COMPMID-617: Add validate support for NEON FullyConnectedLayer

Change-Id: I08987022c8d4cc335c00b8af27bd3edb8fe64d3b
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/111596
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Alexander Gilday <alexander.gilday@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>

6 files changed, 184 insertions, 104 deletions

diff --git a/src/runtime/NEON/functions/NEFlattenLayer.cpp b/src/runtime/NEON/functions/NEFlattenLayer.cpp
index 408eff5746..32edf93b63 100644
--- a/src/runtime/NEON/functions/NEFlattenLayer.cpp
+++ b/src/runtime/NEON/functions/NEFlattenLayer.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -32,6 +32,6 @@ using namespace arm_compute;
 void NEFlattenLayer::configure(const ITensor *input, ITensor *output)
 {
     auto k = arm_compute::support::cpp14::make_unique<NEIm2ColKernel>();
-    k->configure(input, output, Size2D(1, 1), PadStrideInfo(1, 1, 0, 0), false);
+    k->configure(input, output, Size2D(1, 1), PadStrideInfo(1, 1, 0, 0), false, false, true);
     _kernel = std::move(k);
 }
\ No newline at end of file
diff --git a/src/runtime/NEON/functions/NEFullyConnectedLayer.cpp b/src/runtime/NEON/functions/NEFullyConnectedLayer.cpp
index fc04e28972..26b7271710 100644
--- a/src/runtime/NEON/functions/NEFullyConnectedLayer.cpp
+++ b/src/runtime/NEON/functions/NEFullyConnectedLayer.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -23,15 +23,18 @@
  */
 #include "arm_compute/runtime/NEON/functions/NEFullyConnectedLayer.h"
 
+#include "arm_compute/core/Helpers.h"
 #include "arm_compute/core/Size2D.h"
 #include "arm_compute/core/Validate.h"
+#include "arm_compute/core/utils/misc/ShapeCalculator.h"
 #include "arm_compute/runtime/NEON/NEScheduler.h"
 
 #include <algorithm>
 #include <cmath>
 
-namespace arm_compute
-{
+using namespace arm_compute;
+using namespace arm_compute::misc::shape_calculator;
+
 NEFullyConnectedLayerReshapeWeights::NEFullyConnectedLayerReshapeWeights(std::shared_ptr<IMemoryManager> memory_manager)
     : _memory_group(std::move(memory_manager)), _transpose_kernel(), _transpose1xW_kernel(), _transpose_output(), _transpose_weights(false), _is_batched_fc_layer(false)
 {
@@ -39,13 +42,10 @@ NEFullyConnectedLayerReshapeWeights::NEFullyConnectedLayerReshapeWeights(std::sh
 
 void NEFullyConnectedLayerReshapeWeights::configure(const ITensor *input, ITensor *output, bool transpose_weights, bool is_batched_fc_layer)
 {
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32);
-    ARM_COMPUTE_ERROR_ON(input->info()->num_dimensions() > 2);
-    ARM_COMPUTE_ERROR_ON(output == nullptr);
-    ARM_COMPUTE_ERROR_ON(!transpose_weights && !is_batched_fc_layer);
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
 
-    const DataType data_type            = input->info()->data_type();
-    const int      fixed_point_position = input->info()->fixed_point_position();
+    // Perform validate step
+    ARM_COMPUTE_ERROR_THROW_ON(NEFullyConnectedLayerReshapeWeights::validate(input->info(), output->info(), transpose_weights, is_batched_fc_layer));
 
     _transpose_weights   = transpose_weights;
     _is_batched_fc_layer = is_batched_fc_layer;
@@ -56,8 +56,7 @@ void NEFullyConnectedLayerReshapeWeights::configure(const ITensor *input, ITenso
         if(_is_batched_fc_layer)
         {
             // Initialize the output tensor for transpose
-            TensorShape shape_transposed(input->info()->dimension(1), input->info()->dimension(0));
-            _transpose_output.allocator()->init(TensorInfo(shape_transposed, 1, data_type, fixed_point_position));
+            _transpose_output.allocator()->init(input->info()->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(compute_transposed_shape(*input->info())));
             _memory_group.manage(&_transpose_output);
             _transpose_kernel.configure(input, &_transpose_output);
 
