COMPMID-1386: Add support for converting weights for CL.

Change-Id: I62e3ead903366baeeb1488f233a9b8b0c388c9de
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/140403
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Giorgio Arena <giorgio.arena@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>

10 files changed, 201 insertions, 115 deletions

diff --git a/src/core/CL/kernels/CLConvertFullyConnectedWeightsKernel.cpp b/src/core/CL/kernels/CLConvertFullyConnectedWeightsKernel.cpp
index 86858d0c03..69ab590540 100644
--- a/src/core/CL/kernels/CLConvertFullyConnectedWeightsKernel.cpp
+++ b/src/core/CL/kernels/CLConvertFullyConnectedWeightsKernel.cpp
@@ -41,27 +41,32 @@ void CLConvertFullyConnectedWeightsKernel::configure(const ICLTensor *input, ICL
                                                      DataLayout data_layout)
 {
     ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
+
+    // Output tensor auto initialisation if not yet initialized
+    auto_init_if_empty(*output->info(), *input->info()->clone());
+
     ARM_COMPUTE_ERROR_THROW_ON(CLConvertFullyConnectedWeightsKernel::validate(input->info(), output->info(), original_input_shape, data_layout));
 
     _input  = input;
     _output = output;
 
-    const unsigned int num_elems_per_input_plane = original_input_shape.x() * original_input_shape.y();
-    const unsigned int num_channels              = original_input_shape.z();
+    const DataLayout input_data_layout = (data_layout == DataLayout::NCHW) ? DataLayout::NHWC : DataLayout::NCHW;
+
+    const int width_idx   = get_data_layout_dimension_index(input_data_layout, DataLayoutDimension::WIDTH);
+    const int height_idx  = get_data_layout_dimension_index(input_data_layout, DataLayoutDimension::HEIGHT);
+    const int channel_idx = get_data_layout_dimension_index(input_data_layout, DataLayoutDimension::CHANNEL);
+
+    const unsigned int num_elems_per_input_plane = original_input_shape[width_idx] * original_input_shape[height_idx];
+    const unsigned int num_channels              = original_input_shape[channel_idx];
+
+    const unsigned int factor_1 = (data_layout == DataLayout::NCHW) ? num_elems_per_input_plane : num_channels;
+    const unsigned int factor_2 = (data_layout == DataLayout::NCHW) ? num_channels : num_elems_per_input_plane;
 
     // Set build options
     CLBuildOptions build_opts;
     build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type()));
-    if(data_layout == DataLayout::NCHW)
-    {
-        build_opts.add_option("-DFACTOR_1=" + support::cpp11::to_string(num_elems_per_input_plane));
-        build_opts.add_option("-DFACTOR_2=" + support::cpp11::to_string(num_channels));
-    }
-    else
-    {
-        build_opts.add_option("-DFACTOR_1=" + support::cpp11::to_string(num_channels));
-        build_opts.add_option("-DFACTOR_2=" + support::cpp11::to_string(num_elems_per_input_plane));
-    }
+    build_opts.add_option("-DFACTOR_1=" + support::cpp11::to_string(factor_1));
+    build_opts.add_option("-DFACTOR_2=" + support::cpp11::to_string(factor_2));
 
     // Create kernel
     _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("convert_fc_weights", build_opts.options()));
@@ -75,14 +80,22 @@ Status CLConvertFullyConnectedWeightsKernel::validate(const ITensorInfo *input,
                                                       DataLayout data_layout)
 {
     ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input);
-    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8, DataType::S8, DataType::QASYMM8, DataType::U16, DataType::S16, DataType::U32, DataType::S32,
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1,
+                                                         DataType::U8, DataType::S8, DataType::QASYMM8,
+                                                         DataType::U16, DataType::S16,
+                                                         DataType::U32, DataType::S32,
                                                          DataType::F16, DataType::F32);
-    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
-    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, output);
     ARM_COMPUTE_RETURN_ERROR_ON(input->num_dimensions() != 2);
     ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(1) != original_input_shape.total_size_lower(3));
     ARM_COMPUTE_RETURN_ERROR_ON(data_layout == DataLayout::UNKNOWN);
 
+    // Checks performed when output is configured
+    if((output != nullptr) && (output->total_size() != 0))
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, output);
+    }
+
     return Status{};
 }
 
diff --git a/src/core/NEON/kernels/NEConvertFullyConnectedWeightsKernel.cpp b/src/core/NEON/kernels/NEConvertFullyConnectedWeightsKernel.cpp
index be5e6436b3..198565b1d5 100644
--- a/src/core/NEON/kernels/NEConvertFullyConnectedWeightsKernel.cpp
+++ b/src/core/NEON/kernels/NEConvertFullyConnectedWeightsKernel.cpp
@@ -37,25 +37,26 @@ void NEConvertFullyConnectedWeightsKernel::configure(const ITensor *input, ITens
                                                      DataLayout data_layout)
 {
     ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
+
+    // Output tensor auto initialisation if not yet initialized
+    auto_init_if_empty(*output->info(), *input->info()->clone());
+
     ARM_COMPUTE_ERROR_THROW_ON(NEConvertFullyConnectedWeightsKernel::validate(input->info(), output->info(), original_input_shape, data_layout));
 
