arm_compute v18.01

Change-Id: I9bfa178c2e38bfd5fc812e62aab6760d87748e05
diff --git a/src/runtime/NEON/functions/NEConvolution.cpp b/src/runtime/NEON/functions/NEConvolution.cpp
index f10ffa6..b84dfd3 100644
--- a/src/runtime/NEON/functions/NEConvolution.cpp
+++ b/src/runtime/NEON/functions/NEConvolution.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016, 2017 ARM Limited.
+ * Copyright (c) 2016, 2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -54,7 +54,8 @@
 }
 
 template <unsigned int matrix_size>
-void NEConvolutionSquare<matrix_size>::configure(ITensor *input, ITensor *output, const int16_t *conv, uint32_t scale, BorderMode border_mode, uint8_t constant_border_value)
+void NEConvolutionSquare<matrix_size>::configure(ITensor *input, ITensor *output, const int16_t *conv, uint32_t scale, BorderMode border_mode,
+                                                 uint8_t constant_border_value)
 {
     ARM_COMPUTE_ERROR_ON(conv == nullptr);
     ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8);
diff --git a/src/runtime/NEON/functions/NEConvolutionLayer.cpp b/src/runtime/NEON/functions/NEConvolutionLayer.cpp
index 25c639f..8f7d940 100644
--- a/src/runtime/NEON/functions/NEConvolutionLayer.cpp
+++ b/src/runtime/NEON/functions/NEConvolutionLayer.cpp
@@ -44,6 +44,16 @@
 
 namespace arm_compute
 {
+namespace
+{
+TensorShape get_reshaped_weights_shape(const ITensorInfo *weights, bool has_bias)
+{
+    const unsigned int mat_weights_cols = weights->dimension(3);
+    const unsigned int mat_weights_rows = weights->dimension(0) * weights->dimension(1) * weights->dimension(2) + (has_bias ? 1 : 0);
+    return TensorShape(mat_weights_cols, mat_weights_rows);
+}
+} // namespace
+
 NEConvolutionLayerReshapeWeights::NEConvolutionLayerReshapeWeights(std::shared_ptr<IMemoryManager> memory_manager)
     : _memory_group(std::move(memory_manager)), _weights_reshape_kernel(), _weights_transposed_kernel(), _weights_reshaped(), _transpose1xW(false)
 {
@@ -51,18 +61,12 @@
 
 void NEConvolutionLayerReshapeWeights::configure(const ITensor *weights, const ITensor *biases, ITensor *output, bool transpose1xW)
 {
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(weights, 1, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32);
-    ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(weights, output);
-    ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(weights, output);
-    ARM_COMPUTE_ERROR_ON(weights->info()->num_dimensions() > 4);
-
-    if(biases != nullptr)
-    {
-        ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(weights, biases);
-        ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(weights, biases);
-        ARM_COMPUTE_ERROR_ON(biases->info()->dimension(0) != weights->info()->dimension(3));
-        ARM_COMPUTE_ERROR_ON(biases->info()->num_dimensions() > 1);
-    }
+    // Perform validation step
+    ARM_COMPUTE_ERROR_ON_NULLPTR(weights, output);
+    ARM_COMPUTE_ERROR_THROW_ON(NEConvolutionLayerReshapeWeights::validate(weights->info(),
+                                                                          (biases != nullptr) ? biases->info() : nullptr,
+                                                                          output->info(),
+                                                                          transpose1xW));
 
     // Check if bias are present, if yes they will be embedded to the weights matrix
     const bool _has_bias = (biases != nullptr);
@@ -72,10 +76,7 @@
     if(transpose1xW)
     {
         // Create tensor to store the reshaped weights
-        const unsigned int mat_weights_cols = weights->info()->dimension(3);
-        const unsigned int mat_weights_rows = weights->info()->dimension(0) * weights->info()->dimension(1) * weights->info()->dimension(2) + (_has_bias ? 1 : 0);
-        TensorShape        shape_wr(mat_weights_cols, mat_weights_rows);
-        TensorInfo         info_wr(shape_wr, 1, weights->info()->data_type(), weights->info()->fixed_point_position());
+        TensorInfo info_wr = weights->info()->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(get_reshaped_weights_shape(weights->info(), _has_bias));
 
         _weights_reshaped.allocator()->init(info_wr);
         _memory_group.manage(&_weights_reshaped);
@@ -91,6 +92,46 @@
     }
 }
 
+Status NEConvolutionLayerReshapeWeights::validate(const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, bool transpose1xW)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(weights, 1, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32);
+    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(weights, output);
+    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(weights, output);
+    ARM_COMPUTE_RETURN_ERROR_ON(weights->num_dimensions() > 4);
+
+    if(biases != nullptr)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(weights, biases);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(weights, biases);
+        ARM_COMPUTE_RETURN_ERROR_ON(biases->dimension(0) != weights->dimension(3));
+        ARM_COMPUTE_RETURN_ERROR_ON(biases->num_dimensions() > 1);
+    }
+
+    // Check if bias are present, if yes they will be embedded to the weights matrix
+    const bool has_bias = (biases != nullptr);
+
+    // Checks performed when biases are present
+    if(has_bias)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(weights, biases);
+        ARM_COMPUTE_RETURN_ERROR_ON(biases->dimension(0) != weights->dimension(3));
+        ARM_COMPUTE_RETURN_ERROR_ON(biases->num_dimensions() > 1);
+    }
+
+    if(transpose1xW)
+    {
+        TensorInfo weights_reshaped = weights->clone()->set_tensor_shape(get_reshaped_weights_shape(weights, has_bias));
+        ARM_COMPUTE_RETURN_ON_ERROR(NEWeightsReshapeKernel::validate(weights, biases, &weights_reshaped));
+        ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMTranspose1xWKernel::validate(&weights_reshaped, output));
+    }
+    else
+    {
+        ARM_COMPUTE_RETURN_ON_ERROR(NEWeightsReshapeKernel::validate(weights, biases, output));
+    }
+
+    return Status{};
+}
+
 void NEConvolutionLayerReshapeWeights::run()
 {
     _memory_group.acquire();
@@ -105,6 +146,62 @@
     _memory_group.release();
 }
 
