Port CLIm2ColKernel to ClIm2ColKernel

Resolves: COMPMID-4516

Change-Id: I6a6db66797fa801dfe1238fceca413277241d2ec
Signed-off-by: Manuel Bottini <manuel.bottini@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/5946
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>

16 files changed, 237 insertions, 345 deletions

diff --git a/Android.bp b/Android.bp
index f2934cb37d..1d9ec1c9c1 100644
--- a/Android.bp
+++ b/Android.bp
@@ -98,7 +98,6 @@ cc_library_static {
         "src/core/CL/kernels/CLFuseBatchNormalizationKernel.cpp",
         "src/core/CL/kernels/CLGatherKernel.cpp",
         "src/core/CL/kernels/CLGenerateProposalsLayerKernel.cpp",
-        "src/core/CL/kernels/CLIm2ColKernel.cpp",
         "src/core/CL/kernels/CLInstanceNormalizationLayerKernel.cpp",
         "src/core/CL/kernels/CLL2NormalizeLayerKernel.cpp",
         "src/core/CL/kernels/CLMaxUnpoolingLayerKernel.cpp",
@@ -356,6 +355,7 @@ cc_library_static {
         "src/core/gpu/cl/kernels/ClGemmReshapeLhsMatrixKernel.cpp",
         "src/core/gpu/cl/kernels/ClGemmReshapeRhsMatrixKernel.cpp",
         "src/core/gpu/cl/kernels/ClHeightConcatenateKernel.cpp",
+        "src/core/gpu/cl/kernels/ClIm2ColKernel.cpp",
         "src/core/gpu/cl/kernels/ClMulKernel.cpp",
         "src/core/gpu/cl/kernels/ClPermuteKernel.cpp",
         "src/core/gpu/cl/kernels/ClPool2dKernel.cpp",
diff --git a/arm_compute/runtime/CL/functions/CLFullyConnectedLayer.h b/arm_compute/runtime/CL/functions/CLFullyConnectedLayer.h
index 075c5d1f45..82d1621341 100644
--- a/arm_compute/runtime/CL/functions/CLFullyConnectedLayer.h
+++ b/arm_compute/runtime/CL/functions/CLFullyConnectedLayer.h
@@ -96,7 +96,7 @@ private:
 
 /** Basic function to compute a Fully Connected layer on OpenCL. This function calls the following OpenCL kernels:
  *
- *  -# @ref CLIm2ColKernel (called when the input comes from a convolutional layer)
+ *  -# @ref opencl::kernels::ClIm2ColKernel (called when the input comes from a convolutional layer)
  *  -# @ref CLTranspose (if @p are_weights_reshaped is set to false and transpose_weights is set to true ) (called once)
  *  -# @ref opencl::kernels::ClGemmMatrixMultiplyKernel or @ref CLGEMMLowpMatrixMultiplyCore (if quantized asymmetric)
  *
diff --git a/arm_compute/runtime/CL/functions/CLGEMMConvolutionLayer.h b/arm_compute/runtime/CL/functions/CLGEMMConvolutionLayer.h
index 564fb1ecde..e262409ee7 100644
--- a/arm_compute/runtime/CL/functions/CLGEMMConvolutionLayer.h
+++ b/arm_compute/runtime/CL/functions/CLGEMMConvolutionLayer.h
@@ -41,16 +41,16 @@
 
 namespace arm_compute
 {
+class CLWeightsReshapeKernel;
+class ICLTensor;
 namespace opencl
 {
 namespace kernels
 {
+class ClIm2ColKernel;
 class ClCol2ImKernel;
 } // namespace kernels
 } // namespace opencl
-class CLIm2ColKernel;
-class CLWeightsReshapeKernel;
-class ICLTensor;
 
 /** Function to reshape and transpose the weights. This function calls the following kernels:
  * -# @ref CLWeightsReshapeKernel
@@ -173,7 +173,7 @@ private:
 
 /** Basic function to compute the convolution layer. This function calls the following OpenCL kernels/functions:
  *
- * -# @ref CLIm2ColKernel
+ * -# @ref opencl::kernels::ClIm2ColKernel
  * -# @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)
@@ -321,13 +321,14 @@ private:
     IWeightsManager                                                   *_weights_manager;
     CLConvolutionLayerReshapeWeights                                   _reshape_weights;
     weights_transformations::CLConvolutionLayerReshapeWeightsTransform _reshape_weights_managed;
-    std::unique_ptr<CLIm2ColKernel>                                    _im2col_kernel;
+    std::unique_ptr<opencl::kernels::ClIm2ColKernel>                   _im2col_kernel;
     CLGEMM                                                             _mm_gemm;
     CLGEMMLowpMatrixMultiplyCore                                       _mm_gemmlowp;
     std::unique_ptr<opencl::kernels::ClCol2ImKernel>                   _col2im_kernel;
     CLActivationLayer                                                  _activationlayer_function;
 
     const ICLTensor *_original_weights;
+    const ICLTensor *_input;
     const ICLTensor *_gemm_output_to_use;
     ICLTensor       *_output;
 
