CLDepthwiseConvolutionLayer rework - Part 1

Remove the reshaped variant for CLDepthwiseConvolutionLayer 3x3 NHWC Quantized

- Remove kernel selection by GPUTarget
- Remove unused quantized support from the NHWC kernel
- Remove CLDepthwiseConvolutionLayerReshapeWeightsKernel
- Remove OpenCL kernels for reshaped dwc 3x3 quantized and weights reshape
- Remove the "_bifrost" suffix in common OpenCL kernel
- Remove the ICLDepthwiseConvolutionLayer3x3Kernel common interface

Resolve COMPMID-3864, COMPMID-3907

Change-Id: Icfac0fb6c00e214985beb05dad7c0cdbbee7d830
Signed-off-by: Giorgio Arena <giorgio.arena@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/5447
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

17 files changed, 379 insertions, 1916 deletions

diff --git a/Android.bp b/Android.bp
index 92b96848aa..8e4d9fe224 100644
--- a/Android.bp
+++ b/Android.bp
@@ -97,7 +97,6 @@ cc_library_static {
         "src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.cpp",
         "src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.cpp",
         "src/core/CL/kernels/CLDepthwiseConvolutionLayerNativeKernel.cpp",
-        "src/core/CL/kernels/CLDepthwiseConvolutionLayerReshapeWeightsKernel.cpp",
         "src/core/CL/kernels/CLFFTDigitReverseKernel.cpp",
         "src/core/CL/kernels/CLFFTRadixStageKernel.cpp",
         "src/core/CL/kernels/CLFFTScaleKernel.cpp",
diff --git a/arm_compute/core/Types.h b/arm_compute/core/Types.h
index b1f340d18e..b5fd21d29d 100644
--- a/arm_compute/core/Types.h
+++ b/arm_compute/core/Types.h
@@ -1874,12 +1874,6 @@ struct ConvolutionInfo
     Size2D              dilation{ Size2D(1, 1) }; /**< Dilation, in elements, across x and y. Defaults to (1, 1). */
 };
 
-struct DepthwiseConvolutionReshapeInfo
-{
-    unsigned int c0{ 1 };            /**< Number of channels processed by the depth-wise convolution */
-    bool         transpose{ false }; /**< True if the block MxC0 (where M is the area of the filter i.e. KwxKh) has to be transposed */
-};
-
 /** GEMMLowp output stage type */
 enum class GEMMLowpOutputStageType
 {
diff --git a/arm_compute/core/utils/misc/ShapeCalculator.h b/arm_compute/core/utils/misc/ShapeCalculator.h
index ba37f9a61e..8e49c068af 100644
--- a/arm_compute/core/utils/misc/ShapeCalculator.h
+++ b/arm_compute/core/utils/misc/ShapeCalculator.h
@@ -287,30 +287,6 @@ inline TensorShape compute_interleaved_shape(const ITensorInfo &a, int mult_inte
     return shape_interleaved_a;
 }
 
-/** Calculate the reshaped shape of the weights to use in depthwise convolution
- *
- * @param[in] input Input tensor info
- * @param[in] info  Depthwise convolution information to be used for reshaping.
- *
- * @return the calculated shape
- */
-inline TensorShape compute_reshaped_depthwise_weights_shape(const ITensorInfo &input, const DepthwiseConvolutionReshapeInfo &info)
-{
-    const auto  data_layout = input.data_layout();
-    TensorShape weights_shape{};
-
-    const int    width_idx    = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH);
-    const int    height_idx   = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT);
-    const int    channel_idx  = get_data_layout_dimension_index(data_layout, DataLayoutDimension::CHANNEL);
-    const size_t num_channels = input.dimension(channel_idx);
-    const size_t num_rows     = input.dimension(height_idx);
-    const size_t num_cols     = input.dimension(width_idx);
-
-    weights_shape.set(0, num_rows * num_cols * info.c0);
-    weights_shape.set(1, DIV_CEIL(num_channels, info.c0));
-    return weights_shape;
-}
-
 /** Calculate the transposed 1xW shape
  *
  * @param[in] b Input tensor info
diff --git a/arm_compute/runtime/CL/functions/CLDepthwiseConvolutionLayer.h b/arm_compute/runtime/CL/functions/CLDepthwiseConvolutionLayer.h
index e2c5d683cf..1af9e1dc6f 100644
--- a/arm_compute/runtime/CL/functions/CLDepthwiseConvolutionLayer.h
+++ b/arm_compute/runtime/CL/functions/CLDepthwiseConvolutionLayer.h
@@ -35,8 +35,8 @@ namespace arm_compute
 class CLCompileContext;
 class CLFillBorderKernel;
 class CLDepthwiseConvolutionLayerNativeKernel;
-class CLDepthwiseConvolutionLayerReshapeWeightsKernel;
-class ICLDepthwiseConvolutionLayer3x3Kernel;
+class CLDepthwiseConvolutionLayer3x3NCHWKernel;
+class CLDepthwiseConvolutionLayer3x3NHWCKernel;
 class ICLTensor;
 
 /** Function to execute a depthwise convolution
@@ -123,19 +123,17 @@ private:
      * @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 3x3 QASYMM8 supported.
      * @param[in] dilation         (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
-     * @param[in] gpu_target       (Optional) GPU target to validate the kernel for. Defaults to midgard.
      *
      * @return a Depthwise Convolution Function
      */
     static DepthwiseConvolutionFunction get_depthwiseconvolution_function(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), GPUTarget gpu_target = GPUTarget::MIDGARD);
+                                                                          ActivationLayerInfo act_info = ActivationLayerInfo(), const Size2D &dilation = Size2D(1U, 1U));
 
     /** Basic function to execute a depthwise convolution for kernel size 3x3xC (when data layout NCHW) or Cx3x3 (when data layout NHWC). This function calls the following OpenCL kernels:
     *
     * -# @ref CLDepthwiseConvolutionLayer3x3NCHWKernel (if data_layout == NCHW)
     * -# @ref CLDepthwiseConvolutionLayer3x3NHWCKernel (if data_layout == NHWC)
-    * -# @ref CLDepthwiseConvolutionLayerReshapeWeightsKernel (if data_layout == NHWC)
     * -# @ref CLFillBorderKernel (if pad_x or pad_y > 0)
     *
     */
@@ -200,7 +198,7 @@ private:
          * @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(), GPUTarget gpu_target = GPUTarget::MIDGARD, const Size2D &dilation = Size2D(1U, 1U));
+                               ActivationLayerInfo act_info = ActivationLayerInfo(), const Size2D &dilation = Size2D(1U, 1U));
 
         // Inherited methods overriden:
         void run() override;
@@ -212,26 +210,25 @@ private:
         };
 
     private:
-        MemoryGroup                                                      _memory_group;
-        std::unique_ptr<ICLDepthwiseConvolutionLayer3x3Kernel>           _kernel;
-        std::unique_ptr<CLFillBorderKernel>                              _border_handler;
-        CLPermute                                                        _permute_input_to_nchw;
-        CLPermute                                                        _permute_weights_to_nchw;
-        CLPermute                                                        _permute_output_to_nhwc;
-        std::unique_ptr<CLDepthwiseConvolutionLayerReshapeWeightsKernel> _reshape_weights;
-        CLTensor                                                         _permuted_input;
-        CLTensor                                                         _permuted_weights;
-        CLTensor                                                         _permuted_output;
-        CLTensor                                                         _output_multipliers;
-        CLTensor                                                         _output_shifts;
-        const ITensor                                                   *_original_weights;
-        const ITensor                                                   *_input;
-        const ITensor                                                   *_output;
-        bool                                                             _needs_permute;
-        bool                                                             _needs_weights_reshape;
-        bool                                                             _is_prepared;
-        bool                                                             _is_quantized;
-        bool                                                             _is_nhwc;
+        MemoryGroup                                               _memory_group;
+        std::unique_ptr<CLDepthwiseConvolutionLayer3x3NCHWKernel> _kernel_nchw;
+        std::unique_ptr<CLDepthwiseConvolutionLayer3x3NHWCKernel> _kernel_nhwc;
+        std::unique_ptr<CLFillBorderKernel>                       _border_handler;
+        CLPermute                                                 _permute_input_to_nchw;
+        CLPermute                                                 _permute_weights_to_nchw;
+        CLPermute                                                 _permute_output_to_nhwc;
+        CLTensor                                                  _permuted_input;
+        CLTensor                                                  _permuted_weights;
+        CLTensor                                                  _permuted_output;
+        CLTensor                                                  _output_multipliers;
+        CLTensor                                                  _output_shifts;
+        const ITensor                                            *_original_weights;
+        const ITensor                                            *_input;
+        const ITensor                                            *_output;
+        bool                                                      _needs_permute;
+        bool                                                      _is_prepared;
+        bool                                                      _is_quantized;
+        bool                                                      _is_nhwc;
     };
 
     /** Basic function to execute a generic depthwise convolution. This function calls the following OpenCL kernels:
diff --git a/src/core/CL/CLKernelLibrary.cpp b/src/core/CL/CLKernelLibrary.cpp
index 002a14400f..33b5ad0c6e 100644
--- a/src/core/CL/CLKernelLibrary.cpp
+++ b/src/core/CL/CLKernelLibrary.cpp
@@ -231,16 +231,12 @@ const std::map<std::string, std::string> CLKernelLibrary::_kernel_program_map =
     { "dwc_MxN_native_quantized8_nhwc", "depthwise_convolution_quantized.cl" },
     { "dwc_3x3_native_quantized8_nchw", "depthwise_convolution_quantized.cl" },
     { "dwc_3x3_native_quantized8_dot8_nchw", "depthwise_convolution_quantized.cl" },
-    { "dwc_3x3_reshaped_quantized8_nhwc", "depthwise_convolution_quantized.cl" },
-    { "dwc_3x3_reshaped_quantized8_stride1_nhwc", "depthwise_convolution_quantized.cl" },
-    { "dwc_3x3_reshaped_quantized8_dot8_stride1_nhwc", "depthwise_convolution_quantized.cl" },
     { "depth_to_space_nchw", "depth_to_space.cl" },
     { "depth_to_space_nhwc", "depth_to_space.cl" },
-    { "depthwise_convolution_3x3_stridex1_stridey1_bifrost_f16", "depthwise_convolution.cl" },
-    { "depthwise_convolution_3x3_stridex2_stridey2_bifrost_f16", "depthwise_convolution.cl" },
-    { "depthwise_convolution_3x3_stridex1_stridey1_bifrost_f32", "depthwise_convolution.cl" },
-    { "depthwise_convolution_3x3_stridex2_stridey2_bifrost_f32", "depthwise_convolution.cl" },
-    { "depthwise_convolution_reshape_weights", "depthwise_convolution.cl" },
+    { "depthwise_convolution_3x3_stridex1_stridey1_f16", "depthwise_convolution.cl" },
+    { "depthwise_convolution_3x3_stridex2_stridey2_f16", "depthwise_convolution.cl" },
+    { "depthwise_convolution_3x3_stridex1_stridey1_f32", "depthwise_convolution.cl" },
+    { "depthwise_convolution_3x3_stridex2_stridey2_f32", "depthwise_convolution.cl" },
     { "dequantization_layer", "dequantization_layer.cl" },
     { "dequantization_layer_per_channel_nhwc", "dequantization_layer.cl" },
     { "dequantization_layer_per_channel_nchw", "dequantization_layer.cl" },
diff --git a/src/core/CL/CLKernels.h b/src/core/CL/CLKernels.h
index f29c768fa3..63978cea3f 100644
--- a/src/core/CL/CLKernels.h
+++ b/src/core/CL/CLKernels.h
@@ -40,7 +40,6 @@
 #include "src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.h"
 #include "src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.h"
 #include "src/core/CL/kernels/CLDepthwiseConvolutionLayerNativeKernel.h"
-#include "src/core/CL/kernels/CLDepthwiseConvolutionLayerReshapeWeightsKernel.h"
 #include "src/core/CL/kernels/CLFFTDigitReverseKernel.h"
 #include "src/core/CL/kernels/CLFFTRadixStageKernel.h"
 #include "src/core/CL/kernels/CLFFTScaleKernel.h"
@@ -91,6 +90,5 @@
 #include "src/core/CL/kernels/CLWinogradFilterTransformKernel.h"
 #include "src/core/CL/kernels/CLWinogradInputTransformKernel.h"
 #include "src/core/CL/kernels/CLWinogradOutputTransformKernel.h"
-#include "src/core/CL/kernels/ICLDepthwiseConvolutionLayer3x3Kernel.h"
 
 #endif /* ARM_COMPUTE_CLKERNELS_H */
diff --git a/src/core/CL/cl_kernels/depthwise_convolution.cl b/src/core/CL/cl_kernels/depthwise_convolution.cl
index 8ce5617858..22a38e7094 100644
--- a/src/core/CL/cl_kernels/depthwise_convolution.cl
+++ b/src/core/CL/cl_kernels/depthwise_convolution.cl
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2020 Arm Limited.
+ * Copyright (c) 2017-2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -42,110 +42,110 @@ inline __global uchar *ptr_offset(__global uchar *ptr, const int x, const int y,
 
 #if(DILATION_X == 1 && DILATION_Y == 1)
 
-#define CONVOLUTION1x3_BIFROST2X1_STRIDE1(acc, src0, weights_row0) \
-    ({                                                             \
-        acc.s0 = fma(src0.s0, weights_row0.s0, acc.s0);            \
-        acc.s0 = fma(src0.s1, weights_row0.s1, acc.s0);            \
-        acc.s0 = fma(src0.s2, weights_row0.s2, acc.s0);            \
-        acc.s1 = fma(src0.s1, weights_row0.s0, acc.s1);            \
-        acc.s1 = fma(src0.s2, weights_row0.s1, acc.s1);            \
-        acc.s1 = fma(src0.s3, weights_row0.s2, acc.s1);            \
+#define CONVOLUTION1x3_2X1_STRIDE1(acc, src0, weights_row0) \
+    ({                                                      \
+        acc.s0 = fma(src0.s0, weights_row0.s0, acc.s0);     \
+        acc.s0 = fma(src0.s1, weights_row0.s1, acc.s0);     \
+        acc.s0 = fma(src0.s2, weights_row0.s2, acc.s0);     \
+        acc.s1 = fma(src0.s1, weights_row0.s0, acc.s1);     \
+        acc.s1 = fma(src0.s2, weights_row0.s1, acc.s1);     \
+        acc.s1 = fma(src0.s3, weights_row0.s2, acc.s1);     \
     })
 
-#define CONVOLUTION1x3_BIFROST4X1_STRIDE1(acc, src0, weights_row0) \
-    ({                                                             \
-        acc.s0 = fma(src0.s0, weights_row0.s0, acc.s0);            \
-        acc.s0 = fma(src0.s1, weights_row0.s1, acc.s0);            \
-        acc.s0 = fma(src0.s2, weights_row0.s2, acc.s0);            \
-        acc.s1 = fma(src0.s1, weights_row0.s0, acc.s1);            \
-        acc.s1 = fma(src0.s2, weights_row0.s1, acc.s1);            \
-        acc.s1 = fma(src0.s3, weights_row0.s2, acc.s1);            \
-        acc.s2 = fma(src0.s2, weights_row0.s0, acc.s2);            \
-        acc.s2 = fma(src0.s3, weights_row0.s1, acc.s2);            \
-        acc.s2 = fma(src0.s4, weights_row0.s2, acc.s2);            \
-        acc.s3 = fma(src0.s3, weights_row0.s0, acc.s3);            \
-        acc.s3 = fma(src0.s4, weights_row0.s1, acc.s3);            \
-        acc.s3 = fma(src0.s5, weights_row0.s2, acc.s3);            \
+#define CONVOLUTION1x3_4X1_STRIDE1(acc, src0, weights_row0) \
+    ({                                                      \
+        acc.s0 = fma(src0.s0, weights_row0.s0, acc.s0);     \
+        acc.s0 = fma(src0.s1, weights_row0.s1, acc.s0);     \
+        acc.s0 = fma(src0.s2, weights_row0.s2, acc.s0);     \
+        acc.s1 = fma(src0.s1, weights_row0.s0, acc.s1);     \
+        acc.s1 = fma(src0.s2, weights_row0.s1, acc.s1);     \
+        acc.s1 = fma(src0.s3, weights_row0.s2, acc.s1);     \
+        acc.s2 = fma(src0.s2, weights_row0.s0, acc.s2);     \
+        acc.s2 = fma(src0.s3, weights_row0.s1, acc.s2);     \
+        acc.s2 = fma(src0.s4, weights_row0.s2, acc.s2);     \
+        acc.s3 = fma(src0.s3, weights_row0.s0, acc.s3);     \
+        acc.s3 = fma(src0.s4, weights_row0.s1, acc.s3);     \
+        acc.s3 = fma(src0.s5, weights_row0.s2, acc.s3);     \
     })
 
-#define CONVOLUTION1x3_BIFROST2X1_STRIDE2(acc, src0, src1, weights_row0) \
-    ({                                                                   \
-        acc.s0 = fma(src0.s0, weights_row0.s0, acc.s0);                  \
-        acc.s0 = fma(src0.s1, weights_row0.s1, acc.s0);                  \
-        acc.s0 = fma(src0.s2, weights_row0.s2, acc.s0);                  \
-        acc.s1 = fma(src0.s2, weights_row0.s0, acc.s1);                  \
-        acc.s1 = fma(src0.s3, weights_row0.s1, acc.s1);                  \
-        acc.s1 = fma(src1.s0, weights_row0.s2, acc.s1);                  \
+#define CONVOLUTION1x3_2X1_STRIDE2(acc, src0, src1, weights_row0) \
+    ({                                                            \
+        acc.s0 = fma(src0.s0, weights_row0.s0, acc.s0);           \
+        acc.s0 = fma(src0.s1, weights_row0.s1, acc.s0);           \
+        acc.s0 = fma(src0.s2, weights_row0.s2, acc.s0);           \
+        acc.s1 = fma(src0.s2, weights_row0.s0, acc.s1);           \
+        acc.s1 = fma(src0.s3, weights_row0.s1, acc.s1);           \
+        acc.s1 = fma(src1.s0, weights_row0.s2, acc.s1);           \
     })
 
-#define CONVOLUTION1x3_BIFROST4X1_STRIDE2(acc, src0, src1, weights_row0) \
-    ({                                                                   \
-        acc.s0 = fma(src0.s0, weights_row0.s0, acc.s0);                  \
-        acc.s0 = fma(src0.s1, weights_row0.s1, acc.s0);                  \
-        acc.s0 = fma(src0.s2, weights_row0.s2, acc.s0);                  \
-        acc.s1 = fma(src0.s2, weights_row0.s0, acc.s1);                  \
-        acc.s1 = fma(src0.s3, weights_row0.s1, acc.s1);                  \
-        acc.s1 = fma(src0.s4, weights_row0.s2, acc.s1);                  \
-        acc.s2 = fma(src0.s4, weights_row0.s0, acc.s2);                  \
-        acc.s2 = fma(src0.s5, weights_row0.s1, acc.s2);                  \
-        acc.s2 = fma(src0.s6, weights_row0.s2, acc.s2);                  \
-        acc.s3 = fma(src0.s6, weights_row0.s0, acc.s3);                  \
-        acc.s3 = fma(src0.s7, weights_row0.s1, acc.s3);                  \
-        acc.s3 = fma(src1.s0, weights_row0.s2, acc.s3);                  \
+#define CONVOLUTION1x3_4X1_STRIDE2(acc, src0, src1, weights_row0) \
+    ({                                                            \
+        acc.s0 = fma(src0.s0, weights_row0.s0, acc.s0);           \
+        acc.s0 = fma(src0.s1, weights_row0.s1, acc.s0);           \
+        acc.s0 = fma(src0.s2, weights_row0.s2, acc.s0);           \
+        acc.s1 = fma(src0.s2, weights_row0.s0, acc.s1);           \
+        acc.s1 = fma(src0.s3, weights_row0.s1, acc.s1);           \
+        acc.s1 = fma(src0.s4, weights_row0.s2, acc.s1);           \
+        acc.s2 = fma(src0.s4, weights_row0.s0, acc.s2);           \
+        acc.s2 = fma(src0.s5, weights_row0.s1, acc.s2);           \
+        acc.s2 = fma(src0.s6, weights_row0.s2, acc.s2);           \
+        acc.s3 = fma(src0.s6, weights_row0.s0, acc.s3);           \
+        acc.s3 = fma(src0.s7, weights_row0.s1, acc.s3);           \
+        acc.s3 = fma(src1.s0, weights_row0.s2, acc.s3);           \
     })
 
 #else /* DILATION_X==1 && DILATION_Y==1 */
 
-#define CONVOLUTION1x3_BIFROST2X1_STRIDE1(acc, src0_left, src0_mid, src0_right, weights_row0) \
-    ({                                                                                        \
-        acc.s0 = fma(src0_left.s0, weights_row0.s0, acc.s0);                                  \
-        acc.s0 = fma(src0_mid.s0, weights_row0.s1, acc.s0);                                   \
-        acc.s0 = fma(src0_right.s0, weights_row0.s2, acc.s0);                                 \
-        acc.s1 = fma(src0_left.s1, weights_row0.s0, acc.s1);                                  \
-        acc.s1 = fma(src0_mid.s1, weights_row0.s1, acc.s1);                                   \
-        acc.s1 = fma(src0_right.s1, weights_row0.s2, acc.s1);                                 \
+#define CONVOLUTION1x3_2X1_STRIDE1(acc, src0_left, src0_mid, src0_right, weights_row0) \
+    ({                                                                                 \
+        acc.s0 = fma(src0_left.s0, weights_row0.s0, acc.s0);                           \
+        acc.s0 = fma(src0_mid.s0, weights_row0.s1, acc.s0);                            \
+        acc.s0 = fma(src0_right.s0, weights_row0.s2, acc.s0);                          \
+        acc.s1 = fma(src0_left.s1, weights_row0.s0, acc.s1);                           \
+        acc.s1 = fma(src0_mid.s1, weights_row0.s1, acc.s1);                            \
+        acc.s1 = fma(src0_right.s1, weights_row0.s2, acc.s1);                          \
     })
 
