COMPMID-1694: Fuse offset contribution with the output stage when we use NEGEMMLowpMatrixMultiplyCore

Change-Id: Ic1a681e4cc03e1eba3bf8485d9cdb17b3e926047
Signed-off-by: giuros01 <giuseppe.rossini@arm.com>
Reviewed-on: https://review.mlplatform.org/c/561
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>

2 files changed, 244 insertions, 212 deletions

diff --git a/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp b/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp
index be7cc2d0e1..b6c37349c1 100644
--- a/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp
+++ b/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2018 ARM Limited.
+ * Copyright (c) 2017-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -90,16 +90,17 @@ void NEConvolutionLayerReshapeWeights::run()
 }
 
 NEGEMMConvolutionLayer::NEGEMMConvolutionLayer(const 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(), _reshape_layer(), _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(),
+      _reshape_layer(), _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)
 {
 }
 
-void NEGEMMConvolutionLayer::configure_mm(const ITensor *input, const ITensor *weights, ITensor *output, int gemm_3d_depth)
+void NEGEMMConvolutionLayer::configure_mm(const ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const ActivationLayerInfo &act_info, 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 ? nullptr : biases->info(), output == nullptr ? nullptr : output->info(), act_info, 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 */);
@@ -114,7 +115,40 @@ 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, nullptr, output, gemm_info);
+        const QuantizationInfo output_quant_info = (output->info()->total_size() == 0) ? input_quantization_info : output->info()->quantization_info();
+
+        float multiplier = input_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);
+
+        // Merge activation with output stage
+        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)
+        {
+            const int a_const_int = output_quant_info.quantize(act_info.a(), RoundingPolicy::TO_NEAREST_UP);
+            const int b_const_int = output_quant_info.quantize(act_info.b(), RoundingPolicy::TO_NEAREST_UP);
+
+            min_activation = act_info.activation() != ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU ? output_quant_info.offset : b_const_int;
+            max_activation = act_info.activation() == ActivationLayerInfo::ActivationFunction::RELU ? 255 : a_const_int;
+
+            _is_activationlayer_enabled = false;
+        }
+
+        GEMMLowpOutputStageInfo output_info;
+        output_info.type                = GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT;
+        output_info.gemmlowp_offset     = output_quant_info.offset;
+        output_info.gemmlowp_multiplier = output_multiplier;
+        output_info.gemmlowp_shift      = output_shift;
+        output_info.gemmlowp_min_bound  = min_activation;
+        output_info.gemmlowp_max_bound  = max_activation;
+
+        _mm_gemmlowp.configure(input, weights, biases, output, GEMMInfo(false, false, true, gemm_3d_depth, _skip_im2col, false, output_info));
 
         // Revert back QuantizatioInfo as input and weights could be used in other convolution layers
         input->info()->set_quantization_info(input_quantization_info);
@@ -127,9 +161,11 @@ void NEGEMMConvolutionLayer::configure_mm(const ITensor *input, const ITensor *w
     }
 }
 
-Status NEGEMMConvolutionLayer::validate_mm(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *output, int gemm_3d_depth, bool skip_im2col)
+Status NEGEMMConvolutionLayer::validate_mm(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const ActivationLayerInfo &act_info,
+                                           int gemm_3d_depth, bool skip_im2col)
 {
-    const bool is_quantized = is_data_type_quantized_asymmetric(input->data_type());
+    const bool is_quantized          = is_data_type_quantized_asymmetric(input->data_type());
+    const bool is_activation_enabled = act_info.enabled();
 
     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 */);
@@ -145,8 +181,39 @@ Status NEGEMMConvolutionLayer::validate_mm(const ITensorInfo *input, const ITens
         input_qa->set_quantization_info(QuantizationInfo(input_quantization_info.scale, -input_quantization_info.offset));
         weights_qa->set_quantization_info(QuantizationInfo(weights_quantization_info.scale, -weights_quantization_info.offset));
 
