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>

11 files changed, 1008 insertions, 153 deletions

diff --git a/arm_compute/core/CL/kernels/CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel.h b/arm_compute/core/CL/kernels/CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel.h
index 9dd5496c00..7845d244eb 100644
--- a/arm_compute/core/CL/kernels/CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel.h
+++ b/arm_compute/core/CL/kernels/CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2019 ARM Limited.
+ * Copyright (c) 2019-2020 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -25,6 +25,7 @@
 #define ARM_COMPUTE_CLGEMMLOWPMATRIXMULTIPLYRESHAPEDONLYRHSKERNEL_H
 
 #include "arm_compute/core/CL/ICLKernel.h"
+#include "arm_compute/core/KernelDescriptors.h"
 
 namespace arm_compute
 {
@@ -33,6 +34,7 @@ class ICLTensor;
 /** OpenCL kernel to multiply matrices with QASYMM8 data type when only the input matrix RHS (input1) has been reshaped
  *
  * @note The input matrix input1 must be reshaped through @ref CLGEMMReshapeRHSMatrixKernel
+ * @note For fused output stage, only GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT type is supported
  */
 class CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel : public ICLKernel
 {
@@ -49,37 +51,59 @@ public:
     CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel &operator=(CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel &&) = default;
     /** Initialise the kernel's input and output.
      *
-     * @param[in]  input0    Input tensor containing the LHS matrix. Data type supported: QASYMM8/QASYMM8_SIGNED
-     * @param[in]  input1    Input tensor containing the RHS reshaped matrix. Data type supported: same as @p input0
-     * @param[out] output    Output tensor to store the result of matrix multiplication. Data type supported: S32
-     * @param[in]  lhs_info  LHS matrix information used to retrieve the number of rows to be processed by each thread
-     *                       lhs_info.m0: 2,3,4,5,6,7,8
-     *                       lhs_info.k0: 2,3,4,8,16
-     * @param[in]  rhs_info  RHS matrix information used for reshaping the input1 tensor.  Only the following values are supported:
-     *                       rhs_info.n0: 2,3,4,8,16
-     *                       rhs_info.k0: 2,3,4,8,16
-     *                       rhs_info.transpose: true
-     * @param[in]  gemm_info GEMM information used to retrieve the original dimensions of the input matrices
+     * @param[in]  input0             Input tensor containing the LHS matrix. Data type supported: QASYMM8/QASYMM8_SIGNED
+     * @param[in]  input1             Input tensor containing the RHS reshaped matrix. Data type supported: same as @p input0
+     * @param[out] output             Output tensor. Data type supported: QASYMM8/QASYMM8_SIGNED/S32.
+     * @param[in]  gemm_info          GEMM information used to retrieve the original dimensions of the input matrices, output stage information and RHS/LHS info.
+     *                                Only the following values are supported for LHS info:
+     *                                lhs_info.m0: 2,3,4,5,6,7,8
+     *                                lhs_info.k0: 2,3,4,8,16
+     *                                Only the following values are supported for RHS info:
+     *                                rhs_info.n0: 2,3,4,8,16
+     *                                rhs_info.k0: same as lhs_info.k0
+     *                                rhs_info.transpose: true
+     * @param[in]  vector_sum_col     (Optional) Input row-vector of sums of all the entries in each column of matrix B.
+     *                                Note: vector_sum_col can be a nullptr in case a_offset = 0. Data type supported: S32
+     * @param[in]  vector_sum_row     (Optional) Input row-vector of sums of all the entries in each row of matrix A.
+     *                                Note: vector_sum_row can be a nullptr in case b_offset = 0. Data type supported: S32
+     * @param[in]  bias               (Optional) Biases tensor. Only shared biases supported and it can be a nullptr if the addition of biases is not required.
+     *                                Biases are 1D tensor with dimensions [OFM]. Data type supported: S32.
+     * @param[in]  output_multipliers (Optional) Output multipliers tensor. In case of per-channel quantization, the number of multipliers must be equal to the number of filters (OFM).
+     *                                Supported data types: S32.
+     * @param[in]  output_shifts      (Optional) Output shifts tensor. In case of per-channel quantization, the number of multipliers must be equal to the number of filters (OFM).
+     *                                Supported data types: S32.
      */
-    void configure(const ICLTensor *input0, const ICLTensor *input1, ICLTensor *output, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info, const GEMMReshapeInfo &gemm_info);
+    void configure(const ICLTensor *input0, const ICLTensor *input1, ICLTensor *output, const GEMMKernelInfo &gemm_info, const ICLTensor *vector_sum_col = nullptr,
+                   const ICLTensor *vector_sum_row = nullptr, const ICLTensor *bias = nullptr, const ICLTensor *output_multipliers = nullptr, const ICLTensor *output_shifts = nullptr);
     /** Static function to check if given info will lead to a valid configuration of @ref CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel
      *
-     * @param[in] input0    Input tensor info for the LHS matrix. Data type supported: QASYMM8/QASYMM8_SIGNED
-     * @param[in] input1    Input tensor info for the RHS reshaped matrix. Data type supported: same as @p input0
-     * @param[in] output    Output tensor info. Data type supported: S32
-     * @param[in] lhs_info  LHS matrix information used to retrieve the number of rows to be processed by each thread
-     *                      lhs_info.m0: 2,3,4,5,6,7,8
-     *                      lhs_info.k0: 2,3,4,8,16
-     * @param[in] rhs_info  RHS matrix information used for reshaping the input1 tensor.  Only the following values are supported:
-     *                      rhs_info.n0: 2,3,4,8,16
-     *                      rhs_info.k0: same as lhs_info.k0
-     *                      rhs_info.transpose: true
-     * @param[in] gemm_info GEMM information used to retrieve the original dimensions of the input matrices
+     * @param[in] input0             Input tensor info for the LHS matrix. Data type supported: QASYMM8/QASYMM8_SIGNED
+     * @param[in] input1             Input tensor info for the RHS reshaped matrix. Data type supported: same as @p input0
+     * @param[in] output             Output tensor info. Data type supported: QASYMM8/QASYMM8_SIGNED/S32.
+     * @param[in] gemm_info          GEMM information used to retrieve the original dimensions of the input matrices, output stage information and RHS/LHS info.
+     *                               Only the following values are supported for LHS info:
+     *                               lhs_info.m0: 2,3,4,5,6,7,8
+     *                               lhs_info.k0: 2,3,4,8,16
+     *                               Only the following values are supported for RHS info:
+     *                               rhs_info.n0: 2,3,4,8,16
+     *                               rhs_info.k0: same as lhs_info.k0
+     *                               rhs_info.transpose: true
+     * @param[in] vector_sum_col     (Optional) Input row-vector info of sums of all the entries in each column of matrix B.
+     *                               Note: vector_sum_col can be a nullptr in case a_offset = 0. Data type supported: S32
+     * @param[in] vector_sum_row     (Optional) Input row-vector info of sums of all the entries in each row of matrix A.
+     *                               Note: vector_sum_row can be a nullptr in case b_offset = 0. Data type supported: S32
+     * @param[in] bias               (Optional) Biases tensor info. Only shared biases supported and it can be a nullptr if the addition of biases is not required.
+     *                               Biases are 1D tensor with dimensions [OFM]. Data type supported: S32.
+     * @param[in] output_multipliers (Optional) Output multipliers tensor info. In case of per-channel quantization, the number of multipliers must be equal to the number of filters (OFM).
+     *                               Supported data types: S32.
+     * @param[in] output_shifts      (Optional) Output shifts tensor info. In case of per-channel quantization, the number of multipliers must be equal to the number of filters (OFM).
+     *                               Supported data types: S32.
      *
      * @return a status
      */
-    static Status validate(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *output, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info,
-                           const GEMMReshapeInfo &gemm_info);
+    static Status validate(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *output, const GEMMKernelInfo &gemm_info, const ITensorInfo *vector_sum_col = nullptr,
+                           const ITensorInfo *vector_sum_row = nullptr, const ITensorInfo *bias = nullptr, const ITensorInfo *output_multipliers = nullptr,
+                           const ITensorInfo *output_shifts = nullptr);
 
