Rework OpenCL Depthwise Convolution

- Remove dedicated kernels for NCHW. Now we only use NHWC with permute
- Remove specialized kernels for 3x3 NHWC
- Simplify CLDepthwiseConvolutionLayer.cpp to call just the native
  implementation for both floating-point and quantized data types
- Develop two parametric opencl kernels for depthwise convolution layer NHWC
  (floating-point and quantized)
- Add support to export the weights to cl_image
- Extend test for depthwise convolution on opencl

Resolves COMPMID-4417

Change-Id: I253dd5d959a70783c82e62b1771a5e9f91621cb0
Signed-off-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/5806
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Giorgio Arena <giorgio.arena@arm.com>

6 files changed, 139 insertions, 1028 deletions

diff --git a/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.cpp b/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.cpp
deleted file mode 100644
index dda70d2231..0000000000
--- a/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.cpp
+++ /dev/null
@@ -1,432 +0,0 @@
-/*
- * Copyright (c) 2018-2021 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.
- */
-#include "src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.h"
-
-#include "arm_compute/core/CL/CLHelpers.h"
-#include "arm_compute/core/CL/CLKernelLibrary.h"
-#include "arm_compute/core/CL/ICLTensor.h"
-#include "arm_compute/core/Helpers.h"
-#include "arm_compute/core/TensorInfo.h"
-#include "arm_compute/core/Utils.h"
-#include "arm_compute/core/utils/misc/ShapeCalculator.h"
-#include "arm_compute/core/utils/quantization/AsymmHelpers.h"
-#include "src/core/AccessWindowStatic.h"
-#include "src/core/CL/CLValidate.h"
-#include "src/core/CL/ICLKernel.h"
-#include "src/core/helpers/AutoConfiguration.h"
-#include "src/core/helpers/WindowHelpers.h"
-#include "support/StringSupport.h"
-
-namespace arm_compute
-{
-using namespace arm_compute::misc::shape_calculator;
-
-namespace
-{
-Status validate_arguments(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output,
-                          const PadStrideInfo &conv_info, unsigned int depth_multiplier, const ActivationLayerInfo &act_info, const Size2D dilation,
-                          const ITensorInfo *output_multipliers, const ITensorInfo *output_shifts)
-{
-    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_MSG((act_info.enabled()) && (input->data_type() == DataType::QASYMM8 || input->data_type() == DataType::QASYMM8_SIGNED)
-                                    && (act_info.activation() != ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU)
-                                    && (act_info.activation() != ActivationLayerInfo::ActivationFunction::BOUNDED_RELU)
-                                    && (act_info.activation() != ActivationLayerInfo::ActivationFunction::RELU)
-                                    && (act_info.activation() != ActivationLayerInfo::ActivationFunction::LOGISTIC),
-                                    "For QASYMM8 only logistic, relu, lower bounded relu and lower-upper bounded relu are supported");
-    ARM_COMPUTE_RETURN_ERROR_ON(weights->dimension(0) != 3 || weights->dimension(1) != 3);
-    ARM_COMPUTE_RETURN_ERROR_ON(conv_info.stride().first < 1 || conv_info.stride().first > 3);
-
-    ARM_COMPUTE_RETURN_ERROR_ON((dilation.x() < 1) || (dilation.y() < 1));
-
-    const bool is_qasymm = is_data_type_quantized_asymmetric(input->data_type());
-
-    if(biases != nullptr)
-    {
-        if(is_qasymm)
-        {
-            ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(biases, 1, DataType::S32);
-        }
-        else
-        {
-            ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(weights, biases);
-        }
-        ARM_COMPUTE_RETURN_ERROR_ON((biases->dimension(0) != weights->dimension(2)) && (weights->dimension(2) != 1 || biases->dimension(0) != weights->dimension(3)));
-        ARM_COMPUTE_RETURN_ERROR_ON(biases->num_dimensions() > 1);
-    }
-
-    if(is_qasymm)
-    {
-        ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(output_multipliers, output_shifts);
-        ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output_multipliers, 1, DataType::S32);
-        ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output_shifts, 1, DataType::S32);
-        ARM_COMPUTE_RETURN_ERROR_ON(output_multipliers->num_dimensions() > 1);
-        ARM_COMPUTE_RETURN_ERROR_ON(output_shifts->num_dimensions() > 1);
-
-        if(is_data_type_quantized_per_channel(weights->data_type()))
-        {
-            ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(weights, 1, DataType::QSYMM8_PER_CHANNEL);
-            ARM_COMPUTE_RETURN_ERROR_ON(weights->dimension(2) != output_multipliers->dimension(0));
-            ARM_COMPUTE_RETURN_ERROR_ON(weights->dimension(2) != output_shifts->dimension(0));
-        }
-        else
-        {
-            ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights);
-            ARM_COMPUTE_RETURN_ERROR_ON(1 != output_multipliers->dimension(0));
-            ARM_COMPUTE_RETURN_ERROR_ON(1 != output_shifts->dimension(0));
-        }
-    }
-    else
-    {
-        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights);
-    }
-
-    if(output->total_size() != 0)
-    {
-        const ConvolutionInfo info{ conv_info, depth_multiplier, ActivationLayerInfo(), dilation };
-        const TensorShape     output_shape = compute_depthwise_convolution_shape(*input, *weights, info);
-        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), output_shape);
-    }
-
-    return Status{};
-}
-
-std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *weights, ITensorInfo *output, const PadStrideInfo &conv_info,
-                                                        unsigned int depth_multiplier, std::string &kernel_name, const Size2D dilation)
-{
-    // Output auto inizialitation if not yet initialized
-    const ConvolutionInfo info
-    {
-        conv_info, depth_multiplier, ActivationLayerInfo(), dilation
-    };
-    const TensorShape output_shape = compute_depthwise_convolution_shape(*input, *weights, info);
-    auto_init_if_empty(*output, input->clone()->set_tensor_shape(output_shape).set_quantization_info(output->quantization_info()));
-
-    const unsigned int conv_stride_x = conv_info.stride().first;
-    const unsigned int conv_stride_y = conv_info.stride().second;
-    const bool         is_qasymm     = is_data_type_quantized_asymmetric(input->data_type());
-
-    // Configure kernel window
-    unsigned int num_elems_read_per_iteration_x    = 0;
-    unsigned int num_elems_read_per_iteration_y    = 0;
-    unsigned int num_elems_written_per_iteration_x = 0;
-    unsigned int num_elems_written_per_iteration_y = 0;
-
-    if(input->data_type() == DataType::F16)
-    {
-        kernel_name                       = "depthwise_convolution_3x3_f16";
-        num_elems_written_per_iteration_x = 8 / data_size_from_type(input->data_type());
-        num_elems_written_per_iteration_y = 1;
-        num_elems_read_per_iteration_y    = 3;
-        switch(conv_stride_x)
-        {
-            case 1:
-                num_elems_read_per_iteration_x = 8;
-                break;
-            case 2:
-                num_elems_read_per_iteration_x = 9;
-                break;
-            case 3:
-                num_elems_read_per_iteration_x = 16;
-                break;
-            default:
-                num_elems_read_per_iteration_x = 3 + (num_elems_written_per_iteration_x - 1) * conv_stride_x;
-                break;
-        }
-        if(conv_stride_x == 1 && conv_stride_y == 1)
-        {
-            kernel_name                       = "depthwise_convolution_3x3_stridex1_stridey1_f16";
-            num_elems_read_per_iteration_x    = 8;
-            num_elems_written_per_iteration_x = 4;
-            num_elems_read_per_iteration_y    = 6;
-            num_elems_written_per_iteration_y = 4;
-        }
-        else if(conv_stride_x == 2 && conv_stride_y == 2)
-        {
-            kernel_name                       = "depthwise_convolution_3x3_stridex2_stridey2_f16";
-            num_elems_read_per_iteration_x    = 10;
-            num_elems_written_per_iteration_x = 4;
-            num_elems_read_per_iteration_y    = 5;
-            num_elems_written_per_iteration_y = 2;
-        }
-    }
-    else if(input->data_type() == DataType::F32)
-    {
-        if(conv_stride_x == 1 && conv_stride_y == 1)
-        {
-            kernel_name                       = "depthwise_convolution_3x3_stridex1_stridey1_f32";
-            num_elems_read_per_iteration_x    = 4;
-            num_elems_read_per_iteration_y    = 6;
-            num_elems_written_per_iteration_x = 2;
-            num_elems_written_per_iteration_y = 4;
-        }
-        else if(conv_stride_x == 2 && conv_stride_y == 2)
-        {
-            kernel_name                       = "depthwise_convolution_3x3_stridex2_stridey2_f32";
-            num_elems_read_per_iteration_x    = 6;
-            num_elems_read_per_iteration_y    = 5;
-            num_elems_written_per_iteration_x = 2;
-            num_elems_written_per_iteration_y = 2;
-        }
-        else
-        {
-            kernel_name                       = "depthwise_convolution_3x3";
-            num_elems_written_per_iteration_x = 8 / data_size_from_type(input->data_type());
-            num_elems_written_per_iteration_y = 1;
-            num_elems_read_per_iteration_x    = 3 + (num_elems_written_per_iteration_x - 1) * conv_stride_x;
-            num_elems_read_per_iteration_y    = 3;
-        }
-    }
-    else
-    {
-        const bool is_dot8_supported = dot8_supported(CLKernelLibrary::get().get_device()) && !is_data_type_quantized_per_channel(weights->data_type());
-
-        kernel_name = is_qasymm ? "dwc_3x3_native_quantized8" : "depthwise_convolution_3x3";
-        kernel_name += (is_qasymm && is_dot8_supported ? "_dot8" : "");
-        kernel_name += (is_qasymm ? "_nchw" : "");
-
-        num_elems_written_per_iteration_x = 8 / data_size_from_type(input->data_type());
-        num_elems_written_per_iteration_y = (is_qasymm && conv_stride_y == 1 && dilation.y() == 1) ? 2 : 1;
-        num_elems_read_per_iteration_x    = 3 + (num_elems_written_per_iteration_x - 1) * conv_stride_x + (conv_stride_x > 1 ? 1 : 0);
-        num_elems_read_per_iteration_y    = num_elems_written_per_iteration_y + 2;
-    }
-    // The OpenCL routine convolution1x3 does loadn(addr), loadn(addr + dilation_x) and loadn(addr + 2 * dilation_x) on the input.