+        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);
+
+        // Merge activation with output stage
+        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_activation_enabled && supported_acts.count(act_info.activation()) != 0)
+        {
+            const int a_const_int = output_quant_info.quantize(act_info.a(), RoundingPolicy::TO_NEAREST_UP);
+            const int b_const_int = output_quant_info.quantize(act_info.b(), RoundingPolicy::TO_NEAREST_UP);
+
+            min_activation = act_info.activation() != ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU ? output_quant_info.offset : b_const_int;
+            max_activation = act_info.activation() == ActivationLayerInfo::ActivationFunction::RELU ? 255 : a_const_int;
+        }
+
+        GEMMLowpOutputStageInfo output_info;
+        output_info.type                = GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT;
+        output_info.gemmlowp_offset     = output_quant_info.offset;
+        output_info.gemmlowp_multiplier = output_multiplier;
+        output_info.gemmlowp_shift      = output_shift;
+        output_info.gemmlowp_min_bound  = min_activation;
+        output_info.gemmlowp_max_bound  = max_activation;
+
         // Perform validation step on GEMMLowp
-        return NEGEMMLowpMatrixMultiplyCore::validate(input_qa.get(), weights_qa.get(), nullptr, output, gemm_info);
+        return NEGEMMLowpMatrixMultiplyCore::validate(input_qa.get(), weights_qa.get(), biases, output, GEMMInfo(false, false, true, gemm_3d_depth, skip_im2col, false, output_info));
     }
     else
     {
@@ -155,19 +222,18 @@ Status NEGEMMConvolutionLayer::validate_mm(const ITensorInfo *input, const ITens
     }
 }
 
-Status NEGEMMConvolutionLayer::validate_gemm3d(DataType data_type, int gemm_3d_depth, bool skip_im2col)
+Status NEGEMMConvolutionLayer::validate_gemm3d(const ITensorInfo *input_info, const ActivationLayerInfo &act_info, int gemm_3d_depth, bool skip_im2col)
 {
-    const bool         is_quantized          = is_data_type_quantized_asymmetric(data_type);
-    const DataType     output_gemm_data_type = is_quantized ? DataType::S32 : data_type;
-    const unsigned int mult_y                = skip_im2col ? 1U : gemm_3d_depth;
-    const unsigned int mult_z                = skip_im2col ? gemm_3d_depth : 1U;
+    const DataType     data_type = input_info->data_type();
+    const unsigned int mult_y    = skip_im2col ? 1U : gemm_3d_depth;
+    const unsigned int mult_z    = skip_im2col ? gemm_3d_depth : 1U;
 
     // Set dummy tensor shapes for the validation
-    const TensorInfo dummy_input_info(TensorShape(4U, 4U * mult_y, 1U * mult_z), 1, data_type);
+    const TensorInfo dummy_input_info(TensorShape(4U, 4U * mult_y, 1U * mult_z), 1, data_type, input_info->quantization_info());
     const TensorInfo dummy_weights_info(TensorShape(4U, 4U), 1, data_type);
-    const TensorInfo dummy_output_info(TensorShape(4U, 4U, gemm_3d_depth), 1, output_gemm_data_type);
+    const TensorInfo dummy_output_info(TensorShape(4U, 4U, gemm_3d_depth), 1, data_type, input_info->quantization_info());
 
-    return validate_mm(&dummy_input_info, &dummy_weights_info, &dummy_output_info, gemm_3d_depth, skip_im2col);
+    return validate_mm(&dummy_input_info, &dummy_weights_info, nullptr, &dummy_output_info, act_info, gemm_3d_depth, skip_im2col);
 }
 
 void NEGEMMConvolutionLayer::configure(const ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const PadStrideInfo &conv_info, const WeightsInfo &weights_info,
@@ -202,9 +268,8 @@ void NEGEMMConvolutionLayer::configure(const ITensor *input, const ITensor *weig
     _append_bias                = (biases != nullptr) && (!_is_quantized);
     _is_activationlayer_enabled = act_info.enabled();
 
