COMPMID-2847: Fuse output stage in GEMMLowpMatrixMultiplyReshapedOnlyRHS

Change-Id: Icd60eb368a34295434e8c141885b4666973a92a1
Signed-off-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/2732
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

1 files changed, 277 insertions, 34 deletions

diff --git a/src/core/CL/kernels/CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel.cpp b/src/core/CL/kernels/CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel.cpp
index 3fa2fad8fd..c4ed691f2e 100644
--- a/src/core/CL/kernels/CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel.cpp
+++ b/src/core/CL/kernels/CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2019 ARM Limited.
+ * Copyright (c) 2019-2020 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -50,21 +50,27 @@ namespace
 {
 using ElementsProcessed = Steps;
 
-Status validate_arguments(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *output, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info,
-                          const GEMMReshapeInfo &gemm_info)
+Status validate_arguments(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *output, const GEMMKernelInfo &gemm_info,
+                          const ITensorInfo *vector_sum_col, const ITensorInfo *vector_sum_row, const ITensorInfo *bias,
+                          const ITensorInfo *output_multipliers, const ITensorInfo *output_shifts)
 {
     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input0, input1, output);
     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED);
     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input0, input1);
     ARM_COMPUTE_RETURN_ERROR_ON_MSG(input0->num_dimensions() > 4, "The number of dimensions for the LHS matrix must be <= 4");
     ARM_COMPUTE_RETURN_ERROR_ON_MSG(input1->num_dimensions() > 3, "The number of dimensions for the RHS matrix must be <= 3");
+
+    const GEMMRHSMatrixInfo       rhs_info     = gemm_info.rhs_info;
+    const GEMMLHSMatrixInfo       lhs_info     = gemm_info.lhs_info;
+    const GEMMLowpOutputStageInfo output_stage = gemm_info.output_stage;
+
     ARM_COMPUTE_RETURN_ERROR_ON_MSG((((rhs_info.k0 & (rhs_info.k0 - 1)) && rhs_info.k0 != 3) || (rhs_info.k0 > 16)), "Only 2,3,4,8,16 are supported for k0");
     ARM_COMPUTE_RETURN_ERROR_ON(lhs_info.m0 < 1 || lhs_info.m0 > 8);
     ARM_COMPUTE_RETURN_ERROR_ON_MSG((((rhs_info.n0 & (rhs_info.n0 - 1)) && rhs_info.n0 != 3) || rhs_info.n0 > 16), "Only 2,3,4,8,16 are supported for n0");
 
-    const int m = gemm_info.m();
-    const int n = gemm_info.n();
-    const int k = gemm_info.k();
+    const int m = gemm_info.m;
+    const int n = gemm_info.n;
+    const int k = gemm_info.k;
 
     TensorShape tensor_shape1{ input1->tensor_shape() };
     tensor_shape1.set(0, n);
@@ -74,7 +80,7 @@ Status validate_arguments(const ITensorInfo *input0, const ITensorInfo *input1,
     const TensorInfo tensor_info_reshaped1 = input1->clone()->set_tensor_shape(compute_rhs_reshaped_shape(tensor_info1, rhs_info));
 
     ARM_COMPUTE_RETURN_ERROR_ON(input0->dimension(0) != static_cast<unsigned int>(k));
-    if(gemm_info.reinterpret_input_as_3d())
+    if(gemm_info.reinterpret_input_as_3d)
     {
         ARM_COMPUTE_RETURN_ERROR_ON(input0->dimension(1) * input0->dimension(2) != static_cast<unsigned int>(m));
     }
@@ -84,23 +90,118 @@ Status validate_arguments(const ITensorInfo *input0, const ITensorInfo *input1,
     }
     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input1, &tensor_info_reshaped1);
 