-#define CONVOLUTION1x3_BIFROST2X1_STRIDE2(acc, src0_left, src0_mid, src0_right, weights_row0) \
-    ({                                                                                        \
-        acc.s0 = fma(src0_left.s0, weights_row0.s0, acc.s0);                                  \
-        acc.s0 = fma(src0_mid.s0, weights_row0.s1, acc.s0);                                   \
-        acc.s0 = fma(src0_right.s0, weights_row0.s2, acc.s0);                                 \
-        acc.s1 = fma(src0_left.s2, weights_row0.s0, acc.s1);                                  \
-        acc.s1 = fma(src0_mid.s2, weights_row0.s1, acc.s1);                                   \
-        acc.s1 = fma(src0_right.s2, weights_row0.s2, acc.s1);                                 \
+#define CONVOLUTION1x3_2X1_STRIDE2(acc, src0_left, src0_mid, src0_right, weights_row0) \
+    ({                                                                                 \
+        acc.s0 = fma(src0_left.s0, weights_row0.s0, acc.s0);                           \
+        acc.s0 = fma(src0_mid.s0, weights_row0.s1, acc.s0);                            \
+        acc.s0 = fma(src0_right.s0, weights_row0.s2, acc.s0);                          \
+        acc.s1 = fma(src0_left.s2, weights_row0.s0, acc.s1);                           \
+        acc.s1 = fma(src0_mid.s2, weights_row0.s1, acc.s1);                            \
+        acc.s1 = fma(src0_right.s2, weights_row0.s2, acc.s1);                          \
     })
 
-#define CONVOLUTION1x3_BIFROST4X1_STRIDE1(acc, src0_left, src0_mid, src0_right, weights_row0) \
-    ({                                                                                        \
-        acc.s0 = fma(src0_left.s0, weights_row0.s0, acc.s0);                                  \
-        acc.s0 = fma(src0_mid.s0, weights_row0.s1, acc.s0);                                   \
-        acc.s0 = fma(src0_right.s0, weights_row0.s2, acc.s0);                                 \
-        acc.s1 = fma(src0_left.s1, weights_row0.s0, acc.s1);                                  \
-        acc.s1 = fma(src0_mid.s1, weights_row0.s1, acc.s1);                                   \
-        acc.s1 = fma(src0_right.s1, weights_row0.s2, acc.s1);                                 \
-        acc.s2 = fma(src0_left.s2, weights_row0.s0, acc.s2);                                  \
-        acc.s2 = fma(src0_mid.s2, weights_row0.s1, acc.s2);                                   \
-        acc.s2 = fma(src0_right.s2, weights_row0.s2, acc.s2);                                 \
-        acc.s3 = fma(src0_left.s3, weights_row0.s0, acc.s3);                                  \
-        acc.s3 = fma(src0_mid.s3, weights_row0.s1, acc.s3);                                   \
-        acc.s3 = fma(src0_right.s3, weights_row0.s2, acc.s3);                                 \
+#define CONVOLUTION1x3_4X1_STRIDE1(acc, src0_left, src0_mid, src0_right, weights_row0) \
+    ({                                                                                 \
+        acc.s0 = fma(src0_left.s0, weights_row0.s0, acc.s0);                           \
+        acc.s0 = fma(src0_mid.s0, weights_row0.s1, acc.s0);                            \
+        acc.s0 = fma(src0_right.s0, weights_row0.s2, acc.s0);                          \
+        acc.s1 = fma(src0_left.s1, weights_row0.s0, acc.s1);                           \
+        acc.s1 = fma(src0_mid.s1, weights_row0.s1, acc.s1);                            \
+        acc.s1 = fma(src0_right.s1, weights_row0.s2, acc.s1);                          \
+        acc.s2 = fma(src0_left.s2, weights_row0.s0, acc.s2);                           \
+        acc.s2 = fma(src0_mid.s2, weights_row0.s1, acc.s2);                            \
+        acc.s2 = fma(src0_right.s2, weights_row0.s2, acc.s2);                          \
+        acc.s3 = fma(src0_left.s3, weights_row0.s0, acc.s3);                           \
+        acc.s3 = fma(src0_mid.s3, weights_row0.s1, acc.s3);                            \
+        acc.s3 = fma(src0_right.s3, weights_row0.s2, acc.s3);                          \
     })
 
-#define CONVOLUTION1x3_BIFROST4X1_STRIDE2(acc, src0_left, src0_mid, src0_right, weights_row0) \
-    ({                                                                                        \
-        acc.s0 = fma(src0_left.s0, weights_row0.s0, acc.s0);                                  \
-        acc.s0 = fma(src0_mid.s0, weights_row0.s1, acc.s0);                                   \
-        acc.s0 = fma(src0_right.s0, weights_row0.s2, acc.s0);                                 \
-        acc.s1 = fma(src0_left.s2, weights_row0.s0, acc.s1);                                  \
-        acc.s1 = fma(src0_mid.s2, weights_row0.s1, acc.s1);                                   \
-        acc.s1 = fma(src0_right.s2, weights_row0.s2, acc.s1);                                 \
-        acc.s2 = fma(src0_left.s4, weights_row0.s0, acc.s2);                                  \
-        acc.s2 = fma(src0_mid.s4, weights_row0.s1, acc.s2);                                   \
-        acc.s2 = fma(src0_right.s4, weights_row0.s2, acc.s2);                                 \
-        acc.s3 = fma(src0_left.s6, weights_row0.s0, acc.s3);                                  \
-        acc.s3 = fma(src0_mid.s6, weights_row0.s1, acc.s3);                                   \
-        acc.s3 = fma(src0_right.s6, weights_row0.s2, acc.s3);                                 \
+#define CONVOLUTION1x3_4X1_STRIDE2(acc, src0_left, src0_mid, src0_right, weights_row0) \
+    ({                                                                                 \
+        acc.s0 = fma(src0_left.s0, weights_row0.s0, acc.s0);                           \
+        acc.s0 = fma(src0_mid.s0, weights_row0.s1, acc.s0);                            \
+        acc.s0 = fma(src0_right.s0, weights_row0.s2, acc.s0);                          \
+        acc.s1 = fma(src0_left.s2, weights_row0.s0, acc.s1);                           \
+        acc.s1 = fma(src0_mid.s2, weights_row0.s1, acc.s1);                            \
+        acc.s1 = fma(src0_right.s2, weights_row0.s2, acc.s1);                          \
+        acc.s2 = fma(src0_left.s4, weights_row0.s0, acc.s2);                           \
+        acc.s2 = fma(src0_mid.s4, weights_row0.s1, acc.s2);                            \
+        acc.s2 = fma(src0_right.s4, weights_row0.s2, acc.s2);                          \
+        acc.s3 = fma(src0_left.s6, weights_row0.s0, acc.s3);                           \
+        acc.s3 = fma(src0_mid.s6, weights_row0.s1, acc.s3);                            \
+        acc.s3 = fma(src0_right.s6, weights_row0.s2, acc.s3);                          \
     })
 
 #endif /* DILATION_X==1 && DILATION_Y==1 */
@@ -385,8 +385,8 @@ __kernel void depthwise_convolution_3x3(
  * @param[in] weights_addr     Pointer from where to get weights
  * @param[in] weights_stride_y Stride of weights tesnsor in Y dimension
  */
-inline float2 convolution_3x3_dilation_stridex1_stridey1_bifrost_f32(__global uchar *src_addr, const int stride_x_bytes, const int stride_y_bytes,
-                                                                     const int y_offset, __global uchar *weights_addr, const int weights_stride_y)
+inline float2 convolution_3x3_dilation_stridex1_stridey1_f32(__global uchar *src_addr, const int stride_x_bytes, const int stride_y_bytes,
+                                                             const int y_offset, __global uchar *weights_addr, const int weights_stride_y)
 {
     // Load the weights
     float3 weights_row0 = vload3(0, (__global float *)(weights_addr + 0 * weights_stride_y));
@@ -407,9 +407,9 @@ inline float2 convolution_3x3_dilation_stridex1_stridey1_bifrost_f32(__global uc
     float2 src20_mid   = vload2(0, (__global float *)ptr_offset(src_addr, DILATION_X, y_offset + DILATION_Y * 2, stride_x_bytes, stride_y_bytes));
     float2 src20_right = vload2(0, (__global float *)ptr_offset(src_addr, 2 * DILATION_X, y_offset + DILATION_Y * 2, stride_x_bytes, stride_y_bytes));
 
-    CONVOLUTION1x3_BIFROST2X1_STRIDE1(pixels0, src00_left, src00_mid, src00_right, weights_row0);
-    CONVOLUTION1x3_BIFROST2X1_STRIDE1(pixels0, src10_left, src10_mid, src10_right, weights_row1);
-    CONVOLUTION1x3_BIFROST2X1_STRIDE1(pixels0, src20_left, src20_mid, src20_right, weights_row2);
+    CONVOLUTION1x3_2X1_STRIDE1(pixels0, src00_left, src00_mid, src00_right, weights_row0);
+    CONVOLUTION1x3_2X1_STRIDE1(pixels0, src10_left, src10_mid, src10_right, weights_row1);
+    CONVOLUTION1x3_2X1_STRIDE1(pixels0, src20_left, src20_mid, src20_right, weights_row2);
 
     return pixels0;
 }
@@ -423,8 +423,8 @@ inline float2 convolution_3x3_dilation_stridex1_stridey1_bifrost_f32(__global uc
  * @param[in] weights_addr     Pointer from where to get weights
  * @param[in] weights_stride_y Stride of weights tesnsor in Y dimension
  */
-inline float2 convolution_3x3_dilation_stridex2_stridey2_bifrost_f32(__global uchar *src_addr, const int stride_x_bytes, const int stride_y_bytes,
-                                                                     const int y_offset, __global uchar *weights_addr, const int weights_stride_y)
+inline float2 convolution_3x3_dilation_stridex2_stridey2_f32(__global uchar *src_addr, const int stride_x_bytes, const int stride_y_bytes,
+                                                             const int y_offset, __global uchar *weights_addr, const int weights_stride_y)
 {
     // Load the weights
     float3 weights_row0 = vload3(0, (__global float *)(weights_addr + 0 * weights_stride_y));
@@ -445,9 +445,9 @@ inline float2 convolution_3x3_dilation_stridex2_stridey2_bifrost_f32(__global uc
     float3 src20_mid   = vload3(0, (__global float *)ptr_offset(src_addr, DILATION_X, y_offset + DILATION_Y * 2, stride_x_bytes, stride_y_bytes));
     float3 src20_right = vload3(0, (__global float *)ptr_offset(src_addr, 2 * DILATION_X, y_offset + DILATION_Y * 2, stride_x_bytes, stride_y_bytes));
 
-    CONVOLUTION1x3_BIFROST2X1_STRIDE2(pixels0, src00_left, src00_mid, src00_right, weights_row0);
-    CONVOLUTION1x3_BIFROST2X1_STRIDE2(pixels0, src10_left, src10_mid, src10_right, weights_row1);
-    CONVOLUTION1x3_BIFROST2X1_STRIDE2(pixels0, src20_left, src20_mid, src20_right, weights_row2);
+    CONVOLUTION1x3_2X1_STRIDE2(pixels0, src00_left, src00_mid, src00_right, weights_row0);
+    CONVOLUTION1x3_2X1_STRIDE2(pixels0, src10_left, src10_mid, src10_right, weights_row1);
+    CONVOLUTION1x3_2X1_STRIDE2(pixels0, src20_left, src20_mid, src20_right, weights_row2);
 
     return pixels0;
 }
@@ -491,7 +491,7 @@ inline float2 convolution_3x3_dilation_stridex2_stridey2_bifrost_f32(__global uc
  * @param[in] biases_step_x                         (Optional) biases_stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in] biases_offset_first_element_in_bytes  (Optional) The offset of the first element in the biases vector
  */
-__kernel void depthwise_convolution_3x3_stridex1_stridey1_bifrost_f32(
+__kernel void depthwise_convolution_3x3_stridex1_stridey1_f32(
     TENSOR3D_DECLARATION(src),
     TENSOR3D_DECLARATION(dst),
     TENSOR3D_DECLARATION(weights)
@@ -531,29 +531,29 @@ __kernel void depthwise_convolution_3x3_stridex1_stridey1_bifrost_f32(
     float4 src40 = vload4(0, (__global float *)(src_addr + 4 * src_stride_y)); // Row4
     float4 src50 = vload4(0, (__global float *)(src_addr + 5 * src_stride_y)); // Row5
 
-    CONVOLUTION1x3_BIFROST2X1_STRIDE1(pixels0, src00, weights_row0);
-    CONVOLUTION1x3_BIFROST2X1_STRIDE1(pixels0, src10, weights_row1);
-    CONVOLUTION1x3_BIFROST2X1_STRIDE1(pixels0, src20, weights_row2);
-    CONVOLUTION1x3_BIFROST2X1_STRIDE1(pixels1, src10, weights_row0);
-    CONVOLUTION1x3_BIFROST2X1_STRIDE1(pixels1, src20, weights_row1);
-    CONVOLUTION1x3_BIFROST2X1_STRIDE1(pixels1, src30, weights_row2);
-    CONVOLUTION1x3_BIFROST2X1_STRIDE1(pixels2, src20, weights_row0);
-    CONVOLUTION1x3_BIFROST2X1_STRIDE1(pixels2, src30, weights_row1);
-    CONVOLUTION1x3_BIFROST2X1_STRIDE1(pixels2, src40, weights_row2);
-    CONVOLUTION1x3_BIFROST2X1_STRIDE1(pixels3, src30, weights_row0);
-    CONVOLUTION1x3_BIFROST2X1_STRIDE1(pixels3, src40, weights_row1);
-    CONVOLUTION1x3_BIFROST2X1_STRIDE1(pixels3, src50, weights_row2);
+    CONVOLUTION1x3_2X1_STRIDE1(pixels0, src00, weights_row0);
+    CONVOLUTION1x3_2X1_STRIDE1(pixels0, src10, weights_row1);
+    CONVOLUTION1x3_2X1_STRIDE1(pixels0, src20, weights_row2);
+    CONVOLUTION1x3_2X1_STRIDE1(pixels1, src10, weights_row0);
+    CONVOLUTION1x3_2X1_STRIDE1(pixels1, src20, weights_row1);
+    CONVOLUTION1x3_2X1_STRIDE1(pixels1, src30, weights_row2);
+    CONVOLUTION1x3_2X1_STRIDE1(pixels2, src20, weights_row0);
+    CONVOLUTION1x3_2X1_STRIDE1(pixels2, src30, weights_row1);
+    CONVOLUTION1x3_2X1_STRIDE1(pixels2, src40, weights_row2);
+    CONVOLUTION1x3_2X1_STRIDE1(pixels3, src30, weights_row0);
+    CONVOLUTION1x3_2X1_STRIDE1(pixels3, src40, weights_row1);
+    CONVOLUTION1x3_2X1_STRIDE1(pixels3, src50, weights_row2);
 
 #else /* DILATION_X==1 && DILATION_Y==1 */
 
     //3x3 Convolution of elements starting in 0th row
-    pixels0 = convolution_3x3_dilation_stridex1_stridey1_bifrost_f32(src_addr, src_stride_x, src_stride_y, 0, weights_addr, weights_stride_y);
+    pixels0 = convolution_3x3_dilation_stridex1_stridey1_f32(src_addr, src_stride_x, src_stride_y, 0, weights_addr, weights_stride_y);
     //3x3 Convolution of elements starting in 1st row
-    pixels1 = convolution_3x3_dilation_stridex1_stridey1_bifrost_f32(src_addr, src_stride_x, src_stride_y, 1, weights_addr, weights_stride_y);
+    pixels1 = convolution_3x3_dilation_stridex1_stridey1_f32(src_addr, src_stride_x, src_stride_y, 1, weights_addr, weights_stride_y);
     //3x3 Convolution of elements starting in 2nd row
-    pixels2 = convolution_3x3_dilation_stridex1_stridey1_bifrost_f32(src_addr, src_stride_x, src_stride_y, 2, weights_addr, weights_stride_y);
+    pixels2 = convolution_3x3_dilation_stridex1_stridey1_f32(src_addr, src_stride_x, src_stride_y, 2, weights_addr, weights_stride_y);
     //3x3 Convolution of elements starting in 3rd row
-    pixels3 = convolution_3x3_dilation_stridex1_stridey1_bifrost_f32(src_addr, src_stride_x, src_stride_y, 3, weights_addr, weights_stride_y);
+    pixels3 = convolution_3x3_dilation_stridex1_stridey1_f32(src_addr, src_stride_x, src_stride_y, 3, weights_addr, weights_stride_y);
 
 #endif /* DILATION_X==1 && DILATION_Y==1 */
 
@@ -611,7 +611,7 @@ __kernel void depthwise_convolution_3x3_stridex1_stridey1_bifrost_f32(
  * @param[in] biases_step_x                         (Optional) biases_stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in] biases_offset_first_element_in_bytes  (Optional) The offset of the first element in the biases vector
  */
-__kernel void depthwise_convolution_3x3_stridex2_stridey2_bifrost_f32(
+__kernel void depthwise_convolution_3x3_stridex2_stridey2_f32(
     TENSOR3D_DECLARATION(src),
     TENSOR3D_DECLARATION(dst),
     TENSOR3D_DECLARATION(weights)
@@ -654,19 +654,19 @@ __kernel void depthwise_convolution_3x3_stridex2_stridey2_bifrost_f32(
     float4 src40 = vload4(0, (__global float *)(src_addr + 4 * src_stride_y)); // Row4
     float2 src41 = vload2(2, (__global float *)(src_addr + 4 * src_stride_y)); // Row4
 
-    CONVOLUTION1x3_BIFROST2X1_STRIDE2(pixels0, src00, src01, weights_row0);
-    CONVOLUTION1x3_BIFROST2X1_STRIDE2(pixels0, src10, src11, weights_row1);
-    CONVOLUTION1x3_BIFROST2X1_STRIDE2(pixels0, src20, src21, weights_row2);
-    CONVOLUTION1x3_BIFROST2X1_STRIDE2(pixels1, src20, src21, weights_row0);
-    CONVOLUTION1x3_BIFROST2X1_STRIDE2(pixels1, src30, src31, weights_row1);
-    CONVOLUTION1x3_BIFROST2X1_STRIDE2(pixels1, src40, src41, weights_row2);
+    CONVOLUTION1x3_2X1_STRIDE2(pixels0, src00, src01, weights_row0);
+    CONVOLUTION1x3_2X1_STRIDE2(pixels0, src10, src11, weights_row1);
+    CONVOLUTION1x3_2X1_STRIDE2(pixels0, src20, src21, weights_row2);
+    CONVOLUTION1x3_2X1_STRIDE2(pixels1, src20, src21, weights_row0);
+    CONVOLUTION1x3_2X1_STRIDE2(pixels1, src30, src31, weights_row1);
+    CONVOLUTION1x3_2X1_STRIDE2(pixels1, src40, src41, weights_row2);
 
 #else  /* DILATION_X==1 && DILATION_Y==1 */
 
     //3x3 Convolution of elements starting in 0th row
-    pixels0 = convolution_3x3_dilation_stridex2_stridey2_bifrost_f32(src_addr, src_stride_x, src_stride_y, 0, weights_addr, weights_stride_y);
+    pixels0 = convolution_3x3_dilation_stridex2_stridey2_f32(src_addr, src_stride_x, src_stride_y, 0, weights_addr, weights_stride_y);
     //3x3 Convolution of elements starting in 2nd row
-    pixels1 = convolution_3x3_dilation_stridex2_stridey2_bifrost_f32(src_addr, src_stride_x, src_stride_y, 2, weights_addr, weights_stride_y);
+    pixels1 = convolution_3x3_dilation_stridex2_stridey2_f32(src_addr, src_stride_x, src_stride_y, 2, weights_addr, weights_stride_y);
 #endif /* DILATION_X==1 && DILATION_Y==1 */
 
 #ifdef HAS_BIAS
@@ -684,104 +684,6 @@ __kernel void depthwise_convolution_3x3_stridex2_stridey2_bifrost_f32(
 
 #endif // defined(DEPTH_MULTIPLIER) && defined(DST_CHANNELS) && defined(IS_F32)
 