-    const ITensor *gemm_input_to_use         = input;
-    ITensor       *gemm_output_to_use        = output;
-    ITensor       *gemm_output_staged_to_use = output;
+    const ITensor *gemm_input_to_use  = input;
+    ITensor       *gemm_output_to_use = output;
 
     // Get convolved dimensions
     unsigned int conv_w = 0;
@@ -219,7 +284,7 @@ void NEGEMMConvolutionLayer::configure(const ITensor *input, const ITensor *weig
     // Check if GEMM3D is supported
     if(data_layout == DataLayout::NHWC)
     {
-        _skip_col2im = bool(validate_gemm3d(input->info()->data_type(), conv_h, true));
+        _skip_col2im = bool(validate_gemm3d(input->info(), act_info, conv_h, true));
         // If not supported, we need to perform im2col and col2im (or reshape layer)
         if(!_skip_col2im)
         {
@@ -262,26 +327,17 @@ void NEGEMMConvolutionLayer::configure(const ITensor *input, const ITensor *weig
     }
 
     // Create temporary GEMM output tensor in case we cannot skip col2im
-    if(!_skip_col2im || _is_quantized)
+    if(!_skip_col2im)
     {
-        // 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;
-        TensorShape    shape_gemm;
+        TensorShape shape_gemm;
 
-        if(_is_quantized && _skip_col2im)
-        {
-            shape_gemm = output->info()->tensor_shape();
-        }
-        else
-        {
-            // Calculate GEMM output shape
-            shape_gemm = _im2col_output.info()->tensor_shape();
-            shape_gemm.set(0, mat_weights_cols);
-            shape_gemm.set(1, conv_w * conv_h);
-        }
+        // Calculate GEMM output shape
+        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);
@@ -293,62 +349,24 @@ void NEGEMMConvolutionLayer::configure(const ITensor *input, const ITensor *weig
     // Configure GEMM
     // In case we need to skip col2im, GEMM3D (gemm_3d_depth != 0) must be called in order to avoid reshaping the output matrix
     const unsigned int gemm_3d_depth = _skip_col2im ? conv_h : 0;
-    configure_mm(gemm_input_to_use, &_weights_reshaped, gemm_output_to_use, gemm_3d_depth);
+    configure_mm(gemm_input_to_use, &_weights_reshaped, biases, gemm_output_to_use, act_info, gemm_3d_depth);
 
     if(!_skip_im2col)
     {
         _im2col_output.allocator()->allocate();
     }
 
-    // Configure output stage for quantized case
-    if(_is_quantized)
-    {
-        const QuantizationInfo input_quant_info  = input->info()->quantization_info();
-        const QuantizationInfo output_quant_info = (output->info()->total_size() == 0) ? input_quant_info : output->info()->quantization_info();
-
-        float multiplier = input_quant_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);
-
-        if(!_skip_col2im)
-        {
-            _memory_group.manage(&_tmp_output);
-            gemm_output_staged_to_use = &_tmp_output;
-        }
-
-        // Merge activation with output stage
-        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)
-        {
-            const int a_const_int = output_quant_info.quantize(act_info.a(), RoundingPolicy::TO_NEAREST_UP);
-            const int b_const_int = output_quant_info.quantize(act_info.b(), RoundingPolicy::TO_NEAREST_UP);
-
-            min_activation = act_info.activation() != ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU ? output_quant_info.offset : b_const_int;
-            max_activation = act_info.activation() == ActivationLayerInfo::ActivationFunction::RELU ? 255 : a_const_int;
-
-            _is_activationlayer_enabled = false;
-        }
-
-        _gemmlowp_output_stage.configure(gemm_output_to_use, biases, gemm_output_staged_to_use, output_multiplier, output_shift, output_quant_info.offset, min_activation, max_activation);
-    }
-
     if(!_skip_col2im)
     {
         if(_data_layout == DataLayout::NCHW)
         {
             // Configure col2im
-            _col2im_kernel.configure(_is_quantized ? gemm_output_staged_to_use : gemm_output_to_use, output, Size2D(conv_w, conv_h));
+            _col2im_kernel.configure(gemm_output_to_use, output, Size2D(conv_w, conv_h));
         }
         else
         {
             // Configure reshape layer
-            _reshape_layer.configure(_is_quantized ? gemm_output_staged_to_use : gemm_output_to_use, output);
+            _reshape_layer.configure(gemm_output_to_use, output);
         }
     }
 