+namespace
+{
+TensorShape get_reshaped_weights_shape_conv(const ITensorInfo *weights, bool has_bias, bool is_fully_connected_convolution)
+{
+    unsigned int mat_weights_cols = weights->dimension(3);
+    unsigned int mat_weights_rows = weights->dimension(0) * weights->dimension(1) * weights->dimension(2) + (has_bias ? 1 : 0);
+
+    if(is_fully_connected_convolution)
+    {
+        // Create tensor to store the reshaped weights
+        return TensorShape(mat_weights_cols, mat_weights_rows);
+    }
+    else
+    {
+        // Create tensor to store transposed weights
+        const float transpose_width = 16.0f / weights->element_size();
+        return TensorShape(mat_weights_rows * static_cast<unsigned int>(transpose_width), static_cast<unsigned int>(std::ceil(mat_weights_cols / transpose_width)));
+    }
+}
+
+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 &has_bias,
+                                      bool &are_weights_reshaped, unsigned int &kernel_width, unsigned int &kernel_height, bool &is_fully_connected_convolution, unsigned int &mat_weights_cols, unsigned int &mat_weights_rows,
+                                      unsigned int &conv_w, unsigned int &conv_h)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, 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(weights->num_dimensions() > 4);
+
+    if(biases != nullptr)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, biases);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, biases);
+        ARM_COMPUTE_RETURN_ERROR_ON(!weights_info.are_reshaped() && biases->dimension(0) != weights->dimension(3));
+        ARM_COMPUTE_RETURN_ERROR_ON(biases->num_dimensions() > 1);
+    }
+
+    dt                   = input->data_type();
+    has_bias             = (biases != nullptr);
+    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);
+    mat_weights_cols     = weights->dimension(3);
+    mat_weights_rows     = weights->dimension(0) * weights->dimension(1) * weights->dimension(2) + (has_bias ? 1 : 0);
+
+    std::tie(conv_w, conv_h) = scaled_dimensions(input->dimension(0), input->dimension(1), kernel_width, kernel_height,
+                                                 conv_info);
+
+    is_fully_connected_convolution = ((conv_w == 1) && (conv_h == 1));
+
+    return Status{};
+}
+} // namespace
+
 NEConvolutionLayer::NEConvolutionLayer(std::shared_ptr<IMemoryManager> memory_manager)
     : _memory_group(std::move(memory_manager)), _input_im2col_kernel(), _input_interleave_kernel(), _reshape_weights(), _mm_kernel(), _mm_optimised_kernel(nullptr), _output_col2im_kernel(),
       _input_im2col_reshaped(), _input_interleaved_reshaped(), _weights_reshaped(), _gemm_output(), _workspace(), _has_bias(false), _is_fully_connected_convolution(false), _are_weights_reshaped(false)
@@ -113,42 +210,25 @@
 
 void NEConvolutionLayer::configure(const ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const PadStrideInfo &conv_info, const WeightsInfo &weights_info)
 {
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32);
-    ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights);
-    ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, weights);
-    ARM_COMPUTE_ERROR_ON(!weights_info.are_reshaped() && weights->info()->dimension(2) != input->info()->dimension(2));
-    ARM_COMPUTE_ERROR_ON(weights->info()->num_dimensions() > 4);
+    // Perform validate step
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output);
 
-    if(biases != nullptr)
-    {
-        ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, biases);
-        ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, biases);
-        ARM_COMPUTE_ERROR_ON(!weights_info.are_reshaped() && biases->info()->dimension(0) != weights->info()->dimension(3));
-        ARM_COMPUTE_ERROR_ON(biases->info()->num_dimensions() > 1);
-    }
+    DataType     dt{};
+    unsigned int kernel_width     = 0;
+    unsigned int kernel_height    = 0;
+    unsigned int mat_weights_cols = 0;
+    unsigned int mat_weights_rows = 0;
+    unsigned int conv_w           = 0;
+    unsigned int conv_h           = 0;
 
-    const DataType dt                   = input->info()->data_type();
-    const int      fixed_point_position = input->info()->fixed_point_position();
+    Status status = validate_and_initialize_values(input->info(), weights->info(), (biases == nullptr) ? nullptr : biases->info(), conv_info, weights_info, dt, _has_bias, _are_weights_reshaped,
+                                                   kernel_width, kernel_height,
+                                                   _is_fully_connected_convolution,
+                                                   mat_weights_cols, mat_weights_rows, conv_w, conv_h);
 
-    _has_bias             = (biases != nullptr);
-    _are_weights_reshaped = weights_info.are_reshaped();
+    ARM_COMPUTE_ERROR_THROW_ON(status);
 
-    // Get parameters from conv_info
-    unsigned int stride_x = 0;
-    unsigned int stride_y = 0;
-    std::tie(stride_x, stride_y) = conv_info.stride();
-
-    // Get convolved dimensions
-    unsigned int conv_w = 0;
-    unsigned int conv_h = 0;
-
-    const unsigned int kernel_width  = (_are_weights_reshaped) ? weights_info.kernel_size().first : weights->info()->dimension(0);
-    const unsigned int kernel_height = (_are_weights_reshaped) ? weights_info.kernel_size().second : weights->info()->dimension(1);
-    std::tie(conv_w, conv_h) = scaled_dimensions(input->info()->dimension(0), input->info()->dimension(1), kernel_width, kernel_height,
-                                                 conv_info);
-
-    // Check if its a "fully connected" convolution, i.e. the output size is 1x1xnum_kernels
-    _is_fully_connected_convolution = ((conv_w == 1) && (conv_h == 1));
+    const unsigned int fixed_point_position = input->info()->fixed_point_position();
 
 #if defined(__arm__)
     if(NEScheduler::get().cpu_info().CPU == CPUTarget::ARMV7 && dt == DataType::F32)
@@ -162,9 +242,6 @@
     }
 #endif /* defined(__arm__) || defined(__aarch64__) */
 
-    unsigned int mat_weights_cols = weights->info()->dimension(3);
-    unsigned int mat_weights_rows = weights->info()->dimension(0) * weights->info()->dimension(1) * weights->info()->dimension(2) + (_has_bias ? 1 : 0);
-
     // Reshape weights if needed
     if(_mm_optimised_kernel != nullptr)
     {
@@ -230,7 +307,7 @@
     shape_im2col.set(0, mat_input_cols);
     shape_im2col.set(1, mat_input_rows);
     shape_im2col.set(2, 1);
-    _input_im2col_reshaped.allocator()->init(TensorInfo(shape_im2col, 1, dt, fixed_point_position));
+    _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
@@ -239,7 +316,7 @@
         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(TensorInfo(shape_interleaved, 1, dt, fixed_point_position));
+        _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);
     }
 
@@ -247,7 +324,7 @@
     TensorShape shape_gemm(_input_im2col_reshaped.info()->tensor_shape());
     shape_gemm.set(0, mat_weights_cols);
     shape_gemm.set(1, mat_input_rows);
-    _gemm_output.allocator()->init(TensorInfo(shape_gemm, 1, dt, fixed_point_position));
+    _gemm_output.allocator()->init(_input_im2col_reshaped.info()->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(shape_gemm));
     _memory_group.manage(&_gemm_output);
 
     // Configure kernels
@@ -273,14 +350,7 @@
         _memory_group.manage(&_workspace);
 