-#if defined(VEC_SIZE) && defined(DATA_TYPE) && defined(DST_WIDTH)
-/** Reshape the weights for quantized depthwise convolution
- *
- * @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type, e.g. -DDATA_TYPE=uint8
- * @note Output width should be given as a preprocessor argument using -DDST_WIDTH=width, e.g. -DDST_WIDTH=128
- * @note Vector size should be given as a preprocessor argument using -DVEC_SIZE=vec_size, e.g., -DVEC_SIZE=4
- * @attention Input's height and width should be 3
- *
- * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: All
- * @param[in]  src_stride_x                      Stride of the source tensor in X dimension (in bytes)
- * @param[in]  src_step_x                        src_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in]  src_stride_y                      Stride of the source tensor in Y dimension (in bytes)
- * @param[in]  src_step_y                        src_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in]  src_stride_z                      Stride of the source tensor in Z dimension (in bytes)
- * @param[in]  src_step_z                        src_stride_z * number of elements along Y processed per workitem(in bytes)
- * @param[in]  src_offset_first_element_in_bytes The offset of the first element in the source tensor
- * @param[out] dst_ptr                           Pointer to the destination tensor. Supported data types: same as @p src_ptr
- * @param[in]  dst_stride_x                      Stride of the destination tensor in X dimension (in bytes)
- * @param[in]  dst_step_x                        dst_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in]  dst_stride_y                      Stride of the destination tensor in Y dimension (in bytes)
- * @param[in]  dst_step_y                        dst_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in]  dst_offset_first_element_in_bytes The offset of the first element in the destination tensor
- */
-__kernel void depthwise_convolution_reshape_weights(
-    TENSOR3D_DECLARATION(src),
-    IMAGE_DECLARATION(dst))
-{
-    Vector    src = CONVERT_TO_VECTOR_STRUCT(src);
-    const int x   = get_global_id(0);
-
-    // Load 3x3xVEC_SIZE weights
-    VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
-    w0 = VLOAD(VEC_SIZE)(0, src.ptr + 0 * src_stride_y + 0 * src_stride_z);
-    VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
-    w1 = VLOAD(VEC_SIZE)(0, src.ptr + 1 * src_stride_y + 0 * src_stride_z);
-    VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
-    w2 = VLOAD(VEC_SIZE)(0, src.ptr + 2 * src_stride_y + 0 * src_stride_z);
-    VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
-    w3 = VLOAD(VEC_SIZE)(0, src.ptr + 0 * src_stride_y + 1 * src_stride_z);
-    VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
-    w4 = VLOAD(VEC_SIZE)(0, src.ptr + 1 * src_stride_y + 1 * src_stride_z);
-    VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
-    w5 = VLOAD(VEC_SIZE)(0, src.ptr + 2 * src_stride_y + 1 * src_stride_z);
-    VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
-    w6 = VLOAD(VEC_SIZE)(0, src.ptr + 0 * src_stride_y + 2 * src_stride_z);
-    VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
-    w7 = VLOAD(VEC_SIZE)(0, src.ptr + 1 * src_stride_y + 2 * src_stride_z);
-    VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
-    w8 = VLOAD(VEC_SIZE)(0, src.ptr + 2 * src_stride_y + 2 * src_stride_z);
-
-    __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + x * DST_WIDTH * sizeof(DATA_TYPE);
-
-#if defined(TRANSPOSE)
-#if VEC_SIZE != 4
-#error "VEC_SIZE not supported"
-#else  // VEC_SIZE != 4
-    VSTORE(VEC_SIZE)
-    ((VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE))(w0.s0, w1.s0, w2.s0, w3.s0), 0, dst_addr + 0);
-    VSTORE(VEC_SIZE)
-    ((VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE))(w4.s0, w5.s0, w6.s0, w7.s0), 0, dst_addr + 1 * sizeof(DATA_TYPE) * VEC_SIZE);
-    VSTORE(VEC_SIZE)
-    ((VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE))(w8.s0, w0.s1, w1.s1, w2.s1), 0, dst_addr + 2 * sizeof(DATA_TYPE) * VEC_SIZE);
-    VSTORE(VEC_SIZE)
-    ((VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE))(w3.s1, w4.s1, w5.s1, w6.s1), 0, dst_addr + 3 * sizeof(DATA_TYPE) * VEC_SIZE);
-    VSTORE(VEC_SIZE)
-    ((VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE))(w7.s1, w8.s1, w0.s2, w1.s2), 0, dst_addr + 4 * sizeof(DATA_TYPE) * VEC_SIZE);
-    VSTORE(VEC_SIZE)
-    ((VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE))(w2.s2, w3.s2, w4.s2, w5.s2), 0, dst_addr + 5 * sizeof(DATA_TYPE) * VEC_SIZE);
-    VSTORE(VEC_SIZE)
-    ((VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE))(w6.s2, w7.s2, w8.s2, w0.s3), 0, dst_addr + 6 * sizeof(DATA_TYPE) * VEC_SIZE);
-    VSTORE(VEC_SIZE)
-    ((VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE))(w1.s3, w2.s3, w3.s3, w4.s3), 0, dst_addr + 7 * sizeof(DATA_TYPE) * VEC_SIZE);
-    VSTORE(VEC_SIZE)
-    ((VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE))(w5.s3, w6.s3, w7.s3, w8.s3), 0, dst_addr + 8 * sizeof(DATA_TYPE) * VEC_SIZE);
-#endif // VEC_SIZE != 4
-#else  // !defined(TRANSPOSE)
-    VSTORE(VEC_SIZE)
-    (w0, 0, dst_addr + 0);
-    VSTORE(VEC_SIZE)
-    (w1, 0, dst_addr + 1 * sizeof(DATA_TYPE) * VEC_SIZE);
-    VSTORE(VEC_SIZE)
-    (w2, 0, dst_addr + 2 * sizeof(DATA_TYPE) * VEC_SIZE);
-    VSTORE(VEC_SIZE)
-    (w3, 0, dst_addr + 3 * sizeof(DATA_TYPE) * VEC_SIZE);
-    VSTORE(VEC_SIZE)
-    (w4, 0, dst_addr + 4 * sizeof(DATA_TYPE) * VEC_SIZE);
-    VSTORE(VEC_SIZE)
-    (w5, 0, dst_addr + 5 * sizeof(DATA_TYPE) * VEC_SIZE);
-    VSTORE(VEC_SIZE)
-    (w6, 0, dst_addr + 6 * sizeof(DATA_TYPE) * VEC_SIZE);
-    VSTORE(VEC_SIZE)
-    (w7, 0, dst_addr + 7 * sizeof(DATA_TYPE) * VEC_SIZE);
-    VSTORE(VEC_SIZE)
-    (w8, 0, dst_addr + 8 * sizeof(DATA_TYPE) * VEC_SIZE);
-#endif // defined(TRANSPOSE)
-}
-#endif // defined(VEC_SIZE) && defined(DATA_TYPE) && defined(DST_WIDTH)
-
 #if defined(ARM_COMPUTE_OPENCL_FP16_ENABLED) && defined(DEPTH_MULTIPLIER) && defined(DST_CHANNELS) && defined(IS_F16)
 #if defined(CONV_STRIDE_X)
 #if CONV_STRIDE_X == 1
@@ -805,8 +707,8 @@ __kernel void depthwise_convolution_reshape_weights(
  * @param[in] weights_addr     Pointer from where to get weights
  * @param[in] weights_stride_y Stride of weights tesnsor in Y dimension
  */
-inline half4 convolution_3x3_dilation_stridex1_stridey1_bifrost_f16(__global uchar *src_addr, const int stride_x_bytes, const int stride_y_bytes,
-                                                                    const int y_offset, __global uchar *weights_addr, const int weights_stride_y)
+inline half4 convolution_3x3_dilation_stridex1_stridey1_f16(__global uchar *src_addr, const int stride_x_bytes, const int stride_y_bytes,
+                                                            const int y_offset, __global uchar *weights_addr, const int weights_stride_y)
 {
     // Load the weights
     half3 weights_row0 = vload3(0, (__global half *)(weights_addr + 0 * weights_stride_y));
@@ -827,9 +729,9 @@ inline half4 convolution_3x3_dilation_stridex1_stridey1_bifrost_f16(__global uch
     half4 src20_mid   = vload4(0, (__global half *)ptr_offset(src_addr, DILATION_X, y_offset + DILATION_Y * 2, stride_x_bytes, stride_y_bytes));
     half4 src20_right = vload4(0, (__global half *)ptr_offset(src_addr, 2 * DILATION_X, y_offset + DILATION_Y * 2, stride_x_bytes, stride_y_bytes));
 
-    CONVOLUTION1x3_BIFROST4X1_STRIDE1(pixels0, src00_left, src00_mid, src00_right, weights_row0);
-    CONVOLUTION1x3_BIFROST4X1_STRIDE1(pixels0, src10_left, src10_mid, src10_right, weights_row1);
-    CONVOLUTION1x3_BIFROST4X1_STRIDE1(pixels0, src20_left, src20_mid, src20_right, weights_row2);
+    CONVOLUTION1x3_4X1_STRIDE1(pixels0, src00_left, src00_mid, src00_right, weights_row0);
+    CONVOLUTION1x3_4X1_STRIDE1(pixels0, src10_left, src10_mid, src10_right, weights_row1);
+    CONVOLUTION1x3_4X1_STRIDE1(pixels0, src20_left, src20_mid, src20_right, weights_row2);
 
     return pixels0;
 }
@@ -843,8 +745,8 @@ inline half4 convolution_3x3_dilation_stridex1_stridey1_bifrost_f16(__global uch
  * @param[in] weights_addr     Pointer from where to get weights
  * @param[in] weights_stride_y Stride of weights tesnsor in Y dimension
  */
-inline half4 convolution_3x3_dilation_stridex2_stridey2_bifrost_f16(__global uchar *src_addr, const int stride_x_bytes, const int stride_y_bytes,
-                                                                    const int y_offset, __global uchar *weights_addr, const int weights_stride_y)
+inline half4 convolution_3x3_dilation_stridex2_stridey2_f16(__global uchar *src_addr, const int stride_x_bytes, const int stride_y_bytes,
+                                                            const int y_offset, __global uchar *weights_addr, const int weights_stride_y)
 {
     // Load the weights
     half3 weights_row0 = vload3(0, (__global half *)(weights_addr + 0 * weights_stride_y));
@@ -865,9 +767,9 @@ inline half4 convolution_3x3_dilation_stridex2_stridey2_bifrost_f16(__global uch
     half8 src20_mid   = vload8(0, (__global half *)ptr_offset(src_addr, DILATION_X, y_offset + DILATION_Y * 2, stride_x_bytes, stride_y_bytes));
     half8 src20_right = vload8(0, (__global half *)ptr_offset(src_addr, 2 * DILATION_X, y_offset + DILATION_Y * 2, stride_x_bytes, stride_y_bytes));
 
-    CONVOLUTION1x3_BIFROST4X1_STRIDE2(pixels0, src00_left, src00_mid, src00_right, weights_row0);
-    CONVOLUTION1x3_BIFROST4X1_STRIDE2(pixels0, src10_left, src10_mid, src10_right, weights_row1);
-    CONVOLUTION1x3_BIFROST4X1_STRIDE2(pixels0, src20_left, src20_mid, src20_right, weights_row2);
+    CONVOLUTION1x3_4X1_STRIDE2(pixels0, src00_left, src00_mid, src00_right, weights_row0);
+    CONVOLUTION1x3_4X1_STRIDE2(pixels0, src10_left, src10_mid, src10_right, weights_row1);
+    CONVOLUTION1x3_4X1_STRIDE2(pixels0, src20_left, src20_mid, src20_right, weights_row2);
 
     return pixels0;
 }
@@ -1127,7 +1029,7 @@ __kernel void depthwise_convolution_3x3_f16(
  * @param[in] biases_step_x                         (Optional) biases_stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in] biases_offset_first_element_in_bytes  (Optional) The offset of the first element in the biases vector
  */
-__kernel void depthwise_convolution_3x3_stridex1_stridey1_bifrost_f16(
+__kernel void depthwise_convolution_3x3_stridex1_stridey1_f16(
     TENSOR3D_DECLARATION(src),
     TENSOR3D_DECLARATION(dst),
     TENSOR3D_DECLARATION(weights)
@@ -1174,29 +1076,29 @@ __kernel void depthwise_convolution_3x3_stridex1_stridey1_bifrost_f16(
     half8 src40 = vload8(0, (__global half *)(src_addr + 4 * src_stride_y)); // Row4
     half8 src50 = vload8(0, (__global half *)(src_addr + 5 * src_stride_y)); // Row5
 
-    CONVOLUTION1x3_BIFROST4X1_STRIDE1(pixels0, src00, weights_row0);
-    CONVOLUTION1x3_BIFROST4X1_STRIDE1(pixels0, src10, weights_row1);
-    CONVOLUTION1x3_BIFROST4X1_STRIDE1(pixels0, src20, weights_row2);
-    CONVOLUTION1x3_BIFROST4X1_STRIDE1(pixels1, src10, weights_row0);
-    CONVOLUTION1x3_BIFROST4X1_STRIDE1(pixels1, src20, weights_row1);
-    CONVOLUTION1x3_BIFROST4X1_STRIDE1(pixels1, src30, weights_row2);
-    CONVOLUTION1x3_BIFROST4X1_STRIDE1(pixels2, src20, weights_row0);
-    CONVOLUTION1x3_BIFROST4X1_STRIDE1(pixels2, src30, weights_row1);
-    CONVOLUTION1x3_BIFROST4X1_STRIDE1(pixels2, src40, weights_row2);
-    CONVOLUTION1x3_BIFROST4X1_STRIDE1(pixels3, src30, weights_row0);
-    CONVOLUTION1x3_BIFROST4X1_STRIDE1(pixels3, src40, weights_row1);
-    CONVOLUTION1x3_BIFROST4X1_STRIDE1(pixels3, src50, weights_row2);
+    CONVOLUTION1x3_4X1_STRIDE1(pixels0, src00, weights_row0);
+    CONVOLUTION1x3_4X1_STRIDE1(pixels0, src10, weights_row1);
+    CONVOLUTION1x3_4X1_STRIDE1(pixels0, src20, weights_row2);
+    CONVOLUTION1x3_4X1_STRIDE1(pixels1, src10, weights_row0);
+    CONVOLUTION1x3_4X1_STRIDE1(pixels1, src20, weights_row1);
+    CONVOLUTION1x3_4X1_STRIDE1(pixels1, src30, weights_row2);
+    CONVOLUTION1x3_4X1_STRIDE1(pixels2, src20, weights_row0);
+    CONVOLUTION1x3_4X1_STRIDE1(pixels2, src30, weights_row1);
+    CONVOLUTION1x3_4X1_STRIDE1(pixels2, src40, weights_row2);
+    CONVOLUTION1x3_4X1_STRIDE1(pixels3, src30, weights_row0);
+    CONVOLUTION1x3_4X1_STRIDE1(pixels3, src40, weights_row1);
+    CONVOLUTION1x3_4X1_STRIDE1(pixels3, src50, weights_row2);
 
 #else /* DILATION_X==1 && DILATION_Y==1 */
 
     //3x3 Convolution of elements starting in 0th row
-    pixels0 = convolution_3x3_dilation_stridex1_stridey1_bifrost_f16(src_addr, src_stride_x, src_stride_y, 0, weights_addr, weights_stride_y);
+    pixels0 = convolution_3x3_dilation_stridex1_stridey1_f16(src_addr, src_stride_x, src_stride_y, 0, weights_addr, weights_stride_y);
     //3x3 Convolution of elements starting in 1st row
-    pixels1 = convolution_3x3_dilation_stridex1_stridey1_bifrost_f16(src_addr, src_stride_x, src_stride_y, 1, weights_addr, weights_stride_y);
+    pixels1 = convolution_3x3_dilation_stridex1_stridey1_f16(src_addr, src_stride_x, src_stride_y, 1, weights_addr, weights_stride_y);
     //3x3 Convolution of elements starting in 2nd row
-    pixels2 = convolution_3x3_dilation_stridex1_stridey1_bifrost_f16(src_addr, src_stride_x, src_stride_y, 2, weights_addr, weights_stride_y);
+    pixels2 = convolution_3x3_dilation_stridex1_stridey1_f16(src_addr, src_stride_x, src_stride_y, 2, weights_addr, weights_stride_y);
     //3x3 Convolution of elements starting in 3rd row
-    pixels3 = convolution_3x3_dilation_stridex1_stridey1_bifrost_f16(src_addr, src_stride_x, src_stride_y, 3, weights_addr, weights_stride_y);
+    pixels3 = convolution_3x3_dilation_stridex1_stridey1_f16(src_addr, src_stride_x, src_stride_y, 3, weights_addr, weights_stride_y);
 
 #endif /* DILATION_X==1 && DILATION_Y==1 */
 
@@ -1250,7 +1152,7 @@ __kernel void depthwise_convolution_3x3_stridex1_stridey1_bifrost_f16(
  * @param[in] biases_step_x                         (Optional) biases_stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in] biases_offset_first_element_in_bytes  (Optional) The offset of the first element in the biases vector
  */
-__kernel void depthwise_convolution_3x3_stridex2_stridey2_bifrost_f16(
+__kernel void depthwise_convolution_3x3_stridex2_stridey2_f16(
     TENSOR3D_DECLARATION(src),
     TENSOR3D_DECLARATION(dst),
     TENSOR3D_DECLARATION(weights)
@@ -1300,18 +1202,18 @@ __kernel void depthwise_convolution_3x3_stridex2_stridey2_bifrost_f16(
     half8 src40 = vload8(0, (__global half *)(src_addr + 4 * src_stride_y)); // Row4
     half2 src41 = vload2(4, (__global half *)(src_addr + 4 * src_stride_y)); // Row4
 
-    CONVOLUTION1x3_BIFROST4X1_STRIDE2(pixels0, src00, src01, weights_row0);
-    CONVOLUTION1x3_BIFROST4X1_STRIDE2(pixels0, src10, src11, weights_row1);
-    CONVOLUTION1x3_BIFROST4X1_STRIDE2(pixels0, src20, src21, weights_row2);
-    CONVOLUTION1x3_BIFROST4X1_STRIDE2(pixels1, src20, src21, weights_row0);
-    CONVOLUTION1x3_BIFROST4X1_STRIDE2(pixels1, src30, src31, weights_row1);
-    CONVOLUTION1x3_BIFROST4X1_STRIDE2(pixels1, src40, src41, weights_row2);
+    CONVOLUTION1x3_4X1_STRIDE2(pixels0, src00, src01, weights_row0);
+    CONVOLUTION1x3_4X1_STRIDE2(pixels0, src10, src11, weights_row1);
+    CONVOLUTION1x3_4X1_STRIDE2(pixels0, src20, src21, weights_row2);
+    CONVOLUTION1x3_4X1_STRIDE2(pixels1, src20, src21, weights_row0);
+    CONVOLUTION1x3_4X1_STRIDE2(pixels1, src30, src31, weights_row1);
+    CONVOLUTION1x3_4X1_STRIDE2(pixels1, src40, src41, weights_row2);
 
 #else  /* DILATION_X==1 && DILATION_Y==1 */
     //3x3 Convolution of elements starting in 0th row
-    pixels0 = convolution_3x3_dilation_stridex2_stridey2_bifrost_f16(src_addr, src_stride_x, src_stride_y, 0, weights_addr, weights_stride_y);
+    pixels0 = convolution_3x3_dilation_stridex2_stridey2_f16(src_addr, src_stride_x, src_stride_y, 0, weights_addr, weights_stride_y);
     //3x3 Convolution of elements starting in 2nd row
-    pixels1                  = convolution_3x3_dilation_stridex2_stridey2_bifrost_f16(src_addr, src_stride_x, src_stride_y, 2, weights_addr, weights_stride_y);
+    pixels1                  = convolution_3x3_dilation_stridex2_stridey2_f16(src_addr, src_stride_x, src_stride_y, 2, weights_addr, weights_stride_y);
 #endif /* DILATION_X==1 && DILATION_Y==1 */
 
 #ifdef HAS_BIAS
diff --git a/src/core/CL/cl_kernels/depthwise_convolution_quantized.cl b/src/core/CL/cl_kernels/depthwise_convolution_quantized.cl
index c7fe401f80..000dce1590 100644
--- a/src/core/CL/cl_kernels/depthwise_convolution_quantized.cl
+++ b/src/core/CL/cl_kernels/depthwise_convolution_quantized.cl
@@ -334,9 +334,9 @@ __kernel void dwc_3x3_native_quantized8_nchw(
 #else // defined(REAL_MULTIPLIER)
 
 #if defined(PER_CHANNEL_QUANTIZATION)
-    int8 res0_shift_lt0                = ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(values0, output_multiplier, output_shift, 8);
-    int8 res0_shift_gt0                = ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE(values0, output_multiplier, output_shift, 8);
-    values0                            = select(res0_shift_lt0, res0_shift_gt0, (int8)(output_shift) >= 0);
+    int8 res0_shift_lt0 = ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(values0, output_multiplier, output_shift, 8);
+    int8 res0_shift_gt0 = ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE(values0, output_multiplier, output_shift, 8);
+    values0             = select(res0_shift_lt0, res0_shift_gt0, (int8)(output_shift) >= 0);
 #else // defined(PER_CHANNEL_QUANTIZATION)
 #if OUTPUT_SHIFT < 0
     values0 = ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(values0, OUTPUT_MULTIPLIER, OUTPUT_SHIFT, 8);
@@ -360,9 +360,9 @@ __kernel void dwc_3x3_native_quantized8_nchw(
 #else // defined(REAL_MULTIPLIER)
 