@@ -79,11 +78,39 @@ void NEFullyConnectedLayerReshapeWeights::configure(const ITensor *input, ITenso
             // Configure transpose 1xW kernel
             _transpose1xW_kernel.configure(input, output);
         }
+    }
+}
+
+Status NEFullyConnectedLayerReshapeWeights::validate(const ITensorInfo *input, const ITensorInfo *output, bool transpose_weights, bool is_batched_fc_layer)
+{
+    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(input->num_dimensions() > 2);
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(!transpose_weights && !is_batched_fc_layer, "Configuration transpose_weights=false & is_batched_fc_layer=false not supported");
+
+    if(transpose_weights)
+    {
+        if(is_batched_fc_layer)
+        {
+            std::unique_ptr<ITensorInfo> use_output = output->clone();
+            use_output->set_is_resizable(true).reset_padding().set_tensor_shape(compute_transposed_shape(*input));
+
+            ARM_COMPUTE_RETURN_ON_ERROR(NETransposeKernel::validate(input, use_output.get()));
+            ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMTranspose1xWKernel::validate(use_output.get(), output));
+        }
         else
         {
-            ARM_COMPUTE_ERROR("Configuration transpose_weights=false & is_batched_fc_layer=false not supported");
+            ARM_COMPUTE_RETURN_ON_ERROR(NETransposeKernel::validate(input, output));
+        }
+    }
+    else
+    {
+        if(is_batched_fc_layer)
+        {
+            ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMTranspose1xWKernel::validate(input, output));
         }
     }
+
+    return Status{};
 }
 
 void NEFullyConnectedLayerReshapeWeights::run()
@@ -122,26 +149,25 @@ void NEFullyConnectedLayer::configure(const ITensor *input, const ITensor *weigh
     // Weights: flat(In) x Out
     // Biases: Out
     // Output: Out x B (B can be multi-dimensional)
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output);
 
-    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, output);
-    ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT_POSITION(input, weights, output);
+    // Perform validate step
+    ARM_COMPUTE_ERROR_THROW_ON(NEFullyConnectedLayer::validate(input->info(),
+                                                               weights->info(),
+                                                               biases != nullptr ? biases->info() : nullptr,
+                                                               output->info(),
+                                                               transpose_weights,
+                                                               are_weights_reshaped));
 
-    const DataType data_type            = input->info()->data_type();
-    const int      fixed_point_position = input->info()->fixed_point_position();
-    const int      num_batch_dimensions = std::max(0, static_cast<int>(output->info()->tensor_shape().num_dimensions()) - 1);
-    const int      num_input_dimensions = input->info()->tensor_shape().num_dimensions() - num_batch_dimensions;
-    const size_t   linear_input_size    = input->info()->tensor_shape().total_size_lower(num_input_dimensions);
+    const int    num_batch_dimensions = std::max(0, static_cast<int>(output->info()->tensor_shape().num_dimensions()) - 1);
+    const int    num_input_dimensions = input->info()->tensor_shape().num_dimensions() - num_batch_dimensions;
+    const size_t linear_input_size    = input->info()->tensor_shape().total_size_lower(num_input_dimensions);
 
     _linearize_input      = (input->info()->tensor_shape().x() != linear_input_size) || (num_input_dimensions > 1 && linear_input_size == 1);
     _are_weights_reshaped = are_weights_reshaped;
     _accumulate_biases    = biases != nullptr;
     _is_batched_fc_layer  = num_batch_dimensions > 0;
 
-    // Check if number of batches match
-    ARM_COMPUTE_ERROR_ON(input->info()->tensor_shape().total_size_upper(num_input_dimensions) != output->info()->tensor_shape().total_size_upper(1));
-    ARM_COMPUTE_ERROR_ON(weights->info()->num_dimensions() > 2);
-
     const size_t   interleave_width = 16 / input->info()->element_size();
     const ITensor *weights_to_use   = weights;
 