@@ -395,10 +413,9 @@ Status NEGEMMConvolutionLayer::validate(const ITensorInfo *input, const ITensorI
     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;
+    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);
@@ -420,7 +437,7 @@ Status NEGEMMConvolutionLayer::validate(const ITensorInfo *input, const ITensorI
     bool skip_col2im = false;
     if(data_layout == DataLayout::NHWC)
     {
-        skip_col2im = bool(validate_gemm3d(input->data_type(), conv_h, true));
+        skip_col2im = bool(validate_gemm3d(input, act_info, conv_h, true));
         // If not supported, we need to perform im2col and col2im (or reshape layer)
         if(!skip_col2im)
         {
@@ -431,7 +448,7 @@ Status NEGEMMConvolutionLayer::validate(const ITensorInfo *input, const ITensorI
     if(skip_col2im)
     {
         // If not supported, we need to perform im2col and col2im (or reshape layer)
-        if(!bool(validate_gemm3d(input->data_type(), conv_h, skip_im2col)))
+        if(!bool(validate_gemm3d(input, act_info, conv_h, skip_im2col)))
         {
             skip_im2col = false;
             skip_col2im = false;
@@ -495,68 +512,25 @@ Status NEGEMMConvolutionLayer::validate(const ITensorInfo *input, const ITensorI
     }
 
     // Create temporary GEMM output tensor in case we cannot skip col2im
-    const DataType gemm_data_type = is_quantized ? DataType::S32 : data_type;
     if(!skip_col2im)
     {
         TensorShape 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, gemm_data_type);
+        info_gemm = TensorInfo(shape_gemm, 1, data_type);
     }
     else
     {
-        info_gemm = TensorInfo(output->tensor_shape(), 1, gemm_data_type);
+        info_gemm = TensorInfo(output->tensor_shape(), 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 : 0, skip_im2col));
-
-    if(is_quantized)
-    {
-        const QuantizationInfo input_quant_info  = input->quantization_info();
-        const QuantizationInfo output_quant_info = (output->total_size() == 0) ? input_quant_info : output->quantization_info();
-        const float            multiplier        = input_quant_info.scale * weights_to_use->quantization_info().scale / output_quant_info.scale;
-        int                    output_multiplier, output_shift;
-        quantization::calculate_quantized_multiplier_less_than_one(multiplier, &output_multiplier, &output_shift);
-
-        if(!skip_col2im)
-        {
-            tmp_info = TensorInfo(gemm_output_to_use->tensor_shape(), 1, DataType::QASYMM8);
-            tmp_info.set_quantization_info(output->quantization_info()).set_data_layout(data_layout);
-            gemm_output_staged_to_use = &tmp_info;
-        }
-
-        // Merge activation with output stage
-        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_activation_enabled && supported_acts.count(act_info.activation()) != 0)
-        {
-            const int a_const_int = output_quant_info.quantize(act_info.a(), RoundingPolicy::TO_NEAREST_UP);
-            const int b_const_int = output_quant_info.quantize(act_info.b(), RoundingPolicy::TO_NEAREST_UP);
-
-            min_activation = act_info.activation() != ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU ? output_quant_info.offset : b_const_int;
-            max_activation = act_info.activation() == ActivationLayerInfo::ActivationFunction::RELU ? 255 : a_const_int;
-
-            is_activation_enabled = false;
-        }
-
-        // Validate output stage for quantized case
-        NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint::validate(gemm_output_to_use, biases, gemm_output_staged_to_use, min_activation, max_activation);
-    }
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_mm(gemm_input_to_use, weights_to_use, biases, gemm_output_to_use, act_info, skip_col2im ? conv_h : 0, skip_im2col));
 