 #if defined(PER_CHANNEL_QUANTIZATION)
-    int8 res1_shift_lt0      = ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(values1, output_multiplier, output_shift, 8);
-    int8 res1_shift_gt0      = ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE(values1, output_multiplier, output_shift, 8);
-    values1                  = select(res1_shift_lt0, res1_shift_gt0, (int8)(output_shift) >= 0);
+    int8 res1_shift_lt0 = ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(values1, output_multiplier, output_shift, 8);
+    int8 res1_shift_gt0 = ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE(values1, output_multiplier, output_shift, 8);
+    values1             = select(res1_shift_lt0, res1_shift_gt0, (int8)(output_shift) >= 0);
 #else // defined(PER_CHANNEL_QUANTIZATION)
 #if OUTPUT_SHIFT < 0
     values1 = ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(values1, OUTPUT_MULTIPLIER, OUTPUT_SHIFT, 8);
@@ -446,8 +446,8 @@ __kernel void dwc_3x3_native_quantized8_nchw(
         VEC_TYPE(16)                                                                                        \
         temp0 = vload16(0, (__global DATA_TYPE *)(first_value));                                            \
         VEC_TYPE(8)                                                                                         \
-        temp1 = vload8(0, (__global DATA_TYPE *)(first_value + 16 * sizeof(DATA_TYPE))));                   \
-        left = (VEC_TYPE(8))(temp0.s0369, temp0.scf, temp1.s25);                                            \
+        temp1 = vload8(0, (__global DATA_TYPE *)(first_value + 16 * sizeof(DATA_TYPE)));                    \
+        left  = (VEC_TYPE(8))(temp0.s0369, temp0.scf, temp1.s25);                                           \
         \
         temp0  = vload16(0, (__global DATA_TYPE *)(first_value + DILATION_X * sizeof(DATA_TYPE)));          \
         temp1  = vload8(0, (__global DATA_TYPE *)(first_value + (16 + DILATION_X) * sizeof(DATA_TYPE)));    \
@@ -776,835 +776,6 @@ __kernel void dwc_3x3_native_quantized8_dot8_nchw(
 
 #endif // defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8)
 