diff --git a/docs/user_guide/release_version_and_change_log.dox b/docs/user_guide/release_version_and_change_log.dox
index 8b15a384cc..45303e5d87 100644
--- a/docs/user_guide/release_version_and_change_log.dox
+++ b/docs/user_guide/release_version_and_change_log.dox
@@ -567,7 +567,7 @@ v20.08 Public major release
       The default "axis" value for @ref NESoftmaxLayer, @ref NELogSoftmaxLayer is changed from 1 to 0.
       Only axis 0 is supported.
  - The support for quantized data types has been removed from @ref CLLogSoftmaxLayer due to implementation complexity.
- - Removed padding requirement for the input (e.g. LHS of GEMM) and output in CLGEMMMatrixMultiplyNativeKernel, CLGEMMMatrixMultiplyReshapedKernel, CLGEMMMatrixMultiplyReshapedOnlyRHSKernel and @ref CLIm2ColKernel (NHWC only)
+ - Removed padding requirement for the input (e.g. LHS of GEMM) and output in CLGEMMMatrixMultiplyNativeKernel, CLGEMMMatrixMultiplyReshapedKernel, CLGEMMMatrixMultiplyReshapedOnlyRHSKernel and CLIm2ColKernel (NHWC only)
    - This change allows to use @ref CLGEMMConvolutionLayer without extra padding for the input and output.
    - Only the weights/bias of @ref CLGEMMConvolutionLayer could require padding for the computation.
    - Only on Arm® Mali™ Midgard GPUs, @ref CLGEMMConvolutionLayer could require padding since CLGEMMMatrixMultiplyKernel is called and currently requires padding.
@@ -1064,7 +1064,7 @@ v18.08 Public major release
     - @ref CLDirectConvolutionLayer
     - @ref CLConvolutionLayer
     - @ref CLScale
-    - @ref CLIm2ColKernel
+    - CLIm2ColKernel
  - New Arm® Neon™ kernels / functions:
     - @ref NERNNLayer
  - New OpenCL kernels / functions:
@@ -1384,7 +1384,7 @@ v17.02.1 Sources preview
  - New OpenCL kernels / functions:
    - CLLogits1DMaxKernel, CLLogits1DShiftExpSumKernel, CLLogits1DNormKernel / @ref CLSoftmaxLayer
    - CLPoolingLayerKernel / @ref CLPoolingLayer
-   - @ref CLIm2ColKernel, CLCol2ImKernel, CLConvolutionLayerWeightsReshapeKernel / CLConvolutionLayer
+   - 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 c311af459d..68e6aebf4f 100644
--- a/filelist.json
+++ b/filelist.json
@@ -400,7 +400,7 @@
         "files": {
           "kernel": [
             "src/core/gpu/cl/kernels/ClCol2ImKernel.cpp",
-            "src/core/CL/kernels/CLIm2ColKernel.cpp"
+            "src/core/gpu/cl/kernels/ClIm2ColKernel.cpp"
           ]
         }
       },
diff --git a/src/core/CL/CLKernels.h b/src/core/CL/CLKernels.h
index d2d5b928bc..6f6a8642e8 100644
--- a/src/core/CL/CLKernels.h
+++ b/src/core/CL/CLKernels.h
@@ -43,7 +43,6 @@
 #include "src/core/CL/kernels/CLFuseBatchNormalizationKernel.h"
 #include "src/core/CL/kernels/CLGatherKernel.h"
 #include "src/core/CL/kernels/CLGenerateProposalsLayerKernel.h"
-#include "src/core/CL/kernels/CLIm2ColKernel.h"
 #include "src/core/CL/kernels/CLInstanceNormalizationLayerKernel.h"
 #include "src/core/CL/kernels/CLL2NormalizeLayerKernel.h"
 #include "src/core/CL/kernels/CLMaxUnpoolingLayerKernel.h"
diff --git a/src/core/CL/kernels/CLIm2ColKernel.h b/src/core/CL/kernels/CLIm2ColKernel.h
deleted file mode 100644
index 2920c7d138..0000000000
--- a/src/core/CL/kernels/CLIm2ColKernel.h
+++ /dev/null
@@ -1,136 +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_CLIM2COLKERNEL_H
-#define ARM_COMPUTE_CLIM2COLKERNEL_H
-
-#include "arm_compute/core/Size2D.h"
-#include "src/core/CL/ICLKernel.h"
-
-namespace arm_compute
-{
-class ICLTensor;
-
-/** Interface for the im2col reshape kernel.
- *
- * Rearranges image blocks into columns. It is used to strip out each convolution block to a single column.
- * It is used to transform a convolution to a plain matrix multiplication.
- *
- * For example taking into account the image below and assuming 3x3 image blocks with stride of 1 we have:
- * @f[
- * \left( \begin{array}{cccc}
- * a00 & a01 & a02 & a03 \\
- * a10 & a11 & a12 & a13 \\
- * a20 & a21 & a22 & a23 \\
- * a30 & a31 & a32 & a33 \\
- * \end{array} \right)
- * =
- * \left( \begin{array}{ccccccccc}
- * a00 & a01 & a02 & a10 & a11 & a12 & a20 & a21 & a22 \\
- * a01 & a02 & a03 & a11 & a12 & a13 & a21 & a22 & a23 \\
- * a10 & a11 & a12 & a20 & a21 & a22 & a30 & a31 & a32 \\
- * a11 & a12 & a13 & a21 & a22 & a23 & a31 & a32 & a33 \\
- * \end{array} \right)
- * @f]
- */
-class CLIm2ColKernel : public ICLKernel
-{
-public:
-    /** Default constructor */
-    CLIm2ColKernel();
-    /** Prevent instances of this class from being copied (As this class contains pointers) */
-    CLIm2ColKernel(const CLIm2ColKernel &) = delete;
-    /** Prevent instances of this class from being copied (As this class contains pointers) */
-    CLIm2ColKernel &operator=(const CLIm2ColKernel &) = delete;
-    /** Allow instances of this class to be moved */
-    CLIm2ColKernel(CLIm2ColKernel &&) = default;
-    /** Allow instances of this class to be moved */
-    CLIm2ColKernel &operator=(CLIm2ColKernel &&) = default;
-    /** Set the input and output of the kernel.
-     *
-     * @param[in]  input       The input tensor to convert. 3 lower dimensions represent a single input [width, height, IFM],
-     *                         while every optional dimension from 4 and above represent a batch of inputs. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32
-     * @param[out] output      The output tensor. First 2 lower dimensions represent a transform of each 3D input,
-     *                         while every dimension above represents a batch. Data types supported: Same as @p input
-     * @param[in]  kernel_dims The kernel dimensions (width and height).
-     * @param[in]  conv_info   Contains padding and stride information described in @ref PadStrideInfo.
-     * @param[in]  has_bias    In case biases are provided expands the matrix with 1.
-     *                         This is valid only for non-quantized inputs.
-     * @param[in]  dilation    (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
-     * @param[in]  num_groups  (Optional) Number of groups when performing a grouped convolution.
-     *                         Number of groups other than 1 is only supported for NCHW data layout.
-     *                         Number of groups should be multiple to the number of channels.
-     */
-    void configure(const ICLTensor *input, ICLTensor *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, const Size2D &dilation = Size2D(1U, 1U),
-                   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. 3 lower dimensions represent a single input [width, height, IFM],
-     *                             while every optional dimension from 4 and above represent a batch of inputs. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32
-     * @param[out] output          The output tensor. First 2 lower dimensions represent a transform of each 3D input,
-     *                             while every dimension above represents a batch. Data types supported: Same as @p input
-     * @param[in]  kernel_dims     The kernel dimensions (width and height).
-     * @param[in]  conv_info       Contains padding and stride information described in @ref PadStrideInfo.
-     * @param[in]  has_bias        In case biases are provided expands the matrix with 1.
-     * @param[in]  dilation        (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
-     * @param[in]  num_groups      (Optional) Number of groups when performing a grouped convolution. num_groups != 1 is only supported for NCHW data layout
-     */
-    void configure(const CLCompileContext &compile_context, const ICLTensor *input, ICLTensor *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias,
-                   const Size2D &dilation   = Size2D(1U, 1U),
-                   unsigned int  num_groups = 1);
-    /** Static function to check if given info will lead to a valid configuration of @ref CLIm2ColKernel
-     *
-     * @param[in] input       The input tensor to convert. 3 lower dimensions represent a single input [width, height, IFM],
-     *                        while every optional dimension from 4 and above represent a batch of inputs. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32
-     * @param[in] output      The output tensor. First 2 lower dimensions represent a transform of each 3D input,
-     *                        while every dimension above represents a batch. Data types supported: Same as @p input
-     * @param[in] kernel_dims The kernel dimensions (width and height).
-     * @param[in] conv_info   Contains padding and stride information described in @ref PadStrideInfo.
-     * @param[in] has_bias    In case biases are provided expands the matrix with 1.
-     *                        This is valid only for non-quantized inputs.
-     * @param[in] dilation    (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
-     * @param[in] num_groups  (Optional) Number of groups when performing a grouped convolution.
-     *                        Number of groups other than 1 is only supported for NCHW data layout.
-     *                        Number of groups should be multiple to the number of channels.
-     *
-     * @return a status
-     */
-    static Status validate(const ITensorInfo *input, const ITensorInfo *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, const Size2D &dilation = Size2D(1U, 1U),
-                           unsigned int num_groups = 1);
-
-    // Inherited methods overridden:
-    void run(const Window &window, cl::CommandQueue &queue) override;
-
-public:
-    const ICLTensor *_input;
-    ICLTensor       *_output;
-    DataLayout       _data_layout;
-    std::pair<unsigned int, unsigned int> _convolved_dims;
-    unsigned int  _num_elems_processed_per_iteration;
-    Size2D        _kernel_dims;
-    PadStrideInfo _conv_info;
-    unsigned int  _num_groups;
-};
-} // namespace arm_compute
-#endif /*ARM_COMPUTE_CLIM2COLKERNEL_H */
diff --git a/src/core/CL/kernels/CLWeightsReshapeKernel.h b/src/core/CL/kernels/CLWeightsReshapeKernel.h
index 402a60472b..9ac60a7a1a 100644
--- a/src/core/CL/kernels/CLWeightsReshapeKernel.h
+++ b/src/core/CL/kernels/CLWeightsReshapeKernel.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2020 Arm Limited.
+ * Copyright (c) 2017-2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -31,7 +31,7 @@ namespace arm_compute
 /** OpenCL kernel to perform reshaping on the weights used by convolution and locally connected layer
  *
  * Rearranges each 3-dimensional kernel to a single row leading to a matrix with linearized kernels.
- * In combination with the @ref CLIm2ColKernel can transform a convolution to a matrix multiplication.
+ * In combination with the @ref opencl::kernels::ClIm2ColKernel can transform a convolution to a matrix multiplication.
  *
  * For example assuming a 3D weight kernel of 3x3 dimensions and depth of 2 we have:
  * @f[
diff --git a/src/core/gpu/cl/kernels/ClCol2ImKernel.h b/src/core/gpu/cl/kernels/ClCol2ImKernel.h
index 42d0a96075..74a9027628 100644
--- a/src/core/gpu/cl/kernels/ClCol2ImKernel.h
+++ b/src/core/gpu/cl/kernels/ClCol2ImKernel.h
@@ -37,7 +37,7 @@ namespace kernels
 {
 /** Interface for the col2im reshaping kernel.
  *
- * Rearranges each matrix column into image blocks. It's the inverse operation of @ref CLIm2ColKernel.
+ * Rearranges each matrix column into image blocks. It's the inverse operation of @ref opencl::kernels::ClIm2ColKernel.
  *
  * For example, a vector of 9 elements can be reshaped to a block(image) of 3x3:
  *
diff --git a/src/core/CL/kernels/CLIm2ColKernel.cpp b/src/core/gpu/cl/kernels/ClIm2ColKernel.cpp
index 97740e3c34..61ee443aa5 100644
--- a/src/core/CL/kernels/CLIm2ColKernel.cpp
+++ b/src/core/gpu/cl/kernels/ClIm2ColKernel.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/CLIm2ColKernel.h"
+#include "src/core/gpu/cl/kernels/ClIm2ColKernel.h"
 
