the input and output of the kernel.
+     *
+     * @param[in]  compile_context The compile context to be used.
+     * @param[in]  src             The input tensor info to convert. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32
+     * @param[out] dst             The output tensor info. 3 lower dimensions represent a single output [width, height, OFM],
+     *                             while the rest represent batch of outputs. Data types supported: Same as @p input. Data layout: NCHW
+     * @param[in]  convolved_dims  Output convolved dimensions.
+     * @param[in]  num_groups      (Optional) Number of groups when performing a grouped convolution
+     */
+    void configure(const CLCompileContext &compile_context, ITensorInfo *src, ITensorInfo *dst, const Size2D &convolved_dims, unsigned int num_groups = 1);
+    /** Static function to check if given info will lead to a valid configuration
+     *
+     * Similar to ClCol2ImKernel::configure()
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *src, const ITensorInfo *dst, const Size2D &convolved_dims, unsigned int num_groups = 1);
+
+    // Inherited methods overridden:
+    void run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue) override;
+
+public:
+    Size2D _convolved_dims;
+};
+} // namespace kernels
+} // namespace opencl
+} // namespace arm_compute
+#endif /*ARM_COMPUTE_CL_COL2IM_KERNEL_H */
diff --git a/src/runtime/CL/functions/CLGEMMConvolutionLayer.cpp b/src/runtime/CL/functions/CLGEMMConvolutionLayer.cpp
index cef8ad5a0d..f0f45a8659 100644
--- a/src/runtime/CL/functions/CLGEMMConvolutionLayer.cpp
+++ b/src/runtime/CL/functions/CLGEMMConvolutionLayer.cpp
@@ -30,9 +30,9 @@
 #include "arm_compute/core/utils/misc/ShapeCalculator.h"
 #include "arm_compute/core/utils/quantization/AsymmHelpers.h"
 #include "arm_compute/runtime/CL/CLScheduler.h"
-#include "src/core/CL/kernels/CLCol2ImKernel.h"
 #include "src/core/CL/kernels/CLIm2ColKernel.h"
 #include "src/core/CL/kernels/CLWeightsReshapeKernel.h"
+#include "src/core/gpu/cl/kernels/ClCol2ImKernel.h"
 #include "src/core/helpers/AutoConfiguration.h"
 #include "support/Cast.h"
 
@@ -106,8 +106,8 @@ void CLConvolutionLayerReshapeWeights::run()
 
 CLGEMMConvolutionLayer::CLGEMMConvolutionLayer(std::shared_ptr<IMemoryManager> memory_manager, IWeightsManager *weights_manager)
     : _memory_group(memory_manager), _weights_manager(weights_manager), _reshape_weights(), _reshape_weights_managed(), _im2col_kernel(std::make_unique<CLIm2ColKernel>()), _mm_gemm(memory_manager,
-            weights_manager), _mm_gemmlowp(memory_manager), _col2im_kernel(std::make_unique<CLCol2ImKernel>()), _activationlayer_function(), _original_weights(nullptr), _im2col_output(), _weights_reshaped(),
-      _gemm_output(), _skip_im2col(false), _skip_col2im(false), _is_quantized(false), _fuse_activation(true), _is_prepared(false)
+            weights_manager), _mm_gemmlowp(memory_manager), _col2im_kernel(nullptr), _activationlayer_function(), _original_weights(nullptr), _gemm_output_to_use(nullptr), _output(nullptr), _im2col_output(),
+      _weights_reshaped(), _gemm_output(), _skip_im2col(false), _skip_col2im(false), _is_quantized(false), _fuse_activation(true), _is_prepared(false)
 {
 }
 
@@ -238,7 +238,6 @@ void CLGEMMConvolutionLayer::configure(const CLCompileContext &compile_context,
 
     // Set the GPU target for im2col and col2im
     _im2col_kernel->set_target(CLScheduler::get().target());
-    _col2im_kernel->set_target(CLScheduler::get().target());
 
     const ICLTensor *gemm_input_to_use  = input;
     ICLTensor       *gemm_output_to_use = output;
@@ -395,9 +394,14 @@ void CLGEMMConvolutionLayer::configure(const CLCompileContext &compile_context,
 
     if(!_skip_col2im)
     {
+        // Set the GPU target for col2im
+        _col2im_kernel = std::make_unique<opencl::kernels::ClCol2ImKernel>();
+        _col2im_kernel->set_target(CLScheduler::get().target());
         // Configure and tune Col2Im
-        _col2im_kernel->configure(compile_context, gemm_output_to_use, output, Size2D(conv_w, conv_h), num_groups);
+        _col2im_kernel->configure(compile_context, gemm_output_to_use->info(), output->info(), Size2D(conv_w, conv_h), num_groups);
         CLScheduler::get().tune_kernel_static(*_col2im_kernel.get());
+        _gemm_output_to_use = gemm_output_to_use;
+        _output             = output;
     }
 
     if(!_skip_col2im)
@@ -595,7 +599,7 @@ Status CLGEMMConvolutionLayer::validate(const ITensorInfo *input, const ITensorI
     // Validate Col2Im
     if(!skip_col2im)
     {
-        ARM_COMPUTE_RETURN_ON_ERROR(CLCol2ImKernel::validate(gemm_output_to_use, output, Size2D(conv_w, conv_h), num_groups));
+        ARM_COMPUTE_RETURN_ON_ERROR(opencl::kernels::ClCol2ImKernel::validate(gemm_output_to_use, output, Size2D(conv_w, conv_h), num_groups));
     }
 