+    const TensorShape expected_output_shape = compute_mm_shape(*input0, *input1, gemm_info);
     if(output->total_size() != 0)
     {
-        const TensorInfo tensor_info_output = output->clone()->set_tensor_shape(compute_mm_shape(*input0, *input1, gemm_info));
+        const TensorInfo tensor_info_output = output->clone()->set_tensor_shape(expected_output_shape);
         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(output, &tensor_info_output);
-        ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::S32);
+        if(output_stage.type == GEMMLowpOutputStageType::NONE)
+        {
+            ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::S32);
+        }
+        else
+        {
+            ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input0, output);
+        }
+    }
+
+    if(bias != nullptr)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(bias, 1, DataType::S32);
+        ARM_COMPUTE_RETURN_ERROR_ON(bias->num_dimensions() > 1);
+        ARM_COMPUTE_RETURN_ERROR_ON(expected_output_shape[0] != bias->dimension(0));
     }
 
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG((output_stage.type == GEMMLowpOutputStageType::QUANTIZE_DOWN) || (output_stage.type == GEMMLowpOutputStageType::QUANTIZE_DOWN_FLOAT),
+                                    "Only GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT is supported");
+
+    // Checks performed if the output stage needs to be fused
+    if(output_stage.type == GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT)
+    {
+        // If a_offset == 0, vector_sum_col can be a nullptr
+        if(gemm_info.a_offset != 0)
+        {
+            ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(vector_sum_col, 1, DataType::S32);
+            ARM_COMPUTE_RETURN_ERROR_ON(vector_sum_col->dimension(0) != expected_output_shape[0]);
+        }
+
+        // If b_offset == 0, vector_sum_row can be a nullptr
+        if(gemm_info.b_offset != 0)
+        {
+            ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(vector_sum_row, 1, DataType::S32);
+
+            // Check if mm result is a 3D reinterpretation
+            const bool reinterpret_as_3d = expected_output_shape.num_dimensions() > 1 && expected_output_shape.y() != vector_sum_row->tensor_shape().x();
+
+            // Validate input
+            ARM_COMPUTE_RETURN_ERROR_ON(reinterpret_as_3d && vector_sum_row->dimension(0) != (expected_output_shape[1] * expected_output_shape[2]));
+            ARM_COMPUTE_RETURN_ERROR_ON(!reinterpret_as_3d && vector_sum_row->dimension(0) != expected_output_shape[1]);
+
+            if(expected_output_shape.num_dimensions() > 1)
+            {
+                const unsigned int output_batch_idx = reinterpret_as_3d ? 3 : 2;
+
+                TensorShape vector_sum_row_shape = vector_sum_row->tensor_shape();
+                vector_sum_row_shape.collapse_from(1);
+                TensorShape collapsed_output_shape(expected_output_shape);
+                collapsed_output_shape.collapse_from(output_batch_idx);
+
+                ARM_COMPUTE_RETURN_ERROR_ON_MSG(vector_sum_row_shape[1] != collapsed_output_shape[output_batch_idx],
+                                                "vector_sum_row must have the same number of batches of output tensor");
+
+                if(gemm_info.a_offset != 0)
+                {
+                    TensorShape vector_sum_col_shape = vector_sum_col->tensor_shape();
+                    vector_sum_col_shape.collapse_from(1);
+
+                    ARM_COMPUTE_RETURN_ERROR_ON_MSG(vector_sum_col_shape[1] != 1 && vector_sum_col_shape[1] != vector_sum_row_shape[1],
+                                                    "vector_sum_col tensor must have the same number of batches of vector_sum_row_shape or the number of batches must be set to 1");
+                }
+            }
+        }
+
+        PixelValue min_val{};
+        PixelValue max_val{};
+        if(output->total_size() != 0)
+        {
+            ARM_COMPUTE_RETURN_ERROR_ON(output_stage.output_data_type != output->data_type());
+            std::tie(min_val, max_val) = get_min_max(output->data_type());
+            ARM_COMPUTE_RETURN_ERROR_ON(output_stage.gemmlowp_max_bound > max_val.get<int32_t>());
+            ARM_COMPUTE_RETURN_ERROR_ON(output_stage.gemmlowp_min_bound < min_val.get<int32_t>() || output_stage.gemmlowp_min_bound > output_stage.gemmlowp_max_bound);
+        }
+        else
+        {
+            std::tie(min_val, max_val) = get_min_max(output_stage.output_data_type);
+            ARM_COMPUTE_RETURN_ERROR_ON(output_stage.gemmlowp_max_bound > max_val.get<int32_t>());
+            ARM_COMPUTE_RETURN_ERROR_ON(output_stage.gemmlowp_min_bound < min_val.get<int32_t>() || output_stage.gemmlowp_min_bound > output_stage.gemmlowp_max_bound);
+        }
+
+        if(output_multipliers != nullptr && output_shifts != nullptr)
+        {
+            ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output_multipliers, 1, DataType::S32);
+            ARM_COMPUTE_RETURN_ERROR_ON(output_multipliers->num_dimensions() > 1);
+            ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output_shifts, 1, DataType::S32);
+            ARM_COMPUTE_RETURN_ERROR_ON(output_shifts->num_dimensions() > 1);
+            if(output_stage.is_quantized_per_channel)
+            {
+                ARM_COMPUTE_RETURN_ERROR_ON(expected_output_shape[0] != output_shifts->dimension(0));
+                ARM_COMPUTE_RETURN_ERROR_ON(expected_output_shape[0] != output_multipliers->dimension(0));
+            }
+        }
+    }
     return Status{};
 }
 