 #include "arm_compute/core/CL/CLHelpers.h"
 #include "arm_compute/core/CL/CLKernelLibrary.h"
@@ -35,6 +35,7 @@
 #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>
@@ -44,7 +45,10 @@
 namespace arm_compute
 {
 using namespace misc::shape_calculator;
-
+namespace opencl
+{
+namespace kernels
+{
 namespace
 {
 struct Im2ColConfiguration
@@ -55,54 +59,54 @@ struct Im2ColConfiguration
     bool                  is_padding_required_nchw{};
 };
 
-Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, const Size2D &dilation,
+Status validate_arguments(const ITensorInfo *src, const ITensorInfo *dst, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, const Size2D &dilation,
                           unsigned int num_groups)
 {
-    const unsigned int channel_idx = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::CHANNEL);
+    const unsigned int channel_idx = get_data_layout_dimension_index(src->data_layout(), DataLayoutDimension::CHANNEL);
 
-    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(is_data_type_quantized(input->data_type()) && has_bias);
-    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(output);
+    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);
+    ARM_COMPUTE_RETURN_ERROR_ON(is_data_type_quantized(src->data_type()) && has_bias);
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(dst);
     ARM_COMPUTE_RETURN_ERROR_ON((dilation.x() < 1) || (dilation.y() < 1));
-    ARM_COMPUTE_RETURN_ERROR_ON(input->data_layout() == DataLayout::UNKNOWN);
+    ARM_COMPUTE_RETURN_ERROR_ON(src->data_layout() == DataLayout::UNKNOWN);
     ARM_COMPUTE_RETURN_ERROR_ON(num_groups == 0);
-    ARM_COMPUTE_RETURN_ERROR_ON(input->data_layout() == DataLayout::NHWC && num_groups > 1);
-    ARM_COMPUTE_RETURN_ERROR_ON((input->dimension(channel_idx) % num_groups) != 0);
+    ARM_COMPUTE_RETURN_ERROR_ON(src->data_layout() == DataLayout::NHWC && num_groups > 1);
+    ARM_COMPUTE_RETURN_ERROR_ON((src->dimension(channel_idx) % num_groups) != 0);
 
     // Since there's no implicit padding added, check the total input spatial dimensions (with conv paddings) are big enough for the kernel dimensions
-    const unsigned int width_idx    = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::WIDTH);
-    const unsigned int height_idx   = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::HEIGHT);
-    const unsigned     total_width  = input->dimension(width_idx) + conv_info.pad_left() + conv_info.pad_right();
-    const unsigned     total_height = input->dimension(height_idx) + conv_info.pad_top() + conv_info.pad_bottom();
+    const unsigned int width_idx    = get_data_layout_dimension_index(src->data_layout(), DataLayoutDimension::WIDTH);
+    const unsigned int height_idx   = get_data_layout_dimension_index(src->data_layout(), DataLayoutDimension::HEIGHT);
+    const unsigned     total_width  = src->dimension(width_idx) + conv_info.pad_left() + conv_info.pad_right();
+    const unsigned     total_height = src->dimension(height_idx) + conv_info.pad_top() + conv_info.pad_bottom();
     ARM_COMPUTE_RETURN_ERROR_ON((total_width < kernel_dims.width) || (total_height < kernel_dims.height));
 
-    if(output->total_size() > 0)
+    if(dst->total_size() > 0)
     {
-        const TensorInfo tensor_info_output = output->clone()->set_tensor_shape(compute_im2col_conv_shape(input, kernel_dims, conv_info, has_bias, dilation, num_groups == 1, num_groups));
-        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(output, &tensor_info_output);
-        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
-        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(input, output);
+        const TensorInfo tensor_info_output = dst->clone()->set_tensor_shape(compute_im2col_conv_shape(src, kernel_dims, conv_info, has_bias, dilation, num_groups == 1, num_groups));
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(dst, &tensor_info_output);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, dst);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(src, dst);
     }
 
     return Status{};
 }
 
-std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, const Size2D &dilation,
+std::pair<Status, Window> validate_and_configure_window(ITensorInfo *src, ITensorInfo *dst, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, const Size2D &dilation,
                                                         unsigned int num_elems_processed_per_iteration, bool is_padding_required_nchw, unsigned int num_groups)
 {
-    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
+    ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst);
 
     // Output tensor auto initialization if not yet initialized
-    TensorShape expected_output_shape = compute_im2col_conv_shape(input, kernel_dims, conv_info, has_bias, dilation, num_groups == 1, num_groups);
+    TensorShape expected_output_shape = compute_im2col_conv_shape(src, kernel_dims, conv_info, has_bias, dilation, num_groups == 1, num_groups);
 
-    auto_init_if_empty(*output, input->clone()->set_tensor_shape(expected_output_shape));
+    auto_init_if_empty(*dst, src->clone()->set_tensor_shape(expected_output_shape));
 
-    const DataLayout   data_layout  = input->data_layout();
+    const DataLayout   data_layout  = src->data_layout();
     const unsigned int width_idx    = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH);
     const unsigned int height_idx   = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT);
-    const unsigned int input_width  = input->dimension(width_idx);
-    const unsigned int input_height = input->dimension(height_idx);
+    const unsigned int input_width  = src->dimension(width_idx);
+    const unsigned int input_height = src->dimension(height_idx);
 