     //Validate Activation Layer
@@ -634,7 +638,12 @@ void CLGEMMConvolutionLayer::run()
     // Reshape output matrix
     if(!_skip_col2im)
     {
-        CLScheduler::get().enqueue(*_col2im_kernel.get(), false);
+        ITensorPack pack =
+        {
+            { TensorType::ACL_SRC, _gemm_output_to_use },
+            { TensorType::ACL_DST, _output }
+        };
+        CLScheduler::get().enqueue_op(*_col2im_kernel.get(), pack, false);
     }
 
     //Run Activation Layer if we cannot fuse in GEMM
diff --git a/tests/CL/Helper.h b/tests/CL/Helper.h
index b99911e1e6..e3e64c9405 100644
--- a/tests/CL/Helper.h
+++ b/tests/CL/Helper.h
@@ -218,6 +218,39 @@ private:
     CLFillBorderKernel         _border_handler{}; /**< Kernel to handle  borders */
     std::unique_ptr<ICLKernel> _kernel{};         /**< Kernel to run */
 };
+
+/** As above but this also setups a Zero border on the input tensor of the kernel's bordersize */
+template <typename K>
+class ClSynthetizeOperatorWithBorder : public opencl::IClOperator
+{
+public:
+    /** Configure the kernel.
+     *
+     * @param[in] first First configuration argument.
+     * @param[in] args  Rest of the configuration arguments.
+     */
+    template <typename T, typename... Args>
+    void configure(T first, Args &&... args)
+    {
+        auto k = std::make_unique<K>();
+        k->configure(CLKernelLibrary::get().get_compile_context(), first, std::forward<Args>(args)...);
+        _kernel = std::move(k);
+
+        auto b = std::make_unique<CLFillBorderKernel>();
+        b->configure(CLKernelLibrary::get().get_compile_context(), first, BorderSize(_kernel->border_size()), BorderMode::CONSTANT, PixelValue());
+        _border_handler = std::move(b);
+    }
+
+    void run(ITensorPack &tensors) override
+    {
+        CLScheduler::get().enqueue(*_border_handler);
+        CLScheduler::get().enqueue_op(*_kernel, tensors);
+    }
+
+private:
+    std::unique_ptr<ICLKernel> _border_handler{ nullptr }; /**< Kernel to handle  borders */
+    std::unique_ptr<ICLKernel> _kernel{};                  /**< Kernel to run */
+};
 } // namespace test
 } // namespace arm_compute
 #endif /* ARM_COMPUTE_TEST_CL_HELPER_H */
diff --git a/tests/validation/CL/Col2Im.cpp b/tests/validation/CL/Col2Im.cpp
index b651bf8918..96096d314d 100644
--- a/tests/validation/CL/Col2Im.cpp
+++ b/tests/validation/CL/Col2Im.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018-2020 Arm Limited.
+ * Copyright (c) 2018-2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -22,7 +22,7 @@
  * SOFTWARE.
  */
 #include "arm_compute/core/Types.h"
-#include "src/core/CL/kernels/CLCol2ImKernel.h"
+#include "src/core/gpu/cl/kernels/ClCol2ImKernel.h"
 #include "tests/CL/CLAccessor.h"
 #include "tests/CL/Helper.h"
 #include "tests/framework/Asserts.h"
@@ -40,7 +40,7 @@ namespace validation
 TEST_SUITE(CL)
 TEST_SUITE(Col2Im)
 
-using CLCol2Im = CLSynthetizeFunction<CLCol2ImKernel>;
+using ClCol2Im = ClSynthetizeOperatorWithBorder<opencl::kernels::ClCol2ImKernel>;
 
 /** Negative tests
  *
@@ -59,7 +59,7 @@ TEST_CASE(Negative, framework::DatasetMode::ALL)
         const auto input     = TensorInfo(TensorShape(10U, 12U, 1U, 2U), 1, DataType::SIZET);
         const auto output    = TensorInfo(TensorShape(3U, 4U, 10U, 1U, 2U), 1, DataType::F32);
         const auto conv_size = Size2D(3, 4);
-        const auto status    = CLCol2ImKernel::validate(&input, &output, conv_size);
+        const auto status    = opencl::kernels::ClCol2ImKernel::validate(&input, &output, conv_size);
         ARM_COMPUTE_EXPECT(bool(status) == false, framework::LogLevel::ERRORS);
     }
 