     // Inherited methods overridden:
     void run(const Window &window, cl::CommandQueue &queue) override;
@@ -88,10 +112,17 @@ private:
     const ICLTensor *_input0;
     const ICLTensor *_input1;
     ICLTensor       *_output;
+    const ICLTensor *_vector_sum_col;
+    const ICLTensor *_vector_sum_row;
+    const ICLTensor *_bias;
+    const ICLTensor *_output_multipliers;
+    const ICLTensor *_output_shifts;
     bool             _slide_matrix_b;
     bool             _reinterpret_input_as_3d;
     bool             _reinterpret_output_as_3d;
     bool             _use_dummy_work_items;
+    bool             _is_quantized_per_channel;
+    bool             _fuse_output_stage;
 };
 } // namespace arm_compute
 #endif /* ARM_COMPUTE_CLGEMMLOWPMATRIXMULTIPLYRESHAPEDONLYRHSKERNEL_H */
\ No newline at end of file
diff --git a/arm_compute/core/KernelDescriptors.h b/arm_compute/core/KernelDescriptors.h
index 4b04bebdef..58400b190b 100644
--- a/arm_compute/core/KernelDescriptors.h
+++ b/arm_compute/core/KernelDescriptors.h
@@ -54,14 +54,21 @@ struct FFTRadixStageKernelInfo
 /** Descriptor used by the GEMM kernels */
 struct GEMMKernelInfo
 {
-    unsigned int        m{ 0 };                           /**< Number of LHS rows*/
-    unsigned int        n{ 0 };                           /**< Number of RHS columns*/
-    unsigned int        k{ 0 };                           /**< Number of LHS columns or RHS rows */
-    unsigned int        depth_output_gemm3d{ 0 };         /**< Depth of the output tensor in case is reinterpreted as 3D */
-    bool                reinterpret_input_as_3d{ false }; /**< Flag used to reinterpret the input as 3D */
-    bool                broadcast_bias{ false };          /**< Flag used to broadcase the bias addition */
-    bool                fp_mixed_precision{ false };      /**< Flag used to indicate wider accumulators (32 bit instead of 16 for FP16). */
-    ActivationLayerInfo activation_info{};                /**< Activation function to perform after the matrix multiplication */
+    unsigned int            m{ 0 };                           /**< Number of LHS rows*/
+    unsigned int            n{ 0 };                           /**< Number of RHS columns*/
+    unsigned int            k{ 0 };                           /**< Number of LHS columns or RHS rows */
+    unsigned int            depth_output_gemm3d{ 0 };         /**< Depth of the output tensor in case is reinterpreted as 3D */
+    bool                    reinterpret_input_as_3d{ false }; /**< Flag used to reinterpret the input as 3D */
+    bool                    broadcast_bias{ false };          /**< Flag used to broadcast the bias addition */
+    bool                    fp_mixed_precision{ false };      /**< Flag used to indicate wider accumulators (32 bit instead of 16 for FP16). */
+    ActivationLayerInfo     activation_info{};                /**< Activation function to perform after the matrix multiplication */
+    int                     mult_transpose1xW_width{ 1 };     /**< Multiplication factor for the width of the 1xW transposed block */
+    int                     mult_interleave4x4_height{ 1 };   /**< Multiplication factor for the height of the 4x4 interleaved block */
+    GEMMLHSMatrixInfo       lhs_info{};                       /**< LHS matrix information used to retrieve the number of rows processed by each thread */
+    GEMMRHSMatrixInfo       rhs_info{};                       /**< RHS matrix information used for reshaping the RHS matrix */
+    int32_t                 a_offset{ 0 };                    /**< Offset to be added to each element of the matrix A */
+    int32_t                 b_offset{ 0 };                    /**< Offset to be added to each element of the matrix B */
+    GEMMLowpOutputStageInfo output_stage{};                   /**< GEMMLowp output stage information */
 };
 
 /** Descriptor used by the depthwise convolution kernels */
diff --git a/arm_compute/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.h b/arm_compute/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.h
index 66c5e9ee46..c9b1b70c54 100644
--- a/arm_compute/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.h
+++ b/arm_compute/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.h
@@ -135,8 +135,9 @@ private:
     bool    _is_midgard;
     bool    _reshape_b_only_on_first_run;
     bool    _is_prepared;
-    bool    _fuse_output_stage;
+    bool    _run_output_stage;
     bool    _convert_to_qasymm8;
+    bool    _run_offset_contribution;
 };
 } // namespace arm_compute
 #endif /*ARM_COMPUTE_CLGEMMLOWPMATRIXMULTIPLYCORE_H */
\ No newline at end of file
diff --git a/src/core/CL/CLKernelLibrary.cpp b/src/core/CL/CLKernelLibrary.cpp
index 5b59094c81..5d67240299 100644
--- a/src/core/CL/CLKernelLibrary.cpp
+++ b/src/core/CL/CLKernelLibrary.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2019 ARM Limited.
+ * Copyright (c) 2016-2020 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -341,6 +341,7 @@ const std::map<std::string, std::string> CLKernelLibrary::_kernel_program_map =
     { "gemmlowp_mm_native", "gemmlowp.cl" },
     { "gemmlowp_mm_reshaped_lhs_nt_rhs_t", "gemmlowp.cl" },
     { "gemmlowp_mm_reshaped_only_rhs_t", "gemmlowp.cl" },
+    { "gemmlowp_mm_reshaped_only_rhs_t_fused_output_stage_fixedpoint", "gemmlowp.cl" },
     { "gemmlowp_offset_contribution", "gemmlowp.cl" },
     { "gemmlowp_offset_contribution_quantize_down", "gemmlowp.cl" },
     { "gemmlowp_offset_contribution_quantize_down_fixedpoint", "gemmlowp.cl" },
diff --git a/src/core/CL/cl_kernels/gemm_helpers.h b/src/core/CL/cl_kernels/gemm_helpers.h
index 66e83c3558..79a9f094df 100644
--- a/src/core/CL/cl_kernels/gemm_helpers.h
+++ b/src/core/CL/cl_kernels/gemm_helpers.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2019 ARM Limited.
+ * Copyright (c) 2019-2020 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -140,6 +140,118 @@
 #define LOAD_BLOCK(M0, N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z) LOAD_BLOCK_STR(M0, N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y, Z)
 /** @} */ // end of group LOAD_BLOCK
 
