COMPMID-1413 - Improve the performance of GEMMLowp with 8 bit dot product on OpenCL

COMPMID-1424 - Add dot product support for CLDepthwise QASYMM8 3x3 NHWC non-unit stride

With this patch we are able to improve the performance of MobileNet v1-qasymm8 by 37 %
Tried to use the dot product instruction in CLDepthwise QASYMM8 3x3 NHWC non-unit stride
but I have not seen any benefit (maybe because we have few arithemtic operation and we
do not have more load instructions). However Depthwise convolution has been improved by
30%

Change-Id: Id768a99c2e53a04276707e427af5d0ec93419ada
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/155082
Tested-by: bsgcomp <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>

7 files changed, 235 insertions, 134 deletions

diff --git a/src/runtime/CL/functions/CLFullyConnectedLayer.cpp b/src/runtime/CL/functions/CLFullyConnectedLayer.cpp
index 010985db06..c5637dba26 100644
--- a/src/runtime/CL/functions/CLFullyConnectedLayer.cpp
+++ b/src/runtime/CL/functions/CLFullyConnectedLayer.cpp
@@ -49,6 +49,7 @@ Status validate_mm(const ITensorInfo &input, const ITensorInfo &weights, const I
         // Validate gemmlowp function
         ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMLowpMatrixMultiplyCore::validate(&input.clone()->set_quantization_info(input_quantization_info),
                                                                            &weights.clone()->set_quantization_info(weights_quantization_info),
+                                                                           nullptr,
                                                                            &output));
     }
     else
@@ -91,7 +92,7 @@ void CLFullyConnectedLayer::configure_mm(const ICLTensor *input, const ICLTensor
         weights->info()->set_quantization_info(QuantizationInfo(weights_quantization_info.scale, -weights_quantization_info.offset));
 
         // Configure gemmlowp function
-        _mm_gemmlowp.configure(input, weights, output);
+        _mm_gemmlowp.configure(input, weights, nullptr, output);
 
         // Revert back QuantizatioInfo as input and weights could be used in other fully connected layers
         input->info()->set_quantization_info(input_quantization_info);
diff --git a/src/runtime/CL/functions/CLGEMMConvolutionLayer.cpp b/src/runtime/CL/functions/CLGEMMConvolutionLayer.cpp
index 61180fd5d3..67f55d56e2 100644
--- a/src/runtime/CL/functions/CLGEMMConvolutionLayer.cpp
+++ b/src/runtime/CL/functions/CLGEMMConvolutionLayer.cpp
@@ -91,19 +91,21 @@ void CLConvolutionLayerReshapeWeights::run()
 }
 
 CLGEMMConvolutionLayer::CLGEMMConvolutionLayer(std::shared_ptr<IMemoryManager> memory_manager)
-    : _memory_group(memory_manager), _reshape_weights(), _im2col_kernel(), _mm_gemm(memory_manager), _mm_gemmlowp(memory_manager), _gemmlowp_output_stage(), _col2im_kernel(), _activationlayer_function(),
-      _add_bias_kernel(), _original_weights(nullptr), _im2col_output(), _weights_reshaped(), _gemm_output(), _tmp_output(), _data_layout(DataLayout::NCHW), _append_bias(false), _skip_im2col(false),
-      _skip_col2im(false), _is_quantized(false), _is_activationlayer_enabled(false), _is_prepared(false)
+    : _memory_group(memory_manager), _reshape_weights(), _im2col_kernel(), _mm_gemm(memory_manager), _mm_gemmlowp(memory_manager), _col2im_kernel(), _activationlayer_function(), _add_bias_kernel(),
+      _original_weights(nullptr), _im2col_output(), _weights_reshaped(), _gemm_output(), _data_layout(DataLayout::NCHW), _append_bias(false), _skip_im2col(false), _skip_col2im(false), _is_quantized(false),
+      _is_activationlayer_enabled(false), _is_prepared(false)
 {
 }
 
-void CLGEMMConvolutionLayer::configure_mm(const ICLTensor *input, const ICLTensor *weights, ICLTensor *output, int gemm_3d_depth)
+void CLGEMMConvolutionLayer::configure_mm(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const GEMMLowpOutputStageInfo &gemmlowp_output_stage,
+                                          int gemm_3d_depth)
 {
     ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights);
-    ARM_COMPUTE_ERROR_THROW_ON(validate_mm(input->info(), weights->info(), output->info(), gemm_3d_depth, _skip_im2col));
+    ARM_COMPUTE_ERROR_THROW_ON(validate_mm(input->info(), weights->info(), biases != nullptr ? biases->info() : nullptr, output->info(), gemmlowp_output_stage, gemm_3d_depth, _skip_im2col));
 
     const GEMMInfo &gemm_info = GEMMInfo(false, false, true /* Reshape weights only for the first run */,
-                                         gemm_3d_depth, _skip_im2col /* Reinterpret the input as 3D if im2col is skipped */);
+                                         gemm_3d_depth, _skip_im2col /* Reinterpret the input as 3D if im2col is skipped */,
+                                         false, gemmlowp_output_stage);
 
     if(_is_quantized)
     {
@@ -115,7 +117,7 @@ void CLGEMMConvolutionLayer::configure_mm(const ICLTensor *input, const ICLTenso
         input->info()->set_quantization_info(QuantizationInfo(input_quantization_info.scale, -input_quantization_info.offset));
         weights->info()->set_quantization_info(QuantizationInfo(weights_quantization_info.scale, -weights_quantization_info.offset));
 
