arm_compute v18.05
diff --git a/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp b/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp
index a85078c..2888b43 100644
--- a/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp
+++ b/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp
@@ -23,9 +23,6 @@
  */
 #include "arm_compute/runtime/NEON/functions/NEGEMMConvolutionLayer.h"
 
-#include "arm_compute/core/NEON/kernels/arm32/NEGEMMAArch32Kernel.h"
-#include "arm_compute/core/NEON/kernels/arm64/NEGEMMAArch64Kernel.h"
-#include "arm_compute/core/NEON/kernels/arm64/NEGEMMAArch64NativeKernel.h"
 #include "arm_compute/core/PixelValue.h"
 #include "arm_compute/core/Size2D.h"
 #include "arm_compute/core/Utils.h"
@@ -34,13 +31,6 @@
 #include "arm_compute/runtime/NEON/NEScheduler.h"
 #include "support/ToolchainSupport.h"
 
-namespace arm_compute
-{
-#include "arm_compute/core/NEON/kernels/assembly/gemm_interleaved.hpp"
-#include "arm_compute/core/NEON/kernels/assembly/kernels/a32_sgemm_8x6.hpp"
-#include "arm_compute/core/NEON/kernels/assembly/kernels/a64_sgemm_12x8.hpp"
-} // namespace arm_compute
-
 #include <cmath>
 #include <tuple>
 
@@ -175,19 +165,28 @@
     }
 }
 
-Status validate_and_initialize_values(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const PadStrideInfo &conv_info, const WeightsInfo &weights_info, DataType &dt,
-                                      bool &append_bias,
+Status validate_and_initialize_values(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const PadStrideInfo &conv_info, const WeightsInfo &weights_info,
+                                      const ActivationLayerInfo &act_info, DataType &dt,
+                                      bool &append_bias, bool &skip_im2col,
                                       bool &are_weights_reshaped, unsigned int &kernel_width, unsigned int &kernel_height,
-                                      bool &is_fully_connected_convolution, bool &is_interleaved, bool &is_quantized,
+                                      bool &is_fully_connected_convolution, bool &is_interleaved, bool &is_quantized, bool &is_activationlayer_enabled,
                                       unsigned int &mat_weights_cols, unsigned int &mat_weights_rows,
-                                      unsigned int &conv_w, unsigned int &conv_h)
+                                      unsigned int &conv_w, unsigned int &conv_h, const Size2D &dilation)
 {
     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QASYMM8, DataType::QS16, DataType::F16, DataType::F32);
     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights);
     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, weights);
-    ARM_COMPUTE_RETURN_ERROR_ON(!weights_info.are_reshaped() && weights->dimension(2) != input->dimension(2));
+    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(input, weights);
+
+    DataLayout data_layout = input->data_layout();
+    const int  idx_width   = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH);
+    const int  idx_height  = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT);
+    const int  idx_channel = get_data_layout_dimension_index(data_layout, DataLayoutDimension::CHANNEL);
+
+    ARM_COMPUTE_RETURN_ERROR_ON(!weights_info.are_reshaped() && weights->dimension(idx_channel) != input->dimension(idx_channel));
     ARM_COMPUTE_RETURN_ERROR_ON(weights->num_dimensions() > 4);
     ARM_COMPUTE_RETURN_ERROR_ON(weights_info.are_reshaped() && is_data_type_quantized_asymmetric(input->data_type()));
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(data_layout == DataLayout::NHWC && input->data_type() != DataType::F32, "NHWC is only supported for FP32 data type.");
 
     dt           = input->data_type();
     is_quantized = is_data_type_quantized_asymmetric(dt);
@@ -207,28 +206,32 @@
         ARM_COMPUTE_RETURN_ERROR_ON(biases->num_dimensions() > 1);
     }
 
+    // If we have 1x1 convolution and data layout is NHWC we can disable im2col
     append_bias          = (biases != nullptr) && (!is_quantized);
     are_weights_reshaped = weights_info.are_reshaped();
-    kernel_width         = (are_weights_reshaped) ? weights_info.kernel_size().first : weights->dimension(0);
-    kernel_height        = (are_weights_reshaped) ? weights_info.kernel_size().second : weights->dimension(1);
+    kernel_width         = (are_weights_reshaped) ? weights_info.kernel_size().first : weights->dimension(idx_width);
+    kernel_height        = (are_weights_reshaped) ? weights_info.kernel_size().second : weights->dimension(idx_height);
     mat_weights_cols     = weights->dimension(3);
-    mat_weights_rows     = weights->dimension(0) * weights->dimension(1) * weights->dimension(2) + (append_bias ? 1 : 0);
+    mat_weights_rows     = weights->dimension(idx_width) * weights->dimension(idx_height) * weights->dimension(idx_channel) + ((append_bias && !skip_im2col) ? 1 : 0);
+    skip_im2col          = (data_layout == DataLayout::NHWC && kernel_width == 1 && kernel_height == 1);
 