@@ -149,65 +175,33 @@ void NEFullyConnectedLayer::configure(const ITensor *input, const ITensor *weigh
     {
         weights_to_use = &_reshape_weights_output;
 
-        TensorShape reshaped_weights_shape(weights->info()->tensor_shape());
-
-        // Transpose weights if the user hasn't done it
-        if(transpose_weights)
-        {
-            const size_t shape_x = reshaped_weights_shape.x();
-            reshaped_weights_shape.set(0, reshaped_weights_shape.y());
-            reshaped_weights_shape.set(1, shape_x);
-        }
-
-        // If the we run multiple batches we need 1xW transpose, too.
-        if(_is_batched_fc_layer)
-        {
-            const float shape_x = reshaped_weights_shape.x();
-            reshaped_weights_shape.set(0, reshaped_weights_shape.y() * interleave_width);
-            reshaped_weights_shape.set(1, static_cast<unsigned int>(std::ceil(shape_x / interleave_width)));
-        }
-
-        _reshape_weights_output.allocator()->init(TensorInfo(reshaped_weights_shape, 1, data_type, fixed_point_position));
+        _reshape_weights_output.allocator()->init(input->info()->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(compute_fully_connected_reshaped_weights_shape(weights->info(),
+                                                  transpose_weights,
+                                                  _is_batched_fc_layer, interleave_width)));
 
         // Reshape the weights
         _reshape_weights_kernel.configure(weights, &_reshape_weights_output, transpose_weights, _is_batched_fc_layer);
     }
 
-    // Check correct shape of weights
-    if(_is_batched_fc_layer)
-    {
-        // Transpose + Transpose1xW
-        ARM_COMPUTE_ERROR_ON(weights_to_use->info()->tensor_shape().x() != linear_input_size * interleave_width);
-        ARM_COMPUTE_ERROR_ON(weights_to_use->info()->tensor_shape().y() != static_cast<unsigned int>(std::ceil(static_cast<float>(output->info()->tensor_shape().x()) / interleave_width)));
-    }
-    else
-    {
-        // Transpose
-        ARM_COMPUTE_ERROR_ON(weights_to_use->info()->tensor_shape().x() != output->info()->tensor_shape().x());
-        ARM_COMPUTE_ERROR_ON(weights_to_use->info()->tensor_shape().y() != linear_input_size);
-    }
-
     const ITensor *multiply_input = input;
 
     if(_linearize_input)
     {
-        TensorShape shape_im2col(input->info()->tensor_shape());
-        shape_im2col.collapse(num_input_dimensions);
-        _im2col_output.allocator()->init(TensorInfo(shape_im2col, 1, data_type, fixed_point_position));
+        _im2col_output.allocator()->init(input->info()->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(compute_im2col_shape(input->info(), num_input_dimensions)));
 
         // Configure im2col kernel
         _memory_group.manage(&_im2col_output);
-        _im2col_kernel.configure(input, &_im2col_output, Size2D(1, 1), PadStrideInfo(1, 1, 0, 0), false);
+        _im2col_kernel.configure(input, &_im2col_output, Size2D(1, 1), PadStrideInfo(1, 1, 0, 0), false, true);
 
         multiply_input = &_im2col_output;
     }
 
+    int m = multiply_input->info()->dimension(1);
+    int k = multiply_input->info()->dimension(0);
+
     if(_is_batched_fc_layer)
     {
-        TensorShape shape_interleaved(multiply_input->info()->tensor_shape());
-        shape_interleaved.set(0, shape_interleaved.x() * 4);
-        shape_interleaved.set(1, std::ceil(shape_interleaved.y() / 4.f));
-        _interleave4x4_output.allocator()->init(TensorInfo(shape_interleaved, 1, data_type, fixed_point_position));
+        _interleave4x4_output.allocator()->init(input->info()->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(compute_interleaved_shape(*multiply_input->info())));
 
         // Configure interleave4x4 kernel
         _memory_group.manage(&_interleave4x4_output);
@@ -217,13 +211,10 @@ void NEFullyConnectedLayer::configure(const ITensor *input, const ITensor *weigh
     }
 