-        _mm_gemmlowp.configure(input, weights, output, gemm_info);
+        _mm_gemmlowp.configure(input, weights, biases, output, gemm_info);
 
         // Revert back QuantizatioInfo as input and weights could be used in other convolution layers
         input->info()->set_quantization_info(input_quantization_info);
@@ -128,12 +130,14 @@ void CLGEMMConvolutionLayer::configure_mm(const ICLTensor *input, const ICLTenso
     }
 }
 
-Status CLGEMMConvolutionLayer::validate_mm(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *output, int gemm_3d_depth, bool skip_im2col)
+Status CLGEMMConvolutionLayer::validate_mm(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output,
+                                           const GEMMLowpOutputStageInfo &gemmlowp_output_stage, int gemm_3d_depth, bool skip_im2col)
 {
     const bool is_quantized = is_data_type_quantized_asymmetric(input->data_type());
 
     const GEMMInfo &gemm_info = GEMMInfo(false, false, true /* Reshape weights only for the first run */,
-                                         gemm_3d_depth, skip_im2col /* Reinterpret the input as 3D if im2col is skipped */);
+                                         gemm_3d_depth, skip_im2col /* Reinterpret the input as 3D if im2col is skipped */,
+                                         false, gemmlowp_output_stage);
 
     if(is_quantized)
     {
@@ -148,7 +152,7 @@ Status CLGEMMConvolutionLayer::validate_mm(const ITensorInfo *input, const ITens
         weights_qa->set_quantization_info(QuantizationInfo(weights_quantization_info.scale, -weights_quantization_info.offset));
 
         // Perform validation step on GEMMLowp
-        return CLGEMMLowpMatrixMultiplyCore::validate(input_qa.get(), weights_qa.get(), output, gemm_info);
+        return CLGEMMLowpMatrixMultiplyCore::validate(input_qa.get(), weights_qa.get(), biases, output, gemm_info);
     }
     else
     {
@@ -176,27 +180,26 @@ void CLGEMMConvolutionLayer::configure(const ICLTensor *input, const ICLTensor *
     const DataLayout data_layout = input->info()->data_layout();
     const int        idx_width   = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH);
     const int        idx_height  = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT);
-    const int        idx_channel = get_data_layout_dimension_index(data_layout, DataLayoutDimension::CHANNEL);
     const int        idx_kernels = get_data_layout_dimension_index(data_layout, DataLayoutDimension::BATCHES);
 
     const unsigned int kernel_width  = weights->info()->dimension(idx_width);
     const unsigned int kernel_height = weights->info()->dimension(idx_height);
 
-    _is_prepared      = weights_info.retain_internal_weights();
-    _original_weights = weights;
-    _is_quantized     = is_data_type_quantized_asymmetric(input->info()->data_type());
-    _data_layout      = data_layout;
-    _skip_im2col      = (data_layout == DataLayout::NHWC && kernel_width == 1 && kernel_height == 1 && conv_info.stride().first == 1 && conv_info.stride().second == 1);
-    _skip_col2im      = data_layout == DataLayout::NHWC;
-    _append_bias      = (biases != nullptr) && (!_is_quantized);
+    _is_prepared                = weights_info.retain_internal_weights();
+    _original_weights           = weights;
+    _is_quantized               = is_data_type_quantized_asymmetric(input->info()->data_type());
+    _data_layout                = data_layout;
+    _skip_im2col                = (data_layout == DataLayout::NHWC && kernel_width == 1 && kernel_height == 1 && conv_info.stride().first == 1 && conv_info.stride().second == 1);
+    _skip_col2im                = data_layout == DataLayout::NHWC;
+    _append_bias                = (biases != nullptr) && (!_is_quantized);
+    _is_activationlayer_enabled = act_info.enabled();
 
     // Set the GPU target for im2col and col2im
     _im2col_kernel.set_target(CLScheduler::get().target());
     _col2im_kernel.set_target(CLScheduler::get().target());
 
-    const ICLTensor *gemm_input_to_use         = input;
-    ICLTensor       *gemm_output_to_use        = output;
-    ICLTensor       *gemm_output_staged_to_use = output;
+    const ICLTensor *gemm_input_to_use  = input;
+    ICLTensor       *gemm_output_to_use = output;
 
     const ICLTensor *biases_to_use = (_append_bias && !_skip_im2col) ? biases : nullptr;
 
@@ -243,26 +246,17 @@ void CLGEMMConvolutionLayer::configure(const ICLTensor *input, const ICLTensor *
     }
 