-    // Each of the three convolution1x3 gets called by passing addr, (addr + dilation_y) and (addr + 2 * dilation_y)
-    // Hence we must add 2 * dilation.x/y() to the number of elements read in those axes per thread
-    num_elems_read_per_iteration_x += 2 * dilation.x();
-    num_elems_read_per_iteration_y += 2 * dilation.y();
-
-    // Create window and update padding
-    Window win = calculate_max_window(*output, Steps(num_elems_written_per_iteration_x, num_elems_written_per_iteration_y));
-
-    AccessWindowRectangle input_access(input, -conv_info.pad_left(), -conv_info.pad_top(),
-                                       num_elems_read_per_iteration_x, num_elems_read_per_iteration_y,
-                                       conv_stride_x, conv_stride_y);
-    AccessWindowStatic    weights_access(weights, 0, 0, 3, 3);
-    AccessWindowRectangle output_access(output, 0, 0, num_elems_written_per_iteration_x, num_elems_written_per_iteration_y);
-
-    bool window_changed = update_window_and_padding(win, input_access, weights_access, output_access);
-
-    Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
-    return std::make_pair(err, win);
-}
-} // namespace
-
-CLDepthwiseConvolutionLayer3x3NCHWKernel::CLDepthwiseConvolutionLayer3x3NCHWKernel()
-    : _border_size(0), _input(), _output(), _weights(), _biases(), _conv_stride_y(1), _output_multipliers(), _output_shifts(), _is_quantized(false), _conv_stride_x(0), _conv_pad_top(0), _conv_pad_left(0)
-{
-}
-
-BorderSize CLDepthwiseConvolutionLayer3x3NCHWKernel::border_size() const
-{
-    return _border_size;
-}
-
-void CLDepthwiseConvolutionLayer3x3NCHWKernel::configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output,
-                                                         const PadStrideInfo &conv_info, unsigned int depth_multiplier, ActivationLayerInfo act_info, const Size2D &dilation,
-                                                         const ICLTensor *output_multipliers, const ICLTensor *output_shifts)
-{
-    configure(CLKernelLibrary::get().get_compile_context(), input, weights, biases, output, conv_info, depth_multiplier, act_info, dilation, output_multipliers, output_shifts);
-}
-
-void CLDepthwiseConvolutionLayer3x3NCHWKernel::configure(const CLCompileContext &compile_context, const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output,
-                                                         const PadStrideInfo &conv_info, unsigned int depth_multiplier, ActivationLayerInfo act_info, const Size2D &dilation,
-                                                         const ICLTensor *output_multipliers, const ICLTensor *output_shifts)
-{
-    ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output);
-    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), weights->info(), (biases != nullptr) ? biases->info() : nullptr, output->info(),
-                                                  conv_info, depth_multiplier, act_info, dilation,
-                                                  (output_multipliers != nullptr) ? output_multipliers->info() : nullptr,
-                                                  (output_shifts != nullptr) ? output_shifts->info() : nullptr));
-
-    _input              = input;
-    _output             = output;
-    _weights            = weights;
-    _biases             = biases;
-    _conv_stride_x      = conv_info.stride().first;
-    _conv_stride_y      = conv_info.stride().second;
-    _conv_pad_left      = conv_info.pad_left();
-    _conv_pad_top       = conv_info.pad_top();
-    _output_multipliers = output_multipliers;
-    _output_shifts      = output_shifts;
-    _is_quantized       = is_data_type_quantized_asymmetric(input->info()->data_type());
-
-    // Configure kernel window
-    std::string kernel_name;
-
-    auto win_config = validate_and_configure_window(input->info(), weights->info(), output->info(), conv_info, depth_multiplier, kernel_name, dilation);
-    ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
-    ICLKernel::configure_internal(win_config.second);
-
-    _border_size = BorderSize(input->info()->padding());
-
-    // Set build options
-    CLBuildOptions build_opts;
-    build_opts.add_option("-DACTIVATION_TYPE=" + lower_string(string_from_activation_func(act_info.activation())));
-    build_opts.add_option("-DDST_CHANNELS=" + support::cpp11::to_string(_output->info()->tensor_shape().z()));
-    build_opts.add_option("-DDEPTH_MULTIPLIER=" + support::cpp11::to_string(depth_multiplier));
-    build_opts.add_option("-DCONV_STRIDE_X=" + support::cpp11::to_string(_conv_stride_x));
-    build_opts.add_option("-DDILATION_X=" + support::cpp11::to_string(dilation.x()));
-    build_opts.add_option("-DDILATION_Y=" + support::cpp11::to_string(dilation.y()));
-    build_opts.add_option_if(_biases != nullptr, "-DHAS_BIAS");
-
-    if(_is_quantized)
-    {
-        const UniformQuantizationInfo iq_info = _input->info()->quantization_info().uniform();
-        const UniformQuantizationInfo wq_info = _weights->info()->quantization_info().uniform();
-        const UniformQuantizationInfo oq_info = _output->info()->quantization_info().uniform();
-
-        const bool is_quantized_per_channel = is_data_type_quantized_per_channel(weights->info()->data_type());
-        const bool is_dot8_supported        = dot8_supported(CLKernelLibrary::get().get_device()) && !is_quantized_per_channel;
-        build_opts.add_option("-DCONV_STRIDE_Y=" + support::cpp11::to_string(_conv_stride_y));
-        build_opts.add_option("-DINPUT_OFFSET=" + support::cpp11::to_string(-iq_info.offset));
-        build_opts.add_option("-DWEIGHTS_OFFSET=" + support::cpp11::to_string(-wq_info.offset));
-        build_opts.add_option("-DOUTPUT_OFFSET=" + support::cpp11::to_string(oq_info.offset));
-        build_opts.add_option("-DK_OFFSET=" + support::cpp11::to_string(9 * iq_info.offset * wq_info.offset));
-        build_opts.add_option_if(is_quantized_per_channel, "-DPER_CHANNEL_QUANTIZATION");
-        build_opts.add_option_if(is_dot8_supported, "-DIS_DOT8");
-
-        // Compute non-per-channel multiplier and shift anyway to make OpenCL kernel simpler
-        float multiplier        = iq_info.scale * wq_info.scale / oq_info.scale;
-        int   output_multiplier = 0;
-        int   output_shift      = 0;
-        quantization::calculate_quantized_multiplier(multiplier, &output_multiplier, &output_shift);
-        build_opts.add_option("-DOUTPUT_MULTIPLIER=" + support::cpp11::to_string(output_multiplier));
-        build_opts.add_option("-DOUTPUT_SHIFT=" + support::cpp11::to_string(output_shift));
-
-        if(act_info.enabled())
-        {
-            int a_val{};
-            int b_val{};
-            std::tie(b_val, a_val) = get_quantized_activation_min_max(act_info, input->info()->data_type(), oq_info);
-
-            const int o1 = oq_info.offset;
-
-            build_opts.add_option("-DA_VAL=" + support::cpp11::to_string(a_val));
-            build_opts.add_option("-DB_VAL=" + support::cpp11::to_string(b_val));
-            build_opts.add_option("-DCONST_0=" + support::cpp11::to_string(o1));
-
-            const float s1 = iq_info.scale;
-            build_opts.add_option("-DS1_VAL=" + float_to_string_with_full_precision(s1));
-            build_opts.add_option("-DO1_VAL=" + support::cpp11::to_string(o1));
-        }
-
-        build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type()));
-        build_opts.add_option("-DWEIGHTS_TYPE=" + get_cl_type_from_data_type(weights->info()->data_type()));
-        build_opts.add_option("-DWEIGHTS_PROMOTED_TYPE=" + get_cl_promoted_type_from_data_type(weights->info()->data_type()));
-    }
-    else
-    {
-        build_opts.add_option_if(act_info.enabled(), "-DA_VAL=" + float_to_string_with_full_precision(act_info.a()));
-        build_opts.add_option_if(act_info.enabled(), "-DB_VAL=" + float_to_string_with_full_precision(act_info.b()));
-        build_opts.add_option_if(act_info.enabled(), "-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type()));
-        build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(win_config.second.x().step()));
-    }
-
-    build_opts.add_option_if(input->info()->data_type() == DataType::F16, "-DIS_F16");
-    build_opts.add_option_if(input->info()->data_type() == DataType::F32, "-DIS_F32");
-
-    _kernel = create_kernel(compile_context, kernel_name, build_opts.options());
-
-    // Set config_id for enabling LWS tuning
-    _config_id = kernel_name;
-    _config_id += "_";
-    _config_id += lower_string(string_from_data_type(input->info()->data_type()));
-    _config_id += "_";
-    _config_id += support::cpp11::to_string(input->info()->dimension(0));
-    _config_id += "_";
-    _config_id += support::cpp11::to_string(input->info()->dimension(1));
-    _config_id += "_";
-    _config_id += support::cpp11::to_string(input->info()->dimension(2));
-    _config_id += "_";
-    _config_id += support::cpp11::to_string(output->info()->dimension(0));
-    _config_id += "_";
-    _config_id += support::cpp11::to_string(output->info()->dimension(1));
-}
-
-Status CLDepthwiseConvolutionLayer3x3NCHWKernel::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output,
-                                                          const PadStrideInfo &conv_info, unsigned int depth_multiplier, ActivationLayerInfo act_info,
-                                                          const Size2D &dilation, const ITensorInfo *output_multipliers, const ITensorInfo *output_shifts)
-{
-    std::string kernel_name;
-    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, weights, biases, output, conv_info, depth_multiplier, act_info, dilation, output_multipliers, output_shifts));
-    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), weights->clone().get(), output->clone().get(),
-                                                              conv_info, depth_multiplier, kernel_name, dilation)
-                                .first);
-
-    return Status{};
-}
-
-void CLDepthwiseConvolutionLayer3x3NCHWKernel::run(const Window &window, cl::CommandQueue &queue)
-{
-    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
-    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(IKernel::window(), window);
-
-    Window collapsed = window.collapse_if_possible(ICLKernel::window(), Window::DimZ);
-
-    // Create input window and adjust
-    Window collapsed_in = collapsed;
-    collapsed_in.adjust(Window::DimX, -_conv_pad_left, true);
-    collapsed_in.adjust(Window::DimY, -_conv_pad_top, true);
-    collapsed_in.set_dimension_step(Window::DimX, collapsed_in.x().step() * _conv_stride_x);