     _input  = input;
     _output = output;
 
-    const unsigned int num_elems_per_input_plane = original_input_shape.x() * original_input_shape.y();
-    const unsigned int num_channels              = original_input_shape.z();
+    const DataLayout input_data_layout = (data_layout == DataLayout::NCHW) ? DataLayout::NHWC : DataLayout::NCHW;
 
-    // Set build options
-    if(data_layout == DataLayout::NCHW)
-    {
-        _factor1 = num_elems_per_input_plane;
-        _factor2 = num_channels;
-    }
-    else
-    {
-        _factor1 = num_channels;
-        _factor2 = num_elems_per_input_plane;
-    }
+    const int width_idx   = get_data_layout_dimension_index(input_data_layout, DataLayoutDimension::WIDTH);
+    const int height_idx  = get_data_layout_dimension_index(input_data_layout, DataLayoutDimension::HEIGHT);
+    const int channel_idx = get_data_layout_dimension_index(input_data_layout, DataLayoutDimension::CHANNEL);
+
+    const unsigned int num_elems_per_input_plane = original_input_shape[width_idx] * original_input_shape[height_idx];
+    const unsigned int num_channels              = original_input_shape[channel_idx];
+
+    _factor1 = (data_layout == DataLayout::NCHW) ? num_elems_per_input_plane : num_channels;
+    _factor2 = (data_layout == DataLayout::NCHW) ? num_channels : num_elems_per_input_plane;
 
     // Configure kernel window
     Window win = calculate_max_window(*input->info(), Steps());
@@ -65,14 +66,22 @@ void NEConvertFullyConnectedWeightsKernel::configure(const ITensor *input, ITens
 Status NEConvertFullyConnectedWeightsKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const TensorShape &original_input_shape,
                                                       DataLayout data_layout)
 {
-    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8, DataType::S8, DataType::QASYMM8, DataType::U16, DataType::S16, DataType::U32, DataType::S32,
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1,
+                                                         DataType::U8, DataType::S8, DataType::QASYMM8,
+                                                         DataType::U16, DataType::S16,
+                                                         DataType::U32, DataType::S32,
                                                          DataType::F16, DataType::F32);
-    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
-    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, output);
     ARM_COMPUTE_RETURN_ERROR_ON(input->num_dimensions() != 2);
     ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(1) != original_input_shape.total_size_lower(3));
     ARM_COMPUTE_RETURN_ERROR_ON(data_layout == DataLayout::UNKNOWN);
 
+    // Checks performed when output is configured
+    if((output != nullptr) && (output->total_size() != 0))
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, output);
+    }
+
     return Status{};
 }
 
diff --git a/src/graph/GraphBuilder.cpp b/src/graph/GraphBuilder.cpp
index 3b906c0ffd..d26039ec35 100644
--- a/src/graph/GraphBuilder.cpp
+++ b/src/graph/GraphBuilder.cpp
@@ -446,8 +446,13 @@ NodeID GraphBuilder::add_fully_connected_layer(Graph &g, NodeParams params, Node
         b_nid = add_const_node_with_name(g, params, "Bias", b_desc, std::move(bias_accessor));
     }
 
-    // Create convolution node and connect
-    NodeID fc_nid = g.add_node<FullyConnectedLayerNode>(num_outputs);
+    // Add fully connected info
+    // FIXME (COMPMID-1367) : Expose weights layout
+    FullyConnectedLayerInfo fc_info;
+    fc_info.weights_trained_layout = DataLayout::NCHW;
+
+    // Create fully connected node and connect
+    NodeID fc_nid = g.add_node<FullyConnectedLayerNode>(num_outputs, fc_info);
     g.add_connection(input.node_id, input.index, fc_nid, 0);
     g.add_connection(w_nid, 0, fc_nid, 1);
     if(has_bias)
diff --git a/src/graph/nodes/FullyConnectedLayer.cpp b/src/graph/nodes/FullyConnectedLayer.cpp
index d94a7851ff..6ba4eefa88 100644
--- a/src/graph/nodes/FullyConnectedLayer.cpp
+++ b/src/graph/nodes/FullyConnectedLayer.cpp
@@ -31,8 +31,8 @@ namespace arm_compute
 {
 namespace graph
 {
-FullyConnectedLayerNode::FullyConnectedLayerNode(unsigned int num_outputs)
-    : _num_outputs(num_outputs)
+FullyConnectedLayerNode::FullyConnectedLayerNode(unsigned int num_outputs, FullyConnectedLayerInfo fc_info)
+    : _num_outputs(num_outputs), _info(fc_info)
 {
     _input_edges.resize(3, EmptyEdgeID);
     _outputs.resize(1, NullTensorID);
@@ -75,6 +75,11 @@ TensorDescriptor FullyConnectedLayerNode::compute_output_descriptor(const Tensor
     return output_descriptor;
 }
 