-std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input0, ITensorInfo *input1, ITensorInfo *output, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info,
-                                                        const GEMMReshapeInfo &gemm_info, ElementsProcessed &num_elements_processed)
+std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input0, ITensorInfo *input1, ITensorInfo *output, const GEMMKernelInfo &gemm_info,
+                                                        ITensorInfo *vector_sum_col, ITensorInfo *vector_sum_row, ITensorInfo *bias,
+                                                        ITensorInfo *output_multipliers, ITensorInfo *output_shifts, ElementsProcessed &num_elements_processed)
 {
+    const GEMMLowpOutputStageInfo output_stage = gemm_info.output_stage;
+
     unsigned int &num_elems_processed_per_iteration_x = num_elements_processed[0];
     unsigned int &num_elems_processed_per_iteration_y = num_elements_processed[1];
-    bool          reinterpret_input_as_3d             = gemm_info.reinterpret_input_as_3d();
-    bool          reinterpret_output_as_3d            = (gemm_info.depth_output_gemm3d() != 0);
+    bool          reinterpret_input_as_3d             = gemm_info.reinterpret_input_as_3d;
+    bool          reinterpret_output_as_3d            = (gemm_info.depth_output_gemm3d != 0);
 
     Window win{};
     Window win_out{};
@@ -114,7 +215,15 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input0, ITe
     }
 
     // Output tensor auto initialization if not yet initialized
-    auto_init_if_empty(*output, input0->clone()->set_tensor_shape(compute_mm_shape(*input0, *input1, gemm_info)).set_data_type(DataType::S32));
+    const TensorShape expected_output_shape = compute_mm_shape(*input0, *input1, gemm_info);
+    if(output_stage.type != GEMMLowpOutputStageType::NONE)
+    {
+        auto_init_if_empty(*output, input0->clone()->set_tensor_shape(expected_output_shape).set_data_type(output_stage.output_data_type));
+    }
+    else
+    {
+        auto_init_if_empty(*output, input0->clone()->set_tensor_shape(expected_output_shape).set_data_type(DataType::S32));
+    }
 
     TensorInfo tmp_info(*output);
 
@@ -128,19 +237,19 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input0, ITe
     }
 