-    collapsed_in.set_dimension_step(Window::DimY, collapsed_in.y().step() * _conv_stride_y);
-
-    Window slice_in      = collapsed_in.first_slice_window_3D();
-    Window slice_out     = collapsed.first_slice_window_3D();
-    Window slice_weights = window.first_slice_window_3D();
-    slice_weights.set_dimension_step(Window::DimX, 0);
-    slice_weights.set_dimension_step(Window::DimY, 0);
-
-    unsigned int idx = 3 * num_arguments_per_3D_tensor();
-
-    // Set output multipliers in case of quantized data type
-    if(_is_quantized)
-    {
-        Window slice;
-        slice.use_tensor_dimensions(_output_multipliers->info()->tensor_shape());
-        add_1D_tensor_argument(idx, _output_multipliers, slice);
-        add_1D_tensor_argument(idx, _output_shifts, slice);
-    }
-
-    // Set biases
-    if(_biases != nullptr)
-    {
-        Window slice_biases;
-        slice_biases.use_tensor_dimensions(_biases->info()->tensor_shape());
-        add_1D_tensor_argument(idx, _biases, slice_biases);
-    }
-
-    do
-    {
-        idx = 0;
-        add_3D_tensor_argument(idx, _input, slice_in);
-        add_3D_tensor_argument(idx, _output, slice_out);
-        add_3D_tensor_argument(idx, _weights, slice_weights);
-
-        enqueue(queue, *this, slice_out, lws_hint());
-    }
-    while(collapsed.slide_window_slice_3D(slice_out) && collapsed_in.slide_window_slice_3D(slice_in));
-}
-} // namespace arm_compute
diff --git a/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.h b/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.h
deleted file mode 100644
index c4e475f6f2..0000000000
--- a/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.h
+++ /dev/null
@@ -1,131 +0,0 @@
-/*
- * Copyright (c) 2018-2021 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_CLDEPTHWISECONVOLUTIONNCHWKERNEL3x3_H
-#define ARM_COMPUTE_CLDEPTHWISECONVOLUTIONNCHWKERNEL3x3_H
-
-#include "src/core/CL/ICLKernel.h"
-
-namespace arm_compute
-{
-class ICLTensor;
-
-/** Interface for the kernel to run a 3x3 depthwise convolution on a tensor when the data layout is NCHW.
- */
-class CLDepthwiseConvolutionLayer3x3NCHWKernel : public ICLKernel
-{
-public:
-    /** Default constructor */
-    CLDepthwiseConvolutionLayer3x3NCHWKernel();
-    /** Prevent instances of this class from being copied (As this class contains pointers) */
-    CLDepthwiseConvolutionLayer3x3NCHWKernel(const CLDepthwiseConvolutionLayer3x3NCHWKernel &) = delete;
-    /** Prevent instances of this class from being copied (As this class contains pointers) */
-    CLDepthwiseConvolutionLayer3x3NCHWKernel &operator=(const CLDepthwiseConvolutionLayer3x3NCHWKernel &) = delete;
-    /** Default Move Constructor. */
-    CLDepthwiseConvolutionLayer3x3NCHWKernel(CLDepthwiseConvolutionLayer3x3NCHWKernel &&) = default;
-    /** Default move assignment operator */
-    CLDepthwiseConvolutionLayer3x3NCHWKernel &operator=(CLDepthwiseConvolutionLayer3x3NCHWKernel &&) = default;
-    /** Initialize the function's source, destination, conv and border_size.
-     *
-     * @param[in]  input              Source tensor. DataType supported: QASYMM8/QASYMM8_SIGNED/F16/F32.
-     * @param[in]  weights            Weights tensor. A 3D tensor with dimensions [3, 3, IFM].
-     *                                Data type supported: Same as @p input or QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL when @p input is QASYMM8/QASYMM8_SIGNED.
-     * @param[in]  biases             Biases tensor. A 1D tensor with dimensions [IFM]. Must be nullptr if not needed.
-     *                                Data type supported: Same as @p input, S32 when input is QASYMM8/QASYMM8_SIGNED.
-     * @param[out] output             Destination tensor. Data type supported: Same as @p input.
-     * @param[in]  conv_info          Padding and stride information to use for the convolution.
-     * @param[in]  depth_multiplier   (Optional) Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1.
-     * @param[in]  act_info           (Optional) Activation layer information in case of a fused activation. Only RELU, BOUNDED_RELU and LU_BOUNDED_RELU for QASYMM8 supported.
-     * @param[in]  dilation           (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
-     * @param[in]  output_multipliers (Optional) Output multipliers tensor for quantized computations. In case of per-channel quantization,
-     *                                the number of multipliers must be equal to the number of filters (IFM). Supported data types: S32
-     * @param[in]  output_shifts      (Optional) Output shifts tensor for quantized computations. In case of per-channel quantization,
-     *                                the number of multipliers must be equal to the number of filters (IFM). Supported data types: S32
-     */
-    void configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info,
-                   unsigned int depth_multiplier = 1, ActivationLayerInfo act_info = ActivationLayerInfo(), const Size2D &dilation = Size2D(1U, 1U),
-                   const ICLTensor *output_multipliers = nullptr, const ICLTensor *output_shifts = nullptr);
-    /** Initialize the function's source, destination, conv and border_size.
-     *
-     * @param[in]  compile_context    The compile context to be used.
-     * @param[in]  input              Source tensor. DataType supported: QASYMM8/QASYMM8_SIGNED/F16/F32.
-     * @param[in]  weights            Weights tensor. A 3D tensor with dimensions [3, 3, IFM].
-     *                                Data type supported: Same as @p input or QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL when @p input is QASYMM8/QASYMM8_SIGNED.
-     * @param[in]  biases             Biases tensor. A 1D tensor with dimensions [IFM]. Must be nullptr if not needed.
-     *                                Data type supported: Same as @p input, S32 when input is QASYMM8/QASYMM8_SIGNED.
-     * @param[out] output             Destination tensor. Data type supported: Same as @p input.
-     * @param[in]  conv_info          Padding and stride information to use for the convolution.
-     * @param[in]  depth_multiplier   (Optional) Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1.
-     * @param[in]  act_info           (Optional) Activation layer information in case of a fused activation. Only RELU, BOUNDED_RELU and LU_BOUNDED_RELU for QASYMM8 supported.
-     * @param[in]  dilation           (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
-     * @param[in]  output_multipliers (Optional) Output multipliers tensor for quantized computations. In case of per-channel quantization,
-     *                                the number of multipliers must be equal to the number of filters (IFM). Supported data types: S32
-     * @param[in]  output_shifts      (Optional) Output shifts tensor for quantized computations. In case of per-channel quantization,
-     *                                the number of multipliers must be equal to the number of filters (IFM). Supported data types: S32
-     */
-    void configure(const CLCompileContext &compile_context, const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info,
-                   unsigned int depth_multiplier = 1, ActivationLayerInfo act_info = ActivationLayerInfo(), const Size2D &dilation = Size2D(1U, 1U),
-                   const ICLTensor *output_multipliers = nullptr, const ICLTensor *output_shifts = nullptr);
-    /** Static function to check if given info will lead to a valid configuration of @ref CLDepthwiseConvolutionLayer3x3NCHWKernel
-     *
-     * @param[in] input              Source tensor info. DataType supported: QASYMM8/QASYMM8_SIGNED/F16/F32.
-     * @param[in] weights            Weights tensor info. A 3D tensor with dimensions [3, 3, IFM].
-     *                               Data type supported: Same as @p input or QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL when @p input is QASYMM8/QASYMM8_SIGNED.
-     * @param[in] biases             Biases tensor info. A 1D tensor with dimensions [IFM]. Must be nullptr if not needed.
-     *                               Data type supported: Same as @p input, S32 when input is QASYMM8/QASYMM8_SIGNED.
-     * @param[in] output             Destination tensor. Data type supported: Same as @p input.
-     * @param[in] conv_info          Padding and stride information to use for the convolution.
-     * @param[in] depth_multiplier   (Optional) Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1.
-     * @param[in] act_info           (Optional) Activation layer information in case of a fused activation. Only RELU, BOUNDED_RELU and LU_BOUNDED_RELU are supported.
-     * @param[in] dilation           (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
-     * @param[in] output_multipliers (Optional) Output multipliers tensor info for quantized computations. In case of per-channel quantization,
-     *                               the number of multipliers must be equal to the number of filters (IFM). Supported data types: S32
-     * @param[in] output_shifts      (Optional) Output shifts tensor for quantized computations. In case of per-channel quantization,
-     *                               the number of multipliers must be equal to the number of filters (IFM). Supported data types: S32
-     *
-     * @return a status
-     */
-    static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info,
-                           unsigned int depth_multiplier = 1, ActivationLayerInfo act_info = ActivationLayerInfo(),
-                           const Size2D &dilation = Size2D(1U, 1U), const ITensorInfo *output_multipliers = nullptr, const ITensorInfo *output_shifts = nullptr);
-
-    void run(const Window &window, cl::CommandQueue &queue) override;
-    BorderSize border_size() const override;
-
-private:
-    BorderSize       _border_size;
-    const ICLTensor *_input;
-    ICLTensor       *_output;
-    const ICLTensor *_weights;
-    const ICLTensor *_biases;
-    unsigned int     _conv_stride_y;
-    const ICLTensor *_output_multipliers;
-    const ICLTensor *_output_shifts;
-    bool             _is_quantized;
-
-    unsigned int _conv_stride_x;
-    unsigned int _conv_pad_top;
-    unsigned int _conv_pad_left;
-};
-} // namespace arm_compute
-#endif /*ARM_COMPUTE_CLDEPTHWISECONVOLUTIONNCHWKERNEL3x3_H */
diff --git a/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.cpp b/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.cpp
deleted file mode 100644
index 91a2f5745a..0000000000