         // Configure matrix multiplication kernel
-        if(_is_fully_connected_convolution)
-        {
-            _mm_optimised_kernel->configure(&_input_im2col_reshaped, weights, &_gemm_output, &_workspace, 1.f, 0.f);
-        }
-        else
-        {
-            _mm_optimised_kernel->configure(&_input_im2col_reshaped, weights, &_gemm_output, &_workspace);
-        }
+        _mm_optimised_kernel->configure(&_input_im2col_reshaped, weights, &_gemm_output, &_workspace);
 
         _workspace.allocator()->allocate();
     }
@@ -303,8 +373,6 @@
     _output_col2im_kernel.configure(&_gemm_output, output, Size2D(conv_w, conv_h));
     _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");
-
     // Allocate intermediate tensor
     if(!_are_weights_reshaped)
     {
@@ -312,6 +380,128 @@
     }
 }
 
+Status NEConvolutionLayer::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info,
+                                    const WeightsInfo &weights_info)
+{
+    DataType     dt{};
+    bool         has_bias{};
+    bool         are_weights_reshaped{};
+    bool         is_fully_connected_convolution{};
+    unsigned int kernel_width     = 0;
+    unsigned int kernel_height    = 0;
+    unsigned int mat_weights_cols = 0;
+    unsigned int mat_weights_rows = 0;
+    unsigned int conv_w           = 0;
+    unsigned int conv_h           = 0;
+
+    Status status = validate_and_initialize_values(input, weights, biases, conv_info, weights_info, dt, has_bias, are_weights_reshaped, kernel_width, kernel_height,
+                                                   is_fully_connected_convolution, mat_weights_cols, mat_weights_rows,
+                                                   conv_w, conv_h);
+
+    ARM_COMPUTE_RETURN_ON_ERROR(status);
+
+    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)
+    {
+        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, has_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 + (has_bias ? 1 : 0);
+        }
+        else
+        {
+            TensorShape reshaped_weights_shape;
+
+            if(is_fully_connected_convolution)
+            {
+                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, has_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();
+    shape_im2col.set(0, mat_input_cols);
+    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, Size2D(weights->dimension(0), weights->dimension(1)), conv_info, has_bias));
+
+    // Create GEMM output tensor
+    TensorShape shape_gemm(im2_col_info.tensor_shape());
+    shape_gemm.set(0, mat_weights_cols);
+    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_fully_connected_convolution)
+    {
+        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));
+    }
+    else
+    {
+        ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMMatrixMultiplyKernel::validate(&im2_col_info, weights, &gemm_output_info));
+    }
+
+    ARM_COMPUTE_RETURN_ON_ERROR(NECol2ImKernel::validate(&gemm_output_info, output, Size2D(conv_w, conv_h)));
+
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG((output->dimension(0) != conv_w) || (output->dimension(1) != conv_h), "Output shape does not match the expected one");
+
+    return Status{};
+}
+
 void NEConvolutionLayer::run()
 {
     // Run weights reshaping (Runs once for every configure)
diff --git a/src/runtime/NEON/functions/NEDeconvolutionLayer.cpp b/src/runtime/NEON/functions/NEDeconvolutionLayer.cpp
index 7b4e77b..7bce8a6 100644
--- a/src/runtime/NEON/functions/NEDeconvolutionLayer.cpp
+++ b/src/runtime/NEON/functions/NEDeconvolutionLayer.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017, 2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -24,39 +24,43 @@
 #include "arm_compute/runtime/NEON/functions/NEDeconvolutionLayer.h"
 
 #include "arm_compute/core/Helpers.h"
-#include "arm_compute/core/PixelValue.h"
 #include "arm_compute/core/Utils.h"
 #include "arm_compute/core/Validate.h"
+#include "arm_compute/core/utils/misc/ShapeCalculator.h"
 
 using namespace arm_compute;
+using namespace arm_compute::misc::shape_calculator;
 
 NEDeconvolutionLayer::NEDeconvolutionLayer(std::shared_ptr<IMemoryManager> memory_manager) // NOLINT
     : _memory_group(std::move(memory_manager)),
-      _scale_f(),
       _conv_f(),
-      _scaled_output()
+      _scaled_output(),
+      _input(nullptr),
+      _info(),
+      _inner_border()
 {
 }
 
 void NEDeconvolutionLayer::configure(ITensor *input, const ITensor *weights, const ITensor *bias, ITensor *output, const PadStrideInfo &info,
-                                     unsigned int ax, unsigned int ay, float upscalex, float upscaley)
+                                     unsigned int inner_border_right, unsigned int inner_border_top)
 {
     ARM_COMPUTE_ERROR_ON_NULLPTR(output);
     ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32);
     ARM_COMPUTE_ERROR_ON(weights->info()->dimension(0) != weights->info()->dimension(1));
-    ARM_COMPUTE_ERROR_ON(weights->info()->dimension(0) < 1);
+    ARM_COMPUTE_ERROR_ON(weights->info()->dimension(0) != 1 && weights->info()->dimension(0) != 3 && weights->info()->dimension(0) != 5);
 
-    auto out_dims = deconvolution_output_dimensions(input->info()->dimension(0), input->info()->dimension(1), weights->info()->dimension(0), weights->info()->dimension(1),
-                                                    info.pad().first, info.pad().second, ax, ay, upscalex, upscaley, info.round());
+    _input        = input;
+    _info         = info;
+    _inner_border = std::make_pair(inner_border_right, inner_border_top);
+
+    const unsigned int stride_x = info.stride().first;
+    const unsigned int stride_y = info.stride().second;
+    auto               out_dims = deconvolution_output_dimensions(input->info()->dimension(0), input->info()->dimension(1), weights->info()->dimension(0), weights->info()->dimension(1),
+                                                                  info.pad().first, info.pad().second, inner_border_right, inner_border_top, stride_x, stride_y);
 
     const TensorShape output_shape = deconvolution_output_shape(out_dims, input->info()->tensor_shape(), weights->info()->tensor_shape());
 
-    // Output auto initialization if not yet initialized
-    auto_init_if_empty(*output->info(), output_shape, 1, input->info()->data_type(), input->info()->fixed_point_position());
-
-    ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output, weights, bias);
-    ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, output, weights, bias);
-
+    ARM_COMPUTE_UNUSED(output_shape);
     ARM_COMPUTE_ERROR_ON_MSG(output->info()->dimension(Window::DimX) != output_shape.x(), "Output's width is invalid.");
     ARM_COMPUTE_ERROR_ON_MSG(output->info()->dimension(Window::DimY) != output_shape.y(), "Output's height is invalid.");
     ARM_COMPUTE_ERROR_ON_MSG(output->info()->dimension(Window::DimZ) != output_shape.z(), "Output's depth is invalid.");
@@ -64,51 +68,51 @@
     _memory_group.manage(&_scaled_output);
 