     // Configure the execute window based on the selected optimal OpenCL kernel
     bool   window_changed = false;
@@ -110,16 +114,16 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITen
 
     if(data_layout == DataLayout::NHWC)
     {
-        win = calculate_max_window(*input, Steps(num_elems_processed_per_iteration));
+        win = calculate_max_window(*src, Steps(num_elems_processed_per_iteration));
     }
     else
     {
         if(is_padding_required_nchw)
         {
             const BorderSize border(conv_info.pad_top(), conv_info.pad_right(), conv_info.pad_bottom(), conv_info.pad_left());
-            win = calculate_max_window(*input,
+            win = calculate_max_window(*src,
                                        Steps(num_elems_processed_per_iteration * conv_info.stride().first, conv_info.stride().second));
-            AccessWindowStatic input_access(input,
+            AccessWindowStatic input_access(src,
                                             -border.left,
                                             -border.top,
                                             ceil_to_multiple(input_width + border.right, kernel_dims.width * num_elems_processed_per_iteration),
@@ -130,7 +134,7 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITen
         {
             // For the generic case, CLIm2ColKernel doesn't need padding (we do not read out-of-bounds elements) so
             // update_window_and_padding() can be skipped
-            win = calculate_max_window(*input, Steps());
+            win = calculate_max_window(*src, Steps());
         }
     }
 
@@ -141,16 +145,16 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITen
     return std::make_pair(err, win);
 }
 
-Im2ColConfiguration configure_opencl_kernel(const ITensorInfo *input, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, const Size2D &dilation, unsigned int num_groups)
+Im2ColConfiguration configure_opencl_kernel(const ITensorInfo *src, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, const Size2D &dilation, unsigned int num_groups)
 {
-    const DataLayout   data_layout   = input->data_layout();
-    const DataType     data_type     = input->data_type();
+    const DataLayout   data_layout   = src->data_layout();
+    const DataType     data_type     = src->data_type();
     const unsigned int width_idx     = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH);
     const unsigned int height_idx    = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT);
     const unsigned int channel_idx   = get_data_layout_dimension_index(data_layout, DataLayoutDimension::CHANNEL);
-    const unsigned int input_width   = input->dimension(width_idx);
-    const unsigned int input_height  = input->dimension(height_idx);
-    const unsigned int input_channel = input->dimension(channel_idx);
+    const unsigned int input_width   = src->dimension(width_idx);
+    const unsigned int input_height  = src->dimension(height_idx);
+    const unsigned int input_channel = src->dimension(channel_idx);
 
     const std::pair<unsigned int, unsigned int> convolved_dims = scaled_dimensions(input_width, input_height, kernel_dims.width, kernel_dims.height, conv_info, dilation);
 
@@ -159,10 +163,10 @@ Im2ColConfiguration configure_opencl_kernel(const ITensorInfo *input, const Size
     CLBuildOptions                build_opts;
     unsigned int                  num_elems_processed_per_iteration = 1;
     bool                          is_padding_required_nchw          = false;
-    const UniformQuantizationInfo qinfo                             = input->quantization_info().uniform();
+    const UniformQuantizationInfo qinfo                             = src->quantization_info().uniform();
 
     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->element_size()));
+    build_opts.add_option("-DELEMENT_SIZE=" + support::cpp11::to_string(src->element_size()));
     build_opts.add_option("-DKERNEL_WIDTH=" + support::cpp11::to_string(kernel_dims.width));
     build_opts.add_option("-DKERNEL_HEIGHT=" + support::cpp11::to_string(kernel_dims.height));
     build_opts.add_option("-DCONVOLVED_WIDTH=" + support::cpp11::to_string(convolved_dims.first));
@@ -285,43 +289,35 @@ Im2ColConfiguration configure_opencl_kernel(const ITensorInfo *input, const Size
 }
 } // namespace
 
-CLIm2ColKernel::CLIm2ColKernel()
-    : _input(nullptr), _output(nullptr), _data_layout(DataLayout::UNKNOWN), _convolved_dims(), _num_elems_processed_per_iteration(1), _kernel_dims(), _conv_info(), _num_groups()
+ClIm2ColKernel::ClIm2ColKernel()
+    : _data_layout(DataLayout::UNKNOWN), _convolved_dims(), _num_elems_processed_per_iteration(1), _kernel_dims(), _conv_info(), _num_groups()
 {
     _type = CLKernelType::ELEMENTWISE;
 }
 
-void CLIm2ColKernel::configure(const ICLTensor *input, ICLTensor *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, const Size2D &dilation,
-                               unsigned int num_groups)
-{
-    configure(CLKernelLibrary::get().get_compile_context(), input, output, kernel_dims, conv_info, has_bias, dilation, num_groups);
-}
-
-void CLIm2ColKernel::configure(const CLCompileContext &compile_context, const ICLTensor *input, ICLTensor *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias,
+void ClIm2ColKernel::configure(const ClCompileContext &compile_context, ITensorInfo *src, ITensorInfo *dst, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias,
                                const Size2D &dilation,
                                unsigned int  num_groups)
 {
-    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
-    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), kernel_dims, conv_info, has_bias, dilation, num_groups));
+    ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst);
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src, dst, kernel_dims, conv_info, has_bias, dilation, num_groups));
 
-    auto padding_info = get_padding_info({ input, output });
-    _data_layout      = input->info()->data_layout();
+    auto padding_info = get_padding_info({ src, dst });
+    _data_layout      = src->data_layout();
 
     const unsigned int width_idx    = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::WIDTH);
     const unsigned int height_idx   = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::HEIGHT);
-    const unsigned int input_width  = input->info()->dimension(width_idx);
-    const unsigned int input_height = input->info()->dimension(height_idx);
+    const unsigned int input_width  = src->dimension(width_idx);
+    const unsigned int input_height = src->dimension(height_idx);
 
     // Select and configure the optimal OpenCL kernel to run.
     // This function returns the OpenCL kernel's name, the arguments to pass at compile time, the number of elements processed per iteration
     // and the padding requirement flag
-    Im2ColConfiguration im2col_config = configure_opencl_kernel(input->info(), kernel_dims, conv_info, has_bias, dilation, num_groups);
+    Im2ColConfiguration im2col_config = configure_opencl_kernel(src, kernel_dims, conv_info, has_bias, dilation, num_groups);
 
     // Create kernel
     _kernel = create_kernel(compile_context, im2col_config.kernel_name, im2col_config.build_options);
 
-    _input                             = input;
-    _output                            = output;
     _convolved_dims                    = scaled_dimensions(input_width, input_height, kernel_dims.width, kernel_dims.height, conv_info, dilation);
     _num_elems_processed_per_iteration = im2col_config.num_elems_processed_per_iteration;
     _kernel_dims                       = kernel_dims; // Only needed by the Tuner
@@ -329,50 +325,55 @@ void CLIm2ColKernel::configure(const CLCompileContext &compile_context, const IC
     _num_groups                        = num_groups;
 
     // Configure kernel window
-    auto win_config = validate_and_configure_window(input->info(), output->info(), kernel_dims, conv_info, has_bias, dilation, im2col_config.num_elems_processed_per_iteration,
+    auto win_config = validate_and_configure_window(src, dst, kernel_dims, conv_info, has_bias, dilation, im2col_config.num_elems_processed_per_iteration,
                                                     im2col_config.is_padding_required_nchw, 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 = im2col_config.kernel_name;
     _config_id += "_";
-    _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(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));
     _config_id += "_";
     _config_id += lower_string(string_from_data_layout(_data_layout));
 