+/** Loads the elements from 0 to n-1 in the given variables (BASENAME0 to BASENAMEn-1).
+ * @name LOAD_ELEMENT_n
+ *
+ * @param[in] N0        The number of rows to load
+ * @param[in] DATA_TYPE The data type of variables
+ * @param[in] BASENAME  The basename of the destination variables for the loaded rows
+ * @param[in] PTR       The base pointer
+ * @param[in] OFFSET    The offset within a row
+ * @param[in] STRIDE_Y  The stride value in y-axis direction
+ * @{
+ */
+#define LOAD_ELEMENT_1(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y) \
+    VEC_DATA_TYPE(DATA_TYPE, N0)                                       \
+    BASENAME##0 = *((__global DATA_TYPE *)(PTR + OFFSET + 0 * STRIDE_Y));
+
+#define LOAD_ELEMENT_2(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y) \
+    LOAD_ELEMENT_1(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y)     \
+    VEC_DATA_TYPE(DATA_TYPE, N0)                                       \
+    BASENAME##1 = *((__global DATA_TYPE *)(PTR + OFFSET + 1 * STRIDE_Y));
+
+#define LOAD_ELEMENT_3(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y) \
+    LOAD_ELEMENT_2(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y)     \
+    VEC_DATA_TYPE(DATA_TYPE, N0)                                       \
+    BASENAME##2 = *((__global DATA_TYPE *)(PTR + OFFSET + 2 * STRIDE_Y));
+
+#define LOAD_ELEMENT_4(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y) \
+    LOAD_ELEMENT_3(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y)     \
+    VEC_DATA_TYPE(DATA_TYPE, N0)                                       \
+    BASENAME##3 = *((__global DATA_TYPE *)(PTR + OFFSET + 3 * STRIDE_Y));
+
+#define LOAD_ELEMENT_5(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y) \
+    LOAD_ELEMENT_4(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y)     \
+    VEC_DATA_TYPE(DATA_TYPE, N0)                                       \
+    BASENAME##4 = *((__global DATA_TYPE *)(PTR + OFFSET + 4 * STRIDE_Y));
+
+#define LOAD_ELEMENT_6(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y) \
+    LOAD_ELEMENT_5(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y)     \
+    VEC_DATA_TYPE(DATA_TYPE, N0)                                       \
+    BASENAME##5 = *((__global DATA_TYPE *)(PTR + OFFSET + 5 * STRIDE_Y));
+
+#define LOAD_ELEMENT_7(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y) \
+    LOAD_ELEMENT_6(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y)     \
+    VEC_DATA_TYPE(DATA_TYPE, N0)                                       \
+    BASENAME##6 = *((__global DATA_TYPE *)(PTR + OFFSET + 6 * STRIDE_Y));
+
+#define LOAD_ELEMENT_8(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y) \
+    LOAD_ELEMENT_7(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y)     \
+    VEC_DATA_TYPE(DATA_TYPE, N0)                                       \
+    BASENAME##7 = *((__global DATA_TYPE *)(PTR + OFFSET + 7 * STRIDE_Y));
+
+#define LOAD_ELEMENT_9(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y) \
+    LOAD_ELEMENT_8(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y)     \
+    VEC_DATA_TYPE(DATA_TYPE, N0)                                       \
+    BASENAME##8 = *((__global DATA_TYPE *)(PTR + OFFSET + 8 * STRIDE_Y));
+
+#define LOAD_ELEMENT_10(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y) \
+    LOAD_ELEMENT_9(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y)      \
+    VEC_DATA_TYPE(DATA_TYPE, N0)                                        \
+    BASENAME##9 = *((__global DATA_TYPE *)(PTR + OFFSET + 9 * STRIDE_Y));
+
+#define LOAD_ELEMENT_11(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y) \
+    LOAD_ELEMENT_10(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y)     \
+    VEC_DATA_TYPE(DATA_TYPE, N0)                                        \
+    BASENAME##A = *((__global DATA_TYPE *)(PTR + OFFSET + 10 * STRIDE_Y));
+
+#define LOAD_ELEMENT_12(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y) \
+    LOAD_ELEMENT_11(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y)     \
+    VEC_DATA_TYPE(DATA_TYPE, N0)                                        \
+    BASENAME##B = *((__global DATA_TYPE *)(PTR + OFFSET + 11 * STRIDE_Y));
+
+#define LOAD_ELEMENT_13(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y) \
+    LOAD_ELEMENT_12(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y)     \
+    VEC_DATA_TYPE(DATA_TYPE, N0)                                        \
+    BASENAME##C = *((__global DATA_TYPE *)(PTR + OFFSET + 12 * STRIDE_Y));
+
+#define LOAD_ELEMENT_14(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y) \
+    LOAD_ELEMENT_13(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y)     \
+    VEC_DATA_TYPE(DATA_TYPE, N0)                                        \
+    BASENAME##D = *((__global DATA_TYPE *)(PTR + OFFSET + 13 * STRIDE_Y));
+
+#define LOAD_ELEMENT_15(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y) \
+    LOAD_ELEMENT_14(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y)     \
+    VEC_DATA_TYPE(DATA_TYPE, N0)                                        \
+    BASENAME##E = *((__global DATA_TYPE *)(PTR + OFFSET + 14 * STRIDE_Y));
+
+#define LOAD_ELEMENT_16(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y) \
+    LOAD_ELEMENT_15(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y)     \
+    VEC_DATA_TYPE(DATA_TYPE, N0)                                        \
+    BASENAME##F = *((__global DATA_TYPE *)(PTR + OFFSET + 15 * STRIDE_Y));
+
+/** @}*/ // end of group LOAD_ELEMENT_n
+
+/** Load Scalar as Vector (consecutive elements).
+ * @name LOAD_SCALAR_AS_VECTOR
+ *
+ * Supported cases are M0=1,2,3,...,16 and N0=1,2,3,4,8,16
+ * The data to load is expected to have consecutive names for each row.
+ * E.g., for M0=3, and BASENAME=c, the expected data is c0, c1 and c2.
+ *
+ * @param[in] M0        The number of consecutive rows
+ * @param[in] N0        The number of consecutive columns
+ * @param[in] DATA_TYPE The data type of the target
+ * @param[in] BASENAME  The basename of the result variables
+ * @param[in] PTR       The base pointer for the data
+ * @param[in] OFFSET    The offset within a row
+ * @param[in] STRIDE_Y  The stride in y-axis direction
+ * @{
+ */
+#define LOAD_SCALAR_AS_VECTOR_STR(M0, N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y) LOAD_ELEMENT_##M0(N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y)
+#define LOAD_SCALAR_AS_VECTOR(M0, N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y) LOAD_SCALAR_AS_VECTOR_STR(M0, N0, DATA_TYPE, BASENAME, PTR, OFFSET, STRIDE_Y)
+/** @} */ // end of group LOAD_SCALAR_AS_VECTOR
+
 /** Basic macros to calculate Z offset values from Z0 to Zn-1
  * @name CALCULATE_Z_OFFSET_n
  *
diff --git a/src/core/CL/cl_kernels/gemmlowp.cl b/src/core/CL/cl_kernels/gemmlowp.cl
index 8140b8f2a5..a2a47d7823 100644
--- a/src/core/CL/cl_kernels/gemmlowp.cl
+++ b/src/core/CL/cl_kernels/gemmlowp.cl
@@ -814,6 +814,301 @@ __kernel void gemmlowp_mm_reshaped_only_rhs_t(IMAGE_DECLARATION(lhs),
 #undef RHS_OFFSET_X
 #undef RHS_STEP_X
 }
+
+#if defined(RESULT_OFFSET) && defined(RESULT_SHIFT) && defined(RESULT_MULTIPLIER)
+/** This OpenCL kernel computes the matrix multiplication between 2 matrices with fused output stage using fixed-point arithmetic.
+ *  The LHS matrix is NOT reshaped
+ *  The RHS matrix is reshaped with @ref CLGEMMReshapeRHSMatrixKernel and the block K0xN0 is transposed
+ *
+ * @note The input data type must be passed at compile time using -DDATA_TYPE (i.e. -DDATA_TYPE=uchar)
+ * @note The accumulator data type must be passed at compile time using -DACC_DATA_TYPE (i.e. -DACC_DATA_TYPE=uint)
+ * @note The number of columns of LHS matrix must be passed at compile time using -DK (i.e. -DK=64)
+ * @note The block's dimensions used for reshaping the RHS matrix (N0 and K0) must be passed at compile time using -DN0 and -DK0 (i.e. -DN0=8, -DK0=4).
+ * @note The number of M0 rows to process must be passed at compile time using -DM0 (i.e. -DM0=2)
+ * @note The number of K0xN0 horizontal blocks stored on the same output row of the reshaped RHS matrix must be passed at compile time using -DH0 (i.e. -DH0=2)
+ * @note If the K0xN0 blocks in the reshaped RHS matrix have been interleaved, the option -DRHS_INTERLEAVE must passed at compile time.
+ * @note Only the following configurations of M0, N0 and K0 are currently supported:
+ *  - M0 = 1, 2, 3, 4, 5, 6, 7, 8
+ *  - N0 = 2, 3, 4, 8, 16
+ *  - K0 = 2, 3, 4, 8, 16
+ *  - H0 >= 1
+ *
+ * @note In case the input or output have to be reinterpreted as a 3D tensor, the following information must be passed at compile time:
+ *       -# REINTERPRET_INPUT_AS_3D: To reinterpret the input as 3D
+ *       -# REINTERPRET_OUTPUT_AS_3D: To reinterpret the output as 3D
+ *       -# HEIGHT_GEMM3D: The height of the output in case it has to be reinterpreted as a 3D tensor.
+ *       -# DEPTH_GEMM3D: The depth of the output in case it has to be reinterpreted as a 3D tensor
+ *          (HEIGHT_GEMM3D * DEPTH_GEMM3D) = columns LHS matrix
+ *
+ * @note The offset, scalar scale factor and number of bits to shift right of output tensor must be passed at compile time using -DRESULT_OFFSET, -RESULT_MULTIPLIER and -DRESULT_SHIFT
+ * @note In case the addition of int32 biases is required, -DADD_BIAS should be passed at compile time
+ * @note The output datatype should be passed at compile time using -DOUTPUT_DATA_TYPE
+ * @note In case the clamping of the result is required, the min and max bounds can be passed at compile time using -DMIN_BOUND and -DMAX_BOUND.
+ *       These values can be used to implement "rectified linear unit" activation functions
+ * @note In case of per-channel quantization of matrix B, -DPER_CHANNEL_QUANTIZATION must be passed at compile time.
+ *
+ * @param[in]  lhs_ptr                                          Pointer to the LHS reshaped matrix. Supported data type: QASYMM8/QASYMM8_SIGNED
+ * @param[in]  lhs_stride_x                                     Stride of the LHS reshaped matrix in X dimension (in bytes)
+ * @param[in]  lhs_step_x                                       src_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  lhs_stride_y                                     Stride of the LHS reshaped matrix in Y dimension (in bytes)
+ * @param[in]  lhs_step_y                                       src_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  lhs_offset_first_element_in_bytes                The offset of the first element in the LHS reshaped matrix
+ * @param[in]  rhs_ptr                                          Pointer to the RHS reshaped matrix. Supported data type: same as @p lhs_ptr
+ * @param[in]  rhs_stride_x                                     Stride of the RHS reshaped matrix in X dimension (in bytes)
+ * @param[in]  rhs_step_x                                       src_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  rhs_stride_y                                     Stride of the RHS reshaped matrix in Y dimension (in bytes)
+ * @param[in]  rhs_step_y                                       src_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  rhs_offset_first_element_in_bytes                The offset of the first element in the RHS reshaped matrix
+ * @param[out] dst_ptr                                          Pointer to the destination matrix Supported data type: same as @p lhs_ptr
+ * @param[in]  dst_stride_x                                     Stride of the destination matrix in X dimension (in bytes)
+ * @param[in]  dst_step_x                                       dst_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  dst_stride_y                                     Stride of the destination matrix in Y dimension (in bytes)
+ * @param[in]  dst_step_y                                       dst_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  dst_offset_first_element_in_bytes                The offset of the first element in the destination matrix
+ * @param[in]  lhs_stride_z                                     Stride of the LHS reshaped matrix in Z dimension (in bytes)
+ * @param[in]  rhs_stride_z                                     Stride of the RHS reshaped matrix in Z dimension (in bytes)
+ * @param[in]  dst_stride_z                                     Stride of the destination tensor in Z dimension (in bytes)
+ * @param[in]  lhs_cross_plane_pad                              (Optional) Bottom paddings for LHS matrix in unit of elements (only if defined REINTERPRET_INPUT_AS_3D)
+ * @param[in]  dst_cross_plane_pad                              (Optional) Bottom paddings for the output matrix in unit of elements (only if defined REINTERPRET_OUTPUT_AS_3D)
+ * @param[in]  sum_col_ptr                                      (Optional) Pointer to the source tensor. Supported data type: S32
+ * @param[in]  sum_col_stride_x                                 (Optional) Stride of the source tensor in X dimension (in bytes)
+ * @param[in]  sum_col_step_x                                   (Optional) sum_col_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  sum_col_stride_y                                 (Optional) Stride of the source tensor in Y dimension (in bytes)
+ * @param[in]  sum_col_step_y                                   (Optional) sum_col_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  sum_col_offset_first_element_in_bytes            (Optional) The offset of the first element in the source tensor
+ * @param[in]  sum_row_ptr                                      (Optional) Pointer to the source tensor. Supported data type: S32
+ * @param[in]  sum_row_stride_x                                 (Optional) Stride of the source tensor in X dimension (in bytes)
+ * @param[in]  sum_row_step_x                                   (Optional) sum_row_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  sum_row_stride_y                                 (Optional) Stride of the source tensor in Y dimension (in bytes)
+ * @param[in]  sum_row_step_y                                   (Optional) sum_row_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  sum_row_offset_first_element_in_bytes            (Optional) The offset of the first element in the source tensor
+ * @param[in]  biases_ptr                                       (Optional) Pointer to the biases tensor. Supported data type: S32
+ * @param[in]  biases_stride_x                                  (Optional) Stride of the biases tensor in X dimension (in bytes)
+ * @param[in]  biases_step_x                                    (Optional) biases_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  biases_offset_first_element_in_bytes             (Optional) The offset of the first element in the biases tensor
+ * @param[in]  result_multipliers_ptr                           (Optional) Pointer to the output multipliers vector for per-channel quantization. Supported data types: S32
+ * @param[in]  result_multipliers_stride_x                      (Optional) Stride of the output multipliers vector in X dimension (in bytes)
+ * @param[in]  result_multipliers_step_x                        (Optional) output_multipliers_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  result_multipliers_offset_first_element_in_bytes (Optional) The offset of the first element in the output multipliers vector
+ * @param[in]  result_shifts_ptr                                (Optional) Pointer to the output shifts vector for per-channel quantization. Supported data types: S32
+ * @param[in]  result_shifts_stride_x                           (Optional) Stride of the output shifts vector in X dimension (in bytes)
+ * @param[in]  result_shifts_step_x                             (Optional) output_shifts_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  result_shifts_offset_first_element_in_bytes      (Optional) The offset of the first element in the output shifts vector
+ */
+__kernel void gemmlowp_mm_reshaped_only_rhs_t_fused_output_stage_fixedpoint(IMAGE_DECLARATION(lhs),
+                                                                            IMAGE_DECLARATION(rhs),
+                                                                            IMAGE_DECLARATION(dst),
+                                                                            uint lhs_stride_z,
+                                                                            uint rhs_stride_z,
+                                                                            uint dst_stride_z
+#if defined(REINTERPRET_INPUT_AS_3D)
+                                                                            ,
+                                                                            uint lhs_cross_plane_pad
+#endif // REINTERPRET_INPUT_AS_3D
+#if defined(REINTERPRET_OUTPUT_AS_3D)
+                                                                            ,
+                                                                            uint dst_cross_plane_pad
+#endif // REINTERPRET_OUTPUT_AS_3D
+#if defined(A_OFFSET)
+                                                                            ,
+                                                                            IMAGE_DECLARATION(sum_col)
+#endif // defined(A_OFFSET)
+#if defined(B_OFFSET)
+                                                                            ,
+                                                                            IMAGE_DECLARATION(sum_row)
+#endif // defined(B_OFFSET)
+#if defined(ADD_BIAS)
+                                                                            ,
+                                                                            VECTOR_DECLARATION(biases)
+#endif // defined(ADD_BIAS)
+#if defined(PER_CHANNEL_QUANTIZATION)
+                                                                            ,
+                                                                            VECTOR_DECLARATION(result_multipliers),
+                                                                            VECTOR_DECLARATION(result_shifts)
+#endif // defined(PER_CHANNEL_QUANTIZATION)
+                                                                           )
+{
+    // Block size
+#define RHS_BLOCK_SIZE ((K0) * (N0))
+
+    // RHS offset and step X
+#if defined(RHS_INTERLEAVE)
+#define RHS_OFFSET_X (K0)
+#define RHS_STEP_X ((K0) * (H0))
+#define RHS_STEP_LOOP (1)
+#else // defined(RHS_INTERLEAVE)
+#define RHS_OFFSET_X (RHS_BLOCK_SIZE)
+#define RHS_STEP_X (K0)
+#define RHS_STEP_LOOP (H0)
+#endif // defined(RHS_INTERLEAVE)
+
+    uint x = get_global_id(0);
+    uint y = get_global_id(1);
+    uint z = get_global_id(2);
+
+#if defined(DUMMY_WORK_ITEMS)
+    if((x * N0 >= N) || (y * M0 >= M))
+    {
+        return;
+    }
+#endif // defined(DUMMY_WORK_ITEMS)
+
+    // Compute LHS matrix address
+    uint lhs_offset = lhs_offset_first_element_in_bytes + y * M0 * (uint)lhs_stride_y;
+
+    // Compute RHS matrix address
+    uint rhs_offset = rhs_offset_first_element_in_bytes + (x % H0) * (uint)RHS_OFFSET_X + (x / (uint)H0) * rhs_stride_y;
+
+#if defined(MATRIX_B_DEPTH)
+    // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3
+    rhs_offset += (z % MATRIX_B_DEPTH) * rhs_stride_z;
+#else  // defined(MATRIX_B_DEPTH)
+    rhs_offset += z * rhs_stride_z;
+#endif // defined(MATRIX_B_DEPTH)
+
+    REPEAT_VAR_INIT_TO_CONST(8, uint, zlhs, 0); //uint zout0=0,zout1=0,zout2=0,... zout7=0;
+    REPEAT_VAR_INIT_TO_CONST(16, uint, zrhs, 0);
+
+#if defined(REINTERPRET_INPUT_AS_3D)
+    // The plane (zlhs) is calculated dividing M (y * M0) by HEIGHT_GEMM3D
+    CALCULATE_Z_OFFSET(M0, uint, zlhs, y, HEIGHT_GEMM3D, DEPTH_GEMM3D, lhs_cross_plane_pad, lhs_stride_y);
+
+    // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
+    // multiply lhs_stride_z by DEPTH_GEMM3D
+    lhs_offset += z * lhs_stride_z * DEPTH_GEMM3D;
+
+#else // defined(REINTERPRET_INPUT_AS_3D)
+
+    // Add offset for batched GEMM
+    lhs_offset += z * lhs_stride_z;
+
+#endif // defined(REINTERPRET_INPUT_AS_3D)
+
+    // Initialize the accumulators
+    REPEAT_VAR_INIT_TO_CONST(M0, VEC_DATA_TYPE(ACC_DATA_TYPE, N0), c, 0); //VEC_DATA_TYPE(ACC_DATA_TYPE, N0)    c0=0,c1=0,c2=0,... c(N0-1)=0;
+
+    for(int i = 0; i < K; i += K0)
+    {
+        // Load values from LHS matrix
+        LOAD_BLOCK(M0, K0, DATA_TYPE, a, lhs_ptr, lhs_offset, lhs_stride_y, zlhs);
+
+        // Load values from RHS matrix
+        LOAD_BLOCK(N0, K0, DATA_TYPE, b, rhs_ptr, rhs_offset, RHS_STEP_X, zrhs);
+
+        // Partial matrix multiplication M0,N0,K0
+        ARM_MM_K0XN0XM0(M0, N0, K0, a, b, c);
+
+        lhs_offset += K0;
+        rhs_offset += N0 * RHS_STEP_X * RHS_STEP_LOOP;
+    }
+
+    // Result of MM is of type DATA_TYPE
+    __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (x * (uint)N0) * sizeof(DATA_TYPE) + (y * (uint)M0 * dst_stride_y);
+
+    REPEAT_VAR_INIT_TO_CONST(8, uint, zout, 0); //uint zout0=0,zout1=0,zout2=0,... zout7=0;
+
+#if defined(REINTERPRET_OUTPUT_AS_3D)
+    // The plane (zout) is calculated dividing M (y * M0) by HEIGHT_GEMM3D
+    CALCULATE_Z_OFFSET(M0, uint, zout, y, HEIGHT_GEMM3D, DEPTH_GEMM3D, dst_cross_plane_pad, dst_stride_y);
+
+    // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
+    // multiply dst_stride_z by DEPTH_GEMM3D
+    dst_addr += z * dst_stride_z * DEPTH_GEMM3D;
+
+#else // defined(REINTERPRET_OUTPUT_AS_3D)
+
+    // Add offset for batched GEMM
+    dst_addr += z * dst_stride_z;
+
+#endif // defined(REINTERPRET_OUTPUT_AS_3D)
+
+    // Convert result of matrix multiplication to S32
+    REPEAT_VAR_INIT_CONVERT_SAT(M0, VEC_DATA_TYPE(int, N0), c, c_int);
+
+    int batch_id = z;
+#if defined(DEPTH_GEMM3D)
+    batch_id /= (int)DEPTH_GEMM3D;
+#endif // defined(DEPTH_GEMM3D)
+
+    // Offset contribution: c += (A_OFFSET * sum_col) + (B_OFFSET * sum_row) +  K_OFFSET;
+    REPEAT_VAR_INIT_TO_CONST(M0, VEC_DATA_TYPE(int, N0), offset_s32_, K_OFFSET);
+
+#if defined(A_OFFSET)
+    // Compute the offset contribution due to A_OFFSET
+    __global uchar *sum_col_addr = sum_col_ptr + sum_col_offset_first_element_in_bytes + (x * (uint)N0) * sizeof(int);
+
+#if defined(SUM_COL_HAS_BATCHES)
+    sum_col_addr += z * sum_col_stride_y;
+#endif // defined(SUM_COL_HAS_BATCHES)
+    VEC_DATA_TYPE(int, N0)
+    a_offset_s32 = VLOAD(N0)(0, (__global int *)sum_col_addr);
+    a_offset_s32 *= (VEC_DATA_TYPE(int, N0))A_OFFSET;
+
+    REPEAT_ADD_VECTOR_TO_VAR(M0, offset_s32_, a_offset_s32);
+#endif // defined(A_OFFSET)
+
+#if defined(B_OFFSET)
+    // Compute the offset contribution due to B_OFFSET
+    __global uchar *sum_row_addr = sum_row_ptr + sum_row_offset_first_element_in_bytes + (y * (uint)M0) * sizeof(int) + z * sum_row_stride_y;
+
+#if defined(HEIGHT_GEMM3D) && defined(DEPTH_GEMM3D)
+    sum_row_addr += (batch_id % (int)DEPTH_GEMM3D) * (int)HEIGHT_GEMM3D * sizeof(int);
+#endif // defined(HEIGHT_GEMM3D) && defined(DEPTH_GEMM3D)
+    LOAD_SCALAR_AS_VECTOR(M0, N0, int, b_offset_s32_, sum_row_addr, 0, sum_row_stride_x);
+
+    REPEAT_MLA_VAR_WITH_CONST_VEC(M0, offset_s32_, b_offset_s32_, (VEC_DATA_TYPE(int, N0))B_OFFSET);
+#endif // defined(B_OFFSET)
+
+#if defined(ADD_BIAS)
+    // Add bias
+    __global uchar *bias_addr = biases_ptr + biases_offset_first_element_in_bytes + (x * (uint)N0) * sizeof(int);
+
+    VEC_DATA_TYPE(int, N0)
+    bias_values = VLOAD(N0)(0, (__global int *)bias_addr);
+    REPEAT_ADD_VECTOR_TO_VAR(M0, offset_s32_, bias_values);
+#endif // defined(ADD_BIAS)
+
+    REPEAT_ADD_TWO_VARS(M0, c_int, offset_s32_);
+
+    // Multiply by result_mult_int and shift
+#if defined(PER_CHANNEL_QUANTIZATION)
+    __global uchar *result_multipliers_addr = result_multipliers_ptr + result_multipliers_offset_first_element_in_bytes + (x * (uint)N0) * sizeof(int);
+    __global uchar *result_shifts_addr      = result_shifts_ptr + result_shifts_offset_first_element_in_bytes + (x * (uint)N0) * sizeof(int);
+
+    VEC_DATA_TYPE(int, N0)
+    res_mul = VLOAD(N0)(0, (__global int *)result_multipliers_addr);
+    VEC_DATA_TYPE(int, N0)
+    res_shift = VLOAD(N0)(0, (__global int *)result_shifts_addr);
+
+    REPEAT_ASYMM_MULT_BY_QUANT_MULTIPLIER_PER_CHANNEL(M0, N0, c_int, res_mul, res_shift);
+#else // defined(PER_CHANNEL_QUANTIZATION)
+
+#if RESULT_SHIFT < 0
+    REPEAT_ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(M0, N0, c_int, RESULT_MULTIPLIER, RESULT_SHIFT);
+#else  // RESULT_SHIFT >= 0
+    REPEAT_ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE(M0, N0, c_int, RESULT_MULTIPLIER, RESULT_SHIFT);
+#endif // RESULT_SHIFT < 0
+
+#endif // defined(PER_CHANNEL_QUANTIZATION)
+
+    // Add the offset terms to GEMM's result
+    REPEAT_ADD_CONST_TO_VAR(M0, VEC_DATA_TYPE(int, N0), c_int, RESULT_OFFSET);
+
+#if defined(MIN_BOUND)
+    REPEAT_MAX_CONST_VAR(M0, VEC_DATA_TYPE(int, N0), c_int, MIN_BOUND);
+#endif // defined(MIN_BOUND)
+#if defined(MAX_BOUND)
+    REPEAT_MIN_CONST_VAR(M0, VEC_DATA_TYPE(int, N0), c_int, MAX_BOUND);
+#endif // defined(MAX_BOUND)
+
+    // Convert and store output block (does convert saturate)
+    CONVERT_STORE_BLOCK(M0, N0, DATA_TYPE, c_int, dst_addr, dst_stride_y, zout);
+
+#undef RHS_BLOCK_SIZE
+#undef RHS_OFFSET_X
+#undef RHS_STEP_X
+}
+#endif // defined(RESULT_OFFSET) && defined(RESULT_SHIFT) && defined(RESULT_MULTIPLIER)
 #endif // defined(M0) && defined(N0) && defined(K0) && defined(H0) && defined(DATA_TYPE) && defined(K)
 