     // Configure kernel window
-    num_elems_processed_per_iteration_x = rhs_info.n0;
-    num_elems_processed_per_iteration_y = lhs_info.m0;
+    num_elems_processed_per_iteration_x = gemm_info.rhs_info.n0;
+    num_elems_processed_per_iteration_y = gemm_info.lhs_info.m0;
 
     // Note: bottom paddings are calculated manually as the output can be reinterpreted as 3D tensor
     // The only way to set properly the paddings, it is to set those explicitly through the AccessWindowStatic
-    const int m          = reinterpret_output_as_3d ? gemm_info.m() : input0->dimension(1);
+    const int m          = reinterpret_output_as_3d ? gemm_info.m : input0->dimension(1);
     const int bottom_pad = (num_elems_processed_per_iteration_y - (m % num_elems_processed_per_iteration_y)) % num_elems_processed_per_iteration_y;
 
     win     = calculate_max_window(tmp_info, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
     win_out = calculate_max_window(*output, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
 
     AccessWindowStatic input0_access(input0, 0, 0,
-                                     ceil_to_multiple(input0->dimension(0), lhs_info.k0),
+                                     ceil_to_multiple(input0->dimension(0), gemm_info.lhs_info.k0),
                                      input0->dimension(1) + bottom_pad);
     AccessWindowStatic input1_access(input1, 0, 0,
                                      input1->dimension(0),
@@ -152,6 +261,30 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input0, ITe
     window_changed = update_window_and_padding(win, input0_access, input1_access) || // window used by the execute_window_loop
                      update_window_and_padding(win_out, output_access);              // window used to update the padding requirements of output tensor
 
+    if(output_stage.type == GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT)
+    {
+        if(gemm_info.a_offset != 0)
+        {
+            AccessWindowHorizontal vector_sum_col_access(vector_sum_col, 0, num_elems_processed_per_iteration_x);
+            window_changed = window_changed || update_window_and_padding(win_out, vector_sum_col_access);
+        }
+        // No access window needed for vector_sum_row
+        ARM_COMPUTE_UNUSED(vector_sum_row);
+
+        if(bias != nullptr)
+        {
+            AccessWindowHorizontal bias_access(bias, 0, num_elems_processed_per_iteration_x);
+            window_changed = window_changed || update_window_and_padding(win_out, bias_access);
+        }
+
+        if(output_multipliers != nullptr && output_multipliers->dimension(0) > 1)
+        {
+            AccessWindowHorizontal output_multipliers_access(output_multipliers, 0, num_elems_processed_per_iteration_x);
+            AccessWindowHorizontal output_shifts_access(output_shifts, 0, num_elems_processed_per_iteration_x);
+            window_changed = window_changed || update_window_and_padding(win_out, output_multipliers_access, output_shifts_access);
+        }
+    }
+
     output_access.set_valid_region(win_out, ValidRegion(Coordinates(), output->tensor_shape()));
 
     // Collapse along the Z direction
@@ -166,23 +299,56 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input0, ITe
 } // namespace
 
 CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel::CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel()
-    : _input0(nullptr), _input1(nullptr), _output(nullptr), _slide_matrix_b(true), _reinterpret_input_as_3d(false), _reinterpret_output_as_3d(false), _use_dummy_work_items(false)
+    : _input0(nullptr),
+      _input1(nullptr),
+      _output(nullptr),
+      _vector_sum_col(nullptr),
+      _vector_sum_row(nullptr),
+      _bias(nullptr),
+      _output_multipliers(nullptr),
+      _output_shifts(nullptr),
+      _slide_matrix_b(true),
+      _reinterpret_input_as_3d(false),
+      _reinterpret_output_as_3d(false),
+      _use_dummy_work_items(false),
+      _is_quantized_per_channel(false),
+      _fuse_output_stage(false)
 {
 }
 