     // configure scale function
-    //Init and allocate intermmidiate tensor for output, same size as input but the first two axis are the same as the output tensor
-    TensorShape scale_out_shape(input->info()->tensor_shape());
-    scale_out_shape.set(0, output->info()->dimension(0));
-    scale_out_shape.set(1, output->info()->dimension(1));
-    TensorInfo scale_out_info(scale_out_shape, 1, input->info()->data_type(), input->info()->fixed_point_position());
+    // Init and allocate intermmidiate tensor for output, same size as input but the first two axis are the same as the output tensor
+    const TensorInfo scale_out_info(compute_deconvolution_shape(*input->info(), stride_x, stride_y, inner_border_right, inner_border_top, info), 1, input->info()->data_type(),
+                                    input->info()->fixed_point_position());
     _scaled_output.allocator()->init(scale_out_info);
-    const unsigned int kernel_size = weights->info()->dimension(0);
-    // Padding for the upsampled image is calculated with the equiation: p' = k - p - 1, where k is kernel size and p is the input padding
-    ARM_COMPUTE_ERROR_ON(info.pad().first > (kernel_size - 1));
-    const unsigned int  tr_px     = kernel_size - info.pad().first - 1;
-    const unsigned int  tr_py     = kernel_size - info.pad().second - 1;
-    const unsigned int  tr_stride = 1;
-    const PadStrideInfo transposed_info(tr_stride, tr_stride, tr_px, tr_py);
-    _scale_f.configure(input, &_scaled_output, std::make_pair(ax, ay), std::make_pair(info.stride().first - 1u, info.stride().second - 1u), transposed_info);
+
     // setup the function to convolve the upscaled output
-    switch(kernel_size)
-    {
-        case 1:
-        {
-            _conv_f.configure(&_scaled_output, weights, bias, output, PadStrideInfo(1, 1, 0, 0, DimensionRoundingType::CEIL));
-            break;
-        }
-        case 3:
-        {
-            _conv_f.configure(&_scaled_output, weights, bias, output, PadStrideInfo(1, 1, 1, 1, DimensionRoundingType::CEIL));
-            break;
-        }
-        case 5:
-        {
-            _conv_f.configure(&_scaled_output, weights, bias, output, PadStrideInfo(1, 1, 2, 2, DimensionRoundingType::CEIL));
-            break;
-        }
-        default:
-        {
-            ARM_COMPUTE_ERROR("Not supported");
-            break;
-        }
-    }
+    const PadStrideInfo conv_info(1, 1, 0, 0, 0, 0, DimensionRoundingType::CEIL);
+    _conv_f.configure(&_scaled_output, weights, bias, output, conv_info);
     _scaled_output.allocator()->allocate();
 }
 
 void NEDeconvolutionLayer::run()
 {
     _memory_group.acquire();
-    _scale_f.run();
+
+    // Initialize _scaled_output buffer
+    const int width_in      = _input->info()->dimension(0);
+    const int height_in     = _input->info()->dimension(1);
+    const int width_scaled  = _scaled_output.info()->dimension(0);
+    const int height_scaled = _scaled_output.info()->dimension(1);
+    const int num_2d_slices = _input->info()->tensor_shape().total_size() / (width_in * height_in);
+    const int stride_x      = _info.stride().first;
+    const int stride_y      = _info.stride().second;
+
+    std::fill_n(reinterpret_cast<float *>(_scaled_output.buffer()), _scaled_output.info()->tensor_shape().total_size(), 0.f);
+
+    // scaled_output is the input for the forward convolution. We copy the input elements to scaled_output
+    // and insert rows and columns with zeroes depending on the stride values.
+    for(int slice = 0; slice < num_2d_slices; ++slice)
+    {
+        const int start_x = _info.pad().first;
+        const int start_y = _inner_border.second + _info.pad().second;
+        const int end_y   = height_scaled - _info.pad().second;
+        const int end_x   = width_scaled - _inner_border.first - _info.pad().first;
+
+        for(int yi = start_y, in_y = 0; yi < end_y; yi += stride_y, in_y++)
+        {
+            for(int xi = start_x, in_x = 0; xi < end_x; xi += stride_x, in_x++)
+            {
+                const auto in = *(reinterpret_cast<float *>(_input->buffer() + _input->info()->offset_element_in_bytes(Coordinates(in_x, in_y, slice))));
+                *(reinterpret_cast<float *>(_scaled_output.buffer() + _scaled_output.info()->offset_element_in_bytes(Coordinates(xi, yi, slice)))) = in;
+            }
+        }
+    }
+
     _conv_f.run();
     _memory_group.release();
 }
diff --git a/src/runtime/NEON/functions/NEDeconvolutionLayerUpsample.cpp b/src/runtime/NEON/functions/NEDeconvolutionLayerUpsample.cpp
deleted file mode 100644
index 79b9b2d..0000000
--- a/src/runtime/NEON/functions/NEDeconvolutionLayerUpsample.cpp
+++ /dev/null
@@ -1,121 +0,0 @@
-/*
- * Copyright (c) 2016, 2017 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 "arm_compute/runtime/NEON/functions/NEDeconvolutionLayerUpsample.h"
-
-#include "arm_compute/core/Coordinates.h"
-#include "arm_compute/core/Error.h"
-#include "arm_compute/core/Helpers.h"
-#include "arm_compute/core/ITensor.h"
-#include "arm_compute/core/NEON/kernels/NEDeconvolutionLayerUpsampleKernel.h"
-#include "arm_compute/core/PixelValue.h"
-#include "arm_compute/core/TensorInfo.h"
-#include "arm_compute/core/Window.h"
-#include "arm_compute/runtime/NEON/NEScheduler.h"
-#include "arm_compute/runtime/TensorAllocator.h"
-#include "support/ToolchainSupport.h"
-
-#include <cmath>
-#include <cstddef>
-#include <utility>
-
-using namespace arm_compute;
-
-namespace
-{
-inline void precompute_offsets(ITensor *offsets, float wr, size_t input_element_size, const std::pair<unsigned int, unsigned int> &a,
-                               const std::pair<unsigned int, unsigned int> &iz, const PadStrideInfo &info)
-{
-    ARM_COMPUTE_ERROR_ON(nullptr == offsets);
-    Window    win;
-    const int padx          = info.pad().first;
-    const int pady          = info.pad().second;
-    const int ax            = a.first;
-    const int ay            = a.second;
-    const int offset_width  = offsets->info()->dimension(0);
-    const int offset_height = offsets->info()->dimension(1);
-    // The values of ax and ay denote the number of ZEROS to be added on the top and right inner border of the image.
-    // Step value along the XY axis will depend on the number of zeros to be inserted between samples (number of zeros + 1).
-    // Pre-compute the X offset, Y's stride is unknown at this point so we can't precompute Y's offsets
-    for(int yi = ay; yi < (offset_height - pady); yi += (1 + iz.second))
-    {
-        for(int xi = padx; xi < (offset_width - ax); xi += (1 + iz.first))
-        {
-            int         *ptr                  = reinterpret_cast<int *>(offsets->ptr_to_element(Coordinates(xi, yi)));
-            const size_t in_xi                = (xi + 0.5f) * wr;
-            *reinterpret_cast<int32_t *>(ptr) = in_xi * input_element_size;
-        }
-    }
-}
-} // namespace
-
-NEDeconvolutionLayerUpsample::NEDeconvolutionLayerUpsample(std::shared_ptr<IMemoryManager> memory_manager) // NOLINT
-    : _memory_group(std::move(memory_manager)),
-      _offsets(),
-      _border_handler(),
-      _upsample()
-{
-}
-
-void NEDeconvolutionLayerUpsample::configure(ITensor *input, ITensor *output, const std::pair<unsigned int, unsigned int> &a,
-                                             const std::pair<unsigned int, unsigned int> &iz, const PadStrideInfo &info)
-{
-    ARM_COMPUTE_ERROR_ON(nullptr == input);
-    ARM_COMPUTE_ERROR_ON(nullptr == output);
-
-    for(size_t i = 2; i < Coordinates::num_max_dimensions; ++i)
-    {
-        ARM_COMPUTE_ERROR_ON(input->info()->dimension(i) != output->info()->dimension(i));
-    }
-
-    // Get the tensor shape
-    const TensorShape shape(output->info()->dimension(0), output->info()->dimension(1));
-
-    // Compute the ratio between source width/height and destination width/height
-    const auto wr = static_cast<float>(input->info()->dimension(0)) / static_cast<float>(output->info()->dimension(0));
-    const auto hr = static_cast<float>(input->info()->dimension(1)) / static_cast<float>(output->info()->dimension(1));
-    ARM_COMPUTE_UNUSED(hr);
-    // Get the element size of the input image
-    const size_t input_element_size = input->info()->element_size();
-
-    TensorInfo tensor_info_offsets(shape, Format::S32);
-    _offsets.allocator()->init(tensor_info_offsets);
-
-    _upsample.configure(input, &_offsets, output);
-
-    // Allocate once the configure methods have been called
-    _offsets.allocator()->allocate();
-    // Pre-compute offsets for nearest interpolation
-    std::fill_n(reinterpret_cast<int32_t *>(_offsets.buffer()), _offsets.info()->total_size() / sizeof(int32_t), -1 * input_element_size);
-    precompute_offsets(&_offsets, wr, input_element_size, a, iz, info);
-
-    _border_handler.configure(input, _upsample.border_size(), BorderMode::CONSTANT, PixelValue(0.f));
-}
-
-void NEDeconvolutionLayerUpsample::run()
-{
-    NEScheduler::get().schedule(&_border_handler, Window::DimZ);
-    _memory_group.acquire();
-    NEScheduler::get().schedule(&_upsample, Window::DimY);
-    _memory_group.release();
-}
diff --git a/src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp b/src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp
index b890c6f..2d08b45 100644
--- a/src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp
+++ b/src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -26,28 +26,56 @@
 #include "arm_compute/core/Helpers.h"
 #include "arm_compute/core/ITensor.h"
 #include "arm_compute/core/PixelValue.h"
+#include "arm_compute/core/utils/quantization/AsymmHelpers.h"
 #include "arm_compute/runtime/NEON/NEScheduler.h"
 #include "support/ToolchainSupport.h"
 