--- a/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.cpp
+++ /dev/null
@@ -1,238 +0,0 @@
-/*
- * Copyright (c) 2018-2021 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.
- */
-#include "src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.h"
-
-#include "arm_compute/core/CL/CLHelpers.h"
-#include "arm_compute/core/CL/CLKernelLibrary.h"
-#include "arm_compute/core/CL/ICLTensor.h"
-#include "arm_compute/core/Helpers.h"
-#include "arm_compute/core/TensorInfo.h"
-#include "arm_compute/core/Utils.h"
-#include "arm_compute/core/utils/misc/ShapeCalculator.h"
-#include "src/core/AccessWindowStatic.h"
-#include "src/core/CL/CLValidate.h"
-#include "src/core/CL/ICLKernel.h"
-#include "src/core/helpers/AutoConfiguration.h"
-#include "src/core/helpers/WindowHelpers.h"
-#include "support/StringSupport.h"
-
-namespace arm_compute
-{
-namespace
-{
-Status validate_arguments(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output,
-                          const PadStrideInfo &conv_info, unsigned int depth_multiplier, const ActivationLayerInfo &act_info, const Size2D &dilation)
-{
-    ARM_COMPUTE_UNUSED(act_info);
-    ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input);
-    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32);
-    ARM_COMPUTE_RETURN_ERROR_ON(depth_multiplier > 1);
-
-    ARM_COMPUTE_RETURN_ERROR_ON(conv_info.stride().first < 1);
-    ARM_COMPUTE_RETURN_ERROR_ON(std::max(conv_info.pad_top(), conv_info.pad_bottom()) > 4);
-
-    ARM_COMPUTE_RETURN_ERROR_ON((dilation.x() < 1) || (dilation.y() < 1));
-
-    const size_t weights_width  = 3;
-    const size_t weights_height = 3;
-
-    const ConvolutionInfo info{ conv_info, depth_multiplier, ActivationLayerInfo(), dilation };
-
-    const TensorShape output_shape = arm_compute::misc::shape_calculator::compute_depthwise_convolution_shape(
-                                         *input, TensorInfo(TensorShape(weights_width, weights_height), 1, weights->data_type()).set_data_layout(DataLayout::NCHW), info);
-
-    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights);
-    ARM_COMPUTE_RETURN_ERROR_ON((weights->dimension(1) != weights_width) || (weights->dimension(2) != weights_height));
-
-    if(biases != nullptr)
-    {
-        ARM_COMPUTE_RETURN_ERROR_ON(biases->dimension(0) != output_shape[0]);
-        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(weights, biases);
-
-        ARM_COMPUTE_RETURN_ERROR_ON(biases->num_dimensions() > 1);
-    }
-
-    if(output->total_size() != 0)
-    {
-        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), output_shape);
-    }
-
-    return Status{};
-}
-
-std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *weights, ITensorInfo *bias, ITensorInfo *output,
-                                                        const PadStrideInfo &conv_info, unsigned int depth_multiplier, const Size2D &dilation)
-{
-    ARM_COMPUTE_UNUSED(weights, bias);
-    ARM_COMPUTE_UNUSED(depth_multiplier);
-
-    const bool   is_stride_1_dilation_1           = ((conv_info.stride().first == conv_info.stride().second) && (conv_info.stride().first == 1) && dilation.x() == 1 && dilation.y() == 1);
-    unsigned int num_rows_processed_per_iteration = is_stride_1_dilation_1 ? 2 : 1;
-
-    Window win{};
-    Status err{};
-
-    unsigned int num_elems_accessed_per_iteration = adjust_vec_size(4 / input->element_size(), input->dimension(0));
-    win                                           = calculate_max_window(*output, Steps(num_elems_accessed_per_iteration, num_rows_processed_per_iteration));
-
-    return std::make_pair(err, win);
-}
-} // namespace
-
-CLDepthwiseConvolutionLayer3x3NHWCKernel::CLDepthwiseConvolutionLayer3x3NHWCKernel()
-    : _input(), _output(), _weights(), _biases(), _num_planes_processed_per_iteration(1)
-{
-}
-
-void CLDepthwiseConvolutionLayer3x3NHWCKernel::configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output,
-                                                         const PadStrideInfo &conv_info, unsigned int depth_multiplier, ActivationLayerInfo act_info, const Size2D &dilation)
-{
-    configure(CLKernelLibrary::get().get_compile_context(), input, weights, biases, output, conv_info, depth_multiplier, act_info, dilation);
-}
-
-void CLDepthwiseConvolutionLayer3x3NHWCKernel::configure(const CLCompileContext &compile_context, const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output,
-                                                         const PadStrideInfo &conv_info, unsigned int depth_multiplier, ActivationLayerInfo act_info, const Size2D &dilation)
-{
-    ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output);
-    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), weights->info(), (biases != nullptr) ? biases->info() : nullptr, output->info(),
-                                                  conv_info, depth_multiplier, act_info, dilation));
-
-    auto padding_info = get_padding_info({ input, weights, biases, output });
-
-    auto win_config = validate_and_configure_window(input->info(), weights->info(), biases != nullptr ? biases->info() : nullptr, output->info(),
-                                                    conv_info, depth_multiplier, dilation);
-
-    const bool is_stride_1            = ((conv_info.stride().first == conv_info.stride().second) && (conv_info.stride().first == 1));
-    const bool is_stride_1_dilation_1 = (is_stride_1 && dilation.x() == 1 && dilation.y() == 1);
-
-    _input                              = input;
-    _output                             = output;
-    _weights                            = weights;
-    _biases                             = biases;
-    _num_planes_processed_per_iteration = is_stride_1_dilation_1 ? 2 : 1;
-
-    unsigned int num_elems_accessed_per_iteration = adjust_vec_size(4 / input->info()->element_size(), input->info()->dimension(0));
-    unsigned int num_rows_processed_per_iteration = is_stride_1_dilation_1 ? 2 : 1;
-
-    CLBuildOptions build_opts;
-    build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(_input->info()->data_type()));
-    build_opts.add_option("-DACTIVATION_TYPE=" + lower_string(string_from_activation_func(act_info.activation())));
-    build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_accessed_per_iteration));
-    build_opts.add_option("-DSRC_DIM_1=" + support::cpp11::to_string(_input->info()->dimension(1)));
-    build_opts.add_option("-DSRC_DIM_2=" + support::cpp11::to_string(_input->info()->dimension(2)));
-    build_opts.add_option("-DCONV_PAD_TOP=" + support::cpp11::to_string(conv_info.pad_top()));
-    build_opts.add_option("-DCONV_PAD_LEFT=" + support::cpp11::to_string(conv_info.pad_left()));
-    build_opts.add_option("-DPARTIAL_STORE_N0=" + support::cpp11::to_string(input->info()->dimension(0) % num_elems_accessed_per_iteration));
-    build_opts.add_option_if(_biases != nullptr, "-DHAS_BIAS");
-    build_opts.add_option_if(_input->info()->tensor_shape().total_size_upper(3) > 1,
-                             "-DDST_DEPTH=" + support::cpp11::to_string(static_cast<int>(std::ceil(_output->info()->dimension(2) / static_cast<float>(_num_planes_processed_per_iteration)))));
-    build_opts.add_option_if(act_info.enabled(), "-DA_VAL=" + float_to_string_with_full_precision(act_info.a()));
-    build_opts.add_option_if(act_info.enabled(), "-DB_VAL=" + float_to_string_with_full_precision(act_info.b()));
-
-    if(is_stride_1_dilation_1)
-    {
-        build_opts.add_option("-DNUM_ROWS_PROCESSED=" + support::cpp11::to_string(num_rows_processed_per_iteration));
-        build_opts.add_option("-DNUM_PLANES_PROCESSED=" + support::cpp11::to_string(_num_planes_processed_per_iteration));
-        build_opts.add_option("-DDST_DIM_1=" + support::cpp11::to_string(_output->info()->dimension(1)));
-        build_opts.add_option("-DDST_DIM_2=" + support::cpp11::to_string(_output->info()->dimension(2)));
-        build_opts.add_option("-DPARTIAL_STORE_M0=" + support::cpp11::to_string((input->info()->dimension(1) + conv_info.pad_left() + conv_info.pad_right()) % num_rows_processed_per_iteration));
-    }
-    else
-    {
-        build_opts.add_option("-DCONV_STRIDE_X=" + support::cpp11::to_string(conv_info.stride().first));
-        build_opts.add_option("-DCONV_STRIDE_Y=" + support::cpp11::to_string(conv_info.stride().second));
-        build_opts.add_option("-DDILATION_X=" + support::cpp11::to_string(dilation.x()));
-        build_opts.add_option("-DDILATION_Y=" + support::cpp11::to_string(dilation.y()));
-    }
-
-    // Create kernel
-    std::string kernel_name;
-    kernel_name = std::string("depthwise_convolution_3x3_nhwc");
-    kernel_name += (is_stride_1_dilation_1 ? "_stride1" : "");
-
-    ICLKernel::configure_internal(win_config.second);
-    _kernel = create_kernel(compile_context, kernel_name, build_opts.options());
-
-    ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info));
-
-    // Set config_id for enabling LWS tuning
-    _config_id = kernel_name;
-    _config_id += "_";
-    _config_id += support::cpp11::to_string(input->info()->dimension(0));
-    _config_id += "_";
-    _config_id += support::cpp11::to_string(input->info()->dimension(1));
-    _config_id += "_";
-    _config_id += support::cpp11::to_string(input->info()->dimension(2));
-    _config_id += "_";
-    _config_id += support::cpp11::to_string(output->info()->dimension(0));
-    _config_id += "_";
-    _config_id += support::cpp11::to_string(output->info()->dimension(1));
-    _config_id += "_";
-    _config_id += string_from_data_type(input->info()->data_type());
-}
-
-Status CLDepthwiseConvolutionLayer3x3NHWCKernel::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output,
-                                                          const PadStrideInfo &conv_info, unsigned int depth_multiplier, ActivationLayerInfo act_info, const Size2D &dilation)
-{
-    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, weights, biases, output, conv_info, depth_multiplier, act_info, dilation));
-    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), weights->clone().get(),
-                                                              biases != nullptr ? biases->clone().get() : nullptr,
-                                                              output->clone().get(), conv_info, depth_multiplier, dilation)
-                                .first);
-    return Status{};
-}
-
-void CLDepthwiseConvolutionLayer3x3NHWCKernel::run(const Window &window, cl::CommandQueue &queue)
-{
-    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