-    std::tie(conv_w, conv_h) = scaled_dimensions(input->dimension(0), input->dimension(1), kernel_width, kernel_height,
-                                                 conv_info);
+    std::tie(conv_w, conv_h) = scaled_dimensions(input->dimension(idx_width), input->dimension(idx_height), kernel_width, kernel_height,
+                                                 conv_info, dilation);
 
     // Check if its a "fully connected" convolution
     is_fully_connected_convolution = ((conv_w == 1) && (conv_h == 1));
     is_interleaved                 = (!is_fully_connected_convolution && !is_quantized);
+    is_activationlayer_enabled     = act_info.enabled();
 
     return Status{};
 }
 } // namespace
 
 NEGEMMConvolutionLayer::NEGEMMConvolutionLayer(const std::shared_ptr<IMemoryManager> &memory_manager)
-    : _memory_group(memory_manager), _input_im2col_kernel(), _input_interleave_kernel(), _reshape_weights(), _mm_kernel(), _mm_optimised_kernel(nullptr), _mm_gemmlowp(memory_manager),
-      _gemmlowp_output_stage(), _output_col2im_kernel(), _input_im2col_reshaped(), _input_interleaved_reshaped(), _weights_reshaped(), _gemm_output(), _tmp_output(), _workspace(), _append_bias(false),
-      _is_fully_connected_convolution(false), _are_weights_reshaped(false), _is_quantized(false), _is_interleaved(false)
+    : _asm_glue(), _memory_group(memory_manager), _input_im2col_kernel(), _input_interleave_kernel(), _reshape_weights(), _mm_kernel(), _mm_gemmlowp(memory_manager), _gemmlowp_output_stage(),
+      _output_col2im_kernel(), _activationlayer_function(), _add_bias_kernel(), _original_weights(nullptr), _input_im2col_reshaped(), _input_interleaved_reshaped(), _weights_reshaped(), _gemm_output(),
+      _tmp_output(), _workspace(), _B_pretransposed(), _data_layout(DataLayout::NCHW), _append_bias(false), _is_fully_connected_convolution(false), _are_weights_reshaped(false), _is_quantized(false),
+      _is_interleaved(false), _is_activationlayer_enabled(false), _skip_im2col(false)
 {
 }
 
@@ -256,26 +259,8 @@
     }
 }
 
-void NEGEMMConvolutionLayer::configure_asm_mm(const struct CPUInfo &ci, int M, int N, int K)
-{
-    ARM_COMPUTE_UNUSED(ci);
-    ARM_COMPUTE_UNUSED(M);
-    ARM_COMPUTE_UNUSED(N);
-    ARM_COMPUTE_UNUSED(K);
-#if defined(__arm__) || defined(__aarch64__)
-#if defined(__arm__)
-    GemmInterleaved<sgemm_8x6, float, float> gemm(&ci, M, N, K, false, false);
-#elif defined(__aarch64__)
-    GemmInterleaved<sgemm_12x8, float, float> gemm(&ci, M, N, K, false, false);
-#endif /* defined(__arm__) || defined(__aarch64__) */
-
-    constexpr size_t alignment = 4096;
-    _workspace.allocator()->init(TensorInfo(TensorShape{ (gemm.get_working_size() + alignment - 1) * NEScheduler::get().num_threads() }, 1, DataType::U8));
-    _memory_group.manage(&_workspace);
-#endif /* defined(__arm__) || defined(__aarch64__) */
-}
-
-void NEGEMMConvolutionLayer::configure(const ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const PadStrideInfo &conv_info, const WeightsInfo &weights_info)
+void NEGEMMConvolutionLayer::configure(const ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const PadStrideInfo &conv_info, const WeightsInfo &weights_info,
+                                       const Size2D &dilation, const ActivationLayerInfo &act_info)
 {
     // Perform validate step
     ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output);
@@ -288,45 +273,35 @@
     unsigned int conv_w           = 0;
     unsigned int conv_h           = 0;
 