 #if defined(M0) && defined(N0) && defined(K0) && defined(K)
@@ -840,7 +1135,7 @@ __kernel void gemmlowp_mm_reshaped_only_rhs_t(IMAGE_DECLARATION(lhs),
  *       -# DEPTH_GEMM3D: The depth of the output in case it has to be reinterpreted as a 3D tensor
  *          (HEIGHT_GEMM3D * DEPTH_GEMM3D) = columns LHS matrix
  *
- * @param[in]  lhs_ptr                           Pointer to the LHS reshaped matrix. Supported data type: F16/F32
+ * @param[in]  lhs_ptr                           Pointer to the LHS reshaped matrix. Supported data type: QASYMM8
  * @param[in]  lhs_stride_x                      Stride of the LHS reshaped matrix in X dimension (in bytes)
  * @param[in]  lhs_step_x                        src_stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in]  lhs_stride_y                      Stride of the LHS reshaped matrix in Y dimension (in bytes)
@@ -852,7 +1147,7 @@ __kernel void gemmlowp_mm_reshaped_only_rhs_t(IMAGE_DECLARATION(lhs),
  * @param[in]  rhs_stride_y                      Stride of the RHS reshaped matrix in Y dimension (in bytes)
  * @param[in]  rhs_step_y                        src_stride_y * number of elements along Y processed per workitem(in bytes)
  * @param[in]  rhs_offset_first_element_in_bytes The offset of the first element in the RHS reshaped matrix
- * @param[out] dst_ptr                           Pointer to the destination matrix Supported data type: same as @p lhs_ptr
+ * @param[out] dst_ptr                           Pointer to the destination matrix Supported data type: S32
  * @param[in]  dst_stride_x                      Stride of the destination matrix in X dimension (in bytes)
  * @param[in]  dst_step_x                        dst_stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in]  dst_stride_y                      Stride of the destination matrix in Y dimension (in bytes)
@@ -1389,7 +1684,7 @@ __kernel void gemmlowp_offset_contribution(TENSOR3D_DECLARATION(mm_result)
     vstore4(in_s32, 0, (__global int *)mm_result_addr);
 }
 