 using namespace arm_compute;
 
 NEDepthwiseConvolutionLayer3x3::NEDepthwiseConvolutionLayer3x3()
-    : _kernel(), _bias_kernel(), _border_handler(), _has_bias(false)
+    : _kernel(), _output_stage_kernel(), _border_handler(), _accumulator(), _has_bias(false), _is_quantized(false)
 {
 }
 
 void NEDepthwiseConvolutionLayer3x3::configure(ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const PadStrideInfo &conv_info)
 {
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32);
-    ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output, weights);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights);
 
-    // Call convolution kernel
-    _kernel.configure(input, weights, output, conv_info);
-    _border_handler.configure(input, _kernel.border_size(), BorderMode::CONSTANT, PixelValue(static_cast<float>(0.f)));
-    if(biases != nullptr)
+    PixelValue zero_value(0.f);
+
+    _is_quantized = is_data_type_quantized_asymmetric(input->info()->data_type());
+    _has_bias     = biases != nullptr;
+
+    // Allocate the intermediate accumulator tensor in case of fixed point input
+    if(_is_quantized)
     {
-        _bias_kernel.configure(output, biases);
-        _has_bias = true;
+        _accumulator.allocator()->init(TensorInfo(output->info()->tensor_shape(), 1, DataType::S32));
+        _accumulator.info()->set_quantization_info(input->info()->quantization_info());
+        zero_value = PixelValue(static_cast<uint32_t>(input->info()->quantization_info().offset));
+    }
+
+    // Configure depthwise convolution kernel
+    _kernel.configure(input, weights, (_is_quantized) ? &_accumulator : output, conv_info);
+
+    // Configure border handler
+    _border_handler.configure(input, _kernel.border_size(), BorderMode::CONSTANT, zero_value);
+
+    // Configure biases accumulation
+    if(_has_bias || _is_quantized)
+    {
+        if(_is_quantized)
+        {
+            float multiplier = input->info()->quantization_info().scale * weights->info()->quantization_info().scale / output->info()->quantization_info().scale;
+            int   output_multiplier, output_shift;
+            quantization::calculate_quantized_multiplier_less_than_one(multiplier, &output_multiplier, &output_shift);
+            _output_stage_kernel.configure(&_accumulator, biases, output, output_multiplier, output_shift, output->info()->quantization_info().offset);
+            _accumulator.allocator()->allocate();
+        }
+        else
+        {
+            _output_stage_kernel.configure(output, biases);
+        }
     }
 }
 
@@ -55,9 +83,9 @@
 {
     NEScheduler::get().schedule(&_border_handler, Window::DimX);
     NEScheduler::get().schedule(&_kernel, Window::DimX);
-    if(_has_bias)
+    if(_has_bias || _is_quantized)
     {
-        NEScheduler::get().schedule(&_bias_kernel, Window::DimX);
+        NEScheduler::get().schedule(&_output_stage_kernel, Window::DimX);
     }
 }
 
diff --git a/src/runtime/NEON/functions/NEDirectConvolutionLayer.cpp b/src/runtime/NEON/functions/NEDirectConvolutionLayer.cpp
index afa5d97..c26c99a 100644
--- a/src/runtime/NEON/functions/NEDirectConvolutionLayer.cpp
+++ b/src/runtime/NEON/functions/NEDirectConvolutionLayer.cpp
@@ -34,7 +34,7 @@
 using namespace arm_compute;
 