-    Status status = validate_and_initialize_values(input->info(), weights->info(), (biases == nullptr) ? nullptr : biases->info(), conv_info, weights_info, dt, _append_bias, _are_weights_reshaped,
+    _data_layout           = input->info()->data_layout();
+    const bool is_nhwc     = _data_layout == DataLayout::NHWC;
+    const int  idx_width   = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::WIDTH);
+    const int  idx_height  = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::HEIGHT);
+    const int  idx_channel = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::CHANNEL);
+
+    Status status = validate_and_initialize_values(input->info(), weights->info(), (biases == nullptr) ? nullptr : biases->info(), conv_info, weights_info, act_info, dt, _append_bias, _skip_im2col,
+                                                   _are_weights_reshaped,
                                                    kernel_width, kernel_height,
-                                                   _is_fully_connected_convolution, _is_interleaved, _is_quantized,
-                                                   mat_weights_cols, mat_weights_rows, conv_w, conv_h);
+                                                   _is_fully_connected_convolution, _is_interleaved, _is_quantized, _is_activationlayer_enabled,
+                                                   mat_weights_cols, mat_weights_rows, conv_w, conv_h, dilation);
 
     ARM_COMPUTE_ERROR_THROW_ON(status);
 
+    _original_weights                       = weights;
     const unsigned int fixed_point_position = input->info()->fixed_point_position();
     const ITensor     *biases_to_use        = (_append_bias) ? biases : nullptr;
 
-#if defined(__arm__)
-    if(NEScheduler::get().cpu_info().CPU == CPUTarget::ARMV7 && dt == DataType::F32)
-    {
-        _mm_optimised_kernel = support::cpp14::make_unique<NEGEMMAArch32Kernel>();
-    }
-#elif defined(__aarch64__)
-    if(NEScheduler::get().cpu_info().CPU >= CPUTarget::ARMV8 && dt == DataType::F32)
-    {
-        _mm_optimised_kernel = support::cpp14::make_unique<NEGEMMAArch64Kernel>();
-    }
-#endif /* defined(__arm__) || defined(__aarch64__) */
+    bool run_optimised = dt == DataType::F32;
 
     // Reshape weights if needed
-    if(_mm_optimised_kernel != nullptr)
+    if(run_optimised)
     {
-        if(_are_weights_reshaped)
-        {
-            mat_weights_cols = weights_info.num_kernels();
-            mat_weights_rows = weights->info()->dimension(1);
-        }
-        else
-        {
-            TensorShape reshaped_weights_shape{ mat_weights_cols, mat_weights_rows };
+        TensorShape reshaped_weights_shape{ mat_weights_cols, mat_weights_rows };
 
-            // Create tensor to store the reshaped weights
-            _weights_reshaped.allocator()->init(TensorInfo(reshaped_weights_shape, 1, dt, fixed_point_position));
-            _reshape_weights.configure(weights, biases, &_weights_reshaped, false /* 1xW transpose */);
-            weights = &_weights_reshaped;
-        }
+        // Create tensor to store the reshaped weights
+        _weights_reshaped.allocator()->init(TensorInfo(reshaped_weights_shape, 1, dt, fixed_point_position));
+        _reshape_weights.configure(weights, biases, &_weights_reshaped, false /* 1xW transpose */);
+        weights = &_weights_reshaped;
     }
     else
     {
@@ -335,12 +310,12 @@
             if(_is_fully_connected_convolution || _is_quantized)
             {
                 mat_weights_cols = weights_info.num_kernels();
-                mat_weights_rows = weights->info()->dimension(1);
+                mat_weights_rows = weights->info()->dimension(idx_height);
             }
             else
             {
                 mat_weights_cols = weights_info.num_kernels();
-                mat_weights_rows = weights_info.kernel_size().first * weights_info.kernel_size().second * input->info()->dimension(2) + (_append_bias ? 1 : 0);
+                mat_weights_rows = weights_info.kernel_size().first * weights_info.kernel_size().second * input->info()->dimension(idx_channel) + (_append_bias ? 1 : 0);
             }
         }
         else
@@ -366,66 +341,56 @@
         }
     }
 