-    ARM_COMPUTE_ERROR_ON(input->info()->data_layout() == DataLayout::NHWC && has_padding_changed(padding_info));
+    ARM_COMPUTE_ERROR_ON(src->data_layout() == DataLayout::NHWC && has_padding_changed(padding_info));
 }
 
-Status CLIm2ColKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, const Size2D &dilation,
+Status ClIm2ColKernel::validate(const ITensorInfo *src, const ITensorInfo *dst, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, const Size2D &dilation,
                                 unsigned int num_groups)
 {
-    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, kernel_dims, conv_info, has_bias, dilation, num_groups));
-    Im2ColConfiguration im2col_config = configure_opencl_kernel(input, kernel_dims, conv_info, has_bias, dilation, num_groups);
-    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), output->clone().get(), kernel_dims, conv_info, has_bias, dilation, im2col_config.num_elems_processed_per_iteration,
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src, dst, kernel_dims, conv_info, has_bias, dilation, num_groups));
+    Im2ColConfiguration im2col_config = configure_opencl_kernel(src, kernel_dims, conv_info, has_bias, dilation, num_groups);
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(src->clone().get(), dst->clone().get(), kernel_dims, conv_info, has_bias, dilation, im2col_config.num_elems_processed_per_iteration,
                                                               im2col_config.is_padding_required_nchw, num_groups)
                                 .first);
     return Status{};
 }
 
-void CLIm2ColKernel::run(const Window &window, cl::CommandQueue &queue)
+void ClIm2ColKernel::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);
+    ARM_COMPUTE_ERROR_ON(tensors.empty());
 
     // Get initial windows
     // Collapse in order to have (SRC_DEPTH * BATCH_SIZE) on the 3rd dimension
     Window window_collapsed = window.collapse_if_possible(ICLKernel::window(), Window::DimZ);
     window_collapsed.set_dimension_step(Window::DimZ, 1);
 
+    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));
+    ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst);
+
     Window window_output;
-    window_output.use_tensor_dimensions(_output->info()->tensor_shape());
+    window_output.use_tensor_dimensions(dst->info()->tensor_shape());
 
     const Window first_slice_3d = window_collapsed.first_slice_window_3D();
 
@@ -385,7 +386,7 @@ void CLIm2ColKernel::run(const Window &window, cl::CommandQueue &queue)
         const Window tmp_win     = window.collapse_if_possible(ICLKernel::window(), 3);
         const int    num_batches = tmp_win[3].end();
 
-        slice.set(1, Window::Dimension(0, static_cast<int>(_output->info()->tensor_shape()[1]), 1));
+        slice.set(1, Window::Dimension(0, static_cast<int>(dst->info()->tensor_shape()[1]), 1));
         slice.set(2, Window::Dimension(0, static_cast<int>(num_batches), 1));
     }
     else
@@ -407,22 +408,24 @@ void CLIm2ColKernel::run(const Window &window, cl::CommandQueue &queue)
     slice_out.set(Window::DimY, Window::Dimension(0, 0, 0));
 
     unsigned int idx = num_arguments_per_3D_tensor() + (_num_groups == 1 ? num_arguments_per_2D_tensor() : num_arguments_per_3D_tensor());
-    _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_input->info()->strides_in_bytes()[3]));
-    _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_output->info()->strides_in_bytes()[((_num_groups == 1) ? 2 : 3)]));
+    _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(src->info()->strides_in_bytes()[3]));
+    _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(dst->info()->strides_in_bytes()[((_num_groups == 1) ? 2 : 3)]));
     do
     {
         unsigned int idx = 0;
-        add_3D_tensor_argument(idx, _input, slice_in);
+        add_3D_tensor_argument(idx, src, slice_in);
         if(_num_groups == 1)
         {
-            add_2D_tensor_argument(idx, _output, slice_out);
+            add_2D_tensor_argument(idx, dst, slice_out);
         }
         else
         {
-            add_3D_tensor_argument(idx, _output, slice_out);
+            add_3D_tensor_argument(idx, dst, slice_out);
         }
         enqueue(queue, *this, slice, lws_hint());
     }
     while(window_collapsed.slide_window_slice_3D(slice) && window_output.slide_window_slice_2D(slice_out) && window_collapsed.slide_window_slice_3D(slice_in));
 }
+} // namespace kernels
+} // namespace opencl
 } // namespace arm_compute
diff --git a/src/core/gpu/cl/kernels/ClIm2ColKernel.h b/src/core/gpu/cl/kernels/ClIm2ColKernel.h
new file mode 100644
index 0000000000..d1443f0434
--- /dev/null
+++ b/src/core/gpu/cl/kernels/ClIm2ColKernel.h
@@ -0,0 +1,106 @@
+/*
+ * 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_IM2COL_KERNEL_H
+#define ARM_COMPUTE_CL_IM2COL_KERNEL_H
+
+#include "arm_compute/core/KernelDescriptors.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 im2col reshape kernel.
+ *
+ * Rearranges image blocks into columns. It is used to strip out each convolution block to a single column.
+ * It is used to transform a convolution to a plain matrix multiplication.
+ *
+ * For example taking into account the image below and assuming 3x3 image blocks with stride of 1 we have:
+ * @f[
+ * \left( \begin{array}{cccc}
+ * a00 & a01 & a02 & a03 \\
+ * a10 & a11 & a12 & a13 \\
+ * a20 & a21 & a22 & a23 \\
+ * a30 & a31 & a32 & a33 \\
+ * \end{array} \right)
+ * =
+ * \left( \begin{array}{ccccccccc}
+ * a00 & a01 & a02 & a10 & a11 & a12 & a20 & a21 & a22 \\
+ * a01 & a02 & a03 & a11 & a12 & a13 & a21 & a22 & a23 \\
+ * a10 & a11 & a12 & a20 & a21 & a22 & a30 & a31 & a32 \\
+ * a11 & a12 & a13 & a21 & a22 & a23 & a31 & a32 & a33 \\
+ * \end{array} \right)
+ * @f]
+ */
+class ClIm2ColKernel : public IClKernel
+{
+public:
+    /** Default constructor */
+    ClIm2ColKernel();
+    ARM_COMPUTE_DISALLOW_COPY_ALLOW_MOVE(ClIm2ColKernel);
+    /** 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. 3 lower dimensions represent a single input [width, height, IFM],
+     *                             while every optional dimension from 4 and above represent a batch of inputs. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32
+     * @param[out] dst             The output tensor info. First 2 lower dimensions represent a transform of each 3D input,
+     *                             while every dimension above represents a batch. Data types supported: Same as @p input
+     * @param[in]  kernel_dims     The kernel dimensions (width and height).
+     * @param[in]  conv_info       Contains padding and stride information described in @ref PadStrideInfo.
+     * @param[in]  has_bias        In case biases are provided expands the matrix with 1.
+     * @param[in]  dilation        (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
+     * @param[in]  num_groups      (Optional) Number of groups when performing a grouped convolution. num_groups != 1 is only supported for NCHW data layout
+     */
+    void configure(const ClCompileContext &compile_context, ITensorInfo *src, ITensorInfo *dst, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias,
+                   const Size2D &dilation   = Size2D(1U, 1U),
+                   unsigned int  num_groups = 1);
+    /** Static function to check if given info will lead to a valid configuration
+     *
+     * Similar to ClIm2ColKernel::configure()
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *input, const ITensorInfo *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, const Size2D &dilation = Size2D(1U, 1U),
+                           unsigned int num_groups = 1);
+
+    // Inherited methods overridden:
+    void run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue) override;
+
+public:
+    DataLayout _data_layout;
+    std::pair<unsigned int, unsigned int> _convolved_dims;
+    unsigned int  _num_elems_processed_per_iteration;
+    Size2D        _kernel_dims;
+    PadStrideInfo _conv_info;
+    unsigned int  _num_groups;
+};
+} // namespace kernels
+} // namespace opencl
+} // namespace arm_compute
+#endif /* ARM_COMPUTE_CL_IM2COL_KERNEL_H */
diff --git a/src/runtime/CL/functions/CLGEMMConvolutionLayer.cpp b/src/runtime/CL/functions/CLGEMMConvolutionLayer.cpp
index f0f45a8659..f926b1d0a6 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/CLIm2ColKernel.h"
 #include "src/core/CL/kernels/CLWeightsReshapeKernel.h"
 #include "src/core/gpu/cl/kernels/ClCol2ImKernel.h"
+#include "src/core/gpu/cl/kernels/ClIm2ColKernel.h"
 #include "src/core/helpers/AutoConfiguration.h"
 #include "support/Cast.h"
 