+FullyConnectedLayerInfo FullyConnectedLayerNode::info() const
+{
+    return _info;
+}
+
 bool FullyConnectedLayerNode::forward_descriptors()
 {
     if((input_id(0) != NullTensorID) && (output_id(0) != NullTensorID))
diff --git a/src/runtime/CL/functions/CLFullyConnectedLayer.cpp b/src/runtime/CL/functions/CLFullyConnectedLayer.cpp
index 273ef96a03..ccd7813fbc 100644
--- a/src/runtime/CL/functions/CLFullyConnectedLayer.cpp
+++ b/src/runtime/CL/functions/CLFullyConnectedLayer.cpp
@@ -73,8 +73,9 @@ Status CLFullyConnectedLayerReshapeWeights::validate(const ITensorInfo *input, c
 }
 
 CLFullyConnectedLayer::CLFullyConnectedLayer(std::shared_ptr<IMemoryManager> memory_manager)
-    : _memory_group(memory_manager), _im2col_kernel(), _reshape_weights_kernel(), _mm_gemm(memory_manager), _mm_gemmlowp(memory_manager), _gemmlowp_output_stage(), _accumulate_biases_kernel(),
-      _im2col_output(), _gemmlowp_output(), _reshape_weights_output(), _are_weights_reshaped(true), _is_fc_after_conv(true), _accumulate_biases(false), _is_quantized(false), _original_weights(nullptr)
+    : _memory_group(memory_manager), _im2col_kernel(), _convert_weights(), _reshape_weights_kernel(), _mm_gemm(memory_manager), _mm_gemmlowp(memory_manager), _gemmlowp_output_stage(),
+      _accumulate_biases_kernel(), _im2col_output(), _gemmlowp_output(), _converted_weights_output(), _reshape_weights_output(), _are_weights_converted(true), _are_weights_reshaped(true),
+      _is_fc_after_conv(true), _accumulate_biases(false), _is_quantized(false), _is_prepared(false), _original_weights(nullptr)
 {
 }
 
@@ -112,7 +113,7 @@ void CLFullyConnectedLayer::configure_conv_fc(const ICLTensor *input, const ICLT
 
     // Initialize output tensor for im2col
     TensorShape shape_im2col = compute_im2col_fc_shape(input->info());
-    _im2col_output.allocator()->init(input->info()->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(shape_im2col));
+    _im2col_output.allocator()->init(input->info()->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(shape_im2col).set_data_layout(DataLayout::NCHW));
 
     // Configure im2col kernel
     _memory_group.manage(&_im2col_output);
@@ -134,8 +135,8 @@ void CLFullyConnectedLayer::configure_fc_fc(const ICLTensor *input, const ICLTen
     configure_mm(input, weights, output);
 }
 
-void CLFullyConnectedLayer::configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, bool transpose_weights, bool are_weights_reshaped,
-                                      bool retain_internal_weights)
+void CLFullyConnectedLayer::configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output,
+                                      FullyConnectedLayerInfo fc_info)
 {
     ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output);
 
@@ -144,15 +145,15 @@ void CLFullyConnectedLayer::configure(const ICLTensor *input, const ICLTensor *w
                                                                weights->info(),
                                                                biases != nullptr ? biases->info() : nullptr,
                                                                output->info(),
-                                                               transpose_weights,
-                                                               are_weights_reshaped,
-                                                               retain_internal_weights));
+                                                               fc_info));
 
-    _are_weights_reshaped = transpose_weights ? are_weights_reshaped : true;
-    _is_fc_after_conv     = true;
-    _accumulate_biases    = false;
-    _is_quantized         = is_data_type_quantized_asymmetric(input->info()->data_type());
-    _original_weights     = weights;
+    _are_weights_converted = true;
+    _are_weights_reshaped  = fc_info.transpose_weights ? fc_info.are_weights_reshaped : true;
+    _is_fc_after_conv      = true;
+    _accumulate_biases     = false;
+    _is_quantized          = is_data_type_quantized_asymmetric(input->info()->data_type());
+    _is_prepared           = false;
+    _original_weights      = weights;
 
     // Configure gemmlowp output
     if(_is_quantized)
@@ -172,25 +173,16 @@ void CLFullyConnectedLayer::configure(const ICLTensor *input, const ICLTensor *w
         _accumulate_biases_kernel.configure(output, biases);
     }
 
+    const ICLTensor *weights_to_use = weights;
+
     // With the Fully Connected layer we can have 4 different cases:
     //  1) Convolution layer -> Fully Connected layer without batches
     //  2) Fully Connected layer -> Fully Connected layer without batches
     //  3) Convolution layer -> Fully Connected layer with batches
     //  4) Fully Connected layer -> Fully Connected layer with batches
 
-    const ICLTensor *weights_to_use = weights;
-
-    if(!_are_weights_reshaped)
-    {
-        weights_to_use = &_reshape_weights_output;
-
-        // Reshape the weights
-        _reshape_weights_kernel.configure(weights, &_reshape_weights_output);
-    }
-
     // Check if we have a fully connected layer with batches
     const bool is_batched_fc_layer = output->info()->dimension(1) > 1;
-
     if(is_batched_fc_layer)
     {
         _is_fc_after_conv = (TensorShape::num_max_dimensions >= 4) && (std::equal(input->info()->tensor_shape().cbegin() + 3,
@@ -202,6 +194,28 @@ void CLFullyConnectedLayer::configure(const ICLTensor *input, const ICLTensor *w
         _is_fc_after_conv = input->info()->num_dimensions() > 1;
     }
 