-#if defined(RESULT_OFFSET) && defined(RESULT_MULTIPLIER) && defined(RESULT_SHIFT)
+#if defined(RESULT_OFFSET) && defined(RESULT_MULTIPLIER) && defined(RESULT_SHIFT) && defined(OUTPUT_DATA_TYPE)
 /* OpenCL kernel used to add the offset contribution after @ref CLGEMMLowpMatrixMultiplyKernel and it quantizes down to uint8.
  *
  * This kernel takes a final int32 accumulator value (the output of @CLGEMMLowpMatrixMultiplyKernel), adds to it the offset contribution of matrix A and matrix B and quantizes to uint8 through the output stage.
@@ -1742,7 +2037,7 @@ __kernel void gemmlowp_offset_contribution_quantize_down_fixedpoint(TENSOR3D_DEC
     // Store the result
     vstore4(res, 0, (__global OUTPUT_DATA_TYPE *)dst_addr);
 }
-#endif // defined(RESULT_OFFSET) && defined(RESULT_MULTIPLIER) && defined(RESULT_SHIFT)
+#endif // defined(RESULT_OFFSET) && defined(RESULT_MULTIPLIER) && defined(RESULT_SHIFT) && defined(OUTPUT_DATA_TYPE)
 
 #endif // defined(K_OFFSET)
 
diff --git a/src/core/CL/cl_kernels/repeat.h b/src/core/CL/cl_kernels/repeat.h
index 691f7aea01..4e761db073 100644
--- a/src/core/CL/cl_kernels/repeat.h
+++ b/src/core/CL/cl_kernels/repeat.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2019 ARM Limited.
+ * Copyright (c) 2019-2020 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -24,6 +24,8 @@
 #ifndef ARM_COMPUTE_REPEAT_H
 #define ARM_COMPUTE_REPEAT_H
 
+#include "helpers.h"
+
 /** Macros that help in loop unrolling */
 //Repeat macros with 3 param, excluding the implicit ID param
 #define REPEAT_3_1(P_X, P_A, P_B, P_C) P_X##_DEF(0, P_A, P_B, P_C)
@@ -76,8 +78,106 @@
 #define REPEAT_DEF_3_N(P_NUM, P_OP, P_A, P_B, P_C) REPEAT_3_##P_NUM(P_OP, P_A, P_B, P_C) //One level of indirection to ensure order of expansion does not affect preprocessing P_NUM
 #define REPEAT_3_N(P_NUM, P_OP, P_A, P_B, P_C) REPEAT_DEF_3_N(P_NUM, P_OP, P_A, P_B, P_C)
 