     // Configure matrix multiply kernel
-    _mm_kernel.configure(multiply_input, weights_to_use, output, 1.0f);
+    _mm_kernel.configure(multiply_input, weights_to_use, output, 1.0f, _is_batched_fc_layer, GEMMReshapeInfo(m, 0 /* no transpose */, k));
 
     if(_accumulate_biases)
     {
-        ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, biases);
-        ARM_COMPUTE_ERROR_ON(biases->info()->tensor_shape().x() != output->info()->tensor_shape().x());
-
         // Configure accumulate biases kernel
         _accumulate_biases_kernel.configure(output, biases);
     }
@@ -246,6 +237,88 @@ void NEFullyConnectedLayer::configure(const ITensor *input, const ITensor *weigh
     }
 }
 
+Status NEFullyConnectedLayer::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, bool transpose_weights, bool are_weights_reshaped)
+{
+    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, output);
+    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT_POSITION(input, weights, output);
+
+    const int    num_batch_dimensions = std::max(0, static_cast<int>(output->tensor_shape().num_dimensions()) - 1);
+    const int    num_input_dimensions = input->tensor_shape().num_dimensions() - num_batch_dimensions;
+    const size_t linear_input_size    = input->tensor_shape().total_size_lower(num_input_dimensions);
+
+    const bool linearize_input     = (input->tensor_shape().x() != linear_input_size) || (num_input_dimensions > 1 && linear_input_size == 1);
+    const bool accumulate_biases   = biases != nullptr;
+    const bool is_batched_fc_layer = num_batch_dimensions > 0;
+
+    ARM_COMPUTE_RETURN_ERROR_ON(input->tensor_shape().total_size_upper(num_input_dimensions) != output->tensor_shape().total_size_upper(1));
+    ARM_COMPUTE_RETURN_ERROR_ON(weights->num_dimensions() > 2);
+
+    const size_t                 interleave_width       = 16 / input->element_size();
+    const ITensorInfo           *weights_to_use         = weights;
+    std::unique_ptr<ITensorInfo> reshape_weights_output = input->clone();
+
+    if(!are_weights_reshaped && (transpose_weights || is_batched_fc_layer))
+    {
+        reshape_weights_output->set_tensor_shape(compute_fully_connected_reshaped_weights_shape(weights, transpose_weights, is_batched_fc_layer, interleave_width));
+
+        ARM_COMPUTE_RETURN_ON_ERROR(NEFullyConnectedLayerReshapeWeights::validate(weights, reshape_weights_output.get(), transpose_weights, is_batched_fc_layer));
+
+        weights_to_use = reshape_weights_output.get();
+    }
+
+    // Check correct shape of weights
+    if(is_batched_fc_layer)
+    {
+        // Transpose + Transpose1xW
+        ARM_COMPUTE_RETURN_ERROR_ON(weights_to_use->tensor_shape().x() != linear_input_size * interleave_width);
+        ARM_COMPUTE_RETURN_ERROR_ON(weights_to_use->tensor_shape().y() != static_cast<unsigned int>(std::ceil(static_cast<float>(output->tensor_shape().x()) / interleave_width)));
+    }
+    else
+    {
+        // Transpose
+        ARM_COMPUTE_RETURN_ERROR_ON(weights_to_use->tensor_shape().x() != output->tensor_shape().x());
+        ARM_COMPUTE_RETURN_ERROR_ON(weights_to_use->tensor_shape().y() != linear_input_size);
+    }
+
+    const ITensorInfo           *multiply_input       = input;
+    std::unique_ptr<ITensorInfo> im2col_output        = input->clone();
+    std::unique_ptr<ITensorInfo> interleave4x4_output = input->clone();
+
+    if(linearize_input)
+    {
+        im2col_output->set_tensor_shape(compute_im2col_shape(input, num_input_dimensions));
+
+        ARM_COMPUTE_RETURN_ON_ERROR(NEIm2ColKernel::validate(input, im2col_output.get(), Size2D(1, 1), PadStrideInfo(1, 1, 0, 0), false, true));
+
+        multiply_input = im2col_output.get();
+    }
+
+    int m = multiply_input->dimension(1);
+    int k = multiply_input->dimension(0);
+
+    if(is_batched_fc_layer)
+    {
+        interleave4x4_output->set_tensor_shape(compute_interleaved_shape(*multiply_input));
+
+        ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMInterleave4x4Kernel::validate(multiply_input, interleave4x4_output.get()));
+
+        multiply_input = interleave4x4_output.get();
+    }
+
+    ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMMatrixMultiplyKernel::validate(multiply_input, weights_to_use, output, 1.0f, is_batched_fc_layer, GEMMReshapeInfo(m, 0 /* no transpose */, k)));
+
+    if(accumulate_biases)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, biases);
+        ARM_COMPUTE_RETURN_ERROR_ON(biases->tensor_shape().x() != output->tensor_shape().x());
+
+        ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMMatrixAccumulateBiasesKernel::validate(output, biases));
+    }
+
+    return Status{};
+}
+
 void NEFullyConnectedLayer::run()
 {
     // Reshape of the weights (happens only once)
@@ -280,4 +353,3 @@ void NEFullyConnectedLayer::run()
 