     // Validate Col2Im/ReshapeLayer
     if(!skip_col2im && (data_layout == DataLayout::NCHW))
     {
-        ARM_COMPUTE_RETURN_ON_ERROR(NECol2ImKernel::validate(is_quantized ? gemm_output_staged_to_use : gemm_output_to_use,
-                                                             output,
-                                                             Size2D(conv_w, conv_h)));
+        ARM_COMPUTE_RETURN_ON_ERROR(NECol2ImKernel::validate(gemm_output_to_use, output, Size2D(conv_w, conv_h)));
     }
 
     //Validate Activation Layer
@@ -586,9 +560,6 @@ void NEGEMMConvolutionLayer::run()
     {
         // Run gemmlowp
         _mm_gemmlowp.run();
-
-        // Run output stage
-        _gemmlowp_output_stage.run();
     }
     else
     {
diff --git a/src/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.cpp b/src/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.cpp
index 5286f113a5..85e49fd265 100644
--- a/src/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.cpp
+++ b/src/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.cpp
@@ -42,8 +42,8 @@ using namespace arm_compute::misc::shape_calculator;
 
 NEGEMMLowpMatrixMultiplyCore::NEGEMMLowpMatrixMultiplyCore(std::shared_ptr<IMemoryManager> memory_manager)
     : _memory_group(memory_manager), _asm_glue(memory_manager), _mm_kernel(nullptr), _mtx_a_reshape_kernel(nullptr), _mtx_b_reshape_kernel(nullptr), _mtx_a_reduction_kernel(), _mtx_b_reduction_kernel(),
-      _offset_contribution_kernel(), _vector_sum_col(), _vector_sum_row(), _tmp_a(), _tmp_b(), _original_b(nullptr), _a_offset(0), _b_offset(0), _run_vector_matrix_multiplication(false),
-      _dot_product_path(false), _reshape_b_only_on_first_run(false), _is_prepared(false)
+      _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),
+      _run_vector_matrix_multiplication(false), _dot_product_path(false), _reshape_b_only_on_first_run(false), _is_prepared(false), _fuse_output_stage(false)
 {
 }
 
@@ -53,6 +53,9 @@ void NEGEMMLowpMatrixMultiplyCore::configure(const ITensor *a, const ITensor *b,
     ARM_COMPUTE_UNUSED(c);
     ARM_COMPUTE_ERROR_THROW_ON(NEGEMMLowpMatrixMultiplyCore::validate(a->info(), b->info(), c != nullptr ? c->info() : nullptr, output->info(), gemm_info));
 
+    const ITensor *matrix_a = a;
+    const ITensor *matrix_b = b;
+
     // Clear state
     _mtx_a_reshape_kernel = nullptr;
     _mtx_b_reshape_kernel = nullptr;
@@ -65,6 +68,18 @@ void NEGEMMLowpMatrixMultiplyCore::configure(const ITensor *a, const ITensor *b,
     _is_prepared                      = false;
     _original_b                       = b;
 