-#if defined(CONV_STRIDE_X) && defined(CONV_STRIDE_Y) && VEC_SIZE == 4
-/** This function computes the depthwise convolution quantized for NHWC data layout when the stride along the width or height is not 1.
- *
- * @note This kernel assumes VEC_SIZE is 4.
- * @note The weights tensor is expected to be reshaped using @ref CLDepthwiseConvolutionLayerReshapeWeightsKernel.
- * @note The number of elements read per thread must be passed at compile time using -DVEC_SIZE (e.g. -DVEC_SIZE=2)
- * @note Dimension two of the input tensor (height for NHWC data layout) must be passed at compile time using -DSRC_DIM2 (e.g. -DSRC_DIM_2=112)
- * @note The convolution pad top must be passed at compile time using -DCONV_PAD_TOP (e.g. -DCONV_PAD_TOP=1)
- * @note The convolution pad top must be passed at compile time using -DCONV_PAD_LEFT (e.g. -DCONV_PAD_LEFT=1)
- * @note The convolution stride along the width must be passed at compile time using -DCONV_STRIDE_X (e.g. -DCONV_STRIDE_Y=X)
- * @note The convolution stride along the height must be passed at compile time using -DCONV_STRIDE_Y (e.g. -DCONV_STRIDE_Y=1)
- *
- * @param[in] src_ptr                                          Pointer to the source tensor. Supported data types: QASYMM8/QASYMM8_SIGNED
- * @param[in] src_stride_x                                     Stride of the source tensor in X dimension (in bytes)
- * @param[in] src_step_x                                       src_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in] src_stride_y                                     Stride of the source tensor in Y dimension (in bytes)
- * @param[in] src_step_y                                       src_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in] src_stride_z                                     Stride of the source tensor in Z dimension (in bytes)
- * @param[in] src_step_z                                       src_stride_y * number of elements along Z processed per workitem(in bytes)
- * @param[in] src_stride_w                                     Stride of the source tensor in W dimension (in bytes)
- * @param[in] src_step_w                                       src_stride_w * number of elements along W processed per workitem(in bytes)
- * @param[in] src_offset_first_element_in_bytes                The offset of the first element in the source tensor
- * @param[in] dst_ptr                                          Pointer to the destination tensor. Supported data types: same as @p src_ptr
- * @param[in] dst_stride_x                                     Stride of the destination tensor in X dimension (in bytes)
- * @param[in] dst_step_x                                       dst_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in] dst_stride_y                                     Stride of the destination tensor in Y dimension (in bytes)
- * @param[in] dst_step_y                                       dst_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in] dst_stride_z                                     Stride of the destination tensor in Z dimension (in bytes)
- * @param[in] dst_step_z                                       dst_stride_z * number of elements along Y processed per workitem(in bytes)
- * @param[in] dst_stride_w                                     Stride of the destination tensor in W dimension (in bytes)
- * @param[in] dst_step_w                                       dst_stride_w * number of elements along W processed per workitem(in bytes)
- * @param[in] dst_offset_first_element_in_bytes                The offset of the first element in the destination tensor
- * @param[in] weights_ptr                                      Pointer to the weights tensor reshaped. Supported data types: QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL
- * @param[in] weights_stride_x                                 Stride of the weights tensor in X dimension (in bytes)
- * @param[in] weights_step_x                                   weights_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in] weights_stride_y                                 Stride of the weights tensor in Y dimension (in bytes)
- * @param[in] weights_step_y                                   weights_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in] weights_offset_first_element_in_bytes            The offset of the first element in the weights tensor
- * @param[in] output_multipliers_ptr                           Pointer to the output multipliers vector. Supported data types: S32
- * @param[in] output_multipliers_stride_x                      Stride of the output multipliers vector in X dimension (in bytes)
- * @param[in] output_multipliers_step_x                        output_multipliers_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in] output_multipliers_offset_first_element_in_bytes The offset of the first element in the output multipliers vector
- * @param[in] output_shifts_ptr                                Pointer to the output shifts vector. Supported data types: S32
- * @param[in] output_shifts_stride_x                           Stride of the output shifts vector in X dimension (in bytes)
- * @param[in] output_shifts_step_x                             output_shifts_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in] output_shifts_offset_first_element_in_bytes      The offset of the first element in the output shifts vector
- * @param[in] biases_ptr                                       (Optional) Pointer to the biases vector. Supported data types: S32
- * @param[in] biases_stride_x                                  (Optional) Stride of the biases vector in X dimension (in bytes)
- * @param[in] biases_step_x                                    (Optional) biases_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in] biases_offset_first_element_in_bytes             (Optional) The offset of the first element in the biases vector
- * @param[in] max_offset                                       Max offset for the input tensor
- */
-__kernel void dwc_3x3_reshaped_quantized8_nhwc(
-    TENSOR4D_DECLARATION(src),
-    TENSOR4D_DECLARATION(dst),
-    IMAGE_DECLARATION(weights),
-    VECTOR_DECLARATION(output_multipliers),
-    VECTOR_DECLARATION(output_shifts),
-#if defined(HAS_BIAS)
-    VECTOR_DECLARATION(biases),
-#endif /* defined(HAS_BIAS) */
-    int max_offset)
-{
-    const int x = get_global_id(0); // channels
-    const int y = get_global_id(1); // spatial coordinate x
-#if defined(DST_DEPTH)
-    int z = get_global_id(2) % (int)DST_DEPTH; // spatial coordinate y
-    int b = get_global_id(2) / (int)DST_DEPTH; // batch
-#else                                          // defined(DST_DEPTH)
-    int      z                         = get_global_id(2); // spatial coordinate y
-#endif                                         // defined(DST_DEPTH)
-
-    __global uchar *weights_addr = weights_ptr + weights_offset_first_element_in_bytes + x * weights_stride_y;
-
-#if defined(DST_DEPTH)
-    __global uchar *src_addr = src_ptr + src_offset_first_element_in_bytes + x * VEC_SIZE + b * src_stride_w;
-#else  /* defined(DST_DEPTH) */
-    __global uchar *src_addr           = src_ptr + src_offset_first_element_in_bytes + x * VEC_SIZE;
-#endif /* defined(DST_DEPTH) */
-
-    int  z_coord = 0;
-    int4 offset  = 0;
-    int4 y_coord = ((int4)(y * CONV_STRIDE_X) + (int4)(0, DILATION_X * 1, DILATION_X * 2, DILATION_X * 3)) - (int)CONV_PAD_LEFT;
-
-    // Only for y = 0 we can have a negative coordinate. If so, we convert it to SRC_DIM_1
-    y_coord.s0 = min((uint)y_coord.s0, (uint)SRC_DIM_1);
-    y_coord.s1 = min((uint)y_coord.s1, (uint)SRC_DIM_1);
-    y_coord.s2 = min((uint)y_coord.s2, (uint)SRC_DIM_1);
-    y_coord.s3 = min((uint)y_coord.s3, (uint)SRC_DIM_1);
-
-    int4 y_offset = convert_int4(y_coord * (int)src_stride_y);
-
-    // We compute VEC_SIZEx1x1 [C,W,H] elements
-    VEC_INT acc = 0, sum = 0;
-
-    // Load weights
-    VEC_DATA_TYPE(WEIGHTS_TYPE, 16)
-    w0_tmp = VLOAD(16)(0, (__global WEIGHTS_TYPE *)(weights_addr));
-    VEC_DATA_TYPE(WEIGHTS_TYPE, 16)
-    w1_tmp = VLOAD(16)(0, (__global WEIGHTS_TYPE *)(weights_addr + 16));
-    VEC_DATA_TYPE(WEIGHTS_TYPE, 4)
-    w8 = VLOAD(4)(0, (__global WEIGHTS_TYPE *)(weights_addr + 2 * 16));
-
-    VEC_DATA_TYPE(WEIGHTS_TYPE, 4)
-    w0 = w0_tmp.s0123;
-    VEC_DATA_TYPE(WEIGHTS_TYPE, 4)
-    w1 = w0_tmp.s4567;
-    VEC_DATA_TYPE(WEIGHTS_TYPE, 4)
-    w2 = w0_tmp.s89AB;
-    VEC_DATA_TYPE(WEIGHTS_TYPE, 4)
-    w3 = w0_tmp.sCDEF;
-
-    VEC_DATA_TYPE(WEIGHTS_TYPE, 4)
-    w4 = w1_tmp.s0123;
-    VEC_DATA_TYPE(WEIGHTS_TYPE, 4)
-    w5 = w1_tmp.s4567;
-    VEC_DATA_TYPE(WEIGHTS_TYPE, 4)
-    w6 = w1_tmp.s89AB;
-    VEC_DATA_TYPE(WEIGHTS_TYPE, 4)
-    w7 = w1_tmp.sCDEF;
-
-#if INPUT_OFFSET != 0
-    VEC_INT sum_we = CONVERT(w0, VEC_INT) + CONVERT(w1, VEC_INT) + CONVERT(w2, VEC_INT)
-                     + CONVERT(w3, VEC_INT) + CONVERT(w4, VEC_INT) + CONVERT(w5, VEC_INT)
-                     + CONVERT(w6, VEC_INT) + CONVERT(w7, VEC_INT) + CONVERT(w8, VEC_INT);
-#endif /* INPUT_OFFSET != 0 */
-
-    // Load input values
-    // z == 0
-    // Clamp z_coord as for z = 0, it can be negative
-    // z_coord is casted to unsigned int in order to use just a min() operation
-    // A "-1" 32 bit signed variable converted to unsigned gives 4294967295
-    z_coord = z * (int)CONV_STRIDE_Y - (int)CONV_PAD_TOP;
-    z_coord = min((uint)z_coord, (uint)SRC_DIM_2);
-    offset  = select(y_offset + (int4)(z_coord * src_stride_z), (int4)max_offset, (int4)z_coord < 0 || (int4)z_coord >= SRC_DIM_2);
-    VEC_TYPE(VEC_SIZE)
-    values0 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s0));
-    VEC_TYPE(VEC_SIZE)
-    values1 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s1));
-    VEC_TYPE(VEC_SIZE)
-    values2 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s2));
-
-    // z == 1
-    // z_coord can be only negative for z = 0 so we do not need to clamp it
-    // Moreover z_coord cannot be out-of-bound for z = 1 so we do not need to clamp the offset
-    z_coord = z * (int)CONV_STRIDE_Y - (int)CONV_PAD_TOP + DILATION_Y;
-    z_coord = min((uint)z_coord, (uint)SRC_DIM_2);
-    offset  = select(y_offset + (int4)(z_coord * src_stride_z), (int4)max_offset, (int4)z_coord < 0 || (int4)z_coord >= SRC_DIM_2);
-    VEC_TYPE(VEC_SIZE)
-    values3 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s0));
-    VEC_TYPE(VEC_SIZE)
-    values4 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s1));
-    VEC_TYPE(VEC_SIZE)
-    values5 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s2));
-
-    // z == 2
-    // Offset can be out-of-bound so we need to check if it is greater than max_offset
-    z_coord = z * (int)CONV_STRIDE_Y - (int)CONV_PAD_TOP + DILATION_Y * 2;
-    z_coord = min((uint)z_coord, (uint)SRC_DIM_2);
-    offset  = select(y_offset + (int4)(z_coord * src_stride_z), (int4)max_offset, (int4)z_coord < 0 || (int4)z_coord >= SRC_DIM_2);
-    VEC_TYPE(VEC_SIZE)
-    values6 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s0));
-    VEC_TYPE(VEC_SIZE)
-    values7 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s1));
-    VEC_TYPE(VEC_SIZE)
-    values8 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s2));
-
-    MULTIPLY_ADD_ACCUMULATE(values0, w0, acc, sum);
-    MULTIPLY_ADD_ACCUMULATE(values1, w1, acc, sum);
-    MULTIPLY_ADD_ACCUMULATE(values2, w2, acc, sum);
-
-    MULTIPLY_ADD_ACCUMULATE(values3, w3, acc, sum);
-    MULTIPLY_ADD_ACCUMULATE(values4, w4, acc, sum);
-    MULTIPLY_ADD_ACCUMULATE(values5, w5, acc, sum);
-
-    MULTIPLY_ADD_ACCUMULATE(values6, w6, acc, sum);
-    MULTIPLY_ADD_ACCUMULATE(values7, w7, acc, sum);
-    MULTIPLY_ADD_ACCUMULATE(values8, w8, acc, sum);
-
-#if defined(HAS_BIAS)
-    Vector  biases      = CONVERT_TO_VECTOR_STRUCT(biases);
-    VEC_INT bias_values = VLOAD(VEC_SIZE)(0, (__global int *)biases.ptr);
-    acc += bias_values;
-#endif // defined(HAS_BIAS)
-
-#if WEIGHTS_OFFSET != 0
-    acc += WEIGHTS_OFFSET * sum;
-#endif /* WEIGHTS_OFFSET != 0 */
-
-#if INPUT_OFFSET != 0
-    acc += INPUT_OFFSET * sum_we;
-#endif /* INPUT_OFFSET != 0 */
-
-#if K_OFFSET != 0
-    acc += (VEC_INT)K_OFFSET;
-#endif /* K_OFFSET != 0 */
-
-#if defined(REAL_MULTIPLIER)
-
-    acc = CONVERT(round(CONVERT(acc, VEC_FLOAT) * (VEC_FLOAT)REAL_MULTIPLIER), VEC_INT);
-
-#else // defined(REAL_MULTIPLIER)
-
-#if defined(PER_CHANNEL_QUANTIZATION)
-    Vector          output_multipliers = CONVERT_TO_VECTOR_STRUCT(output_multipliers);
-    Vector          output_shifts      = CONVERT_TO_VECTOR_STRUCT(output_shifts);
-    VEC_INT         output_multiplier  = VLOAD(VEC_SIZE)(0, (__global int *)output_multipliers.ptr);
-    VEC_INT         output_shift       = VLOAD(VEC_SIZE)(0, (__global int *)output_shifts.ptr);
-
-    VEC_INT res_shift_lt0              = ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(acc, output_multiplier, output_shift, VEC_SIZE);
-    VEC_INT res_shift_gt0              = asymm_mult_by_quant_multiplier_less_than_one(acc, output_multiplier, output_shift);
-    acc                                = select(res_shift_lt0, res_shift_gt0, output_shift >= 0);
-#else // defined(PER_CHANNEL_QUANTIZATION)
-#if OUTPUT_SHIFT < 0
-    acc     = ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(acc, OUTPUT_MULTIPLIER, OUTPUT_SHIFT, VEC_SIZE);
-#else  // OUTPUT_SHIFT < 0
-    acc     = asymm_mult_by_quant_multiplier_less_than_one(acc, OUTPUT_MULTIPLIER, OUTPUT_SHIFT);
-#endif // OUTPUT_SHIFT < 0
-#endif // defined(PER_CHANNEL_QUANTIZATION)
-
-#endif // defined(REAL_MULTIPLIER)
-
-    acc += (VEC_INT)OUTPUT_OFFSET;
-
-    VEC_TYPE(VEC_SIZE)
-    res = CONVERT_SAT(acc, VEC_TYPE(VEC_SIZE));
-
-#if defined(DST_DEPTH)
-    __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + x * dst_step_x + y * dst_step_y + z * dst_step_z + b * dst_stride_w;
-#else  /* defined(DST_DEPTH) */
-    __global uchar *dst_addr           = dst_ptr + dst_offset_first_element_in_bytes + x * dst_step_x + y * dst_step_y + z * dst_step_z;
-#endif /* defined(DST_DEPTH) */
-
-    VSTORE(VEC_SIZE)
-    (ACTIVATION_FUNC(res), 0, (__global DATA_TYPE *)(dst_addr));
-}
-#endif // defined(CONV_STRIDE_X) && defined(CONV_STRIDE_Y)
-
-#if defined(NUM_ROWS_PROCESSED) && defined(NUM_PLANES_PROCESSED) && VEC_SIZE == 4
-/** This function computes the depthwise convolution quantized for NHWC data layout when the stride along the width and height is 1.
- *
- * @note This kernel assumes VEC_SIZE is 4.
- * @note The weights tensor is expected to be reshaped using @ref CLDepthwiseConvolutionLayerReshapeWeightsKernel.
- * @note The number of elements read per thread must be passed at compile time using -DVEC_SIZE (e.g. -DVEC_SIZE=2)
- * @note Dimension two of the input tensor (height for NHWC data layout) must be passed at compile time using -DSRC_DIM2 (e.g. -DSRC_DIM_2=112)
- * @note The number of rows processed per thread must be passed at compile time using -DNUM_ROWS_PROCESSED (i.e. -DNUM_ROWS_PROCESSED=2)
- * @note The number of planes processed per thread must be passed at compile time using -DNUM_PLANES_PROCESSED (i.e. -DNUM_PLANES_PROCESSED=2)
- * @note The convolution pad top must be passed at compile time using -DCONV_PAD_TOP (e.g. -DCONV_PAD_TOP=1)
- * @note The convolution pad top must be passed at compile time using -DCONV_PAD_LEFT (e.g. -DCONV_PAD_LEFT=1).
- *
- * @param[in] src_ptr                                          Pointer to the source tensor. Supported data types: QASYMM8/QASYMM8_SIGNED
- * @param[in] src_stride_x                                     Stride of the source tensor in X dimension (in bytes)
- * @param[in] src_step_x                                       src_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in] src_stride_y                                     Stride of the source tensor in Y dimension (in bytes)
- * @param[in] src_step_y                                       src_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in] src_stride_z                                     Stride of the source tensor in Z dimension (in bytes)
- * @param[in] src_step_z                                       src_stride_y * number of elements along Z processed per workitem(in bytes)
- * @param[in] src_stride_w                                     Stride of the source tensor in W dimension (in bytes)
- * @param[in] src_step_w                                       src_stride_w * number of elements along W processed per workitem(in bytes)
- * @param[in] src_offset_first_element_in_bytes                The offset of the first element in the source tensor
- * @param[in] dst_ptr                                          Pointer to the destination tensor. Supported data types: same as @p src_ptr
- * @param[in] dst_stride_x                                     Stride of the destination tensor in X dimension (in bytes)
- * @param[in] dst_step_x                                       dst_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in] dst_stride_y                                     Stride of the destination tensor in Y dimension (in bytes)
- * @param[in] dst_step_y                                       dst_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in] dst_stride_z                                     Stride of the destination tensor in Z dimension (in bytes)
- * @param[in] dst_step_z                                       dst_stride_z * number of elements along Y processed per workitem(in bytes)
- * @param[in] dst_stride_w                                     Stride of the destination tensor in W dimension (in bytes)
- * @param[in] dst_step_w                                       dst_stride_w * number of elements along W processed per workitem(in bytes)
- * @param[in] dst_offset_first_element_in_bytes                The offset of the first element in the destination tensor
- * @param[in] weights_ptr                                      Pointer to the weights tensor. Supported data types: QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL
- * @param[in] weights_stride_x                                 Stride of the weights tensor in X dimension (in bytes)
- * @param[in] weights_step_x                                   weights_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in] weights_stride_y                                 Stride of the weights tensor in Y dimension (in bytes)
- * @param[in] weights_step_y                                   weights_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in] weights_offset_first_element_in_bytes            The offset of the first element in the weights tensor
- * @param[in] output_multipliers_ptr                           Pointer to the output multipliers vector. Supported data types: S32
- * @param[in] output_multipliers_stride_x                      Stride of the output multipliers vector in X dimension (in bytes)
- * @param[in] output_multipliers_step_x                        output_multipliers_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in] output_multipliers_offset_first_element_in_bytes The offset of the first element in the output multipliers vector
- * @param[in] output_shifts_ptr                                Pointer to the output shifts vector. Supported data types: S32
- * @param[in] output_shifts_stride_x                           Stride of the output shifts vector in X dimension (in bytes)
- * @param[in] output_shifts_step_x                             output_shifts_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in] output_shifts_offset_first_element_in_bytes      The offset of the first element in the output shifts vector
- * @param[in] biases_ptr                                       (Optional) Pointer to the biases vector. Supported data types: S32
- * @param[in] biases_stride_x                                  (Optional) Stride of the biases vector in X dimension (in bytes)
- * @param[in] biases_step_x                                    (Optional) biases_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in] biases_offset_first_element_in_bytes             (Optional) The offset of the first element in the biases vector
- * @param[in] max_offset                                       Max offset for the input tensor
- */
-
-__kernel void dwc_3x3_reshaped_quantized8_stride1_nhwc(
-    TENSOR4D_DECLARATION(src),
-    TENSOR4D_DECLARATION(dst),
-    IMAGE_DECLARATION(weights),
-    VECTOR_DECLARATION(output_multipliers),
-    VECTOR_DECLARATION(output_shifts),
-#if defined(HAS_BIAS)
-    VECTOR_DECLARATION(biases),
-#endif /* defined(HAS_BIAS) */
-    int max_offset)
-{
-    int x = get_global_id(0);
-    int y = get_global_id(1);
-#if defined(DST_DEPTH)
-    int z = get_global_id(2) % (int)DST_DEPTH; // spatial coordinate y
-    int b = get_global_id(2) / (int)DST_DEPTH; // batch
-#else                                          // defined(DST_DEPTH)
-    int             z                  = get_global_id(2); // spatial coordinate y
-#endif                                         // defined(DST_DEPTH)
-
-    __global uchar *weights_addr = weights_ptr + weights_offset_first_element_in_bytes + x * weights_stride_y;
-
-#if defined(DST_DEPTH)
-    __global uchar *src_addr = src_ptr + src_offset_first_element_in_bytes + x * VEC_SIZE + b * src_stride_w;
-#else  /* defined(DST_DEPTH) */
-    __global uchar *src_addr           = src_ptr + src_offset_first_element_in_bytes + x * VEC_SIZE;
-#endif /* defined(DST_DEPTH) */
-
-    int  z_coord = 0;
-    int4 offset  = 0;
-    int4 y_coord = ((int4)(y * NUM_ROWS_PROCESSED) + (int4)(0, 1, 2, 3)) - (int)CONV_PAD_LEFT;
-
-    // Only for y = 0 we can have a negative coordinate. If so, we convert it to SRC_DIM_1
-    y_coord.s0 = min((uint)y_coord.s0, (uint)SRC_DIM_1);
-    y_coord.s1 = min((uint)y_coord.s1, (uint)SRC_DIM_1);
-    y_coord.s2 = min((uint)y_coord.s2, (uint)SRC_DIM_1);
-    y_coord.s3 = min((uint)y_coord.s3, (uint)SRC_DIM_1);
-
-    int4 y_offset = convert_int4(y_coord * (int)src_stride_y);
-
-    // We compute 4x2x2 [C,W,H] elements
-    VEC_INT acc0 = 0, sum0 = 0;
-    VEC_INT acc1 = 0, sum1 = 0;
-    VEC_INT acc2 = 0, sum2 = 0;
-    VEC_INT acc3 = 0, sum3 = 0;
-
-    // Load weights
-    VEC_DATA_TYPE(WEIGHTS_TYPE, 16)
-    w0_tmp = VLOAD(16)(0, (__global WEIGHTS_TYPE *)(weights_addr));
-    VEC_DATA_TYPE(WEIGHTS_TYPE, 16)
-    w1_tmp = VLOAD(16)(0, (__global WEIGHTS_TYPE *)(weights_addr + 16));
-    VEC_DATA_TYPE(WEIGHTS_TYPE, 4)
-    w8 = VLOAD(4)(0, (__global WEIGHTS_TYPE *)(weights_addr + 2 * 16));
-
-    VEC_DATA_TYPE(WEIGHTS_TYPE, 4)
-    w0 = w0_tmp.s0123;
-    VEC_DATA_TYPE(WEIGHTS_TYPE, 4)
-    w1 = w0_tmp.s4567;
-    VEC_DATA_TYPE(WEIGHTS_TYPE, 4)
-    w2 = w0_tmp.s89AB;
-    VEC_DATA_TYPE(WEIGHTS_TYPE, 4)
-    w3 = w0_tmp.sCDEF;
-
-    VEC_DATA_TYPE(WEIGHTS_TYPE, 4)
-    w4 = w1_tmp.s0123;
-    VEC_DATA_TYPE(WEIGHTS_TYPE, 4)
-    w5 = w1_tmp.s4567;
-    VEC_DATA_TYPE(WEIGHTS_TYPE, 4)
-    w6 = w1_tmp.s89AB;
-    VEC_DATA_TYPE(WEIGHTS_TYPE, 4)
-    w7 = w1_tmp.sCDEF;
-
-#if INPUT_OFFSET != 0
-    VEC_INT sum_we = CONVERT(w0, VEC_INT) + CONVERT(w1, VEC_INT) + CONVERT(w2, VEC_INT)
-                     + CONVERT(w3, VEC_INT) + CONVERT(w4, VEC_INT) + CONVERT(w5, VEC_INT)
-                     + CONVERT(w6, VEC_INT) + CONVERT(w7, VEC_INT) + CONVERT(w8, VEC_INT);
-#endif /* INPUT_OFFSET != 0 */
-
-    // Load input values
-    // z == 0
-    // Clamp z_coord as for z = 0, it can be negative
-    // z_coord is casted to unsigned int in order to use just a min() operation
-    // A "-1" 32 bit signed variable converted to unsigned gives 4294967295
-    z_coord = z * (int)NUM_PLANES_PROCESSED - (int)CONV_PAD_TOP;
-    z_coord = min((uint)z_coord, (uint)SRC_DIM_2);
-    offset  = select(y_offset + (int4)(z_coord * src_stride_z), (int4)max_offset, (int4)z_coord < 0 || (int4)z_coord >= SRC_DIM_2);
-    VEC_TYPE(VEC_SIZE)
-    values0 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s0));
-    VEC_TYPE(VEC_SIZE)
-    values1 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s1));
-    VEC_TYPE(VEC_SIZE)
-    values2 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s2));
-    VEC_TYPE(VEC_SIZE)
-    values3 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s3));
-
-    // z == 1
-    z_coord = z * (int)NUM_PLANES_PROCESSED - (int)CONV_PAD_TOP + 1;
-    z_coord = min((uint)z_coord, (uint)SRC_DIM_2);
-    offset  = select(y_offset + (int4)(z_coord * src_stride_z), (int4)max_offset, (int4)z_coord < 0 || (int4)z_coord >= SRC_DIM_2);
-    VEC_TYPE(VEC_SIZE)
-    values4 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s0));
-    VEC_TYPE(VEC_SIZE)
-    values5 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s1));
-    VEC_TYPE(VEC_SIZE)
-    values6 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s2));
-    VEC_TYPE(VEC_SIZE)
-    values7 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s3));
-
-    // z == 2
-    z_coord = z * (int)NUM_PLANES_PROCESSED - (int)CONV_PAD_TOP + 2;
-    z_coord = min((uint)z_coord, (uint)SRC_DIM_2);
-    offset  = select(y_offset + (int4)(z_coord * src_stride_z), (int4)max_offset, (int4)z_coord < 0 || (int4)z_coord >= SRC_DIM_2);
-    VEC_TYPE(VEC_SIZE)
-    values8 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s0));
-    VEC_TYPE(VEC_SIZE)
-    values9 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s1));
-    VEC_TYPE(VEC_SIZE)
-    values10 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s2));
-    VEC_TYPE(VEC_SIZE)
-    values11 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s3));
-
-    // z == 3
-    z_coord = z * (int)NUM_PLANES_PROCESSED - (int)CONV_PAD_TOP + 3;
-    z_coord = min((uint)z_coord, (uint)SRC_DIM_2);
-    offset  = select(y_offset + (int4)(z_coord * src_stride_z), (int4)max_offset, (int4)z_coord < 0 || (int4)z_coord >= SRC_DIM_2);
-    VEC_TYPE(VEC_SIZE)
-    values12 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s0));
-    VEC_TYPE(VEC_SIZE)
-    values13 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s1));
-    VEC_TYPE(VEC_SIZE)
-    values14 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s2));
-    VEC_TYPE(VEC_SIZE)
-    values15 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s3));
-
-    MULTIPLY_ADD_ACCUMULATE(values0, w0, acc0, sum0);
-    MULTIPLY_ADD_ACCUMULATE(values1, w1, acc0, sum0);
-    MULTIPLY_ADD_ACCUMULATE(values2, w2, acc0, sum0);
-    MULTIPLY_ADD_ACCUMULATE(values1, w0, acc1, sum1);
-    MULTIPLY_ADD_ACCUMULATE(values2, w1, acc1, sum1);
-    MULTIPLY_ADD_ACCUMULATE(values3, w2, acc1, sum1);
-
-    MULTIPLY_ADD_ACCUMULATE(values4, w3, acc0, sum0);
-    MULTIPLY_ADD_ACCUMULATE(values5, w4, acc0, sum0);
-    MULTIPLY_ADD_ACCUMULATE(values6, w5, acc0, sum0);
-    MULTIPLY_ADD_ACCUMULATE(values5, w3, acc1, sum1);
-    MULTIPLY_ADD_ACCUMULATE(values6, w4, acc1, sum1);
-    MULTIPLY_ADD_ACCUMULATE(values7, w5, acc1, sum1);
-
-    MULTIPLY_ADD_ACCUMULATE(values8, w6, acc0, sum0);
-    MULTIPLY_ADD_ACCUMULATE(values9, w7, acc0, sum0);
-    MULTIPLY_ADD_ACCUMULATE(values10, w8, acc0, sum0);
-    MULTIPLY_ADD_ACCUMULATE(values9, w6, acc1, sum1);
-    MULTIPLY_ADD_ACCUMULATE(values10, w7, acc1, sum1);
-    MULTIPLY_ADD_ACCUMULATE(values11, w8, acc1, sum1);
-
-    MULTIPLY_ADD_ACCUMULATE(values4, w0, acc2, sum2);
-    MULTIPLY_ADD_ACCUMULATE(values5, w1, acc2, sum2);
-    MULTIPLY_ADD_ACCUMULATE(values6, w2, acc2, sum2);
-    MULTIPLY_ADD_ACCUMULATE(values5, w0, acc3, sum3);
-    MULTIPLY_ADD_ACCUMULATE(values6, w1, acc3, sum3);
-    MULTIPLY_ADD_ACCUMULATE(values7, w2, acc3, sum3);
-
-    MULTIPLY_ADD_ACCUMULATE(values8, w3, acc2, sum2);
-    MULTIPLY_ADD_ACCUMULATE(values9, w4, acc2, sum2);
-    MULTIPLY_ADD_ACCUMULATE(values10, w5, acc2, sum2);
-    MULTIPLY_ADD_ACCUMULATE(values9, w3, acc3, sum3);
-    MULTIPLY_ADD_ACCUMULATE(values10, w4, acc3, sum3);
-    MULTIPLY_ADD_ACCUMULATE(values11, w5, acc3, sum3);
-
-    MULTIPLY_ADD_ACCUMULATE(values12, w6, acc2, sum2);
-    MULTIPLY_ADD_ACCUMULATE(values13, w7, acc2, sum2);
-    MULTIPLY_ADD_ACCUMULATE(values14, w8, acc2, sum2);
-    MULTIPLY_ADD_ACCUMULATE(values13, w6, acc3, sum3);
-    MULTIPLY_ADD_ACCUMULATE(values14, w7, acc3, sum3);
-    MULTIPLY_ADD_ACCUMULATE(values15, w8, acc3, sum3);
-
-#if defined(HAS_BIAS)
-    Vector biases = CONVERT_TO_VECTOR_STRUCT(biases);
-
-    VEC_INT bias_values = VLOAD(VEC_SIZE)(0, (__global int *)biases.ptr);
-
-    acc0 += bias_values;
-    acc1 += bias_values;
-    acc2 += bias_values;
-    acc3 += bias_values;
-#endif /* defined(HAS_BIAS) */
-
-#if WEIGHTS_OFFSET != 0
-    acc0 += WEIGHTS_OFFSET * sum0;
-    acc1 += WEIGHTS_OFFSET * sum1;
-    acc2 += WEIGHTS_OFFSET * sum2;
-    acc3 += WEIGHTS_OFFSET * sum3;
-#endif /* WEIGHTS_OFFSET != 0 */
-
-#if INPUT_OFFSET != 0
-    VEC_INT offs = INPUT_OFFSET * sum_we;
-
-    acc0 += offs;
-    acc1 += offs;
-    acc2 += offs;
-    acc3 += offs;
-#endif /* INPUT_OFFSET != 0 */
-
-#if K_OFFSET != 0
-    acc0 += (VEC_INT)K_OFFSET;
-    acc1 += (VEC_INT)K_OFFSET;
-    acc2 += (VEC_INT)K_OFFSET;
-    acc3 += (VEC_INT)K_OFFSET;
-#endif /* K_OFFSET != 0 */
-
-#if defined(REAL_MULTIPLIER)
-
-    acc0 = CONVERT(round(CONVERT(acc0, VEC_FLOAT) * (VEC_FLOAT)REAL_MULTIPLIER), VEC_INT);
-    acc1 = CONVERT(round(CONVERT(acc1, VEC_FLOAT) * (VEC_FLOAT)REAL_MULTIPLIER), VEC_INT);
-    acc2 = CONVERT(round(CONVERT(acc2, VEC_FLOAT) * (VEC_FLOAT)REAL_MULTIPLIER), VEC_INT);
-    acc3 = CONVERT(round(CONVERT(acc3, VEC_FLOAT) * (VEC_FLOAT)REAL_MULTIPLIER), VEC_INT);
-
-#else // defined(REAL_MULTIPLIER)
-
-#if defined(PER_CHANNEL_QUANTIZATION)
-    Vector          output_multipliers = CONVERT_TO_VECTOR_STRUCT(output_multipliers);
-    Vector          output_shifts      = CONVERT_TO_VECTOR_STRUCT(output_shifts);
-    VEC_INT         output_multiplier  = VLOAD(VEC_SIZE)(0, (__global int *)output_multipliers.ptr);
-    VEC_INT         output_shift       = VLOAD(VEC_SIZE)(0, (__global int *)output_shifts.ptr);
-
-    VEC_INT res0_shift_lt0   = ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(acc0, output_multiplier, output_shift, VEC_SIZE);
-    VEC_INT res1_shift_lt0   = ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(acc1, output_multiplier, output_shift, VEC_SIZE);
-    VEC_INT res2_shift_lt0   = ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(acc2, output_multiplier, output_shift, VEC_SIZE);
-    VEC_INT res3_shift_lt0   = ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(acc3, output_multiplier, output_shift, VEC_SIZE);
-    VEC_INT res0_shift_gt0   = asymm_mult_by_quant_multiplier_less_than_one(acc0, output_multiplier, output_shift);
-    VEC_INT res1_shift_gt0   = asymm_mult_by_quant_multiplier_less_than_one(acc1, output_multiplier, output_shift);
-    VEC_INT res2_shift_gt0   = asymm_mult_by_quant_multiplier_less_than_one(acc2, output_multiplier, output_shift);
-    VEC_INT res3_shift_gt0   = asymm_mult_by_quant_multiplier_less_than_one(acc3, output_multiplier, output_shift);
-    acc0                     = select(res0_shift_lt0, res0_shift_gt0, output_shift >= 0);
-    acc1                     = select(res1_shift_lt0, res1_shift_gt0, output_shift >= 0);
-    acc2                     = select(res2_shift_lt0, res2_shift_gt0, output_shift >= 0);
-    acc3                     = select(res3_shift_lt0, res3_shift_gt0, output_shift >= 0);
-#else // defined(PER_CHANNEL_QUANTIZATION)
-#if OUTPUT_SHIFT < 0
-    acc0    = ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(acc0, OUTPUT_MULTIPLIER, OUTPUT_SHIFT, VEC_SIZE);
-    acc1    = ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(acc1, OUTPUT_MULTIPLIER, OUTPUT_SHIFT, VEC_SIZE);
-    acc2    = ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(acc2, OUTPUT_MULTIPLIER, OUTPUT_SHIFT, VEC_SIZE);
-    acc3    = ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(acc3, OUTPUT_MULTIPLIER, OUTPUT_SHIFT, VEC_SIZE);
-#else  // OUTPUT_SHIFT < 0
-    acc0    = asymm_mult_by_quant_multiplier_less_than_one(acc0, OUTPUT_MULTIPLIER, OUTPUT_SHIFT);
-    acc1    = asymm_mult_by_quant_multiplier_less_than_one(acc1, OUTPUT_MULTIPLIER, OUTPUT_SHIFT);
-    acc2    = asymm_mult_by_quant_multiplier_less_than_one(acc2, OUTPUT_MULTIPLIER, OUTPUT_SHIFT);
-    acc3    = asymm_mult_by_quant_multiplier_less_than_one(acc3, OUTPUT_MULTIPLIER, OUTPUT_SHIFT);
-#endif // OUTPUT_SHIFT < 0
-#endif // defined(PER_CHANNEL_QUANTIZATION)
-
-#endif // defined(REAL_MULTIPLIER)
-
-    acc0 += (VEC_INT)OUTPUT_OFFSET;
-    acc1 += (VEC_INT)OUTPUT_OFFSET;
-    acc2 += (VEC_INT)OUTPUT_OFFSET;
-    acc3 += (VEC_INT)OUTPUT_OFFSET;
-
-    VEC_TYPE(VEC_SIZE)
-    res0 = CONVERT_SAT(acc0, VEC_TYPE(VEC_SIZE));
-    VEC_TYPE(VEC_SIZE)
-    res1 = CONVERT_SAT(acc1, VEC_TYPE(VEC_SIZE));
-    VEC_TYPE(VEC_SIZE)
-    res2 = CONVERT_SAT(acc2, VEC_TYPE(VEC_SIZE));
-    VEC_TYPE(VEC_SIZE)
-    res3 = CONVERT_SAT(acc3, VEC_TYPE(VEC_SIZE));
-
-#if defined(DST_DEPTH)
-    __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + x * dst_step_x + y * dst_step_y + (z * NUM_PLANES_PROCESSED) * dst_step_z + b * dst_stride_w;
-#else  /* defined(DST_DEPTH) */
-    __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + x * dst_step_x + y * dst_step_y + (z * NUM_PLANES_PROCESSED) * dst_step_z;
-#endif /* defined(DST_DEPTH) */
-
-    VSTORE(VEC_SIZE)
-    (ACTIVATION_FUNC(res0), 0, dst_addr + 0 * dst_stride_y);
-    VSTORE(VEC_SIZE)
-    (ACTIVATION_FUNC(res1), 0, dst_addr + 1 * dst_stride_y);
-
-#if((DST_DIM_2 % NUM_PLANES_PROCESSED) != 0)
-    if((z * NUM_PLANES_PROCESSED + 1) < DST_DIM_2)
-#endif // ((DST_DIM_2 % NUM_PLANES_PROCESSED) != 0)
-    {
-        VSTORE(VEC_SIZE)
-        (ACTIVATION_FUNC(res2), 0, (__global DATA_TYPE *)(dst_addr + 0 * dst_stride_y + 1 * dst_stride_z));
-        VSTORE(VEC_SIZE)
-        (ACTIVATION_FUNC(res3), 0, (__global DATA_TYPE *)(dst_addr + 1 * dst_stride_y + 1 * dst_stride_z));
-    }
-}
-
-#if defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8) && VEC_SIZE == 4
-/** This function computes the depthwise convolution quantized for NHWC data layout when the stride along the width and height is 1 using dot product.
- *
- * @note Per-channel quantization is not supported by this kernel.
- * @note This kernel assumes VEC_SIZE is 4.
- * @note The weights tensor is expected to be reshaped using @ref CLDepthwiseConvolutionLayerReshapeWeightsKernel.
- * @note The number of elements read per thread must be passed at compile time using -DVEC_SIZE (e.g. -DVEC_SIZE=2)
- * @note Dimension two of the input tensor (height for NHWC data layout) must be passed at compile time using -DSRC_DIM2 (e.g. -DSRC_DIM_2=112)
- * @note The number of rows processed per thread must be passed at compile time using -DNUM_ROWS_PROCESSED (i.e. -DNUM_ROWS_PROCESSED=2)
- * @note The number of planes processed per thread must be passed at compile time using -DNUM_PLANES_PROCESSED (i.e. -DNUM_PLANES_PROCESSED=2)
- * @note The convolution pad top must be passed at compile time using -DCONV_PAD_TOP (e.g. -DCONV_PAD_TOP=1)
- * @note The convolution pad top must be passed at compile time using -DCONV_PAD_LEFT (e.g. -DCONV_PAD_LEFT=1).
- * @note If REAL_MULTIPLIER is passed at compile time (i.e. -DREAL_MULTIPLIER=1.355f), the final quantization is performed using a floating point multiplication.
- *       If not, the quantization will be performed using a fixed point multiplication
- *
- * @param[in] src_ptr                                          Pointer to the source tensor. Supported data types: QASYMM8/QASYMM8_SIGNED
- * @param[in] src_stride_x                                     Stride of the source tensor in X dimension (in bytes)
- * @param[in] src_step_x                                       src_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in] src_stride_y                                     Stride of the source tensor in Y dimension (in bytes)
- * @param[in] src_step_y                                       src_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in] src_stride_z                                     Stride of the source tensor in Z dimension (in bytes)
- * @param[in] src_step_z                                       src_stride_y * number of elements along Z processed per workitem(in bytes)
- * @param[in] src_stride_w                                     Stride of the source tensor in W dimension (in bytes)
- * @param[in] src_step_w                                       src_stride_w * number of elements along W processed per workitem(in bytes)
- * @param[in] src_offset_first_element_in_bytes                The offset of the first element in the source tensor
- * @param[in] dst_ptr                                          Pointer to the destination tensor. Supported data types: same as @p src_ptr
- * @param[in] dst_stride_x                                     Stride of the destination tensor in X dimension (in bytes)
- * @param[in] dst_step_x                                       dst_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in] dst_stride_y                                     Stride of the destination tensor in Y dimension (in bytes)
- * @param[in] dst_step_y                                       dst_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in] dst_stride_z                                     Stride of the destination tensor in Z dimension (in bytes)
- * @param[in] dst_step_z                                       dst_stride_z * number of elements along Y processed per workitem(in bytes)
- * @param[in] dst_stride_w                                     Stride of the destination tensor in W dimension (in bytes)
- * @param[in] dst_step_w                                       dst_stride_w * number of elements along W processed per workitem(in bytes)
- * @param[in] dst_offset_first_element_in_bytes                The offset of the first element in the destination tensor
- * @param[in] weights_ptr                                      Pointer to the weights tensor. Supported data types: same as @p src_ptr
- * @param[in] weights_stride_x                                 Stride of the weights tensor in X dimension (in bytes)
- * @param[in] weights_step_x                                   weights_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in] weights_stride_y                                 Stride of the weights tensor in Y dimension (in bytes)
- * @param[in] weights_step_y                                   weights_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in] weights_offset_first_element_in_bytes            The offset of the first element in the weights tensor
- * @param[in] output_multipliers_ptr                           Pointer to the output multipliers vector. Supported data types: S32
- * @param[in] output_multipliers_stride_x                      Stride of the output multipliers vector in X dimension (in bytes)
- * @param[in] output_multipliers_step_x                        output_multipliers_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in] output_multipliers_offset_first_element_in_bytes The offset of the first element in the output multipliers vector
- * @param[in] output_shifts_ptr                                Pointer to the output shifts vector. Supported data types: S32
- * @param[in] output_shifts_stride_x                           Stride of the output shifts vector in X dimension (in bytes)
- * @param[in] output_shifts_step_x                             output_shifts_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in] output_shifts_offset_first_element_in_bytes      The offset of the first element in the output shifts vector
- * @param[in] biases_ptr                                       (Optional) Pointer to the biases vector. Supported data types: S32
- * @param[in] biases_stride_x                                  (Optional) Stride of the biases vector in X dimension (in bytes)
- * @param[in] biases_step_x                                    (Optional) biases_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in] biases_offset_first_element_in_bytes             (Optional) The offset of the first element in the biases vector
- * @param[in] max_offset                                       The maximum allowed offset for the input tensor
- */
-__kernel void dwc_3x3_reshaped_quantized8_dot8_stride1_nhwc(
-    TENSOR4D_DECLARATION(src),
-    TENSOR4D_DECLARATION(dst),
-    IMAGE_DECLARATION(weights),
-    VECTOR_DECLARATION(output_multipliers),
-    VECTOR_DECLARATION(output_shifts),
-#if defined(HAS_BIAS)
-    VECTOR_DECLARATION(biases),
-#endif // defined(HAS_BIAS)
-    int max_offset)
-{
-    int x = get_global_id(0);
-    int y = get_global_id(1);
-#if defined(DST_DEPTH)
-    int z = get_global_id(2) % (int)DST_DEPTH; // spatial coordinate y
-    int b = get_global_id(2) / (int)DST_DEPTH; // batch
-#else                                          // defined(DST_DEPTH)
-    int      z               = get_global_id(2); // spatial coordinate y
-#endif                                         // defined(DST_DEPTH)
-
-    __global uchar *weights_addr = weights_ptr + weights_offset_first_element_in_bytes + x * weights_stride_y;
-
-#if defined(DST_DEPTH)
-    __global uchar *src_addr = src_ptr + src_offset_first_element_in_bytes + x * VEC_SIZE + b * src_stride_w;
-#else  /* defined(DST_DEPTH) */
-    __global uchar *src_addr = src_ptr + src_offset_first_element_in_bytes + x * VEC_SIZE;
-#endif /* defined(DST_DEPTH) */
-
-    int  z_coord = 0;
-    int4 offset  = 0;
-    int4 y_coord = ((int4)(y * NUM_ROWS_PROCESSED) + (int4)(0, 1, 2, 3)) - (int)CONV_PAD_LEFT;
-
-    // Only for y = 0 we can have a negative coordinate. If so, we convert it to SRC_DIM_1
-    y_coord.s0 = min((uint)y_coord.s0, (uint)SRC_DIM_1);
-    y_coord.s1 = min((uint)y_coord.s1, (uint)SRC_DIM_1);
-    y_coord.s2 = min((uint)y_coord.s2, (uint)SRC_DIM_1);
-    y_coord.s3 = min((uint)y_coord.s3, (uint)SRC_DIM_1);
-
-    int4 y_offset = convert_int4(y_coord * (int)src_stride_y);
-
-    // We compute 4x2x1 [C,W,H] elements
-    VEC_INT acc0 = 0;
-    VEC_INT acc1 = 0;
-    VEC_INT sum0 = 0;
-    VEC_INT sum1 = 0;
-
-    // Load weights
-    VEC_TYPE(16)
-    w0 = VLOAD(16)(0, (__global WEIGHTS_TYPE *)(weights_addr));
-    VEC_TYPE(16)
-    w1 = VLOAD(16)(0, (__global WEIGHTS_TYPE *)(weights_addr + 16));
-    VEC_TYPE(4)
-    w2 = VLOAD(4)(0, (__global WEIGHTS_TYPE *)(weights_addr + 32));
-
-#if INPUT_OFFSET != 0
-    // Initilize the final result with the weights reduction multiplied by INPUT_OFFSET
-    DOT_PRODUCT_REDUCTION_WEIGHTS(acc0.s0, w0.s01234567, w0.s8);
-    DOT_PRODUCT_REDUCTION_WEIGHTS(acc0.s1, (VEC_TYPE(8))((w0.s9ABC), (w0.sDEF), w1.s0), w1.s1);
-    DOT_PRODUCT_REDUCTION_WEIGHTS(acc0.s2, w1.s23456789, w1.sA);
-    DOT_PRODUCT_REDUCTION_WEIGHTS(acc0.s3, (VEC_TYPE(8))((w1.sBCD), (w1.sEF), (w2.s012)), w2.s3);
-
-    // Multiply the weights reduction with INPUT_OFFSET
-    acc0 = INPUT_OFFSET * acc0;
-
-    acc1 = acc0;
-#endif // INPUT_OFFSET != 0
-
-    // Load input values
-    // z == 0
-    // Clamp z_coord as for z = 0, it can be negative
-    // z_coord is casted to unsigned int in order to use just a min() operation
-    // A "-1" 32 bit signed variable converted to unsigned gives 4294967295
-    z_coord = z - (int)CONV_PAD_TOP;
-    z_coord = min((uint)z_coord, (uint)SRC_DIM_2);
-    offset  = y_offset + (int4)(z_coord * src_stride_z);
-    offset  = min(offset, (int4)max_offset);
-
-    VEC_TYPE(VEC_SIZE)
-    values0 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s0));
-    VEC_TYPE(VEC_SIZE)
-    values1 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s1));
-    VEC_TYPE(VEC_SIZE)
-    values2 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s2));
-    VEC_TYPE(VEC_SIZE)
-    values3 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s3));
-
-    // z == 1
-    // z_coord can be only negative for z = 0 so we do not need to clamp it
-    // Moreover z_coord cannot be out-of-bound for z = 1 so we do not need to clamp the offset
-    z_coord = z - (int)CONV_PAD_TOP + 1;
-    offset  = y_offset + (int4)(z_coord * src_stride_z);
-    VEC_TYPE(VEC_SIZE)
-    values4 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s0));
-    VEC_TYPE(VEC_SIZE)
-    values5 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s1));
-    VEC_TYPE(VEC_SIZE)
-    values6 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s2));
-    VEC_TYPE(VEC_SIZE)
-    values7 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s3));
-
-    // z == 2
-    // After z = 1 we can simply add src_stride_z to offset without updating z_coord
-    // However offset can be out-of-bound so we need to check if it is greater than max_offset
-    offset += (int4)src_stride_z;
-    offset = min(offset, (int4)max_offset);
-    VEC_TYPE(VEC_SIZE)
-    values8 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s0));
-    VEC_TYPE(VEC_SIZE)
-    values9 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s1));
-    VEC_TYPE(VEC_SIZE)
-    values10 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s2));
-    VEC_TYPE(VEC_SIZE)
-    values11 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s3));
-
-    DOT_PRODUCT_REDUCTION(sum0.s0, values0.s0, values1.s0, values2.s0, values4.s0, values5.s0, values6.s0, values8.s0, values9.s0, values10.s0);
-    DOT_PRODUCT_REDUCTION(sum1.s0, values1.s0, values2.s0, values3.s0, values5.s0, values6.s0, values7.s0, values9.s0, values10.s0, values11.s0);
-    DOT_PRODUCT(acc0.s0, values0.s0, values1.s0, values2.s0, values4.s0, values5.s0, values6.s0, values8.s0, values9.s0, values10.s0, w0.s01234567, w0.s8);
-    DOT_PRODUCT(acc1.s0, values1.s0, values2.s0, values3.s0, values5.s0, values6.s0, values7.s0, values9.s0, values10.s0, values11.s0, w0.s01234567, w0.s8);
-
-    DOT_PRODUCT_REDUCTION(sum0.s1, values0.s1, values1.s1, values2.s1, values4.s1, values5.s1, values6.s1, values8.s1, values9.s1, values10.s1);
-    DOT_PRODUCT_REDUCTION(sum1.s1, values1.s1, values2.s1, values3.s1, values5.s1, values6.s1, values7.s1, values9.s1, values10.s1, values11.s1);
-    DOT_PRODUCT(acc0.s1, values0.s1, values1.s1, values2.s1, values4.s1, values5.s1, values6.s1, values8.s1, values9.s1, values10.s1, (VEC_TYPE(8))((w0.s9ABC), (w0.sDEF), w1.s0), w1.s1);
-    DOT_PRODUCT(acc1.s1, values1.s1, values2.s1, values3.s1, values5.s1, values6.s1, values7.s1, values9.s1, values10.s1, values11.s1, (VEC_TYPE(8))((w0.s9ABC), (w0.sDEF), w1.s0), w1.s1);
-
-    DOT_PRODUCT_REDUCTION(sum0.s2, values0.s2, values1.s2, values2.s2, values4.s2, values5.s2, values6.s2, values8.s2, values9.s2, values10.s2);
-    DOT_PRODUCT_REDUCTION(sum1.s2, values1.s2, values2.s2, values3.s2, values5.s2, values6.s2, values7.s2, values9.s2, values10.s2, values11.s2);
-    DOT_PRODUCT(acc0.s2, values0.s2, values1.s2, values2.s2, values4.s2, values5.s2, values6.s2, values8.s2, values9.s2, values10.s2, w1.s23456789, w1.sA);
-    DOT_PRODUCT(acc1.s2, values1.s2, values2.s2, values3.s2, values5.s2, values6.s2, values7.s2, values9.s2, values10.s2, values11.s2, w1.s23456789, w1.sA);
-
-    DOT_PRODUCT_REDUCTION(sum0.s3, values0.s3, values1.s3, values2.s3, values4.s3, values5.s3, values6.s3, values8.s3, values9.s3, values10.s3);
-    DOT_PRODUCT_REDUCTION(sum1.s3, values1.s3, values2.s3, values3.s3, values5.s3, values6.s3, values7.s3, values9.s3, values10.s3, values11.s3);
-    DOT_PRODUCT(acc0.s3, values0.s3, values1.s3, values2.s3, values4.s3, values5.s3, values6.s3, values8.s3, values9.s3, values10.s3, (VEC_TYPE(8))((w1.sBCD), (w1.sEF), (w2.s012)), w2.s3);
-    DOT_PRODUCT(acc1.s3, values1.s3, values2.s3, values3.s3, values5.s3, values6.s3, values7.s3, values9.s3, values10.s3, values11.s3, (VEC_TYPE(8))((w1.sBCD), (w1.sEF), (w2.s012)), w2.s3);
-
-#if defined(HAS_BIAS)
-    Vector biases = CONVERT_TO_VECTOR_STRUCT(biases);
-
-    VEC_INT bias_values = VLOAD(VEC_SIZE)(0, (__global int *)biases.ptr);
-
-    acc0 += bias_values;
-    acc1 += bias_values;
-
-#endif // defined(HAS_BIAS)
-
-#if WEIGHTS_OFFSET != 0
-    acc0 += WEIGHTS_OFFSET * sum0;
-    acc1 += WEIGHTS_OFFSET * sum1;
-#endif // WEIGHTS_OFFSET != 0
-
-#if K_OFFSET != 0
-    acc0 += (VEC_INT)K_OFFSET;
-    acc1 += (VEC_INT)K_OFFSET;
-
-#endif // K_OFFSET != 0
-
-#if defined(REAL_MULTIPLIER)
-
-    acc0 = CONVERT(round(CONVERT(acc0, VEC_FLOAT) * (VEC_FLOAT)REAL_MULTIPLIER), VEC_INT);
-    acc1 = CONVERT(round(CONVERT(acc1, VEC_FLOAT) * (VEC_FLOAT)REAL_MULTIPLIER), VEC_INT);
-
-#else // defined(REAL_MULTIPLIER)
-
-#if OUTPUT_SHIFT < 0
-    acc0                     = ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(acc0, OUTPUT_MULTIPLIER, OUTPUT_SHIFT, VEC_SIZE);
-    acc1                     = ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(acc1, OUTPUT_MULTIPLIER, OUTPUT_SHIFT, VEC_SIZE);
-#else  // OUTPUT_SHIFT < 0
-    acc0    = asymm_mult_by_quant_multiplier_less_than_one(acc0, OUTPUT_MULTIPLIER, OUTPUT_SHIFT);
-    acc1    = asymm_mult_by_quant_multiplier_less_than_one(acc1, OUTPUT_MULTIPLIER, OUTPUT_SHIFT);
-#endif // OUTPUT_SHIFT < 0
-
-#endif // defined(REAL_MULTIPLIER)
-    acc0 += (VEC_INT)OUTPUT_OFFSET;
-    acc1 += (VEC_INT)OUTPUT_OFFSET;
-
-    VEC_TYPE(VEC_SIZE)
-    res0 = CONVERT_SAT(acc0, VEC_TYPE(VEC_SIZE));
-    VEC_TYPE(VEC_SIZE)
-    res1 = CONVERT_SAT(acc1, VEC_TYPE(VEC_SIZE));
-
-#if defined(DST_DEPTH)
-    __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + x * dst_step_x + y * dst_step_y + z * dst_step_z + b * dst_stride_w;
-#else  /* defined(DST_DEPTH) */
-    __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + x * dst_step_x + y * dst_step_y + z * dst_step_z;
-#endif /* defined(DST_DEPTH) */
-
-    VSTORE(VEC_SIZE)
-    (ACTIVATION_FUNC(res0), 0, (__global DATA_TYPE *)(dst_addr + 0 * dst_stride_y));
-    VSTORE(VEC_SIZE)
-    (ACTIVATION_FUNC(res1), 0, (__global DATA_TYPE *)(dst_addr + 1 * dst_stride_y));
-}
-#endif // defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8) && VEC_SIZE==4
-
-#endif // defined(NUM_ROWS_PROCESSED) && defined(NUM_PLANES_PROCESSED)
-
 #endif // defined(VEC_SIZE) && defined(SRC_DIM_1) && defined(SRC_DIM_2) && defined(CONV_PAD_TOP) && defined(CONV_PAD_LEFT)
 