-    // Create tensor to store im2col reshaped inputs
-    const unsigned int mat_input_cols = mat_weights_rows;
-    const unsigned int mat_input_rows = conv_w * conv_h;
-
-    TensorShape shape_im2col(input->info()->tensor_shape());
-    shape_im2col.set(0, mat_input_cols);
-    shape_im2col.set(1, mat_input_rows);
-    shape_im2col.set(2, 1);
-    _input_im2col_reshaped.allocator()->init(input->info()->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(shape_im2col));
-    _memory_group.manage(&_input_im2col_reshaped);
-
-    // Create tensor (interleave) to prepare input tensor for GEMM
-    if(!_is_fully_connected_convolution && _mm_optimised_kernel == nullptr)
+    // In case we skip im2col we have to add bias
+    if(!_skip_im2col)
     {
-        TensorShape shape_interleaved(shape_im2col);
-        shape_interleaved.set(0, shape_interleaved.x() * 4);
-        shape_interleaved.set(1, std::ceil(shape_interleaved.y() / 4.f));
-        _input_interleaved_reshaped.allocator()->init(input->info()->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(shape_interleaved));
-        _memory_group.manage(&_input_interleaved_reshaped);
+        const unsigned int mat_input_cols = mat_weights_rows;
+        const unsigned int mat_input_rows = conv_w * conv_h;
+
+        // Create tensor to store im2col reshaped inputs
+        TensorShape shape_im2col(input->info()->tensor_shape());
+        shape_im2col.set(0, mat_input_cols);
+        shape_im2col.set(1, mat_input_rows);
+        shape_im2col.set(2, 1);
+        _input_im2col_reshaped.allocator()->init(input->info()->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(shape_im2col));
+        _memory_group.manage(&_input_im2col_reshaped);
+
+        // Create tensor (interleave) to prepare input tensor for GEMM
+        if(!_is_fully_connected_convolution && !run_optimised && _is_interleaved)
+        {
+            TensorShape shape_interleaved(shape_im2col);
+            shape_interleaved.set(idx_width, shape_interleaved.x() * 4);
+            shape_interleaved.set(idx_height, std::ceil(shape_interleaved[idx_height] / 4.f));
+            _input_interleaved_reshaped.allocator()->init(input->info()->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(shape_interleaved));
+            _memory_group.manage(&_input_interleaved_reshaped);
+        }
+
+        // Create GEMM output tensor
+        TensorShape shape_gemm(_input_im2col_reshaped.info()->tensor_shape());
+        shape_gemm.set(0, mat_weights_cols);
+        shape_gemm.set(1, mat_input_rows);
+        const DataType gemm_data_type = _is_quantized ? DataType::S32 : dt;
+        // GEMM output should be S32 for acquiring raw integer accumulator without quantized postprocessing for quantized asymmetric input.
+        TensorInfo info_gemm(shape_gemm, 1, gemm_data_type, input->info()->fixed_point_position());
+        info_gemm.set_quantization_info(output->info()->quantization_info());
+        _gemm_output.allocator()->init(info_gemm);
+
+        // Configure im2col
+        _input_im2col_kernel.configure(input, &_input_im2col_reshaped, Size2D(kernel_width, kernel_height), conv_info, _append_bias, false, false, dilation);
+    }
+    else if(_append_bias)
+    {
+        // Configure add bias kernel
+        _add_bias_kernel.configure(output, biases, output, ConvertPolicy::SATURATE);
     }
 