 NEDirectConvolutionLayer::NEDirectConvolutionLayer(std::shared_ptr<IMemoryManager> memory_manager)
-    : _memory_group(std::move(memory_manager)), _accumulate_bias_kernel(), _conv_kernel(), _input_border_handler(), _accumulator(), _has_bias(false)
+    : _memory_group(std::move(memory_manager)), _output_stage_kernel(), _conv_kernel(), _input_border_handler(), _accumulator(), _has_bias(false), _is_fixed_point(false)
 {
 }
 
@@ -50,16 +50,16 @@
     _has_bias = (bias != nullptr);
 
     // Allocate the intermediate accumulator tensor in case of fixed point input
-    if(is_data_type_fixed_point(input->info()->data_type()))
+    _is_fixed_point = is_data_type_fixed_point(input->info()->data_type());
+    if(_is_fixed_point)
     {
         const DataType promoted_dt = (input->info()->data_type() == DataType::QS8) ? DataType::QS16 : DataType::QS32;
         _accumulator.allocator()->init(TensorInfo(output->info()->tensor_shape(), 1, promoted_dt, output->info()->fixed_point_position()));
         _memory_group.manage(&_accumulator);
         _conv_kernel.configure(input, weights, &_accumulator, conv_info);
-        if(_has_bias)
-        {
-            _accumulate_bias_kernel.configure(&_accumulator, bias, output);
-        }
+
+        // When no bias is provided, we need to downscale the accumulator tensor
+        _output_stage_kernel.configure(&_accumulator, bias, output);
         _accumulator.allocator()->allocate();
     }
     else
@@ -67,7 +67,7 @@
         _conv_kernel.configure(input, weights, output, conv_info);
         if(_has_bias)
         {
-            _accumulate_bias_kernel.configure(output, bias);
+            _output_stage_kernel.configure(output, bias);
         }
     }
 
@@ -90,20 +90,17 @@
     // Validate Convolution kernel
     ARM_COMPUTE_RETURN_ON_ERROR(NEDirectConvolutionLayerKernel::validate(input, weights, &accumulator, conv_info));
 
-    // Validate bias
-    ARM_COMPUTE_RETURN_ERROR_ON_MSG((bias == nullptr) && is_data_type_fixed_point(data_type),
-                                    "Biases should be provided for fixed point inputs");
     if(bias != nullptr)
     {
         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(weights, bias);
         ARM_COMPUTE_RETURN_ERROR_ON_MSG(bias->dimension(0) != weights->dimension(3),
                                         "Biases size and number of input feature maps should match");
         ARM_COMPUTE_RETURN_ERROR_ON_MSG(bias->num_dimensions() > 1, "Biases should be one dimensional");
-
-        // Validate bias kernel
-        ARM_COMPUTE_RETURN_ON_ERROR(NEDirectConvolutionLayerBiasAccumulateKernel::validate(&accumulator, bias, output));
     }
 
+    // Validate bias kernel
+    ARM_COMPUTE_RETURN_ON_ERROR(NEDirectConvolutionLayerOutputStageKernel::validate(&accumulator, bias, output));
+
     return Status{};
 }
 
@@ -114,10 +111,9 @@
     _memory_group.acquire();
 
     NEScheduler::get().schedule(&_conv_kernel, Window::DimZ);
-    if(_has_bias)
+    if(_has_bias || _is_fixed_point)
     {
-        NEScheduler::get().schedule(&_accumulate_bias_kernel, Window::DimY);
+        NEScheduler::get().schedule(&_output_stage_kernel, Window::DimY);
     }
-
     _memory_group.release();
 }
diff --git a/src/runtime/NEON/functions/NEGEMM.cpp b/src/runtime/NEON/functions/NEGEMM.cpp
index 950f4c9..e640b06 100644
--- a/src/runtime/NEON/functions/NEGEMM.cpp
+++ b/src/runtime/NEON/functions/NEGEMM.cpp
@@ -50,15 +50,17 @@
 {
 NEGEMM::NEGEMM(std::shared_ptr<IMemoryManager> memory_manager)
     : _memory_group(std::move(memory_manager)), _interleave_kernel(), _transpose_kernel(), _mm_kernel(), _mm_optimised_kernel(nullptr), _ma_kernel(), _tmp_a(), _tmp_b(), _workspace(),
-      _run_vector_matrix_multiplication(false), _run_addition(false)
+      _run_vector_matrix_multiplication(false), _run_addition(false), _is_first_run(true), _reshape_b_only_on_first_run(false)
 {
 }
 
-void NEGEMM::configure(const ITensor *a, const ITensor *b, const ITensor *c, ITensor *d, float alpha, float beta)
+void NEGEMM::configure(const ITensor *a, const ITensor *b, const ITensor *c, ITensor *d, float alpha, float beta, const GEMMInfo &gemm_info)
 {
     ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(a, 1, DataType::F32, DataType::F16, DataType::QS8, DataType::QS16);
     ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(a, b, d);
     ARM_COMPUTE_ERROR_ON_MSG(a->info()->dimension(0) != b->info()->dimension(1), "The product AB is defined only if the number of columns in A is equal to the number of rows in B");
+    ARM_COMPUTE_ERROR_ON_MSG(gemm_info.is_a_reshaped(), "Matrix A already reshaped is not supported");
+    ARM_COMPUTE_ERROR_ON_MSG(gemm_info.is_b_reshaped(), "Matrix B already reshaped is not supported");
 
     if(c != nullptr)
     {
@@ -70,6 +72,8 @@
         ARM_COMPUTE_ERROR_ON_MSG(c->info()->dimension(1) != d->info()->dimension(1), "The C matrix must have the same number of columns as the output matrix");
     }
 
+    // Check if we need to reshape the matrix B only on the first run
+    _reshape_b_only_on_first_run      = gemm_info.reshape_b_only_on_first_run();
     _run_vector_matrix_multiplication = a->info()->dimension(1) < 2;
 
     // Check if the first input tensor is a vector.
@@ -142,7 +146,7 @@
             _memory_group.manage(&_workspace);
 
             // Configure matrix multiplication kernel
-            _mm_optimised_kernel->configure(a, b, d, &_workspace, alpha, 0.f);
+            _mm_optimised_kernel->configure(a, b, d, &_workspace, alpha, 0.f, false /* is_transposed_0 */, false /* is_transposed_1 */);
             _workspace.allocator()->allocate();
         }
         else
@@ -207,8 +211,18 @@
             // Run interleave kernel
             NEScheduler::get().schedule(&_interleave_kernel, Window::DimY);
 
-            // Run transpose kernel
-            NEScheduler::get().schedule(&_transpose_kernel, Window::DimY);
+            if(_is_first_run)
+            {
+                // Run transpose kernel
+                NEScheduler::get().schedule(&_transpose_kernel, Window::DimY);
+
+                _is_first_run = false;
+            }
+            else if(!_reshape_b_only_on_first_run)
+            {
+                // Run transpose kernel
+                NEScheduler::get().schedule(&_transpose_kernel, Window::DimY);
+            }
         }
 