+    // 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);
+
+        TensorInfo info_mm_result_s32(output->info()->tensor_shape(), 1, DataType::S32);
+
+        _mm_result_s32.allocator()->init(info_mm_result_s32);
+    }
+
 #ifdef __aarch64__
     switch(a->info()->data_type())
     {
@@ -72,7 +87,7 @@ void NEGEMMLowpMatrixMultiplyCore::configure(const ITensor *a, const ITensor *b,
         case DataType::U8:
         case DataType::S8:
         {
-            _asm_glue.configure(a, b, output, 1.f, 0.f, _reshape_b_only_on_first_run);
+            _asm_glue.configure(a, b, _fuse_output_stage ? &_mm_result_s32 : output, 1.f, 0.f, _reshape_b_only_on_first_run);
             _dot_product_path = _asm_glue.is_configured();
             break;
         }
@@ -83,51 +98,35 @@ void NEGEMMLowpMatrixMultiplyCore::configure(const ITensor *a, const ITensor *b,
         }
     }
 #endif /* __aarch64__ */
-    if(!_dot_product_path)
+    if(!(_dot_product_path || _run_vector_matrix_multiplication))
     {
-        if(_run_vector_matrix_multiplication)
+        matrix_a = &_tmp_a;
+        matrix_b = &_tmp_b;
+
+        // The interleaved output matrix will have the following shape: [ a_height * 4, ceil(a_width / 4.0f) ]
+        TensorInfo a_info(compute_interleaved_shape(*a->info()), 1, a->info()->data_type());
+        // The transpose1xW output matrix will have the following shape: [ b_height * 16, ceil(b_width / 16.0f) ]
+        TensorInfo b_info(compute_transpose1xW_shape(*b->info()), 1, b->info()->data_type());
+        _tmp_a.allocator()->init(a_info);
+        _tmp_b.allocator()->init(b_info);
+        _memory_group.manage(&_tmp_a);
+        if(!_reshape_b_only_on_first_run)
         {
-            // Configure matrix multiply kernel
-            {
-                auto k = arm_compute::support::cpp14::make_unique<NEGEMMLowpMatrixMultiplyKernel>();
-                k->configure(a, b, output);
-                _mm_kernel = std::move(k);
-            }
+            _memory_group.manage(&_tmp_b);
         }
-        else
-        {
-            // The interleaved output matrix will have the following shape: [ a_height * 4, ceil(a_width / 4.0f) ]
-            TensorInfo info_a = a->info()->clone()->set_tensor_shape(compute_interleaved_shape(*a->info())).set_is_resizable(true);
-            // The transpose1xW output matrix will have the following shape: [ b_height * 16, ceil(b_width / 16.0f) ]
-            TensorInfo info_b = b->info()->clone()->set_tensor_shape(compute_transpose1xW_shape(*b->info())).set_is_resizable(true);
-            _tmp_a.allocator()->init(info_a);
-            _tmp_b.allocator()->init(info_b);
-            _memory_group.manage(&_tmp_a);
-            if(!_reshape_b_only_on_first_run)
-            {
-                _memory_group.manage(&_tmp_b);
-            }
 
-            // Configure interleave kernel
-            {
-                auto k = arm_compute::support::cpp14::make_unique<NEGEMMInterleave4x4Kernel>();
-                k->configure(a, &_tmp_a);
-                _mtx_a_reshape_kernel = std::move(k);
-            }
-
-            // Configure transpose kernel
-            {
-                auto k = arm_compute::support::cpp14::make_unique<NEGEMMTranspose1xWKernel>();
-                k->configure(b, &_tmp_b);
-                _mtx_b_reshape_kernel = std::move(k);
-            }
+        // Configure interleave kernel
+        {
+            auto k = arm_compute::support::cpp14::make_unique<NEGEMMInterleave4x4Kernel>();
+            k->configure(a, &_tmp_a);
+            _mtx_a_reshape_kernel = std::move(k);
+        }
 
-            // Configure matrix multiply kernel
-            {
-                auto k = arm_compute::support::cpp14::make_unique<NEGEMMLowpMatrixMultiplyKernel>();
-                k->configure(&_tmp_a, &_tmp_b, output);
-                _mm_kernel = std::move(k);
-            }
+        // Configure transpose kernel
+        {
+            auto k = arm_compute::support::cpp14::make_unique<NEGEMMTranspose1xWKernel>();
+            k->configure(b, &_tmp_b);
+            _mtx_b_reshape_kernel = std::move(k);
         }
     }
 
@@ -158,8 +157,33 @@ void NEGEMMLowpMatrixMultiplyCore::configure(const ITensor *a, const ITensor *b,
         _mtx_a_reduction_kernel.configure(a, &_vector_sum_row, a->info()->dimension(0), false);
     }
 