     // Create GEMM output tensor
-    if(!_skip_col2im || _is_quantized)
+    if(!_skip_col2im)
     {
         TensorShape shape_gemm;
-        if(_skip_col2im)
-        {
-            shape_gemm = input->info()->tensor_shape();
-            shape_gemm.set(idx_width, conv_w);
-            shape_gemm.set(idx_height, conv_h);
-            shape_gemm.set(idx_channel, mat_weights_cols);
-        }
-        else
-        {
-            shape_gemm = _im2col_output.info()->tensor_shape();
-            shape_gemm.set(0, mat_weights_cols);
-            shape_gemm.set(1, conv_w * conv_h);
-        }
-        // GEMM output should be S32 for acquiring raw integer accumulator without quantized postprocessing for quantized asymmetric input.
-        const DataType gemm_data_type = _is_quantized ? DataType::S32 : data_type;
+
+        // If we cannot skip col2im it means we run im2col as well
+        shape_gemm = _im2col_output.info()->tensor_shape();
+        shape_gemm.set(0, mat_weights_cols);
+        shape_gemm.set(1, conv_w * conv_h);
+
         // FIXME: input->clone() doesn't work with subtensors for grouped convolutions.
-        TensorInfo info_gemm(shape_gemm, 1, gemm_data_type);
+        TensorInfo info_gemm(shape_gemm, 1, data_type);
         info_gemm.set_quantization_info(output->info()->quantization_info()).set_data_layout(input->info()->data_layout());
         _gemm_output.allocator()->init(info_gemm);
         _memory_group.manage(&_gemm_output);
@@ -271,42 +265,64 @@ void CLGEMMConvolutionLayer::configure(const ICLTensor *input, const ICLTensor *
         gemm_output_to_use = &_gemm_output;
     }
 
-    // Configure and tune GEMM
-    configure_mm(gemm_input_to_use, &_weights_reshaped, gemm_output_to_use, (data_layout == DataLayout::NHWC) ? conv_h : 1);
-
-    if(!_skip_im2col)
-    {
-        _im2col_output.allocator()->allocate();
-    }
+    GEMMLowpOutputStageInfo gemmlowp_output_stage;
+    gemmlowp_output_stage.type                = GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT;
+    gemmlowp_output_stage.gemmlowp_offset     = 0;
+    gemmlowp_output_stage.gemmlowp_multiplier = 0;
+    gemmlowp_output_stage.gemmlowp_shift      = 0;
 
     // Configure output stage for quantized case
     if(_is_quantized)
     {
         const QuantizationInfo output_quant_info = (output->info()->total_size() == 0) ? input->info()->quantization_info() : output->info()->quantization_info();
 
-        if(!_skip_col2im)
+        float multiplier = input->info()->quantization_info().scale * weights->info()->quantization_info().scale / output_quant_info.scale;
+        int   output_multiplier, output_shift;
+        quantization::calculate_quantized_multiplier_less_than_one(multiplier, &output_multiplier, &output_shift);
+
+        int min_activation = 0;
+        int max_activation = 0;
+
+        const std::set<ActivationLayerInfo::ActivationFunction> supported_acts = { ActivationLayerInfo::ActivationFunction::RELU,
+                                                                                   ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU
+                                                                                 };
+
+        if(_is_activationlayer_enabled && supported_acts.count(act_info.activation()) != 0)
         {
-            _memory_group.manage(&_tmp_output);
-            gemm_output_staged_to_use = &_tmp_output;
+            const int a_const_int = input->info()->quantization_info().quantize(act_info.a(), RoundingPolicy::TO_NEAREST_UP);
+            const int b_const_int = input->info()->quantization_info().quantize(act_info.b(), RoundingPolicy::TO_NEAREST_UP);
+
+            min_activation = act_info.activation() == ActivationLayerInfo::ActivationFunction::RELU ? input->info()->quantization_info().offset : b_const_int;
+            max_activation = act_info.activation() == ActivationLayerInfo::ActivationFunction::RELU ? 255 : a_const_int;
+
+            // If the activation layer is RELU, BOUNDED_RELU or LU_BOUNDED_RELU, we can use the GEMMLowp output stage to perform this operation
+            _is_activationlayer_enabled = false;
         }
 
-        float multiplier = input->info()->quantization_info().scale * weights->info()->quantization_info().scale / output_quant_info.scale;
-        _gemmlowp_output_stage.configure(gemm_output_to_use, biases, gemm_output_staged_to_use, multiplier, output_quant_info.offset);
+        // Set the GEMMLowp output stage info
+        gemmlowp_output_stage.gemmlowp_offset     = output_quant_info.offset;
+        gemmlowp_output_stage.gemmlowp_multiplier = output_multiplier;
+        gemmlowp_output_stage.gemmlowp_shift      = output_shift;
+        gemmlowp_output_stage.gemmlowp_min_bound  = min_activation;
+        gemmlowp_output_stage.gemmlowp_max_bound  = max_activation;
     }
 
-    if(!_skip_col2im)
+    // Configure and tune GEMM
+    configure_mm(gemm_input_to_use, &_weights_reshaped, biases, gemm_output_to_use, gemmlowp_output_stage, (data_layout == DataLayout::NHWC) ? conv_h : 1);
+
+    if(!_skip_im2col)
     {
-        // Configure and tune Col2Im
-        _col2im_kernel.configure(_is_quantized ? gemm_output_staged_to_use : gemm_output_to_use, output, Size2D(conv_w, conv_h), num_groups);
-        CLScheduler::get().tune_kernel_static(_col2im_kernel);
+        _im2col_output.allocator()->allocate();
     }
 
     if(!_skip_col2im)
     {
-        _tmp_output.allocator()->allocate();
+        // Configure and tune Col2Im
+        _col2im_kernel.configure(gemm_output_to_use, output, Size2D(conv_w, conv_h), num_groups);
+        CLScheduler::get().tune_kernel_static(_col2im_kernel);
     }
 