-    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(IKernel::window(), window);
-
-    const size_t total_batches = _input->info()->tensor_shape().total_size_upper(3);
-
-    Window win = window.collapse_if_possible(ICLKernel::window(), Window::DimZ);
-    win.set(Window::DimZ, Window::Dimension(0, std::ceil(_output->info()->dimension(2) / static_cast<float>(_num_planes_processed_per_iteration)) * total_batches, 1));
-
-    unsigned int idx = 2 * num_arguments_per_4D_tensor() + num_arguments_per_3D_tensor();
-
-    if(_biases != nullptr)
-    {
-        Window win_biases;
-        win_biases.use_tensor_dimensions(_biases->info()->tensor_shape());
-        win_biases.set_dimension_step(Window::DimX, window.x().step());
-        add_1D_tensor_argument(idx, _biases, win_biases);
-    }
-
-    Window slice = win.first_slice_window_4D();
-    do
-    {
-        unsigned int idx = 0;
-        add_4D_tensor_argument(idx, _input, slice);
-        add_4D_tensor_argument(idx, _output, slice);
-        add_3D_tensor_argument(idx, _weights, slice);
-
-        enqueue(queue, *this, slice, lws_hint());
-    }
-    while(win.slide_window_slice_4D(slice));
-}
-} // namespace arm_compute
diff --git a/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.h b/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.h
deleted file mode 100644
index ee47d98807..0000000000
--- a/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.h
+++ /dev/null
@@ -1,110 +0,0 @@
-/*
- * Copyright (c) 2018-2021 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_CLDEPTHWISECONVOLUTIONNHWCKERNEL3x3_H
-#define ARM_COMPUTE_CLDEPTHWISECONVOLUTIONNHWCKERNEL3x3_H
-
-#include "src/core/CL/ICLKernel.h"
-
-namespace arm_compute
-{
-class ICLTensor;
-
-/** Interface for the kernel to run a 3x3 depthwise convolution on a tensor when the data layout is NHWC.
- */
-class CLDepthwiseConvolutionLayer3x3NHWCKernel : public ICLKernel
-{
-public:
-    /** Default constructor */
-    CLDepthwiseConvolutionLayer3x3NHWCKernel();
-    /** Prevent instances of this class from being copied (As this class contains pointers) */
-    CLDepthwiseConvolutionLayer3x3NHWCKernel(const CLDepthwiseConvolutionLayer3x3NHWCKernel &) = delete;
-    /** Prevent instances of this class from being copied (As this class contains pointers) */
-    CLDepthwiseConvolutionLayer3x3NHWCKernel &operator=(const CLDepthwiseConvolutionLayer3x3NHWCKernel &) = delete;
-    /** Default Move Constructor. */
-    CLDepthwiseConvolutionLayer3x3NHWCKernel(CLDepthwiseConvolutionLayer3x3NHWCKernel &&) = default;
-    /** Default move assignment operator */
-    CLDepthwiseConvolutionLayer3x3NHWCKernel &operator=(CLDepthwiseConvolutionLayer3x3NHWCKernel &&) = default;
-    /** Default move assignment operator. */
-    /** Initialize the function's source, destination, conv and border_size.
-     *
-     * @param[in]  input            Source tensor. DataType supported: F16/F32.
-     * @param[in]  weights          Weights tensor. A 3D tensor with dimensions [IFM, 3, 3].
-     *                              Data type supported: Same as @p input.
-     * @param[in]  biases           Biases tensor. A 1D tensor with dimensions [IFM]. Must be nullptr if not needed.
-     *                              Data type supported: Same as @p input.
-     * @param[out] output           Destination tensor. Data type supported: Same as @p input.
-     * @param[in]  conv_info        Padding and stride information to use for the convolution.
-     * @param[in]  depth_multiplier (Optional) Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1.
-     * @param[in]  act_info         (Optional) Activation layer information in case of a fused activation. Only RELU, BOUNDED_RELU and LU_BOUNDED_RELU are supported.
-     * @param[in]  dilation         (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
-     */
-    void configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info,
-                   unsigned int depth_multiplier = 1, ActivationLayerInfo act_info = ActivationLayerInfo(), const Size2D &dilation = Size2D(1U, 1U));
-    /** Initialize the function's source, destination, conv and border_size.
-     *
-     * @param[in]  compile_context  The compile context to be used.
-     * @param[in]  input            Source tensor. DataType supported: F16/F32.
-     * @param[in]  weights          Weights tensor. A 3D tensor with dimensions [IFM, 3, 3].
-     *                              Data type supported: Same as @p input.
-     * @param[in]  biases           Biases tensor. A 1D tensor with dimensions [IFM]. Must be nullptr if not needed.
-     *                              Data type supported: Same as @p input.
-     * @param[out] output           Destination tensor. Data type supported: Same as @p input.
-     * @param[in]  conv_info        Padding and stride information to use for the convolution.
-     * @param[in]  depth_multiplier (Optional) Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1.
-     * @param[in]  act_info         (Optional) Activation layer information in case of a fused activation. Only RELU, BOUNDED_RELU and LU_BOUNDED_RELU are supported.
-     * @param[in]  dilation         (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
-     */
-    void configure(const CLCompileContext &compile_context, const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info,
-                   unsigned int depth_multiplier = 1, ActivationLayerInfo act_info = ActivationLayerInfo(), const Size2D &dilation = Size2D(1U, 1U));
-    /** Static function to check if given info will lead to a valid configuration of @ref CLDepthwiseConvolutionLayer3x3NHWCKernel
-     *
-     * @param[in] input            Source tensor info. DataType supported: F16/F32.
-     * @param[in] weights          Weights tensor info. A 3D tensor with dimensions [IFM, 3, 3].
-     *                             Data type supported: Same as @p input.
-     * @param[in] biases           Biases tensor info. A 1D tensor with dimensions [IFM]. Must be nullptr if not needed.
-     *                             Data type supported: Same as @p input.
-     * @param[in] output           Destination tensor info. Data type supported: Same as @p input.
-     * @param[in] conv_info        Padding and stride information to use for the convolution.
-     * @param[in] depth_multiplier (Optional) Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1.
-     * @param[in] act_info         (Optional) Activation layer information in case of a fused activation. Only RELU, BOUNDED_RELU and LU_BOUNDED_RELU are supported.
-     * @param[in] dilation         (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
-     *
-     * @return a status
-     */
-    static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info,
-                           unsigned int depth_multiplier = 1, ActivationLayerInfo act_info = ActivationLayerInfo(), const Size2D &dilation = Size2D(1U, 1U));
-
-    // Inherited methods overridden:
-    void run(const Window &window, cl::CommandQueue &queue) override;
-
-private:
-    const ICLTensor *_input;
-    ICLTensor       *_output;
-    const ICLTensor *_weights;
-    const ICLTensor *_biases;
-
-    unsigned int _num_planes_processed_per_iteration;
-};
-} // namespace arm_compute
-#endif /*ARM_COMPUTE_CLDEPTHWISECONVOLUTIONNHWCKERNEL3x3_H */
diff --git a/src/core/CL/kernels/CLDepthwiseConvolutionLayerNativeKernel.cpp b/src/core/CL/kernels/CLDepthwiseConvolutionLayerNativeKernel.cpp
index fcfa7f878d..67ca341b0d 100644
--- a/src/core/CL/kernels/CLDepthwiseConvolutionLayerNativeKernel.cpp
+++ b/src/core/CL/kernels/CLDepthwiseConvolutionLayerNativeKernel.cpp
@@ -31,8 +31,10 @@
 #include "arm_compute/core/Utils.h"
 #include "arm_compute/core/utils/misc/ShapeCalculator.h"
 #include "arm_compute/core/utils/quantization/AsymmHelpers.h"
+#include "src/core/CL/CLUtils.h"
 #include "src/core/CL/CLValidate.h"
 #include "src/core/CL/ICLKernel.h"
+#include "src/core/gpu/cl/kernels/gemm/ClGemmHelpers.h"
 #include "src/core/helpers/AutoConfiguration.h"
 #include "src/core/helpers/WindowHelpers.h"
 #include "support/StringSupport.h"
@@ -41,25 +43,28 @@ namespace arm_compute
 {
 namespace
 {
-Status validate_arguments(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const DWCWeightsKernelInfo &dwc_weights_info,
-                          const DWCKernelInfo &dwc_info, const PadStrideInfo &conv_info, unsigned int depth_multiplier, const Size2D &dilation,
+Status validate_arguments(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const DWCComputeKernelInfo &dwc_info,
+                          const ConvolutionInfo &conv_info,
                           const ITensorInfo *output_multipliers, const ITensorInfo *output_shifts)
 {
-    ARM_COMPUTE_UNUSED(dwc_info);
     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, weights, output);
     ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input);
     ARM_COMPUTE_RETURN_ERROR_ON_DATA_LAYOUT_NOT_IN(input, DataLayout::NHWC);
     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(depth_multiplier > 1 && dwc_weights_info.n0 != 1);
-    ARM_COMPUTE_RETURN_ERROR_ON(conv_info.stride().first < 1);
-    ARM_COMPUTE_RETURN_ERROR_ON(conv_info.stride().second < 1);
-    ARM_COMPUTE_RETURN_ERROR_ON((dilation.x() < 1) || (dilation.y() < 1));
+    ARM_COMPUTE_RETURN_ERROR_ON(conv_info.depth_multiplier > 1 && dwc_info.n0 != 1);
+    ARM_COMPUTE_RETURN_ERROR_ON(conv_info.pad_stride_info.stride().first > 1 && dwc_info.m0 != 1);
+    ARM_COMPUTE_RETURN_ERROR_ON(conv_info.dilation.x() > 1 && dwc_info.m0 != 1);
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG((dwc_info.export_weights_to_cl_image == true) && (export_weights_to_cl_image(weights) == false), "Export to cl_image not supported!");
+    ARM_COMPUTE_RETURN_ERROR_ON((dwc_info.export_weights_to_cl_image == true) && (conv_info.depth_multiplier > 1));
+    ARM_COMPUTE_RETURN_ERROR_ON((dwc_info.export_weights_to_cl_image == true) && ((dwc_info.n0 % 4) != 0));
+    ARM_COMPUTE_RETURN_ERROR_ON(conv_info.pad_stride_info.stride().first < 1);
+    ARM_COMPUTE_RETURN_ERROR_ON(conv_info.pad_stride_info.stride().second < 1);
+    ARM_COMPUTE_RETURN_ERROR_ON((conv_info.dilation.x() < 1) || (conv_info.dilation.y() < 1));
     const size_t idx_c = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::CHANNEL);
     ARM_COMPUTE_UNUSED(idx_c);
-    ARM_COMPUTE_RETURN_ERROR_ON(weights->dimension(idx_c) != (input->dimension(idx_c) * depth_multiplier));
+    ARM_COMPUTE_RETURN_ERROR_ON(weights->dimension(idx_c) != (input->dimension(idx_c) * conv_info.depth_multiplier));
 