         NEScheduler::get().schedule(&_mm_kernel, _run_vector_matrix_multiplication ? Window::DimX : Window::DimY);
diff --git a/src/runtime/NEON/functions/NEGEMMLowpAssemblyMatrixMultiplyCore.cpp b/src/runtime/NEON/functions/NEGEMMLowpAssemblyMatrixMultiplyCore.cpp
index 6e03ffa..9b36e81 100644
--- a/src/runtime/NEON/functions/NEGEMMLowpAssemblyMatrixMultiplyCore.cpp
+++ b/src/runtime/NEON/functions/NEGEMMLowpAssemblyMatrixMultiplyCore.cpp
@@ -74,7 +74,7 @@
 #endif /* __aarch64__ */
 
 #ifdef ARM_COMPUTE_AARCH64_V8_2
-    if(ci.CPU == CPUTarget::A75_DOT)
+    if(ci.CPU == CPUTarget::A75_DOT || ci.CPU == CPUTarget::A55_DOT)
     {
         // Configure matrix multiply kernel
         GemmInterleaved<gemm_s8_12x8, int8_t, int32_t> gemm(&ci, M, N, K, false, false);
@@ -87,10 +87,6 @@
         _mm_kernel = std::move(k);
         _workspace.allocator()->allocate();
     }
-    else if(ci.CPU == CPUTarget::A55_DOT)
-    {
-        ARM_COMPUTE_ERROR_ON("WIP");
-    }
     else
 #elif defined(ARM_COMPUTE_AARCH64_V8A)
     if(ci.CPU == CPUTarget::A53)
diff --git a/src/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.cpp b/src/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.cpp
index 50aa5b6..c4028dc 100644
--- a/src/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.cpp
+++ b/src/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.cpp
@@ -84,7 +84,7 @@
 
         // Configure matrix multiplication kernel
         auto k = arm_compute::support::cpp14::make_unique<NEGEMMLowpAArch64V8P4Kernel>();
-        k->configure(a, b, output, &_workspace, 1.f, 1.f);
+        k->configure(a, b, output, &_workspace, 1.f, 1.f, false, false);
         _mm_kernel = std::move(k);
     }
     else
diff --git a/src/runtime/NEON/functions/NEPoolingLayer.cpp b/src/runtime/NEON/functions/NEPoolingLayer.cpp
index 530c7fc..8a32507 100644
--- a/src/runtime/NEON/functions/NEPoolingLayer.cpp
+++ b/src/runtime/NEON/functions/NEPoolingLayer.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -43,9 +43,14 @@
     // Configure pooling kernel
     _pooling_layer_kernel.configure(input, output, pool_info);
 
-    // Configure border depending on operation required
+    // Configure border depending on operation required (quantize border in case of asymmetric data_type)
     BorderMode border_mode = (pool_info.pool_type() == PoolingType::MAX) ? BorderMode::REPLICATE : BorderMode::CONSTANT;
-    _border_handler.configure(input, _pooling_layer_kernel.border_size(), border_mode, PixelValue(static_cast<float>(0.f)));
+    PixelValue zero_value(0.f);
+    if(is_data_type_quantized_asymmetric(input->info()->data_type()) && !pool_info.exclude_padding())
+    {
+        zero_value = PixelValue(static_cast<uint32_t>(input->info()->quantization_info().offset));
+    }
+    _border_handler.configure(input, _pooling_layer_kernel.border_size(), border_mode, zero_value);
 }
 
 Status NEPoolingLayer::validate(const ITensorInfo *input, const ITensorInfo *output, const PoolingLayerInfo &pool_info)
diff --git a/src/runtime/NEON/functions/NEWinogradLayer.cpp b/src/runtime/NEON/functions/NEWinogradLayer.cpp
index 3251de4..da46f87 100644
--- a/src/runtime/NEON/functions/NEWinogradLayer.cpp
+++ b/src/runtime/NEON/functions/NEWinogradLayer.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -43,7 +43,8 @@
 namespace arm_compute
 {
 NEWinogradLayer::NEWinogradLayer(std::shared_ptr<IMemoryManager> memory_manager)
-    : _memory_group(std::move(memory_manager)), _winograd_kernel(), _weights_workspace(), _workspace(), _kernel_storage(), _input(), _weights(), _output(), _reshaped_kernel(false), _conv()
+    : _memory_group(std::move(memory_manager)), _winograd_kernel(), _permute_input(), _permute_weights(), _permute_output(), _input_workspace(), _output_workspace(), _kernel_storage(), _input_nhwc(),
+      _output_nhwc(), _weights_hwio(), _input(), _weights(), _output(), _reshaped_kernel(false), _conv()
 {
 } /* arm_compute */
 
@@ -71,85 +72,107 @@
     ARM_COMPUTE_ERROR_ON_MSG(stride_y != 1 || stride_x != 1, "Winograd layer only supports unit strides.");
 
     // Get convolved dimensions
-    auto      padding     = PADDING_VALID;
-    const int in_channels = input->info()->dimension(2);
+    const int in_channels  = input->info()->dimension(2);
+    const int out_channels = output->info()->dimension(2);
 
-    const int out_channels   = output->info()->dimension(2);
-    const int weights_width  = weights->info()->dimension(0);
-    const int weights_height = weights->info()->dimension(1);
-
-    const KernelShape   kernel_shape({ out_channels, weights_height, weights_width, in_channels });
     const Tensor4DShape in_shape(internal_get_input_shape(input));
 