+    // Reshape weights if needed
+    if(!_are_weights_reshaped)
+    {
+        // Reshape the weights
+        _reshape_weights_kernel.configure(weights, &_reshape_weights_output);
+        weights_to_use = &_reshape_weights_output;
+    }
+
+    // Convert weights if needed
+    if(_is_fc_after_conv && (input->info()->data_layout() != fc_info.weights_trained_layout))
+    {
+        // Convert weights
+        _convert_weights.configure(weights_to_use,
+                                   &_converted_weights_output,
+                                   input->info()->tensor_shape(),
+                                   fc_info.weights_trained_layout);
+
+        weights_to_use         = &_converted_weights_output;
+        _are_weights_converted = false;
+    }
+
+    // Configure fc core
     ICLTensor *tmp_output = (_is_quantized) ? &_gemmlowp_output : output;
     if(_is_fc_after_conv)
     {
@@ -224,26 +238,26 @@ void CLFullyConnectedLayer::configure(const ICLTensor *input, const ICLTensor *w
         _gemmlowp_output.allocator()->allocate();
     }
 
-    _are_weights_reshaped = _are_weights_reshaped || retain_internal_weights;
+    _are_weights_reshaped = _are_weights_reshaped || fc_info.retain_internal_weights;
 }
 
-Status CLFullyConnectedLayer::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, bool transpose_weights, bool are_weights_reshaped,
-                                       bool retain_internal_weights)
+Status CLFullyConnectedLayer::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output,
+                                       FullyConnectedLayerInfo fc_info)
 {
-    ARM_COMPUTE_UNUSED(retain_internal_weights);
     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, weights, output);
     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F16, DataType::F32);
     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights, output);
     ARM_COMPUTE_RETURN_ERROR_ON(weights->num_dimensions() > 2);
 
-    bool            weights_reshaped = transpose_weights ? are_weights_reshaped : true;
+    bool            weights_reshaped = fc_info.transpose_weights ? fc_info.are_weights_reshaped : true;
     bool            is_fc_after_conv = true;
     bool            is_quantized     = is_data_type_quantized_asymmetric(input->data_type());
     const GPUTarget gpu_target       = CLScheduler::get().target();
 
-    const ITensorInfo &im2col_input     = TensorInfo(input->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(compute_im2col_fc_shape(input)));
-    const ITensorInfo &reshaped_weights = TensorInfo(weights->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(compute_transposed_shape(*weights)));
-    const ITensorInfo &gemmlowp_output  = TensorInfo(output->clone()->set_is_resizable(true).reset_padding().set_data_type(DataType::S32));
+    const ITensorInfo &im2col_input      = TensorInfo(input->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(compute_im2col_fc_shape(input)).set_data_layout(DataLayout::NCHW));
+    const ITensorInfo &reshaped_weights  = TensorInfo(weights->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(compute_transposed_shape(*weights)));
+    const ITensorInfo &converted_weights = TensorInfo(reshaped_weights.clone()->set_is_resizable(true).reset_padding());
+    const ITensorInfo &gemmlowp_output   = TensorInfo(output->clone()->set_is_resizable(true).reset_padding().set_data_type(DataType::S32));
 
     // Configure accumulate biases kernel for non quantized asymmetric types
     if(biases != nullptr && !is_quantized)
@@ -262,16 +276,8 @@ Status CLFullyConnectedLayer::validate(const ITensorInfo *input, const ITensorIn
     const ITensorInfo *weights_to_use = weights;
     const ITensorInfo *tmp_output     = (is_quantized) ? &gemmlowp_output : output;
 
-    if(!weights_reshaped)
-    {
-        // Validate reshape weights kernel
-        ARM_COMPUTE_RETURN_ON_ERROR(CLFullyConnectedLayerReshapeWeights::validate(weights, &reshaped_weights));
-        weights_to_use = &reshaped_weights;
-    }
-
     // Check if we have a fully connected layer with batches
     const bool is_batched_fc_layer = output->dimension(1) > 1;
-
     if(is_batched_fc_layer)
     {
         is_fc_after_conv = (TensorShape::num_max_dimensions >= 4) && (std::equal(input->tensor_shape().cbegin() + 3,
@@ -283,6 +289,23 @@ Status CLFullyConnectedLayer::validate(const ITensorInfo *input, const ITensorIn
         is_fc_after_conv = input->num_dimensions() > 1;
     }
 