 #endif // defined(WEIGHTS_PROMOTED_TYPE)
@@ -1612,7 +783,7 @@ __kernel void dwc_3x3_reshaped_quantized8_dot8_stride1_nhwc(
 #endif // defined(WEIGHTS_OFFSET) && defined(INPUT_OFFSET) && defined(K_OFFSET) && ((defined(OUTPUT_OFFSET) && defined(OUTPUT_MULTIPLIER) && defined(OUTPUT_SHIFT)) || defined(REAL_MULTIPLIER))
 
 #if defined(SRC_DIM1) && defined(SRC_DIM2) && defined(KERNEL_WIDTH) && defined(KERNEL_HEIGHT) && defined(N0) && defined(DILATION_X) && defined(DILATION_Y) && defined(CONV_STRIDE_X) && defined(CONV_STRIDE_Y) && defined(CONV_PAD_LEFT) && defined(CONV_PAD_TOP) && defined(INPUT_OFFSET) && defined(WEIGHTS_OFFSET) && defined(OUTPUT_OFFSET) && defined(OUTPUT_SHIFT) && defined(OUTPUT_MULTIPLIER) && defined(VEC_SIZE_LEFTOVER)
-/** This function computes the depthwise convolution for NHWC data layout. This kernel assumes that the weights tensor is NOT reshaped
+/** This function computes the depthwise convolution for NHWC data layout.
  *
  * @note The number of elements processed must be passed at compile time using -DN0 (e.g. -DN0=2)
  * @note The depth multiplier must be passed at compile time using -DDEPTH_MULTIPLIER (e.g. -DDEPTH_MULTIPLIER=1)
diff --git a/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.cpp b/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.cpp
index 287a965f5b..dda70d2231 100644
--- a/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.cpp
+++ b/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.cpp
@@ -114,20 +114,19 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *weights,
 }
 
 std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *weights, ITensorInfo *output, const PadStrideInfo &conv_info,
-                                                        unsigned int depth_multiplier, GPUTarget gpu_target, std::string &kernel_name, const Size2D dilation)
+                                                        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);
+    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());
-    const bool         is_bifrost    = get_arch_from_target(gpu_target) == GPUTarget::BIFROST;
 
     // Configure kernel window
     unsigned int num_elems_read_per_iteration_x    = 0;
@@ -156,31 +155,28 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITen
                 num_elems_read_per_iteration_x = 3 + (num_elems_written_per_iteration_x - 1) * conv_stride_x;
                 break;
         }
-        if(is_bifrost)
+        if(conv_stride_x == 1 && conv_stride_y == 1)
         {
-            if(conv_stride_x == 1 && conv_stride_y == 1)
-            {
-                kernel_name                       = "depthwise_convolution_3x3_stridex1_stridey1_bifrost_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_bifrost_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;
-            }
+            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 && is_bifrost)
+    else if(input->data_type() == DataType::F32)
     {
         if(conv_stride_x == 1 && conv_stride_y == 1)
         {
-            kernel_name                       = "depthwise_convolution_3x3_stridex1_stridey1_bifrost_f32";
+            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;
@@ -188,7 +184,7 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITen
         }
         else if(conv_stride_x == 2 && conv_stride_y == 2)
         {
-            kernel_name                       = "depthwise_convolution_3x3_stridex2_stridey2_bifrost_f32";
+            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;
@@ -239,7 +235,7 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITen
 } // namespace
 
 CLDepthwiseConvolutionLayer3x3NCHWKernel::CLDepthwiseConvolutionLayer3x3NCHWKernel()
-    : _conv_stride_x(0), _conv_pad_top(0), _conv_pad_left(0)
+    : _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)
 {
 }
 
@@ -278,10 +274,9 @@ void CLDepthwiseConvolutionLayer3x3NCHWKernel::configure(const CLCompileContext
     _is_quantized       = is_data_type_quantized_asymmetric(input->info()->data_type());
 
     // Configure kernel window
-    std::string     kernel_name;
-    const GPUTarget gpu_target = get_target();
+    std::string kernel_name;
 
-    auto win_config = validate_and_configure_window(input->info(), weights->info(), output->info(), conv_info, depth_multiplier, gpu_target, kernel_name, dilation);
+    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);
 
@@ -372,13 +367,13 @@ void CLDepthwiseConvolutionLayer3x3NCHWKernel::configure(const CLCompileContext
 }
 
 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, GPUTarget gpu_target,
+                                                          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, gpu_target, kernel_name, dilation)
+                                                              conv_info, depth_multiplier, kernel_name, dilation)
                                 .first);
 
     return Status{};
diff --git a/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.h b/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.h
index 45b5869676..c4e475f6f2 100644
--- a/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.h
+++ b/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018-2020 Arm Limited.
+ * Copyright (c) 2018-2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -24,7 +24,7 @@
 #ifndef ARM_COMPUTE_CLDEPTHWISECONVOLUTIONNCHWKERNEL3x3_H
 #define ARM_COMPUTE_CLDEPTHWISECONVOLUTIONNCHWKERNEL3x3_H
 