-void CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel::configure(const ICLTensor *input0, const ICLTensor *input1, ICLTensor *output, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info,
-                                                              const GEMMReshapeInfo &gemm_info)
+void CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel::configure(const ICLTensor *input0, const ICLTensor *input1, ICLTensor *output, const GEMMKernelInfo &gemm_info,
+                                                              const ICLTensor *vector_sum_col, const ICLTensor *vector_sum_row, const ICLTensor *bias,
+                                                              const ICLTensor *output_multipliers, const ICLTensor *output_shifts)
 {
     ARM_COMPUTE_ERROR_ON_NULLPTR(input0, input1, output);
-
-    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input0->info(), input1->info(), output->info(), lhs_info, rhs_info, gemm_info));
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input0->info(),
+                                                  input1->info(),
+                                                  output->info(),
+                                                  gemm_info,
+                                                  vector_sum_col != nullptr ? vector_sum_col->info() : nullptr,
+                                                  vector_sum_row != nullptr ? vector_sum_row->info() : nullptr,
+                                                  bias != nullptr ? bias->info() : nullptr,
+                                                  output_multipliers != nullptr ? output_multipliers->info() : nullptr,
+                                                  output_shifts != nullptr ? output_shifts->info() : nullptr));
+
+    const GEMMRHSMatrixInfo       rhs_info     = gemm_info.rhs_info;
+    const GEMMLHSMatrixInfo       lhs_info     = gemm_info.lhs_info;
+    const GEMMLowpOutputStageInfo output_stage = gemm_info.output_stage;
+    const int32_t                 a_offset     = gemm_info.a_offset;
+    const int32_t                 b_offset     = gemm_info.b_offset;
 
     _input0                   = input0;
     _input1                   = input1;
     _output                   = output;
-    _reinterpret_input_as_3d  = gemm_info.reinterpret_input_as_3d();
-    _reinterpret_output_as_3d = (gemm_info.depth_output_gemm3d() != 0);
+    _vector_sum_col           = vector_sum_col;
+    _vector_sum_row           = vector_sum_row;
+    _bias                     = bias;
+    _output_multipliers       = output_multipliers;
+    _output_shifts            = output_shifts;
+    _reinterpret_input_as_3d  = gemm_info.reinterpret_input_as_3d;
+    _reinterpret_output_as_3d = (gemm_info.depth_output_gemm3d != 0);
     _use_dummy_work_items     = preferred_dummy_work_items_support(CLKernelLibrary::get().get_device());
+    _is_quantized_per_channel = output_stage.is_quantized_per_channel;
 
     // In case both input and output have to be reinterpreted as 3D tensors,
     // force reinterpret_input_as_3d and reinterpret_output_as_3d to be false.
@@ -199,7 +365,16 @@ void CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel::configure(const ICLTensor *i
     ElementsProcessed num_elements_processed{};
 
     // Configure kernel window
-    auto win_config = validate_and_configure_window(input0->info(), input1->info(), output->info(), lhs_info, rhs_info, gemm_info, num_elements_processed);
+    auto win_config = validate_and_configure_window(input0->info(),
+                                                    input1->info(),
+                                                    output->info(),
+                                                    gemm_info,
+                                                    vector_sum_col != nullptr ? vector_sum_col->info() : nullptr,
+                                                    vector_sum_row != nullptr ? vector_sum_row->info() : nullptr,
+                                                    bias != nullptr ? bias->info() : nullptr,
+                                                    output_multipliers != nullptr ? output_multipliers->info() : nullptr,
+                                                    output_shifts != nullptr ? output_shifts->info() : nullptr,
+                                                    num_elements_processed);
     ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
     ICLKernel::configure_internal(win_config.second);
 