-    // 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(_fuse_output_stage)
+    {
+        // Configure matrix multiply kernel
+        if(!_dot_product_path)
+        {
+            auto k = arm_compute::support::cpp14::make_unique<NEGEMMLowpMatrixMultiplyKernel>();
+            k->configure(matrix_a, matrix_b, &_mm_result_s32);
+            _mm_kernel = std::move(k);
+        }
+
+        _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
+        if(!_dot_product_path)
+        {
+            auto k = arm_compute::support::cpp14::make_unique<NEGEMMLowpMatrixMultiplyKernel>();
+            k->configure(matrix_a, matrix_b, output);
+            _mm_kernel = std::move(k);
+        }
+        // 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);
+    }
 
     // Allocate tensors
     if(!_dot_product_path && !_run_vector_matrix_multiplication)
@@ -185,43 +209,53 @@ void NEGEMMLowpMatrixMultiplyCore::configure(const ITensor *a, const ITensor *b,
 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_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::S32, DataType::QASYMM8);
     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(c != nullptr && gemm_info.gemmlowp_output_stage().type == GEMMLowpOutputStageType::NONE, "Bias addition not supported in NEGEMMLowpMatrixMultiplyCore for output S32");
     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(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");
 
+    const ITensorInfo *matrix_a_info = a;
+    const ITensorInfo *matrix_b_info = b;
+
+    TensorInfo tmp_a_info{};
+    TensorInfo tmp_b_info{};
+    TensorInfo mm_result_s32_info{};
+
     int32_t    a_offset                    = a->quantization_info().offset;
     int32_t    b_offset                    = b->quantization_info().offset;
     const bool reshape_b_only_on_first_run = gemm_info.reshape_b_only_on_first_run();
 
+    bool fuse_output_stage = gemm_info.gemmlowp_output_stage().type != GEMMLowpOutputStageType::NONE;
+    if(fuse_output_stage)
+    {
+        auto_init_if_empty(mm_result_s32_info, a->clone()->set_tensor_shape(output->tensor_shape()).set_data_type(DataType::S32));
+    }
+
     // Check if we need to run the optimized assembly kernel
-    const bool run_optimised = bool(NEGEMMAssemblyDispatch::validate(a, b, output, 1.f, 0.f, reshape_b_only_on_first_run));
+    const bool run_optimised = bool(NEGEMMAssemblyDispatch::validate(a, b, fuse_output_stage ? &mm_result_s32_info : output, 1.f, 0.f, reshape_b_only_on_first_run));
 
     if(run_optimised)
     {
-        if(output->total_size() != 0)
+        ARM_COMPUTE_RETURN_ERROR_ON(b->dimension(0) != output->dimension(0));
+        if(gemm_info.depth_output_gemm3d() != 0)
         {
-            ARM_COMPUTE_RETURN_ERROR_ON(b->dimension(0) != output->dimension(0));
-            if(gemm_info.depth_output_gemm3d() != 0)
+            if(gemm_info.reinterpret_input_as_3d())
             {
-                if(gemm_info.reinterpret_input_as_3d())
-                {
-                    ARM_COMPUTE_RETURN_ERROR_ON(a->dimension(1) != output->dimension(1));
-                    ARM_COMPUTE_RETURN_ERROR_ON(a->dimension(2) != output->dimension(2));
-                }
-                else
-                {
-                    ARM_COMPUTE_RETURN_ERROR_ON(a->dimension(1) != output->dimension(1) * output->dimension(2));
-                }
+                ARM_COMPUTE_RETURN_ERROR_ON(a->dimension(1) != output->dimension(1));
+                ARM_COMPUTE_RETURN_ERROR_ON(a->dimension(2) != output->dimension(2));
             }
             else
             {
-                ARM_COMPUTE_RETURN_ERROR_ON(a->dimension(1) != output->dimension(1));
+                ARM_COMPUTE_RETURN_ERROR_ON(a->dimension(1) != output->dimension(1) * output->dimension(2));
             }
         }
+        else
+        {
+            ARM_COMPUTE_RETURN_ERROR_ON(a->dimension(1) != output->dimension(1));
+        }
     }
     else
     {
@@ -231,6 +265,9 @@ Status NEGEMMLowpMatrixMultiplyCore::validate(const ITensorInfo *a, const ITenso
         const bool run_vector_matrix_multiplication = a->dimension(1) < 2;
         if(!run_vector_matrix_multiplication)
         {
+            matrix_a_info = &tmp_a_info;
+            matrix_b_info = &tmp_b_info;
+
             // The interleaved output matrix will have the following shape: [ a_height * 4, ceil(a_width / 4.0f) ]
             TensorShape shape_tmp_a = a->tensor_shape();
             shape_tmp_a.set(0, a->dimension(0) * 4);
@@ -241,16 +278,12 @@ Status NEGEMMLowpMatrixMultiplyCore::validate(const ITensorInfo *a, const ITenso
             shape_tmp_b.set(0, b->dimension(1) * 16);
             shape_tmp_b.set(1, std::ceil(b->dimension(0) / 16.f));
 