+    if(!weights_reshaped)
+    {
+        // Validate reshape weights kernel
+        ARM_COMPUTE_RETURN_ON_ERROR(CLFullyConnectedLayerReshapeWeights::validate(weights, &reshaped_weights));
+        weights_to_use = &reshaped_weights;
+    }
+
+    if(is_fc_after_conv && (input->data_layout() != fc_info.weights_trained_layout))
+    {
+        // Validate convert weights kernel
+        ARM_COMPUTE_RETURN_ON_ERROR(CLConvertFullyConnectedWeights::validate(weights_to_use,
+                                                                             &converted_weights,
+                                                                             input->tensor_shape(),
+                                                                             fc_info.weights_trained_layout));
+        weights_to_use = &converted_weights;
+    }
+
     if(is_fc_after_conv)
     {
         // Fully Connected layer after a Convolution Layer without batches
@@ -349,27 +372,57 @@ void CLFullyConnectedLayer::run()
 
 void CLFullyConnectedLayer::prepare()
 {
-    // Reshape of the weights (happens only once)
-    if(!_are_weights_reshaped)
+    if(!_is_prepared)
     {
         ARM_COMPUTE_ERROR_ON(!_original_weights->is_used());
 
-        // Run reshape weights kernel and mark weights as unused
-        _reshape_weights_output.allocator()->allocate();
-        _reshape_weights_kernel.run();
-        _original_weights->mark_as_unused();
+        auto release_unused = [](CLTensor * w)
+        {
+            if(!w->is_used())
+            {
+                CLScheduler::get().queue().finish();
+                w->allocator()->free();
+            }
+        };
+
+        // Pointer to current weights
+        const ICLTensor *cur_weights = _original_weights;
+
+        // Reshape of the weights if needed (happens only once)
+        if(!_are_weights_reshaped)
+        {
+            // Run reshape weights kernel and mark weights as unused
+            _reshape_weights_output.allocator()->allocate();
+            _reshape_weights_kernel.run();
+
+            cur_weights->mark_as_unused();
+            cur_weights           = &_reshape_weights_output;
+            _are_weights_reshaped = true;
+        }
+
+        // Convert weights if needed (happens only once)
+        if(!_are_weights_converted)
+        {
+            _converted_weights_output.allocator()->allocate();
+            _convert_weights.run();
+
+            cur_weights->mark_as_unused();
+            _are_weights_converted = true;
+        }
+
+        // Release reshaped weights if unused
+        release_unused(&_reshape_weights_output);
 
         // Prepare GEMM prepare and release unused weights
         if(!_is_quantized)
         {
             _mm_gemm.prepare();
-            if(!_reshape_weights_output.is_used())
-            {
-                _reshape_weights_output.allocator()->free();
-            }
         }
 
-        CLScheduler::get().queue().finish();
-        _are_weights_reshaped = true;
+        // Release converted weights if unused
+        release_unused(&_reshape_weights_output);
+        release_unused(&_converted_weights_output);
+
+        _is_prepared = true;
     }
 }
diff --git a/src/runtime/CL/functions/CLLSTMLayer.cpp b/src/runtime/CL/functions/CLLSTMLayer.cpp
index 872325175d..d384400ed3 100644
--- a/src/runtime/CL/functions/CLLSTMLayer.cpp
+++ b/src/runtime/CL/functions/CLLSTMLayer.cpp
@@ -90,7 +90,7 @@ void CLLSTMLayer::configure(const ICLTensor *input, const ICLTensor *input_to_fo
     // Configure block that calculates the forget gate
     // forget_gate = Activation(input * input_to_forget_weights + output_state * recurrent_to_forget_weights + PixelWiseMul(cell_state, cell_to_forget_weights) + forget_gate_bias)
     _memory_group.manage(&_forget_gate_out1);
-    _fully_connected_forget_gate.configure(input, input_to_forget_weights, forget_gate_bias, &_forget_gate_out1, true, false);
+    _fully_connected_forget_gate.configure(input, input_to_forget_weights, forget_gate_bias, &_forget_gate_out1);
     _memory_group.manage(&_forget_gate_out2);
     _transpose_forget_gate.configure(recurrent_to_forget_weights, &_forget_gate_out2);
     _memory_group.manage(&_forget_gate_out3);
@@ -142,7 +142,7 @@ void CLLSTMLayer::configure(const ICLTensor *input, const ICLTensor *input_to_fo
         _input_gate_out5.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
 