-    if(!_skip_col2im || _is_quantized)
+    if(!_skip_col2im)
     {
         _gemm_output.allocator()->allocate();
     }
@@ -314,9 +330,6 @@ void CLGEMMConvolutionLayer::configure(const ICLTensor *input, const ICLTensor *
     ARM_COMPUTE_ERROR_ON_MSG((output->info()->dimension(idx_width) != conv_w) || (output->info()->dimension(idx_height) != conv_h),
                              "Output shape does not match the expected one");
 
-    //Configure Activation Layer
-    _is_activationlayer_enabled = act_info.enabled();
-
     if(_is_activationlayer_enabled)
     {
         _activationlayer_function.configure(output, nullptr, act_info);
@@ -347,16 +360,16 @@ Status CLGEMMConvolutionLayer::validate(const ITensorInfo *input, const ITensorI
     const unsigned int kernel_width  = weights->dimension(idx_width);
     const unsigned int kernel_height = weights->dimension(idx_height);
 
-    TensorInfo         im2col_reshaped_info, info_gemm, tmp_info, weights_reshaped_info;
-    const ITensorInfo *gemm_input_to_use         = input;
-    const ITensorInfo *gemm_output_to_use        = output;
-    const ITensorInfo *gemm_output_staged_to_use = output;
-    const ITensorInfo *weights_to_use            = weights;
+    TensorInfo         im2col_reshaped_info, info_gemm, weights_reshaped_info;
+    const ITensorInfo *gemm_input_to_use  = input;
+    const ITensorInfo *gemm_output_to_use = output;
+    const ITensorInfo *weights_to_use     = weights;
 
-    const bool is_quantized = is_data_type_quantized_asymmetric(data_type);
-    const bool append_bias  = (biases != nullptr) && (!is_quantized);
-    const bool skip_im2col  = (data_layout == DataLayout::NHWC && kernel_width == 1 && kernel_height == 1 && conv_info.stride().first == 1 && conv_info.stride().second == 1);
-    const bool skip_col2im  = data_layout == DataLayout::NHWC;
+    const bool is_quantized               = is_data_type_quantized_asymmetric(data_type);
+    const bool append_bias                = (biases != nullptr) && (!is_quantized);
+    const bool skip_im2col                = (data_layout == DataLayout::NHWC && kernel_width == 1 && kernel_height == 1 && conv_info.stride().first == 1 && conv_info.stride().second == 1);
+    const bool skip_col2im                = data_layout == DataLayout::NHWC;
+    bool       is_activationlayer_enabled = act_info.enabled();
 
     ARM_COMPUTE_RETURN_ERROR_ON((weights->dimension(idx_channel) * num_groups) != input->dimension(idx_channel));
     ARM_COMPUTE_RETURN_ERROR_ON(weights->num_dimensions() > 4);
@@ -418,52 +431,80 @@ Status CLGEMMConvolutionLayer::validate(const ITensorInfo *input, const ITensorI
     }
 
     // Create GEMM output tensor
-    if(!skip_col2im || is_quantized)
+    if(!skip_col2im)
     {
-        const DataType gemm_data_type = is_quantized ? DataType::S32 : data_type;
-        TensorShape    shape_gemm;
-        if(skip_col2im)
-        {
-            shape_gemm = input->tensor_shape();
-            shape_gemm.set(idx_width, conv_w);
-            shape_gemm.set(idx_height, conv_h);
-            shape_gemm.set(idx_channel, mat_weights_cols);
-        }
-        else
-        {
-            shape_gemm = gemm_input_to_use->tensor_shape();
-            shape_gemm.set(0, mat_weights_cols);
-            shape_gemm.set(1, conv_w * conv_h);
-        }
-        // GEMM output should be S32 for acquiring raw integer accumulator without quantized postprocessing for quantized asymmetric input.
-        info_gemm = TensorInfo(shape_gemm, 1, gemm_data_type);
+        TensorShape shape_gemm;
+
+        shape_gemm = gemm_input_to_use->tensor_shape();
+        shape_gemm.set(0, mat_weights_cols);
+        shape_gemm.set(1, conv_w * conv_h);
+
+        info_gemm = TensorInfo(shape_gemm, 1, data_type);
         info_gemm.set_quantization_info(output->quantization_info()).set_data_layout(input->data_layout());
         gemm_output_to_use = &info_gemm;
     }
 
-    ARM_COMPUTE_RETURN_ON_ERROR(validate_mm(gemm_input_to_use, weights_to_use, gemm_output_to_use, skip_col2im ? conv_h : 1, skip_im2col));
+    GEMMLowpOutputStageInfo gemmlowp_output_stage;
+    gemmlowp_output_stage.type                = GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT;
+    gemmlowp_output_stage.gemmlowp_offset     = 0;
+    gemmlowp_output_stage.gemmlowp_multiplier = 0;
+    gemmlowp_output_stage.gemmlowp_shift      = 0;
 