-//Macro for initializing N variables. generates N statements that defines VAR##N = RHS_ACCESSOR_DEF(...)
+// Repeat macros with 4 param, excluding the implicit ID param
+#define REPEAT_4_1(P_X, P_A, P_B, P_C, P_D) P_X##_DEF(0, P_A, P_B, P_C, P_D)
+#define REPEAT_4_2(P_X, P_A, P_B, P_C, P_D) \
+    P_X##_DEF(1, P_A, P_B, P_C, P_D);       \
+    REPEAT_4_1(P_X, P_A, P_B, P_C, P_D)
+#define REPEAT_4_3(P_X, P_A, P_B, P_C, P_D) \
+    P_X##_DEF(2, P_A, P_B, P_C, P_D);       \
+    REPEAT_4_2(P_X, P_A, P_B, P_C, P_D)
+#define REPEAT_4_4(P_X, P_A, P_B, P_C, P_D) \
+    P_X##_DEF(3, P_A, P_B, P_C, P_D);       \
+    REPEAT_4_3(P_X, P_A, P_B, P_C, P_D)
+#define REPEAT_4_5(P_X, P_A, P_B, P_C, P_D) \
+    P_X##_DEF(4, P_A, P_B, P_C, P_D);       \
+    REPEAT_4_4(P_X, P_A, P_B, P_C, P_D)
+#define REPEAT_4_6(P_X, P_A, P_B, P_C, P_D) \
+    P_X##_DEF(5, P_A, P_B, P_C, P_D);       \
+    REPEAT_4_5(P_X, P_A, P_B, P_C, P_D)
+#define REPEAT_4_7(P_X, P_A, P_B, P_C, P_D) \
+    P_X##_DEF(6, P_A, P_B, P_C, P_D);       \
+    REPEAT_4_6(P_X, P_A, P_B, P_C, P_D)
+#define REPEAT_4_8(P_X, P_A, P_B, P_C, P_D) \
+    P_X##_DEF(7, P_A, P_B, P_C, P_D);       \
+    REPEAT_4_7(P_X, P_A, P_B, P_C, P_D)
+#define REPEAT_4_9(P_X, P_A, P_B, P_C, P_D) \
+    P_X##_DEF(8, P_A, P_B, P_C, P_D);       \
+    REPEAT_4_8(P_X, P_A, P_B, P_C, P_D)
+#define REPEAT_4_10(P_X, P_A, P_B, P_C, P_D) \
+    P_X##_DEF(9, P_A, P_B, P_C, P_D);        \
+    REPEAT_4_9(P_X, P_A, P_B, P_C, P_D)
+#define REPEAT_4_11(P_X, P_A, P_B, P_C, P_D) \
+    P_X##_DEF(A, P_A, P_B, P_C, P_D);        \
+    REPEAT_4_10(P_X, P_A, P_B, P_C, P_D)
+#define REPEAT_4_12(P_X, P_A, P_B, P_C, P_D) \
+    P_X##_DEF(B, P_A, P_B, P_C, P_D);        \
+    REPEAT_4_11(P_X, P_A, P_B, P_C, P_D)
+#define REPEAT_4_13(P_X, P_A, P_B, P_C, P_D) \
+    P_X##_DEF(C, P_A, P_B, P_C, P_D);        \
+    REPEAT_4_12(P_X, P_A, P_B, P_C, P_D)
+#define REPEAT_4_14(P_X, P_A, P_B, P_C, P_D) \
+    P_X##_DEF(D, P_A, P_B, P_C, P_D);        \
+    REPEAT_4_13(P_X, P_A, P_B, P_C, P_D)
+#define REPEAT_4_15(P_X, P_A, P_B, P_C, P_D) \
+    P_X##_DEF(E, P_A, P_B, P_C, P_D);        \
+    REPEAT_4_14(P_X, P_A, P_B, P_C, P_D)
+#define REPEAT_4_16(P_X, P_A, P_B, P_C, P_D) \
+    P_X##_DEF(F, P_A, P_B, P_C, P_D);        \
+    REPEAT_4_15(P_X, P_A, P_B, P_C, P_D)
+
+#define REPEAT_DEF_4_N(P_NUM, P_OP, P_A, P_B, P_C, P_D) REPEAT_4_##P_NUM(P_OP, P_A, P_B, P_C, P_D) //One level of indirection to ensure order of expansion does not affect preprocessing P_NUM
+#define REPEAT_4_N(P_NUM, P_OP, P_A, P_B, P_C, P_D) REPEAT_DEF_4_N(P_NUM, P_OP, P_A, P_B, P_C, P_D)
+
+// Macro for initializing N variables. Generates N statements that defines VAR##N = RHS_ACCESSOR_DEF(...)
 #define VAR_INIT_TO_CONST_DEF(ID, TYPE, VAR, VAL) TYPE VAR##ID = VAL
 #define REPEAT_VAR_INIT_TO_CONST(N, TYPE, VAR, VAL) REPEAT_3_N(N, VAR_INIT_TO_CONST, TYPE, VAR, VAL)
 
+// Macro for initializing N variables by converting the data type. Generates N statements that defines VAR##N = RHS_ACCESSOR_DEF(...)
+#define VAR_INIT_CONVERT_SAT_DEF(ID, TYPE_OUT, VAR_IN, VAR_OUT) TYPE_OUT VAR_OUT##ID = CONVERT_SAT(VAR_IN##ID, TYPE_OUT)
+#define REPEAT_VAR_INIT_CONVERT_SAT(N, TYPE_OUT, VAR_IN, VAR_OUT) REPEAT_3_N(N, VAR_INIT_CONVERT_SAT, TYPE_OUT, VAR_IN, VAR_OUT)
+
+// Macro for adding a constant to N variables. Generates N statements that defines VAR##N =RHS_ACCESSOR_DEF(...)
+#define ADD_CONST_TO_VAR_DEF(ID, TYPE, VAR, VAL) VAR##ID += (TYPE)VAL
+#define REPEAT_ADD_CONST_TO_VAR(N, TYPE, VAR, VAL) REPEAT_3_N(N, ADD_CONST_TO_VAR, TYPE, VAR, VAL)
+
+// Macro for multiplying N variables (VAR_B) by a constant (VAL) and adding to other N variables (VAR_A). Generates N statements that defines VAR_A##N =RHS_ACCESSOR_DEF(...)
+#define MLA_VAR_WITH_CONST_VEC_DEF(ID, VAR_A, VAR_B, VAL) VAR_A##ID += VAR_B##ID * VAL
+#define REPEAT_MLA_VAR_WITH_CONST_VEC(N, VAR_A, VAR_B, VAL) REPEAT_3_N(N, MLA_VAR_WITH_CONST_VEC, VAR_A, VAR_B, VAL)
+
+// Macro for adding a vector to N-variables. Generates N statements that defines VAR##N =RHS_ACCESSOR_DEF(...)
+#define ADD_VECTOR_TO_VAR_DEF(ID, TYPE, VAR, VEC) VAR##ID += VEC
+#define REPEAT_ADD_VECTOR_TO_VAR(N, VAR, VEC) REPEAT_3_N(N, ADD_VECTOR_TO_VAR, "", VAR, VEC)
+
+// Macro for adding a two N-variables. Generates N statements that defines VAR##N =RHS_ACCESSOR_DEF(...)
+#define ADD_TWO_VARS_DEF(ID, TYPE, VAR_A, VAR_B) VAR_A##ID += VAR_B##ID
+#define REPEAT_ADD_TWO_VARS(N, VAR_A, VAR_B) REPEAT_3_N(N, ADD_TWO_VARS, "", VAR_A, VAR_B)
+
+// Macro for performing Max between a constant and N variables. Generates N statements that defines VAR##N =RHS_ACCESSOR_DEF(...)
+#define MAX_CONST_VAR_DEF(ID, TYPE, VAR, VAL) VAR##ID = max(VAR##ID, (TYPE)VAL)
+#define REPEAT_MAX_CONST_VAR(N, TYPE, VAR, VAL) REPEAT_3_N(N, MAX_CONST_VAR, TYPE, VAR, VAL)
+
+// Macro for performing Min between a constant and N variables. Generates N statements that defines VAR##N =RHS_ACCESSOR_DEF(...)
+#define MIN_CONST_VAR_DEF(ID, TYPE, VAR, VAL) VAR##ID = min(VAR##ID, (TYPE)VAL)
+#define REPEAT_MIN_CONST_VAR(N, TYPE, VAR, VAL) REPEAT_3_N(N, MIN_CONST_VAR, TYPE, VAR, VAL)
+
+// Macro for performing ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE to N variables. Generates N statements that defines VAR##N =RHS_ACCESSOR_DEF(...)
+#define ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE_DEF(ID, SIZE, VAR, RES_MUL, RES_SHIFT) VAR##ID = ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(VAR##ID, RES_MUL, RES_SHIFT, SIZE)
+#define REPEAT_ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(N, SIZE, VAR, RES_MUL, RES_SHIFT) REPEAT_4_N(N, ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE, SIZE, VAR, RES_MUL, RES_SHIFT)
+
+// Macro for performing ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE to N variables. Generates N statements that defines VAR##N =RHS_ACCESSOR_DEF(...)
+#define ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE_DEF(ID, SIZE, VAR, RES_MUL, RES_SHIFT) VAR##ID = ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE(VAR##ID, RES_MUL, RES_SHIFT, SIZE)
+#define REPEAT_ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE(N, SIZE, VAR, RES_MUL, RES_SHIFT) REPEAT_4_N(N, ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE, SIZE, VAR, RES_MUL, RES_SHIFT)
+
+// Macro for performing per-channel ASYMM_MULT_BY_QUANT_MULTIPLIER to N variables.
+#define ASYMM_MULT_BY_QUANT_MULTIPLIER_PER_CHANNEL_DEF(ID, SIZE, VAR, RES_MUL, RES_SHIFT)                     \
+    ({                                                                                                        \
+        VEC_DATA_TYPE(int, N0)                                                                                \
+        VAR##ID_shift_lt0 = ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(VAR##ID, RES_MUL, RES_SHIFT, N0); \
+        VEC_DATA_TYPE(int, N0)                                                                                \
+        VAR##ID_shift_gt0 = ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE(VAR##ID, RES_MUL, RES_SHIFT, N0);    \
+        VAR##ID           = select(VAR##ID_shift_lt0, VAR##ID_shift_gt0, RES_SHIFT >= 0);                     \
+    })
+#define REPEAT_ASYMM_MULT_BY_QUANT_MULTIPLIER_PER_CHANNEL(N, SIZE, VAR, RES_MUL, RES_SHIFT) REPEAT_4_N(N, ASYMM_MULT_BY_QUANT_MULTIPLIER_PER_CHANNEL, SIZE, VAR, RES_MUL, RES_SHIFT)
+
 #endif // ARM_COMPUTE_REPEAT_H
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));
diff --git a/src/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.cpp b/src/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.cpp
index cdb78c291d..54b63df924 100644
--- a/src/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.cpp
+++ b/src/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.cpp
@@ -75,8 +75,9 @@ CLGEMMLowpMatrixMultiplyCore::CLGEMMLowpMatrixMultiplyCore(std::shared_ptr<IMemo
       _is_midgard(false),
       _reshape_b_only_on_first_run(false),
       _is_prepared(false),
-      _fuse_output_stage(false),
-      _convert_to_qasymm8(false)
+      _run_output_stage(false),
+      _convert_to_qasymm8(false),
+      _run_offset_contribution(false)
 {
 }
 