@@ -105,8 +105,8 @@ 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(nullptr), _activationlayer_function(), _original_weights(nullptr), _gemm_output_to_use(nullptr), _output(nullptr), _im2col_output(),
+    : _memory_group(memory_manager), _weights_manager(weights_manager), _reshape_weights(), _reshape_weights_managed(), _im2col_kernel(nullptr), _mm_gemm(memory_manager, weights_manager),
+      _mm_gemmlowp(memory_manager), _col2im_kernel(nullptr), _activationlayer_function(), _original_weights(nullptr), _input(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)
 {
 }
@@ -229,6 +229,7 @@ void CLGEMMConvolutionLayer::configure(const CLCompileContext &compile_context,
 
     _is_prepared      = weights_info.retain_internal_weights();
     _original_weights = weights;
+    _input            = input;
     _is_quantized     = is_data_type_quantized_asymmetric(input->info()->data_type());
     _skip_im2col      = (data_layout == DataLayout::NHWC && kernel_width == 1 && kernel_height == 1 && conv_info.stride().first == 1 && conv_info.stride().second == 1);
     _skip_col2im      = data_layout == DataLayout::NHWC;
@@ -236,9 +237,6 @@ void CLGEMMConvolutionLayer::configure(const CLCompileContext &compile_context,
     // Only for quantize there are few cases where we cannot fuse the activation function in GEMM
     _fuse_activation = true;
 
-    // Set the GPU target for im2col and col2im
-    _im2col_kernel->set_target(CLScheduler::get().target());
-
     const ICLTensor *gemm_input_to_use  = input;
     ICLTensor       *gemm_output_to_use = output;
 
@@ -299,7 +297,11 @@ void CLGEMMConvolutionLayer::configure(const CLCompileContext &compile_context,
         _memory_group.manage(&_im2col_output);
 
         // Configure and tune im2col. im2col output shape is auto-initialized
-        _im2col_kernel->configure(compile_context, input, &_im2col_output, Size2D(kernel_width, kernel_height), conv_info, append_bias, dilation, num_groups);
+        _im2col_kernel = std::make_unique<opencl::kernels::ClIm2ColKernel>();
+
+        // Set the GPU target for im2col
+        _im2col_kernel->set_target(CLScheduler::get().target());
+        _im2col_kernel->configure(compile_context, input->info(), _im2col_output.info(), Size2D(kernel_width, kernel_height), conv_info, append_bias, dilation, num_groups);
 
         // Set quantization info
         _im2col_output.info()->set_quantization_info(input->info()->quantization_info());
@@ -525,7 +527,7 @@ Status CLGEMMConvolutionLayer::validate(const ITensorInfo *input, const ITensorI
 
         auto_init_if_empty(im2col_reshaped_info, input->clone()->set_tensor_shape(expected_output_shape));
 
-        ARM_COMPUTE_RETURN_ON_ERROR(CLIm2ColKernel::validate(input, &im2col_reshaped_info, kernel_dims, conv_info, append_bias, dilation, num_groups));
+        ARM_COMPUTE_RETURN_ON_ERROR(opencl::kernels::ClIm2ColKernel::validate(input, &im2col_reshaped_info, kernel_dims, conv_info, append_bias, dilation, num_groups));
         gemm_input_to_use = &im2col_reshaped_info;
     }
 
@@ -620,7 +622,12 @@ void CLGEMMConvolutionLayer::run()
     // Run im2col
     if(!_skip_im2col)
     {
-        CLScheduler::get().enqueue(*_im2col_kernel);
+        ITensorPack pack =
+        {
+            { TensorType::ACL_SRC, _input },
+            { TensorType::ACL_DST, &_im2col_output }
+        };
+        CLScheduler::get().enqueue_op(*_im2col_kernel, pack, false);
     }
 
     // Runs CLGEMM or CLGEMMLowpMatrixMultiplyCore functions
diff --git a/src/runtime/CL/functions/CLGEMMDeconvolutionLayer.cpp b/src/runtime/CL/functions/CLGEMMDeconvolutionLayer.cpp
index bab29a5095..7b98b524c1 100644
--- a/src/runtime/CL/functions/CLGEMMDeconvolutionLayer.cpp
+++ b/src/runtime/CL/functions/CLGEMMDeconvolutionLayer.cpp
@@ -30,7 +30,6 @@
 #include "arm_compute/runtime/CL/CLScheduler.h"
 #include "src/core/CL/kernels/CLDeconvolutionReshapeOutputKernel.h"
 #include "src/core/CL/kernels/CLFillBorderKernel.h"
-#include "src/core/CL/kernels/CLIm2ColKernel.h"
 #include "src/core/CL/kernels/CLWeightsReshapeKernel.h"
 
 #include <tuple>
diff --git a/tests/validation/CL/Im2Col.cpp b/tests/validation/CL/Im2Col.cpp
index c6006efcba..041f549777 100644
--- a/tests/validation/CL/Im2Col.cpp
+++ b/tests/validation/CL/Im2Col.cpp
@@ -22,7 +22,7 @@
  * SOFTWARE.
  */
 #include "arm_compute/core/Types.h"
-#include "src/core/CL/kernels/CLIm2ColKernel.h"
+#include "src/core/gpu/cl/kernels/ClIm2ColKernel.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(Im2Col)
 
-using CLIm2Col = CLSynthetizeFunction<CLIm2ColKernel>;
+using ClIm2Col = ClSynthetizeOperatorWithBorder<opencl::kernels::ClIm2ColKernel>;
 
 /** Negative tests
  *
@@ -63,7 +63,7 @@ TEST_CASE(Negative, framework::DatasetMode::ALL)
         const auto output    = TensorInfo(TensorShape(9U, 10U, 12U, 2U), 1, DataType::F32);
         const auto conv_size = Size2D(3, 3);
         const bool has_bias  = false;
-        const auto status    = CLIm2ColKernel::validate(&input, &output, conv_size, PadStrideInfo(), has_bias);
+        const auto status    = opencl::kernels::ClIm2ColKernel::validate(&input, &output, conv_size, PadStrideInfo(), has_bias);
         ARM_COMPUTE_EXPECT(bool(status) == false, framework::LogLevel::ERRORS);
     }
 