-    const ConvolutionInfo info{ conv_info, depth_multiplier, ActivationLayerInfo(), dilation };
-    const TensorShape     output_shape = arm_compute::misc::shape_calculator::compute_depthwise_convolution_shape(*input, *weights, info);
+    const TensorShape output_shape = arm_compute::misc::shape_calculator::compute_depthwise_convolution_shape(*input, *weights, conv_info);
 
     const bool is_quantized = is_data_type_quantized(input->data_type());
 
@@ -134,111 +139,131 @@ CLDepthwiseConvolutionLayerNativeKernel::CLDepthwiseConvolutionLayerNativeKernel
       _depth_multiplier(1),
       _output_multipliers(nullptr),
       _output_shifts(nullptr),
+      _export_to_cl_image(false),
       _is_quantized(false)
 {
 }
 
-void CLDepthwiseConvolutionLayerNativeKernel::configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const DWCWeightsKernelInfo &dwc_weights_info,
-                                                        const DWCKernelInfo &dwc_info, const PadStrideInfo &conv_info, unsigned int depth_multiplier, const Size2D &dilation,
+void CLDepthwiseConvolutionLayerNativeKernel::configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output,
+                                                        const DWCComputeKernelInfo &dwc_info, const ConvolutionInfo &conv_info,
                                                         const ICLTensor *output_multipliers, const ICLTensor *output_shifts)
 {
-    configure(CLKernelLibrary::get().get_compile_context(), input, weights, biases, output, dwc_weights_info, dwc_info, conv_info, depth_multiplier, dilation, output_multipliers, output_shifts);
+    configure(CLKernelLibrary::get().get_compile_context(), input, weights, biases, output, dwc_info, conv_info, output_multipliers, output_shifts);
 }
 
 void CLDepthwiseConvolutionLayerNativeKernel::configure(const CLCompileContext &compile_context, const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output,
-                                                        const DWCWeightsKernelInfo &dwc_weights_info,
-                                                        const DWCKernelInfo &dwc_info, const PadStrideInfo &conv_info, unsigned int depth_multiplier, const Size2D &dilation,
+                                                        const DWCComputeKernelInfo &dwc_info, const ConvolutionInfo &conv_info,
                                                         const ICLTensor *output_multipliers, const ICLTensor *output_shifts)
 {
     ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output);
     ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), weights->info(), (biases != nullptr) ? biases->info() : nullptr, output->info(),
-                                                  dwc_weights_info, dwc_info, conv_info, depth_multiplier, dilation,
-                                                  (output_multipliers != nullptr) ? output_multipliers->info() : nullptr, (output_shifts != nullptr) ? output_shifts->info() : nullptr));
+                                                  dwc_info, conv_info, (output_multipliers != nullptr) ? output_multipliers->info() : nullptr, (output_shifts != nullptr) ? output_shifts->info() : nullptr));
 
     auto padding_info = get_padding_info({ input, output });
 
-    const ConvolutionInfo info{ conv_info, depth_multiplier, ActivationLayerInfo(), dilation };
-    const TensorShape     output_shape = arm_compute::misc::shape_calculator::compute_depthwise_convolution_shape(*(input->info()), *(weights->info()), info);
+    const TensorShape output_shape = arm_compute::misc::shape_calculator::compute_depthwise_convolution_shape(*(input->info()), *(weights->info()), conv_info);
     auto_init_if_empty(*(output->info()), input->info()->clone()->set_tensor_shape(output_shape).set_quantization_info(output->info()->quantization_info()));
 
     _input              = input;
     _output             = output;
     _weights            = weights;
     _biases             = biases;
-    _depth_multiplier   = depth_multiplier;
+    _depth_multiplier   = conv_info.depth_multiplier;
     _output_multipliers = output_multipliers;
     _output_shifts      = output_shifts;
+    _export_to_cl_image = dwc_info.export_weights_to_cl_image;
     _is_quantized       = is_data_type_quantized(input->info()->data_type());
 
-    const unsigned int n0 = adjust_vec_size(dwc_weights_info.n0, input->info()->dimension(0));
+    const unsigned int n0          = adjust_vec_size(dwc_info.n0, input->info()->dimension(0));
+    const unsigned int m0          = std::min(dwc_info.m0, (unsigned int)output->info()->dimension(1));
+    std::string        kernel_name = "";
 
     CLBuildOptions build_opts;
-    build_opts.add_option_if(_biases != nullptr, "-DHAS_BIAS");
-    build_opts.add_option_if(_input->info()->tensor_shape().total_size_upper(3) > 1, "-DDST_DEPTH=" + support::cpp11::to_string(static_cast<int>(_output->info()->dimension(2))));
-    build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(_input->info()->data_type()));
-    build_opts.add_option("-DACTIVATION_TYPE=" + lower_string(string_from_activation_func(dwc_info.activation_info.activation())));
-    build_opts.add_option("-DDEPTH_MULTIPLIER=" + support::cpp11::to_string(depth_multiplier));
+
+    // Update the padding for the weights tensor if we can export to cl_image
+    if(_export_to_cl_image)
+    {
+        arm_compute::opencl::kernels::gemm::update_padding_for_cl_image(weights->info());
+    }
+
+    build_opts.add_option("-cl-fast-relaxed-math");
+    build_opts.add_option("-DACTIVATION_TYPE=" + lower_string(string_from_activation_func(conv_info.act_info.activation())));
+    build_opts.add_option("-DDEPTH_MULTIPLIER=" + support::cpp11::to_string(conv_info.depth_multiplier));
+    build_opts.add_option("-DSRC_TENSOR_TYPE=BUFFER");
+    build_opts.add_option("-DSRC_WIDTH=" + support::cpp11::to_string(_input->info()->dimension(1)));
+    build_opts.add_option("-DSRC_HEIGHT=" + support::cpp11::to_string(_input->info()->dimension(2)));
+    // Note: SRC_DATA_TYPE must have the same data type of WEI_DATA_TYPE. In quantized, we could
+    // have a case where the data types for the activation and weights are different. However, since the implementation
+    // only works when both have same data type, we have to change the offset to take into account this aspect
+    build_opts.add_option("-DSRC_DATA_TYPE=" + get_cl_type_from_data_type(_weights->info()->data_type()));
+    build_opts.add_option("-DDST_TENSOR_TYPE=BUFFER");
+    build_opts.add_option("-DDST_WIDTH=" + support::cpp11::to_string(_output->info()->dimension(1)));
+    build_opts.add_option("-DDST_HEIGHT=" + support::cpp11::to_string(_output->info()->dimension(2)));
+    build_opts.add_option("-DDST_DATA_TYPE=" + get_cl_type_from_data_type(_output->info()->data_type()));
+    build_opts.add_option_if_else(_export_to_cl_image, "-DWEI_TENSOR_TYPE=IMAGE", "-DWEI_TENSOR_TYPE=BUFFER");
+    build_opts.add_option("-DWEI_WIDTH=" + support::cpp11::to_string(weights->info()->dimension(1)));
+    build_opts.add_option("-DWEI_HEIGHT=" + support::cpp11::to_string(weights->info()->dimension(2)));
+    build_opts.add_option("-DWEI_DATA_TYPE=" + get_cl_type_from_data_type(weights->info()->data_type()));
+    build_opts.add_option("-DPAD_TOP=" + support::cpp11::to_string(conv_info.pad_stride_info.pad_top()));
+    build_opts.add_option("-DPAD_LEFT=" + support::cpp11::to_string(conv_info.pad_stride_info.pad_left()));
+    build_opts.add_option("-DSTRIDE_X=" + support::cpp11::to_string(conv_info.pad_stride_info.stride().first));
+    build_opts.add_option("-DSTRIDE_Y=" + support::cpp11::to_string(conv_info.pad_stride_info.stride().second));
+    build_opts.add_option("-DDILATION_X=" + support::cpp11::to_string(conv_info.dilation.x()));
+    build_opts.add_option("-DDILATION_Y=" + support::cpp11::to_string(conv_info.dilation.y()));
     build_opts.add_option("-DN0=" + support::cpp11::to_string(n0));
-    build_opts.add_option("-DSRC_DIM1=" + support::cpp11::to_string(_input->info()->dimension(1)));
-    build_opts.add_option("-DSRC_DIM2=" + support::cpp11::to_string(_input->info()->dimension(2)));
-    build_opts.add_option("-DKERNEL_WIDTH=" + support::cpp11::to_string(weights->info()->dimension(1)));
-    build_opts.add_option("-DKERNEL_HEIGHT=" + support::cpp11::to_string(weights->info()->dimension(2)));
-    build_opts.add_option("-DCONV_PAD_TOP=" + support::cpp11::to_string(conv_info.pad_top()));
-    build_opts.add_option("-DCONV_PAD_LEFT=" + support::cpp11::to_string(conv_info.pad_left()));
-    build_opts.add_option("-DCONV_STRIDE_X=" + support::cpp11::to_string(conv_info.stride().first));
-    build_opts.add_option("-DCONV_STRIDE_Y=" + support::cpp11::to_string(conv_info.stride().second));
-    build_opts.add_option("-DDILATION_X=" + support::cpp11::to_string(dilation.x()));
-    build_opts.add_option("-DDILATION_Y=" + support::cpp11::to_string(dilation.y()));
-    build_opts.add_option("-DVEC_SIZE_LEFTOVER=" + support::cpp11::to_string(_input->info()->dimension(0) % n0));
-
-    std::string kernel_name = (_is_quantized) ? "dwc_MxN_native_quantized8_nhwc" : "dwc_MxN_native_fp_nhwc";
+    build_opts.add_option("-DM0=" + support::cpp11::to_string(m0));
+    build_opts.add_option("-DM0_A=" + support::cpp11::to_string(weights->info()->dimension(1) + m0 - 1));
+    build_opts.add_option("-DPARTIAL_N0=" + support::cpp11::to_string(_input->info()->dimension(0) % n0));
+    build_opts.add_option_if(_input->info()->num_dimensions() > 3, "-DBATCHED_EXECUTION");
+    if(biases != nullptr)
+    {
+        build_opts.add_option(std::string("-DHAS_BIAS"));
+        build_opts.add_option(std::string("-DBIA_DATA_TYPE=" + get_cl_type_from_data_type(biases->info()->data_type())));
+    }
 