     if(is_quantized)
     {
-        if(!skip_col2im)
+        const QuantizationInfo output_quant_info = (output->total_size() == 0) ? input->quantization_info() : output->quantization_info();
+
+        float multiplier = input->quantization_info().scale * weights->quantization_info().scale / output_quant_info.scale;
+        int   output_multiplier, output_shift;
+        quantization::calculate_quantized_multiplier_less_than_one(multiplier, &output_multiplier, &output_shift);
+
+        int min_activation = 0;
+        int max_activation = 0;
+
+        const std::set<ActivationLayerInfo::ActivationFunction> supported_acts = { ActivationLayerInfo::ActivationFunction::RELU,
+                                                                                   ActivationLayerInfo::ActivationFunction::BOUNDED_RELU,
+                                                                                   ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU
+                                                                                 };
+
+        if(is_activationlayer_enabled && supported_acts.count(act_info.activation()) != 0)
         {
-            tmp_info = TensorInfo(gemm_output_to_use->tensor_shape(), 1, DataType::QASYMM8);
-            tmp_info.set_quantization_info(output->quantization_info());
-            gemm_output_staged_to_use = &tmp_info;
+            const int a_const_int = input->quantization_info().quantize(act_info.a(), RoundingPolicy::TO_NEAREST_UP);
+            const int b_const_int = input->quantization_info().quantize(act_info.b(), RoundingPolicy::TO_NEAREST_UP);
+
+            min_activation = b_const_int;
+            max_activation = a_const_int;
+
+            if(act_info.activation() != ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU)
+            {
+                min_activation = input->quantization_info().offset;
+            }
+            if(act_info.activation() == ActivationLayerInfo::ActivationFunction::RELU)
+            {
+                max_activation = 255;
+            }
+
+            // If the activation layer is RELU, BOUNDED_RELU or LU_BOUNDED_RELU, we can use the GEMMLowp output stage to perform this operation
+            is_activationlayer_enabled = false;
+
+            // Set the GEMMLowp output stage info
+            gemmlowp_output_stage.gemmlowp_offset     = output_quant_info.offset;
+            gemmlowp_output_stage.gemmlowp_multiplier = output_multiplier;
+            gemmlowp_output_stage.gemmlowp_shift      = output_shift;
+            gemmlowp_output_stage.gemmlowp_min_bound  = min_activation;
+            gemmlowp_output_stage.gemmlowp_max_bound  = max_activation;
         }
-        // Validate output stage for quantized case
-        CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint::validate(gemm_output_to_use, biases, gemm_output_staged_to_use);
     }
 
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_mm(gemm_input_to_use, weights_to_use, biases, gemm_output_to_use, gemmlowp_output_stage, skip_col2im ? conv_h : 1, skip_im2col));
+
     // Validate Col2Im
     if(!skip_col2im)
     {
-        ARM_COMPUTE_RETURN_ON_ERROR(CLCol2ImKernel::validate(is_quantized ? gemm_output_staged_to_use : gemm_output_to_use, output,
-                                                             Size2D(conv_w, conv_h), num_groups));
+        ARM_COMPUTE_RETURN_ON_ERROR(CLCol2ImKernel::validate(gemm_output_to_use, output, Size2D(conv_w, conv_h), num_groups));
     }
 
     //Validate Activation Layer
-    if(act_info.enabled())
+    if(is_activationlayer_enabled)
     {
         ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayer::validate(output, nullptr, act_info));
     }
@@ -488,9 +529,6 @@ void CLGEMMConvolutionLayer::run()
     {
         // Run gemmlowp
         _mm_gemmlowp.run();
-
-        // Run output stage
-        _gemmlowp_output_stage.run();
     }
     else
     {
diff --git a/src/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.cpp b/src/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.cpp
index f79fb43073..f2efb3249b 100644
--- a/src/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.cpp
+++ b/src/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.cpp
@@ -42,7 +42,7 @@ inline bool is_interleaved_transposed(int m, int n, int k, bool reshape_b_only_o
     bool flag = true;
 
     if(gpu_target_is_in(gpu_target,
-                        GPUTarget::G71, GPUTarget::G72, GPUTarget::G76,
+                        GPUTarget::G71, GPUTarget::G72,
                         GPUTarget::G51, GPUTarget::G51BIG, GPUTarget::G51LIT,
                         GPUTarget::G52, GPUTarget::G52LIT))
     {
@@ -56,6 +56,10 @@ inline bool is_interleaved_transposed(int m, int n, int k, bool reshape_b_only_o
             flag = false;
         }
     }
+    else
+    {
+        flag = m > 1;
+    }
 
     return flag;
 }
@@ -69,24 +73,26 @@ CLGEMMLowpMatrixMultiplyCore::CLGEMMLowpMatrixMultiplyCore(std::shared_ptr<IMemo
       _mtx_a_reduction_kernel(),
       _mtx_b_reduction_kernel(),
       _offset_contribution_kernel(),
+      _offset_contribution_output_stage_kernel(),
       _vector_sum_col(),
       _vector_sum_row(),
       _tmp_a(),
       _tmp_b(),
+      _mm_result_s32(),
       _original_b(nullptr),
       _a_offset(0),
       _b_offset(0),
       _is_interleaved_transposed(true),
       _reshape_b_only_on_first_run(false),
-      _is_prepared(false)
+      _is_prepared(false),
+      _fuse_output_stage(false)
 {
 }
 