-    // Create GEMM output tensor
-    TensorShape shape_gemm(_input_im2col_reshaped.info()->tensor_shape());
-    shape_gemm.set(0, mat_weights_cols);
-    shape_gemm.set(1, mat_input_rows);
-    const DataType gemm_data_type = _is_quantized ? DataType::S32 : dt;
-    // GEMM output should be S32 for acquiring raw integer accumulator without quantized postprocessing for quantized asymmetric input.
-    TensorInfo info_gemm(shape_gemm, 1, gemm_data_type, input->info()->fixed_point_position());
-    info_gemm.set_quantization_info(output->info()->quantization_info());
-    _gemm_output.allocator()->init(info_gemm);
-    _memory_group.manage(&_gemm_output);
-
-    // Configure kernels
-    // Configure im2col
-    _input_im2col_kernel.configure(input, &_input_im2col_reshaped, Size2D(kernel_width, kernel_height), conv_info, _append_bias);
-
     // Configure matrix multiply
-    if(_mm_optimised_kernel != nullptr)
+    if(run_optimised)
     {
-        struct CPUInfo ci = NEScheduler::get().cpu_info();
-
-        const int M = _gemm_output.info()->tensor_shape().y();
-        const int N = _gemm_output.info()->tensor_shape().x();
-        const int K = _input_im2col_reshaped.info()->tensor_shape().x();
-
-#if defined(__aarch64__)
-        if((N <= 128) && (K <= 128))
+        if(!setup_assembly_kernel(_skip_im2col ? input : &_input_im2col_reshaped, weights, is_nhwc ? output : &_gemm_output, 1.f, 0.f, true, _workspace, _B_pretransposed, _memory_group, _asm_glue))
         {
-            _mm_optimised_kernel = support::cpp14::make_unique<NEGEMMAArch64NativeKernel>();
+            ARM_COMPUTE_ERROR("setup_assembly_kernel failed.");
         }
-        else
-#endif /* defined(__aarch64__) */
-        {
-            configure_asm_mm(ci, M, N, K);
-        }
-
-        // Configure matrix multiplication kernel
-        _mm_optimised_kernel->configure(&_input_im2col_reshaped, weights, &_gemm_output, &_workspace);
-
-        _workspace.allocator()->allocate();
     }
     else
     {
@@ -435,8 +400,8 @@
             _input_interleave_kernel.configure(&_input_im2col_reshaped, &_input_interleaved_reshaped);
 
             // Configure GEMM
-            configure_mm(&_input_interleaved_reshaped, weights, &_gemm_output, _is_interleaved, GEMMReshapeInfo(_input_im2col_reshaped.info()->dimension(1), 0 /* no transpose */,
-                                                                                                                _input_im2col_reshaped.info()->dimension(0)));
+            configure_mm(&_input_interleaved_reshaped, weights, &_gemm_output, _is_interleaved, GEMMReshapeInfo(_input_im2col_reshaped.info()->dimension(idx_height), 0 /* no transpose */,
+                                                                                                                _input_im2col_reshaped.info()->dimension(idx_width)));
             _input_interleaved_reshaped.allocator()->allocate();
         }
         else
@@ -445,48 +410,63 @@
         }
     }
 
-    _input_im2col_reshaped.allocator()->allocate();
-
-    // Configure output stage for quantized case
-    if(_is_quantized)
+    if(!_skip_im2col)
     {
-        const QuantizationInfo output_quant_info = (output->info()->total_size() == 0) ? input->info()->quantization_info() : output->info()->quantization_info();
+        _input_im2col_reshaped.allocator()->allocate();
 
-        float multiplier = input->info()->quantization_info().scale * weights->info()->quantization_info().scale / output_quant_info.scale;
-        int   output_multiplier, output_shift;
-        quantization::calculate_quantized_multiplier_less_than_one(multiplier, &output_multiplier, &output_shift);
-        _memory_group.manage(&_tmp_output);
-        _gemmlowp_output_stage.configure(&_gemm_output, biases, &_tmp_output, output_multiplier, output_shift, output_quant_info.offset);
+        // Configure output stage for quantized case
+        if(_is_quantized)
+        {
+            const QuantizationInfo output_quant_info = (output->info()->total_size() == 0) ? input->info()->quantization_info() : output->info()->quantization_info();
+
+            float multiplier = input->info()->quantization_info().scale * weights->info()->quantization_info().scale / output_quant_info.scale;
+            int   output_multiplier, output_shift;
+            quantization::calculate_quantized_multiplier_less_than_one(multiplier, &output_multiplier, &output_shift);
+            _memory_group.manage(&_tmp_output);
+            _gemmlowp_output_stage.configure(&_gemm_output, biases, &_tmp_output, output_multiplier, output_shift, output_quant_info.offset);
+        }
+
+        // Configure Col2Im
+        if(!is_nhwc)
+        {
+            _output_col2im_kernel.configure(_is_quantized ? &_tmp_output : &_gemm_output, output, Size2D(conv_w, conv_h));
+        }
+
+        if(_is_quantized)
+        {
+            _tmp_output.allocator()->allocate();
+        }
+        _gemm_output.allocator()->allocate();
     }
 