-#include "src/core/CL/kernels/ICLDepthwiseConvolutionLayer3x3Kernel.h"
+#include "src/core/CL/ICLKernel.h"
 
 namespace arm_compute
 {
@@ -32,11 +32,19 @@ class ICLTensor;
 
 /** Interface for the kernel to run a 3x3 depthwise convolution on a tensor when the data layout is NCHW.
  */
-class CLDepthwiseConvolutionLayer3x3NCHWKernel : public ICLDepthwiseConvolutionLayer3x3Kernel
+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.
@@ -56,7 +64,7 @@ public:
      */
     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) override;
+                   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.
@@ -77,7 +85,7 @@ public:
      */
     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) override;
+                   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.
@@ -89,7 +97,6 @@ public:
      * @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] gpu_target         (Optional) GPU target to validate the kernel for. Defaults to midgard.
      * @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
@@ -99,13 +106,23 @@ public:
      * @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(), GPUTarget gpu_target = GPUTarget::MIDGARD,
+                           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;
diff --git a/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.cpp b/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.cpp
index f7603e6397..2a1365e6e2 100644
--- a/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.cpp
+++ b/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.cpp
@@ -30,7 +30,6 @@
 #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"
@@ -43,17 +42,11 @@ 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,
-                          const ITensorInfo *output_multipliers, const ITensorInfo *output_shifts)
+                          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, DataType::QASYMM8, DataType::QASYMM8_SIGNED);
-    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_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32);
     ARM_COMPUTE_RETURN_ERROR_ON(depth_multiplier > 1); // COMPMID-1071 Add depth multiplier support for NHWC
 
     ARM_COMPUTE_RETURN_ERROR_ON(conv_info.stride().first < 1);
@@ -61,54 +54,21 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *weights,
 
     ARM_COMPUTE_RETURN_ERROR_ON((dilation.x() < 1) || (dilation.y() < 1));
 
-    const bool   is_qasymm      = is_data_type_quantized_asymmetric(input->data_type());
     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);
-    if(is_qasymm)
-    {
-        DepthwiseConvolutionReshapeInfo info;
-        info.c0 = 4;
-        ARM_COMPUTE_RETURN_ERROR_ON((weights->dimension(0) / info.c0) != weights_width * weights_height);
-
-        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(output_shape[0] != output_multipliers->dimension(0));
-            ARM_COMPUTE_RETURN_ERROR_ON(output_shape[0] != 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);
-        ARM_COMPUTE_RETURN_ERROR_ON((weights->dimension(1) != weights_width) || (weights->dimension(2) != weights_height));
-    }
+
+    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]);
-        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_MISMATCHING_DATA_TYPES(weights, biases);
 
         ARM_COMPUTE_RETURN_ERROR_ON(biases->num_dimensions() > 1);
     }
@@ -122,10 +82,9 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *weights,
 }
 
 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,
-                                                        ITensorInfo *output_multipliers, ITensorInfo *output_shifts)
+                                                        const PadStrideInfo &conv_info, unsigned int depth_multiplier, const Size2D &dilation)
 {
-    ARM_COMPUTE_UNUSED(weights);
+    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);
@@ -134,115 +93,46 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITen
     Window win{};
     Status err{};
 
-    if(is_data_type_quantized_asymmetric(input->data_type()))
-    {
-        const unsigned int num_elems_accessed_per_iteration = 4;
-        const unsigned int num_rows_read_per_iteration      = num_rows_processed_per_iteration + 2;
-        const unsigned int num_rows_written_per_iteration   = std::ceil(num_rows_processed_per_iteration / static_cast<float>(conv_info.stride().first));
-
-        BorderSize border_size;
-        border_size = BorderSize(conv_info.pad_left(), 0, std::max(std::max(conv_info.pad_right(), conv_info.pad_bottom()), conv_info.pad_top()), 0);
-
-        // Configure kernel window
-        win = calculate_max_window(*output, Steps(num_elems_accessed_per_iteration, num_rows_written_per_iteration));
-
-        AccessWindowStatic input_access(input, 0, -border_size.top, ceil_to_multiple(input->dimension(0), num_elems_accessed_per_iteration),
-                                        ceil_to_multiple(input->dimension(1) + border_size.bottom, num_rows_read_per_iteration));
-        AccessWindowRectangle output_access(output, 0, 0, num_elems_accessed_per_iteration, num_rows_written_per_iteration);
-
-        bool window_changed = false;
-
-        if((output_multipliers != nullptr) && (output_shifts != nullptr))
-        {
-            AccessWindowHorizontal output_multipliers_access(output_multipliers, 0, num_elems_accessed_per_iteration);
-            AccessWindowHorizontal output_shifts_access(output_shifts, 0, num_elems_accessed_per_iteration);
-            window_changed = window_changed || update_window_and_padding(win, input_access, output_access, output_multipliers_access, output_shifts_access);
-        }
-        else
-        {
-            Status err = ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "output_multipliers and output_shifts must be non-nullptr for quantized input");
-            return std::make_pair(err, win);
-        }
-
-        if(bias != nullptr)
-        {
-            AccessWindowHorizontal bias_access(bias, 0, num_elems_accessed_per_iteration);
-            window_changed = window_changed || update_window_and_padding(win, bias_access);
-        }
-
-        err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
-    }
-    else
-    {
-        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));
-    }
+    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()
-    : _num_planes_processed_per_iteration(1)
-{
-}
-
-BorderSize CLDepthwiseConvolutionLayer3x3NHWCKernel::border_size() const
+    : _input(), _output(), _weights(), _biases(), _num_planes_processed_per_iteration(1)
 {
-    return _border_size;
 }
 
 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,
-                                                         const ICLTensor *output_multipliers, const ICLTensor *output_shifts)
+                                                         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, output_multipliers, output_shifts);
+    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,
-                                                         const ICLTensor *output_multipliers, const ICLTensor *output_shifts)
+                                                         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,
-                                                  (output_multipliers != nullptr) ? output_multipliers->info() : nullptr,
-                                                  (output_shifts != nullptr) ? output_shifts->info() : nullptr));
+                                                  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,
-                                                    (output_multipliers != nullptr) ? output_multipliers->info() : nullptr,
-                                                    (output_shifts != nullptr) ? output_shifts->info() : nullptr);
+                                                    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);
-    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;
+    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;
-    _conv_stride_y                      = conv_info.stride().second;
     _num_planes_processed_per_iteration = is_stride_1_dilation_1 ? 2 : 1;
-    _output_multipliers                 = output_multipliers;
-    _output_shifts                      = output_shifts;
-    _is_quantized                       = is_data_type_quantized_asymmetric(input->info()->data_type());
 
-    if(_is_quantized)
-    {
-        _border_size = BorderSize(input->info()->padding());
-
-        // If QASYMM8 and the 8 bit dot product is available, force _num_planes_processed_per_iteration to 1
-        if(is_dot8_supported)
-        {
-            _num_planes_processed_per_iteration = 1;
-        }
-    }
-
-    unsigned int num_elems_accessed_per_iteration = _is_quantized ? 4 : adjust_vec_size(4 / input->info()->element_size(), input->info()->dimension(0));
+    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;
@@ -257,54 +147,8 @@ void CLDepthwiseConvolutionLayer3x3NHWCKernel::configure(const CLCompileContext
     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)))));
-
-    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();
-
-        build_opts.add_option("-DSRC_DIM_1=" + support::cpp11::to_string(_input->info()->dimension(1)));
-        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("-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(), "-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)
     {
@@ -317,30 +161,20 @@ void CLDepthwiseConvolutionLayer3x3NHWCKernel::configure(const CLCompileContext
     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_stride_y));
+        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()));
     }
 
-    std::string kernel_name;
     // Create kernel
-    if(_is_quantized)
-    {
-        kernel_name = std::string("dwc_3x3_reshaped_quantized8");
-        kernel_name += (is_dot8_supported && is_stride_1_dilation_1 ? "_dot8" : "");
-        kernel_name += (is_stride_1_dilation_1 ? "_stride1" : "");
-        kernel_name += "_nhwc";
-    }
-    else
-    {
-        kernel_name = std::string("depthwise_convolution_3x3_nhwc");
-        kernel_name += (is_stride_1_dilation_1 ? "_stride1" : "");
-    }
+    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(!_is_quantized && has_padding_changed(padding_info));
+    ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info));
 
     // Set config_id for enabling LWS tuning
     _config_id = kernel_name;
@@ -359,15 +193,12 @@ void CLDepthwiseConvolutionLayer3x3NHWCKernel::configure(const CLCompileContext
 }
 
 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,
-                                                          const ITensorInfo *output_multipliers, const ITensorInfo *output_shifts)
+                                                          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, output_multipliers, output_shifts));
+    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,
-                                                              (output_multipliers != nullptr) ? output_multipliers->clone().get() : nullptr,
-                                                              (output_shifts != nullptr) ? output_shifts->clone().get() : nullptr)
+                                                              output->clone().get(), conv_info, depth_multiplier, dilation)
                                 .first);
     return Status{};
 }
@@ -382,16 +213,7 @@ void CLDepthwiseConvolutionLayer3x3NHWCKernel::run(const Window &window, cl::Com
     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() + (_is_quantized ? num_arguments_per_2D_tensor() : num_arguments_per_3D_tensor());
-
-    if(_is_quantized)
-    {
-        Window slice;
-        slice.use_tensor_dimensions(_output_multipliers->info()->tensor_shape());
-        slice.set_dimension_step(Window::DimX, window.x().step());
-        add_1D_tensor_argument(idx, _output_multipliers, slice);
-        add_1D_tensor_argument(idx, _output_shifts, slice);
-    }
+    unsigned int idx = 2 * num_arguments_per_4D_tensor() + num_arguments_per_3D_tensor();
 
     if(_biases != nullptr)
     {
@@ -401,62 +223,14 @@ void CLDepthwiseConvolutionLayer3x3NHWCKernel::run(const Window &window, cl::Com
         add_1D_tensor_argument(idx, _biases, win_biases);
     }
 
-    if(_is_quantized)
-    {
-        // Calculate the max_offset.
-        // max_offset is the offset for the last NOT valid value in the Z dimension (spatial dimension Y for NHWC)
-        //  |******************|
-        //  |     pad_top      |
-        //  |******************|
-        //  |                  |
-        //  |      plane0      |
-        //  |      batch0      |
-        //  |__________________|
-        //  |******************|       Batch 0
-        //  |    pad_bottom    |
-        //  |     pad_top      |
-        //  |******************|
-        //  |                  |
-        //  |      plane1      |
-        //  |      batch0      |
-        //  |__________________|-----> max_offset
-        //  |******************|
-        //  |    pad_bottom    |
-        //  |     pad_top      |
-        //  |******************|
-        //  |                  |
-        //  |      plane0      |
-        //  |      batch1      |
-        //  |__________________|
-        //  |******************|       Batch 1
-        //  |    pad_bottom    |
-        //  |     pad_top      |
-        //  |******************|
-        //  |                  |
-        //  |      plane1      |
-        //  |      batch1      |
-        //  |__________________|
-        //  |     pad_bottom   |
-        //  |******************|
-        const int max_offset = ((_input->info()->dimension(1) * _input->info()->dimension(2)) + (_input->info()->padding().bottom + _input->info()->padding().top) * (_input->info()->dimension(
-                                    2) - 1)) * _input->info()->strides_in_bytes().y();
-        _kernel.setArg(idx, max_offset);
-    }
-
     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);
-        if(_is_quantized)
-        {
-            add_2D_tensor_argument(idx, _weights, slice);
-        }
-        else
-        {
-            add_3D_tensor_argument(idx, _weights, slice);
-        }
+        add_3D_tensor_argument(idx, _weights, slice);
+
         enqueue(queue, *this, slice, lws_hint());
     }
     while(win.slide_window_slice_4D(slice));
diff --git a/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.h b/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.h
index ce0bf5ceb3..ee47d98807 100644
--- a/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.h
+++ b/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018-2020 Arm Limited.
+ * Copyright (c) 2018-2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -24,7 +24,7 @@
 #ifndef ARM_COMPUTE_CLDEPTHWISECONVOLUTIONNHWCKERNEL3x3_H
 #define ARM_COMPUTE_CLDEPTHWISECONVOLUTIONNHWCKERNEL3x3_H
 
-#include "src/core/CL/kernels/ICLDepthwiseConvolutionLayer3x3Kernel.h"
+#include "src/core/CL/ICLKernel.h"
 
 namespace arm_compute
 {
@@ -32,81 +32,78 @@ class ICLTensor;
 
 /** Interface for the kernel to run a 3x3 depthwise convolution on a tensor when the data layout is NHWC.
  */
-class CLDepthwiseConvolutionLayer3x3NHWCKernel : public ICLDepthwiseConvolutionLayer3x3Kernel
+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: QASYMM8/QASYMM8_SIGNED/F16/F32.
-     * @param[in]  weights            Weights tensor. A 3D tensor with dimensions [IFM, 3, 3].
-     *                                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 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 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
+     * @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),
-                   const ICLTensor *output_multipliers = nullptr, const ICLTensor *output_shifts = nullptr) override;
+                   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: QASYMM8/QASYMM8_SIGNED/F16/F32.
-     * @param[in]  weights            Weights tensor. A 3D tensor with dimensions [IFM, 3, 3].
-     *                                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 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 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
+     * @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),
-                   const ICLTensor *output_multipliers = nullptr, const ICLTensor *output_shifts = nullptr) override;
+                   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: QASYMM8/QASYMM8_SIGNED/F16/F32.
-     * @param[in] weights            Weights tensor info. A 3D tensor with dimensions [IFM, 3, 3].
-     *                               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 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).
-     * @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
+     * @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),
-                           const ITensorInfo *output_multipliers = nullptr, const ITensorInfo *output_shifts = nullptr);
+                           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;
-    BorderSize border_size() const override;
 
 private:
+    const ICLTensor *_input;
+    ICLTensor       *_output;
+    const ICLTensor *_weights;
+    const ICLTensor *_biases;
+
     unsigned int _num_planes_processed_per_iteration;
 };
 } // namespace arm_compute
diff --git a/src/core/CL/kernels/CLDepthwiseConvolutionLayerReshapeWeightsKernel.cpp b/src/core/CL/kernels/CLDepthwiseConvolutionLayerReshapeWeightsKernel.cpp
deleted file mode 100644
index 386d634cea..0000000000
--- a/src/core/CL/kernels/CLDepthwiseConvolutionLayerReshapeWeightsKernel.cpp
+++ /dev/null
@@ -1,131 +0,0 @@
-/*
- * Copyright (c) 2019-2021 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/CLDepthwiseConvolutionLayerReshapeWeightsKernel.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 *output, const DepthwiseConvolutionReshapeInfo &info)
-{
-    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
-    const size_t idx_w = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::WIDTH);
-    const size_t idx_h = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::HEIGHT);
-
-    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(info.c0 != 4);
-    ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(idx_h) != 3);
-    ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(idx_w) != 3);
-    ARM_COMPUTE_RETURN_ERROR_ON(input->data_type() == DataType::UNKNOWN);
-
-    if(output->total_size() != 0)
-    {
-        auto reshaped_weights_shape = arm_compute::misc::shape_calculator::compute_reshaped_depthwise_weights_shape(*input, info);
-        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
-        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), reshaped_weights_shape);
-        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(input, output);
-    }
-
-    return Status{};
-}
-
-std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *output, const DepthwiseConvolutionReshapeInfo &info)
-{
-    auto reshaped_input_shape = arm_compute::misc::shape_calculator::compute_reshaped_depthwise_weights_shape(*input, info);
-    auto_init_if_empty(*output, reshaped_input_shape, 1, input->data_type(), input->quantization_info());
-
-    Window                 win = calculate_max_window(*input, Steps(info.c0));
-    AccessWindowHorizontal weights_access(input, 0, info.c0);
-    const bool             window_changed = update_window_and_padding(win, weights_access);
-
-    Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
-    return std::make_pair(err, win);
-}
-} // namespace
-
-CLDepthwiseConvolutionLayerReshapeWeightsKernel::CLDepthwiseConvolutionLayerReshapeWeightsKernel()
-    : _input(nullptr), _output(nullptr)
-{
-}
-
-void CLDepthwiseConvolutionLayerReshapeWeightsKernel::configure(const ICLTensor *input, ICLTensor *output, const DepthwiseConvolutionReshapeInfo &info)
-{
-    configure(CLKernelLibrary::get().get_compile_context(), input, output, info);
-}
-
-void CLDepthwiseConvolutionLayerReshapeWeightsKernel::configure(const CLCompileContext &compile_context, const ICLTensor *input, ICLTensor *output, const DepthwiseConvolutionReshapeInfo &info)
-{
-    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
-    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), info));
-    auto win_config = validate_and_configure_window(input->info(), output->info(), info);
-    ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
-
-    ICLKernel::configure_internal(win_config.second);
-
-    _input  = input;
-    _output = output;
-
-    // Build the kernel
-    CLBuildOptions build_opts;
-    build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(info.c0));
-    build_opts.add_option("-DDST_WIDTH=" + support::cpp11::to_string(_output->info()->dimension(0)));
-    build_opts.add_option_if(info.transpose, "-DTRANSPOSE");
-    build_opts.add_option("-DDATA_TYPE=" + get_cl_unsigned_type_from_element_size(input->info()->element_size()));
-
-    _kernel = create_kernel(compile_context, "depthwise_convolution_reshape_weights", build_opts.options());
-}
-
-Status CLDepthwiseConvolutionLayerReshapeWeightsKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const DepthwiseConvolutionReshapeInfo &info)
-{
-    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, info));
-    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), output->clone().get(), info).first);
-    return Status{};
-}
-
-void CLDepthwiseConvolutionLayerReshapeWeightsKernel::run(const Window &window, cl::CommandQueue &queue)
-{
-    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
-    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(IKernel::window(), window);
-
-    unsigned int idx = 0;
-    add_3D_tensor_argument(idx, _input, window);
-    add_2D_tensor_argument(idx, _output, window);
-    enqueue(queue, *this, window, lws_hint());
-}
-} // namespace arm_compute
diff --git a/src/core/CL/kernels/CLDepthwiseConvolutionLayerReshapeWeightsKernel.h b/src/core/CL/kernels/CLDepthwiseConvolutionLayerReshapeWeightsKernel.h
deleted file mode 100644
index 650fe9a11b..0000000000
--- a/src/core/CL/kernels/CLDepthwiseConvolutionLayerReshapeWeightsKernel.h
+++ /dev/null
@@ -1,85 +0,0 @@
-/*
- * Copyright (c) 2019-2020 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_CLDEPTHWISECONVOLUTIONLAYERRESHAPEWEIGHTSKERNEL_H
-#define ARM_COMPUTE_CLDEPTHWISECONVOLUTIONLAYERRESHAPEWEIGHTSKERNEL_H
-
-#include "src/core/CL/ICLKernel.h"
-
-namespace arm_compute
-{
-class ICLTensor;
-
-/** Interface for the kernel to reshape the weights of depthwise convolution. */
-class CLDepthwiseConvolutionLayerReshapeWeightsKernel : public ICLKernel
-{
-public:
-    /** Default constructor */
-    CLDepthwiseConvolutionLayerReshapeWeightsKernel();
-    /** Prevent instances of this class from being copied (As this class contains pointers) */
-    CLDepthwiseConvolutionLayerReshapeWeightsKernel(const CLDepthwiseConvolutionLayerReshapeWeightsKernel &) = delete;
-    /** Prevent instances of this class from being copied (As this class contains pointers) */
-    CLDepthwiseConvolutionLayerReshapeWeightsKernel &operator=(const CLDepthwiseConvolutionLayerReshapeWeightsKernel &) = delete;
-    /** Default Move Constructor. */
-    CLDepthwiseConvolutionLayerReshapeWeightsKernel(CLDepthwiseConvolutionLayerReshapeWeightsKernel &&) = default;
-    /** Default move assignment operator */
-    CLDepthwiseConvolutionLayerReshapeWeightsKernel &operator=(CLDepthwiseConvolutionLayerReshapeWeightsKernel &&) = default;
-
-    /** Initialize the function's source and destination.
-     *
-     * @param[in]  input  The input tensor of dimension [IFM, W, H]. Data types supported: All. Data layouts supported: NHWC
-     * @param[out] output The output tensor of dimension [W*H*C0, ceil(IFM/C0)]. C0 is the number of channels read by each thread. Data types supported: same as @p weights.
-     * @param[in]  info   Depthwise convolution information to reshape the input tensor.
-     */
-    void configure(const ICLTensor *input, ICLTensor *output, const DepthwiseConvolutionReshapeInfo &info);
-    /** Initialize the function's source and destination.
-     *
-     * @param[in]  compile_context The compile context to be used.
-     * @param[in]  input           The input tensor of dimension [IFM, W, H]. Data types supported: All. Data layouts supported: NHWC
-     * @param[out] output          The output tensor of dimension [W*H*C0, ceil(IFM/C0)]. C0 is the number of channels read by each thread. Data types supported: same as @p weights.
-     * @param[in]  info            Depthwise convolution information to reshape the input tensor.
-     */
-    void configure(const CLCompileContext &compile_context, const ICLTensor *input, ICLTensor *output, const DepthwiseConvolutionReshapeInfo &info);
-
-    /** Static function to check if given info will lead to a valid configuration of @ref CLDepthwiseConvolutionLayer3x3NHWCKernel
-     *
-     * @param[in] input  The input tensor info of dimension [IFM, W, H]. Data types supported: All. Data layouts supported: NHWC
-     * @param[in] output The output tensor info of dimension [W*H*C0, ceil(IFM/C0)]. C0 is the number of channels read by each thread. Data types supported: same as @p weights.
-     * @param[in] info   Depthwise convolution information to reshape the input tensor.
-     *
-     * @return a Status
-     */
-    static Status validate(const ITensorInfo *input, const ITensorInfo *output, const DepthwiseConvolutionReshapeInfo &info);
-
-    // Inherited methods overridden:
-    void run(const Window &window, cl::CommandQueue &queue) override;
-
-private:
-    const ICLTensor *_input;
-    ICLTensor       *_output;
-
-    void configure_dot_product(const DepthwiseConvolutionReshapeInfo &info);
-    void configure_generic(const DepthwiseConvolutionReshapeInfo &info);
-};
-} // namespace arm_compute
-#endif /* ARM_COMPUTE_CLDEPTHWISECONVOLUTIONLAYERRESHAPEWEIGHTSKERNEL_H */
diff --git a/src/core/CL/kernels/CLL2NormalizeLayerKernel.cpp b/src/core/CL/kernels/CLL2NormalizeLayerKernel.cpp
index d9f293ba73..c688951d57 100644
--- a/src/core/CL/kernels/CLL2NormalizeLayerKernel.cpp
+++ b/src/core/CL/kernels/CLL2NormalizeLayerKernel.cpp
@@ -36,8 +36,6 @@
 