@@ -172,44 +173,67 @@ void CLGEMMLowpMatrixMultiplyCore::configure(const ICLTensor *a, const ICLTensor
         _mtx_a_reduction_kernel.configure(a, &_vector_sum_row);
     }
 
+    GEMMKernelInfo gemm_kernel_info;
+    gemm_kernel_info.m                       = m;
+    gemm_kernel_info.n                       = n;
+    gemm_kernel_info.k                       = k;
+    gemm_kernel_info.depth_output_gemm3d     = depth_output_gemm3d;
+    gemm_kernel_info.reinterpret_input_as_3d = reinterpret_input_as_3d;
+    gemm_kernel_info.lhs_info                = lhs_info;
+    gemm_kernel_info.rhs_info                = rhs_info;
+    gemm_kernel_info.a_offset                = _a_offset;
+    gemm_kernel_info.b_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;
+        // Configure offset contribution kernel
+        const size_t num_filters = (gemm_info.gemmlowp_output_stage().is_quantized_per_channel) ? gemm_info.gemmlowp_output_stage().gemmlowp_multipliers.size() : 1;
 
-        _memory_group.manage(&_mm_result_s32);
+        _gemm_output_stage_multipliers.allocator()->init(TensorInfo(TensorShape(num_filters), 1, DataType::S32));
+        _gemm_output_stage_shifts.allocator()->init(TensorInfo(TensorShape(num_filters), 1, DataType::S32));
 
-        if(_is_gemm_reshaped)
+        GEMMLowpOutputStageInfo gemmlowp_output_stage = gemm_info.gemmlowp_output_stage();
+        gemmlowp_output_stage.output_data_type        = _matrix_a->info()->data_type();
+
+        gemm_kernel_info.output_stage = gemmlowp_output_stage;
+
+        if(_is_gemm_reshaped && gemmlowp_output_stage.type == GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT)
         {
-            // Configure and tune matrix multiply kernel
-            _mm_reshaped_only_rhs_kernel.configure(_matrix_a, matrix_b, &_mm_result_s32, lhs_info, rhs_info, GEMMReshapeInfo(m, n, k, 1, 1, depth_output_gemm3d, reinterpret_input_as_3d));
+            // Configure and tune matrix multiply kernel with fused output stage
+            _mm_reshaped_only_rhs_kernel.configure(_matrix_a, matrix_b, output, gemm_kernel_info, _a_offset == 0 ? nullptr : &_vector_sum_col,
+                                                   _b_offset == 0 ? nullptr : &_vector_sum_row, c, &_gemm_output_stage_multipliers, &_gemm_output_stage_shifts);
         }
         else
         {
-            if(_is_midgard)
+            _run_output_stage = true;
+
+            _memory_group.manage(&_mm_result_s32);
+
+            if(_is_gemm_reshaped)
             {
-                // Configure matrix multiply kernel
-                _mm_midgard_kernel.configure(_matrix_a, matrix_b, &_mm_result_s32, GEMMReshapeInfo(m, n, k, 1, 1, depth_output_gemm3d, reinterpret_input_as_3d));
+                _mm_reshaped_only_rhs_kernel.configure(_matrix_a, matrix_b, &_mm_result_s32, gemm_kernel_info);
             }
             else
             {
-                // Pick up the GEMM configuration
-                std::tie(lhs_info, rhs_info) = CLGEMMNativeKernelConfigurationFactory::create(gpu_target)->configure(m, n, k, batch_size, DataType::QASYMM8);
-
-                // Configure matrix multiply kernel
-                _mm_native_kernel.configure(_matrix_a, matrix_b, &_mm_result_s32, lhs_info, rhs_info, GEMMReshapeInfo(m, n, k, 1, 1, depth_output_gemm3d, reinterpret_input_as_3d));
+                if(_is_midgard)
+                {
+                    // Configure matrix multiply kernel
+                    _mm_midgard_kernel.configure(_matrix_a, matrix_b, &_mm_result_s32, GEMMReshapeInfo(m, n, k, 1, 1, depth_output_gemm3d, reinterpret_input_as_3d));
+                }
+                else
+                {
+                    // Pick up the GEMM configuration
+                    std::tie(lhs_info, rhs_info) = CLGEMMNativeKernelConfigurationFactory::create(gpu_target)->configure(m, n, k, batch_size, DataType::QASYMM8);
+
+                    // Configure matrix multiply kernel
+                    _mm_native_kernel.configure(_matrix_a, matrix_b, &_mm_result_s32, lhs_info, rhs_info, GEMMReshapeInfo(m, n, k, 1, 1, depth_output_gemm3d, reinterpret_input_as_3d));
+                }
+                _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, gemmlowp_output_stage, &_gemm_output_stage_multipliers, &_gemm_output_stage_shifts);
+
+                _mm_result_s32.allocator()->allocate();
             }
         }
-        // Configure offset contribution kernel
-        const size_t num_filters = (gemm_info.gemmlowp_output_stage().is_quantized_per_channel) ? gemm_info.gemmlowp_output_stage().gemmlowp_multipliers.size() : 1;
-
-        _gemm_output_stage_multipliers.allocator()->init(TensorInfo(TensorShape(num_filters), 1, DataType::S32));
-        _gemm_output_stage_shifts.allocator()->init(TensorInfo(TensorShape(num_filters), 1, DataType::S32));
-
-        GEMMLowpOutputStageInfo gemmlowp_output_stage = gemm_info.gemmlowp_output_stage();
-        gemmlowp_output_stage.output_data_type        = _matrix_a->info()->data_type();
-        _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, gemmlowp_output_stage, &_gemm_output_stage_multipliers, &_gemm_output_stage_shifts);
 
         _gemm_output_stage_multipliers.allocator()->allocate();
         _gemm_output_stage_shifts.allocator()->allocate();
@@ -220,15 +244,14 @@ void CLGEMMLowpMatrixMultiplyCore::configure(const ICLTensor *a, const ICLTensor
         std::memcpy(_gemm_output_stage_shifts.ptr_to_element(Coordinates(0)), gemm_info.gemmlowp_output_stage().gemmlowp_shifts.data(), num_filters * sizeof(int32_t));
         _gemm_output_stage_multipliers.unmap();
         _gemm_output_stage_shifts.unmap();
-
-        _mm_result_s32.allocator()->allocate();
     }
     else
     {
+        _run_offset_contribution = true;
         if(_is_gemm_reshaped)
         {
             // Configure and tune matrix multiply kernel
-            _mm_reshaped_only_rhs_kernel.configure(_matrix_a, matrix_b, output, lhs_info, rhs_info, GEMMReshapeInfo(m, n, k, 1, 1, depth_output_gemm3d, reinterpret_input_as_3d));
+            _mm_reshaped_only_rhs_kernel.configure(_matrix_a, matrix_b, output, gemm_kernel_info);
         }
         else
         {
@@ -350,61 +373,85 @@ Status CLGEMMLowpMatrixMultiplyCore::validate(const ITensorInfo *a, const ITenso
         ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMLowpMatrixAReductionKernel::validate(a, &info_vector_sum_row));
     }
 
+    GEMMKernelInfo gemm_kernel_info;
+    gemm_kernel_info.m                       = m;
+    gemm_kernel_info.n                       = n;
+    gemm_kernel_info.k                       = k;
+    gemm_kernel_info.depth_output_gemm3d     = depth_output_gemm3d;
+    gemm_kernel_info.reinterpret_input_as_3d = reinterpret_input_as_3d;
+    gemm_kernel_info.lhs_info                = lhs_info;
+    gemm_kernel_info.rhs_info                = rhs_info;
+    gemm_kernel_info.a_offset                = a_offset;
+    gemm_kernel_info.b_offset                = b_offset;
     if(gemm_info.gemmlowp_output_stage().type != GEMMLowpOutputStageType::NONE)
     {
-        TensorInfo mm_result_s32_info{};
+        const size_t num_filters = (gemm_info.gemmlowp_output_stage().is_quantized_per_channel) ? gemm_info.gemmlowp_output_stage().gemmlowp_multipliers.size() : 1;
 
-        if(reshape_matrix_b)
-        {
-            // 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_info)).set_data_type(DataType::S32));
+        const TensorInfo gemm_output_stage_multipliers_shifts_info(TensorInfo(TensorShape(num_filters), 1, DataType::S32));
 