-    // Configure Col2Im
-    _output_col2im_kernel.configure(_is_quantized ? &_tmp_output : &_gemm_output, output, Size2D(conv_w, conv_h));
-    if(_is_quantized)
-    {
-        _tmp_output.allocator()->allocate();
-    }
-    _gemm_output.allocator()->allocate();
-
-    ARM_COMPUTE_ERROR_ON_MSG((output->info()->dimension(0) != conv_w) || (output->info()->dimension(1) != conv_h), "Output shape does not match the expected one");
+    ARM_COMPUTE_ERROR_ON_MSG((output->info()->dimension(idx_width) != conv_w) || (output->info()->dimension(idx_height) != conv_h), "Output shape does not match the expected one");
 
     // Allocate intermediate tensor
     if(!_are_weights_reshaped)
     {
         _weights_reshaped.allocator()->allocate();
     }
+
+    //Configure Activation Layer
+    if(_is_activationlayer_enabled)
+    {
+        _activationlayer_function.configure(output, nullptr, act_info);
+    }
 }
 
 Status NEGEMMConvolutionLayer::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info,
-                                        const WeightsInfo &weights_info)
+                                        const WeightsInfo &weights_info, const Size2D &dilation, const ActivationLayerInfo &act_info)
 {
     ARM_COMPUTE_UNUSED(output);
 
     DataType     dt{};
     bool         append_bias{};
+    bool         skip_im2col{};
     bool         are_weights_reshaped{};
     bool         is_fully_connected_convolution{};
     bool         is_interleaved{};
     bool         is_quantized{};
+    bool         is_activationlayer_enabled{};
     unsigned int kernel_width     = 0;
     unsigned int kernel_height    = 0;
     unsigned int mat_weights_cols = 0;
@@ -494,9 +474,14 @@
     unsigned int conv_w           = 0;
     unsigned int conv_h           = 0;
 
-    Status status = validate_and_initialize_values(input, weights, biases, conv_info, weights_info, dt, append_bias, are_weights_reshaped, kernel_width, kernel_height,
-                                                   is_fully_connected_convolution, is_interleaved, is_quantized, mat_weights_cols, mat_weights_rows,
-                                                   conv_w, conv_h);
+    const DataLayout data_layout = input->data_layout();
+    const bool       is_nhwc     = data_layout == DataLayout::NHWC;
+    const int        idx_width   = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH);
+    const int        idx_height  = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT);
+
+    Status status = validate_and_initialize_values(input, weights, biases, conv_info, weights_info, act_info, dt, append_bias, skip_im2col, are_weights_reshaped, kernel_width, kernel_height,
+                                                   is_fully_connected_convolution, is_interleaved, is_quantized, is_activationlayer_enabled, mat_weights_cols, mat_weights_rows,
+                                                   conv_w, conv_h, dilation);
 
     const Size2D kernel_weights = Size2D(kernel_width, kernel_height);
 
@@ -505,68 +490,11 @@
     std::unique_ptr<ITensorInfo> reshaped_weights = weights->clone();
     bool                         optimised_kernel = false;
 
-#if defined(__arm__)
-    if(NEScheduler::get().cpu_info().CPU == CPUTarget::ARMV7 && dt == DataType::F32)
+    if(dt == DataType::F32)
     {
         optimised_kernel = true;
     }
-#elif defined(__aarch64__)
-    if(NEScheduler::get().cpu_info().CPU >= CPUTarget::ARMV8 && dt == DataType::F32)
-    {
-        optimised_kernel = true;
-    }
-#endif /* defined(__arm__) || defined(__aarch64__) */
 