         _memory_group.manage(&_input_gate_out1);
-        _fully_connected_input_gate.configure(input, lstm_params.input_to_input_weights(), lstm_params.input_gate_bias(), &_input_gate_out1, true, false);
+        _fully_connected_input_gate.configure(input, lstm_params.input_to_input_weights(), lstm_params.input_gate_bias(), &_input_gate_out1);
         _memory_group.manage(&_input_gate_out2);
         _transpose_input_gate.configure(lstm_params.recurrent_to_input_weights(), &_input_gate_out2);
         _memory_group.manage(&_input_gate_out3);
@@ -169,7 +169,7 @@ void CLLSTMLayer::configure(const ICLTensor *input, const ICLTensor *input_to_fo
     // Configure block that calculates the cell state
     // cell_state = Clip((PixelwiseMul(input_gate, Activation(input * input_to_cell_weights + output_state * recurrent_to_cell_weights + cell_bias)) + PixelwiseMul(forget_gate, cell_state)), cell_threshold)
     _memory_group.manage(&_cell_state_out1);
-    _fully_connected_cell_state.configure(input, input_to_cell_weights, cell_bias, &_cell_state_out1, true, false);
+    _fully_connected_cell_state.configure(input, input_to_cell_weights, cell_bias, &_cell_state_out1);
     _memory_group.manage(&_cell_state_out2);
     _transpose_cell_state.configure(recurrent_to_cell_weights, &_cell_state_out2);
     _memory_group.manage(&_cell_state_out3);
@@ -204,7 +204,7 @@ void CLLSTMLayer::configure(const ICLTensor *input, const ICLTensor *input_to_fo
     // Configure block that calculates the output
     // output_state = Activation(input * input_to_output_weights + output_state * recurrent_to_output_weights + PixelWiseMul(cell_state, cell_to_output_weights) + output_gate_bias)
     _memory_group.manage(&_output1);
-    _fully_connected_output.configure(input, input_to_output_weights, output_gate_bias, &_output1, true, false);
+    _fully_connected_output.configure(input, input_to_output_weights, output_gate_bias, &_output1);
     _memory_group.manage(&_output2);
     _transpose_output.configure(recurrent_to_output_weights, &_output2);
     _memory_group.manage(&_output3);
@@ -255,7 +255,7 @@ void CLLSTMLayer::configure(const ICLTensor *input, const ICLTensor *input_to_fo
         _has_projection_weights = true;
         _output_projection1.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
         _memory_group.manage(&_output_projection1);
-        _fully_connected_output_state.configure(output_state, lstm_params.projection_weights(), lstm_params.projection_bias(), &_output_projection1, true, false);
+        _fully_connected_output_state.configure(output_state, lstm_params.projection_weights(), lstm_params.projection_bias(), &_output_projection1);
         // Perform clipping
         if(projection_threshold != 0.f)
         {
@@ -326,7 +326,7 @@ Status CLLSTMLayer::validate(const ITensorInfo *input, const ITensorInfo *input_
     const TensorInfo num_units_transposed_info  = TensorInfo(num_units_transposed_shape, 1, input->data_type());
 
     // Validate forget gate
-    ARM_COMPUTE_RETURN_ON_ERROR(CLFullyConnectedLayer::validate(input, input_to_forget_weights, forget_gate_bias, cell_state, true, false));
+    ARM_COMPUTE_RETURN_ON_ERROR(CLFullyConnectedLayer::validate(input, input_to_forget_weights, forget_gate_bias, cell_state));
     ARM_COMPUTE_RETURN_ON_ERROR(CLGEMM::validate(output_state, &units_out_transposed_info, nullptr, cell_state, 1.f, 0.f, GEMMInfo()));
     ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAdditionKernel::validate(cell_state, cell_state, cell_state, ConvertPolicy::SATURATE));
     if(lstm_params.has_peephole_opt())
@@ -344,7 +344,7 @@ Status CLLSTMLayer::validate(const ITensorInfo *input, const ITensorInfo *input_
         ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.recurrent_to_input_weights()->num_dimensions() > 2);
         ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.cell_to_input_weights()->num_dimensions() > 1);
         ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.input_gate_bias()->num_dimensions() > 1);
-        ARM_COMPUTE_RETURN_ON_ERROR(CLFullyConnectedLayer::validate(input, lstm_params.input_to_input_weights(), lstm_params.input_gate_bias(), cell_state, true, false));
+        ARM_COMPUTE_RETURN_ON_ERROR(CLFullyConnectedLayer::validate(input, lstm_params.input_to_input_weights(), lstm_params.input_gate_bias(), cell_state));
         ARM_COMPUTE_RETURN_ON_ERROR(CLGEMM::validate(cell_state, &num_units_transposed_info, nullptr, &gemmv_shape_info, 1.f, 0.f, GEMMInfo()));
         ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAddition::validate(cell_state, &gemmv_shape_info, cell_state, ConvertPolicy::SATURATE));
         ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayerKernel::validate(cell_state, nullptr, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC)));
@@ -355,7 +355,7 @@ Status CLLSTMLayer::validate(const ITensorInfo *input, const ITensorInfo *input_
     }
 
     // Validate cell state
-    ARM_COMPUTE_RETURN_ON_ERROR(CLFullyConnectedLayer::validate(input, input_to_cell_weights, cell_bias, cell_state, true, false));
+    ARM_COMPUTE_RETURN_ON_ERROR(CLFullyConnectedLayer::validate(input, input_to_cell_weights, cell_bias, cell_state));
     ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayerKernel::validate(cell_state, nullptr, activation_info));
     ARM_COMPUTE_RETURN_ON_ERROR(CLPixelWiseMultiplicationKernel::validate(cell_state, cell_state, cell_state, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN));
 
@@ -364,7 +364,7 @@ Status CLLSTMLayer::validate(const ITensorInfo *input, const ITensorInfo *input_
         ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayerKernel::validate(cell_state, nullptr, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, -cell_threshold, cell_threshold)));
     }
 