@@ -68,7 +68,7 @@ TEST_CASE(Negative, framework::DatasetMode::ALL)
         const auto input     = TensorInfo(TensorShape(10U, 12U, 1U, 2U), 1, DataType::F32);
         const auto output    = TensorInfo(TensorShape(3U, 4U, 10U, 1U, 2U), 1, DataType::F32, DataLayout::NHWC);
         const auto conv_size = Size2D(3, 4);
-        const auto status    = CLCol2ImKernel::validate(&input, &output, conv_size);
+        const auto status    = opencl::kernels::ClCol2ImKernel::validate(&input, &output, conv_size);
         ARM_COMPUTE_EXPECT(bool(status) == false, framework::LogLevel::ERRORS);
     }
 
@@ -77,13 +77,13 @@ TEST_CASE(Negative, framework::DatasetMode::ALL)
         const auto input     = TensorInfo(TensorShape(10U, 12U, 1U, 2U), 1, DataType::F32);
         const auto output    = TensorInfo(TensorShape(3U, 4U, 10U, 2U, 2U), 1, DataType::F32);
         const auto conv_size = Size2D(3, 4);
-        const auto status    = CLCol2ImKernel::validate(&input, &output, conv_size);
+        const auto status    = opencl::kernels::ClCol2ImKernel::validate(&input, &output, conv_size);
         ARM_COMPUTE_EXPECT(bool(status) == false, framework::LogLevel::ERRORS);
     }
 }
 
 template <typename T>
-using CLCol2ImFixture = Col2ImValidationFixture<CLTensor, CLAccessor, CLCol2Im, T, true>;
+using ClCol2ImFixture = Col2ImOpValidationFixture<CLTensor, CLAccessor, ClCol2Im, T, true>;
 
 /** Test kernel for single-precision floating point
  *
@@ -99,7 +99,7 @@ using CLCol2ImFixture = Col2ImValidationFixture<CLTensor, CLAccessor, CLCol2Im,
  *  Kernel tested col2im
  */
 FIXTURE_DATA_TEST_CASE(FP32,
-                       CLCol2ImFixture<float>,
+                       ClCol2ImFixture<float>,
                        framework::DatasetMode::ALL,
                        combine(combine(combine(combine(
                                                    framework::dataset::make("InputShape", { TensorShape(8U, 16U, 3U, 1U), TensorShape(17U, 16U, 3U, 1U), TensorShape(7U, 16U, 3U, 1U) }),
@@ -125,7 +125,7 @@ FIXTURE_DATA_TEST_CASE(FP32,
  *  Kernel tested col2im
  */
 FIXTURE_DATA_TEST_CASE(F16,
-                       CLCol2ImFixture<half>,
+                       ClCol2ImFixture<half>,
                        framework::DatasetMode::ALL,
                        combine(combine(combine(combine(
                                                    framework::dataset::make("InputShape", TensorShape(17U, 16U, 3U, 1U)),
@@ -151,7 +151,7 @@ FIXTURE_DATA_TEST_CASE(F16,
  *  Kernel tested col2im
  */
 FIXTURE_DATA_TEST_CASE(QASYMM8,
-                       CLCol2ImFixture<uint8_t>,
+                       ClCol2ImFixture<uint8_t>,
                        framework::DatasetMode::ALL,
                        combine(combine(combine(combine(
                                                    framework::dataset::make("InputShape", TensorShape(17U, 16U, 3U, 1U)),
diff --git a/tests/validation/fixtures/Col2ImFixture.h b/tests/validation/fixtures/Col2ImFixture.h
index 88d420aa3d..ee7a14f844 100644
--- a/tests/validation/fixtures/Col2ImFixture.h
+++ b/tests/validation/fixtures/Col2ImFixture.h
@@ -45,7 +45,7 @@ namespace validation
 using namespace arm_compute::misc::shape_calculator;
 
 template <typename TensorType, typename AccessorType, typename FunctionType, typename T, bool batch_size_on_z>
-class Col2ImValidationFixture : public framework::Fixture
+class Col2ImOpValidationFixture : public framework::Fixture
 {
 public:
     template <typename...>
@@ -74,7 +74,7 @@ protected:
 
         // Create and configure function
         FunctionType col2im_func;
-        col2im_func.configure(&src, &dst, convolved_dims, num_groups);
+        col2im_func.configure(src.info(), dst.info(), convolved_dims, num_groups);
 
         ARM_COMPUTE_ASSERT(src.info()->is_resizable());
         ARM_COMPUTE_ASSERT(dst.info()->is_resizable());
@@ -89,8 +89,13 @@ protected:
         // Fill tensors
         fill(AccessorType(src), 0);
 
+        arm_compute::ITensorPack pack =
+        {
+            { arm_compute::TensorType::ACL_SRC, &src },
+            { arm_compute::TensorType::ACL_DST, &dst }
+        };
         // Compute function
-        col2im_func.run();
+        col2im_func.run(pack);
 
         return dst;
     }