     if(_is_quantized)
     {
-        const UniformQuantizationInfo iq_info = _input->info()->quantization_info().uniform();
-        const UniformQuantizationInfo wq_info = _weights->info()->quantization_info().uniform();
-        const UniformQuantizationInfo oq_info = _output->info()->quantization_info().uniform();
+        kernel_name                          = "dwc_native_quantized_nhwc";
+        const UniformQuantizationInfo iqinfo = input->info()->quantization_info().uniform();
+        const UniformQuantizationInfo wqinfo = weights->info()->quantization_info().uniform();
+        const UniformQuantizationInfo oqinfo = output->info()->quantization_info().uniform();
 
-        build_opts.add_option("-DINPUT_OFFSET=" + support::cpp11::to_string(-iq_info.offset));
-        build_opts.add_option("-DWEIGHTS_OFFSET=" + support::cpp11::to_string(-wq_info.offset));
-        build_opts.add_option("-DOUTPUT_OFFSET=" + support::cpp11::to_string(oq_info.offset));
-        build_opts.add_option_if(is_data_type_quantized_per_channel(weights->info()->data_type()), "-DPER_CHANNEL_QUANTIZATION");
+        PixelValue zero_value = PixelValue(0, input->info()->data_type(), input->info()->quantization_info());
+        int        zero_value_s32;
+        zero_value.get(zero_value_s32);
 
-        // Compute non-per-channel multiplier and shift anyway to make OpenCL kernel simpler
-        float multiplier        = iq_info.scale * wq_info.scale / oq_info.scale;
+        float multiplier        = iqinfo.scale * wqinfo.scale / oqinfo.scale;
         int   output_multiplier = 0;
         int   output_shift      = 0;
         quantization::calculate_quantized_multiplier(multiplier, &output_multiplier, &output_shift);
-        build_opts.add_option("-DOUTPUT_MULTIPLIER=" + support::cpp11::to_string(output_multiplier));
-        build_opts.add_option("-DOUTPUT_SHIFT=" + support::cpp11::to_string(output_shift));
-
-        if(dwc_info.activation_info.enabled())
-        {
-            int a_val{};
-            int b_val{};
-            std::tie(b_val, a_val) = get_quantized_activation_min_max(dwc_info.activation_info, input->info()->data_type(), oq_info);
-
-            const int o1 = oq_info.offset;
-
-            build_opts.add_option("-DA_VAL=" + support::cpp11::to_string(a_val));
-            build_opts.add_option("-DB_VAL=" + support::cpp11::to_string(b_val));
-            build_opts.add_option("-DCONST_0=" + support::cpp11::to_string(o1));
-
-            const float s1 = iq_info.scale;
-            build_opts.add_option("-DS1_VAL=" + float_to_string_with_full_precision(s1));
-            build_opts.add_option("-DO1_VAL=" + support::cpp11::to_string(o1));
-        }
-
-        build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type()));
-        build_opts.add_option("-DWEIGHTS_TYPE=" + get_cl_type_from_data_type(weights->info()->data_type()));
+        build_opts.add_option("-DDST_MULTIPLIER=" + support::cpp11::to_string(output_multiplier));
+        build_opts.add_option("-DDST_SHIFT=" + support::cpp11::to_string(output_shift));
+        build_opts.add_option("-DSRC_OFFSET=" + support::cpp11::to_string(-iqinfo.offset));
+        build_opts.add_option("-DWEI_OFFSET=" + support::cpp11::to_string(-wqinfo.offset));
+        build_opts.add_option("-DDST_OFFSET=" + support::cpp11::to_string(oqinfo.offset));
+        build_opts.add_option("-DZERO_VALUE=" + support::cpp11::to_string(zero_value_s32));
+        build_opts.add_option("-DACC_DATA_TYPE=" + get_cl_type_from_data_type(DataType::S32));
+        build_opts.add_option("-DDST_MULTIPLIERS_DATA_TYPE=" + get_cl_type_from_data_type(_output_multipliers->info()->data_type()));
+        build_opts.add_option("-DDST_SHIFTS_DATA_TYPE=" + get_cl_type_from_data_type(_output_shifts->info()->data_type()));
+        build_opts.add_option_if_else(weights->info()->data_type() == DataType::QSYMM8_PER_CHANNEL, "-DQUANTIZATION_TYPE=PER_CHANNEL", "-DQUANTIZATION_TYPE=PER_TENSOR");
+        // Note: We expect the input and output tensors to always adopt a per-tensor quantization approach
+        int a_val{};
+        int b_val{};
+        std::tie(b_val, a_val) = get_quantized_activation_min_max(conv_info.act_info, input->info()->data_type(), oqinfo);
+
+        build_opts.add_option_if(conv_info.act_info.enabled(), "-DA_VAL=" + support::cpp11::to_string(a_val));
+        build_opts.add_option_if(conv_info.act_info.enabled(), "-DB_VAL=" + support::cpp11::to_string(b_val));
     }
     else
     {
-        build_opts.add_option_if(dwc_info.activation_info.enabled(), "-DA_VAL=" + float_to_string_with_full_precision(dwc_info.activation_info.a()));
-        build_opts.add_option_if(dwc_info.activation_info.enabled(), "-DB_VAL=" + float_to_string_with_full_precision(dwc_info.activation_info.b()));
+        kernel_name = "dwc_native_fp_nhwc";
+        build_opts.add_option("-DACC_DATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type()));
+        build_opts.add_option("-DZERO_VALUE=" + support::cpp11::to_string(0));
+        build_opts.add_option_if(conv_info.act_info.enabled(), "-DA_VAL=" + float_to_string_with_full_precision(conv_info.act_info.a()));
+        build_opts.add_option_if(conv_info.act_info.enabled(), "-DB_VAL=" + float_to_string_with_full_precision(conv_info.act_info.b()));
     }
 
-    Window win = calculate_max_window(*(output->info()), Steps(n0));
+    Window win = calculate_max_window(*(output->info()), Steps(n0, m0));
     ICLKernel::configure_internal(win);
 
     _kernel = create_kernel(compile_context, kernel_name, build_opts.options());
@@ -264,10 +289,9 @@ void CLDepthwiseConvolutionLayerNativeKernel::configure(const CLCompileContext &
 }
 
 Status CLDepthwiseConvolutionLayerNativeKernel::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output,
-                                                         const DWCWeightsKernelInfo &dwc_weights_info, const DWCKernelInfo &dwc_info, const PadStrideInfo &conv_info,
-                                                         unsigned int depth_multiplier, const Size2D &dilation, const ITensorInfo *output_multipliers, const ITensorInfo *output_shifts)
+                                                         const DWCComputeKernelInfo &dwc_info, const ConvolutionInfo &conv_info, const ITensorInfo *output_multipliers, const ITensorInfo *output_shifts)
 {
-    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, weights, biases, output, dwc_weights_info, dwc_info, conv_info, depth_multiplier, dilation, output_multipliers, output_shifts));
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, weights, biases, output, dwc_info, conv_info, output_multipliers, output_shifts));
     return Status{};
 }
 