     // Get the memory required to instantiate a new Winograd operator.
-    constexpr size_t kstore_alignment          = 64;
-    const size_t     kernel_storage_per_thread = NEWinogradLayerKernel::get_kernel_storage_size(kernel_shape);
-    _kernel_storage.allocator()->init(TensorInfo(TensorShape{ (kernel_storage_per_thread + kstore_alignment - 1) }, 1, DataType::U8));
+    constexpr size_t storage_alignment   = 64;
+    const size_t     kernel_storage_size = NEWinogradLayerKernel::get_weight_storage_size(out_channels, in_channels) * sizeof(float);
+    _kernel_storage.allocator()->init(TensorInfo(TensorShape{ (kernel_storage_size + storage_alignment - 1) }, 1, DataType::U8));
     _memory_group.manage(&_kernel_storage);
-
-    // Get workbench size and allocate memory
-    constexpr size_t wspace_alignment = 64;
-    const size_t     ws_size          = NEWinogradLayerKernel::get_working_space_size(in_shape, kernel_shape, padding);
-    _workspace.allocator()->init(TensorInfo(TensorShape{ (ws_size + wspace_alignment - 1) }, 1, DataType::U8));
-    _memory_group.manage(&_workspace);
-
-    // Workspace for weights transform
-    const size_t weights_transform_size = NEWinogradLayerKernel::get_kernel_transform_working_size(kernel_shape);
-    _weights_workspace.allocator()->init(TensorInfo(TensorShape{ (weights_transform_size + wspace_alignment - 1) }, 1, DataType::U8));
-    _memory_group.manage(&_weights_workspace);
-
+    _memory_group.manage(&_input_nhwc);
     _kernel_storage.allocator()->allocate();
-    _workspace.allocator()->allocate();
-    _weights_workspace.allocator()->allocate();
+    // Input storage
+    const size_t input_storage_size = NEWinogradLayerKernel::get_input_storage_size(in_shape.n_batches, in_shape.n_channels, in_shape.n_rows, in_shape.n_cols, false) * sizeof(float);
+    _input_workspace.allocator()->init(TensorInfo(TensorShape{ (input_storage_size + storage_alignment - 1) }, 1, DataType::U8));
+    _memory_group.manage(&_input_workspace);
+    _input_workspace.allocator()->allocate();
+
+    // Output storage
+    const size_t output_storage_size = NEWinogradLayerKernel::get_output_storage_size(in_shape.n_batches, in_shape.n_rows, in_shape.n_cols, out_channels, false) * sizeof(float);
+    _output_workspace.allocator()->init(TensorInfo(TensorShape{ (output_storage_size + storage_alignment - 1) }, 1, DataType::U8));
+    _memory_group.manage(&_output_workspace);
+    _output_workspace.allocator()->allocate();
+
+    // configure and allocate dst tensor to be used to convert from winograd domain to spatial domain when calling to reshape_output()
+    TensorInfo info(TensorShape(_output->info()->dimension(2), _output->info()->dimension(0),
+                                _output->info()->dimension(1), _output->info()->dimension(3)),
+                    1, _output->info()->data_type());
+    _output_nhwc.allocator()->init(info);
+
+    _output_nhwc.allocator()->allocate();
+
+    // Re-order a weight tensor from [Output feature map x Input feature map x Height x Width] to [Height x Width x Input feature map x Output feature map]
+    switch(weights->info()->num_dimensions())
+    {
+        case 3:
+        {
+            _permute_weights.configure(weights, &_weights_hwio, PermutationVector(2U, 0U, 1U));
+            break;
+        }
+        case 4:
+        {
+            _permute_weights.configure(weights, &_weights_hwio, PermutationVector(3U, 2U, 0U, 1U));
+            break;
+        }
+        default:
+        {
+            ARM_COMPUTE_ERROR("Not supported.");
+            break;
+        }
+    }
+
+    _weights_hwio.allocator()->allocate();
+
+    // configure the kernel to transform the input tensor from NCHW -> NHWC
+    _permute_input.configure(input, &_input_nhwc, PermutationVector(2U, 0U, 1U));
+
+    _input_nhwc.allocator()->allocate();
 
     // Create Winograd operator object
-    _conv = support::cpp14::make_unique<Winograd3x3F32>(kernel_shape, in_shape, padding, _kernel_storage.buffer());
+    _conv = support::cpp14::make_unique<Winograd3x3F32>(
+                in_shape.n_batches,
+                in_shape.n_channels,
+                in_shape.n_rows,
+                in_shape.n_cols,
+                out_channels,
+                false,
+                reinterpret_cast<const float *>(_weights_hwio.buffer()),
+                reinterpret_cast<float *>(_kernel_storage.buffer()),
+                reinterpret_cast<float *>(_input_nhwc.buffer()),
+                reinterpret_cast<float *>(_input_workspace.buffer()),
+                reinterpret_cast<float *>(_output_nhwc.buffer()),
+                reinterpret_cast<float *>(_output_workspace.buffer()));
 
     // Configure the kernel, padding not needed so it's safe to call configure after allocare
-    _winograd_kernel.configure(output, _conv.get());
+    _winograd_kernel.configure(_conv.get());
+
+    // Reorder the convoluted output to ACL's ordering NCHW
+    _permute_output.configure(&_output_nhwc, _output, PermutationVector(1U, 2U, 0U));
+
 }
 
 void NEWinogradLayer::run()
 {
-#if defined(__aarch64__)
     _memory_group.acquire();
     if(!_reshaped_kernel)
     {
-        _conv->transform_weights(reinterpret_cast<const float *>(_weights->buffer()), reinterpret_cast<float *>(_weights_workspace.buffer()));
         _reshaped_kernel = true;
+        _permute_weights.run();
+        _conv->transform_weights();
     }
-    const Tensor4DShape in_shape(internal_get_input_shape(_input));
-    auto                padding = PADDING_VALID;
-
     //Bring channels to the front as Winograd code expects the tensor to be in the format NHWC
-    _conv->nchw2nhwc(in_shape, padding, _workspace.buffer(), reinterpret_cast<const float *>(_input->buffer()));
-
-    //Get ptrs into the workspace
-    std::pair<void *, void *> nhwc_ptrs = _conv->get_nhwc_ptrs(in_shape, padding, _workspace.buffer());
-
-    //Setup matrices ptrs and transfor the input tensor to the appropriate form before running GEMM.
-    _conv->reshape_input(in_shape, padding, nhwc_ptrs.second, _workspace.buffer());
-
+    _permute_input.run();
+    // Transform input tensor to the winograd domain
+    _conv->transform_input();
     //Run 16 GEMMs in multiple threads, each kernel runs one or more GEMMs
-    NEScheduler::get().schedule(&_winograd_kernel, Window::DimY);
-
-    //Transform the output to the appropriate form
-    _conv->reshape_output(in_shape, padding, nhwc_ptrs.first);
-
-    //Transform back to NCHW
-    _conv->nhwc2nchw(in_shape, padding, _workspace.buffer(), reinterpret_cast<float *>(_output->buffer()));
-
+    NEScheduler::get().schedule(&_winograd_kernel, Window::DimX);
+    // Transform output tensor to the spatial domain
+    _conv->transform_output();
+    // Reorder the convoluted output to ACL's ordering NCHW
+    _permute_output.run();
     _memory_group.release();
-#else  /* __aarch64__ */
-    ARM_COMPUTE_UNUSED(_winograd_kernel);
-    ARM_COMPUTE_UNUSED(_workspace);
-    ARM_COMPUTE_UNUSED(_kernel_storage);
-    ARM_COMPUTE_UNUSED(_input);
-    ARM_COMPUTE_UNUSED(_weights);
-    ARM_COMPUTE_UNUSED(_output);
-    ARM_COMPUTE_UNUSED(_reshaped_kernel);
-    ARM_COMPUTE_UNUSED(_conv);
-    ARM_COMPUTE_ERROR("Winograd only supported for aarch64, recompile with arch=arm64-v8a.");
-#endif /* __aarch64__ */
 }
 } // namespace arm_compute