-void CLGEMMLowpMatrixMultiplyCore::configure(const ICLTensor *a, const ICLTensor *b, ICLTensor *output, const GEMMInfo &gemm_info)
+void CLGEMMLowpMatrixMultiplyCore::configure(const ICLTensor *a, const ICLTensor *b, const ICLTensor *c, ICLTensor *output, const GEMMInfo &gemm_info)
 {
     ARM_COMPUTE_ERROR_ON_NULLPTR(a, b, output);
-    ARM_COMPUTE_UNUSED(gemm_info);
-    ARM_COMPUTE_ERROR_THROW_ON(CLGEMMLowpMatrixMultiplyCore::validate(a->info(), b->info(), output->info(), gemm_info));
+    ARM_COMPUTE_ERROR_THROW_ON(CLGEMMLowpMatrixMultiplyCore::validate(a->info(), b->info(), c != nullptr ? c->info() : nullptr, output->info(), gemm_info));
 
     _is_prepared                 = false;
     _original_b                  = b;
@@ -108,6 +114,7 @@ void CLGEMMLowpMatrixMultiplyCore::configure(const ICLTensor *a, const ICLTensor
     // If we pass the matrix A and matrix B reshaped to CLGEMMMatrixMultiplyKernel, we need to pass m, n, k, mult_transpose1xW_width and mult_interleave4x4_height to CLGEMMReshapeInfo
     // in order to know how the matrices have been reshaped
     bool          reinterpret_input_as_3d   = gemm_info.reinterpret_input_as_3d();
+    const bool    unroll_block              = dot8_supported(CLKernelLibrary::get().get_device());
     const int     m                         = reinterpret_input_as_3d ? (a->info()->dimension(1) * a->info()->dimension(2)) : a->info()->dimension(1);
     const int     n                         = b->info()->dimension(0);
     const int     k                         = a->info()->dimension(0);
@@ -133,15 +140,11 @@ void CLGEMMLowpMatrixMultiplyCore::configure(const ICLTensor *a, const ICLTensor
         }
 
         // Configure interleave kernel
-        _mtx_a_reshape_kernel.configure(a, &_tmp_a, mult_interleave4x4_height, gemm_info.reinterpret_input_as_3d());
+        _mtx_a_reshape_kernel.configure(a, &_tmp_a, mult_interleave4x4_height, gemm_info.reinterpret_input_as_3d(), unroll_block);
 
         // Configure transpose kernel
         _mtx_b_reshape_kernel.configure(b, &_tmp_b, mult_transpose1xW_width);
     }
-    // Configure matrix multiply kernel
-    _mm_kernel.configure(matrix_a, matrix_b, output, _is_interleaved_transposed, GEMMReshapeInfo(m, n, k,
-                                                                                                 mult_transpose1xW_width, mult_interleave4x4_height,
-                                                                                                 depth_output_gemm3d, reinterpret_input_as_3d));
 
     // Initialize matrix B reduction kernel only if _a_offset is not equal to 0
     if(_a_offset != 0)
@@ -168,8 +171,34 @@ void CLGEMMLowpMatrixMultiplyCore::configure(const ICLTensor *a, const ICLTensor
         _mtx_a_reduction_kernel.configure(a, &_vector_sum_row);
     }
 
-    // Configure offset contribution kernel
-    _offset_contribution_kernel.configure(output, _a_offset == 0 ? nullptr : &_vector_sum_col, _b_offset == 0 ? nullptr : &_vector_sum_row, a->info()->dimension(0), _a_offset, _b_offset);
+    // If GEMMLowpOutputStage != NONE, fuse the offset contribution with the output stage
+    if(gemm_info.gemmlowp_output_stage().type != GEMMLowpOutputStageType::NONE)
+    {
+        _fuse_output_stage = true;
+
+        _memory_group.manage(&_mm_result_s32);
+
+        // Configure matrix multiply kernel
+        _mm_kernel.configure(matrix_a, matrix_b, &_mm_result_s32, _is_interleaved_transposed, GEMMReshapeInfo(m, n, k,
+                                                                                                              mult_transpose1xW_width, mult_interleave4x4_height,
+                                                                                                              depth_output_gemm3d, reinterpret_input_as_3d));
+
+        // Configure offset contribution kernel
+        _offset_contribution_output_stage_kernel.configure(&_mm_result_s32, _a_offset == 0 ? nullptr : &_vector_sum_col, _b_offset == 0 ? nullptr : &_vector_sum_row, c, output, a->info()->dimension(0),
+                                                           _a_offset, _b_offset, gemm_info.gemmlowp_output_stage());
+
+        _mm_result_s32.allocator()->allocate();
+    }
+    else
+    {
+        // Configure matrix multiply kernel
+        _mm_kernel.configure(matrix_a, matrix_b, output, _is_interleaved_transposed, GEMMReshapeInfo(m, n, k,
+                                                                                                     mult_transpose1xW_width, mult_interleave4x4_height,
+                                                                                                     depth_output_gemm3d, reinterpret_input_as_3d));
+
+        // Configure offset contribution kernel
+        _offset_contribution_kernel.configure(output, _a_offset == 0 ? nullptr : &_vector_sum_col, _b_offset == 0 ? nullptr : &_vector_sum_row, c, a->info()->dimension(0), _a_offset, _b_offset);
+    }
 
     // Allocate tensors
     if(_is_interleaved_transposed)
@@ -192,10 +221,9 @@ void CLGEMMLowpMatrixMultiplyCore::configure(const ICLTensor *a, const ICLTensor
     }
 }
 