@@ -278,37 +302,46 @@ void CLDepthwiseConvolutionLayerNativeKernel::run(const Window &window, cl::Comm
 
     // Collapse window
     Window window_collapsed = window.collapse(ICLKernel::window(), Window::DimZ);
-    Window slice_in         = window.first_slice_window_4D();
-    Window slice_out        = window_collapsed.first_slice_window_4D();
+
+    Window slice = window_collapsed.first_slice_window_4D();
 
     if(_depth_multiplier != 1)
     {
-        ARM_COMPUTE_ERROR_ON(slice_out.x().step() != 1);
-        slice_out.set(Window::DimX, Window::Dimension(0, _input->info()->tensor_shape()[0], 1));
+        // If the depth multiplier > 1, we need to use the input channels rather than the output channels
+        ARM_COMPUTE_ERROR_ON(slice.x().step() != 1);
+        slice.set(Window::DimX, Window::Dimension(0, _input->info()->tensor_shape()[0], 1));
     }
 
-    unsigned int idx = 2 * num_arguments_per_4D_tensor() + num_arguments_per_3D_tensor();
+    cl::Image2D weights_cl_image;
 
-    // Set output multipliers in case of quantized data type
-    if(_is_quantized)
+    if(_export_to_cl_image)
     {
-        add_1D_tensor_argument(idx, _output_multipliers, slice_in);
-        add_1D_tensor_argument(idx, _output_shifts, slice_in);
+        const size_t      image_w = _weights->info()->dimension(0) / 4;
+        const size_t      image_h = _weights->info()->dimension(1) * _weights->info()->dimension(2) * _weights->info()->dimension(3);
+        const TensorShape shape2d(image_w, image_h);
+        const size_t      image_row_pitch = _weights->info()->strides_in_bytes()[1];
+
+        // Export cl_buffer to cl_image
+        weights_cl_image = create_image2d_from_buffer(CLKernelLibrary::get().context(), _weights->cl_buffer(), shape2d, _weights->info()->data_type(), image_row_pitch);
     }
 
-    if(_biases != nullptr)
+    unsigned int idx = 0;
+    add_4D_tensor_argument(idx, _input, slice);
+    add_4D_tensor_argument(idx, _output, slice);
+    if(_export_to_cl_image)
     {
-        add_1D_tensor_argument(idx, _biases, slice_in);
+        _kernel.setArg(idx++, weights_cl_image);
     }
-
-    do
+    add_4D_tensor_argument(idx, _weights, slice);
+    if(_is_quantized)
+    {
+        add_1D_tensor_argument(idx, _output_multipliers, slice);
+        add_1D_tensor_argument(idx, _output_shifts, slice);
+    }
+    if(_biases != nullptr)
     {
-        idx = 0;
-        add_4D_tensor_argument(idx, _input, slice_in);
-        add_4D_tensor_argument(idx, _output, slice_out);
-        add_3D_tensor_argument(idx, _weights, slice_out);
-        enqueue(queue, *this, slice_out, lws_hint());
+        add_1D_tensor_argument(idx, _biases, slice);
     }
-    while(window_collapsed.slide_window_slice_4D(slice_out) && window.slide_window_slice_4D(slice_in));
+    enqueue(queue, *this, slice, lws_hint());
 }
 } // namespace arm_compute
diff --git a/src/core/CL/kernels/CLDepthwiseConvolutionLayerNativeKernel.h b/src/core/CL/kernels/CLDepthwiseConvolutionLayerNativeKernel.h
index 325f4e7067..68e4ccfc1e 100644
--- a/src/core/CL/kernels/CLDepthwiseConvolutionLayerNativeKernel.h
+++ b/src/core/CL/kernels/CLDepthwiseConvolutionLayerNativeKernel.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2019-2020 Arm Limited.
+ * Copyright (c) 2019-2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -55,19 +55,15 @@ public:
      * @param[in]  biases             Biases tensor. A 1D tensor with dimensions [IFM]. Must be nullptr if not needed.
      *                                Data type supported: Same as @p input, S32 when input is QASYMM8/QASYMM8_SIGNED.
      * @param[out] output             Destination tensor. Data type supported: Same as @p input.
-     * @param[in]  dwc_weights_info   Depthwise convolution layer weights info to retrieve the number of output elements processed by each thread
      * @param[in]  dwc_info           Depthwise convolution layer info
-     * @param[in]  conv_info          Padding and stride information to use for the convolution.
-     * @param[in]  depth_multiplier   (Optional) Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1.
-     * @param[in]  dilation           (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
+     * @param[in]  conv_info          Convolution info (padding, stride, dilation, ...)
      * @param[in]  output_multipliers (Optional) Output multipliers tensor for quantized computations. In case of per-channel quantization,
      *                                the number of multipliers must be equal to the number of filters (IFM). Supported data types: S32
      * @param[in]  output_shifts      (Optional) Output shifts tensor for quantized computations. In case of per-channel quantization,
      *                                the number of multipliers must be equal to the number of filters (IFM). Supported data types: S32
      */
-    void configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const DWCWeightsKernelInfo &dwc_weights_info,
-                   const DWCKernelInfo &dwc_info, const PadStrideInfo &conv_info, unsigned int depth_multiplier = 1, const Size2D &dilation = Size2D(1U, 1U),
-                   const ICLTensor *output_multipliers = nullptr, const ICLTensor *output_shifts = nullptr);
+    void configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const DWCComputeKernelInfo &dwc_info,
+                   const ConvolutionInfo &conv_info, const ICLTensor *output_multipliers = nullptr, const ICLTensor *output_shifts = nullptr);
     /** Initialize the function's source, destination and parameters
      *
      * @param[in]  compile_context    The compile context to be used.
@@ -77,19 +73,15 @@ public:
      * @param[in]  biases             Biases tensor. A 1D tensor with dimensions [IFM]. Must be nullptr if not needed.
      *                                Data type supported: Same as @p input, S32 when input is QASYMM8/QASYMM8_SIGNED.
      * @param[out] output             Destination tensor. Data type supported: Same as @p input.
-     * @param[in]  dwc_weights_info   Depthwise convolution layer weights info to retrieve the number of output elements processed by each thread
      * @param[in]  dwc_info           Depthwise convolution layer info
-     * @param[in]  conv_info          Padding and stride information to use for the convolution.
-     * @param[in]  depth_multiplier   (Optional) Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1.
-     * @param[in]  dilation           (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
+     * @param[in]  conv_info          Convolution info (padding, stride, dilation, ...)
      * @param[in]  output_multipliers (Optional) Output multipliers tensor for quantized computations. In case of per-channel quantization,
      *                                the number of multipliers must be equal to the number of filters (IFM). Supported data types: S32
      * @param[in]  output_shifts      (Optional) Output shifts tensor for quantized computations. In case of per-channel quantization,
      *                                the number of multipliers must be equal to the number of filters (IFM). Supported data types: S32
      */
-    void configure(const CLCompileContext &compile_context, const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const DWCWeightsKernelInfo &dwc_weights_info,
-                   const DWCKernelInfo &dwc_info, const PadStrideInfo &conv_info, unsigned int depth_multiplier = 1, const Size2D &dilation = Size2D(1U, 1U),
-                   const ICLTensor *output_multipliers = nullptr, const ICLTensor *output_shifts = nullptr);
+    void configure(const CLCompileContext &compile_context, const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const DWCComputeKernelInfo &dwc_info,
+                   const ConvolutionInfo &conv_info, const ICLTensor *output_multipliers = nullptr, const ICLTensor *output_shifts = nullptr);
     /** Static function to check if given info will lead to a valid configuration of @ref CLDepthwiseConvolutionLayerNativeKernel
      *
      * @param[in] input              Source tensor info. Data type supported: QASYMM8/QASYMM8_SIGNED/FP32/FP16. Data layout supported: NHWC
@@ -98,11 +90,8 @@ public:
      * @param[in] biases             Biases tensor info. A 1D tensor with dimensions [IFM]. Must be nullptr if not needed.
      *                               Data type supported: Same as @p input, S32 when input is QASYMM8/QASYMM8_SIGNED.
      * @param[in] output             Destination tensor info. Data type supported: Same as @p input.
-     * @param[in] dwc_weights_info   Depthwise convolution layer weights info to retrieve the number of output elements processed by each thread
      * @param[in] dwc_info           Depthwise convolution layer info
-     * @param[in] conv_info          Padding and stride information to use for the convolution.
-     * @param[in] depth_multiplier   (Optional) Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1.
-     * @param[in] dilation           (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
+     * @param[in] conv_info          Convolution info (padding, stride, dilation, ...)
      * @param[in] output_multipliers (Optional) Output multipliers tensor for quantized computations. In case of per-channel quantization,
      *                               the number of multipliers must be equal to the number of filters (IFM). Supported data types: S32
      * @param[in] output_shifts      (Optional) Output shifts tensor for quantized computations. In case of per-channel quantization,
@@ -110,9 +99,8 @@ public:
      *
      * @return a status
      */
-    static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const DWCWeightsKernelInfo &dwc_weights_info,
-                           const DWCKernelInfo &dwc_info, const PadStrideInfo &conv_info, unsigned int depth_multiplier = 1, const Size2D &dilation = Size2D(1U, 1U),
-                           const ITensorInfo *output_multipliers = nullptr, const ITensorInfo *output_shifts = nullptr);
+    static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const DWCComputeKernelInfo &dwc_info,
+                           const ConvolutionInfo &conv_info, const ITensorInfo *output_multipliers = nullptr, const ITensorInfo *output_shifts = nullptr);
 
     // Inherited methods overridden:
     void run(const Window &window, cl::CommandQueue &queue) override;
@@ -125,6 +113,7 @@ private:
     unsigned int     _depth_multiplier;
     const ICLTensor *_output_multipliers;
     const ICLTensor *_output_shifts;
+    bool             _export_to_cl_image;
     bool             _is_quantized;
 };
 } // namespace arm_compute