-    ARM_COMPUTE_RETURN_ON_ERROR(CLFullyConnectedLayer::validate(input, input_to_output_weights, output_gate_bias, cell_state, true, false));
+    ARM_COMPUTE_RETURN_ON_ERROR(CLFullyConnectedLayer::validate(input, input_to_output_weights, output_gate_bias, cell_state));
     if(lstm_params.has_peephole_opt())
     {
         ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAddition::validate(cell_state, cell_state, cell_state, ConvertPolicy::SATURATE));
@@ -376,7 +376,7 @@ Status CLLSTMLayer::validate(const ITensorInfo *input, const ITensorInfo *input_
     ARM_COMPUTE_RETURN_ON_ERROR(CLPixelWiseMultiplicationKernel::validate(cell_state, output, output_state, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN));
     if(lstm_params.has_projection())
     {
-        ARM_COMPUTE_RETURN_ON_ERROR(CLFullyConnectedLayer::validate(output_state, lstm_params.projection_weights(), lstm_params.projection_bias(), cell_state, true, false));
+        ARM_COMPUTE_RETURN_ON_ERROR(CLFullyConnectedLayer::validate(output_state, lstm_params.projection_weights(), lstm_params.projection_bias(), cell_state));
         if(projection_threshold != 0.f)
         {
             ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayerKernel::validate(cell_state, output_state, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, -projection_threshold,
diff --git a/src/runtime/CL/functions/CLRNNLayer.cpp b/src/runtime/CL/functions/CLRNNLayer.cpp
index 0e1b9d5b58..1809e6e64e 100644
--- a/src/runtime/CL/functions/CLRNNLayer.cpp
+++ b/src/runtime/CL/functions/CLRNNLayer.cpp
@@ -58,7 +58,7 @@ Status CLRNNLayer::validate(const ITensorInfo *input, const ITensorInfo *weights
 
     auto shape_info = TensorInfo(compute_rnn_shape(recurrent_weights, hidden_state->dimension(idx_height)), 1, input->data_type());
 
-    ARM_COMPUTE_RETURN_ON_ERROR(CLFullyConnectedLayer::validate(input, weights, bias, &shape_info, true, false));
+    ARM_COMPUTE_RETURN_ON_ERROR(CLFullyConnectedLayer::validate(input, weights, bias, &shape_info));
     ARM_COMPUTE_RETURN_ON_ERROR(CLGEMM::validate(hidden_state, recurrent_weights, nullptr, &shape_info, 1.f, 0.f));
     ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAdditionKernel::validate(&shape_info, &shape_info, &shape_info, ConvertPolicy::SATURATE));
     ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayerKernel::validate(&shape_info, &shape_info, info));
@@ -82,7 +82,7 @@ void CLRNNLayer::configure(const ICLTensor *input, const ICLTensor *weights, con
 
     // Manage intermediate buffers and configure
     _memory_group.manage(&_fully_connected_out);
-    _fully_connected_kernel.configure(input, weights, bias, &_fully_connected_out, true, false);
+    _fully_connected_kernel.configure(input, weights, bias, &_fully_connected_out);
 
     _memory_group.manage(&_gemm_output);
     _gemm_state_f.configure(hidden_state, recurrent_weights, nullptr, &_gemm_output, 1.f, 0.f);
diff --git a/src/runtime/GLES_COMPUTE/functions/GCFullyConnectedLayer.cpp b/src/runtime/GLES_COMPUTE/functions/GCFullyConnectedLayer.cpp
index ab2c6c2813..6b8e341b14 100644
--- a/src/runtime/GLES_COMPUTE/functions/GCFullyConnectedLayer.cpp
+++ b/src/runtime/GLES_COMPUTE/functions/GCFullyConnectedLayer.cpp
@@ -80,14 +80,14 @@ void GCFullyConnectedLayer::configure_fc_fc(const IGCTensor *input, const IGCTen
 }
 
 void GCFullyConnectedLayer::configure(const IGCTensor *input, const IGCTensor *weights, const IGCTensor *biases, IGCTensor *output,
-                                      bool transpose_weights, bool are_weights_reshaped, bool retain_internal_weights)
+                                      FullyConnectedLayerInfo fc_info)
 {
     ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32, DataType::F16);
     ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights, output);
     ARM_COMPUTE_ERROR_ON(weights->info()->num_dimensions() > 2);
 
     _original_weights     = weights;
-    _are_weights_reshaped = transpose_weights ? are_weights_reshaped : true;
+    _are_weights_reshaped = fc_info.transpose_weights ? fc_info.are_weights_reshaped : true;
     _is_fc_after_conv     = true;
     _accumulate_biases    = false;
 
@@ -142,8 +142,8 @@ void GCFullyConnectedLayer::configure(const IGCTensor *input, const IGCTensor *w
         configure_fc_fc(input, weights_to_use, output);
     }
 
-    ARM_COMPUTE_ERROR_ON(retain_internal_weights && _reshape_weights_output.gc_buffer() == 0);
-    _are_weights_reshaped = _are_weights_reshaped || retain_internal_weights;
+    ARM_COMPUTE_ERROR_ON(fc_info.retain_internal_weights && _reshape_weights_output.gc_buffer() == 0);
+    _are_weights_reshaped = _are_weights_reshaped || fc_info.retain_internal_weights;
 }
 
 void GCFullyConnectedLayer::run()
diff --git a/src/runtime/NEON/functions/NEFullyConnectedLayer.cpp b/src/runtime/NEON/functions/NEFullyConnectedLayer.cpp
index 3126823e9c..1aab3a05e0 100644
--- a/src/runtime/NEON/functions/NEFullyConnectedLayer.cpp
+++ b/src/runtime/NEON/functions/NEFullyConnectedLayer.cpp
@@ -136,7 +136,8 @@ NEFullyConnectedLayer::NEFullyConnectedLayer(std::shared_ptr<IMemoryManager> mem
 {
 }
 