@@ -73,7 +73,7 @@ TEST_CASE(Negative, framework::DatasetMode::ALL)
         const auto output    = TensorInfo(TensorShape(9U, 80U, 2U), 1, DataType::QASYMM8);
         const auto conv_size = Size2D(3, 3);
         const bool has_bias  = true;
-        const auto status    = CLIm2ColKernel::validate(&input, &output, conv_size, PadStrideInfo(), has_bias);
+        const auto status    = opencl::kernels::ClIm2ColKernel::validate(&input, &output, conv_size, PadStrideInfo(), has_bias);
         ARM_COMPUTE_EXPECT(bool(status) == false, framework::LogLevel::ERRORS);
     }
 
@@ -84,7 +84,7 @@ TEST_CASE(Negative, framework::DatasetMode::ALL)
         const auto conv_size = Size2D(3, 3);
         const auto dilation  = Size2D(0, 1);
         const bool has_bias  = false;
-        const auto status    = CLIm2ColKernel::validate(&input, &output, conv_size, PadStrideInfo(), has_bias, dilation);
+        const auto status    = opencl::kernels::ClIm2ColKernel::validate(&input, &output, conv_size, PadStrideInfo(), has_bias, dilation);
         ARM_COMPUTE_EXPECT(bool(status) == false, framework::LogLevel::ERRORS);
     }
 
@@ -96,7 +96,7 @@ TEST_CASE(Negative, framework::DatasetMode::ALL)
         const auto         dilation   = Size2D(1, 1);
         const bool         has_bias   = false;
         const unsigned int num_groups = 2;
-        const auto         status     = CLIm2ColKernel::validate(&input, &output, conv_size, PadStrideInfo(), has_bias, dilation, num_groups);
+        const auto         status     = opencl::kernels::ClIm2ColKernel::validate(&input, &output, conv_size, PadStrideInfo(), has_bias, dilation, num_groups);
         ARM_COMPUTE_EXPECT(bool(status) == false, framework::LogLevel::ERRORS);
     }
 
@@ -108,7 +108,7 @@ TEST_CASE(Negative, framework::DatasetMode::ALL)
         const auto         dilation   = Size2D(1, 1);
         const bool         has_bias   = false;
         const unsigned int num_groups = 2;
-        const auto         status     = CLIm2ColKernel::validate(&input, &output, conv_size, PadStrideInfo(), has_bias, dilation, num_groups);
+        const auto         status     = opencl::kernels::ClIm2ColKernel::validate(&input, &output, conv_size, PadStrideInfo(), has_bias, dilation, num_groups);
         ARM_COMPUTE_EXPECT(bool(status) == false, framework::LogLevel::ERRORS);
     }
 
@@ -118,7 +118,7 @@ TEST_CASE(Negative, framework::DatasetMode::ALL)
         const auto output    = TensorInfo(TensorShape(9U, 81U, 2U), 1, DataType::F32);
         const auto conv_size = Size2D(3, 3);
         const bool has_bias  = false;
-        const auto status    = CLIm2ColKernel::validate(&input, &output, conv_size, PadStrideInfo(), has_bias);
+        const auto status    = opencl::kernels::ClIm2ColKernel::validate(&input, &output, conv_size, PadStrideInfo(), has_bias);
         ARM_COMPUTE_EXPECT(bool(status) == false, framework::LogLevel::ERRORS);
     }
 
@@ -128,13 +128,13 @@ TEST_CASE(Negative, framework::DatasetMode::ALL)
         const auto output    = TensorInfo(TensorShape(1U, 1U, 1U, 2U), 1, DataType::F32, DataLayout::NHWC);
         const auto conv_size = Size2D(9, 9);
         const bool has_bias  = false;
-        const auto status    = CLIm2ColKernel::validate(&input, &output, conv_size, PadStrideInfo(), has_bias);
+        const auto status    = opencl::kernels::ClIm2ColKernel::validate(&input, &output, conv_size, PadStrideInfo(), has_bias);
         ARM_COMPUTE_EXPECT(bool(status) == false, framework::LogLevel::ERRORS);
     }
 }
 
 template <typename T>
-using CLIm2ColFixture = Im2ColValidationFixture<CLTensor, CLAccessor, CLIm2Col, T, true>;
+using ClIm2ColFixture = Im2ColOpValidationFixture<CLTensor, CLAccessor, ClIm2Col, T, true>;
 
 TEST_SUITE(NHWC)
 