@@ -213,8 +388,8 @@ void CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel::configure(const ICLTensor *i
     build_opts.add_option_if(rhs_info.interleave, "-DRHS_INTERLEAVE");
     build_opts.add_option_if(_use_dummy_work_items, "-DDUMMY_WORK_ITEMS");
     build_opts.add_option("-DM=" + support::cpp11::to_string(input0->info()->dimension(1)));
-    build_opts.add_option("-DN=" + support::cpp11::to_string(gemm_info.n()));
-    build_opts.add_option("-DK=" + support::cpp11::to_string(gemm_info.k()));
+    build_opts.add_option("-DN=" + support::cpp11::to_string(gemm_info.n));
+    build_opts.add_option("-DK=" + support::cpp11::to_string(gemm_info.k));
     build_opts.add_option("-DM0=" + support::cpp11::to_string(lhs_info.m0));
     build_opts.add_option("-DN0=" + support::cpp11::to_string(rhs_info.n0));
     build_opts.add_option("-DK0=" + support::cpp11::to_string(rhs_info.k0));
@@ -225,6 +400,35 @@ void CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel::configure(const ICLTensor *i
     std::string kernel_name("gemmlowp_mm_reshaped_only_rhs_");
     kernel_name += rhs_info.transpose ? "t" : "nt";
 
+    if(output_stage.type == GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT)
+    {
+        kernel_name += "_fused_output_stage_fixedpoint";
+        _fuse_output_stage = true;
+        // If a_offset == 0, vector_sum_col can be a nullptr
+        if(a_offset != 0)
+        {
+            build_opts.add_option("-DA_OFFSET=" + support::cpp11::to_string(a_offset));
+            build_opts.add_option_if(vector_sum_col->info()->tensor_shape().num_dimensions() > 1, "-DSUM_COL_HAS_BATCHES");
+        }
+        // If b_offset == 0, vector_sum_row can be a nullptr
+        build_opts.add_option_if(b_offset != 0, "-DB_OFFSET=" + support::cpp11::to_string(b_offset));
+        build_opts.add_option("-DK_OFFSET=" + support::cpp11::to_string(a_offset * b_offset * input0->info()->dimension(0)));
+        build_opts.add_option_if(bias != nullptr, "-DADD_BIAS");
+        build_opts.add_option("-DRESULT_OFFSET=" + support::cpp11::to_string(output_stage.gemmlowp_offset));
+        build_opts.add_option("-DRESULT_MULTIPLIER=" + support::cpp11::to_string(output_stage.gemmlowp_multipliers[0]));
+        build_opts.add_option("-DRESULT_SHIFT=" + support::cpp11::to_string(output_stage.gemmlowp_shifts[0]));
+        build_opts.add_option_if(_is_quantized_per_channel, "-DPER_CHANNEL_QUANTIZATION");
+
+        const int min = output_stage.gemmlowp_min_bound;
+        const int max = output_stage.gemmlowp_max_bound;
+
+        PixelValue min_val{};
+        PixelValue max_val{};
+        std::tie(min_val, max_val) = get_min_max(output->info()->data_type());
+        build_opts.add_option_if((min != min_val.get<int32_t>()) && (min != max), "-DMIN_BOUND=" + support::cpp11::to_string(min));
+        build_opts.add_option_if((max != max_val.get<int32_t>()) && (min != max), "-DMAX_BOUND=" + support::cpp11::to_string(max));
+    }
+
     // Create kernel
     _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name, build_opts.options()));
 
@@ -239,7 +443,7 @@ void CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel::configure(const ICLTensor *i
     _config_id += "_";
     _config_id += support::cpp11::to_string(output->info()->dimension(0));
     _config_id += "_";
-    _config_id += support::cpp11::to_string(gemm_info.k());
+    _config_id += support::cpp11::to_string(gemm_info.k);
     _config_id += "_";
     _config_id += support::cpp11::to_string(output->info()->dimension(2));
     _config_id += "_";
@@ -254,17 +458,21 @@ void CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel::configure(const ICLTensor *i
     _config_id += support::cpp11::to_string(rhs_info.interleave);
 }
 