-    // Reshape weights if needed
-    if(optimised_kernel)
-    {
-        if(are_weights_reshaped)
-        {
-            mat_weights_cols = weights_info.num_kernels();
-            mat_weights_rows = weights->dimension(1);
-        }
-        else
-        {
-            TensorShape reshaped_weights_shape{ mat_weights_cols, mat_weights_rows };
-
-            // Create tensor to store the reshaped weights
-            reshaped_weights->set_tensor_shape(get_reshaped_weights_shape_conv(weights, append_bias, is_fully_connected_convolution));
-            ARM_COMPUTE_RETURN_ON_ERROR(NEConvolutionLayerReshapeWeights::validate(weights, biases, reshaped_weights.get(), !is_fully_connected_convolution /* 1xW transpose */));
-            weights = reshaped_weights.get();
-        }
-    }
-    else
-    {
-        if(are_weights_reshaped)
-        {
-            const unsigned int transpose_width = 16 / input->element_size();
-            mat_weights_cols                   = weights_info.num_kernels();
-            mat_weights_rows                   = weights->dimension(0) / transpose_width + (append_bias ? 1 : 0);
-        }
-        else
-        {
-            TensorShape reshaped_weights_shape;
-
-            if(is_fully_connected_convolution || is_quantized)
-            {
-                reshaped_weights_shape = TensorShape{ mat_weights_cols, mat_weights_rows };
-            }
-            else
-            {
-                // Create tensor to store transposed weights
-                const float transpose_width = 16.0f / input->element_size();
-                reshaped_weights_shape      = TensorShape{ mat_weights_rows *static_cast<unsigned int>(transpose_width),
-                                                           static_cast<unsigned int>(std::ceil(mat_weights_cols / transpose_width)) };
-            }
-
-            // Create tensor to store the reshaped weights
-            reshaped_weights->set_tensor_shape(get_reshaped_weights_shape_conv(weights, append_bias, is_fully_connected_convolution));
-            ARM_COMPUTE_RETURN_ON_ERROR(NEConvolutionLayerReshapeWeights::validate(weights, biases, reshaped_weights.get(), !is_fully_connected_convolution /* 1xW transpose */));
-            weights = reshaped_weights.get();
-        }
-    }
-
-    // Validate im2col
     const unsigned int mat_input_cols = mat_weights_rows;
     const unsigned int mat_input_rows = conv_w * conv_h;
     TensorShape        shape_im2col   = input->tensor_shape();
@@ -574,7 +502,17 @@
     shape_im2col.set(1, mat_input_rows);
     shape_im2col.set(2, 1);
     TensorInfo im2_col_info = input->clone()->set_tensor_shape(shape_im2col);
-    ARM_COMPUTE_RETURN_ON_ERROR(NEIm2ColKernel::validate(input, &im2_col_info, kernel_weights, conv_info, append_bias, false));
+
+    if(!skip_im2col)
+    {
+        // Validate im2col
+        ARM_COMPUTE_RETURN_ON_ERROR(NEIm2ColKernel::validate(input, &im2_col_info, kernel_weights, conv_info, append_bias, false, false, dilation));
+    }
+    else if(append_bias)
+    {
+        // Validate add bias kernel
+        ARM_COMPUTE_RETURN_ON_ERROR(NEArithmeticAdditionKernel::validate(output, biases, output, ConvertPolicy::SATURATE));
+    }
 
     // Create GEMM output tensor
     TensorShape shape_gemm(im2_col_info.tensor_shape());
@@ -582,19 +520,63 @@
     shape_gemm.set(1, mat_input_rows);
     TensorInfo gemm_output_info = input->clone()->set_tensor_shape(shape_gemm);
 