-Status CLGEMMLowpMatrixMultiplyCore::validate(const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *output, const GEMMInfo &gemm_info)
+Status CLGEMMLowpMatrixMultiplyCore::validate(const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *c, const ITensorInfo *output, const GEMMInfo &gemm_info)
 {
     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(a, 1, DataType::QASYMM8);
-    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::S32);
     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(a, b);
     ARM_COMPUTE_RETURN_ERROR_ON_MSG(gemm_info.is_a_reshaped(), "Matrix A already reshaped is not supported");
     ARM_COMPUTE_RETURN_ERROR_ON_MSG(gemm_info.is_b_reshaped(), "Matrix B already reshaped is not supported");
@@ -241,9 +269,6 @@ Status CLGEMMLowpMatrixMultiplyCore::validate(const ITensorInfo *a, const ITenso
         ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMTranspose1xWKernel::validate(b, &tmp_b_info, mult_transpose1xW_width));
     }
 
-    // Validate matrix multiply
-    ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMLowpMatrixMultiplyKernel::validate(matrix_a_info, matrix_b_info, output, reshape_matrices, reshape_info));
-
     TensorInfo info_vector_sum_col, info_vector_sum_row;
 
     // Validate matrix B reduction kernel only if _a_offset is not equal to 0
@@ -264,11 +289,37 @@ Status CLGEMMLowpMatrixMultiplyCore::validate(const ITensorInfo *a, const ITenso
         ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMLowpMatrixAReductionKernel::validate(a, &info_vector_sum_row));
     }
 
-    // Validate offset contribution kernel
-    ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMLowpOffsetContributionKernel::validate(output,
-                                                                             a_offset == 0 ? nullptr : &info_vector_sum_col,
-                                                                             b_offset == 0 ? nullptr : &info_vector_sum_row,
-                                                                             a_offset, b_offset));
+    if(gemm_info.gemmlowp_output_stage().type != GEMMLowpOutputStageType::NONE)
+    {
+        TensorInfo mm_result_s32_info{};
+
+        // Output tensor auto inizialitation if not yet initialized
+        auto_init_if_empty(mm_result_s32_info, a->clone()->set_tensor_shape(compute_mm_shape(*matrix_a_info, *matrix_b_info, reshape_matrices, reshape_info)).set_data_type(DataType::S32));
+
+        // Validate matrix multiply
+        ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMLowpMatrixMultiplyKernel::validate(matrix_a_info, matrix_b_info, &mm_result_s32_info, reshape_matrices, reshape_info));
+
+        // Validate offset contribution kernel
+        ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMLowpOffsetContributionOutputStageKernel::validate(&mm_result_s32_info,
+                                                                                            a_offset == 0 ? nullptr : &info_vector_sum_col,
+                                                                                            b_offset == 0 ? nullptr : &info_vector_sum_row,
+                                                                                            c,
+                                                                                            output,
+                                                                                            a_offset, b_offset,
+                                                                                            gemm_info.gemmlowp_output_stage()));
+    }
+    else
+    {
+        // Validate matrix multiply
+        ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMLowpMatrixMultiplyKernel::validate(matrix_a_info, matrix_b_info, output, reshape_matrices, reshape_info));
+
+        // Validate offset contribution kernel
+        ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMLowpOffsetContributionKernel::validate(output,
+                                                                                 a_offset == 0 ? nullptr : &info_vector_sum_col,
+                                                                                 b_offset == 0 ? nullptr : &info_vector_sum_row,
+                                                                                 c,
+                                                                                 a_offset, b_offset));
+    }
 
     return Status{};
 }
@@ -306,8 +357,16 @@ void CLGEMMLowpMatrixMultiplyCore::run()
         CLScheduler::get().enqueue(_mtx_a_reduction_kernel, false);
     }
 
-    // Run offset contribution kernel
-    CLScheduler::get().enqueue(_offset_contribution_kernel, true);
+    if(_fuse_output_stage)
+    {
+        // Run offset contribution/output stage kernel
+        CLScheduler::get().enqueue(_offset_contribution_output_stage_kernel, true);
+    }
+    else
+    {
+        // Run offset contribution kernel
+        CLScheduler::get().enqueue(_offset_contribution_kernel, true);
+    }
 
     _memory_group.release();
 }
diff --git a/src/runtime/CL/functions/CLGEMMLowpOutputStage.cpp b/src/runtime/CL/functions/CLGEMMLowpOutputStage.cpp
index f5dc655776..f1c24626dc 100644
--- a/src/runtime/CL/functions/CLGEMMLowpOutputStage.cpp
+++ b/src/runtime/CL/functions/CLGEMMLowpOutputStage.cpp
@@ -45,17 +45,17 @@ Status CLGEMMLowpQuantizeDownInt32ToUint8Scale::validate(const ITensorInfo *inpu
 
 void CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint::configure(const ICLTensor *input, const ICLTensor *bias, ICLTensor *output,
                                                                     int result_fixedpoint_multiplier, int result_shift, int result_offset_after_shift,
-                                                                    int min, int max, unsigned int output_3d_depth)
+                                                                    int min, int max)
 {
     auto k = arm_compute::support::cpp14::make_unique<CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel>();
-    k->configure(input, bias, output, result_fixedpoint_multiplier, result_shift, result_offset_after_shift, min, max, output_3d_depth);
+    k->configure(input, bias, output, result_fixedpoint_multiplier, result_shift, result_offset_after_shift, min, max);
     _kernel = std::move(k);
 }
 