-Status CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel::validate(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *output, const GEMMLHSMatrixInfo &lhs_info,
-                                                               const GEMMRHSMatrixInfo &rhs_info, const GEMMReshapeInfo &gemm_info)
+Status CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel::validate(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *output, const GEMMKernelInfo &gemm_info,
+                                                               const ITensorInfo *vector_sum_col, const ITensorInfo *vector_sum_row, const ITensorInfo *bias,
+                                                               const ITensorInfo *output_multipliers, const ITensorInfo *output_shifts)
 {
     ElementsProcessed num_elements_processed{};
-    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input0, input1, output, lhs_info, rhs_info, gemm_info));
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input0, input1, output, gemm_info, vector_sum_col, vector_sum_row, bias, output_multipliers, output_shifts));
     ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input0->clone().get(),
                                                               input1->clone().get(),
                                                               output->clone().get(),
-                                                              lhs_info,
-                                                              rhs_info,
                                                               gemm_info,
+                                                              vector_sum_col != nullptr ? vector_sum_col->clone().get() : nullptr,
+                                                              vector_sum_row != nullptr ? vector_sum_row->clone().get() : nullptr,
+                                                              bias != nullptr ? bias->clone().get() : nullptr,
+                                                              output_multipliers != nullptr ? output_multipliers->clone().get() : nullptr,
+                                                              output_shifts != nullptr ? output_shifts->clone().get() : nullptr,
                                                               num_elements_processed)
                                 .first);
 
@@ -304,6 +512,21 @@ void CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel::run(const Window &window, cl
         _kernel.setArg<cl_uint>(idx0, static_cast<unsigned int>(total_cross_plane_pad));
     }
 
+    // Set window for vector_sum_col
+    Window win_vector_sum_col = slice;
+    win_vector_sum_col.set(Window::DimY, Window::Dimension(0, 0, 0));
+    win_vector_sum_col.set(Window::DimZ, Window::Dimension(0, 0, 0));
+
+    // Set window for vector_sum_row
+    Window win_vector_sum_row = slice;
+    win_vector_sum_row.set(Window::DimX, Window::Dimension(0, 0, 0));
+    win_vector_sum_row.set(Window::DimY, Window::Dimension(0, 0, 0));
+    win_vector_sum_col.set(Window::DimZ, Window::Dimension(0, 0, 0));
+
+    Window biases_slice = slice;
+    biases_slice.set(Window::DimY, Window::Dimension(0, 1, 1));
+    biases_slice.set(Window::DimZ, Window::Dimension(0, 1, 1));
+
     do
     {
         Window slice_b = slice;
@@ -321,6 +544,26 @@ void CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel::run(const Window &window, cl
         _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_input0->info()->strides_in_bytes()[2]));
         _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_input1->info()->strides_in_bytes()[2]));
         _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_output->info()->strides_in_bytes()[2]));
+        if(_reinterpret_input_as_3d)
+        {
+            // Pass bottom paddings to the kernel if the input has to be reinterpreted as 3D tensor
+            idx++;
+        }
+
+        if(_reinterpret_output_as_3d)
+        {
+            // Pass bottom paddings to the kernel if the output has to be reinterpreted as 3D tensor
+            idx++;
+        }
+
+        if(_fuse_output_stage)
+        {
+            add_2D_tensor_argument_if((_vector_sum_col != nullptr), idx, _vector_sum_col, win_vector_sum_col);
+            add_2D_tensor_argument_if((_vector_sum_row != nullptr), idx, _vector_sum_row, win_vector_sum_row);
+            add_1D_tensor_argument_if((_bias != nullptr), idx, _bias, biases_slice);
+            add_1D_tensor_argument_if(_is_quantized_per_channel, idx, _output_multipliers, biases_slice);
+            add_1D_tensor_argument_if(_is_quantized_per_channel, idx, _output_shifts, biases_slice);
+        }
         enqueue(queue, *this, slice, lws_hint(), _use_dummy_work_items);
     }
     while(window.slide_window_slice_3D(slice));