-            TensorInfo info_a = a->clone()->set_tensor_shape(shape_tmp_a).set_is_resizable(true);
-            TensorInfo info_b = b->clone()->set_tensor_shape(shape_tmp_b).set_is_resizable(true);
+            // Validate interleave kernel
+            auto_init_if_empty(tmp_a_info, a->clone()->set_tensor_shape(shape_tmp_a));
+            auto_init_if_empty(tmp_b_info, b->clone()->set_tensor_shape(shape_tmp_b));
 
-            ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMInterleave4x4Kernel::validate(a, &info_a));
-            ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMTranspose1xWKernel::validate(b, &info_b));
-            ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMLowpMatrixMultiplyKernel::validate(&info_a, &info_b, output));
-        }
-        else
-        {
-            ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMLowpMatrixMultiplyKernel::validate(a, b, output));
+            ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMInterleave4x4Kernel::validate(a, &tmp_a_info));
+            ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMTranspose1xWKernel::validate(b, &tmp_b_info));
         }
     }
 
@@ -274,12 +307,32 @@ Status NEGEMMLowpMatrixMultiplyCore::validate(const ITensorInfo *a, const ITenso
         ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMLowpMatrixAReductionKernel::validate(a, &info_vector_sum_row, a->dimension(0), false));
     }
 
-    // Validate offset contribution kernel
-    ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMLowpOffsetContributionKernel::validate(output,
-                                                                             a_offset == 0 ? nullptr : &info_vector_sum_col,
-                                                                             b_offset == 0 ? nullptr : &info_vector_sum_row,
-                                                                             a_offset, b_offset));
+    if(fuse_output_stage)
+    {
+        if(!run_optimised)
+        {
+            ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMLowpMatrixMultiplyKernel::validate(matrix_a_info, matrix_b_info, &mm_result_s32_info));
+        }
 
+        // Validate offset contribution kernel
+        ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMLowpOffsetContributionOutputStageKernel::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
+    {
+        if(!run_optimised)
+        {
+            ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMLowpMatrixMultiplyKernel::validate(matrix_a_info, matrix_b_info, output));
+        }
+        // Validate offset contribution kernel
+        ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMLowpOffsetContributionKernel::validate(output,
+                                                                                 a_offset == 0 ? nullptr : &info_vector_sum_col,
+                                                                                 b_offset == 0 ? nullptr : &info_vector_sum_row,
+                                                                                 a_offset, b_offset));
+    }
     return Status{};
 }
 
@@ -321,8 +374,16 @@ void NEGEMMLowpMatrixMultiplyCore::run()
         NEScheduler::get().schedule(&_mtx_b_reduction_kernel, Window::DimX);
     }
 
-    // Run offset contribution kernel
-    NEScheduler::get().schedule(&_offset_contribution_kernel, Window::DimY);
+    if(_fuse_output_stage)
+    {
+        // Run offset contribution kernel
+        NEScheduler::get().schedule(&_offset_contribution_output_stage_kernel, Window::DimY);
+    }
+    else
+    {
+        // Run offset contribution kernel
+        NEScheduler::get().schedule(&_offset_contribution_kernel, Window::DimY);
+    }
 
     _memory_group.release();
 }