 Status CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint::validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output,
-                                                                     int min, int max, unsigned int output_3d_depth)
+                                                                     int min, int max)
 {
-    return CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::validate(input, bias, output, min, max, output_3d_depth);
+    return CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::validate(input, bias, output, min, max);
 }
 
 void CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloat::configure(const ICLTensor *input, const ICLTensor *bias, ICLTensor *output,
diff --git a/src/runtime/NEON/functions/NEFullyConnectedLayer.cpp b/src/runtime/NEON/functions/NEFullyConnectedLayer.cpp
index 60f6294394..45e21b53d1 100644
--- a/src/runtime/NEON/functions/NEFullyConnectedLayer.cpp
+++ b/src/runtime/NEON/functions/NEFullyConnectedLayer.cpp
@@ -50,6 +50,7 @@ Status validate_mm(const ITensorInfo &input, const ITensorInfo &weights, const I
         // Validate gemmlowp function
         ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMLowpMatrixMultiplyCore::validate(&input.clone()->set_quantization_info(input_quantization_info),
                                                                            &weights.clone()->set_quantization_info(weights_quantization_info),
+                                                                           nullptr,
                                                                            &output));
     }
     else
@@ -93,7 +94,7 @@ void NEFullyConnectedLayer::configure_mm(const ITensor *input, const ITensor *we
         weights->info()->set_quantization_info(QuantizationInfo(weights_quantization_info.scale, -weights_quantization_info.offset));
 
         // Configure gemmlowp function
-        _mm_gemmlowp.configure(input, weights, output);
+        _mm_gemmlowp.configure(input, weights, nullptr, output);
 
         // Revert back QuantizatioInfo as input and weights could be used in other fully connected layers
         input->info()->set_quantization_info(input_quantization_info);
diff --git a/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp b/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp
index fb6d4a1847..fc65469488 100644
--- a/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp
+++ b/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp
@@ -111,7 +111,7 @@ void NEGEMMConvolutionLayer::configure_mm(const ITensor *input, const ITensor *w
         input->info()->set_quantization_info(QuantizationInfo(input_quantization_info.scale, -input_quantization_info.offset));
         weights->info()->set_quantization_info(QuantizationInfo(weights_quantization_info.scale, -weights_quantization_info.offset));
 
-        _mm_gemmlowp.configure(input, weights, output, GEMMInfo(false, false, true /* Reshape weights only for the first run*/));
+        _mm_gemmlowp.configure(input, weights, nullptr, output, GEMMInfo(false, false, true /* Reshape weights only for the first run*/));
 
         // Revert back QuantizatioInfo as input and weights could be used in other convolution layers
         input->info()->set_quantization_info(input_quantization_info);
@@ -143,7 +143,7 @@ Status NEGEMMConvolutionLayer::validate_mm(const ITensorInfo *input, const ITens
         weights_qa->set_quantization_info(QuantizationInfo(weights_quantization_info.scale, -weights_quantization_info.offset));
 
         // Perform validation step on GEMMLowp
-        return NEGEMMLowpMatrixMultiplyCore::validate(input_qa.get(), weights_qa.get(), output, gemm_info);
+        return NEGEMMLowpMatrixMultiplyCore::validate(input_qa.get(), weights_qa.get(), nullptr, output, gemm_info);
     }
     else
     {
diff --git a/src/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.cpp b/src/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.cpp
index 828011d019..80f5ab0c93 100644
--- a/src/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.cpp
+++ b/src/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.cpp
@@ -47,10 +47,11 @@ NEGEMMLowpMatrixMultiplyCore::NEGEMMLowpMatrixMultiplyCore(std::shared_ptr<IMemo
 {
 }
 
-void NEGEMMLowpMatrixMultiplyCore::configure(const ITensor *a, const ITensor *b, ITensor *output, const GEMMInfo &gemm_info)
+void NEGEMMLowpMatrixMultiplyCore::configure(const ITensor *a, const ITensor *b, const ITensor *c, ITensor *output, const GEMMInfo &gemm_info)
 {
     ARM_COMPUTE_ERROR_ON_NULLPTR(a, b, output);
-    ARM_COMPUTE_ERROR_THROW_ON(NEGEMMLowpMatrixMultiplyCore::validate(a->info(), b->info(), output->info(), gemm_info));
+    ARM_COMPUTE_UNUSED(c);
+    ARM_COMPUTE_ERROR_THROW_ON(NEGEMMLowpMatrixMultiplyCore::validate(a->info(), b->info(), c != nullptr ? c->info() : nullptr, output->info(), gemm_info));
 
     // Clear state
     _mtx_a_reshape_kernel = nullptr;
@@ -181,11 +182,12 @@ void NEGEMMLowpMatrixMultiplyCore::configure(const ITensor *a, const ITensor *b,
     }
 }
 
-Status NEGEMMLowpMatrixMultiplyCore::validate(const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *output, const GEMMInfo &gemm_info)
+Status NEGEMMLowpMatrixMultiplyCore::validate(const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *c, const ITensorInfo *output, const GEMMInfo &gemm_info)
 {
     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(a, 1, DataType::QASYMM8);
     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::S32);
     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(a, b);
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(c != nullptr, "Bias addition not supported in NEGEMMLowpMatrixMultiplyCore");
     ARM_COMPUTE_RETURN_ERROR_ON_MSG((a)->dimension(0) != (b)->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_RETURN_ERROR_ON_MSG((a)->dimension(1) != (output)->dimension(1),