-            // Validate matrix multiply
-            ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel::validate(matrix_a_info, matrix_b_info, &mm_result_s32_info, lhs_info, rhs_info, reshape_info));
+        GEMMLowpOutputStageInfo gemmlowp_output_stage = gemm_info.gemmlowp_output_stage();
+        gemmlowp_output_stage.output_data_type        = a->data_type();
+
+        gemm_kernel_info.output_stage = gemmlowp_output_stage;
+        if(reshape_matrix_b && gemm_info.gemmlowp_output_stage().type == GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT)
+        {
+            ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel::validate(matrix_a_info, matrix_b_info, output, gemm_kernel_info,
+                                                                                                a_offset == 0 ? nullptr : &info_vector_sum_col,
+                                                                                                b_offset == 0 ? nullptr : &info_vector_sum_row,
+                                                                                                c,
+                                                                                                &gemm_output_stage_multipliers_shifts_info,
+                                                                                                &gemm_output_stage_multipliers_shifts_info));
         }
         else
         {
-            // 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, false, reshape_info)).set_data_type(DataType::S32));
+            TensorInfo mm_result_s32_info{};
 
-            if(is_midgard)
+            if(reshape_matrix_b)
             {
+                // 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_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_info));
+                ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel::validate(matrix_a_info, matrix_b_info, &mm_result_s32_info, gemm_kernel_info));
             }
             else
             {
-                // Pick up the GEMM configuration
-                std::tie(lhs_info, rhs_info) = CLGEMMNativeKernelConfigurationFactory::create(gpu_target)->configure(m, n, k, batch_size, DataType::QASYMM8);
-
-                // Validate matrix multiply
-                ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMLowpMatrixMultiplyNativeKernel::validate(matrix_a_info, matrix_b_info, &mm_result_s32_info, lhs_info, rhs_info, reshape_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, false, reshape_info)).set_data_type(DataType::S32));
+
+                if(is_midgard)
+                {
+                    // Validate matrix multiply
+                    ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMLowpMatrixMultiplyKernel::validate(matrix_a_info, matrix_b_info, &mm_result_s32_info, reshape_info));
+                }
+                else
+                {
+                    // Pick up the GEMM configuration
+                    std::tie(lhs_info, rhs_info) = CLGEMMNativeKernelConfigurationFactory::create(gpu_target)->configure(m, n, k, batch_size, DataType::QASYMM8);
+
+                    // Validate matrix multiply
+                    ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMLowpMatrixMultiplyNativeKernel::validate(matrix_a_info, matrix_b_info, &mm_result_s32_info, lhs_info, rhs_info, reshape_info));
+                }
             }
-        }
-
-        // Validate offset contribution kernel
-        const size_t num_filters = (gemm_info.gemmlowp_output_stage().is_quantized_per_channel) ? gemm_info.gemmlowp_output_stage().gemmlowp_multipliers.size() : 1;
 
-        const TensorInfo gemm_output_stage_multipliers_shifts_info(TensorInfo(TensorShape(num_filters), 1, DataType::S32));
-
-        GEMMLowpOutputStageInfo gemmlowp_output_stage = gemm_info.gemmlowp_output_stage();
-        gemmlowp_output_stage.output_data_type        = a->data_type();
-        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,
-                                                                                            gemmlowp_output_stage,
-                                                                                            &gemm_output_stage_multipliers_shifts_info,
-                                                                                            &gemm_output_stage_multipliers_shifts_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,
+                                                                                                gemmlowp_output_stage,
+                                                                                                &gemm_output_stage_multipliers_shifts_info,
+                                                                                                &gemm_output_stage_multipliers_shifts_info));
+        }
     }
     else
     {
         if(reshape_matrix_b)
         {
             // Validate matrix multiply
-            ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel::validate(matrix_a_info, matrix_b_info, output, lhs_info, rhs_info, reshape_info));
+            ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel::validate(matrix_a_info, matrix_b_info, output, gemm_kernel_info));
         }
         else
         {
@@ -458,6 +505,12 @@ void CLGEMMLowpMatrixMultiplyCore::run()
         CLScheduler::get().enqueue(_mtx_b_reduction_kernel, false);
     }
 
+    // Run matrix A reduction kernel only if _b_offset is not equal to 0
+    if(_b_offset != 0)
+    {
+        CLScheduler::get().enqueue(_mtx_a_reduction_kernel, false);
+    }
+
     // Run matrix multiply
     if(_is_gemm_reshaped)
     {
@@ -474,19 +527,12 @@ void CLGEMMLowpMatrixMultiplyCore::run()
             CLScheduler::get().enqueue(_mm_native_kernel, false);
         }
     }
-
-    // Run matrix A reduction kernel only if _b_offset is not equal to 0
-    if(_b_offset != 0)
-    {
-        CLScheduler::get().enqueue(_mtx_a_reduction_kernel, false);
-    }
-
-    if(_fuse_output_stage)
+    if(_run_output_stage)
     {
         // Run offset contribution/output stage kernel
         CLScheduler::get().enqueue(_offset_contribution_output_stage_kernel, true);
     }
-    else
+    if(_run_offset_contribution)
     {
         // Run offset contribution kernel
         CLScheduler::get().enqueue(_offset_contribution_kernel, true);
diff --git a/tests/validation/CL/GEMMLowpMatrixMultiplyReshapedOnlyRHS.cpp b/tests/validation/CL/GEMMLowpMatrixMultiplyReshapedOnlyRHS.cpp
index 30c91b7091..f4083bfd95 100644
--- a/tests/validation/CL/GEMMLowpMatrixMultiplyReshapedOnlyRHS.cpp
+++ b/tests/validation/CL/GEMMLowpMatrixMultiplyReshapedOnlyRHS.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2019 ARM Limited.
+ * Copyright (c) 2019-2020 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -130,7 +130,12 @@ void validate_configuration(unsigned int m_value, unsigned int n_value, unsigned
     rhs_info.interleave = i_value_rhs;
     rhs_info.transpose  = true;
 
-    GEMMReshapeInfo gemm_info(M, N, K);
+    GEMMKernelInfo gemm_info;
+    gemm_info.m = M;
+    gemm_info.n = N;
+    gemm_info.k = K;
+    gemm_info.lhs_info = lhs_info;
+    gemm_info.rhs_info = rhs_info;
 
     const TensorShape lhs_shape(K, M, b_value);
     const TensorShape rhs_shape(N, K, b_value);
@@ -152,7 +157,7 @@ void validate_configuration(unsigned int m_value, unsigned int n_value, unsigned
 
     // Create and configure function
     CLGEMMLowpMatrixMultiplyReshapedOnlyRHS gemm;
-    gemm.configure(&lhs, &rhs_reshaped, &dst, lhs_info, rhs_info, gemm_info);
+    gemm.configure(&lhs, &rhs_reshaped, &dst, gemm_info);
 }
 } // namespace
 
diff --git a/tests/validation/fixtures/GEMMLowpFixture.h b/tests/validation/fixtures/GEMMLowpFixture.h
index d13ada9d64..0207f4c5ae 100644
--- a/tests/validation/fixtures/GEMMLowpFixture.h
+++ b/tests/validation/fixtures/GEMMLowpFixture.h
@@ -24,6 +24,7 @@
 #ifndef ARM_COMPUTE_TEST_GEMMLOWP_FIXTURE
 #define ARM_COMPUTE_TEST_GEMMLOWP_FIXTURE
 
+#include "arm_compute/core/KernelDescriptors.h"
 #include "arm_compute/core/TensorShape.h"
 #include "arm_compute/core/Types.h"
 #include "arm_compute/core/utils/quantization/AsymmHelpers.h"
@@ -1012,13 +1013,19 @@ protected:
         const unsigned int N = rhs_shape[0];
         const unsigned int K = lhs_shape[0];
 
+        GEMMKernelInfo gemm_info;
+        gemm_info.m        = M;
+        gemm_info.n        = N;
+        gemm_info.k        = K;
+        gemm_info.lhs_info = lhs_info;
+        gemm_info.rhs_info = rhs_info;
         // The output tensor will be auto-initialized within the function
 
         // Create and configure function
         ReshapeRHSFunctionType reshape_rhs;
         GEMMFunctionType       gemm;
         reshape_rhs.configure(&rhs, &rhs_reshaped, rhs_info);
-        gemm.configure(&lhs, &rhs_reshaped, &dst, lhs_info, rhs_info, GEMMReshapeInfo(M, N, K));
+        gemm.configure(&lhs, &rhs_reshaped, &dst, gemm_info);
 
         ARM_COMPUTE_EXPECT(lhs.info()->is_resizable(), framework::LogLevel::ERRORS);
         ARM_COMPUTE_EXPECT(rhs.info()->is_resizable(), framework::LogLevel::ERRORS);
@@ -1148,13 +1155,20 @@ protected:
         const unsigned int N = rhs_shape[0];
         const unsigned int K = lhs_shape[0];
 
+        GEMMKernelInfo gemm_info;
+        gemm_info.m                   = M;
+        gemm_info.n                   = N;
+        gemm_info.k                   = K;
+        gemm_info.depth_output_gemm3d = m_h;
+        gemm_info.lhs_info            = lhs_info;
+        gemm_info.rhs_info            = rhs_info;
         // The output tensor will be auto-initialized within the function
 
         // Create and configure function
         ReshapeRHSFunctionType reshape_rhs;
         GEMMFunctionType       gemm;
         reshape_rhs.configure(&rhs, &rhs_reshaped, rhs_info);
-        gemm.configure(&lhs, &rhs_reshaped, &dst, lhs_info, rhs_info, GEMMReshapeInfo(M, N, K, 1, 1, m_h));
+        gemm.configure(&lhs, &rhs_reshaped, &dst, gemm_info);
 
         ARM_COMPUTE_EXPECT(lhs.info()->is_resizable(), framework::LogLevel::ERRORS);
         ARM_COMPUTE_EXPECT(rhs.info()->is_resizable(), framework::LogLevel::ERRORS);