-    // Validate GEMM interleave and multiply
-    if(is_interleaved)
+    // Reshape weights if needed
+    if(optimised_kernel)
     {
-        TensorShape shape_interleaved = shape_im2col;
-        shape_interleaved.set(0, shape_interleaved.x() * 4);
-        shape_interleaved.set(1, std::ceil(shape_interleaved.y() / 4.f));
-        TensorInfo input_interleaved_info = input->clone()->set_tensor_shape(shape_interleaved);
-        ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMInterleave4x4Kernel::validate(&im2_col_info, &input_interleaved_info));
-        ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMMatrixMultiplyKernel::validate(&input_interleaved_info, weights, &gemm_output_info, 1.f, is_interleaved, GEMMReshapeInfo()));
+        ARM_COMPUTE_RETURN_ERROR_ON(are_weights_reshaped);
+
+        // Create tensor to store the reshaped weights
+        reshaped_weights->set_tensor_shape(get_reshaped_weights_shape_conv(weights, append_bias, is_fully_connected_convolution));
+        ARM_COMPUTE_RETURN_ON_ERROR(NEConvolutionLayerReshapeWeights::validate(weights, biases, reshaped_weights.get(), !is_fully_connected_convolution /* 1xW transpose */));
     }
-    else
+    else if(!is_quantized)
     {
-        ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMMatrixMultiplyKernel::validate(&im2_col_info, weights, &gemm_output_info, 1.f, is_interleaved, GEMMReshapeInfo()));
+        TensorShape reshaped_weights_shape;
+
+        if(is_fully_connected_convolution || is_quantized)
+        {
+            reshaped_weights_shape = TensorShape{ mat_weights_cols, mat_weights_rows };
+        }
+        else
+        {
+            // Create tensor to store transposed weights
+            const float transpose_width = 16.0f / input->element_size();
+            reshaped_weights_shape      = TensorShape{ mat_weights_rows *static_cast<unsigned int>(transpose_width),
+                                                       static_cast<unsigned int>(std::ceil(mat_weights_cols / transpose_width)) };
+        }
+
+        // Create tensor to store the reshaped weights
+        reshaped_weights->set_tensor_shape(get_reshaped_weights_shape_conv(weights, append_bias, is_fully_connected_convolution));
+        ARM_COMPUTE_RETURN_ON_ERROR(NEConvolutionLayerReshapeWeights::validate(weights, biases, reshaped_weights.get(), !is_fully_connected_convolution /* 1xW transpose */));
+        weights = reshaped_weights.get();
+
+        // Validate GEMM interleave and multiply
+        if(is_interleaved)
+        {
+            TensorShape shape_interleaved = shape_im2col;
+            shape_interleaved.set(idx_width, shape_interleaved.x() * 4);
+            shape_interleaved.set(idx_height, std::ceil(shape_interleaved.y() / 4.f));
+            TensorInfo input_interleaved_info = input->clone()->set_tensor_shape(shape_interleaved);
+            ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMInterleave4x4Kernel::validate(&im2_col_info, &input_interleaved_info));
+            ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMMatrixMultiplyKernel::validate(&input_interleaved_info, weights, &gemm_output_info, 1.f, is_interleaved, GEMMReshapeInfo(shape_im2col[1],            // m
+                                                                             weights->tensor_shape()[0], // n
+                                                                             shape_im2col[0]) /* k */));
+        }
+        else
+        {
+            ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMMatrixMultiplyKernel::validate(&im2_col_info, weights, &gemm_output_info, 1.f, is_interleaved, GEMMReshapeInfo()));
+        }
+    }
+    if(!is_nhwc)
+    {
+        ARM_COMPUTE_RETURN_ON_ERROR(NECol2ImKernel::validate(&gemm_output_info, output, Size2D(conv_w, conv_h)));
+    }
+
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG((output->dimension(idx_width) != conv_w) || (output->dimension(idx_height) != conv_h), "Output shape does not match the expected one");
+
+    if(act_info.enabled())
+    {
+        ARM_COMPUTE_RETURN_ON_ERROR(NEActivationLayer::validate(output, nullptr, act_info));
     }
 
     return Status{};
@@ -605,19 +587,33 @@
     // Run weights reshaping (Runs once for every configure)
     if(!_are_weights_reshaped)
     {
+        ARM_COMPUTE_ERROR_ON(!_original_weights->is_used());
+
         _are_weights_reshaped = true;
         _reshape_weights.run();
+
+        // Mark original weights tensor as unused
+        _original_weights->mark_as_unused();
     }
 
     _memory_group.acquire();
 
-    // Run input reshaping
-    NEScheduler::get().schedule(&_input_im2col_kernel, Window::DimY);
+    if(!_skip_im2col)
+    {
+        // Run input reshaping
+        unsigned int _y_dim = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::HEIGHT);
+        NEScheduler::get().schedule(&_input_im2col_kernel, _y_dim);
+    }
 
     // Runs matrix multiply on reshaped matrices
-    if(_mm_optimised_kernel != nullptr)
+    if(_asm_glue._optimised_kernel != nullptr)
     {
-        NEScheduler::get().schedule(_mm_optimised_kernel.get(), Window::DimY);
+        _asm_glue.run();
+        // Release weights in case buffer is pretransposed
+        if(!_weights_reshaped.is_used())
+        {
+            _weights_reshaped.allocator()->free();
+        }
     }
     else
     {
@@ -638,6 +634,11 @@
         }
     }
 
+    if(_skip_im2col && _append_bias)
+    {
+        NEScheduler::get().schedule(&_add_bias_kernel, Window::DimY);
+    }
+
     // Run output stage for quantized case
     if(_is_quantized)
     {
@@ -645,7 +646,15 @@
     }
 
     // Reshape output matrix
-    NEScheduler::get().schedule(&_output_col2im_kernel, Window::DimY);
+    if(_data_layout == DataLayout::NCHW)
+    {
+        NEScheduler::get().schedule(&_output_col2im_kernel, Window::DimY);
+    }
+
+    if(_is_activationlayer_enabled)
+    {
+        _activationlayer_function.run();
+    }
 
     _memory_group.release();
 }