 #include "support/StringSupport.h"
 
-#include "utils/TypePrinter.h"
-
 namespace arm_compute
 {
 namespace
diff --git a/src/core/CL/kernels/ICLDepthwiseConvolutionLayer3x3Kernel.h b/src/core/CL/kernels/ICLDepthwiseConvolutionLayer3x3Kernel.h
deleted file mode 100644
index 4c92ae417f..0000000000
--- a/src/core/CL/kernels/ICLDepthwiseConvolutionLayer3x3Kernel.h
+++ /dev/null
@@ -1,105 +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_ICLDEPTHWISECONVOLUTIONKERNEL3x3_H
-#define ARM_COMPUTE_ICLDEPTHWISECONVOLUTIONKERNEL3x3_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.
- */
-class ICLDepthwiseConvolutionLayer3x3Kernel : public ICLKernel
-{
-public:
-    /** Default constructor */
-    ICLDepthwiseConvolutionLayer3x3Kernel()
-        : _border_size(0), _input(), _output(), _weights(), _biases(), _conv_stride_y(1), _output_multipliers(), _output_shifts(), _is_quantized(false)
-    {
-    }
-    /** Prevent instances of this class from being copied (As this class contains pointers) */
-    ICLDepthwiseConvolutionLayer3x3Kernel(const ICLDepthwiseConvolutionLayer3x3Kernel &) = delete;
-    /** Prevent instances of this class from being copied (As this class contains pointers) */
-    ICLDepthwiseConvolutionLayer3x3Kernel &operator=(const ICLDepthwiseConvolutionLayer3x3Kernel &) = delete;
-    /** Default Move Constructor. */
-    ICLDepthwiseConvolutionLayer3x3Kernel(ICLDepthwiseConvolutionLayer3x3Kernel &&) = default;
-    /** Default move assignment operator */
-    ICLDepthwiseConvolutionLayer3x3Kernel &operator=(ICLDepthwiseConvolutionLayer3x3Kernel &&) = default;
-    /** Initialize the function's source, destination, conv and border_size.
-     *
-     * @param[in]  input              Source tensor. DataType supported: QASYMM8/F16/F32.
-     * @param[in]  weights            Weights tensor. A 3D tensor with dimensions [3, 3, IFM].
-     *                                Data type supported: Same as @p input, QASYMM8/QSYMM8_PER_CHANNEL when input is QASYMM8.
-     * @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.
-     * @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 for QASYMM8.
-     * @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
-     */
-    virtual 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) = 0;
-    /** 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/F16/F32.
-     * @param[in]  weights            Weights tensor. A 3D tensor with dimensions [3, 3, IFM].
-     *                                Data type supported: Same as @p input, QASYMM8/QSYMM8_PER_CHANNEL when input is QASYMM8.
-     * @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.
-     * @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 for QASYMM8.
-     * @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
-     */
-    virtual 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) = 0;
-
-protected:
-    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;
-};
-} // namespace arm_compute
-#endif /*ARM_COMPUTE_ICLDEPTHWISECONVOLUTIONKERNEL3x3_H */
diff --git a/src/runtime/CL/functions/CLDepthwiseConvolutionLayer.cpp b/src/runtime/CL/functions/CLDepthwiseConvolutionLayer.cpp
index 8e3d010786..6467caffef 100644
--- a/src/runtime/CL/functions/CLDepthwiseConvolutionLayer.cpp
+++ b/src/runtime/CL/functions/CLDepthwiseConvolutionLayer.cpp
@@ -30,13 +30,9 @@
 #include "arm_compute/core/utils/quantization/AsymmHelpers.h"
 #include "arm_compute/runtime/CL/CLScheduler.h"
 #include "src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.h"
-#include "src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.h"
-#include "src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.h"
 #include "src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.h"
 #include "src/core/CL/kernels/CLDepthwiseConvolutionLayerNativeKernel.h"
-#include "src/core/CL/kernels/CLDepthwiseConvolutionLayerReshapeWeightsKernel.h"
 #include "src/core/CL/kernels/CLFillBorderKernel.h"
-#include "src/core/CL/kernels/ICLDepthwiseConvolutionLayer3x3Kernel.h"
 
 namespace arm_compute
 {
@@ -46,23 +42,18 @@ using namespace arm_compute::misc::shape_calculator;
 namespace
 {
 Status validate_arguments_3x3(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info,
-                              unsigned int depth_multiplier, ActivationLayerInfo act_info, GPUTarget gpu_target, const Size2D &dilation)
+                              unsigned int depth_multiplier, ActivationLayerInfo act_info, const Size2D &dilation)
 {
     // This function should be removed and incorporated inside CLDepthwiseConvolutionLayerInternal3x3 once CLDepthwiseConvolutionLayer3x3 is properly removed
     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, weights, output);
     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F16, DataType::F32);
     ARM_COMPUTE_RETURN_ERROR_ON(input->data_layout() == DataLayout::UNKNOWN);
 
-    const bool                      is_quantized           = is_data_type_quantized_asymmetric(input->data_type());
-    const bool                      is_nhwc                = input->data_layout() == DataLayout::NHWC;
-    const bool                      needs_permute          = is_nhwc && (depth_multiplier > 1);
-    const bool                      needs_weights_reshape  = is_nhwc && (depth_multiplier == 1) && is_quantized;
-    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);
-    const bool                      is_dot8_supported      = dot8_supported(CLKernelLibrary::get().get_device());
-    DepthwiseConvolutionReshapeInfo info;
-    info.c0        = 4;
-    info.transpose = is_stride_1_dilation_1 && is_dot8_supported;
+    const bool is_quantized  = is_data_type_quantized_asymmetric(input->data_type());
+    const bool is_nhwc       = input->data_layout() == DataLayout::NHWC;
+    const bool needs_permute = is_nhwc && (depth_multiplier > 1);
+
+    ARM_COMPUTE_RETURN_ERROR_ON(is_quantized && is_nhwc && !needs_permute);
 
     TensorInfo output_multipliers_shifts_info(TensorInfo(TensorShape(1U), 1, DataType::S32));
     if(is_quantized)
@@ -96,27 +87,17 @@ Status validate_arguments_3x3(const ITensorInfo *input, const ITensorInfo *weigh
         const TensorInfo permuted_output  = output->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(permuted_output_shape).set_data_layout(DataLayout::NCHW);
 
         ARM_COMPUTE_RETURN_ON_ERROR(CLDepthwiseConvolutionLayer3x3NCHWKernel::validate(&permuted_input, &permuted_weights, biases, &permuted_output,
-                                                                                       conv_info, depth_multiplier, act_info, gpu_target,
+                                                                                       conv_info, depth_multiplier, act_info,
                                                                                        dilation, &output_multipliers_shifts_info, &output_multipliers_shifts_info));
     }
     else if(is_nhwc)
     {
-        if(needs_weights_reshape)
-        {
-            auto reshaped_weights_shape = arm_compute::misc::shape_calculator::compute_reshaped_depthwise_weights_shape(*weights, info);
-            ARM_COMPUTE_RETURN_ON_ERROR(CLDepthwiseConvolutionLayer3x3NHWCKernel::validate(input, &weights->clone()->set_tensor_shape(reshaped_weights_shape), biases,
-                                                                                           output, conv_info, depth_multiplier, act_info,
-                                                                                           dilation, &output_multipliers_shifts_info, &output_multipliers_shifts_info));
-        }
-        else
-        {
-            ARM_COMPUTE_RETURN_ON_ERROR(CLDepthwiseConvolutionLayer3x3NHWCKernel::validate(input, weights, biases, output, conv_info, depth_multiplier, act_info,
-                                                                                           dilation, &output_multipliers_shifts_info, &output_multipliers_shifts_info));
-        }
+        ARM_COMPUTE_RETURN_ON_ERROR(CLDepthwiseConvolutionLayer3x3NHWCKernel::validate(input, weights, biases, output, conv_info, depth_multiplier, act_info,
+                                                                                       dilation));
     }
     else
     {
-        ARM_COMPUTE_RETURN_ON_ERROR(CLDepthwiseConvolutionLayer3x3NCHWKernel::validate(input, weights, biases, output, conv_info, depth_multiplier, act_info, gpu_target,
+        ARM_COMPUTE_RETURN_ON_ERROR(CLDepthwiseConvolutionLayer3x3NCHWKernel::validate(input, weights, biases, output, conv_info, depth_multiplier, act_info,
                                                                                        dilation, &output_multipliers_shifts_info, &output_multipliers_shifts_info));
     }
     return Status{};
@@ -351,12 +332,12 @@ void CLDepthwiseConvolutionLayer::CLDepthwiseConvolutionLayerGeneric::prepare()
 
 CLDepthwiseConvolutionLayer::CLDepthwiseConvolutionLayerInternal3x3::CLDepthwiseConvolutionLayerInternal3x3(std::shared_ptr<IMemoryManager> memory_manager)
     : _memory_group(std::move(memory_manager)),
-      _kernel(nullptr),
+      _kernel_nchw(nullptr),
+      _kernel_nhwc(nullptr),
       _border_handler(std::make_unique<CLFillBorderKernel>()),
       _permute_input_to_nchw(),
       _permute_weights_to_nchw(),
       _permute_output_to_nhwc(),
-      _reshape_weights(std::make_unique<CLDepthwiseConvolutionLayerReshapeWeightsKernel>()),
       _permuted_input(),
       _permuted_weights(),
       _permuted_output(),
@@ -366,7 +347,6 @@ CLDepthwiseConvolutionLayer::CLDepthwiseConvolutionLayerInternal3x3::CLDepthwise
       _input(nullptr),
       _output(nullptr),
       _needs_permute(false),
-      _needs_weights_reshape(false),
       _is_prepared(false),
       _is_quantized(false),
       _is_nhwc(false)
@@ -383,8 +363,6 @@ void CLDepthwiseConvolutionLayer::CLDepthwiseConvolutionLayerInternal3x3::config
                                                                                     ICLTensor           *output,
                                                                                     const PadStrideInfo &conv_info, unsigned int depth_multiplier, ActivationLayerInfo act_info, const Size2D &dilation)
 {
-    const GPUTarget gpu_target = CLScheduler::get().target();
-
     // Perform validation step
     ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output);
     ARM_COMPUTE_ERROR_THROW_ON(CLDepthwiseConvolutionLayerInternal3x3::validate(input->info(),
@@ -394,13 +372,11 @@ void CLDepthwiseConvolutionLayer::CLDepthwiseConvolutionLayerInternal3x3::config
                                                                                 conv_info,
                                                                                 depth_multiplier,
                                                                                 act_info,
-                                                                                gpu_target,
                                                                                 dilation));
 
-    _is_nhwc               = input->info()->data_layout() == DataLayout::NHWC;
-    _is_quantized          = is_data_type_quantized_asymmetric(input->info()->data_type());
-    _needs_permute         = _is_nhwc && (depth_multiplier > 1);
-    _needs_weights_reshape = _is_nhwc && (depth_multiplier == 1) && _is_quantized;
+    _is_nhwc       = input->info()->data_layout() == DataLayout::NHWC;
+    _is_quantized  = is_data_type_quantized_asymmetric(input->info()->data_type());
+    _needs_permute = _is_nhwc && (depth_multiplier > 1);
 
     _is_prepared      = false;
     _original_weights = weights;
@@ -412,13 +388,6 @@ void CLDepthwiseConvolutionLayer::CLDepthwiseConvolutionLayerInternal3x3::config
     ICLTensor       *output_to_use  = output;
 
     const bool is_quantized_per_channel = is_data_type_quantized_per_channel(weights->info()->data_type());
-    const bool is_stride_1              = ((conv_info.stride().first == conv_info.stride().second) && (conv_info.stride().first == 1));
-    const bool is_dot8_supported        = dot8_supported(CLKernelLibrary::get().get_device()) && !is_quantized_per_channel;
-    const bool is_stride_1_dilation_1   = (is_stride_1 && dilation.x() == 1 && dilation.y() == 1);
-
-    DepthwiseConvolutionReshapeInfo info;
-    info.c0        = 4;
-    info.transpose = is_stride_1_dilation_1 && is_dot8_supported;
 
     if(_needs_permute)
     {
@@ -438,20 +407,15 @@ void CLDepthwiseConvolutionLayer::CLDepthwiseConvolutionLayerInternal3x3::config
         weights_to_use = &_permuted_weights;
         output_to_use  = &_permuted_output;
 
-        _kernel = std::make_unique<CLDepthwiseConvolutionLayer3x3NCHWKernel>();
+        _kernel_nchw = std::make_unique<CLDepthwiseConvolutionLayer3x3NCHWKernel>();
     }
     else if(_is_nhwc)
     {
-        if(_needs_weights_reshape)
-        {
-            _reshape_weights->configure(compile_context, weights, &_permuted_weights, info);
-            weights_to_use = &_permuted_weights;
-        }
-        _kernel = std::make_unique<CLDepthwiseConvolutionLayer3x3NHWCKernel>();
+        _kernel_nhwc = std::make_unique<CLDepthwiseConvolutionLayer3x3NHWCKernel>();
     }
     else
     {
-        _kernel = std::make_unique<CLDepthwiseConvolutionLayer3x3NCHWKernel>();
+        _kernel_nchw = std::make_unique<CLDepthwiseConvolutionLayer3x3NCHWKernel>();
     }
 
     CLTensor *output_multipliers_to_use = nullptr;
@@ -469,9 +433,16 @@ void CLDepthwiseConvolutionLayer::CLDepthwiseConvolutionLayerInternal3x3::config
     }
 
     // Configure kernel
-    _kernel->set_target(gpu_target);
-    _kernel->configure(compile_context, input_to_use, weights_to_use, biases, output_to_use, conv_info, depth_multiplier,
-                       act_info, dilation, output_multipliers_to_use, output_shifts_to_use);
+    if(_is_nhwc && !_needs_permute)
+    {
+        _kernel_nhwc->configure(compile_context, input_to_use, weights_to_use, biases, output_to_use, conv_info, depth_multiplier,
+                                act_info, dilation);
+    }
+    else
+    {
+        _kernel_nchw->configure(compile_context, input_to_use, weights_to_use, biases, output_to_use, conv_info, depth_multiplier,
+                                act_info, dilation, output_multipliers_to_use, output_shifts_to_use);
+    }
 
     if(_is_quantized)
     {
@@ -496,13 +467,16 @@ void CLDepthwiseConvolutionLayer::CLDepthwiseConvolutionLayerInternal3x3::config
     {
         zero_value = PixelValue(static_cast<uint8_t>(input->info()->quantization_info().uniform().offset));
     }
-    _border_handler->configure(compile_context, input_to_use, _kernel->border_size(), BorderMode::CONSTANT, zero_value);
+    if(!_is_nhwc || _needs_permute)
+    {
+        _border_handler->configure(compile_context, input_to_use, _kernel_nchw->border_size(), BorderMode::CONSTANT, zero_value);
+    }
 }
 
 Status CLDepthwiseConvolutionLayer::CLDepthwiseConvolutionLayerInternal3x3::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output,
-                                                                                     const PadStrideInfo &conv_info, unsigned int depth_multiplier, ActivationLayerInfo act_info, GPUTarget gpu_target, const Size2D &dilation)
+                                                                                     const PadStrideInfo &conv_info, unsigned int depth_multiplier, ActivationLayerInfo act_info, const Size2D &dilation)
 {
-    return validate_arguments_3x3(input, weights, biases, output, conv_info, depth_multiplier, act_info, gpu_target, dilation);
+    return validate_arguments_3x3(input, weights, biases, output, conv_info, depth_multiplier, act_info, dilation);
 }
 
 void CLDepthwiseConvolutionLayer::CLDepthwiseConvolutionLayerInternal3x3::run()
@@ -516,7 +490,14 @@ void CLDepthwiseConvolutionLayer::CLDepthwiseConvolutionLayerInternal3x3::run()
         _permute_input_to_nchw.run();
     }
     CLScheduler::get().enqueue(*_border_handler);
-    CLScheduler::get().enqueue(*_kernel);
+    if(_is_nhwc && !_needs_permute)
+    {
+        CLScheduler::get().enqueue(*_kernel_nhwc);
+    }
+    else
+    {
+        CLScheduler::get().enqueue(*_kernel_nchw);
+    }
 
     if(_needs_permute)
     {
@@ -552,14 +533,6 @@ void CLDepthwiseConvolutionLayer::CLDepthwiseConvolutionLayerInternal3x3::prepar
             _original_weights->mark_as_unused();
         }
 
-        if(_needs_weights_reshape)
-        {
-            ARM_COMPUTE_ERROR_ON(_needs_permute);
-            ARM_COMPUTE_ERROR_ON(!_original_weights->is_used());
-            _permuted_weights.allocator()->allocate();
-            CLScheduler::get().enqueue(*_reshape_weights);
-            _original_weights->mark_as_unused();
-        }
         _is_prepared = true;
     }
 }
@@ -580,9 +553,8 @@ void CLDepthwiseConvolutionLayer::configure(const CLCompileContext &compile_cont
                                             unsigned int         depth_multiplier,
                                             ActivationLayerInfo act_info, const Size2D &dilation)
 {
-    const GPUTarget gpu_target = CLScheduler::get().target();
-    _depth_conv_func           = get_depthwiseconvolution_function(input->info(), weights->info(), (biases != nullptr) ? biases->info() : nullptr, output->info(), conv_info, depth_multiplier, act_info,
-                                                                   dilation, gpu_target);
+    _depth_conv_func = get_depthwiseconvolution_function(input->info(), weights->info(), (biases != nullptr) ? biases->info() : nullptr, output->info(), conv_info, depth_multiplier, act_info,
+                                                         dilation);
     switch(_depth_conv_func)
     {
         case DepthwiseConvolutionFunction::OPTIMIZED:
@@ -603,12 +575,11 @@ void CLDepthwiseConvolutionLayer::configure(const CLCompileContext &compile_cont
 Status CLDepthwiseConvolutionLayer::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 GPUTarget              gpu_target      = CLScheduler::get().target();
-    DepthwiseConvolutionFunction depth_conv_func = get_depthwiseconvolution_function(input, weights, biases, output, conv_info, depth_multiplier, act_info, dilation, gpu_target);
+    DepthwiseConvolutionFunction depth_conv_func = get_depthwiseconvolution_function(input, weights, biases, output, conv_info, depth_multiplier, act_info, dilation);
     switch(depth_conv_func)
     {
         case DepthwiseConvolutionFunction::OPTIMIZED:
-            return CLDepthwiseConvolutionLayerInternal3x3::validate(input, weights, biases, output, conv_info, depth_multiplier, act_info, gpu_target, dilation);
+            return CLDepthwiseConvolutionLayerInternal3x3::validate(input, weights, biases, output, conv_info, depth_multiplier, act_info, dilation);
         case DepthwiseConvolutionFunction::GENERIC:
             return CLDepthwiseConvolutionLayerGeneric::validate(input, weights, biases, output, conv_info, depth_multiplier, act_info, dilation);
         default:
@@ -618,10 +589,9 @@ Status CLDepthwiseConvolutionLayer::validate(const ITensorInfo *input, const ITe
 
 DepthwiseConvolutionFunction CLDepthwiseConvolutionLayer::get_depthwiseconvolution_function(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, GPUTarget gpu_target)
+                                                                                            unsigned int depth_multiplier, ActivationLayerInfo act_info, const Size2D &dilation)
 {
-    if(bool(CLDepthwiseConvolutionLayerInternal3x3::validate(input, weights, biases, output, conv_info, depth_multiplier, act_info, gpu_target, dilation)) && (is_data_type_float(input->data_type())
-            || get_arch_from_target(gpu_target) == GPUTarget::MIDGARD))
+    if(bool(CLDepthwiseConvolutionLayerInternal3x3::validate(input, weights, biases, output, conv_info, depth_multiplier, act_info, dilation)))
     {
         return DepthwiseConvolutionFunction::OPTIMIZED;
     }