     _memory_group.release();
 }
-} // namespace arm_compute
diff --git a/src/runtime/NEON/functions/NEGEMM.cpp b/src/runtime/NEON/functions/NEGEMM.cpp
index 48a0d2af1c..05907bab07 100644
--- a/src/runtime/NEON/functions/NEGEMM.cpp
+++ b/src/runtime/NEON/functions/NEGEMM.cpp
@@ -120,7 +120,7 @@ void NEGEMM::configure(const ITensor *a, const ITensor *b, const ITensor *c, ITe
 #endif /* defined(__aarch64__) */
         {
             // Configure the matrix multiply kernel
-            _mm_kernel.configure(a, b, d, alpha);
+            _mm_kernel.configure(a, b, d, alpha, false);
         }
 
         // Configure matrix addition kernel
@@ -212,6 +212,10 @@ void NEGEMM::configure(const ITensor *a, const ITensor *b, const ITensor *c, ITe
             _memory_group.manage(&_tmp_a);
             _memory_group.manage(&_tmp_b);
 
+            int m = a->info()->dimension(1);
+            int n = b->info()->dimension(0);
+            int k = a->info()->dimension(0);
+
             // Configure interleave kernel
             _interleave_kernel.configure(a, &_tmp_a);
 
@@ -219,7 +223,7 @@ void NEGEMM::configure(const ITensor *a, const ITensor *b, const ITensor *c, ITe
             _transpose_kernel.configure(b, &_tmp_b);
 
             // Configure matrix multiplication kernel
-            _mm_kernel.configure(&_tmp_a, &_tmp_b, d, alpha);
+            _mm_kernel.configure(&_tmp_a, &_tmp_b, d, alpha, true, GEMMReshapeInfo(m, n, k));
 
             // Allocate once the all configure methods have been called
             _tmp_a.allocator()->allocate();
diff --git a/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp b/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp
index d0a16ef40d..a85078cf71 100644
--- a/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp
+++ b/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp
@@ -178,7 +178,7 @@ TensorShape get_reshaped_weights_shape_conv(const ITensorInfo *weights, bool app
 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,
                                       bool &are_weights_reshaped, unsigned int &kernel_width, unsigned int &kernel_height,
-                                      bool &is_fully_connected_convolution, bool &is_interleaved_transposed, bool &is_quantized,
+                                      bool &is_fully_connected_convolution, bool &is_interleaved, bool &is_quantized,
                                       unsigned int &mat_weights_cols, unsigned int &mat_weights_rows,
                                       unsigned int &conv_w, unsigned int &conv_h)
 {
@@ -219,7 +219,7 @@ Status validate_and_initialize_values(const ITensorInfo *input, const ITensorInf
 
     // Check if its a "fully connected" convolution
     is_fully_connected_convolution = ((conv_w == 1) && (conv_h == 1));
-    is_interleaved_transposed      = (!is_fully_connected_convolution && !is_quantized);
+    is_interleaved                 = (!is_fully_connected_convolution && !is_quantized);
 
     return Status{};
 }
@@ -228,11 +228,11 @@ Status validate_and_initialize_values(const ITensorInfo *input, const ITensorInf
 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_transposed(false)
+      _is_fully_connected_convolution(false), _are_weights_reshaped(false), _is_quantized(false), _is_interleaved(false)
 {
 }
 