-void NEFullyConnectedLayer::configure(const ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, bool transpose_weights, bool are_weights_reshaped)
+void NEFullyConnectedLayer::configure(const ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output,
+                                      FullyConnectedLayerInfo fc_info)
 {
     // With the Fully Connected layer we can have 4 different cases:
     //  1) Convolution layer -> Fully Connected layer without batches
@@ -156,8 +157,7 @@ void NEFullyConnectedLayer::configure(const ITensor *input, const ITensor *weigh
                                                                weights->info(),
                                                                biases != nullptr ? biases->info() : nullptr,
                                                                output->info(),
-                                                               transpose_weights,
-                                                               are_weights_reshaped));
+                                                               fc_info));
 
     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;
@@ -167,7 +167,7 @@ void NEFullyConnectedLayer::configure(const ITensor *input, const ITensor *weigh
     _linearize_input     = (input->info()->tensor_shape().x() != linear_input_size) || (num_input_dimensions > 1 && linear_input_size == 1);
     _accumulate_biases   = biases != nullptr;
     _is_batched_fc_layer = num_batch_dimensions > 0;
-    _is_prepared         = are_weights_reshaped || (!transpose_weights && !_is_batched_fc_layer);
+    _is_prepared         = fc_info.are_weights_reshaped || (!fc_info.transpose_weights && !_is_batched_fc_layer);
 
     const size_t   interleave_width = 16 / input->info()->element_size();
     const ITensor *weights_to_use   = weights;
@@ -177,11 +177,11 @@ void NEFullyConnectedLayer::configure(const ITensor *input, const ITensor *weigh
         weights_to_use = &_reshape_weights_output;
 
         _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,
+                                                  fc_info.transpose_weights,
                                                   _is_batched_fc_layer, interleave_width)));
 
         // Reshape the weights
-        _reshape_weights_function.configure(weights, &_reshape_weights_output, transpose_weights, _is_batched_fc_layer);
+        _reshape_weights_function.configure(weights, &_reshape_weights_output, fc_info.transpose_weights, _is_batched_fc_layer);
     }
 
     const ITensor *multiply_input = input;
@@ -231,7 +231,8 @@ 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)
+Status NEFullyConnectedLayer::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output,
+                                       FullyConnectedLayerInfo fc_info)
 {
     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32);
     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights, output);
@@ -251,11 +252,11 @@ Status NEFullyConnectedLayer::validate(const ITensorInfo *input, const ITensorIn
     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))
+    if(!fc_info.are_weights_reshaped && (fc_info.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));
+        reshape_weights_output->set_tensor_shape(compute_fully_connected_reshaped_weights_shape(weights, fc_info.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));
+        ARM_COMPUTE_RETURN_ON_ERROR(NEFullyConnectedLayerReshapeWeights::validate(weights, reshape_weights_output.get(), fc_info.transpose_weights, is_batched_fc_layer));
 
         weights_to_use = reshape_weights_output.get();
     }
diff --git a/src/runtime/NEON/functions/NERNNLayer.cpp b/src/runtime/NEON/functions/NERNNLayer.cpp
index 08017e20c3..f77566a108 100644
--- a/src/runtime/NEON/functions/NERNNLayer.cpp
+++ b/src/runtime/NEON/functions/NERNNLayer.cpp
@@ -57,7 +57,7 @@ Status NERNNLayer::validate(const ITensorInfo *input, const ITensorInfo *weights
 
     auto shape_info = TensorInfo(misc::shape_calculator::compute_rnn_shape(recurrent_weights, hidden_state->dimension(idx_height)), 1, input->data_type());
 
-    ARM_COMPUTE_RETURN_ON_ERROR(NEFullyConnectedLayer::validate(input, weights, bias, &shape_info, true, false));
+    ARM_COMPUTE_RETURN_ON_ERROR(NEFullyConnectedLayer::validate(input, weights, bias, &shape_info));
     ARM_COMPUTE_RETURN_ON_ERROR(NEArithmeticAdditionKernel::validate(&shape_info, &shape_info, &shape_info, ConvertPolicy::SATURATE));
     ARM_COMPUTE_RETURN_ON_ERROR(NEActivationLayerKernel::validate(&shape_info, &shape_info, info));
 
@@ -79,7 +79,7 @@ void NERNNLayer::configure(const ITensor *input, const ITensor *weights, const I
     // Manage intermediate buffers and configure
     _fully_connected_out.allocator()->init(TensorInfo(shape, 1, input->info()->data_type()));
     _memory_group.manage(&_fully_connected_out);
-    _fully_connected_kernel.configure(input, weights, bias, &_fully_connected_out, true, false);
+    _fully_connected_kernel.configure(input, weights, bias, &_fully_connected_out);
 
     _gemm_output.allocator()->init(TensorInfo(shape, 1, input->info()->data_type()));
     _memory_group.manage(&_gemm_output);