@@ -150,7 +150,7 @@ TEST_SUITE(NHWC)
  *  Kernel tested im2col3x3_nhwc
  */
 FIXTURE_DATA_TEST_CASE(W3x3,
-                       CLIm2ColFixture<float>,
+                       ClIm2ColFixture<float>,
                        framework::DatasetMode::ALL,
                        combine(combine(combine(combine(combine(combine(
                                                                    framework::dataset::make("InputShape",
@@ -180,7 +180,7 @@ framework::dataset::make("Groups", 1)))
  *  Kernel tested im2col9x9_nhwc
  */
 FIXTURE_DATA_TEST_CASE(W9x9,
-                       CLIm2ColFixture<float>,
+                       ClIm2ColFixture<float>,
                        framework::DatasetMode::ALL,
                        combine(combine(combine(combine(combine(combine(
                                                                    framework::dataset::make("InputShape",
@@ -210,7 +210,7 @@ framework::dataset::make("Groups", 1)))
  *  Kernel tested im2col_generic_nhwc
  */
 FIXTURE_DATA_TEST_CASE(Generic,
-                       CLIm2ColFixture<float>,
+                       ClIm2ColFixture<float>,
                        framework::DatasetMode::ALL,
                        combine(combine(combine(combine(combine(combine(
                                                                    framework::dataset::make("InputShape",
@@ -243,7 +243,7 @@ TEST_SUITE(NCHW)
  *  Kernel tested im2col1x1_stridex1_nchw
  */
 FIXTURE_DATA_TEST_CASE(W1x1_Stride1_NoPad,
-                       CLIm2ColFixture<float>,
+                       ClIm2ColFixture<float>,
                        framework::DatasetMode::ALL,
                        combine(combine(combine(combine(combine(combine(
                                                                    framework::dataset::make("InputShape", { TensorShape(4U, 4U, 3U, 2U), TensorShape(5U, 4U, 3U, 2U), TensorShape(3U, 4U, 3U, 2U) }),
@@ -267,7 +267,7 @@ FIXTURE_DATA_TEST_CASE(W1x1_Stride1_NoPad,
  *  Kernel tested im2col3x3_nchw
  */
 FIXTURE_DATA_TEST_CASE(W3x3,
-                       CLIm2ColFixture<float>,
+                       ClIm2ColFixture<float>,
                        framework::DatasetMode::ALL,
                        combine(combine(combine(combine(combine(combine(
                                                                    framework::dataset::make("InputShape", TensorShape(4U, 4U, 3U, 2U)),
@@ -291,7 +291,7 @@ FIXTURE_DATA_TEST_CASE(W3x3,
  *  Kernel tested im2col5x5_nchw
  */
 FIXTURE_DATA_TEST_CASE(W5x5,
-                       CLIm2ColFixture<float>,
+                       ClIm2ColFixture<float>,
                        framework::DatasetMode::ALL,
                        combine(combine(combine(combine(combine(combine(
                                                                    framework::dataset::make("InputShape", TensorShape(7U, 4U, 3U, 2U)),
@@ -317,7 +317,7 @@ FIXTURE_DATA_TEST_CASE(W5x5,
  * Kernel tested im2col11x11_padx0_pady0_nchw
  */
 FIXTURE_DATA_TEST_CASE(W11x11_NoPad,
-                       CLIm2ColFixture<float>,
+                       ClIm2ColFixture<float>,
                        framework::DatasetMode::ALL,
                        combine(combine(combine(combine(combine(combine(
                                                                    framework::dataset::make("InputShape", { TensorShape(11U, 11U, 2U, 2U), TensorShape(14U, 13U, 1U, 2U) }),
@@ -341,7 +341,7 @@ FIXTURE_DATA_TEST_CASE(W11x11_NoPad,
  * Kernel tested im2col_generic_padx0_pady0_nchw
  */
 FIXTURE_DATA_TEST_CASE(GenericZeroPad,
-                       CLIm2ColFixture<float>,
+                       ClIm2ColFixture<float>,
                        framework::DatasetMode::ALL,
                        combine(combine(combine(combine(combine(combine(
                                                                    framework::dataset::make("InputShape", TensorShape(13U, 11U, 2U, 2U)),
@@ -367,7 +367,7 @@ TEST_SUITE_END() // NCHW
  * Kernel tested im2col_generic_(nchw|nhwc)
  */
 FIXTURE_DATA_TEST_CASE(Generic,
-                       CLIm2ColFixture<float>,
+                       ClIm2ColFixture<float>,
                        framework::DatasetMode::ALL,
                        combine(combine(combine(combine(combine(combine(
                                                                    framework::dataset::make("InputShape", TensorShape(13U, 11U, 5U, 2U)),
@@ -393,7 +393,7 @@ FIXTURE_DATA_TEST_CASE(Generic,
  *  - im2col9x9_nhwc
  */
 FIXTURE_DATA_TEST_CASE(Quantized,
-                       CLIm2ColFixture<uint8_t>,
+                       ClIm2ColFixture<uint8_t>,
                        framework::DatasetMode::ALL,
                        combine(combine(combine(combine(combine(combine(
                                                                    framework::dataset::make("InputShape", TensorShape(13U, 11U, 11U, 2U)),
@@ -419,7 +419,7 @@ FIXTURE_DATA_TEST_CASE(Quantized,
  *  - im2col9x9_nhwc
  */
 FIXTURE_DATA_TEST_CASE(FP16,
-                       CLIm2ColFixture<half>,
+                       ClIm2ColFixture<half>,
                        framework::DatasetMode::ALL,
                        combine(combine(combine(combine(combine(combine(
                                                                    framework::dataset::make("InputShape", TensorShape(13U, 11U, 11U, 2U)),
diff --git a/tests/validation/CL/UNIT/DynamicTensor.cpp b/tests/validation/CL/UNIT/DynamicTensor.cpp
index ad2d4892ba..f83a92ec2f 100644
--- a/tests/validation/CL/UNIT/DynamicTensor.cpp
+++ b/tests/validation/CL/UNIT/DynamicTensor.cpp
@@ -29,7 +29,6 @@
 #include "arm_compute/runtime/MemoryManagerOnDemand.h"
 #include "arm_compute/runtime/PoolManager.h"
 #include "src/core/CL/kernels/CLFillBorderKernel.h"
-#include "src/core/CL/kernels/CLIm2ColKernel.h"
 #include "src/core/CL/kernels/CLL2NormalizeLayerKernel.h"
 #include "src/core/CL/kernels/CLReductionOperationKernel.h"
 #include "src/core/CL/kernels/CLWeightsReshapeKernel.h"
diff --git a/tests/validation/fixtures/Im2ColFixture.h b/tests/validation/fixtures/Im2ColFixture.h
index 38970116f6..a0732c3eb3 100644
--- a/tests/validation/fixtures/Im2ColFixture.h
+++ b/tests/validation/fixtures/Im2ColFixture.h
@@ -134,92 +134,6 @@ protected:
     bool             _has_bias{};
     unsigned int     _num_groups{};
 };
-
-template <typename TensorType, typename AccessorType, typename FunctionType, typename T, bool batch_size_on_z>
-class Im2ColValidationFixture : public framework::Fixture
-{
-public:
-    template <typename...>
-    void setup(TensorShape input_shape, DataType data_type, const Size2D &kernel_dims, const PadStrideInfo &conv_info, const QuantizationInfo &quant_info, const DataLayout &data_layout,
-               unsigned int num_groups)
-    {
-        _kernel_dims = kernel_dims;
-        _conv_info   = conv_info;
-        _quant_info  = quant_info;
-        _data_layout = data_layout;
-        _has_bias    = data_type != DataType::QASYMM8;
-        _num_groups  = num_groups;
-
-        if(_data_layout == DataLayout::NHWC)
-        {
-            permute(input_shape, PermutationVector(2U, 0U, 1U));
-        }
-
-        TensorInfo input_info(input_shape, 1, data_type);
-        input_info.set_data_layout(_data_layout);
-
-        const TensorShape output_shape = compute_im2col_conv_shape(&input_info, _kernel_dims, _conv_info, _has_bias, Size2D(1U, 1U), batch_size_on_z && _num_groups == 1, _num_groups);
-        _target                        = compute_target(input_shape, output_shape, data_type);
-
-        compute_reference(input_shape, output_shape, data_type);
-    }
-
-protected:
-    template <typename U>
-    void fill(U &&tensor)
-    {
-        library->fill_tensor_uniform(tensor, 0);
-    }
-
-    TensorType compute_target(const TensorShape &input_shape, const TensorShape &output_shape, DataType data_type)
-    {
-        // Create tensors
-        TensorType src = create_tensor<TensorType>(input_shape, data_type, 1, _quant_info, _data_layout);
-        TensorType dst = create_tensor<TensorType>(output_shape, data_type, 1, _quant_info);
-
-        // Create and configure function
-        FunctionType im2col_func;
-        im2col_func.configure(&src, &dst, _kernel_dims, _conv_info, _has_bias, Size2D(1U, 1U), _num_groups);
-
-        ARM_COMPUTE_ASSERT(src.info()->is_resizable());
-        ARM_COMPUTE_ASSERT(dst.info()->is_resizable());
-
-        // Allocate tensors
-        src.allocator()->allocate();
-        dst.allocator()->allocate();
-
-        ARM_COMPUTE_ASSERT(!src.info()->is_resizable());
-        ARM_COMPUTE_ASSERT(!dst.info()->is_resizable());
-
-        // Fill tensors
-        fill(AccessorType(src));
-
-        // Compute function
-        im2col_func.run();
-
-        return dst;
-    }
-
-    void compute_reference(const TensorShape &input_shape, const TensorShape &output_shape, DataType data_type)
-    {
-        // Create reference
-        SimpleTensor<T> src{ input_shape, data_type, 1, _quant_info, _data_layout };
-        _reference = SimpleTensor<T>(output_shape, data_type, 1, _quant_info, DataLayout::NCHW);
-
-        // Fill reference
-        fill(src);
-
-        reference::im2col<T>(src, _reference, _kernel_dims, _conv_info, _has_bias, _num_groups);
-    }
-    TensorType       _target{};
-    SimpleTensor<T>  _reference{};
-    Size2D           _kernel_dims{};
-    PadStrideInfo    _conv_info{};
-    DataLayout       _data_layout{};
-    QuantizationInfo _quant_info{};
-    bool             _has_bias{};
-    unsigned int     _num_groups{};
-};
 } // namespace validation
 } // namespace test
 } // namespace arm_compute