-void NEGEMMConvolutionLayer::configure_mm(const ITensor *input, const ITensor *weights, ITensor *output)
+void NEGEMMConvolutionLayer::configure_mm(const ITensor *input, const ITensor *weights, ITensor *output, bool is_interleaved, const GEMMReshapeInfo &reshape_info)
 {
     if(_is_quantized)
     {
@@ -252,7 +252,7 @@ void NEGEMMConvolutionLayer::configure_mm(const ITensor *input, const ITensor *w
     }
     else
     {
-        _mm_kernel.configure(input, weights, output, 1.f);
+        _mm_kernel.configure(input, weights, output, 1.f, is_interleaved, reshape_info);
     }
 }
 
@@ -290,7 +290,7 @@ void NEGEMMConvolutionLayer::configure(const ITensor *input, const ITensor *weig
 
     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,
                                                    kernel_width, kernel_height,
-                                                   _is_fully_connected_convolution, _is_interleaved_transposed, _is_quantized,
+                                                   _is_fully_connected_convolution, _is_interleaved, _is_quantized,
                                                    mat_weights_cols, mat_weights_rows, conv_w, conv_h);
 
     ARM_COMPUTE_ERROR_THROW_ON(status);
@@ -339,9 +339,8 @@ void NEGEMMConvolutionLayer::configure(const ITensor *input, const ITensor *weig
             }
             else
             {
-                const unsigned int transpose_width = 16 / input->info()->element_size();
-                mat_weights_cols                   = weights_info.num_kernels();
-                mat_weights_rows                   = weights->info()->dimension(0) / transpose_width + (_append_bias ? 1 : 0);
+                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);
             }
         }
         else
@@ -362,7 +361,7 @@ void NEGEMMConvolutionLayer::configure(const ITensor *input, const ITensor *weig
 
             // 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_to_use, &_weights_reshaped, _is_interleaved_transposed /* 1xW transpose */);
+            _reshape_weights.configure(weights, biases_to_use, &_weights_reshaped, _is_interleaved /* 1xW transpose */);
             weights = &_weights_reshaped;
         }
     }
@@ -430,18 +429,19 @@ void NEGEMMConvolutionLayer::configure(const ITensor *input, const ITensor *weig
     }
     else
     {
-        if(_is_interleaved_transposed)
+        if(_is_interleaved)
         {
             // Configure GEMMInterleave4x4. _input_interleaved_reshaped will be auto configured in the kernel
             _input_interleave_kernel.configure(&_input_im2col_reshaped, &_input_interleaved_reshaped);
 
             // Configure GEMM
-            configure_mm(&_input_interleaved_reshaped, weights, &_gemm_output);
+            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)));
             _input_interleaved_reshaped.allocator()->allocate();
         }
         else
         {
-            configure_mm(&_input_im2col_reshaped, weights, &_gemm_output);
+            configure_mm(&_input_im2col_reshaped, weights, &_gemm_output, _is_interleaved);
         }
     }
 
@@ -479,11 +479,13 @@ void NEGEMMConvolutionLayer::configure(const ITensor *input, const ITensor *weig
 Status NEGEMMConvolutionLayer::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info,
                                         const WeightsInfo &weights_info)
 {
+    ARM_COMPUTE_UNUSED(output);
+
     DataType     dt{};
     bool         append_bias{};
     bool         are_weights_reshaped{};
     bool         is_fully_connected_convolution{};
-    bool         is_interleaved_transposed{};
+    bool         is_interleaved{};
     bool         is_quantized{};
     unsigned int kernel_width     = 0;
     unsigned int kernel_height    = 0;
@@ -493,9 +495,11 @@ Status NEGEMMConvolutionLayer::validate(const ITensorInfo *input, const ITensorI
     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_transposed, is_quantized, mat_weights_cols, mat_weights_rows,
+                                                   is_fully_connected_convolution, is_interleaved, is_quantized, mat_weights_cols, mat_weights_rows,
                                                    conv_w, conv_h);
 
+    const Size2D kernel_weights = Size2D(kernel_width, kernel_height);
+
     ARM_COMPUTE_RETURN_ON_ERROR(status);
 
     std::unique_ptr<ITensorInfo> reshaped_weights = weights->clone();
@@ -570,7 +574,7 @@ Status NEGEMMConvolutionLayer::validate(const ITensorInfo *input, const ITensorI
     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, append_bias));
+    ARM_COMPUTE_RETURN_ON_ERROR(NEIm2ColKernel::validate(input, &im2_col_info, kernel_weights, conv_info, append_bias, false));
 
     // Create GEMM output tensor
     TensorShape shape_gemm(im2_col_info.tensor_shape());
@@ -579,24 +583,20 @@ Status NEGEMMConvolutionLayer::validate(const ITensorInfo *input, const ITensorI
     TensorInfo gemm_output_info = input->clone()->set_tensor_shape(shape_gemm);
 
     // Validate GEMM interleave and multiply
-    if(is_interleaved_transposed)
+    if(is_interleaved)
     {
         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));
+        ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMMatrixMultiplyKernel::validate(&input_interleaved_info, weights, &gemm_output_info, 1.f, is_interleaved, GEMMReshapeInfo()));
     }
     else
     {
-        ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMMatrixMultiplyKernel::validate(&im2_col_info, weights, &gemm_output_info));
+        ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMMatrixMultiplyKernel::validate(&im2_col_info, weights, &gemm_output_info, 1.f, is_interleaved, GEMMReshapeInfo()));
     }
 
-    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{};
 }
 
@@ -621,7 +621,7 @@ void NEGEMMConvolutionLayer::run()
     }
     else
     {
-        if(_is_interleaved_transposed)
+        if(_is_interleaved)
         {
             // Run interleave
             NEScheduler::get().schedule(&_input_interleave_kernel, Window::DimY);
diff --git a/src/runtime/NEON/functions/NEGEMMTranspose1xW.cpp b/src/runtime/NEON/functions/NEGEMMTranspose1xW.cpp
index 571bf2bc74..802b94650e 100644
--- a/src/runtime/NEON/functions/NEGEMMTranspose1xW.cpp
+++ b/src/runtime/NEON/functions/NEGEMMTranspose1xW.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -38,3 +38,7 @@ void NEGEMMTranspose1xW::configure(const ITensor *input, ITensor *output)
     k->configure(input, output);
     _kernel = std::move(k);
 }
+Status NEGEMMTranspose1xW::validate(const ITensorInfo *input, const ITensorInfo *output)
+{
+    return NEGEMMTranspose1xWKernel::validate(input, output);
+}
diff --git a/src/runtime/NEON/functions/NEIm2Col.cpp b/src/runtime/NEON/functions/NEIm2Col.cpp
index 8e90e66dcc..b962db9144 100644
--- a/src/runtime/NEON/functions/NEIm2Col.cpp
+++ b/src/runtime/NEON/functions/NEIm2Col.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -28,14 +28,14 @@
 
 using namespace arm_compute;
 
-void NEIm2Col::configure(const ITensor *input, ITensor *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias)
+void NEIm2Col::configure(const ITensor *input, ITensor *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, bool is_fully_connected)
 {
     auto k = arm_compute::support::cpp14::make_unique<NEIm2ColKernel>();
-    k->configure(input, output, kernel_dims, conv_info, has_bias);
+    k->configure(input, output, kernel_dims, conv_info, has_bias, is_fully_connected);
     _kernel = std::move(k);
 }
 
-Status NEIm2Col::validate(const ITensorInfo *input, const ITensorInfo *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias)
+Status NEIm2Col::validate(const ITensorInfo *input, const ITensorInfo *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, bool is_fully_connected)
 {
-    return NEIm2ColKernel::validate(input, output, kernel_dims, conv_info, has_bias);
+    return NEIm2ColKernel::validate(input, output, kernel_dims, conv_info, has_bias, is_fully_connected);
 }